E.

ON Energy Research Center Series

Econometric Estimation of Energy Demand

Elasticities
Reinhard Madlener, Ronald Bernstein,
Miguel ngel Alva Gonzlez

Volume 3, Issue 8
Table of Contents
Executive Summary ................................................................................................................................................. 1
1 Introduction ..................................................................................................................................................... 3
1.1 Goals of the project ................................................................................................................................ 4
1.2 Positioning of the project within the E.ON ERC strategy ........................................................................ 4
2 Theoretical framework .................................................................................................................................... 4
2.1 Disaggregate sectoral approach ............................................................................................................ 4
2.2 Industrial electricity demand ................................................................................................................... 5
2.3 Residential electricity demand ............................................................................................................... 5
2.4 Residential natural gas demand ............................................................................................................. 6
2.5 Interfuel substitution ............................................................................................................................... 6
3 Literature overview on empirical econometric research .................................................................................. 8
3.1 Energy demand elasticities .................................................................................................................... 8
3.1.1 Industrial electricity demand .......................................................................................................... 8
3.1.2 Residential electricity demand ....................................................................................................... 9
3.1.3 Residential natural gas demand .................................................................................................. 10
3.2 Causality in the energy/growth nexus .................................................................................................. 11
3.3 Interfuel substitution ............................................................................................................................. 12
4 Econometric methodologies used ................................................................................................................. 13
4.1 Unit root tests ....................................................................................................................................... 14
4.1.1 Unit root tests based on time series data .................................................................................... 14
4.1.2 Unit root tests based on panel data ............................................................................................. 15
4.2 Approaches to cointegration analysis................................................................................................... 15
4.2.1 Johansen Maximum Likelihood ................................................................................................... 15
4.2.2 Panel cointegration ...................................................................................................................... 17
4.2.3 ARDL bounds testing approach................................................................................................... 19
4.3 Granger causality testing ..................................................................................................................... 20
4.4 Translog model .................................................................................................................................... 21
5 Empirical analysis ......................................................................................................................................... 22
5.1 Industrial electricity demand on the subsectoral level in Germany ....................................................... 22
5.1.1 Data ............................................................................................................................................. 22
5.1.2 Unit root tests .............................................................................................................................. 24
5.1.3 VAR specification ........................................................................................................................ 25
5.1.4 Cointegration rank tests .............................................................................................................. 26
5.1.5 VECM estimation (long-run elasticities) ....................................................................................... 26
5.1.6 Granger causality ........................................................................................................................ 28
5.1.7 Impulse response functions ......................................................................................................... 29
5.1.8 Short-run elasticities .................................................................................................................... 31
5.2 Residential electricity demand in OECD countries ............................................................................... 32
5.2.1 Data ............................................................................................................................................. 32
5.2.2 Panel unit root tests ..................................................................................................................... 34
5.2.3 Panel cointegration tests ............................................................................................................. 35
5.2.4 Long- and short-run elasticities ................................................................................................... 35
5.2.5 Panel Granger causality testing................................................................................................... 39
5.3 Residential natural gas demand in OECD countries ............................................................................ 40
5.3.1 Data ............................................................................................................................................. 40
5.3.2 Bounds test to cointegration ........................................................................................................ 43
5.3.3 Long-run relationships and short-run dynamics........................................................................... 43
5.3.4 Constancy of cointegration space ............................................................................................... 46
5.4 Interfuel substitution in major European countries ............................................................................... 48
5.4.1 Data ............................................................................................................................................. 48
5.4.2 Translog estimations ................................................................................................................... 48
6 Conclusions .................................................................................................................................................. 50
7 Literature....................................................................................................................................................... 52
8 Attachments .................................................................................................................................................. 56
8.1 List of Figures....................................................................................................................................... 56
8.2 List of Tables ........................................................................................................................................ 56
8.3 Publications generated by the project .................................................................................................. 57
8.4 Short CV of scientists involved in the project ....................................................................................... 57
8.5 Project timeline..................................................................................................................................... 58
8.6 Activities within the scope of the project ............................................................................................... 58

i
Executive Summary
In this study we aim at shedding light on the responsiveness of energy demand to measures of
economic activity and energy price. Furthermore, we also aim at analyzing the causal relationship
between energy use and economic activity.
The methodological focus of the study lies in the application of econometric methods based on time
series and panel data. As most economic time series contain stochastic trends, the application of
simple regression techniques will likely lead to spurious results. Hence, the application of cointegration
analysis, which takes the non-stationarity of the data explicitly into account, is required. The concept of
cointegration was introduced by the seminal paper of Engle and Granger (1987). Since then, a
number of different approaches to cointegration have been introduced in the econometrics literature.
We apply three of the most common state-of-the-art techniques, which all have their own merits: (1)
the maximum likelihood system approach (Johansen, 1988 and 1995), (2) the fully modified OLS
(FMOLS) and dynamic OLS (DOLS) group-means panel estimation framework (Pedroni, 1999 and
2004), and (3) the autoregressive distributed lag (ARDL) bounds testing procedure (Pesaran and Shin,
1999; Pesaran et al. 2001).
A further focus of this study is the use of data at the lowest level of aggregation as possible, which
stands in contrast to most similar studies in the energy demand literature. The reason for this
approach is that analyzing data aggregated over widely heterogeneous sectors will most likely result in
crude inference concerning economic relationships and consumer behavior. To this end, our aim is to
reap the benefits of additional information otherwise blurred through aggregation, by analyzing
sectoral/subsectoral demand functions for a single energy carrier. More specifically, in this study, we
focus on industrial electricity demand, residential electricity demand and residential natural gas
demand. For many developed countries, industry data is even available on a subsectoral level, which
allows to further reduce the heterogeneity of the consumer groups analyzed.
The report is organized as follows. Section 1 gives an introduction on the goals and on the content
of the project. Section 2 provides the theoretical framework for the empirical analysis, while Section 3
gives an overview of the relevant empirical literature. Section 4 explains the different econometric
methods used in the analysis. Section 5 describes the data used, the application of the estimation
methods, and the results. Section 6 concludes.
In the first part of the empirical analysis (Section 5.1), we use the Johansen maximum likelihood
system approach to estimate industrial electricity demand elasticities at the subsectoral level. This
enables us to reap the benefits of lower heterogeneity within the electricity-consuming sectors
investigated and of retaining additional information otherwise blurred by aggregation. The annual data
set used covers eight subsectors of the German economy for the period 1970 to 2007. By employing a
cointegrated VAR model specification and accounting for structural breaks we find cointegration
relationships for five of the eight subsectors studied. The long-run elasticities range between 0.70 and
1.90 for economic activity and between 0.52 and zero for the price of electricity. The short-run
elasticities are estimated by single-equation error correction modeling and found to be between 0.17
to 1.02 for economic activity and 0.57 to zero for electricity price. Granger-causality tests indicate
that, in the long term, causality runs from both economic activity and electricity price to electricity
consumption, while Granger-causality from electricity price and electricity consumption to economic
activity is detected in only two subsectors. Electricity price is found to be Granger-caused neither in
the long nor the short run. Finally, an impulse response analysis yields plausible results with regard to
the demand adjustment to a price shock, confirming the usefulness of the approach adopted.
In the second part (Section 5.2), we estimate residential electricity demand elasticities and conduct
an analysis of the causal relationship between electricity demand, disposable income and electricity
price for a group of several OECD member countries. More specifically, we apply panel cointegration
and Granger causality testing to a data set consisting of eighteen countries in the cross-sectional
dimension and the years 1981 to 2008 in the time domain. Our results for the whole panel indicate a
near unity income elasticity and an inelastic price elasticity of approximately 0.4 in the long run.
These results are robust with regard to the estimation methods employed (i.e. group-means panel
FMOLS and DOLS, respectively). In the short run, our estimates from an ECM indicate an income

1
elasticity of 0.2 and a price elasticity of approximately 0.1. Moreover, our tests on Granger causality
provide an indication for a bidirectional causal relationship between electricity consumption and
economic growth. Hence, our findings are in favor of the feedback hypothesis.
In the third part (Section 5.3), we analyze residential natural gas demand for twelve OECD
countries using available time series data from 1980 to 2008. We estimate long-run demand
elasticities with regard to real disposable income and real residential natural gas price using the
autoregressive distributed lag (ARDL) bounds testing procedure. By employing an error correction
framework we also obtain estimates for the speeds of adjustment to long-run equilibrium and short-run
elasticities for the single countries. The effect of weather conditions on natural gas demand in a given
year is accounted for by including heating degree days as a control variable. On average, the long-run
elasticities are 0.94 with regard to income, 0.51 with regard to price, and 1.35 with regard to heating
degree days. The short-run dynamics assessed by estimation of the error correction models indicate
an average adjustment coefficient of 0.58, a short-run income elasticity of 0.45, a short-run price
elasticity of 0.24, and a short-run elasticity with regard to heating degree days of 0.72. Hence, on
average, the short-run elasticities have approximately half the magnitude of their long-run
counterparts.
In the fourth part (Section 5.4), we supplement the research with the inclusion of an analysis of
interfuel substitution by using a translog model specification. In particular, we estimate to what extent
there is substitution among different kinds of energy inputs (e.g., electricity, gas, oil) in the economy of
five major European countries (i.e., Germany, France, Italy, Spain, and the UK). To this end, we apply
a cost share model specification of energy consumption for four fuels (where data on coal were lacking
three fuels), obtaining a set of own-price and cross-price elasticity estimates for the aggregate
economy as well as the industrial and the residential household sector. The models are estimated both
with and without a deterministic time trend as a proxy for technical change and other influences
otherwise not picked up by the model specification, found to be statistically significant (and thus
relevant to be included) in almost all cases. While many of the estimated coefficients are statistically
significant and have the expected sign, the great range of results and lack of consistent patterns of
results (across sectors, fuels and countries) makes it difficult to come up with specific
recommendations for policy-makers. In this respect, we find the results obtained from the cointegration
analysis as more robust and useful.
Overall, regardless of the sector or energy type considered, our results imply that the steering
effect of tax-induced price increases on energy demand has a very limited potential for energy
conservation, and hence a reduction of GHG emissions. Furthermore, our findings suggest that
reductions in electricity consumption are associated with a trade-off concerning economic growth in
the residential sector, as well as some subsectors of the manufacturing industry.

2
1 Introduction
Energy-related policy decisions have far-reaching and long-term consequences for the structure of the
prevailing energy system. Hence also the awareness of the mechanisms of the energy markets is an
important issue. In order to enable a more environmentally and socially benign use of energy in the
future, and to provide guidance for policy-makers in designing appropriate policies, the accurate
prediction of how energy consumers react to changes in price, income and other explanatory variables
is paramount.
Elasticity estimates provide some information as to how sensitive consumer behavior is with
respect to changes in important explanatory variables (e.g. energy price, income, energy-using capital
stock). In the field of energy use, it is a major challenge for applied energy economists/
econometricians and modelers to determine energy demand elasticities that are useful for policy
design and strategic decision-making of utilities, technology manufacturers, and other relevant
stakeholders.
Furthermore, the assessment of the causal relationship between energy use and economic activity
is valuable for an appraisal of potentially conflicting policy objectives, such as the trade-off between
energy conservation and economic growth. There are four alternative economic hypotheses regarding
the causal mechanisms underlying the energy consumption economic growth nexus, which are
currently heavily under debate in the energy economics literature (for a useful survey see Payne,
2010): the conservation hypothesis, the growth hypothesis, the feedback hypothesis and the neutrality
hypothesis. While the conservation hypothesis implies causality running from economic growth to
energy consumption, the growth hypothesis implies the opposite causal relationship. The feedback
hypothesis combines the latter hypotheses, by claiming an interdependent causal relationship
between both variables. Finally, the neutrality hypothesis states that both variables are only of little
importance in determining each other.
Econometric work still seems to dominate the applied research done in quantitative analysis of
energy demand. In recent years, importantly, new econometric techniques (e.g. panel data
econometrics, cointegration analysis) have become more and more popular among researchers and
analysts alike, and increasingly applied, while at the same time interest in the estimation of energy
demand elasticities and changes thereof seems to have diminished. However, both policy-makers and
decision-makers in industry alike, and energy modelers providing model-based insights, rely on
reliable estimates of the sensitivity of energy demand in reaction to changes in important variables
shaping energy demand.
Reliable econometric estimates of energy demand elasticities are rare, and research interest
seems to have waned in recent years, despite an increasing stock of data and the acknowledgement
that a better understanding of energy consumer behavior is crucial for decision support in the energy
domain. This is all the more surprising in view of dynamic structural change in the energy markets,
especially in the electricity markets, and the availability of improved time series, cross-sectional, and
micro data as well as econometric modeling techniques (such as cointegration analysis and panel data
econometrics).
The main aims of the proposed research project are at least threefold: (1) to describe and apply the
state-of-the-art in using econometric techniques for estimating elasticities of energy demand; (2) to
estimate sectoral and aggregate elasticities of energy demand in selected European countries for
own- and cross-prices, income, temperature, and other explanatory variables; and (3) to compare and
contrast the empirical results obtained from different model specifications, estimation techniques and
data samples used with the existing literature. The results will provide new insights into the sensitivity
of different classes of energy consumers with regard to changes in important explanatory variables.
Furthermore, they will also shed new light on the robustness of the econometric estimates if different
methods are employed.

3
1.1 Goals of the project
The main goal of this research project is to discuss the state-of-the-art of estimating energy demand
elasticities, and to estimate elasticities of energy demand for different sectors of selected European
economies, based on state-of-the-art econometric estimation techniques.
The following objectives are to be accomplished:
Overview of the state-of-the-art of the literature on econometric studies of energy demand,
and in particular the estimation of elasticities of demand for energy relative to a selection of
influencing variables of interest (price, income, other).
Collection and detailed descriptive analysis of the data available for the empirical study
(annual, quarterly and monthly, if possible).
Exploration and statistical testing of a set of different model specifications (e.g. translog and
error correction) and econometric estimation techniques, and systematic comparison and
interpretation of the results gained from the different model specifications.
Selection of the best (most promising) models and results and testing for robustness and
possible model misspecification.
Systematic compilation of and reflection on the results obtained for the different sectors and
time periods studied.
Formulation of a Final Report.

1.2 Positioning of the project within the E.ON ERC strategy

This project helps to improve the understanding of energy demand responsiveness in different sectors
of the economy of selected OECD countries with regard to changes in price, income and weather, as
well as the understanding of causality of the relationship between economic activity and energy use.
With regard to capacity building at the E.ON ERC, the project helps to establish and improve both
theoretical and applied competence in state-of-the-art time series and panel econometrics analysis.
The results obtained are useful inputs to various kinds of energy models used within the E.ON ERC
and by external modelers globally. Finally, the results are helpful for energy policy design aiming at
influencing energy demand patterns by price-based incentives (subsidies or taxes).

2 Theoretical framework
2.1 Disaggregate sectoral approach
Energy demand modeling on the basis of historical time-series data has traditionally been conducted
for a specific country, at an aggregate or disaggregate level and in two dimensions. One dimension
concerns the type of energy (i.e. mainly electricity, natural gas or gasoline), while the other dimension
concerns different types of major end-use sectors: industry, commerce and public services, residential,
and transportation. At one extreme, there is the analysis on the basis of data aggregated over all
energy carriers and sectors (i.e. at the economy-wide level), whereas at the other extreme there is the
analysis for only one energy carrier for one sector or subsector of the economy. The analysis of data
aggregated over widely heterogeneous sectors will most likely result in crude inference concerning
economic relationships and consumer behavior. In this respect, we share the view of Pesaran et al.
(1998, p.46), viz. that it is important for a valid econometric demand analysis to be aimed at as
homogenous a grouping of consumers as is feasible. This implies that studies on energy demand
should use data at the lowest level of aggregation as possible.1 To this end, our aim in the present
study is to reap the benefits of additional information otherwise blurred through aggregation, by

1
Implying that aggregation in general is aimed at stepwise compiling entities with similar characteristics and, hence, also similar
consumption behavior and technologies.

4
analyzing (sub-)sectoral demand functions for a single energy carrier. More specifically, in the
following, we focus on industrial electricity demand, residential electricity demand and residential
natural gas demand. For many developed countries, data for industry is even available on a
subsectoral level, which allows us to reduce the heterogeneity of the consumer groups analyzed still
further.

2.2 Industrial electricity demand

A generic long-run electricity demand relationship for the industrial sectors of an economy can be
characterized by the general function

Et f Vt , Pt , X t , At ,
(1)

where electricity consumption (Et) is contemporaneously dependent on the level of real economic
activity (Vt), real electricity price (Pt), other endogenous or exogenous variables (Xt) (which may
include, for example, the real price of an electricity substitute and/or weather variables), and
exogenous factors (At), such as a sector-specific coefficient for autonomous technical change, energy-
saving technological progress or shifts/changes in the structure of industrial production. The latter may
comprise structural changes due to substitution of labor by electricity-using capital and the offshoring
of labor-intensive production processes to other countries. In contrast to energy-saving technological
progress, both changes tend to increase the electricity intensity of the respective national sectors.
These factors affect the relationships between the other variables and can be indirectly accounted for
by the inclusion of deterministic terms.
Various econometric studies have found that other energy inputs are generally poor substitutes for
electricity in industrial processes (for a survey, see Barker et al., 1995). Thus, we refrain from
controlling for interfuel substitution by including prices of other energy carriers. Moreover, the inclusion
of heating and cooling degree day variables, which were available to us only from 1975 onwards,
would have considerably reduced the number of degrees of freedom in our analysis.2 Specifically, for
the empirical analysis we chose the simple standard constant elasticity (Cobb-Douglas type) functional
form

Et C0 exp dt Vt v Pt p ,

(2)

where At = C0 exp(dt) is the deterministic term, C0 is a constant, dt is a linear time trend and v and p
are the constant demand elasticities with regard to economic activity and electricity price, respectively.
As stated by Amarawickrama and Hunt (2008), this standard log-linear specification, apart from the
obvious advantages, such as its simplicity, its straightforward interpretation and the limited data
requirements, according to Pesaran et al. (1998) outperforms more complex models.

2.3 Residential electricity demand

The long-run relationship between residential electricity demand and its determinants can be
characterized by the general function

Et f Yt , Pt , X t . (3)

Eq. (3) states that residential electricity consumption per capita (Et) is a function of real (disposable)
income per capita (Yt) and real residential electricity price (Pt). Previous studies on residential
electricity demand have included further control variables (Xt), as for example the real price of an

2
Other studies have also found that weather variables tend to be insignificant in industrial energy demand functions, especially
in electricity demand functions (see Kamerschen and Porter, 2004).

5
electricity substitute (e.g., Narayan et al., 2007), heating and cooling degree days (e.g., Zachariadis
and Pashourtidou, 2007), urbanization (e.g., Holtedahl and Joutz, 2004), and capital variables (e.g.,
Silk and Joutz, 1997).
Following the principle of Ockhams Razor (see Ariew, 1976), we choose a parsimonious
specification, which only includes real disposable income per capita and real residential electricity
price as determinants of electricity demand. Furthermore, as the following analysis comprises a panel
of various countries, problems with the availability of data on additional variables would have brought
further restrictions to the data set with regard to the cross-sections and/or the time periods studied.3
Finally, having less parameters to estimate has the advantage of attaining more degrees of freedom in
the estimation of the core explanatory variables coefficients. More specifically, the demand model
upon which our econometric analysis is based, takes the following standard constant elasticity
functional form:

y
Et C0 Yt Pt p , (4)

where C0 is a country-specific drift term, and y and p are the long-run elasticities to be estimated with
regard to income and electricity price, respectively. A higher income is expected to increase electricity
demand on account of higher economic activity, whereas a higher electricity price is naturally expected
to decrease electricity demand. Moreover, the price elasticity is expected to be inelastic, as in general
electricity is characterized by a lack of substitutability.

2.4 Residential natural gas demand

Similarly to the residential electricity demand function from the last section, natural gas demand can
generally be considered to be a function of several determinants, such as

Gt f Yt , Pt , X t , (5)

where Gt is residential natural gas consumption per capita, Yt is real net disposable income, Pt is real
residential natural gas price and Xt stands for further control variables, all at time t. As most of the
natural gas in the residential sector is used for heating purposes4, a variable which controls for the
temperature is included as well (heating degree days, HDD). More specifically, for the long-run natural
gas demand relationship in the residential sector we chose the following constant elasticity functional
form:

Gt 0 exp 1t Yt 2 Pt 3 HDDt4 , (6)

where HDDt are heating degree days and the s are the coefficients to be estimated. Following other
studies, we initially also considered the price of electricity as a substitute for natural gas. But as the
estimates of the respective cross-price elasticities were not significant, we omitted these in order to
gain degrees of freedom.

2.5 Interfuel substitution

Assuming that production is characterized by a production function, Y = f(x), the solution to the
problem of minimizing the cost of producing a specified output, given a set of factor prices, leads to the
cost-minimizing set of factor demands x = x(Y, p). The total cost of production is then given by the cost
function, which can then be used as a starting point for interfuel substitution modeling.

3
Also, not all variables, such as cooling degree days or the price of natural gas, are likely to have the same relevance (or, more
specifically, similar explanatory power) for all the countries considered.
4
A minor share of natural gas consumption is attributable to cooking.

6
 The use of a cost function rather than a production function for estimating production parameters
has certain advantages. First, having prices as independent variables, the inversion of the matrix of
estimates is not necessary (i.e., less complexity is involved). Second, all estimation equations are
linear (in logarithms, for the translog case) and, third, since there is usually little multicollinearity
among factor prices, multicollinearity is a problem that does not arise in a cost function.
In particular, the use of a transcendental logarithmic (translog) function can provide accurate global
approximations of many cost frontiers used in econometric analysis (Christensen et al., 1973). Thus,
differentiating the translog frontier while holding the necessary conditions for equilibrium constant, we
obtain the factor cost shares for each kind of energy input i.
The derivation is as follows. We start from a cost minimization problem

n
min C X i Pi s.t. Y f X 1 , X 2 ,..., X n ; i 1, 2,..., n. (7)
i 1

where C is production cost, Xi is input of fuel i, and Pi is the real price of fuel input i.

The minimum cost function is specified as

C * g Y , P1 ,..., Pn (8)

or in natural logarithms as

ln C * f ln Y , ln P1 ,..., ln Pn . (9)

Using a logarithmic Taylor series expansion (to the second term) on (9), we obtain the twice
differentiable analytic cost function:

1
ln C * v0 vY ln Y vi ln Pi i, j ln Pi ln Pj i i,Y ln Pi ln Y . (10)
i 2 i j

Equation (10) can be directly estimated, or estimated in its first derivatives (by applying Shephards
Lemma), as factor shares (i.e., cost shares that sum up to unity), that is:

ln C * 1
Si vi i , j ln Pj i ,Y ln Y (11)
ln Pi 2 j

Thus, partial elasticities of substitution (among fuels) are defined as:

PX i i
2C
i , j j ,i i 1
. (12)
xi x j Pi Pj

The i,j parameters are related to elasticities of substitution and of factor demand shares as follows:

i, j
i, j 1 (13)
Si S j

Finally, the own-price elasticities and cross-price elasticities of input demand are defined as:

7
 i ,i Si Si 1 i, j S j S j
Ei ,i 2
, and Ei , j (14)
S i Si S j

3 Literature overview on empirical econometric research

3.1 Energy demand elasticities
The econometric estimation of energy demand elasticities has a long tradition, going back as far as
the early 1950s. Among the first studies are Houthakker (1951), Fisher and Kaysen (1962), Halvorsen
(1975) and Pindyck (1979). Similarly, the related literature strand aiming at determining the causal
relationship in the energy-growth nexus was sparked by the early work of Kraft and Kraft (1978), and
has received considerable interest especially since the turn of the century. However, it was only with
the introduction of cointegration5 analysis, which was triggered by the seminal paper of Engle and
Granger (1987), that the problem of spurious regressions6 was starting to be adequately dealt with in
econometric applications.

3.1.1 Industrial electricity demand

Despite the crucial relevance of sound elasticity estimates in energy modeling used for policy advice,
scholarly literature on the econometric estimation of energy demand elasticities in industry is
surprisingly scarce, and this is even more so with regard to electricity. Table 1 summarizes recent
studies in which electricity demand elasticities of economic activity and/or electricity price in industry
are estimated (Beenstock et al., 1999; Bose and Shukla, 1999; Kamerschen and Porter, 2004;
Polemis, 2007). These studies differ with regard to the model specification, the econometric method
used and time span covered, the data frequency, and the country analyzed. Beenstock et al. (1999)
use dynamic regression and cointegration techniques to analyze electricity demand in the household
and industry sector in Israel. For the industrial sector they estimate long-run elasticities of 0.99 to 1.12
with regard to economic activity and 0.31 to 0.44 with regard to electricity price, depending on the
estimation method applied. Using time series data for nine years from 19 states in India, Bose and
Shukla (1999) estimate sectoral elasticities including industry (split into small/medium and large
industries) by employing a pooled regression approach. The estimated elasticities of economic activity
and price are 0.49 and 0.04 (the latter not significant), respectively, for the small- and medium-sized
industries, and 1.06 and 0.45 for the large industries. Kamerschen and Porter (2004) employ a
simultaneous equation approach for estimating price elasticities of electricity demand by the U.S.
industry.7 Depending on the specification their estimates vary between 0.34 and 0.55. Polemis
(2007) uses a multivariate cointegration technique (the Johansen maximum likelihood approach) to
estimate aggregate oil and electricity demand functions for the Greek industry. His estimates for long-
run elasticities regarding economic activity and price are 0.85 and 0.85, while in the short-run they
amount to 0.61 and 0.35, respectively.
The only energy demand study known to us that uses disaggregated industrial data at the two-digit
level of the NACE taxonomy is Agnolucci (2009). In contrast to our study however, he focuses on
aggregate energy in the British and German industry. Moreover, the analysis is based on a panel data
approach, as the time series estimates mostly failed to show intuitive results. This presumably is due
to the short time spans covered by the data. Finally, although information from disaggregated data is
used in the estimation, the panel approach does not deliver subsector-specific estimates of energy
demand elasticities.

5
If a linear combination of two or more non-stationary stochastic processes exists, which itself is stationary, the processes are
said to be cointegrated. Tests on cointegration are introduced in Section 4.
6
A regression, which reveals a statistically significant relationship not due to a causal relationship is said to be spurious.
7
Kamerschen & Porter (2004) also consider a partial-adjustment approach, which, however, had to be dropped due to counter-
intuitive estimates.

8
Table 1
Industrial electricity demand studies and elasticity estimates
Study Country Method Data Elasticity estimates
Econ. activity Price

Beenstock et. Time series (quarterly), L: 0.31 to

Israel Cointegration L: 0.99 to 1.12
al. (1999)* 1975q21994q4 0.44

Bose & Shukla Pooled Panel data (annual),

India 0.49 to 1.06 0.04 to 0.45
(1999)* regression 1985/861993/94

Kamerschen & Simultaneous Time series (annual),

USA 0.34 to0.55
Porter (2004)* equations 19731998

Polemis Time series (annual), L: 0.85 L: 0.85

(2007)** Greece Cointegration
19702004 S: 0.61 S: 0.35

Notes: * also estimate demand for other sectors. ** also estimates an oil demand function separately. S and L denote estimates
for the short and the long run, respectively.

3.1.2 Residential electricity demand

Table 2 gives an overview of selected recent studies on residential electricity demand elasticities, most
of which employ pure time series methods for single countries:8 Athukorala and Wilson (2010) for Sri
Lanka; Dergiadis and Tsoulfidis (2008) for the US; Halicioglu (2007) for Turkey; Holtedahl and Joutz
(2004) for Taiwan; Hondroyiannis (2004) for Greece; Nakajima (2010) for Japan; Nakajima and
Hamori (2010) for the US; Narayan and Smyth (2005) for Australia; Narayan et al. (2007) for the G7
countries; Zachariadis and Pashourtidou (2007) for Cyprus and Ziramba (2008) for South Africa. The
long-run demand elasticities from these studies range between 0.25 and 1.57 with regard to income
and between 0.14 and 1.56 with regard to own price. For the short run, the elasticity estimates
range between 0.10 and 0.44 with regard to income and between 0.11 and 0.46 with regard to own
price.
Unlike the majority of the studies reported, Nakajima and Hamori (2010), Nakajima (2010) and
Narayan et al. (2007) all use cointegration techniques that are based on panel data in their analyses.
In order to attain a cross-sectional dimension, the former two studies make use of available data on
geographical regions within Japan and the US, respectively. The latter study is closest to ours with
regard to the countries9 analyzed and the econometric approach applied, but interestingly comes to
different conclusions. Specifically, Narayan et al. (2007) analyze a panel consisting of the G7 countries
as the cross-sectional dimension for the time period 19782003. They employ both the ordinary least
squares (OLS) and the dynamic OLS (DOLS) method (Stock and Watson, 1993; Kao and Chiang,
2000) to estimate the cointegration relationship between electricity demand, income, electricity price,
and natural gas price. For the whole panel they estimate an inelastic income elasticity of 0.25 (0.31)
and an elastic price elasticity of 1.56 (1.45) using DOLS (OLS). Consequently, they come to the
conclusion that pricing policies aimed at reducing residential electricity consumption and hence GHG
emissions are bound to be successful.

8
We only consider studies here that employ time-series or panel estimation techniques, that account for non-stationarity in the
data generating process (DGP), and that were published after the year 2000.
9
Our set of countries also includes the G7 countries, except for Canada, for which recent data on the residential electricity price
were lacking.

9
Table 2
Residential electricity demand studies and elasticity estimates
Study Country Method Data Elasticity estimates
Income Price

Athukorala & Johansen / Time series, L: 0.78 L: 0.62

Sri Lanka
Wilson (2010) VECM 19602007 (annual) S: 0.32 S: 0.16

Dergiades &
Bounds testing Time series, L: 0.27 L: 1.07
Tsoulfidis US
/ ARDL 19652006 (annual) S: 0.10 S: 0.39
(2008)

Halicioglu Bounds testing Time series, L: 0.49 to 0.70 L: 0.52 to 0.63

Turkey
(2007) / ARDL 19682005 (annual) S: 0.37 to 0.44 S: 0.33 to 0.46

Time series,
Holtedahl & Johansen / L: 1.04 to 1.57 L: 0.15
Taiwan 19551995 (annual)
Joutz (2004) VECM S: 0.22 S: 0.15

Hondroyiannis Johansen / Time series, L: 1.56

Greece L: 0.41
(2004) VECM 19861999 (monthly) S: 0.20

Panel Panel data,

Nakajima
Japan cointegration, 19752005 (annual), L: 0.60 to 0.65 L: 1.13 to 1.20
(2010)
DOLS T N: 31 46 = 1426

Panel Panel data,

Nakajima &
US cointegration, 19932008 (quarterly), L: 0.38 to 0.85 L: 0.14 to 0.33
Hamori (2010)
DOLS T N: 32 49 = 1568

Narayan & Bounds testing Time series, L: 0.32 to 0.41 L: 0.47 to 0.54
Australia
Smyth (2005) / ARDL 19692000 (annual) S: 0.01 to 0.04 S: 0.26 to 0.27

Panel Panel data,

Narayan et al. L: 0.25 to 0.31 L: 1.45 to 1.56
G7 Cointegration, 19782003 (annual),
(2007) S: 0.19 S: 0.11
OLS & DOLS T N: 26 7 = 182

Zachariadis &
Johansen / Time series,
Pashourtidou Cyprus L: 1.18 L: 0.43
VECM 19602004 (annual)
(2007)

Ziramba South Bounds testing Time series, L: 0.31 to 0.87 L: 0.01 to 0.04
(2008) Africa / ARDL 19782005 (annual) S: 0.30 S: 0.02

Notes: S and L denote estimates for the short and the long run, respectively. Elasticity estimates which are not statistically
significantly different from zero on conventional levels are printed in italics. T: Number of time series observations; N: Number of
cross-sections. DOLS: Dynamic OLS; ARDL: Autoregressive Distributed Lag.

3.1.3 Residential natural gas demand

There are many studies on the econometric analysis of (residential) natural gas demand, but nearly all
of these are from the 1960s to the 1980s (for a useful survey see Madlener, 1996). More recent
studies, especially from the 2000s, are very rare, and none of them take the aforementioned problems
related to non-stationarity in the data / spurious regressions into account. Table 3 provides an
overview of recent studies known to us: Asche et al. (2008) analyze residential natural gas demand in
12 EU member countries, using panel data for the time period 1978 to 2002. Their shrinkage estimator
reveals an income elasticity of 3.32 in the long run and 0.81 in the short run. For the price elasticity the
estimates are 0.10 and 0.03 for the long run and the short run, respectively. Berkhout et al. (2004)
use fixed effects to estimate residential natural gas demand elasticities in the Netherlands. The

10
estimates are (surprising) 0.27 and 0.19 (the latter not significant) for the long-run income and price
elasticity, respectively. Using an error-components and seemingly unrelated regression (SUR)
approach, Lin et al. (1987) estimate elasticities for the US based on panel data from 1960 to 1983.
Their estimates for the income elasticities are 0.57 and 0.11 for the long run and the short run,
respectively. For the price elasticity, their estimates are 1.22 for the long run and 0.15 for the short
run. Using the shrinkage estimator, Joutz et al. (2008) estimate elasticities for the US based on panel
data. Their of price elasticity estimates are 0.18 in the long run and 0.09 in the short run.

Table 3
Residential natural gas demand studies
Study Country Method Data Elasticity estimates
Income Price

Asche et al. 12 EU Shrinkage Panel data (annual), L: 3.32 L: 0.10

(2008) countries estimator* 19782002 S: 0.81 S: 0.03

Berkhout et al. Panel data (annual), L: 0.27 L: 0.19

Netherlands Fixed effects
(2004) 19921999

Joutz et al. Shrinkage Panel data (monthly), L: 0.18

US
(2008) estimator 1980unclear S: 0.09

Notes: S and L denote estimates for the short and the long run, respectively. * Asche et al. (2008) also use fixed effects, random
effects and OLS estimators, but the results appear to be rather implausible.

3.2 Causality in the energy/growth nexus

Table 4 displays the four studies that conduct Granger causality analyses between industrial or
residential electricity consumption, economic activity and electricity price.

Table 4
Results from causality analyses on residential electricity demand in the literature
Direction of causality
Study Country Data
Long-run Short-run

Dergiades & Time series,

R US Y, P E PY
Tsoulfidis (2008) 19652006 (annual)

Time series,
Halicioglu (2007)R Turkey Y, P E Y, P E
19682005 (annual)

Time series,
Polemis (2007)I Greece Y, P E E, Y P
19702004 (annual)

Zachariadis &
Time series, Y, P E
Pashourtidou Cyprus EY
R 19602004 (annual) E, P Y
(2007)
R I
Notes: : Residential sector; : Industrial sector; Y: Economic activity; P: Electricity price; E: Electricity consumption; denotes
the direction of causality.

For the industrial sector, Polemis (2007) finds unidirectional long-run causality from economic
activity and electricity price to electricity consumption, while in the short run he finds causality running
from economic activity and income to electricity price. Hence, empirical evidence for the industrial
sector so far is in favor of the conservation hypothesis in the long run.

11
For the residential sector, Dergiades and Tsoulfidis (2008) and Halicioglu (2007) both find
unidirectional long-run causality from income and price to electricity consumption. For the short run,
the former find causality from electricity price to income, while the latter finds causality from income
and price to electricity consumption.10 Zachariadis and Pashourtidou (2007) find evidence for a long-
run causal relationship, running from income and price to electricity consumption, and from electricity
consumption and price to income. For the short run, they only find causality running from electricity
consumption to income. Hence, so far there is empirical evidence for the conservation as well as the
feedback hypothesis concerning the causal relationship between residential electricity demand and
real income.

3.3 Interfuel substitution

For the case of energy demand with different kinds of fuel inputs (e.g., coal, gas, electricity) we use an
approach that is based on production theory, where energy is the output of several factors (i.e., fuels).
Such an approach fits into the context of a flexible functional form that is often used in econometrics to
obtain elasticities (unrestricted ones, i.e., all elasticities of factor substitution do not necessarily add up
to one), which are functions of the second derivatives of production, cost or utility functions.
Commonly, the most popular flexible functional forms are the transcendental logarithmic (translog) and
logit cost share models (e.g. Jones, 1995; Urga and Walter, 2003). Here we focus on the translog
model specification.
Since the 1970s, many authors have used a cost function rather than a production function for
estimating production parameters. More complex model structures (where, e.g., electricity expenditure
in each time period considered is determined according to separate utility functions) have been
proposed recently but most of which present very sophisticated empirical specifications whose
application is beyond the scope of this study.
The work of Christensen et al. (1973) paved the way for the use of translog production possibility
frontiers in energy demand studies. There, the authors set the theoretical basis for further econometric
studies where the translog frontiers provide global approximations to the input and output relations.
Along this avenue, Berndt and Wood (1975) explicitly investigated cross-substitution between energy
and non-energy inputs by employing a translog cost function (i.e., a KLEM model) to investigate
elasticities in US manufacturing (for the period 19471971), and in particular how the energy input
does substitute and complement with other inputs (e.g., labor and capital). Griffin and Gregory (1976)
explored the generality of those results and applied the same methodology with a pooled international
time series data set (time period 19551969) for the manufacturing sector of nine industrialized
countries. Pindyck (1979) conducted further research on the determination of input substitution and,
more specifically, also dealt with interfuel substitution, i.e., he divided expenditures on energy into
expenditures on oil, gas, coal, and electricity, respectively, and studied the industrial demand of ten
countries. Such studies were the first ones to explore the possibilities of the translog method and
triggered a lively debate about appropriate model specifications. Bohi and Zimmerman (1984)
surveyed a number of econometric studies of energy demand behavior and conclude that (at that time)
investigations of commercial and industrial energy use are very much constrained by the lack of
detailed information on how energy is used within these sectors.
During the 1970s the oil price shocks led to considerable increases in energy prices which spurred
research in this field due to the increased policy relevance of energy elasticities. Hesse and Tarkka
(1986) focused their research on the impact of such disturbances and separated the analysis for 9
Western European countries into two periods: 19601972 and 19731980. Kim and Labys (1988)
applied the translog approach for interfactor and interfuel substitutiton in the Korean industrial sector
(by disaggregating it into manufacturing and non-manufacturing sectors). Ibrahim and Hurst (1990)
analyzed energy demand and in particular oil demand in (thirteen) developing countries, in particular,
oil demand, during the 1970s and early 1980s.

10
Dergiades and Tsoulfidis (2008) also consider other variables, such as oil price and cooling and heating degree days, for
which we do not report the results here.

12
 Harris et al. (1993) continued the estimation of demand and substitution elasticities for the UK
economy (disaggregated into 19 sectors), but in comparison with previous studies, extended the
period of time studied to 19601980. Taheri (1994) did an analysis taking into account the marked
price increases of the 1970s and investigates substitution among fuels such as coal, gas and
electricity during that period, by identifying inter- and intra-industry variations of industrial fuel use in
the US. Also for the case of the US, Jones (1995) analyzed interfuel substitution by contrasting the
translog model with a dynamic linear logit model (using data for 19601992), where the latter provided
superior global properties, i.e., a direct, unbiased estimate of the rate of dynamic adjustment.
Filippini (1994), estimated electricity demand by time-of-use with a share equation approach, thus
using a decomposition of the cost function, in order to be able to compute the elasticities of
substitution in a simpler way.
In the tradition at that time, i.e. to use (mostly geographically) aggregated data, Ryan et al. (1996)
examined residential energy demand in a single province of Canada in order to better be able to
observe the asymmetric price response of energy demand; specifically they allowed for interfuel
substitution and used annual data for 19621989. Henley and Peirson (1998) investigated consumer
responses conditional on temperature levels, specifiying a translog model of residential electricity
demand and modeling price and temperature interactively.
Concerning comparisons between models, Zarnikau (2003) compared linear, log-linear and
translog functional forms against non-parametric ones; his results, for cross-sectional household-level
data for the US, suggested that the parametric forms may not be sufficiently flexible to provide valid
results in certain applications. Urga and Walters (2003) contrasted dynamic formulations of the
translog and the logit cost share model and confirmed a poor performance of the former, concluding
that this one cannot be explained only by a mis-specification of the model or the inclusion of non-
energy fuels price-unresponsiveness. Recently, with pooled data from several countries, Roy et al.
(2006) estimated long-run substitution and own-price elasticities for the paper and the iron and steel
sector, and aggregate manufacturing industries (i.e., no interfuel substitution), and reported a wide
variation across countries and industries. Another recent study of interfuel substitution using (static
and dynamic versions of the) translog and logit cost share models is Madlener (2004), with a focus on
Switzerland. The price elasticities and elasticities of fuel substitution for the private and industrial
sector from that study are reported in Table 5 as an illustration and comparison with our results.

Table 5
Price-elasticities and elasticities of substitution for Switzerland, private and industrial sector
Private sector Industrial sector
Input
(19802000) (19701998)
Price elasticities
Oil 0.02 (0.55) 0.43** (5.17)
Electricity 0.08* (2.54) 0.12** (3.46)
Gas 0.61** (2.58) 0.43** (4.64)
Elasticities of substitution
OilElectricity 0.01 (1.00) 0.45** (5.09)
OilGas 0.01 (0.47) 0.02 (0.74)
ElectricityOil 0.02 (1.00) 0.16** (5.09)
ElectricityGas 0.10** (2.87) 0.04** (6.30)
GasOil 0.03 (0.47) 0.05 (0.74)
GasElectricity 0.58** (2.87) 0.38** (6.30)
Notes: t-statistics in parentheses; bold font for those elasticities statistically significant at the 5% (*) and 10% (**) level,
respectively.

4 Econometric methodologies used

Ever since the seminal paper by Engle and Granger (1987), cointegration analysis has increasingly
become the favored methodological approach for analyzing time series data containing stochastic

13
trends. If the data generating processes (DGPs) underlying the time series are integrated of order one,
I(1) (which is the case for most economic variables), or higher, usual regression analysis can lead to
spurious results. Instead of taking first differences of the data, which was the common prior solution
but leads to a loss of long-run information, this problem can be tackled by identifying possibly existing
stationary linear combinations of two or more non-stationary time series. Such stationary linear
combinations indicate common stochastic trends (i.e. cointegration), which can be interpreted as long-
run equilibrium relationships between the variables considered and, therefore, according to the
Granger representation theorem (Engle and Granger, 1987), can be characterized by being generated
through an error correction mechanism.
A methodology for cointegration analysis that has received considerable attention is the maximum
likelihood (ML) system estimation and testing procedure developed by Johansen (1988, 1995). In
contrast to single-equation methods, the Johansen approach does not impose the assumption of a
unique cointegrating vector a priori and efficiently estimates the short-run dynamics simultaneously
along with the long-run relationship. Moreover, restrictions to the cointegration space can be applied
and tested for. Furthermore, Johansen et al. (2000) provide an extension for incorporating structural
breaks in the cointegrating vectors. In a simulation study, Gonzalo (1994) finds superior finite sample
properties of the Johansen ML estimator when compared to four other commonly used estimation
methods in cointegration analysis, even when the dynamics are not known and the errors are non-
Gaussian. This approach is explained in more detail in Section 4.2.1 and applied to industrial
electricity demand on a subsectoral level in Section 5.1.
Most previous studies on the estimation of energy demand elasticities are based on time series
data. Since the introduction of cointegration analysis, the spurious regression problem has been
accounted for in most of these studies. However, traditional unit root and cointegration tests in a pure
time series context are known to suffer from the problem of very low power and size. Hence,
increasing the number of observations by including a cross-sectional dimension helps to reduce this
problem. The added cross-sections can be interpreted as repeated draws from the same distribution
that increase power and hence permit more reliable statistical inference. Amongst others, Pedroni
(1999, 2004) has developed an estimation and testing framework for cointegration analysis based on
panel data. This approach is explained in more detail in Section 4.2.2 and applied to residential
electricity demand in Section 5.2.
An alternative way of circumventing the problem associated with a very low power of unit root tests
is the autoregressive distributed lag (ARDL) bounds testing approach to cointegration. This method,
which was introduced by Pesaran and Shin (1999) and Pesaran et al. (2001), has received
considerable attention over the past years. The advantage of this approach is that information
regarding the order of integration of the variables included in the analysis is not necessarily needed.
Hence, the pretesting for unit roots, which is required for other cointegration approaches, can be
omitted. Rather, the significance of a long-run relationship is tested using critical value bounds, which
are determined by the two extreme cases that all variables are I(0) (the lower bound) and that all
variables are I(1) (the upper bound). This approach is explained in more detail in Section 4.2.3 and
applied to residential natural gas demand in Section 5.3.

4.1 Unit root tests

Before turning to the single cointegration approaches in more detail, this section briefly outlines the
methods used for checking the unit root properties of the variables included in the analysis. This is
necessary as an order of integration of at least one, I(1), is a prerequisite for including them in most
cointegration tests.

4.1.1 Unit root tests based on time series data

It is well known that the standard ADF (Augmented-Dickey-Fuller) and PP (Phillips-Perron) tests suffer
from a considerable loss of power in cases where the DGP underlying a series is a near-unit root
(trend-)stationary process. Furthermore, the existence of structural breaks, if not accounted for,

14
distorts unit root test results (see Perron, 1989). Therefore, we apply more recent efficient unit root
tests, which to a certain extent overcome the deficiencies of the traditional unit root tests: the ERS
(Elliott-Rothenberg-Stock, see Elliott et al., 1996) test for non-breaking series and the LLS (Lanne-
Ltkepohl-Saikkonen, see Saikkonen and Ltkepohl, 2002; and Lanne et al., 2002) test for variables
containing structural breaks, both of which have good power properties compared to alternative tests.
The ERS test is a modification of the ADF test in that the series under consideration is detrended by
using a GLS (Generalized-Least-Squares) regression before the actual unit root test is conducted. The
LLS test works in a similar manner, but also takes into account a level shift in the deterministic term. In
a Monte Carlo simulation study, Lanne and Ltkepohl (2002) show that their LLS test enables
remarkable gains in size and power and performs best in comparison to a number of other unit root
tests that incorporate level shifts at some known points in time.

4.1.2 Unit root tests based on panel data

Recently a number of panel unit root tests have been developed by various authors. Amongst the
most common ones are the LLC (Levin et al., 2002), the UB (Breitung, 2000), the IPS (Im et al., 2003),
the ADF-Fisher (Maddala and Wu, 1999), and the PP-Fisher (Maddala and Wu, 1999).
All the mentioned tests are based on the following AR(1) panel regression model:

xi ,t i xi ,t 1 i X i ,t ui ,t , i = 1, 2, , N; t = 1, 2, , T, (15)

where i are the autoregressive parameters, Xi,t represents exogenous variables and/or fixed effects
and cross-section-specific time trends and ui,t are stationary error terms. In the case that |i| < 1, xi is
referred to be weakly trend-stationary. In contrast, if |i| = 1, xi is considered to be a unit root process.
The five aforementioned tests can be divided into two different groups with respect to the
assumptions about the i. On the one hand, the LLC test and the UB test assume that all cross-
sections have a common unit root, i.e. i = for all i. On the other hand, the IPS test, the ADF-Fisher
test and the PP-Fisher test all assume that the i can be heterogeneous across the cross-sections.
Due to space limitations we refer the reader to the original articles for further details on these tests.
In the case where the test results for a variable in levels indicate a rejection of the null hypothesis,
whereas the test results for the same variable in first differences does not reject the null at
conventional significance levels, this variable is assumed to be integrated of order one, denoted I(1).

4.2 Approaches to cointegration analysis

4.2.1 Johansen Maximum Likelihood

For estimating the industrial electricity demand function outlined in Section 2.2, we apply the Johansen
ML approach. Taking the natural logarithm of Eq. (2) and adding a stochastic error term yields the
linear double-log specification of our econometric model for the long-run electricity demand function

et c0 dt v vt p pt t ,
(16)

where et = ln(Et), vt = ln(Vt), pt = ln(Pt), c0 is a constant, dt is a deterministic time trend and t is the
error term. v and p are the constant elasticities of economic activity and price, respectively, with
regard to electricity demand.
The Johansen ML system approach is briefly outlined as follows: As the basic DGP, consider the
unrestricted three-variable vector autoregression model of order p, VAR(p), defined as

p
Yt Tt St I t Aj Yt j U t , (17)
j 1

15
where Yt is a (3 x 1)-dimensional vector containing the endogenous I(1) variables,
Yt = [et, vt, pt], the Aj are (3 x 3)-dimensional parameter matrices, and Ut is a three-dimensional
Gaussian white-noise process representing the error terms. The deterministics are a constant (), a
linear trend (Tt), a shift dummy (St), and impulse dummies (It). The vectors and and the matrix
contain the corresponding parameters. Eq. (17) can be rewritten as a vector error correction model of
order (p 1), VECM(p 1), specified as

p 1
Yt Yt 1 j Yt j I t U t , (18)
j 1

where = (I A1 Ap) and j = (Aj+1 + + Ap) for j = 1, 2, , (p 1). If the variables in Yt are
indeed cointegrated, has reduced rank, rk() = r, and can be decomposed into = . Here
spans the space of r cointegrating vectors, so that Yt-1 represents up to (k 1) cointegration
relationships, whereas contains the corresponding adjustment coefficients.11
Using the trace test provided by Johansen (1994, 1995) and Johansen et al. (2000), respectively,
depending on whether structural breaks are incorporated or not, it is then possible to determine how
many r (k 1) distinct eigenvalues (i) exist that are significantly different from zero, and hence how
many cointegrating relationships are present in . The likelihood ratio trace statistic (Trace) is given by

k
Trace T ln 1 i , (19)
i r 1

where k is the number of endogenous variables and T is the number of observations. The null
hypothesis is the existence of at most r cointegrating relations (0 r k) against the alternative of (r +
1) cointegrating relations.
If the null hypothesis is rejected on the first level (i.e. r = 0), but accepted on the second level
(i.e. r 1), we can conclude that there exists only one cointegrating vector, which leads to the
following most general specification of the VECM(p 1):12

et 0,e 1,e p 1
et j e ,t

vt 0,v 1,v et 1 v vt 1 p pt 1 Dt 1 j vt j I t v ,t , (20)
p j 1
p
t 0, p 1. p t j p ,t

where the term in squared brackets is the error correction term (ECTt-1). The deterministic term (Dt) in
the most general case is Dt = c + d1t + d2s. The restricted constant (c) is always included, the linear
time trend (t) is included whenever significant, and the shift dummy (s) is included only when a
structural break occurs in one of the series. It is a vector of impulse dummies and is a matrix
containing the corresponding parameters. The j are the (p 1) coefficient matrices for the lagged
differences of the three endogenous variables. Following Ltkepohl and Krtzig (2004, pp.116-119),
we include a shift dummy for the time of any break date (tau) and further intervention dummies for ,
+ i, , + (m 1), where m is equal to the lag length of the corresponding VAR (m = p).
Estimation of Eq. (20) yields the estimates of the long-run equilibrium demand relationship, i.e. the
estimated cointegrating vector, which equals zero in the long-run equilibrium

ECTt et v vt p pt D t , (21)

where ECTt stands for the error correction term, which represents the deviation from the long-run
equilibrium in any period t. In order to attain the short-run elasticities we can then proceed by

11
Note that in the software package JMulTi 4.24, which we use in cases where shift dummies are included in the cointegrating
vector (otherwise we use EViews 6), the constant is restricted to the cointegrating vector.
12
Note that this specification is sufficient, since in all cases where we did find a long-run relationship, our inference has always
led to a model with one cointegrating vector only.

16
estimating a standard single-equation error correction model (ECM) based on the long-run relationship
obtained from estimation of Eq. (20):

l m n
et 0 ECTt 1 e,i et i 1 v ,i vt i p ,i pt i t , (22)
i 0 i 0 i 0

where 0 is a constant, is the loading coefficient, e,i, v,i and p,i are the short-run parameters and t
is a white-noise error term. By deleting insignificant coefficients, a parsimonious specification of the
short-run dynamic equation can be searched for.
An additional advantage of a cointegrated VAR setting is the possibility of further analysis on the
dynamics of the examined relationship. Hence, before estimating the short-run elasticities in the
single-equation ECM framework, we employ tests on Granger-causality and examine the impulse
response functions.

4.2.2 Panel cointegration

For the residential electricity demand relationship described in Section 2.3, we examine the possibility
of using the panel cointegration approach by Pedroni (1999, 2004).
Taking natural logarithms of Eq. (4) and adding an error term and the cross-sectional dimension
(i > 1) yields the econometric specification of our long-run residential electricity demand function:

ei ,t ci y ,i yi ,t p ,i pi ,t i ,t , i = 1, 2, , N; t = 1, 2, , T, (23)

where ei,t = ln(Ei,t), yi,t = ln(Yi,t), pi,t = ln(Pi,t), ci are country-specific fixed effects and i,t are the error
terms, which are interpreted as deviations from long-run equilibria. The country-specific slope
coefficients y,i and p,i are the long-run elasticities to be estimated with regard to income and
electricity price, respectively. Hence, this specification allows for the cointegrating vectors to vary
across the single countries of our panel.
Pedroni (1999, 2004) extends the cointegration testing approach of Engle and Granger (1987),
which is based on examining the stationarity properties of the residuals from a regression using I(1)
variables, to a panel data setting. Following this approach, Eq. (23) is estimated by OLS, and the
residuals obtained, i ,t , are used for the following auxiliary autoregression for every i:

ni
i ,t i i ,t 1 i ,t i ,t j wi ,t , (24)
j 1

Where the i are autoregressive parameters, ni are the lag lengths in the augmented case, and wi are
stationary error terms.
Under the null hypothesis of no cointegration, the i ,t should be found to be I(1). This is the case if

H 0 : i 1, i 1,..., N

is not rejected. For each of the seven statistics provided by Pedroni (1999, 2004) this is the null
hypothesis. Concerning the alternative hypothesis, the tests can be divided into two classes. For the
so-called (within-dimension) panel statistics tests (i.e. the Panel-v, the Panel-PP-, the Panel-PP-t and
the Panel-ADF-t test) the alternative hypothesis is

H1 : i 1, i 1,..., N ,

whereas for the so-called (between-dimension) group statistics tests (i.e. the Group-PP-, the Group-
PP-t and the Group-ADF-t test) the alternative hypothesis is

17
 H1 : i 1, i 1,..., N .

Hence, the group statistics tests are less restrictive in the sense that they allow for heterogeneity
across countries.
Given that the panel cointegration tests indicate a significant cointegration relationship, we apply
the fully modified OLS (FMOLS) and the dynamic OLS (DOLS) group-means panel estimators
proposed by Pedroni (2000, 2001) for estimating the long-run demand relationship characterized by
Eq. (4). Both estimators allow for standard normal inference through incorporating corrections for
endogeneity bias and serial correlation.13 While the FMOLS estimator employs a semi-parametric
correction using yi,t, pi,t and i ,t , the DOLS estimator employs a parametric approach by
augmenting Eq. (23) with lead and lag dynamics of yi,t and pi,t as follows:

li li
ei ,t i y ,i yi ,t p ,i pi ,t y,i,l yi,t l p,i,l pi,t l i,t
l li l li
(25)

where li is the lead and lag length, i is the country-specific fixed effect and i,t is the error term.
The group-means FMOLS and DOLS panel estimates for the slope coefficients, GFMOLS and GDOLS ,

and their corresponding t-statistics,

t GFMOLS and t GDOLS , are calculated as follows:

1 N m
Gm, y y ,i
N i 1
(26)

t
N
1
t Gm, y m
y ,i (27)
N i 1

1 N
Gm, p
N

i 1
m
p ,i (28)

t
N
1
t Gm, p m
p ,i (29)
N i 1

where the superscript m is a place holder denoting either the FMOLS or the DOLS estimation method;
y ,i and p ,i are the country-specific estimates of income and price elasticity, respectively.
Considering both the FMOLS and the DOLS panel approach has the advantage of being able to
provide some evidence on the robustness of our results with regard to the estimation method.
In order to estimate the short-run elasticities and the speed of adjustment to long-run equilibrium,
the residuals from the cointegrating regressions, which resemble the deviation from long-run
equilibrium in any given period t, are used as error correction terms (ECT) in country-specific and
panel error correction models (ECM). The latter takes on the form

ei ,t 0,mi im ECTi ,mt 1 ym,i yi ,t pm,i pi ,t im,t , (30)

where m denotes the estimation method (FMOLS or DOLS), 0,i is a country-specific constant, i is the
speed of adjustment coefficient, ECTi is the aforementioned error correction term lagged by one
period, y,i and p,i are the short-run income and price elasticities, respectively, and i,t the error terms.
Note that in order to ensure an error correction mechanism via adjustments of electricity consumption,
i has to be negative.

13
Harris and Sollis (2003) provide an excellent exposition on this topic.

18
4.2.3 ARDL bounds testing approach

A well-known drawback of testing for nonstationarity in the data generating process of time series is
the very low power of unit root tests. With regard to this problem, a method which has received
considerable attention over the past years is the autoregressive distributed lag (ARDL) bounds testing
approach to cointegration developed by Pesaran and Shin (1999) and Pesaran et al. (2001). The
advantage of this approach is that information regarding the order of integration of the variables
included in the analysis is not necessarily needed. Hence, the pretesting for unit roots, which is
required for other cointegration approaches, can be omitted. Rather, the significance of a long-run
relationship is tested using critical value bounds, which are determinded by the two extreme cases that
all variables are I(0) (the lower bound) and that all variables are I(1) (the upper bound).
Taking natural logarithms of Eq. (6) and adding an error term yields the econometric specification
of our long-run residential natural gas demand function:

gt 0 1t 2 yt 3 pt 4 hddt t (31)

where gt = ln(Gt); yt = ln(Yt); pt = ln(Pt); and hddt = ln(HDDt). The s are the long-run coefficients and t
is a white noise error term.
The first step of the bounds testing approach is to estimate the following unrestricted error
correction model using OLS:

gt c dt 1 gt 1 2 yt 1 3 pt 1 4 hddt 1
k l m n (32)
1,i gt i 2,i yt i 3,i pt i 4,i hddt i ut
i 1 i 1 i 1 i 1

where the are the long-run multipliers, c is a drift term, are the short-run coefficients and ut is a
white noise error term. Due to the fact that it is not clear a priori whether y, p and hdd are the long-run
forcing variables for natural gas consumption, current values of y, p and hdd are excluded from
Eq. (32).
As a second step, an F-test on the joint hypothesis that the long-run multipliers of the lagged level
variables are all equal to zero against the alternative hypothesis that at least one long-run multiplier is
non-zero is conducted, i.e.:

H0: 1 = 2 = 3 = 4 = 0;

H1: 1 0, or 2 0, or 3 0, or 4 0.
Critical values are provided by Pesaran and Pesaran (2009, p.544). These depend on the number of
regressors and the deterministic terms included. For each of the conventional significance levels, two
sets of critical values are given, which constitute the lower and the upper bound. The lower bound
represents the critical values for the case in which all included variables are assumed to be I(0), while
the upper bound assumes all the variables to be I(1). Hence, all possible combinations of orders of
integration for the single variables are covered. If the calculated F-statistic lies above the upper bound,
the null hypothesis of no cointegration can be rejected, irrespective of the number of unit roots in the
single variables. On the other hand, if it lies below the lower bound, the null hypothesis is not rejected.
Only if the F-statistic lies between the bounds, the result of the inference is inconclusive, given that the
order of integration of the single variables is unknown.
If the existence of a significant cointegration relationship is identified by the bounds F-test, the next
step is to select the optimal ARDL specification of Eq. (32). This process is guided by the Akaike
Information Criterion (AIC) and the Schwarz Bayesian Criterion (SBC). Furthermore, the properties of
the residuals are checked to ensure the absence of serial correlation.
A representation of the ARDL(k,l,m,n) model in the general case is

19
 k l m n
gt c d t 1,i g t i 2,i yt i 3,i pt i 4,i hddt i wt , (33)
i 1 i 0 i 0 i 0

where wt is an error term and k, l, m, and n are the lag lengths of the single variables.
The long-run coefficients are constructed as non-linear functions of the parameter estimates of
Eq. (33) as follows:

c
0 , (34)
k

1,i
1
i 1

d
1 and (35)
k

1,i
1
i 1
q

j ,i
j i
(36)
k

1,i
1
i 1

with j = 2, 3, 4 and q = k, l, m, n. 0 and 1 are the constant and the deterministic trend in the long-run
model, Eq. (31), respectively. The j are the long-run slope coefficients.
Finally, the (dynamic) short-run coefficients for the error correction representation are estimated
according to

k l m n
gt c d t ect ECTt 1 1,i gt i 2,i yt i 3,i pt i 4,i hdd t i vt , (37)
i 1 i 1 i 1 i 1

where ECTt-1 is the error correction term resulting from the estimated long-run equilibrium relationship,
Eq. (31), and ect is the coefficient reflecting the speed of adjustment to long-run equilibrium, i.e. the
percental annual correction of a deviation from the long-run equilibrium the year before.

4.3 Granger causality testing

A cointegration relationship between a set of variables necessarily implies Granger causality in at least
one direction. For the industrial electricity demand relationship in Eq. (2) and the residential electricity
demand relationship in Eq. (4), the direction of Granger causality between the single variables is
tested by the two-step Engle-Granger procedure (Engle and Granger, 1987) on the basis of the
following (panel)VECM (PVECM):14

k k k
ei ,t 1,mi 1,mi ECTi ,mt 1 11,
m
i , j ei ,t j 12,i , j yi ,t j 13, i , j pi , t j 1, i , t
m m m
(38a)
j 1 j 1 j 1

k k k
yi ,t 2,mi 2,mi ECTi ,mt 1 21,
m
i , j ei , t j 22, i , j yi , t j 23,i , j pi , t j 2, i , t
m m m
(38b)
j 1 j 1 j 1

14
Note, that for the industrial electricity demand the cross-sectional dimension equals one, i = 1.

20
 k k k
pi ,t 3,mi 3,mi ECTi ,mt 1 31,
m
i , j ei ,t j 32, i , j yi , t j 33, i , j pi , t j 3, i ,t ,
m m m
(38c)
j 1 j 1 j 1

where m again denotes the estimation method, ECTi is the error correction term (obtained from
estimation of Eq. (4)), k is the lag length, the i are the speed of adjustment coefficients, the i are
country-specific constants, the i are short-run coefficients and the i,t are serially uncorrelated error
terms. In the panel case, i > 1, the equation system (38a-c) is estimated by the pooled mean group
estimator (PMGE) method proposed by Pesaran et al. (1999).
For the long run, the direction of Granger causality can be checked by testing whether the
adjustment coefficient of the respective equation is significantly different from zero. Besides long-run
Granger causality, short-run Granger causality can occur through the lagged first differences of the
independent variables in each equation of the system. Hence we check for short-run Granger causality
from income and electricity price to electricity consumption by testing H0 : 12,i,j = 0, i,j and H0 : 13,i,j =
0, i,j in Eq. (38a). For short-run Granger causality from electricity consumption and electricity price
to income in Eq. (38b), we test H0 : 21,i,j = 0, i,j and H0 : 23,i,j = 0, i,j, respectively. In Eq. (38c) we
test H0 : 31,i,j = 0, i,j and H0 : 32,i,j = 0, i,j for short-run Granger causality running from electricity
consumption and income to electricity price, respectively. Finally, we test on the joint significance
between the lagged ECT and the lagged differences of the independent variables in the respective
equations in order to check for strong causality in each equation.

4.4 Translog model

The system of share equations from the translog model can be estimated by a seemingly unrelated
regression (SUR) model. To make the model operational we must impose the restrictions in the
parameter matrix (i.e., column sums must be equal to zero and row sums equal to zero) and solve the
problem of singularity of the disturbance covariance matrix of the share equations (Greene, 2003). We
compute maximum likelihood estimates of the coefficients to ensure invariance of the results with
respect to the choice of which share equation to drop.
More specifically, the estimable system of cost share equations S, derived from the translog cost
function having four inputs (electricity, E; oil, O; gas, G; and coal, C) is obtained by dropping one share
equation (e.g. coal), yielding:

P PO PG
S E E E , E ln E E ,O ln E ,G ln (39a)
PC PC PC

P PO PG
SO O O , E ln E O ,O ln O ,G ln (39b)
PC PC PC

P P P
SG G G , E ln E G ,O ln O G ,G ln G . (39c)
PC PC PC

Since the elasticities will differ at every data point, we decided to compute them at the means of the
cost-share data.
In order to avoid bias in estimation, we take into account neutral and non-neutral efficiency
differences among the sectors studied, e.g., those related to time (as a proxy variable for technological
change in energy consumption). This is achieved by including time as a proxy for technical change (i,t
ln t) in each of the three share equations. Hence, the coefficient i,t is an estimator of the rate of
technical change: if the coefficients were zero, time alone would not affect the factor shares. Also, as
the equations are related by their error terms, we estimate the model using generalized least squares

21
(GLS) estimation, which gives more efficient estimators (compared to ordinary least squares
regression).
As can be seen from Eqs. (39ac) the regressors are the natural logarithms of the ratio of the price
considered divided by the price of the fuel whose share equation was dropped (in our case the price of
coal). To obtain the cost shares for each factor we use the final energy consumption and the real index
of end-use prices for each fuel in the respective sector or the whole economy (see Section 5.4).

5 Empirical analysis

5.1 Industrial electricity demand on the subsectoral level in Germany

5.1.1 Data

The data required for our analysis encompasses long time series of electricity consumption, a
measure for the level of economic activity, such as the real value added, and the real electricity price.
The International Energy Agency (IEA) provides data for electricity consumption at a subsectoral level
as well as industrial electricity prices, whereas the EU-KLEMS database (November 2009 release;
www.eu-klems.org) offers subsectoral data on gross value added (VA) volume indices (1995 = 100).

Table 6
Data availability, aggregational groupings and matches
Sector IEA: Electricity consumption EU-KLEMS: VA, VA price index
identifier NACE Description NACE Description
Total Manufacturing Industry: 15-22, 24-37
[1] 15-16 Food & tobacco 15-16 Food, beverages & tobacco
15 Food & beverages
16 Tobacco
[2] 17-19 Textile & leather 17-19 Textiles, textile, leather & footwear
17 Textiles
18 Wearing apparel, dressing & dying of fur
19 Leather, leather products & footwear
[3] 20 Wood & wood products 20 Wood & of wood & cork
[4] 21-22 Paper, pulp & printing 21-22 Pulp, paper, printing & publishing
21 Pulp, paper & paper products
22 Printing, publishing & reproduction
[5] 24 Chemical & petrochemical 24 Chemicals & chemical products
[6] 26 Non-metallic minerals 26 Other non-metallic minerals
[7] 27.1-3; Iron & steel 27-28 Basic metals & fabricated metal
27.51-52 27 Basic metals
28 Fabricated metal
27.4; Non-ferrous metals
27.53-54
28-32 Machinery

29 Machinery, NEC
30-33 Electrical & optical equipment
30 Office accounting & computing machinery
31 Electrical machinery & apparatus
32 Radio, television & communic. equipment
MISMATCH 33 Medical, precision & optical instruments
[8] 34-35 Transport equipment 34-35 Transport equipment
Notes: Data on sectors printed in italics are only available as part of a higher aggregation level.

Most studies on energy demand use the consumer price index (CPI) for deflating nominal energy
prices. For an analysis at the aggregate (national) level, this is an appropriate approximation.
However, the price underlying the decision-making process of an economic agent is the energy price
relative to the prices of all other goods and services relevant to its economic activity. On a sectoral
level, this is best approximated by the sector-specific value added price index. Consequently, we

22
23
obtain real electricity prices through deflating the nominal industrial electricity price by using the
sector-specific value added price indices from the EU-KLEMS database.15
Our aim, i.e. to analyze industrial branches on the lowest aggregational level possible, was guided
by data availability. As two different databases were used as sources, matches concerning groupings
had to be constructed, using NACE (Rev. 1) classifications. Table 6 provides an overview of the data
availability, short descriptions of the different subsectors investigated, the aggregational groupings,
and the resulting matches with regard to the two databases. Due to differing aggregational groupings
concerning the iron, metals and machinery subsectors (i.e. NACE codes 27-33) in the two data sets,
these subsectors were aggregated to what we label as the Metal & Machinery [7] sector. Moreover, it
turned out to be impossible to construct a perfect data match for this sector. More specifically, and in
contrast to the value added data, the aggregate electricity consumption data does not cover the
Medical, precision & optical instruments sector (NACE code 33).
The time series for nominal electricity price is available from 1960 onwards. The same applies to
the electricity consumption variable for most of the industries. Nevertheless, data on real value added
and the value added price index only start in 1970. Hence, the longest time span completely covered
by the data is from 1970 until 2007. The German reunification in 1990 is handled by imposing the
trends for West Germany to the series of unified Germany in 1991.
Figure 1 displays the individual time series for all the subsectors considered. Visual inspection
reveals the following trends in the series: Both electricity consumption and real value added generally
show an increasing trend over the time span under consideration and for nearly all the subsectors.
Exceptions are the Textile & Leather [2] industry, where both variables decrease and the Food &
Tobacco [1] industry, where value added more or less stagnates over the entire period. In addition,
some of the series show large sudden level shifts. This is the case for electricity consumption in the
Chemicals [5] sector and the Metal & Machinery [7] sector in 1991, and for real value added in the
Metal & Machinery [7] sector and the Transport equipment [8] sector in 1993. The reasons for these
level shifts are not apparent, but at least the break in electricity consumption could be due to the
closure of energy-intensive industries in East Germany after German reunification in 1990. These
breaks will be accounted for in the unit root and cointegration analysis of the respective sectors.
From 1973 till 1975, and then again from 1979 till 1982/83, real electricity prices experienced
strong increases, coinciding with the oil crises of 1973/74 and 1979/80. In April 1998 the German
electricity market was liberalized, resulting in a gradual decrease of industrial electricity price of
approximately 50% by the year 2000. After this initial liberalization, the power supply industry
experienced a phase of market concentration, leaving the supply side with four major players (E.ON,
RWE, EnBW and Vattenfall) to which approximately 80% of power generation can be attributed, and a
number of minor regional utilities. On account of this oligopoly formation and hence a rise in market
power, but also due to some newly introduced or increased taxes and charges, electricity prices have
surged ever since.

5.1.2 Unit root tests

To check for the non-stationary behavior of the individual time series, as a first step we apply unit root
tests. Specifically, we apply the ERS test to the non-breaking series and the LLS test to the breaking
series. The break dates for the LLS test are chosen according to the arguments in the last section.
The results are summarized in Table 7. Based on the visual inspection of Figure 1, we include a
constant and a deterministic trend in the test regression for electricity consumption (e) and value
added (v) variables in levels, and only a constant for the tests on the first differences. For the price (p)
variable we include a constant in levels and no deterministic term in first differences. The lag lengths
were chosen according to the conventional information criteria. For none of the series in levels the null
hypothesis of a unit root can be rejected at the 10% level. For the first differences, the null hypothesis

15
Hence, the same double deflation measure is used for both the nominal electricity price and nominal gross value added, as
the value added volume indices are constructed by deflating nominal value added with the value added price index and then
normalizing to 1995 = 100.

24
is rejected on the 1% level. Thus, we conclude that all examined series are integrated of order one,
I(1).

Table 7
Unit root tests
ERS test LLS test
Sector Variable
Levels Differences Levels Differences
[1] e 1.365 (2) 4.094*** (1)
v 1.310 (2) 4.490*** (1)
p 1.900 (1) 4.048*** (0)
[2] e 1.365 (0) 4.309*** (0)
v 2.009 (1) 4.142*** (0)
p 1.916 (1) 4.161*** (0)
[3] e 2.498 (2) 7.385*** (1)
v 1.548 (0) 5.560*** (0)
p 2.394 (1) 4.365*** (0)
[4] e 1.939 (3) 4.707*** (2)
v 2.427 (1) 4.477*** (0)
p 1.820 (1) 4.432*** (0)
[5] eSB 1.271 (3) 7.158*** (1)
v 2.196 (2) 7.592*** (0)
p 2.570 (2) 5.430*** (0)
[6] e 2.281 (2) 6.325*** (1)
v 2.394 (0) 5.176*** (0)
p 1.934 (1) 4.363*** (0)
SB
[7] e 1.198 (0) 4.156*** (0)
vSB 2.363 (0) 5.316*** (0)
p 1.286 (0) 4.373*** (0)
[8] e 2.833 (1) 7.942*** (0)
SB
v 2.589 (1) 8.486*** (0)
p 1.911 (1) 4.200*** (0)
SB
Notes: We report t-statistics. *** denotes significance at the 1% level. Lag lengths are in parentheses. denotes series
containing structural breaks. Critical values for the LLS test are from Lanne et al. (2002). For the ERS test the critical values for
the models in levels are taken from Elliott et al. (1996, Table 1), while the critical values for the models in differences are from
MacKinnon (1996).

5.1.3 VAR specification

Having detected non-stationary behavior for all the series, we include all of them in the corresponding
sector-specific cointegration models. We start by specifying a VAR model in levels for each subsector
as in Eq. (17), with the aim to ensuring Gaussian residuals. For the optimal choice of lag length, the
conventional information criteria were only used as a rough guideline.16 Instead, the main focus was
on ensuring that the single VAR models pass the diagnostic tests for autocorrelation. Moreover, we
also conducted tests for non-normality and heteroscedasticity. To attain normality of the residuals, we
accounted for outliers by using impulse dummies (I) for the respective years. Shift dummies (S) and
corresponding impulse dummies are inserted following the arguments provided in Section 5.1.1.
Table 8 shows the VAR specifications and the results of the respective diagnostic tests.17 Most
models pass the set of diagnostic tests. Exceptions are models [3] and [8], where the Jarque-Bera test
indicates problems with the normality assumption. As this is due to residual kurtosis, which in contrast
to residual skewness (Paruolo, 1996) does not invalidate the test results, we conclude that there is no
significant deviation from the theoretical model assumptions.

16
On account of the overparameterization problem associated with VAR models, we set the maximum lag length to three.
17
We also run tests on heteroscedasticity. Here only model [1] shows slight signs of model defects. As the trace test is quite
robust to heteroscedasticity, however, we decided to ignore this deficiency.

25
Table 8
Vector autoregression (VAR) specification and diagnostic tests
Diagnostic tests
Sector VAR(p), Deterministic terms
LM(2): LM(4): PAR(8): JB:
[1] VAR(1), C, T, I94 0.1873 0.7334 0.3689 0.3664
[2] VAR(3), C 0.1150 0.2905 0.2319 0.5629
[3] VAR(1), C, I93 0.8668 0.5030 0.8430 0.0347
[4] VAR(1), C, I75, I91 0.1188 0.8530 0.3526 0.7432
SB
[5] VAR(2), C, T, S91, I91, I92 0.3285 0.4291 0.3261 0.3564
[6] VAR(1), C, T, I93, I94, I00 0.1582 0.1305 0.1856 0.1969
[7]SB VAR(2), C, T, S93, I93, I94 0.1320 0.1245 0.6810 0.7830
[8]SB VAR(1), C, S93, I93, I00 0.5565 0.7643 0.2388 0.0943
SB
Notes: indicates models which account for structural breaks by inclusion of shift dummies. p denotes the lag length. C and T
denote a constant and linear deterministic time trend, respectively. IYY and SYY indicate the year (YY) when an impulse dummy
and shift dummy is included, respectively. The diagnostic tests are a Lagrange Multiplier test (LM) for no serial correlation at 2nd
and 4th lag, a Portmanteau test (PAR) for no autocorrelation up to lag 8, and the Jarque-Bera test (JB) for multivariate normally
distributed residuals (orthogonalization method: Cholesky of covariance). For brevity we only report p-values.

5.1.4 Cointegration rank tests

In a next step we apply the trace test outlined in Section 4.2.1 to the VAR models specified in the
previous section. Table 9 provides an overview of the test results. According to the trace statistics, for
five of the eight models, the null hypothesis of no cointegrating vector (r = 0) can be rejected at the 5%
level, while the null hypothesis of at least one cointegrating vector (r 0) cannot be rejected at the
10% level. Thus we conclude that the rank is one, i.e. a unique cointegration relationship, for five of
the eight sectors, namely Food & Tobacco [1], Pulp & Paper [4], Chemicals [5], Non-metallic minerals
[6] and Transport equipment [8]. For all other models the null of no cointegration cannot be rejected at
the 10% significance level. We do not infer that the lack of significant cointegration relationships for
some sectors necessarily contradicts pure economic theory. Rather, this probably reflects the missing
strength of the long-run relations in some sectors. Furthermore, problems with data quality might also
be an issue.

5.1.5 VECM estimation (long-run elasticities)

After having determined the sectors with significant cointegration relationships, the next step is to
estimate the corresponding VECMs given by Eq. (20), in order to attain the long-run parameter
estimates. These and the equation-specific adjustment coefficients are reported in Table 10, along
with the corresponding t-statistics. Overall, the statistical significance and the plausibility of the
cointegrating vector estimates in terms of sign and magnitude indicate that we have obtained
reasonable estimates of the true equilibrium relationship.
First, for most models the elasticity estimates turn out to be economically reasonable with regard to
sign and magnitude. However, for sectors [1] and [5] the price coefficient is far from significant at
conventional levels. Thus, as a double check we perform tests for beta restrictions on the cointegrating
vectors of these two models, in order to assess the adequacy of restricting price to zero (p = 0). The
results indicate that these restrictions are valid and, hence, we estimate the restricted models that are
also reported in Table 5. The elasticity estimates of economic activity are 0.70 in sector [1] (Food &
Tobacco), 1.90 in sector [4] (Pulp & Paper), 1.11 in sector [5] (Chemicals), 1.01 in sector [6] (Non-
metallic minerals) and 1.00 in sector [8] (Transport equipment). The corresponding elasticity estimates
of electricity price are zero in sectors [1] (Food & Tobacco) and [5] (Chemicals), 0.52 in sector [4]
(Pulp & Paper), 0.30 in sector [6] (Non-metallic minerals) and 0.30 in sector [8] (Transport
equipment). This corroborates estimates from previous studies, where elasticities of economic activity
are normally found to be close to unity and price elasticities range between approximately zero and
0.50 (see Table 1).

26
Table 9
Cointegration tests with and without level shifts
Panel A: Without level shift
Sector H0: H1: Trace [p-value] Critical values: 95% Inference
[1] r=0 r1 49.65*** [0.0093] 42.92
r1 r2 17.62 [0.3701] 25.87 Rank() = 1
r2 r=3 3.05 [0.8700] 12.52
[2] r=0 r1 26.77 [0.1073] 29.80
r1 r2 7.65 [0.5032] 15.49 Rank() = 0
r2 r=3 1.41 [0.2344] 3.84
[3] r=0 r1 24.54 [0.1784] 29.80
r1 r2 6.85 [0.5952] 15.49 Rank() = 0
r2 r=3 1.99 [0.1578] 3.84
[4] r=0 r1 45.57*** [0.0004] 29.80
r1 r2 10.46 [0.2471] 15.49 Rank() = 1
r2 r=3 3.20 [0.0736] 3.84
[6] r=0 r1 77.07*** [0.0000] 42.92
r1 r2 23.88 [0.0868] 25.87 Rank() = 1
r2 r=3 6.35 [0.4166] 12.52
Panel B: With level shift
Sector H0: H1: Trace [p-value] Critical values: 95% Inference
[5] r=0 r1 50.08** [0.0208] 46.44
r1 r2 26.40 [0.0840] 28.30 Rank() = 1
r2 r=3 6.39 [0.5384] 13.91
[7] r=0 r1 42.82 [0.1098] 46.50
r1 r2 19.24 [0.4096] 28.39 Rank() = 0
r2 r=3 6.20 [0.5742] 13.93
[8] r=0 r1 41.68** [0.0317] 40.28
r1 r2 22.68 [0.0935] 24.45 Rank() = 1
r2 r=3 9.28 [0.1843] 12.79
Notes: * and ** denote significance on the 1% and 5% level, respectively. Critical values are obtained by computing the
response surfaces according to MacKinnon et al. (1999) in Panel A and Johansen et al. (2000) in Panel B.

Table 10
Estimated long-run relationships
Panel A: Without level shifts
Cointegrating vector Adjustment coefficients
Sector
et vt pt Const. Trend e v p
1.000 0.776 0.070 19.40 0.024 0.487 0.258 0.177
[1]
(5.815) (1.247) (17.889) (5.959) (1.920) (0.747)
[1]R 1.000 0.703 _______ 19.40 0.025 0.492 0.305 0.265
(6.610) (34.413) (5.472) (2.187) (1.069)

[4]
1.000 1.899 0.516 17.07 _______ 0.201 0.085 0.239
(6.300) (4.906) (2.839) (2.497) (1.052)
1.000 1.011 0.300 19.81 0.007 0.772 0.077 0.028
[6] (5.413) (3.634) (5.404) (6.163) (0.602) (0.351)
Panel B: With level shifts
Cointegrating vector Adjustment coefficients
Sector
et vt pt Const. Trend Shift e v p
1.000 1.045 0.089 20.44 0.027 0.233 0.733 0.450 0.090
[5] (5.629) (1.382) (29.08) (4.915) (7.538) (3.709) (1.428) (0.576)

[5]R 1.000 1.106 ______ 20.56 0.028 0.242 0.733 0.450 0.090
(5.976) (28.75) (5.032) (7.782) (3.709) (1.428) (0.576)

[8]
1.000 0.998 0.302 19.96 ______ 0.107 0.308 0.104 0.200
(8.113) (1.705) (17.89) (1.430) (2.193) (0.937) (1.524)
Notes: t-statistics in parentheses. Results in Panel A are computed with Johansens maximum likelihood estimation using
EViews 6. Results in Panel B are computed with a simple two-step method (see Ahn and Reinsel, 1990), which allows to apply
R
tests on beta restrictions using JMulTi 4.24. denotes the restricted models: In sector [1] the binding restriction that p = 0 was
tested through a LR test for beta restrictions with one degree of freedom, resulting in a test statistic of 1.290 and a
corresponding p-value of 0.256. In sector [4] the binding restriction that p = 0 was tested through a Wald test for binding beta
restrictions with one degree of freedom, resulting in a test statistic of 1.036 and a corresponding p-value of 0.309. Hence, the
restricted models are not rejected.

27
 Second, the time trends are retained only in the models of sectors [1] (Food & Tobacco), [5]
(Chemicals) and [6] (Non-metallic minerals) where they show significance but with differing signs. In
sectors [1] (Food & Tobacco) and [6] (Non-metallic minerals) the time trend has a positive sign,
implying an increasing net effect on electricity intensity by exogenous factors, such as technical
progress or changes in the structure of industry over time. In contrast, in sector [5] (Chemicals) the
trend picks up a decreasing net effect on the electricity intensity of these factors.
Third, the adjustment coefficients also reveal plausible signs and magnitude whenever they are
significantly different from zero, thereby establishing the error correction mechanism, which forces the
system to its long-run equilibrium. In all the VECMs, electricity consumption adjusts negatively to a
deviation from long-run equilibrium, with the speed of adjustment ranging between an annual error
correction of about 20% and 77% per year, depending on the sector. Hence, a near-complete
adjustment (of at least 95%) to long-run equilibrium induced by electricity consumption would take
approximately fourteen years in the slowest case and three in the fastest case. The level of economic
activity, on the contrary, adjusts (positively) to errors from equilibrium only in the sectors [1] (Food &
Tobacco) and [4] (Pulp & Paper), though with a less pronounced speed. In all other models,
adjustments take place only via electricity consumption.

5.1.6 Granger causality

A cointegration relationship necessarily implies the existence of Granger-causality in at least one

direction. Besides long-run Granger-causality, which is reflected by the adjustment coefficients, short-
run Granger-causality can occur through the lagged differences of the independent variables in each
equation of the system. Table 11 summarizes the results and inference on long- and short-run
Granger-causality. Besides the t-statistics of the ECT for long-run Granger-causality, we present Wald
test statistics on the joint significance of the lagged independent variables in the respective equations.
Note that this only applies to the model of sector [5], as in all other cases the VECMs have an order of
zero. For all models Granger-causality from value added and electricity price to electricity consumption
can be established in the long run. Further, a Granger-causal relation in the same direction is detected
by the Wald test in model [5] for the short run as well. Conversely, value added is Granger-caused by
electricity consumption and electricity price in the long run, only in models [1] (Food & Tobacco) and
[4] (Pulp & Paper). The only variable that is Granger-caused neither in the short nor in the long run is

Table 11
Granger-causality analysis
Long-run cause Joint short-run cause Inference
Effect {vt & pt}e
ECT {et & pt}v Long-run Short-run
variables
{et & vt}p causality causality
[1] et 5.959***
vt 1.920* v e

pt 0.747 e v

[4] et 2.839***
2.497** v&pe
vt
pt 1.052 e&pv

[5] et 3.709*** {3.483** [0.012]}e
vt 0.576 {1.971 [0.109]}v v e v&pe
pt 1.428 {1.576 [0.192]}p
[6] et 6.163***
vt 0.351 v&pe
pt 0.602
[8] et 2.193**
vt 0.937 v&pe
pt 1.524
Notes: ECT stands for error correction term. The second column (long-run cause) presents t-statistics, while the third column
(joint short-run cause) presents F-statistics with the corresponding p-values in brackets. denotes the direction of Granger-
causality. ***, ** and * denote significance on the 1%, 5% and 10% level, respectively.

28
electricity price. Thus, electricity price can be considered to be strongly exogenous in the subsectoral
demand functions. This of course complies with intuition, as subsectoral demand is only a small
fragment of aggregate demand and, hence, is not likely to affect the process of electricity price
formation.
In reference to the aforementioned hypotheses regarding the energy-growth nexus, our results
indicate evidence for the feedback hypothesis in sectors [1] (Food & Tobacco) and [4] (Pulp & Paper),
and the conservation hypothesis in the sectors [5] (Chemicals), [6] (Non-metallic minerals) and [8]
(Transport equipment). This means that in the three latter sectors conservation policies would not
adversely affect economic growth, whereas in the former two sectors such policies would imply a
trade-off concerning growth.

5.1.7 Impulse response functions

In order to trace out the dynamic behavior of electricity demand in response to one-time innovations in
the two other variables, we perform an impulse response analysis on the basis of the estimated
VECMs. The impulse responses are displayed in Figure 2. All of them show a plausible behavior.
Electricity consumption responds positively to an impulse in value added and negatively to an impulse
in electricity price. Shocks do not appear transitory but have long-term impacts, which makes sense as
the variables are found to be integrated of order one, I(1). Given that no further shocks hit the system,
convergence to the new long-run level occurs after approximately three to fourteen periods, depending
on the sector-specific speed of adjustment to long-run equilibrium, as reflected by the loading
coefficients reported in Table 10.
.03 .06

.05
.02 .04

.03
.02
.02

.01
.01
.00

.01 -.01

-.02

.00 -.03
2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20

(a) Response of e to an impulse in v [1] (b) Response of e to an impulse in v and p [4]

.04 .04

.03
.03

.03
.02

.02
.01
.02

.00
.01

-.01
.01

.00 -.02
2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20

(c) Response of e to an impulse in v [5] (d) Response of e to an impulse in v and p [6]

.03

.03

.02

.02

.01

.01

.00

-.01

-.01
2 4 6 8 10 12 14 16 18 20

(e) Response of e to an impulse in v and p [8]

Fig. 2: Orthogonalized impulse responses of electricity consumption (e) to Cholesky one standard deviation
innovations in real value added (v) and real electricity price (p), based on the estimated VECMs for the subsectors
[1], [4] - [6] and [8].

29
30
 12 1.6

8
1.2

4
0.8

0.4
-4

0.0
-8

-12 -0.4
96 97 98 99 00 01 02 03 04 05 06 07 96 97 98 99 00 01 02 03 04 05 06 07

(a) Food & Tobacco [1]

12 1.6

8
1.2

4
0.8

0.4
-4

0.0
-8

-12 -0.4
92 93 94 95 96 97 98 99 00 01 02 03 04 05 06 07 92 93 94 95 96 97 98 99 00 01 02 03 04 05 06 07

(b) Pulp & Paper [4]

12 1.6

8
1.2

4
0.8

0.4
-4

0.0
-8

-12 -0.4
94 95 96 97 98 99 00 01 02 03 04 05 06 07 94 95 96 97 98 99 00 01 02 03 04 05 06 07

(c) Chemicals [5]

8 1.6

6
1.2
4

2
0.8

0.4
-2

-4
0.0
-6

-8 -0.4
01 02 03 04 05 06 07 01 02 03 04 05 06 07

(d) Non-metallic minerals [6]

12 1.6

8
1.2

4
0.8

0.4
-4

0.0
-8

-12 -0.4
96 97 98 99 00 01 02 03 04 05 06 07 96 97 98 99 00 01 02 03 04 05 06 07

(e) Transport equipment [8]

Fig. 3: CUSUM and CUSUMSQ plots for the estimated ECMs

5.1.8 Short-run elasticities

Given the results of the preceding estimation and testing procedure, the estimated cointegrating
vectors can be used in the estimation of the ECMs according to Eq. (22), in order to quantify the short-
run elasticities. As the first differences as well as the EC terms are all I(0), the equations are balanced.
We start by estimating unrestricted ECMs with a maximum lag length of three and stepwise delete
insignificant covariates. Outliers are modeled using impulse dummies. The model properties are
checked using a battery of diagnostic tests. More precisely, we apply a Lagrange Multiplier test for
serial correlation of order one and two, a Portmanteau test for autocorrelation up to four lags, the
Breusch-Pagan-Godfrey test for heteroscedasticity, the Jarque-Bera test for normality and the Ramsey

31
RESET test for functional form misspecification. All models pass the diagnostic tests, indicating no
significant deviations from the desired model properties. The estimated equations and diagnostic test
results are summarized in Table 12. Further, we check for parameter instability by applying the
cumulative sum of recursive residuals (CUSUM) and CUSUM of squares (CUSUMSQ) tests (see
Brown et al., 1975). Figure 3 displays the results for the single ECMs, which overall indicate parameter
stability. As can be seen only the CUSUMSQ plot for the Transport equipment [8] model slightly
crosses the lower 5% critical bound at one point.
The coefficients of the error correction terms have the right sign and are significant in all the
models. The magnitudes vary between 0.20 in the model for sector [8] (Transport equipment) and
0.83 in the model for sector [6] (Non-metallic minerals). The short-run demand elasticity of economic
activity has the expected sign and is significant in all models. Here the magnitudes are 0.17 in sector
[1] (Food & Tobacco), 1.02 in sector [4] (Pulp & Paper), 0.74 in sector [5] (Chemicals), 0.51 in sector
[6] (Non-metallic minerals) and 0.48 in sector [8] (Transport equipment). The sizes of the short-run
elasticities are in accordance with the corresponding long-run elasticities, in the sense that when the
long-run elasticity is relatively high in a sector, the short-run elasticity is relatively high as well. The
short-run elasticity of electricity price, in contrast, is significant only in the models for sectors [6] (Non-
metallic minerals) and [8] (Transport equipment), with magnitudes of 0.57 and 0.31, respectively.
This leads to a peculiar situation in sector [6] (Non-metallic minerals), in that the price elasticity in the
short run is of a higher magnitude than in the long run. The constant in the model of sector [8]
(Transport equipment) is only significant at the 12% level. However, we decide to keep it in the
specification for technical reasons.

5.2 Residential electricity demand in OECD countries

5.2.1 Data

We make use of annual data on residential electricity consumption, net disposable income and
residential electricity price, which are all available for a reasonable length of time (19782008) for the
eighteen OECD countries already mentioned in the introduction.
The time series are transformed as follows: In order to obtain real values, disposable income and
residential electricity price are deflated to the 2000 levels using the consumer price index (CPI) of the
respective country. After that, the 2000 exchange rate conversion factors are used to standardize both
the series for every non-Euro country to . Electricity consumption and disposable income are divided
by total population in order to get per capita values for each country. Finally, all the series are log-
transformed.18
Data on residential electricity consumption and residential electricity price are taken from the
International Energy Agency (IEA) database (http://www.iea.org/stats) Energy Balances of OECD
Countries and Energy Prices & Taxes, respectively, while the currency exchange rates are from the
European Central Bank (ECB, http://www.ecb.int/stats). All other variables, i.e. national net disposable
income, total population and CPIs, are adopted from the OECD database (http://stats.oecd.org).
The second oil price shock in 1979/1980 will most likely have introduced exogenous structural
breaks to the DGP of the time series at hand. In order to circumvent this problem, we truncate the
initially available period (19782008) accordingly, which leaves us with a sample ranging from 1981 to
2008 for every country. This yields a panel with the dimensions N = 18 and T = 28. Hence, our
analysis is based on 504 observations in total.
Figure 4, plots (a) (c), displays the individual time series of electricity consumption (measured in
tons of oil equivalent, toe) per capita, real disposable income (measured in constant 1000 , base year
2000 = 100) per capita, and real electricity price (measured in constant 1000 /toe, base year 2000 =
100), respectively.

18
Note that the German reunification in October 1990 is treated by imposing trends for West Germany to the series of unified
Germany in 1991.

32
33
 Both, electricity consumption and real income show an overall upward trend in all countries
considered. The decreasing effect of the latest financial crisis in 2008 on the level of these variables
can be seen in nearly all of the income series. One anomaly worth mentioning is the Norwegian
electricity consumption, which features very high per capita values and reveals a very high volatility
compared to other countries. Both characteristics presumably are due to the fact that Norwegian
households predominantly (approximately 70%) use electricity for heating purposes, which of course is
highly subject to the weather conditions in a given year.
For the price of electricity it is difficult to identify an overall trend for all countries. One price series
that stands out is the one for Japan, which, compared internationally, starts on a very high level in the
beginning of the 1980s and then shows a sharp decline throughout the time period under
consideration. One of the reasons for this high initial level is a very high proportion of oil in the power
generation mix at that time and the all-time peak in oil prices following the two oil crises. The
subsequent decline of electricity prices can, amongst other possible reasons, be attributed to decrease
of oil prices that followed. Moreover, since the early 1990s, the Japanese government has undertaken
several efforts with regard to power market liberalization in order to lower electricity price to an
internationally more competitive level (see IEA, 2008).

5.2.2 Panel unit root tests

To check for the properties of the time series, we apply the panel unit root tests outlined in Section
4.1.2.19 The test statistics and the corresponding p-values are reported in Table 13. While the test
regressions for the variables in levels (e, y, p) contain an intercept and a time trend, the test
regressions for the variables in first differences (e, y, p) only contain an intercept. For the first four
tests, the lag lengths are selected according to the usual information criteria.

Table 13
Panel unit root tests
Method e e y y p p
0.67 15.09*** 0.29 3.45*** 0.12 9.77***
LLC
[0.25] [0.00] [0.61] [0.00] [0.55] [0.00]
1.07 6.09*** 5.37 1.15 3.65 1.53*
UB
[0.86] [0.00] [1.00] [0.87] [1.00] [0.06]
0.23 15.02*** 0.33 7.25*** 0.26 10.76***
IPS
[0.59] [0.00] [0.37] [0.00] [0.40] [0.00]
31.47 179.25*** 46.62 123.87*** 39.29 95.69***
ADF-Fisher
[0.68] [0.00] [0.11] [0.00] [0.32] [0.00]
43.88 282.98*** 19.63 132.09*** 44.41 218.85***
PP-Fisher
[0.17] [0.00] [0.99] [0.00] [0.16] [0.00]
Notes: *** and * denote significance at the 1% and 10% level, respectively. p-values are reported in squared brackets.
LLC: Levin et al. (2002); UB: Breitung (2000); IPS: Im et al. (2003); ADF-Fisher and PP-Fisher: Maddala and Wu (1999).

The results of the first five tests on the variables in levels do not allow for a rejection of the panel unit
root hypothesis at the conventional significance levels and hence indicate an order of integration of at
least one for all three variables. However, the tests on the first differences of the variables reject the
null hypothesis at least at the 10% level. An exception is the UB test statistic, which indicates a second
panel unit root for real income. As this order of integration is very unlikely for variables such as
income, and the other tests clearly reject the panel unit root hypothesis for the first differences of the
income variable, we conclude for all three variables that they are indeed integrated of order one, I(1).

19
We also considered the test of Hadri (2000), but the results, which we do not report here, indicate the existence of a second
unit root for all three variables. In a simulation study, Hlouskova and Wagner (2006) find that the Hadri test tends to overreject
the null hypothesis of stationarity, which is consistent with the findings of most empirical applications.

34
5.2.3 Panel cointegration tests

Having established a non-stationary behavior for all the time series considered, we now proceed to
testing for a long-run cointegrating relationship by applying Pedronis panel cointegration tests
described in Section 4.2.2. The results for both the within- and the between-dimension tests are
summarized in Table 14. Except for the Panel PP- and the Group PP- tests, the null hypothesis of
no cointegration is rejected at the 1% level. Using Monte Carlo simulations, Pedroni (2004)
investigates the small-sample properties of the Panel v, the Panel PP-, the Panel PP-t, the Group
PP- and the Group PP-t statistics for different dimensions of the panel. In the case where N = 20,
which is fairly close to our case, the Panel PP-t and Group PP-t perform best regarding size and
power when T < 130 and T < 40, respectively. Hence, we have some evidence for a stationary
behavior of the residuals from Eq. (4) and conclude that there exists a panel-cointegrating relationship
between residential electricity consumption, real disposable income, and real residential electricity
price.

Table 14
Pedronis panel cointegration test
Unweighted Weighted
Method Inference
Statistic [Prob.] Statistic [Prob.]
Panel A: Alternative hypothesis: common AR coefficients (within-dimension)
Panel v 8.78*** [0.00] 4.52*** [0.00] i ,t I 0
Panel PP- 0.51 [0.30] 0.67 [0.25] i ,t I 1
Panel PP-t 3.10*** [0.00] 3.59*** [0.00] i ,t I 0
Panel ADF-t 4.39*** [0.00] 4.46*** [0.00] i ,t I 0
Panel B: Alternative hypothesis: individual AR coefficients (between-dimension)
Group PP- 0.44 [0.67] i ,t I 1
Group PP-t 3.11*** [0.00] i ,t I 0
Group ADF-t 4.14*** [0.00] i ,t I 0
Notes: Normalization: ei,t = ci + y,i yi,t + p,i pi,t + i,t. Null hypothesis: no cointegration. Tests assume individual intercepts. Lag
lengths selected by the Schwarz Information Criterion (SIC). *** denotes significance at the 1% level. p-values are reported in
squared brackets.

5.2.4 Long- and short-run elasticities

Long-run elasticities
As a next step, we employ Pedronis group-means FMOLS and DOLS panel estimators described in
Section 4.2.2 for estimating the long-run demand relationship characterized by Eq. (4). The coefficient
estimates and the corresponding t-statistics from both the individual tests and the panel tests are
summarized in Table 15. Comparing the FMOLS and DOLS estimates of the slope coefficients with
each other, we find that most of them are in agreement concerning sign and magnitude, rendering our
results robust with regard to the estimation method. Furthermore, the plausibility of the cointegrating
vector estimates in terms of sign and magnitude, as well as the statistical significance, indicate that we
have indeed attained a reasonable approximation of the true equilibrium relationship. In the following,
we summarize the results in more detail.
For the whole panel the estimate of the income elasticity is 0.96 (0.91) from the FMOLS (DOLS)
estimator. This implies that a 1% rise in income is associated with a 0.9%1.0% increase in electricity
demand. The estimate of the price elasticity is 0.39 (0.38) from the FMOLS (DOLS) estimator,
implying a change in the magnitude of 0.39% (0.38%) in electricity demand in response to a 1%
increase in electricity price.
In the country-specific regressions, the coefficients on real disposable income from both FMOLS
and DOLS have a positive sign and are significant at the 1% level in sixteen out of the eighteen
countries considered. The magnitudes of the significant coefficients vary between 0.38 (0.35) in the
UK and a high 2.04 (1.92) in Austria from the FMOLS (DOLS) estimator. The price elasticities have a

35
negative sign and are significant at the 10% level in ten out of eighteen countries for both the FMOLS
and DOLS estimations. The magnitudes of the significant coefficients vary between 0.12 (0.14) in
the UK and 1.36 (1.37) in Japan from the FMOLS (DOLS) estimator.
The only country which exhibits neither a significant income nor a significant price elasticity is
Norway. A possible explanation can be found in the observations from the graphical inspection of
Figure 4, plot (a), in Section 5.2.1. Accordingly, in contrast to other countries, weather conditions
presumably have a much higher explanatory power for the variation in electricity consumption in
Norway, as compared to, say, income or the price of electricity.

Table 15
Long-run elasticities by (group-means panel) FMOLS and DOLS estimation
Fully modified OLS (FMOLS) Dynamic OLS (DOLS)
Countries
y p y p
0.96*** 0.39*** 0.91*** 0.38***
Group
(38.33) (9.76) (27.96) (13.19)
2.04*** 0.37 1.92*** 0.27
Austria
(6.53) (1.24) (5.81) (0.81)
0.82*** 0.87*** 0.79*** 0.80***
Denmark
(9.15) (7.63) (7.25) (5.37)
0.87*** 1.03** 0.99*** 0.85
Finland
(5.70) (2.32) (4.95) (1.51)
1.59*** 0.08 1.64*** 0.03
France
(3.10) (0.23) (3.23) (0.10)
0.48*** 0.04 0.42*** 0.16**
Germany
(8.80) (0.65) (6.01) (2.25)
1.14*** 0.58*** 0.98*** 0.56***
Greece
(7.89) (6.60) (4.82) (6.04)
0.53*** 0.03 0.53*** 0.06
Ireland
(15.46) (0.31) (10.93) (0.42)
0.99*** 0.04 0.90*** 0.10
Italy
(34.75) (0.57) (25.09) (0.97)
0.09 1.36*** 0.05 1.37***
Japan
(0.43) (11.32) (0.55) (27.51)
1.61*** 1.28*** 1.57*** 1.16**
Mexico
(5.92) (2.88) (5.00) (2.21)
0.74*** 0.00 0.56*** 0.01
Netherlands
(20.55) (0.09) (6.13) (0.13)
0.10 0.13 0.14 0.25
Norway
(0.95) (0.63) (0.87) (0.97)
1.42*** 0.80*** 1.60*** 0.67*
Portugal
(7.68) (2.88) (4.91) (1.88)
0.97*** 0.58** 0.86*** 0.66**
South Korea
(6.78) (2.23) (6.63) (2.83)
1.21*** 0.34** 1.24*** 0.30**
Spain
(9.79) (2.35) (9.61) (2.30)
1.72*** 0.09 1.34*** 0.23
Switzerland
(4.38) (0.33) (3.11) (0.70)
0.38*** 0.12* 0.35*** 0.14**
UK
(10.12) (1.93) (9.17) (2.13)
0.72*** 0.23** 0.66*** 0.21**
US
(5.50) (2.08) (5.64) (2.41)
Notes: Dependent variable: Electricity consumption (e). t-statistics appear in parentheses below the respective coefficients. ***,
** and * denote significance at the 1%, 5% and 10% level, respectively. Coefficient estimates significant at the 10% level are
highlighted in boldface.

Table 16 compares country-specific estimates of long-run elasticities from our analysis to estimates
from previous studies:

36
 For France, our estimate of the income elasticity is fairly high (1.64) and of similar magnitude as
the one obtained in Narayan et al. (2007) of 1.49. However, our estimate of the price elasticity is
not significant, while their estimate is, and amounts to 0.50.
For Germany, our income elasticity is 0.42 compared to 0.54 (not significant) in Narayan et al.
(2007). Concerning the price elasticity, however, we find a considerable difference: While we
estimate an elasticity of 0.16, their estimate takes on a value of 4.20, which seems implausibly
high in magnitude, as electricity price elasticities in general are expected to be fairly inelastic due
to limited substitution possibilities for electricity.
For Greece, we have very similar results with regard to the price elasticity as Hondroyiannis
(2004), i.e. 0.56 versus 0.41, respectively. However, while we estimate a near unity income
elasticity, his estimate is somewhat higher (1.56).
For Italy, our estimate of the income elasticity amounts to 0.90, while the estimate of the price
elasticity is statistically not significant. In Narayan et al. (2007) both estimates are not significant.
For Japan, our estimate for the price elasticity, which appears to be fairly high in magnitude
(1.37), is supported by both Nakajima (2010) and Narayan et al. (2007), whose estimates are
1.20 and 1.49, respectively.
For the UK, our estimate of the income elasticity amounts to 0.35, while the estimate of the price
elasticity amounts to 0.14. The estimates from Narayan et al. (2007) are both not significant.
For the US, our results are much closer to Nakajima and Hamori (2010) than to Dergiades and
Tsoulfidis (2008). While our estimate for the income elasticity (price elasticity) is 0.66 (0.21),
which is fairly close to the 0.85 (0.33) from Nakajima and Hamori, Dergiades and Tsoulfidis find
an income elasticity (price elasticity) of 0.27 (1.07). In Narayan et al. (2007) both estimates are
statistically not significant.
Comparing our group estimates of electricity demand elasticities to the ones from Narayan et al.
(2007), we find a striking difference in magnitudes, despite the similarity with regard to the time span
analyzed and the estimation method used. While we find a near unity income elasticity (0.9 to 1) and
an inelastic price elasticity (0.4) for the whole panel, their estimates reveal an inelastic income
elasticity (0.2 to 0.3) and an elastic price elasticity (1.6 to 1.5). We suspect that their group estimate
for price elasticity is most likely biased to a considerable extent by the unreasonably high estimate for
Germany (4.20), especially bearing in mind that the sample only consists of seven countries in the
cross-sectional dimension.

Table 16
Comparison of long-run country-specific demand elasticities
Countries BM NSP H DS N NH
Y P Y P Y P Y P Y P Y P
France 1.64 0.03 1.49 0.50 __ __ __ __ __ __ __ __

Germany __ __ __ __ __ __ __ __
0.42 0.16 0.54 4.20

Greece __ __ __ __ __ __
0.98 0.56 __ __ 1.56 0.41

Italy __ __ __ __ __ __ __ __
0.90 0.10 0.49 0.08

Japan __ __ __ __ __ __
0.05 1.37 0.89 1.49 0.65 1.20

UK __ __ __ __ __ __ __ __
0.35 0.14 0.66 0.60

US __ __ __ __
0.66 0.21 0.40 0.33 0.27 1.07 0.85 0.33

Notes: Y and P denote income and price elasticities, respectively. Elasticity estimates statistically significantly different from zero
at conventional levels are highlighted in boldface. BM: our estimations; NSP: Narayan et al. (2007); H: Hondroyiannis (2004);
DS: Dergiades and Tsoulfidis (2008); N: Nakajima (2010); NH: Nakajima and Hamori (2010). For simplicity of illustration we only
report the DOLS estimates from BM and only the most recent estimates from NH.

37
 Hence, quite contrary to the findings of Narayan et al. (2007), the implications from our results
indicate that in general the potential effectiveness of pricing policies aimed at reducing residential
electricity consumption are very limited.

Short-run elasticities
Using the residuals from the FMOLS and DOLS cointegrating regressions, we estimate two ECMs for
each country and the whole panel according to Eq. (30). The coefficients and the corresponding
t-statistics are reported in Table 17. As expected, electricity consumption adjusts negatively to
deviations from the long-run equilibrium both in the panel and in the country-specific models.

Table 17
Short-run elasticities
ECTs from FMOLS estimation ECTs from DOLS estimation
Countries
Constant y p ECTt1 Constant y p ECTt1
0.02*** 0.23*** 0.06*** 0.17*** 0.02*** 0.17*** 0.05** 0.27***
Group
(8.04) (4.99) (2.68) (6.40) (8.38) (4.28) (2.40) (6.81)
0.02 0.79 0.04 0.36** 0.02 0.73 0.10 0.35**
Austria
(1.57) (1.22) (0.18) (2.21) (1.10) (1.14) (0.49) (2.22)
0.00 0.45* 0.21** 0.45*** 0.00 0.40 0.19* 0.43***
Denmark
(0.62) (1.84) (2.14) (3.74) (0.62) (1.66) (1.99) (3.61)
0.03*** 0.03 0.15 0.23*** 0.03*** 0.03 0.10 0.19***
Finland
(3.54) (0.19) (0.91) (3.57) (3.68) (0.17) (0.58) (2.99)
0.03*** 0.40 0.22 0.36*** 0.03*** 0.45 0.24 0.29**
France
(2.88) (0.98) (0.80) (2.92) (3.05) (1.06) (1.05) (2.36)
0.01 0.24 0.04 0.33 0.01* 0.05 0.05 0.49*
Germany
(1.19) (0.85) (0.36) (1.63) (1.70) (0.23) (0.50) (1.75)
0.03*** 0.19 0.17 0.24* 0.03*** 0.13 0.07 0.52***
Greece
(4.42) (0.86) (1.52) (1.90) (4.72) (0.67) (0.74) (3.06)
0.02** 0.10 0.19* 0.49*** 0.02*** 0.01 0.14 0.56**
Ireland
(2.43) (0.86) (1.93) (3.09) (3.29) (0.08) (1.30) (2.41)
0.01*** 0.34** 0.02 0.29** 0.01*** 0.28** 0.04 0.55***
Italy
(3.91) (2.43) (0.45) (2.25) (4.02) (2.20) (0.74) (2.78)
0.02*** 0.16 0.30 0.26** 0.02** 0.16 0.30 0.26**
Japan
(3.69) (0.78) (1.02) (2.15) (3.69) (0.78) (1.02) (2.15)
0.04*** 0.23* 0.08 0.05 0.04*** 0.22 0.09 0.16*
Mexico
(5.13) (1.74) (0.90) (1.08) (4.88) (1.65) (1.03) (1.89)
0.01* 0.18 0.02 0.27 0.01* 0.03 0.02 0.60
Netherlands
(1.88) (0.77) (0.33) (1.61) (1.85) (0.11) (0.37) (1.63)
0.01 0.07 0.12** 0.29*** 0.01 0.06 0.11* 0.40***
Norway
(1.14) (0.49) (2.10) (3.00) (0.75) (0.42) (1.85) (2.96)
0.04*** 0.27 0.02 0.12 0.05*** 0.05 0.16 0.33***
Portugal
(4.73) (1.37) (0.09) (1.37) (6.67) (0.32) (0.83) (3.01)
South 0.03*** 0.53*** 0.27 0.23** 0.04*** 0.45*** 0.29* 0.46***
Korea (3.76) (4.53) (1.57) (2.10) (4.51) (4.43) (1.91) (3.09)
0.03*** 0.30 0.01 0.27* 0.03*** 0.14 0.14 0.52***
Spain
(3.02) (0.89) (0.06) (1.72) (3.53) (0.48) (0.64) (2.81)
0.02*** 0.31** 0.61*** 0.27*** 0.02*** 0.28** 0.45** 0.21***
Switzerland
(4.54) (2.76) (3.84) (4.53) (4.07) (2.16) (2.64) (3.29)
0.02** 0.21 0.16* 0.31* 0.01* 0.16 0.15* 0.33*
UK
(2.03) (0.86) (2.00) (1.76) (1.78) (0.66) (1.94) (1.95)
0.01 0.29* 0.21 0.47** 0.01 0.29* 0.23* 0.85***
US
(1.13) (1.80) (1.52) (2.72) (1.58) (2.01) (1.85) (4.03)
Notes: Dependent variable: First difference of electricity consumption (e). t-statistics appear in parentheses below the
respective coefficients. ***, ** and * denote significance at the 1%, 5% and 10% level, respectively. Slope coefficient estimates
significant at the 10% level are highlighted in boldface.

For the panel, the adjustment coefficients have the expected negative sign to ensure error
correction to the long-run equilibrium. The short-run income elasticity estimate is 0.23 and 0.17, while
the short-run price elasticity is estimated at 0.06 and 0.05, depending on whether the ECTs are

38
estimated by FMOLS or DOLS. Hence, the elasticities are lower in magnitude than their long-run
counterparts, which complies with intuition and the results from previous studies.
In the country-specific ECMs the adjustment coefficients are all negative and mostly significant in
both the FMOLS-based ECMs and the DOLS-based ECMs. In both models for the Netherlands,
significance is only found at a level of approximately 12%. For the most part, the country-specific
short-run elasticities are not significantly different from zero. Where they are, the signs and
magnitudes are reasonable and, in magnitude, well below the corresponding long-run elasticities. The
only exception is the short-run price elasticity of Switzerland, which takes on a significant positive
value.

5.2.5 Panel Granger causality testing

As the last step in our analysis of residential electricity demand in OECD countries, we conduct tests
on Granger causality between electricity consumption, disposable income and electricity price. As
described in Section 4.3, we estimate two PVECMs according to Eq. (38a-c), both of which are based
on the residuals of the FMOLS and DOLS regressions, respectively, using the PMGE method
proposed by Pesaran et al. (1999). The lag length is chosen such that serially uncorrelated residuals
are ensured. The Schwarz Information Criterion (SIC) points at an optimal lag length of one, k = 1. The
results are reported in Table 18.

Table 18
Panel Granger causality tests
Panel A: ECTs from FMOLS
Independent variables (sources of causation)
Dependent
variable Short-run Long-run Joint
e y p ECT ECT & e ECT & y ECT & p
(12a) e 2.87*** 0.56 6.52*** 26.00*** 21.65***
[0.00] [0.58] [0.00] [0.00] [0.00]
(12b) y 1.40 0.68 8.48*** 38.24*** 36.89***
[0.16] [0.49] [0.00] [0.00] [0.00]

(12c) p 0.22 0.83 1.82* 1.67 2.07

[0.83] [0.40] [0.07] [0.19] [0.13]

Panel B: ECTs from DOLS

Independent variables (sources of causation)
Dependent
variable Short-run Long-run Joint
e y p ECT ECT & e ECT & y ECT & p
(12a) e 2.15** 0.03 9.58*** 50.53*** 46.31***
[0.03] [0.98] [0.00] [0.00] [0.00]
(12b) y 1.97** 0.26 4.13*** 11.34*** 8.73***
[0.05] [0.79] [0.00] [0.00] [0.00]

(12c) p 0.06 1.26 0.76 0.30 1.18

[0.95] [0.21] [0.45] [0.74] [0.31]

Notes: t-statistics are displayed for short-run and long-run causation. Partial F-statistics are displayed for joint causation. p-
values are in brackets. ***, ** and * denote significance at the 1%, 5% and 10% level, respectively. Tests are based on a
PVECM(1).

First of all, in the short run we find evidence for causality running from disposable income to
electricity consumption in both, the FMOLS- and DOLS-based models. Vice versa, the null hypothesis
of no short-run causality from electricity consumption to disposable income is only rejected in the
FMOLS-based PVECM. In both models, the ECTs in the electricity consumption and disposable
income equations are significant and have the right sign in order to ensure error correction to long-run
equilibrium after a shock occurs. This bidirectional long-run Granger causality between electricity
consumption and disposable income is also supported by the results of the F-tests on joint
significance. The magnitudes of the adjustment coefficients (not reported in Table 18) are 0.13
(0.25) and 0.17 (0.12) for the electricity consumption and disposable income equation in the FMOLS-

39
based (DOLS-based) model. Thus, near-complete adjustments (of at least 95%) to long-run
equilibrium induced by changes in electricity consumption and disposable income take approximately
seven to nine years.
For the electricity price equation there is only evidence for an adjustment to equilibrium, and hence
long-run causality, in the FMOLS-based model and only at the 10% level of significance. However, the
F-tests on joint significance are unable to reject the null hypotheses. This indicates that electricity price
in our models is to be viewed as strongly exogenous.
Altogether, our results suggest that the feedback hypothesis holds at least in the long run, while in
the short run the results are mixed. This implies that policy measures aiming at electricity conservation
at the same time will involve a trade-off in the form of a decreasing effect on economic growth at least
in the long run.
Comparing our results to the three studies listed in Table 4, we find support for the results of
Zachariadis and Pashourtidou (2007), who also find evidence for the feedback hypothesis. Halicioglu
(2007) and Dergiades and Tsoulfidis (2008), on the other hand, both find evidence for the
conservation hypothesis in the long run.

5.3 Residential natural gas demand in OECD countries

5.3.1 Data

For the following analysis we gather data on residential natural gas consumption, net disposable
income, residential natural gas price, CPI and heating degree days for as many OECD countries as
possible. Residential natural gas consumption and nominal natural gas prices for the time period 1978
to 2008 are obtained from the IEA database Energy Balances of OECD Countries and Energy Prices
& Taxes, respectively, while net disposable income and total population are from the OECD database
(http://stats.oecd.org). The nominal price and income data are deflated using the CPI also provided by
the OECD, while natural gas consumption and real disposable income are divided by total population
in order to attain per capita values. Finally, the heating degree day indexes are taken from Eurostat
(http://epp.eurostat.ec.europa.eu). Data (non-)availability leaves us with the following countries:
Austria, Finland, France, Germany, Ireland, Japan, Luxembourg, the Netherlands, Spain, Switzerland,
the UK and the US. For Japan and the US, heating degree days are not provided by Eurostat.
Figures 58 display the time series of residential natural gas consumption (measured in tons of oil
equivalent, toe), real disposable income (measured in constant 1000 , base year 2000 = 100), real
residential natural gas price (measured in constant 1000 / toe, base year 2000 = 100) and heating
degree days, respectively.
Visual inspection of the single time series reveals the following trends:
Most countries reveal an overall upward trend in residential natural gas consumption.
Exceptions are the Netherlands and the US. Most of the gas consumption series have a local
peak in the year 1996, which coincides with a peak in the heating degree days.
All countries reveal an overall upward trend in real disposable income. In nearly all the series,
the decreasing effect of the financial crisis in 2008 on the variables can be seen.
Coinciding with the second oil price shock, real natural gas prices reach their highest levels in
the beginning of the 1980s and, then, fall to relatively low levels during the late 1980s. Only
since the year 2000 an overall upward trend for most of the series can be identified.

40
 0,8

0,7
Austria
Finland
0,6
France
Germany
0,5
Ireland

0,4 Japan
Luxembourg
0,3 Netherlands
Spain
0,2 Switzerland
UK
0,1 US

0
1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008
Fig. 5: Residential natural gas consumption per capita (in ktoe). Source: IEA, own illustration.

42
Austria
37 Finland
France
32 Germany
Ireland
27 Japan
Luxembourg
22 Netherlands
Spain
17 Switzerland
UK
12 US

7
1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008

Fig. 6: Real net disposable income per capita (in 1000 ), 2000 = 100. Source: OECD, own illustration.

41
 2050

1850
Austria
1650
Finland
1450 France
Germany
1250
Ireland

1050 Japan
Luxembourg
850
Netherlands

650 Spain
Switzerland
450
UK

250 US

50
1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008

Fig. 7: Real residential natural gas prices (in 1000 /toe), 2000 = 100. Source: IEA, own illustration.

6400

Austria
5400
Finland
France

4400 Germany
Ireland
Luxembourg
3400 Netherlands
Spain
Switzerland
2400 UK

1400
1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008

Fig. 8: Heating degree days. Source: Eurostat, own illustration.

42
5.3.2 Bounds test to cointegration

As a first step of the ARDL / bounds testing procedure we estimate Eq. (32) for each country using
OLS. As our analysis is based on annual data, we consider lag lengths of one and two. A deterministic
trend is included whenever significant. Then we conduct an F-test on the joint significance of the
lagged variables in levels. The results of the F-tests for all countries are given in Table 19. Only for
France and Spain the F-statistic indicates no joint significance. For all other countries the null
hypothesis of no long-run relationship is rejected at least at the 10% level.

Table 19
Bounds F-tests for a cointegration relationship
Country Lag length: 1 Lag length: 2
Austria Fg(g | y, p, hdd) = 3.749** [0.022] Fg(g | y, p, hdd) = 5.013** [0.011]
Finland Fg(g | y, p, hdd) = 4.779*** [0.008] Fg(g | y, p, hdd) = 5.231*** [0.010]
France Fg(g | y, p, hdd) = 1.743 [0.182] Fg(g | y, p, hdd) = 1.882 [0.170]
T
Germany Fg(g | y, p, hdd) = 2.915* [0.067] Fg(g | y, p, hdd) = 4.387** [0.043]
Ireland Fg(g | y, p, hdd) = 3.935** [0.019] Fg(g | y, p, hdd) = 7.236*** [0.003]
T
Japan Fg(g | y, p) = 13.869*** [0.000] Fg(g | y, p, hdd) = 3.188** [0.050]
Luxembourg Fg(g | y, p, hdd) = 4.134** [0.016] Fg(g | y, p, hdd) = 3.777** [0.033]
T
Netherlands Fg(g | y, p, hdd) = 3.581** [0.027] Fg(g | y, p, hdd) = 4.060** [0.026]
Spain Fg(g | y, p, hdd) = 1.160 [0.363] Fg(g | y, p, hdd) = 1.623 [0.232]
T
Switzerland Fg(g | y, p, hdd) = 6.114*** [0.003] Fg(g | y, p, hdd) = 4.198** [0.024]
UK Fg(g | y, p, hdd) = 2.601* [0.067] Fg(g | y, p, hdd) = 2.832 [0.062]
T
US Fg(g | y, p) = 6.052*** [0.004] Fg(g | y, p, hdd) = 6.543*** [0.004]
T
Notes: indicates the inclusion of a deterministic trend. p-values are reported in brackets. ***, ** and * denote significance at the
1%, 5% and 10% level, respectively.

5.3.3 Long-run relationships and short-run dynamics

According to the test results from the preceding section, we proceed to estimate the long-run
elasticities and the corresponding error correction models for ten of the twelve initially considered
countries, viz. Austria, Finland, Germany, Ireland, Japan, Luxembourg, the Netherlands, Switzerland,
the UK, and the US.
As described in Section 4.2.3, for each country a model is specified according to Eq. (33). The
model selection is guided by the Akaike Information Criterion (AIC) and the Schwarz Bayesian
Criterion (SBC). Furthermore, we make sure that the residuals are not serially correlated. The resulting
parameter estimates are used to construct the long-run elasticities according to Eqs. (34)(36). Finally,
in order to obtain the short-run dynamics, the corresponding error correction models according to
Eq. (37) are estimated using the lagged ECTs resulting from the long-run relationships estimated.
Tables 2022 summarize the estimated long-run coefficients, the error correction estimation results
and the diagnostic tests (for serial correlation, normality, and heteroscedasticity) of the underlying
ARDL models for the individual countries, respectively.20
The order of the individual country-specific ARDLs and the specification with regard to the
deterministic term are given along with the estimated long-run coefficients in Table 20. The signs of
the statistically significant income and price elasticities are positive and negative as expected. The
magnitudes of the income elasticities range between 0.44 in Japan and 1.72 in Ireland with a near
unity (0.94) average. The demand elasticities with regard to price range between 0.14 in the
Netherlands and 1.62 in Ireland, with a cross-country average of 0.51. The long-run elasticities with
regard to heating degree days range between 0.95 in the Netherlands and 2.01 in Austria with a mean
of 1.36.

20
Tests on constancy of the cointegration space are delivered in Section 4.4.

43
 The coefficient estimates of the error correction models are presented in Table 21. The speeds of
adjustment toward long-run equilibria range between an annual correction of 31% in the UK and 96%
in the US. On average, 57% of any deviation from long-run equilibrium is corrected every year. Hence,
a near complete adjustment (of at least 95%) is achieved after three to four years on average. The
short-run elasticities with regard to income, price and heating degree days are generally lower in
magnitude than their long-run counterparts. Short-run income elasticities range between 0.27 in Japan
and 0.68 in the Netherlands, with an average of 0.44. The short-run price elasticities range between
0.54 in Ireland and 0.12 in the Netherlands, with a mean of 0.23. Finally, the short-run coefficients
for heating degree days range between 0.55 in Luxembourg and 0.90 in the Netherlands, with an
average of 0.71. Hence, on average, the short-run elasticities have approximately half the magnitude
of their long-run counterparts.
Finally, Table 22 summarizes a set of diagnostic test statistics for the individual ARDL models. The
tests indicate problems with the assumptions of no serial correlation, normality and homoscedasticity
of the residuals in the model for Finland. Furthermore, the models for Ireland and the UK have
deficiencies with regard to normality and homoscedasticity of the residuals, respectively. For all other
models the tests show no deviations from theoretical model assumptions.

Table 20
Long-run coefficients for country-specific ARDLs
Country Long-run coefficients
Order of ARDL y p hdd Trend Constant

Austria 1.196** 0.360** 2.007** 18.259***

ARDL(1,2,2,0) [0.021] [0.045] [0.014] [0.008]

Finland 0.797 0.267 0.551 12.250

ARDL(1,1,0,0) [0.127] [0.214] [0.742] [0.357]

Germany 0.080 0.233** 1.047*** 0.043*** 15.541***

ARDL(2,1,1,1) [0.904] [0.011] [0.003] [0.000] [0.000]

Ireland 1.715*** 1.621** 1.945 7.607

ARDL(1,0,0,0) [0.001] [0.021] [0.500] [0.700]

Japan 0.482** 0.350 0.009*** 5.551***

ARDL(1,0,0) [0.039] [0.110] [0.001] [0.000]

Luxembourg 0.653*** 0.117 1.131** 15.160***

ARDL(1,1,1,0) [0.000] [0.228] [0.041] [0.001]

Netherlands 0.720*** 0.144*** 0.950*** 0.022*** 10.941***

ARDL(1,0,1,0) [0.005] [0.000] [0.000] [0.000] [0.000]

Switzerland 1.121** 0.804*** 1.129*** 0.026*** 9.861***

ARDL(1,2,0,0) [0.012] [0.000] [0.001] [0.000] [0.000]

UK 0.711*** 0.350** 1.897** 18.048***

ARDL(1,0,0,0) [0.000] [0.014] [0.028] [0.004]

US 0.187 0.184*** 0.005** 7.264***

ARDL(1,0,1) [0.451] [0.002] [0.023] [0.000]
MEAN 0.936 0.514 1.360
MIN 0.435 1.621 0.950
MAX 1.715 0.144 2.007
Notes: ***, ** and * denote significance at the 1%, 5% and 10% level, respectively. p-values are reported in brackets. MEAN,
MIN and MAX values refer to significant estimates only. Moreover, Finland is not included in the calculation of the summary
statistics on account of model deficiencies.

44
45
Table 22
Diagnostic tests for the underlying ARDL models
Lagrange multiplier statistics
Country
Serial correlation: SC
2
(1) Normality: N2 (2) Heteroscedasticity: H2 (1)
Austria 0.149 1.094 3.315
[0.699] [0.579] [0.069]
Finland 3.410 21.215 3.963
[0.065] [0.000] [0.047]
Germany 1.941 1.880 0.208
[0.164] [0.391] [0.649]
Ireland 0.270 16.694 2.665
[0.604] [0.000] [0.103]
Japan 2.677 1.525 2.035
[0.102] [0.466] [0.154]
Luxembourg 0.252 1.716 0.000
[0.616] [0.424] [0.993]
Netherlands 0.139 0.494 2.141
[0.709] [0.781] [0.143]
Switzerland 0.006 0.211 5.665
[0.938] [0.900] [0.017]
UK 0.677 0.065 9.197
[0.411] [0.798] [0.010]
US 0.027 0.722 0.006
[0.870] [0.697] [0.937]

Notes: p-values are reported in brackets.

5.3.4 Constancy of cointegration space

In order to check for parameter constancy, we employ the CUSUM and the CUSUMSQ stability tests
to the estimated ARDL model of each country.
The CUSUM and CUSUMSQ plots for each model are shown in Figure 9, plots (a)(j). As can be
seen, the plots are within the 5% critical bounds in all the models, except for the case of Japan. Here
the CUSUMSQ plot crosses the upper critical bound, indicating some instability of the estimated
coefficients in the years 1993 to 1997. For all other models the stability tests show an overall
constancy of the cointegration space.

46
47
5.4 Interfuel substitution in major European countries

5.4.1 Data

The data used for the analysis reported in this section comes from the International Energy Agency
(IEA) databases Energy Prices & Taxes and Energy Balances of OECD Countries (IEA, 2008),
respectively. We have obtained a coherent data set on prices and energy consumption of different
types of fuel on the aggregate and sectoral (industry and residential) level for five European countries
(Germany, France, Italy, Spain and the UK). Some restrictions apply: While we have managed to
collect data on electricity, oil and gas for all countries, data on coal for both the industrial and
residential sector was only available for France and the UK. Furthermore, data on coal consumption
for the Italian residential sector was also unavailable, whereas for Germany and Spain data on coal
prices were missing for all sectors. As a consequence, we were unable to determine the same
amounts of elasticity estimates. The time span covered is from 1978 to 2006. The variables used in
the econometric estimation are: (i) total final consumption, (ii) industry consumption and (iii) residential
consumption of (a) electricity, (b) petroleum products (summarized as oil), (c) natural gas, and (d)
coal products. Moreover, as price variables we use the real end-use price index of the respective
fuels. As explained in Section 2.5 we calculate the cost shares using translog functions.

5.4.2 Translog estimations

For each country considered several seemingly unrelated regression (SUR) estimates have been
generated with and without including time as explanatory variable for the industrial and residential
sectors, and the total economy. That is, based on the available data, we have specified a system of
cost-share functions for each country according to the model system described in Eqs. (39a-c). The
own- and cross-price elasticities of demand are calculated according to Eq. (14) with the
corresponding parameters estimated, s, and the average of the factor-shares for each fuel (and
sector) between 1978 and 2006.
Table 23 gives a summary of the elasticities obtained (calculated from the coefficient estimates (s)
and the average cost shares for each fuel, see Eq. (14)). While a number of the elasticities are
statistically significant, they do show a wide range of values across fuels, sectors and countries
studied (and sometimes have counterintuitive signs), making it difficult to come up with general
conclusions based on certain patterns identified. Likewise, whether fuels are substitutes or
complements to each other seems to differ a lot by sector and country (e.g. gas and electricity in
industry are complements in France but substitutes in the other countries considered, while in the
residential sector fuel mixes for heating and cooling vary a lot by country and fuel flexibility can be
expected to be lower than in industry), rendering it again difficult to come up with firm patterns in the
estimates or conclusions for policy-makers.
Table 24 summarizes the coefficient estimates obtained for the deterministic time trend included in
the cost share equations, providing evidence for its usefulness to improve the fit of the model (most
are significant at the 1% level). Note that in several cases the inclusion of the time trend changes the
sign of the elasticity derived from the cost shares and coefficients estimated, so that the choice and
interpretation of these results should be made with some caution. Since the coefficient of the time
trend is only found to be insignificant in two cases (fuel share of natural gas in the aggregate UK
economy, share of electricity in the industry in France), it seems more appropriate to rely on the
columns elasticities with time included in Table 23.

48
49
Table 24
Estimates of the coefficients of the proxy variable, time, for technical change, 1978-2006
Country Share of fuel Total economy Industry Households
Electricity 0.49** (3.01) 0.77 (1.64) 0.11** (2.99)
France Gas 1.79** (4.77) 1.49 (1.63) 0.04** (25.01)
Oil 0.06** (62.95) 0.01** (7.16) 0.03** (10.09)
Electricity 6.66** (6.76) 7.59** (4.96) 2.77** (2.91)
UK Gas 0.80 (1.04) 4.55** (3.17) 13.19** (6.70)
Oil 14.57** (9.12) 18.69** (8.54) 14.88** (5.38)
Electricity 4.29** (21.79) 5.52** (9.14) 5.66** (10.76)
1
Italy Gas 6.72** (9.11) 10.73** (5.22) 20.40** (18.69)
Oil 9.10** (14.63) 9.54** (4.89)
Electricity 2.27** (3.73) 13.16** (14.69) 7.15** (3.54)
Germany
Gas 6.41** (22.99) 7.04** (11.00) 41.15** (12.08)
Electricity 4.76** (6.46) 8.04** (2.14) 20.53** (2.85)
Spain
Gas 11.18** (17.91) 32.77** (11.69) 24.85** (5.82)
1
Notes: Complete data only for total economy and industry; z-values in parentheses; * and ** denote statistically significant on
the 5% level and 10% level, respectively. More detailed estimation results can be obtained from the authors upon request.

6 Conclusions
In this study we have investigated the responsiveness of energy demand to measures of economic
activity and energy price. Furthermore, we also aimed at assessing the direction of causality between
energy use and economic activity. This was done using a number of econometric time series and
panel data techniques.
In Section 5.1 we have undertaken what, to the best of our knowledge, is the first attempt to
estimate subsector-specific electricity demand elasticities with regard to economic activity and
electricity price. This subsectoral approach aims at reducing the heterogeneity concerning
consumption behavior of the analyzed consumer groups and thereby reaping the benefits of additional
information otherwise veiled through aggregation. Making use of German annual data on a
subsectoral level from 1970 to 2007, we were successful in finding statistically significant cointegration
relationships in five of the eight analyzed subsectors employing a multivariate cointegrated VAR
setting. We take account of structural breaks by incorporating shift dummies into the cointegrating
vectors (see Johansen et al., 2000). Furthermore, we check for the direction of Granger-causality and
conduct an impulse response analysis. In order to attain the corresponding short-run elasticities of
economic activity and price, we also estimate single-equation error correction models in which we
include the error correction terms resulting from the VECM estimation.
Our findings for the long-run elasticity estimates are economically reasonable both in terms of sign
and magnitude. For three of the subsectors, long-run demand elasticities of economic activity are
found to be near unity. The most extreme estimates are 0.70 in the Food & Tobacco industry and 1.90
in the Pulp & Paper industry. The estimates of price elasticities vary between zero (statistically
insignificant), to which the corresponding coefficients in two of the models can be restricted, and
0.52. In the short run, the elasticities of economic activity range between 0.17 and 1.02, whereas the
price elasticities range between zero (insignificant) and 0.57. The deterministic time trends included
in the long-run demand relationships pick up differing effects from exogenous factors, such as
structural changes in the industry and technological progress, in three of the subsectors. In the Food &
Tobacco and the Non-metallic minerals industries the trend shows a positive sign, implying an
increasing net effect on electricity intensity, whereas in the Chemicals industry the sign is negative,
implying the opposite net effect. Adjustments to long-run equilibrium take place with widely differing
speeds in the single sectors. Hence, depending on the sector, near-complete adjustments are
achieved after approximately three to fourteen years. This is also reflected by the impulse response
functions.
The tests on Granger-causality indicate evidence for the feedback hypothesis in the Food &
Tobacco and the Pulp & Paper industry, and the conservation hypothesis in the Chemicals, the Non-
metallic minerals and the Transport equipment sectors. Hence, only in the two former sectors

50
conservation policies would be confronted with a certain trade-off in terms of decreasing output
growth.
In Section 5.2 we have conducted an analysis on the responsiveness of residential electricity
demand for a set of OECD countries. We estimate electricity demand elasticities with regard to
disposable income and own price, thereby differentiating between the short and the long run in an
error correction framework. Furthermore, we conduct tests on Granger causality between the
variables. In order to circumvent the problem of low power and size of more traditional unit root and
cointegration tests based on time series data, we make use of available panel data for eighteen
countries in the cross-sectional dimension and the years 19812008 in the time domain. The methods
employed for estimating the cointegrating vectors are Pedronis group-means fully modified OLS
(FMOLS) and dynamic OLS (DOLS) panel estimators. The Granger causality tests are based on a
panel vector error correction model estimated using the pooled mean group estimator (Pesaran et al.,
1999).
Our findings for the whole panel indicate a near unity income elasticity and an inelastic price
elasticity of approximately 0.4 in the long run. These results are robust with respect to the two
estimation methods employed. In the short run, our estimates from error correction models indicate an
income elasticity of 0.2 and a price elasticity of approximately 0.1. When compared with the results of
previous studies on residential electricity demand, our elasticity estimates bear some resemblance to
other estimates on the individual country level.
Overall, our results imply that the steering effect of tax-induced price increases on residential
electricity demand has a very limited potential for energy conservation, and hence a reduction of GHG
emissions. These policy implications are in stark contrast, e.g., to the ones reported in Narayan et al.
(2007).
Furthermore, our tests on Granger causality indicate that both electricity consumption and income
adjust toward the long-run equilibrium after a shock hits the system. Thus, a bidirectional causal
relationship between electricity consumption and economic growth exists in the long run. Our findings
are, therefore, in favor of the feedback hypothesis. Hence, a reduction of electricity consumption will
be associated with a trade-off with regard to per capita income.
In Section 5.3 we have analyzed residential natural gas demand for twelve OECD countries using
available time series data from 1980 to 2008. We estimate long-run demand elasticities with regard to
real disposable income and real residential natural gas price using the autoregressive distributed lag
(ARDL) bounds testing procedure developed by Pesaran and Shin (1999) and Pesaran et al. (2001).
In contrast to other cointegration approaches, this procedure has the advantage of needing no
pretesting on the time series properties of the single variables, and thereby circumvents the problem
associated with the low power of unit root tests. By employing an error correction framework we also
obtain estimates for the speeds of adjustment to long-run equilibrium and short-run elasticities for the
single countries. The effect of weather conditions on natural gas demand in a given year is accounted
for by including heating degree days as a control variable.
For ten of the twelve countries we find a significant long-run relationship. On average, the long-run
elasticities are 0.94 with regard to income, 0.51 with regard to price and 1.36 with regard to heating
degree days. For the single countries, the long-run income elasticities range between 0.44 in Japan
and 1.72 in Ireland, while the long-run price elasticities range between 1.62 in Ireland and 0.14 in
the Netherlands. The long-run elasticities with regard to heating degree days range between 0.95 in
the Netherlands and 2.01 in Austria.
The short-run dynamics assessed by estimation of the error correction models indicate an average
adjustment coefficient of 0.57, a short-run income elasticity of 0.44, a short-run price elasticity of
0.23, and a short-run elasticity with regard to heating degree days of 0.71. Hence, on average, the
short-run elasticities have approximately half the magnitude of their long-run counterparts.
In Section 5.4 a static translog model specification is used as a complementary approach to
calculate energy elasticities. In particular, we want to find out to what extent there is substitution or
complementarity among the different types of fuel inputs considered (e.g., electricity, gas, oil).
Specifically, we apply a cost share model of energy consumption for four fuels (where data on coal
were lacking three fuels) to five major European countries (i.e., Germany, France, Italy, Spain, and the
UK), obtaining a set of own-price and cross-price elasticity estimates for the aggregate economy as

51
well as the industrial and the residential household sector. The models are estimated both with and
without a deterministic time trend as a proxy for technical change and further influences ignored
otherwise by the model specification. The estimated coefficients are mostly found to be significant,
justifying their inclusion in the model. While many of the indirectly estimated elasticities are statistically
significant and have the expected sign, the great range of results and lack of easily observable
patterns in the estimates (across sectors, fuels and countries) makes it difficult to come up with
recommendations for policy-makers from this part of the study. A comparison between the results
obtained from the cointegration analyses and those from the translog model specification is difficult,
not least to the often less intuitive signs and magnitudes of the estimated coefficients gained from the
latter approach. In this respect, we find the results obtained from the cointegration analysis as more
plausible and robust and thus also more useful.
Overall, regardless of the sector or energy type considered, our results imply that the steering
effect of tax-induced price increases on energy demand has a very limited potential for energy
conservation, and hence a reduction of GHG emissions. Furthermore, our findings suggest that
reductions in electricity consumption are associated with a trade-off concerning economic growth in
the residential sector, as well as some subsectors of the manufacturing industry.

7 Literature
Agnolucci, P. (2009). The energy demand in the British and German industrial sectors: Heterogeneity
and common factors. Energy Economics 31(1): 175187.
Ahn, S.K., Reinsel, G.C. (1990). Estimation for partially nonstationary multivariate autoregressive
models. Journal of the American Statistical Association 85(411): 813823.
Amarawickrama, H.A., Hunt, L.C. (2008). Electricity demand for Sri Lanka: A time series analysis.
Energy 33(5): 724739.
Ariew, R. (1976). Ockham's Razor: A Historical and Philosophical Analysis of Ockham's Principle of
Parsimony. Champaign-Urbana, University of Illinois.
Asche, F., Nilsen, O.B., Tveteras, R. (2008). Natural gas demand in the European household sector.
The Energy Journal 29(3): 2746.
Athukorala, P.P.A.W, Wilson, C. (2010). Estimating short and long-term residential demand for
electricity: New evidence from Sri Lanka. Energy Economics 32(S1): S34S40.
Barker, T., Ekins, P., Johnstone, N. (1995). Global Warming and Energy Demand. Routledge, Taylor &
Francis Group.
Beenstock, M., Goldin, E., Nabot, D. (1999). The demand for electricity in Israel. Energy Economics
21(2): 168183.
Berkhout, P.H.G., Ferrer-i-Carbonell, A., Muskens, J.C. (2004). The ex post impact of an energy tax
on household energy demand. Energy Economics 26(3): 297317.
Berndt, E.R., Wood, D.O. (1975). Technology, Prices, and the Derived Demand for Energy. The
Review of Economics and Statistics 57(3): 259268.
Bohi, D.R., Zimmerman, M.B. (1984). An Update on Econometric Studies of Energy Demand
Behavior. Annual Review of Energy 9:105-154.
Bose, R.K., Shukla, M. (1999). Elasticities of electricity demand in India. Energy Policy 27(3): 137
146.
Breitung, J. (2000). The local power of some unit root tests for panel data. In: Baltagi, B.H., Formby,
T.B., Hill, R.C. (Eds.), Advances in Econometrics: Nonstationary Panels, Panel Cointegration and
Dynamic Panels, JAI Press, Amsterdam, Vol. 15: 161178.
Brown, R. L., J. Durbin, and J. M. Evans (1975). Techniques for testing the constancy of regression
relationships over time. Journal of the Royal Statistical Society. Series B (Methodolgical) 37(2): 149
192.

52
Christensen, L.R., Jorgenson, D.W. and Lau, L.J. (1973). Trascendental Logarithmic Production
Frontiers. The Review of Economics and Statistics 55(1): 2845.
Dahl, C. (1993). A survey of energy demand elasticities in support of the development of the NEMS.
(prepared for United States Department of Energy???), October.
Dergiades, T., Tsoulfidis, L. (2008). Estimating residential demand for electricity in the United States,
19652006. Energy Economics 30(5): 27222730.
Elliott, G., Rothenberg, T.J., Stock, J.H. (1996). Efficient tests for an autoregressive unit root.
Econometrica 64(4): 813836.
Engle, R.F., Granger, C.W.J. (1987). Co-integration and error correction: representation, estimation
and testing. Econometrica 55(2): 251276.
Erdogdu, E. (2010). Natural gas demand in Turkey. Applied Energy 87(1): 211219.
Filippini, M. (1994). Estimating Electricity Demand by Time-of-Use: A Share Equation Approach.
Economia delle fonti di energia e dellambiente 37(1): 3343.
Fisher, F.M., Kaysen, C. (1962). A Study in Econometrics: The demand for electricity in the United
States. North-Holland Publishing Company, Amsterdam.
Gonzalo, J. (1994). Five alternative methods of estimating long-run equilibrium relationships. Journal
of Econometrics 60(1/2): 203233.
Granger, C.W.J. (1969). Investigating causal relations by econometric models and cross-spectral
methods. Econometrica 37(3): 424438.
Greene, W. (2003). Econometric Analysis. Fifth edition, Pearson Education.
Griffin, J.M., Gregory, P.R. (1976). An Intercountry Translog Model of Energy Substitution Responses.
The American Economic Review 66(5):845857.
Hadri, K. (2000). Testing for stationarity in heterogeneous panel data. Econometrics Journal 3(2):
148161.
Halicioglu, F. (2007). Residential electricity demand dynamics in Turkey. Energy Economics 29(2):
199210.
Halvorsen, R. (1975). Residential demand for electric energy. The Review of Economics and Statistics
75(1): 1218.
Harris, A., McAvinchey, I.D., Yannopoulos, A. (1993). The Demand for Labour, Capital, Fuels and
Electricity: A Sectoral Model of the United Kingdom Economy. Journal of Economic Studies 20(3):24
35.
Harris, R., Sollis, R. (2003). Applied Time Series Modelling and Forecasting. John Wiley & Sons,
Chichester.
Henley, A., Peirson, J. (1998). Residential Energy Demand and the Interaction of Price and
Temperature: British Experimental Evidence. Energy Economics 20(2): 157171.
Hesse, D.M., Tarkka, H. (1986). The Demand for Capital, Labor and Energy in European
Manufacturing Industry before and after Oil Price Shocks. The Scandinavian Journal of Economics
88(3):529546.
Hlouskova, J., Wagner, M. (2006). The performance of panel unit root and stationarity tests: results
from a large scale simulation study. Econometric Reviews 25(1): 85116.
Holtedahl, P., Joutz, F.L. (2004). Residential electricity demand in Taiwan. Energy Economics 26(2):
201224.
Hondroyiannis, G. (2004). Estimating residential demand for electricity in Greece. Energy Economics
26(3): 319334.
Houthakker, H.S. (1951). Some calculations on electricity consumption in Great Britain. Journal of the
Royal Statistical Society, Series A (General), CXIV, part III: 359371.
Huntington, H.G. (2007). Industrial natural gas consumption in the United States: An empirical model
for evaluating future trends. Energy Economics 29(4): 743759.

53
Ibrahim, B.I., Hurst, C. (1990). Estimating Energy and Oil Demand Functions: A study of Thirteen
Developing Countries. Energy Economics 12(2):93102.
IEA (2008). Energy policies of IEA countries Japan, 2008 Review. [Online] URL:
http://www.iea.org/textbase/nppdf/free/2008/japan2008.pdf.
Im, K.S., Pesaran, M.H., Shin, Y. (2003). Testing for unit roots in heterogeneous panels. Journal of
Econometrics 115(1): 5374.
Johansen, S. (1988). Statistical analysis of cointegrating vectors. Journal of Economic Dynamics and
Control 12(2/3): 231254.
Johansen, S. (1994). The role of the constant and linear terms in cointegration analysis of
nonstationary variables. Econometric Reviews 13(2): 205229.
Johansen, S. (1995). Likelihood-based Inference in Cointegrated Vector Autoregressive Models.
Oxford University Press, Oxford.
Johansen, S., Mosconi R., Nielsen B. (2000). Cointegration analysis in the presence of structural
breaks in the deterministic trend. Econometrics Journal 3(2): 216249.
Jones, C.T. (1995). A Dynamic Analysis of Interfuel Substitution in US Industrial Energy Demand.
Journal of Business & Economic Statistics 13(4):459465.
Joutz, F.L., Shin, D., McDowell, B., Trost, R.P. (2008). Estimating regional short-run and long-run price
elasticities of residential natural gas demand in the U.S. 28th USAEE/IAEE Annual North American
Conference, New Orleans, LA, December 35.
Kamerschen, D.R., Porter, D.V. (2004). The demand for residential, industrial and total electricity,
1973-1998. Energy Economics 26(1): 87100.
Kao, C., Chiang, M-H. (2000). On the estimation and inference of a cointegrated regression in panel
data. In: Baltagi, B.H., Formby, T.B., Hill, R.C, (Eds.), Advances in Econometrics: Nonstationary
Panels, Panel Cointegration and Dynamic Panels, JAI Press, Amsterdam, Vol. 15: 179222.
Kim, B.C., Labys, W.C. (1988). Application of the Translog Model of Energy Substitution to Developing
Countries: The Case of Korea. Energy Economics 10(4): 313323.
Kraft, J., Kraft, A. (1978). On the relationship between energy and GNP. Journal of Energy and
Development 3(2): 401403.
Krichene, N. (2002). World crude oil and natural gas: a demand and supply model. Energy Economics
24(6): 557576.
Lanne, M., Ltkepohl, H. (2002). Unit root tests for time series with level shifts: a comparison of
different proposals. Economics Letters 75(1): 109114.
Lanne, M., Ltkepohl, H., Saikkonen, P. (2002). Comparison of unit root tests for time series with level
shifts. Journal of Time Series Analysis 23(6): 667685.
Lanne, M., Ltkepohl, H., Saikkonen, P. (2003). Test procedures for unit roots in time series with level
shifts at unknown time. Oxford Bulletin of Economics and Statistics 65(1): 91115.
Levin, A., Lin, C.F., Chu, C.S. (2002). Unit root tests in panel data: asymptotic and finite-sample
properties. Journal of Econometrics 108(1): 124.
Lin, W.T., Chen, Y.H., Chatov R. (1987). The demand for natural gas, electricity and heating oil in the
United States. Resources and Energy 9(3): 233258.
Liu, B. (1983). Natural gas price elasticities: Variations by region and sector in the USA. Energy
Economics 5(3): 195201.
Ltkepohl, H., Krtzig, M. (2004). Applied Time Series Econometrics. Cambridge University Press,
Cambridge.
MacKinnon, J.G. (1996). Numerical distribution functions for unit root and cointegration test. Journal of
Applied Econometrics 11(6): 601618.
MacKinnon, J.G., Haug, A.A., Michelis, L. (1999). Numerical distribution functions of likelihood ratio
tests for cointegration. Journal of Applied Econometrics 14(5): 563577.

54
Maddala, G.S., Wu, S.A. (1999). A comparative study of unit root tests with panel data and a new
simple test. Oxford Bulletin of Economics and Statistics 61(S1): 631652.
Madlener R. (1996). Econometric analysis of residential energy demand: A survey. Journal of Energy
Literature II(2): 332.
Madlener R. (2004). Chapter 3: Substitution and Diffusion of Natural Gas, in Jochem, E. and Jakob, M.
(eds.) Energy Perspectives and CO2 Mitigation Potentials in Switzerland until 2010. Energy Efficiency
and Substitution by Natural Gas and Renewable Energies, Vdf-Hochschulverlag, Zurich, pp. 89114.
Nakajima, T. (2010). The residential demand for electricity in Japan: An examination using empirical
panel analysis techniques. Journal of Asian Economics 21(4): 412420.
Nakajima, T., Hamori, S. (2010). Change in consumer sensitivity to electricity prices in response to
retail deregulation: a panel empirical analysis of the residential demand for electricity in the United
States. Energy Policy 38(5): 24702476.
Narayan, P.K., Smyth, R. (2005). The residential demand for electricity in Australia: an application of
the bounds testing approach to cointegration. Energy Policy 33(4): 467474.
Narayan, P.K., Smyth, R., Prasad, A. (2007). Electricity consumption in G7 countries: A panel
cointegration analysis of residential demand elasticities. Energy Policy 35(9): 44854494.
Paruolo, P. (1996). On the determination of integration indices in I(2) systems. Journal of
Econometrics 72(1-2): 313356.
Payne, J.E. (2010). Survey of the international evidence on the causal relationship between energy
consumption and growth. Journal of Economic Studies 37(1): 5395.
Pedroni, P. (1999). Critical values for cointegration tests in heterogeneous panels with multiple
regressors. Oxford Bulletin of Economics and Statistics 61(1): 653670.
Pedroni, P. (2000). Fully modified OLS for heterogeneous cointegrated panels. In: Baltagi, B.H.,
Formby, T.B. and Hill, R.C. (Eds.), Advances in Econometrics: Nonstationary Panels, Panel
Cointegration and Dynamic Panels, JAI Press, Amsterdam, Vol. 15: 93130.
Pedroni, P. (2001). Purchasing power parity tests in cointegrated panels. The Review of Economics
and Statistics 83(4): 727731.
Pedroni, P. (2004). Panel cointegration: asymptotic and finite sample properties of pooled time series
tests with an application to the PPP hypothesis. Econometric Theory 20(3): 597625.
Perron, P. (1989). The great crash, the oil price shock and the unit root hypothesis. Econometrica
57(6): 13611401.
Pesaran, P., Pesaran, M.H., (2009). Time series econometrics using Microfit 5.0. Oxford University
Press, Oxford.
Pesaran, M.H., Smith, R.P., Akiyama, T. (1998). Energy Demand in Asian Developing Economies.
Oxford University Press, Oxford.
Pesaran, M.H., Shin, Y. (1999). An autoregressive distributed-lag modelling approach to cointegration
analysis. In: Strom, S. (Ed.), Econometrics and Economic Theory in the 20th Century. Cambridge
University Press, Cambridge.
Pesaran, M.H., Shin, Y., Smith, R.P. (1999). Pooled mean group estimation of dynamic
heterogeneous panels. Journal of the American Statistical Association 94(446): 621634.
Pesaran, M.H., Shin, Y., Smith, R.J. (2001). Bounds testing approaches to the analysis of level
relationships. Journal of Applied Econometrics 16: 289326.
Pindyck, R.S. (1979). Interfuel substitution and the industrial demand for energy: an international
comparison. The Review of Economics and Statistics 61(2): 169179.
Polemis, M.L. (2007). Modeling industrial energy demand in Greece using cointegration techniques.
Energy Policy 35(8): 40394050.
Roy, J., Sanstad, A.H., Sathaye, A., Khaddaria, R. (2006). Substitution and Price Elasticity Estimates
Using Inter-country Pooled Data in a Translog Cost Model. Energy Economics 28:706719.

55
Ryan, D.L., Wang, Y., Plourde, A. (1996). Asymmetric Price Responses of Residential Energy
Demand in Ontario. Canadian Journal of Economics 29(1):317323.
Saikkonen, P. (1992). Estimation and testing of cointegrated systems by an autoregressive
approximation. Econometric Theory 8(1): 127.
Saikkonen, P., Ltkepohl, H. (2002). Testing for a unit root in a time series with a level shift at
unknown time. Econometric Theory 18(2): 313348.
Silk, J.I., Joutz, F.L. (1997). Short and long-run elasticities in US residential electricity demand: a co-
integration approach. Energy Economics 19(4): 493513.
Stock, J.H., Watson, M.W. (1993). A simple estimator of cointegrating vectors in higher order
integrated systems. Econometrica 61(4): 783820.
Taheri, A.A. (1994). Oil Shocks and the Dynamics of Substitution Adjustments of Industrial Fuels in the
US. Applied Economics 26(8):751756.
Urga, G., Walters, C. (2003). Dynamic Trasnlog and Linear Logit Models: A Factor Demand Analysis
of Interfuel Substitution in US Industrial Energy Demand. Energy Economics 25(2003):121.
Zachariadis, T., Pashourtidou, N. (2007). An empirical analysis of electricity consumption in Cyprus.
Energy Economics 29(2): 183198.
Zarnikau, J. (2003). Functional Forms in Energy Demand Modeling. Energy Economics 25:603613.
Ziramba, E. (2008). The demand for residential electricity in South Africa. Energy Policy 36(9): 3460
3466.

8 Attachments

8.1 List of Figures

Figure 1: Subsectoral industrial electricity consumption, real value added and real electricity price
Figure 2: Orthogonalized impulse response functions
Figure 3: CUSUM and CUSUMQ plots for the estimated ECMs
Figure 4: Residential electricity consumption, real disposable income and real electricity price
Figure 5: Residential natural gas consumption
Figure 6: Real net disposable income
Figure 7: Real residential natural gas prices
Figure 8: Heating Degree Days
Figure 9: CUSUM and CUSUM of squares plots for the estimated ARDL models

8.2 List of Tables

Table 1: Industrial electricity demand studies and elasticity estimates
Table 2: Residential electricity demand studies and elasticity estimates
Table 3: Residential natural gas demand studies and elasticity estimates
Table 4: Results from causality analyses on residential electricity demand in the literature
Table 5: Price-elasticities and elasticities of substitution for Switzerland (private & industrial sector)
Table 6: Data availability, aggregational groupings and matches
Table 7: Unit root tests
Table 8: Vector autoregression (VAR) specification and diagnostic tests
Table 9: Cointegration tests with and without level shifts
Table 10: Estimated long-run relationships
Table 11: Granger causality analysis
Table 12: Estimated ECMs and diagnostic tests
Table 13: Panel unit root tests
Table 14: Pedronis panel cointegration tests

56
Table 15: Long-run elasticities by (group-means panel) FMOLS and DOLS estimation
Table 16: Comparison of long-run country-specific demand elasticities
Table 17: Short-run elasticities
Table 18: Panel Granger causality tests
Table 19: Bounds F-tests for a cointegration relationship
Table 20: Long-run coefficients for country-specific ARDL models
Table 21: Error correction representations for the underlying ARDL models
Table 22: Diagnostic tests for the underlying ARDL models
Table 23: Price elasticities and elasticities of substitution
Table 24: Estimates of the coefficients of the proxy variable, time, for technical change, 1978-2006

8.3 Publications generated by the project

Bernstein R., Madlener R. (2011). Natural Gas Demand Elasticities in OECD Countries: An ARDL
Bounds Testing Approach, FCN Working Paper No. 15/2011, Institute for Future Energy Consumer
Needs and Behavior, RWTH Aachen University, October.
Bernstein R., Madlener R. (2011). Responsiveness of Residential Electricity Demand in OECD
Countries: A Panel Cointegration and Causality Analysis, FCN Working Paper No. 8/2011, Institute
for Future Energy Consumer Needs and Behavior, RWTH Aachen University, April.
Bernstein R., Madlener R. (2010). Short- and Long-Run Electricity Demand Elasticities at the
Subsectoral Level: A Cointegration Analysis for German Manufacturing Industries, FCN Working
Paper No. 19/2010, Institute for Future Energy Consumer Needs and Behavior, RWTH Aachen
University, November.

8.4 Short CV of scientists involved in the project

Prof. Reinhard Madlener studied Commerce and Finance as well as Paedagogics at the Vienna
University of Economics and Business Administration (WU Wien) and then also Economics at the
Institute for Advanced Studies Vienna (IHS). He obtained his PhD at WU Wien in the Economics and
Social Sciences (Dr. rer. soc. oec.), specializing in General Economics, Environmental Economics,
and Statistics. Before taking up his position at RWTH Aachen University in June 2007, he was
Managing Director of the Institute for Advanced Studies Carinthia (1999-2000), Assistant Professor at
the Centre for Energy Policy and Economics (CEPE), ETH Zurich (2001-2007), Lecturer at the Faculty
of Economics, University of Zurich (since 2003), and Senior Researcher at the German Institute of
Economic Research / DIW Berlin (2007). Among others, he was Visiting Fellow at the University of
Illinois (Urbana-Champaign), the European University Institute (Florence, Italy), and the University of
Warwick (Coventry, UK). Prof. Madlener is one of five full professors of the E.ON Energy Research
Center (E.ON ERC), established at RWTH Aachen University end of 2006, and Director of the Institute
for Future Energy Consumer Needs and Behavior (FCN) founded by him in June 2007. He is also
RWTH Director of the Juelich-Aachen Research Alliance Energy (JARA-Energy) (since January 2008)
and Research Professor with the German Institute of Economic Research (DIW Berlin) (since
December 2010).
Ronald Bernstein Diplom-Volkswirt (Freie Universitt Berlin, Germany), joined the Institute for Future
Energy Consumer Needs and Behavior (FCN) at RWTH Aachen University in May 2008 as a
Research Associate. His main research interests lie in energy modeling with a focus on applied time
series econometrics. His studies concentrate on the econometric estimation of energy demand
elasticities and on the analysis of causality in the energy-growth nexus.
Dr. Miguel Angel Alva-Gonzalez Since September 2009 is a post-doctoral Research Associate with
FCN, RWTH Aachen University. He studied Economics at the Technological Institute of Monterrey
(ITESM), Mexico, earned his MSc in Environmental and Resource Economics from the University
College London, and obtained a PhD in Agricultural Economics from the University of Bonn. His recent
research focuses on sustainable transport, consumption behavior, and urban development.

57
8.5 Project timeline

8.6 Activities within the scope of the project

Bernstein R., Madlener, R. (2011). Short- and Long-Run Electricity Demand Elasticities at the
Subsectoral Level: A Cointegration Analysis for German Manufacturing Industries. 10th Workshop of
the GEE Student Chapter, 8-9 May 2011, RWI Essen, Germany.

58
Econometric Estimation of Energy Demand Elasticities

Project Synopsis

Authors: Reinhard Madlener, Ronald Bernstein, Miguel ngel Alva Gonzlez

FCN Institute for Future Energy Consumer Needs and Behavior

E.ON Energy Research Center (E.ON ERC), RWTH Aachen University
Mathieustr. 6
52074 Aachen, Germany

Univ.-Prof. Dr. rer. soc. oec. Reinhard Madlener

Tel.: +49 241/80 49820
Fax.: +49 241/80 49829
RMadlener@eonerc.rwth-aachen.de

Dipl.-Volksw. Ronald Bernstein

Tel.: +49 241/80 49832
Fax.: +49 241/80 49829
RBernstein@eonerc.rwth-aachen.de

Dr. Miguel ngel Alva Gonzlez

Tel.: +49 241/80 49839
Fax.: +49 241/80 49829
MAlva-gonzalez@eonerc.rwth-aachen.de

Categories E.ON ERC focus:

Small Scale CHP Large Power Plants
Energy Storage Energy Efficiency
Consumer Behavior Energy Economics Modeling
Energy and Buildings Power Electronics
Distribution Networks Renewable Energy
Carbon Storage (CCS) Others

Type of project report: Final Project Report

Start and end date of project: July 2008 - October 2011
Project in planned timelines: yes no (see section 8.5)

Participating Chairs of E.ON ERC:

Automation of Complex Power Systems (ACS)
Energy Efficient Buildings and Indoor Climate (EBC)
Future Energy Consumer Needs and Behavior (FCN)
Applied Geophysics and Geothermal Energy (GGE)
Power Generation and Storage Systems (PGS)

Acknowledgments:
This project was supported by a grant of E.ON ERC gGmbH, Project No. 03-015

59
E.ON Energy Research Center Series

ISSN: 1868-7415
First Edition: Aachen, October 2011

E.ON Energy Research Center,

RWTH Aachen University

Mathieustrae
52074 Aachen
Germany

T +49 (0)241 80 49667

F +49 (0)241 80 49669
post_erc@eonerc.rwth-aachen.de
www.eonerc.rwth-aachen.de