A dynamic state-level analysis of carbon dioxide emissions in the United States

Energy Policy 59 (2013) 931–937

Contents lists available at SciVerse ScienceDirect

Energy Policy journal homepage: www.elsevier.com/locate/enpol

Communication

A dynamic state-level analysis of carbon dioxide emissions in the United States Travis Roach n Department of Economics, Texas Tech University, PO Box 41014, Lubbock, TX 79409-1014, United States

H I G H L I G H T S







State-level analysis of carbon dioxide emissions. Dynamic panel estimation to account for time series properties. Feasible environmental Kuznets curve for carbon dioxide emissions. Implications for state environmental policy discussed.

art ic l e i nf o

a b s t r a c t

Article history: Received 29 October 2012 Accepted 11 April 2013 Available online 14 May 2013

As climate change and the regulation of carbon dioxide emissions play an increasingly important role in the global policy debate, careful consideration of the state-level determinants driving emissions must be considered. The importance of state-level determinants in the transmission of carbon dioxide matters especially for a country that differs from coast to coast in energy use and industry makeup like the United States. To add to the policy debate this paper estimates two models that account for the dynamic nature of emissions of carbon dioxide emissions at the state-level from 1980–2010 while taking account of scale, technique, and composition effects. When stochastic trends are taken account of, an environmental Kuznets curve relationship with a feasible turning point is found for carbon dioxide emissions. Published by Elsevier Ltd.

Keywords: Carbon dioxide Dynamic panel estimation State-level analysis

1. Introduction The regulation of carbon dioxide (CO2) and other greenhouse gases has come to the forefront of policy debates taking place on the international and state-level alike. Although there is a robust literature on CO2 emissions from an international perspective, very few authors have considered CO2 emissions at a sub-national level for the United States. Many states are in fact larger than some developed countries and thus produce more emissions through the scale effect alone. There are many underlying factors in aggregate reported data that matter when it comes to CO2 emissions. As Stanton et al. (2010) point out, if per capita CO2 emissions were equal across the United States to that of California, global emissions would fall by 8%. If per capita emissions were equal to Texas, however, global emissions would increase by 7%. Furthermore, state-level analyses of emissions are important because many states have begun to pursue more stringent environmental standards unilaterally than they are required by Federal law (Prasad and Munch, 2012). As a policy tool, some states have

n

Tel.: +1 817 673 2662; fax: +1 806 742 1137. E-mail address: [email protected]

0301-4215/$ - see front matter Published by Elsevier Ltd. http://dx.doi.org/10.1016/j.enpol.2013.04.029

implemented renewable portfolio standards that dictate the amount of electricity generated by renewable means. More stringent policies have been adapted, though, such as the cap-andtrade program for CO2 emissions in California and the Regional Greenhouse Gas Initiative (RGGI). RGGI (2012) is in fact a perfect example of states going beyond Federal regulations to limit CO2 emissions because it is a coalition of nine Northeast and MidAtlantic states that are trying to reduce carbon emissions from electricity by 10% in 2018. This paper adds to the diminutive literature on state CO2 emissions by estimating a dynamic state-level model of the determinants of CO2 emissions while accounting for effects due to the scale, technique, and composition of economic activity that stem from the literature on free trade and the environment. Understanding the determinants of CO2 emissions from a statelevel is certainly important in the crafting of future legislation, and reconsideration of current policies. This paper also contributes to the wider literature on estimating carbon dioxide emissions because it explicitly takes into account the time series properties of emissions, which are seldom addressed. Continuing on, Section 2 briefly discusses the scale, composition, and technique effects that will be used to add structure to the model as well as discusses papers that specifically account for the time-series properties of

932

T. Roach / Energy Policy 59 (2013) 931–937

Fig. 1. Map of subsample states.

emissions. Section 2 then concludes with a discussion of works that look at emissions from a state-level perspective. Section 3 develops the dynamic panel model. Section 4 concludes with a discussion of policy implications that stem naturally from the model estimates Fig. 1.

2. Background The environmental impact of trade liberalization and economic growth in general has been at the vanguard of attention in the environmental economic literature for quite some time and has led to many theoretical and empirical mainstays that are useful for analyzing regional effects of economic growth on the environment. Despite the robust literature on the income–environment relationship at international and country levels, and the variance in state and federal policies regarding CO2 emissions, very little has been written on state-level CO2 emissions. For the sake of brevity, then, only a few papers that use national or international data will be discussed in this section. In order to motivate and add structure to the empirical application that follows elements from the free trade and the environment will be discussed here in modest detail.1 The net effect that economic growth has with regard to the environment can be decomposed into three separate effects: the composition, technique, and scale effects (Copeland and Taylor, 2003; Antweiler, et al., 2001; Managi et al., 2009). The composition effect demonstrates how emissions are affected by changes in the composition of output in an economy. The composition of output would likely change due to the degree of trade liberalization in a country and its relative comparative advantages compared to other countries. Naturally, the composition effect is country or state specific and can lead to a net increase or decrease in emissions. In their seminal work on the environmental Kuznets curve (discussed below), Grossman and Krueger (1991) estimate that the composition effect is what drives lower pollution levels in Mexico. They argue that instead of Mexico being turned into a 1 For a more comprehensive review of the literature on free trade and the environment Frankel (2009) and Copeland and Taylor (2003) are both excellent sources.

pollution haven, higher demand for environmental quality from the United States will mitigate this worry. The technique effect measures how growing incomes change the intensity of emissions. Typically, rising income will result in a decrease in emissions because more environmentally sound production methods are demanded. Thus, as trade and income growth gives consumers a greater variety of goods to consume, this will afford countries the opportunity to attain higher levels of welfare, for a given level of domestic output, which will in turn increase the demand for better environmental quality (Frankel, 2003). The technique effect itself has fostered a robust body of research on what is commonly known as the environmental Kuznets curve (EKC).2 The hypothesis behind the EKC is that pollution transitions from being a normal good to being an inferior good as income increases. That is to say, carbon dioxide emissions follow an “inverted-U” shape: increasing as income increases for a time, and eventually falling with increases in income after a threshold level is reached. It is interesting to note that the EKC does not seem to hold in the case of global pollutants like CO2 (Dinda, 2004; Müller-Fürstenberger and Wagner, 2007). Many authors find that CO2 emissions do not follow the inverted parabolic trend that can be found in pollutants like ozone and sulfur dioxide (Cole et al., 1997; Frankel, 2003; Kim, 2013). Vollebergh et al. (2009) suggest, though, that the lack of robustness for EKC findings may be due to under-identification. Luzzati and Orsini (2009) suggest that differences in EKC findings can be attributed to including outliers, using measures of per capita CO2 instead of CO2 in levels, and including time trends. The present study uses CO2 emissions in (log) levels because it is better justified theoretically (Luzzati and Orsini, 2009). Carson (2010) suggests that instead of income being the causal agent for increasing environmental quality it is instead good government regulations which are correlated with higher incomes that motivate emissions reductions. The scale effect measures the basic effect that an increase in production, or GDP, has on emissions. According to theory, if the scale effect is positive then we can conclude that economic growth is a driver of pollution. There is also the so called “scale-technique effect” which measures the combined scale and technique effects

2

A thorough background on the EKC can be found in Dinda (2004).

T. Roach / Energy Policy 59 (2013) 931–937

(e.g., GDP per capita). Managi et al. (2009) and Frankel (2003) estimate the scale-technique effect and composition effect at the national level and find the scale-technique effect to be positive in the case of CO2 emissions. While an appropriate theoretical background is extremely important in selecting variables to describe the income–environment relationship, so too are the intrinsic properties of the variables that are selected. Recent literature has highlighted the importance of accounting for the time-series properties of emissions as a cause for bias in the estimation process (Perman and Stern, 2003; Aldy, 2005; Müller-Fürstenberger and Wagner, 2007; Volleberg et al., 2009; Galeotti et al., 2009; Piaggio and Padilla, 2012). Piaggio and Padilla (2012) offer a particularly good account of the history of controlling for non-stationarity when estimating carbon dioxide emissions. In addition to non-stationary CO2 data, and adding insult to injury as it were, worries over stochastic trends in the independent variable, real disposable income per capita, have been raised as well with regard to spurious regression results. Many studies that explicitly model the time-series dimension look to cointegration and Granger causality tests to describe the relationship between income and emissions (Dinda and Coondoo, 2006, Soytas et al., 2007, Piaggio and Padilla, 2012). In fact, Soytas et al. (2007) find a relationship between energy consumption and CO2 emissions, but not one between income and CO2 emissions in the United States. For this reason energy consumption in the industrial and transportation sectors are used as explanatory variables in the empirical analysis that follows. 2.1. State-level analyses As previously mentioned, there are very few works that estimate the determinants of CO2 emissions at the state-level. In an early look at carbon dioxide emissions and the environmental Kuznets curve, Carson et al. (1997) find support for the EKC hypothesis in multiple pollutants at the state-level. In a similar study that uses data on sulfur dioxide and nitrogen oxide emissions, List and Gallet (1999) find that although there is evidence in favor of an EKC relationship, restricting cross-sections to a common estimate may be providing biased results. Their work is important in addressing the fact that while some states may be at similar levels of development, state-level heterogeneity is an important factor to account for. Typically state-specific differences are accounted for by estimating fixed effects models, but the dynamic panel model that follows accounts for state-level heterogeneity by removing these fixed effects through differencing. In another early study, Casler and Rose (1998) approached the problem of state-level carbon dioxide emissions with the use of an input-output structural decomposition. These authors find that CO2 emissions due to economic growth have been offset by changes in the mix of final demand and interfuel substitution. Both Jiusto (2008) and Prasad and Munch (2012) take note of the aforementioned differences in regional or state climate initiatives. Jiusto (2008) expands on these discrepancies in environmental policies and provides a framework from which to analyze carbon emissions at the state-level. Prasad and Munch (2012), however, dive more deeply into the differences in policies across states and estimate the reductions in CO2 emissions that one can expect based on whether or not a state has a policy in place, and the tenure of that policy. The authors' main conclusion from this study is that a public benefits fund is associated with large decreases in carbon emissions. In perhaps the most well-known account of state CO2 emissions, Aldy (2005) considers state-level emissions by constructing his own index of CO2 emissions for the time period 1960–1999. Aldy separates his analysis by estimating separate consumption and production based EKCs finding that consumption based EKCs peak at much higher levels than production based EKCs. At the conclusion of his paper, however, Aldy (2005) notes that,

933

“failure to account for potential stochastic trends in the data may yield misleading results. Continued work applying time-series methods to evaluate the robustness of estimated EKCs would be worthwhile”. Davidsdottir and Fisher (2011) determine the role that renewable energy can have on the economy by using a novel index of “carbon emissions economic intensity” at the state-level. While their work considers the possibility of spurious results, the dependent variable that they use is an index of emissions which varies from the present paper.

3. Empirical model 3.1. Data To appropriately estimate the determinants of carbon dioxide emissions in the United States many factors must be accounted for. In order to add structure to the model, this study incorporates variables that have been shown to be important in measuring the different effects economic growth has on the environment: the scale, technique, and composition effects (Copeland and Taylor, 2003; Frankel, 2003). To account for the scale effect, real disposable personal-income (RDPI) per capita and its square are used as independent variables. Typically the capital-to-labor ratio has been used to measure the composition effect. This paper, however, deviates from this standard by focusing on the energy demand in typically “dirty” sectors of the economy by including data (in logarithm) on energy demand from the industrial and transportation sectors to proxy for the composition of economic activity. The inclusion of energy demand variables is also in line with the Granger-causality findings of Soytas et al. (2007). Along with the scale and composition effects that drive CO2 emissions, the burning of coal has been shown to be an important determinant (Aldy, 2005; Jiusto, 2008). Indeed, industrial energy demand may be particularly high in a state, but if the majority of that energy comes from relatively clean sources the impact on CO2 emissions will be lower than that of a state that uses mostly coal for power. For this reason, two variables measuring a state's reliance on coal are used to represent the technique effect: Coal intensity is a binary variable that measures whether or not the majority of emissions in a state come from coal, as opposed to petroleum or natural gas combustion; coal share measures the percentage of a state's CO2 emissions that come from coal sources. In fact, these variables capture an interesting amount of variation in the data because many states change over time in whether or not they are coal-intensive. Too, these variables serve to partially explain changes in pollution at the intensive margin. All data is collected for the 50 U. S. states from 1980–2010 forming a balanced panel. Data on state CO2 emissions has been collected from the Energy Information Administration, data on disposable income per capita was collected from the Bureau of Economic Analysis and transformed to real values using CPI data by region from the Bureau of Labor Statistics, energy demand data comes from the U.S. Energy Information Administration's state energy data system. A more detailed description of these variables, and their descriptive statistics, can be found below in Table 1.3 3.2. Unit-root and cointegration tests It is important to test for a unit root in both CO2 and the independent variables. To do so, the following simple panel model 3 It would seem that the energy demand variables may be too closely related to the scale of economic activity, but they are only moderately correlated with RDPI per capita. Correlations between all variables are available upon request.

934

T. Roach / Energy Policy 59 (2013) 931–937

tests when the time series dimension is low. As Table 3 shows, we fail to reject the hypothesis of no cointegration with each test.

is considered yit ¼ ρyi;t−1 þ Z′it γ i þ εit

ð1Þ

where yit is the variable of concern, and Z′it γ i represents panelspecific means. The typical hypothesis being tested is whether or not ρi ¼ 1 for all states, i. Müller-Fürstenberger and Wagner (2007) considers this a “first-generation” type test because it does not take into account the heterogeneous nature of the data. Breuer et al. (2001) also take note of this issue and show that misleading results may be obtained when cross-sectional heterogeneity is not allowed. Still, though, a “first-generation” Harris and Tzavalis (1999) test is estimated for explorative purposes (HT test here forward). This test is the most appropriate for the data at hand, compared to a Levin–Lin–Chu test, due to the low time-series dimension of only 30 periods of observation for 50 states (i.e. because there are more cross-sectional units than time periods). It must be noted, however, that the HT test is asymptotically efficient as N-∞ which is not feasible for the United States, but can be justified if one considers adding other countries to the data. The results of the HT test, which can be found in Table 2, should be interpreted as follows—a high p-value fails to reject the hypothesis that the panel contains unit roots. For all of the variables with the exception of coal share we fail to reject the null hypothesis that the panels contain unit roots. Another method of computing unit root tests in panel data is to allow ρi to differ across cross sectional units. This method has been developed using an augmented Dickey–Fuller test by Choi (2001, 2006) and results in four possible test statistics. For this paper, only the inverse normal Z statistic will be used because it offers the best size and power tradeoff. The Choi specification is recommended more highly by Müller-Fürstenberger and Wagner (2007) as it is part of the “second generation” of test statistics that account for heterogeneity between cross-sectional units. The Choi test accommodates state heterogeneity by calculating individual augmented D–F tests on each cross sectional unit. The results of the Choi D–F test should be interpreted as follows—a low p-value supports the hypothesis that at least one panel is stationary. According to the two panel unit-root tests performed here, for both GDP per capita and CO2 we can say with confidence that every cross sectional unit is non-stationary, but the same cannot be said for energy demand in the industrial and transportation sectors. Because RDPI and CO2 are non-stationary, the presence of a cointegrating relationship must be examined in order to properly specify the empirical model that follows. To test for a cointegrating relationship in RDPI and CO2, the error-correction-based cointegration tests of Westerlund (2007) are implemented. These tests offer many appealing properties in that “each test is able to accommodate individual-specific short-run dynamics, including serially correlated error terms, non-strictly exogenous regressors, individual specific intercept and trend terms, and individualspecific slope parameters” (Westerlund, 2007). Moreover, the Westerlund tests are well suited for the data in this paper because they are more powerful than residual-based panel cointegration

3.3. Estimating the determinants of CO2 emissions In order to calculate CO2 emissions while taking account of the heterogeneous auto-regressive nature of the data, the following dynamic linear regression model is proposed yit ¼ αyi;t−1 þ x′it β þ ui þ ∈it

ð2Þ

where yit is CO2 emissions for state i at time t, and xit is a vector of explanatory variables that account for the scale, technique, and composition effects. The error terms, ui and ∈it represent the typical error term decomposition that includes unobserved individual heterogeneity. The model will be estimated in two ways to account for autoregressive processes: Arellano and Bond's system GMM estimator, and fixed effects estimation assuming an autoregressive pattern is present in the error term. The fixed effects estimation is also used here to offer another specification that allows for cross-sectional heterogeneity. As will be seen, the model estimates are similar in interpretation between the two specifications, though the marginal effect of each determinant is lower in the fixed effects representation. In the dynamic panel model differencing removes any fixed individual effects and corrects for spurious regressions due to non-stationary data. This can be seen in the first-step of the estimation, Eq. (3). Δyit ¼ αΔyi;t−1 þ Δx′it β þ Δ∈it

ð3Þ

In the fixed effects model with autoregressive errors, the error term is assumed to take the form. ∈it ¼ ρ∈i;t−1 þ υit

ð4Þ

Table 2 Panel unit-root tests. Variable

p-Value (HT)

p-Value (Choi)

RDPI CO2 Industry Transportation Coal share

1 0.6240 0.7180 0.9932 0

1 0.4705 0.0399 0 0

Notes: Panel-specific means included, p-values shown.

Table 3 Westerlund cointegration tests. Statistic

Value

p-value

Gt Gα Pt Pα

−2.363 −9.469 −15.334 −8.999

0.511 0.996 0.325 0.470

Notes: Null-hypothesis is of no cointegration.

Table 1 Summary statistics. Variable

Mean

Std. dev.

Min.

Max

Description

CO2 RDPI Industry Transp. Coal Int. Coal share

4.23 13,394.74 12.63 12.30 .49 34.48

.984 2,512.38 1.15 .97 .50 21.74

1.41 7,695.97 9.49 9.85 0  .19

6.57 20,919.21 15.60 14.86 1 82.19783

Log of carbon dioxide emissions from energy Per capita RDPI in 1982-84 dollars Log of industrial energy consumption measured in BTUs Log energy consumption in the transportation sector in BTUs Dummy variable: 1 if state is coal-intensive, 0 otherwise % of emissions derived from coal

Notes: 15,500 observations: annual data on 50 states from 1980–2010. Negative value for the minimum of coal share is a result of EIA net calculations, see EIA (2013).

T. Roach / Energy Policy 59 (2013) 931–937

935

where υit is an independently and identically distributed error term with mean zero and variance s2υ , and ρ is estimated and accounted for. Many authors have shown that the Arellano and Bond generalized method of moments (GMM) estimation technique corrects for bias found in OLS and 2SLS models (Baltagi and Kao, 2000). Estimation bias is corrected by solving for a set of exogenous instrumental variables which are correlated with the independent variables, but orthogonal to the errors, from the moment conditions. For dynamic panel models, this is achieved by exploiting the time series nature of the data and obtaining instruments from lagged values. As stated above, state-level heterogenetiy is of concern since there are natural differences among the states in this study, e.g. carbon-intensity or efficiency of the electricity sector, proximity of the state to sources of combustible fuels or viability of renewable electricity. These differences may be time invariant or random in nature which traditionally leads to fixed or random effects estimation, though given that the path of each state is the object of this study, a random effects model would not be conceptually correct. In the dynamic panel model, individual effects are accounted for through the method of differencing (Arellano and Bond, 1991, 1995; Blundell and Bond, 1998; Blundell et al., 2000). Among the GMM estimation techniques, the system GMM estimator (Arellano and Bover, 1995; Blundell and Bond, 1998) has been proven less biased in small sample scenarios than the first differencing GMM estimator proposed by Arellano and Bond (Hayakawa, 2007; Soto, 2009). Hayakawa shows that when an AR (1) panel data model is estimated, the system GMM estimator is least biased. Roodman (2008) shows, however, that the when the instruments obtained in the difference GMM estimates are not weak that the additional instruments obtained using system GMM estimation may over-specify the model. Moreover, the system GMM estimator may have a larger absolute bias than the difference GMM estimator when the dependent variables show a fair amount of persistence, as is the case with CO2 emissions (Bun and Windmeijer, 2010). Another source of bias in the system GMM estimator can be detected as the coefficient on the lagged dependent variable approaches one. Each of the previously mentioned potential downfalls of the system estimator were found to be problems for the system GMM estimates. Hence, the difference estimator is the preferred model specification and is the only GMM model presented in this paper. Blundell et al. (2000) show that the differenced estimator performs poorly as α-1, and Blundell and Bond (1998) show that in a first differenced model with weak instruments the estimates will be biased towards the fixed effects estimator. Neither of these problems occurs in the differenced model presented below.

Sargan test shows, the difference GMM specifications fail to reject the hypothesis that the over-identifying restrictions are valid. Hence, the dynamic panel models are correctly specified and thoroughly account for the time-series properties of the data that many previous studies have dismissed. The estimates for the difference GMM models show that there is a positive scale-technique effect (RDPI per capita) in certain income ranges for both model specifications. Specifically, the data shows that CO2 emissions do in fact follow the EKC type trend of an inverted U. This is important to note because many authors reject the EKC hypothesis for CO2. The estimated “turning point” in the EKC occurs at about $16,134 (in 1982-4 dollars) in the first model, and $17,092 in the second model. These values are relatively high for the sample period because they are more than one standard deviation away from the mean, but are ultimately feasible because there are states that have surpassed both levels. For the fixed effects specification the implied EKC turning point is $16,372 and $16,577, respectively, which follows the same pattern as the GMM specification in that the implied turning point is higher when coal share is considered as an independent variable. As one might expect, states that are relatively coal intensive in their energy use are expected to emit 8.7% more emissions on average. From the fixed effects model a 2.7% increase is expected. This shows that policies aimed at changing a state's energy source portfolio could lead to significant declines in CO2 emissions, all else equal. There are quite a few states that have already pursued such policies, many of which have passed the estimated turning point in the EKC. Similarly, when there is an increase in the share of coal in total CO2 emissions we can expect overall CO2 emissions to increase, all else equal. From the GMM specification, a 10% increase in the share of coal is expected to result in a 7% increase in overall emissions. Increases in energy demand from the industrial sector will also increase CO2 emissions, all else equal. Using the first dynamic panel specification, in the industrial sector it is estimated that a 10% increase in energy use will result in about a 3.9% increase in CO2 emissions. This finding is in fact significant at the 6% level. For both specifications, we cannot say with traditional statistical confidence that changes in transportation sector energy demand will result in changes in CO2 emissions. This finding may seem contrary to expectations, but it is robust to multiple model specifications including models that interact energy use variables with the coal intensity of a state using both dynamic panel and fixed effects estimation. Potential reasons for this finding include increased gas mileage standards and inter-fuel substitution that the data at hand cannot account for.

3.4. Estimation results

4. Conclusions and discussion

The empirical results presented in Table 4 shows both methods of estimating the determinants of CO2 emissions. The first two regressions are estimated by the two-step difference GMM estimator. The latter two regressions show the estimates from the fixed effects model with auto-regressive errors. For consistent estimation in the GMM models the error term, ∈it , must be serially uncorrelated. The Arellano–Bond test for zero autocorrelation in first-differenced errors shows this is not the case for 2nd order autocorrelation. In other words, we fail to reject the null hypothesis of no autocorrelation. Arellano and Bond (1991) note that the estimates would only be inconsistent if 2nd order serial correlation was present. Hence, the estimates from the GMM estimator are consistent because serial correlation in the error term is not present. Another diagnostic that must be checked is whether or not the instruments obtained are weak. As the

This paper has furthered research on the relationship between economic growth and the environment by focusing on how CO2 emissions are related to per capita income changes and energy demand changes at the state-level for the United States. Moreover, this paper has accounted for the time-series properties of CO2 emissions that are seldom addressed by estimating both a dynamic panel model, and a fixed effects model with autoregressive errors. Another way in which this paper has advanced research on CO2 emissions is by using a sub-national panel because much of the research thus far utilizes country-level data. By focusing on CO2 emissions at the state-level this study is able to account for the vast differences among states in resource endowments, composition of economic activity, and societal preferences which are important with regard to policy decisions at the state-level.

936

T. Roach / Energy Policy 59 (2013) 931–937

Table 4 Regression estimates. Variable

Difference-GMM (1)

CO2

.5968025 .0474479 .0001323 .0000227 −4.10e-09 7.62e-10 .0389944 .0208187 −.0167681 .0255807 .0878761 .0378957

t-1

RDPI RDPI2 Industry Transportation Coal int. Coal share Obs EKC Within R2 Instruments Sargan AB(1) AB(2)

– – 1450 $16,134.15 – 63 .7316 .00 .8036

Difference-GMM (2) nnn

nnn

nnn

n

nn

Fixed effects (3)

.528923 .1216889 .0001217 .0000511 −3.56e-09 1.70e-09 .0244125 .0467485 −.0118124 .0416993 – –

nnn

.0075772 .0040751 1450 $17,092.69 – 92 .9994 .001 .995

n

nn

nn

.7595196 .0159384 .0000501 6.36e-06 −1.53e-09 2.10e-10 .0237869 .0079946 .0176319 .0112668 .0273827 .0076497 – – 1450 $16,372.55 .8792 – – – –

Fixed effects (4) nnn

nnn

nnn

nnn

nnn

.750329 .0158741 .0000494 6.34e-06 −1.49e-09 2.09e-10 .0258777 .0079588 .0183855 .0112087 – –

nnn

.0026509 .0003727 1450 $16,577.18 .8801 – – – –

nnn

nnn

nnn

nnn

Notes: p-values shown for AB test for autocorrelation and Sargan test for over-identification; clustered robust standard errors shown below estimates; statistical significance at the 10, 5, and 1% level marked n,nn,nnn.

The results from this paper can be interpreted in many ways and have very important policy implications. First, consider the implied turning point of $16,134 (again, in 1982-4 dollars) for the EKC found in the first GMM specification. Twenty-eight states within the sample surpass this point estimate in at least one period between 1980 and 2010. For the EKC turning point of $17,092 found in the second GMM specification, 21 states pass this threshold in at least one period.4 Many states only reached this level at or near the end of the period, or only surpassed this level for 1–2 periods which includes observations before and after the great recession which began in the fourth quarter of 2007. Because it is unlikely that climate initiatives were discussed in earnest during the wake of the great recession, consider a subsample of states including only those that surpassed the EKC turning point for at least three consecutive periods. Of the 16 states included in this subsample, eight have already passed substantial legislation limiting carbon dioxide emissions: California, and seven of the ten states participating in the Regional Greenhouse Gas Initiative. As a clarification note, New Jersey was still a part of the RGGI in the sample period studied here, but left the coalition in 2011. Other states in this subsample have also pursued climate initiatives, though they have not been as far reaching as that of California or the RGGI. While this is anecdotal evidence, it seems that the turning point of $17,092 is reasonable considering that these states have begun to take action to reduce CO2 emissions. A limitation of the model must be noted here, though, in that the implied turning may not be exactly the same across all states, and may not exist at all for some states due to path heterogeneity. Evidence of differing EKC turning points is discussed thoroughly at the national level in Piaggio and Padilla (2012), and the same rationale can be extended to the present study. The potentiality of differing EKC turning points can be seen anecdotally in a number of ways; first by the amount of states that have reached this level and not crafted legislation, second by the states who have joined the RGGI but not surpassed the estimated

4 States with RDPI per capita above $17,092 at some point within the sample period, subsample states marked by an asterisk: AK, CAn, COn, CTn, DEn, FLn, ILn, KS, MDn, MAn, MNn, NE, NHn, NJn, NYn, NDn, SD, TX, VAn, WAn, WYn.

EKC threshold, and finally by the recent exit of states within the subsample from coalition climate initiatives: Washington and Wyoming from the Western Climate Initiative, and New Jersey from the RGGI. Further, when a 95% confidence interval is formed around the $17,092 EKC estimate the upper bound is $21,698, a level which is not reached by any state during the sample studied here. Although much effort has been taken to allow for state-level heterogeneity, it should not be a policy aim to reach and surpass this specific level of per capita income. This is especially true because for the most part, any efforts to increase per capita income will increase CO2 emissions in the short-run. Policy makers should consider this when crafting legislation and perhaps consider the early adoption of CO2 reduction mechanisms to “lower and flatten” their EKC (Dasgupta et al., 2002). The estimates of the composition effect show that there are many ways in which CO2 emissions can be reduced. Moreover, the estimates from the first regression show that emissions will be reduced by a larger amount if efforts are taken to decrease energy demand in the industrial sector than in the transportation sector, all else equal. Achieving reductions in energy demand for either sector may be easier said than done in some states, though, depending on regional differences and the political landscape. Also, because the United States differs so greatly in terms of industrial makeup and societal preferences a single policy recommendation would not be warranted here. There are many ways in which to reduce energy demand in these sectors such as: providing subsidies for new and more efficient technology, taxes on energy use beyond a specified level, tradable permits, or any market-based incentive scheme. Another effort that would be beneficial in reducing CO2 emissions, and relatively easy to implement from a policy perspective, is to target cleaner burning fuels in energy production as the estimates for whether or not a state is coal intensive and the share of coal in CO2 emissions indicate. Narrowing the focus of policy to drive changes in pollution at the intensive margin, e.g. the polluting source rather than the extensive margin, e.g. overall pollution levels, may in fact serve to reduce pollution on the whole. Indeed, energy portfolio standards are an interesting way in which to reduce emissions that many states have already adopted. Further state-level research in the vein of Prasad and Munch (2012) that

T. Roach / Energy Policy 59 (2013) 931–937

focuses on specific industries ought to be pursued to address the efficacy of such policies.

Acknowledgments I am thankful for the many thoughtful and helpful comments of two anonymous referees, Dr. Timothy Hubbard, and session participants of the Western Economic Association 87th annual conference. Any remaining mistakes or omissions are naturally my own. References Aldy, J., 2005. An environmental Kuznets curve analysis of U.S. state-level carbon dioxide emissions. Environment & Development 14 (1), 48–72. Antweiler, W., Copeland, B., Taylor, M., 2001. Is free trade good for the environment? American Economic Review 91, 877–908. Arellano, M., Bond, S., 1991. Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. Review of Economic Studies, 58; 277–297. Arellano, M., Bover, O., 1995. Another look at the instrumental-variable estimation of error components models. Journal of Econometrics 68, 29–52. Baltagi, B., Kao, C., 2000. Nonstationary panels, cointegration in panels and dynamic panels: a survey. In: Baltagi, B. (Ed.), Nonstationary Panels, Panel Cointegration, and Dynamic Panels. Advances in Econometrics, vol. 15. JAI Press, New York. Blundell, R., Bond, S., 1998. Initial conditions and moment restrictions in dynamic panel models. Journal of Econometrics 87, 115–143. Blundell, R., Bond, S., Windmeijer, F., 2000. Estimation in dynamic panel data models: improving on the performance of the standard GMM estimator. In: Baltagi, B (Ed.), Nonstationary Panels, Panel Cointegration, and Dynamic Panels. Advances in Econometrics, vol. 15. JAI Press, New York. Breuer, J.B., McNown, R., Wallace, M., 2001. Misleading inferences from panel unitroot tests with an illustration from purchasing power parity. Review of International Economics 9, 482–493. Bun, M., Windmeijer, F., 2010. The weak instrument problem of the system GMM estimator in dynamic panel models. The Econometrics Journal 13, 95–126. Carson, R.T., 2010. The environmental Kuznets curve: seeking empirical regularity and theoretical structure. Review of Environmental Economics and Policy 4 (1), 3–23. Carson, R., Yongil, J., McCubbin, R., 1997. The relationship between air pollution emissions and income: U.S. data. Environment and Development Economics 2 (4), 433–450. Casler, S., Rose, A., 1998. Carbon dioxide emissions in the U.S. economy a structural decomposition analysis. Environmental & Resource Economics 11 (3–4), 349–363. Choi, I., 2001. Unit root tests for panel data. Journal of International Money and Finance 20 (2), 249–272. Choi, I., 2006. Combination unit root tests for cross-sectionally correlated panels. In: Corbae, D., Durlauf, S., Hansen, B. (Eds.), Econometric Theory and Practice: Frontiers of Analysis and Applied Research. Cambridge University Press, New York. Cole, M., Rayner, A., Bates, J., 1997. The environmental Kuznets curve: an empirical analysis. Environment and Development Economics 2, 401–416. Copeland, B., Taylor, M., 2003. Trade and the Environment: Theory and Evidence. Princeton University Press, Princeton.

937

Davidsdottir, B., Fisher, M., 2011. The odd couple: the relationship between state economic performance and carbon emissions economic intensity. Energy Policy 39, 4551–4562. Dasgupta, S., Laplante, B., Wang, H., Wheeler, D., 2002. Confronting the environmental Kuznets curve. Journal of Economic Perspectives 16, 147–168. Dinda, S., 2004. Environmental Kuznets curve hypothesis: a survey. Ecological Economics 49 (4), 431–455. Dinda, S., Coondoo, D., 2006. Income and emission: a panel-data based cointegration analysis. Ecological Economics 57, 167–181. EIA (2013) State Methodology, 〈http:/www.eia.gov/〉 cited March 2013. Frankel, J., 2003. The environment and globalization. In: Weinstein, M. (Ed.), Globalization: What's New. Columbia University Press, New York. Frankel, J. 2009. Environmental Effects of International Trade. Working Paper for Presentation. 〈http:/web.hks.harvard.edu/publications/getFile.aspx?Id=335〉 accessed September 13 2011. Galeotti, M., Manera, M., Lanza, A., 2009. On the robustness of robustness checks of the environmental Kuznets curve hypothesis. Environmental & Resource Economics 42, 551–574. Grossman, G., Krueger, A., 1991. Environmental Impacts of a North American Free Trade Agreement NBER Working Paper no. 3914. Hayakawa, K., 2007. Small sample bias properties of the system GMM estimator in dynamic panel models. Economics Letters 95, 32–38. Harris, R., Tzavalis, E., 1999. Inference for unit roots in dynamic panels where the time dimension is fixed. Journal of Econometrics 91, 201–226. Jiusto, S., 2008. An indicator framework for assessing US state carbon emissions reduction efforts (with baseline trends from 1990–2001). Energy Policy 36, 2234–2252. Kim, K., 2013. Inference of the environmental Kuznets curve. Applied Economics Letters 20, 119–122. List, J., Gallet, C., 1999. The environmental Kuznets curve: does one fit all? Ecological Economics 31, 409–423. Luzzati, T., Orsini, M., 2009. Investigating the energy-environmental Kuznets curve. Energy 34, 291–300. Managi, S., Hibiki, A., Tsurumi, T., 2009. Does trade openness improve environmental quality? Journal of Environmental Economics and Management 58, 346–363. Müller-Fürstenberger, G., Wagner, M., 2007. Exploring the environmental Kuznets hypothesis: theoretical and econometric problems. Ecological Economics 62 (3), 648–660. Piaggio, M., Padilla, E., 2012. CO2 emissions and economic activity: heterogeneity across countries and non-stationary series. Energy Policy 46, 370–381. Perman, R., Stern, D.I., 2003. Evidence from panel unit root and cointegration tests that the environmental Kuznets curve does not exist. Australian Journal of Agricultural and Resource Economics 47 (3), 325–347. Prasad, M., Munch, S., 2012. State-level renewable electricity policies and reductions in carbon emissions. Energy Policy 45, 237–242. RGGI, 2012. 〈http:/www.rggi.org/〉 cited July 2012. Roodman, D., 2008. A Note on the Theme of Too Many Instruments. Centre for Global Development, Working Paper no. 125. Stanton, A., Ackerman, A., Sheeran, K., 2010. Why Do State Emissions Differ So Widely? Economics for Equity & Environment December 2010 accessed online 12/24/2012. 〈http:/www.e3network.org/papers/Why_do_state_emissions_dif fer_so_widely.pdf〉. Soto, M., 2009 System GMM Estimation with a Small Sample. Barcelona Economics Working Paper Series no 395: July, 2009. Soytas, U., Sari, R., Ewing, B., 2007. Energy consumption, income, and carbon emissions in the United States. Ecological Economics 62, 482–489. Vollebergh, R.J., Melenberg, B., Dijkgraaf, E., 2009. Identifying reduced-form relations with panel data: the case of pollution and income. Journal of Environmental Economics and Management 58, 27–42. Westerlund, J., 2007. Testing for error correction in panel data. Oxford Bulletin of Economics and Statistics 69, 709–748.