Abstract
This research explores relationships between the spatial distribution of industrial facilities that release lead and lead compounds, community characteristics, and levels of violent and property crime in Hillsborough County, Florida. The spatial distribution of selected Toxics Release Inventory (TRI) facilities was modeled using Getis-Ord Gi* to classify census tracts into three groups: (1) tracts with statistically significant spatial clusters of TRIs; (2) tracts where there was a non-random-absence of TRIs; and (3) tracts with a random distribution of TRIs. Results of one-way ANOVA comparing demographics of clustered locations of TRIs to random and dispersed areas found significant differences, but local prediction models from geographically weighted regression (GWR) were no more useful in understanding TRI concentrations in areas where they are more clustered than in random or dispersed areas. GWR models predicting property crime were improved when TRI concentration was used in random and dispersed areas, but TRI concentration did not improve models predicting violent crime rates.
Notes
1In environmental justice research, exposure to risk is commonly operationalized as the location or density of toxic sites, the number of toxic releases, or the amount of toxins released. The approach in the current study offers an alternative by examining the spatial distribution of facilities that release lead and lead compounds.
2For smoother presentation, future references to the locations of Toxic Release Inventory (TRI) facilities within Hillsborough County, Florida, that reported releases in excess of the thresholds for total on-site disposal or other releases of lead and lead compounds for 2008 will be referred to as “selected TRI facilities.”
3These data provide five-year average estimates of focal variables, which more accurately reflect Hillsborough County's population in 2008 (i.e., the year for which crime data were obtained) than data produced from either the 2000 or 2010 decennial census.
4In addition to the more traditional county demographics that have been used by environmental justice researchers, we also included measures of social disorganization (Sampson and Groves Citation1989; Shaw and McKay Citation1942) in our analysis. It is possible that these facilities disrupt the social cohesion of the local neighborhoods, resulting in higher crime rates and other social problems.
5Data from Plant City (population of approximately 34,000) and Temple Terrace (population of approximately 25,000) were requested, but were not made available to include in the current investigation.
6Getis-Ord Gi* was calculated with a 9,860 meter threshold, using the fixed distance band spatial conceptualization.
7Our approach to classification TRI facilities is consistent with other criminological studies that examine the spatial distribution of risk, most notably the techniques used in Risk Terrain Modeling (see, for example, Caplan Citation2011; Caplan and Kennedy Citation2010; and Caplan, Kennedy, and Miller Citation2011).
Note. Post hoc comparisons used Fisher's least significant difference (LSD) test.
8The measure of spatial concentration of TRI locations used in the prediction models that follow is based on the observed z-scores produced by the Getis-Ord Gi* analysis, which is described in the previous section. Z-scores for the 318 tracts range from −2.10 to 3.80 (M = 0.24; SD = 1.09).
9Research shows that GWR is an appropriate technique for exploring spatial nonstationarity (Brunsdon, Fotheringham, and Charlton Citation1996). However, GWR does not account for spatially autocorrelated residuals that are often produced by this technique (Leung, Mei, and Zhang Citation2000). Although spatially autocorrelated residuals caused by spatial dependency (e.g., when error terms across spatial units are correlated or when the dependent variable at one place is affected by the independent variables in that same location and by the independent variables at a proximate location) can be modeled (for a detailed discussion, see Getis Citation2005), the current study uses GWR to model spatially heterogeneous relationships among variables.
*p < .05; **p < .01.
10Exploratory regression was also performed in ArcGIS. Results suggested that in addition to the demographic factors we already included, our model would be improved by adding the percent white variable. However, GWR was unable to produce local estimates for a model containing the percent white variable due to localized dependency with other predictors (i.e., percent black). Therefore, the percent white variable was excluded from our final OLS/GWR models predicting the spatial concentration of TRI facilities.
Note. Predictor variables include percent Black, percent Hispanic, and percent with a bachelor's degree (AIC = 433.00; Adjusted R 2 = .8857). Results based on an adaptive kernel and AICc bandwidth that utilized 24 neighbors.
11An adaptive kernel was used for each GWR model; and the corrected Akaike Information Criterion (AICc), which corrects for small sample sizes, was selected as the bandwidth method.
*p < .05; **p < .01.
Note. Other predictor variables in both models include percent Black, percent Hispanic, percent with a bachelor's degree. For the GWR model including the TRI predictor, AIC = 2983.00 and Adjusted R 2 = .6277. For the GWR model without the TRI predictor, AIC = 2958.95 and Adjusted R 2 = .7044. Results for Model 1 and Model 2 are based on an adaptive kernel and AICc bandwidth that utilized 49 and 29 neighbors, respectively.
*p < .05; **p < .01.
Note. Other predictor variables in both models include percent Black, percent Hispanic, percent with a bachelor's degree. For the GWR model including the TRI predictor, AIC = 2999.72.00 and Adjusted R 2 = .4448. For the GWR model without the TRI predictor, AIC = 3008.50 and Adjusted R 2 = .4290. Results for Model 1 and Model 2 are based on an adaptive kernel and AICc bandwidth that utilized 64 and 54 neighbors, respectively.