Energy Consumption and GDP: Causality Relationship in G-7 Countries and Emerging Markets. Ugur Soytas

Transcription

1 Energy Consumption and GDP: Causality Relationship in G-7 Countries and Emerging Markets Ugur Soytas Department of Economics and Geography, Texas Tech University, Box Lubbock, TX , U.S.A. Tel: (806) Fax: (806) Ramazan Sari Department of Economics and Geography, Texas Tech University, Box Lubbock, TX , U.S.A Tel: (806) Fax: (806)

2 ABSTRACT The causality relationship between energy consumption and income is a well-studied topic in energy economics. This paper studies the time series properties of energy consumption and GDP and reexamines the causality relationship between the two series in top-ten emerging markets -excluding China due to lack of data- and G-7 countries. We discover bi-directional causality in Argentina, causality running from GDP to energy consumption in Italy and Korea, and from energy consumption to GDP in Turkey, France, Germany, and Japan. Hence, energy conservation may harm economic growth in the last four countries. Keywords: Energy consumption, Economic Growth, cointegration, vector error correction

3 1) INTRODUCTION The causal relationship between energy consumption and income is a well-studied topic in the literature of energy economics. Kraft and Kraft [12], in their pioneering study, found unidirectional causality running from GNP to energy consumption for the United States. They utilized Sims [19] technique and used annual data for the period. However, Akarca and Long [2] pointed out that the Kraft-Kraft results are spurious by changing the time period by two years. Other studies utilizing different time periods and different techniques have either confirmed or contradicted Kraft-Kraft results (Abosedra and Baghestani [1]; Yu and Choi [22]; Cheng [3]; Hwang and Gum [9]; Erol and Yu [6]). The large number of studies in this area, unfortunately, found different results for different countries as well as for different time periods within the same country. In most recent studies, researchers have focused on the cointegrating relationship between energy consumption and income for a few countries (Yu and Jin [21]; Masih and Masih [14]; Masih and Masih [15]; Glasure and Lee [7]). This paper reexamines the causal relationship between GDP and energy consumption in the top ten emerging markets and the G-7 countries, using cointegration and vector error correction techniques. We find evidence of a causal relationship in three of the nine emerging markets and four of the seven developed countries. The causality is in the sense of Granger causality (Granger [8]).

4 The organization of the paper is as follows. In the next section we discuss the data and the methodology. In the third section we introduce the results and the last section is a brief concluding section. 2) DATA We use annual energy consumption (EC hereafter) and GDP per capita data. EC is million metric tons of coal equivalent and is sourced from various issues of United Nation s Statistical Yearbook. GDP per capita data is from Penn World Tables. For all countries the time period used is , except for Argentina ( ), Indonesia ( ), Korea ( ), and Poland ( ). Note that, we dropped China from the analysis because its energy consumption data was combined with Taiwan for a long period of time and the remaining data was not long enough. Also note that we used the natural logs of both variables. 3) METHODOLOGY Stock and Watson [20] argued that the causality tests are very sensitive to the stationarity of the series. Also, Nelson and Plosser [16] stated the fact that many macroeconomic series are nonstationary. Therefore, it is necessary to check for the stationarity of energy consumption and GDP series.

5 3.A Unit Root Tests We used Dickey-Fuller (DF), augmented Dickey-Fuller (ADF), and Phillips- Perron (PP) tests to assess the degree of integration of the two series (Dickey and Fuller, [4]; Phillips and Perron [18]). If the series are nonstationary in levels and stationary when first differenced, then they are said to be integrated of order one. We can test for cointegration between series integrated of the same order. We provide the results of the unit root tests in full, but we do not specifically discuss the tests due to space constraints. 3.B Cointegration Tests Next, we utilized Johansen [10],and Johansen and Juselius [11] (JJ hereafter) maximum likelihood procedure to test for cointegration. The same methodology also provides estimates of the cointegrating vectors. The existence of cointegration rules out Granger noncausality. The maximum likelihood methodology is based on the following VAR model: X t = µ + Π 1 X t-1 + Π 2 X t Π p X t-p + ε t (1) where X is a (n x 1) vector of variables, µ is a (n x 1) vector of constant terms, Π 1, Π 2,, Π p are (n x n) coefficient matrices and ε is a (n x 1) vector of error terms with zero mean and constant variance. The reparameterization of (1) can be written as:

6 X t = µ + Σ i=1 p-1 Γ i X t-i + Γ X t-p + e t (2) where Γ I = - I + Π Π i (i= 1, 2,, p-1), and Γ = - I + Π Π p. The rank of the matrix Γ, the matrix determining the long-run relationships between variables, is equal to the number of independent cointegrating vectors denoted by r. If r = 0, then the elements of X are nonstationary, and (2) is a usual VAR in first differences. Instead, if the rank(γ) is n and r = n, then the elements of X are stationary. Γ X t-p is error-correction factor, if r = 1. For other cases, 1 < r < n, there are multiple cointegrating vectors. The number of distinct cointegrating vectors can be obtained by checking the significance of characteristic roots of Γ. Johansen [10] and JJ [11] suggest two statistics to test for cointegration. The trace statistic is for testing the hypothesis that the cointegrating rank is at most r against a general alternative. The trace statistic can be calculated from (3). λ trace (r) = - T n i= r+ 1 ln (1- λ i ) (3) The max statistic is employed for testing the alternative hypothesis r + 1 and can be derived from (4). λ max (r,r+1) = - T ln(1- λ i+1 ) (4)

7 where λ i are characteristic roots obtained from estimated matrix Γ and T is the number of usable observations. It should be clear that if characteristic roots are close to zero, both λ trace and λ max statistics will be small. 3.C Vector Error Correction Model Cointegration implies the existence of Granger causality, however it does not point out the direction of the causality relationship. Therefore, we employed the vector error correction modeling to detect the direction of the causality. Engle and Granger [5] argued that if there is cointegration between the series, then the vector error-correction model can be written as: LGDP t = Ψ 10 + Σ i=1 5 Ψ 11i LGDP t-i + Σ i=1 5 Ψ 12i LEC t-i + γ ECT t-1 + η 1t (5) LEC t = Ψ 10 + Σ i=1 5 Ψ 11i LEC t-i + Σ i=1 5 Ψ 12i LGDP t-i + γ ECT t-1 +η 1t (6) Where the error correction term (ECT) represents the error terms derived from the long run cointegrating relationship. The error-correction representation allows for causality to emerge via two avenues. First, testing the joint significance of the coefficients (Ψ 11i ) of the independent variable we can check for short run causality. The joint significance of Ψ 11i indicates that the dependent variable is responding to shortterm shocks to the stochastic environment. Secondly, the long run causality can be tested by looking at the significance of the speed of adjustment (γ), which is the coefficient of

8 the error correction term. The significance of γ indicates that the long run equilibrium relationship is directly driving the dependent variable. In addition to the extra way for causality to emerge, the VEC offers another advantage that the lost information due to differencing is brought back into the system through the error correction term. 3.D Variance Decomposition The causality tests we discussed so far are valid only within the sample period. We utilize variance decompositions (VDCs) in order to assess the validity of causality beyond the sample period. The variance of the forecast error of a variable can be partitioned (in this case into two) with respect to the innovations in each variable in the system. For example, the variance of the forecast error in GDP can be attributable to innovations in energy consumption as well as to its own innovations. In that sense, the VDCs can be viewed as out of sample causality tests. 4) RESULTS The results of the DF, PP, and ADF unit root tests for levels and first differences are reported in Table 1A and 1B, for G-7 countries and emerging markets respectively. In all countries, LEC and LGDP appear to be I(1) variables. Among the top ten emerging markets, only in the case of Indonesia, the ADF test indicates nonexistence of a unit root in levels when we assume trend and intercept in the series. However, the test statistics are

9 on the margin and the trend term did not appear significant in the ADF regression. We make similar assumptions for LGDP in UK and W. Germany and for LEC in Italy. Therefore, we conclude that LEC and LGDP are I(1) in all sixteen countries. The next step is to test for the presence of cointegration. The results of the Johansen [10] and JJ [11] methodology are reported in Table 2A and 2B, for G-7 countries and emerging markets respectively. Given the importance of selecting the appropriate lag length, we chose the lag length according to Akaike information criterion. Note that in all equations the chosen lag length satisfies absence of serial correlation. We found evidence of a cointegrating vector for only seven out of sixteen countries: Argentina, Turkey, Korea, France, Italy, West Germany, and Japan. The estimates of the cointegrating vectors normalized to LGDP are listed in Table 3. We extracted the ECT for the VEC model from these underlying cointegrating vectors. The next step is to examine the direction of the causality. The results of the VEC model are summarized in Table 4A and 4B. There is evidence of long run uni-directional causality running from LEC to LGDP for Turkey, France, W. Germany, and Japan. The long run causality is reversed for Italy, and Korea, and there is bi-directional long run causality in Argentina. In addition, there is evidence of short run bi-directional causality in Argentina. The bi-directional relationship also holds for Turkey in the short run. In none of the other countries, LGDP and LEC enter significantly in each other s equation, implying lack of short run causality.

10 We utilized Breusch-Godfrey test to check for serial correlation in all equations and found no evidence of serial correlation. We also tested the structural stability of the equations by Ramsey Reset test and found no evidence against stability. We were not able to reject the null hypothesis of structural stability for all equations in all seven countries at 1% significance level. Therefore, we conclude that all our equations are structurally stable. In this step we examine whether the causality relationships we discovered for our sample also holds beyond our sample period. For that purpose we utilized VDCs. The results of VDCs are reported in Table 5. The VDC results are consistent with our findings in VEC model analysis. In the case of Italy, LEC does not appear to explain more than 1.55% of an innovation in LGDP even after thirty years. In Argentina s case, the portion explained by LEC goes up to 9.08% from 0% in four years and stabilizes at around 7.5% after five years. The LEC in Turkey and Germany accounts for 61-63% of the shock in LGDP in thirty years. The proportions explained by LEC increase to 90-91% in France and Japan. Thus, indicating causality running from energy consumption to income for Turkey, France, Germany, and Japan. In Argentina and Italy, LGDP accounts for 42% and 31% of the innovation in LEC after thirty years, respectively. In the case of Korea, the portion of a shock in LEC explained by LGDP fluctuates between 5% to10%. The results of VDCs may be viewed as an indication of causality running from LGDP to LEC in these countries. Hence, we conclude that, overall, VDCs confirm our findings in VEC model analysis.

11 5) CONCLUDING REMARKS In this paper we reexamined the causal relationship between GDP and energy consumption in sixteen countries. In all countries both series appear to be nonstationary in levels, but stationary in first differences. For seven countries, there exists a stationary linear cointegrating relationship between the variables. In Turkey, France, Germany, and Japan the causality runs from energy consumption to GDP. We discovered bi-directional causality in the case of Argentina. This indicates that in the long run energy conservation may harm economic growth in these countries. The causality relationship appears to be reversed for Italy and Korea. The results of VDCs support the causal relationships we discovered using the VEC model.

12 Table 1A. Unit Root Tests for G-7 Countries LGDP D(LGDP) LEC D(LEC) ADF DF PP ADF DF PP ADF DF PP ADF DF PP Trend U. S. A (1) (2)a a a (1) (1)a a a and U. K (2)c (2)b a a (1) (1)b a a constant France (1) (1)a a a (1) (1)c a a Italy (1) (1)a a a (1)b a a (1)a a a Germany (2)c c (1)a a a (1) (1)a a a Japan (1) (1)c a a (1) (1)b a a Canada (1) (1)a a a (1) (1)a a a Constant U. S. A (1) (1)a a a (1) (4) a a U. K (3) (2)a a a (1) (1)a a a France (1) a b (1)c a a (1) (1)b a a Italy (1)c b a (1)a a a (1) b c (1)a a a Germany (2)a a a (1)a a a (1) (1)a a a Japan (1) a b (2) a a (1) (2)c a a Canada (1) (1)a a a (1) (1)a a a None U. S. A (1) (1)a a a (1) (4) a a U. K (3) (2)b a a (1) (1)a a a France (1) (2) b c (1) (1)a a a Italy (1) (2) a b (2) (1)a a a Germany (1) (4)b b b (1) (1)a a a Japan (1) (1) b b (1) (2)c a a Canada (1) (4) a a (1) (1)a a a 1 a, b, c represent significance at 1%, 5%, and 10% respectively 2 Critical values are sourced from Mackinnon (1991)

13 Table 1B. Unit Root Tests for Top-Ten Emerging Markets LGDP D(LGDP) LEC D(LEC) ADF DF PP ADF DF PP ADF DF PP ADF DF PP Trend Argentina (2) (1)a a a (2) (1)a a a and Brazil (2) (1) a a (2) (1)b a a constant India (1) (1)a a a (1) (1)a a a Indonesia (1)c (2) b b (1)c (1) a a S. Korea (2) (1)c a a (1) (2)a c Mexico (3) (3)c a a (1) (2)b a a Poland (1) (1)c (1) (1)b c c S. Africa (2) (1)a a a (1) (2)a a a Turkey (1) b b (1)a a a (1) (2)b a a Constant Argentina (2) (1)a a a (2) b (1)a a a Brazil (2) (1) a a (1) (1)a a a India (1) (1)a a a (1) (1)a a a Indonesia (3) (2)c b a (2) (1)b a a S. Korea (1) (1)c a a (1) (2)a a b Mexico (4) (3)b a a (1) (2)b a a Poland (1) (1)c (1) (1)c c c S. Africa (2) (2) a a (1) (1)a a a Turkey (1) (3)a a a (1) (2)a a a None Argentina (2) (1)a a a (2) (4) a a Brazil (2) (1)b a a (2) (1)a a a India (1) (1)a a a (1) (1)b a a Indonesia (3) (2) b b (2) (1)c a a S. Korea (1) (2) b c (1) (3)c a a Mexico (4) (3)b a a (1) (2) a a Poland (1) (1)c c c (1) (1)a a a S. Africa (1) (2)b a a (1) (1)a a a Turkey (1) (1)a a a (1) (2)c a a 1 a, b, c represent significance at 1%, 5%, and 10% respectively 2 Critical values are sourced from Mackinnon (1991)

14 Table 2A. Cointegration Test Results for G-7 Countries U. S. A. U. K. France Italy W. Germany Japan Canada H 0 H 1 λ trace 1% 5% 10% r = 0 r c a b c r 1 r b a b c H 0 H 1 λ max r = 0 r > a c c r 1 r = b a b c Lag lengths are determined by Akaike information criterion 2 a, b, c represent significance at 1%, 5%, and 10% respectively 3 Critical values are sourced from Osterwald-Lenum (1992)

15 Table 2B. Cointegration Test Results for Top-Ten Emerging Markets Korea Mexico Poland S. Africa Turkey H 0 H 1 λ trace 1% 5% 10% r = 0 r a c r 1 r a H 0 H 1 λ max r = 0 r > c r 1 r = a Indonesia India Brazil Argentina H 0 H 1 λ trace 1% 5% 10% r = 0 r c r 1 r H 0 H 1 λ max r = 0 r > r 1 r = a, b, c represent significance at 1%, 5%, and 10% respectively. 2 Lag lengths are determined by Akaike information criterion 3 Critical values are sourced from Osterwald-Lenum (1992)

16 Table 3. Estimates of the Cointegrating Vectors France Italy W. Germany Japan Argentina Turkey Korea LGDP LEC ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) C Standard deviations and t-stats are in parenthesis

17 Table 4A. VEC Results for G-7 Countries Eq. ECT ΣcLGDP ΣcLEC Adj. R 2 BG RR France LGDP (+) (0.0113)b (0.4693) (0.6188) (0.9174) LEC (-) (0.3075) (0.7554) (0.7331) (0.1026) Italy LGDP (+) (0.5508) (0.1852) (0.7577) (0.0454)b LEC (+) (0.0000)a (0.1750) (0.1341) (0.8565) Germany LGDP (-) ( ) (0.0015)a (0.3444) (0.1756) (0.6122) LEC (+) (0.8865) (0.1537) (0.5577) (0.8158) Japan LGDP (+) (0.0023)a (0.4525) (0.8519) (0.6749) LEC (+) (0.5303) (0.1991) (0.2253) (0.3707) 1 ECT: speeds of adjustment (coefficients of the error correction terms), t-statistics, and in parenthesis are the probabilities. 2 ΣcLGDP, ΣcLEC: sign of sum of coefficients, the joint significance test F-stat, and in parenthesis are the probabilities. 3 BG, RR: Breusch-Godfrey serial correlation, Ramsey-Reset stability (1 fitted term) tests, the F-statistics, and in parenthesis are probabilities. 4 a, b, c represent significance at 1%, 5%, and 10% respectively.

18 Table 4B. The Vector Error Correction Results for Top-Ten Emerging Markets Eq. ECT ΣcLGDP ΣcLEC Adj. R 2 BG RR Argentina LGDP (-) ( ) ( ) (0.0492)b (0.0366)b (0.3447) (0.9109) LEC (+) ( ) ( ) (0.0079)a (0.0984)c (0.2985) (0.9117) Korea LGDP (+) ( ) ( ) (0.1626) (0.6368) (0.4699) (0.1729) LEC (-) ( ) ( ) (0.0348)b (0.6324) (0.8276) (0.0146)b Turkey LGDP (+) ( ) ( ) (0.0040)a (0.0459)b (0.8245) (0.3655) LEC (-) ( ) ( ) (0.7797)a (0.0370)b (0.1785) (0.0289)b 1 ECT: speeds of adjustment (coefficients of the error correction terms), t-statistics, and in parenthesis are the probabilities. 2 ΣcLGDP, ΣcLEC: sign of sum of coefficients, the joint significance test F-stat, and in parenthesis are the probabilities. 3 BG, RR: Breusch-Godfrey serial correlation, Ramsey-Reset stability (1 fitted term) tests, the F-statistics, and in parenthesis are probabilities. 4 a, b, c represent significance at 1%, 5%, and 10% respectively.

19 Table 5A. VDC Results for G-7 Countries France Italy Germany Japan Variance Decomposition of LGDP Periods LGDP LEC LGDP LEC LGDP LEC LGDP LEC Variance Decomposition of LEC Periods LGDP LEC LGDP LEC LGDP LEC LGDP LEC Order: LGDP LEC

20 Table 5B. VDC Results for Top-Ten Emerging Markets Korea Turkey Argentina Variance Decomposition of LGDP Periods LGDP LEC LGDP LEC LGDP LEC Variance Decomposition of LEC Periods LGDP LEC LGDP LEC LGDP LEC Order: LGDP LEC

21 REFERENCES 1. Abosedra, S., Baghestani, H., New Evidence on the Causal Relationship between U.S. Energy Consumption and Gross National Product, Journal of Energy and Development, Vol 14, No 2, Spring, 1989, pp Akarca, A.T., Long, T.V., On the Relationship between Energy and GNP: A Reexamination, Journal of Energy and Development, Vol 5, No 2, Spring, 1980, pp Cheng, B. S., An Investigation of Cointegration and Causality between Energy Consumption and Economic Growth, Journal of Energy and Development, Vol 21, No 1, Autumn, 1995, pp Dickey, D., Fuller, W., Distribution of the Estimators for Autoregressive Time Series with a Unit Root, Journal of the American Statistical Association, Vol. 74, No 366, Part 1, June, 1979, pp Engle, R.F., Granger, C.W.J, Co-integration and Error Correction: Representation, Estimation, and Testing, Econometrica, Vol 55, No 2, March, 1987, pp Erol, U., Yu, E.S.H, On the Causal Relationship between Energy and Income for Industrialized Countries, Journal of Energy and Development, Vol 13, No 1, Autumn, 1987, pp Glasure, Y. U., Lee, A.R., Cointegration, Error Correction, and the Relationship Between GDP and Energy: The Case of South Korea and Singapore, Resource and, Vol 20, No 1, March, 1997, pp Granger, C.W.J., Investigating Causal Relations by Econometric Models and Crossspectral Methods, Econometrica, Vol 37, No 3, July, 1969, pp Hwang, D., Gum, B., The Causal Relationship between Energy and GNP: The case of Taiwan, Journal of Energy and Development, Vol 16, No 2, Spring, 1991, pp Johansen, S., Statistical Analysis of Cointegrating Vectors, Journal of Economic Dynamics and Control, Vol 12, No 2/3, June-September, 1988, pp Johansen, S., Juselius, K., Maximum Likelihood Estimation and Inference on Cointegration with Applications to the Demand for Money, Oxford Bulletin of Economics and Statistics, Vol 52, No 2, May, 1990, pp Kraft, J., Kraft A., On the Relationship Between Energy and GNP, Journal of Energy and Development, Vol 3, No 2, Spring, 1978, pp

22 13. Mackinnon, J.G., Critical Values for Cointegration Tests in Long-run Economic Relationships, in R.F. Engle and C.W.J. Granger (eds.), Long-run economic relationships: Readings in cointegration. Advanced Texts in Econometrics Oxford; New York; Toronto and Melbourne: Oxford University Press, 1991, pp Masih, A.M.M., Masih, R., Energy Consumption, Real Income and Temporal Causality: Results from a multi-country study based on cointegration and error-correction modeling techniques,, Vol 18, No 3, July, 1996, pp Masih, A.M.M., Masih, R., On the Temporal Causal Relationship Between Energy Consumption, Real Income, and Prices: Some new evidence from Asian-energy dependent NICs based on a multivariate cointegration/vector error-correction approach, Journal of Policy Modeling, Vol 19, No 4, August, 1997, pp Nelson, C.R., Plosser, C.I., Trends and Random Walks in Macroeconomic Time Series, Journal of Monetary Economics, Vol 10, No 2, September, 1982, pp Osterwald-Lenum, M., A note with Quantiles of the Asymptotic Distribution of the Maximum Likelihood Cointegration Rank Test Statistics, Oxford Bulletin of Economics and Statistics, Vol 54, No 3, August, 1992, pp Phillips, P.C.B., Perron, P., Testing for Unit Root in Time Series Regression, Biometrica, Vol 75, 1988, pp Sims, C.A., Money, Income, and Causality, American Economic Review, September, 1972, Stock, J.H., Watson, M.W., Interpreting the Evidence on Money-Income Causality, Journal of Econometrics, Vol 40, No 1, January, 1989, pp Yu, S.H., Jin, Jang C., Cointegration Tests of Energy Consumption, Income, and Employment, Resources and Energy, Vol 14, No 3, September, 1992, pp Yu, S.H., Choi, J.Y., The Causal Relationship Between Energy and GNP: An International Comparison, Journal of Energy and Development, Vol 10, No 2, Spring, 1985, pp