Data

15 This study uses annual data from 1990 to 2014 for 34 OECD countries (except for Latvia due to the unavailability of data). Data on GDP per capita in constant 2000 U.S. dollars is used as a proxy for measuring economic growth (RY). The energy consumption from non-renewable sources is expressed in kilograms of oil equivalent per capita (NE) while renewable energy consumption is in % of total final energy consumption (RE). Data on gross fixed capital formation (K – measured in constant 2000 U.S. dollars per capita) and on energy efficiency (EI – captured by the energy intensity level) were used as a set of control variables in the estimated models to avoid the omitted variable biais identified in a bivariate framework (Lutkepohl [1982]). Energy intensity level of primary energy is defined as the ratio between energy supply and gross domestic product measured at purchasing power parity. Energy intensity is an indication of how much energy is engaged to produce one unit of economic output. Lower ratio indicates that less energy is used to produce one unit of output. Since all variables are in natural logarithms, each estimated coefficient should be interpreted as a constant elasticity of the dependent variable with respect to the independent variable. All data have been obtained from the World Bank’s World Development Indicators. Following Apergis and Payne [2012] and Giraud and Kahraman [2014], the production function takes the following form:

RY_it = f (RE_it, NE_it, K_it, EI_it ) [1]

16 where RY_it is real GDP per capita, RE_it is renewable energy consumption, NE_it is non-renewable energy consumption per capita, K_it is gross fixed capital formation per capita, and EI_it is energy intensity level, respectively. Subscripts i refer to the country, and t to the time period. This model has its origins in the theoretical Cobb-Douglas production function where RY can be written as: RY = A K^α E^λ

17 In this function, RY is real output per capita, E and K correspond to energy consumption and gross fixed capital formation per capita. The first right-term A is called the technology parameter. α and λ are the production elasticities with respect to the capital formation and energy consumption. Overall, the model illustrates that the gross domestic product (GDP) is explained by a set of economic factors such as: the capital and the energy consumption. If the sum of production elasticities related to capital and energy consumption equals 1 (i.e., α + λ = 1), the Cobb-Douglas production function gets constant returns to scale. The log-linear form of this production function is given by:

αλ
Ln (Y_it ) = ln (A) + ln (K_it)
+ ln (E_it)
+ e_it [2]

18 Table 2 displays the country lists that will be used in the estimations in the next sections.

Table 2

Country list

OECD_34	OECD_24 + Chile, Czech Rep., Estonia, Hungary, Israel, Korea Rep., Mexico, Poland, Slovenia, Sweden
OECD_24	OECD_15
OECD_15 + Australia, Canada, Denmark, Iceland, Luxembourg, New Zeeland, Norway, Switzerland and Turkey	Austria Belgium Finland France Germany	Greece Ireland Italy Netherlands Portugal	Slovak Rep. Spain United Kingdom Japan United States

Country list

19 The table 3 shows the matrix correlation between the independent variables for OECD_34. Since there is no significant correlation between them (except for the correlation between K and NE, and respectively, NE and EI), they can be simultaneously included in the model.

Table 3

Matrix correlation (variables in logs)

Independent variables	K	RE	NE	EI
K	1
RE	0,02	1
NE	0.68	0.034	1
EI	– 0,19	0,003	0,53	1

Matrix correlation (variables in logs)

Note: i) The fixed brut capital formation variable is expressed, here, in % of GDP;
ii) Matrix correlation is qualitatively similar for OECD_24 and OECD_15.

20 The descriptive statistics on the selected variables for each country are displayed in Table 4 (see appendix). The lowest level of renewable energy consumption is given by Korea (0.4 % of total final energy consumption) followed by United Kingdom (0.6 % of total final energy use). The highest value of renewable energy use corresponds to Island (57.8 % of total final energy consumption) closely followed by Norway (56.3 % of total final energy use). Island has also the highest level on non-renewable energy use (459 kilograms of oil equivalent) while Switzerland reaches the lowest level of non-renewable energy use (42.5 kilograms of oil equivalent per capita).

Cross-section dependence and unit root tests

21 To avoid spurious results, it is necessary to test for the cross-sectional independence in the errors, for the stationarity and the cointegration of variables. I apply two tests that rely on the assumption of cross-sectional dependence in the errors: the Pesaran [2004] test and the Baltagi, Feng and Kao [2012] test. These tests help to identify the appropriate panel unit roots tests the first generation unit root tests (relying on the assumption of cross-sectional independence between countries) versus the second generation unit root tests (based on the hypothesis of cross-section dependence). The cross-sectional dependence in the errors may arise due to the presence of common shocks (e.g., the recent global financial crisis, diverse oil shocks) and other unobserved components. This hypothesis is more likely to be validated for OECD countries because they have experienced a higher economic and financial integration process during the last decade. The Pesaran [2004] test is based on the pair-wise correlation coefficients and its statistics (CD) has the following form:

CD = √
2T N − 1 N
∑i = 1 ∑j
= i + 1^ρ̂ιJ^[3]
N (N−1)

22 The Baltagi, Feng and Kao [2012] biais-corrected scaled LM provides a simple asymptotic bias correction for the scaled LM test statistic and is asymptotically standard normal. Its LM statistic has the form:

LM_BC = √
1 N−1 N ̂2 N
∑i = 1 ∑j = i + 1^{( ρ} ιJ^Tij^{− 1 ) −}
[4]
N (N
−1) 2 (T−
1)

23 As the estimation results presented in the next section validate the cross-sectional hypothesis, I apply the Pesaran [2007] CIPS test allowing for heterogeneity in the autoregressive coefficient of the Dickey-Fuller (DF) regression and for the presence of a single unobserved common factor with heterogeneous factor loadings in the data. The test statistics is constructed from the results of panel-member-specific (A) DF regressions with the cross-section averages of the dependent and independent variables (with the lagged differences to account for serial correlation). The null hypothesis is (homogenous stationarity) H0: b_i = 0 for all i (i.e., the countries) against the possibly of heterogeneous alternatives: H1: b_i < 0, i = 1, 2,..., N1 or b_i = 0 and i = N1 + 1, N1 + 2,..., N in the following cross-sectional augmented DF (CADF) regression:

ΔlnY_it = a_i + b_i *lnY_{i, t − 1} + c_i *lnY_{t − 1} + d_i *ΔlnY_t + ε_it [5]

24 where i = 1,..., N for each country of the panel and t = 1,..., T refers to the time period. Y represents the real per capita GDP. The parameters a_i capture the possibility of country-specific fixed effects and the ε_it are the estimated residuals (assumed to be independent and identically distributed with zero mean and constant variance). The Pesaran [2007] CIPS test proposes three specifications that can be gradually tested: i) models without constant and trend; ii) models with individual specific intercepts (i.e., Eq. (5)) and iii) models with incidental linear trends. The major benefit of applying this panel unit root test is its high power of exploring the cross-sectional dependence which induces strong interdependencies between countries. The results of the table 5 strongly reject the null hypothesis of no cross-sectional dependence at the 1 % level of significance for all variables, which reflects a potentially regular dynamics to the countries.

Table 5

Cross section dependence results of Pesaran (CD) and Baltagi, Feng and Kao (LM)

H0: no cross-section dependence in residuals	PANEL: VARIABLES IN LOG
Period 1990-2014	CD	p-value	Biais corrected scaled LM	p-value
OECD_34
Ln NE	25.67a	0.00	95.824a	0.00
Ln RE	40.44a	0.00	171.18a	0.00
Ln Y	110.37a	0.00	348.07a	0.00
Ln EI	81.13a	0.00	253.25a	0.00
Ln K	70.85a	0.00	217.06a	0.00

Cross section dependence results of Pesaran (CD) and Baltagi, Feng and Kao (LM)

Note: a means statistically significant at 1 % level.

25 The panel unit root results obtained by using the Pesaran [2007] CIPS test are shown in the table 6. Results indicate that variables are integrated of order one. These variables are non-stationary in levels, but stationary in first-differences at the 1 % level of significance.

Table 6

The Pesaran [2007] CIPS results

H0: series is I (1)	OECD_34 / Period: 1990-2014
	Pesaran [2007] CIPS test
Variable	Level			First difference
	Statistic p-value			Statistic p-value
Ln NE	2.333 (2)	(0.990)		– 1.875*** (2)	(0.000)
Ln RE	0.101 (0)	(0.540)		– 7.837*** (0)	(0.000)
Ln Y	0. 390 (1)	(0.652)		– 5.033*** (1)	(0.000)
Ln EI	2.577 (1)	(0.995)		– 246.673** (1)	(0.000)
Ln K	– 0.098 (2)	(0.461)		– 3.270*** (2)	(0.000)
(N, T)	(34, 25)

The Pesaran [2007] CIPS results

Notes: All panel unit roots include an intercept and a trend. But, specifications without trend are qualitatively similar. The Lag length of variables is displayed in first small parenthesis; the values in brackets are the associated probabilities; *, ** and *** indicate significance at the 1 %, 5 % and 10 % levels, respectively.

Panel cointegration analysis

26 Since the unit root test show evidence in favor of cross section dependence assumption, cross-section cointegration test has to be analyzed. In this sense, I perform the Westerlund [2007] cointegration test relying on the cross-sectional dependence hypothesis. The test explores whether an error correction model has or not an error correction term (individual group or full panel) using the following model:

ΔlnY_it = a_i + α_i (lnY_it−1 − β_i lnX_it−1 ) + ∑_j^pi₌₁ α_ij ΔlnY_i,t−j +
q−1
∑
y_ij ΔlnX_i,t−j + e_it [6]
j=0

27 where i = 1,..., N for each country of the panel and t = 1,..., T refers to the time period. Y is the real GDP and X is the vector of the explanatory variables (the (non)-renewable energy consumption, capital stock and labor force). The parameters a_i capture the possibility of country-specific fixed effects and the e_it are the estimated residuals (assumed to be independent and identically distributed with zero mean and constant variance). Imposing as null hypothesis the absence of cointegration, the test assumes the existence of an error correction for individual countries in the panel (Gt and Ga being the group-mean statistics) and/or for the panel as a whole (Pt and Pa being the panel statistics) without any common-factor restriction. The test consents to a large degree of heterogeneity, both, in the long-run cointegrating relationship and in the short-run dynamic, and for dependence within, as well as across, the cross-sectional units (as shown by Westerlund [2007]).

28 The table 7 reports the results of the Westerlund [2007] test for the model integrating the constant and when the dependent variable is the renewable energy consumption. Only two statistics assume the existence of an error correction for individual panel members at 10 % significance level (reflected by the group-mean statistics- Gt) and for the panel as a whole (shown by panel statistics- Pt) at 1 % level. Overall, the results show that variables are somewhat cointegrated. These results enable to test the long-run effect of energy consumption from both sources of energy, capital formation and energy efficiency on GDP growth. For this purpose, I apply the Dynamic Ordinary Least Squares (DOLS) for the whole sample (OECD_34) and for two sub-samples (OECD_24 and OECD_15) to identify potential differences between the three groups.

Table 7

The Westerlund [2007] cointegration test results for OECD: RE (dep. var)

Statistics with constant and trend	Value	Z-value	P-value
Gt	– 1.970*	– 1.505	0.066
Ga	– 1.481	5.995	1.000
Pt	– 10.774***	– 2.276	0.011
Pa	– 6.361	0.854	0.803

The Westerlund [2007] cointegration test results for OECD: RE (dep. var)

Note: i) *** p < 0.01, ** p < 0.05, * p < 0.10; ii) the p-values are based on the normal distribution;
ii) the average AIC selected lag length is 1.77 and the average AIC lead length is 1.

29 The DOLS estimation technique implies regressing the dependent variable on the level, leads and lags of the independent variables, and therefore, offers answers for the small sample selection bias, endogeneity and serial correlation problems (Stock and Watson [1993]). The corresponding results are shown in tables from 8 to 10 and confirm that coefficients associated with the explanatory variables remain globally and qualitatively similar with those obtained in the PMG specifications.

30 In summary, the results of this set of estimations show that macroeconomic variables included in the models have a long-run impact on the GDP (except for RE). DOLS model indicates a positive and significant effect of non-renewable energy use on real GDP (validating the growth hypothesis). The outcomes suggest that a 1 % increase in non-renewable energy consumption raises GDP by a value between 0.82 % (for OECD_24) and 0.90 % (for OECD_15). Hence, the energy from non-renewable sources seems to be an important ingredient for the economic activity, and consequently, strong energy policies will have a stimulant effect on economic growth. Conversely, the effect of renewable energy consumption on the real GDP is negative and non-significant and it does not enhance growth. This might be the result of still small investments in renewable energy technologies or of a renewable energy market at its beginnings. In the long run, capital formation has a positive effect on GDP while, for energy efficiency, we see a negative impact. Overall, the results are qualitatively similar across these three samples. Notice also that the sum of our four estimated output elasticities is lesser than 1 in the case of OECD_34 suggesting that global returns to scale with respect to both types of energy use, energy efficiency and capital formation are slightly decreasing (i.e., the output increases less than the proportional changes in production factors which involves higher marginal costs) while for the two groups (OECD_24 and OECD_15) they are increasing little.

Table 8

Long-run panel estimators (dependent variable – real GDP: lnRY)

Variables	DOLS – OECD_34	DOLS – OECD_24	DOLS – OECD_15
Ln RE	– 0.005 (0.004)	– 0.006 (0.004)	– 0.005 (0.005)
Ln NE	0.849*** (0.016)	0.824*** (0.013)	0.904*** (0.029)
Ln K	0.0553*** (0.006)	0.067*** (0.005)	0.030** (0.013)
Ln EI	– 0.902*** (0.018)	– 0.883*** (0.015)	– 0.908*** (0.030)
Nb of observations	753	531	330

Long-run panel estimators (dependent variable – real GDP: lnRY)

Note: *** p < 0.01, ** p < 0.05, * p < 0.10; Models are with linear trends for DOLS; I used pooled weighted and heterogeneous first-stage long-run coefficients; The standard errors are in the parenthesis; Akaike criterion for leads and lags; Even without adding K in the model (given the correlation), the results are qualitatively similar.

31 The results reported in the table 9 shows that a 1 % increase in real GDP raises the renewable energy consumption by a value of 1.4 % (for OECD_34), 5.48 % (for OECD_24) and 5.67 % (for OECD_15). The sign of coefficients are quasi-similar in all three models. Furthermore, the capital formation’s coefficients are qualitatively the same in the selected samples and have a negative and significant effect on renewable energy consumption maybe suggesting weak investments in renewable energy sources. These results show also an inverse relationship between energy use from renewables and from non-renewables which underlines a potential substitutability between these sources of energy.

Table 9

Long-run panel estimators (dependent variable – RE)

Variables	DOLS – OECD_34	DOLS – OECD_24	DOLS – OECD_15
Ln RY	1.395* (0.807)	5.474*** (1.854)	5.672** (2.473)
Ln NE	– 0.456 (0.815)	– 4.778*** (1.939)	– 5.364** (2.411)
Ln K	– 0.738*** (0.061)	– 0.755*** (0.217)	– 0.742*** (0.200)
Ln EI	– 0.529 (0.819)	3.290* (1.853)	3.028 (2.424)
Nb of observations	753	532	330

Long-run panel estimators (dependent variable – RE)

Note: *** p < 0.01, ** p < 0.05, * p < 0.10; Models with linear trends (except for OECD_15 – model with constant). I used pooled weighted and heterogeneous first-stage long-run coefficients; The standard errors are in the parenthesis; Akaike criterion for leads and lags. Even without adding K in the model (given the correlation with RY and NE), the results are qualitatively the same. Estimations are done with Eviews program.

32 In the same line, the results of the table 10 indicate that a 1 % rise in real GDP enhance energy consumption from non-renewables by a value between 0.99 % (for OCDE_34) and 0.96 % (for OCDE_24 and OCDE_15) confirming the conservation hypothesis. Again, the results show that there is substituability relationship between renewable and nonrenewable energy use since coefficients are negative and significant in almost all samples. Estimates also suggest that energy intensity has a positive effect on nonrenewable energy use meaning that when energy intensity grows by 1 %, non-renewable energy use goes up by about 0.96 % – 0.98 %.

Table 10

Long-run panel estimators (dependent variable – NE)

Variables	DOLS – OECD_34	DOLS – OECD_24	DOLS – OECD_15
Ln Y	0.991*** (0.009)	0.962*** (0.015)	0.968*** (0.019)
Ln RE	– 0.010*** (0.002)	– 0.018*** (0.003)	– 0.003 (0.003)
Ln K	– 0.003 (0.004)	0.016* (0.010)	0.009 (0.012)
Ln EI	0.978*** (0.007)	0.970*** (0.012)	0.958*** (0.015)
Nb of observations	763	542	330

Long-run panel estimators (dependent variable – NE)

Note: *** p < 0.01, ** p < 0.05, * p < 0.10; Models with constant only; I used pooled weighted and heterogeneous first-stage long-run coefficients; The standard errors are in the parenthesis; Akaike criterion for leads and lags; Even without adding K in the equation (given the correlation), the results are qualitatively similar.

Granger causality

33 Because the cointegration hypothesis has been validated, the next step is to search for the existence of causality between output growth and energy consumption. For this purpose, I apply the pooled mean group model (PMG) for dynamic heterogeneous panels by Pesaran, Shin and Smith [1999]. Based on the autoregressive distributed lag (ARDL) model for time periods t = 1, 2,..., 25 and groups i = 1, 2,..., 35, the first model can be written as follows:

lnY_it = μ_i + ∑_j^p₌₁ λ_ij lnY_i,t−j + ∑_j^q₌₀ y ′_ij lnX_i,t−j + ε_it [7]

ΔlnY_it = φ_i [lnY_{i, t − 1} − θ′_i lnX_{i, t − 1}] + ∑_j^p₌⁻₁¹ λ^*_ijΔlnY_i,t−j +
∑jq=−01^γ′ ij^ΔlnXi, t − j^{+ u}it^[8]

ΔlnE_it = φ_i [lnE_{i, t − 1} − θ ′_ilnX_{i, t − 1} ] + ∑_j^p₌⁻₁¹ λ ″_ijΔlnE_{i, t − j} +
∑jq=−01^{γ ″}ij^ΔlnXi, t − j^{+ u′}it^[9]

34 where Y_it is the GDP_it dependent variable, X_it is the k × 1 vector of explanatory variables for group I, μ_i denotes the fixed effects, λ_ij’s are scalar coefficients of the lagged dependent variables, y_ij’s are k × 1 coefficient vectors. I re-parameterize the Eq. (7) and obtain:

35 where u_it are independently distributed across i and t, with zero means and variances 

36 ; and

37 .

38 The equation, in which the energy consumption is the dependent variable, is computed in the same way and is given by:

39 The PMG model is used to identify the short-run and the long-run causality among considered variables. It allows intercept, slope coefficient and error variance to vary across countries, and therefore, to identify heterogeneity among countries of the panel. An alternative panel specification would be the Mean Group (MG) estimator by Pesaran and Smith [1995] that does not account for the fact that some economic conditions may be the same across countries in the long run. The efficiency gain of PMG estimator comes from the hypothesis of heterogeneous short-run dynamics and identical long-run coefficient across countries. The PMG estimator allows assessing two types of causality: a short-run causality by testing the significance of the coefficients related to the lagged differences of economic and energy variables ( and respectively, γ ′_ij) and a long-run causality associated to the speed of adjustment coefficient or the error correction term (φ_i). The former has to be negative to validate long-run equilibrium among the variables. A larger value of (φ_i) implies a greater response of the variable to the deviation from long-run equilibrium while a low value indicates that any deviation from long-run equilibrium needs much longer time to force the variables back to it. If the speed of adjustment coefficient is significant in both equations, bidirectional causality between energy use and economic growth takes place.

40 Table 11 reports the results helping to choose the best estimation method by comparing the PMG estimates with those obtained by using MG method. The test of difference in these models is performed with the well-known Hausman test. The results are shown for the entire sample (the Hausman test results being qualitatively the same for the last two groups) and show that the PMG estimators are consistent and more efficient than the MG estimators.

Table 11

Selection of the estimation model for OECD_34.

Model_OCDE_34	ΔlnY	ΔlnRE	ΔlnNE	ΔlnK	ΔlnEI	ECT	Hausman test
PMG	-	– 0.01 (0.01)	0.23*** (0.12)	0.10*** (0.02)	– 0.27*** (0.12)	– 0.39*** (0.12)	0.527 (p-value)
MG	-	– 0.01 (0.01)	– 0.13 (0.41)	0.12*** (0.02)	0.09 (0.04)	– 0.74 (0.42)

Selection of the estimation model for OECD_34.

41 Table 12 presents the panel causality estimates of the panel vector error correction models. The number of lags of the selected variables was set according to the Akaike criterion. With respect to the long-run causality relationship between variables, the error correction terms suggest that there is a bi-directional causality between (non) renewable energy consumption and GDP growth in the long run (because the sign of the error correction term is negative and highly significant in all estimations which means that the system moves towards equilibrium). In other words, the speed of adjustment coefficients indicate that when the disequilibrium does occur, adjustments comming back to equilibrium take between 2,22 years and 12,5 years (computed as the inverse of the absolute value of the error correction parameter).

42 Looking at the outcomes for the short-run, it turns out that the energy consumption from non-renewable sources Granger causes growth for all selected samples: OCDE_34, OECD_24 and OCDE_15 which validates the growth hypothesis (the impact being slightly larger for OECD_24). Furthermore, the findings suggest that the GDP growth has a positive and significant effect on the energy consumption from non-renewable sources. This fact validates the conservation hypothesis (the impact being greater for the OECD_15). Hence, the results confirm the feed-back hypothesis concerning the relationship between NE and Y for both the long-run and the short-run. These results are slightly different from those of Giraud et Kahraman [2014] which found only an unidirectional causality from energy use to growth in the short and the long-run or from those of Coers and Sanders [2013] identifying an unidirectional causality from growth to energy use. In terms of policy implications, investors and managers need to understand that the long-run elasticity of income is positive and statistically significant (in all samples) suggesting the relevance of the income in the dynamics of nonrenewable sources of energy.

43 The economic growth has also an influence on renewable sources of energy. The estimates show that the effects of energy consumption from renewable sources on economic growth are negative and not significant in the short-run. Differently, the economic growth influences negatively and significantly renewable energy consumption only in the OECD_15 country sample. This result means that a 1 % increase in the GDP growth per capita decreases the renewable energy use, on average, by a value equal to 4.56 %. Overall, “the feed-back hypothesis” is validated in the long-run for both sources of energy use. Regarding the behavior of energy consumption from renewable sources, estimates validate the conservation hypothesis only for OECD_15 sample.

44 The short and long-run causality estimates show interesting results on the relationship between renewable and non-renewable sources of energy. It could be argued that renewable energy consumption positively affects the non-renewable ones for two sub-samples chosen for testing the sensitivity of the estimates: the OECD_15 and OECD_24. For the short-run, this finding means that a 1 % increase in renewable energy use raises, on average, the non-renewable energy use by a value between 0.03 % and 0.04 %. In this case, they can be seen as complements rather than substitutes. Furthermore, the short-run impact of non-renewable energy use on renewable energy use is also positive and significant for the same sub-samples: OECD_24 and OECD_15 (the last group having a greater impact). An increase of 10 % of energy use per capita from non-renewable sources induces, on average, a rise of about 8 % or 32 % of renewable energy consumption. This result shows evidence again for a potential complementarity between nonrenewable energy and renewable energy use. It is interesting to remark that, DOLS estimations have suggested substitutability between the two types of energy sources in the long-run. A possible explanation, widely accepted in the literature, is that strong substitution relationship might be possible in the (very) long-run as the innovation expands the range of technological possibilities. But, for the medium to short run horizons, these possibilities are less plausible since renewable energy has difficulty to replace fossil fuels used as intermediate products. Kumar et al. [2014] argue that this elasticity could be explained by the inertia of existing equipment and the technical constraints imposed on the system by the energy carriers that transform primary into final energy and into specific energy services. Finally, the effects of the other explanatory variables (the capital formation and the energy efficiency) are not clear in the short-run; they are positive or negative, significant or non-significant among the three selected models. For example, it could be observed a bi-directional and positive causality between capital and growth meaning that higher growth rates tend to increase capital formation (and vice-versa) which is in line with theoretical predictions. Furthermore, energy consumption from non-renewable sources indirectly influences economic growth through its positive and significant effects on gross fixed capital formation. The results also show that energy intensity has a negative and significant impact on economic growth and on energy consumption from renewable sources, whilst its effects on non-renewable energy use are positive and significant.

45 Because gross fixed capital formation is correlated with non-renewable energy consumption, I performed another sensitivity analysis by estimating the models that do not include this variable. From the table 13, results show that the feed-back hypothesis is again validated, in the long-run, for all samples (except for OECD_15 when the dependent variable is the GDP growth). In the short-run time period, estimates indicate a negative and significant effect of GDP growth on renewable energy use as in the previous estimations (the magnitude of impact being significant and greater for OECD_24 and OECD_15). Conversely, the impact of GDP growth on nonrenewable energy consumption is positive and significant in all three OECD groups, validating once more the conservation hypothesis. Finally, the growth hypothesis is confirmed only for the sub-samples: OECD_24 and OECD_15. Overall, the outcomes confirm the feed-back hypothesis in the long-run for both sources of energy. In the short-run, this assumption is validated only for non-renewable energy consumption. Hence, economic policies should address growth and expansion of renewable energy sector concurrently by financing R & D activities on promising renewable technologies and related infrastructure to make renewable energy sources in an enhanced market position than fossil fuels, and also by promoting regional cooperation and sustainable development between countries.

46 Substitution relationship from non-renewable energy use to renewable energy use holds in the long-run, for all panel groups, whilst in the short-run, this is true only for the OECD_34. However, the positive and significant effect of renewable energy use on non-renewable energy consumption indicates also a complementarily link between these two forms of energy which is more likely to happen in the short-run because fossil fuels, often used as intermediate products, cannot be so quickly replaced by renewable sources. This small flexibility might be explained by the inertia of existing equipments or by other technical constraints of the industry system. Policy makers should, at least, in a transition phase, examine the location layout of industrial sector and consider substitution elasticity between renewable energy and fossil fuel to reduce CO2 emission and to better protect the environment.

Table 12

Panel causality results

Dep. variables

Model 1: ΔlnY

Model 2: ΔlnRE

Model 3: ΔlnNE

Model 4: ΔlnK

Model 5: ΔlnEI

Method PMG

S1

S2

S3

S1

S2

S3

S1

S2

S3

S1

S2

S3

S2

S3

Sources of causation (independent variables)

Long-run

PANEL A: Panel causality test with both RE and NE

ECT (φi)

– 0.39a (0.12)

– 0.26a (0.07)

– 0.33a (0.09)

– 0.28a (0.08)

– 0.45a (0.07)

– 0.23a (0.10)

– 0.31a (0.07)

– 0.42a (0.12)

– 0.17a (0.04)

– 0.18a (0.05)

– 0.16a (0.06)

– 0.20a (0.06)

– 0.14b (0.06)

– 0.08a (0.03)

Short-run

ΔlnY

-

-

-

– 0.70 (0.50)

– 0.57 (0.76)

– 4.56a (1.14)

0.66a (0.10)

0.48a (0.15)

0.70a (0.10)

1.89a (0.30)

1.87a (0.39)

0.65 (0.66)

– 0.01 (0.02)

– 0.74a* (0.08)

ΔlnRE

– 0.01 (0.01)

– 0.002 (0.10)

– 0.01 (0.01)

-

-

-

– 0.01 (0.01)

0.04a (0.01)

0.03a (0.01)

– 0.06c (0.03)

– 0.05 (0.04)

– 0.01 (0.04)

– 0.07a (0.01)

– 0.06a* (0.03)

ΔlnNE

0.23b (0.12)

0.33a (0.06)

0.27a (0.08)

– 0,06 (0.28)

0.78b (0.41)

3.24a (1.22)

-

-

-

0.16 (0.23)

0.17 (0.27)

1.24b (0.52)

0.68a (0.06)

0.64a (0.06)

ΔlnK

0.10a (0.02)

0.11a (0.02)

0.13a (0.03)

– 0.06 (0.12)

– 0.10 (0.19)

0.23 (0.28)

– 0.05a (0.02)

– 0.03 (0.02)

– 0.03 (0.03)

-

-

-

– 0.01 (0.02)

0.03 (0.03)

ΔlnEI

– 0.27b (0.12)

– 0.35a (0.06)

– 0.30a (0.08)

– 0.56a (0.44)

– 1.21a (0.44)

– 3.97a (1.07)

0.64a (0.07)

0.53a (0.12)

0.70a (0.07)

– 0.16 (0.27)

– 0.20 (0.30)

– 1.24b (0.57)

-

-

No. obs.

780

550

345

780

550

345

814

550

345

780

550

345

550

345

Panel causality results

Note: (i) all equations includes a constant country specific term +/- a trend; (ii) a, b, c means p < 0.01, p < 0.05, and p < 0.10; iii) the standard errors are in the parenthesis; iv) the selected ARDL (1,2,2,2,2) in almost all equations; v) lag terms are not shown; vi) S1 – OECD_34; S2-OECD_24 and S3-OECD_15; vii) a* or b*- means that the lags coefficients of the related independent variable are significant and have the same sign.

Table 13

Panel causality results – Robustness checks

Dep. variables

Model 1: Δln Y

Model 2: ΔlnRE

Model 3: ΔlnNE

Model 4: ΔlnEI

Method PMG

S1

S2

S3

S1

S2

S3

S1

S2

S3

S1

S2

S3

Sources of causation (independent variables)

Long-run

PANEL B: Panel causality test with both RE and NE

ECT (φi)

– 6.62c (3.90)

– 0.02c (0.01)

0.004 (0.01)

– 0.19a (0.04)

– 0.15a (0.04)

– 0.23a (0.07)

– 0.31a (0.07)

– 0.14a (0.06)

– 0.20a (0.06)

– 0.12b (0.03)

– 0.11b (0.05)

– 0.03c- (0.02)

Short-run

Δln Y

-

-

-

– 0.21 (0.32)

– 0.84b (0.34)

– 1.70a (0.28)

0.57a (0.09)

0.74a (0.07)

0.70a (0.08)

0.78a (0.05)

– 0.71a (0.06)

– 0.80a (0.06)

Δln RE

– 0.02b (0.01)

– 0.02c (0.01)

– 0.02c (0.01)

-

-

-

0.01 (0.01)

0.03a (0.01)

0.04a (0.01)

– 0.03b (0.01)

– 0.04a (0.01)

– 0.04a (0.01)

Δln NE

– 5.84 (3.89)

0.76a (0.04)

0.81a (0.04)

– 0.49c (0.26)

0.11 (0.26)

0.74a (0.27)

-

-

-

0.74a (0.06)

0.74a (0.05)

0.78a (0.05)

Δln EI (3.89)

5.79 (0.05)

– 0.78a (0.06)

– 0.89a (0.32)

– 0.09 (0.30)

– 1.10a (0.24)

– 1.58a (0.07)

0.64a (0.06)

0.81a (0.06)

0.78a

-

-

-

No. obs.

780

550

345

814

574

360

814

550

360

780

550

345

ARDL

(2,2,2,2)

(2,2,2,2)

(2,2,2,2)

(1,1,1,1)

(1,1,1,1)

(1,1,1,1)

(1,1,1,1)

(2,2,2,2)

(1,1,1,1)

(3,3,3,3)

(2,2,2,2)

(2,2,2,2)

Panel causality results – Robustness checks

Note: (i) all equations includes a constant country specific term +/- a trend; (ii) a, b, c means p < 0.01, p < 0.05 and p < 0.10; iii) the standard errors are in the parenthesis; iv) lag terms are not shown; v) S1-OECD_34; S2-OECD_24 and S3-OECD_15; v) the ARDL models in italic are most efficient by using 2 or 3 lags of independent variables, but results are qualitatively similar with ARDL (1,1,1,1) as in Giraud and Kahraman [2014]; vii) c^- means that p < 0.1063.