The impact of economic growth on CO2 emissions in Azerbaijan

This paper investigates the relationship between the economic growth and CO2 emissions in Azerbaijan. A cointegration analysis is conducted over the period 1992e2013. For getting more robust results, Johansen, ARDLBT, DOLS, FMOLS and CCR methods are employed to explore cointegration and estimate long-run coefficients. We use cubic, quadratic and linear specifications and conclude that the last one is an adequate representation for the impact of the economic growth on CO2 emissions in Azerbaijan. The results from the different cointegration methods are consistent with each other and show that the economic growth has a positive and statistically significant impact on the emissions in the long-run implying that the EKC hypothesis does not hold for Azerbaijan. The income elasticity of CO2 emissions, using different methods, is found to be between 0.7 and 0.8. Moreover, we find that any short-run imbalance can be adjusted towards the long-run equilibrium path within less than one year. The paper concludes that measures to increase energy efficiency, carbon pricing instruments in production and international-domestic trade activities, and nationwide social awareness programs to instruct about the negative consequences of pollution can be considered as relevant environmental policies aimed at reducing carbon emissions. © 2019 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). 1 Transforming our world: the 2030 Agenda for Sustainable Development, page 23. 2


Introduction
It is well known that greenhouse gases (GHGs) are required to keep the Earth's temperature at levels so as to sustain life. However, increasing amounts of GHG emissions due to man-made activities, such as burning fossil fuels, absorb heat and cause global warming, giving rise to changes in the climate system. Arguably, this is one of the greatest problems humanity is facing today. Global climate change is, therefore, one of the main policy concerns of the century for all governments since it threatens societies' well-being, challenges the process of economic development and alters the natural environment. As it was noted in the "Transforming our world: the 2030 Agenda for Sustainable Development", according to the 13th Sustainable Development Goal, countries should "take urgent action to combat climate change and its impacts, strengthen resilience and adaptive capacity to climate-related hazards and natural disasters in all countries and integrate climate change measures into national policies, strategies and planning". 1 According to the World Bank (2007), CO2 emissions stemming from the burning of fossil fuels and the manufacture of cement are responsible for almost 60% of GHGs. 2 Moreover, according to the Intergovernmental Panel on Climate Change (IPCC) (2014), CO2 emissions from fossil fuel combustion and industrial processes shared about 78% of the increase in total GHG emissions for the period of 1970e2010, with analogous fraction contribution from 2000 to 2010. In order to attain environmental sustainability around the world in terms of GHGs, in 1997 the Kyoto protocol was signed by many governments of developed countries as well as developing and least developed countries. According to the IPCC (2007), these countries accounted for about 76.7% of total anthropogenic GHG emissions in 2004. The Protocol contained obligatory emission reduction goals for industrialized countries.
Although the reduction obligations of CO2 emissions referred to developed countries, based on the fact that they are the main contributors to global CO2 emissions, there have been calls on developing economies to take an active part in global emissions reduction (Winkler et al., 2002). CO2 emissions share from developing economies was nearly 50% of the world's total CO2 emissions in 2003 (Martinez-Zarzoso andMaruotti, 2011). If the current level of energy consumption continues, today's CO2 trend is expected to grow. It is thus a common interest for all policymakers of any level to adopt those policy measures that will be most effective in mitigating CO2 emissions. In this vein, in order to meet the sustainable development goal of maintaining "global temperature to increase below 2 C", the Paris agreement (COP21, 2015) has launched compulsory commitments by all parties, regardless of their development stage, consisting of "nationally determined contributions" (NDCs) intended at reaching the temperature goal. However, because of the differences between developed and developing countries and even dissimilarities between different countries within the same group, those policy measures will generally not be identical and should be investigated for individual countries (Stern et al., 1996;de Bruyn, 1998;Dijkgraaf and Vollebergh, 1998;Stern and Common, 2001;Dinda, 2004, inter alia).
Many number of studies have been conducted on the relationship between CO2 emissions and their main drivers for different individual countries. Just to mention a few, these include Canada (Hamit-Haggar, 2012), China (Du et al., 2012), France (Iwata et al., 2010), India (Tiwari et al., 2013), Malaysia , Russia (Pao et al., 2011), Spain (Esteve and Tamarit, 2012), Turkey (Yavuz, 2014), and for Brazil, China, Egypt, Japan, Mexico, Nigeria, South Korea, and South Africa (Onafowora and Owoye, 2014). To the best of our knowledge, no time series study has been conducted in the case of Azerbaijan. 3 The current study investigates the relationship between CO2 emissions and economic growth for Azerbaijan. Four main reasons led us to select this country.
First, as a resource-rich (mainly with oil and gas) country, Azerbaijan has been characterized by a considerable achievement in economic growth and it has been passed through different development stages (Hasanov and Hasanli, 2011;Hasanov et al., 2016). As pointed out by Winkler et al. (2002) among others, it is important to investigate the effect of economic growth on environmental degradation if a developing country experiences significant economic growth. The Azerbaijani economy has shown considerable economic growth since 2006. This new trend can be explained by the coming into force of oil contracts signed since the middle of 1990s. After the "Contract of the Century", which was signed in 1994 with 13 recognized world oil companies, 41 oil companies from 19 countries have signed 27 additional contracts. Moreover, the Baku-Tbilisi-Ceyhan oil export pipeline construction process was completed in 2005 and Azerbaijani oil began to be exported through this pipeline to reach the world oil markets (EIT, 2018). Second, energy consumption and economic growth, however, can determine negative impacts on the environment by increasing CO2 emissions. In turn, a damaged environment and environmental resources have negative impacts on people, society, and nature. In order to keep the balance among the elements of development, that is, to achieve a sustainable development, resources have to be used environmentally friendly. To reach this goal, some considerable activities, measures, and programs have been implemented by the Azerbaijani government agencies since the second half of the 1990s. The country signed the Kyoto Protocol in 2000. Development concept called "Azerbaijan 2020: Look into the Future" was released on December 29, 2012. In this concept, one of the main directions is to provide appropriate programs and activities to reach sustainable development. It is noteworthy that the concept planned to bring the amount of carbon dioxide in line with the appropriate level of member countries of the Organization for Economic Cooperation and Development by 2020 (Azerbaijan, 2020). Therefore, it is important to investigate how CO2 and economic growth relationship evolves over time in light of the implemented policy measures.
Third, the relationship between economic growth and CO2 emissions has been studied for different countries in the literature. However, only a few panel studies have included Azerbaijan in their CO2 analysis and, to the best of our knowledge, there is no time series study investigating this issue for Azerbaijan.
Fourth, investigating the relationship in the case of Azerbaijan, a resource-rich developing country, would be an example for other similar countries and thus may provide some understandings which are common across such kind of economies.
All the above-discussed considerations motivate the fact that building a well-designed econometric model relating CO2 emissions to economic and social factors is very important for Azerbaijan. In this study, we investigate the relationship between economic growth and CO2 emissions in the case of Azerbaijan employing different time series cointegration methods. Our study may contribute to the existing literature in a number of ways. First, considering that all previous CO2 studies for Azerbaijan have employed panel data methods, which might suffer from ignoring country-specific features, this is the first time series study specifically for the country. Second, we test the Environmental Kuznets Curve (EKC) hypothesis for Azerbaijan using time series data. Furthermore, we test different possible representations of the relationship between CO2 emissions and income based on the EKC framework. Third, we employ five different cointegration methods as well as account for small sample bias correction as a robustness check. Fourth, we make extensive and constructive literature review of CO2-income studies for Azerbaijan and other oil exporting small open developing economies. Finally, the findings of this study might be a roadmap for the similar countries.
The study concludes that implementation of carbon pricing instruments and mechanisms for cleaner production and international-domestic trade activities, nationwide social awareness programs to instruct on the negative consequences of the pollution, initiating the promotion of government policies to intensify the use of modern, less energy/pollution intensive technologies can be considered as the first-to-be-done policies in similar economies.
The rest of the paper is organized as follows: Section 2 provides a selected review of the related literature. Section 3 presents the conceptual framework of the study and the data employed. Section 4 discusses the methodology. The empirical results are reported in Section 5. Section 6 discusses the empirical results and Section 7 concludes the study and provides policy implications.

Literature review
In the early 1990s, three empirical studies independently analyzed the relationship between environmental degradation and per capita income (Grossman and Krueger, 1995;Shafik and Bandyopadhyay, 1992;Panayotou, 1993). All three studies concluded that the relationship between pollution variables and per capita income exhibited an inverted U-shaped curve, called the Environmental Kuznets Curve (EKC) by Panayotou (1993) to emphasize its similarity to the well-known Kuznets Curve between income distribution and per capita income levels (Kuznets and Simon, 1955).
Since then there have been a plethora of studies investigating the EKC, but the relationship between environmental degradation or quality and income remains hotly debated. There is a group of studies that discuss the theoretical underpinnings of the EKC, analyzing the potential explanations of a bell-shaped relationship and the reasons for the 'turning back' of environmental impacts after some threshold level of income. Following Kijima et al. (2010), the theoretical models of the EKC can be divided into two groups: static and dynamic. 4 An example of static EKC approach is the paper by L opez (1994) that discussed the theory of the EKC utilizing a production function approach. McConnell (1997), Andreoni and Levinson (2001), Lieb (2002), and Di Vita (2008) employ instead utility functions to explain the rationale for the EKC. Examples of the dynamic approach to the EKC include John and Pecchenino (1994), Selden and Song (1995), Dinda (2005), Chimeli and Braden (2009), and Prieur (2009) that consider resource allocation between consumption and abatement expenditures. Stokey (1998) and Tahvonen and Salo (2001) consider the effect of production technologies on the environment, whereas Jones and Manuelli (2001) and Egli and Steger (2007) propose models that take into account the effect of tax policy on pollution regulation. Finally, Wirl (2006) and Kijima et al. (2010) are examples of dynamic models dealing with the case of uncertainty in the economy.
In addition to the above mentioned theoretical studies, the empirical literature abounds with studies that investigate the environmental effects of energy use and economic growth for both developed and developing countries using different datasets, model specifications, methodologies, and functional forms.
The theoretical underpinnings suggest a non-linear relationship between environmental degradation and income. The empirical studies have therefore generally focused on quadratic and cubic EKC functional specifications, as originally proposed by Shafik and Bandyopadhyay (1992). 5 Table A1 in Appendix summarizes the empirical contributions to the EKC limiting the attention to those for small open oil-exporting developing economies (SOOEDE) since 2010. These are countries that are similar to Azerbaijan which is the focus of this paper. While there are no time series estimates of the EKC for Azerbaijan, only a limited number of panel studies include this country: as such they might not be able to capture the countryspecific features of the relationship among the variables of interest. For example, three studies (Tamazian and Rao, 2010;Apergis and Payne, 2010;Al-Mulali et al., 2016), which include Azerbaijan, find an inverted U-shaped curve, two papers obtain a U-shaped curve (Brizga et al., 2013;Narayan et al., 2016), one ends up with no specific patterns (Perez Suarez and Lopez-Menendez, 2015), and two studies obtain a monotonically increasing relationship (Ito, 2017;Mitic et al., 2017) between income and CO2 emissions.
One notable aspect is that many of the studies in the table (13 out of 24) employed the linear functional form, which can cause misspecification problem and misleading results. For example, using a linear specification, different studies found significantly different income elasticities for KSA (Alkhathlan and Javid, 2013: 0.45;Alshehry and Belloumi, 2016: 0.0025;Bekhet and Yasmin, 2017: À0.024;Narayan and Narayan, 2010: 0.40;Shahbaz et al., 2016: 0.26). 6 As said, there is not time series study devoted to Azerbaijan. In this regard, there is a need to conduct a time series study by taking into account all the above mentioned limitations including those noted in Table A1 of the existing studies in order to well understand the CO2-income relationship in Azerbaijan. To the best of our knowledge, the present one is the first study for Azerbaijan, investigating drivers of CO2 emissions and employing different time-series cointegration methods.

Employed functional form
The following is the traditional and widely used functional form for analyzing the relationship between CO2 emissions and GDP (Shafik and Bandyopadhyay, 1992;Grossman and Krueger, 1995;Lieb, 2003;Dinda, 2004, inter alia): Where co 2 is CO2 emissions measured per capita, y is GDP per capita, x is a vector of additional explanatory variables and u is the error term. Often (1) is estimated with a time trend in order to capture the effects on CO2 emissions caused by technological progress or enhanced environmental awareness (Shafik and Bandyopadhyay, 1992;Lieb, 2003). Denoting a time trend with t, the model used for the present analysis is: Equation (2) can be run in levels or in log form of the variables. In some cases, for example for the quadratic formulation, it may be preferable to estimate the model in log form (Cole et al., 1997), although the best formulation should be chosen in principle on the basis of the estimation results. We will use variables in the logarithmic form. Equation (2) enables us to test several forms of the CO2 emissions-economic development relationship (Dinda, 2004): i.e., an Environmental Kuznets Curve (EKC); 7 5. b 1 < 0; b 2 > 0 and b 3 ¼ 0 : a U-shaped relationship; 8 6. b 1 > 0; b 2 < 0 and b 3 > 0 : a cubic polynomial or N-shaped figure; 7. b 1 < 0; b 2 > 0 and b 3 < 0: the opposite to the N-shaped curve. 4 See the review of some of the theoretical EKC studies by Lieb (2003), Dinda (2004, and Kijima et al. (2010), among others. 5 Additional detailed information on the empirical studies devoted to the EKC and its different aspects can be found in Lieb (2003), Stern (2004), Dinda (2004), and Uchiyama (2016). 6 Several contributions in the literature investigate causality and impulse response effects between GDP and emissions (see, for instance, Magazzino, 2016aMagazzino, , 2016bMagazzino, , 2016cMagazzino, , 2016dMagazzino, , 2017. The focus of this paper is instead on the long-run relationship between those variables. 7 Strictly speaking the relationship is concave, implying decoupling of emissions from income. 8 In principle, we could also have the case of a convex relationship, Note that all abovementioned equalities and inequalities are considered to hold with statistical significance. In the case of EKC, the turning point, where the emission level starts to fall, should be within a reasonable range (Uchiyama, 2016). From the abovementioned cases, it is clear that the EKC is only one of the possible shapes implied by model (2).
Following Shafik and Bandyopadhyay (1992), we will adopt the following testing procedure: if the cubic term is not statistically significant, it can be dropped. Likewise, if the squared term is also insignificant, we conclude the linear relationship between emissions and income. As it can be found in Table 1 of Shafik and Bandyopadhyay (1992), one can get all b 1 ; b 2 and b 3 to be insignificant in the cubic form, but b 1 and b 2 are significant in the case of the quadratic specification. Moreover, following Kaufmann et al. (1998), Scruggs (1998), Dinda et al. (2000), Harbaugh et al. (2000), Millimet and Stengos (2000) and Lieb (2003) we interpret a U-shaped emission-income relationship as evidence of a N-shaped curve. Additionally, following Cole et al. (1997) and Stern and Common (2001), we interpret an EKC with an estimated turning point outside the sample range as evidence for a monotonically increasing emission-income relationship.

Data
Our study uses annual data on carbon dioxide emissions and GDP over the period 1992e2013 for Azerbaijan. 9 CO2 emissions (CO 2 ) are measured in kilotons (kt) of carbon dioxide and are those stemming from the burning of fossil fuels and the manufacture of cement. This is our dependent variable, which we converted into per capita terms using population data measured in persons. The data on CO 2 and population are retrieved from the World Bank Development Indicators Database (WB, 2016) over the period indicated above. Only the values of CO2 for 2012 and 2013 are taken from the official webpage of EnerData (http://www.enerdata.net/) since they are not available from the World Bank Development Indicators Database. GDP per capita in 2005 constant USD is retrieved from the World Bank Development Indicators Database (WB, 2015) over the period indicated above. Fig. 1 below shows the time profile of the above variables, both levels and growth rates, over the period 1992e2013.
As a general tendency for the chosen period co 2 and y increased, though the first variable has demonstrated volatility in the time profile since 2004. Before 1996 the level of CO2 emissions was decreasing due to the collapse of the previous economic system and only after that time the economy started to recover. Azerbaijani CO2 emissions increased from 31510 kilotons in 1996e37513 kilotons in 2013 with an annual growth rate of 1.12% over the period. GDP instead increased by 11% annually. The jump in 2006 was most likely due to the effect of oil revenues following the completion of the Baku-Tbilisi-Ceyhan main export oil pipeline. Table 1 presents the descriptive statistics of the used variables. As it can be seen from that table, in terms of coefficient of variation, both variables means' represent the variables quite well. In addition, minimum and maximum values are within the three standard deviation, indicating that there are not outliers for either variables.
In paper we use the natural logarithm of the variables, which are represented by small letters, i.e., co 2 , y, y 2 and y 3 .

Methodology
We use the Johansen cointegration approach as a main method. To get more robust results, we also employ the Bounds Testing approach to Auto Regressive Distributed Lag models (ARDLBT) (Pesaran and Shin, 1999;Pesaran et al., 2001), Dynamic OLS (DOLS) (Hansen, 1992a(Hansen, , 1992bPhillips and Hansen, 1990), Fully Modified Ordinary Least Squares (FMOLS) (Saikkonen, 1992;and Stock and Watson, 1993) and Canonical Cointegration Regression (CCR) (Park, 1992) methods. Moreover, we account for small sample bias in order to rule out misleading results.

Unit root tests
Since most socio-economic and environmental variables are non-stationary, first we check this property of our variables before proceeding to the cointegration analysis. We use two types of unit root tests for robustness purposes: (a) the Augmented Dickey-Fuller (ADF, Dickey and Fuller, 1981) and the Phillips-Perron (PP, Phillips and Perron, 1988) tests, in which the null hypothesis is a unit root and (b) Kwiatkowski et al. (1992) test, KPSS hereafter, a test with the null hypothesis of (trend) stationarity. We run the tests in two options, i.e., with intercept and trend and with intercept only. Perron (1989) and Hansen (2001), among others, state that conventional unit root tests might be biased towards fail to reject a false unit root null hypothesis in the case of data is trend stationary with a structural break as they do not account for such breaks. Given this point in mind and following the comment from a unanimous referee, we also perform unit root tests with structural breaks. For this purpose, we employ the ADF with structural breaks (ADFBP hereafter) developed by Perron (1989), Vogelsang (1992a, 1992b), and Vogelsang and Perron (1998) 10 . This test will further robustify our interferences and conclusions on the integration orders of the variables. Since these tests are widely used ones, we do not describe them here. Interested readers can refer to the above given references as well as Enders (2010), Perron (2006), Zivot and Andrews (1992), and Banerjee et al. (1992).

Cointegration methods
The current paper employs Johansen (1988) and Johansen and Juselius (1990) full information maximum likelihood method called Vector Error Correction Model (VECM) as a main tool. Since this a widely used method in similar studies, we will not describe it here. The detailed discussion can be found in Johansen and Juselius (1990), Johansen (1992aJohansen ( , 1992b.
The study also employs ARDLBT, FMOLS, CCR and DOLS cointegration methods for robustness check. Due to space limitations, we do not describe ARDLBT, FMOLS, CCR and DOLS cointegration techniques in this section, but the detailed discussions can be found in Pesaran and Shin (1999), Pesaran et al. (2001), Hansen (1992a, co2 ¼ logarithm of per capita CO2 emissions, y ¼ logarithm of per capita GDP, SD ¼ standard deviation, CoV ¼ coefficient of variation, Mean-3*SD ¼ 3 standard deviation from the mean to left, Mean þ 3*SD ¼ 3 standard deviation from the mean to right. 9 Note that selection of the period is based on data availability. 10 The advantage of this test is that it can be applied regardless of whether a break happened immediately or gradually, the break is in trend or level of a given variable or in both, the break date is known or unknown. 1992b), Phillips and Hansen (1990), Hamilton (1994), Park (1992), Saikkonen (1992), and Stock and Watson (1993), inter alia.

Empirical results
This section first discusses the results of testing for unit roots and cointegration, and presents long-run estimation results of the Johansen, ARDLBT, FMOLS, DOLS, and CCR methods.

Unit root tests
The results of the unit root tests are reported in Table 2. We can see that for co 2 emissions in the more general specification with intercept and trend, all conventional tests indicate that the variable is stationary in first differences, i.e. it is I (1). Note that we also run the ADFBP for co 2 to see whether there is any statistically significant breakpoints either at its level or trend or in both but we could not find such an evidence. The situation for y and its powers is not straightforward. When only the intercept is included, the ADF and KPSS tests say that the income variables are I (1), while the PP test rejects the stationarity of the first differences. In the case of the intercept and trend are included in the test equation, all the conventional tests reject the unit root hypothesis in the level of the variables in favor of trend stationarity. According to the discussion in the methodological section above, such a conclusion from the conventional unit root tests might be misleading given that Fig. 1 illustrates that the GDP per capita might have a shift in its level, which starts in 2005 and ends in 2007. Such a break can be caused by expansion in the oil sector of the economy (see Hasanov, 2013 inter alia). We apply the ADFBP test to y and its powers using both test specifications, i.e., with intercept and trend as well as with intercept only. The results tabulated in Table 2 support this evidence.
In other words, the test results indicate that the variables are non-stationary at their levels with the level break and the first differences of them are stationary with one time break in 2005. We further note that the coefficients of the lagged dependent variables of the ADF specification with intercept and trend for y and its powers are found to be À0.2. This means that p in the original ADF specification is 0.8 (¼1e0.2), which is closer to unity, indicating a unit root process. Moreover, it is known that in small samples these tests tend to reject the null. Considering the above mentioned facts, the graphical inspection and common theoretical sense, we conclude that y and its powers are stationary in first differences. We thus conclude that our variables are non-stationary in levels but stationary in their first differences. In other words, they follow integrated of order one, I (1), processes.

Cointegration analysis
Before testing the significance of the cubed and squared terms, we performed cointegration tests for the cubic and squared specifications. The tests concluded co-movement of the variables for either specifications in the long-run. To save the space we do not report the results here but they are available from the authors upon the request. Detailed discussion of the cointegration test results for the final specification are presented below.
As discussed in Section 3.1, if the cubic term is insignificant then it should be dropped from the model. Table A2 in Appendix shows that for all employed methods, except DOLS, the cubic term is insignificant. For DOLS it is significant only at 5% significance level. Moreover, the magnitude of the coefficient is completely different in comparison with other methods. Therefore, based on the fact that we have only one single weak evidence out of five cases, we  Maximum lag order is taken equal to two and optimal lag order (k) is selected based on Schwarz criterion in the ADF test; *, ** and *** indicate rejection of the null hypotheses at the 10%, 5% and 1% significance levels correspondingly; The critical values are taken from Mackinnon (1996) for ADF and PP tests and from Kwiatkowski et al. (1992) for KPSS test.
drop the cubic GDP term from the model. Interestingly, the same holds for the quadratic income variable as well, that is the quadratic GDP term is statistically insignificant in all econometric methods employed, except for VECM. In the VECM results with squared income term, the sign of the trend is opposite to the conventional one. Moreover, based on the estimation results the turning point (3.533/2 Â 0.200 ¼ 8.833) occurs outside of the U-shaped relationship, which can be interpreted as monotonically decreasing relationship. But these two facts are opposite to the conventional known facts. Therefore, based on the estimation results from the five different methods we conclude that the squared term should also be dropped from the model. The coefficient b 1 is statistically significant in the case of the linear model. Therefore, we will proceed in our empirical analysis on the basis of a linear specification. Mathematically speaking the equation (2) reduces to the following form: The following discussion is based on this specification. We first test for the existence of a long-run relationship among the variables involved and then turn to the estimated parameters of such relationship. We first apply the Johansen cointegration approach to equation (3) to see if there is one cointegration vector, because it is known that in the case of n variables there are at most n-1 long-run relationships. To apply the Johansen procedure, the optimal lag number should first be chosen. A Vector Auto Regressive (VAR) model was initially specified with the endogenous variables of co 2 and y, and exogenous variables intercept, trend and a pulse dummy. 11 The trend is included in order to see whether or not it has any power in explaining the behavior of the variables, especially y, as they are trending over time: if we excluded it, then our VAR would have instability problems. 12 A maximum of two lags was initially considered and both lag selection criteria and lag exclusion tests statistics proposed that indeed a lag of order two was optimal. It is worth nothing that the VAR with 2 lags well passes all the residual diagnostics tests, as shown in Panels A through C of Table 3, as well as the stability test. The Johansen cointegration test results are given in Panels D and E of Table 3.
Although Table A2 has shown that the cointegrating equation is mainly linear, as in columns (c) or (d) of Panel D of Table 3, we test for the presence of long-run relationship in all potential 5 test types. We can see that type (e) reports no cointegration equation. It is difficult to believe that the variables are cointegrated with a quadratic trend because, first, the unit root tests do not show any non-stationarity with a quadratic trend and, second, it is a very rare case for socio-economic variables and, finally, it is hard to interpret economically. For cases (b), (c) and (d) both the trace and the maxeigenvalue test statistics indicate one cointegration relationship among the variables. The results of the small sample corrected version of the trace and max-eigenvalue tests are given in the Panel E. Here, again both tests point in favor of the existence of one cointegrating relationship, trace at 10% and max-eigenvalue at 5% significance levels. Hence, we conclude that there is a long-run association among the variables.
The Johansen method outperforms all its alternative methods in the case of more than two variables in terms of properly determining the number of long-run relations. This is why we adopted it. However, we also employed the ARDLBT and Engle-Granger type DOLS, FMOLS and CCR methods to test whether the variables are cointegrated. The results from the ARDLBT, even after correcting for small sample bias using Narayan (2005) critical values, and other three methods also indicate that there is a cointegrating relation among the variables. This indicates that the cointegration results from the Johansen method are robust. 13 The ARDLBT, FMOLS, DOLS and CCR methods were also employed as a robustness check alongside the VECM in estimating the long-run coefficients. The  Urzua (1997) system normality test's null hypothesis is residuals are multivariate normal; c The used White Heteroscedasticity Test's null is no cross terms heteroscedasticity in the residuals; c 2 is the Chi-square distribution; d.f. ¼ degree of freedom; C is intercept and t stands for trend. r is the number of long-run equations; l trace is Trace statistics while l max is Max-Eigenvalue statistics, l a trace and l a max are their adjusted forms; *** , ** and * represent rejection of null hypothesis at the 1, 5 and 10% significance levels correspondingly; Critical values for the cointegration test are from Mackinnon et al. (1999). 11 The pulse dummy, equal to one in 2007 and zero otherwise, is included in order to capture the jump of co 2 in 2007. 12 Juselius (2006) discusses that one should take care of other variables along with a variable of interest since VAR is a system of variables. All the intermediate results are not given here to save on space, but are available from the authors under request. 13 The results are not reported here but are available from the authors under request.
H 0 is rejected.

2.
H 0 is accepted.  results are presented in Table 4. 14 As it can be seen from Table 4 the long-run coefficients from the five different techniques are very close to each other in terms of sign and magnitude and they all are statistically significant. Additionally, the residuals of the estimated specifications well pass the residuals diagnostics tests which is another sign of the robustness of the estimation results. The long-run elasticity of carbon emission with respect to income is around 0.7, as the magnitude of the estimated income coefficients ranges from 0.697 to 0.823. In the models with trend the estimated coefficient of this variable lies between À0.067 and À0.078.

Discussion of the empirical results
The results from the unit root tests, given in Table 2, indicate that the levels of the variables follow a unit root process. This implies that any shock to these variables will have a permanent effect and therefore they will deviate from their underlying development path. As a piece of evidence of this fact, for example, the global financial crisis in 2008 has changed significantly the development path of GDP as can be seen in Fig. 1. The figure also shows that there has not been a permanent change in the development path of carbon emission. This does not necessarily mean that this variable is not non-stationary, but rather indicates that the crisis has not had a permanent effect on it. Different variables simply can react to shocks differently. The concept is also true when the variables are positively shocked. Moreover, having a unit root process implies that the variables contain a stochastic trend, so that it is difficult to predict futures values of them. The implication for policy makers and forecasters is that they should consider growth rates rather than levels of the variables in their policy analysis and projections.
The finding of a cointegrating relationship among the variables, as reported in Table 3, implies that there is a stochastic trend which is the same for all of them. In other words, there is a long-run relationship among carbon emission and GDP. Since such a relationship exists, it is useful for policy analysis and forecasting purposes to estimate numerical values, i.e., parameters (especially elasticities) of this long-run relationship. To this end, we estimated the dependence of the carbon emission from GDP employing the five different cointegration methods as a robustness check. The desirable outcome from these estimations, reported in Table 3, is that they produce consistent results numerically, statistically and conceptually: as the magnitude of the corresponding coefficients are close to each other, they all are statistically significant and have the same expected signs.
For all methods, the estimated coefficient of the income variable has a positive sign, which implies a linear relationship (monotonically increasing) among income and emissions. This says that the EKC does not hold for Azerbaijan over the period analyzed. It is noteworthy that some previous panel studies also found the similar results. For example, recent studies such as Ito (2017) and Mitic et al. (2017) find a monotonically increasing relationship for the panel with Azerbaijan. Thus, we conclude that there is a linear relationship between GDP and CO2 emissions. This implies that an increase in GDP results in an increase in environmental pollution in Azerbaijan. Such a conclusion is quite reasonable in the sense that the EKC usually holds for developed countries and Azerbaijan is a developing economy.
As can be seen from Table 4, based on the VECM approach, the income elasticity of CO2 emissions is 0.823. Hence, all other things being equal, a 1% increase in GDP leads to 0.823% increase in carbon emissions. The estimated income elasticity obtained by Brizga et al.  H 0 is rejected.
Notes: Dependent variable is co 2t ; Coef. and Std. Er. denote coefficient and standard error; *, ** and *** indicate significance levels at 10%, 5% and 1%; Probabilities are in brackets; SoA ¼ Speed of adjustment; Q ARð2Þ ¼ Q-statistic from testing AR(2) process; LM SC ¼ Lagrange multiplier statistic of serial correlation test; c 2 HETR ¼ Chi-squared statistic for heteroscedasticity test; JB N ¼ Jarque-Bera statistic for testing normality; In VECM, Jarque-Bera statistic was taken from the option of Orthogonalization: Residual Correlation (Doornik-Hansen). Intercepts of the long-run equations are not reported for simplicity. 14 Note that we set a maximum lag order equal to two in the ARDLBT estimation as we did for the VAR. Then optimum lag order for dependent variable and regressors is selected with the Schwarz criterion. In the DOLS estimation, we set maximum number of lag and lead order to one and the optimal order is selected by the Schwarz criterion because of the same reason. Since there is no dynamic part in FMOLS and CCR methods, we used a pulse dummy taking on unity in 2007 and zero otherwise, to capture the sharp decrease in CO2 emission in 2007. Also, note that we included a time trend in the ARDLBT, FMOLS, DOLS and CCR to capture technological and other changes and it appeared to be significant in all estimators.
(2013) for the panel of former Soviet countries including Azerbaijan was equal to 0.86%. The difference between our finding and previous results in terms of magnitudes might be due to the use of individual countries data in our case, while all previous studies employed panel of countries. If we compare the obtained income elasticity of CO2 emissions with the results previous studies in similar country cases, we see that magnitude wise our findings are close to the results of Narayan and Narayan (2010) for Bahrain and Syria cases, Al-Mulali and Tang (2013) for Oman, Omri (2013) for Oman, Qatar, and Saudi Arabian cases and Bekhet and Yasmin (2017) for Kuwait case. Since, in many of the reviewed papers the income elasticities are not reported we cannot compare our findings with the results of those studies. Furthermore, since some of the previous studies employ in their specification of CO2 emissions, which is calculated based on energy use data, energy consumption as an independent variable which in its turn causes biases in the coefficients as discussed in Jaforullah and King (2017), the findings of those papers suffer from being compared with.
As a further robustness check, we applied a Threshold Regression (TR) model to our data and the results showed that there is not a threshold value, i.e. a turning point in the relationship, which also can be seen as an evidence against non-linear relationship between income and CO2 emissions.
The finding of a monotonically increasing relationship between emissions and income can be explained as follows. Since, after the "Contract of the Century" the Azerbaijan economy has been benefiting from sizeable oil revenues. This in turn caused an increase in environmental degradation as result of the revitalization of the economy after the collapse of the Soviet Union. Moreover, CO2 emissions started to increase again after the beginning of the reinvigoration process of the industrial sector in 2009. The construction and launching of a number of new factories and technocity, as well as recovering of the old ones, might be the main reason of this increase. 15 Finally, long-run estimates show that carbon emission declines on average 7% per annum over the period 1992e2013, which can be considered as a result of technological improvement with other implicit factors. The negative sign of the trend variable is in line with the emission-income relationship. In other words, the effect of development in employed technologies on the environmental degradation is expected to be positive, which implies the expected sign of the trend variable is negative.
Among others, one of the solutions to reduce carbon emission is to decrease energy intensity. To shed further light on this point, Azerbaijan was able to reduce energy intensity -which is calculated as energy for every dollar of GDP output at market exchange rates from 0.42 in 1990 to 0.24 in 2013, while CO2 emissions has decreased by 56.8% over the period (Vazim et al., 2016). These numbers show that Azerbaijan gained considerable achievement in the reduction of the CO2 emissions and energy intensity during the period of investigation. It has to be noted that the reduction in energy intensity in Azerbaijan can be the result of two different issues. On the one hand, as the country develops, modern technologies are used in sectors of the economy, which leads to decrease in the carbon emission. On the other hand, over the period a lot of plants and factories, which were mainly the large carbon emitters remained from the former Soviet Union period were shut down. Instead, other less carbon emitting sectors, like services, tourism etc., developed and grew (Hasanov, 2013;Oomes and Kalcheva, 2007). Nevertheless, compared with the world average, the CO2 intensity in Azerbaijan was 1.1 mt in 2011, while the world average figure was 0.6 mt, which is 1.8 times smaller. Energy intensity was 19376 Btu in Azerbaijan in 2011 and this figure is almost 3 times higher than the world average of 9905 Btu. 16 To put it differently, Azerbaijan spends three times more energy than the world average to generate each dollar value added. Menyah and Wolde-Rufael (2010) among others point out that some countries endowed with abundant energy resources experience inefficiency in their energy use.
The above discussion highlights that energy inefficiency is one of the main challenges for the country. The energy intensity can be reduced by two channels: (a) using less energy intensive equipment and technologies in cement production and power generation to save energy and decrease a loss during distribution and transmission; (b) implementing different tariff mechanisms and cutting subsidies.

Conclusion and policy implications
This study investigates carbon dioxide emission effects of economic growth in Azerbaijan using annual data for the period 1992e2013. The Johansen, ARDLBT, FMOLS, DOLS and CCR cointegration methods are employed to analyze the long-run relationships between the variables. The methods produced consistent results, which can be considered as a sign of robustness of our findings. The results endorse the validity of a cointegrating relationship among the variables. The estimation results point to the invalidity of the EKC hypothesis in Azerbaijan. The relationship between CO2 emissions and income is found to be monotonically increasing. The other finding of our study is that economic growth has a positive and statistically significant impact on carbon emissions in the long-run. In comparison with the World's average figures, in terms of CO2 emissions, each dollar costs 1.8 times more than the World's average. The country has a potential to materialize economic development implementing energy conservative measures without causing an increase in CO2 emissions (Opitz et al., 2015 inter alia). Moreover, the Azerbaijani government planned to bring down the amount of carbon dioxide in line with the appropriate figures of the OECD countries by the end of 2020. Therefore, a suitable environmental policy to reduce total CO2 emissions without harming economic growth is to improve energy efficiency, which can be obtained by increasing optimal infrastructure investment and employing energy conservative policies to avoid unnecessary use of energy. Put differently, using less energy intensive technologies, minimizing the loss of power during distribution and transmission processes, and employing different tariff mechanisms to control energy use are some applicable policies that are capable to increase energy efficiency. Although, the results of the current study do not explicitly enlighten the sources of the factors which causes increasing response of the CO2 emissions to economic growth, as a developing country case some measures need to be considered to reach the sustainable development path. Implementation of carbon pricing instruments and mechanisms in production and international-domestic trade activities, nationwide social awareness programs to enlighten the negative consequences of the pollution, initiating government/ nation sourced promotion policies to intensify the use of modern, less energy/pollution intensive technologies can be considered as the first-to-be-done policies in the similar economies.

Acknowledgements
We are grateful to anonymous referees, and the editors for their comments and suggestions that have helped to improve considerably the paper; nonetheless, we are of course responsible for all errors and omissions. Finally, the views expressed in this paper are those of the authors and do not necessarily represent the views of their affiliated institutions.    Notes: Dependent variable is co 2 . ** and *** mean significance at the 5% and 1% levels, respectively.