Asymmetries in the dynamic interrelationship between energy consumption and economic growth_ Evidence from Turkey

Asymmetries in the dynamic interrelationship between energyconsumption and economic growth: Evidence from Turkey Ay şen Araça,⁎ , Mübariz Hasanovb,1 a Department of Economics, Hacettepe

Asymmetries in the dynamic interrelationship between energy

consumption and economic growth: Evidence from Turkey

Ay şen Araça,⁎ , Mübariz Hasanovb,1

a

Department of Economics, Hacettepe University, Beytepe, Ankara, Turkey

b Department of Banking and Finance, Okan University, İstanbul, Turkey

a b s t r a c t

a r t i c l e i n f o

Article history:

Received 23 September 2013

Received in revised form 14 April 2014

Accepted 17 April 2014

Available online 27 April 2014

JEL classification:

Q43

C32

C51

Keywords:

Energy

Output growth

Nonlinearity

In this study we examine possible nonlinearities in dynamic interrelationship between energy consumption and economic growth in Turkey for the 1960–2010 period by using a smooth transition vector autoregressive model

In order to trace the effects of one variable on another, we calculate Generalized Impulse Response Functions (GIRFs) The computed impulse response functions demonstrate asymmetric effects of positive versus negative and small versus large energy consumption shocks on output growth and vice versa Specifically, we find that negative energy shocks have a greater effect on output growth than positive energy shocks, and that big negative energy shocks affect output much more than small negative energy shocks Similarly, wefind that positive output shock has a greater effect on energy consumption whereas negative shocks have almost no effect on energy consumption The results of this study have clear and important implications for energy economists and policymakers in Turkey

© 2014 Elsevier B.V All rights reserved

1 Introduction

The knowledge of the dynamic interaction between energy

con-sumption and economic growth plays a crucial role in design and

imple-mentation of energy policies If, for instance, a decrease in energy

consumption hampers economic growth, then adopting of energy

conserving policies designed to reduce energy consumption will not

be desirable On the other hand, if reducing energy consumption does

not affect economic growth, energy conserving policies may be

imple-mented without deteriorating economic growth In this study we aim

to analyze the dynamic interaction between energy consumption and

economic growth in Turkey

There are different views on interrelationship between energy

con-sumption and economic growth in the energy economics literature

(see, for example,Ozturk, 2010) Proponents of the so-called“neutrality

hypothesis” argue that there is no relationship between energy use

and output growth (Yu and Jin, 1992) This hypothesis is supported by

the absence of causality between energy consumption and output

growth rate, and implies that energy conversing policies will not affect

output and hence employment adversely Supporters of the“growth

hypothesis” view energy as a compliment to labor and capital in the

production function Hence, reducing energy use will hamper output (Beaundreau, 2005; Ghali and El-Sakka, 2004; Lee and Chang, 2008;

Oh and Lee, 2004; Stern, 2000) Supporters of the “conservation hypothesis”, on the other hand, argue that positive relationship between energy use and output growth stems from positive effect

of output on energy, but not vice-versa Therefore, energy conversing policies may be implemented without hampering employment and output (Apergis and Payne, 2009; Lee and Chang, 2008) Finally,

“feedback hypothesis” implies that there is bidirectional causality between energy use and output growth Hence, reducing energy use may hamper output growth.2

Due to the importance of the issue both for policymakers and economists, the dynamic interaction between energy consumption and economic growth has been intensively investigated in energy economics literature since the seminal work ofKraft and Kraft (1978) However, the empirical evidence is mixed (see also literature surveys byOzturk, 2010; Payne, 2010) For example,Kraft and Kraft (1978), Akarca and Long (1979, 1980), Yu and Hwank (1984),

Abosedra and Baghestani (1989),Yu and Choi (1985),Erol and Yu (1987),Masih and Masih (1996),Cheng and Lai (1997),Ang (2008),

Zhang and Cheng (2009),Zamani (2007)andMehrara (2007)found unidirectional causality running from economic growth to energy con-sumption On the other hand,Yu and Choi (1985),Masih and Masih

⁎ Corresponding author Tel.: +90 312 2978651/164; fax: +90 312 299 2003.

E-mail addresses: aysens@hacettepe.edu.tr (A Araç), muhariz.hasanov@okan.edu.tr

(M Hasanov).

1

Tel.: +90 216 677 1630; fax: +90 216 677 1667.

2

For a thorough discussion of the issue, see, for example, literature surveys by Ozturk (2010) and Payne (2010)

http://dx.doi.org/10.1016/j.eneco.2014.04.013

Contents lists available atScienceDirect

Energy Economics

j o u r n a l h o m e p a g e : w w w e l s e v i e r c o m / l o c a t e / e n e c o

(1996),Asafu-Adjaye (2000),Bowden and Payne (2009),Belloumi

(2009),Stern (2000),Oh and Lee (2004),Wolde-Rufael (2004)and

Ho and Siu (2007)found unidirectional causality running from energy

consumption to economic growth.Glasure (2002),Erdal et al (2008)

andBelloumi (2009)found bidirectional causal relationship between

energy consumption and economic growth, whereasHalicioglu (2009)

andPayne (2009)found no causality between them

A common feature of the aforementioned studies is that all of them

used linear models Conflicting findings regarding the dynamic

interac-tions between energy consumption and economic growth may be

attributed to the assumption that the relationship between energy

con-sumption and economic growth is linear In linear models the

parame-ters assumed to be constant over the sample period which implies

that the relationship between energy consumption and economic

growth is stable However, some events such as changes in the policies,

energy crises and economic crises could affect the parameters Hence, in

time series framework these changes must be taken into consideration

in order to avoid spurious results

Recently, a growing number of theoretical and empirical studies

have taken into consideration nonlinearity to analyze the dynamic

interactions between the macroeconomic series in question.Moon

and Sonn (1996),Lee and Chang (2007),Chiou-Wei et al (2008),

Huang et al (2008),Cheng-Lang et al (2010),Rahman and Serletis

(2010), among others, have investigated possible nonlinear

relation-ships between energy use and macroeconomic variables By introducing

an endogenous growth model that emphasizes energy requirements to

support potential growth,Moon and Sonn (1996)claim that at the

beginning, economic growth rate increases with productive energy

expenditures but it subsequently decreases They estimated their

theoretical model with Korean data to confirm the validity of their

hypotheses Taking account of the fact that level of development

may affect the interrelationship between energy use and economic

growth,Lee and Chang (2007)examined energy consumption

out-put growth causality by categorizing countries into different groups

by level of development Their results suggest that the causality

be-tween energy consumption and output level is not linear, and varies

with output level

Chiou-Wei et al (2008)used nonlinear causality tests besides

lin-ear causality tests to examine causality between energy

consump-tion and economic growth in the case of eight Asian countries and

the USA They argue that changes in the economic events and

re-gime changes such as changes in energy policy orfluctuations in

en-ergy price can cause structural changes in the pattern of enen-ergy

consumption, which in turn, creates a room for a nonlinear

relation-ship between energy use and economic growth When they take

into account nonlinearity in the interrelationship between energy

consumption and output, the direction of causality between the

var-iables is reversed in the cases of Taiwan, Singapore, Malaysia and

Indonesia On the other hand, in the cases of Korea, Hong Kong,

Philippines, Thailand and the USA both linear and nonlinear

causal-ity tests imply the same direction of causalcausal-ity or non-causalcausal-ity

Based on the arguments ofChiou-Wei et al (2008)that changes in

economic environment, policy alterations and world energy prices

may lead to a nonlinear interrelationship among electricity

con-sumption and economic activity,Cheng-Lang et al (2010)analyzed

causality between total and sectoral electricity consumption levels

and output in Taiwan They concluded that there is bidirectional

nonlinear causality between total electricity consumption and real

output In addition, theyfind that there is unidirectional nonlinear

causality from output level to residential electricity consumption

In order to investigate nonlinear relationships between energy

con-sumption and economic growth for 82 countries,Huang et al (2008)

employed threshold regression models Their results suggest a signi

fi-cant positive relationship between energy consumption and output

growth for regimes associated with lower threshold values However,

when the threshold variables are higher than certain threshold levels,

they found either no significant relationship or a significant but negative relationship between energy consumption and economic growth

Rahman and Serletis (2010)examined asymmetric effects of oil price and monetary shocks using data for the United States In particular, they employ a nonlinear VAR model by using realized oil price volatility as a regime switching variable Theyfind that both oil prices and oil price volatility have impacts on macroeconomic activity In addition, they find that monetary policy not only reinforces the effects of oil price shocks on output, but it also contributes to the asymmetries in the effects of oil price shocks on output

Hasanov and Telatar (2011)analyzed stationarity properties of per capita total primary energy consumption across 178 countries around the world allowing for both structural breaks and nonlinearities

in the data generating process, and found that allowing for breaks and nonlinearities in the data generating process leads to more frequent rejection of the null hypothesis of unit root They also tested linearity

of the series under investigation and found that all series under consid-eration are characterized by some type of nonlinearity They suggested taking account of possible nonlinear dynamics in analyzing relationship between energy use and macroeconomic variables

The results obtained in these empirical studies imply that nonlinear-ity may stem from level of development or changes in energy policies or fluctuations in world energy markets In fact,Hasanov and Telatar (2011)argue thatfluctuations in energy prices may lead nonlinear dynamics in presence of adjustment costs As they point out, a change

in input prices affectsfirms' input demands Firms react to increases in energy prices by reducing energy use in the short run, and adopting energy saving production technologies in the long run However, adoption of new technologies is costly Hence, if the costs of adoption

of new technology are greater than the costs of operation with energy-intensive technology, thenfirms shall not adjust their produc-tion processes On the other hand, if the gains from adopting new technology cover the costs of adjustment, thenfirms will incur adjust-ment costs and adopt energy-saving technology This implies that the adjustment of energy consumption to the desired level might be inherently nonlinear

The purpose of this study is to examine possible nonlinearities in the dynamic interaction between energy consumption and economic growth in Turkey Several authors have examined energy consumption and economic growth nexus and reported conflicting results in the case

of Turkey.Soytaş and Sarı (2003)employed a vector error correction model (VECM) and concluded that unidirectional causality runs from energy consumption to economic growth for the 1960–1995 period

Altinay and Karagol (2004), using Hsiao' version of Granger method over the period 1950–2000, found no causal relationship between

ener-gy consumption and economic growth.Altinay and Karagol (2005)

focused on the 1960–2003 period and used a VAR model and standard Granger test and found unidirectional causality running from electricity consumption to economic growth.Jobert and Karanfil (2007)employed

a cointegration and Granger causality analysis and found no evidence of causality between energy consumption and economic growth in the long run.Lise and Van Montfort (2007), using an error correction model (ECM) for the 1970–2003 period, concluded that unidirectional causality runs from economic growth to energy consumption.Erdal et

al (2008), employing a pairwise Granger causality for the 1970–2006 period, concluded that bidirectional causality exists between energy consumption and economic growth.Halicioglu (2009)undertook an autoregressive distributed lag (ARDL) approach for the 1960–2005 period and found that energy consumption and economic growth are neutral to each other.Yalta (2011), using a maximum entropy bootstrap over the period 1950–2006, found an evidence supporting the neutrality between energy consumption and economic growth Our approach in this paper differs from previous researches on energy-output relationship for the case of Turkey As briefly discussed

inSection 2, Turkey has undergone serious structural changes during the analyzed period In addition, Turkey has limited energy sources

and heavily depends on imported energy, which increases vulnerability

of economy of the country to energy shocks These specific features of

the Turkish economy coupled with massive reforms in the energy sector

imply that simple linear models might be inadequate to analyze

energy-output causality in Turkey Therefore, we prefer to use a nonlinear

model to examine possible nonlinearities in the interrelationship

be-tween energy use and output growth Specifically, we employ smooth

transition (STR) vector autoregressive model, as detailed inTeräsvirta

and Anderson (1992)andTeräsvirta (1994) The STR modeling has

several advantages over the competing nonlinear models First, STR

models are theoretically more appealing over simple threshold and

Markov regime switching models, which impose an abrupt change in

coefficients Instantaneous changes in regimes are possible only if all

economic agents act simultaneously and in the same direction Second,

the STR model allows for modeling different types of nonlinear and

asymmetric dynamics depending on the type of the transition function

In particular, a STR model with afirst-order logistic transition function

is more convenient for modeling the interaction between energy

consumption and output growth rate if the dynamic interrelationships

between the variables depend on the phases of business cycles On the

other hand, a STR model with an exponential or second-order logistic

transition function is more convenient if, for example, the interaction

between the variables depend not on the sign but on the size of

fluctu-ations in variables Finally, the STR modeling allows one to choose both

the most appropriate switching variable and type of transition function

unlike other regime-switching models In passing we note thatRahman

and Serletis (2010)also used a multivariate STR model and found that

these models capture possible asymmetries in the effects of energy

shocks on macroeconomic variables quite well

The main novelty of this paper is that we examine possible

asymmetries in the effects of big versus small and negative versus

positive shocks to one variable on another Traditional linear models

assume that negative energy shocks have the same effect on output

level as the positive energy shocks of the same magnitude in absolute

terms and that a twofold energy shock affects output twice larger

than a onefold energy shock However, wefind asymmetries in the

ef-fects of positive versus negative and small versus large shocks In

partic-ular, our computations show that negative energy shocks have greater

effect on output growth rate in absolute terms than the positive energy

shocks In addition, wefind that big negative energy shocks have

rela-tively larger effects on output than small negative energy shocks On

the other hand, wefind no asymmetry in the effects of big and small

positive energy shocks As regards the effects of output shocks on

ener-gy use, wefind that small and big negative output shocks have

negligi-ble effects on energy use whereas positive output shocks affect energy

use Interestingly, our results imply that small output shocks have

relatively greater effects on energy use than larger output shocks These

results shed light on design of appropriate energy policies in Turkey

The remaining of this paper is organized as follows.Section 2

pro-vides the general overview of output growth and energy policies in

Turkey over the period 1960–2010.Section 3specifies a brief

descrip-tion of smooth transidescrip-tion vector autoregressive (STR-VAR) model and

computations of the generalized impulse response function.Section 4

presents the data and empirical evidence Thefinal section discusses

policy implications of the results and concludes the paper

2 Economic performance and energy in Turkey in the

period 1960–2010

Turkey has implemented quite different macroeconomic policies

and has undergone several economic and political crises since 1960.3

Turkey pursued an import-substitution industrialization and growth

strategy during the period 1960–1980 This period is also known as

“planned development period” With the establishment of the State Planning Organization in 1960 and adoption of a new constitution in

1961, state enterprises started to play a key role in industrialization

of the economy Many state enterprises were established during this period, which have made huge investments in heavy manufacturing and capital goods sectors in order to boost economic growth and contribute to industrialization of the economy

The main target of the import-substitution industrialization policies implemented until 1980 was to achieve foreign trade balance and reduce dependence on imported goods However, these policies were not successful in achieving these targets as industrial production was heavily dependent on intermediate goods imports As a result, Turkey abandoned inward oriented growth strategies and adopted export-oriented growth strategies in 1980 in a response to a severe balance of payment crisis

Turkey has implemented massive economic reforms throughout the 1980s aimed at liberalization of import regime and encouraging exports and foreign direct investments, adopting a flexible exchange rate regime, liberalization of current account transactions, and removing price controls While private enterprises were considered as main actors

of economic activity, the government continued to undertake huge investments in order to bolster economic growth Unlike the 1960s and 1970s, public investments were directed mainly to infrastructure rather than industry The engine of economic growth during the 1980s was huge public investments in infrastructure and growth in exports Although economy grew well during the 1980s, government's concern has shifted from external competitiveness to domestic stability in the face of accelerated inflation Therefore, Turkey abandoned real ex-change rate depreciation policy aimed at export promotion in order to prevent inflationary effects, as a consequence output growth slowed with the loss of external competitiveness The exchange rate policy aimed at stabilization of exchange rates was unsustainable as foreign trade deficits grew drastically As a result, Turkey was obliged to aban-don exchange rate policy in 1994 and Turkey underwent a severe crisis

in that year

Although stabilization program announced in that year was success-ful in stabilizingfinancial markets and reducing budget deficits, the government once more switched to expansionary policies due to political considerations Since comprehensive reforms were not under-taken, the stabilization policies had limited effects, and the economic condition began to deteriorate starting from 1995 The economic growth slowed down in 1999 when two earthquakes hit the most industrialized region of Turkey Turkey adopted an IMF-backed exchange rate stabilization program in December 1999 However, the failure of the government to implement massive structural reforms and privatization has diluted the credibility of the stabilization program

As a result, the stabilization program failed, and an economic and financial crisis outburst in early 2001

Turkey abandoned exchange rate targeting and implemented massive reforms starting from 2001 The Central Bank was given independence and adopted inflation targeting policies Fiscal policy was tightened, goods andfinancial markets were further liberalized, and massive privatization has been implemented As a result, Turkey enjoyed relatively high growth rates and low inflation post-2001 period Even though Turkey underwent several crises and outputfluctuated widely during the analysed period, economy grew considerably during the analysed period and Turkey succeeded to transform from agricultural economy to industrial economy In fact, the annual compounded GDP growth rate from 1960 to 2010 was 4.3% Total energy consumption rose by 4.6% per annum during the same period.Table 1below presents average annual growth rates of per capita GDP and energy consumption

as well as share of energy imports in total energy use and output volatility First note that output volatility was relatively high and increased through-out the period as a result of the economic crises in 1994 and in 2001 The table suggests that per capita energy use grew with per capita income

3

See Müslümov et al (2002) , Dibooglu and Kibritcioglu (2004) , Hasanov and

Omay (2008) , Hasanov et al (2010) , among others, for a brief discussion on economic

However, the pattern of energy consumption growth was not the same

over the decades In particular, energy growth was lower than output

growth during the 1960–1969 and post 2000 periods, but higher than

output growth 1970 to 1990, and almost the same as the output growth

in the 1990s This implies that the relationship between output and

energy was not the same over the analysed period

Energy mix also changed drastically during the analysed period In

particular, the share of fuels with higher calorific values increased

over the pastfive decades.Table 2below presents share of fuels in

total energy consumption of Turkey As the table suggests, the share of

low calorific fuels (wood, bio-fuels and wastes) dropped drastically

from 32% in 1970 to 3% by 2010, whereas share of hydrocarbons

(oil and gas) rose to almost 60% by 2010 The share of coal and

hydro-electric power also rose during this period This is an expected result

as use of hydrocarbon resources grows with industrial output In the

face of poor endowment with hydrocarbon resources, energy imports

of Turkey also rose drastically as industrial production increased

While energy imports constituted around 12.3% of total energy use in

1960, the share of energy imports in total energy use rose to almost

70% by 2010 Growing dependence on energy imports has increased

vulnerability of the Turkish economy to world energy shocks

3 Econometric methodology

We followWeise (1999)andRothman et al (2001), who generalize

the smooth transition autoregressive (STAR) model ofTeräsvirta (1994)

to vector autoregressive (VAR) models Let lentand lytdenote the

energy consumption and the output growth in logarithm, respectively

Then the smooth transition vector autoregressive (STR-VAR) model

for energy consumption and output growth can be written as follows:

xt¼ Ψ1 ;0þXp

i¼1

Ψ1 ;ixt−iþ Ψ2 ;0þXp

i¼1

Ψ2 ;ixt−i

!

 F sðt; γ; cÞ þ εt ð2:1Þ

where xtis a (2x1) column vector given by xt= (lent, lyt) ', Ψj,0, j = 1, 2

are (2x1) vector of constants,Ψj,i, j = 1, 2, i = 1, 2,… p are (2x2)

matrices of parameters, andεt= (εent,εyt) ' is a (2x1) vector of residuals with mean zero and (2x2) covariance matrixΣ The transition function F(st;γ,c) is a continuous function bounded between zero and one The STR-VAR model can be interpreted as a regime switching model that allows for two regimes, associated with the extreme values of the tran-sition function, F(st;γ,c) = 0 and F(st;γ,c) = 1, whereas the transition from one regime to the other is gradual The regime that occurs at time t can be determined by the observable variable stand associated value of F(st;γ,c)

One of the popular choices of the transition function F(st;γ,c) is the logistic function as given below:

F sðt; γ; cÞ ¼ 1 þ exp −γ sð f ðt−cÞgÞ−1; γ N0 ð2:2Þ This choice of the transition functions gives rise to the logistic smooth transition vector autoregressive (LSTVAR) model Here, stis a transition variable, andγ and c are slope and threshold parameters, respectively The restrictionγ N 0 is an identifying restriction As st

increases, the logistic function F(st;γ,c) changes monotonically from

0 to 1 around threshold parameter c with F(c;γ,c) = 0.5 The slope parameterγ determines the smoothness of transition from one regime

to another This function can be convenient for modeling, for example, business cycle asymmetries to distinguish expansions and recessions (Teräsvirta and Anderson, 1992)

In the LSTVAR model, the regimes are associated with small and large values of the transition variable stwith respect to the threshold

c In order to allow for regime-changing behavior that depends on the absolute value but not on the sign of the transition variable, one may use the following exponential transition function:

F sðt; γ; cÞ ¼ 1−exp −γ sn ðt−cÞ2o

The exponential transition function equals zero when st= c and approaches 1 for larger deviations of the transition variable stfrom the threshold value c Thus, in the exponential smooth transition vector autoregressive (ESTVAR) models, regime changes are associated with small and big values of the transition variable stwith respect to the threshold value c One of the shortcomings of the exponential function

is that ifγ approaches either 0 or infinity, the function F(st;γ,c) becomes

a constant In this case, one may use quadratic or second-order logistic transition function (see, for example,van Dijk, 1999)

One may adopt a“specific-to-general” approach to specify an appropriate STR-VAR model In thefirst step of this approach, one first estimates an appropriate linear model and then tests linearity against a smooth-transition type nonlinearity The linearity tests, however, are complicated by the presence of unidentified nuisance parameters In particular, under the null hypothesis of linearity,

Ψ2,i, i = 0, 1, 2,… p are unidentified nuisance parameters, which renders standard asymptotic inference invalid FollowingLuukkonen

et al (1988), one may circumvent this problem by approximating the transition functions in Eq.(2.1) by a third-order Taylor expansion

Table 1

GDP and energy use in Turkey, 1960–2010.

Periods

Source: World Development Indicators.

Notes: The annual growth rates of per capita GDP and energy use are calculated as compounded annual growth rates for the relevant period Net energy imports are as of the end-of-period Output volatility is measured as the standard deviation of annual growth rates of per capita GDP for the relevant decade.

Table 2

Share of fuels in total energy consumption.

Wood, agricultural residues

and animal wastes

Source: Our own calculations using data from Total Energy Balance 2010, and Primary

Energy Consumption by Source, 1970–2006 Ministry of Energy and Natural Resources

of Turkey.

Raw data on primary energy consumption by source is available online at ministry's website

at: http://www.enerji.gov.tr/index.php?dil=tr&sf=webpages&b=y_istatistik&bn=

around the null hypothesis After replacing the transition function in

Eq.(2.1)by a third-order Taylor expansion, one obtains the following

auxiliary regression model:

xt¼ Φ1 ;0þXp

i¼1

Φ1 ;ixt−iþXp

i¼1

Φ2 ;ixt−istþXp

i¼1

Φ3 ;ixt−is2t

þXp

i¼1

Φ4 ;ixt −is3tþ et

ð2:4Þ

The vector etcomprises the original shocks εtas well as the

error arising from the Taylor approximation In Eq (2.4) it is

assumed that the transition variable stis one of the elements of xt If this

is not the case, then additional regressorsΦ5st, i = 1, 2, 3 enter the

aux-iliary regression model(2.4) The parameters inΦj,i, j = 1, 2 , 4, i =

0, 1, 2,… p are functions of the parameters Ψj,i, j = 1, 2, i =

0, 1, 2,… p, γ and c in the bivariate STR-VAR model in Eq.(2.1)

In Eq.(2.4), it is clear thatΦ1=Ψ1andΦ2,i=Φ3,i=Φ4,i= 0 if

and only ifγ = 0 in Eq.(2.1) Therefore, the null hypothesis of

linearity in the auxiliary regression model(2.4)can be written as

H0:Φ2,i=Φ3,i=Φ4,i= 0, which can be tested directly by a likelihood

ratio (LR) test LetΩ0¼ ∑^et^e0t=T and Ω1¼ ∑^et^e0t=T be the estimated

variance–covariance matrices of residuals from the restricted and

unrestricted regressions, respectively Then the LR test statistic for

linearity of a k variable VAR model with p lags is given by LR = T

{log |Ω0|− log |Ω1|}, which is asymptotically distributedχ2(3pk2)

In particular, the empirical specification procedure for STR-VAR

models consists of the following steps:

1 Specify an appropriate linear VAR model for time series xtunder

investigation

2 Test the null hypothesis of linearity against the alternative of

STR-type nonlinearity To identify the appropriate transition

variable st, the LR test can be computed for several candidates,

and the one for which the p-value of the test statistics is smallest

can be selected If the null of linearity is not rejected against

either alternative, then retain the linear VAR model

3 If the linearity tests described above suggest that the appropriate

model is a STR-VAR model, then one may follow the procedure of

Teräsvirta (1994)to decide whether logistic or exponential functions

are more convenient transition function.Teräsvirta (1994)suggests

using a decision rule based upon a sequence of tests In particular,

he proposes to test the hypotheses

H03: Φ4¼ 0

H02: Φ3¼ 0jΦ4¼ 0

in Eq.(2.4)by means of LM type test The decision rule is as follows:

(i)Φ4is always nonzero only if the model is an LSTVAR model, (ii)Φ3

is always nonzero if the transition function is“exponential” function,

(iii)Φ2is always nonzero if the transition function is a logistic

function If the p-value corresponding H02is the smallest, a model

with“exponential” function should be chosen, otherwise the logistic

transition function should be preferred as the transition function

4 Once the appropriate model is chosen, the model can be

esti-mated by nonlinear least squares, and used for descriptive or

forecasting purposes

After estimating the model, we use generalized impulse response

functions to examine the dynamic relationships between energy

consumption and economic growth The method of computation

and features of the generalized impulse response functions are

discussed byKoop et al (1996)in great detail.4

The impulse response measure which is commonly used in the analysis of linear models is defined as the difference between two realizations of xt + nwhich starts from identical histories of the time series up to time t− 1, denoted as wt − 1 In one realization, the process

is‘hit’ by a shock of size vt=δ at time t, while in the other realization no shock occurs at time t All shocks in intermediate periods between t and

t + n are set equal to zero in both realizations That is, the traditional impulse response function [TIRF] is given by

TIRFxðn; vt; wt−1Þ ¼ E x tþnjvt¼ δ; vtþ1¼ 0; …; vtþn¼ 0; wt−1

−E x tþnjvt¼ 0; vtþ1¼ 0; …; vtþn¼ 0; wt−1 ð2:6Þ for n = 0, 1,…

The traditional impulse response function as defined above has some characteristic properties in case the model is linear First, the TIRF is symmetric, in the sense that a shock of−δ has exactly the opposite effect as a shock of size +δ Furthermore, it might be called linear, as the impulse response is proportional to the size of the shock Finally, the impulse response is history independent as it does not depend on the particular history wt − 1 Because of these properties, TIRF is not appropriate for nonlinear models In nonlinear models, the impact of a shock depends on the sign and the size of the shock, as well as on the history of the process Furthermore, if the effect of

a shock on the time series nN 1 periods ahead is to be analyzed, the assumption that no shocks occur in intermediate periods might give rise to quite misleading inference concerning the propagation mechanism of the model (van Dijk, 1999)

The Generalized Impulse Response Function [GIRF], introduced by

Koop et al (1996)provides a natural solution to the problems involved

in defining impulse responses in nonlinear models The GIRF for an arbitrary shock vt=δ and history wt − 1is defined as

GIRFxðn; vt; wt −1Þ ¼ E x tþnjvt; wt −1

−E x tþnjwt−1

for n

The GIRF is a function of vtand wt − 1 It is natural to treat vt, and wt − 1

as realizations from the same stochastic process that generates the realizations of {xt} (Koop et al., 1996) The generalized impulse response function (GIRF) is designed to solve the problems categorized

as follows: What types of shocks (e.g., variable-specific or system-wide shocks) hit the system at time t? What is the“history” of the system at time t− 1 (e.g., expansionary or recessionary) before the shock hits? What future shocks are assumed to hit the system from t + 1 to

t + n? The problem of treatment of the future is circumvented by using the expectation operator conditioned on only the history and/or shock Thus, the response constructed is an average of what might happen given the present and past The natural baseline for the im-pulse response function is then defined as the conditional expectations, given only the history

Koop et al (1996)suggest that the impulse response functions are to

be computed by simulating the model In order to compute the impulse response functions, the following algorithm might be used (see also

Weise, 1999):

1 Pick a history wtr− 1 The history is the actual value of the lagged endogenous variables at a particular date

2 Pick a sequence of (k-dimensional) shocks vt + nb , n = 0,…, q The shocks are drawn with replacement from the estimated residuals of the model The shocks are assumed to be jointly distributed, so if date t's shock is drawn, all k residuals for date t are collected

3 Using wtr− 1and vt + nb , simulate the evolution of xt + nover q + 1 pe-riods Denote the resulting baseline path Xt + n(wtr− 1, vt + nb ), n = 0, , q

4 Substitute vi0 for the i,0 element of vt + nb and simulate the evolution of Xt + nover q + 1 periods Denote the resulting path

X (v , wr , vb ), n = 0, , q

4

For a thorough discussion of generalized impulse response functions and comparison

to traditional impulse response functions, see also Chapter 2.6 in van Dijk (1999)

5 Repeat steps 2 to 4 B times.

6 Repeat steps 1 to 5 R times and compute Xt + na (vi0) =

[Xt + n(vi0, wtr− 1, vt + nb ) − Xt + n(wtr− 1, vt + nb )]/BR for

the average impulse response function, or Xt + nrn (vi0) =

median[Xt + n(vi0, wtr− 1, vt + nb ) − Xt + n(wtr− 1, vt + nb )]

for the median response

In this paper, we compute impulse responses for ten periods (n = 10)

for all available data points (R = 36)5with B = 1000 replications

4 Data and estimation results

In this paper, we use the annual data spanning the period

1960–2010 for Turkey This time period is dictated by data

avail-ability Energy consumption is proxied by energy use (kg of oil

equivalent per capita) The output level here is GDP per capita

All data are obtained from World Development Indicators

data-base of The World Bank Wefirst present the data and then give

estimation results

4.1 Preliminary data analysis

Table 3presents basic descriptive statistics of the growth rates of

output and energy consumption As the table reveals, average growth

rate of energy consumption was higher than the output growth rate,

mainly due to industrialization of the economy Although both output and energy consumption exhibited large fluctuations during the analysed period,fluctuations in per capita energy use were larger than per capita income Large falls in both variables were observed during the economic crisis in 2001 Large increases in both variables, on the other hand, were observed during the early industrialization periods The statistics presented in theTable 3also suggest that although both variables are negatively skewed, the null hypothesis that the series are normally distributed cannot be rejected In addition,Ljung and Box's (1978)Q test for autocorrelation does not reject the null hypothesis of

no correlation of order one Similarly,Engle's (1982)ARCH-LM test indicates no conditional heteroscedasticity in both of the series

In order to better understand thefluctuations in energy use and output growth we present a graph of the growth rates of output and energy consumption inFig 1below As can be seen fromFig 1, energy consumption and output moved in the same direction for most of the period On average, both per capita output and energy use grew rapidly during the 1960s and at the beginning of the 1970s as a consequence of rapid industrialization during that period Furthermore, as thefigure suggests, energy use and output moved closely with each other during the 1990s On the other hand, this close co-movement weakened during other sub-periods In fact, as further elaborated inSection 2, per capita output and energy use grew on average by 1.5% during the 1990 whereas growth rate of energy use exceeded that of output growth dur-ing the period 1970–1990 This pattern of co-movement of the variables suggests that the relationship between energy use and output might

be nonlinear Now, we turn to formal statistical tests to determine the appropriate model for examining the dynamic interrelationship between the variables

5 Note that we use the annual data covering the period 1960–2010 After preliminary

data transformations and lags used to estimate the model, we have 46 data points Since

we estimate impulse responses for ten periods, there remain only 36 data points for

com-Table 3

Descriptive statistics.

Output growth rate 0.025 0.039 −0.070 0.088 −0.705 (0.048) 0.010 (0.989) 4.138 (0.126) 0.000 (0.993) 0.002 (0.961) Energy use growth rate 0.028 0.041 −0.091 0.105 −0.664 (0.063) 0.518 (0.486) 4.231 (0.121) 0.214 (0.644) 0.525 (0.469) Notes: S.E denotes standard error, Sk denotes excess skewness, Ku denotes excess kurtosis, and J–B denotes Jarque–Berra's test for normality of series Q(1) is Ljung–Box's Q test for autocorrelation of order one ARCH(1) is Engel's (1982) LM test for first order autoregressive conditional heteroscedasticity p-values of diagnostic tests are provided in parenthesis.

1961 1965 1969 1973 1977 1981 1985 1989 1993 1997 2001 2005 2009 -0.100

-0.075 -0.050 -0.025 -0.000 0.025 0.050 0.075 0.100

0.125

OUTPUT ENERGY

Fig 1 Output and energy consumption growth rates in Turkey, 1961–2010.

4.2 Stochastic properties of the series

The specification procedures described in the previous section rely

on the assumption that both the output growth (lyt) and energy

consumption (lent) are I(0) processes Therefore, prior to estimation of

the linear model, wefirst tested stationarity of the variables concerned

Taking account of the low power of conventional Augmented Dickey–

Fuller (ADF) test against alternative data generating processes, we

also carried out the Phillips–Perron (PP), Kwiatkowski, et al (KPSS),

and Ng and Perron (NG–Perron) unit root tests The KPSS test differs

from other tests in that it assumes that the series under investigation

are stationary under the null hypothesis whereas the other tests assume

that series have a unit root under the null hypothesis We apply all tests

both with and without a time trend

Table 4presents the unit root test results As can be seen from

the table, the ADF, PP and NG–Perron tests suggest that both energy

consumption and output level contain a unit root On the other hand,

the KPSS test suggests that per capita output may be trend stationary

In order to check if there is a cointegration relationship between

the variables, wefirst employedJohansen and Juselius (1990)(JJ)

cointegration tests, results of which appear below inTable 5 The

re-sults of the JJ cointegration test suggest no cointegration relationship

between energy use and output level

In addition to the conventional JJ cointegration test, we also applied

Pesaran et al (2001)'s bounds testing procedure and nonlinear

cointegration tests ofKapetanios et al (2006) The bounds testing

approach proposed byPesaran et al (2001)is applicable irrespective

of whether the underlying regressors are purely I(0), purely I(1) or

mutually cointegrated This test became popular recently in the face of

increasing concerns about power and size properties of the unit root

tests In fact, the results presented inTable 4above lead to conflicting

conclusions about order of integration of the per capita income

There-fore, the bounds testing procedure ofPesaran et al (2001)provide a

better tool for testing relationship between levels of per capita income

and energy use

The statistic underlying the bound testing procedure is the familiar F-test with new critical values that they tabulate The statistic is used

to test the significance of lagged levels of the variables under consider-ation in a conditional unrestricted equilibrium correction model (ECM) The critical values consist of a lower bound on the assumption that all variables are integrated of order zero and an upper bound on the as-sumption that all variables are integrated of order one If the computed F-statistics falls behind the lower bound, it indicates no cointegration If the computed F-statistics exceeds the upper bound, the conclusive deci-sion can be made in favor of the cointegration If, however F-statistics falls within these bounds, inference would be inconclusive The critical values of the F-statistic for upper bound and the lower bound with one regressor are 4.04 and 4.78 respectively at the 10% level of signi fi-cance for the model with unrestricted intercept and no trend and, 5.59 and 6.26 for the model with unrestricted intercept and unrestricted trend, respectively The computed F-statistics are given inTable 6 The computed F-statistics fall behind the lower bounds, suggesting that energy consumption and economic growth are not cointegrated

As briefly described above, the dynamics of energy consumption and output suggest that the interrelationship between these variables might not be linear Hence, we also applied the cointegration test procedure proposed byKapetanois et al (2006) Unlike other cointegration tests, this test allow for possible nonlinearities in the adjustment to the equi-librium level Kapetanois et al (2006) propose four different test statis-tics, namely FNEC, FNEC⁎ , tNECand tNEG Computed test statistics are given in

Table 7 As none of these test statistics are significant even at 10% significance level, we conclude that energy consumption and output are not cointegrated

All considered cointegration tests suggest that the variables under consideration are not cointegrated Thisfinding implies that there exists

no long-run relationship between energy use and output level Hence, time paths of the variables are independent of each other in the long-run and the relationship between them are restricted to the short long-run 4.3 Linearity tests and smooth-transition VAR model

Since we found no cointegration relationship between the variables,

we estimated a VAR model in differences The nonlinearity tests are sensitive to autocorrelation So, the autoregressive structure of the

Table 4

Unit root test results.

Note: The lag lengths of each variable in each equation are selected by applying conventional AIC Figures in parenthesis are p-values of the test statistics *, **, *** denote rejection of the null hypothesis of unit root for NG–Perron test and rejection of the null hypothesis of stationarity for KPSS test at 10%, 5% and 1% significance levels, respectively.

Table 5

Cointegration test results.

Unrestricted cointegration rank test (trace)

Hypothesized

no of CE(s)

Eigenvalue Trace statistic Critical value(0.05) Probability

Unrestricted cointegration rank test (maximum eigenvalue)

Hypothesized

no of CE(s)

Eigenvalue Max-eigen statistic Critical value(0.05) Probability

Table 6

Pesaran et al (2001) 's Bounds testing results.

Dependent variable

Lag order F-statistics

(unrestricted intercept,

no trend)

F-statistics (unrestricted intercept, unrestricted trend)

Note: The lag lengths of each variable in each equation are selected by applying

model should be specified so as to capture significant autocorrelation in

the linear model The lag lengths of each variable in each equation were

selected by applying conventional Akaike Information Criterion (AIC),

and then the resultant model was tested against autocorrelation of

re-siduals Accordingly, one lag of each variable was used in each equation

Moreover, we added dummy variables into each equation of the VAR

model for outliers evident in the residuals in order to ensure that

rejec-tion of the null hypothesis of linearity is not due to the presence of big

outliers Wefirst estimated the equations without dummy variables,

and computed standard deviations of the residuals Then, we defined

positive outliers as those observations which are two times larger

than the standard deviation whereas negative outliers as those

observa-tions which are two times larger than the negative value of the standard

deviation for each equation

After estimating a linear VAR model, we tested linearity of the model

against STR type nonlinearity The results of the linearity tests are

provided inTable 8

In panel A ofTable 8, we report system-wide linearity tests for

several candidate transition variables As can readily be seen from the

table, the null hypothesis of linearity is rejected for many candidate

transition variables This result implies that the dynamic interaction

between energy use and output growth rate might be inherently

non-linear The null hypothesis is more convincingly rejected when the

third lag of the change in energy consumption is used as a candidate

transition variable Therefore, we chose this variable as the appropriate

switching variable and proceed to select the form of the transition

func-tion The tests of the null hypotheses H03, H02 and H01 are reported in

panel B ofTable 3 Since H02 is not rejected whereas both H03 and H01

are rejected at conventional significance levels, we choose logistic tran-sition function

After selecting both appropriate switching variable and the form

of the transition function, we estimated a LSTVAR for energy use and output growth rates The parameter estimates of the LSTVAR model are provided inTable 9

4.4 Asymmetries in the dynamic interrelationship between variables The parameters of the estimated LSTVAR model are difficult to interpret (e.g.,Rahman and Serletis, 2010) However, one may use impulse-response functions tofigure out the dynamic relationship between the variables in the estimated LSTVAR model Accordingly, the generalized impulse response functions (GIRF) introduced by

Koop et al (1996)are used to compute impulse responses

The computation of GIRFs in the case of multivariate nonlinear models is made difficult by the inability to construct analytical expressions for the conditional expectations, E[xt + n|vt, wt − 1] and E[xt + n|wt − 1] in Eq.(2.7) To deal with this problem, following suggestions ofKoop et al (1996), we carry out stochastic simulation

to construct the generalized impulse responses We use all available initial data points as histories, which leave 36 data points, and compute impulse responses for 10 consecutive periods Shocks for a particular horizon are randomly drawn from the residuals of the estimated nonlinear model

In order to assess possible asymmetries in the effects of positive ver-sus negative and small verver-sus big shocks, we compute impulse response

of one variable to positive and negative small and large shocks to the other variable In particular, small positive energy shock is set to one standard error of the residuals from the energy equation, whereas large energy shock is defined as two standard errors Negative small and large energy shocks are defined as negative of one and two standard errors, respectively In a similar way, positive and negative small and large output shocks are defined similarly as positive and negative one and two standard error of the residuals of the output equation,

Table 7

Nonlinear cointegration tests results.

Dependent

variable

Intercept only

Intercept and time trend

Intercept only

Intercept and time trend

Critical values of the F NEC , F ⁎ , t NEC NEC and t NEG tests at 10% significance level for the

intercept (intercept and trend) case are 11.79(13.95), 10.13(12.83), −2.92(−3.30),

and −2.98(−3.41), respectively The lag lengths of each variable in each equation

are selected by applying conventional AIC.

Table 8

Linearity test results.

Panel A System-wide linearity tests against STR-type nonlinearity

Candidate transition variable LR test statistic Probability

Panel B Transition function specification test

Table 9 The estimates of the LSTVAR model.

Output equation Energy equation

F(Δlen t − 3 ) Δly t − 1 1.696 (0.959)* −1.089 (0.893) F(Δlen t − 3 ) Δlen −1.137 (0.623)* –

F(Δlen t − 3 ) Δlen t − 1 −1.556 (0.807)* 0.931 (0.699)

Estimated transition function: F(Δlen t − 3 ) = (1 + exp

{−5.204(Δlen t − 3 + 0.028)}) −1

(5.246) (0.009)***

Residual diagnostic tests

Kurtosis (excess) −0.483 [0.536] −0.130 [0.867] J–B normality test 1.941 [0.379] 1.025 [0.599]

Note: D y+, D y − , D e+, D e − denote positive and negative dummy variables for the output and energy equations, respectively J–B is Jarque–Berra's test for normality of residuals Ljung–Box Q(j) denotes Ljung and Box (1978) Q-test for residual autocorrelation of order j ARCH(1) is Engel's (1982) LM test for first order autoregressive conditional heteroscedasticity Figures in parenthesis are standard errors of parameter estimates Sig-nificance levels of the diagnostic tests are provided in square brackets ***, **, and * denote significance at 1%, 5%, and 10% significance levels, respectively.

respectively The difference between these forecast values and baseline

model is the impulse response for a given shock and particular history

In this way, 1000 realizations of impulse responses are calculated for

each available history, and average impulse responses are obtained

Computed impulse responses are plotted below inFigs 2 and 3 In

thesefigures, responses to negative shocks are plotted with reversed

sign so as to compare them with the responses to positive shocks

Similarly, responses to two standard error shocks are divided by two

so as to compare them to responses to one standard error shocks

Fig 2presents the responses of output growth rate to positive and

negative small and large energy shocks As can be seen inFig 2, the

magnitude of the effects of negative energy shocks on output is greater

than the effects of positive energy shocks Thisfinding implies that a

reduction in energy consumption decreases output much more than

an increase in energy consumption boosts output In addition, the

com-puted impulse responses suggest that there is no asymmetry in the

effects of big versus small energy consumption shocks on output On

the other hand, large negative energy shocks affect output growth rate

much more than small negative energy shocks Thisfinding implies

that the effects of energy shocks depend not only on the sign, but on

the magnitude as well

The computed responses of energy consumption to output shocks

are plotted inFig 3

Fig 3suggests that while positive shocks to output growth rate

increases energy consumption, a negative shock to output has almost

no effect on energy use Thisfinding implies that although energy

con-sumption increases with increasing output growth rate, a decrease in

output growth rate does not affect energy use In addition, wefind

that small shocks to output growth rate have greater effects on energy use when compared to large output shocks Thisfinding implies that although energy consumption is increasing with output growth rate, the rate of growth of energy consumption is decreasing with higher output growth rates Finally, wefind no asymmetry in the effects of big versus small negative output shocks on the energy consumption

5 Policy implications and conclusion

In this paper we have examined possible asymmetries in the

dynam-ic interrelationship between energy consumption and economdynam-ic growth

in Turkey for the 1960–2010 period For this purpose, we first estimated

a linear VAR model for energy consumption and output growth rate, and tested for linearity of the model The results of the linearity tests suggest that the dynamic interrelationship between energy consump-tion and output growth rate is inherently nonlinear Then we estimated

a nonlinear smooth transition vector autoregressive model, and used generalized impulse response functions to assess dynamic effects of one variable on the other In order to examine nonlinearities in the effects of big versus small and negative versus positive shocks, we have computed generalized responses of variables to one and two stan-dard deviation positive and negative shocks This approach allows us to examine possible effects of moderate and aggressive energy conversing

or energy-promoting policies as well as shed light on the dynamics of energy consumption in the future as the economy grows further The computed impulse responses suggest that negative and positive energy shocks affect output growth rate asymmetrically We alsofind asymmetry in the big versus small negative energy shocks On the

Response of Output Growth to Positive and Negative 1 sd Energy Consumption Shock

-0.016

0.000

0.016

0.032

0.048

0.064

0.080

0.096

+1 sd shock -1 sd shock

Response of Output Growth to Positive 1 and 2sd Energy Consumption Shock

-0.016

0.000

0.016

0.032

0.048

0.064

0.080

0.096

+1 sd shock +2 sd shock

Response of Output Growth to Negative 1 and 2sd Energy Consumption Shock

1

-0.016

0.000

0.016

0.032

0.048

0.064

0.080

0.096

-1 sd shock -2 sd shock

Fig 2 Response of output growth to energy consumption shocks.

other hand, wefind no asymmetry in the effects of big versus small

pos-itive energy shocks Similarly, wefind that negative and positive output

shocks affect energy consumption asymmetrically Although wefind

asymmetry in the effects of small versus big positive output shocks,

wefind no asymmetry in the effects of negative output shocks

The results of this paper shed light on the nature of dynamic

interre-lationship between energy use and economic growth in Turkey, and

hence provide important information for the design of energy policies

in Turkey First, wefind that energy consumption increases with output,

but at declining rates This result implies that higher output growth

rates are achieved mainly by increasing employment rather than energy

or energy using capital inputs Such an outcome may be observed if

firms believe that higher growth rates will not be sustained in the

long run and hence hesitate to invest in capital goods In fact, as briefly

discussed inSection 2, outputfluctuated widely in Turkey as a result of

economic crises, which reduced investors' beliefs in Turkey's future

eco-nomic performance Thefinding that energy use increases with output

at declining rates is also important for policymakers and energy

special-ists, suggesting that they must take account of such a nonlinear behavior

in forecasting energy demand in the future

Second, wefind that a decrease in output growth rate has a

negligi-ble effect on the energy consumption Thisfinding suggests that energy

consumption exhibits a downward inertia in the case of Turkey This

re-sult implies that producers tend to substitute other production factors,

especially labor force, with energy and/or energy intensive capital

goods in bad economic conditions In fact, it is a well-established fact

that unemployment rises sharply in recession periods but declines

very slowly and far after economic recovery starts This was also the case during and after the economic crises in Turkey Therefore, it must not be surprising that a decline in output does not reduce energy use Third, our results imply that a decline in energy use decreases output much more than an increase in energy consumption increases output In addition, wefind that a larger decrease in energy consumption affects output much more than a small decrease in energy consumption One of the possible reasons of such an asymmetric behavior may be the use of energy intensive production processes In fact, although

ener-gy intensity in Turkey was far below when compared to developed countries, it stayed relatively stable throughout the analysed period In particular, energy intensity measured as kg of oil equivalent energy use per thousand constant dollars GDP fall only to 115 in 2010 from

120 in 1980 Despite successes in industrialization and structural trans-formation of the Turkish economy over the lastfive decades, technolog-ical progress in Turkey has not been energy saving and hence energy intensity stayed stable over longer periods of time As regards energy policies, thisfinding implies that aggressive energy conserving policies might be detrimental for output growth and hence for employment

as well Therefore, policy authorities must assess the effects of energy conserving policies very carefully, and adopt a gradual, rather than an aggressive energy conserving policies Possibly, most viable energy conserving policies in Turkey are those that promote energy-saving production technologies In particular, incentives for adopting more environment-friendly production processes might be more appropriate and effective than penalizing greenhouse gas emissions in the case of Turkey, and probably, other countries with similar economic structure

-0.016

0.000

0.016

0.032

0.048

0.064

0.080

0.096

+1 sd shock -1 sd shock

-0.016

0.000

0.016

0.032

0.048

0.064

0.080

0.096

+1 sd shock +2 sd shock

1

-0.016

0.000

0.016

0.032

0.048

0.064

0.080

0.096

-1 sd shock -2 sd shock

Fig 3 Response of energy consumption to output growth shocks.