Energy Economics 34 (2012) 511–517

Contents lists available at SciVerse ScienceDirect

Energy Economics
journal homepage: www.elsevier.com/locate/eneco

Energy consumption and economic growth nexus in Portugal, Italy, Greece, Spain and
Turkey: An ARDL bounds test approach (1965–2009)☆
José Alberto Fuinhas ⁎, António Cardoso Marques
NECE, University of Beira Interior, Management and Economics Department, Estrada do Sineiro, 6200-209 Covilhã, Portugal

a r t i c l e i n f o a b s t r a c t

Article history: The paper examines the nexus between primary energy consumption and growth in Portugal, Italy, Greece,
Received 29 November 2010 Spain and Turkey (PIGST), with annual time series data, from 1965 to 2009. PIGST are southern European
Received in revised form 1 July 2011 economies which have experienced several episodes that make them of particular interest to the study
Accepted 1 October 2011
of periods of economic expansion and stagnation. An ARDL bounds test approach is a suitable technique
Available online 10 October 2011
to examine energy-growth nexus, within the context of countries with both sporadic shocks (outliers)
JEL classification:
and permanent shocks (structural breaks). Empirical results suggest bidirectional causality between energy
C22 and growth in both the long-run and short-run, supporting the feedback hypothesis. The results reveal
C23 themselves to be robust to panel framework. An energy conservation policy will reduce GDP growth,
Q43 while a saving phenomenon is observed, since one additional unit of product requires less than one unit
O52 of energy.
© 2011 Elsevier B.V. All rights reserved.
Keywords:
ARDL bounds test
Short and long-run elasticities
Energy-growth nexus
PIGST

1. Introduction and cointegration between variables and, ultimately, lead to wrong

policy measures.
Working with lengthy time series has the advantage of enabling us to We use the autoregressive distributed lag (ARDL) bounds test
identify the nature of the relationships over time, especially if we are in approach. This well tested econometric approach: (i) guarantees robust
the presence of long-term phenomena. Although energy-growth nexus results; (ii) allows handling of a diverse number of optimal lags for
has been the object of numerous empirical papers, the literature has not different variables; and (iii) is not affected by the inclusion of “one-zero”
taken advantage of series beginning in the mid-1960s. For most countries, dummy variables.
working on long time spans is advised to deal with shocks and economic Portugal, Italy, Greece, Spain and Turkey (PIGST) are countries of
regime shifts, such as political instability and international commitments. southern Europe that have experienced periods of turbulence. This
Indeed, if we do not take into account these outliers and structural breaks, turbulence results from a mix of political and economic facts. Indeed,
then cointegration relations in energy-growth nexus could be masked. these countries had suffered coups, such as “The Carnation Revolution”
This is the reason why the models should to incorporate both impulse (1974) in Portugal, “The Regime of the Colonels” (1974) in Greece,
and shift dummies. “Coup by Memorandum” (1971) and “The 12 September 1980 Turkish
In general, the literature does not take into account that economic coup d'état” in Turkey; or attempted coups, such as Italy's “The
regime shifts, such as the membership of economic blocks or monetary Golpe Borghese” (1970), Spain's “Operación Galaxia” (1978), “23-F”
union, could permanently change the nature of the energy–growth (1981), and “The October 27, 1982 coup d'état attempt”. Most of them
relationship (cointegration in the presence of a structural break), but have undergone regime shifts from an authoritarian dictatorship to
not destabilize it (lack of cointegration). Indeed, the absence of these a democracy (Portugal, Greece, Spain and Turkey). Greece (1981),
dummies could lead to the misinterpretation of causality, elasticities and Portugal and Spain (1986) become members of the European
Economic Community and are, together with Italy, founding members
of the Economic and Monetary Union (1999). These economies have
☆ Research supported by: NECE, I&D unity funded by the FCT—Portuguese Foundation experienced periods of strong growth and lengthy stagnation. Moreover,
for the Development of Science and Technology, Ministry of Science, Technology and they also share strong energy dependency, which exposed them to
Higher Education.
⁎ Corresponding author. Tel.: + 351 275 319 600; fax: + 351 275 319 601.
the oil shocks of 1973 and 1979. It is highly probable that these former
E-mail addresses: fuinhas@ubi.pt, jafuinhas@gmail.com (J.A. Fuinhas), occurrences are being manifested in the forms of either shocks or struc-
amarques@ubi.pt (A.C. Marques). tural breaks. For PIGST, the relationship between primary energy

0140-9883/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.eneco.2011.10.003
512 J.A. Fuinhas, A.C. Marques / Energy Economics 34 (2012) 511–517

consumption (hereafter energy) and economic growth (hereafter have the advantage of allowing straight interpretation of relationships
growth) has been poorly studied. Their characteristics make it interest- between variables and dummies.
ing to focus our analysis on these countries. The ARDL bounds test approach, introduced by Pesaran and Shin
We innovate by exploring the potential of dummies to control for (1999) and later extended by Pesaran et al. (2001), is extensively used
outliers and structural breaks through the use of ARDL bounds test. in the literature of energy-growth (e.g. Odhiambo, 2009; Wolde-Rufael,
With impulse dummies and shift dummies, we show how it is possible 2010). This technique allows data to be handled in a flexible way,
and desirable to work upon distinctive characteristics of data. We assess enabling it to overcome most of the shortcomings of alternative meth-
the direction of causality, calculate the magnitude of short and long-run odologies. The ARDL bounds test is robust for finite samples, even in the
elasticities, and compute the cointegration equations between variables. presence of phenomena of shocks and regime shifts. In practice, these
Moreover, we add empirical evidence for a new set of countries. phenomena could be modeled by using “one-zero” dummy variables.
As a whole, results suggest that growth is highly driven by energy Furthermore, the technique allows conclusions to be drawn about
in PIGST, since growth appears to be elastic in both the short and long cointegration among variables even with these dummies.
term. The variables are cointegrated. Empirical support was found for In the energy-growth nexus the presence of complex dynamic
the feedback hypothesis, since we obtained two cointegrating vectors: effects was expected. The interaction among the variables can be
(i) one running from primary energy consumption to growth; and different in time, and may be delayed to a lesser or greater extent.
(ii) the other running from growth to primary energy consumption. In other words, the time responsiveness of variables may be different,
The remainder of the paper is organized as follows: Section 2 which implies that the optimal lags for different variables could be
provides a brief review of energy-growth nexus; Section 3 describes the dissimilar. The ARDL bounds test also allows a diverse number of optimal
data, sets out the method and exposes the models; Section 4 analyzes lags to be handled. Moreover, it does not impose a restrictive assumption
the empirical results; Section 5 debates the robustness of the results; that all variables should have the same integration order, which is very
and Section 6 concludes. useful when the integration of variables is borderline I(0)/I(1).

2. Energy-growth nexus 3. Data, methodology and models

There is a large body of literature focusing on the nexus between The study uses annual data for Gross Domestic Product (GDP) and
energy consumption and growth. The recent papers of Odhiambo primary energy consumption, for the time span 1965 to 2009. The
(2010), Ozturk (2010), and Payne (2010) summarize that literature. sources of data are:—European Commission—Economic and Financial
Some of those studies focus on the analysis of a particular country (e.g. Affairs—Indicators—AMECO database, for GDP, at 2000 market prices;
Lee and Chang, 2007; Wolde-Rufael, 2009), while others analyze groups and—BP Statistical Review of World Energy (June 2010), for the primary
of countries (e.g. Akinlo, 2008; Chiou-Wei et al., 2008). Nevertheless, as energy consumption, in million tons oil equivalent. The data series for
pointed out by Ozturk (2010), empirical outcomes on the direction of PIGST economies has 45 years of observations (a moderate number),
causality, and short and long-run impacts, seem to depend on data, revealing idiosyncratic outliers and structural breaks. As pointed out
countries’ characteristics, and econometric methodologies. above, an adequate technique to handle these handicaps is the ARDL
The causal relationship between energy and growth has strong bounds test approach.
implications from the theoretical, practical and policy points of view. Let Y denote GDP, E denote primary energy consumption, L denote
From an empirical standpoint, there are four possible hypotheses the natural logarithm, and D denote the first difference operator.
(Ozturk, 2010): (i) growth hypothesis, states a unidirectional causality Eqs. (1) and (2) are general equations in relating LE and LY, once
running from energy to growth, implying that growth requires energy, stationarity or cointegration are verified:
and a drop off in energy will possibly restrain growth; (ii) conservation
hypothesis, specifies a unidirectional causality running from growth to LYt ¼ θ0 þ θ1 t þ θ2 LEt þ μ1t ; ð1Þ
energy, denoting that a country is not fully dependent on energy for
growth, and that energy conservation policies can be put into practice LEt ¼ φ0 þ φ1 t þ φ2 LYt þ μ2t ; ð2Þ
with few or no adverse effects on growth; (iii) feedback hypothesis,
assumes bidirectional causality between energy and growth, since where θ0 and φ0 means the intercepts, t the trends, and μ1t and μ2t are
when the economy grows, energy demand increases, and the reverse the disturbance terms assuming white noise and normal distribution.
is also true; and (iv) neutrality hypothesis, asserts that energy and If these relationships prove to be cointegrated, this assures the presence
growth are neutral with respect to each other, and implies that energy of causality and its direction. Furthermore, they provide information
conservation policies have no effect on growth. about long-run elasticities. Eqs. (1) and (2) could be converted into
In the literature on energy-growth nexus, the option for either their equivalent autoregressive distributed lag and, if variables are
bivariate or multivariate model (e.g. Lean and Smyth, 2010; Wolde- cointegrated, into an unrestricted error-correction model (UECM).
Rufael, 2010) is usual. Empirical literature that considers long periods The general UECM could be specified in its equivalent ARDL bounds
is scarce, which is possibly a consequence of a poor understanding of test Eqs. (3) and (4), as follows:
the countries' political and economic past. Their history could reveal
shocks and economic regime shifts, i.e., the data could include outliers X
m X
n

and structural breaks that must be corrected. The literature analyzing DLYt ¼ α0 þ α1 t þ α2i DLYt−i þ α3i DLEt−i þ α4 LYt−1
i¼1 i¼0
the energy-growth nexus, controlling for particular sporadic and þ α5 LEt−1 μ3t ; ð3Þ
permanent shocks by country (e.g. Zachariadis, 2007), is still scarce.
When structural breaks are found, it means that the coefficients of
where the expected signs of the parameters are: α0 ≠ 0; α1 ≠ 0;
variables have changed and this occurrence cannot be overcome by
α2i ≠ 0; α3i ≠ 0; α4 b 0; and α5 ≠ 0. The parameters α2i and α3i
the inclusion of more variables. In line with recent literature (e.g.
explain the short-run dynamic coefficients, while α4 and α5 explain
Odhiambo, 2009; Ozturk et al., 2010; Paul and Uddin, 2010; Tsani,
the long-run multipliers of the equation.
2010), we use a bivariate model. In principle, multivariate models
are only applied when it is not possible to detect causal relationships X
m X
n
in bivariate models. A bivariate model may not detect the causality, but DLEt ¼ β0 þ β1 t þ β2i DLEt−i þ β3i DLYt−i þ β4 LEt−1
if it is detected, then there is no reason to incorporate more variables i¼1 i¼0

when the aim is to assess causality. Moreover, the bivariate models þ β5 LYt−1 þ μ4t ; ð4Þ
 J.A. Fuinhas, A.C. Marques / Energy Economics 34 (2012) 511–517 513

where the expected signs of the parameters are: β0 ≠ 0; β1 ≠ 0; Table 1

β2i ≠ 0; β3i > 0; β4 b 0; and β5 ≠ 0. Summary statistics.

LY LE DLY DLE
X
m X
n
DLYt ¼ γ0 þ γ1i DLYt−i þ γ2i DLEt−i þ μ5t : ð5Þ Portugal Mean 4.309886 2.579033 0.032260 0.037623
i¼1 i¼0 Maximum 4.882336 3.229737 0.099744 0.127821
Minimum 3.435731 1.449973 − 0.052295 − 0.058876
where the expected signs of the parameters are: γ0 ≠ 0; γ1i ≠ 0; and Std. dev. 0.448707 0.523746 0.030840 0.046136
Observations 45 45 44 44
γ2i > 0.
Italy Mean 6.759128 4.973685 0.024639 0.016398
Maximum 7.161586 5.209729 0.069326 0.099476
X
m X
n
DLEt ¼ δ0 þ δ1i DLEt−i þ δ2i DLYt−i þ μ6t ; ð6Þ Minimum 6.012516 4.374629 − 0.051694 − 0.077656
Std. dev. 0.328548 0.194432 0.025107 0.035158
i¼1 i¼0
Observations 45 45 44 44
Greece Mean 4.628805 2.993518 0.032276 0.035935
with the expected signs of the parameters being: δ0 ≠ 0; δ1i ≠ 0; and
Maximum 5.222692 3.544830 0.109426 0.175715
δ2i > 0. Minimum 3.782714 1.906575 − 0.066555 − 0.131775
Eqs. (3)–(6) represent Model 1, Model 2, Model 3, and Model 4, Std. dev. 0.360095 0.456649 0.034929 0.055071
respectively. We will test the presence of short-run causal relations Observations 45 45 44 44
Spain Mean 6.051577 4.391679 0.032322 0.036233
in Model 3 and Model 4, if and only if the presence of cointegration
Maximum 6.689751 5.005978 0.085329 0.130180
in Model 1 and Model 2 is not detected. Minimum 5.230534 3.293055 − 0.037066 − 0.096934
In short, to examine the energy-growth nexus, the UECM versions Std. dev. 0.398979 0.455007 0.023754 0.040464
of the ARDL model (Bahmani-Oskooee and Nasir, 2004; Pesaran et al., Observations 45 45 44 44
2001) are used. A first approach to the optimal lag structure using Turkey Mean 4.578910 3.579486 0.042754 0.052599
Maximum 5.459814 4.623127 0.110733 0.166609
significance of parameters was made. The aptness of use of the
Minimum 3.530323 2.215937 − 0.058662 − 0.090421
ARDL model is pointed out by Pahlavani et al. (2005), when the Std. dev. 0.550657 0.738662 0.040730 0.057815
specification of a model advises a diverse number of optimal lags Observations 45 45 44 44
for different variables. We proceed by analyzing residuals for detection
of outliers and structural breaks in relationships among variables. We
use dummy variables (with values zero and one) to capture the effects tests and stationarity tests is globally inconclusive, regarding the
of outliers and structural breaks. This approach is supported by Pesaran order of integration of variables. In fact, for some countries, variables
et al. (2001), who argued that the asymptotic theory developed in the appear to be borderline I(0)/I(1), which supports the use of ARDL
ARDL bounds test approach is not affected by the inclusion of such models (see Table A2, which summarizes ADF, PP unit root tests,
“one-zero” dummy variables. This approach was recently used (Hoque and KPSS stationarity tests).
and Yusop, 2010; Zachariadis, 2007). The econometric analyses of the energy-growth nexus suggest
We follow the Hendry (1995) general-to-specific modeling ap- that Model 1 and Model 2 are the better choices. Only for Turkey is
proach in order to select a model that is as parsimonious as possible, Model 3 (see Table A3) superior to Model 1, because the cointegration
and passes the diagnostic tests: (i) Jarque–Bera normality test; from energy to growth is rejected. The results for growth (Model 1)
(ii) Breusch–Godfrey serial correlation LM test; (iii) ARCH test for are presented in Table 2, and for energy (Model 2) are presented in
heteroscedasticity; (iv) Ramsey RESET test for model specification; Table 3.
and (v) stability tests of CUSUM and CUSUM of squares. The Jarque–Bera statistic confirms the normality behavior of the
We begin by analyzing the statistical properties of the series. We estimated residuals. The Breusch–Godfrey LM test statistic rejects
then proceed to find the order of integration of series making use of the first, second and third order serial correlation for the equations.
visual inspection of series and examining the statistics ADF, PP and The ARCH test confirms that the residuals are homoscedastic in all
KPSS (see Table A2). After that, the ARDL model was estimated, equations, and the RESET test confirms the correct functional form of
selecting the optimal number of lags and dummies, taking into con- the equations. The CUSUM and CUSUM of squares tests (see Fig. A1),
sideration the significance of parameters and carefully analyzing the suggest that the parameters were stable over the sample period for
residuals. Having performed the diagnostic tests, we achieved the op- all the equations (with the exception of Model 1 for Turkey, which
timal model. Then, we made the bounds test to conclude about the slightly violates the CUSUM test). Globally, the battery of diagnostic
presence of cointegration. Finally, we calculated the short and long- tests indicates that Model 1 and Model 2 have the desired econometric
run elasticities, and estimated long-run cointegrating equations. The properties.
robustness of achievements was checked using panel data analysis. Overall, the models are parsimonious and their specifications
require few impulse or shift dummies, which limits the loss of degrees
4. Results of freedom. Results show, as the main evidence, that DLE has a positive
and contemporary effect on DLY and vice versa, for all models, regardless
Initial assessment of the relationship between E and Y suggests of the country. In the same way, in both models, the error-correction
that developments in energy and GDP patterns are strongly correlated. mechanism (ECM) reveals itself to be statistically highly significant for
The correlation coefficients between LE and LY (see Table A1) vary in all countries.
range from the minimum of 0.9471 (Italy) and the maximum of Impulse dummies appear only in the first half of the time span
0.9955 (Portugal). For DLE and DLY, the correlation varies in range analyzed and shift dummies in the second. This is in line with what
from the minimum of 0.5485 (Greece) and the maximum of 0.7782 was expected, inasmuch as the first half of the period was marked by
(Spain). These high positive correlations indicate that higher values political and economic instability. For Model 1, the coups, in the case
of growth are associated with higher values of increase in energy. of Greece and Portugal, and the instability of the early 70s (years of
This correlation is just a condition to be fulfilled in order for any causal lead), in the case of Italy, are statistically significant. For Spain, there
relationships to be present. However, correlation does not assure the are two positive outliers that correspond to years of strong economic
presence of any causal relationships or, if it exists, its nature. Table 1 growth (1968 and 1972). For Turkey, the impulse dummy of the year
shows the summary of statistics. 1966 was probably required to absorb a statistical measurement error.
The visual inspections of data suggest that variables (in natural Indeed, it is improbable for a developing economy to grow 11.1%
logarithms) are non-stationary. However, the analysis of unit root while energy contracts by 0.2%.
514 J.A. Fuinhas, A.C. Marques / Energy Economics 34 (2012) 511–517

Table 2
Estimated ARDL—model 1.

Portugal Italy Greece Spain Turkey

Constant 1.174090*** 0.175789** 0.362811*** 0.535078***

Trend 0.006288***
DLY(− 1) 0.368887*** 0.379525***
DLE 0.190105*** 0.548983*** 0.171939** 0.453999*** 0.471468***
DLE(− 1) 0.160191**
LY(− 1) − 0.443305*** − 0.128361*** − 0.115350*** − 0.032533*** − 0.253199***
LE(− 1) 0.231966*** 0.138680*** 0.071428** 0.045032*** 0.176635***

Time dummies
ID66 0.078823***
ID68 0.033873***
ID71 − 0.038743***
ID72 0.036799***
ID74 − 0.109881***
ID75 − 0.058174***
ID79 − 0.032569***
SD8198 − 0.036190***
SD8788 0.018181**
SD7909 0.025529***
SD0109 − 0.022341**
SD0208 0.045467***

Diagnostic tests
ARS 0.802107 0.758986 0.702243 0.827927 0.616530
SER 0.013855 0.012326 0.019060 0.009668 0.025222
JB [0.716120] [0.761208] [0.868334] [0.685256] [0.218066]
LM (1) [0.2847] (1) [0.6388] (1) [0.5716] (1) [0.4889] (1) [0.4016]
(2) [0.2329] (2) [0.2935] (2) [0.2173] (2) [0.1802] (2) [0.6655]
(3) [0.3551] (3) [0.1796] (3) [0.3188] (3) [0.3327] (3) [0.4007]
ARCH (1) [0.3390] (1) [0.3460] (1) [0.2130] (1) [0.7444] (1) [0.4980]
(2) [0.6820] (2) [0.5998] (2) [0.3314] (2) [0.1962] (2) [0.7621]
(3) [0.7527] (3) [0.7674] (3) [0.5247] (3) [0.2263] (3) [0.8572]
RESET [0.3918] [0.1719] [0.1094] [0.6628] [0.9320]

Notes: Diagnostic tests results are based on F-statistic. In [ ] we represent p-values. In ( ) we represent both lags for variables and order for tests. ARS means Adjusted R-squared. SER
means standard error of regression. JB means Jarque–Bera normality test. LM means Breusch–Godfrey serial correlation LM test. ARCH means ARCH test. RESET means Ramsey
RESET test. Estimated method: least squares. *** Significant at 1% level. ** Significant at 5% level.

In general, the regime shifts occurred after the later 1970s. On the Following Banerjee et al. (1998), the ECM test is employed based
one hand, the shift dummies suggest that for Portugal and Greece, the on the OLS coefficient of the respective lag one level dependent vari-
adhesion to the Economic and Monetary Union and European Economic able in the UECM. From Table 2, for Model 1, the t-statistic for the co-
Community (EEC), respectively, do not translate into energy savings. On efficient on LY(−1) exceeds the critical value at the 1% significance
the other hand, Italy and Turkey reacted successfully to second oil shock level for all countries. The significant and negative magnitude of the
and economic policy measures (macroeconomic strategy of the AKP ECM, ranging from a minimum of −0.032533 (Spain) to a maximum
government) respectively. For Spain, only a small period of adjustment of − 0.443305 (Portugal), reflects a very low to moderate speed of
was detected in the aftermath of joining the EEC. In general, for Model 2, adjustment from short-run disequilibrium towards the state of
the results are symmetrical to Model 1. long-run equilibrium (a little over two years), which reaffirms the
Table 4 shows the likelihood ratio exclusion test (LR) for variables cointegration results. The t-statistic for the coefficient on LE(−1) is
in natural logarithms lagged once. All LR are highly significant. This significant at 1% level, for Model 2, for all countries (Table 3). The
former result and the fact that all parameters reveal the right sign ECM is significant, negative and shows a magnitude ranging from a
and are highly significant (see Tables 2 and 3), constitute additional minimum of −0.058840 (Turkey) to a maximum of −0.237701
support for choosing the ARDL bounds test approach, with an error- (Italy). This reflects a very low to low speed of adjustment from
correction mechanism. short-run disequilibrium towards the state of long-run equilibrium
In order to ascertain the presence of a long-run relationship (almost two and a half years), which is in line with the presence of
among variables, bounds test were conducted (Table 5). two cointegration vectors. The detection of a cointegration relationship,
The bounds test results suggest, at the 1% significance level for in Model 1 and Model 2, shows that, with the exception of Turkey
Portugal, Italy Spain and Turkey, and the 5% significance level for (for Model 1), no estimation of Model 3 and/or Model 4 is required.
Greece, that the null hypothesis of a non-cointegrating relationship The long-run elasticities are calculated from the estimated coeffi-
is rejected for all models, with the exception of Model 1 for Turkey. cients of the respective lag one level independent variable divided by
Therefore, the long-run relationship is confirmed between both LY the coefficient of the lag one level dependent variable (LY or LE) of
and LE and between LE and LY. The former evidence implies the pres- the respective model, from Table 2 or from Table 3, and then multiplied
ence of two cointegration vectors for PIGST, with the exception of by a negative sign. The short-run elasticities are represented by the
Turkey, which only has one (from LY to LE). Cointegration implies coefficients of the respective first difference variables. When there is
that those variables move together, so they cannot diverge from more than one coefficient for a particular variable in the short-run,
each other independently. In other words, any disequilibrium they are added and their joint significance is tested using the Wald-
among variables is a short-run phenomenon. Following the assertion coefficient test. Table 6 shows elasticities.
by Tang (2003), the existence of a cointegrating relationship indicates All elasticities have the expected signs. The short-run elasticity in
that the nexus between growth and energy for PIGST was stable over Model 1 shows that energy exerts a positive impact on growth. A 1%
the sample period. increase in energy could lead to an increase, running from (a minimum
 J.A. Fuinhas, A.C. Marques / Energy Economics 34 (2012) 511–517 515

Table 3
Estimated ARDL—Model 2.

Portugal Italy Greece Spain Turkey

DLE(− 1) − 0.640310*** 0.293624**

DLY 1.096903*** 1.048848*** 0.712931*** 1.421281*** 0.876542***
DLY(− 1) − 0.649656***
LE(− 1) − 0.187388*** − 0.237701*** − 0.076864*** − 0.085458*** − 0.058840***
LY(− 1) 0.114608*** 0.177072*** 0.053247*** 0.062342*** 0.049262***

Time dummies
ID66 − 0.142843***
ID68 − 0.087163*** − 0.053128**
ID70 0.050913***
ID71 0.051177***
ID72 − 0.058641***
ID76 − 0.072304***
ID78 − 0088361***
ID79 − 0.169483*** 0.060777***
SD7909 − 0.036441***
SD8688 0.075977***
SD8997 0.045412***
SD9803 0.079435***

Diagnostic tests
ARS 0.796210 0.816205 0.577964 0.792903 0.724038
SER 0.020532 0.015073 0.035777 0.017396 0.035052
JB [0.820622] [0.950505] [0.727620] [0.516819] [0.677021]
LM (1) [0.8022] (1) [0.6774] (1) [0.1618] (1) [0.7875] (1) [0.1787]
(2) [0.9670] (2) [0.2681] (2) [0.1639] (2) [0.9613] (2) [0.4096]
(3) [0.1231] (3) [0.4545] (3) [0.2392] (3) [0.3603] (3) [0.4620]
ARCH (1) [0.1009] (1) [0.3305] (1) [0.3053] (1) [0.7206] (1) [0.8496]
(2) [0.1298] (2) [0.6089] (2) [0.2538] (2) [0.9213] (2) [0.9756]
(3) [0.1704] (3) [0.7420] (3) [0.4076] (3) [0.9139] (3) [0.9594]
RESET [0.2031] [0.4531] [0.3225] [0.1791] [0.3599]

Notes: Diagnostic tests results are based on F-statistic. In [ ] we represent p-values. In ( ) we represent both lags for variables and order for tests. ARS means Adjusted R-squared. SER
means standard error of regression. JB means Jarque–Bera normality test. LM means Breusch–Godfrey serial correlation LM test. ARCH means ARCH test. RESET means Ramsey
RESET test. Estimated method: least squares. *** Significant at 1% level. ** Significant at 5% level.

of) 0.172% for Greece to (a maximum of) 0.549% for Italy, in growth. The The estimated long-run cointegrating equations (Table 7) reveal
long-run elasticity shows that energy exerts a permanent and a positive different patterns when comparing models 1 and 2. In Model 2, the
impact on growth. A 1% increase in energy could lead to a significant equations are similar for the countries and compatible with the ener-
increase in growth running from 0.523% (Portugal) to 1.384% (Spain). gy saving growth. Turkey has the minor energy saving effect. In its
In Model 2, the short-run elasticity reveals that growth exerts a pos- turn, in Model 1, estimated equations suggest that countries realities
itive impact on energy running from 0.713% (Greece) to 1.097% (Por- are heterogeneous. As a consequence, in order to accommodate the
tugal). The long-run elasticity shows a positive and permanent effect permanent effects, growth policy design should be framed into the
running from 0.612% (Portugal) to 0.837% (Turkey). For Turkey, the specific context of each country.
short-run elasticity, in Model 3, shows that energy exerts a positive
and high significant impact on growth of 0.6218% (see Table A3).
The previously indicated presence of long-run elasticity running 5. Robustness of results
from growth to energy and long-run elasticity running from energy
to growth reveals that the so-called feedback hypothesis is present in The results from the previous section strongly suggest the existence,
PIGST. This former conclusion has strong implications in framing pol- among countries, of: (i) heterogeneity in the relationship from energy
icies. In fact, the process of making economic policy based upon the to growth; and (ii) homogeneity in the relationship from growth to
relationship between energy and growth requires deep knowledge
about both the direction of causality and its magnitude. By estimating
long-run cointegrating equations, we obtained a concise, but robust Table 5
picture about constraints faced by countries’ governments. Bounds test results.

Portugal Italy Greece Spain Turkey

Model 1
Table 4 F-statistic 9.959489*** 12.02013*** 6.519481** 9.223593*** 4.074366
Likelihood ratio exclusion test.
Model 2
Portugal Italy Greece Spain Turkey
F-statistic 39.62019*** 16.93826*** 5.655285** 8.290624*** 7.236175***
Model 1
Notes: k = 1 (number of independent variables in equation estimated). Critical values
LY(− 1) 19.65031*** 20.10959*** 9.781551*** 10.25125*** 8.400284***
were obtained from Pesaran et al. (2001). Critical values for no intercept and no
LE(− 1) 15.54366*** 13.05846*** 6.557295** 11.00363*** 8.070416***
trend for bottom and for top are, respectively, 4.81 and 6.02, for 1%; and 3.15 and
4.11, for 5%. Critical values for unrestricted intercept and no trend for bottom and for
Model 2
top are, respectively, 6.84 and 7.84, for 1%; and 4.94 and 5.73, for 5%. Critical values
LE(− 1) 50.85432*** 27.62444*** 9.486287*** 12.34565*** 13.10957***
for unrestricted intercept and unrestricted trend for bottom and for top are,
LY(− 1) 50.71281*** 27.40136*** 10.08749*** 11.70641*** 13.74920***
respectively, 8.74 and 9.63, for 1%; and 6.56 and 7.30, for 5%. *** Significant at 1%
Notes: χ2-statistic. *** Significant at 1% level. ** Significant at 5% level. level. ** Significant at 5% level.
516 J.A. Fuinhas, A.C. Marques / Energy Economics 34 (2012) 511–517

Table 6 Table 8
Short-run and long-run elasticities. Estimated panel.

Portugal Italy Greece Spain Turkey Model 1 Model 2

Model 1 Random effects Fixed effects GEE population-averaged

Short-run 0.350296*** 0.548983*** 0.171939** 0.453999***
Constant 0.0309772*** 0.1466245***
DLE
DLY 0.9162957***
Long-run 0.523265*** 1.080392*** 0.619225*** 1.384176***
DLE 0.4270593*** 0.3966262***
LE
LY(− 1) − 0.0068948 − 0.0483591*** 0.0308573***
LE(− 1) 0.0061806 0.0341443** − 0.0426624***
Model 2
Short-run 1.096903*** 1.048848*** 0.7129301*** 0.771625*** 0.876542***
Time dummy
DLY
ID79G − 0.1711943***
Long-run 0.611605*** 0.7443535*** 0.692739*** 0.729500*** 0.837219***
LY
Diagnostic tests
Note: *** Significant at 1% level. ** Significant at 5% level. Observations 220 220 220
Wald (χ2) 172.96*** 294.2***
F-test N(0,1)) 57.98***
Hausman(χ2) 9.53**
energy. In order to check the robustness of the results, we recur to a JSE 5.39***
panel data analysis for a panel of those five countries. In accordance, Notes: The Wald test has χ2 distribution and tests the null hypothesis of non-significance
we proceed by estimating, for Model 1, Random Effects (RE) and of all coefficients of explanatory variables; The F-test is normally distributed N(0,1) and
Fixed Effects (FE). For Model 2, which is a model without a constant, tests the null hypothesis of non-significance as a whole of the estimated parameters;
the Hausman test has χ2 distribution and tests that the H0 of unobservable individual
we provide a Population Average (PA) model that uses a Generalized
effects are not correlated with the explanatory variables; JSE—Joint Significant Test is a
Estimating Equation (GEE) approach. The general form of the panel to Wald test and as the null hypothesis of HO : α4 = α5 = 0 , with α4 and α5 meaning the
estimate Model 1 and Model 2 are respectively Eqs. (7) and (8): coefficient of LY(−1) and LE(−1) respectively. In GEE PA option corr(exchangeable)
was used. Asymptotic robust standard deviations are reported in brackets. *** Significant
X
k at 1% level. ** Significant at 5% level.
DLYct ¼ α0 þ αk Xkct þ dt þ ηc þ μ ct ; ð7Þ
k¼1
energy consumption on growth, given the existence of fixed country
specific effects; and (ii) the structural similitude among countries,
X
k
DLEct ¼ βk Xkct þ dt þ ηc þ μ ct ; ð8Þ when the impact of growth on energy is analyzed. Furthermore, the
k¼1 signals and magnitude of parameters confirm previous evidence
from Model 1 and Model 2. Globally, the results of panels are in line
where α0 is the constant, Xct is the vector of explanatory variables, dt is with the presence of cointegration. This evidence shows that the ARDL
the time effects, c represents each one of the countries, and t represents bounds test approach is a suitable model for studying the specific reality
the year. of each country.
For the RE model, ηc is purely random, assuming that it is uncorre-
lated with the regressors. RE estimates are inconsistent if the FE esti-
mator is appropriate. In the FE model, the countries specific effects 6. Conclusion
are assumed to be correlated with explanatory variables. In this
way, the unobserved countries heterogeneity is controlled and the Using the ARDL bounds test approach, with annual time series
bias of the estimated coefficients is eliminated. We test the adequacy data from 1965 to 2009, the study has estimated and discussed the
of random vs. fixed effects by providing the Hausman test. Table 8 nexus between energy and growth in PIGST. It was found that the
reveals the estimations results. variables are cointegrated with two cointegrating vectors: (i) one
For Model 1, both the Wald test and the F-test were provided. The running from energy to growth (with the exception of Turkey); and
non-significance, as a whole, of the parameters of explanatory vari- (2) the other running from growth to energy. The former result assures
able is strongly rejected, both for random and fixed effects estimators. the presence of bidirectional causality, i.e., supports the feedback
The Hausman test was performed under the null hypothesis that hypothesis. This hypothesis is globally reinforced by the presence
countries’ effects are random. At 5% significance, we reject that the of short-run and long-run elasticities, running from energy to
RE estimator provides consistent estimates. As a consequence, the growth and vice versa. Indeed, in PIGST, energy significantly causes
FE estimator is appropriate and assures robust estimates. For Model growth and vice versa.
2, the generalized estimating equations estimator was performed. The elasticities found suggest important economic policy implica-
We assume that errors are equicorrelated and, given the short panel tions for PIGST. Policy measures that constrain energy consumption
under analysis, cluster-robust standard errors are provided (Cameron slow down economic growth. When the economy grows, energy de-
and Triverdi, 2009). Indeed, this option is asymptotically equivalent mand also increases. However, this path is energy saving, i.e., one ad-
to the RE estimator, without constant. ditional unit of product requires less than one unit of energy.
Results from the panel analysis check the robustness of: (i) opting The ARDL bounds test approach proves to be a suitable tool to exam-
to focus on each country individually, to study the impact of primary ine the nexus between growth and energy within the context of

Table 7
Estimated long-run cointegrating equations.

Model 1 Model 2

Portugal LY ¼ 2:64849138ð0:09165095Þ þ 0:01418535ð0:00249620Þ t þ 0:52326478ð0:05643788Þ LE LE ¼ 0:61160540ð0:00637558Þ LY

Italy LY ¼ 1:36948801ð0:66400508Þ þ 1:08039154ð0:13405992Þ LE LE ¼ 0:7443535ð0:00445490Þ LY
Greece LY ¼ 3:14529869ð0:31523848Þ þ 0:61922493ð0:07723380Þ LE LE ¼ 0:69273875ð0:02194465Þ LY
Spain LY ¼ 1:38417609ð0:02582105Þ LE LE ¼ 0:72949967ð0:01171889Þ LY
Turkey LE ¼ 0:83721936ð0:02558675Þ LY

Note: Standard errors in parentheses.

 J.A. Fuinhas, A.C. Marques / Energy Economics 34 (2012) 511–517 517

countries with both sporadic and permanent shocks. The ARDL reveals Lean, H.H., Smyth, R., 2010. On the dynamics of aggregate output, electricity consumption
and exports in Malaysia: evidence from multivariate Granger causality tests. Appl.
idiosyncratic sporadic and permanent shocks, i.e., there are fixed effects Energy 87 (6), 1963–1971.
by country, which reinforces the adequate treatment of countries indi- Lee, C.-C., Chang, C.-P., 2007. The impact of energy consumption on economic growth:
vidually, namely in explaining the relationship between energy and evidence from linear and nonlinear models in Taiwan. Energy 32 (12), 2282–2294.
Odhiambo, N.M., 2009. Energy consumption and economic growth nexus in Tanzania:
growth. an ARDL bounds testing approach. Energy Policy 37 (2), 617–622.
Odhiambo, N., 2010. Energy consumption, prices and economic growth in three SSA
Acknowledgment countries: a comparative study. Energy Policy 38 (5), 2463–2469.
Ozturk, I., 2010. A literature survey on energy-growth nexus. Energy Policy 38 (1),
340–349.
We gratefully acknowledge the generous financial support of the Ozturk, I., Aslan, A., Kalyoncu, H., 2010. Energy consumption and economic growth
NECE—Research Unit in Business Science and Economics, sponsored relationship: evidence from panel data for low and middle income countries. Energy
Policy 38 (8), 4422–4428.
by the Portuguese Foundation for the Development of Science and Pahlavani, M., Wilson, E., Worthington, A., 2005. Trade-GDP nexus in Iran: an application
Technology. We thank two anonymous reviewers for useful comments. of the autoregressive distributed lag (ARDL) model. Am. J. Appl. Sci. 2 (7), 1158–1165.
Paul, B.P., Uddin, G.S., 2010. Energy and output dynamics in Bangladesh. Energy Econ.
doi:10.1016/j.eneco.2010.11.011.
Appendix A. Supplementary data Payne, J.E., 2010. Survey of the international evidence on the causal relationship between
energy consumption and growth. J. Econ. Stud. 37 (1), 53–95.
Supplementary data to this article can be found online at doi:10. Pesaran, M., Shin, Y., 1999. An autoregressive distributed lag modelling approach to
cointegration analysis. In: Strom, S. (Ed.), Econometrics and Economic Theory in
1016/j.eneco.2011.10.003. the 20th Century: The Ragnar Frisch centennial Symposium. Cambridge University
Press, Cambridge.
Pesaran, M., Shin, Y., Smith, R., 2001. Bounds testing approaches to the analysis of level
References
relationships. J. Appl. Econ. 16 (3), 289–326.
Tang, T., 2003. Japanese aggregate import demand function: reassessment from
Akinlo, A., 2008. Energy consumption and economic growth: evidence from 11
‘bound’ testing approach. Jpn. World Econ. 15 (4), 419–436.
Sub-Sahara African countries. Energy Econ. 30 (5), 2391–2400.
Tsani, S.Z., 2010. Energy consumption and economic growth: a causality analysis for
Bahmani-Oskooee, M., Nasir, A., 2004. ARDL approach to test the productivity Bias
Greece. Energy Econ. 32 (3), 582–590.
Hypothesis. Rev. Dev. Econ. 8 (3), 483–488.
Wolde-Rufael, Y., 2009. Energy consumption and economic growth: the experience of
Banerjee, A., Dolado, J., Mestre, R., 1998. Error-correction mechanism tests for cointegration
African countries revisited. Energy Econ. 31 (2), 217–224.
in a single-equation framework. J. Time Ser. Anal. 19 (3), 267–283.
Wolde-Rufael, Y., 2010. Bounds test approach to cointegration and causality between
Cameron, A., Triverdi, P., 2009. Microeconometrics using Stata. StataPress, Texas.
nuclear energy consumption and economic growth in India. Energy Policy 38 (1),
Chiou-Wei, S., Chen, C.-F., Zhu, Z., 2008. Economic growth and energy consumption
52–58.
revisited—evidence from linear and nonlinear Granger causality. Energy Econ.
Zachariadis, T., 2007. Exploring the relationship between energy use and economic
30 (6), 3063–3076.
growth with bivariate models: new evidence from G-7 countries. Energy Econ.
Hendry, D., 1995. Dynamic econometrics. Oxford University Press, Oxford.
29 (6), 1233–1253.
Hoque, M., Yusop, Z., 2010. Impacts of trade liberalisation on aggregate import in
Bangladesh: an ARDL Bounds test approach. J. Asian Econ. 21 (1), 37–52.