Figure 1.2 Research Flow Chart
Chapter 2 Literature Review
2.1 Issues on Energy Consumption and Economic Growth
Ever since 1970s numerous studies have examined the relationship between energy consumption and economic growth. A major question concerning this issue is which variable leads to the other: Is energy consumption a stimulus for economic growth or does economic growth lead to energy consumption? One of the time series methodologies to employ is the concept of Granger causality. Following Kraft and Kraft (1978) who provide pioneering evidence in support of causality from GNP to energy consumption in the United States, many empirical studies later extend to cover other industrial countries such as the United Kingdom, Canada, Germany, Italy, Japan, and France (e.g., Yu and Choi, 1985; Erol and Yu, 1987). However, the related literature on developed and developing countries, with diverse methodologies, and using various time periods fails to reach a unanimous conclusion.
Because of the critical role played by energy in the economic growth, an energy conservation policy (whether or not it can successfully be propagated within an individual country) has been a striking topic widely explored. The directions of the causal relationship between energy consumption and economic growth can be categorized into four types and evidence on either direction has important implications for an energy policy. First, if there is a unidirectional causality from economic growth to energy consumption, then policies for reducing energy consumption may be implemented with only little adverse or no effect on economic growth, such as in a less energy-dependent economy (Lise and Montfort, 2007; Oh and Lee, 2004; Yoo and Kim, 2006). Second, if there is unidirectional causality from energy consumption to economic growth, then restrictions on the use of energy
may have significantly adverse effects on economic growth, while an increase in energy consumption may contribute to economic growth (Altinay and Karagol, 2005;
Lee, 2005; Narayan and Singh, 2007; Shiu and Lam, 2004; Wolde-Rufael, 2004; Yuan et al., 2007). Third, if there is a bidirectional causal relationship, then economic growth may demand more energy whereas more energy consumption may also induce economic growth. Energy consumption and economic growth complement each other such that radical energy conservation measures may significantly hinder economic growth (Jumbe, 2004; Yang, 2000; Yoo, 2005). Finally, if there is no causality in either direction, which is known as the ‘neutrality hypothesis’, then neither conservative nor expansive energy consumption has any effect on economic growth (Asafu-Adjaye, 2000; Wolde-Rufael, 2005).
Another time series methodology explaining the relationship between energy consumption and economic growth is the co-integration technique with a bivariate (e.g., Yang, 2000; Zachariadis, 2007; Zamani, 2007) or multivariate (e.g., Masih and Masih, 1997; Oh and Lee, 2004; Soytas and Sari, 2007) framework. Stern (1993) adopts a multivariate vector autoregression (VAR) model to explore the causal relationship between GDP, energy use, capital, and labor inputs in the United States, where using a quality-adjusted index of energy input in place of gross energy use.
Compared to the bivariate VAR analysis, the multivariate context is important because changes in energy inputs are more frequently countered by the substitution of other production factors, resulting in an insignificant overall impact on output. Stern (2000) further extends his previous analysis by incorporating the co-integration analysis with some relevant variables. The results show that there is co-integration in a relationship among GDP, capital, labor, and energy.
Ghali and El-Sakka (2004) employ the Johansen co-integration technique to analyze the relationship among output, capital, labor, and energy use in Canada on the
basis of neo-classical one-sector aggregate production technology. Their results indicate that the long-run movements of output, capital, labor, and energy use are related by two co-integrating vectors.
Lise and Montfort (2007) undertake a co-integration analysis not only to explore the link between energy consumption and GDP, but also to take into account environmental protection and economic development for Turkey. Co-integration is found between energy consumption and GDP, while the energy Kuznets curve (EKC) hypothesis is rejected.
The aforementioned literature strengthens Stern’s conclusions that energy can be considered a limiting factor in economic growth. Shocks to the energy supply tend to reduce output. Table 2.1 summarizes more details about these studies of causality and co-integration analysis between energy consumption and economic growth.
2.2 Issues on Oil Shocks and Economic Activity
The important role of crude oil in the global economy has attracted a great deal of attention among politicians and economists. Since the first oil shock in 1973-74, many studies have been undertaken into the oil price-macroeconomy relationship.
These studies have reached different conclusions over time. As such, Hamilton (1983), Burbidge and Harrison (1984), Gisser and Goodwin (1986), Mork (1989), Hamilton (1996), Bernanke et al. (1997), Hamilton (2003), and several others have concluded that there is a negative correlation between increases in oil prices and the subsequent economic downturns in the United States. Nevertheless, the relationship seems to lose significance as data from 1985 onwards are covered. In fact, the declines in oil prices occur over the second half of the 1980s are found to have smaller positive effects on economic activity than predicted by linear models
Table 2.1 A Comparison of Earlier Studies about Causality and Co-integration Analysis Between Energy Consumption and GDP
Authors Countries Study period Causality Co-integration relationship
Cheng and Lai (1997) Taiwan 1955-1993 GDP→ EC No co-integration Ghali and El-Sakka
(2004)
Canada 1961-1997 GDP↔ EC Co-integration
Hondroyiannis et al.
(2002)
Greece 1960-1996 No causality Co-integration
Hwang and Gum (1992) Taiwan 1955-1993 GDP↔ EC Lee (2005) 18 developing
countries
1975-2001 EC→ GDP Co-integration
Lee and Chang (2005)
Taiwan 1954-2003 GDP↔ EC No co-integration
Lise and Montfort (2007)
Turkey 1970-2003 GDP→ EC Co-integration
Masih and Masih Oh and Lee (2004) South Korea 1961-1990 No causality Co-integration
Soytas and Sari (2003) 16 countries 1950-1992 EC→ GDP in Turkey
Co-integration for 7 out of 16 countries
Stern (2000) U.S. 1948-1994 EC→ GDP Co-integration Yang (2000) Taiwan 1954-1997 GDP↔ EC
Zamani (2007) Iran 1967-2003 GDP→ EC
considered up to then. After taking into account the role of the breakdate 1985-1986, some researchers argue that the instability observed in this relationship may be due to a mis-specification of the functional form used. The linear specification might mis-represent the relationship between economic growth and oil prices.
The mis-specification of linear function form has led to different attempts to reestablish the measures of the relationship between oil price changes and output.
On the one hand, Mork (1989) separates out oil price changes into negative and positive oil price changes, concluding that the decreases are not statistically significant. Thus, the results confirm that the negative correlation between GDP
growth and oil price increases remain when data from 1985 onwards are included.
Mory (1993) follow Mork’s (1989) measures and separated the oil price into negative and positive oil price changes. He finds that the positive oil price shocks Granger-caused the macroeconomic variables, but that negative shocks do not.
Mork et al. (1994) also find the asymmetric effects for seven industrialized countries.
On the other hand, Lee et al. (1995) report that the response to an oil price shock by the economic growth depends on the environment of oil price stability. An oil shock in a price stability environment is more likely to have larger effects on GDP growth than those occur in a price volatile environment. These researchers propose a measure that takes the volatility into account through a GARCH-based on oil price transformation. This transformation scales estimated oil price shocks by their conditional variance. They find asymmetry in the effects of positive and negative oil price shocks, but they also reestablish the significance of the above-mentioned negative correlation. Using the same way, Hamilton (1996) shows that it seems more appropriate to compare the prevailing oil price with what it is during the previous year, rather than the previous quarter. Finally, Hamilton (2003) provides evidence of a non-linear representation and states that the functional form that relates GDP growth to oil price changes is similar what has been suggested in earlier studies.
He specially analyzes the three non-linear transformations of oil prices proposed in the literature (i.e., Mork, 1989, Lee et al., 1995 and Hamilton, 1996), indicating that the formulation of Lee et al. (1995) has the best work of summarizing the non-linearity.
Afterwards, there are several works to study the impacts of oil price shocks, and the related issues can be divided into two parts. The first one part is related to macroeconomic level. Papapetrou (2001) analyzes the dynamic interactions among interest rates, real oil prices, real stock returns, industrial production and the
employment for Greece. The evidence suggests that oil price changes affect real economic activity and employment. Cunado and Pérez de Gracia (2003) analyze the oil price-macroeconomy relationship by analyzing the impact of oil prices on inflation and industrial production for European countries. Using the transformation of oil price data, they find that oil prices have permanent effects on inflation and short run with asymmetric effects on production growth. More recently, Farzanegan and Markwardt (2009) find a strong positive relationship between positive oil price changes and industrial output growth in the Iranian economy.
As to the Asian developing countries studies, Cunado and Pérez de Gracia (2005) find that oil prices have a significant effect on both economic activity and price indexes, although the impact is limited to the short run and more significant when oil price shocks are measured in local currencies. Moreover, they find evidence of asymmetries in the oil price-macroeconomy relationship across some of the Asian countries. Chang and Wong (2003) suggest that the impact of an oil price shock on the Singapore economy is marginal and small.
Another part involves in stock markets. Asset prices are determined on the stock market depending on information about future prospects as well as current economic conditions facing firms. Jones and Kaul (1996) examine stock market efficiency, focusing on the extent to which stock prices change in response to oil price changes, (i.e., whether changes in stock prices reflect current and future real cash flows). By using a cash-flow/dividend valuation model, they find that oil prices can predict stock returns and output on their own. Sadorsky (1999) identifies that oil price shocks and its volatility play an important part in explaining US stock returns and the movements of oil price explained more than interest rates for the forecasting variance. Cong et al. (2008) find that oil price shocks do not show statistically significant impact on the real stock returns of most Chinese stock market indices.
Park and Ratti (2008) show that oil price shocks have s statistically significant impact on real stock returns contemporaneously and within the following month in US and 13 European countries. Besides, they show that there is little evidence of asymmetric effects on stock returns of positive and negative oil price shocks.
Apergis and Miller (2009) also show that different oil market structural shocks play significant role in explaining the adjustment in stock returns. However, the magnitude of such effects proves small. Bjørnland (2009) analyzes the effect of oil price shocks on stock returns in Norway. He finds that following a 10% increase in oil prices, stock returns increase by 2.5%. Table 2.2 summarizes the aforementioned and existing literature about the effects of oil price changes on macroeconomic activities and stock markets.
Table 2.2An Overview of Previous Studies of the Impacts of Oil Price Shocks on Stock Markets and Macroeconomics Activities
Authors Periods Countries Variables Methodology Main Conclusions Apergis and Miller respond in a large way to oil market shocks
Bjørnland (2009) 1993-2005 Norway Oil Price;
Stock Price;
1978-2000 Singapore Oil price;
GDP;
Table 2.2 An Overview of Previous Studies of the Impacts of Oil Price Shocks on Stock Markets and Macroeconomics Activities (Continued)
Authors Periods Countries Variables Methodology Main Conclusions Cunado and Pérez
1975-2006 Iran Oil Price;
GDP;
1993-2007 Tunisia Oil price;
Inflation rate;
Papapetrou (2001) 1989-1999 Greece Oil Price;
Stock Return;
1947-2005 US Oil Price;
GDP;
Note: VDC denotes the variance decomposition.
Chapter 3 Methodology
In this chapter the threshold co-integration and multivariate threshold autoregrresive models will be introduced to address two issues. To more clearly express the utilization of methods, we outline the research process with respect to each issue in Figure 3.1.
Unit Root Test ADF Test
KPSS Test
Threshold Cointegration Threshold Cointegration Tests Threshold VECM
Issue Two Stock Market Returns Oil Price
Industrial Production Interest Rate
Impulse Response Analysis Variance Decomposition One-Regime VAR
Cointegration Test Maximum Eigenvalues Test Trace Test
Issue One
Disaggregated Energy Consumption GDP
Two-Regime VAR
Figure 3.1 Methodology Flow Chart
3.1 Unit Root Tests
A time series is a set of y observations, each one being record at a specific t time t with stochastic process. To aid in identification, we know that a covariance stationary series need to be satisfied:
(1) Exhibits mean reversion in that it fluctuates around a constant long-run mean.
(2) Has a finite variance that is time-invariant.
(3) Has a theoretical correlogram that diminishes as lag length increases.
On the other hand, a non-stationary series necessarily has permanent components. The mean and variance of non-stationary series are time-dependent.
To aid in identification of a non-stationary series, we know that:
(1) There is no long-run mean to which the series returns.
(2) The variance is time-dependent and goes to infinity as time approaches infinity.
(3) Theoretical autocorrelations do not decay, but the sample correlogram dies out slowly in finite samples.
Although the traditional OLS approach often assumes the time series are stationary and its disturbances all white noise. If we assume the non-stationary time series as stationary, it may cause spurious regression proposed by Granger and Newbold (1974). Its result may have higher coefficient of determinant and much significant t value, implying non-reject the null hypothesis and though meaningless under spurious regression. Before proceeding analysis, we should test whether these variables have the stationarity property. If the time series variable is stationary with d-times differencing, it can be called the integrated of order d and denoted as I(d).
We adopt two applicable unit root methods for examining the existence of unit roots.
3.1.1 Augmented Dickey Fuller (ADF) Test
Dickey and Fuller (1979) consider a autoregressive process AR(1),
1 1
However, simple unit root test described above is valid only if the series is an AR(1) process. If the series is correlated at higher order lags, the assumption of
white noise disturbances is violated. Dickey and Fuller (1981) make a parametric correction for higher order correlation by assuming that the
{ }
yt follows an AR(p) process and adjusting the test methodology, the general form can be expressed as follows:0 1 1 2 2
t t t p t p t
y =α α+ y− +α y− + +α y− + (1) ε To best understand the methodology of the augmented Dickey-Fuller test, add
and subtract αp t py− +1 to obtain: Continuing in this fashion, we get:
0 1 1 be used by the Akaike information criterion (AIC):
AIC=T ln(residual sum of squares)+2n (5) where n is the number of parameters estimated and T is the number of usable
observations.
Three ADF test actually consider three different regression equations that can be used to test for the presence of a unit root:
1 1 The differences between the three regressions concerns the presence of the deterministic elements α0 and α2t. The first considers a pure random walk plus lagged dependent variables, the second adds an intercept (or drift term), and the third includes an additional linear time trend. The parameter of interest in all the regression equations is γ. If the null hypothesis γ = cannot be rejected, then the 0 {yt} sequence contains a unit root; otherwise, this sequence is stationary.
3.1.2 The Kwiatkowski, Phillips, Schmidt and Shin (KPSS) Test
The standard conclusion that is drawn from this empirical evidence is that many or most aggregate economic time series contain a unit root. However, it is important to note that in this empirical work the unit root is the null hypothesis to be tested, and the way in which classical hypothesis testing is carried out ensures that the null hypothesis is accepted unless there is strong evidence against it. Therefore, an alternative explanation for the common failure to reject a unit root is simply that most economic time series are not very informative about whether or not there is a unit root, or equivalently, that standard unit root tests are not very powerful against relevant alternatives.
Kwiatkowski et al. (1992) use a parameterization which provides a plausible representation of both stationary and non-stationary variables and which leads naturally to a test of the hypothesis of stationarity. Specifically, they choose a
component representation in which the time series under study is written as the sum of a deterministic trend, a random walk, and a stationary error. The KPSS test differs from the other unit root tests described here in that the {yt} sequence is assumed to be (trend) stationary under the null. The KPSS statistic is based on the residuals from the OLS regression of yt on the exogenous variables xt:
t t t
y =x′δ ε+
The Lagrange Multiplier (LM) statistic can be defined as:
2 2
( ) /( o)
t
LM =
∑
S t T fwhere ( )S t is a cumulative residual function (i.e.,
1
( ) t ˆi, 1, 2, ,
i
S t ε t T
=
=
∑
= ), andf is an estimator of the residual spectrum at frequency zero. We point out that the o
estimator of δ used in this calculation differs from the estimators for δ used by detrended GLS since it is based on a regression involving the original data and not on the quasi-differenced data.
3.2 Cointegration Analysis
Co-integration theory is definitely the innovation in theoretical econometrics that has created the most interest among economists in the last decade. Co-integration is an econometric property of time series variables. If two or more time series variables are non-stationary, but a linear combination of them is stationary, then the series are said to be co-integrated.
The Johansen co-integration method is provided by Johansen (1988) and Johansen and Juselius (1990). This procedure applying maximum likelihood to the vector autoregressive (VAR) model, and consider the relationships among more than two variables. Let yt denotes an (n× vector. The maintained hypothesis is that 1) yt follows a VAR(P) in levels and all of the elements for yt are I(1) process. In
addition, the errors are Gaussian.
1 1 2 2
+ + + + + , 1, 2, ,
t t t p t p t
y =µ Π x− Π x− Π x− ε t= T (9)
where μis constant term and εti i d. . .∼ N(0, )Ω . Moreover, VAR(p) in levels can be written as:
1 1 2 2 1 1 1
t t t p t p t t
y µ ς y− ς y− ς − y− + ςy− ε
∆ = + ∆ + ∆ + + ∆ + + (10) where ς = −(In − Π − Π − − Π = −Π1 2 p) (1)
1 2
( ) 1, 2 , 1
i In i i p
ς = − − Π − Π − − Π = −
Suppose that each individual variable yit is I(1) and linear combinations of yt are stationary. That implies ς can be showed as
ς = −αβ′
whereβis the cointegrating matrices, and α is the adjustment coefficients for both α and β (r n× matrices. The number of cointegrating relations relies on the ) rank of ς , and the rank of ς is :
(1) rank( )ς = , n ς is full rank means that all components of yt is a stationary process.
(2) rank( ) 0ς = , ς is null matrix meaning that there is no co-integration relationships.
(3) 0 rank( )< ς = < , the variables for yr n t are co-integrated and the number of cointegrating vectors is r.
To determine the number of co-integrating vectors, Johansen proposes two different likelihood ratio tests of the significance of these canonical correlations and thereby the reduced rank of the Π matrix: the trace test and maximum eigenvalue test, shown as follows:
(1) Trace test:
0: rank( )
H ς ≤ , i.e., there are at most r cointegrating vectors r
1: rank( ) H ς > r The test statistic is
1
ln(1 ˆ)
n
trace i
i r
λ T λ
= +
= −
∑
− ,where r is the cointegrating vector, T is the sample size, and λˆi is the ith largest canonical correlation. The statistic has a limit distribution which can be expressed in terms of a (n-r)-dimensional Brownian motion.
(2) Maximum eigenvalues test:
H : there are r co-integrating vectors 0
H : there are 1 r+ co-integrating vectors 1
The test statistic is λmax = −Tln(1−λˆr+1). If the absolute value of eigenvalue, ˆi
λ , is larger, then the test statistic will be higher and tend to reject the null hypothesis.
Neither of these test statistics follows a chi-square distribution in general; asymptotic critical values can be found in Johansen and Juselius (1990). Since the critical
Neither of these test statistics follows a chi-square distribution in general; asymptotic critical values can be found in Johansen and Juselius (1990). Since the critical