See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/266370484

Electricity Consumption, Employment and Real Income in Australia: Evidence

from Multivariate Granger Causality Tests

Article  in  Energy Policy · October 2005

DOI: 10.1016/j.enpol.2003.11.010

CITATIONS READS

390 2,670

2 authors:

Paresh Kumar Narayan Russell Smyth

Deakin University Monash University (Australia)
359 PUBLICATIONS   12,722 CITATIONS    520 PUBLICATIONS   8,176 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

The effects of the motor vehicle industry on employment and research innovation in Australia View project

Asset Pricing and Price Discovery View project

All content following this page was uploaded by Russell Smyth on 06 October 2017.

The user has requested enhancement of the downloaded file.

 ARTICLE IN PRESS

Energy Policy 33 (2005) 1109–1116

Electricity consumption, employment and real income in Australia

evidence from multivariate Granger causality tests
Paresh Kumar Narayan, Russell Smyth*
Department of Economics, Monash University, P.O. Box 11E, 3800 Vic., Australia

Abstract

This paper examines the relationship between electricity consumption, employment and real income in Australia within a
cointegration and causality framework. We find that electricity consumption, employment and real income are cointegrated and that
in the long-run employment and real income Granger cause electricity consumption, while in the short run there is weak
unidirectional Granger causality running from income to electricity consumption and from income to employment.
r 2003 Elsevier Ltd. All rights reserved.

Keywords: Electricity consumption; Causality; Bounds test; Cointegration

1. Introduction electricity consumption could lead to a fall in income

and/or employment.
The electricity sector is one of Australia’s major Methodologically, a contribution of the paper is that
industries. In 1997–1998 it generated $13 billion in sales we employ the bounds testing approach to cointegra-
to final consumers, which accounted for around 1.6% of tion, within an autoregressive distributed lag (ARDL)
GDP, and employed 32,000 people (ESAA, 1999, ABS framework suggested by Pesaran and others (Pesaran
Cat. No. 5206). Table 1 shows growth of electricity and Shin, 1999; Pesaran et al., 2001). With the exception
consumption and GDP since the beginning of the 1970s. of Fatai et al. (2001), who use this approach in a study
The average growth in electricity consumption over this for New Zealand, the bounds test has not been used in
period has been 4.8%, which has outpaced growth in the energy consumption-growth literature. The bounds
real GDP, which has been 3.4%. The purpose of this test approach has the advantage that it has better small
paper is twofold. The first objective is to determine sample properties than other popular methods of
whether there is a stationary, long-run equilibrium cointegration such as the Engle and Granger (1987)
relationship between electricity consumption, employ- and Johansen and Juselius (1990) test that have been
ment and income in Australia. The second is to examine widely used in the past.
the temporal causality between these variables. From a Since the widespread adoption of cointegration
policy viewpoint, the direction of causality between techniques, the evidence on whether there is a long-run
these variables has important implications (Asafu- equilibrium relationship between energy consumption
Adjaye, 2000, pp. 616–617; Ghosh, 2002, p. 125). If and output is mixed. Some studies have found that
unidirectional Granger causality runs from income or energy consumption and output are cointegrated (see
employment to electricity consumption or if there is no e.g. Nachane et al., 1988; Masih and Masih, 1996 (for
causality in either direction this implies that electricity India, Malaysia and Pakistan)), but many have failed to
conservation policies would not affect economic growth. find evidence of a long-run equilibrium relationship (see
However, if unidirectional causality runs from electri- e.g. Yu and Jin, 1992; Cheng and Lai, 1997). In a
city consumption to income or employment reducing relatively recent study Soytas and Sari (2003) found a
cointegrating vector for only seven out of 16 countries in
a study of the G-7 countries and nine emerging markets
*Corresponding author. Tel.: +613-9905-1560; fax: +613-9905- in Asia, Europe and South America. The literature is
5476. equally divided about the relationship between energy
E-mail address: russell.smyth@buseco.monash.edu.au (R. Smyth). consumption and employment. Some studies support

0301-4215/$ - see front matter r 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/j.enpol.2003.11.010
 ARTICLE IN PRESS
1110 P.K. Narayan, R. Smyth / Energy Policy 33 (2005) 1109–1116

Table 1 implement energy conservation policies designed to

Growth of electricity generated and real GDP, Australia 1969/70– reduce greenhouse emissions and to move towards
1998/99 generating electricity from renewable resources. In
Growth in electricity Growth in 1997, the Australian Prime Minister described this aim
consumption real GDP as follows: ‘‘The possibilities for fuel substitution and
1969/70–1974/75 3.9 3.7 innovation will be enhanced as we continue to accelerate
1976/77–1979/80 2.3 2.5 the process of energy market reform’’ (Howard, 1997).
1980/81–1984/85 5.2 3.7 Prior to the Kyoto conference, the Australian govern-
1985/86–1989/90 4.8 3.2 ment forecast that competition policy reforms to the
1990/91–1994/95 5.9 3.1
energy market would result in greenhouse emission
1995/96–1998/99 6.6 4.1
1969/70–1998/99 4.8 3.4 reductions of 14 million t/annum (Hamilton and Den-
niss, 2001, p. 16). Measures to achieve this have included
Source: ABS Cat. No. 5206.0, Abbott (2001).
mandatory renewable energy targets and some financial
incentives to promote the use of renewable energy
technologies (for an extensive discussion see Bradbrook
the neutrality thesis (Yu et al., 1988; Erol and Yu, 1989). and Wawryk, 2002).
Others suggest that unidirectional causality runs from To this point, competition policy has not delivered in
employment to energy consumption (Murray and Nan, this area. In fact greenhouse emissions from the
1992) or that causality runs from energy consumption to electricity sector have been increasing throughout the
employment (Akarca and Long, 1979). 1990s (Hamilton and Denniss, 2001). In 1998, net
There is only one study of this sort of which we are carbon dioxide emissions from electricity generation
aware for Australia, which is by Fatai et al. (2001a). were 30% higher than in 1990 (AGO, 2000). The
This study differs from Fatai et al. (2001a) in two amount of electricity produced from renewable energy
important respects. First Fatai et al. (2001a) use the in Australia is very small. More than 92% of Australia’s
Johansen approach to cointegration, but, unlike their electricity is produced by fossil fuels—coal and natural
later paper focusing solely on New Zealand (Fatai et al., gas—and it is expected that fossil fuels will still account
2001), do not use the bounds test approach. Second, for 90% of electricity generation in 2010 (ESAA, 2002).
more importantly, these authors test for Granger To put this in perspective, if Australia’s 8% Kyoto
causality between total final energy consumption and target was applied equally across all sectors then the
its components and real GDP using a bivariate frame- electricity sector’s net emissions could increase by 10.3
work. We use a multivariate cointegration and vector million t over the period 1990–2010. However, Hamilton
error-correction approach. A common view in the and Denniss (2001) estimate that competition policy,
literature now is that studies which focus only on two through focusing on short-term cost minimization, was
variable cases may be biased due to the omission of responsible for an increase in greenhouse emissions from
relevant variables (see e.g. Chang et al., 2001, p. 1046; the electricity sector of 11 million t in 1998, the year that
Stern, 2000, p. 268). Thus, beginning with Stern (1993), the National Electricity Market was introduced, alone.
the most common approach in recent studies is to The Australian government is putting the onus on the
employ Granger causality tests using some sort of electricity industry to increase the contribution of
multivariate approach. Studies typically include in renewable energy sources in Australia’s energy mix
addition to energy and output variables one or more through setting mandatory renewable energy targets. In
other variables. Recent studies have included a variable response, the Electricity Supply Association of Australia
for employment (Chang et al., 2001; Fatai et al., 2001) is urging the Commonwealth and State governments to
or energy prices (Masih and Masih, 1997) or capital and implement more effective electricity end-use efficiency
labour inputs (Stern, 1993). programs. It claims that ‘‘such programs have the
The paper is set out as follows. The next section potential to reduce demand in 2010 by as much as
briefly outlines attempts at energy conservation in 20,000 GW h a year and to cut greenhouse gas emissions
Australia. Section three outlines the data and econo- by 20 million t a year—three times the abatement
metric methodology. The results are presented in section targeted under the renewable energy scheme’’ (ESAA,
four. The final section concludes. 2002).
There has been some attempt to curtail electricity use
at the state level through the implementation of demand
2. Energy conservation policies in Australia management programs. This has been the main response
of the Victorian government to the problem of shortages
Since the beginning of the 1990s the electricity in that state. In 1999/2000, the states of Victoria
industry has undergone a series of structural reforms. and South Australia experienced power outages, due
One of the main objectives of the reforms has been to to excess demand. A general shortage of electricity
 ARTICLE IN PRESS
P.K. Narayan, R. Smyth / Energy Policy 33 (2005) 1109–1116 1111

generation resulting from the failure of regulatory Model C takes the following form:
reform is regarded as one of the main causes of the
California electricity crisis (Borenstein, 2002). Sparked Dyt ¼ k þ fyt1 þ bt þ y1 DUt
by fears of what happened in California and acting in X
k
response to a Victoria Security of Electricity Taskforce þ g1 DTt þ dj Dytj þ et : ð2Þ
Report (2000), in 2000–2001 the Victorian government j¼1
initiated a program of encouraging voluntary reductions
in electricity consumption by industry during peak Here, D is the first difference operator, et is a white noise
periods to ensure security of supply. As Australia comes disturbance term with variance s2 ; and t ¼ 1;y,T is an
under increasing international pressure to ratify the index of time. The Dyt2j terms on the right-hand side of
Kyoto protocol it is important to know what the effect of Eqs. (1) and (2) allow for serial correlation and ensure
further implementation of energy conservation policies that the disturbance term is white noise. DUt is an
along these lines will be on employment and output. indicator dummy variable for a mean shift occurring at
time TB and DTt is the corresponding trend shift
variable, where
3. Data and econometric methodology (
1 if t > TB;
DUt ¼
3.1. Data 0 otherwise:

The study employs annual time series data from 1966 and
to 1999. Total electricity consumption per capita (
(measured in kWh per capita), real GDP per capita t  TB if t > TB;
DTt ¼
(1995=100) and an index of manufacturing sector 0 otherwise:
employment are from the International Energy Agency
and IMF International Financial Statistics. The period Zivot and Andrews (1992) provide asymptotic critical
for the analysis was dictated by data availability. values for their test, but it is well known that these are
Following the extant literature, all variables were not reliable in small sample sizes. Thus, we calculate
transformed into natural logs because this helps to exact critical values for our sample size of 34 observa-
induce stationarity in the variance–covariance matrix tions for model A and model C following Zivot
(Chang et al., 2001; Fatai et al., 2001). and Andrews’ (1992) methodology. We estimate an
ARMA (p; q) model for each Dyit ; with p and q
3.2. Unit root tests selected according to the Schwarz Bayesian criterion.
The implied ARMA process is then used as the
A three-stage procedure was followed to test the data generating process for generation of 5000 34-
direction of causality. In the first stage the order of observation series under the null hypothesis of a unit
integration was tested using the Augmented Dickey– root with no structural breaks. We obtain a minimum
Fuller (ADF), Phillips–Perron (PP) and Zivot and ADF statistic for each of the 5000 series. The
Andrews’ (1992) unit root tests. While one of the critical values are then constructed from this empirical
advantages of the bounds test for cointegration is that it distribution.
can be applied irrespective of whether the variables are
integrated of order zero Ið0Þ or integrated of order 1
3.3. Cointegration
Ið1Þ; to implement the Granger causality tests all
variables must be Ið1Þ: The lag length in the ADF test
The second stage involves testing for the existence of a
was selected to minimize the Schwarz Bayesian criterion,
long-run equilibrium relationship between electricity
while the bandwidth for the PP test was selected with the
consumption, real income and employment growth
Newey-West Bartlett kernel.
using the bounds test. This involves investigating the
We used two versions of the Zivot and Andrews
existence of a long-run relationship using the following
(1992) test. These are in (Zivot and Andrews (1992),
unrestricted error correction model (UECM):
terminology) model A, which allows for a one break in
the intercept of the trend function and model C, which X
n X
n
allows for one break in intercept and slope. D ln ECt ¼ aoEC þ biEC D ln ECti þ ciEC D ln GNIti
Model A has the following form: i¼1 i¼1
X
n
X
k þ diEC D ln EMti þ s1EC ln ECt1 þ s2EC ln GNIt1
Dyt ¼ k þ fyt1 þ bt þ y1 DUt þ dj Dytj þ et : ð1Þ i¼1

j¼1 þ s3EC ln EMt1 þ e1t ; ð3Þ

 ARTICLE IN PRESS
1112 P.K. Narayan, R. Smyth / Energy Policy 33 (2005) 1109–1116

X
n X
n
relationship (this term is not included if the variables are
D ln GNIt ¼ aoGNI þ biGNI D ln GNIti þ ciGNI D ln ECti
i¼1 i¼1 not cointegrated) and e1t ; e2t and e3t are serially
X
n independent random errors with mean zero and finite
þ diGNI D ln EMti þ s1GNI ln GNIt1 covariance matrix. The dependent variable is regressed
i¼1 against past values of itself and other variables. The lag
þ s2GNI ln ECt1 þ s3GNI ln EMt1 þ e1t ; length P is based on the Schwarz Bayesian criterion.
ð4Þ
X
n X
n
D ln EMt ¼ aoEM þ biEM D ln EMti þ ciEM DlnGNIti 4. Empirical results
i¼1 i¼1
X
n
4.1. Order of integration and cointegration
þ diEM D ln ECti þ s1EM ln EMt1
i¼1
The results for the ADF and PP unit root tests for
þ s2EM ln GNIt1 þ s3EM ln Ect1 þ e1t : electricity consumption per capita (ln EC), real income
ð5Þ per capita (ln GNI) and employment (ln EM) are
Here, D is the first difference operator, ln EC is the log reported in Table 2. The ADF and Phillips–Perron
of per capita electricity consumption, ln GNI is the log tests give the same results. The null hypothesis that the
of per capita real income and ln EM is the log of the series contain a unit root cannot be rejected for any of
manufacturing sector employment index. The F test is the series in levels at the 5% level, but when the data are
used to determine whether a long-run relationship exists first differenced, the null of nonstationarity can be
between the variables through testing the significance of rejected for all series at the 5% level. This indicates that
the lagged levels of the variables. With small sample electricity consumption per capita, employment and real
sizes the relevant critical values potentially deviate income per capita are Ið1Þ:
substantially from the critical values reported in Pesaran The results of the Zivot and Andrews (1992) model A
et al. (2001). Thus, we calculate exact critical value and model C unit root tests and corresponding exact
bounds for T ¼ 34 with two regressors, based on 40,000 critical values are reported in Table 3. They find no
replications for the F -statistic. If the computed F additional evidence against the unit root hypothesis
statistics fall outside the critical bounds, it follows that relative to the unit root tests without a structural break.
a conclusive decision can be made regarding cointegra- In each case we are unable to reject the unit root null
tion without knowing the order of integration of the hypothesis at the 10% level or better, confirming that
regressors. the series are Ið1Þ: The break dates are statistically
significant for model A in each case. With electricity
3.4. Granger causality consumption only the break in the slope of the trend is
statistically significant in model C and neither the break
The third stage involves constructing standard Gran- in the intercept nor slope is statistically significant in
ger-type causality tests augmented with a lagged error model C for employment or real income. The statisti-
correction term in the event that the series are cally significant breaks are 1971 and 1990 (electricity
cointegrated. Where such a long-run relationship exists, consumption), 1980 (employment) and 1983 (real
Granger causality tests involve specifying a multivariate income). While 1980 was a period of strong labour
pth-order vector error correction model (VECM) as
follows: Table 2
2 3 2 3 Augmented Dickey–Fuller (ADF) and Phillips–Perron (PP) tests for
ln ECt a1 unit roots
6 7 6 7
ð1  LÞ4 ln GNIt 5 ¼ 4 a2 5 Variables ADF statistic CV (LL) PP statistic CV (BW)
ln EMt a3 Ln ECt 1.813 3.552 (0) 2.123 3.559 (7)
2 32 3
b11i b12i b13i ln ECti Ln GNIt 2.839 3.558 (1) 2.045 3.553 (1)
Xp
6 76 7 Ln ECt 2.365 3.553 (0) 2.552 3.553 (2)
þ ð1  LÞ4 b21i b22i b23i 54 ln GNIti 5
Dln ECt 3.625 2.957 (0) 3.625 2.957 (0)
i¼1
b31i b32i b33i ln EMti Dln GNIt 2.973 2.957 (0) 2.713 2.617 (8)
2 3 2 3 Dln ECt 5.007 2.957 (0) 5.007 2.957 (0)
y e1t
6 7 6 7 Notes: CV stands for critical values, which are at the 5% level. The
þ 4 W 5½ECTt1  þ 4 e2t 5: ð6Þ
critical values are calculated from MacKinnon (1991). LL stands for
c e3t lag length and BW stands for bandwidth. The lag lengths are selected
using the Schwarz Bayseian criterion while the bandwidth is selected
In addition to the variables defined above, (12L) is using the Newey–West Bartlett kernel. Both tests were conducted
the difference operator, ECTt21 is the lagged error- including an intercept and linear deterministic trend. Tests using an
correction term derived from the long-run cointegrating intercept only were quantitatively similar.
 ARTICLE IN PRESS
P.K. Narayan, R. Smyth / Energy Policy 33 (2005) 1109–1116 1113

Table 3
Zivot–Andrews test for unit roots in the presence of one structural break

Electricity consumption Real income Employment

Model A Model C Model A Model C Model A Model C

TB 1990 1971 1983 1979 1980 1980

f 0.2175 0.2136 0.9350 0.3222 0.9787 0.9621
(3.4530) (3.0078) (5.0609) (4.3465) (4.9066) (4.6926)
y 0.0330 0.0333 0.1841 0.1202 0.0682 0.0642
(2.4727) (0.0033) (3.5950) (0.0059) (3.2119) (0.0016)
g — 2.1156 — 3.6400 — 2.8084
(6.0910) (1.3454) (0.5323)
Lag length 1 0 6 0 2 2

Exact critical values for tf

1% 6.8233 6.8568 6.9783 6.7508 6.7878 6.6471
5% 5.8716 5.7933 5.9359 5.8404 5.7924 5.8540
10% 5.4353 5.3787 5.4539 5.3865 5.3657 5.4378

Notes:  denotes statistical significance at the 1% level. The critical values for the structural break dummy variables follow the asymptotic standard
normal distribution. The critical values for tf are calculated based on 5000 replications of a Monte Carlo simulation as described in the text.

Table 4 within the error correction mechanism (ECM). The

Bounds tests for cointegration F-statistics on the lagged explanatory variables of the
F-statistics Exact 5% critical Exact 1% ECM indicates the significance of the short-run causal
value bounds critical value effects. The t-statistics on the coefficients of the lagged
bounds error-correction term indicates the significance of the
I(0) I(1) I(0) I(1) long-run causal effect. Beginning with the short-run
FEC ðECjGNI; EMÞ ¼ 5:408 3.990 4.538 4.943 6.128 effects, real income is significant at the 10% level in the
FGNI ðGNIjEC; EMÞ ¼ 1:478 electricity consumption and employment equations, but
FEM ðEMjEC; GNIÞ ¼ 3:347 neither electricity consumption nor employment is
Notes: Critical values are calculated using stochastic simulations for significant. This suggests that in the short-run there is
T ¼ 34 and two regressors based on 40,000 replications. weak unidirectional Granger causality running from
income to electricity consumption and income to
employment, but there is neutrality between electricity
growth, 1983 and 1990 were periods of recession in consumption and employment.
Australia. Turning to the t-statistic on the coefficient of
The results of the bounds test for cointegration, the lagged error-correction term, the coefficient on the
together with exact critical values for T ¼ 34; are lagged error-correction term is significant in the
reported in Table 4. The bounds test indicates that electricity consumption equation at 1% with a negative
cointegration is only present when electricity consump- sign, which confirms the result of the bounds test for
tion is the dependent variable. This is because cointegration. We do not include a lagged error-
FEC ðECjGNI; EMÞ is higher than the upper bound correction term when either employment or real income
critical value at the 5% critical value. However, the is the dependent variable because the variables are not
bounds test indicates that when real income and cointegrated in these cases. The results from the Granger
employment are the dependent variables FGNI ðGNIj causality test implies that changes in electricity con-
EC; EMÞ and FEM ðEMjEC; GNIÞ are lower than the sumption are a function of disequilibrium in the
lower bound critical value at the 5% level. Therefore, cointegrating relationship, but this is not the case for
there is no cointegration when these variables are employment and income. In other words, in the long run
treated as the dependent variables, meaning that there both employment and income Granger cause electricity
is one single long-run relationship. consumption, meaning that causality runs interactively
The existence of a cointegrating relationship among through the error correction term from employment and
electricity consumption, employment and real income income to electricity consumption.1
suggests that there must be Granger causality in at least
one direction, but it does not indicate the direction of 1
Fatai et al. (2001) use the ARDL approach to infer causality. We
temporal causality between the variables. Table 5 also used the ARDL approach to infer causality and this gave the same
examines short-run and long-run Granger causality results as the Granger causality tests, which are reported.
 ARTICLE IN PRESS
1114 P.K. Narayan, R. Smyth / Energy Policy 33 (2005) 1109–1116

Table 5 1.6
Results of Granger causality F-tests
1.2
Dependent D ln ECt D ln GNIt D ln EMt ECTt (t-statistic)
variable
0.8
D ln ECt — 2.9682 1.2445 0.0415
(0.0514) (0.2881) (4.5669)
0.4
D ln GNIt 0.8048 — 1.3221 —
(0.4472) (0.2666)
D ln EMt 0.1533 2.7710 — — 0.0
(0.8578) (0.0624)
-0.4
Note: () denotes statistical significance at the 1% and 10% levels 78 80 82 84 86 88 90 92 94 96 98
respectively. Figures in parenthesis below D ln ECt ; D ln GNIt and
D ln EMt are probability values. CUSUM of Squares 5% Significance

Fig. 1. Plot of the CUSUM of squares when ln ECt is the dependent

variable.
4.2. Constancy of cointegration space

One problem with time series regression models is that Table 6

the estimated parameters may change over time. Hansen test for parameter stability
Unstable parameters can result in model misspecifica- Tests Test statistic Probability value
tion and, if left undetected, have the potential to bias the
ln EC is the dependent variable
results. To account for this, here, we examine whether LC 0.0837 >0.20
the estimated elasticities are stable over time. To do this MeanF 0.8256 >0.20
we use the Pesaran and Pesaran (1997) and Hansen SupF 1.6386 >0.20
(1992) tests for parameter instability. Note: The test program is available from http://www.ssc.wisc.edu/
The Pesaran and Pesaran (1997) test amounts to bhansen/.
estimating the error correction models through taking
each differenced variable as a dependent variable
together with the lagged error correction term. This (1997) CUSUMSQ test. This indicates that the structure
exercise is only possible for the equation, which has of the parameters have not diverged abnormally over
electricity consumption as the dependent variable, given the period of the analysis.
that this is the only equation for which there is
cointegration in this model. The error correction term 4.3. Forecasting electricity consumption
is calculated from the long run cointegrating vector.
Once the model has been estimated, Pesaran and We finish through using the Box–Jenkins autoregres-
Pesaran (1997) suggest applying the cumulative sum of sive integrated moving average (ARIMA) model to
recursive residuals (CUSUM) and the CUSUM of forecast per capita electricity consumption for Australia
square (CUSUMSQ) tests proposed by Brown et al. for the decade 2000–2010. To gauge the accuracy of the
(1975) to assess the parameter constancy. The models forecasting performance we also calculate in-sample
were estimated by OLS and the residuals subjected to forecasts of electricity consumption and use the Theil
the CUSUMSQ test. Fig. 1 plots the CUSUMSQ index and the mean absolute percentage error as
statistic when electricity consumption is the dependent measures of forecasting accuracy. Since these methods
variable. The results indicate no instability in the are widely known, to conserve space, we do not repeat
coefficients as the plot of the CUSUMSQ statistics are them here.
confined within the 5% critical bounds of parameter The forecasts of per capita electricity consumption
stability. over the 2000–2010 period together with the actual
As a further check on parameter stability the consumption data from 1990 to 1999 are presented in
parameter non-constancy tests for I(1) processes advo- Table 7. The actual, fitted and residuals of electricity
cated by Hansen (1992) were employed. Hansen (1992) consumption are plotted in Fig. 2. Visual inspection
proposes three tests—SupF, Meanf, and LC —which all suggests that our forecasts are highly accurate. Upon
have the same null hypothesis that the parameters are using robust forecasting accuracy means we find that the
stable. The Hansen suite of test results, together with Theil index is very small (0.0098) and that the MAPE is
their probability values, are reported in Table 6. They 1.58. Given that the closer the Theil index moves to zero
indicate parameter stability, since the probability values the more accurate are the forecasts and that if the
for each test are greater than 0.05. The Hansen test MAPE is less than 10%, this indicates accurate forecasts
statistics corroborate those of the Pesaran and Pesaran our forecasting performance is strong.
 ARTICLE IN PRESS
P.K. Narayan, R. Smyth / Energy Policy 33 (2005) 1109–1116 1115

Table 7 et al.’s (2001a) finding of unidirectional causality

Forecasts of per capita electricity consumption for Australia, 1990– running from real GDP to electricity consumption
2010 within a bivariate cointegration/vector error-correction
Year Actual Forecasts framework.
1990 7572 7253.12
1991 7621 7436.62
1992 7601 7619.27 Acknowledgements
1993 7750 7801.07
1994 7877 7982.04
We thank Hashem Pesaran and Yongcheol Shin for
1995 8021 8162.17
1996 8135 8341.48 sharing the GAUSS codes they used to produce the
1997 8303 8519.95 original set of critical values for the bounds testing
1998 8703 8697.61 approach to cointegration as reported in Pesaran et al.
1999 8884 8874.44 (2001). We also thank Heather Anderson, Mita Bhatta-
2000 — 9050.46 charya and an anonymous referee for suggestions on an
2001 — 9225.67
2002 — 9400.07 earlier version of this paper. We alone are responsible
2003 — 9573.67 for the paper’s contents.
2004 — 9746.46
2005 — 9918.46
2006 — 10089.67
References
2007 — 10260.08
2008 — 10429.71
2009 — 10598.56 Abbott, M., 2001. Market power and the problem of security of supply
2010 — 10766.63 in the National Electricity Market. Economic Papers 20 (3), 1–14.
Akarca, A.T., Long, T.V., 1979. Energy and employment: a time series
analysis of the causal relationship. Resources Energy 2, 151–162.
Asafu-Adjaye, J., 2000. The relationship between energy consumption,
energy prices and economic growth: time series evidence from
10000 Asian developing countries. Energy Economics 22, 615–625.
Australian Greenhouse Office (AGO), 2000. National Greenhouse Gas
8000
Inventory: Analysis of Trends and Greenhouse Indicators 1990 to
300
6000 1998. AGO, Canberra.
200
Borenstein, S., 2002. The trouble with electricity markets: under-
4000 standing California’s restructuring disaster. Journal of Economic
100 Perspectives 16, 191–211.
2000 Bradbrook, A.J., Wawryk, A.S., 2002. Government initiatives
0
promoting renewable energy for electricity generation in Australia.
-100 University of New South Wales Law Journal 25, 124–159.
-200 Brown, R.L., Durbin, J., Evans, J.M., 1975. Techniques for testing the
1970 1975 1980 1985 1990 1995 constancy of regression relationships over time. Journal of the
Residual Actual Fitted Royal Statistical Society, Series B 37, 149–192.
Chang, T., Fang, W., Wen, L.-F., 2001. Energy consumption,
Fig. 2. Actual, fitted and residuals from ARIMA (1,1,1) model of per employment, output and temporal causality: evidence from Taiwan
capita electricity consumption. based on cointegration and error-correction modelling techniques.
Applied Economics 33, 1045–1056.
Cheng, B.S., Lai, T.W., 1997. An investigation of cointegration and
5. Conclusions causality between energy consumption and economic activity in
Taiwan. Energy Economics 19, 435–444.
Our main findings are that electricity consumption, Electricity Supply Association of Australia (ESAA), 1999. Electricity
Australia 1999. Gotham Press, Sydney.
employment and income are cointegrated and that in the Electricity Supply Association of Australia (ESAA), 2002. Strong
long run unidirectional causality runs from employment electricity demand continues. Media release June 13.
and income to electricity consumption. This has Engle, R.F., Granger, C.W.J., 1987. Cointegration and error correc-
important consequences for energy conservation policies tion representation: estimation and testing. Econometrica 55,
as Australia grapples with reform to the electricity 251–276.
Erol, U., Yu, E.S.H., 1989. Spectral analysis of the relationship
industry and measures to reduce greenhouse gas between energy consumption, employment and business cycles.
emissions. The results in this paper suggest that Resources and Energy 11, 395–412.
electricity conservation policies through efficiency im- Fatai, K., Oxley, L., Scrimgeour, F., 2001. Energy consumption and
provement measures and demand management policies employment in New Zealand: searching for causality. Unpublished
manuscript.
which are designed to reduce the wastage of electricity
Fatai, K., Oxley, L., Scrimgeour, F., 2001a. Modeling the causal
and curtail consumption can be implemented without relationship between energy consumption and GDP in New
having an adverse effect on Australia’s employment or Zealand, Australia, India, Indonesia, The Philippines and Thai-
real income. This conclusion is consistent with Fatai land. MODSIM 2001, Australian National University, Canberra.
 ARTICLE IN PRESS
1116 P.K. Narayan, R. Smyth / Energy Policy 33 (2005) 1109–1116

Ghosh, S., 2002. Electricity consumption and economic growth in Nachane, D.M., Nadkarni, R.M., Karnik, A.V., 1988. Cointegration
India. Energy Policy 30, 125–129. and causality testing of the energy–GDP relationship: a cross-
Hamilton, C., Denniss, R., 2001. Generating emissions? The impact of country study. Applied Economics 20, 1511–1531.
microeconomic reform on the electricity industry. Economic Pesaran, M.H., Pesaran, B., 1997. Working with Microfit 4.0:
Papers 20 (3), 15–28. Interactive Econometric Analysis. Oxford University Press, Oxford.
Hansen, B., 1992. Tests for parameter instability in regressions with Pesaran, M.H., Shin, Y., 1999. An autoregressive distributed lag
I(1) processes. Journal of Business and Economics Statistics 10, modelling approach to cointegration analysis. In: Strom, S. (Ed.),
321–335. Econometrics and Economic Theory in the 20th Century: The
Howard, J., 1997. Safeguarding the future: Australia’s response to Ragnar Frisch Centennial Symposium. Cambridge University
climate change. Statement by the Prime Minister, November 20. Press, Cambridge.
Johansen, S., Juselius, K., 1990. Maximum likelihood estimation Pesaran, M.H., Shin, Y., Smith, R., 2001. Bounds testing approaches
and inference on cointegration with applications to the demand to the analysis of level relationships. Journal of Applied Econo-
for money. Oxford Bulletin of Economics and Statistics 52, metrics 16, 289–326.
169–210. Soytas, U., Sari, R., 2003. Energy consumption and GDP: causality
MacKinnon, J., 1991. Critical values for cointegration tests. In: relationship in G-7 countries and emerging markets. Energy
Engle, R.F., Granger, C.W.J. (Eds.), Long-run Economic Re- Economics 25, 33–37.
lationships: Readings in Cointegration. Oxford University Press, Stern, D.I., 1993. Energy growth in the USA: a multivariate approach.
Oxford. Energy Economics 15, 137–150.
Masih, A.M.M., Masih, R., 1996. Energy consumption, real income Stern, D.I., 2000. A multivariate cointegration analysis of the role of
and temporal causality: results from a multi-country study based energy in the US macroeconomy. Energy Economics 22, 267–283.
on cointegration and error-correction modelling techniques. Victoria Security of Electricity Taskforce Report, 2000. East Mel-
Energy Economics 18, 165–183. bourne.
Masih, A.M.M., Masih, R., 1997. On the temporal causal relationship Yu, E.S.H., Jin, J.C., 1992. Cointegration tests of energy consumption,
between energy consumption, real income and price: some new income and employment. Resources Energy 14, 259–266.
evidence from Asian NICs based on a multivariate cointegration/ Yu, E.S.H., Choi, P.C.Y., Choi, J.Y., 1988. The relationship between
vector error-correction approach. Journal of Policy Modeling 19, energy and employment: a re-examination. Energy Systems Policy
417–440. 11, 287–295.
Murray, D.A., Nan, G.D., 1992. The energy consumption and Zivot, E., Andrews, D., 1992. Further evidence of the great crash, the
employment relationship: a clarification. Journal of Energy oil-price shock and the unit-root hypothesis. Journal of Business
Development 16, 121–131. and Economic Statistics 10, 251–270.

View publication stats