CEEP-BIT WORKING PAPER SERIES. Is China s carbon reduction target allocation reasonable? An analysis based on carbon intensity convergence

Transcription

1 CEEP-BIT WORKING PAPER SERIES Is China s carbon reduction target allocation reasonable? An analysis based on carbon intensity convergence Yu Hao Hua Liao Yi-Ming Wei Working Paper 71 Center for Energy and Environmental Policy Research Beijing Institute of Technology No.5 Zhongguancun South Street, Haidian District Beijing September 2014 This paper can be cited as: Hao Y, Liao H and Wei Y-M Is China s carbon reduction target allocation reasonable? An analysis based on carbon intensity convergence. CEEP-BIT Working Paper. The authors acknowledge the financial support from the National Natural Science Foundation of China ( ) and Fundamental Research Fund of Beijing Institute of Technology ( ). The views expressed herein are those of the authors and do not necessarily reflect the views of the Center for Energy and Environmental Policy Research by Yu Hao, Hua Liao and Yi-Ming Wei. All rights reserved.

2 The Center for Energy and Environmental Policy Research, Beijing Institute of Technology (CEEP-BIT), was established in CEEP-BIT conducts researches on energy economics, climate policy and environmental management to provide scientific basis for public and private decisions in strategy planning and management. CEEP-BIT serves as the platform for the international exchange in the area of energy and environmental policy. Currently, CEEP-BIT Ranks 82, top 5% institutions in the field of Energy Economics at IDEAS( and Ranks 104, top 5% institutions in the field of Environmental Economics at IDEAS ( top/top.env.html). Yi-Ming Wei Director of Center for Energy and Environmental Policy Research, Beijing Institute of Technology For more information, please contact the office: Address: Director of Center for Energy and Environmental Policy Research Beijing Institute of Technology No.5 Zhongguancun South Street Haidian District, Beijing , P.R. China Access: Tel: Fax: Website:

3 Is China s Carbon Reduction Target Allocation Reasonable? An Analysis Based on Carbon Intensity Convergence Yu Hao a,b, Hua Liao a,b, Yi-Ming Wei a,b,* a Center for Energy and Environmental Policy Research, Beijing Institute of Technology, Beijing , China b School of Management and Economics, Beijing Institute of Technology, Beijing , China Abstract: To curb CO 2 emissions, the Chinese government has announced ambitious goals to reduce the CO 2 intensity of GDP, and the total target has been allocated to all Chinese provinces during the twelfth Five-year Plan period ( ). Although setting the target allocation plan is an efficient way to achieve this goal, some key questions, including how the plan is designed, remained unanswered. From an economic perspective, this requires us to test for the existence of convergence in the CO 2 intensity of GDP because the convergence is one of the most important intrinsic economic characteristics that policy makers should take into account: if the convergence exists, the provinces with a higher CO 2 intensity of GDP tend to experience a more rapid reduction in the intensity and therefore could share a heavier burden of the intensity reduction. The existence of stochastic convergence and β-convergence is verified by employing different estimation methods and using various estimation specifications. As a result, the direct policy implication is that provinces with high CO 2 intensity should be assigned tougher reduction targets to cut CO 2 intensity at higher speeds, while the provinces with low carbon intensity should be allowed to reduce the CO 2 intensity at a relatively lower speed. Because some social and economic indicators such as GDP per capita, industrial structure and population density may influence CO 2 intensity, the policy makers should take all these factors into consideration to design reasonable reduction target allocation plan. Key Words: CO 2 intensity of GDP; convergence; China; panel data * Corresponding author addresses: haoyuking@gmail.com (Yu Hao), liaohua5@gmail.com (Hua Liao), wei@bit.edu.cn (Yi-Ming Wei) -1-

4 1. Introduction In recent years, as China s economy developed rapidly, China s carbon dioxide (CO 2 ) emissions have boomed. According to the statistics of Carbon Dioxide Information Analysis Center (CDIAC), since 2007, China has overtaken the U.S. to become the largest CO 2 emitter in the world, and in 2010, China s CO 2 emissions contributed as much as one-quarter of global emissions. The huge amount and rapid growth of CO 2 emissions have drawn increasing attention from the international community and have brought increasing pressure to the Chinese government, especially at this critical time when global warming has become a serious threat to human beings. The Chinese government has indeed heeded the mounting pressure and taken action. In recent years, the Chinese government has set several carbon control targets. However, partly due to worries that economic growth would be severely affected, China s carbon control targets address the carbon intensity of GDP. For instance, in 2009, China s central government announced an ambitious target to reduce its CO 2 intensity of GDP (the ratio of total CO 2 emissions to GDP) by 40%-45% of its 2005 level by In 2011, the State Council of China drew up a plan for CO 2 emissions reduction during the twelfth Five-Year Plan period ( ). In this plan, China vowed to decrease its CO 2 intensity by 17% of its 2010 level by the year Additionally, this plan for the first time disaggregates the national target into separate targets for each province. 2 1 As a reference, see It should be noted that carbon intensity targets are less binding compared with the total emissions targets because carbon intensity would decrease as long as GDP grows more rapidly than absolute emissions. 2 There are at present 23 provinces, four Centrally Administered Municipalities, and five autonomous regions in -2-

5 According to the official target, Guangdong, a wealthy southeast coastal province with an average CO 2 intensity of GDP of 1.45 ton/10000 yuan (at 2000 constant price) between 2009 and 2011 should cut its CO 2 intensity of GDP by 19.5 percent, the highest target of all the provinces. In contrast, Qinghai, a less-developed western province with an average CO 2 intensity of GDP of more than three times of Guangdong s level for the same period, only needs to cut its CO 2 intensity of GDP by 10 percent, the lowest target among all the provinces. 3 Although the State Council of China did not reveal the reasons for or the basis of calculation for the provincial disaggregated target, some pilot studies have been made on the reasonability and feasibility of provincial allowance allocation (Yi et al., 2011; Wei et al., 2012; Wang et al., 2013). Yi et al. (2011) developed a factor model to explore the importance of several key factors that determine the provincial deduction ability, and they designed a reduction target allocation plan based on the model s estimation results. Wei et al. (2012) and Wang et al. (2013) utilized a Data Envelopment Analysis (DEA) approach to estimate each province s efficiency of the CO 2 intensity of GDP, which determines the deduction allowance for each province. The basic principle of the DEA approach is that provinces with higher carbon emissions efficiency should shoulder a heavier burden of the carbon intensity cuts. For instance, according to Wang et al. s (2013) calculation, the more economically developed provinces, such as Beijing, Tianjin, Shanghai and Jiangsu, also generally have higher carbon emission efficiency. Therefore, these provinces should undertake more reduction tasks in reducing carbon intensity. China. For simplicity and consistency, throughout this study, the term province is used to represent all of these sub-national administrative entities that are administratively equal. 3 The detailed reduction targets for all provinces and each province s average CO 2 intensity of GDP between 2009 and 2011 are presented in Table A1 in the Appendix. -3-

6 One important potential flaw of previous studies is that an innate principle of economic development might be simply ignored: there may be convergence in the CO 2 intensity of GDP across provinces. In economic studies, the original concept of convergence indicates that the poor economies have a tendency to catch up with rich economies in terms of per capita income. This concept was first raised in Solow s (1956) seminal study on neoclassical growth theory. There are three main notions of convergence: β-convergence, σ-convergence and stochastic convergence. In the context of this study, β-convergence implies that the provinces with higher CO 2 intensity of GDP would experience a more rapid decrease in carbon intensity compared with the provinces with lower carbon intensity; σ-convergence means that the dispersion of the CO 2 intensity of GDP diminishes over time. Both β-convergence and σ-convergence could be further classified into two types: conditional convergence and absolute (unconditional) convergence. If the convergence takes place after controlling for province-specific characteristics other than the targeted indicator (in our case, the CO 2 intensity of GDP), the convergence is conditional; otherwise, it is absolute (unconditional). On the contrary, stochastic convergence indicates that the shocks in CO 2 intensity of GDP relative to the sample average are only temporary. The most common method used to examine the stochastic convergence is the unit root test. 4 To date, there has been a growing body of research on the topic of convergence in pollutant emissions across different countries. Carbon dioxide is the pollutant that has received the most 4 To date, various types of unit root test have been developed for stochastic convergence. In recent years, the unit root tests for panel data and tests incorporating structural breaks have attracted increasing attention. For instance, in a recent study, Romero-Ávila (2008) reviewed the unit root tests used to test for the stochastic convergence of CO 2 emissions per capita and utilized a bootstrap method that allows for an unknown number of structure breaks to investigate the stochastic and deterministic convergence of CO 2 per capita across industrialized countries. -4-

7 attention, partly because global warming is generally considered to be an increasing concern to the international community. After Strazicich and List s (2003) pioneering work, there is a growing body of literature on the existence of convergence in per capita CO 2 emissions across different groups of countries. The representative studies include Ezcurra (2007), Romero-Ávila (2008), Westerlund and Basher (2008), Jobert et al. (2010), Ordás Criado and Grether (2011), and Camarero et al. (2013). Strazicich and List (2003) and Ezcurra (2007) utilized panel unit root tests and cross-sectional/panel regressions to test for stochastic and conditional convergence. Ezcurra (2007) employed non-parametric techniques (stochastic kernels) to examine cross-country convergence in per capita CO 2 emissions by estimating the dynamic distribution of CO 2 emissions; Jobert et al. (2010) utilized the Bayesian shrinkage estimation approach to estimate the speed of convergence of per capita CO 2 emissions for each of 22 European countries. Using a long-run panel data of OECD countries between 1960 and 2008, Camarero et al. (2013) studied convergence in CO 2 emission intensity based on its determinants, and their results indicate that differences in emission intensity convergence are more determined by differences in convergence of the carbonisation index rather than by differences in the dynamic convergence of energy intensity. Recently, Pettersson et al. (2014) reviewed research on convergence of carbon dioxide emissions among countries. They found that the existing estimation results are sensitive to the choice of econometric approach and data set. Two very recent studies have focused on the convergence of CO 2 emissions in China. Huang and Meng (2013) utilized a specification of the panel regression model with full consideration of spatial-temporal dependency to verify the convergence of per capita CO 2 emissions across Chinese provinces. Wang and Zhang (2014) employed panel unit root tests and a panel regression model to find evidence for the existence of -5-

8 stochastic convergence and β-convergence. Besides, given the important policy implications, in recent years an increasing number of researches have investigated the convergence in energy intensity across countries or within a specific country. The representative researches include Markandya et al. (2006), Liddle (2009, 2010), Le Pen and Sévi (2010), Stern (2012), Herrerias (2012), Herrerias and Liu (2013), and Adhikari and Chen (2014). Markandya et al. (2006) verified the existence of β-convergence in energy intensity toward the EU average across transition countries. Liddle (2009, 2010) found evidence for σ-convergence and γ-convergence among different groups of countries. Contrarily, Le Pen and Sévi (2010) utilized relatively newly developed pair-wise econometric approach to reject the hypothesis of global stochastic convergence in energy intensity among 97 countries during the period Using a stochastic production frontier, Stern (2012) modeled energy efficiency trends and verified its convergence in 85 countries over a 37-year period. Herrerias (2012) utilized a distribution dynamics approach to find evidence for the convergence in the intensity of various energy sources in a group of countries. Adhikari and Chen (2014) employed spatial panel data approach to examine the convergence in Asian countries. They found weak evidence of σ-convergence for all sample countries, while their estimation results support the existence of β-convergence for the whole set of sample countries and North East Asia. In a recent study, Herrerias and Liu (2013) for the first time examined the convergence in electricity intensity in China using provincial panel data. They found evidence for club convergence: the convergence pattern exists within groups of regions, while there are regions that still diverge. -6-

9 The existing literature concerning carbon convergence mainly focused on the convergence of per capita CO 2 emissions. To the best of our knowledge, there has been very few research to date on the convergence of CO 2 intensity of GDP, although some researches on the convergence in energy intensity have been conducted. As a result, the main contribution of this study is to examine whether there is β-convergence in the CO 2 intensity of GDP across Chinese provinces for the first time. Because the CO 2 intensity of GDP is broadly used as an indicator of carbon efficiency, this research would shed some light on several important policy issues for China. First of all, because the official target of reducing CO 2 emissions focuses on carbon efficiency rather than an absolute emissions level, the β-convergence in the CO 2 intensity of GDP, if it indeed exists, has important policy implications for China s CO 2 control: the provinces with low carbon efficiency (a high level of carbon intensity) could shoulder a higher reduction allowance because they have a tendency to experience a more rapid reduction in carbon intensity. Second, the existence of convergence in carbon intensity is a necessary condition for the existence of the turning point for the national total CO 2 emissions (Jobert et al., 2010). If the convergence in carbon intensity across provinces does not exist, some provinces with positive and high growth rates of carbon intensity would see increasing carbon intensity over time. Eventually these provinces will dominate the provinces with low and negative growth rates of carbon intensity in terms of carbon intensity, hence the carbon intensity and therefore total CO 2 emissions for the whole nation would continue to grow without any turning point. Additionally, the convergence in CO 2 intensity of GDP is to some extent coherent to the convergence in GDP per capita. Because the CO 2 intensity of GDP to a certain degree reflects the efficiency of energy usage and technology level, the convergence in -7-

10 carbon intensity suggests that the energy efficiency and technology progress in relatively poor provinces would catch up with richer provinces, which implies a favorable trend of decreasing regional difference in economic development. In addition, to examine the reasonability of carbon intensity reduction plan in China also has international implications. Given China s considerable contribution to global CO 2 emissions, China s effort to limit CO 2 emissions is essential to address the problem of climate change. Considering the fact that China is a typical unitary country with powerful central government, the reasonable plan for carbon intensity reduction allocation among provinces would restrain the CO 2 emissions in China effectively and therefore help to deal with global warming more efficiently. Specifically, as China has made public commitments to see its turning point in CO 2 emissions by 2030 during the APEC summit in November 2014, the reasonable and optimal CO 2 reduction allocation plan would undoubtedly help China to achieve this ambitious goal. 5 The remainder of this paper is arranged as follows. In section 2, the estimation methodology and data utilized are briefly introduced. In section 3, the estimation results are reported and discussed. In section 4, further discussions about the estimation results are made. Finally, section 5 concludes the study, and corresponding policy implications are raised. 2. Methods 2.1. Methodology 5 As a reference to China s commitment to peak CO 2 emissions by 2030, see -8-

11 Because the main purpose of this research is to investigate whether there is catch-up in the CO 2 intensity of GDP across Chinese provinces (i.e., whether the provinces with higher level of carbon intensity would experience a more rapid reduction of the carbon intensity), the main methodology is to employ the appropriate regression model to test for β-convergence. The early literature on β-convergence generally utilized cross-sectional data such as country-level GDP and the GDP growth rate (Baumol, 1986; Delong, 1988; Barro, 1991). However, as Jobert et al. (2010) stressed, the cross-sectional analysis bears a significant drawback: only the initial and the final periods of the sample are considered in the empirical study; therefore, rich information between these two periods is simply ignored. Such ignorance may harm the credibility of the estimation results and sometimes even lead to biased estimations (Maddala, 1999). Compared with a conventional cross-sectional analysis, panel data analysis has more freedoms and is able to take consideration of provinces heterogeneity into the convergence regression. More importantly, the panel data framework makes it possible to correct omitted variable bias by controlling for the individual (provincial) effects, which is the biggest merit of panel data analysis (Islam, 1995). As a result, in this study, panel data analysis is utilized. Following the conventional estimation approach for convergence analysis, the benchmark regression model is to regress the growth rate of the CO 2 intensity of GDP on its original level. Under the framework of panel data, when the time lag is chosen as one year, the regression model is as follows: ln( c c ) ln( c ) z (1) it it it 1 it 1 it -9-

12 where c represents the CO 2 intensity of GDP and z stands for the control variables that may affect the growth rate of the CO 2 intensity of GDP. The subscripts i and t represent the specific province i and the time period t, respectively. The coefficient that we are most interested in is β 1 because a significant negative estimate of β 1 with the value between -1 and 0 indicates that the β-convergence exists (Romer, 2011). The absolute convergence is tested for when the control variables z are absent; otherwise, the conditional convergence is estimated. As noted by Jobert et al. (2010), the choice of control variables is key to determining conditional convergence because the different choices of control variables may cause discrepant estimation results. In this study, GDP per capita, population density and the industrial structure indicator (the ratio of secondary industry s value added to GDP) are selected as control variables. The choice of GDP per capita is to control for the level of economic development of a province because the more rapid economic growth means lower CO 2 intensity of GDP with all else being equal. The introduction of GDP per capita also has a theoretical basis: in a neoclassical model developed by Ordás Criado and Grether (2011) to explain the convergence of pollutant emissions, the initial level of per capital output appears in the regression model after considering the higher-order Taylor expansions. A population factor is often used in a convergence regression for economic indicators or carbon emissions (Islam, 1995; Jobert et al. 2011; Wang and Zhang, 2014). 6 An industrial structure indicator is introduced because of its importance in generating carbon emissions. In the context of China, the carbon emissions from the secondary industry account for approximately 70% of total carbon emissions (Gregg et al. 2008). Eq. (1) is estimated by panel data methods that could 6 In the convergence regression deducted from the Solow growth model, the term (n+g+δ) is often used (generally in logarithmic form) as a regressor (for instance, see Eq. (9) in Islam (1995)). n, g and δ stand for the population growth rate, technology growth rate and capital depreciation ratio. Because in most empirical studies, g+δ is usually set to be as a constant value of 0.05, the population growth rate (n) is the actual variable of the regressor. -10-

13 control for cross-sectional individual effects and heterogeneity (such as the fixed effects (FE) and Feasible Generalized Least Squares (FGLS) approaches). Even after controlling for cross-sectional individual effects and heterogeneity, there might still exist an endogeneity problem, which may be caused by omitted variables and simultaneity. The potential simultaneity arises because GDP per capita is used as an explanatory variable in Eq. (1). In fact, the CO 2 intensity of GDP could also affect economic growth and therefore GDP per capita. To address the potential endogeneity, Eq. (1) is rearranged by adding ln(c it-1 ) to both sides of the equation to get Eq. (2): ln( c ) ln( c ) z (2) it 0 1 it 1 it 1 it where γ 1 =1+β 1 Like in Eq. (1), z is a vector that contains all of the control variables. Note that Eq. (2) is a dynamic equation because the first-order lag of the dependent variable is used as an explanatory variable. The traditional approach to address the endogeneity problem is to use instrumental variable (IV) estimates. However, due to the difficulty in finding proper IVs, we utilize the Generalized Method of Moments (GMM) method developed by Arellano and Bond (1998). GMM is essentially an IV estimation method because it uses predetermined and exogenous variables as instruments to account for the potential dynamics in the determination of the dependent variable and address potential endogeneity caused by reverse causality. In Eq. (2), we assume that the lagged dependent variable as well as GDP per capita and the industrial structure indicator are endogenous and treat population density as an exogenous variable. There are two main GMM methods, namely, first-difference GMM and system GMM. The main -11-

14 difference between the two notions is whether the regressors are at first differenced to eliminate the initial effect of efficiency. The technical details of the two GMM approaches could be found in Bond et al. (2001) (first-difference GMM), Bover (1995) and Blundell and Bond (1998) (system GMM). After β 1 or γ 1 is estimated, the convergence speed λ (i.e., the adjustment speed of a province s CO 2 intensity of GDP toward its steady-state level) can be obtained as ln 1 ln( ) (3) 1 1 Note that γ 1 =1+β 1. The half-life period, or the time needed for a province to shorten the distance between the current situation and the steady state by half, could be calculated by dividing approximately 0.69 by the convergence speed λ (Romer, 2011). Therefore, the higher convergence speed corresponds to a shorter half-life. To sum up, to test for the convergence of CO 2 intensity of GDP, Eq. (1) is first estimated with conventional panel data approaches by controlling for individual provincial effects and cross-sectional heterogeneity, and then the dynamic Eq. (2) is estimated with GMM methods. The convergence speed λ and corresponding half-life are calculated on the basis of the estimates of β 1 and γ 1. The estimation and calculation results from different estimation approaches are compared as a robustness test of our empirical study. In addition, the stochastic convergence of the CO 2 intensity of GDP is tested by employing various -12-

15 panel unit root tests to check whether the CO 2 intensity of GDP is present in the I(1) process. If so, the shocks influencing the intensity are permanent, and therefore, stochastic convergence does not exist. Contrarily, if the series follow the I(0) process, the shocks to the intensity are only transitory, which implies that stochastic convergence exists because the policy of CO 2 intensity reduction is less necessary. Additionally, as Lee and Chang (2008) noted, the stochastic convergence is a necessary but not sufficient condition for β-convergence; therefore, testing for stochastic convergence could be treated as a necessity for the examination of β-convergence. 2.2 Research framework To show the research focus of this study in an intuitive way, in Figure 1, we represent the innate logic of the stochastic convergence and β-convergence that are discussed and tested in this empirical study, and the corresponding regression estimations used to examine these convergences are also briefly interpreted. -13-

16 Figure 1. Sketch of research framework Shocks to CO 2 intensity are only transitory More responsibility for Stochastic convergence Necessary condition β-convergence the provinces with high CO 2 intensity CO 2 intensity reduction allocation plan Unit root tests Other economic and social factors also affect convergence speed Multiple regressions analysis 2.3 Data To gauge the CO 2 intensity of GDP, the CO 2 emissions and GDP data should be known. Because there have been no official statistics for China s CO 2 emissions to date, we have to conduct the calculation ourselves. Because the majority of CO 2 emissions in China are generated from the combustion of fossil fuels and the production of cement, the CO 2 emissions could be calculated by utilizing the following formula: m i ij ij ij ij ij i j 1 EC F CF CC SC OF CE (4) where the subscripts i and j represent province i and the energy of type j, respectively. EC stands for CO 2 emissions, F is the fuel consumption, CF is the fuel s calorific factor, CC is the fuel s potential carbon emission factor, SC is the fuel s carbon sequestration, and OF is the fuel s fraction of carbon oxidized. The first item on the right side of Eq. (4) computes the CO 2 emissions -14-

17 produced due to fossil fuel combustion in province i, which follows the suggestions of IPCC (2006) and is frequently used in other studies concerning China s CO 2 emissions (Wei et al., 2010; Huang and Meng, 2013; Wang and Zhang, 2014). However, most previous studies simply ignored the CO 2 emissions generated in cement production (the second term of Eq. (4)). 7 According to Du s (2010) estimation, the CO 2 emissions from cement production account for approximately 10% of the total CO 2 emissions in China; therefore, ignoring the CO 2 emissions from cement production would underestimate the total CO 2 emissions and most likely cause biased estimates in empirical studies. Provincial energy consumption statistics are recorded in China Energy Statistical Yearbooks, and the provincial cement production data are from the China Economic Information Network. Because there are no statistics for provincial energy consumption before 1995, our sample period is from 1995 to The provincial GDP data are collected from China Statistic Yearbooks and converted into real GDP at 2000 constant prices. Because the CO 2 intensity of GDP equals the ratio of CO 2 emissions to GDP, we calculate the intensity for each province during the sample period, which is plotted in Figure 2 as follows. 7 During the production of cement, CO 2 emissions are generated as the carbonates are resolved by heat. The corresponding chemical equations are CaCO 3 = CaO + CO 2 and MgCO 3 = MgO + CO

18 Figure 2. CO 2 intensity of GDP for each province, (tons CO 2 /10000 yuan at 2000 price) Note: The series are calculated by authors. Tibet is excluded because the energy consumption data are unavailable. Because Chongqing became a municipality city only in 1997, it is merged into Sichuan province to keep the continuity of the series. An initial glance at Figure 2 permits several preliminary observations. First of all, there is an apparent trend of downward convergence in the CO 2 intensity of GDP across provinces because the mean of CO 2 intensity decreased by 33%, from 5.57 in 1995 to 3.75 in After excluding the outlier of the Ningxia province, the reduction is as high as 38%. Second, the changing trend of CO 2 intensity for different regions and provinces is quite different. For instance, the province with the highest CO 2 intensity of GDP in 1995, Shanxi, saw a reduction in the intensity by 46.7%, while Hainan, the southeastern island province that had the lowest original CO 2 intensity of GDP at 1995 experienced a significant increase in the CO 2 intensity of GDP by 69.2%. If we look into the situations of different regions, the eastern region (middle region) had the lowest (highest) original CO 2 intensity and saw the most rapid decrease in intensity; the western region s CO 2-16-

19 intensity of GDP was slightly lower than that of middle region, but the reduction rate of the intensity was lowest there among all three regions (as shown in Figure 3). As a result, it seems that there is no convergence across different regions. It is necessary to utilize a more formal framework to test the existence of the convergence nationwide. Figure 3. The CO 2 intensity of GDP for three regions and the whole country, (tons CO 2 /10000 yuan at 2000 price) Note: The same as in Figure 2. As mentioned previously, following the studies concerning CO 2 emissions, the control variables used in this study include GDP per capita, population density and an industrial structure indicator. The basic data are collected from China Statistical Yearbooks (various years), and the population density and industrial structure indicator are calculated accordingly. The summary statistics of all variables used in the study are shown in Table 1 below. -17-

20 Table 1. Summary Statistics of the variables used in the empirical study Variable Unit Mean Median Maximum Minimum Std. Dev. Obs ln(c t /c t-1 ) % c ton/ 2000-price yuan GDPperc Yuan (constant price) Popden Person/ square-kilometers Indstr % Note: c t denotes CO 2 intensity of GDP at year t, therefore ln(c t /c t-1 ) is the annual growth rate of CO 2 intensity of GDP. GDPperc, Popden and Indstr stand for GDP per capita, population density and industrial structure, respectively. 3. Results 3.1 Stochastic convergence: the shocks to the CO 2 intensity of GDP are not permanent The most commonly used methods to test for stochastic convergence are various unit root tests. In this study, we employ the following unit root tests: the Im, Pesaran and Shin (IPS) test developed by Im et al. (2003), the Augmented Dickey Fuller (ADF) test developed by Said and Dicky (1984), the Phillips Perron (PP) test proposed by Phillips and Perron (1988), and the Levin-Lin-Chu (LLC) test suggested by Levin et al. (2002). The first three tests assume a common unit root process for the whole sample, while the last test presumes that there is a unit root process for each individual. The technical details can be found in corresponding papers. The test results are depicted in Table 2 below. -18-

21 Table 2. Panel unit root test results for stochastic convergence Method IPS ADF PP LLC Individual intercept Individual intercept and trend c (0.000) b (0.043) c (0.000) a (0.061) c (0.000) b (0.019) c (0.000) c (0.000) Note: The series being tested are lnc, the logarithmic CO 2 intensity of GDP. The lag length is automatically selected by Schwarz Information Criterion. p-values of each test are reported in the parentheses. c, b and a indicate significance at the 1%, 5% and 10% levels. As shown clearly in Table 2, at significance level of 5%, all test results except for the ADF test when individual intercepts and trends are included indicate that the null hypothesis of no stationarity should be rejected. When the significance level is chosen to be 10%, then all test results support the rejection of the null hypothesis. As a result, we have sufficient confidence to conclude that the CO 2 intensity of GDP is a I(0) series, and therefore, stochastic convergence exists. The existence of stochastic convergence in the CO 2 intensity of GDP indicates that the shocks to the time path of CO 2 intensity for a specific province to the average level are only transitory. As time passes, the transitory shocks may die out and the provincial CO 2 intensity would tend to move closer to the national average level. As discussed previously, the existence of stochastic convergence is also a precondition of the existence of β-convergence, which will be tested in the next subsection β-convergence: the provinces with higher CO 2 intensity of GDP would see a more rapid -19-

22 decrease in CO 2 intensity As discussed in the methodology section, to test for the β-convergence of the CO 2 intensity of GDP, Eq. (1) is first estimated when the growth rate of the CO 2 intensity of GDP is used as the dependent variable. In Table 3, the estimation results of absolute and conditional β-convergence for several different specifications and regression approaches of Eq. (1) are represented. Table 3. Estimation results for Eq. (1) Dep. Var: ln(c t /c t-1 ) Absolute convergence Conditional convergence (1) (2) (3) (4) (5) (6) Regression approach FE FGLS FE FGLS FE FGLS ln(c t-1 ) c (0.043) c (0.050) c (0.033) c (0.051) c (0.032) ln(gdpperc t-1 ) (0.040) ln(popden) Indstr t Implied convergence speed λ (0.052) c (0.124) c (0.124) (0.198) (0.097) c (0.120) c (0.048) b (0.066) a (0.085) c (0.157) b (0.007) Province fixed effects Yes Yes Yes Yes Yes Yes Time fixed effects Yes No Yes No Yes No Obs Provinces R AIC BIC Note: Robust standard errors are in parentheses. The meanings of the abbreviations are the same as in Table 1. t represents time trend. c, b and a indicate significance at the 1%, 5% and 10% levels, respectively. In the first two columns of Table 2, the absolute convergence is tested. Because there most likely -20-

23 are some province-specific time-invariant factors that may influence the CO 2 intensity of GDP such as energy usage habits, energy endowment, and the topography and climate of the province, the province-fixed effects should be controlled for. In addition, because of the technological progress and policy changes that occur over time that may affect the CO 2 intensity of GDP in all provinces, the time-temporal effect should be accounted for. In the first column, the two-way fixed effects (FE) model that controls both province- and time-fixed effects is utilized. Because of potential autocorrelation within the group, in the second column, the Feasible Generalized Least Squares (FGLS) estimator is employed and the time effect is estimated by introducing the time trend. 8 Because the coefficients of ln(c t-1 ) in the first two columns are significantly negative, the absolute convergence of the CO 2 intensity of GDP indeed exists, and the implied convergence speed is approximately The existence of absolute convergence and the relatively rapid convergence speed suggest that for the CO 2 intensity of GDP, there is a very strong trend toward convergence across provinces, no matter at which economic development level a specific province is found. The estimation results shown in columns 3-6 are for conditional convergence. Similar to the estimation for absolute convergence, two-way FE and FGLS methods are used to estimate the same specifications for a robustness check. In the third and fourth columns, industrial structure is added as a control variable. Because most CO 2 emissions in China are generated in the secondary industry (Wei et al., 2010), the more important the secondary industry in the economy is, the more difficult it is for the CO 2 intensity of GDP to decrease. As a result, the coefficient of the 8 Using the serial correlation test suggested by Woodbridge (2010, pp ), the F-statistic (with two degrees of freedom being 1 and 28, respectively) is , and the corresponding p-value is As a result, the null hypothesis of no autocorrelation within group could be comfortably rejected at the 1% significance level. -21-

24 industrial structure is expected to be positive. The estimation results in the third and fourth columns verify this inference, and the coefficients of ln(c t-1 ) are also significant and negative, implying that the conditional convergence exists after controlling for industrial structure. In the last columns, all three control variables are introduced, and the results also indicate the existence of conditional convergence. These control variables turn out to be significant in the FGLS results (column 6), where within-group autocorrelation is accounted for. In the sixth column, the industrial structure is still positive, as expected; the first lag of GDP per capita is negative, suggesting that richer provinces tend to cut the CO 2 intensity of GDP more rapidly than do poorer provinces. An interesting observation is that the coefficient of population density is negative, which indicates that the concentration of population is favorable to the reduction in the CO 2 intensity of GDP. One possible explanation for this observation is that the concentrated population makes it possible to utilize energy more intensively (for instance, central heating systems for megacities in winter) and therefore reduce CO 2 emissions per capita. The coefficients of ln(c t-1 ) in the last two columns are also significantly positive, which again verify the existence of conditional convergence. Compared with the results of absolute convergence, the convergence speed for conditional convergence is more rapid and the corresponding half-life is shorter. This finding is consistent with the empirical results of studies focused on convergence in the GDP per capita (Barro, 1991, Peng, 2005) and CO 2 emissions per capita (Jobert et al., 2010). Because Eq. (1) is essentially a dynamic equation, it is necessary to transform the equation into Eq. (2) and therefore introduce the dynamics. Moreover, there might be a potential endogeneity problem caused by possible bidirectional causality between the CO 2 intensity of GDP and the -22-

25 control variables (such as industrial structure) and omitted variables. As a result, we also estimate Eq. (2) by employing GMM approaches. The estimation results are shown in Table 4 as follows. Dep. ln(c t ) Regression approach Var: Table 4. Estimation results for Eq. (2) (9) (10) (11) (12) (13) (14) (15) FE DIFF-GMM SYS-GMM DIFF-GMM SYS-GMM DIFF-GMM DIFF-GMM Area National National National National National Eastern ln(c t-1 ) ln(gdpperc t-1 ) c (0.025) a (0.015) ln(popden) Indstr Implied convergence (0.105) c (0.145) c (0.056) a (0.059) (0.445) b (0.509) c (0.053) (0.026) (0.039) a (0.232) c (0.044) c (0.026) b (0.361) c (0.052) (0.028) b (0.304) provinces c (0.096) (0.054) (1.213) Inland provinces c (0.080) (0.059) c (0.547) speed λ Hansen Test A-B test of AR(1) A-B test of AR(2) Obs/ instruments /45 464/93 435/30 464/62 165/30 270/30 Provinces Note: Robust standard errors are in parentheses. For all tests the probabilities are reported. For all test results the p-values are presented. According to the suggestions of National Bureau of Statistics of China, East China includes 11 provinces: Beijing, Tianjin, Hebei, Liaoning, Shanghai, Jiangsu, Zhejiang, Fujian, Shandong, Guangdong and Hainan. All the other 18 provinces in our sample are classified as inland provinces. In models (9)-(13) ln(c t-1 ), ln(popden) and Indstr are treated as endogenous variables, while ln(gdpperc t-1 ) is considered as a predetermined variable. In models (14) and (15) the endogenous variables include ln(c t-1 ) and Indstr, while ln(gdpperc t-1 ) is treated as a predetermined variable. a, b and c indicate significance at the 1%, 5% and 10% levels, respectively. -23-

26 In the first column of Table 4 (model (9)), the estimates of conventional fixed effects (FE) method are consistent with the estimates of model (6) shown in Table 3, which verifies that Eq. (2) is in fact equivalent to Eq. (1). 9 However, because the conventional FE estimator could not address potential endogeneity problem, the GMM estimators are utilized. From the second to the sixth columns of Table 4 (models 10-13), both first-difference GMM and system GMM approaches are employed to estimate two different specifications of Eq. (2) for all 29 provinces nationwide. First, all of the control variables are included, but because the population density turned out to be insignificant, the variable ln(popden) is dropped and the Eq. (2) is estimated with only two control variables. In these four models, the estimated coefficients of ln(c t-1 ) by first-difference GMM are consistently lower than those by system GMM, implying that the corresponding convergence speed calculated by first-difference GMM is higher than that obtained from system GMM approach. Moreover, the implied convergence speed from first-difference GMM estimates is significantly higher than that deduced from FE and FGLS estimates. This finding is in line with Weeks et al. s (2003) studies concerning the β-convergence in GDP per capita across Chinese provinces. Because the province-specific individual effects are eliminated through a differencing calculation, the first-difference GMM could control for provincial individual effects and is therefore preferred. 10 Given the dramatic difference in economic development and the CO 2 intensity of GDP among different regions in China, especially between the eastern and inland 9 The main advantage of the FGLS estimates (model 8 in Table 3) compared with the FE estimates (model 7 in Table 3) is that by FGLS, the within-group autocorrelation is controlled for. After allowing for dynamics by introducing the first lag of ln(c t ) as an explanatory variable (model 9 in Table 4), the autocorrelation is to some extent accounted for. 10 It is noteworthy that Bond et al. (2001) argued that the first-differenced GMM estimator is subject to a large downward finite sample bias, particularly when the number of time-series observations is small. They therefore recommend the system GMM estimator instead. However, in practice, choosing which GMM estimator depends on the issues under investigation. For instance, Weeks et al (2003) compared the estimates of OLS, Within Group (WG), first-difference GMM and system GMM for β-convergence in the GDP per capita across Chinese provinces and finally chose system GMM estimates as benchmark results. Contrarily, Halkos and Paizanos (2013) treated first-differenced GMM estimates as benchmark results because in their research, the country-specific individual effects are the main causes of estimation bias and therefore should be controlled for. -24-

27 provinces (as shown in Figures 1 and 2), the first-difference GMM is also used to test for the convergence in the two regions in China (models 14 and 15 in the last two columns of Table 4). The results indicate that the eastern provinces have higher convergence speed in the CO 2 intensity of GDP compared to inland provinces. This finding is consistent with the decrease in the absolute level of the CO 2 intensity of GDP in three regions in China: during our sample period from 1995 to 2011, the CO 2 intensity of GDP decreased by 42% in eastern provinces, while the percentage of decrease in inland provinces was 38%. These primary results suggest that there might be club convergence in the CO 2 intensity of GDP across different regions of China, as is the case for GDP per capita (Cai et al. 2002; Shen and Ma, 2002; Zou and Zhou, 2007; Maasoumi and Wang, 2008) Sensitivity analysis: extreme values and outlier provinces do not affect the estimation stability To check the robustness of our estimation results, a sensitive analysis is needed. We carry out a sensitive analysis for two reasons: to check the influence of outlier provinces and to examine the stability of the estimation results over different sample periods. First of all, because the differences in the provincial CO 2 intensity of GDP and its growth rate are considerable (see Figure 2), it is necessary to check whether a certain province with extreme values or a distinctive growth path of CO 2 intensity of GDP would significantly affect the estimation results. Accordingly, we exclude the only province that saw an increase in the CO 2 intensity of GDP (Ningxia) and the 11 In early researches on club convergence, the provinces are classified into different clubs on the basis of their geographic locations, such as Cai et al. (2002) and Shen and Ma (2002). This classification method has the merit of simplicity and intuition, but the geographical factor may not capture all of the inner economic characteristics of the provinces in a specific club, therefore, recently, the clustering technique is frequently utilized to classify the provinces in an endogenous way (Zou and Zhou, 2007; Maasoumi and Wang, 2008). As a primary study, the classification is made here based on geographical location. -25-

28 province with highest deduction rate in CO 2 intensity of GDP (Beijing) during the sample period, respectively. The estimation results with the rest of the sample are very similar to the results obtained from the whole sample, as shown in Tables 3 and 4. Moreover, because 14 out of all 29 provinces have experienced an increase in the CO 2 intensity of GDP after 2008, the growth pattern of the CO 2 intensity of GDP might have changed systemically since As a result, we also repeated the estimations for a shorter period Again, the estimation results and implied convergence speed for the shorter period are roughly the same as those obtained for the whole sample period Discussion As a result of the existence of β-convergence in the carbon intensity of GDP across provinces, a reasonable and effective CO 2 emissions reduction allocation target should require the provinces with a higher CO 2 intensity of GDP to decrease the carbon intensity by a higher percentage because the intrinsic economic feature of β-convergence suggests that these provinces would in essence see a higher reduction in the CO 2 intensity of GDP even if there is not any reduction target. As a result, the current allocation plan is obviously inexpedient for some provinces because the plan does not reflect this economically innate principle of convergence. For example, some underdeveloped western provinces such as Ningxia, Inner Mongolia, Shanxi and Guizhou had the highest CO 2 intensity on the eve of the 12th Five-year Plan, but the reduction targets assigned to these provinces are at most moderate under current allocation plan: 16% for Ningxia, 16% for Inner Mongolia, 17% for Shanxi and 16% for Guizhou (see Table A1). These arrangements 12 Due to limitations on space, the estimation results for the sensitive analysis are not depicted in this paper, but they are available upon requests. -26-

29 are contradictory to the innate economic principle revealed by the existence of convergence in the CO 2 intensity of GDP. The consequence would be serious. The western provinces would introduce more energy-intensive and pollution-intensive industries to boost their economic growth because the relatively low carbon reduction targets in effect encourage these provinces to produce far more pollution than they would following the rule of convergence. Moreover, because western China is an ecologically fragile area, the rapid development of the energy-intensive and pollution-intensive industries is extraordinarily harmful to local environment. On the other hand, the economically developed provinces with low carbon intensity have to share much heavier burdens of cutting carbon intensity than they should, which would inevitably affect the economic development in these regions. To sum up, a reasonable carbon reduction allocation plan should be primarily based on the current level of carbon intensity for one specific province. Because the level of economic development and the industrial structure (the relative importance of a secondary industry in the economy) are important to the speed of convergence in carbon intensity, the allocation targets for different provinces with similar carbon intensity can be differentiated when these factors are taken into consideration. In fact, the finding that the richer provinces are more likely to see a more rapid decrease in CO 2 intensity may complicate the policy-making, because more care is needed to set reasonable reduction targets for the rich provinces with relatively low CO 2 intensities. For instance, Beijing, Shanghai, Guangdong and Tianjin, China s main economic powerhouses, had relatively high levels of per capita GDP and considerably low carbon intensity when the 12th Five-year Plan started. For these provinces, -27-