Are Driving Forces of Co2 Emissions Different Across Countries?: Insights from Identity and Econometric Analyses

This paper investigates factors behind the growth of carbon dioxide emissions over the 35 years between 1980 and 2015 in more than 100 countries, using an index decomposition technique (the Logarithmic Mean Divisia Index). The results are further confirmed using an econometric technique (the general method of moments). The study finds that economic growth, measurred in per capita gross domestic product, and population growth are the main drivers of the growth of carbon dioxide emissions during 1980?2015. Although economic growth is mainly responsible for the growth of emissions in high-, upper-middle-, and lower-middle-income countries, population growth that is primarily responsilble for it in low-income countries. More than 70 percent of the global growth in carbon dioxide emissions over the past 35 years was contributed by upper-middle-income countries. Improved energy efficiency, reflected in the declining energy intensity of gross domestic product, has substantially contributed to limit global carbon dioxide emissions at the current level; otherwise, the world's current carbon dioxide emissions would have been 40 percent higher. Despite the recent rapid expansion of renewable energy, its contribution to slowing the growth of global carbon dioxide emissions is not noticeable yet, due to its small share in the global energy supply mix.


Are Driving Forces of CO2 Emissions Different across Countries? Insights from
Identity and Econometric Analyses

Introduction
Over the past several decades, CO2 emissions have been increasing steadily in most countries around the world along with economic growth, population growth, industrialization and urbanization. At the global level ( Fig. 1), the annual energy consumption and CO2 emissions have been increasing by, respectively 2.0% and 1.7%, on average, over the 35 years during the 1980-2015 period. During the same period, the world economy has been growing at the rate of 2.9% and population is growing at the rate of 1.5%, on average, annually [1].
In this paper, we investigate the driving forces behind the rapid growth of CO2, emissions in more than 100 countries for the 35 years during the 1980-2015 period. We also analyze how the roles of the driving factors change in different groups of countries differentiated by their income. Data source: [1,2]. Note: the value of each item in 1980 is equal to 100. Two different approaches are found to be used in the literature to determine the causes of CO2, growth. Broadly, these approaches can be classified into two groups: identity approach and econometric approach. The first decomposes the independent variables to calculate their relative roles in driving the emission growth; it is also called index decomposition analysis, pioneered by Professor B.W. Ang of Singapore National University [3]. The second uses the standard econometric approach to check the causality between the dependent and independent variables.
Within the decomposition or identity approach, there are several specific techniques, of which the Logarithmic Mean Divisia Index (LMDI) method introduced by Ang and Zhang in 2000 [4] has been widely used. The popularity of this technique has increased due to its ability to handle cases with zero values without leaving residuals, consistency in aggregation, and path independency [5][6][7][8][9]. Since the 2000s, a growing body of studies has employed the LMDI approach to identify the factors impacting CO2 at the country and regional levels (see Table A1). For instance, at the regional level, González et al. [10] decompose CO2 in the European Union (EU) from 2000 to 2010 into five factors. Other scholars investigated CO2 at the country level (Hatzigeorgiou et al. [11] for Greece, Tunç et al. [12] for Turkey, Oh et al. [13] for the Republic of Korea, de Freitas and Kaneko [14] for Brazil, O'Mahony et al. [15] for Ireland, Zhang et al. [16] for China, Feng et al. [17] for the United States, and Mousavi et al. [18] for the Islamic Republic of Iran, among others).
On the econometric side, various methods are employed where the focus of these studies is usually the relation between CO2 and income and carried out at the national, regional and global levels. At the national level, Bento and Moutinho [19] explore the linkage between renewables energy adoption and CO2 emissions in Italy using Toda-Yamamoto causality tests. Similar studies have been conducted by Bélaïd and Youssef [20] for Algeria and Danish et al. [21] for Pakistan. At the regional level, Dong et al. [22,23] use the VECM panel Granger causality method to estimate how expansion of renewable energy consumption would reduce CO2 emissions in the BRICS countries (i.e., Brazil, Russia Federation, India, China, and South Africa). Other similar regional studies include Dogan and Seker [24] and Jebli et al. [25].
The added value of this paper is to complement the analysis using the LMDI approach with the econometric approach --the generalized method of moments so that results derived from the former technique can be validated or strengthened with the latter thereby enhancing the robustness of the findings. Unlike many existing studies that determine the drivers of CO2 growth at the national or regional level, this study considers a global analysis using data from 100 plus countries. 3 The data span  considered in this study is much longer as compared to other existing literature.
The remainder of this paper is structured as follows. Section 2 outlines the methodology and data used. Section 3 discusses the results from the decomposition analysis. Section 4 discusses the results from the econometric analysis. Section 5 conducts a comparison of the results between the two approaches: statistical vs.
numerical. Section 6 offers policy discussion and concluding remarks.

Methodology and data
Let us define global CO2 emissions from energy consumption in a country for a given year with the following identity: where C denotes the global CO2 emissions; ij C represents the amount of CO2 emissions of energy type j in country i ; ij E stands for the energy consumption by energy type j consumed in country i ; i E , i GDP , and i P refer to the total primary 3 Two other global studies include Wang et al. [26] and Bacon et al. [27]. However, their period of analysis is much shorter than the one used in this paper.
[26] Wang S, Li G, Fang C. Urbanization, economic growth, energy consumption, and CO2 emissions:  Equations 4 to 8). This causes the results to be biased towards larger countries and effects of driving factors from small countries do not show up. For example, if the energy intensity of GDP in many countries is improving significantly, but decreasing in the large countries, the improvements in many countries will be overshadowed by deteriorations in the aggregate or the global result. The GMM method, on the other hand, corrects this bias by putting an equal weight on each country for a given driver. The results under the GMM are more representative of all countries instead of large countries only. Therefore, complementing LMDI analysis with GMM brings additional insights in the analysis.

The LMDI decomposition method
Using the LMDI decomposition technique, the following relationship can be derived from Eq. (2): , and subscripts i and j are energy and country types, 7 respectively.

Econometric methodology
For the GMM approach, we convert Eq. 2 to the relationship expressed in Eq. 9. In addition, we added one more factor, square of GDP per capita to investigate if the environmental Kuznets curve (EKC) ( [28] & [29]) occurs. Note in Eq. 9 that we assumed a log-log relation between CO2 emissions and the driving factors.
where subscripts i and t denote country and year, respectively; 1 6 -  are the parameters to be estimated; 2 CO represents the amount of CO2 emissions (measured in million tonnes, Mt); EC indicates the emission coefficient (measured in tonnes/toe); ECS denotes energy mix (measured in the share of dirty fossil fuels (i.e., coal and petroleum) in total final energy consumption, %); EI describes energy intensity (measured in tonnes/10,000 US$); PC G ( 2 P C G ) stands for per capita GDP (squared) (measured in 1,000 US$); P is population size (measured in billions); 0  is a constant term; and  is a random error term.
An endogenous problem due to the correlation between the independent variable and error term results in biased estimates when conventional panel data estimation methods, such as pooled ordinary least squares (OLS), fixed effect, and random effect, are adopted. However, the GMM estimator proposed by Arellano and Bond [30] and developed by Arellano and Bover [31] and Blundell and Bond [32], provides a solution to the endogeneity and also controls individual-and time-specific effects.
Two different types of GMM estimators are usually used in the literature: (i) difference GMM (i.e., first-difference GMM and orthogonal-difference GMM) developed by Arellano and Bond [30]; and, (ii) system GMM (Arellano and Bover [31] and Blundell and Bond [32]). The main difference between the two methods is in the choice of the endogenous variables and corresponding instrumental variables. As reported by Sung et al. [33], the system GMM estimator is theoretically more efficient than the difference GMM estimator because it allows for a richer set of instruments.
The difference GMM estimator may also suffer from weak finite sample bias when the cross-section dimension (N) is large enough. In our sample, the cross-section dimension (N=110) and time dimension (T=8). In this paper we adopt the system GMM estimator.
The efficiency of the system GMM depends on the following two specification tests: (i) the difference-in-Hansen test for too many instruments; and (ii) the Arellano and Bond test for second-order autocorrelation (i.e., AR(2) test). When employing our data, the analysis suggests that there is no instruments proliferation or we do not have too many instruments.

Data
We use a balanced panel data set for 110 countries, with eight data points covering 1980-2015 corresponding to 1980,1985,1990,1995,2000,2005,2010, and 2015. Prior to the Kyoto Protocol adopted in 1997, CO2 emissions where not much of a concern.
The 110 countries are classified into four groups based on their per capita gross national income (GNI) calculated using the World Bank Atlas method 2016 [34]: lowincome countries (LI countries, less than $1,005), lower-middle-income countries (LMI countries, $1,006-$3,955), upper-middle-income countries (UMI countries, $3,956-$12,235), and high-income countries (HI countries, more than $12,236). The lowincome subpanel in this study consists of data for 10 countries, while the lower-middle-income, upper-middle-income, and high-income subpanels comprise data for 28 countries, 28 countries, and 44 countries, respectively (see Table A2 in Appendix A).
The data on CO2 emissions from fuel combustion are collected from the CO2 Emissions from Fuel Combustion 2017 published by the International Energy Agency (IEA) [35], while IEA's World Energy Balances [36] provides the data on energy consumption across the globe. In this study, we use primary energy supplied defined as energy production plus energy imports plus increase in stocks (which is negative if the stock decreases), minus energy exports and international bunkers. In addition, the data on GDP and population are obtained from the World Development Indicators (WDI) published by the World Bank [2], in which the data on GDP are in 2010 constant prices.

Decomposition results of global CO2 emissions changes
The decomposition results are depicted in Although change in energy intensity slowed down global emissions by 98.9%, the change was overwhelmed by changes in income (148.9%) and population (48.5%) resulting in an increase in global emissions of 103.1% (see Fig. 2b) -see also Table A4. It would be interesting to see the relationship between CO2 emission growth and driving factors at different periods. For this, we divide the entire period  into two periods: early period (1980)(1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990) (1980)(1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990). It is interesting to note that despite the rapid expansion of renewable energy more recently, its share in the global energy supply mix is still small and has not contributed much to limit global CO2 emissions.

Decomposition results in countries with different income levels
What role would different drivers play across different groups of countries to influence CO2 emission growth? Figure 5 plots the impacts of various factors on CO2 growth in different groups of countries. Change in the income level was the primary factor of increasing emissions in the countries with high income levels, such as the HI, UMI, and LMI countries; while change in population size was the main effect behind the increase in emissions for the countries with low income levels, such as the LI countries. Improved energy efficiency helped to slow down emission growth in all groups of countries in both periods. Its effect is more prominent in the HI and UMI groups of countries. In addition, changes in energy mix were a significant factor in reducing CO2 emissions in HI countries.

The estimation results
The results of the analysis using the GMM technique are presented in Table 1. In the table, column 1 presents the empirical results for 110 countries, while columns 2-5 depict the empirical results for the HI, UMI, LMI, and LI groups of countries, respectively. As shown in the bottom of Table 1, the null hypotheses of the Hansen and AR (2) tests cannot be rejected, indicating that the instruments remain valid and that there is no evidence for second-order serial correlation.
The coefficient of the lag stock of CO2 (i.e., 2 1 ln it CO  ) is positive and strongly significant. If countries emitted large amounts of CO2 in the past, then they are likely to continue emitting large amounts of CO2 in the future in the absence of any policy intervention; this finding is consistent with Lee et al. [38].

Results from the 110-country panel
The estimated coefficients of ln GDP and   2 ln GDP are positive and negative, respectively, suggesting the presence of the EKC curve. Note that the EKC is a phenomenon of the long run [39,40], further suggesting that the countries with different income levels should follow a long-term strategy that achieves the targeted balance between reduced CO2 emissions and high economic growth.
The results further indicate that, for the whole sample of 110 countries, the estimated coefficients of ln EC , ln ECS , ln E I , ln P C G , and l n P are positive and significant at a 1% significance level. Similar to the decomposition analysis presented

Differences in impacts across income levels
The theoretical EKC is an inverted U-curve, with the coefficients of ln GDP and    The effects of the five factors on emissions vary (see Fig. 8). The common factors are the effects of energy intensity, income growth, and population growth, the former mitigating CO2, whereas the latter two increasing CO2.

Time and the importance of the various factors
We

The 110-country panel
According to the first column in Table 2, for the 110-country panel, the estimated coefficients of ln EC , ln ECS , ln E I , ln P C G , and l n P between 1980 and 1990 are positive and significant at a 1% significance level. GDP per capita and population were the primary drivers leading to an increase in CO2 emissions between 1980 and 1990, while the factors that slowed the increase of the emissions over the same period were emission coefficient, energy mix, and energy intensity.
The first column in Table 3     However, it is noteworthy that over time, the impact of changes in energy mix and economic growth, i.e., ln ECS and ln PCG , would become more significant in developing countries, especially in the UMI countries.

Comparing the two approaches: LMDI versus system GMM
To further compare the two methods, we calculated the marginal impacts of various factors on CO2 growth using Eq. (11); the detailed results are listed in Table A8. We compare the results estimated using the LMDI and the system GMM methods through the comparison of the marginal impacts of various factors on CO2 growth. The same marginal impact is calculated once under the LMDI method and once under the GMM method. We do this for the whole sample but also for the different income groups (Table 4). From this table, we can see that the level of the marginal impacts of various factors on CO2 growth obtained by the LMDI method are significantly bigger than that of the GMM. This may be because, as addressed previously (i.e., section 2), the system GMM estimator takes into account each country's characteristics when calculating the marginal impact whereas the LMDI method ignores the individual characteristics. The difference in the results between LMDI and GMM are important to understand how CO2 growth responds to change in a given driver. Since the LMDI results are biased towards the country size (i.e., they are skewed due to China and the United States, the larger countries' results), the response indicators (or marginal impact here) based on LMDI may be biased. The response indicators estimated based on GMM analysis are more credible.

Conclusions
In this study, we measured the role of various factors that have driven CO2 emissions over the past 35 years at the global level and different groups of countries by their current income level (i.e., high income, upper-middle income, lower-middle income, and low income). We employ data from 110 countries in two different techniques: Index decomposition technique (identity analysis) and an econometric technique. We also examined if a factor has influenced CO2 growth differently during different time intervals: early period of the study horizon  and late interval of the study horizon (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015).
The results from our identity and econometric analysis reveal that income (measured in terms of per capita GDP) growth and population growth are the major drivers for the increased CO2 emissions during the period. Improved energy efficiency (i.e., decreasing energy intensity of GDP) has slowed down the growth; otherwise, CO2 emissions would have increased even further. Improvement of energy efficiency has contributed in all groups of countries to slow down their CO2 emission growth. This factor is more prominent in the highincome and upper-middle-income groups of countries. The fuel mix factor or the increasing share of low carbon (e.g., natural gas) or no carbon (e.g., renewable) fuels in the total primary energy supply mix contributed to slow down CO2 growth in the high-       Note: EC, ECS, EI, PCG, and P refer to emission coefficient, energy mix, energy intensity, income, and population, respectively. Note: EC, ECS, EI, PCG, and P refer to emission coefficient, energy mix, energy intensity, income, and population, respectively. Note: EC indicates emission coefficient; ECS denotes energy mix; EI describes energy intensity; P C G stands for per capita GDP; and P is population size.