Spatiotemporal Perspectives on Air Pollution and Environmental Justice in Hamilton, Canada, 1985–1996

Abstract

This article addresses two questions: (1) How do spatiotemporal changes in air pollution levels—specifically, total suspended particulates (TSP)—rise or fall with socioeconomic status? (2) A critical equity interpretation of environmental policy then motivates this question: does the pursuit of average regional reductions in pollution benefit those who need improvements least, benefit those who need improvements most, or maintain the status quo? TSP data are drawn from networks of monitoring stations operated in 1985, 1990, and 1995. The monitoring data are interpolated with a kriging algorithm to produce estimates of likely pollution distribution throughout Hamilton. Exposure is related to socioeconomic status (SES) variables at the census tract level for corresponding years—1986, 1991, and 1996—and associations are tested with ordinary least squares (OLS) and spatial regression models. The results show that whether TSP rises or falls, injustice persists but becomes less pronounced over time. Among all SES indicators, dwelling value consistently predicts TSP levels for all years, suggestive of a land-rent/spatial-externalities dynamic. As we move forward in time, it becomes increasingly difficult to differentiate air-pollution exposure among Hamilton neighborhoods, as industrial TSP sources become more dispersed in the region and transportation pollution becomes relatively more important. We conjecture that more equitable distributions of air pollution have resulted more from post-Fordist industrial and spatial restructuring than from environmental policy intervention. Injustice in Hamilton and its apparent relationship with changing industrial structure appear similar to results in the United States and speak to a continental, intraurban environmental-justice experience.

Key Words:

• air pollution
• environmental justice
• GIS
• Hamilton
• kriging

Acknowledgments

The authors would like to acknowledge the support of the Toxic Substances Research Initiative, a joint program of Environmental and Health Canada. We would also like to thank Dr. Brian Klinkenberg, four anonymous reviewers, and Dr. Michael Goodchild for their comments. All remaining errors/omissions remain the authors'.

Notes

Source: Statistics Canada censuses of 1986, 1991, and 1996. For TSP estimates, see text.

^a Y ₁=a variable representing estimates of total suspended particulate matter. X ₁=average dwelling value. X ₂=a variable representing percentage of economic families or unattached individuals over 15 years of age below Statistics Canada's low-income cut-off. X ₃=percentage of population 15 years of age or older that has less than grade-nine education. X ₄=median household income. X ₅=unemployment rate among members of population who are 15 years of age or older. X ₆=percent of population employed in the manufacturing sector. X ₇=lone-parent families as a proportion of all families. X ₈=proportion of total income for the population 15 years of age or older, derived from government transfer payments.

^bCanadian dollars.

Note: For variable specification, see Table 1.

^aSee text for a discussion of the indexing of dwelling value and testing of pooled residuals in this model.

^bPseudo-R²

1. These years provided the largest air-monitoring networks close to the years of the censuses.

2. In 1995, one station (station 29017) produced a suspect spike in the pollution surface. Inspection of the site reveals that the particulate matter was likely the effluence of local fugitive sources, including an asphalt plant. This kind of source is unlikely to produce hazardous particulate matter, due to a high proportion of large particles that are not inhalable. Frank Dobroff, an air quality analyst with MOE, confirmed this and suggested the data from this station be discarded (personal communication via email 29 November/3 December 2001). The station was shut down by the MOE the next year.

3. We began with a larger set of variables, but multicollinearity was found in tolerance tests of all potential predictors. We thus proceeded with the set of covariates shown in Table 1.

4. All variables measured in dollars were indexed to 1986 constant dollars for the pooled regression analysis.

5. All SAR models employed a first-order adjacency matrix.

6. As mentioned earlier, several other variables were removed from testing because they were found to be collinear in tolerance tests. Many of the variables included in the analysis also show high intercorrelations, though the modeling procedures guard against this. The improvement to F-ration tests guard against inflated standard errors, a common symptom of collinear predictors. In addition, Mallow's Cp selects optimal models by including the strongest—and orthogonal—predictors.

7. Based on Statistics Canada's CANSIM GDP Implicit Price Index (SDDS 1902 STC 13-213) for the period 1981–2001, using 1986 as the base year.

8. The test for spatial autocorrelation in the pooled OLS model was based on the application of Moran's I to Σ(e ₁₉₈₆+e ₁₉₉₁+e ₁₉₉₆, i…j)/3, for each ith case and jth year, representing the average residual of the error term for each tract.

9. As noted earlier, sensitivity analyses were undertaken with the 1996 data. The first involved the use of the suspect air-monitoring station (see endnote 2). The analysis showed that dwelling value remains the only significant covariate with a coefficient of −6.39E-07 (t=−3.53, p<0.01), though model fit does diminish (R ²=0.11). Given the results of other models and qualitative input from the MOE, it was decided to leave station 29017 out of the final analysis. The second sensitivity analysis of the reported 1996 model removed the suburban census tracts in Hamilton where the air-monitoring network is sparse and the kriging predictions show a rise in TSP (and standard errors) where this is unlikely. As a result, the dwelling-value variable actually diminished in significance (b=−1.03e-06, t=−3.79) as compared with that reported in Table 3, suggesting that the reported model best captures disparities in TSP exposure.

10. Since the 1996 data form part of the pooled model, the sensitivity analysis using the suspect air-monitoring station was extended to this model. The analysis shows little change in the set of SES covariates, their relative significance, and their overall goodness of fit (adj. R ²=0.51, F=52.77, p<0.01). For the second sensitivity analysis, most suburban tracts had already been removed, since they were not commensurate between the comparison years. Removal of three other suburban tracts changed neither model fit (adj. R ²=0.52, F=53.55, p<0.01) nor the importance of the covariates.

11. The loess-smoothing term with a spatial span of 10 percent of observed points (centroids) slightly overfilters the spatial dependency of the covariates. Several other neighborhood spans were used for the loess term, but only a 10-percent span could remove spatial dependency. The resultant Moran's I is very small and negative (>−0.000), and the dwelling-value variable changes sign in the final model. But this is unlikely to affect the relative importance of the covariates.