Available online at www.sciencedirect.

com

Energy Economics 30 (2008) 2302 – 2313

www.elsevier.com/locate/eneco

The relationship between disaggregate energy

consumption and industrial production in the
United States: An ARDL approach
Ramazan Sari a,b , Bradley T. Ewing c , Ugur Soytas b,⁎
a
Abant Izzet Baysal University, Department of Economics, 14280 Bolu, Turkey
b
Middle East Technical University, Department of Bus. Admin., 06531 Ankara, Turkey
c
Texas Tech Univ., Rawls College of Business, Lubbock, TX 79409-2101, USA
Received 11 May 2007; received in revised form 9 October 2007; accepted 14 October 2007
Available online 22 October 2007

Abstract

We re-examine the relationship between disaggregate energy consumption and industrial output, as well
as employment, in the United States using the autoregressive distributed lag (ARDL) approach developed
by Pesaran and Pesaran [Pesaran, M.H., Pesaran, B., 1997. Working with Microfit 4.0. Camfit Data Ltd,
Cambridge] and Pesaran, Shin and Smith [Pesaran, M.H., Shin, Y., Smith, R.J., 2001. Bounds testing
approaches to the analysis of level relationships. Journal of Applied Econometrics 16; 289–326] In
particular, we focus attention on the following energy consumption variables: coal, fossil fuels,
conventional hydroelectric power, solar energy, wind energy, natural gas, wood, and waste. The sample
period covers 2001:1–2005:6. Our results imply that real output and employment are long run forcing
variables for nearly all measures of disaggregate energy consumption.
© 2007 Elsevier B.V. All rights reserved.

JEL classification: Q 42; Q 43

Keywords: Disaggregate energy consumption; Industrial output; Autoregressive distributed lag

1. Introduction

Energy plays a unique role in the supply chain as it is both a final good for end-users as well
as an input into the production processes of many businesses. The decisions households and

⁎ Corresponding author. Tel.: +90 312 210 2048; fax: +90 312 2107962.
E-mail addresses: [email protected] (R. Sari), [email protected] (B.T. Ewing), [email protected] (U. Soytas).

0140-9883/$ - see front matter © 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.eneco.2007.10.002
 R. Sari et al. / Energy Economics 30 (2008) 2302–2313 2303

businesses must make regarding energy use are influenced by, and have implications for, short run
changes in economic activity as well as longer term trends. For this reason, considerable attention has
been placed on estimating the relationship between energy consumption and output. In the United
States total energy expenditures account for more than 7% of Gross Domestic Product (Economic
Report of the President, 2006). In fact, while petroleum comprises the largest imported energy source
for the US, diversification of energy is generally considered important since it would lessen the
dependence on foreign energy and thus dampen the effects of oil disruptions. Key strategies for
smoothing price fluctuations often include more efficient energy production and improvements in
demand management practices. Consequently, potential suppliers in the renewable sector, as well as
in deregulated markets, may play a significant role in meeting the future energy requirements of the
United States. Of course, newer technologies are developed and adopted as market signals dictate
(Economic Report of the President, 2006). Given these recent trends and concerns, we examine the
dynamic relationship between disaggregate energy sectors and real output in the US. In particular,
our primary focus is on the consumption of renewable energy and the role that these greener energy
sources may play in determining real economic activity.
We re-examine the link between disaggregate energy use, employment, and income in the US
employing the ARDL method. We use the autoregressive distributed lag (ARDL) approach of
Pesaran and Pesaran (1997) and Pesaran et al. (2001) to test for the existence of a relationship
between the disaggregate energy data and industrial production (i.e., real output) in level form.
Additionally, based on prior energy-related research which found the labor market to play a
significant role in the determination of output, our study also incorporates total non-farm
employment in the analysis. The ARDL approach may be applied to time series variables
irrespective of whether they are I(0), I(1), or mutually cointegrated.
The paper proceeds as follows. First, we provide a brief review of related research followed by
a description of the data and the ARDL approach. We then discuss the results and offer some
concluding comments regarding the implication for energy-related policy.

2. A brief literature review

A number of studies have focused on the relationship between energy and output or income
and the outcomes have varied considerably. Several factors may have led to this lack of
consensus. For instance, the varied results may be due to the different economic structures of the
particular countries studied, especially those that are at different stages of development. It is also
possible that the use of aggregate energy data has led to differences in outcomes. The use of
aggregate energy data does not capture the degree or extent to which different countries depend
on different energy resources (Yang, 2000). Studies that focus solely on aggregate data may not be
able to identify the impact of a specific type of energy on output. The use of disaggregate energy
consumption in a study of the energy–output relationship allows us to identify the impact of
different energy sources on income in the US.
Early work on the relationship between income and energy provided evidence of Granger
causality running from income to energy in the US (Kraft and Kraft, 1978). Building on this,
researchers have examined different countries over various time periods, using a number of
different methodologies. Some of the studies that found evidence of Granger causality running
from income to energy consumption include: Abosedra and Baghestani (1989) for the US, Yu and
Choi (1985) and Soytas and Sari (2003) for South Korea, Erol and Yu (1987) for West Germany,
Masih and Masih (1996) for Indonesia, Soytas and Sari (2003) for Italy, Wolde-Rufael (2005) for
five African countries, Narayan and Smyth (2005) for Australia, and Lee (2006) for France, Italy
2304 R. Sari et al. / Energy Economics 30 (2008) 2302–2313

and Japan. Additionally, Stern and Cleveland (2003) provide an excellent review of the literature
on the topic.
In contrast, evidence of causality running from energy consumption to income has also been
found. For example, Stern (1993, 2000) for the US, Erol and Yu (1987) for Japan, Yu and Choi
(1985) for Philippines, Masih and Masih (1996) for India and Indonesia, Soytas and Sari (2003)
for Turkey, France, Germany and Japan, Wolde-Rufael (2004) for Shangai, and Lee (2005) for
eighteen developing countries.
It is well known that economic systems may exhibit feedback effects or bi-directional
causality. In fact, a number of researchers have reported that the energy-income relationship may
be characterized by bi-directional causality. Erol and Yu (1987) report bi-directional causality for
Japan and Italy. Similar results have been reported for Taiwan (Hwang and Gum, 1992), while
additional evidence of bi-directional causality was found by Masih and Masih (1996) for
Pakistan, Soytas and Sari (2003) for Argentina, Ghali and El-Sakka (2004) for Canada, Wolde-
Rufael (2005) for Gabon and Zambia, and Lee (2006) for the US.
In stark contrast to the findings of causality, regardless of direction or the existence of
feedback, is the neutrality of energy hypothesis. Under this hypothesis, energy consumption and
income/output are unrelated and evolve independently from one another. Those finding evidence
in favor of the neutrality hypothesis include Akarca and Long (1980), Yu and Hwang (1984), Yu
and Choi (1985), Erol and Yu (1987) for the case of the US, Masih and Masih (1996) for
Malaysia, Singapore and Philippines, Soytas and Sari (2003) for nine countries including the US,
Asafu-Adjaye (2000) for Indonesia and India, Altinay and Karagol (2004) for Turkey, Wolde-
Rufael (2005) for eleven African countries, Lee (2006) for the UK, Germany and Sweden, and
Soytas and Sari (2006) for China.
Although there are studies that employ the ARDL technique to cointegration in the energy sector
(e.g De Vita et al., 2006), few studies have explicitly examined the relationship between energy
consumption and income using disaggregate data for the US (and to the extent of our knowledge
none did so using ARDL) which may provide information as to the source, if any, of an energy-
income relationship. Additionally, the disaggregate data allow for comparisons of the relative
strengths of the relationship by energy source. The few studies that did utilize disaggregate data
include Yang (2000), Wolde-Rufael (2004), Sari and Soytas (2004), and Ewing, Sari and Soytas
(2007) who highlight the importance of this new avenue of research. Thus, our approach is to utilize
the disaggregate data in conjunction with a methodology that does not impose the additional
restriction that the underlying series be integrated of the same order. Our focus is on investigating the
inter-temporal link between the consumption of energy from coal, fossil fuel, natural gas,
hydroelectric power, solar, wind, wood, waste, and industrial production incorporating total non-
farm employment in the US. The results shed light on at least one market signal, that is, the relative
role that different energy sources may play in explaining movements in output. The findings have
particular relevance for establishing research and development related policy in the US.

3. Data and methods

We examine monthly data for the United Sates over the period of 2001:1–2005:6.1 We use
industrial production, base year 2002, employment in thousands, coal, fossil fuels, conventional
hydroelectric power, solar energy, wind energy, natural gas, wood, and waste consumption.
Energy use data are in trillion British thermal units (Btu). Total renewable energy accounts for

1
Data are from Economagic.com. The sample period is constrained by data availability across all series.
 R. Sari et al. / Energy Economics 30 (2008) 2302–2313 2305

about 6% of total energy in the U.S. Fossil fuels make up about 86% of total energy with coal and
natural gas each being around 23%. In terms of renewables, solar is the least utilized and
comprises less than one-tenth of 1% and hydroelectric power, the most utilized, nearly 3% of total
energy. All data are in natural logarithms and are seasonally adjusted.
We employ the autoregressive distributed lag (ARDL) approach of Pesaran and Pesaran (1997)
and Pesaran et al. (2001) to test for existence of a relationship between the energy data,
employment, and industrial production in levels. As noted, this approach can be applied to series
irrespective of whether they are I(0), I(1), or mutually cointegrated. The methods adopted in the
literature in previous years mainly concentrate on cases in which the underlying variables are
integrated of order one (Pesaran et al., 2001).
One of the most frequently adopted methods that requires variables to be integrated of order one
was developed by Engle and Granger (1987). This approach consists of a procedure with two steps.
In the first step, a test of the cointegration is tested in which a regression of one non-stationary
series on another is run and the residuals are examined for stationarity. If the two non-stationary
series form a stable linear relationship, then they are said to be cointegrated. According to the
Representation Theorem, if the variables are cointegrated, then there is an error correction
representation. The next step is to estimate the error correction model which identifies the short run
dynamics of the system as well as the long run linkage. This approach is inefficient in multivariate
cases. Another commonly used approach was developed by Johansen (1988, 1991) and Johansen
and Juselius (1990) and is more efficient in multivariate systems. The ARDL approach has some
advantages over these other approaches. First, the series used do not have to be I(1) (Pesaran and
Pesaran, 1997). Second, even with small samples, more efficient cointegration relationships can be
determined (Ghatak and Siddiki, 2001). Finally, Laurenceson and Chai (2003) state that the ARDL
approach overcomes the problems resulting from non-stationary time series data. For instance,
non-stationary time series data leads to spurious regression coefficients that are biased towards
zero (Stock and Watson, 2003).

4. ARDL testing procedure and results

The ARDL method involves three steps (see Pesaran and Pesaran (1997) for more details and
an application using MICROFIT econometric software.). The first step is to test for the presence
of cointegration among the variables by employing the bounds testing procedure (Pesaran and
Pesaran, 1997; Pesaran, Shin and Smith, 2001). This test can identify the long run relationship
with a dependent variable followed by its forcing variables. Without having any prior information
about the direction of the long run relationship between industrial production and disaggregate
energy consumption, we construct the following regressions. Thus,
X n Xn X
n
DlnINDPRt ¼ a0indpr þ biindpr DlnINDPRti þ ciindpr DlnEMPti þ diindpr DlnENti
i¼1 i¼0 i¼0

þk1indpr lnINDPRt1 þ k2indpr lnEMPt1 þ k3indpr lnENt1 þ e1t ð1Þ

X
n X
n X
n
DlnENt ¼ a0ien þ bien DlnINDPRti þ cien DlnEMPti þ dien DlnENti
i¼0 i¼0 i¼1
þ k1en lnINDPRt1 þ k2en lnEMPt1 þ k3en lnENt1 þ e2t ð2Þ

where EN represents the particular disaggregate energy variable under investigation, INDPR
denotes industrial production, and EMP denotes employment. The parameters b, c and d are the
2306 R. Sari et al. / Energy Economics 30 (2008) 2302–2313

short run coefficients and λs are the corresponding long run multipliers of the underlying ARDL
model. The null hypothesis of “no cointegration” in Eqs. (1) and (2) is λ1 = λ2 = λ3 = 0. The
hypotheses are tested by computing the general F-statistics and comparing them with critical
values in Pesaran and Pesaran (1997) and Pesaran et al. (2001).
To determine the order of the series, we conducted six different unit root tests. We used the
augmented Dickey and Fuller (1979) (ADF), Phillips and Perron (1988) (PP), Elliot et al. (1996)
Dickey-Fuller GLS detrended (DF–GLS) and Point Optimal (ERS–SPO), Kwiatkowski et al.
(1992) (KPSS), and Ng and Perron's (2001) MZα (NP) tests. To conserve space, we do not
discuss the details of the unit root tests here (see Maddala and Kim (1998) for a review of ADF,
PP, KPSS, and DF–GLS; and Ng and Perron (2001) for more on NP). The results of these unit
root tests are available from the authors upon request.
The unit root test results indicate that for models including conventional hydroelectric power,
waste and wind energy consumption, we should use upper bounds for determination of cointe-
gration, while for the rest of the models, we should use both lower and upper bound critical values
reported in Pesaran, Shin and Smith (2001), and Pesaran and Pesaran (1997). The calculated
F-statistics are reported in Table 1.
The comparisons indicate that there are unique cointegrating relationships between the
variables in the models and that the long run forcing variables are employment and industrial
production in all relationships with the exception being when coal is the disaggregated energy
consumption variable. F(Industrial Productiont|Employmentt, Coalt) = 4.15 indicates that there is
a cointegrating relationship when the dependent variable is industrial production. In this case, the
forcing variables are employment and coal consumption. These results indicate that in all
relationships, except for coal, employment and industrial production are the forcing variables that
move first when a common stochastic shock hits the system. Then, energy consumption follows

Table 1
Bounds-testing procedure results
Cointegration hypotheses F-statistics
F(Industrial Production t|Employment t, Coal t) 4.15⁎⁎⁎
F(Coal t|Employment t, Industrial Production t) 2.79
F(Industrial Production t|Employment t, Fossil Fuel t) 2.30
F(Fossil Fuel t|Employment t, Industrial Productiont) 30.19⁎
F(Industrial Production t|Employment t, Hydroelectric t) 2.40
F(Hydroelectric t|Employment t, Industrial Production t) 9.10⁎
F(Industrial Production t|Employment t, Natural Gas t) 2.40
F(Natural Gas t|Employment t, Industrial Production t) 9.10⁎
F(Industrial Production t|Employment t, Solar t) 2.05
F(Solart|Employment t, Industrial Production t) 40.16⁎
F(Industrial Production t|Employment t, Waste t) 3.06
F(Waste t|Employment t, Industrial Production t) 11.72⁎
F(Industrial Production t|Employment t, Wind t) 2.65
F(Wind t|Employment t, Industrial Production t) 25.99⁎
F(Industrial Production t|Employment t, Wood t) 2.93
F(Wood t|Employment t, Industrial Production t) 32.34⁎
⁎represents significance at 1%, ⁎⁎ at 5%, and ⁎⁎⁎ at 10%. The critical values from Pesaran and Pesaran (1997) are 3.182–
4.126, 3.793–4.855, 4.404–5.524, 5.288–6.309 for 10%, 5%, 2.5%, and 1% significance level, respectively. Similarly,
from Pesaran, Shin and Smith (2001) we have critical values 3.17–4.15, 3.79–4.85, 4.41–5.52, 5.15–6.36 for 10%, 5%,
2.5%, and 1% significance level, respectively.
 R. Sari et al. / Energy Economics 30 (2008) 2302–2313 2307

Table 2
Estimated autoregressive distributed lag models, long run coefficients, and short run error correction model (Coal
Consumption, dependent variable: industrial production, ARDL(1,1,1))
Regressor Coefficient Standard error T-ratio[prob]
Panel A. Estimated long run coefficients
Employment −2.2153 11.2983 − 0.1961[.845]
Coal −1.6786 5.5031 − 0.3050[.762]
Intercept 43.3908 171.8561 0.2525[.802]

Panel B. Error correction representation for the selected ARDL

ΔEmployment 2.4292 0.8100 2.9993[.004]
ΔCoal 0.0524 0.0306 1.7134[.093]
Intercept 0.9131 1.3455 0.6786[.501]
ecm(− 1) −0.0210 0.0566 − 0.3718[.712]

the changes in the employment and industrial production. In the case of coal, it is the industrial
production that follows due to the unexpected shocks to the system.
The second step is to estimate the coefficient of the long run relationships identified in the
first step.2 Having found long run relationships (i.e. cointegration) among industrial production,
employment, and the various disaggregate energy consumption variables, in the next step
the long run relationship are estimated using the following selected ARDL(h, z, r) models
(See Tables 2–9).
X
h X
z X
r
INDPRt ¼ a0ien þ ai lnCOALti þ bi lnINDPRti þ gi lnEMPti þ e2t ð3Þ
i¼0 i¼1 i¼0

X
h X
z X
r
EVt ¼ a0ien þ ai lnEVti þ bi lnINDPRti þ gi lnEMPti þ e2t ð4Þ
i¼1 i¼0 i¼0

where EV denotes the particular energy variable under investigation except for coal
consumption which will not be a dependent variable based on results reported above. The
lag lengths h, z and r are determined by Schwartz Bayesian Criteria (SBC) criterion following
the suggestion of Pesaran and Pesaran (1997). Taking into consideration the limited number of
observations, a maximum of 3 lags was used. Tests for models of fossil fuel, wind, and solar
energy include minimum of 1 lag for dependent variable to ensure lagged explanatory variables
are present in the error correction model (ECM).
The long run test results (Panel A in Tables 2–9) reveal that industrial production and
employment are the key determinants of fossil fuel, conventional hydroelectric power, solar,
waste and wind energy consumption. In contrast, neither employment nor industrial production is
found to have a significant long run impact on consumption of natural gas and wood energy. The
industrial production equation indicates that neither employment nor coal energy consumption
has a significant long run impact on real output.
The long run impact of industrial production on energy consumption is generally positive as
expected, with the sole exception being a negative impact on solar energy. Mixed results are
obtained for the energy and employment relationship. Usually the relationship is negative and

2
See Pesaran and Pesaran (1997) for more details and an application using MICROFIT econometric software.
2308 R. Sari et al. / Energy Economics 30 (2008) 2302–2313

Table 3
Estimated autoregressive distributed lag models, long run coefficients, and short run error correction model (Fossil Fuels
Consumption, dependent variable: Fossil Fuels, ARDL(1,0,1))
Regressor Coefficient Standard error T-ratio[prob]
Panel A. Estimated long run coefficients
Industrial Production 1.1628 0.2491 4.6691[.000]
Employment − 1.7707 0.5814 − 3.0455[.004]
Intercept 24.3477 5.9890 4.0654[.000]

Panel B. Error correction representation for the selected ARDL

ΔIndustrial Production 1.0282 0.2513 4.0908[.000]
ΔEmployment − 8.2547 4.0875 − 2.0195[.049]
Intercept 21.5288 6.0223 3.5748[.001]
ecm(− 1) − 0.8842 0.1381 − 6.4039[.000]

significant indicating that energy use and employment are substitutes in the production processes.
However, in wood consumption the relationship is positive because labor intensive technologies
tend to require the wood use.
The third step is to estimate the short run dynamic coefficients. The short run dynamics are
provided in Panel B of Tables 2–9. In terms of signs and significances, the results are generally
consistent with the long run findings. However, for the model including natural gas in which we
found no significant long run relationship, we do find significance in the short run. Once again,
the relationship between employment, industrial production and conventional hydroelectric
power, solar, waste, and wind energy use are found to be statistically significant.
Employment seems to negatively affect the consumption of energy from all sources, except for
solar energy. The sign is reversed for the coefficient of industrial production, solar energy again
being an exception. The reason may be due in part to technology as solar power does not require
turning of turbines to produce electricity. The economic impact on jobs and output may thus differ
as other sources (i.e., wind, hydro, wood, waste) require manufacturing of turbines and related
equipment which is not required in the case of solar. Accordingly, it is possible that the economic
multipliers associated with construction and ongoing operations may differ. In fact, the Natural
Renewable Energy Laboratory (NREL) has developed an economic impact model (referred to as

Table 4
Estimated autoregressive distributed lag models, long run coefficients, and short run error correction model (Hydroelectric
Power, dependent variable: Hydroelectric, ARDL(1,0,0))
Regressor Coefficient Standard error T-ratio[prob]
Panel A. Estimated long run coefficients
Industrial production 3.3266 1.0144 3.2795[.002]
Employment − 11.5452 3.5276 − 3.2728[.002]
Intercept 126.0499 38.4305 3.2799[.002]

Panel B. Error correction representation for the selected ARDL

ΔIndustrial production 1.2967 0.6111 2.1218[.039]
ΔEmployment − 4.5002 1.9527 − 2.3045[.026]
Intercept 49.1326 21.0665 2.3323[.024]
ecm(− 1) − 0.3898 0.1221 − 3.1919[.003]
 R. Sari et al. / Energy Economics 30 (2008) 2302–2313 2309

Table 5
Estimated autoregressive distributed lag models, long run coefficients, and short run error correction model (Natural Gas
Consumption, dependent variable: Natural Gas, ARDL(1,1,2))
Regressor Coefficient Standard error T-ratio[prob]
Panel A. Estimated long run coefficients
Industrial production −0.2946 1.2218 −0.2411[.811]
Employment 0.0438 2.4670 .017737[.986]
Intercept 8.3864 24.3973 0.3437[.733]

Panel B. Error correction representation for the selected ARDL

ΔINDPR 2.2886 1.2127 1.8871[.065]
ΔEmployment −15.5954 7.7633 − 2.0088[.050]
ΔEmployment(−1) 16.7650 7.4350 2.2549[.029]
Intercept 4.1217 12.5619 0.3281[.744]
ecm(− 1) −0.4915 0.13924 − 3.5298[.001]

JEDI) that is unique to the wind energy industry. Future research may examine differences in
economic impacts based on energy source.
The findings in Table 2 reveal that coal consumption and employment are positively and
significantly related in the short run but not in the long run. A similar relationship exists between
coal consumption and output. These findings may be due to the many investment issues related to
future expectations of energy markets in general and, in particular, the associated risks. The
market may see alternative sources as the future for energy but recognizes the more immediate
role that coal plays in the economy. Further, the current production technologies for alternatives,
including oil-to-liquids, are not yet mature and production costs are high relative to conventional
sources (Economic Report of the President, 2006).
In all energy equations, except coal consumption, the error correction term (denoted ecm(− 1)
in Tables 2–9) is found to be negative and statistically significant. This term indicates the speed of
adjustment process to restore equilibrium following a disturbance in the long run equilibrium
relationship. A negative and significant error correction term implies how quickly variables return
to equilibrium. A relatively high ecm coefficient (in absolute magnitude) implies a faster
adjustment process. For instance, the model with fossil fuels implies that almost 89% (ecm

Table 6
Estimated autoregressive distributed lag models, long run coefficients, and short run error correction model (Solar Energy
Consumption, dependent variable: Solar, ARDL(1,0,0))
Regressor Coefficient Standard error T-ratio[prob]
Panel A. Estimated long run coefficients
Industrial production − 0.6441 0.1152 − 5.5902[.000]
Employment 1.3181 0.4069 3.2397[.002]
Intercept − 10.8810 4.4311 − 2.4556[.018]

Panel B. Error correction representation for the selected ARDL

ΔIndustrial production − 0.5155 0.1315 − 3.9199[.000]
ΔEmployment 1.0549 0.3679 2.8672[.006]
Intercept − 8.7080 3.7882 − 2.2987[.026]
ecm(− 1) − 0.8003 0.1468 − 5.4530[.000]
2310 R. Sari et al. / Energy Economics 30 (2008) 2302–2313

Table 7
Estimated autoregressive distributed lag models, long run coefficients, and short run error correction model (Waste,
dependent variable: Waste, ARDL(1,0,0))
Regressor Coefficient Standard error T-ratio[prob]
Panel A. Estimated long run coefficients
Industrial production 1.3971 0.4649 3.0048[.004]
Employment −4.0774 1.6548 − 2.4640[.017]
Intercept 45.4342 18.0099 2.5227[.015]

Panel B. Error correction representation for the selected ARDL

ΔIndustrial produciton 0.5107 0.2477 2.0622[.045]
ΔEmployment −1.4905 0.8284 − 1.7993[.078]
Intercept 16.6084 9.0920 1.8267[.074]
ecm(− 1) −0.3656 0.1109 − 3.2970[.002]

coefficient = − 0.8842) of the disequilibrium of the previous month's shocks adjust back to the
long run equilibrium in the current month. This value is approximately 39%, 49%, 80%, 37%,
79%, and 68% for equations including conventional hydroelectric power, natural gas, solar,
waste, wind and wood energy use, respectively.
The last issue we address is related to the goodness of fit of the ARDL models. For this purpose
we perform a series of diagnostic and stability tests. The diagnostic tests examine serial correlation
using the Lagrange multiplier test of residual serial correlation, functional form by employing
Ramsey's RESET test using the square of the fitted values, and heteroscedasticity based on the
regression of squared residuals on squared fitted values. The diagnostic tests reveal no evidence of
misspecification and, additionally, we find no evidence of autocorrelation. To test for structural
stability we utilize the cumulative sum of recursive residuals (CUSUM) and the cumulative sum of
squares of recursive residuals (CUSUMSQ). The results of CUSUM and CUSUMSQ stability test
indicate that the estimated coefficients of all models are stable.3

5. Conclusion and policy implications

In this paper we examined the relationship between disaggregate energy consumption and
industrial output, as well as employment, in the United States using the autoregressive
distributed lag (ARDL) approach developed by Pesaran and Pesaran (1997) and Pesaran et al.
(2001). This research contributes to the field of energy economics in two important ways. First,
following the recent trend, we utilize measures of disaggregate energy consumption thus
providing a comprehensive analysis. Second, we employ a relatively new time series approach
capable of uncovering relationships that might otherwise be missed using more conventional
methods.
The results of the bounds testing procedure confirm the presence of cointegration between
the energy measures, employment and industrial output. Over the long run output and labor
are the key determinants of fossil fuel, conventional hydroelectric power, solar, waste and
wind energy consumption. Employment and output are not found to have significant long run
impacts on natural gas, and wood energy. Our results also reveal information on the short run

3
Results of all diagnostic and stability tests are available upon request.
 R. Sari et al. / Energy Economics 30 (2008) 2302–2313 2311

Table 8
Estimated autoregressive distributed lag models, long run coefficients, and short run error correction model (Wind Energy,
dependent variable: Wind, ARDL(1,0,0))
Regressor Coefficient Standard error T-ratio[prob]
Panel A. Estimated long run coefficients
Industrial production 11.0669 0.9072 12.1984[.000]
Employment − 22.3938 3.1882 − 7.0240[.000]
Intercept 214.9290 34.7046 6.1931[.000]

Panel B. Error correction representation for the selected ARDL

ΔIndustrial production 8.7216 1.6647 5.2392[.000]
ΔEmployment − 17.6481 3.9285 − 4.4923[.000]
Intercept 169.3808 39.8057 4.2552[.000]
ecm(− 1) −0.7881 0.1306 − 6.0366[.000]

speed of adjustment process to restore long run equilibrium. The model with fossil fuels,
conventional hydroelectric power, natural gas, solar, waste, wind and wood energy use
implies that about 89%, 39%, 49%, 80%, 37%, 79%, and 68%, respectively, of the
disequilibrium of the previous month's shocks adjust back to the long run equilibrium in the
current month.
The results presented in this paper have important implications for public US energy policy
and private sector investment in energy production. For example, the results for the renewable
energy measures help identify the particular sectors in which economic growth is tied to
energy consumption over long periods of time. Thus, developers of renewable energy, both
public and private, can take this information into account when conducting benefit-cost
analyses and economic impact studies. Furthermore, payback periods for investment in wind
energy farms, etc. often take years, and knowledge about the long run relationships with
output and employment should help to provide more accurate forecasts of future energy
trends. Finally, for all the energy sectors (i.e., conventional and renewable), the speed of
adjustment estimates provide direct information as to the behavior of short run fluctuations
that can be incorporated into the demand management strategies of energy market producers
and policy makers.

Table 9
Estimated autoregressive distributed lag models, long run coefficients, and short run error correction model (Wood Energy,
dependent variable: Wood, ARDL(1,0,0))
Regressor Coefficient Standard error T-ratio[prob]
Panel A. Estimated long run coefficients
Industrial production 0.4045 0.2566 1.5764[.122]
Employment 0.6397 0.9094 .70341[.485]
Intercept − 4.3177 9.8922 − .43647[.664]

Panel B. Error correction representation for the selected ARDL

ΔIndustrial production 0.2744 0.1727 1.5885[.119]
ΔEmployment 0.4340 0.6420 .67598[.502]
Intercept − 2.9290 6.8621 − .42684[.671]
ecm(− 1) − 0.6784 0.1396 − 4.8606[.000]
2312 R. Sari et al. / Energy Economics 30 (2008) 2302–2313

References

Abosedra, S., Baghestani, H., 1989. New evidence on the causal relationship between United States energy consumption
and gross national product. Journal of Energy and Development 14, 285–292.
Akarca, A.T., Long, T.V., 1980. On the relationship between energy and GNP: a reexamination. Journal of Energy and
Development 5, 326–331.
Altinay, G., Karagol, E., 2004. Structural break, unit root, and the causality between energy consumption and GDP in
Turkey. Energy Economics 26, 985–994.
Asafu-Adjaye, J., 2000. The relationship between energy consumption, energy prices and economic growth: time series
evidence from Asian developing countries. Energy Economics 22, 615–625.
De Vita, G., Endressen, K., Hunt, L.C., 2006. An empirical analysis of energy demand in Namibia. Energy Policy 34, 3447–3463.
Dickey, D.A., Fuller, W.A., 1979. Distribution of the estimators for autoregressive time series with a unit root. Journal of
the American Statistical Society 75, 427–431.
Economic Report of the President. United State Government Printing Office: Washington, DC; 2006.
Elliott, G., Rothenberg, T.J., Stock, J.H., 1996. Efficient tests for an autoregressive unit root. Econometrica 64, 813–836.
Engle, R., Granger, C., 1987. Co-integration and error correction: representation, estimation, and testing. Econometrica 55,
251–276.
Erol, U., Yu, E.S.H., 1987. On the causal relationship between energy and income for industrialized countries. Journal of
Energy and Development 13, 113–122.
Ewing, B.T., Sari, R., Soytas, U., 2007. Disaggregate energy consumption and industrial output in the United States.
Energy Policy 35, 1274–1281.
Ghali, K.H., El-Sakka, M.I.T., 2004. Energy use and output growth in Canada: a multivariate cointegration analysis.
Energy Economics 26, 225–238.
Ghatak, S., Siddiki, J., 2001. The use of ARDL approach in estimating virtual exchange rates in India. Journal of Applied
Statistics 28, 573–583.
Hwang, D.B.K., Gum, B., 1992. The causal relationship between energy and GNP: the case of Taiwan. The Journal of
Energy and Development 16, 219–226.
Johansen, S., 1988. Statistical analysis of cointegrating vectors. Journal of Economic Dynamics and Control 12, 231–254.
Johansen, S., 1991. Estimation and hypothesis testing of cointegration vectors in Gaussian vector autoregressive models.
Econometrica 59, 1551–1580.
Johansen, S., Juselius, K., 1990. Maximum likelihood estimation and inference on cointegration: with application to the
demand for money. Oxford Bulletin of Economics and Statistics 52, 169–210.
Kraft, J., Kraft, A., 1978. On the relationship between energy and GNP. Journal of Energy and Development 3, 401–403.
Kwiatkowski, D., Phillips, P.C.B., Schmidt, P., Shin, Y., 1992. Testing the null hypothesis of stationary against the
alternative of a unit root. Journal of Econometrics 54, 159–178.
Laurenceson, J., Chai, J.C.H., 2003. Financial Reform and Economic Development in China. Edward Elgar, Cheltenham,
UK.
Lee, C.C., 2005. Energy consumption and GDP in developing countries: a cointegrated panel analysis. Energy Economics
27, 415–427.
Lee, C.C., 2006. The causality relationship between energy consumption and GDP in G-11 countries revisited. Energy
Policy 34, 1086–1093.
Maddala, G.S., Kim, I., 1998. Unit roots, cointegration, and structural change. Cambridge University Press, Cambridge.
Masih, A., Masih, R., 1996. Energy consumption, real income and temporal causality: Results from a multi-country study
based on cointegration and error-correction modeling techniques. Energy Economics 18, 165–183.
Narayan, P.K., Smyth, R., 2005. Electricity consumption, employment and real income in Australia: evidence from
multivariate Granger causality tests. Energy Policy 33, 1109–1116.
Ng, S., Perron, P., 2001. Lag length selection and the construction of unit root tests with good size and power.
Econometrica 69, 1519–1554.
Pesaran, M.H., Pesaran, B., 1997. Working with Microfit 4.0. Camfit Data Ltd, Cambridge.
Pesaran, M.H., Shin, Y., Smith, R.J., 2001. Bounds testing approaches to the analysis of level relationships. Journal of
Applied Econometrics 16, 289–326.
Phillips, P.C.B., Perron, P., 1988. Testing for a unit root in time series regressions. Biometrica 75, 335–346.
Sari, R., Soytas, U., 2004. Disaggregate energy consumption, employment, and income in Turkey. Energy Economics 26,
335–344.
Soytas, U., Sari, R., 2003. Energy consumption and GDP: causality relationship in G-7 countries and emerging markets.
Energy Economics 25, 33–37.
 R. Sari et al. / Energy Economics 30 (2008) 2302–2313 2313

Soytas, U., Sari, R., 2006. Can China contribute more to the fight against global warming? Journal of Policy Modeling 28,
837–846.
Stern, D.I., 1993. Energy and economic growth in the USA, a multivariate approach. Energy Economics 15, 137–150.
Stern, D.I., 2000. A multivariate cointegration analysis of the role of energy in the US economy. Energy Economics 22,
267–283.
Stern, D.I., Cleveland, C.J., 2003. Energy and economic growth. Rensselaer Working Papers in Economics, p. 0410.
Stock, J.H., Watson, M.W., 2003. Introduction to econometrics. Addison Wesley, Boston.
Yang, H.Y., 2000. A note on the causal relationship between energy and GDP in Taiwan. Energy Economics 22, 309–317.
Yu, E.S.H., Hwang, B.K., 1984. The relationship between energy and GNP: further results. Energy Economics 6,
186–190.
Yu, E.S.H., Choi, J.Y., 1985. The causal relationship between energy and GNP: an international comparison. Journal of
Energy and Development 10, 249–272.
Wolde-Rufael, Y., 2004. Disaggregated energy consumption and GDP, the experience of Shangai 1952–1999. Energy
Economics 26, 69–75.
Wolde-Rufael, Y., 2005. Energy demand and economic growth: the African experience 19 countries. Journal of Policy
Modeling 27, 891–903.