The impact of the transition towards renewable energy
consumption on economic growth
Lammert Dijkstra
*Faculty of Economics and Business, University of Groningen
June 2018
Abstract:
This paper aims to investigate the impact of the energy transition on economic growth, which is measured by the net effect of replacing non-renewable energy consumption by renewable energy consumption. The starting point is the Cobb-Douglas production function, which enables the calculation of the marginal effects for both renewable and non-renewable energy consumption once the output is known. The output is obtained by estimating several estimators, each correcting for different econometrical problems. The FMOLS estimator is considered to be the most accurate. The results favour the case of a positive net effect of the energy transition for 26 OECD countries over the period 1971-2014. Especially, high income countries and countries with a low share of renewables can greatly benefit from this transition. Furthermore, as time progresses the energy transition is expected to become increasingly profitable. It is also concluded that the impact of the transition on economic growth increases as the distance to the optimal share enlarges. Keywords: energy transition, renewable energy, economic growth, OECD members, panel data. JEL classification: C23, O44, Q42, Q43.
1. INTRODCUTION
Fossil fuels are the main resource for the production of energy, therefore the extraction of coal, gas and oil is essential to meet the energy demand of today. For instance, the United States relies for about 80 percent on these resources to produce energy. However, these resources are extracted at larger amounts than is created by nature, implying that these resources are categorized as non-renewable. This also means that these resources will eventually run out, and the world has to search for alternative sources of energy. In
* E-mail address: lammertdijkstra6655@gmail.com
Student number: 2729199
2 recent times there has been a lot of debate on the exact moment these resources will be fully depleted. Shafiee and Topal (2009) estimated that oil will be fully depleted in 35 years, coal in 107 years and gas in 37 years. Klass (2003) even suggest that five times the amount of proven reserves, for oil and natural gas, cannot stop complete depletion before the end of the 21th century. The strong dependence on non-renewable
sources to supply energy makes these numbers even more alarming, and confirms the need for alternative energy sources. Optimally, the quest has to go towards renewable energy sources, of which hydro, wind, solar, bio and geothermal power are important sources. In contrast to fossil fuels these renewable energy sources will not exhaust by extraction. However, tremendous investments in technologies are needed to make renewable energy affordable and applicable.
However, due to recent discoveries of fossil fuels the scarcity argument becomes less relevant in the discussion concerning the energy transition from non-renewable energy sources towards renewable energy sources. The damage caused to the environment by burning fossil fuels is overtaking scarcity as the main argument for the energy transition. Energy consumption, cattle breeding and deforestation are mainly held responsible for the increase in emissions of greenhouse gases, which is mostly seen as the origin of climate change around the globe (European Commission, 2015). Consequences of climate change are rising sea levels, rising temperatures and droughts. In addition, human health is affected by climate change through e.g. air pollution and spread of diseases. Therefore, it is necessary to keep climate change and the associated, negative, consequences within boundaries. Kaygusuz (2007) argues that the transition towards renewable energy is an efficient, clean and sustainable solution to address these negative consequences in the case of Turkey, but this is likely to hold in general.
3 the transition from non-renewable energy consumption towards renewable energy consumption impact economic growth? This is estimated by simultaneously including renewable energy consumption and non-renewable energy consumption into the Cobb-Douglas (1928) production function. This enables the possibility to capture the effect of replacing non-renewables by renewables by equal amounts, using their marginal effects to compute the net effect. The log-linearized Cobb-Douglas function is estimated in order to obtain the elasticities needed to calculate the marginal effect. The results indicate that increasing energy consumption is associated with increasing economic growth for both renewables and non-renewables, consistent with literature of Apergis and Payne (2012), Salim et al. (2014) and Bhattacharya et al. (2017). Furthermore, the results suggest a positive effect of the energy transition on economic growth. Also, the future impact of the energy transition on economic growth is discussed on the basis of trends from different subsamples.
This paper starts with a literature review, where existing literature is presented and discussed. This is followed by a section devoted to the methodology and a section discussing the data. Thereafter, the empirical analysis is executed, where both the data and the model are tested for econometrical problems. Then the results are presented and the hypotheses are examined. Furthermore, the implications of the results are discussed and subsamples are examined. Then a section is devoted to limitations and possibilities for further research. Lastly, the paper is concluded.
2. LITERATURE REVIEW
The literature related to (renewable) energy consumption is mostly focused on the direction of causality. Therefore, this topic is discussed in detail, for both renewable and total energy consumption. Furthermore, the usefulness of a Cobb-Douglas production function, including energy related variables, as starting point is discussed. In addition, the expected consequences of the energy transition, so the replacement of non-renewable energy by non-renewable energy, on economic growth is discussed. All papers mentioned here relate to (a subset of) OECD countries, unless stated otherwise.
2.1 Economic growth and energy consumption:
4 Cleveland et al. (1984) are among the first discussing the relationship between economic growth and energy use. Their findings show that there is a clear positive association between these two variables for the Unites States from 1890-1980. However, even though there is a positive relationship between economic growth and energy use, the direction of causality has to be determined before strong conclusions and recommendations can be made. Belke et al. (2011) provide a clear overview of the direction of causality observed in previous papers, where bi-directional causality between economic growth and energy consumption is most common for OECD countries. Lee and Chang (2007); Lee et al. (2008); Lee and Lee (2010); Costantini and Martine (2010); and Belke et al. (2011) conclude bi-directional causality using samples consisting of at least 22 OECD countries. Contrastingly, Huang et al. (2008) find unidirectional causality, from economic growth to energy consumption, for 26 high income countries, of which 24 countries are member of the OECD. They find evidence that economic growth reduces energy consumption in high income countries, because environmental protection becomes more important. Chontanawat et al. (2006) test the direction of causality for 30 OECD countries individually, and find that bi-directional causality is present in 40 percent of all cases. Hence, the existing literature favours the case of bi-directional causality between economic growth and energy consumption, rather than the alternatives.
The previous studies have in common that they apply a log-linear functional form to quantify the results. However, different variables and estimators are used, resulting in different results. For instance, Belke et al. (2011) estimate that a one percent increase in energy consumption leads to a 0.6 percent increase in GDP, using the dynamic OLS estimator (DOLS) for the period 1981-2007. Contrastingly, the coefficient found by Lee et al. (2008) is much lower, where the percentage change in GDP is estimated to be only 0.25 percent, using the fully modified OLS estimator (FMOLS) for the period 1960-2001. Furthermore, Lee and Chang (2007) use a panel VAR model for the period 1965-2002. The results indicate that the lag of energy consumption has a positive, but insignificant effect on GDP. The second lag is negative and significant at the 10% level, but is small in magnitude as the elasticity is only -0.05. Finally, Lee and Lee (2010) only estimate the coefficient relating GDP with total energy demand, for the period 1978-2004 using FMOLS. They state that a one percent increase in GDP implies a 0.52 percent increase in total energy demand.
2.2 Economic growth and renewable energy consumption:
In the literature there are four hypotheses which can describe the causal relation between economic growth and renewable energy consumption. These are the growth hypothesis, the feedback hypothesis, the conservation hypothesis and the neutrality hypothesis. Note that these hypotheses can also be applied to the relationship between total energy consumption and economic growth.
5 and long run causality in both directions, for the period 1985-2005. Furthermore, Apergis and Payne (2012) conclude that increasing the sample size to 80 countries and adjusting the period to 1990-2007 does not affect the results. Therefore, the feedback hypothesis does not only apply to OECD members. Contrastingly, Salim et al. (2014) only find evidence for the feedback hypothesis in the long run, and unidirectional causality, from renewable energy consumption to GDP growth, in the short run, for the period 1980-2011. Omri et al. (2015) test causality for countries separately, and concludes bi-directional causality for the United States, Belgium, Canada and France.
Secondly, in the case of the growth hypothesis renewable energy consumption results in economic growth, implying unidirectional causality. This is observed for Hungary, Japan, the Netherlands and Sweden (Omri et al., 2015) and for Germany (Chang et al., 2015) over the period 1990-2011. Another interesting result is that of Yildirim et al. (2012), who test this hypothesis for different renewable energy sources for the United States for the period 1970-2010. The results indicate that only energy consumption derived from biomass waste Granger causes GDP, whereas the opposite does not hold. The other sources, including hydropower, geothermal and biomass-wood, show no evidence for a causal relationship.
Thirdly, the conservation hypothesis corresponds to the other case of unidirectional causality, i.e. economic growth causes changes in renewable energy consumption. This is the case for the United Kingdom and France (Chang et al., 2015); and for Spain and Switzerland (Omri et al., 2015). Also, Menyah and Wolde-Rufael (2010) find evidence supporting this hypothesis for the United States for period 1960-2007. Furthermore, it can be concluded that some conclusions of these authors conflict with previous ones. For instance, Omri et al. (2015) concluded bi-directional causality for France and the United States, instead of unidirectional. Therefore, there is no decisive conclusion on the direction of causality for specific countries. Finally, the neutrality hypothesis states that there is no causal relationship in either direction. This is the case for Finland (Omri et al., 2015), Canada, Italy, and the United States (Chang et al., 2015; Tugcu et al., 2012). Also, Menegaki (2011) does not find evidence for a causal relationship for 27 European countries, of which 23 are member of the OECD, for the period 1997-2007. Again, conflicting conclusions are exposed. For example, the United States is associated with the neutrality hypothesis (Chang et al., 2015; Tugcu et al., 2012), the conservation hypothesis (Menyah and Wolde-Rufael, 2010) and the feedback hypothesis (Omri et al. 2015).
6 2.3 Economic models used in existing literature:
The common starting point to empirically test the relationship between economic growth and (renewable) energy consumption is the neoclassical Cobb-Douglas production function. This production function represents the relationship between output, i.e. production, and the inputs, which are capital and labour. Equation (1) presents the original Cobb-Douglas production function.
(1) 𝑌𝑖𝑡 = 𝐴𝐾𝑖𝑡𝛼𝐿𝑖𝑡
𝛽
, 𝛼, 𝛽 > 0,
where 𝑌𝑖𝑡 is the output/production, 𝐴 is the state of technology, 𝐾𝑖𝑡 is the capital stock, 𝐿𝑖𝑡 is the labour force and the parameters 𝛼 and 𝛽 represent the elasticity of output with respect to capital and labour.
The Cobb-Douglas production function has multiple unrealistic assumptions. For instance, the assumption of the elasticity of substitution equalling unity is made, which turns out to be unrealistic. This is more extensive discussed in Section 3. Furthermore, Equation (1) assumes Hicks-neutral technological progress. This implies that A changes with technological progress, but the balance of the inputs remains unaffected. Furthermore, this implies that the marginal effect of all inputs change in equal proportions, given a change in the state of technology. So, this does not allow for specific technological change.
However, energy plays an essential role in the production process, as is argued by Stern (1997, 2011). Moreover, it is argued by Salim et al. (2014) that energy can constrain or enable production. Hence, energy should be considered as an important input of the production process. Therefore, an energy component is included into Equation (1), resulting in:
(2) 𝑌𝑖𝑡 = 𝐴𝐾𝑖𝑡𝛼𝐿𝑖𝑡
𝛽
𝐸𝑖𝑡𝜃 , 𝛼, 𝛽, 𝜃 > 0,
where 𝐸𝑖𝑡 is the energy component and 𝜃 is the elasticity of output with respect to energy. By taking the natural logarithm of Equation (2) an equation suitable for linear regression is derived:
(3) ln 𝑌𝑖𝑡 = ln 𝐴 + 𝛼 ln 𝐾𝑖𝑡+ 𝛽 ln 𝐿𝑖𝑡+ 𝜃 ln 𝐸𝑖𝑡, 𝛼, 𝛽, 𝜃 > 0.
7 There are also models including both renewable energy consumption and non-renewable consumption (e.g. Apergis and Payne, 2012; Salim et al., 2014; and Bhattacharya et al., 2017). Consequently, the model will adjust to:
(4) ln 𝑌𝑖𝑡 = ln 𝐴 + 𝛼 ln 𝐾𝑖𝑡+ 𝛽 ln 𝐿𝑖𝑡+ 𝛾 ln 𝑅𝐸𝑁𝑖𝑡+ 𝛿 ln 𝑁𝑅𝐸𝑁𝑖𝑡, 𝛼, 𝛽, 𝛾, 𝛿 > 0,
where 𝑅𝐸𝑁𝑖𝑡 is renewable energy consumption, 𝑁𝑅𝐸𝑁𝑖𝑡 is non-renewable consumption, 𝛾 and 𝛿 are the elasticities of output with respect to renewables and non-renewables. These elasticities are not able to answer the research question directly. This is because a one percent change in renewable energy consumption does not equal a one percent change in non-renewables, since the levels are different. However, these elasticities are useful in calculating the marginal effect of renewable and non-renewable energy consumption, which will be used to answer the research question. The calculation of the marginal effects of (non-)renewable energy consumption has not been done in literature yet. However, it is important to quantify the precise impact of the energy transition on economic growth. Hence, this functional form is the fundament of this paper, since it allows for the calculation of the net effect of the energy transition on economic growth.
Another alternative is to include the share of renewables in total energy consumption into the model (Menegaki, 2011; and Inglesi-Lotz, 2016). This allows to observe the impact of changes in the share of renewable in total energy consumption on economic growth. In this case, the model is similar to:
(5) ln 𝑌𝑖𝑡 = ln 𝐴 + 𝛼 ln 𝐾𝑖𝑡+ 𝛽 ln 𝐿𝑖𝑡+ 𝜑 ln 𝜔𝑅𝐸𝑁, 𝛼, 𝛽 > 0,
where 𝜔𝑅𝐸𝑁 is the share of renewables in total energy consumption and 𝜑 is the elasticity of output with respect to this share and can both take positive and negative values. The results by Inglesi-Lotz (2016) show a positive and significant coefficient for the share of renewables in total. This seems to suggest that the transition from non-renewables towards renewable energy is beneficial for economic growth. However, this model has a pitfall, because there are multiple ways to increase this share. First, by decreasing non-renewable energy consumption the denominator will decline and consequently the share will increase. Second, only the numerator can increase when non-renewables are replaced by renewables. Third, increasing renewables and leaving non-renewables unaffected.1 Although this model looks promising it
cannot be used for the purpose of this paper, since the origin of an increasing share cannot be determined. Hence, the log-linear functional form which includes both renewable and non-renewable energy consumption is used in this paper. This allows for the calculation of the marginal effect of the variables, but also of the net effect of the energy transition on economic growth.
1 Given an fixed amount of non-renewable energy consumption: lim
8 2.4 Economic growth and the energy transition:
Although the relationship between economic growth and the energy transition has not been tested empirically, there are scholars addressing this problem theoretically. The argument mentioned most is the price difference between renewables and non-renewables. Energy derived from renewable sources are more expensive than that of non-renewable sources. As a result, economic growth will fall, since valuable resources are not used in the most effective manner. However, there are social cost, due to negative externalities, associated with non-renewable energy consumption, which are not internalized. Owen (2006) argues that energy prices of renewable- and non-renewable energy would be similar if these externalities are taken into account. Therefore, the government can play an important role. By implementing a tax on non-renewable energy both sources can become competitive, and economic growth will not suffer from choosing the environmentally friendly option.
Marques and Fuinhas (2012) also stress the importance of cost differential between renewables and non-renewables. They state that the positive effect of renewable energy, i.e. job creation and wealth generation caused by local production, do not outweigh the negative effects, i.e. the opportunity costs. It is argued that the cost of promoting renewable energy sources will be transferred upon the economy and therefore reduces economic growth.
Next to the cost differentials and externalities, Fang (2011) also argues that the energy transition increases imports, if knowledge and technology is not adequately available. Furthermore, it is argued that governments should implement complementary policies, which both enhance economic growth and environmental protection. Even though Fang (2011) focuses on China, these conclusions will also hold for other countries trying to simultaneously improve economic growth and the environmental situation.
Roelofsen et al. (2016) argue that the transition towards renewable energy can also facilitate economic growth. It is discussed that the transition creates jobs, also in the long run, since renewable energy sources are more labour intensive than non-renewable energy sources. Furthermore, the possibility to become market leader in renewable technologies might result in increased exports. However, renewable energy also requires huge investments in infrastructure, i.e. the installation of the decentralized renewable energy sources to the energy grid. Again, this paper circumscribes a specific country, the Netherlands, but could be applied to other knowledge-abundant countries as well.
3. METHODOLOGY
9 are based on average numbers. Thus, for small countries these marginal effects are huge, whereas it is small for the larger countries. By dividing the variables by the population, these differences are reduced and this allows for a broader application of the results. As a result, Equation (2), with 𝑟𝑒𝑛 en 𝑛𝑟𝑒𝑛 replacing 𝐸, modifies to: (6) 𝑦𝑖𝑡 = 𝐴𝑘𝑖𝑡𝛼𝑙𝑖𝑡 𝛽 𝑟𝑒𝑛𝑖𝑡𝛾𝑛𝑟𝑒𝑛𝑖𝑡𝛿𝑝𝑜𝑝𝑖𝑡 𝛼+𝛽+𝛾+𝛿−1 , 𝛼, 𝛽, 𝛾, 𝛿 > 0.
Here, 𝑦𝑖𝑡 is the output per capita, where GDP/population is used as proxy. Furthermore, 𝑟𝑒𝑛𝑖𝑡 is the amount of renewable energy consumption per capita and 𝑛𝑟𝑒𝑛𝑖𝑡 is non-renewable energy consumption per capita. Moreover, 𝑙𝑖𝑡 is the share of the labour force in the population of a country at time t. 𝑘𝑖𝑡 represents the capital stock per capita and 𝑝𝑜𝑝𝑖𝑡 is the population of country i at time t. The exponents are the elasticities of output with respect to the corresponding variable.
In contrast to Lee et al. (2008) and Omri et al. (2015), the assumption of constant returns to scale (i.e. 𝛼 + 𝛽 + 𝛾 + 𝛿 = 1) is avoided in this paper, because for most estimations this assumptions is rejected based upon statistical analysis. Therefore, this paper allows for varying returns to scale, where increasing returns seem to be the most realistic. However, this raises a potential problem, namely the elasticity of output with respect to population is unequal to the sum of all other elasticities minus 1 in most estimations if this elasticity is unrestricted. Therefore, population has also some explaining power, which is not grounded in literature when the labour force is also included. This is because population can be seen as a demand factor, whereas the variables in the Cobb-Douglas production function are all inputs of the production process. However, this can be seen as a minor problem compared to the unrealistic assumption of constant returns to scale.
Since the Cobb-Douglas production function is used, the elasticity of substitution is assumed to equal 1. However, as previously discussed, the substitutability of energy is low compared to capital and labour (Stern, 1997; 2011). This indicates that this assumption is unrealistic, since the elasticity of substitution is probably lower than one. Consequently, a change in the price ratio will have an effect on the ratio between the inputs, but with proportionally less than the change in the price ratio. Therefore, energy and the other inputs are called gross complements, as there is limited substitutability. This might have an impact on the results, which will be discussed in Section 6.
Before anything can be said about the impact of the energy transition on economic growth, the elasticities needs to be estimated. This is done by estimating Equation (7), which is the log-linearized functional form of Equation (6), since it allows for linear regression:
10 where 𝜌 is the elasticity of output with respect to population. The estimators used in this paper do not allow for restrictions and therefore 𝜌 is allowed to move freely. Hence, population might have some explaining power as well. Moreover, based on existing literature the expectation is that increases in renewable or non-renewable energy consumption are likely to positively contribute to economic growth (i.a. Apergis and Payne, 2012; Salim et al., 2014; and Bhattacharya et al. 2017). Therefore, the elasticities are expected to be positive and significant. This introduces the first two hypotheses of this paper, namely:
Hypothesis 1: Increasing non-renewable energy consumption positively impacts economic growth. Hypothesis 2: Increasing renewable energy consumption positively impacts economic growth.
Moreover, the elasticities of output with respect to capital, labour and population are expected to be positive. Once the elasticities are estimated, the research question can be answered by observing and comparing the marginal effect of renewable (𝑀𝐸𝑟𝑒𝑛) and non-renewable energy consumption (𝑀𝐸𝑛𝑟𝑒𝑛) on output. The marginal effect of these variables, using mean values to represent the sample, equals:
(8) 𝑀𝐸𝑟𝑒𝑛 = 𝜕𝑦/𝜕𝑟𝑒𝑛 = 𝛾(𝑦̅/𝑟𝑒𝑛̅̅̅̅̅),
(9) 𝑀𝐸𝑛𝑟𝑒𝑛= 𝜕𝑦/𝜕𝑛𝑟𝑒𝑛 = 𝛿(𝑦̅/𝑛𝑟𝑒𝑛̅̅̅̅̅̅̅),
where 𝑦̅ is the mean of GDP per capita and (𝑛)𝑟𝑒𝑛̅̅̅̅̅̅̅̅̅ is the mean of (non-)renewable energy consumption per capita. These marginal effects represent the changes in output per capita, i.e. economic growth, resulting from changes in per capita energy consumption. Given equal changes in renewable and non-renewable energy consumption, but of opposite sign, the impact on economic growth by the energy transition can be investigated, using Equation (10):
(10) 𝐸[∆𝑦| ∆𝑟𝑒𝑛 = −∆𝑛𝑟𝑒𝑛] = 𝑀𝐸𝑟𝑒𝑛∙ ∆𝑟𝑒𝑛 + 𝑀𝐸𝑛𝑟𝑒𝑛∙ ∆𝑛𝑟𝑒𝑛 = [𝑀𝐸𝑟𝑒𝑛− 𝑀𝐸𝑛𝑟𝑒𝑛] ∙ ∆𝑟𝑒𝑛.
The expectation is that the replacement of non-renewable energy consumption by renewable energy consumption negatively affects the economy, with cost differentials being the most decisive reason (Marques and Fuinhas, 2012; and Owen, 2006). Thus, finding negative values for Equation (10) would support this statement. This immediately states the most important hypothesis of the paper, which is:
Hypothesis 3: The replacement of non-renewable energy consumption by renewable energy consumption lowers economic growth.
11 2006). However, the CCEMG will yield inconsistent results, because the CCEMG fails to take bi-directional causality into account. The only estimator that provides accurate estimates in the presence of both serial correlation and endogeneity is the FMOLS (Pedroni, 2000). Here the group mean FMOLS is used, which estimates all panels individually and takes the average of these coefficients to get a sample coefficient. The results of this paper will be mainly focused upon the estimations of the FMOLS, as it provides accurate results in the presence of the assumed bi-directional causality. However, heteroscedasticity is not taken into account, but in Section 5 it is shown that this might be less of a problem. Furthermore, cross-sectional dependence is not solved.
4. DATA
The annual data from 1971 to 2014 is derived from the International Energy Agency (IEA), the World Development Indicators’ database made available by the World Bank, the Investment and Capital Stock Dataset of the IMF and the OECD Indicators. The data for the variables GDP and capital are expressed in constant 2011 international dollars per capita, which implies the data is corrected for inflation and for purchasing power differences. The energy consumption variables are defined as the tonne of oil equivalent (toe) in per capita terms, where one toe equals the energy released by burning one tonne of oil. Renewable energy consumption consist of energy consumption from renewable energy sources, such as hydro, wind, solar and geothermal etc.; and from waste sources, such as industrial and municipal waste. Non-renewable energy consumption is the sum of all other sources of energy consumption, including i.a. (crude) oil, natural gas, nuclear and coal. The labour force includes the employed and the unemployed actively looking for work and is also divided by the population.
In many papers, e.g. Apergis and Payne (2010); and Salim et al. (2014), real gross fixed capital formation (GFCF) is used as a proxy for capital. However, this is problematic, since capital is a stock variable whereas GFCF is a flow variable. They justify the use of GFCF by stating the close relationship between the variance of capital and the change in investment. Despite this justification, I will use the estimation of the capital stock made available by the IMF (2017).2 This paper only includes the private and
the public capital stock, since the inclusion of the private-public partnership capital stock would substantially reduce the sample size. This is unlikely to impact the results, because the private-public partnership stock represents a small fraction of total capital stock.
2 This dataset is constructed using this perpetual inventory equation: 𝐾
𝑖,𝑡+1= (1 − 𝛿𝑖𝑡)𝐾𝑖,𝑡+ (1 − 0.5𝛿𝑖𝑡)𝐼𝑖,𝑡 for the period
12 Furthermore, a requirement on the number of observations is imposed to effectively test for stationarity and cointegration. All countries having 20 or less observations are excluded. Consequently, nine OECD countries are left outside the analysis, and 26 countries remain.3 Even after the exclusion of these
countries, the panel data remains unbalanced. For the number of available observations per country see Appendix A. The potential amount of observations is 1144, but due to missing values this number reduces to 1049. Moreover, renewable energy consumption per capita is zero at some moments. This reduces the sample size to 950, since the natural logarithm cannot be taken of these observations. So, this sample includes approximately 36.54 observations per country on average where 44 is the maximum.
In Table 1 the main statistics of the variables specified in Equation (6) are presented. Panel data makes it difficult to interpret these statistics, since the countries to which these statistics apply are not known. However, it shows that indeed some countries had zero renewable energy consumption, from closer inspection these values are in the early years of the sample. There is also a large variation in the variables per capita GDP and capital, which might indicate differences in development. Also a large variation is observed for population. This confirms the need for the variables being in per capita terms, as the differences in country size differs extremely.
5. EMPIRICAL ANALYSIS
In this section potential problems relating to the estimation of Equation (7) are discussed. First, stationarity and the order of integration is determined for all variables, where the order is found to be one. Therefore, these variables have to be cointegrated to safely estimate the model. Also, the presence of heteroscedasticity and serial correlation is examined, and it is discussed how to take this into account. Thereafter, the problem of cross-sectional dependence is analysed. Finally, the plausible endogeneity problem is discussed.
3 Included countries: Australia, Austria, Belgium, Canada, Chili, Czech Republic, Denmark, Finland, France, Germany, Greece,
Hungary, Iceland, Ireland, Italy, Japan, Netherlands, New Zealand, Norway, Poland, Portugal, Republic of Korea, Spain, Sweden, United Kingdom and United States.
Excluded countries: Estonia, Israel, Latvia, Luxembourg, Mexico, Slovak Republic, Slovenia, Switzerland and Turkey.
Table 1: Descriptive statistics for the variables in levels for the period 1971-2014. Descriptive statistics:
Description Obs. Mean Minimum Maximum St. Dev.
13 5.1 Stationarity and cointegration:
As a consequence of unbalanced data, only two stationarity tests can be applied. That is the Im-Pesaran-Shin test (2003) and the Fisher-type test. Both tests allow for a heterogeneous coefficient of the lag term to test for the dependence of past values. However, the asymptotic assumption differs between the tests, because in the Im-Pesaran-Shin test N tends to go to infinity and in the Fisher-type test T tends to go to infinity. The number of countries cannot grow infinitely, whereas time is infinite. Therefore, based on this sample the Fisher-type stationarity test is applied to the variables.4 This test is constructed as a meta-analysis,
since all panels are individually tested for stationarity and their p-values are combined to get the test statistics. The null hypothesis states no stationarity, and the alternative hypothesis states that at least one panel is stationary. The results of the stationary test are shown in Table 2, also the first-differences of all variables are tested for stationarity. The results indicate that the null hypothesis cannot be rejected for all variables using any test statistic, except for the labour force and population for which two statistics indicate stationarity in at least one panel at the 10% level. The first-differences of the variables are firmly able to reject the null hypothesis, and therefore point to the presence of a stochastic trend in the variables. Hence, the variables are first-difference stationary and are said to be integrated of order one, i.e. I(1).
One of the problems related to non-stationary variables is the potential of spurious regressions, i.e. deriving significant coefficients with unrelated data. To circumvent this problem, the data is tested for cointegration, which means that the variables share a long-term relationship. If cointegration is found, then the possibility of a spurious regression no longer exist. The tests proposed by Pedroni (1999, 2004) are used to test for cointegration. These tests can be split in two categories, namely the within-dimension (homogenous cointegration parameter) and the between-dimension (heterogeneous cointegration parameter). The null hypothesis states no cointegration, whereas the alternative hypothesis states that all panels are cointegrated. The test results are shown in Table 2. The results indicate that three out of seven tests reject the null hypothesis of no cointegration at the 10% significance level. Especially, the within-dimension favours the case of cointegration. For the between-within-dimension only the Modified Phillips-Perron statistic supports cointegration slightly, as the p-value is close to the 10% level. Hence, I proceed with the presumption that a linear combination of them yields residuals which are integrated of order zero, i.e. I(0). So, these variables share a long-term relationship and a spurious regression is unlikely.
5.2 Serial correlation and heteroscedasticity:
Serial correlation and heteroscedasticity are problems occurring during the estimation and relate to the structure of the error term. If present, these problems indicate that the estimates are no longer the most efficient. The coefficients are still unbiased and consistent, but it creates inconsistent standard errors. Serial
14 correlation refers to the case where the error terms are correlated over time, i.e. 𝑐𝑜𝑣(𝑒𝑡, 𝑒𝑠)≠ 0. Heteroscedasticity implies that the variance of the error term is not the same for all observations, which often occurs with cross-sectional data.
To test for serial correlation the Wooldridge (2002) serial correlation test for linear panel data is employed to the FE estimation (see Table 4, pg. 17, for the estimation results). The FE estimation is used, because of the unavailability of test methods for the other estimators. The Wooldridge test can only detect first-order serial correlation, so it is unable to provide information of the presence of higher-order serial correlation. If the null hypothesis is rejected, then there is significant evidence to conclude the existence of first-order serial correlation. The test results are shown in Table 3, and it shows the presence of first-order serial correlation. This implies that the error term of last period is correlated with the error term of the current period.
Also, heteroscedasticity is tested for, by means of two methods. The first method uses three versions of heteroscedasticity tests of Breusch and Pagan (1979) and Cook and Weisberg (1983), and combines these results to get one test statistic. The null hypothesis for this method is constant variance. The second method tests the null hypothesis of no groupwise heteroscedasticity, i.e. each unit has the same variance structure. Moreover, this test allows for heterogeneous constants, so allowing for country fixed effects. Both tests are again based on the FE estimation of Equation (7). The test statistics of these tests can be found in Table 3. The results indicate the presence of groupwise heteroscedasticity, implying different variance structures
Table 2: Test results of the stationarity and cointegration tests. Fisher-type stationarity test:
ln 𝑦 ∆ln 𝑦 ln 𝑟𝑒𝑛 ∆ln 𝑟𝑒𝑛 ln 𝑛𝑟𝑒𝑛 ∆ ln 𝑛𝑟𝑒𝑛 ln 𝑘 ∆ln 𝑘 ln 𝑙 ∆ln 𝑙 ln 𝑝𝑜𝑝 ∆ln 𝑝𝑜𝑝 P 0.62 0.00 0.83 0.00 0.22 0.00 0.27 0.00 0.09 0.00 0.01 0.00 Z 0.75 0.00 0.99 0.00 0.36 0.00 0.25 0.00 0.21 0.00 0.99 0.00 L* 0.71 0.00 1.00 0.00 0.38 0.00 0.33 0.00 0.16 0.00 0.92 0.00 Pm 0.64 0.00 0.83 0.00 0.23 0.00 0.28 0.00 0.08 0.00 0.00 0.00
The four test statistics are: inverse chi-squared (P), inverse normal (Z), inverse logit t (L*), modified inverse chi-squared (Pm). The p-values are shown in this table, so significance can be directly observed. Null hypothesis: all panels are non-stationary, alternative hypothesis: at least one panel is stationary. For all variables one lag is used and only panel means are included. The augmented dickey fuller test is applied to all panels individually, rather than the Phillips-Perron test.
Pedroni cointegration test:
Within dimensions: p-value Between dimension: p-value
Modified variance ratio 0.019 Modified Phillips-Perron t 0.118
Modified Phills-Perron t 0.053 Phillips-Perron t 0.468
Phillips-Perron t 0.206 Augmented Dickey-Fuller t 0.314
Augmented Dickey-Fuller t 0.094
15 among panels. However, there is less evidence for heteroscedasticity among individual observations, as the null hypothesis is not rejected at the 5% significance level.
The common approach to address serial correlation is including lags of the dependent variable. However, by solving one issue it creates another, namely the Nickell bias. Therefore, the solution has to be found elsewhere. Robust standard errors can be used to make the standard errors efficient in the presence of serial correlation and heteroscedasticity, and this is relatively simple to apply. However, this does not make the estimates more efficient, because it does not solve the presence of these problems. Therefore, the GLS estimator is also estimated, which takes into account first-order serial correlation and heteroscedasticity.
5.3 Cross-sectional dependence:
Most of the investigated countries are European countries located close to each other. This raises the concern for (spatial) spillovers or common factors, such as cross-border investment or European Union regulation. Therefore, there might be some degree of cross-sectional dependence. If present and not addressed, FE could yield inefficient estimates.
The Pesaran (2004) test is used to check the existence of cross-sectional dependence and is also employed to the FE estimation. This test has the advantage that it can handle unbalanced panels and is robust in the presence of unit roots. Under the null hypothesis there is cross-sectional independence. The alternative hypothesis states the presence of cross-sectional dependence. The test result is shown in Table 3, and indicates the presence of cross-sectional dependence. Hence, the error terms are also correlated across units. The only estimator, which is able to take this into account is the CCEMG estimator. However, it does not take into account bi-directional causality and this estimator finds counterintuitive elasticities (e.g. the insignificant elasticity of capital, see Table 4).
Table 3: Test results of the serial correlation, the heteroscedasticity and the cross-sectional dependence tests based on the FE estimation.
Serial correlation Heteroscedasticity
Wooldridge test: Breusch-Pagan & Cook-Weisberg test:
F-statistic 462.54 𝜒2(1) 3.67
p-value 0.00 p-value 0.06
Cross-sectional dependence
Pesaran test: Groupwise heteroscedasticity:
Test statistic 14.60 𝜒2(26) 2581.75
p-value 0.00 p-value 0.00
16 5.4 Endogeneity:
The problem of endogeneity arises when the error term is correlated with at least one of the explanatory variables. This problem can be divided into three causes, namely omitted variable bias, measurement error and reverse causality. Measurement error is likely for this paper, since the variable capital is constructed using a perpetual inventory equation (Footnote 2, pg. 11). Omitted variable bias causes the coefficients to be biased and inconsistent, since the coefficients take over the effect of the omitted variable in explaining the variance of the dependent variable. A potential omitted variable could be human capital. However, including fixed effects controls for unobserved heterogeneity and reduces substantially the omitted variable bias. The last case is reverse causality, which is also likely to be present. In Section 2 the direction of causality is discussed for the energy variables, where most literature found bi-directional causality. This indicates that reverse causality is present, since the explanatory variables are also affected by the dependent variable. Also, capital and labour are expected to be endogenous, since a stronger economy attracts more labour and stimulates capital growth. Hence, there is strong suspicion for the presence of endogeneity.
Testing for and handling the endogeneity problem is difficult, mainly because instruments are hard to find. The instruments should be relevant, i.e. the instruments are able to explain the endogenous variables, and they should be valid, i.e. the instruments should be exogenous. This paper fails to meet the condition of valid instruments, since no specific instruments are found and the Nickell bias causes the lags to be invalid. Therefore, I cannot test for the presence of endogeneity, but it is highly likely based on literature.
6. RESULTS
Here, the research question is answered stepwise. First, Equation (7) is estimated, which will provide us the elasticities of output with respect to the variables. Furthermore, it allows us to evaluate the first two hypothesis. So, it is tested whether increases in non-renewable energy consumption positively impacts economic growth (hypothesis 1). Moreover, this is also investigated for renewable energy consumption (hypothesis 2). Thereafter, the estimated elasticities are used to calculate the marginal effects of renewable and non-renewable energy consumption, see Equation (8) and (9). Based on these marginal effects and Equation (10) hypothesis 3, i.e. the replacement of non-renewable by renewable energy consumption negatively impacts economic growth, can be analysed. Moreover, the research question can be answered after analysing hypothesis 3. Hence, these one-tailed hypotheses are examined:
1) 𝐻0: 𝛿 ≤ 0, 𝐻1: 𝛿 > 0, hypothesis 1
2) 𝐻0: 𝛾 ≤ 0, 𝐻1: 𝛾 > 0, hypothesis 2
17 The FE and FMOLS estimator use t-tests to check the significance of the coefficients, whereas the GLS and CCEMG estimator use the z-statistic. Fortunately, the critical values of both statistics are asymptotically equivalent, whenever the sample size is large. This is the case, so the corresponding critical values for both statistics are 1.282 for the 10% significance level, 1.645 for the 5% level and 2.326 for the 1% level. These values are used to test for the significance of hypothesis 1 and 2. In order to test hypothesis 3 the variance of the net effect is manually calculated, where the variance of the marginal effects is an important element.
(11) 𝑣𝑎𝑟[𝑀𝐸𝑟𝑒𝑛] = 𝑣𝑎𝑟[𝛾(𝑦̅/𝑟𝑒𝑛̅̅̅̅̅)] = (𝑦̅/𝑟𝑒𝑛̅̅̅̅̅)2∙ 𝑣𝑎𝑟(𝛾).
(12) 𝑣𝑎𝑟[∆𝑦] = 𝑣𝑎𝑟[(𝑀𝐸𝑟𝑒𝑛− 𝑀𝐸𝑛𝑟𝑒𝑛) ∙ ∆𝑟𝑒𝑛] = (∆𝑟𝑒𝑛)2∙ (𝑣𝑎𝑟[𝑀𝐸𝑟𝑒𝑛] + 𝑣𝑎𝑟[𝑀𝐸𝑛𝑟𝑒𝑛]). Equation (11) shows the variance for the marginal effect of renewables, but the same method applies for the marginal effect of non-renewables. Equation (12) presents the variance of the net effects assuming the marginal effects of renewables and non-renewables are independent. After the variance of the net effect is determined the t-statistic can be calculated and statistical inferences can be made. Note that the change in renewables, i.e. ∆𝑟𝑒𝑛, does not affect the t-statistic, since it will cancel out. Moreover, the critical values used to test for the one-tailed test regarding hypothesis 3 are -1.282 for the 10% significance level, -1.645 for the 5% level and -2.326 for the 1% level.
The evidence provided in Table 4 support both hypothesis 1 and hypothesis 2 in general, since mainly positive and significant values are found for the elasticities 𝛿 and 𝛾. The null hypotheses, i.e. no or a negative impact on economic growth, can all be rejected at the 5% level or 1% level in favour of the alternative hypotheses, except for 𝛾 in the CCEMG estimation. Hence, increasing renewable or non-renewable energy consumption enables economic growth. This is consistent with literature of Apergis and Payne (2012), Salim et al. (2014) and Bhattacharya et al. (2017). Furthermore, the support for both hypotheses is most likely the result of the positive relationship between energy consumption in general and
Table 4: Estimation results of Equation (7), for 26 OECD countries for the period 1971-2014. Dependent variable: ln 𝑦𝑖𝑡 (1) FE (2) GLS (3) CCEMG (4) FMOLS ln 𝑟𝑒𝑛𝑖𝑡 (𝛾) 0.045** [2.70] 0.012*** (2.90) 0.014 (1.07) 0.04*** [5.04] ln 𝑛𝑟𝑒𝑛𝑖𝑡 (𝛿) 0.104** [1.72] 0.243*** (13.60) 0.306*** (4.86) 0.32*** [26.78] ln 𝑘𝑖𝑡 (𝛼) 0.713*** [10.46] 0.541*** (20.11) -0.002 (-0.01) 0.42*** [30.35] ln 𝑙𝑖𝑡 (𝛽) 0.157 [0.63] 0.413*** (7.49) 0.310*** (2.90) 0.76*** [22.84] ln 𝑝𝑜𝑝𝑖𝑡 (𝜌) 0.654*** [4.21] 0.046*** (4.32) 0.709** (1.71) 1.40*** [31.06] Obs. 950 950 950 950
18 economic growth. The elasticity of non-renewable energy consumption in the FMOLS estimation, which is considered the most precise estimator, is 0.32. This implies that a one percent change in non-renewable energy consumption increases GDP per capita with 0.32%, which can be interpreted as economic growth. The elasticity of renewable energy consumption is 0.04. Hence, a one percent increase in renewable energy consumption enables 0.04% economic growth. Furthermore, the elasticities of capital (0.42) and labour (0.76) are reasonable, as they show the expected sign and show similarities with the results found by Salim et al. (2014). In addition, evidence is found for increasing returns to scale, as the sum of 𝛼 + 𝛽 + 𝛾 + 𝛿 is significantly different from one at the 1% level for the FMOLS estimation in Table 4. Moreover, it can be concluded that the variable population has more explaining power than desired, since the elasticity 𝜌 equals 1.40. However, if this elasticity would have been restricted it should equal 0.54 (𝛼 + 𝛽 + 𝛾 + 𝛿 − 1). It can be rejected at the 1% level that these two values are equal. Therefore, the elasticity of output with respect to population controls not only for varying returns to scale, but it also has some explaining power.
The marginal effects can be calculated using Equation (8) and (9) and the estimated elasticities. The mean values of per capita GDP, renewable and non-renewable energy consumption can be found in Table 1. These mean values are used instead of the mean values of the 950 observations used to calculate the elasticities. This is because the mean values in Table 1 provide the best reflection of the true mean values. The calculated marginal effects can be found in Table 5. Based on these results hypothesis 3 can be answered by subtracting the marginal effect of non-renewables from the marginal effect of renewables, see Equation (10), and using Equation (12) to find the corresponding variance. Finding a significant and negative value would support the hypothesis, whereas an insignificant value shows that the energy transition is not or positively associated with economic growth. The values of the net effect can be found in the third row of Table 5, and the corresponding test statistic in fourth row. The FMOLS estimation finds a positive value for the net effect, thus it does not support hypothesis 3. This value is even significantly larger than zero, if a one-tailed test was executed with the alternative hypothesis being positive, i.e. 𝐻1: ∆𝑦 > 0, which is the opposite of hypothesis 3. Therefore, the results are in line with literature of Roelofsen et al. (2016), who indicate that the benefits outweigh the costs. Contrastingly, the findings of the GLS and CCEMG are more consistent with the arguments of Owen (2006) and Marques and Fuinhas (2012), since they indicate a negative net effect. However, these less precise estimators fail to conclude that the impact of the energy transition on economic growth is significantly smaller than zero.
19 the energy transition on economic growth is positive, but is only small to moderate of magnitude. Therefore, hypothesis 3 is not supported by the estimates provided here. This implies that renewable energy projects are beneficial, at least at the aggregate level. Therefore, the government should actively support the transition as the benefits outweigh the costs. A possible reason a positive impact is observed might be the results of complementary policies both enhancing economic growth and environmental protection, which is also advocated by Fang (2011). Furthermore, these findings argue for a faster implementation of the transition than the current pace.
However, there is a large concern regarding the robustness of the previous findings, namely the elasticities and the marginal effects are expected to be constant. As changes in renewable and/or non-renewable energy consumption occur, the elasticities are also expected to change. Furthermore, the energy transition causes changes in the mean values of renewables, non-renewables and GDP, since the transition does not occur overnight. Therefore, the marginal effects are not expected to be constant throughout the energy transition.
This can be taken into account by assuming that the mean of total energy consumption per capita is constant. Consequently, different marginal effects can be obtained for each point in the distribution of renewables and non-renewables, given a fixed value of total energy consumption. This might also indicate the existence of a turning point, which implies that the energy transition goes from profitable to unprofitable at a certain share of renewables in total. By creating a system of equations the optimal share of renewables can be calculated. At the optimum the marginal effects of renewables and non-renewables equalize, i.e. the costs of replacing additional non-renewable consumption equals the benefits of the additional renewable energy consumption. The optimal share of renewable energy in total energy consumption equals:
(13) (𝑟𝑒𝑛/𝑡𝑜𝑡𝑎𝑙)∗ = (𝑟𝑒𝑛/(𝑟𝑒𝑛 + 𝑛𝑟𝑒𝑛))∗ = 𝛾/(𝛾 + 𝛿),
Table 5: Marginal effects of renewable and non-renewable energy consumption, and the impact of the energy transition on economic growth, i.e. the change in GDP per capita.
(1) FE (2) GLS (3) CCEMG (4) FMOLS 𝑀𝐸𝑟𝑒𝑛 6566.72 1789.37 2125.96 5837.09 𝑀𝐸𝑛𝑟𝑒𝑛 1057.04 2464.73 3107.28 3252.42 ∆𝑦| ∆𝑟𝑒𝑛 = −∆𝑛𝑟𝑒𝑛 = 1 5509.68 -675.36 -981.33 2584.67 t-statistic [2.19] [-1.05] [-0.47] [2.22] (𝑟𝑒𝑛/𝑡𝑜𝑡𝑎𝑙)* 30.20% 4.71% 4.38% 11.11%
20 where 𝑡𝑜𝑡𝑎𝑙 is the mean value of total energy consumption per capita. The derivation of Equation (13) is shown in Appendix B. There it is also stated that the mean of GDP is allowed to vary, as it cancels out on both sides. The only assumptions made are constant elasticities and constant total energy consumption. Moreover, the research question can be answered more precisely now. Notice that above the optimal share the energy transition reduces economic growth, whereas below the optimal share it increases economic growth. The results of the optimal share can be found in Table 5. The FMOLS indicates that the optimal share is 11.11%, but the optimal shares among the estimators are disperse, ranging from 4.38% to 30.20%. However, it can be concluded that the further away from the optimal share, the larger the impact of the energy transition is on economic growth. Hence, this allows for a more precise answer to the research question, since it can be concluded that the extent to which the energy transition impacts economic growth is a function of the proximity to the optimal share. This also implies that policymakers should at least strive for the optimal share, such that both economic growth and environmental protection can increase. After the turning point further stimulation of environmental protection is associated with declining economic growth. Therefore, policymakers face a difficult trade-off, where the outcome will largely depend on the willingness to incur losses, the size of the loss and the social value of the environment which is a positive externality.
21 7. TRENDS AND ROBUSTNESS CHECKS
In this section the variability of the results is investigated. Firstly, it is evaluated if the United Nations Framework Convention on Climate Change (UNFCCC) has changed the results, which might be the result of increased environmental awareness. This can also be seen as a subsample which investigates the results over time. Thereafter, the degree of development is used to see whether the results change with income. Also, the existent share of renewables in total energy consumption is used to divide the dataset. Lastly, residential energy consumption is excluded, since this does not directly contribute to the production process and GDP. If a net effect is found to be positive and significant, then the opposite of hypothesis 3 is examined, i.e.
𝐻
0:
∆𝑦≤ 0
and𝐻
1:
∆𝑦> 0.
7.1 United Nations Framework Convention on Climate Change:
The results presented in Section 6 might be sensitive to time for multiple reasons. The most important reasons are technological progress and environmental awareness. From 1971 towards 2014 the world has been subject to high amount of technological change, upon which renewable technologies are no exception. This immediately brings forward the problem of the assumption of Hicks-neutral technological progress being unrealistic. Moreover, there has been an enormous increase in environmental awareness by the public over the investigated period. Hence, the expectation is that renewable energy is increasingly profitable over time, compared to non-renewable energy. To test for this the sample is divided into two: a sample from 1971-1993 and a sample from 1994-2014. This cutting point is chosen, since the UNFCCC treaty came into force in the year 1994. Although it is drafted and signed in 1992, it only became effective in 1994. This treaty is signed by all countries in the sample, except for Chili and the Republic of Korea. Furthermore, environmental awareness and responsibility increased with this treaty, which leads to even more investments in renewable technology. Also, the Kyoto Protocol (1997) and the Paris Agreement (2015) are based and extended upon the UNFCCC. Therefore, it is believed that this is the right cutting point in the sample.
22 insignificant in the late sample, so this implies no impact of the energy transition on economic growth. The FE estimation shows an insignificant value in the early sample, but shows a positive and significant value in the late sample. Hence, this indicates that the profitability of renewable energy consumption has increased over time as expected, which is most likely the result of technological progress and environmental awareness. This statement is also supported by the increasing optimal share over time.
The elasticities of the FMOLS estimator are the same in both periods. This is because this estimator builds on the same long-term relationship in both sample periods. However, the net effect is different, where in both periods the net effect is significant and positive. It can be observed that the net effect increases over time, similar to the conclusion of the other estimators.
7.2 Different stages of development:
Although, the sample consist of developed countries these can be further subdivided. This is done according to their stage of development, where GDP per capita is used as proxy for development. The countries can be categorized as either a low, middle or high income country. Interesting to evaluate is the effect of the energy transition along the stages of development. To rank the countries in the sample the average of GDP per capita (in 2011 international dollars) from 1990 up to and including 2012 is used. This has been chosen as all 26 countries have complete data during this period, which makes it possible to compare these averages.
Table 6: The main results of the subsamples 1971-1993 and 1994-2014, where the cutting point is based upon the UNFCCC treaty. Dependent variable: ln 𝑦𝑖𝑡.
(1) FE (2) GLS (3) CCEMG (4) FMOLS 1971-1993 ln 𝑟𝑒𝑛𝑖𝑡 (𝛾) 0.012 [1.00] 0.003 (0.67) -0.019 (-1.17) 0.04*** [5.04] ln 𝑛𝑟𝑒𝑛𝑖𝑡 (𝛿) 0.148** [1.68] 0.201*** (11.04) 0.311*** (6.39) 0.32*** [26.78] Obs. 412 412 382 412 ∆𝑦 245.11 ∙ ∆𝑟𝑒𝑛 -1285.46 ∙ ∆𝑟𝑒𝑛*** -4969.95 ∙ ∆𝑟𝑒𝑛*** 2153.35 ∙ ∆𝑟𝑒𝑛 t-statistic [0.15] [-2.83] [-2.42] [3.31] (𝑟𝑒𝑛/𝑡𝑜𝑡𝑎𝑙)* 7.50% 1.47% 0.00% 11.11% 1994-2014 ln 𝑟𝑒𝑛𝑖𝑡 (𝛾) 0.061*** [2.44] 0.015** (2.15) -0.003 (-0.22) 0.04*** [5.04] ln 𝑛𝑟𝑒𝑛𝑖𝑡 (𝛿) 0.205** [2.00] 0.153*** (8.82) 0.137*** (3.47) 0.32*** [26.78] Obs. 538 538 538 538 ∆𝑦 7979.51 ∙ ∆𝑟𝑒𝑛 716.26 ∙ ∆𝑟𝑒𝑛 -2079.60 ∙ ∆𝑟𝑒𝑛 3067.23 ∙ ∆𝑟𝑒𝑛 t-statistic [1.81] [0.61] [-0.90] [3.37] (𝑟𝑒𝑛/𝑡𝑜𝑡𝑎𝑙)* 22.93% 8.93% 0.00% 11.11%
The means used here can be found in Appendix C.The numbers between brackets are t-statistics and between parentheses are z-statistics. *** Significant at the 1% level, ** at the 5% level and * at the 10% level. Robust standard errors are used in the FE estimation. The row ‘t-statistic’ applies to hypothesis 3, with the alternative hypothesis stating a negative
23 Furthermore, it is believed that the average performs better than any arbitrary chosen year, since this partly excludes the effect of outliers. Both the low and middle income category contain nine countries and the high income category contains eight countries.
The results can be found in Table 7. In general both hypothesis 1 and 2 are supported by the income categories low and high, whereas hardly no evidence is found for hypothesis 2 for the category middle income. Hypothesis 3 is again not supported for the FMOLS estimation in any category, since positive net effects are found. However, it can be observed that only in high income countries the net effect is significantly larger than zero. Furthermore, the high income countries show higher optimal shares than the low and middle income countries, for all estimations. This suggest that renewable energy consumption can be easier and more widely adopted in high income countries compared to the other countries. Hence, the energy transition will be most successful in high income countries. Furthermore, comparison between the low and middle income countries is inconclusive, since both show insignificant net effects. However, the absolute value is higher for low income countries, but this might be due to the fact that low income countries
Table 7: The main results of the subsamples categorized by income. Dependent variable: ln 𝑦𝑖𝑡.
(1) FE (2) GLS (3) CCEMG (4) FMOLS Low income ln 𝑟𝑒𝑛𝑖𝑡 (𝛾) 0.041** [1.98] 0.016*** (2.60) 0.008 (0.21) 0.04** [1.87] ln 𝑛𝑟𝑒𝑛𝑖𝑡 (𝛿) 0.180*** [2.69] 0.350*** (11.68) 0.422*** (5.06) 0.44*** [19.75] Obs. 298 298 298 298 ∆𝑦 5097.66 ∙ ∆𝑟𝑒𝑛 -1166.49 ∙ ∆𝑟𝑒𝑛 -3351.21 ∙ ∆𝑟𝑒𝑛 2090.60 ∙ ∆𝑟𝑒𝑛 t-statistic [1.39] [-1.06] [-0.53] [0.56] (𝑟𝑒𝑛/𝑡𝑜𝑡𝑎𝑙)* 18.55% 4.37% 1.86% 8.33% Middle income ln 𝑟𝑒𝑛𝑖𝑡 (𝛾) 0.066*** [3.40] 0.007 (0.99) 0.007 (0.47) 0.03 [-0.79] ln 𝑛𝑟𝑒𝑛𝑖𝑡 (𝛿) 0.393*** [4.04] 0.271*** (9.37) 0.077** (2.24) 0.34*** [18.05] Obs. 345 345 345 345 ∆𝑦 3270.10 ∙ ∆𝑟𝑒𝑛 -1770.31 ∙ ∆𝑟𝑒𝑛*** 15.18 ∙ ∆𝑟𝑒𝑛 22.43 ∙ ∆𝑟𝑒𝑛 t-statistic [1.49] [-2.39] [0.01] [0.01] (𝑟𝑒𝑛/𝑡𝑜𝑡𝑎𝑙)* 14.38% 2.52% 8.33% 8.11% High income ln 𝑟𝑒𝑛𝑖𝑡 (𝛾) 0.064*** [2.95] 0.012* (1.38) 0.026* (1.52) 0.05*** [7.93] ln 𝑛𝑟𝑒𝑛𝑖𝑡 (𝛿) 0.253* [1.46] 0.198*** (6.26) 0.199 (1.20) 0.16*** [8.18] Obs. 307 307 307 307 ∆𝑦 10341.95 ∙ ∆𝑟𝑒𝑛 263.01 ∙ ∆𝑟𝑒𝑛 3180.31 ∙ ∆𝑟𝑒𝑛 8424.67 ∙ ∆𝑟𝑒𝑛 t-statistic [2.16] [0.15] [0.81] [6.50] (𝑟𝑒𝑛/𝑡𝑜𝑡𝑎𝑙)* 20.19% 5.71% 11.56% 23.08%
The means used here can be found in Appendix C.The numbers between brackets are t-statistics and between parentheses are z-statistics. *** Significant at the 1% level, ** at the 5% level and * at the 10% level. Robust standard errors are used in the FE estimation. The row ‘t-statistic’ applies to hypothesis 3, with the alternative hypothesis stating a negative
24 are further away from their optimal shares than the middle income countries. Recall from Section 6 that the proximity to the optimal share determines the extent to which the energy transition is beneficial.
These results are based upon relatively small differences between the categories, since all countries are developed. Therefore, including developing countries as well, as is done by Narayan and Doytch (2017), can yield more convincing evidence with respect to this division by development. Also, note that the sign of the t-statistic of the elasticity 𝛾 in the FMOLS estimation for middle income countries does not properly match with the sign of the elasticity. This is most likely the result of the method used to estimate the elasticity and the t-statistic.
7.3 Existent integration of renewables:
The division by the existent degree of integration of renewables in the economy can yield powerful insights concerning the profitability along the energy transition. The share of renewable in total energy consumption is used to create three different categories, namely countries with a low, moderate or high share. The countries are ranked in a similar way as done by GDP per capita. However, the average is taken of the period 1971-2014, since data on energy consumption is fully available. Again, the categories low and moderate share include nine countries, whereas high share includes eight countries.
Based on the results for the FMOLS presented in Table 8 these subsamples do not always support hypothesis 1 and 2. Furthermore, the net effect of the energy transition differs among these subsamples. Countries with an initial low share of renewables can substantially benefit from the energy transition, since the net effects is significant and positive in the FMOLS estimation. However, the net effect falls with an increasing share, and is significant and negative in the subsample including countries with a high initial share of renewables. Therefore, this category supports hypothesis 3. A potential explanation combines both GDP and the share of renewables. The moderate countries are, on average, the richest countries, which can be concluded from Appendix C. Therefore, the opportunity costs of renewables is high to invest in renewables, resulting in a net effect not significantly different from zero (Marques and Fuinhas, 2012). Countries with a high share are on average the poorest, which might indicate that the resources are not available to implement additional renewables effectively, e.g. knowledge, technologies and financing. This inefficiency might be the reason for the negative net effect. Low share countries are medium regarding GDP, this might indicate the availability of resources to switch towards renewables and lower opportunity costs than high income countries. So, this might explain the large positive sign of the net effect. However, this reasoning partly contradicts the results found in Section 7.2, since the highest income countries, not the middle income countries, showed the highest net effect.
25 inconsistent with literature and the expectation. Especially, the CCEMG estimator estimates negative values, which support the argument of this estimator being less consistent than the FMOLS. Again, a sign of the test statistic does not match with the sign of the elasticity of renewables (𝛾) in the FMOLS estimation for the moderate share.
7.4 Excluding residential energy consumption:
The energy consumption data used in the Section 4-6 includes both residential energy consumption and energy consumption used in the production process. However, residential energy consumption does not directly contribute to GDP, whereas energy used in the production process does. So, it might be the case that hypothesis 3 is not supported, because of the inclusion of residential energy consumption. Therefore, the robustness of the results are tested using data which excludes residential consumption. The remaining data consist of energy consumption in the different sectors (industry, services and agriculture), but also transport is included.
Table 8: The main results of the subsamples categorized by the share of renewables in total energy consumption. Dependent variable: ln 𝑦𝑖𝑡.
(1) FE (2) GLS (3) CCEMG (4) FMOLS Low share ln 𝑟𝑒𝑛𝑖𝑡 (𝛾) 0.044*** [2.44] 0.013** (2.18) 0.009 (0.46) 0.07*** [2.48] ln 𝑛𝑟𝑒𝑛𝑖𝑡 (𝛿) 0.072 [0.44] 0.174*** (5.44) 0.185* (1.64) 0.05*** [10.26] Obs. 280 280 280 280 ∆𝑦 35175.23 ∙ ∆𝑟𝑒𝑛 8636.14 ∙ ∆𝑟𝑒𝑛 4984.83 ∙ ∆𝑟𝑒𝑛 56395.36 ∙ ∆𝑟𝑒𝑛 t-statistic [2.36] [1.76] [0.33] [2.46] (𝑟𝑒𝑛/𝑡𝑜𝑡𝑎𝑙)* 37.93% 6.95% 4.64% 58.33% Moderate share ln 𝑟𝑒𝑛𝑖𝑡 (𝛾) 0.028 [0.51] -0.0002 (-0.02) -0.004 (-0.14) 0.03 [-2.66] ln 𝑛𝑟𝑒𝑛𝑖𝑡 (𝛿) -0.011 [-0.06] 0.199*** (6.87) 0.293*** (2.92) 0.58*** [23.22] Obs. 342 342 342 342 ∆𝑦 5710.69 ∙ ∆𝑟𝑒𝑛 -1945.04 ∙ ∆𝑟𝑒𝑛 -3597.51 ∙ ∆𝑟𝑒𝑛 409.75 ∙ ∆𝑟𝑒𝑛 t-statistic [0.52] [-0.99] [-0.63] [0.18] (𝑟𝑒𝑛/𝑡𝑜𝑡𝑎𝑙)* 100% 0.00% 0.00% 4.92% High share ln 𝑟𝑒𝑛𝑖𝑡 (𝛾) 0.063** [1.85] 0.028*** (2.90) -0.006 (-0.30) 0.02*** [9.28] ln 𝑛𝑟𝑒𝑛𝑖𝑡 (𝛿) 0.177 [1.25] 0.317*** (10.63) 0.262*** (2.85) 0.32*** [12.77] Obs. 328 328 328 328 ∆𝑦 2264.00 ∙ ∆𝑟𝑒𝑛 -1281.11 ∙ ∆𝑟𝑒𝑛** -2921.86 ∙ ∆𝑟𝑒𝑛** -1823.17 ∙ ∆𝑟𝑒𝑛*** t-statistic [0.89] [-1.90] [-1.83] [-6.58] (𝑟𝑒𝑛/𝑡𝑜𝑡𝑎𝑙)* 35.59% 8.12% 0.00% 5.88%
The means used here can be found in Appendix C.The numbers between brackets are t-statistics and between parentheses are z-statistics. *** Significant at the 1% level, ** at the 5% level and * at the 10% level. Robust standard errors are used in the FE estimation. The row ‘t-statistic’ applies to hypothesis 3, with the alternative hypothesis stating a negative
26 The result can be found in Table 9. Again, hypothesis 1 and 2 are supported in this subsample by the FMOLS estimation. However, more interesting to see is, whether there is enough evidence for hypothesis 3 in this case. Again, the results indicate that the net effect of replacing non-renewables by renewables is positive and significant in the FMOLS estimation. Therefore, there is again no evidence for hypothesis 3, based solely on the FMOLS estimation. However, the GLS and CCEMG estimations do find significant and negative net effects. Although, these estimations are argued to be less precise estimators they do support hypothesis 3 in this sample, which was not the case in the main sample. Furthermore, the optimal shares in this sample are lower than the optimal shares presented in Table 5. This indicates that the turning point of profitable to unprofitable occurs sooner in this dataset, compared to the dataset which also includes residential energy consumption. Hence, it is less beneficial for producers to increase the share of renewable energy consumption, than it is for residents. Again, this result relies heavily on the assumption of constant elasticities, which is unlikely to hold. So, it can be concluded that the results are robust against the exclusion of residential energy consumption for the FMOLS, since results are similar compared to the results in Section 6. However, the GLS and CCEMG estimator do show support for hypothesis 3 in this sample.
8. LIMITATIONS AND FUTURE RESEARCH
To obtain the results of this paper multiple assumptions have been made, which lowers the degree to which reality is reflected. Furthermore, the chosen research design involves additional limitations. However, these limitations also provide possibilities for future research, which is also discussed here.
Although, the Cobb-Douglas production function is relatively easy to estimate it entails many strong assumptions. For instance, in this paper the elasticity of substitution is an important issue. The Cobb-Douglas production function assumes the elasticity of substitution to equal one, whereas Stern (1997;2011)
Table 9: Main estimation results when excluding residential energy consumption, for 26 OECD countries for the period 1971-2014. Dependent variable: ln 𝑦𝑖𝑡.
(1) FE (2) GLS (3) CCEMG (4) FMOLS Production ln 𝑟𝑒𝑛𝑖𝑡 (𝛾) 0.044*** [3.71] 0.005** (1.75) 0.004 (0.63) 0.03*** [7.49] ln 𝑛𝑟𝑒𝑛𝑖𝑡 (𝛿) 0.173*** [3.09] 0.295*** (17.47) 0.292*** (5.23) 0.41*** [39.44] Obs. 937 937 937 937 ∆𝑦 9568.78 ∙ ∆𝑟𝑒𝑛 -2540.34 ∙ ∆𝑟𝑒𝑛*** -2802.70 ∙ ∆𝑟𝑒𝑛** 2722.19 ∙ ∆𝑟𝑒𝑛 t-statistic [2.93] [-3.44] [-1.70] [2.54] (𝑟𝑒𝑛/𝑡𝑜𝑡𝑎𝑙)* 20.28% 1.66% 1.35% 6.82%
The means used here can be found in Appendix C.The numbers between brackets are t-statistics and between parentheses are z-statistics. *** Significant at the 1% level, ** at the 5% level and * at the 10% level. Robust standard errors are used in the FE estimation. The row ‘t-statistic’ applies to hypothesis 3, with the alternative hypothesis stating a negative
