The impact of income inequality on environmental quality: a sectoral-level analysis

Many studies in the literature examine the income inequality-environment nexus at the country level. In this paper, we argue that the impact of inequality on sectoral emissions might vary and should be examined by considering sectoral-level differences. We focus on 28 OECD economies and use DOLSMG, BA-OLS, and CUP-FM estimators. Our findings reveal that a cointegration relationship exists among the series in the long run, indicating that both income and income inequality are crucial factors in sectoral emissions. The estimates show that a 1% increase in the Gini index leads to an increase in emissions from the power and building sectors by about 1.4%. On the other hand, a 1% rise in the Gini index positively contributes to the environment in the transport, other industrial combustion, and other sectors by about 0.05%, 0.05%, and 0.02%, respectively. Policies aimed at reducing carbon emissions should be designed at the sectoral level.


Introduction
Income inequality and climate change are two major threats faced by humankind in the twenty-first century. They undoubtedly play a key role in shaping our ecosystem and future. Therefore, both of them have gained significant attention from researchers and policymakers worldwide during the last decades. Recent model projections and data also confirm their importance. For example, according to the Intergovernmental Panel on Climate Change (IPCC 2018), global warming due to human activities might cause further changes in the climate system if global net anthropogenic carbon emissions have not been reduced by about 45% from 2010 levels by 2030. Similarly, absolute income disparities continue to increase. As shown in the United Nations (2020) report, the per capita income gap between high and low-income countries increased from $27,600 to $42,800 between 1990 and 2018. Therefore, reducing inequality within and among countries, which is integral to achieving the Sustainable Development Goals (SDGs), still remains a distant goal by 2030.
The simultaneous worsening of both income distribution and environmental outcomes raises the following question: Does a relationship exist between these two indicators? Or, more specifically, does income inequality have significant implications for climate change? The answer to this question is regarded as highly important in the existing literature from the economic and environmental policy perspectives. This is because the balance of power between the poor and the rich is considered to have a substantial potential to determine the level of environmental degradation (Berthe and Elie 2015;Borghesi 2006;Boyce 1994;Chen et al. 2020;Hailemariam, Dzhumashev, and Shahbaz 2020;Ravallion, Heil, and Jalan 2000). The importance of this issue is also strongly supported by recent data. According to the Oxfam (2020) report, while the richest 10% are responsible for 46% of total emissions growth, this is 49% and 6% for the middle 40% and the poorest 50%, respectively.
The theoretical arguments and empirical tests of the relationship between income inequality and environmental degradation date back to the mid-1990s (Berthe and Elie 2015;Borghesi 2000;Boyce 1994;Cushing et al. 2015;Grunewald et al. 2017;Scruggs 1998). Although findings generally confirm the strong link between these variables, a clear consensus regarding its sign has not yet been reached. While some scholars argue that a negative or statistically insignificant link exists between income inequality and carbon dioxide (CO2) emissions (Heerink, Mulatu, and Bulte 2001;Ravallion, Heil, and Jalan 2000;Scruggs 1998), some others highlight the positive association and suggest that greater income inequality leads to more environmental deterioration (Boyce 1994(Boyce , 2007Marsiliani and Renstroem 2000;Torras and Boyce 1998). Based on the vast literature on the inequality-emissions nexus, it can be concluded that the empirical estimates produce mixed results more likely for the following reasons: (i) differences in the country sample, period, dataset, or econometric techniques (Borghesi 2006;Clement and Meunie 2010;Coondoo and Dinda 2008;Grunewald et al. 2017;Mittmann and de Mattos 2020;Morse 2018;Zhu et al. 2018); (ii) the exclusion of some important explanatory variables from the model specification Drabo 2011;Kashwan 2017;Kasuga and Takaya 2017;You et al. 2020) (iii) the lack of reliable historical data (Atkinson and Brandolini 2009;Berthe and Elie 2015;Hailemariam, Dzhumashev, and Shahbaz 2020;Uddin, Mishra, and Smyth 2020); and (iv) differences in data aggregation (Guo 2014;Hao, Chen, and Zhang 2016;Mushtaq et al. 2020;Zhang and Zhao 2014).
In this paper, we aim to contribute to this growing literature on the inequalityemissions nexus by considering sectoral-level differences. Put differently, although the impact of economic inequality on environmental degradation has been examined by many researchers, so far in the existing literature Grunewald et al. 2017;Hailemariam, Dzhumashev, and Shahbaz 2020;Yang et al. 2020), these country-level studies have mostly ignored the sectoral level differences in carbon emissions. 1 We argue that this highly preferred aggregated-level perspective might be one reason for obtaining conflicting results in the empirical literature. This is because the share of sectoral emissions in total CO 2 emissions in the countries significantly differs, implying that the contribution of each sector to the total carbon emissions of the country is not the same. Besides, while CO 2 emissions in some sectors have a downward trend over time, this trend displays a worrying trend for others. Therefore, the impact of income inequality on sectoral emissions might vary.
The numbers depicted in Figure 1 also strongly support our argument. While panel (a) of Figure 1 shows the evolution of sectoral CO 2 emissions in the OECD countries over the period between 1990 and 2018, panel (b) calculates the percentage change for the same period. As can be seen in panel (a), the buildings and industrial combustion sectors have a downward trend over the whole period. However, the power and transport sectors do not show the same gradual decline in emissions. Contrary to the buildings and industrial combustion sectors, the power and transport sectors have an increasing trend over the whole period. Although CO 2 emissions for these two sectors decrease over the short period between 2007 and 2009, their current value in 2018 is still higher than in 1990. As shown in panel (b), while the percentage change in CO 2 emissions between 1990 and 2018 is negative for the buildings and industrial combustion sectors, it is positive for the rest. Given this observation, assuming that the impact of income inequality on emissions is the same for all sectors does not seem to be a completely suitable approach to identify this link and prevents us from developing sector-specific strategies. Therefore, we consider that the environmental effect of income inequality is expected to vary from sector to sector. 2 Source: Emissions Database for Global Atmospheric Research (EDGAR) (Crippa et al. 2019).
Given the premises above, we investigate the nexus between income inequality and CO 2 emissions at the sectoral level. To this end, we perform an empirical analysis for five different sectors: power industry, buildings, transport, other industrial combustion, and other sectors. To specify our model, we augment the well-known environmental Kuznets curve (EKC) framework with income inequality Krueger 1991, 1995). 3 We also control the impact of globalization and urbanization on sectoral emissions (Inglesi-Lotz 2019). Our country sample consists of 28 OECD economies 4 for the period between 1990 and 2018. We consider these countries deserve special interest from researchers and policymakers as they are the most important players in industrial production and trade, and take strong measures at different levels to mitigate CO 2 emissions ( Figure 1).
Our study contributes to the literature in three ways. First, we examine the relationship between income inequality and CO 2 emissions at the sectoral level. To the best of our knowledge, this is the first study to empirically test the inequality-emissions nexus at the sectoral level. We believe such a sectoral cross-country empirical investigation is highly important and relevant in some respects. As discussed briefly above, a crosscountry analysis not allowing for sectoral differences might make the establishment of a uniform relationship between inequality and emissions difficult and lead to conflicting empirical results in the existing literature. From this point of view, the sectoral level analysis conducted in this paper is expected to address the limitations of these previous studies and reveal the importance of sectoral differences for correctly identifying the link between inequality and emissions. Besides, as sectors differ from each other in many ways, such as the amount of CO 2 emissions, government intervention, production process, the impact of income inequality on emissions might be expected to vary across sectors. Therefore, an empirical analysis ignoring sectoral differences in emissions does not seem to be a completely suitable approach to identify this link and prevents us from developing sector-specific strategies. From this point of view, to implement suitable policies and evaluate whether they are successful and effective, it is crucial to correctly identify the link in specific sectors, where actions are directed (Cialani and Mortazavi 2021;Morales-Lage et al. 2019a). Second, as our model specification is based on the EKC framework, we can test the validity of the inverted Ushaped relationship between growth and emissions at the sectoral level, which is also rarely discussed in the literature (Amin, Altinoz, and Dogan 2020;Aslan, Destek, and Okumus 2018). Therefore, we will not only discuss the sectoral link between inequality and emissions, but also income and emissions. Third, we methodologically use the second-generation panel data techniques to account for cross-sectional dependence and test the slope homogeneity of model estimates for five sectors. Based on the results, we choose cointegration tests and estimators allowing for heterogeneity in the slope parameters and cross-sectional dependence. We compare the weighted difference between the cross-sections and decide which assumption (homogeneity and heterogeneity) is implemented to estimate more powerful panel cointegration tests and estimators. Our findings reveal that the nexus between income inequality and emissions varies significantly across sectors. While income inequality has a statistically significant and positive effect on CO 2 emissions from the power and building sectors, this effect is found to be negative for the transport, other industrial combustion, and other sectors.
The remainder of the paper is organized as follows. Section 2 provides a literature review on the nexus between income inequality and emissions. Section 3 introduces the data, model specification, and econometric framework, while Section 4 presents our empirical results. Section 5 concludes with policy implications.

Theoretical mechanisms
Theoretically, several different mechanisms have been identified so far in the existing literature to explain the relationship between income inequality and environmental degradation. In this regard, the first hypothesis in the literature is known as the political economy approach (Boyce 1994(Boyce , 2007Torras and Boyce 1998). According to this approach, higher inequality results in a higher level of environmental degradation, meaning that a positive link exists between income inequality and environmental degradation. This effect mainly occurs in three different ways, i.e., the asymmetries in the political power, the cost-benefit analysis, and the rate of time preference. The asymmetries in the political power mainly indicate that the wealthy class, or the winners, benefitting more from the environmentally degrading activities have a greater influence on environmental policies than the poor due to their economic and political power. As the relative power of rich people increases the possibility of environmentally degrading activities, widening the income gap between groups harms the environment. Therefore, unequal distribution of power leads to environmental degradation. On the other hand, while the cost-benefit analysis mainly refers to the purchasing power differences across individuals, groups, and classes in willingness to pay to avoid pollution, the third transmission channel is the rate of time preference and compares the willingness to trade present benefits with future costs (Borghesi 2006).
The political economy approach, from a more technical point of view, suggests that environmental costs and benefits are not equally distributed across the different groups of society. Therefore, as the marginal benefit of the winners is not equal to the marginal cost for the losers, the actual environmental degradation level is higher than the socially efficient one and varies depending on the relative political power of the two groups. And if a regressive income redistribution exists from losers to winners, it will increase both the socially efficient and actual level of environmental degradation. Thus, an increase in the income gap leads to greater environmental degradation (Boyce 1994).
Another study conducted by Marsiliani and Renstroem (2000) also theoretically confirms the political economy approach proposed by Boyce (1994) but in a different way. The main intuition behind the argument of Marsiliani and Renstroem (2000) is the distorted or less stringent environmental policies. Using the overlapping generations models, this approach mainly suggests that countries relatively more equal in terms of income are more likely to enact stricter environmental policies, and environmental degradation reduces in these societies. Therefore, according to this hypothesis, income inequality is expected to be positively correlated with carbon emissions. Scruggs (1998) opposes the arguments advanced by Boyce (1994) and Marsiliani and Renstroem (2000) by highlighting the differences in consumer preferences. Scruggs (1998) mainly argues that consumers might prefer to use environmentallyfriendly goods and tend to promote environmental regulations as their income increases. Therefore, unlike the hypotheses proposed by Boyce (1994) and Marsiliani and Renstroem (2000), the wealthy class is expected to increase their demand for a cleaner environment. As a consequence, increasing income inequality reduces carbon emissions and contributes to environmental quality. In other words, Scruggs (1998) suggests that as the environment is a normal or superior good, it should be expected that a slight income increase in the wealthy class increases environmental deterioration less than does the poor class due to the income effect.
The positive association between inequality and environmental quality advanced by Scruggs (1998) is also shown by Ravallion, Heil, and Jalan (2000). However, Ravallion, Heil, and Jalan (2000) mainly focus on the marginal propensity to emit (MPE) concept, which measures the change in emissions with respect to the change in income. According to this approach, the losers have higher MPE values than winners. Therefore, if an income redistribution exists from winners to losers, or the income inequality narrows in a society, it harms the environment. 5

Empirical literature
The competing mechanisms theoretically explaining the inequality-emissions nexus have motivated many researchers to empirically verify this link. Consequently, since the early 2000s, the growing interest of scholars has made this discussion a highlystudied topic in the existing empirical literature. However, as stated earlier, a clear empirical consensus has not yet been reached. While some studies support the arguments of Boyce (1994), Marsiliani and Renstroem (2000), Torras and Boyce (1998) by reporting positive estimates for the relationship between inequality and emissions Hailemariam, Dzhumashev, and Shahbaz 2020;Knight, Schor, and Jorgenson 2017;Ridzuan 2019), some others highlight a negative or statistically insignificant linkage (Borghesi 2006;Heerink, Mulatu, and Bulte 2001;Huang and Duan 2020;H€ ubler 2017;Ravallion, Heil, and Jalan 2000;Scruggs 1998; Wolde-Rufael and Idowu 2017). 6 Magnani (2000) finds that the increasing income gap reduces environmental care for the 19 OECD countries over the period between 1980 and 1991. Padilla and Serrano (2006) confirm this positive link between inequality and emissions for a longer time dimension, i.e., 1971-1999. The studies conducted by Clement and Meunie (2010), Drabo (2011), and Ridzuan (2019) use alternative indicators as a proxy for environmental degradation, such as sulfur dioxide emissions (SO 2 ) and organic water pollution. The results largely support the positive linkage once again for the larger country samples and highlight the sensitivity of parameter estimates to the variable selection for environmental quality (Morse 2018). Similar findings are also reported by Holland, Peterson, and Gonzalez (2009), Qu and Zhang (2011), and Knight, Schor, and Jorgenson (2017) for various environmental indicators, such as the proportion of threatened plant and vertebrate species, consumption-based CO 2 emissions, and oxides of nitrogen (NOX) for different country samples. The existence of a positive linkage between income inequality and environmental degradation is robust, even with alternative indicators as a proxy for income inequality (Coondoo and Dinda 2008;Hailemariam, Dzhumashev, and Shahbaz 2020;Kashwan 2017;Magnani 2000) or different estimators Uddin, Mishra, and Smyth 2020;You et al. 2020;Zhu et al. 2018) have been used. Baek and Gweisah (2013), Uzar and Eyuboglu (2019), Ali (2022), and Hundie (2021) confirm the positive impact of the Gini index on CO 2 emissions at the single country level, i.e., for the USA, Turkey, Egypt, and Ethiopia, by using different econometric techniques. Some other studies highlight the importance of other factors, such as income level, geographical area, institutional quality, to identify the correct link between study variables (Coondoo and Dinda 2008;Grunewald et al. 2017;Kashwan 2017;Mittmann and de Mattos 2020).
The empirical studies challenging the positive association between inequality and emissions also support their findings with alternative indicators, periods, methods, and country samples. The cross-section analysis by Heerink, Mulatu, and Bulte (2001) shows that while the Gini coefficient has a statistically significant and negative effect on CO 2 emissions, it is statistically insignificant for SO 2 and suspended particulate matter (SPM). H€ ubler (2017), Huang and Duan (2020), and Yang et al. (2020) confirm this negative linkage for different country groups by using alternative estimators. Alataş (2022) reveals that a widening income gap leads to a decrease in per capita emissions for a panel of 120 countries for the period between 1980 and 2019. Wolde-Rufael and Idowu (2017) employ three different estimators to test the inequality-emissions nexus in China and India. The empirical results reveal that income distribution is not an important factor in explaining the changes in CO 2 emissions.
It is equally important to note that all of the studies discussed above have been conducted at the country level, regardless of what they find about the direction of the link between income distribution and environmental quality. However, the studies in the literature are not only limited to these aggregated level studies. Some other papers also investigate the same nexus at the disaggregated level, such as at the regional (Guo 2014), state (Pattison, Habans, and Clement 2014), or household level (Y. Liu, Ren, et al. 2020). 7 For the US states, while Boyce et al. (1999), Bouvier (2014) Mader (2018) shows that alternative indicators, techniques, periods, and regions do not support the findings of Jorgenson, Schor, and Huang (2017). The empirical estimates of Kasuga and Takaya (2017) reveal a positive and statistically significant impact of inequality (90 th /10 th ) on SO 2 , NOX, and SPM for Japanese cities in the residential and commercial areas. Based on provincial panel data for China, Hao, Chen, and Zhang (2016) emphasize the importance of regional differences in analyzing the link between income inequality and carbon emissions. Mushtaq et al. (2020) confirm the regional difference in China for the larger time-period by using alternative estimators and emphasizing the moderating role of innovation. Similar analyses are also performed for Chinese provinces by Golley and Meng (2012), Zhang and Zhao (2014), Guo (2014), Q. Liu, Jiang, et al. (2019), and . Based on the Swedish household cross-sectional data, Br€ annlund and Ghalwash (2008) suggest that not only income but also income distribution is an important factor in explaining the changes in environmental pollution in Sweden.
In summary, the current literature is unclear whether the impact of income distribution on environmental quality is positive or negative. While the vast majority of the empirical studies are conducted at the country level, only a limited number of papers focus on the disaggregated data, such as the state or household level. Our study contributes to this growing literature by considering differences in sectoral CO 2 emissions. To the best of our knowledge, no other study is available in the literature to investigate the link between income inequality and environmental degradation at the sectoral level for OECD countries. Therefore, this study is believed to offer a significant contribution to the literature.

Data description, model specification, and econometric framework 3.1. Data description
In this study, we investigate the nexus between income inequality (GINI) and CO 2 emissions (CO2) at the sectoral level for OECD countries. 8 The sectors covered in the study are the power industry (POW), buildings (BUI), transport (TRA), other industrial combustion (OIC), and other (OTH) sectors. In addition to income inequality, following the EKC framework, we have included gross domestic product (GDP) and the square of it (GDP 2 ) into our regressions as a determinant of sectoral emissions. We add two control variables, i.e., globalization (KOF), and urbanization (URB). We perform our empirical investigation for the period between 1990 and 2018. While we retrieve sectoral per capita CO 2 emissions (metric units) from Crippa et al. (2019), data for the Gini coefficient of income inequality come from Solt (2020). The data source for GDP per capita (constant 2010 USD) and urban population (% total population) is the World Bank's World Development Indicators (WDI)) (2020). We compile globalization data from the Swiss Economic Institute (Gygli et al. 2019). Table 1 reports the descriptive statistics for all the variables. As can be seen, there are 812 observations for all variables. The difference between the minimum and maximum values of the explanatory variables is low, which indicates the similarity in the panel countries. Besides, dependent variables have a high standard deviation, indicating that their values expand over a broader range. There is no strong correlation between explanatory variables, indicating that there is no multicollinearity problem. 9

Model specification
We investigate the nexus between CO 2 emissions and income inequality under the well-known EKC framework (Grossman and Krueger 1991;Panayotou 1994). The EKC hypothesis mainly posits that as income increases in a country, it affects environmental quality negatively and harms the environment in the first stage. However, after reaching a certain income level (turning point), this negative impact diminishes, and environmental quality improves. Therefore, it suggests an inverted U-shaped relationship between environmental deterioration and income level (Bilgili et al. 2019).
Based on the EKC framework and following Chen et al. (2020), Hailemariam, Dzhumashev, and Shahbaz (2020), Mushtaq et al. (2020), Torras andBoyce (1998), Wolde-Rufael andIdowu (2017), and many others, we specify our empirical model to be estimated in this study as follows where, while CO 2 is the dependent variable representing sectoral per capita CO 2 emissions, GINI, GDP, and GDP 2 are our key independent variables and stand for income inequality, income per capita, and the square of income per capita, respectively. u it is the error term. The subscripts i and t denote country and time period, respectively. In order to control the potential impact of other variables on sectoral CO 2 , we added two control variables to our model specification: globalization (Inglesi-Lotz 2019; Ulucak, Danish, and Khan 2020) and urbanization (Amin, Altinoz, and Dogan 2020;You et al. 2020). 10 All series used in the estimates are in natural logarithm (ln) form in the estimations. It is worth noting that as we estimate Equation (1)

Econometric framework
The empirical investigation in this study consists of four parts. We first test the existence of the cross-sectional dependence of our study variables. To this end, we apply three different cross-sectional dependence tests. The first one is the Lagrange multiplier (LM) test developed by Breusch and Pagan (Breusch and Pagan 1980). This test performs well when T is larger than N. However, it has substantial size distortions when N is large, and T is small. Pesaran (2004) overcomes this weakness and proposes the CD test designed for large N and small T panels. Therefore, we also apply the CD test by Pesaran (2004). The last one is the modified version of the LM test. The bias-adjusted LM test proposed by Pesaran, Ullah, and Yamagata (2008) examines the sustainable power of exogenous regressors and normal errors in the panel. Therefore, it produces more robust results than the other cross-sectional dependence tests. Rejection of the null hypothesis for all three tests implies that the residuals are crosssectionally dependent (Akın 2019;Burdisso and Sangi acomo 2016;De Hoyos and Sarafidis 2006).
Second, we examine the stationarity properties of data. As we find the cross-sectional dependence for all variables in the previous step of the empirical investigation, we perform the second-generation unit root tests considering cross-sectional dependence. We apply two unit root tests: bootstrap-IPS (Smith et al. 2004) and cross-sectionally augmented IPS (CIPS) (Pesaran 2007). Both tests are based on the Augmented Dickey-Fuller (ADF) test. To consider cross-sectional dependence, while Smith et al. (2004) improve the ADF test by limiting distribution in a bootstrap-based approach, Pesaran (2007) augments the ADF test with the cross-section averages of lagged levels and first-differences of the individual series. CIPS test is a factor modeling (FM) approach and assumes the presence of an unobserved common factor. The null hypothesis for both tests is the existence of the unit root for the panel. It is worth noting that the CIPS test (especially the three-dimensional version) has a better power performance than the Bootstrap-IPS test if high levels of cross-sectional dependence exist in the data (Giulietti, Otero, and Smith 2008, 191). Therefore, we consider these issues when interpreting our unit root test results.
Third, we investigate the cointegration relationship. To this end, we use the Westerlund (2007b) panel ECM cointegration approach. The panel cointegration test suggested by Westerlund (2007b) is a technique that is often used in the existing literature to analyze the long-run cointegration relationship. Westerlund (2007b) develop four different panel cointegration tests. All these tests consider cross-sectional dependence. As the group-mean statistics G s , G a ð Þare calculated under the heterogeneity assumption, the alternative hypothesis of these tests is that at least one unit is cointegrated. On the other hand, the panel statistics P s , P a ð Þ are calculated under the homogeneity assumption. Thus, unlike the group-mean test statistics, the alternative hypothesis suggests that the panel is cointegrated as a whole (Persyn and Westerlund 2008).
It is worth noting that G a and P a depend on T values. In other words, when the number of lags is large, normalization of G a and P a by T may lead to the Type I error (Westerlund 2007b). In this study, we prefer to interpret G s and P s as our sample size to be large enough. We also examine the testing slope homogeneity of the panel with the Delta test (Pesaran and Yamagata 2008) based on a standardized version of Swamy's test (Swamy 1970) to decide test statistics (G s or P s ).
In the fourth step, we estimate the long-run parameters. We employ three different estimators in the study: the group-mean panel-dynamic ordinary least-squares (DOLSMG) (Pedroni 2001), the bias-adjusted OLS (BA-OLS) (Westerlund 2007a), and the continuous updated fully modified (CUP-FM) (Bai and Kao 2005).
The DOLSMG estimator proposed by Pedroni (2001) is the augmented version of the individual time-series DOLS estimator. It can be applied to the non-stationary data showing the cointegrating relationship between variables. This estimator has an important advantage for between-dimension panel time-series estimators in the case of slope heterogeneity (Neal 2014;Pedroni 2001). We, therefore, mainly focus on the DOLSMG estimator for models 1, 2, and 3 as the slope parameters of these models are found to be heterogeneous. However, as the delta test results confirm the slope homogeneity for models 4 and 5, we pay special attention to the long-run coefficients of the CUP-FM and BA-OLS estimators for these models (Tato glu 2020, 230-232).
The CUP-FM estimator uses the principal component and FM methods to calculate common factors and estimate the cointegrating vector. Besides, as this estimator assumes that the number of common factors is known, it significantly reduces the dimension of the cross-sectional correlation. The CUP-FM estimator thus performs well in small samples compared to the OLS. On the other hand, the BA-OLS uses several panel information criteria to estimate the number of factors. The simulation results show that the BA estimator outperforms in terms of precision and size accuracy (Bai and Kao 2005;Westerlund 2007a).

Empirical results and discussion
Before examining the empirical investigation of the relationship between inequality and emissions, we first test the existence of cross-sectional dependence among countries in the sample. The results are reported in Table 2. The findings reveal that the null hypothesis of independence among cross-sections is strongly rejected for all variables, implying the existence of dependence between cross-sections. It means that a shock occurring in one of the OECD countries might spill over into other economies. From the methodological perspective, this result helps us perform more appropriate tests and estimates in the following steps of the empirical investigation (Henningsen and Henningsen 2019;Tugcu 2018).
After analyzing and confirming the cross-sectional dependence for all variables, we secondly test the stationarity properties of data. As all variables are cross-sectionally dependent, we apply the unit root tests allowing for cross-sectional dependence, known as the second-generation unit root tests in the literature. In doing so, we have performed two different unit root tests, i.e., IPS bootstrap and CIPS. The results are presented in Table 3. As can be seen, the unit root test results unveil a uniform order of integration for all variables. While the p-values of the variables for both unit root tests are generally found to be greater than 0.05 at the level; they are below 0.05 when the first difference of them is taken. It means that our study variables are not stationary at I(0). However, they turn out to be stationary at I(1). Therefore, a cointegration analysis seems ideal for further empirical analysis (Jalil and Rao 2019).
As the integration order of all variables is one, we thirdly investigate whether sectoral CO 2 emissions and their determinants are cointegrated or not in the long run. However, it is equally important to note that the selection of an appropriate cointegration test and estimator largely depends on the homogeneity and cross-sectional dependence test results. Therefore, prior to the cointegration test, we first test the homogeneity and cross-sectional dependence for each model. The homogeneity and cross-sectional dependence test results are given in panel (a) of Table 4. The homogeneity test (D and D adj Þ results show that the null hypothesis of homogeneity is rejected for models 1, 2, and 3, meaning that the slope coefficients are heterogeneous. On the other hand, the slope parameters are found to be homogenous in models 4 and 5. Besides, the test statistics of LM, LM adj , and CD strongly reject the null of no crosssectional dependence in line with our previous results for variables. Given these outcomes, we perform a cointegration test considering the cross-sectional dependence and heterogeneity. To this end, we apply the Westerlund (Westerlund 2007b) panel ECM cointegration approach as we find cross-sectional dependence in all models. The test results for each model are given in panel (b) of Table 4. As clearly seen, the results strongly reject the null hypothesis of no cointegration and verify the cointegration relationship between sectoral CO 2 emissions, income, the square of income, income inequality, globalization, and urbanization in OECD countries.
Following the cointegration testing, we finally analyze the relationship between study variables by estimating the long-run parameters. To this end, we employ three different estimators, i.e., DOLSMG, BA-OLS, and CUP-FM. As indicated earlier, while the DOLSMG estimator performs well in the case of parameter heterogeneity, the BA-OLS and the CUP-FM estimators consider slope homogeneity. Therefore,    and leads are 1 for the DOLSMG estimates. The maximum common factor number is 2 for the BA-OLS and CUP-FM estimates. While models 1 and 2 correspond to the power industry and buildings sectors, models 3, 4, and 5 are for the transport, other industrial combustion, and other sectors, respectively.
when interpreting the estimates, we consider the homogeneity test results reported in Table 4. The long-run parameter estimates of our models are presented in Table 5. Based on these estimation results reported in Table 5, we obtain the following outcomes. First of all, the nexus between income and emissions significantly varies across sectors. For example, the DOLSMG estimates confirm the validity of the EKC hypothesis for the power and building sectors. As can be seen, while the relationship between income and CO 2 emissions is statistically significant and positive for models 1 and 2 (9.824 and 13.230), it turns out to be negative for the square of income (À0.4496 and À0.6183). This means that rising income initially increases CO 2 emissions from the power and building sectors in the OECD countries. However, after reaching a turning point, it reduces emissions in these sectors and positively contributes to the environment. The validity of the inverted U-shaped EKC hypothesis clearly shows us the importance of income level for mitigating emissions in these high-emitting sectors (see Figure 1). Our finding is not consistent with Erdo gan et al. (2020), who find a statistically insignificant linkage for the selected G20 countries using the CCEMG and AMG estimators, more likely due to the differences in the study period, estimators, and country sample.
At this point, we should emphasize an important issue. A significant and negative estimated parameter of the quadratic term in the regression alone might not be a reliable criterion for the existence of an inverted U-shaped relationship in models 1 and 2. Therefore, such a conventional test of the non-linear relationship might be flawed. To address this issue, Lind and Mehlum (2010) have proposed a test, which gives the exact necessary and sufficient conditions. The proposed test consists of three steps. In the first step, the estimated parameter of the quadratic term needs to be statistically significant and should have the expected sign. Second, the slope must be sufficiently steep at both ends of the data range. Therefore, both slope tests need to be statistically significant. Besides, an inverted U-shaped relationship requires a positive slope at the lower bound and a negative slope at the upper bound of the given data range. In the third step, the turning point needs to be located with this data range (Haans, Pieter, and He 2016). In this study, we have employed the Lind and Mehlum u-test to the DOLSMG estimated parameters. However, the results are not consistent with what we have found by using conventional techniques. Thus, we should approach the conventional technique results cautiously. 11 Another important issue to be considered for the inverted U-shaped relationship is the calculation of the turning points, implying the existence of local maxima for models 1 and 2. 12 The results reveal that the turning point of income for models 1 and 2 are $55.562 and $44.297, respectively.
It is equally important to note that the inverted U-shaped relationship found for the power and building sectors does not hold for the transport sector. Although the DOLSMG results produce a positive and significant parameter estimate for the relationship between income and emissions in this sector (model 3) (8.023), the estimated coefficient of the square of income is not statistically significant in the same model (À0.2902). While this result is confirmed by Aslan, Destek, and Okumus (2018) and Chandran and Tang (2013), it is not consistent with the findings of Amin, Altinoz, and Dogan (2020) and Ozkan, Yanginlar, and Kalayci (2019). Similarly, the BA-OLS and CUP-FM estimates reject the validity of the inverted U-shaped EKC hypothesis for models 4 and 5. For example, as the estimated parameters of income and the square of income are, respectively, negative or statistically insignificant (À0.015, À0.003, À0.021, and À0.005) for model 5, this finding verifies a linearly decreasing growth-pollution association in the other sector (Sinha, Shahbaz, and Balsalobre 2019). The calculated turning point based on the DOLSMG estimates (local minima) where emissions are at a minimum is $12.575 for model 5.
Second, we find that income inequality has a heterogeneous effect on sectoral CO 2 emissions. In other words, as shown in the fifth column of Table 5, the Gini coefficient has a statistically significant effect on emissions for all sectors (DOLSMG estimates for all models; BA-OLS and CUP-FM estimates for models 3, 4, and 5), varying in terms of magnitude depending on the sector. However, while this effect is positive for the power and building sectors (models 1 and 2), it is found to be negative for the transport, other industrial combustion, and other sectors (models 3, 4, and 5). For example, the DOLSMG estimates reveal that a 1% increase in the Gini index leads to an increase in emissions from the power and building sectors by about 1.4% (1.542% and 1.451%, respectively). On the other hand, BA-OLS and CUP-FM estimates show that an increase in income inequality is negatively associated with CO 2 emissions other industrial combustion (À0.069 and À0.046), and other sectors (À0.028 and À0.022). Our positive estimates for the power and building sectors are in line with the political economy approach (Boyce 1994(Boyce , 2007Marsiliani and Renstroem 2000;Torras and Boyce 1998). Therefore, it can be concluded that more equally distributed income might significantly reduce carbon emissions in the power and building sectors. Although our country sample differs significantly, the results obtained from this study are largely consistent with the findings of Bale zentis et al. (2020), Baloch et al. (2020), Chen et al. (2020), Hailemariam, Dzhumashev, and Shahbaz (2020), Hundie (2021), Knight, Schor, and Jorgenson (2017), and Ridzuan (2019). For example, while DOLS estimates of Hundie (2021) for Ethiopia reveal that a 1% increase in income inequality increases carbon emissions by 0.21% in the long run, Chen et al. (2020) obtain a similar outcome for developing countries using simultaneous quantile regression. Baloch et al. (2020) suggest that increasing income inequality has a detrimental effect on environmental degradation in Sub-Sharan African countries, whereas Bale zentis et al. (2020) show that the increasing inequality increases the consumption-based carbon footprint per capita in the low-income inequality countries. 13 Yet, this positive finding does not hold for the transport, other industrial combustion, and other sectors. The negative estimates for these sectors verify the arguments and findings of Scruggs (1998), Ravallion, Heil, andJalan (2000), H€ ubler (2017), Heerink, Mulatu, and Bulte (2001), Duan (2020), andWolde-Rufael andIdowu (2017), and suggest that environmental quality increases with a rising income gap in the OECD countries. For example, while Wu and Xie (2020) reveal that income inequality and carbon emissions are negatively associated in OECD and high-income non-OECD countries, Alataş (2022) and Rojas-Vallejos and Lastuka (2020) report estimates that support the tradeoff hypothesis between inequality and emissions.
The empirical findings presented above, and observations depicted in Figure 1 clearly show us that the power, building, and transport sectors deserve special attention. In other words, these three sectors play a crucial role in determining the total CO 2 emissions of the OECD countries due to their larger share in total carbon emissions. For example, the share of CO 2 emissions from the power, building, and transport sectors is about 35%, 15%, and 25% over the years, respectively. More importantly, the percentage change in CO 2 emissions between 1990 and 2018 is positive for the power and transport sectors (Crippa et al. 2019). When we combine this important information with our findings discussed above, the heterogeneous impact of income, the square of income, and income inequality on sectoral carbon emissions produce a highly interesting outcome for these sectors. For instance, we confirm the validity of the U-shaped EKC hypothesis for the power and building sectors, meaning that income level is an important factor affecting sectoral emissions. It is also equally important to note that the rise in emissions in the initial stages of the EKC curve is expected to be more pronounced than the decrease during the latter stages. We also find that improvements in income distribution might significantly reduce emissions in these sectors. Therefore, we can conclude that the reduction of emissions in the power and building sectors depends on both income level and how income is distributed. In this regard, policies to reduce carbon emissions in these sectors should be designed not only to increase income level but also to narrow the income gap. This implies a challenge to policymakers who seek to reduce both income inequality and carbon emissions, as also found by Rojas-Vallejos and Lastuka (2020).
On the other hand, the results reveal the opposite outcome for the transport sector. As can be seen, CO 2 emissions from the transport sector reduce as income decreases or income equality increases. However, we should approach this result cautiously. This is because, as also stated by Scruggs (1998) and Mittmann and de Mattos (2020), this result does not necessarily imply that reducing income level or maintaining income inequality might significantly contribute to environmental quality. It instead signals a challenge from the sustainability perspective for this sector. From this point of view, policies aimed at mitigating carbon emissions in the transport sector by reducing income inequality might not produce the expected effective results for improving environmental outcomes. This also holds for the other industrial combustion and other sectors (Wolde-Rufael and Idowu 2017). In this regard, transmission channels explaining the link between income inequality and emissions should be correctly identified. One of them is social norms consisting of individualism, short-termism, and consumerism. According to this transmission channel, increasing income inequality across individuals in the community harms social cohesion and leads to a decrease in people's sensitivity toward the environment. As people start to prioritize their short-term benefits in society, they tend to put more pressure on the environment, and their conspicuous consumption increases. The demand for better environmental quality significantly decreases due to less social resistance to environmental degrading activities (Alataş 2022;Awaworyi, Ivanovski, and Ephraim 2021;Berthe and Elie 2015;Laurent 2015;Uzar 2020;Wilkinson and Pickett 2010). Therefore, policies aimed at making changes in the consumption style of individuals, especially in terms of transportation modes, such as increasing zero-emissions or use of public vehicles, might play an important role as road transportation has a significant share in the total transport sector GHG emissions. However, it is equally important to note that these policies are not sufficient and should be supported with some other policies, such as raising public awareness toward the transportation modes and their effect on the environment.
Third, we discuss the results for other control variables, i.e., globalization and urbanization. As it is understood, the impact of globalization on sectoral emissions is statistically significant at the conventional significance level for all models. However, as also found for income inequality, this effect varies across the sectors. While the coefficient of globalization (lnKOF) is positive for model 2 (0.324), it is found to be negative for models 1, 3, 4, and 5. This means that increased trade liberalization has a detrimental effect on environmental quality as it increases CO 2 emissions from the building sector, implying a risk for sustainability problems. On the other hand, globalization has a negative effect on CO 2 emissions from power, transport, other industrial combustion, and other sectors, indicating that trade openness contributes to the reduction of carbon emissions. Our negative estimates are compatible with Shahbaz et al. (2017), You and Lv (2018), and M. Liu, Ren, et al. (2020) in terms of the overall emission-globalization nexus.
Regarding the effect of urbanization on sectoral CO 2 emissions, we find that urbanization is an important factor in explaining changes in CO 2 emissions at the sectoral level. Except for the power and building sectors, urbanization positively affects carbon emissions in all sectors, implying the negative role of urbanization on environmental quality. This result is not consistent with Amin, Altinoz, and Dogan (2020), who find a statistically insignificant linkage between emissions and urbanization for the transport sector in European countries. However, it is compatible with Xu and Lin (2015) confirming the significant effect of urbanization on China's transport sector.

Conclusion
A growing number of studies empirically investigate the inequality-emissions nexus in the literature. However, they ignore the sectoral differences in CO 2 emissions. In this study, we criticize this practice by showing the varying environmental effect of inequality across sectors. The sectors covered in the study are the power industry, building, transport, other industrial combustion, and other sectors. Our model specification is based on the well-known EKC framework. We perform our empirical investigation for the 28 OECD countries over the period 1990 to 2018. We applied the second-generation panel unit root tests and estimators, which produce robust results against cross-sectional dependence.
We obtained the following outcomes: (i) the cointegration test results confirm the long-run cointegration relationship between sectoral CO 2 emissions, income, the square of income, income inequality, globalization, and urbanization in OECD countries; (ii) the parameter estimates find a statistically significant association between sectoral emissions and its determinants for almost all sectors; (iii) we confirm the validity of the EKC hypothesis for the power and building sectors, but not for the transport, other industrial combustion, and other sectors. While the effect of income on emissions from the transport sector is found to be positive, it is negative for the other industrial combustion and other sectors; (iv) income inequality has a heterogeneous effect on sectoral emissions. This effect is positive for the power and building sectors, whereas it is found to be negative for transport, other industrial combustion, and other sectors; (v) as in the case of inequality, the effect of globalization varies across the sectors. Yet, except for the building sector, urbanization positively affects carbon emissions in all sectors.
This paper provides some policy recommendations to reduce sectoral CO 2 emissions in OECD countries based on these findings. First, as it is clearly shown, not only income but also income distribution plays a key role in explaining the changes in sectoral emissions. Therefore, focusing on policies designed to increase income level might not be sufficient for controlling climate change and improving environmental quality. Policymakers should also pay appropriate attention to income distribution. This implies a challenge to policymakers who seek to reduce both income inequality and carbon emissions.
Second, as the environmental effect of income inequality varies in terms of sign depending on the sector examined, designing sector-specific policies seems to be a more suitable approach for improving environmental outcomes. For example, more equally distributed income might be beneficial for the power and building sectors to reduce CO 2 emissions from these sectors as there is no tradeoff between narrowing the income gap and achieving environmental goals. On the other hand, the negative link between inequality and emissions, especially for the high-emitting sectors, i.e., transport and other industrial combustion, confirms the tradeoff and implies a big challenge from the sustainability perspective. In this regard, it is recommended that alternative policies should be formulated and implemented for these sectors, especially for the transport sector. For example, policies aimed at making changes in the consumption style of individuals, especially in terms of transportation modes, such as increasing the use of zero-emissions or public vehicles, might play an important role as road transportation has a significant share in the total transport sector GHG emissions. However, it is equally important to note that these policies are insufficient and they should be supported by other policies, such as raising public awareness toward transportation modes and their effect on the environment. Similarly, increasing investment in renewable energy sources or facilitating the diffusion of technological development might significantly reduce their contribution to total emissions.
Third, in line with some other studies, observations and findings for the transport sector produce worrying outcomes. Therefore, it is quite obvious that this sector deserves special attention by governments. From this point of view, promoting sustainable transportation modes and increasing the environmental awareness of the urban population might significantly help to reduce CO 2 emissions as a large part of transport emissions come from road vehicles.
In this paper, we, for the first time, argue that the impact of inequality on sectoral emissions might vary and should be examined by considering sectoral-level differences in carbon emissions. In this respect, further studies that produce other more reliable and robust estimates with alternative econometric techniques, or that use different indicators for carbon emissions at the sectoral level, or that specifically focus on one sector in more detail, can further enrich our findings.
Notes have excluded some countries from the sample due to data unavailability, especially for both income inequality and control variables. 9. For the pair-wise correlation matrix, please see Table 7 in the supplementary file. 10. For recent studies extending the EKC framework by adding control variables, please see Inglesi-Lotz (2019). 11. The results are available upon request from the authors. 12. To calculate the local maxima (or minima) points (turning point of income where emissions are at a maximum or minimum), we use the following formula: s ¼ exp Àb 2 =2b 3 ½ , where b 2 and b 3 are the estimated coefficients of income per capita and its square presented in Equation (1), respectively. 13. The assumption that the slope coefficients are homogeneous or heterogeneous can also affect the sign of the coefficients. Therefore, simply ignoring slope heterogeneity leads to biased results (Pesaran and Smith 1995). Due to assumption differences, some coefficients estimated from alternative estimators for the same model can obtain opposite signs.