You are on page 1of 20

Energy 25 (2000) 149–168

www.elsevier.com/locate/energy

Modeling demand for coal in India: vector autoregressive


models with cointegrated variables
Mudit Kulshreshtha *, Jyoti K. Parikh
Indira Gandhi Institute of Development Research, Gen A.K. Vaidya Marg, Goregaon (East), Mumbai 400 065, India
Received 8 March 1999

Abstract

In this paper, long-run structural relationships of coal demand with price and income variables have
been estimated for the four major coal consuming sectors in India using annual time series data from 1970–
1995. Some of the recent developments in multivariate dynamic econometric time series modeling tech-
niques have been used, including the estimation of long-run cointegrating relationships, short-run dynamics
and measurement of the effects of shocks and their effect on the evolution of dynamic coal demand system.
The models have been estimated using cointegrating VAR framework, which allows for endogeneity of
regressors. Results indicate that coal demand is likely to grow more than proportionately with economic
growth due to high GDP elasticities and low price elasticities. Further, coal prices are found to be weakly
exogenous in all the sectors except cement. Persistence profiles indicate that coal demand systems in the
four sectors seem to be stable and converge to equilibrium within a period of around 4–7 years after a
typical system-wide shock.  2000 Elsevier Science Ltd. All rights reserved.

1. Introduction

During the last two decades significant movements in energy prices and rapid economic growth
in developing countries have created a unique experimental framework to assess the responsive-
ness of energy demand to economic factors. In India, over 55% of commercial energy consumption
comes from coal which is a major energy resource. Estimating the demand for coal is important
not only for the coal sector per se but also from the standpoint of overall economy and the local
and global environment. The problems associated with modeling of energy demand involve cap-
turing the dynamics of demand, determining the appropriate level of aggregation, identifying the
separate influences on demand, separating supply from demand effects and using appropriate

* Corresponding author. Fax: +91 022 8402752.


E-mail address: mudit@igidr.ac.in (M. Kulshreshtha)

0360-5442/00/$ - see front matter  2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 3 6 0 - 5 4 4 2 ( 9 9 ) 0 0 0 5 9 - 6
150 M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168

Nomenclature1
CP Coal consumption (’000 tonnes) by power sector, after adjustment for declining
quality of coal
CC Coal consumption in cement industry (’000 tonnes)
CS Coal consumption in steel industry including imported coal (’000 tonnes)
CO Coal consumption in other industries including textile, paper and fertilizer (’000
tonnes)
GDP Gross domestic product at market price at 1980–81 prices (Rs.crores, 1crore=10
million)
PNCK Real price of non-coking coal
PCK Real price of coking coal
PCEM Real price of cement
PST Real price of steel
ALT Generation of alternative sources of power, viz. hydro and nuclear
IIP Industrial index of production (1980–81=100)

equation forms and estimation techniques [1,2]. Specifically, the issue concerning dynamics of
demand pertain to the distinction between short-run and long-run adjustments (or conventionally
the distinction between stock utilization and stock adjustments). The traditional approaches of
modeling energy demand have included the use of panel data models [3], variants of the distrib-
uted lag models [4] market share framework of models, multivariate structural demand models
[5] and a new generation of dynamic factor demand models [6–8].
The traditional approaches are data intensive and often associated with several empirical prob-
lems such as the possible endogeneity of regressors as well as the nonstationarity of the variables.
The standard classical methods such as the ordinary least squares (OLS) and hypotheses testing
are based on the assumption that the time series are stationary.2 Since the distribution theory that
applies to non-stationary series is different from the standard Gaussian asymptotic theory, use of
classical estimation methods such as OLS for estimating relationships between non-stationary
variables may give rise to misleading inferences or spurious regressions. The problems with esti-
mation of single equation framework with integrated or nonstationary variables are: non standard
distributions of coefficient estimates, error process not being stationary, explanatory variables
generated by processes that display autocorrelation, existence of more than one cointegrating
vector and failure of weak exogeneity [9,10]. The remedy from problematic regressions with
integrated variables is to test for cointegration and estimation of vector error-correction models
to distinguish between short-run and long-run responses. The cointegrating relationship may be
thought of as a long-run steady state of dynamic relationship though there can be finite short-run

1
All the series are converted to natural logarithms for the purpose of the analysis.
2
A series is stationary if its mean and variance are constant over time and the value of covariance between two time periods
depends only on the distance or lag between the two time periods and not on the actual time at which the covariance is computed.
A nonstationary series is said to be integrated of order d or I(d) if it must be differenced d times to make it stationary.
M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168 151

variations around the long-run equilibrium. The existence of the cointegrating or long-run relation-
ship also implies and is implied by the existence of some adjustment process that prevents the
errors in the long-run relationship from becoming larger and larger. This is modelled through the
popular econometric specification of error-correction model (ECM) which integrates the long-run
equilibrium analysis and short-run dynamic adjustment by including in the short-run dynamic
models a measure of disequilibrium (measured by deviations from long-run relationship) in the
previous period.
The cointegration and ECM framework have been used for modeling energy demand in a
number of studies, e.g. modeling of Danish gasoline demand [11], road transport energy demand
for Australia [12], coal demand in China [13], UK’s final user energy demand [14], and to estimate
gasoline demand in India [15]. In these studies the multivariate Johansen–Juselius cointegration
framework [18–21] is used only to ascertain the cointegration rank and the univariate approaches
to cointegration such as the Engle–Granger residual based two step method [16], or dynamic OLS
[17] are used to estimate the models (see Section 3.1).
In this paper, the long-run structural relationships for coal demand in India is estimated using
the maximum likelihood procedure in multivariate cointegrating vector autoregressive (VAR)
framework. Annual time series data for the period 1970–1995 is used for the four major coal
consuming sectors, viz. power, cement, steel and others, which includes rest of the industries.
Both coking and non-coking coal are considered with the demand for coking coal attributed only
to the steel sector.
Estimation of models in cointegrating VAR in the present analysis provides certain advantages
over methods employed in previous studies on energy demand. First, it takes into account the
endogeneity and simultaneity of variables by allowing the error term to be contemporaneously
correlated across equations. Second, it allows one to identify which of the variables in the demand
system adjusts towards equilibrium. Third, since long-run structural estimates and parametric
restrictions are explicitly incorporated in multivariate short-run equations, it allows one to model
the short-run dynamics of adjustment. This is done in terms of the demand system’s response to
system-wide and variable-specific shocks (typically known as “innovations”) to obtain meaningful
inferences about weak exogeneity through innovation-accounting tools such as impulse responses
and variance decompositions.
In this paper, the effect of a “typical” sytemwide shock and the speed with which the coal
demand system responds to disequilibrium and finally converges is studied through persistence
profiles [22,23]. The effect of variable specific shocks or innovations is studied through gen-
eralized impulse responses which are invariant with respect to ordering [24] and thus avoids
problems associated with “ordering” in orthogonalized shocks [25].
Past attempts of projecting coal demand in India have relied on simple techniques not taking
into account non-stationarity of data series. The Government of India uses the normative method
based on consumption norm [26] or at times based on input–output models. Parikh [27] used
conventional regression techniques to model energy demand in major sectors. The MEDEE—S
approach [28] was also a disaggregated, end use energy-demand analysis. In this paper, the appli-
cation of cointegrating VAR methodology has been demonstrated, which could be applied to
estimate a number of useful policy parameters for calibrating large-scale energy models. The
policy issues, which we seek to address, are:
1. What is the long-run relationship between economic growth, coal prices and coal consumption
152 M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168

in the major coal consuming sectors of India? What are the long-run price and income elas-
ticities?
2. What is the nature of feedback between economic growth and coal demand in each of the sec-
tors?
3. Is the coal demand system stable? What is the response of sectoral coal demand to system
shocks and variable specific shocks?
4. What are the factors that influence the dynamics of short-run adjustments of the coal demand
system? Which variables adjust to shocks and which do not?
5. What kind of policy conclusions can one derive from the response of coal demand to changes
in policy variables such as coal price and economic growth?

The paper is organized as follows. Section 2 gives a brief overview of coal demand in India and
its consumption patterns. Section 3 gives an account of time series econometric methodology
employed. Section 4 discusses the results. In Section 5, a conclusion and policy implications have
been provided. The data sources are provided in Appendix A.

2. Overview of coal demand

India is the fourth largest coal producing country in the world after China, the USA and Russia.
It has 7% of the world’s proven coal reserves, which occur, in thick seams and shallow depths.
The total coal reserve in 1995 was assessed at nearly 200 bt of which 34% was proven reserve.
85% of the total reserve was of non-coking coal and the rest of coking coal. The coal production
has risen from 70 mt in 1970–71 to 300 mt in 1997–98. The power sector is the largest consumer
of coal in India. Over 67% of electricity is generated from the thermal power stations which
primarily use coal. In addition, critical infrastructure industries like steel, cement and large number
of fertilizer, chemicals, paper and other medium and small-scale industries are dependent on coal.
Of the total consumption, 66% is consumed by the power sector, 13% by steel, 4% by cement
and 17% by the other industries. Presently, India is not a major exporter of coal but imports
around 10 mt of coking coal principally for use by steel plants. The Ninth Five-Year Plan (1997–
2001) has projected the coal demand to be 416 mt by 2001–02. Consumption demand for coal
has risen at an annual rate of 6% since 1992–93. Although oil can be considered as the closest
substitute for coal, oil conservation has always been the mainstay of India’s energy policy due
to the volatile nature of international energy markets and the foreign exchange constraints faced
by India. To counter any threat of oil price hikes over the next decades, an energy policy evolving
around the abundantly available domestic coal resources has become a major priority in India
[29]. Due to the stagnation of domestic crude oil production, uncertain reservoir behaviour in
Bombay offshore basins and problems in North East India, the Approach paper to the Ninth plan
has emphasized the importance of utilizing the abundant domestic resource base of coal in the
country. It has projected that only 34.3% of total consumption of oil would be met by domestic
production in 2001–02 in comparison to 45% in 1996–97. Hence, meeting energy demand through
coal would reduce the vulnerability of the economy to the uncertainties of external markets as
well as the threat of a balance of payment crisis.
M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168 153

3. Methodology and econometric modeling issues

3.1. Approaches to cointegration

A number of methods for the estimation of long-run cointegration vector (or long-run equilib-
rium relationship) have been proposed since the introduction of the concept of cointegration.
These are the two-step residual-based method of Engle and Granger [16], non linear least squares
(NLS) by Stock [30], principal components by Stock and Watson [31], maximum likelihood model
by Johansen [18,19] and Johansen and Juselius [20,21], Engle and Yoo three step procedure [32],
the auxiliary regression procedure [33] and Stock and Watson dynamic OLS [17]. Based on Monte
Carlo evidence, Gonzalo [34] has shown that though the estimates can vary significantly, despite
being super consistent, the maximum likelihood estimates in a fully specified error-correction
model (ECM) or Johansen’s approach has better properties (such as smallest mean squared error
across a range of parameter values) than other estimators and their finite sample properties are
consistent with asymptotic results.
The estimation and identification of cointegrating relationships in the multivariate framework
avoids many of the assumptions implicit in studies that use static or dynamic univariate framework
like the Engle Granger and ADL (Autoregressive Distributed Lags) approaches. The single-equ-
ation representations assume the existence of at most a single cointegrating relationship and do
not take into account the possible endogeneity of explanatory variables. The single equation frame-
work is estimated under the highly restrictive assumption that any disequilibrium would be cor-
rected by only one variable, i.e. the dependent variable, which will adjust in a direction opposite
to that of disequilibrium.3 This limitation is overcome in a cointegrating VAR approach, which
allows a number of variables to adjust and respond at different rates to disturbances in a way
such that the system converges to long-run equilibrium. Further, Monte Carlo experiments suggest
that the omitted dynamics in static univariate approaches, such as EG, may result in large finite
sample biases in estimates [9,10,35]. Dynamic single-equation approaches such as ADL and fully-
modified OLS approaches [36], give efficient estimates only when there is at most one cointegrat-
ing relationship and when (n⫺1) variables are weakly exogenous4 [10]. The multivariate cointe-
grating VAR procedure [20,21] does not, a priori, assume the existence of at most one cointegrat-
ing vector and explicitly tests for the number of cointegrating relationships. Typically, n variables
integrated of the same order can have at most n⫺1 linearly independent cointegrating vectors.

3.2. Vector autoregressive models with cointegrated variables

Prior to estimation of the cointegration space and determination of the cointegration rank in a
multivariate framework, it is important to test for the order of integration of each variable and
ensure that the variables are at most I(1). The order of integration is ascertained by testing the

3
This is equivalent to assuming that the other variables are weakly exogenous (see Section 3.2).
4
The works of Inder [35], Urbain [37] suggest that Wu–Hausman type orthogonality tests are unlikely to provide clear results
on exogeneity.
154 M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168

null hypothesis that a series contains a unit root against the alternative of stationarity through
Augmented Dickey Fuller (ADF) tests.5 The basic equations for ADF tests are:


p

⌬Y⫽a0⫹a1Yt−1⫹ gj ⌬Yt−j ⫹et (1)


j⫽1


p

⌬Y⫽a0⫹a1Yt−1⫹a2t⫹ gj ⌬Yt−j ⫹et (2)


j⫽1

where ⌬ represents the first difference operator and et is Gaussian white noise term. The null of
non-stationarity is equivalent to testing for the significance of a1=0.
We consider a K-dimensional vector autoregressive (VAR) model of the form
Yt⫽A1Yt−1⫹…ApYt−p⫹m⫹⌿Dt⫹ut (3)
where Yt=(y1t, …, yKt)⬘, Ai are K×K coefficient matrices, ut is Gaussian white noise; ut and us are
independent for s⫽t and ut苲N(0,⌺) (where ⌺ is a K×K positive definite matrix), yi are jointly
determined variables which are at most I(1). Dt is a vector of non stochastic variables [39]. Under
the hypothesis of cointegration the model can be reparameterized in equivalent error correction
form6
⌬Yt⫽⌫1⌬Yt−1⫹…⫹⌫p−1⌬Yt−p−1⫹⌸Yt−1⫹m⫹⌿Dt⫹et. (4)
The K×K matrix ⌸ can be expressed as ⌸=ab⬘ where both a and b are K×r matrix of full rank.
b⬘ is a matrix representing cointegrating relations such that b⬘Yt is stationary and is interpreted
as the long-run equilibrium relationship between the jointly determined variables.
The empirical implementation of the cointegration analysis has been done in the econometric
software package Microfit 4.0 [41]. The statistical framework used in the present analysis follows
from the vector error correction model used in Pesaran et al. [24], which is fundamentally similar
to the Johansen framework, but allows for a vector xt of random variables in Yt in Eq. (3) that
are weakly exogenous to the system. These I(1) variables are referred to as the “long-run forcing
variables”, in the sense that in the long-run they are not “caused” by other variables in the model,
although contemporaneous or short-run interaction exists. Therefore in the cointegration analysis,
the error correction terms do not enter in the sub-system for xt. The VECM can be written as:


p⫺1

⌬yt⫽ ⌫iy⌬Yt−i⫹⌸yYt−i⫹m⫹⌿yDt⫹et (5)


i⫽1

where, Yt=(yt⬘, xt⬘)⬘, yt is an my×1 vector of jointly determined (endogenous) I(1) variables, so that,

5
The inclusion of constant and trend terms in the tests were decided on the basis of joint tests as in SHAZAM 7.0 [38]. The
tests on first difference of the series were considered without trend. Often there are problems with these tests as regard to its size
and power. To check the possibility of serial correlation, further examination of residuals through correlogram and Ljung–Box Q-
statistics was done.
6
This follows from Granger representation theorem [16].
M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168 155


p⫺1

⌬xt⫽ ⌫ix⌬Yt−i⫹⌿xDt⫹nt (6)


i⫽1

Dt is a q×1 vector of exogenous/deterministic I(0) variables excluding the intercepts and/or trends.
Eq. (6) allows for feedbacks from ⌬y to ⌬x, but does not allow for level feedbacks, and hence
assumes that xts are not themselves cointegrated.
The word equilibrium in cointegration does not refer to market clearing condition but rather
the relation to which the system will converge over time. During short-run, there may be stochastic
shocks forcing the system to deviate from the equilibrium relation. However, in the presence of
cointegrating relationship, systematic co-movement of economic variables will restore long-run
relationship. The deviations from equilibrium relations b⬘Yt, form a stationary process and a
contains the speed of adjustment coefficients for the equation.
For the model used for each of the sectors in this paper, K=3, Yt=(coal demand variable, income
and coal price)⬘,7 while Dt refers to the I(0) exogenous variables. GDP has been chosen as the
uniform income variable across all the models instead of industry specific value added (VA)
because of the greater cyclical variation in the latter.8 VA being more disaggregated, is likely to
have more noise which would reduce the precision of the estimates [40]. The estimated elasticity
of sectoral energy demand to GDP can be thought of as the product of elasticity of energy demand
to sectoral value added and the elasticity of value added to GDP. All the industries considered
here are input industries. Hence, their outputs are demand driven and India being a developing
country, demand for these crucial infrastructure industries is determined by economic growth. In
the near future their share in total national output is most likely to follow the present structural pat-
tern.
The sector specific information in the disaggregated equations for the four industries, comes
from the use of sector specific I(0) exogenous variables. It is growth in alternative power gener-
ation (⌬ALT) in the model for power sector, price of cement (PCEM) in the model for cement,
price of steel (PST) and growth in industrial index of production (⌬IIP) in the model for steel
and ⌬IIP in others. The basic purpose of including these variables is to allow for the short-run
movements in the I(1) variables which move them away from their long-run equilibrium [41].
The price of power has not been considered in the model for power because power prices are
administered and are not homogenous across consumers. Instead, growth in alternative power
generation is included in the model for power sector as I(0) exogenous variable. Rise in alternative
power generation is likely to reduce requirement of thermal power generation and thus reduce
coal demand by the power sector. Growth in industrial index of production is also introduced as
I(0) variable in the model for steel industry and in the model for other industries. Rise in industrial
activity is likely to affect steel demand and in turn coal demand in the steel industry.
In the four models, for each of the sectors the (3×1) a=(a1, a2, a3)⬘ matrix represents the speed
of adjustment coefficients. a1 shows how y1 or coal demand responds to deviations; a2 shows
how y2 or the income variable responds to deviations and a3 shows how y3 or coal price responds

7
Price of oil was initially included in the demand function as substitute price, but was found insignificant in the long-run model
and hence dropped to prevent over parameterization of the model.
8
While ascertaining VA, both output and input prices are taken into account and the latter have considerable cyclical variations.
In GDP the impact of such cyclical component gets diluted due to aggregation.
156 M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168

to deviations. The simultaneous equation system, thus, refers to Eq. (4). In this equation, there
exists contemporaneous correlation between the error terms, thus involving simultaneous bi-direc-
tional feedbacks between the endogenous variables coal demand, GDP and coal price. The I(0)
exogenous variables in each of the models, while causing fluctuations in the demand system in
short-run does not influence the relationship in the long-run due to adjustments brought about by
endogenous variables.
Under the reduced rank hypothesis of ⌸ matrix, the maximum eigenvalue and the trace statistics
advanced by Johansen [18,19] are employed to ascertain the number of cointegrating vectors. The
structural estimates of the long-run model are estimated by the maximum likelihood procedure.
Several authors like Canova [42], Pesaran and Smith [43] and Pesaran [44] have emphasized the
validity of economic theory only in the long-run and its inadequacy in the short-run. It is difficult
to identify the structure behind the VAR through meaningful modeling of the contemporaneous
influences. Since economic theory is usually inadequate about short-run adjustment processes,
which involve complex bi-directional and contemporaneous feedback, the short-run dynamics
are modelled through vector error-correction model (VECM) in an unrestricted VAR modeling
framework. The VECMs allow a number of variables to adjust simultaneously at different rates
in response to short-run disequilibrium. The approach provides a good approximation to the
unknown data generating process. Any deviation in the long-run relationship feeds back into the
cointegrating variables, which in turn adjust to bring the system back to long-run equilibrium.
This adjustment is captured through the error correction term (ECT) in the vector error correction
models (VECMs) and the error correction coefficient is known as the speed of adjustment. It is
possible to infer about the weak exogeneity of a variable from the VECMs as variables not
responding to the ECT are deemed as weakly exogenous. The test for weak exogeneity of variable
YK with respect to b in a multivariate cointegrated VAR of dimension K, when the parameters
of interest are a and b, is that aKj=0 ∀ j where ⌸=ab⬘ and both a and b are K×r matrix of full
rank [Eq. (4)] [39]. The final long-run conditional model is estimated by incorporating the findings
regarding weak exogeneity of the I(1) variables from the VECMs. The model is re-estimated with
VECMs as given in Eq. (5). The weakly exogenous I(1) variables are specified as the long-run
forcing variable xt. Conditioning on the weakly exogenous variables is considered advantageous
as a means of improving the stochastic properties of the model [39].

3.3. Short-run dynamics: persistence profiles and impulse response analysis

The time profile of a typical system-wide shock 9 on the evolution of dynamic coal demand
system is analyzed using the persistence profiles approach [23]. The persistence profile is also
indicative of the stability of the long-run relationship. Since a number of variables can adjust in a
cointegrating VAR (by increasing or decreasing at different rates), the cointegrated or equilibrium
relations act as stabilizer for the system. Hence, shocks to the cointegrating relationship will die
out though there might be persistent effect of shocks on individual variables. The persistence
profile is therefore illustrative of the speed with which the coal demand system for each of the

9
That is all elements of ut vector of VAR are shocked simultaneously.
M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168 157

sectors converges to equilibrium in response to a “typical” system-wide shock that occurred during
the sample period.
The effect of innovations on individual equations in a VAR is analyzed through the impulse
response functions and the forecast error variance decomposition. Impulse responses measure the
time profile of the effect of shocks at a given point of time, on the (expected) future values of
variables in a dynamic system. The variance decomposition tells us how much of the average
squared forecast error variance of one variable at the kth step ahead is associated with surprise
movements in each variable of the model. Partitioning the forecast error of a certain variable into
proportions attributable to innovations in each of the variables in the system including its own
can provide indication of the relativities among the variables in the system. Both the innovation
accounting tools can be used to make inferences regarding the nature of dynamic interactions
between the variables. If the variance of a particular variable is explained primarily by its own
innovations then the variable is likely to be weakly exogenous to the system. The impulse
responses or dynamic multipliers can be obtained from infinite moving average representation of
a K-dimensional VAR model [45] as follows:
Yt⫽AtYt−1⫹…⫹ApYt−p⫹ut (7)


n

⌽n⫽(fik,n)⫽ ⌽n−j Aj (8)


j⫽1

where n=1, 2, …, ⌽0=IK, Aj=0 for j⬎p and jik,n is the ikth element of ⌽n represents the response
of variable yi to a unit shock in variable k, n periods ago. Since the covariance matrix of VAR
⌺u is positive-definite, it is essential to transform the innovations of the system into a contempor-
aneously uncorrelated form. If disturbances across equations are correlated, an innovation in one
of the equations describes its dynamic response to a complex combination of several economically
interpretable shocks. In order to have economically interpretable shocks, the orthogonalization
requires imposition of restrictions on contemporaneous coefficients of the underlying structural
VAR and hence involves imposing a particular “causal” order on the relationship. The orthog-
onalised impulse responses thus vary with re-ordering of the variables in the VAR model. Owing
to this, the generalized impulse response function is used, which is invariant to the reordering of
the variables in the VAR. It takes into account the historical patterns of correlation observed
amongst the different shocks10 and hence does not require orthogonalization. The generalized
impulse responses have the advantage that they take into account the properties of data generating
process into account and hence simulate the model’s response to conditions similar to that, which
occurred in the sample period. The orthogonalized impulse response on the other hand “turns off”
the contemporaneous effect of some of the shocks depending on the ordering of the variables.

4. Empirical results and discussion


4.1. Order of integration: unit root tests and tests of multivariate cointegration
The results of augmented Dickey Fuller tests are presented in Table 1. The description of data
series used is given in the list of nomenclatures. ADF tests were performed on the natural logarith-

10
The history invariance property is preserved only for the linear multivariate systems.
158 M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168

Table 1
Augmented Dickey–Fuller tests (sample 1970–1995)
a b c
Series Level First difference

CP ⫺2.8484 (1)d ⫺4.5467*


CC ⫺2.3953 (2) ⫺5.0002*
CS ⫺1.5363 (2) ⫺4.1268*
CO ⫺1.6151 (1) ⫺6.2843*
GDP ⫺2.3421 (1) ⫺5.7003*
PNCK ⫺2.7724 (1) ⫺5.2939*
PCK ⫺1.0996 (2) ⫺4.2263*
PCEM ⫺5.6235*(1) ⫺6.2920*
PST ⫺4.0057*(1) ⫺5.1220*
IIP ⫺1.7730(1) ⫺3.5126*
ALT ⫺2.9328(1) ⫺6.8078*
a
All the variables are in natural logarithmic form.
b
The Dicky Fuller regressions include an intercept and a linear trend.
c
The Dicky Fuller regressions include an intercept but no trend.
d
(1) Figures in parentheses indicate the number of lags chosen for the ADF test. (2) * Denotes rejection of the null
of non-stationarity at 5% significance level. The critical value of the ADF statistic at 5% level of significance for
regressions with intercept but no trend was ⫺3.0819 and the same with both intercept and trend was ⫺3.7612. (3) To
check the possibility of serial correlation we examined correlogram of residuals and Ljung–Box Q-statistics. Both these
post regression diagnostic checks were done to ensure that there was no evidence of serial correlation in the residuals.
The critical values were from MacKinnon [46].

mic values of each of the variables on full sample for the period 1970–1995, both on levels as
well as on the differenced forms to ensure the order of integration. With the exception of PCEM
and PST, in case of all other variables in levels, the null of a unit root cannot be rejected. Further,
any evidence of serial correlation in the residuals was not found after exploring the correlogram
of residuals and Ljung–Box Q statistics. To ascertain the existence and number of cointegrating
relationships for each of the coal-consuming sectors, likelihood ratio tests based on maximal
eigenvalue and trace of the stochastic matrix in the multivariate framework were performed.
Results of these tests with 95% critical values are reported in Table 2 hypothesis of r=0 for all
the coal consuming sectors is rejected at 5% significance level using both tests, except in the case
of ltrace test for steel and power sectors, where the null hypothesis of r=0 is rejected at a 10%
level of significance. Thus for all the coal consuming sectors, there exists considerable evidence
for the existence of a single cointegrating vector.

4.2. Long-run structural relationships: identification of long-run elasticities

Having established the presence of single long-run equilibrating relationship for each of the
industries, maximum likelihood estimates of the long-run coal demand functions are presented in
Table 3. Results indicate that though economic growth is the major determinant of coal demand
in the long-run, the long-run income elasticities vary across sectors. In cement and steel, long-
run income elasticity is found to be around unity while in the power sector income elasticity is
greater than unity, about 1.6. All these three industries are crucial infrastructure industries having
M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168 159

Table 2
Likelihood ratio tests for multiple cointegrating vectors

H0 lmax ltrace

95% critical 95% critical


Test statistic Test statistic
value value

Power r=0c 26.70* 21.12 29.97** 31.54a


rⱕ1 3.20 14.88 3.26 17.86
rⱕ2 0.41 8.07 0.41 8.07
Cement r=0 28.88* 22.04 40.30* 34.87
rⱕ1 10.71 15.87 11.42 20.18
rⱕ2 0.71 9.16 0.71 9.16
Steel r=0 22.47* 22.04 33.15** 34.87b
rⱕ1 7.34 15.87 10.68 20.18
rⱕ2 3.33 9.16 3.33 9.16
Others r=0 31.61* 21.12 41.15* 31.54
rⱕ1 8.43 14.88 9.55 17.86
rⱕ2 1.11 8.07 1.11 8.07
a
The critical value at 10% significance level is 28.78.
b
The critical value at 10% significance level is 31.93. The critical values are as computed by Pesaran et al. [47].
Econometric software package used was Microfit for Windows, Version 4.0.
c
r indicates the number of cointegrating relationships. * and ** denote rejection of the null at 5% and 10%
significance levels respectively. The optimal lag length of VAR was selected by minimizing Akaike’s FPE criterion.
The optimal order of VAR was found to be 2. The alternative hypothesis for lmax is r=n and for ltrace is rⱖn, where
n is the hypothesized number of cointegrating relationships.

Table 3
Long-run structural relationships (ML estimates) for the conditional models

Power CP = 1.573 GDP ⫺ 0.125 PNCK


(0.251)a (0.105)
Cement CC = 9.256 + 1.196 GDP ⫺ 0.665 PNCK
(1.278) (0.100) (0.205)
Steel CS = 3.103 + 1.045 GDP ⫺ 0.421 PCK
(1.470) (0.150) (0.067)
Other CO = 0.675 GDP ⫺ 0.214 PNCK
(0.067) (0.103)
a
The values in parentheses indicate standard error.

strong forward and backward linkages with the rest of the economy. These industries together
accounted for 82.6% of the total coal consumption in 1995–96. The power sector alone contributes
to 66% of the total coal consumption. The relatively higher income elasticity for the power sector
may be attributable to the fact that India is a developing economy where early economic growth
is associated with shifts from the use of non-commercial to commercial energy forms like elec-
tricity. India’s major source of power generation is thermal power plants, which are mostly coal
160 M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168

based.11 The income elasticity is comparatively lower in others because this sector includes minor
industries with low energy requirements.
Price responsiveness for coal demand varies from negligible elasticity in power sector12 to 0.6
for cement. The weak responsiveness of coal demand to coal prices can be attributed to the
administered nature of the prices. This is particularly true for the power sector, where there is
large cross subsidization of power generation through prices of power grade coal. In India, conser-
vation of oil rather than coal has always been actively pursued through energy policy in wake of
the foreign exchange constraints and the abundance of domestic coal resources. The relatively
high price elasticity for cement could be because of the fact that the cost of energy input in
cement manufacture is considerably higher relative to other inputs used. In case of the steel sector
where coking coal is used, price elasticity, although moderate (⫺0.4) is found to be significant.
Demand is moderately responsive to price of coking coal because coking coal price is set relatively
higher and is more reflective of market price. The low price elasticities of coal demand in the
major coal consuming sectors indicate that the effect of any coal price increase will have small
impact on coal consumption. Also, the high GDP elasticity would more than offset any negative
impact of coal price rise.

4.3. Short-run dynamics

4.3.1. Short term responses from VECMs


A battery of diagnostic tests was applied to each of the VECMs estimated in the multivariate
VAR framework to check the robustness of the models to various departures from standard
regression assumptions like serial correlation, heteroscedasticity, misspecification of functional
forms or non normality of residuals. The results of these tests along with the error correction
(EC) coefficients for each of the endogenous variables for all the sectors are reported in Table
4. The cumulative sum of recursive residuals (CUSUM) test and the cumulative sum of squares
of recursive residuals (CUSUMSQ) tests were conducted to investigate the stability of the model
parameters for the equation for coal demand. The null hypothesis of stability could not be rejected
since the CUSUM and CUSUMSQ were found to lie within critical lines of 5% significance level.
The error correction coefficients (EC) are important for identifying the variables, which adjust
to restore long-run equilibrium, whenever there is any deviation from it. The ECs were found
significant for coal demand in all the sectors. In cement, price of non-coking coal and GDP also
have significant EC terms while in power and steel, only the EC for GDP is statistically significant
besides coal demand. In others, the GDP and price variables do not seem to respond to any
disturbance in the long-run equilibrium, with statistically insignificant ECs. This implies that all
the variables in cement, coal demand and GDP in power and steel and only the demand variable
in other industries adjust to shocks in the system and revert the system back to long-run equilib-
rium. The VECMs do not give the structural coefficients for the cointegrating variables, hence
we do not attempt to explain them [48].
The statistical insignificance of EC implies the possibility of the variable being weakly exogen-

11
In 1995–96 around 72% of the gross power generation was coal based.
12
The overidentifying restriction of price elasticity being insignificant could not be rejected. The likelihood ratio test statistic (c2)
for this hypothesis was 0.9053, producing a P-value of 0.331.
M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168 161

Table 4
Estimates of the error correction coefficients and diagnostic statistics for coal demand system in different sectors

Equation ECTt⫺1 SCa FF NN HET


c2(1) c2(1) c2(2) c2(1)

Power ⌬CP ⫺0.340* 2.905 0.194 0.325 0.006


[0.148]b (0.088) (0.659) (0.850) (0.938)
⌬PNCK ⫺0.193 0.665 0.822 2.386 0.503
[0.256] (0.415) (0.774) (0.303) (0.478)
⌬GDP 0.246* 2.522 2.508 4.004 1.565
[0.054] (0.112) (0.113) (0.135) (0.282)
Cement ⌬CC 0.357* 0.156 1.478 0.117 0.177
[0.185] (0.693) (0.224) (0.943) (0.674)
⌬GDP 0.271* 0.013 3.846 0.163 0.162
[0.099] (0.907) (0.060) (0.921) (0.687)
⌬PNCK ⫺0.616* 0.015 0.043 0.882 0.464
[0.192] (0.900) (0.835) (0.643) (0.496)
Steel ⌬CS ⫺0.783* 1.922 0.381 0.827 0.760
[0.207] (0.166) (0.537) (0.661) (0.383)
⌬GDP 0.165** 0.025 0.415 2.398 0.554
[0.100] (0.872) (0.519) (0.301) (0.457)
⌬PCK 0.013 0.009 0.877 2.829 0.966
[0.597] (0.922) (0.767) (0.243) (0.325)
Others ⌬CO ⫺0.556* 1.598 0.772 4.026 1.672
[0.102] (0.206) (0.379) (0.134) (0.196)
⌬GDP 0.115 0.885 0.299 2.220 0.412
[0.117] [0.347] (0.584) (0.136) (0.521)
⌬PNCK ⫺0.634 0.386 0.741 2.180 0.559
[0.323] (0.534) (0.389) (0.336) (0.454)
a
SC denotes the Lagrange multiplier statistic for testing the null of no serial correlation, FF is Ramsey’s RESET
test statistic, NN is Jarque–Bera statistic for testing the null of Gaussian errors, and HET is the statistic for testing the
null of no heteroscedasticity.
b
The figures in [ ] are asymptotic standard errors whereas those in ( ) are P-values. * denotes statistical significance
at 95% level. ** The corresponding P-value was found to be 0.11 indicating statistical significance at 89%.

ous to the system. Table 5 confirms that the null of weak exogeneity for the long-run parameters
b, cannot be rejected for the coal price variable in power, steel and others, and income variable
in others, based on a significance level of 5%.13 Price of coal not responding to the EC term in any
of the cases excepting cement seems reasonable given the fact that price of coal is administered in
India and is not driven by market forces. The magnitude of the EC is found to be around 0.3 in
the coal demand equations of both power and cement implying moderate rates of adjustment by
coal demand in both sectors. In steel and others, the speeds of adjustment of coal consumption
are relatively more rapid of around 0.7 and 0.55 respectively. Although the EC for coal demand
in cement is positive, positive and significant EC for GDP and negative and significant EC for

13
The null hypothesis can be rejected at 10% significance level for GDP in the steel sector.
162 M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168

Table 5
c2 Test for weak exogeneity of variables in the coal demand system for different sectors

Sector I(1) variables c2 statistic Sector I(1) variables c2 statistic

Power CP 6.32*a Steel CS 11.12*


GDP 14.86* GDP 3.47**
PNCK 1.71 PCK 0.12
Cement CC 9.40* Others CO 6.53*
GDP 13.71* GDP 0.06
PNCK 7.49* PNCK 0.06
a
*Denotes statistical significance at 95% level. The critical value is 3.84. **Denotes statistical significance at
90% level.

coal price ensure that the joint process of adjustment of the three variables brings the system
back to long-run equilibrium.14

4.3.2. Persistence profiles and impulse responses of variables


The joint process of adjustment is illustrated in the persistence profiles for the four sectors in
Fig. 1. Fig. 1 suggests that the demand system is stable in all the four sectors. The major impact

Fig. 1. Persistence profiles for the four sectors.

14
Since the estimation is done in a multivariate framework, where all the endogenous cointegrating variables adjust simultaneously,
it is not necessary for the sign of EC term to be opposite to that of the variable’s coefficient. This condition is however required for
a univariate framework to be stable.
M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168 163

Fig. 2. Generalized impulse responses for the power sector.

of any “typical” systemwide shock to the cointegrating relation dissipates within a period of
around 5 years in the power and steel sectors and a slightly longer period of around 6–8 years
in the cement sector. In case of others, the system fully returns to long-run equilibrium within 3
years. Since the persistence profile is indicative of the time during which the major impact of
typical systemwide shock gets diluted, it can also be used to interpret the span of the long-run,
e.g. it is around 4–5 years in the power and steel sectors.
The short-run dynamics is studied through impulse responses for power and steel sectors which
are the largest consumers of non-coking and coking coal respectively. Plots of generalized impulse
responses from shocks in the equations for GDP and coal prices have been shown for the two
sectors in Figs. 2 and 3. The figures indicate that a unit positive shock to GDP would result in
overshooting of demand by more than 1.5% in both industries in the first few periods after the
shock. Fig. 3 shows that an output shock has more persistent effects on coal demand in the case
of steel where it takes around ten periods for the shock to die down. The same in the case of
power is about five periods. In response to a positive shock on coal price, coal demand shows a
contemporaneous decline of around 1% in steel, while in power, there is a positive impulse
response of around 1%. In case of power, demand finally steadies at around 3% above initial
level, by the fourth period. In steel, coal demand declines further and finally stabilizes at around
5% below the original level by the ninth period. The impulse responses observed cannot be

Fig. 3. Generalized impulse responses for the steel sector.


164 M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168

entirely attributed to the shocked variable. Since complex bi-directional feedback exists among
all the variables, the propagation mechanism of the shock is not always obvious. The impulse
responses are of interest because they capture the overall impact of a shock in one variable of
the system on another, taking into account all the intermediate feedback among all the cointegrat-
ing variables of the system.

4.3.3. Error variance decomposition


Forecast error variance decomposition (FEVD) is studied along with impulse responses to
examine how much of the variance is attributable to each of the innovation processes. Different
lead lag orderings of the I(1) variables imply different implicit causal chains between the variables
in the short-run. Coal price (Table 5) being found weakly exogenous, was placed last in ordering
of the I(1) variables and the error variance decompositions in the power and steel industries were
studied for the two causal orderings of coal demand and GDP from innovations in each of these
variables. The variance decompositions are shown in Table 6 for power and Table 7 for steel
industry respectively. Results suggest that in both the sectors, innovations in coal demand are
explained more by coal prices than GDP for both the causal orderings of coal demand and GDP.
However, innovations to GDP are explained more by coal demand for both the causal orderings.
The error variance decompositions indicate that short-run variations in coal demand in both the
sectors can be explained substantially by coal prices.

5. Conclusions and policy implications

In this study, long-run structural relationships between coal demand and its influencing econ-
omic variables have been estimated using annual time series data for the period 1970–1995 for
Table 6
Error variance decompositions for the power sector

Ordering Horizona Due to innovations in CP Due to innovations in GDP

CP→GDP→PNCK CP GDP PNCK CP GDP PNCK

0 100.00 0.00 0.00 4.34 95.65 0.00


5 69.22 0.18 30.59 70.41 20.19 9.39
10 65.88 0.24 33.87 75.07 11.76 13.16
15 64.33 0.26 35.44 76.48 9.29 14.22
20 63.42 0.28 36.30 77.17 8.08 14.74

GDP→CP→PNCK GDP CP PNCK GDP CP PNCK

0 4.34 95.65 0.00 100.00 0.00 0.00


5 3.45 65.95 30.59 32.34 58.26 9.39
10 3.91 62.21 33.87 23.24 63.59 13.16
15 4.12 60.48 35.40 20.61 65.16 14.22
20 4.24 59.45 36.30 19.33 65.92 14.74
a
Horizon refers to the frequency of data, i.e. years. All figures are rounded to two places of decimal.
M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168 165

Table 7
Error variance decompositions for steel

Ordering Horizona Due to innovation in CS Due to innovation in GDP

CS→GDP→PCK CS GDP PCK CS GDP PCK

0 100.00 0.00 0.00 14.91 85.08 0.00


5 25.70 0.16 74.13 45.31 45.47 9.05
10 22.56 0.13 77.30 47.26 43.71 9.03
15 20.69 0.11 79.19 47.64 43.10 9.26
20 19.42 00.10 80.48 47.76 42.81 9.43

GDP→CS→PCK GDP CS PCK GDP CS PCK

0 14.91 85.08 0.00 100.00 0.00 0.00


5 4.83 21.03 74.13 73.34 17.44 9.22
10 4.32 18.37 77.30 73.96 17.00 9.03
15 3.98 16.82 79.19 74.17 16.57 9.26
20 3.75 15.76 80.48 74.24 16.31 9.44
a
Horizon refers to the frequency of data, i.e. years. All figures are rounded to two places of decimal.

the four major coal consuming sectors in India. The analysis was done in a multivariate cointegrat-
ing VAR framework to avoid the shortcomings associated with single equation approaches to
cointegration, employed in previous studies. Johansen’s maximum likelihood approach was used
to ascertain the cointegration rank and estimates, as its finite sample properties are consistent with
asymptotic results [34]. The results indicate that in the cement and steel industries, coal demand
can be expected to grow proportionately with economic growth, while in the power sector the
rise in coal demand can be more than proportionate in the long-run. Long-run price elasticities
were found to be low, especially in power where it is negligible. Persistence profiles indicated
that the coal demand system is fairly stable and that the effect of a typical system-wide shock
gets dissipated within a period of around 5 years in the power and steel sectors, 6 years in the
cement sector and only around 3 years in the other industries. These periods could also be inter-
preted as the span of the long-run for each of the sectors. Study of the error correction coefficients
in the VECMs and test for weak exogeneity indicated the coal prices to be weakly exogenous to
the system except in the cement sector. Both the coal demand and the income variable in all the
sectors, except the “others” sector, adjust to ensure long-run equilibrium. The effects of variable
specific innovations in the short-run were studied for the power and the steel industries through
the generalized impulse responses and the forecast error variance decompositions. Positive output
shocks in both the sectors result in positive impulse responses by coal demand. Study of FEVD
indicated that while innovations in GDP are explained more by coal demand, innovations in coal
demand are likely to be explained more by coal price and demand itself than by GDP in the sort
run. Negligible price elasticity in power and small price elasticity in steel and other sectors imply
that the overall effect of any coal price hike would be offset by positive response to income. The
intensity of coal consumption could rise for overall economy as the power sector currently
accounts for 66% of total coal consumption. Thus for sustained economic growth, India needs to
shift towards more efficient coal utilization technologies.
166 M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168

The energy markets in India exhibits the typical features of the early stages of a developing
economy, with the administrated pricing regime yet to be dismantled. This is reflected in the high
responsiveness of coal demand to economic growth and low responsiveness to coal prices. The
low own price elasticities (insignificant in the case of power sector) indicate little scope of substi-
tutability of coal by other fuels and that the prices of alternative fuels will not have much explana-
tory power in determining coal demand. Further, given our span of the long-run as indicated by
persistence profile, there is little likelihood of future behaviour being much different from the
present. Energy demand estimates are sometimes considered sensitive to changing institutional
environment and changing technologies [1]. However, there are reasons to believe that our esti-
mates would be applicable at least in the “long-run” time span of the model. Owing to the political
economy involved with deregulation of energy markets, energy supply security issues and various
market imperfections typical to a developing country, which are unfriendly to rapid penetration
of new technologies, the near future would be like the past.
The aim of this paper was to illustrate the application of multivariate cointegrating VAR meth-
odology to draw inferences about the responsiveness of coal demand to derive useful policy
parameters for calibration of large scale macro energy policy models. The methodological frame-
work used in this study is based on past trends and movements inherent in the time series data.
Hence, the findings of the study would be useful in establishing the business-as-usual (BAU) or
the base scenarios. However, the present methodological framework is flexible enough to generate
alternative policy scenarios by incorporating normative assumptions in alternative functional
forms, for example, declining income elasticities and rising price elasticities to capture economic
structure associated with different stages of economic growth.

Acknowledgements

The authors are thankful to the three anonymous referees for their comments and suggestions.
The usual disclaimer applies.

Appendix A. Data sources

Data for coal consumption by the various sectors and price series of coking and non coking
coal were obtained from Coal Directory of India [49] of various years. Quality of coal consumed
by the power sector declined steadily in the last two decades. Since energy from coal is what is
actually demanded, we have adjusted the coal consumed in physical terms for the declining gross
calorific value of coal (given in kcal/kg) consumed by this sector. Data for coal quality was
obtained from Public Electricity Supply, All India Statistics [50]. Data series for GDP at 1980–
81 prices and Industrial Index of Production (1980–81=100) were available from various issues
of National Accounts Statistics of India [51]. All the other price series were obtained from Index
of Wholesale Prices in India [52].
M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168 167

References

[1] Bohi DR. Analyzing Demand Behaviour: A Study of Energy Elasticities. Baltimore, MD: Johns Hopkins Univer-
sity Press, 1981.
[2] Watkins SC. The economic analysis of energy demand: Perspectives of a practitioner. In: Hawdon D, editor.
Energy Demand: Evidence and Expectations. UK: Guildford, 1992.
[3] Pindyck RS. The Structure of World Energy Demand. Cambridge, MA: MIT Press, 1980.
[4] Engsted T, Bentzen J. Expectation, adjustment costs and energy demand. Resources and Energy Economics
1993;15:371–85.
[5] McFadden D, Dubin J. Econometric analysis of residential electric appliance holdings and consumption. Econo-
metrica 1994;52:345–62.
[6] Berndt ER, Morrison C, Watkins GC. Dynamic models of energy demand: an assessment and comparison. In:
Berndt ER, Fields B, editors. Modeling and Measuring Natural Resource Substitution. Cambridge, MA: MIT
Press, 1981.
[7] Watkins GC. Short-and long-run equilibria: relationship between first and third generation dynamic factor demand
models. Energy Economics 1991;13:2–9.
[8] Hogan WW. A dynamic putty-semi-putty model of aggregate energy demand. Energy Economics 1989;11:53–69.
[9] Banerjee A, Dolado JJ, Hendry DF, Smith GW. Exploring equilibrium relationships in econometrics through static
models: some Monte-Carlo evidence. Oxford Bulletin of Economics and Statistics 1986;48:253–77.
[10] Banerjee A, Dolado JJ, Galbraith JW, Hendry DF. Cointegration, Error-Correction, and the Econometric Analysis
of Non-Stationary Data. Oxford: Oxford University Press, 1993.
[11] Bentzen J. A empirical analysis of gasoline demand in Denmark using cointegration techniques. Energy Economics
1994;16:139–43.
[12] Samimi R. Road transport energy demand in Australia: a cointegration approach. Energy Economics
1995;17:329–39.
[13] Chan HL, Lee SK. Modeling and forecasting the demand for coal in China. Energy Economics 1997;19:271–87.
[14] Fouquet R, Pearson A, Hawdon PD, Robinson C, Stevens P. The future of UK final user energy demand. Energy
Policy 1997;25:231–40.
[15] Ramanathan R. Short-run and long-run elasticities of gasoline demand in India: an empirical analysis using cointe-
gration technique. Energy Economics 1999;21:321–30.
[16] Engle RF, Granger CWJ. Cointegration and error-correction: representation, estimation and testing. Econometrica
1987;55:251–76.
[17] Stock JH, Watson MW. A simple estimator of cointegrated vectors in higher order integrated systems. Econo-
metrica 1993;61:783–820.
[18] Johansen S. Statistical analysis of cointegration vectors. Journal of Economic Dynamics and Control
1988;12:231–54.
[19] Johansen S. Estimation and hypothesis testing of cointegrated vectors in Gaussian vector autoregressive models.
Econometrica 1991;59:1551–80.
[20] Johansen S, Juselius K. The full information maximum likelihood procedure for inference on cointegration—with
application to the demand for money. Oxford Bulletin of Economics and Statistics 1990;52:169–210.
[21] Johansen S, Juselius K. Identification of the long-run and the short-run structure: an application to the ISLM
model. Journal of Econometrics 1994;63:7–36.
[22] Lee KC, Pesaran MH, Pierce RG. Persistence of shocks and its sources in a multisectoral model of UK output
growth. Economic Journal 1993;102:342–56.
[23] Pesaran MH, Shin Y. Cointegration and speed of convergence to equilibrium. Journal of Econometrics
1996;71:117–43.
[24] Pesaran MH, Shin Y. Generalized impulse response analysis in linear multivariate models. Economics Letters
1998;58:17–29.
[25] Sims CA. Macroeconomics and reality. Econometrica 1980;48:1–48.
[26] Sengupta RP. Perspective planning and policy for commercial energy, New Delhi: Planning Commission, Govern-
ment of India, 1989.
168 M. Kulshreshtha, J.K. Parikh / Energy 25 (2000) 149–168

[27] Parikh J. Modeling energy demand for policy analysis, New Delhi: Planning Commission, Government of
India, 1981.
[28] MEDEE. Sectoral energy demand in India, New Delhi: Planning Commission, Government of India, 1991.
[29] Murthy NS, Panda M, Parikh J. Economic development, poverty reduction and carbon emissions in India. Energy
Economics 1997;19:327–54.
[30] Stock JH. Asymptotic properties of least square estimation of cointegrating vectors. Econometrica
1987;55:1035–56.
[31] Stock JH, Watson MW. Testing for common trends. Journal of the American Statistical Association
1988;83:1097–107.
[32] Engle RF, Yoo BS. Cointegrated economic time series; an overview with new results. In: Engle RF, Granger CWJ,
editors. Long-run Economic Relationships: Readings in Cointegration. Oxford: Oxford University Press, 1991.
[33] Park JY. Canonical cointegrating regressions. Econometrica 1992;60:119–43.
[34] Gonzalo J. Five alternative methods of estimating long-run equilibrium relationships. Journal of Econometrics
1994;60:203–33.
[35] Inder B. Estimating long-run relationships in economics: a comparison of different approaches. Journal of Econo-
metrics 1993;57:53–68.
[36] Phillips PC, Hansen BE. Statistical inference in instrumental variables regression with I(1) processes. Review of
Economic Studies 1990;57:99–125.
[37] Urbain JP. On weak exogeneity in error correction models. Oxford Bulletin of Economics and Statistics
1992;54:187–207.
[38] White KJ, Wong SD, Whistler D, Haun SA. SHAZAM Users’ Reference Manual Version 7.0. New York:
McGraw–Hill, 1993.
[39] Hansen H, Juselius K. CATS in RATS: Cointegration Analysis of Time Series. Evanston, IL: Estima, 1994.
[40] Pesaran MH, Akiyama T, Smith RP. Energy Demand in Asian Developing Economies. Oxford: Oxford University
Press, 1998.
[41] Pesaran MH, Pesaran B. Working with Microfit 4.0: An Interactive Econometric Software Package. Oxford:
Oxford University Press, 1997.
[42] Canova F. Vector autoregressive models: specification, estimation, inference and forecasting. In: Pesaran M, Wick-
ens MR, editors. Handbook of Applied Econometrics. Oxford: Oxford University Press, 1995.
[43] Pesaran MH, Smith R. The role of theory in econometrics. Journal of Econometrics 1995;67:61–79.
[44] Pesaran MH. The role of economic theory in modeling the long-run. Economic Journal 1997;107:178–91.
[45] Lutkepohl H, Reimers HE. Impulse response analysis of cointegrated systems. Journal of Economic Dynamics
and Control 1992;16:53–78.
[46] MacKinnon JG. Critical values for cointegration tests. In: Engle RF, Granger CWJ, editors. Long-run Equilibrium
Relationships: Readings in Cointegration. Oxford: Oxford University Press, 1991.
[47] Pesaran MH, Shin Y, Smith RJ. Structural analysis of vector error correction models with exogenous I(1) variables.
DAE Working Paper No. 9706, Department of Applies Economics, University of Cambridge, 1996.
[48] Kulshreshtha M, Nag B. Some methodological comments on “Public investment and private capital formation in
a vector error correction model of growth” by K.H. Ghali. Applied Economics Letters (In press).
[49] CCO. Coal directory of India. Coal Controller’s Organization, Ministry of Coal, Government of India (several
issues).
[50] CEA. Public electricity supply: all India statistics: general review. Central Electricity Authority, Government of
India (several issues).
[51] CSO. National accounts statistics. Central Statistics Organization, Ministry of Planning and Programme Implemen-
tation, Government of India (several issues).
[52] GOI. Index of wholesale prices in India. Office of the economic advisor, Government of India (several issues).

You might also like