Modelling the nonlinear relationship between CO2 emissions from oil and economic growth

Economic Modelling 29 (2012) 1537–1547

Contents lists available at SciVerse ScienceDirect

Economic Modelling journal homepage: www.elsevier.com/locate/ecmod

Modelling the nonlinear relationship between CO2 emissions from oil and economic growth Kuan-Min Wang ⁎ Department of Finance, Overseas Chinese University, 100 Chiao Kwang Road, Taichung, 40721, Taiwan

a r t i c l e

i n f o

Article history: Accepted 2 May 2012 JEL classifications: C33 C22 O10 Q53 Q56 Keywords: Carbon dioxide emissions from oil Economic growth Environmental Kuznets curve Dynamic panel threshold model

a b s t r a c t The purpose of this paper is to examine the relationship between carbon dioxide (CO2) emissions from oil and GDP, using panel data from 1971 to 2007 of 98 countries. Previous studies have discussed the environmental Kuznets curve (EKC) hypothesis, but little attention has been paid to the existence of a nonlinear relationship between these two variables. We argue that there exists a threshold effect between the two variables: different levels of economic growth bear different impacts on oil CO2 emissions. Our empirical results do not support the EKC hypothesis. Additionally, the results of short-term analyses of static and dynamic panel threshold estimations suggest the efficacy of a double-threshold (three-regime) model. In the low economic growth regime, economic growth negatively affects oil CO2 emissions growth; in the medium economic growth regime, however, economic growth positively impacts oil CO2 emissions growth; and in the high economic growth regime, the impact of economic growth is insignificant. © 2012 Elsevier B.V. All rights reserved.

1. Introduction As Azomahou et al. (2006) point out, there are two reasons to study carbon dioxide (CO2) emissions. First, the greenhouse gas effect is considered a huge threat to the health of the environment; of the many greenhouse gases, CO2 is the most problematic and the most difficult to manage. In addition, CO2 has a very long lifespan (i.e. 50–200 years). 1 Second, CO2 emissions mostly derive from energy consumption — a crucial factor relating to modern production and consumption in the world economy. CO2 is produced mostly as a result of fossil-fuel consumption — including that of coal, petroleum, and natural gas — and cement production.2 For these two reasons, the relationship between CO2 emissions and economic growth is of concern to most economists and environmentalists.

⁎ Tel.: + 886 4 27016855x2383; fax: + 886 4 24521646. E-mail address: [email protected]. 1 Scientists estimate that CO2 remains stable in the atmosphere for anywhere from 50 to 200 years (see: http://greennature.com/article281.html). 2 As per the CO2 Emissions from Fuel Combustion Highlights (2010 Edition), total fuel CO2 emissions across the globe in 1971 were 14096.3 million tons, which increased to 29381.4 million tons in 2008, representing a growth rate of 108.4%. The oil CO2 emissions in 1971 were 6837.8 million tons, which increased to 10821.0 million tons in 2008, representing a growth rate of 58.25%. In 2008, 43% of the CO2 emissions from fuel combustion were produced from coal, 37% from oil, and 20% from gas. The growth rates for these fuels in 2008 were quite different, reflecting varying trends that are expected to continue in the future. 0264-9993/$ – see front matter © 2012 Elsevier B.V. All rights reserved. doi:10.1016/j.econmod.2012.05.001

Regarding studies of the relationship between CO2 emissions and GDP, most of the literature focuses on discussions of the environmental Kuznets curve (EKC). These studies argue that the relationship between the two variables takes the shape of an inverted-U curve. This means that at a relatively low income level, and as income increases, energy consumption will increase as well — which in turn will raise CO2 emission levels and environmental pollution. Therefore, at a relatively low income level, CO2 emissions and income correlate positively. As income increases to a certain higher level, awareness of environmental protection is enhanced, and both people and the government become more willing to spend more resources on enforcing regulations and creating environmental policies, thus leading to decreases in environmental pollution and CO2 emissions. The above discussion highlights why the EKC is an inverted-U curve: as income increases, the relationship between income and CO2 emissions switches from a positive number to 0, and then to a negative number. Thus, the income elasticity related to CO2 emissions also changes from a positive number to 0, and then to a negative number, as income increases. When applied to econometric analysis, this phenomenon implies that the estimated relationship between CO2 emissions and GDP could be positive, 0, or negative, depending on the unique economic and environmental status of the country under examination. The income elasticity related to CO2 emissions is a good indicator in studying the relationship between CO2 emissions and GDP. If the income elasticity is greater than 1, it means that the growth rate of CO2 emissions is higher than the GDP growth rate. In such a case, it is said that CO2 emissions and GDP are cross-coupling with each

1538

K-M. Wang / Economic Modelling 29 (2012) 1537–1547

other. If the income elasticity is positive but less than 1, however, the growth rate of CO2 emissions is smaller than that of GDP, and it is said that the two variables are relative-decoupling with each other. On the other hand, if the income elasticity is 0 or a negative number, then the CO2 growth rate is not affected by the economic growth rate — or, the CO2 growth rate is decreasing with the economic growth rate. In this case, the two variables are said to be absolute-decoupling with each other. In the literature in this field, most studies use a single country or region as the sample under examination. For instance, Grossman and Krueger (1991), Gallagher (2004), and Stern (2007) focus on Mexico; Carson et al. (1997) and Aldy (2005) study the United States; List and Gallet (1999) look at both OECD and non-OECD countries in their sample; Vincent (1997) focuses on Malaysia; and Auffhammer and Carson (2008) utilise Chinese data. Recent studies (e.g. Azomahou et al., 2006; Diao et al., 2009; Romero-Ávila, 2008) analyse the problem with new econometric models or tests, such as the nonparametric panel data model, state-of-the-art panel stationarity test, linear model, quadratic function model, or cubic function model. However, most studies employ a linear rather than nonlinear model and utilise the levels, rather than the growth rates, of the variables. The research methodologies adopted by many of these previous studies bear several shortcomings. First, most studies utilise the ordinary least squares (OLS) model, whose estimation result is an averaged outcome that is very sensitive to outliers. In addition, if the data are characterised by heteroscedasticity, the estimation result may be biased. Second, as discussed above, the inverted-U curve EKC contains a time factor: the time trend of CO2 emissions as income increases. In comparison, most studies analyse the EKC by using methods involving cross-sectional data (e.g. Grossman and Krueger, 1991; Panayotou, 1993; Robers and Grimes, 1997; Shafik, 1994). Cross-sectional analytical methods cannot reveal the impact of time; therefore, the estimation result may not be able to offer a ‘big picture’ perspective. Third, the aforementioned articles do not include in the regressions the lagged CO2 emission.3 This lagged variable is crucial to the study of the phenomenon, because the lagged variable could impact the current variable over time. In addition to the aforementioned methodology problems, there is one more issue we wish to address. Previous studies on this topic usually assume that at a certain time, each country is found at a unique and different point on the same EKC (Dinda, 2004). Some studies employ time-series data to conduct single-country case studies; examples include those of Jalil and Mahmud (2009) and Zhang and Cheng (2009). These studies can analyse the time trend inherent in the relationship between CO2 emissions and GDP; however, since they are single-country case studies, the results are not generalisable, and it is impossible to use the results to compare countries. To overcome the limitations inherent in the use of cross-sectional and time-series data, some studies have started to utilise panel data to conduct empirical studies (e.g. Apergis and Payne, 2009; Aslanidis and Iranzo, 2009; Lantz and Feng, 2006; Seldon and Song, 1994; Wagner, 2008). Panel data could be considered a combination of cross-sectional and time-series data; therefore, panel data could reveal the impact of individual differences or time transitions, which in turn have greater explanatory power. In addition, since panel data have both individual and time dimensions, there are more observations in a panel data sample; this increases the degrees of freedom, enhances the estimation efficiency, and reduces the possibility

3 The dynamic model includes as one of the explanatory variables the lagged dependent variable. If the explanatory variables of a regression model include only exogenous variables, the model is called a ‘static model’. The advantage of a dynamic model is that it considers the impact of time; its disadvantage is that there would be an endogenous heteroscedasticity problem with the residual. Since this problem has remained unresolved, most previous studies have employed static models. In this study, we use both static and dynamic models and compare the estimation results thereof.

of multicollinearity. Especially when conducting multi-country analysis, if one uses time-series data from a single country, it is possible that the analysis would ignore the impact of economic integration, and the estimation result could suffer from an inefficient-testing power problem. On the other hand, cross-sectional data ignore the impact of time (e.g. the impact of the business cycle). For this reason, recent studies in the field of international economic integration often utilise panel data. Carson (2010) surveys previous theoretical and empirical studies and concludes that several factors could affect the estimation results of EKC models; these include econometrics models, sample countries, pollution indices, explanatory variables, data attributes, and income and pollution functions. However, there is no mention of the possible nonlinear relationship between CO2 emissions and GDP. The purpose of the current study is to investigate whether there is a nonlinear relationship between oil-based CO2 emissions and GDP. Our sample includes 98 countries, and the time span is from 1971 to 2007. In the discussion of the relationship between oil CO2 emissions and GDP, previous studies have focused on the inverted-U curve-shaped relationship between the two variables — that is, most previous studies have focused on the establishment of the EKC. Unlike these studies, the current study examines the existence of a threshold effect in the relationship between the two variables. In other words, we would like to see if different levels of economic growth yield different effects vis-à-vis this relationship. This study focuses on the following three points. First, we investigate whether the long-term relationship between oil CO2 emissions and GDP is stable. Concurrently, we examine the nature of that relationship: crosscoupling, relative-decoupling, or absolute-decoupling. The answer can be used to test the EKC hypothesis. Second, we utilise the economic growth rate as the threshold variable, to construct static and dynamic panel threshold models (PTMs). The estimation result of the PTMs could be used to examine the existence of a nonlinear relationship between oil CO2 emissions and GDP. Third, we investigate the impact of population growth on the growth rate of oil CO2 emissions and on the convergence velocity from the short-term disequilibrium to the long-term equilibrium. 4 The empirical findings can be summarised as follows. First, the cointegration test result indicates that in the long term, the relationship between oil CO2 emissions and GDP is steady. The long-term income elasticity shows that the two variables relative-decouple with each other. Second, the estimation results of the panel static and dynamic models show that in the short term, the three-regime model is appropriate for describing the relationship between the two variables. In the low economic growth regime, GDP negatively affects the growth rate of oil CO2 emissions; however, in the medium economic growth regime, GDP has a significantly positive impact on the growth rate of oil CO2 emissions, and in the high economic growth regime, the

4 Previous EKC studies include some important macroeconomic variables in their linear models. The effects of these macroeconomic variables on the environment are mixed. Cropper and Griffiths (1994), Bruvoll and Medin (2003), Shi (2003), and Lantz and Feng (2006) each include the population variable in the regression and obtain a significantly monotonic relationship between the population and various forms of environmental variables. This finding suggests that a larger population would increase the demand for goods and services, which in turn would consume more natural resources. On the other hand, Seldon and Song (1994), Kahn and McDonald (1994), Tiffen et al. (1994), and Patel et al. (1995) each find population to relate negatively to environmental variables. One possible explanation for this finding is that as the population increases, more people become aware of the importance of environmental protection, and so the government raises environmental protection standards (Seldon and Song, 1994). In the current study, unlike in previous studies, we include the population growth in the short-term PTM model, to investigate the impact of population growth on the growth rate of oil CO2 emissions and on the short-term convergence adjustment of the model. The current study therefore is the first such attempt within the field, on this research topic.

K-M. Wang / Economic Modelling 29 (2012) 1537–1547

relationship between the two variables is insignificant. Third, population growth stimulates increases in oil CO2 emissions. By including the population growth variable, there is an increase in the adjustment velocity of the short-term error correction term. The remainder of this paper is organised as follows. The second section introduces the research methodology. The third section explains the model, data, empirical study procedures, and empirical results thereof. The fourth section provides concluding remarks.

xi;t ðγ Þ ¼ 4

The EKC was first discovered by Grossman and Krueger (1991), but the curve was not so named until the publication of the 1992 World Bank Development Report (Shafik and Bandyopadhyay, 1992). Many studies try to pinpoint the turning point — the income level at which the income elasticity equals 0 — of the inverted-U curve-shaped EKC, using cross-sectional or time-series data. Recent studies that utilise modern econometric methods can be categorised into one of three groups. Studies in the first group estimate the long- or short-term models by constructing vector autoregressive (VAR) models, autoregressive distributed lag (ARDL) models, or vector error correction models (VECMs). Their authors examine the relationship between CO2 emissions and GDP by using causality tests. Examples include Ang (2007), Halicioglu (2009), Jalil and Mahmud (2009), Sari and Soytas (2009), Soytas and Sari (2009), Zhang and Cheng (2009), Fodha and Zaghdoud (2010), and Menyah and Wolde-Rufael (2010). Papers in the second group examine the existence of the EKC via panel data analysis; examples include Azomahou et al. (2006) and RomeroÁvila (2008). The third group utilises polynomial or parameter models to investigate the relationship; examples include Diao et al. (2009) and He and Richard (2010). Most of the empirical results of the aforementioned studies support the relationship between GDP and CO2 emissions, as well as the sustainment of the EKC. The current study focuses on the growth rate of the variables to examine this relationship, as well as the existence of the EKC. We utilise the dynamic panel threshold model (DPTM) to conduct research, and we are the first to use this research method with respect to this topic. In the following, we explain our application of DPTM and the ideas therein. There are two problems that need to be addressed when constructing the DPTM. First, in our model, one of the regressors is the lagged dependent variable; this makes our model a dynamic panel data model (DPDM). If we use the regular fixed-effect model of the panel data analysis, then there would be bias created by the correlation between the lagged dependent variable and the residual. To resolve this problem, we utilise the generalised method of moment (GMM) estimation proposed by Arellano and Bond (1991) to conduct the estimation. Second, if the data are panel data and are nonlinear in nature, then one must use the PTM to reveal the data properties. Basically, these two problems can be resolved individually; if they were to be resolved simultaneously, then the regression residual may have an autocorrelation problem. This would cause unidentifiable variance and, of course, it could not be used in the following hypothesis test. Because of this, we estimate the model in two steps. In step one, we estimate the static PTM, and in step two, we estimate the DPTM. In the following, we explain the two models in detail. In step one, we utilise the static PTM proposed by Hansen (1999), to exploit the nonlinear characteristic of the panel data. If we were to assume that there is only one threshold in the model, the model can be expressed as
 qi;t ≤γ yi;t ¼ μ i þ
 ; þβ′2 xi;t I qi;t > γ þ ei;t ′ β1 xi;t I

and t = 1,…, T denoting the time; and γ is the threshold value. The vector of explanatory variables is partitioned into a subset x1i,t, of exogenous variables uncorrelated with ei,t, and a subset of endogenous variables x2i,t, correlated with ei,t. I in Eq. (1) is the indicator function. If the condition in the parentheses holds, then I = 1; otherwise, I = 0. For example, if the value of the threshold variable is less than γ, then I(qi,t≤γ) = 1. Assume β = (β′1,β′2)’ and define xi,t(γ) as 2

2. Research methodology

where μi is the country fixed effect; xi,t is the strictly exogenous variables; qi,t is the threshold variable, i = l,…, N denoting the country,


 3 xi;t I qi;t ≤γ
 5: xi;t I qi;t > γ

ð2Þ

Then, we can simplify Eq. (1) as ′

yi;t ¼ μ i þ β xi;t ðγ Þ þ ei;t :

ð3Þ

For each i, average Eq. (3) over time; we then derive ′ y
i: ¼ μ i þ β x
i: ðγ Þ þ e
i: ;

ð4Þ

where T 1X y ; T t¼1 i;t T 1X x ðγÞ; x
i: ðγÞ ¼ T t¼1 i;t

y
i: ¼

e
i ¼

T X ei;t : t¼1

Define yi;t ¼ yi;t −y
i: ; xi;t ðγ Þ ¼ xi;t ðγ Þ−x
i ðγÞ, e*i, t = ei, t − ēi. and 2 3 2  3 2  3 xi;2 ðγ Þ0 yi;2 ei;2 6 7 yi ¼ 4 ⋮ 5, xi ðγ Þ ¼ 4 ⋮ 5, ei ðγ Þ ¼ 4 ⋮ 5, and we can write yi;T ei;T x ðγ Þ0 i;T

Eq. (5) as 





Y ¼ X ðγ Þβ þ e ;

ð5Þ

where 2

3 y1 6 7 6 ⋮ 7  7 Y ¼6 6 yi 7; 4 5 ⋮ yN

2

3 x1 ðγ Þ 6 7 6 ⋮ 7  7; X ðγÞ ¼ 6 ð γ Þ x 6 i 7 4 5 ⋮  xN ðγÞ

2

3 e1 6 7 6 ⋮ 7  7 e ¼6 6 ei 7: 4 5 ⋮ eN

^ and the sum of the squared error, Residual e^  ðγÞ ¼ Y  −X  ðγÞβ S1*(γ), is 



′ 

S1 ðγÞ ¼ e ðγ Þ e ðγ Þ:

ð6Þ

To obtain the optimal threshold value, one can repeat the estimation process described above, using the γ within the possible value The γ value that corresponds to the minimum range to achieve S*(γ). 1 ^ . That is, S*(γ) is the optimal threshold value, γ 1 ^ ¼ arg min S1 ðγ Þ: γ γ

ð7Þ

^ Þ, the residThe estimator matrix of the threshold regression is θ^ðγ ^ Þ, and the residual variance is ual is e^ ðγ 2 σ^ ¼

ð1Þ

1539

1 1 1  ^ Þe^ ðγ ^Þ ¼ ^ Þ: e^ ðγ S ðγ N ðT−1Þ NðT−1Þ 1

ð8Þ

^ , we need to test for the existence of the threshAfter we obtain γ old effect. The null hypothesis is H0 : β1 ¼ β2 :

ð9Þ

1540

K-M. Wang / Economic Modelling 29 (2012) 1537–1547 γ

^ we can search the second optimal threshold value, γ ^2: Given γ, ^ γ2 ¼ arg min S2 ðγ 2 Þ; γ

ð10Þ

γ2



γ

^ 1 ; γ 2 Þ; if γ ^ 1 bγ2 S ðγ ;  ^ 1 Þ; if γ2 bγ ^1 S ðγ2 ; γ

and test the existence of the second threshold effect. Through this method, we can construct a multi-threshold model to reveal the nonlinear characteristic. For a detailed description of the model and the estimation process, please refer to Hansen (1999). In the second step, we use the GMM method proposed by Arellano and Bond (1991) to conduct the estimation. We transform the PTM in the first step into a linear panel data model. By using the dummy variable, we can still reveal the nonlinear characteristic identified in step 1. In addition, we add yi,t-1 as one of the regressors, and so our model is the DPDM with a nonlinear characteristic. Following the specification of Eq. (5) and by adding yi,1*t(=y*i, t − 1) as one of the regressors, we specify the model as follows:

1

ð11Þ



R

Y ¼Y ρþX βþe ;

ð12Þ

where DV(R1) and DV(R2) are the dummy variables that denote regimes 1 and 2, respectively, and the values of the dummy variables come from the PTM estimation in step 1. As to the definition of the other variables in Eqs. (11) and (12), 2

1 y
i: ¼

T 1 X y ; T−1 t¼2 i;t−1

2

 3 y1 6 7 6 ⋮ 7 7 Y ¼ 6 6 yi 7; 4 5 ⋮ yN



1

2

Y 1

1

yi;t−1 ¼ yi;t−1 −y
i: ¼ yi;t ; 3 y1 1 6 ⋮ 7 6 1 7 7 ¼6 6 yi 7; 4 ⋮ 5 y1 N

1

yi

3

6 yi;2 7 6 7 6 7 6 ⋮ 7 ¼ 6  7; 6y 0 7 6 i;T 7 4 5

2

X R



3 xR 1 6 ⋮ 7 6 R 7 7 ¼6 6 xi 7; 4 ⋮ 5 R xN

0

2

 3 e1 6 7 6 ⋮ 7 7 e ¼ 6 6 ei 7: 4 5 ⋮ eN

Before conducting the GMM estimation, we need to ascertain whether the orthogonal condition holds. The condition is: h
i   ¼0 E y1 i;t−s ei;t −ei;t−1 h
i R   E xi;s ei;t −ei;t−1 ¼ 0

s ¼ 2; …; t−1; t ¼ 3; …; T;

ð13Þ

s ¼ 1; …; T; t ¼ 3; ⋯; T:

It is obvious that as t gets larger, the number of orthogonal conditions increases. Rewriting Eq. (13) into a matrix, we get Eq. (14): h
i    E W i ′ ei;t −ei;t−1 ¼ 0

i ¼ 1; ⋯; N;

ð14Þ

where Wi* is the instrumental variable, and 2
6 6 6 Wi ¼ 6 6 4

R R y1 i;1 ; xi;1 ; xi;2

0 ⋮ 0




0 1 R R R y1 i;1 ; yi;2 ; xi;1 ; xi;2 ; xi;3

⋮ 0

0 ⋯ 0 ⋯ ⋮ ⋱ 0 ⋯ 0 ⋯

3 0 0 0 0 7 7 ⋮ ⋮ 7 7: 2 −1 5 −1 2

Please refer to Arellano and Bond (1991) and Arellano and Bover (1995) for detailed descriptions of the estimation method. The above discussion can be summarised as follows. In the first-step estimation, we use PTM to capture the nonlinear characteristic from among the variables. In the second step, we transform the model and conduct the DPDM estimation. After these two steps, we can obtain a complete DPTM estimation result. The estimation methods utilised in the two steps are standard econometric methods that have been validated by many studies; using them in this way can preclude the aforementioned autocorrelation residual problem. 3. Models, data, and empirical results

′  yi;t ¼ μ i þ ρ y1 i;t þ β1 xi;t DV ðR1 Þ ; ′   þβ2 xi;t DV ðR2 Þ þ ei;t 

2

2 −1 6 −1 2 6 H¼6 ⋮ 6 ⋮ 4 0 0 0 0

where S2 ðγ2 Þ ¼

In Eq. (15), AN = [(1/N) ∑ iN= 1Wi*′ HWi*] − 1, where H is a (T − 2) × (T − 2) matrix:



3

⋯

0

⋯

0

⋱
⋮ R R 0 yi1 ; yi2 ; …; yi;T−2 ; xR i;1 ; xi;2 ; …; xi;T−1

7 7 7 7 7 5


 ^ ′ , and ^0; β Let the GMM estimator matrix be θ^ ρ
0  R   R −1 R 0   θ^ ¼ X W AN W ′X X W AN W ′Y:

ð15Þ

3.1. Models and data Following Ang (2007) and Apergis and Payne (2009), we specify the long-term relationship between oil CO2 emissions and real percapita GDP as follows: 2

co2it ¼ α i þ β1i yit þ β2i yit þ εit

ð16Þ

where i stands for the country and i = 1, ….., N; t denotes the time and t = 1, ….., T; co2 is the CO2 emissions from oil (in metric tons per capita); and y is real GDP per capita measured in 2000 US dollars. The parameters β1 and β2 are the long-term elasticity estimators of oil CO2 emissions with respect to real GDP and to the square of real GDP, respectively. The EKC hypothesis suggests that β1 > 0 and β2 b 0. β1 being positive reveals the phenomenon wherein as income increases, the oil CO2 emissions increase as well; β2 being negative reflects the inverted-U curve-shaped pattern of the EKC, where once income passes the threshold, the oil CO2 emissions will decrease. In addition, if the value of β1 is greater than 1, which indicates that the growth rate of oil CO2 emissions is greater than that of GDP, then the two variables are said to be cross-coupling. If the value of β1 is between 0 and 1, which indicates that the oil CO2 emissions increase with GDP growth, then the two variables are said to be relative-decoupling. If the value of β1 is less than or equal to 0 — which indicates that oil CO2 emissions decrease with GDP growth — then the two variables are said to be absolutedecoupling. When conducting the short-term analysis, we add the population growth rate as the exogenous variable, to examine whether adding this variable better explains the oil CO2 emissions. In all, 98 countries are represented in our sample; please see Appendix A for a list of them. The variables we employ include oil CO2 emissions per capita (co2), real GDP per capita (y), the square of real GDP per capita (y 2), and population (pop). All the data are annual data from 1971 to 2007 and obtained from CO2 Emissions from Fuel Combustion (2009 Edition) published by the International Energy Agency (IEA) in Paris. We take the natural log of all data when conducting empirical analyses. Let us first present the basic statistic characteristics of the variables. Table 1 reports the statistics of the level and first-differenced (i.e. the growth rate) values of the three variables. The mean of co2 is 0.171, a global multi-year average of CO2 emissions. The global multi-year average per-capita GDP is 0.727, and the global multiyear average population is 2.609. As to the medium values, we find

K-M. Wang / Economic Modelling 29 (2012) 1537–1547 Table 1 Summary of basic statistics.

1541

Table 3 Panel cointegration test.

Level

co2

y

pop

Null hypothesis: no cointegration

Statistic

Prob.

Mean Median Std. Dev. Observations Cross sections

0.171 0.489 1.500 3626 98

0.727 0.730 1.650 3626 98

2.609 2.573 1.271 3626 98

Alternative hypothesis: Common AR coefs. (within-dimension) Panel v-statistic Panel rho-statistic Panel PP-statistic

3.59 − 4.37 − 6.19

0.00 0.00 0.00

First difference (%)

Δco2

Δy

Δpop

Mean Median Std. Dev. Observations Cross sections

1.185 0.936 14.58 3528 98

1.605 2.429 19.32 3528 98

1.784 1.777 9.047 3528 98

Alternative hypothesis: Individual AR coefs. (between-dimension) Panel ADF-statistic Group rho-statistic Group PP-statistic Group ADF-statistic

− 5.53 0.62 − 2.40 − 1.93

0.00 0.73 0.01 0.03

Note: the Pedroni's (1999) statistics are asymptotically distributed as normal.

that the difference between the medium and mean of co2 is larger, while that of y is smaller. Regarding the statistic values of the first-differenced variables, the means are all positive, indicating that per-capita oil CO2 emissions, per-capita GDP, and population are all increasing. The population growth rate is the largest, and the growth rate of oil CO2 emissions is the smallest. This phenomenon suggests that different economic growth status levels may have differential impacts on the growth rate of oil CO2 emission; for this reason, we believe that the use of threshold variables to divide the economic growth status levels into several regimes will help us better understand the relationship between GDP and oil CO2 emissions.

3.2. Empirical results To avoid the spurious regression problem, we test the stationarity of the variables. We employ five panel unit-root tests: the t*-statistic test of Levin et al. (2002), the W-statistic test of Im et al. (2003), the t-statistic test of Breitung (2000), and the augmented Dickey–Fuller (ADF)–Fisher Chi-square and the Phillips–Perron (PP)–Fisher Chisquare tests of Maddala and Wu (1999). Table 2 reports the unitroot test results. The four variables (co2, y, y 2 and pop) are all I(1), which means that we need to first-difference them to derive stationary variables. Since the variables are I(1), we further examine whether co2, y, and y 2 are cointegrated. We employ seven test methods: the panel v-statistic test of Pedroni (1999), the panel rho-statistic test, the panel PP-statistic test, the panel ADF-statistic test, the group rho-

statistic test, the group PP-statistic test, and the group ADF-statistic test. Table 3 reports the panel cointegration test results. We find that there are long-term cointegration relationships among co2, y, and y 2. Estimating the relationship via the fully modified OLS (FMOLS) method, we obtain: co2it ¼ 0:393 þ 0:632yit þ 0:032y2it p  valueð0:000Þð0:000Þ ð0:000Þ:

ð17Þ

The coefficients in Eq. (17) are all significant. Since β1 and β2 are positive, the implication is that the EKC hypothesis cannot hold. However, the per-capita GDP has a positive impact on oil CO2 emissions: petroleum consumption rises as income increases, which in turn causes an increase in oil CO2 emissions. In addition, a positive β1 indicates that oil CO2 emissions and per-capita GDP are relative-decoupling. To understand Eq. (17) better, we plot in Fig. 1 the estimates of co2 and y. It is obvious that the relationship between the two variables is not an inverted-U shape; therefore, the EKC hypothesis cannot hold here. In the short-term estimation, we want to address the following three issues: whether different economic growth rates would lead to different levels of oil CO2 emissions growth, whether population growth would affect oil CO2 emissions growth, and whether the error correction adjustment process is stable and convergent. To answer these three questions, we first employ the economic growth rate as the threshold variable, to determine the existence of a nonlinear relationship. If this relationship exists, it means that different economic growth status levels would influence the relationships among variables. In addition, since the long-term cointegration relationship exists, we

Table 2 Panel unit-root test. Level

co2 Statistic

LLC t*-stat Breitung t-stat IPS W-stat ADF–Fisher Chi-square PP–Fisher Chi-square

− 0.195 0.701 − 0.377 201.5 208.3

First difference

Δco2

LLC t*-stat Breitung t-stat IPS W-stat ADF–Fisher Chi-square PP–Fisher Chi-square

− 19.17 − 16.21 − 24.07 927.9 2786.4

y2

y p-value (0.42) (0.76) (0.35) (0.38) (0.26)

Statistic 0.594 − 0.535 1.948 201.2 199.1

p-value (0.72) (0.30) (0.97) (0.38) (0.42)

Δy (0.00) (0.00) (0.00) (0.00) (0.00)

− 31.05 − 15.31 − 26.15 1359.9 2259.1

Statistic 1.517 4.019 3.439 195.2 247.5

pop p-value (0.94) (1.00) (1.00) (0.50) (0.01)

Δpop (0.00) (0.00) (0.00) (0.00) (0.00)

− 26.76 − 12.31 − 25.43 1431.7 2673.2

Statistic − 6.490 8.536 6.042 210.9 214.3

p-value (0.05) (1.00) (1.00) (0.22) (0.18)

Δco2 (0.00) (0.00) (0.00) (0.00) (0.00)

− 3.123 − 1.351 − 7.424 4341 1869.9

(0.00) (0.05) (0.00) (0.00) (0.00)

Note: LLC, and IPS represent the panel unit root tests of Levin et al. (2002) and Im et al. (2003), respectively. Fisher–ADF and Fisher–PP represent the Maddala and Wu (1999) Fisher–ADF and Fisher–PP panel unit root tests, respectively. Exogenous variables: Individual effects, individual linear trends. Fisher tests are computed using an asymptotic Chi-square distribution. All other tests assume asymptotic normality. When carrying out the test as well as the estimation, all variables are formed in natural logarithm.

1542

K-M. Wang / Economic Modelling 29 (2012) 1537–1547

5

Table 4 Threshold effect test.

4 3

co2f

2 1 0 -1 -2 -3 -4 -6

-4

-2

0

2

4

6

8

y Fig. 1. The relationship between oil CO2 emission (co2f) and income (y).

must consider the adjustment process from the short-term disequilibrium to the long-term equilibrium; therefore, we must estimate a panel threshold error correction model with an error correct term: 2

ECMit1 ¼ co2it1 þ 0:393  0:632yit1  0:032yit1 :

ð18Þ

We follow Hansen (1999) and first confirm the number of thresholds. Table 4 reports the threshold effect test result. F1 and F2 are significant, but F3 is not; this indicates that there are two threshold points. 5 The same result can be seen in Figs. 2–5. Fig. 2 presents the single-threshold confidence interval and shows that there is one threshold point. Figs. 3 and 4 present the doublethreshold confidence intervals and show that there are two threshold points. Fig. 5 is the triple-threshold confidence interval and shows that there is no third threshold point. 6 The two threshold values are γ1 = − 0.1925 and γ2 = 0.1197. The model of our firststep estimation is the static panel double-threshold (three regimes) model 7: Δco2it ¼ ai þ b1 ECM it1 þ b2 Δy2it−1 þ b3 Δpopit−1 þ b4 Δyit−1 IðΔyit−1 ≤γ 1 Þ þb5 Δyit−1 I ðγ 1 < Δyit−1 ≤γ2 Þ þ b6 Δyit−1 Iðγ2 < Δyit−1 Þ þ eit ;

LR test

P-value

Test for single threshold F1 (10%, 5%, 1% critical values)

33.01 (8.78, 11.09, 15.06)

0.00

Test for double threshold F2 (10%, 5%, 1% critical values)

18.34 (7.01, 9.52, 15.54)

0.01

Test for triple threshold F3 (10%, 5%, 1% critical values)

1.17 (8.51, 10.82, 18.51)

0.93

Note: The likelihood ratio (LR) testing for a threshold is based on the statistic F1 (test for single threshold). The asymptotic distribution of F1 is non-standard, and strictly dominates the χ2 distribution. Because F1 rejects the null of no threshold (Eq. (9)), a further tests to discriminate between one and two thresholds. Thus an approximate likelihood ratio test of one versus two thresholds based on the statistic F2 (test for double threshold). Because F2 rejects the null of one threshold, then the statistic F3 (test for triple threshold), an approximate likelihood ratio test of two versus three thresholds (see Hansen, 1999). Because F1 and F2 are significant but F3 is not, which indicates that there are two threshold points.

where a negative b1 indicates that that the adjustment process is stable and convergent; Δyit2 − 1 is the square of the change in the per-capita GDP; b2 denotes the relationship between the square of the economic growth rate and the growth rate of oil CO2 emissions, where a significant b2 indicates the existence of a nonlinear relationship between the two variables; Δpop is the population growth rate; the significance of b3 shows the impact of population growth on the growth rate of oil CO2 emissions; γ1 and γ2 are the threshold values; and I(.) is the index variable to differentiate among regimes, where I(.) = 1 if the condition in the parentheses holds, I(.) = 0, otherwise. Let I(Δyit − 1 ≤ γ1) indicate regime 1, a low economic growth regime; I(γ1 b Δyit − 1 ≤ γ2) regime 2, a medium economic growth regime; and I(γ2 b Δyit − 1) regime 3, a high economic growth regime. Coefficients b4, b5, and b6 represent the impacts of the three regimes on the growth rate of oil CO2 emissions, and their significance also reveals the existence of the nonlinear relationship.

ð19Þ

where ECMit is the error correction term; b1 is the adjustment speed from the short-term disequilibrium to the long-term equilibrium,

5 The likelihood ratio testing for a threshold is based on the statistic F1 (test for single threshold). The asymptotic distribution of F1 is nonstandard and strictly dominates the χ2 distribution. If F1 rejects the null of no threshold (Eq. (9)), further testing to discriminate between one and two thresholds is required. Thus an approximate likelihood ratio test of one versus two thresholds based on the statistic F2 (test for double threshold) is executed. If F2 rejects the null of one threshold, then the statistic F3 (test for triple threshold) — an approximate likelihood ratio test of two versus three thresholds — is required (see Hansen, 1999). 6 More information can be learned about the threshold estimates from plots of the concentrated likelihood ratio function LR1(γ),LR2r (γ)¸LR1r (γ) in Figs. 2–5 (corresponding ^ 1 and the refinement estimators γ ^ r1 ). The point es^ r2 and γ to the first-stage estimate γ timates are the value of γ at which the likelihood ratio hits the zero axis, which is in the far left part of the graph. The 95% confidence intervals for γ2 and γ1 can be found from LR2r (γ) and LR1r (γ) by the values of γ for which the likelihood ratio lies beneath the blued line. It is interesting to examine the unrefined first-step likelihood ratio functionLR1(γ), which is computed when estimating a single threshold model. The first-step threshold estimate is the point where the LR1(γ) equals zero, which occurs at γ1 = − 19.25%. There is a second major dip in the likelihood ratio around the second-step estimate 11.97%. Thus the single threshold likelihood conveys information that suggests that there is a second threshold in the regression. 7 Since the estimated coefficient and standard deviation are very close to 0, we do not include Δyit3 − 1 in the model. In addition, we choose the lagged period equal to 1, according to Akaike's information criterion.

60

Likelihood Ratio

50

40

30

20

10

0 -30

-20

-10

0

10

20

Threshold Parameter Fig. 2. The confidence interval of the single‐threshold model.

30

K-M. Wang / Economic Modelling 29 (2012) 1537–1547

20

Likelihood Ratio

8

18

1543

Likelihood Ratio

7

16

6

14

5

12 4 10 3 8 2

6 4

1

2

0 -6

0 -15

-4

-2

0

2

4

6

8

Third Threshold Parameter -10

-5

0

5

10

15

20

25

30

Second Threshold Parameter

Fig. 5. The confidence interval of the triple‐threshold model.

Fig. 3. The confidence interval of the double‐threshold model.

70

Likelihood Ratio

60

50

40

30

20

10

0 -30

-20

-10

0

10

20

30

First Threshold Parameter Fig. 4. The confidence interval of the double‐threshold model.

3.2.1. Estimation result of the static panel threshold ECM model Table 5 reports the estimation results. In regime 1 where the growth rate of per-capita GDP is lower than −19.25%, since the economy is in recession or depression, individuals can focus only on basic life needs and must curb other, unnecessary consumption; this, in turn, reduces petroleum consumption and oil CO2 emissions. Therefore, the relationship between per-capita GDP and oil CO2 emissions is significantly negative. In regime 2, where the growth rate of percapita GDP is between −19.25% and 11.97%, since the per-capita GDP is not too low or even increases, petroleum consumption on account of transportation, electric devices, or appliances will increase as income grows; this, in turn, will increase oil CO2 emissions. Therefore, the relationship between per-capita GDP and oil CO2 emissions is significantly positive. In regime 3, the growth rate of per-capita GDP is greater than 11.97%; in this regime, the oil CO2 emissions do

not increase with per-capita GDP. Compared to the impact of GDP growth on the oil CO2 emissions growth rate in regime 2, the impact in regime 3 is smaller. This suggests that in an economy featuring a high economic growth rate, people care more about environmental protections; therefore, although per-capita GDP growth still positively correlates with oil CO2 emissions growth, the correlation is lowered and insignificant. As to the estimation of other variables, the coefficient of Δyit2 − 1 is significantly negative; this indicates that there exists a nonlinear relationship between the growth rate of oil CO2 emissions and the economic growth rate. However, the coefficient is extremely small (−0.001); this means that the impact of Δyit2 − 1 on the growth rate of oil CO2 emissions is quite miniscule. The coefficient of ECMit is negative (b1 = −7.411), which shows that the adjustment process from the short-term disequilibrium to the long-term equilibrium is stable and convergent. The coefficient of the population growth shows that the impact of this variable on oil CO2 emissions growth is not significant. We pay special attention to the population growth rate because, intuitively, a larger population should lead to more consumption of petroleum and therefore greater oil CO2 emissions; however, the result here is contrary to intuition. We think the insignificance of the relationship between population growth and oil CO2 emissions growth may be caused by the estimation method used. In the next section, we estimate the dynamic model to reconfirm the significance and the direction of this relationship.

Table 5 Estimation result of the static panel threshold ECM model. Dependent variable: Δco2t Repressor

Coefficient

P-value

ECMit-1 Δyit2 − 1 Δpopit − 1 Δyit − 1I(Δyit − 1 ≤ − 19.25%) Δyit − 1I(− 19.252% b Δyit − 1 ≤ 11.97%) Δyit − 1I(Δyit − 1 > 11.97%)

− 7.411 − 0.001 0.011 − 0.207 0.414 0.081

(0.00) (0.00) (0.98) (0.03) (0.00) (0.83)

Note:γ1 and γ2 are the threshold values; I(.) is the index variable to differentiate among regimes, where I(.) = 1 if the condition in the parenthesis holds, I(.) = 0, otherwise. Let I(Δyit − 1 ≤ γ1) indicate regime 1, a low economic growth regime; I(γ1 b Δyit − 1 ≤ γ2) regime 2, a medium economic growth regime; and I(γ2 b Δyit − 1) regime 3, a high economic growth regime.

1544

K-M. Wang / Economic Modelling 29 (2012) 1537–1547

Table 6 Estimation result of the dynamic panel threshold ECM model. Dependent variable: Δco2t

Model-1

Regressor

Coeff.

Δco2it − 1 ECMit-1 Δyit2 − 1 Δpopit − 1 Δyit − 1 × DV1(R1 : Δyit − 1 ≤ − 19.25%) Δyit − 1 × DV2(R2 : − 19.25% b Δyit − 1 ≤ 11.97%) Δyit − 1 × DV3(R3 : Δyit − 1 > 11.97%) S.E. of regression Sum squared residual

Model-2 P-value

Coeff.

P-value

0.043 − 19.20 − 0.001

(0.00) (0.00) (0.00)

− 0.193 0.703 0.001 19.798 1265190

(0.00) (0.00) (0.89)

0.042 − 19.29 − 0.001 0.057 − 0.148 0.687 0.002 19.795 1264522

(0.00) (0.00) (0.000) (0.00) (0.00) (0.00) (0.88)

Note: DV1(R1), DV2(R2), and DV3(R3) are the dummy variables denoting regime 1, regime 2, and regime 3, respectively. The values of the dummy variables come from the PTM estimation of Table 5.

3.2.2. Estimation result of the dynamic panel threshold ECM The model of our second-step estimation is the dynamic panel double-threshold (three regimes) model: Δco2it ¼ ai þ φΔco2it1 þ b1 ECM it1 þ b2 Δy2it−1 þ b3 Δpopit−1 þb4 Δyit−1 DV1ðR1 : Δyit−1 ≤  19:25%Þ þb5 Δyit−1 DV2ðR2 : 19:25% < Δyit−1 ≤11:97%Þ;

ð20Þ

þb6 Δyit−1 DV3ðR3 : Δyit−1 > 11:97%Þ þ eit where DV1(R1), DV2(R2), and DV3(R3) are the dummy variables denoting regime 1, regime 2, and regime 3, respectively. The values of the dummy variables come from the PTM estimation. Table 6 reports the estimation result using Arellano and Bond's (1991) GMM method. In the estimation, the dummy variables represent the three regimes identified in the static model. In addition, we estimate two dynamic models, model 1 and model 2, without and with the population growth rate, respectively. The estimation results for the three regimes are very similar to the results in Table 5. The direction of the impact of the economic growth rate on the oil CO2 emissions growth is the same in the three regimes of the static and of the two dynamic models: the impact is positively significant in regimes 1 and 2 and insignificant in regime 3. In the dynamic model, we have Δco2it − 1 as one of the regressors to reveal the dynamic adjustment process of Δco2t. With this specification, we can draw more information from the data; this cannot be done with the static model. The coefficient of Δyit2 − 1 is significantly negative, as in Table 5. As to the impact of the population growth rate, the estimation result of model 2 shows that the impact is significantly positive, which confirms our argument that a larger population leads to greater consumption of petroleum and therefore higher oil CO2 emissions. In addition, the standard error of the regression coefficient (S.E. of regression) and the sum of the squared residual are both smaller in model 2 than in model 1; this indicates that the inclusion of the population growth rate (model 2) is a superior specification, and that the population growth rate is a crucial variable here. Finally, the value of b1 is significantly negative in both model 1 and model 2, which suggests that the adjustment process from the short-term disequilibrium to the longterm equilibrium is stable and convergent. Since the absolute value of b1 is larger in model 2, this means that the adjustment speed is higher in model 2.

difficult for government authorities to reduce emissions, because the two variables still have a relative-decoupling status in most countries today. In addition, our long-term estimation result does not support the EKC hypothesis; this was also the case with the results of He and Richard (2010). However, in each of Diao et al. (2009), Jalil and Mahmud (2009), and Fodha and Zaghdoud (2010), the EKC hypothesis is still sustained. One possible explanation for the different empirical results is that our focus is on oil CO2 emissions, while other studies use data pertaining to all CO2 emissions. Moreover, in the real world today, there is no perfect substitute for petroleum, and the world's population is still growing; therefore, it will be a long time before oil CO2 emissions are reduced. These facts explain why our estimation results do not to support the EKC hypothesis. 8 Table 7 describes the figures for country and proportion in the three regions during 1972–2007. On average, region 1 has the smallest proportion at 1.19%, region 2 has the largest proportion at 95.35%, and region 3 is in the middle at 3.46%. These results show that most countries exhibit a positive relationship between economic growth and oil CO2 emissions growth in the regime 2, and that most countries still pursue economic growth, which then causes oil CO2 emissions growth. Therefore, the integrated panel threshold regression analysis shows that decreases in oil CO2 emissions do not have a remarkable effect in most counties. On the other hand, Romero-Ávila (2008) points out that the structure change will affect the sustainment of the EKC hypothesis. However, by ‘structural change’, the author is referring to timedependent structural change, rather than the relationship between oil CO2 emissions and GDP. In our opinion, it is the structural change of the economic growth status that counts, because the economic growth process, which also contains a time factor, endogenously impacts oil CO2 emissions. As long as the relationship between oil CO2 emissions and GDP is stable, it is possible for government authorities to consider whether a CO2-reduction environmental protection policy would work. When the long-term relationship between the two variables exists and is stable, how could the government authority implement an efficient policy? In such a case, one could utilise the error correction model to analyse the short-term relationships among the variables. In the current study, by considering the three regimes, our empirical results tell us that the short-term disequilibrium could adjust to the longterm equilibrium.

3.3. Discussion and policy implications In this paper, we use the panel regression estimation to investigate the relationship between oil CO2 emissions and GDP. We find that, in the long term, an increase in GDP will increase oil CO2 emissions. According to our estimation result, it is unfortunately

8 The reviewer provides an ideal that in future there is the third way of CO2 reduction: a technological breakdown such as fusion power. In such case, economic growth would be associated with the reduction of CO2, and EKC–inverted‐U‐shape curve — might hold.

K-M. Wang / Economic Modelling 29 (2012) 1537–1547 Table 7 The country numbers and proportion in the 3 regions during 1972–2007. Regimes Low economic growth regime

Medium economic growth regime

High economic growth regime

Year

Δyit − 1 ≤ − 19.25%

− 19.252% b Δyit − 1 ≤ 11.97%

Δyit − 1 > 11.97%

1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 Average

0 1 3 2 1 0 0 0 2 4 1 2 1 2 3 0 3 2 3 3 0 2 0 0 1 0 1 0 0 1 1 3 0 0 0 0 1.19%

87 90 89 89 86 93 94 94 92 89 97 93 95 94 95 96 94 94 92 92 94 93 97 97 96 96 95 95 93 95 95 93 93 97 96 96 95.35%

11 7 6 7 11 5 4 4 4 5 1 3 2 2 0 2 1 3 3 3 4 3 1 1 1 2 2 3 5 2 2 2 5 1 2 2 3.46%

As to the population growth variable, why not consider it in the long-term estimation? From the viewpoint of statistics, we could do this; however, from the viewpoints of economics and econometrics, cointegrated variables are usually endogenous and it is unreasonable to think that increases in CO2 could cause the population to increase. We treat the population growth variable as the exogenous variable and find that population growth leads to an increase in oil CO2 emissions growth. Studies such as those of Cropper and Griffiths (1994), Bruvoll and Medin (2003), Shi (2003), and Lantz and Feng (2006) have similar findings as well. In addition, we have one more interesting finding: the inclusion of the population growth variable raises the convergent speed in the long term. This implies that increases in GDP increase oil CO2 emissions, and that population growth further increases the growth of oil CO2 emissions. Therefore, reducing population growth could be an efficient policy in decreasing the growth of oil CO2 emissions. Compared with past literature research, De Bruyn (1997) investigates the roles of structural change and environmental policy in explaining the EKC hypothesis and finds that the downward sloping part of the EKC is better explained by environmental policy than by structural change. He concludes that international cooperation may play an important role in providing encouragement to countries that have not yet reached their turning point on the hypothesised EKC. De Bruyn et al. (1998) look at emissions for

1545

the UK, USA, Netherlands, and West Germany for various time intervals between 1960 and 1993. They find that the effect of economic growth on emissions is positive and significant in most cases, and their findings indicate the existence of an inverted‐Ushaped curve. However, Harbaugh et al. (2002) look for the existence of the EKC using data from cities worldwide and find little or no evidence of an inverted‐U-shaped relationship between pollution and income. Stern (2004) stating that the EKC hypothesis is based on weak econometric framework and is therefore not a suitable approach to the environment income relationship; on the other there are studies like Costantini and Martini (2010) that find evidence of the EKC even after using various robust econometric techniques. In our study, we focus on threshold effect and the impact of population growth on oil CO2 emissions from oil and economic growth. We find that the positive relationship only exists in the middle economic growth region, but not in low and high growth region. In addition, we conduct static and dynamic panel threshold analyses; the research method has the advantage of enhancing the estimation credibility and enlarging the sample. Finally, and most importantly, our empirical results show that there are three regimes and that only in the low economic growth regime, where the economic growth rate is negative, does oil CO2 emissions growth not increase. Although the major factor driving the increase in oil CO2 emissions growth in the medium economic growth regime does not significantly impact oil CO2 emissions growth in the high economic growth regime, it also reflects the fact that the authorities in high economic growth regime countries are not sufficiently active in reducing oil CO2 emissions.

4. Conclusions The current study's data sample comprises 98 countries, an examination of which allows us to see the relationship between oil CO2 emissions and GDP. We use the economic growth rate as the threshold variable, to conduct static and dynamic panel threshold analyses. This research method has the advantage of enhancing the estimation credibility and enlarging the sample. In our empirical study, we first examine the long-term relationship; we then investigate the short-term error correction dynamic adjustment process. Our findings can be summarised as follows. The longterm estimation result shows that oil CO2 emissions and GDP are relative-decoupling with each other and that the EKC hypothesis cannot hold. As to our short-term estimation, we find that there exists a nonlinear double‐threshold (three regimes) effect in our model. The static and dynamic estimations show that in the low economic growth regime, where the economic growth rate is negative, oil CO2 emissions growth is reduced. In the medium economic growth regime, oil CO2 emissions growth increases with GDP growth. In the high economic growth regime, GDP growth does not significantly impact the growth of oil CO2 emissions. Moreover, population growth is another factor that increases oil CO2 emissions growth, and the inclusion of this variable speeds up the adjustment velocity from the short-term disequilibrium to the longterm equilibrium. The current study is innovative in two respects. First, this is the first study to use two-stage static and dynamic panel threshold analyses to examine the dynamic nonlinear relationship between oil CO2 emissions and GDP. Second, this is also the first study to include the population growth variable as an exogenous variable in studying its impact on the error correction adjustment speed. From the two findings mentioned in the previous paragraph, we can say that there are several ways to reduce oil CO2 emissions growth: search for substitute energy sources, directly reduce oil CO2 emissions growth in medium and high economic growth countries, and reduce population growth.

1546

K-M. Wang / Economic Modelling 29 (2012) 1537–1547

Appendix A. List of 98 countries

Name

Continental

Name

Continental

Name

Continental

Name

Continental

ALGERIA ANGOLA ARGENTINA AUSTRALIA AUSTRIA BANGLADESH BELGIUM BRAZIL BULGARIA CAMEROON CANADA CHILE CHINA COLOMBIA COSTA RICA COTEDIVOIRE CUBA CYPRUS CZECH DEM REP CONGO DENMARK DOMINICAN ECUADOR EGYPT ELSALVADOR

AFRICA AFRICA America Asia & Ocean Europe Asia & Ocean Europe America Europe AFRICA America America Asia & Ocean America America AFRICA America Asia & Ocean Europe AFRICA Europe America America AFRICA America

LEBANON LIBYANARABJAMA MALAYSIA MEXICO MOROCCO MOZAMBIQUE MYANMAR NEPAL NETHERLANDS NEWZEALAND NIGERIA NORWAY OMAN PAKISTAN PANAMA PARAGUAY PERU PHILIPPINES POLAND PORTUGAL QATAR ROMANIA SAUDI ARABIA SENEGAL SINGAPORE

Asia & Ocean AFRICA Asia & Ocean America AFRICA AFRICA Asia & Ocean Asia & Ocean Europe Asia & Ocean AFRICA Europe Asia & Ocean Asia & Ocean America America America Asia & Ocean Europe Europe Asia & Ocean Europe Asia & Ocean AFRICA Asia & Ocean

ETHIOPIA FINLAND FRANCE GERMANY GHANA GREECE GUATEMALA HAITI HONDURAS HONG KONG HUNGARY INDIA INDONESIA IRAN IRAQ IRELAND ISRAEL ITALY JAPAN JORDAN KENYA KOREA KOREA DPR KUWAIT

AFRICA Europe Europe Europe AFRICA Europe America America America Asia & Ocean Europe Asia & Ocean Asia & Ocean Asia & Ocean Asia & Ocean Europe Asia &Ocean Europe Asia & Ocean Asia & Ocean AFRICA Asia & Ocean Asia & Ocean Asia & Ocean

SLOVAK SOUTHAFRICA SPAIN SRILANKA SUDAN SWEDEN SWITZERLAND SYRIAN Arab TAIWAN TANZANIA THAILAND TOGO TRINID & TOBAGO TUNISIA TURKEY UNITEDARABEMI UNITED KINGDOM UNITED STATES URUGUAY VENEZUELA VIETNAM YEMEN ZAMBIA ZIMBABWE

Europe AFRICA Europe Asia & Ocean AFRICA Europe Europe Asia & Ocean Asia & Ocean AFRICA Asia & Ocean AFRICA America AFRICA Asia & Ocean Asia & Ocean Europe America America America Asia & Ocean Asia & Ocean AFRICA AFRICA

Source: CO2 Emissions from Fuel Combustion (2009 Edition), IEA, Paris.

References Aldy, J.E., 2005. An environmental Kuznets curve analysis of U.S. state-level carbon dioxide emissions. Environment and Development Economics 14, 48–72. Ang, J.B., 2007. CO2 emissions, energy consumption, and output in France. Energy Policy 35 (10), 4772–4778. Apergis, N., Payne, J.E., 2009. CO2 emissions, energy usage, and output in Central America. Energy Policy 37, 3282–3286. Arellano, M., Bond, S.R., 1991. Some tests of specification for panel data: Monte Carlo evidence and an application to employment equation. Review of Economic Studies 58, 277–297. Arellano, M., Bover, O., 1995. Another look at the instrumental variable estimation of error-components models. Journal of Econometrics 68, 29–52. Aslanidis, N., Iranzo, S., 2009. Environment and development: is there a Kuznets Curve for CO2 emissions? Applied Economics 41 (6), 803–810. Auffhammer, M., Carson, R.T., 2008. Forecasting the path of China's CO2 emissions using province level information. Journal of Environmental Economics and Management 55, 229–247. Azomahou, T., Laisney, F., Van Phu, N., 2006. Economic development and CO2 emissions: a nonparametric panel approach. Journal of Public Economics 90, 1347–1363. Breitung, J., 2000. The local power of some unit root tests for panel data. In: Baltagi, B. (Ed.), Advances in Econometrics. : Nonstationary Panels, Panel Cointegration, and Dynamic Panels, Vol. 15. JAI Press, Amsterdam, pp. 161–178. Bruvoll, A., Medin, H., 2003. Factors behind the environmental Kuznets curve. Environmental and Resource Economics 24, 27–48. Carson, R.T., 2010. The environmental Kuznets curve: seeking empirical regularity and theoretical structure. Review of Environmental Economics and Policy 4 (1), 3–23. Carson, R.T., Jeon, Y., McCubbin, D., 1997. The relationship between air pollution and emissions: U.S. data. Environment and Development Economics 2, 433–450. Costantini, V., Martini, C., 2010. A modified environmental Kuznets curve for sustainable development assessment using panel data. International Journal of Global Environmental Issues 10 (1/2), 84–122. Cropper, M., Griffiths, C., 1994. The interaction of population growth and environmental quality. American Economic Review 84, 250–254. De Bruyn, S.M., 1997. Explaining the environmental Kuznets curve: structural change and international agreements in reducing sulphur dioxide emissions. Environment and Development Economics 2, 485–503. De Bruyn, S.M., Van Den Bergh, J.C.J.M., Opschoor, J.B., 1998. Economic growth and emissions: reconsidering the empirical basis of environmental Kuznets curves. Ecological Economics 25, 161–175. Diao, X.D., Zeng, S.X., Tamb, C.M., Tamc, V.W.Y., 2009. EKC analysis for studying economic growth and environmental quality: a case study in China. Journal of Cleaner Production 17, 541–548.

Dinda, S., 2004. Environmental Kuznets curve hypothesis: a survey. Ecological Economics 49, 431–455. Fodha, M., Zaghdoud, O., 2010. Economic growth and pollutant emissions in Tunisia: an empirical analysis of the environmental Kuznets curve. Energy Policy 38, 1150–1156. Gallagher, K., 2004. Free Trade and the Environment: Mexico, NAFTA, and Beyond. Stanford University Press, Palo Alto. Grossman, G.M., Krueger, A.B., 1991. Environmental impacts of a North American Free Trade Agreement. Working paper No. 3914. National Bureau of Economic Research. Halicioglu, F., 2009. An econometric study of CO2 emissions, energy consumption, income and foreign trade in Turkey. Energy Policy 37 (3), 1156–1164. Hansen, B.E., 1999. Threshold effects in non-dynamic panel: estimation, testing and inference. Journal of Econometrics 93, 345–368. Harbaugh, W., Levinson, A., Wilson, D., 2002. Reexamining the empirical evidence for an environmental Kuznets curve. Review of Economics and Statistics 84 (3), 541–551. He, H., Richard, P., 2010. Environmental Kuznets curve for CO2 in Canada. Ecological Economics 69 (5), 1083–1093. Im, K.S., Pesaran, M.H., Shin, Y., 2003. Testing for unit roots in heterogeneous panels. Journal of Econometrics 115, 53–74. Jalil, A., Mahmud, S.F., 2009. Environment Kuznets curve for CO2 emissions: a cointegration analysis for China. Energy Policy 37, 5167–5172. Kahn, H., McDonald, A., 1994. Deforestation in Post-war Philippines. University of Chicago Press, Chicago, IL. Lantz, V., Feng, Q., 2006. Assessing income, population, and technology impacts on CO2 emissions in Canada: where's the EKC? Ecological Economics 57, 229–238. Levin, A., Lin, C.F., Chu, C., 2002. Unit root tests in panel data: asymptotic and finitesample properties. Journal of Econometrics 108, 1–24. List, J.A., Gallet, C.A., 1999. The environmental Kuznets curve: does one size fit all? Ecological Economics 31, 409–423. Maddala, G.S., Wu, S., 1999. A comparative study of unit root tests with panel data and a new simple test. Oxford Bulletin of Economics and Statistics 61, 631–652. Menyah, K., Wolde-Rufael, Y., 2010. CO2 emissions, nuclear energy, renewable energy and economic growth in the US. Energy Policy 38 (6), 2911–2915. Panayotou, T., 1993. Empirical tests and policy analysis of environmental degradation at different stages of economic development. Working Paper WP238. International Labour Office, Geneva, Switzerland. Patel, S., Pinckney, T., Jaeger, W., 1995. Smallholder wood production and population pressure in East Africa: evidence of an environmental Kuznets curve? Land Economics 71 (4), 516–533. Pedroni, P., 1999. Critical values for cointegration tests in heterogeneous panels with multiple regressors. Oxford Bulletin of Economics and Statistics 61, 653–670. Robers, J.T., Grimes, P.K., 1997. Carbon intensity and economic development 1962–1991: a brief exploration of the environmental Kuznets curve. World Development 25 (2), 191–198.

K-M. Wang / Economic Modelling 29 (2012) 1537–1547 Romero-Ávila, D., 2008. Convergence in carbon dioxide emissions among 37 industrialized countries revisited. Energy Economics 30, 2265–2282. Sari, R., Soytas, U., 2009. Are global warming and economic growth combatable? Evidence from five OPEC countries. Applied Energy 86, 1887–1893. Seldon, T., Song, D., 1994. Environmental quality and development: is there a Kuznets curve for air pollution emissions? Journal of Environmental Economics and Management 27, 147–162. Shafik, N., 1994. Economic development and environmental quality: an econometric analysis. Oxford Economic Papers 46, 757–773. Shafik, N., Bandyopadhyay, S., 1992. Economic growth and environmental quality: time series and cross section evidence. Working paper. World Development Report 1992. Oxford University Press, New York. Shi, A., 2003. The impact of population pressure on global carbon dioxide emissions, 1975–1996: evidence from pooled cross-country data. Ecological Economics 44, 29–42.

1547

Soytas, U., Sari, R., 2009. Energy consumption, economic growth and carbon emissions: challenges faced by a EU candidate member. Ecological Economics 68, 1667–1675. Stern, D.I., 2004. The rise and fall of the environmental Kuznets curve. World Development 32 (8), 1419–1439. Stern, N.H., 2007. Stern review: the economics of climate change. Cambridge University Press, Cambridge. Tiffen, M., Mortimore, M., Gichuki, F., 1994. More People, Less Erosion: Environmental Recovery in Kenya. J. Wiley, Chichester, New York. Vincent, J.R., 1997. Testing for environmental Kuznets curves within a developing country. Environment and Development Economics 2, 417–431. Wagner, M., 2008. The carbon Kuznets curve: a cloudy picture emitted by bad econometrics? Resource and Energy Economics 30, 388–408. Zhang, X.P., Cheng, X.M., 2009. Energy consumption, carbon emissions, and economic growth in China. Ecological Economics 68, 2706–2712.