Environmental justice is the principle that all

The Distribution of Environmental Equity: Exploring Spatial Nonstationarity in Multivariate Models of Air Toxic Releases Jeremy L. Mennis* and Lisa Jo...
Author: Grace Whitehead
3 downloads 0 Views 6MB Size
The Distribution of Environmental Equity: Exploring Spatial Nonstationarity in Multivariate Models of Air Toxic Releases Jeremy L. Mennis* and Lisa Jordan** *Department of Geography and Urban Studies, Temple University **Department of Geography, University of Colorado at Boulder

Conventional multivariate regression can hide important local variations in the relationships among independent and dependent variables in models of environmental equity. Geographically weighted regression (GWR), in combination with choropleth mapping, can reveal this spatial nonstationarity and shed light on its form. We use GWR, in combination with conventional univariate and multivariate statistics, to model the density of air toxic releases in New Jersey, as listed in the U.S. Environmental Protection Agency’s Toxic Release Inventory (TRI). The GWR analysis shows that the relationships among race, class, employment, urban concentration, and land use with air toxic release density in New Jersey vary significantly over space. Generally, there is a positively significant relationship of minorities with air toxic releases over a large swath of urban and suburban New Jersey, although this pattern is not evident for all urban areas. Northeast New Jersey, the most densely populated part of the state, contains areas of both significantly positive and negative relationships between concentrations of minorities and air toxic releases. The association of minorities with concentrations of air toxic releases, where observed, is often mediated by other factors, though the role of these mediating factors also varies from place to place. In some of these areas the minority–air-toxic-release association is mediated by high poverty rates, in other areas, by the presence of industrial, commercial, and transportation land uses. Key Words: environmental justice, environmental racism, geographically weighted regression, spatial statistics, spatial analysis.

E

nvironmental justice is the principle that all people, regardless of race or socioeconomic status, are entitled to equal protection under environmental laws and to participate in environmental decision making in their community (Bullard 1996; EPA 2003). A fundamental question in environmental justice research concerns environmental equity—whether the spatial distribution of environmental risk is indeed equitable among different racial and socioeconomic groups. There have been many studies that have found statistical evidence both for and against environmental equity (Bullard 1983; UCC 1987; Mohai and Bryant 1992; Anderton et al. 1994; Bowen et al. 1995; Atlas 2002). The majority of these studies have focused on the locations of hazardous facilities as a proxy for environmental risk. The operative question for these studies is: Is there evidence of racial inequity in the distribution of hazardous facilities, and, if so, are there other nonracial factors that may statistically explain the variation in hazardous facility location? Many environmental equity studies employ some form of multivariate statistical analysis to disentangle the interrelationships among variables that may predict hazardous facility location (e.g., Ringquist 1997; Hird and Reese 1998; Sadd et al. 1999a). Multivariate sta-

tistical models of spatial data may exhibit spatial nonstationarity—the relationships among the independent and dependent variables vary over space (Fotheringham, Charlton, and Brunsdon 1996). While the presence of spatial nonstationarity in multivariate regression can, in one sense, be considered a methodological problem to be overcome, investigating spatial nonstationarity can also be a useful avenue of analysis in itself. This research investigates spatial nonstationarity in multivariate models of environmental equity using geographically weighted regression (GWR; Brunsdon, Fotheringham, and Charlton 1996) in combination with more conventional univariate and multivariate techniques. Whereas conventional multivariate regression produces a global predictive model, GWR expresses the spatial variation in model parameter estimates. To demonstrate the application of GWR to environmental equity analysis, we investigate the spatial distribution of air toxic release facilities in New Jersey listed in the U.S. Environmental Protection Agency’s (EPA) Toxic Release Inventory (TRI) (Figure 1). We model the association between the concentration of TRI facilities and socioeconomic status using U.S. Bureau of the Census tract-level data. We focus on New Jersey because of its relatively high number of TRI facilities and

Annals of the Association of American Geographers, 95(2), 2005, pp. 249–268 r 2005 by Association of American Geographers Initial submission, November 2003; revised submission, May 2004; final acceptance, May 2004 Published by Blackwell Publishing, 350 Main Street, Malden, MA 02148, and 9600 Garsington Road, Oxford OX4 2DQ, U.K.

250

Mennis and Jordan

Figure 1. The New Jersey study area: Locations of major and selected minor cities, locations of TRI facilities, and locations of PBT facilities. Small boxes indicate locations of larger-scale maps in Figures 7, 8, and 9.

minority population. New Jersey also contains large industrial, urban, and rural areas that make it ideal for investigating the role of urban concentration and industrial land use in the context of environmental equity. In addition, this statewide analysis complements the many environmental equity studies of TRI data that have focused on either a national (Ringquist 1997; Daniels and Friedman 1999) or local, metropolitan-area (Burke 1993; Sadd et al. 1999a) scope. Our specific objectives are twofold. First, we aim to advance statistical methodology in the analysis of environmental equity. Because spatial data analysis often differs from the statistical analysis of nonspatial data, issues of methodology have been a primary concern (and critique) of environmental justice research (Zimmerman 1994; Cutter, Holm, and Clark 1996; McMaster, Leitner, and Sheppard 1997; Bowen and Wells 2002). Improving the statistical assessment of environmental equity demands analytical techniques adapted for spatial data, such as GWR. Our second objective is to contribute to a substantive understanding of the various socioeconomic factors that are associated with hazardous facility location in New Jersey. We are particularly interested in the role of urban concentration and land use in mediating the relationship between race and ethnicity with TRI facility location, as well as how the role of these mediating factors might vary over space.

Environmental Equity, the TRI, and Spatial Nonstationarity There have been a number of U.S. national-level environmental equity studies focusing on TRI data.

Perlin et al. (1995) used bivariate analyses of countylevel socioeconomic data in finding that TRI releases were positively associated with the presence of minorities. In a multivariate, regression-based analysis estimating the presence, number, and release amount of TRI facilities in residential ZIP codes, Ringquist (1997) found minorities to be associated with greater environmental risk even when the influence of other factors indicating employment and class is removed. Daniels and Friedman (1999) used county-level, socioeconomic data in a multivariate regression model to predict pounds of toxic release per square mile. They found that counties with a high percentage of African Americans are associated with higher concentrations of toxic releases, although this relationship is mediated to a large degree by factors of urban concentration and industrial location. Other regional environmental equity studies of TRI data have focused on urban areas. In a tract-level study of Los Angeles, Burke (1993) found that the number of TRI facilities is positively associated with percent minority and negatively associated with income and population density. In another study focusing on southern California, Sadd et al. (1999a) used a combination of logit and tobit regression analyses at the tract level to predict the presence and toxicity ranking of TRI facilities. They found that industrial land use, employment in manufacturing, and population density are most important in predicting the presence of TRI facilities. However, when areas proximate to TRI-hosting tracts and rankings of toxicity are considered, percent minority emerges as a significant variable, even when the influence of these other variables is accounted for. In a study of Santa Clara, California, Szasz and Meuser (2000) found that environmental inequity in TRI facility

The Distribution of Environmental Equity location resulted from nonracially motivated processes of economic development as well as differences in education and employment among racial groups. A handful of spatially nested environmental equity studies have placed a local analysis of TRI data within the context of a more regional analysis. Bowen et al. (1995) used multivariate statistics to examine a series of explanatory variables predicting various TRI-based toxicity indices for both the county level in Ohio and the tract level in Cuyahoga County, Ohio, which includes the city of Cleveland. They found that at the county level, high percent minority is associated with high environmental risk, even when the influence of other variables indicating class has been removed. At the tract level, however, the relationship of race to environmental risk actually reverses when the influence of the other variables is accounted for. Mennis (2002a, b) performed a nested study that used multivariate regression to predict distance to TRI facilities in the Philadelphia area and Pennsylvania as a whole. He found much stronger evidence of environmental inequity in the Philadelphia area as compared to the entire state. The above research, as a whole, suggests that, whatever the mediating role of class and other factors may be, there is racial inequity in the distribution and toxic release amounts of TRI facilities, both nationally and in select urban areas. Unfortunately, a universal and recognizable pattern of the interaction of race with other explanatory factors of toxic releases remains elusive. Perhaps the most important conclusion that may be drawn from these studies is that the relationships among race, class, employment, and land use with regard to environmental risk vary from place to place; that is, there is spatial nonstationarity. There are a number of potential causes of spatial nonstationarity in the context of environmental equity analyses. One cause may be a misspecification of the model in the form of missing variables. For example, local urban, suburban, or rural character may mediate the relationships among race, class, and hazardous facility location. Or such relationships may be influenced by access to natural resources demanded by certain polluting industries. While researchers have attempted to account for such local character by incorporating dummy variables to represent individual subregions (e.g., Hird and Reese 1998), this approach is relatively coarse and cannot capture complex spatial variation. In addition, adding dummy variables assumes the researcher has a priori knowledge of all relevant geographic characteristics, which is unlikely. An alternative interpretation of spatial nonstationarity is that it does not represent a misspecification of a global multivariate regression

251

model, but is rather a reflection of what is unique about individual places (Fotheringham and Brunsdon 1999). For example, local political and economic environments may produce relationships among environmental risk and indicators of socioeconomic status that are unique to a particular community.

Data and Preprocessing The TRI data were acquired from the EPA. The TRI database is self-reported and provides the identity, location, and release amounts for facilities that treat or release any of approximately 650 toxic chemicals into the environment. We used the TRI database for 2000 and selected only those facilities with greater than zero pounds of air toxic releases, for which there are 483 facilities (Figure 1). We also extracted a subset of these TRI facilities that the EPA has recorded as releasing ‘‘persistent bioaccumulative toxins’’ (PBTs), chemicals that the EPA has recognized to be particularly dangerous because of their propensity to remain in the environment over long periods of time and accumulate in body tissue, including dioxin, lead, and mercury compounds (EPA 2004). There are sixty-two PBT facilities (Figure 1). Hereafter, we refer to these two groups of facilities as TRI and PBT facilities, respectively. We collected data on TRI and those TRI facilities classified as PBT facilities to distinguish between different levels of environmental hazard, assuming that PBT facilities carry greater potential hazards than other TRI facilities. We considered a number of approaches for representing the environmental hazard associated with the TRI and PBT facilities for each tract. The vast majority of environmental justice research has simply calculated a binary variable indicating whether a tract hosts, or is within a certain distance of, a hazardous facility (Anderton et al. 1994, Sadd et al. 1999a). A number of environmental equity studies have sought to improve on the quantification of environmental hazard by numerical modeling of chemical dispersion (Chakraborty and Armstrong 1997), modeling the distribution of air pollution from monitoring station data (Jerrett et al. 2001), or calculating the pounds of chemical release per unit area of the host spatial unit (Daniels and Friedman 1999). Unfortunately, numerical modeling of chemical dispersion is not scalable to large numbers of facilities, and air pollution monitoring station data is not available over large areas at a high enough sampling density for accurate modeling. Calculating the pounds of release per unit area is a promising approach, but it is still a relatively coarse indicator of toxicity that is subject to the arbitrary area of the host spatial unit and does not

252

Mennis and Jordan

account for the impact of chemical dispersion across spatial unit boundaries. We chose to model hazards from air toxic releases by calculating the density of TRI and PBT facilities within each tract. Our approach to calculating facility density is similar to that described by Mennis (2002b), Bolin et al. (2002), and Downey (2003). It captures the clustering of hazardous facilities and accounts for the fact that facilities may lie near the borders of tracts. By analyzing all TRI facilities with air toxic releases as well as only those TRI facilities classified as PBT facilities, we are able to differentiate categorically between different levels of hazard—tracts having high densities of PBT facilities being particularly hazardous. While we acknowledge that the density of facilities is not a direct indicator of environmental hazard, it does capture the relative concentration of hazardous facilities. And as Sadd et al. (1999b) point out, simple proximity to a hazardous facility can have negative effects on the perception of environmental quality in a neighborhood. The TRI and PBT facility data were aggregated to the tract level by calculating the density of facilities within each tract using the following technique. First, an empty 30-m resolution grid was overlaid with the facility location data. Second, the number of TRI (or PBT) facilities within a 1 km search radius of the center of each grid cell was calculated. Third, the number of facilities found within 1 km of a given grid cell was divided by the area of search to derive a facility density value for that grid cell. This simple density technique is a common approach to modeling point density (Bailey and Gatrell 1995). The tract map was then overlaid on top of the grid of facility density, and the mean facility density value of all the grid cells falling within each tract was calculated and assigned as an attribute of that tract. We selected a set of tract-level, socioeconomic variables from the 2000 U.S. Census that have been shown to be significant in previous environmental equity research (Figure 2). Concerning race and ethnicity, we focused on the percentage of persons who self-identified as black or Hispanic (note that Hispanic can be of any race) because these are the largest minority groups in New Jersey. We also calculated the percentage of people living below the poverty line and the percentage of people over sixteen years of age employed in manufacturing, as these variables have been shown to have significant relationships with hazardous facility location (Anderton et al. 1994). Population density (total population divided by the area of the tract, in persons/km2) is used as a proxy for urban concentration, which has been demonstrated to be significantly related to TRI facility location (Daniels and Friedman 1999). Educa-

tional attainment (percentage of persons over the age of twenty-five with a high school diploma) was also considered but was ultimately discarded because of its high correlation with poverty rate. There are 1,944 tracts in New Jersey. This analysis focused on the 1,933 tracts in New Jersey that contained people in 2000 and for which data on all the socioeconomic variables listed above were available. Of these, 883 had TRI facility density values greater than zero. A number of studies have found industrial land use to be significantly related to TRI facility location (Boer et al. 1997; Sadd et al. 1999a). As a proxy for industrial land use, we extracted land classified as industrial, commercial, or transportation (one class) for New Jersey from the U.S. Geological Survey’s (USGS) National Land Cover Data (NLCD) program. These 30-m resolution raster data were created using 1990s Landsat Thematic Mapper (TM) satellite imagery. While it would have been beneficial to have industrial land use as a category separate from commercial and transportation land uses, the NLCD data do not differentiate among these three classes, ostensibly because they were indistinguishable within the source imagery. We calculated the percent industrial, commercial, and transportation land use within each tract. For simplicity, we refer to this variable as simply ‘‘percent industrial.’’

Methods As a first step in the analysis, we simply review choropleth maps of all variables. We then derive the correlation between TRI and PBT density and each of the explanatory variables. Kendall’s tau-b correlation is used because of the nonnormal distribution of many of the variables. Conventional multivariate regression is then employed to estimate TRI and PBT density using combinations of the other variables. In order to improve the goodness of fit of the models and better approach a normal distribution of the residuals, population density and each of the facility density variables were transformed prior to the regression by taking the natural log and square root, respectively. Note that percent Hispanic has Pearson correlations of between 0.5–0.6 with both population density (natural log) and poverty rate. These correlations are not high enough to prohibit entry of all three variables into a single regression, but are worth noting as an aid to statistical interpretation. Following the conventional multivariate regression, GWR and choropleth mapping were used to explore spatial nonstationarity. It is useful to point out briefly the differences between spatial autocorrelation and spatial nonstationarity. Spatial autocorrelation, also referred to

The Distribution of Environmental Equity









253











− − −













− −







− −











Figure 2. Explanatory variables used in the analysis.

as spatial dependency, occurs when the distribution of the values of georeferenced observations is not spatially random; rather, observations located near one another

tend to have similar (or particularly dissimilar) values. Spatial autocorrelation can be measured for a single spatially distributed variable, for instance, using the well-

254

Mennis and Jordan

known Moran’s I and Geary’s c statistics (Cliff and Ord 1973). Spatial nonstationarity, on the other hand, occurs when the measurement of the spatial relationship among variables differs depending on where the measurement is taken (Fotheringham, Brunsdon, and Charlton 2002). GWR can be considered an exploratory statistical technique that allows a researcher to investigate the nature of spatial nonstationarity. There are other methods besides GWR for investigating spatial nonstationarity, including multilevel modeling (Jones 1991) and the expansion method (Casetti 1972, 1997). Multilevel models are defined by discrete hierarchies that attempt to account for effects at the level of both the individual and the place within which that individual resides. However, the hierarchies used in multilevel models tend to ignore the continuity of space, and it must be assumed that the spatial hierarchy captures the spatial process being modeled. The spatial expansion method allows model parameters to be defined by secondary equations that can represent spatial variation, but is limited to the representation of broad spatial trends (Fotheringham, Brunsdon, and Charlton 2000). GWR builds on the expansion method to account for more complex local variation in model parameter estimates. Consider the conventional regression equation X yi ¼ a0 þ ak xik þ ei : ð1Þ k

Geographically weighted regression modifies this equation so that there is an individual parameter estimate for each observation’s location. The equation may thus be rewritten as X yi ¼ a0 ðui ; vi Þ þ ak ðui ; vi Þxik þ ei ð2Þ k

where (ui, vi) represents the coordinate location of observation i (Brunsdon, Fotheringham, and Charlton 1996; Fotheringham, Brunsdon, and Charlton 1998; Fotheringham, Brunsdon, and Charlton 2002). Calibration of GWR takes place by weighting all observations according to a distance decay function away from observation i. Whereas conventional regression generates a single equation to represent global relationships among variables, GWR calibrates the regression equation differently for each observation based on a unique weighting of all observations. Note that for areal data, as is used in the present analysis, each observation’s location in the GWR is represented by the coordinates of its centroid. We used the GWR 2.2 software package for GWR analysis. GWR 2.2 provides a variety of options for choosing the weighting function that determines the weight wij of an observation j for the GWR calibration

centered on observation i. A Gaussian function may be used to calculate wij as a continuous function of distance such that wij ¼ expðd2ij =b2 Þ

ð3Þ

where dij is the distance between observations i and j and b is the bandwidth. Note that at distances greater than the bandwidth, the weight of observation j rapidly approaches zero. Thus, the bandwidth can be considered the sphere of influence around observation i. The bandwidth may be chosen by the user or determined by selecting the bandwidth that minimizes a cross-validation (CV) score given by n X CV ¼ ðyi  ^yi Þ2 ð4Þ i¼1

where n is the number of observations and ðyi  ^yi Þ is the difference between the observed and estimated values for observation i (Charlton, Fotheringham, and Brunsdon n.d.). The Golden Section search technique (Greig 1980) is used to minimize CV. GWR results can vary greatly with different bandwidth settings. One could posit that there is an optimal bandwidth that reflects the ‘true’ sphere of influence surrounding each observation. However, the idea of an optimal bandwidth is perhaps misleading. GWR is an exploratory technique, and investigating the range of GWR results under different bandwidth settings may reveal interesting patterns that may be missed were one to focus exclusively on a single bandwidth. We experimented with a number of different GWR bandwidth settings in this research. However, in the interests of bounding this research and generating a manageable volume of results, we used this experimentation to identify a single bandwidth on which to focus our interpretations, which we describe in detail in the following section. We would like to emphasize that both conventional multivariate regression and GWR are not used in this research to infer causality in the distribution of TRI facilities, but rather to explore the relationship between TRI facility location and race and ethnicity, while accounting for the confounding effects of other related socioeconomic and land-use characteristics. As Jerrett et al. (2001) note, the use of regression in environmental equity analysis does not necessarily imply the suggestion that minorities, poverty, or other socioeconomic characteristics cause the presence of TRI facilities. The distribution of TRI facilities may result from a number of interrelated factors, such as the availability of inexpensive and appropriately zoned land, access to transportation corridors, and the political empowerment of neighborhood

The Distribution of Environmental Equity residents, in addition to outright racism on the part of hazardous facility developers or environmental policy makers (Szasz and Meuser 1997; Helfand and Peyton 1999). Environmental inequity may be caused by the intersection of hazardous facility siting decisions with the residential choices made by members of various racial and ethnic groups, which may in turn be influenced by land prices and the availability of certain types of employment (Anderton et al. 1994; Been 1994). A number of authors have also argued that the causes of environmental inequity are rooted in the complex histories of regional industrial development, urban development, and race relations (Gelobter 1994; Pulido 2000; Szasz and Meuser 2000; Cutter, Hodgson, and Dow 2001). However, this historical perspective does not preclude the value of quantitative analyses that seek to describe the association of environmental hazards with socioeconomic and related factors as a means of investigating the presence, degree, and nature of environmental inequity, as is the case in the present research.

Results Mapping Analysis, Univariate Analysis, and Conventional Multivariate Regression A review of the maps presented in Figure 2 clearly shows spatial relationships among many of the variables. The highest values of percent black and percent Hispanic are generally found where population density is high, in major cities such as Camden, Trenton, and Newark, although concentrations of blacks and Hispanics also occur in areas of lower population density in central and southern New Jersey. Industrial, commercial, and transportation land uses are primarily concentrated in the northeastern part of the state surrounding the satellite cities of New York, as well as in the southwestern part of the state along the Delaware River, which forms the western state border. High degrees of poverty are also concentrated in the cities, although there are some higher poverty rates in the central and southern parts of the state broadly coincident with higher percent black and Hispanic values. High percent manufacturing employment is concentrated in the very southern end of the state and dispersed across northern New Jersey. Table 1 reports the results of the Kendall tau-b correlation for each of the explanatory variables with TRI and PBT density. All explanatory variables have a significant and positive relationship with each facility density variable, with the exception of population density with PBT density. Percent industrial stands out as

255

Table 1. Kendall Tau-b Correlation Coefficients for TRI and PBT Density Variable Percent Black Percent Hispanic Population Density Percent Industrial Percent Living below the Poverty Line Percent Manufacturing Employment

TRI Density PBT Density 0.098** 0.284** 0.190** 0.280** 0.149** 0.287**

0.080** 0.130** 0.007 0.190** 0.104** 0.120**

**po0.01

being highly correlated with both TRI and PBT density, as does percent Hispanic and percent manufacturing employment to a lesser degree. For the conventional regressions of each of the facility density variables, we first looked at each ethnicity variable in combination with population density and percent industrial to control for urban concentration and industrial land use. We then added in percent living in poverty and percent manufacturing employment to control for economic factors. This procedure was done for each ethnicity variable separately because we did not want the variation in one ethnic variable to obscure a significant relationship evident in the other. Table 2 shows the results for the regression of TRI density. Higher densities of TRI facilities are associated primarily with industrial land cover, and to a lesser degree with higher population density and rates of manufacturing employment. Percent black and percent living in poverty are not significant, indicating that the relationships of these two variables with TRI density that are evident in Table 1 are explained by other explanatory variables. Percent Hispanic is significant, although its contribution is marginal when all other explanatory variables are included. The results for the regression of PBT density are reported in Table 3. Goodness-of-fit values are quite low as compared to the models of TRI density. Percent industrial explains the greatest amount of variation in PBT density. Percent Hispanic retains a positively significant relationship with PBT density even after the influence of the other explanatory variables is accounted for. Note that population density flips sign to become significantly negative in Models 3 and 4, most likely due to collinearity with percent Hispanic.

Geographically Weighted Regression We experimented with both CV- and user-defined bandwidth settings. Bandwidths derived from crossvalidation typically converged on a bandwidth of less than

256

Mennis and Jordan Table 2. Standardized Coefficients for Regression of TRI Density (Square Root)

Independent Variables Constant Percent Black

Model 1

Model 2

Model 3

Model 4

 0.178 (  7.148)***  0.028 (  1.343)

 0.278 (  10.973)***  0.005 (  0.200)

 0.088 (  3.357)***

 0.250 (  8.722)***

0.190 (8.050)*** 0.132 (5.791)*** 0.362 (17.698)***

0.056 (2.110)* 0.170 (7.419)*** 0.354 (17.386)*** 0.005 (0.213) 0.259 (12.220)*** 1,933 0.325

Percent Hispanic Population Density (natural log)

0.234 (11.011)*** 0.407 (20.146)***

Percent Industrial Percent Living below the Poverty Line Percent Manufacturing Employment N Adjusted R2

1,933 0.247

0.189 (9.969)*** 0.359 (17.738)*** 0.023 (0.881) 0.278 (14.435)*** 1,933 0.323

1,933 0.271

*po 0.05, ***p o0.005; t-values appear in parentheses.

5 km. This small bandwidth is most likely an indication of the high spatial variability in the relationships among the independent variables with the facility density variables, particularly in urban areas where the tracts tend to be relatively small. However, a handful of the rural tracts have diameters over 10 km, making the use of a bandwidth less than 5 km impractical. Note that in the most densely populated areas, and hence areas with the smallest tracts (e.g., Newark), a 5 km radius circle can contain well over one hundred tracts, while in the most sparsely populated areas with large tracts, a 5 km radius circle can be contained wholly within one tract.

We also generated results from GWR runs using 10 km, 25 km, and 50 km user-defined bandwidths. Resulting choropleth maps indicated that bandwidths greater than 10 km produced a large degree of smoothing; the spatial variation in parameter estimates appeared similar to a second-order trend surface with the primary trend axis running in a north–south orientation. Consequently, more subtle local variations in parameter estimates could not be recognized. The 10 km bandwidth produced the most informative results, in terms of capturing the spatial variation of parameter estimates according to recognizable areas such as particular cities

Table 3. Standardized Coefficients for Regression of PBT Density (Square Root) Independent Variables Constant Percent Black

Model 1 0.007 (0.749)  0.021 (  0.903)

Model 2 0.001 (0.116)  0.057 (  2.031)*

Percent Hispanic Population Density (natural log) Percent Industrial

0.000 (  0.005) 0.249 (11.004)***

Percent Living below the Poverty Line Percent Manufacturing Employment N Adjusted R2

1,933 0.059

*po0.05, **po0.01, ***po0.005; t-values appear in parentheses.

 0.036 (  1.476) 0.214 (9.032)*** 0.097 (3.188)*** 0.094 (4.159)*** 1,933 0.073

Model 3

Model 4

0.028 (3.010)***

0.017 (1.637)

0.142 (5.335)***  0.076 (  2.950)*** 0.215 (9.310)***

1,933 0.072

0.097 (3.132)***  0.076 (  2.862)*** 0.206 (8.673)*** 0.037 (1.377) 0.067 (2.679)** 1,933 0.076

The Distribution of Environmental Equity and rural regions. This is likely because a 10 km bandwidth is small enough to capture local socioeconomic and facility density relationships yet is large enough to accommodate the irregular distribution of tract centroids in New Jersey. We therefore used a 10 km GWR bandwidth for this study. Figure 3 shows the results of the GWR of TRI density with percent black as the only independent variable, including choropleth maps of the distribution of the constant, percent-black coefficient, and local R2. Note that maps of the constant and coefficient show negatively and positively significant t-values (po0.05) in light gray and black, respectively; the darker gray indicates areas that are not significant. Percent black is positively significant in the cities of Camden and Vineland, in the northeastern corner of the state, and in the area stretching from Trenton through New Brunswick. Interestingly, in the Newark area, which hosts large numbers of both blacks and TRI facilities, percent black is negatively significant. The constant is positively significant over a large area covering the northeastern portion of the state, through Trenton, Camden, and south along the Delaware River, indicating that these areas have elevated values of TRI density even after the variation in percent black has been accounted for. Local R2 values are highest in the south and along the Delaware River northwest of Trenton. When the other explanatory variables (except percent Hispanic) are added into the GWR (as in the conventional regression presented in Table 2, Model 2), the area of positive significance for percent black shrinks considerably (Figure 4). Percent black is now significant primarily in two areas: south of Camden along the Delaware River and in the area surrounding New Brunswick. Those areas that were positively significant for percent black in

Figure 3. Results of GWR of TRI facility density with percent black as the independent variable. Maps of the constant and variable coefficients show t-values where light gray is negatively significant, dark gray is not significant, and black is positively significant (po0.05).

257

Figure 3 but not in Figure 4 are explained by other variables. In the northeast corner of the state, TRI density is explained by percent manufacturing employment, percent industrial, and population density. And in the Camden area it is explained by percent industrial and percent living below the poverty line. Also of note is the large area of positive significance for percent industrial, which extends across the northeastern portion of the state to Trenton and south along the Delaware River. Recall that because this variable includes industrial, commercial, and transportation land uses, it is impossible to distinguish which among the three land uses is of particular relevance. It may be that commercial land use is related to hazardous facility location in one area and industrial land use is the basis for the relationship in another area. Note also that the area of significance for the constant is largely reduced as compared to Figure 3, and the local R2 values are generally higher. Results of the GWR of TRI density with percent Hispanic as the only independent variable are broadly similar to that of percent black, although there is a larger and more dispersed pattern of positive significance, and the area of negative significance in Newark is notably absent (Figure 5). When the other explanatory variables are added in the GWR, it is apparent that TRI density is better explained by percent industrial in Trenton and by percent living below the poverty line in Camden (Figure 6). Percent Hispanic retains its positive significance in the area extending northwest from Newark, the area between Trenton and New Brunswick, and south of Camden along the Delaware River. Results for the GWR of PBT density for both percent black and Hispanic (both in isolation and with the other explanatory variables) are similar to that of the GWR of TRI density, though there are some notable differences.

− − − −

258

Mennis and Jordan

Figure 4. Results of GWR of TRI facility density with percent black, population density, percent industrial, percent poverty, and percent manufacturing as the independent variables. Maps of the constant and variable coefficients are as in Figure 3.

− − − −

The Distribution of Environmental Equity



Figure 5. Results of GWR of TRI facility density with percent Hispanic as the independent variable. Maps of the constant and variable coefficients are as in Figure 3.

The relationship of both percent black and Hispanic with PBT density in Camden is mediated by poverty rate. In addition, population density has a negatively significant relationship with PBT density in Camden and Newark, whereas the relationship in those areas is positively significant in the GWR of TRI density. Also of note is that although the goodness-of-fit values in the conventional multivariate regressions of PBT density were all less than 0.08 (Table 3), in the analogous GWR runs, several tracts in the southern part of the state had goodness-of-fit values ranging between 0.75 and 0.98. Previous research has suggested that minorities tend to reside near hazardous facilities, but not too close, forming something of a bull’s-eye pattern where percent minority may be low at the actual facility location but rises rapidly at a proximate distance (Anderton et al. 1994; Sadd et al. 1999a; Mennis 2002b). To investigate this pattern, we recalculated the TRI and PBT density variables using a 5 km radius instead of a 1 km radius. We then repeated all the analytical steps presented thus far for the two new facility-density variables. Results of the analysis of 5 km radius TRI and PBT density are broadly similar to those for the 1 km radius versions, with some notable exceptions. Kendall tau-b correlations are higher for all explanatory variables, and all variables are significant (po0.01) (Table 4). Population density and percent Hispanic are the most highly correlated with 5 km radius TRI and PBT facility density, respectively. All the conventional regression models for the 5 km radius facility density variables also show improved goodness of fit as compared to their 1 km radius counterparts (Tables 5 and 6). These models differ from their 1 km radius analogs primarily in the elevated role of population density, which is positively significant in all regression models of 5 km radius TRI and PBT density

259

− − −

(po0.005). This supports the notion of a bull’s-eye pattern where population is, in fact, concentrated near (i.e., within 5 km) but not too near (i.e., within 1 km) TRI facilities. This pattern appears to be particularly strong for PBT facilities. Another difference of note is that percent black is positively significant in multivariate regressions of 5 km radius TRI density. Because of space considerations, we do not show maps of the results of the GWRs of 5 km radius TRI and PBT density, but briefly discuss how they differ with their 1 km radius counterparts. One notable difference between the GWRs of 5 km versus 1 km radius TRI density with percent black in isolation is that the negatively significant relationship of percent black for Newark shown in Figure 3 becomes insignificant. Generally, in the GWRs of TRI and PBT density where percent black and percent Hispanic are entered into the regression in isolation, the 5 km radius versions result in a larger area of positive significance for each ethnicity variable as compared to the 1 km radius models. This does not always transfer into a similarly larger area in a multivariate setting, however, because many of these new areas of significance are explained by population density and percent living below the poverty line, which take more prominent roles than they do in the 1 km radius GWR models. A notable exception is that of the GWR of 5 km radius PBT density, for which percent Hispanic remains significant across a large area covering Camden, Trenton, New Brunswick, and Newark.

Discussion Our results generally support the notion that there is environmental inequity in New Jersey. The correlation tests show that both blacks and Hispanics are dispro-

260

Mennis and Jordan

Figure 6. Results of GWR of TRI facility density with percent Hispanic, population density, percent industrial, percent poverty, and percent manufacturing as the independent variables. Maps of the constant and variable coefficients are as in Figure 3.

− − − −

portionately concentrated near air toxic releases and for those facilities releasing particularly dangerous toxins, PBT facilities. While we do not subscribe to the notion that racial inequity in environmental risk is somehow

made acceptable in the presence of urban poverty or other related factors, it is certainly of interest to understand how factors of poverty, population density, manufacturing employment, and industrial land cover

The Distribution of Environmental Equity Table 4. Kendall Tau-b Correlation Coefficients for 5 km Radius TRI and PBT Density Variable

TRI Density

PBT Density

0.221** 0.354** 0.440** 0.255** 0.231** 0.265**

0.166** 0.385** 0.329** 0.255** 0.267** 0.202**

Percent Black Percent Hispanic Population Density Percent Industrial Percent Poverty Percent Manufacturing **po0.01.

interact in the context of environmental inequity. After these other factors are accounted for, Hispanics are still likely to reside near (within 1 km of ) concentrations of TRI and PBT facilities. When the definition of near is expanded to 5 km, blacks are likely to live near concentrations of TRI facilities, and Hispanics are likely to live near concentrations of PBT facilities. Population density and land use explain much of the variation in the relationship of percent black with TRI and PBT density, while poverty and land use explain much of the variation in the relationship of percent Hispanic with TRI and PBT density. In contrast to the conventional correlation and multivariate regression analysis, the GWR analysis shows that the relationships between percent black and percent Hispanic with TRI and PBT concentration are not uniform across New Jersey, but are more pronounced in some areas and not evident in other areas. It is perhaps intuitive that one would expect environmental inequity

261

to be concentrated in urban areas. Certainly, minorities, hazardous facilities, and concentrations of poverty tend to co-locate in cities and their extended metropolitan areas (Figure 2). A number of researchers have suggested that statistical evidence of racial inequity in hazardous facility location can be explained in large measure by the tendency of minorities and industrial land to co-locate in urban areas (Daniels and Friedman 1999; Bowen and Wells 2002). Others have noted the interrelationship between the process of urban, industrial development and historical settlement patterns among various ethnic groups that underlies currently observed patterns of environmental equity (Gelobter 1994; Pulido 2000). Our results indicate, however, that while urban concentration (as indicated by population density) is an important factor in explaining environmental inequity, its influence varies from location to location. Though many urban areas in New Jersey exhibit evidence of racial inequity in the distribution of TRI and PBT facilities (e.g., Trenton and Camden), there are also urban areas that do not. Atlantic City, for example, is home to a large concentration of minorities but relatively few TRI facilities. There are also relatively rural areas that exhibit environmental inequity. Vineland, for example, is a small city and is surrounded by relatively low population density, yet it is home to environmental inequity, particularly for Hispanics. Newark has the distinction of being a city that exhibits a negative relationship between percent black and TRI facilities (when TRI density is calculated using a 1 km radius), even though the greater Newark area hosts large concentrations of both.

Table 5. Standardized Coefficients for Regression of 5 km Radius TRI Density (Square Root) Independent Variables Constant Percent Black

Model 1

Model 2

Model 3

Model 4

 0.259 (  15.967)*** 0.066 (3.580)***

 0.344 (  21.101)*** 0.123 (5.940)***

 0.250 (  14.506)***

 0.386 (  20.885)***

0.064 (3.052)*** 0.485 (23.980)*** 0.303 (16.742)***

 0.103 (  4.482)*** 0.518 (26.253)*** 0.278 (15.854)*** 0.053 (2.677)** 0.295 (16.084)*** 1,933 0.496

Percent Hispanic Population Density (natural log) Percent Industrial

0.494 (26.760)*** 0.310 (17.595)***

Percent Living below the Poverty Line Percent Manufacturing Employment N Adjusted R2

1,933 0.430

*po0.05, **po0.01, ***po0.005; t-values appear in parentheses.

0.469 (25.833)*** 0.273 (15.684)***  0.047 (  2.106)* 0.273 (16.489)*** 1,933 0.500

1,933 0.429

262

Mennis and Jordan Table 6. Standardized Coefficients for Regression of 5 km Radius PBT Facility Density (Square Root)

Independent Variables Constant Percent Black

Model 1

Model 2

Model 3

Model 4

 0.099 (  13.237)***  0.109 (  5.269)***

0.110 (  13.954)***  0.136 (  5.526)***

 0.006 (  7.548)***

 0.007 (  8.381)***

0.235 (10.201)*** 0.234 (10.525)*** 0.278 (13.972)***

0.222 (8.319)*** 0.259 (11.249)*** 0.284 (13.840)***  0.58 (  2.497)* 0.060 (2.790)** 1,933 0.314

Percent Hispanic Population Density (natural log) Percent Industrial

0.383 (18.487)*** 0.343 (17.356)***

Percent Living below the Poverty Line Percent Manufacturing Employment N Adjusted R2

1,933 0.282

0.351 (16.352)*** 0.301 (14.615)*** 0.082 (3.108)*** 0.122 (6.213)*** 1,933 0.300

1,933 0.309

*po0.05, **po0.01, ***po0.005; t-values appear in parentheses.

There are also differences between blacks and Hispanics in the way population density, industrial land use, and other factors interact with these different groups in the context of environmental inequity. These differences are highlighted in the area surrounding Newark. Recall that Newark has a positively significant relationship of percent Hispanic with PBT density, whereas for percent black, the relationship is negatively significant. Figure 7 provides a series of close-up maps of Newark and the surrounding area (marked as the small box in the city location map in Figure 1) for all explanatory variables, as well as TRI and PTB density (1 km radius). Note the high spatial heterogeneity in nearly all the variables; for example, very high population density tracts are adjacent to very low population density tracts. This is particularly notable in the northeastern portion of the population density map in which a strip of very low population density separates areas of high population density— Newark to the west and Jersey City to the east. This strip of relatively sparse population consists of Newark Bay and the waterfront area, which is home to a large concentration of industrial, commercial, and transportation land uses, as indicated by the percent industrial map. Maps of TRI and PBT density in Figure 7 show that most of the TRI and PBT facilities in the area are located within these industrial lands of Newark Bay, as well as along the shoreline of Arthur Kill (which forms the border between New Jersey and Staten Island, New York) in places such as Elizabeth and Perth Amboy. This pattern is particularly strong for PBT density. A review of the percent black and percent Hispanic maps in Figure 7

shows that while high concentrations of blacks live near these industrial-dominated tracts, Hispanics are more prominent within those tracts. High percent manufacturing employment is also coincident with industrial land cover. Percent poverty, however, is concentrated primarily in Newark, coincident with concentrations of blacks. These maps demonstrate that the negative relationship between percent black and TRI and PBT density found for Newark results from the concentration of blacks within Newark proper, although Newark is near (in fact, surrounded by), but not coincident with, high concentrations of facilities in adjacent industrial lands. Hispanics, however, are distributed more homogeneously throughout the Newark area, including on those industrial lands that host many of the TRI and PBT facilities. Because these industrial lands are typically sparsely populated, the relationship between percent Hispanic and PBT density remains strongly positive even when the influence of population density is accounted for. Compare Figure 7 to Figure 8, which provides an analogous set of large-scale maps for the boxed area surrounding Camden in Figure 1. TRI and PBT facilities are concentrated in Camden, as well as in an arc just to the east of Camden and south of Gloucester City along the Delaware River. Blacks and Hispanics are also concentrated in Camden, though concentrations of blacks dot the entire area. Camden is also home to the highest concentrations of people, industrial land, and poverty in the area. Consequently, although the GWRs indicate that Camden is racially inequitable in the distribution of TRI and PBT facilities (Figures 3 and 5), the relationship is explained by other factors in the

The Distribution of Environmental Equity

263

































− −





− − − −













Figure 7. Large-scale map of variables for Newark and surrounding area, shown in small box in Figure 1.

multivariate GWRs (Figures 4 and 6). The areas that emerge as positively significant for blacks and Hispanics in the multivariate GWRs are the areas immediately

outside Camden—south of Camden along the Delaware River and southeast of Camden (Figures 4 and 6). While home to lesser concentrations of minorities, these areas

264

Mennis and Jordan

− − − −

− − − −

− − − −

− − − −

− − − −

− − − −

− − −

− − −

exhibit evidence of environmental inequity in a multivariate context because of the reduced influence of population density and poverty. Finally, compare Newark and Camden to the Trenton area. Figure 9 shows a set of maps similar to those in Figures 7 and 8 but for the boxed area surrounding Trenton in Figure 1. As in Camden, Trenton exhibits evidence of racial inequity in the distribution of TRI facilities for both blacks and Hispanics (Figures 3 and 5). But in the multivariate GWRs, the inequity is explained

Figure 8. Large-scale map of variables for Camden and surrounding area, shown in small box in Figure 1.

by other factors (Figures 4 and 6). Figure 9 shows that concentrations of blacks, Hispanics, industrial lands, high poverty rates, and high population density are largely coincident in a single area within Trenton. TRI facilities in the area are concentrated in Trenton, but also in the area extending south of Trenton along the Delaware River. PBT facilities in the area are concentrated exclusively southeast of Trenton. Thus, any spatial relationship of concentrations of minorities with TRI facility location is explained by other factors.

The Distribution of Environmental Equity

265





− −

− −











− −











− −

− − − −

− −

− − −



Figure 9. Large-scale map of variables for Trenton and surrounding area, shown in small box in Figure 1. Note that the scale is larger than that depicted in Figures 7 and 8.

Conclusion Nearly all previous environmental equity studies employing multivariate regression have assumed that the relationships among the independent and dependent

variables are stationary across the study region. Consequently, the idea that different places within a single study region might differ from one another with regard to the nature of environmental inequity has been ignored in the quest for finding global relationships that hold

266

Mennis and Jordan

over the entire data set. The present research demonstrates that conventional multivariate regression can hide important local variations in the relationships among independent and dependent variables in models of environmental equity. This spatial nonstationarity can be significant within a single region, and even within a single urban area (e.g., Newark and surrounding cities). GWR, in combination with choropleth mapping, can reveal spatial nonstationarity and shed light on its form. For example, recall that the conventional regression of TRI and PBT density in New Jersey suggests that although the presence of minorities is associated with concentrations of TRI and PBT facilities, this relationship is largely mediated by urban concentration associated with industrial, commercial, and transportation land uses. The GWR, however, indicates that this regionwide pattern is really a broad summarization that masks extensive local variation. TRI and PBT facilities are indeed associated with minorities, but this pattern is evident in only some, not all, areas of the state. While many of these areas are indeed cities, not all cities have a positive association between hazardous facility concentration and minorities, and there are also relatively sparsely populated areas where this relationship is evident. Where concentrations of minorities and hazardous facilities do coincide, their association is often mediated by land use. In some places, however, the presence of poverty is most influential in estimating hazardous facility concentration, while in other places, high population density has the greatest influence. These local variations in the global, positive relationship of hazardous facilities with socioeconomic status are obscured by conventional multivariate regression, but are revealed using GWR. We hope that other researchers will recognize the utility of GWR as an exploratory data analysis technique for environmental equity analysis. The sensitivity of statistical assessments of environmental equity to scale is also shown to be important here. A number of authors have shown that the quantitative analysis of environmental equity, like many analyses utilizing demographic data aggregated to spatial units, is subject to the modifiable areal unit problem— the resolution and partitioning of the spatial units influences the analytical results (Fotheringham and Wong 1991; Cutter, Holm, and Clark 1996; Sui 1999). There has been much debate over the most appropriate spatial unit, whether tract, block group, or ZIP code, and thus the scale of analysis, for environmental equity studies (Anderton et al. 1994; Williams 1999). We believe that the purpose of investigating the impact of scale on multivariate analysis of environmental equity is not to prove (or disprove) the robustness of a given multivari-

ate model to scale variation, but is rather a means toward exploring the spatial relationships among factors that may be associated with hazardous facility location. We partially addressed this issue in this research by using both 1 km and 5 km search radii in the simple density function used to generate the TRI and PBT facility density variables. GWR offers another approach to investigate the issue of scale through the use of the bandwidth setting. While we chose in this research to use a bandwidth of 10 km, an iterative and incremental variation of different bandwidth settings would reveal the sensitivity of the evidence of environmental equity to different scales of analysis. In addition, the GWR software provides an option to allow the bandwidth to be optimized for each observation based on the density of surrounding observations; the bandwidth shrinks in densely sampled areas and expands where sampling is sparse. We chose not to use this variable bandwidth option in this research because the bandwidth would vary based on the relatively arbitrary partitioning of space into tracts, and we wanted to hold the bandwidth constant so as not to introduce another component of variation to the analysis beside that of the spatial nonstationarity present in the multivariate models. However, the use of a spatially varying bandwidth adds an interesting dimension to a GWR analysis as it allows the sphere of influence around each observation to vary, and it may thus capture the fact that regions of relative homogeneity in the relationships among independent and dependent variables may vary in size across space. In future research, we intend to explore the impact of scale on evidence of environmental inequity in a more structured manner by varying the simple density function radius (e.g., 1 km versus 5 km), spatial unit of analysis (e.g., tract versus block group), and GWR bandwidth in concert. This research does not indicate the temporal ordering of TRI and PBT facility siting or minority settlement— we cannot say whether hazardous facilities developed in minority neighborhoods or minorities moved near existing facilities. Consequently, this research does not imply the presence of intentional discrimination by hazardous facility developers or environmental policymakers or other potential causes of environmental inequity. Our results do, however, suggest particular lines of investigation for exploring the causes of environmental inequity in the context of certain places along the lines of the qualitative research exemplified by Pulido, Sidawi, and Vos (1996) and Boone and Modarres (1999) for certain neighborhoods in Los Angeles. For instance, the results of the GWR show that the relationships among socioeconomic status and environ-

The Distribution of Environmental Equity mental hazards differ from place to place. One can therefore conclude that the industrial development and urban demographic processes that have created the observed patterns of environmental inequity are also different in these different places. Consider, for instance, the case of Newark, where there is strong evidence of environmental inequity in the distribution of PBT facilities for Hispanics, but not for blacks. One might ask why Hispanics, as opposed to blacks or any other group, are more prominent in the relatively sparsely populated industrial lands, which host the densest concentrations of PBT facilities. Why is poverty more strongly associated with blacks in close proximity to PBT facilities than Hispanics? Why does this particular spatial pattern that produces environmental inequity for Hispanics occur in Newark but not in other industrial cities in New Jersey, such as Trenton and Camden? From a policy perspective, this research indicates particular areas that are racially inequitable in the distribution of air toxic releases. This information may aid state and federal environmental regulatory agencies in advancing environmental justice through the restriction of the development of new facilities with planned air toxic releases in targeted communities. Our results can also play a role in the ongoing legal battles in environmental justice that typically pit environmental justice activists, community groups, and legal aid organizations against environmental regulatory agencies or hazardous facility developers. This research gives credence to activist claims that certain minority neighborhoods do indeed bear a disproportionate burden of environmental risk via proximity to air toxic releases. In many areas, this racial inequity is significant even when other factors that are associated with hazardous facility location, such as urban concentration and industrial, commercial, and transportation land uses, are accounted for.

Acknowledgments The authors would like to thank Stewart Fotheringham, Chris Brunsdon, and Martin Charlton for developing and distributing the GWR software. Thanks also to Gary Gaile for providing helpful comments on a previous draft.

References Anderton, Douglas L., Andy B. Anderson, John M. Oakes, and Michael R. Fraser. 1994. Environmental equity: The demographics of dumping. Demography 31:229–48. Atlas, Mark. 2002. Few and far between? An environmental equity analysis of the geographic distribution of hazardous waste generation. Social Science Quarterly 83:365–78.

267

Bailey, T. C., and A. Gatrell. 1995. Interactive spatial data analysis. Harlow, Essex, U.K.: Longman Scientific and Technical. Been, Vicki. 1994. Locally undesirable land uses in minority neighborhoods: Disproportionate siting or market dynamics? Yale Law Journal 103:1383–422. Boer, J. Thomas, Manuel Pastor Jr., Jim L. Sadd, and Lori D. Snyder. 1997. Is there environmental racism? The demographics of hazardous waste in Los Angeles County. Social Science Quarterly 78:793–810. Bolin, Bob, Amy Nelson, Edward J. Hackett, K. David Pijawka, C. Scott Smith, Diane Sicotte, Edward K. Sadalla, Eric Matranga, and Maureen O’Donnell. 2002. The ecology of technological risk in a sunbelt city. Environmental and Planning A 34:317–39. Boone, Christopher G., and Ali Modarres. 1999. Creating a toxic neighborhood in Los Angeles County: A historical examination of environmental inequity. Urban Affairs Review 35:163–87. Bowen, William M., Mark J. Salling, Kingsley E. Haynes, and Ellen J. Cyran. 1995. Toward environmental justice: Spatial equity in Ohio and Cleveland. Annals of the Association of American Geographers 85:541–663. Bowen, William M., and Michael V. Wells. 2002. The politics and reality of environmental justice: A history and considerations for public administrators and policy makers. Public Administration Review 62:687–98. Brunsdon, Chris, Stewart A. Fotheringham, and Martin Charlton. 1996. Geographically weighted regression: A method for exploring spatial nonstationarity. Geographical Analysis 28:281–98. Bullard, Robert D. 1983. Solid waste sites and the black Houston community. Sociological Inquiry 53:273–88. FFF. 1996. Environmental justice: It’s more than waste facility siting. Social Science Quarterly 77:493–99. Burke, Lauretta M. 1993. Race and environmental equity: A geographic analysis in Los Angeles. GeoInfo Systems October: 44–50. Casetti, E 1972. Generating models by the expansion method: Applications to geographic research. Geographical Analysis 4:81–91. FFF. 1997. The expansion method, mathematical modeling, and spatial econometrics. International Regional Science Review 20:9–32. Chakraborty, Jayajit, and Marc P. Armstrong. 1997. Exploring the use of buffer analysis for the identification of impacted areas in environmental equity assessment. Cartography and Geographic Information Systems 24:145–57. Charlton, Martin, Stewart Fotheringham, and Chris Brunsdon. n.d. Geographically weighted regression version 2.x, User’s manual and installation guide. Cliff, Andrew D., and J. Keith Ord. 1973. Spatial autocorrelation. London: Pion. Cutter, Susan L., Michael E. Hodgson, and Kirstin Dow. 2001. Subsidized inequities: The spatial patterning of environmental risks and federally assisted housing. Urban Geography 22:29–53. Cutter, Susan L., Danika Holm, and Lloyd Clark. 1996. The role of geographic scale in monitoring environmental justice. Risk Analysis 16:517–25. Daniels, Glynis, and Samantha Friedman. 1999. Spatial inequality and the distribution of industrial toxic releases: Evidence from the 1990 TRI. Social Science Quarterly 80: 244–62.

268

Mennis and Jordan

Downey, Liam. 2003. Spatial measurement, geography, and urban racial inequality. Social Forces 81:937–52. EPA (U.S. Environmental Protection Agency). 2003. Environmental justice fact sheet: EPA’s commitment to environmental justice. Washington, DC: U.S. EPA Office of Environmental Justice. FFF. 2004. Toxic release inventory (TRI) program: persistent, bioaccumulative, and toxic (PBT) chemicals rules. http://www.epa.gov/tri/lawsandregs/pbt/pbtrule.htm (last accessed January 28, 2004). Fotheringham, Stewart A., and Chris Brunsdon. 1999. Local forms of spatial analysis. Geographical Analysis 31: 340–58. Fotheringham, Stewart A., Chris Brunsdon, and Martin E. Charlton. 1998. Geographically weighted regression: A natural evolution of the expansion method for spatial data analysis. Environment and Planning A 30:1905–27. FFF. 2000. Quantitative geography: Perspectives on spatial data analysis. London: Sage Publications. FFF. 2002. Geographically weighted regression: The analysis of spatially varying relationships. Chichester, U.K.: Wiley. Fotheringham, Stewart A., Martin E. Charlton, and Chris F. Brunsdon. 1996. The geography of parameter space: An investigation into spatial nonstationarity. International Journal of Geographical Information Systems 10:605–27. Fotheringham, Stewart A., and David W. S. Wong. 1991. The modifiable areal unit problem in multivariate statistical analysis. Environment and Planning A 23:1025–44. Gelobter, Michael 1994. The meaning of urban environmental justice. Fordham Urban Law Journal 21:841–56. Greig, D. M. 1980. Optimisation. Harlow, U.K.: Longman. GWR 2.2. Martin Charlton, Stewart Fotheringham, and Chris Brunsdon. University of Newcastle, Newcastle, U.K. Helfand, Gloria E., and L. James Peyton. 1999. A conceptual model of environmental justice. Social Science Quarterly 80:68–83. Hird, John A., and Michael Reese. 1998. The distribution of environmental quality: An empirical analysis. Social Science Quarterly 79:693–716. Jerrett, Michael, Richard T. Burnett, Pavlos Kanaroglou, Paul Eyles, Norm Finkelstein, Chris Giovis, and Jeffrey Brook. 2001. A GIS-environmental justice analysis of particulate air pollution in Hamilton, Canada. Environment and Planning A 33:955–73. Jones, Kelvyn. 1991. Specifying and estimating multilevel models for geographical research. Transactions of the Institute of British Geographers 16:148–59. McMaster, Robert B., Helga Leitner, and Eric Sheppard. 1997. GIS-based environmental equity and risk assessment: Methodological problems and prospects. Cartography and Geographic Information Systems 24:172–89.

Mennis, Jeremy. 2002a. Socioeconomic disadvantage and environmentally hazardous facility location in Pennsylvania. The Pennsylvania Geographer XL (2): 113–24. FFF. 2002b. Using geographic information systems to create and analyze statistical surfaces of population and risk for environmental justice analysis. Social Science Quarterly 83:281–97. Mohai, Paul, and Bunyan Bryant. 1992. Environmental racism: Reviewing the evidence. Race and the incidence of environmental hazards, ed. Bunyan Bryant and Paul Mohai, 163–76. Boulder, CO: Westview Press. Perlin, Susan A., Woodrow R. Setzer, John Creason, and Ken Sexton. 1995. Distribution of industrial air emissions by income and race in the United States: An approach using the Toxic Release Inventory. Environmental Science and Technology 29:69–80. Pulido, Laura. 2000. Rethinking environmental racism: White privilege and urban development in southern California. Annals of the Association of American Geographers 90:12–40. Pulido, Laura, Steve Sidawi, and Robert O. Vos. 1996. An archaeology of environmental racism in Los Angeles. Urban Geography 17:419–39. Ringquist, Evan J. 1997. Equity and the distribution of environmental risk: The case of TRI facilities. Social Science Quarterly 78:811–29. Sadd, James L., Manuel Pastor Jr., J. Thomas Boer, and Lori D. Snyder. 1999a. ‘‘Every breath you take . . . ’’: The demographics of toxic air releases in southern California. Economic Development Quarterly 13:107–23. FFF. 1999b. Response to comments by William M. Bowen. Economic Development Quarterly 13:135–40. Sui, Daniel. 1999. GIS, environmental equity analysis, and the modifiable areal unit problem (MAUP). Geographic information research: Trans-Atlantic perspectives, ed. M. Craglia and H. Onsrud, 41–54. London: Taylor and Francis. Szasz, Andrew, and Michael Meuser. 1997. Environmental inequalities: Literature review and proposals for new directions in research and theory. Current Sociology 45:99–120. FFF. 2000. Unintended, inexorable: The production of environmental inequalities in Santa Clara County, California. The American Behavioral Scientist 43:602–32. UCC (United Church of Christ’s Commission for Racial Justice). 1987. Toxic wastes and race in the United States: A national report on the racial and socio-economic characteristics of communities with hazardous waste sites. New York: Public Data Access. Williams, Robert W. 1999. The contested terrain of environmental justice research: Community as a unit of analysis. The Social Science Journal 36:313–28. Zimmerman, Rae. 1994. Issues of classification in environmental equity: How we manage is how we measure. Fordham Urban Law Journal 21:633–70.

Correspondence: Department of Geography and Urban Studies, Temple University, 1115 W. Berks St., 309 Gladfelter Hall, Philadelphia, PA 19122, e-mail: [email protected](Mennis); Department of Geography, University of Colorado at Boulder, Boulder, CO 80309, e-mail: [email protected] (Jordan).