Carbon Emission Intensity, Economic Development and Energy Factors in 19 G20 Countries: Empirical Analysis Based on a Heterogeneous Panel from 1990 to 2015
The increasing global climate problem caused by excessive carbon emissions results in global carbon emission reduction governance becoming a top priority and requires close international coordination. Group of Twenty (G20) is gradually becoming the leading agency of global carbon emission reduction governance, but the unbalanced development among G20 countries has hindered the full play of G20's role. Thus, this paper aims to examine the interrelationships among economic development mode,
... ic development level, and energy factors including energy use efficiency and structure in 19 G20 countries over the period 1990-2015. Considering the panel heterogeneity and the endogeneity of variables, a series of heterogeneous panel analysis techniques are employed in this paper. The empirical findings suggest that for the panel, the improvement of energy use efficiency and the optimization of energy use structure can help to achieve a low-carbon development mode, implying that some international agreements such as the Copenhagen Accord and Paris Agreement on Climate Change are necessary, binding, and effective. However, for individuals, energy factors and development level influence development mode differently across countries, revealing that each country should formulate specific policies that are consistent with its own actual situation. Finally, this paper discusses the role that G20 can play in the global carbon emissions reduction governance, which provides a reference for global low-carbon and sustainable development. 2 of 26 hard work to employ appropriate methods to describe the quantitative relationships among the three variables. Normally, most investigators tend to use econometric analysis techniques to complete this work, because of the objectivity of the results and the rigor of the theory [9, 10] . Among the studies focusing on describing the quantitative relations among CO 2 emissions, economic growth, and energy consumption, gross national product (GNP) was first adopted to measure the economic growth [1, 2] . After that, gross domestic product (GDP) was widely used as the representative variable of economic growth  . Besides, some studies used per capita GDP (PCGDP) to express economic growth [4, 7] . For energy consumption, the early researchers usually applied the total energy consumption (TEC) or TEC per capita in their studies [6, 9] , whose results were a bit indistinct and lack more in-depth descriptions. Hence, some scholars began to use more specific variables to represent energy consumption, like total fossil energy consumption (FEC) [10, 11] , non-renewable energy consumption (NEC)  , and renewable energy consumption (REC)  . Furthermore, some researchers took the selected type of energy consumption as the analyzed variable, such as coal consumption  , natural gas consumption , nuclear  , and electricity  , making the results more concrete and more realistic to reflect the specific relations among CO 2 emissions, energy use, and economic growth. For CO 2 emissions, most investigators employed the total CO 2 emissions from energy  , while some other scholars considered the total anthropogenic CO 2 emissions  or the CO 2 emissions per capita  . In fact, no matter which variable is used, a reasonable explanation of the results is essential. There are two sets of common econometric methods used to describe the relationship among CO 2 emissions, energy use, and economic growth, namely time-series analysis methods and panel analysis methods. Every set of methods has been widely used by many researchers for different research objects. For time-series analysis methods, which aim to describe the quantitative relationship among the variables of one object, like China , America  , and Japan , a unit root test is necessary before investigating the nexus among the variables, in order to prevent spurious regression. After that, the two kinds of cointegration tests, namely the E-G cointegration test used for two variables and the Johansen test used for over two variables, are widely adopted to examine the long-term equilibrium among the variables. Finally, the causal relationships among variables are tested via Granger causality test technique    . Furthermore, some advanced approaches, like Zivot and Andrews (ZA) unit root test, autoregressive distributed lag (ARDL) approach, Gregory and Hansen cointegration test, and Toda and Yamamoto (T-Y) causality test, are also widely used in some specific situations [9, 27] . For panel analysis methods, which aim to examine the quantitative relationship among the variables in multiple objects, like EU countries , BRICS countries , APEC countries , Central American countries  , and so on [6, 29] , the cross-sectional dependence test [30, 31] should firstly be completed, and then the two types of panel unit root test methods, namely the first and the second generation panel unit root tests, will be performed according to the results of the cross-sectional dependence test, in order to explore whether the panel data are stationary      . Next, the presence of a cointegration relationship among variables is usually examined by Pedroni panel cointegration test [36, 37] . Once the presence of cointegration relationship is proven, several estimation methods, like ordinary least square (OLS), dynamic OLS (DOLS), and fully modified OLS (FMOLS) models, are widely applied to obtain the exact nexus among variables. Normally speaking, the likelihood ratio (LR) test and the Hausman test can be performed to define the form of the panel model before using the OLS approach to estimate the panel model  . However, because of the endogenous issues and serial correlation problems, the OLS approach is sometimes inaccurate  . Therefore, the FMOLS and DOLS methods are employed to obtain an accurate estimation result in some studies [12, 39] . Besides, several panel Granger causality test approaches, like panel stacked Granger causality test, panel vector error correction model (VECM) causality test, and Dumitrescu-Hurlin panel causality test, are applied by many scholars [39, 40] , according to the characteristics of examined panel data. In order to reflect the relevant research more intuitively, the main research results are listed in Table 1 , in which the methods, sample, and main findings are reported.