Ecological Modernization or Aristocratic Conservation? Exploring the Impact of Affluence on Carbon Emissions at the Local Level

Abstract

Social scientists have debated how affluence impacts carbon emissions at the national level. We conduct an exploratory study at the subnational level to expose another dimension of the affluence–emissions debate. Based on the notion of aristocratic conservation, we hypothesize that affluence is positively related to carbon emissions from consumption activities but negatively related to emissions from production activities. We test these hypotheses using county-level data in the United States for the year 2002. A spatial regression analysis demonstrates that median household income is positively associated with consumption-based emissions; nevertheless, we find evidence of an environmental inequality Kuznets curve in the relationship between median household income and production-based emissions. This finding suggests that the wealthiest counties are able to displace certain types of emissions, specifically those related to energy and industrial production. We discuss the theoretical and political implications of these results.

Keywords:

• carbon emissions
• environmental inequality
• environmental Kuznets curve
• growth machine
• spatial regression

Notes

^a Gurney et al. (2009).

^b U.S. Census Bureau (2010).

^c USDA Economic Research Service (2013).

*Pseudo p ≤ .001 based on 999 permutations.

Note. Standard errors in parentheses. All variables except the metropolitan dummy variable are natural log transformations.

^a Pseudo-R-squared is reported for the spatial regressions; this is not directly comparable with the R-squared reported for OLS.

**p ≤ .01.

Note. Standard errors in parentheses.

^a Pseudo-R-squared is reported for the spatial regressions; this is not directly comparable with the R-squared reported for OLS.

*p ≤ .05, **p ≤ .01.

Using stepwise regression, Luna (2008) did not find a significant slope estimate for household income. We caution against the use of stepwise regression. See York and Rosa (2005) for a discussion about how this procedure can produce misleading results in the case of modeling environmental outcomes.

We note that much of the EKC literature (for a review see Dinda 2004) does not address the legacy of Kuznets's (1955) original research on inequality and development.

In analyses not reported, we also include a control for the percent of economic output that is from the industrial sector. Including this control does not change the slope estimates of our models.

We also explored the potential presence of “spatial heterogeneity” by mapping local indicators of spatial autocorrelation. While the Mountain West contained a large cluster of high values for consumption emissions per capita, these counties are primarily rural areas. The source of variation, thus, is likely captured by the measures of urbanization, population density, and metropolitan designation. In other words, exploratory spatial data analysis based on local indicators of spatial autocorrelation did not reveal clear “spatial regimes” within the continental United States that would clearly suggest taking steps to allow for structural differences in regression relationships across different regions. This issue is further discussed later in conjunction with regression diagnostics.

Both Breusch–Pagan and Koenker–Basset tests for heteroskedasticity were significant.

That all variables have been transformed into their natural logarithms facilitates interpretation of slope estimates, where the coefficient represents the percent change in the dependent variable for a 1% change in the independent variable, holding the rest of the model constant.

The coefficient estimates for classic OLS and the spatial lag model were generally similar, and overall model fit improved slightly with the inclusion of the spatial lag term. Since the traditional R-squared is not appropriate for a spatial regression model, fit is measured through log-likelihood, the Akaike information criterion, and the Schwarz criterion. All three indicated better fit in the spatial regression.

When using a weights matrix based on second-order contiguity, the spatial error model was actually revealed to be the appropriate alternative, based on a significant robust LM-error statistic. While this did not substantively alter the interpretation of the structural variables, it is notable that the spatial “effect” disappeared when expanding the relationship beyond immediate neighbors. This may suggest that the cross-border spatial “diffusion” process apparent for immediate neighbors evaporates when expanding the scale spatial relationship rule, leaving only spatial “disturbance” in the residuals.

Overall model fit also improved with the inclusion of the spatial error term. Log-likelihood, the Akaike information criterion, and the Schwarz criterion indicated that the spatial regression provides better model fit than classic OLS. As with Model 1, the signs and levels of significance generally remained similar across the classic OLS and spatial error models for consumption emissions.

We note that without the squared term of household income the maximum variance inflation factor (VIF) is 2.13 in our models. This indicates that multicollinearity is likely not a concern in establishing significance.