European Journal of Operational Research 220 (2012) 844852

Contents lists available at SciVerse ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor

Interfaces with Other Disciplines

Measuring productivity growth under factor non-substitution: An application to US

steam-electric power generation utilities
Margarita Genius a, Spiro E. Stefanou b,c,, Vangelis Tzouvelekas a
a

Department of Economics, University of Crete, Greece

Department of Agricultural Economics & Rural Sociology, Pennsylvania State University, PA, USA
c
Business Economics Group, Wageningen University, Netherlands
b

a r t i c l e

i n f o

Article history:
Received 23 January 2011
Accepted 16 February 2012
Available online 3 March 2012
Keywords:
Productivity and competitiveness
OR in energy
Productivity growth
Technical efciency
Factor non-substitution
Leontief-type technologies

a b s t r a c t
A theoretical framework is developed for decomposing partial factor productivity and measuring technical inefciency when the underlying technology is characterized by factor non-substitution. With
Farrells (1957) radial index of technical inefciency being inappropriate in this case, Russell non-radial
indices are adapted to measure technical inefciency in a Leontief-type model. A system of factor demand
equations with a regime specic technical inefciency term is proposed and estimated allowing for
dependence across inputs using a copula approach. Then the paper presents a complete decomposition
of partial factor productivity changes using a dataset of US steam-power electric generation utilities.
 2012 Elsevier B.V. All rights reserved.

1. Introduction
The decomposition of productivity growth has been explored
and measured extensively to include efciency changes over time
in addition to scale effects and technical change components (see
Fried et al., 2008, for a recent overview). This partitioning of the
different contributions is important since different incentives
inuence different components. For example, expansionary investment involves impacting the scale effect of the growth decomposition, while replacement investment acts on the technical change
effect. Decisions and incentives associated with learning to extract
the full potential of implemented technologies are acting on the
efciency change component of promoting growth.
The core theoretical concept for building these measures is the
production technology, where one can dene formally the notions
of technical efciency (operating on the boundary of the feasible
technology set), technical progress (shifting the boundary of this
set) and scale effects (moving along the boundary of an existing
set). The abundant economic literature on the estimation of stochastic production frontier functions and the subsequent measurement
of technical inefciency has assumed, in general, that the underlying production technology displays some degree of substitutability
Corresponding author. Address: 208B Armsby Building, Pennsylvania State
University, University Park, PA 16802, USA. Tel.: +1 814 86 8635.
E-mail address: spiros@psu.edu (S.E. Stefanou).
0377-2217/$ - see front matter  2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2012.02.023

between factors of production.1 This is not unusual as a production

technology with zero input elasticity of substitution would imply that
the cost-minimizing inputs are independent of their prices, which is a
restrictive assumption in many real world applications.
However, certain types of production activities may exhibit a
zero or limited elasticity of substitution among inputs. In some
sectors or industries technological conditions are characterized
by ex ante limited substitutability between inputs for a given level
of output but also by substantial economies of scale. A classical
example is the transportation sector (i.e., air-carriers, buses, railways), where one expects to nd rather limited substitution possibilities between cost driving inputs. Buses, airplanes or trains
cannot be operated without drivers or pilots, fuel is worth nothing
without planes, buses or trains, etc. Apart from the specicities of
the technology itself, it is also possible that these may, in practice,
never be substituted for each other in some sectors where outputs
or inputs are technologically substitutable, (Barnum and Gleason,
2011). This may arise when managers follow a common industry
practice to avoid risk or because local norms or regulations imply
the use of a decision heuristic.
1
Indeed, Bravo-Ureta et al. (2007) in their meta-regression analysis reviewing 167
empirical studies for measuring productive efciency in agricultural applications, in
both developed and developing countries, found that the vast majority of those hinge
either on a Cobb-Douglas or a Translog functional specication to approximate the
underlying production technology allowing for substitution possibilities among
factors of production.

M. Genius et al. / European Journal of Operational Research 220 (2012) 844852

The rst study questioning the classical substitution model was

Komiya (1962) who addressed economies of scale and technical
progress in the generation of steam electric power utilities in the
US with two alternative specications of the production technology; namely, the traditional CobbDouglas and a factor limitation
model.2 His study was inspired by the engineering literature which
made use of various decision heuristics for estimating the capital
cost of plants of different sizes.3 Komiya (1962) found that the substitution model was unsatisfactory and the limitation model provided a better t to the production data of steam-red electric
utilities. Several studies followed examining the substitution
hypothesis in several countries and different sectors. Lau and
Tamura (1972) proposed the use of a non-homothetic Leontief production function to analyze the Japanese petrochemical industry,
while Haldi and Whitcomb (1967) and Ozaki (1970) used a similar
approach based on a homothetic Leontief production function in
their analysis of economies of scale in US and Japanese steel industries, respectively. Nakamura (1990) utilized a non-homothetic generalized Leontief technological structure for empirically analyzing
the Japanese iron and steel industry. Buccola and Sil (1996) measured productivity in the agricultural marketing sector and recently
Holvad et al. (2004) assumed a Leontief-type technology for the Norwegian bus industry. Furthermore, a stream of the literature in agricultural economics addressing the modeling of crop response to
different fertilizers nutrients levels, has maintained zero substitution among crop nutrients using a linear plateau specication motivated by the von Liebig farm technology.4
Measuring technical inefciency in the case of Leontief-type
technologies is of interest given that Farrells (1957) radial measures are the basis for most applied work on the measurement of
efciency. However, radial measures can be inadequate under such
technologies as they may classify inefcient input combinations as
being efcient, while input and output measures might not coincide even under constant returns-to-scale.5 Fre and Lovell
(1978), Russell (1985, 1987), Portela and Thanassoulis (2005) and
Cherchye and Van Puyenbroek (2009) focus on the non-radial technical efciency measures that provide a price rationale to evaluate a
mixed efciency factor. Once the technology is governed by a Leontief-type structure, it is plausible to have inefciency displayed by
none, all or a subset of the inputs, rendering radial measures unsatisfactory. In addition, output measures may fail to recognize inefciencies when they affect a subset of the inputs only.
This paper develops a framework for modeling productivity
growth under factor non-substitution that accounts for technical
inefciency and technical progress. The econometric modeling
framework accommodates the absence of substitution possibilities
among inputs where inefciency between factors can be correlated. This Leontief frontier model adapts the copula approach to
modeling the joint distribution between the one-sided error terms
that capture factor-specic technical inefciencies. Factor-specic
technical efciencies are specied and measured using Kopps
(1981) orthogonal indices of technical efciency, combined into
an overall technical efciency measure using Russells non-radial
2
Komiya (1962) utilized an older dataset than ours of US steam electric generation
utilities.
3
The economic-engineering rationale for the Leontief technology specication is
that changes in the rate of utilization impacts the thermal efciency of a steam-power
plant after it has started operation. Otherwise a plants input requirements change
little once the plant has been built.
4
Paris (1992) presents an excellent overview of the historical literature and an
econometric estimation approach, and Holloway and Paris (2002) address productive
efciency in the context of the von Liebig specication. Guan et al. (2006) introduce
an alternative modeling framework when encountering limited substitution
grounded in an agronomic model of nutrient exchange.
5
Fre and Lovell (1978) proved that if a regular production technology is linear
homogeneous then input technical efciency coincides with output technical
efciency. However, this is not true in Leontief-type production technologies.

845

index of productive efciency. Then, we proceed to developing a

tractable approach for the analysis of partial factor productivity
growth. An application to a panel data set of 72 fossil-fuel red
steam electric power generation utilities in the US observed during
the 19861996 period follows.
The next section develops the theoretical framework for measuring technical efciency in production structures that exhibit
zero elasticity of substitution among inputs, while Section 3 presents the empirical model discussing briey the econometric
methods used. Section 4 presents the estimation results of an
application to US electric utilities and nally, Section 5 provides
some concluding remarks and suggestions for future extensions.
2. Theoretical framework
Assume that producers in period t utilize a vector of variable inputs x 2 RJ together with a vector of quasi-xed inputs z 2 RK to
produce a single output y 2 R through a technology described by
the following closed, nonempty production possibilities set
Tt fx; z; y : x 2 RJ ; z 2 RK can produce y at year tg. Accordingly, for every y 2 R we can dene the input correspondence
set as all the input combinations capable of producing y, i.e.,

Ly; z; t fx 2 RJ : x; z; y 2 Ttg

If we assume that the above dened production technology is characterized by ex ante limited substitutability between factors of production, we can dene the cost function for all y such that
L(y, z, t) :

Cy; w; z; t minfw0 x : x 2 Ly; z; tg

x

which is the minimum cost of producing output quantities y with

periods t technology, when the factor prices w 2 RK are strictly
positive. Applying Shephards lemma to (2) we obtain the system
of derived demand equations as:

@Cy; w; z; t
g j y; z; t xj :
@wj

Since the elasticity of substitution between any pair of factors of

production, holding output constant, is assumed to be zero, the derived demand functions are independent of factor prices. Such a
system is used by Komiya (1962) who refers to it as plant base
factor limitational production function and by Haldi and Whitcomb
(1967), Ozaki (1970) and Lau and Tamura (1972). The function
gj(  ) is a positive real-valued convex function dened and nite
for all y > 0 with gj(0, z, t) = 0.
The production function f x; z; t : RJ ! R corresponding to
the dual cost function dened in (2) is given by

y maxfy : w0 x P Cy; w; z; tg
y

which means that, for any given set of factor prices, the maximum y
is obtained such that the observed cost of production is greater than
or equal to the optimum factor cost. The solution of the above optimization problem requires xj P g j y; z; t "j.6 Assuming that gj(  ) is
non-decreasing and lower semi-continuous in y, we may dene its
generalized inverse, and hence the production function may be
reformulated as

y maxfy : xj P g j y; z; t8jg g 1
j xj ; z; t
y

The maximum y satisfying the above optimization problem is then

given by

y minfg 1
j xj ; z; tg
j

See Lau and Tamura (1972) pp. 117172.

846

M. Genius et al. / European Journal of Operational Research 220 (2012) 844852

Fuel

F1

F*

L*

L2

L1

Labor

Fig. 1. Farrell and Russell measures of input technical inefciency under factor nonsubstitution.

which is a non-homothetic Leontief production function corresponding to the dual cost function dened in (2). It is non-homothetic as
the expansion path is not necessarily a ray through the origin and
the elasticities of substitution are zero between any pair of factors
of production. Given (6), the input requirement set for this nonhomothetic technology structure may be restated as

Ly; z; t fx : min g 1
j xj ; z; t P y; 8jg
xj

where L(y, z, t) satises the correspondence RJ ! R. In addition to

the production function and the input correspondence set the following two subsets are important: (a) the isoquant and, (b) the
technically efcient subset. In the case of the non-homothetic Leontief technology both sets are dened, respectively, as

Q y; z; t fx : x 2 Ly; z; t; xk P g k y; z; t ^ xj
g j y; z; t; 8k; jk jg

and

Ey; z; t fx : x 2 Ly; z; t; xj g j y; z; t; 8jg

Unlike conventional technologies where substitution possibilities

among factors of production exist, the efcient subset of the input
correspondence is a subset of the isoquant for each output level
y.7 Actually, the efcient subset coincides with the right angle point
of the Leontief-type technology isoquants (i.e., L-shaped). In these
instances, where the production technology exhibits L-shaped isoquants, technical efciency coincides with productive efciency as dened by Farrell (1957) since allocative efciency is always
maintained (i.e., the cost-minimizing input bundle is always on the
left angles of the isoquants).8
In the case of the non-homothetic Leontief production technology, Farrells (1957) measure could well classify an inefcient
input bundle as being efcient since its a radial measure that constrains the input contraction to be the same across inputs. In contrast, Russells non-radial measure of technical efciency,
suggested by Fre and Lovell (1978), allowing for different inputs
to display different reduction levels is suitable for technologies
that exhibit non-substitution among factors of production.9 Fig. 1
7
The variable elasticity of substitution (VES) and weak input disposability
functions are also examples of production functions whose isoquants are not
contained in their efcient subsets (Fre and Lovell, 1978).
8
However, this presumes that any change in factor prices does not affect the xed
proportion in which inputs are combined in the production process.
9
The same problem with the Farrell index occurs when the approximated
technology exhibits input slacks. In this case, again, the Russell index can serve as
an alternative in measuring correctly input technical efciency (Fre et al., 2008).

illustrates the non-substitution between two inputs (e.g., fuel and labor) and a production function given by (6), where the production
 using an input combinaunit is producing a given level of output y
tion dened by point A, with L1 units of labor and F1 units of fuel. The
same level of output can be produced by reducing the use of both inputs until point B which lies on the isoquant associated with the
. Farrells denition
minimum level of inputs required to produce y
of a radial measure of input-oriented technical inefciency is the ratio 0C/0A. In this case both input contractions are the same, i.e., 0L2/
0L1 = 0F/0F1. However, point C is not the minimum level of inputs
, as labor is used in excessive quantities. This
required to produce y
 but it does not lie in the efcient set of inpoint is on the isoquant y
puts. Therefore technical inefciency is due to the excess use of the
labor input only. If we decrease its use until point B and leave the
.
fuel input constant we produce the same output y
On the other hand, Russell non-radial index can appropriately
measure technical inefciency of this type of production technology. Using the input correspondence dened in (7), the Russell
non-radial technical inefciency index can be dened as

8
9
P1
>
>
nj
<Y
=
j
TE min
hj
: h1 x1 ; . . . hj xj 2 Ly; z; t ^ hj 2 0; 1; 8j
hj >
>
: j
;
R

10
where G denotes the product over j, nj = 1 if xj > 0 and nj = 0 if xj = 0.
The index in (10) is the ratio of two distances computed along
diverging rays. The Russell measure clearly generalizes the Farrell
input measure of technical efciency, with the latter being the special case for hj 
h8j. Fig. 1 illustrates how inputs F and L are contracted by different proportions to reach the technical efcient
input mix to reach the efcient point B.
In this case, technical inefciency should be measured non-radially and is dened as the distance DB/DA which is different from the
Farrell (radial) measure of 0C/0A.Labor needs to be reduced by 0L/
0L1, while fuel needs to be reduced by 0F/0F1 and 0L/0L1 0F/0F1.
Given the nature of the underlying production technology, Russells
measure is actually the simple geometric average of the orthogonal
non-radial factor-specic technical efciency indices suggested by
Kopp (1981).10 Formally, they are dened as:
1

TEKP
j minfhj : hj > 0; ming j hj xj ; z; t P yg
hj

11

or using (3) under technical inefciency as:

TEKP
j

g j y; z; t
xj

12

where hj e (0, 1] is the orthogonal factor-specic measure of technical efciency. The factor-specic technical efciency dened in (11),
(12) has an input-conserving interpretation, which however, cannot
be converted into a cost saving measure due to its non-radial nature. From the above index of factor-specic technical efciency,
we may derive a complete decomposition formula for partial factor
productivity growth. The partial factor productivity growth approach is appropriate when dealing with a production system
where signicant capital structures are involved and this system
is at a long-run equilibrium.11
First, following Reifschneider and Stevenson (1991) and Battese
and Coelli (1995) inefciency effects model, we may assume that
10
Instead of the simple average, Russells technical efciency index can be obtained
using an unweighted geometric mean.
11
When estimating a system allowing for dynamic adjustment that is manifested as

a linear accelerator, optimal net investment is dened as dz
mz  z, where z is
dt
the long-run optimal capital stock (that necessarily depends on arguments taken as
xed such as prices), z is the current capital stock and m is the adjustment rate. When
dz
mz  z I  dz and m d, then I dz.
dt

M. Genius et al. / European Journal of Operational Research 220 (2012) 844852

factor-specic technical inefciency dened in (12), is affected by

the utilization of the available capacity by individual rms as well
as by time (i.e., autonomous changes due to learning-by-doing effects). Then, taking logarithms and totally differentiating with respect to time relation (12) we obtain12

T E_ KP
j z; t

ln TEKP
j

X@
k

@ ln zk

 z_ k

@ ln g j y; z; t
@ ln g j y; z; t
 z_ k
@ ln zk
@ ln y
 y_

@ ln g j y; z; t
 x_ j
@t

13

3. Econometric model
Building on the previous section, we can rewrite (12) by taking
logarithms and rearranging terms as

ln xj ln g j y; z; t  ln TEKP
j z; t 8j

PFP j T E_ KP
j z; t

X @ ln TEKP
j
k

@ ln zk

 z_ k 

@ ln g j y; z; t
C
 z_ k  C tj 1  ej y  y_
@ ln zk
14

where the rst two terms constitute the technical efciency changes,
which contribute positively (negatively) to PFP growth as long as efciency changes are associated with movements towards (away from)
the production frontier. These changes may be due to two factors: (a)
the passage of time, i.e., autonomous changes (rst term) and, (b)
changes in the quantity of quasi-xed inputs (second term). If the
passage of time does not affect technical efciency levels or if the level of quasi-xed input remains constant both terms are equal to
zero. The third term incorporates the sub-equilibrium effects associated with the existence of quasi-xed inputs (Luh and Stefanou,
1991; Morrison, 1992). If the market price of quasi-xed inputs coincides with their shadow price then the third term vanishes. In any
other case it is positive under capacity over- (under-) utilization as
long as the stock of capital increases (decreases) over time. The
fourth term is the factor-specic dual rate of technical change which
is positive (negative) under progressive (regressive) technical
change.14 The nal term is the scale effect where the sign depends
on both the magnitude of the scale elasticity and the changes of the
aggregate output over time. In the context of the non-homothetic
Leontief production function adopted in this study, the degree of returns to scale can be different for each variable factor of production.15
It is positive (negative) under increasing (decreasing) returns to scale
as long as output produced increases. This term vanishes when either
the technology is characterized by constant returns to scale or the
aggregate output quantity remains unchanged over time.
12
This formulation implicitly assumes a deterministic frontier. We have adopted
this formulation in order for our results to be directly comparable with those of Bauer
(1990) and Lovell (1996). However, in implementing the proposed model empirically,
it is necessary to take into account the stochastic nature of output and to make
additional distributional assumptions to obtain estimates of TEKP
j z; t. Without loss of
generality, these elements are added into the model in the next sections, where
specic functional forms for g(  ) as well as the mean of TEKP
j z; t are imposed.
13
As noted by Silberberg (1972, p. 944) and Hulten, 1973, pp. 102122 the use of
partial productivity Divisia indices are path dependent when the production
technology is non-homothetic. This will create inconsistent aggregate measures of
total factor productivity. However, in our case we focus on partial factor productivity
measures without aggregating over all inputs and therefore the use of partial
productivity Divisia indices is free of aggregation errors as long as our analysis is
restricted to partial productivity measures (see Diewert and Nakamura, 1993, pp.
217218).
14
One of the properties of the non-homothetic Leontief production function refers
that the optimal relative factor intensities may vary across rms if the output levels
differ even in the case of Hicks-neutral technical change and in the absence of price
changes.
15
The degree of returns-to-scale could further vary even for the same input
depending on the choice of g(  ).

15

Substituting TEKP
j z; t with hj and assuming an additive two-sided
error term in each equation, vj, capturing unobserved random factors affecting input demands (e.g., exogenous shocks, measurement
errors), the econometric model is then given by

ln xj ln g j y; z; t v j  ln hj 8j
where a dot over a function or variable indicates its time rate of
change. Substituting into (13) the conventional Divisia index of partial factor productivity growth, i.e., PFPj y_  x_ j we obtain13

847

16

where ej = vj - ln hj is the familiar composite error term presented in

the stochastic frontier literature.
An important issue in the above system is whether dependencies exist between the composite error terms of the different equations. Lau and Tamura (1972) specify only the two-sided error
term, i.e., vj, which was appended to the system of derived demand
equation in (16) to capture exogenous shocks and measurement
errors. Hence, their model was a simple SURE system which,
depending on the choice of functional forms for the factor demand
equations, can be estimated even with a simple OLS. However,
when the presence of technical inefciency is allowed, the onesided error term is introduced into (16) which complicates the
econometric estimation of the model. Given that technical efciency arises from a producers know-how and his/her ability to
organize the production process effectively, it is reasonable to assume that dependencies may exist between the one-sided error
terms of the different equations. If a producer fails to use efciently
the labor input due to lack of knowledge and/or entrepreneurial
ability the same should happen with capital. Furthermore, dependence can also arise between the two-sided error terms simply because contemporaneous random shocks to different inputs for rm
i are correlated or dependent. This study does not distinguish between these two cases, allowing for dependence between the overall composite error terms.16
Allowing for dependencies requires the specication of a joint
distribution and it is not obvious which joint distribution one
should specify given the structure of the composite error term.
Researchers estimating stochastic frontier models are interested
in imposing different distributional assumptions for the efciency
term. While the two-sided error term is assumed to be normal it is
difcult to nd and justify any given joint distribution for the composite error terms.
A useful direction to address this concern is the copula approach
to modeling joint distributions.17 A copula is a multivariate distribution with uniform margins or it can also be dened as a function
that links a multidimensional distribution to its univariate margins.
In effect, by using a set of given marginal distribution functions as its
arguments it produces a joint distribution whose margins will coincide with the above and with a given dependence structure. Indeed,
in a situation where the marginal distributions, Fj of the J variates xj
are known to the researcher, Sklars theorem (Sklar, 1959) establishes: ifC:[0, 1]J ? [0, 1] is a copula function then the function
H(x1, . . ., xj) = C(F1(x1), . . ., FJ(xj); q) is a well dened joint distribution
function with margins given by Fj.
16
Our model can be extended in a straightforward manner to allow for different
types of dependencies between the composite error terms. However, this is not the
scope of the present study.
17
Copulas have been applied especially in the eld of nance, where normality
could be an untenable assumption when modeling asset returns and asymmetries in
the dependence structure of different returns or markets exclude the application of a
multivariate normal distribution. For an excellent survey on copulas and their
applications the reader is referred to Trivedi and Zimmer (2007) while a mathematical treatment of copulas can be found in Nelsen (1999).

848

M. Genius et al. / European Journal of Operational Research 220 (2012) 844852

The advantage of copulas is that they allow the modeling of the

marginal distributions separately from that of the dependence
structure, making them especially useful in situations where a researcher has some knowledge about the marginal distributions but
needs to specify their joint distribution, as in the case posed by the
system of derived demand equations above.
4. The case of US electric utility rms
The application is to a panel of fossil fuel-red electric utility
power generating rms in the United States over the period
19861996. The production technology is represented by one output and three inputs (fuels, the aggregate of labor and maintenance, and capital stocks). Fuels and the aggregate of labor and
maintenance are specied as variable inputs whereas the capital
stocks are treated as an aggregate quasi-xed input in the production. Variables used in the estimation consist of output, prices and
quantities of fuels, the aggregate of labor and maintenance, and
capital stock.
Electricity deregulation and restructuring emerged on the policy agenda in many states in the USA in the early 1990s. Under regulation, electric utilities receive a guaranteed prot for the
generation of electricity, with a customer paying one regulated
price for electricity to a single vertically integrated utility responsible for generation, transmission, distribution, and marketing. National energy policy initiatives forced utilities with transmission
networks to deliver power to third parties at nondiscriminatory
cost-based rates. To date, deregulation in the electricity markets
has been incomplete with continued regulation in some of its segments and it should give rms the incentives to lower costs by
improving technical and input allocative efciency to maximize
their prots. Under the regulatory era, electric utilities had a guaranteed prot for the generation of electricity leading to strong
incentives to overinvest in capital as well as operating at an inefcient level of production. Since an underlying goal of the electricity
deregulation plan is to improve the efciency level of electric utilities, the core issues associated with evaluating the deregulation of
electricity generation surround measuring the potential saving in
production costs due to the deregulation of this stage across the
country, and the contribution of efciency gains (Athanassopoulos
et al., 1999; Granderson and Linvill (2002)).
4.1. Model specication
We assume that the derived factor demand equations have the
following general form under factor-specic technical inefciency
(Lau and Tamura, 1972)18

ln xfit bf0 bfy ln yit bfd Dt bfdd Dt2 bfz ln zit  ln hfit v fit
ln xlit bl0 bly ln yit bld Dt bldd Dt2 blz ln zit  ln hlit v lit

17

where subscripts i = 1, . . ., N and t = 1, . . ., T correspond to rms and

time, respectively; superscripts f, l are the input indices for fuel and
labor; y is the volume of output produced; z is the quasi-xed input;
Dt is a simple time index capturing technical change; xf and xl are
the levels of fuel and labor use; uj = ln(hj), j = f, l are the one-sided
error terms capturing factor-specic technical inefciency; and, vs
are the two-sided error terms. Given the above specication, factorspecic returns-to-scale are determined by the magnitude of the
parameter bjy . Specically, if bjy > 1 input j exhibits decreasing returns to scale; if bjy 1 constant returns to scale and; if bjy < 1
increasing returns to scale.
18
Note that in the case of the non-homothetic Leontief production technology the
functional specication of the derived demand equations in (18) may differ across
factor of production. For simplicity we keep the same functional specication herein.

The error terms are assumed to satisfy the following: (a) for
each j = f, l, v jit is assumed to be independently and identically distributed according to a normal distribution with mean zero and
unknown variance r2v j ; (b) for the technical inefciency terms
hjit expujit , it is assumed that ujit are independently distributed
according to a normal distribution with mean ljit and unknown
variance r2uj truncated at zero so that ujit is non-negative; and, (c)
0
ujit is independent from v ji0 t0 , as it is traditionally done in the stochastic frontier literature, j, j = f, l, "t, t0 = 1, . . . , T. The above structure of the inefciency random term is related to that suggested by
Reifschneider and Stevenson (1991) and Battese and Coelli (1995).
The composed error term for each equation is given by
ejit ujit v jit and its density function, which can be derived in a
straightforward manner from Battese and Coelli (1995), by noting
that ujit enters additively in our case, is given by,

0
10 0
111


j
j
j
c
e

1

c

l
1
el
CB B l CC
j
j
fej e u
UB
@ q A@U@qAA
rj
rj
cj 1  cj r2j
cj r2j
where

r2j r2v j r2uj and cj

18

r2Uj
.
r2j

The functional form of the pre-truncation mean of the efciency

terms and that of the copula functions need to be specied in order
to fully specify the log-likelihood function. Specically, we allow
the pre-truncation mean of each efciency term to be time varying
through the use of a second degree polynomial in time,19 to depend
on the level of the quasi-xed factor, capital, and on a dummy indicating deregulation

ljit dj0 djd Dt djdd Dt2 djD DUMi djz ln zit

19

where DUMi is a dummy reecting whether the utility is located in a

state that has some deregulation plan. We investigate three different copulas, namely the Gaussian, Clayton and Gumbel which display
a disparity of dependence structures. The Gaussian copula is dened
by CG(u, v) = U2(U-1(u), U1(v); q), where U2 is the standard bivariate normal, U1 is the inverse of the standard univariate normal
and the parameter q is the correlation coefcient. This copula
exhibits symmetry in its dependence and allows only linear dependence, therefore it will not provide a good t in cases where the
type of dependence is different from correlation. The Clayton copula
1
is dened as C C u; v uq v q  1q ; q 2 0; 1 which exhibits
asymmetric dependence with a clustering of values in the left
tail and it can best t data displaying higher degree of dependence in the lower left quadrant than in the upper right quadrant.
 
The Gumbel copula is dened as C B u; v exp  ln uq
1

 ln uq q q 2 1; 1 which exhibits asymmetric dependence with

a clustering of values in the right tail and therefore displays higher
dependence in the upper right quadrant than in the lower left quadrant. The Independence copula given by C(u, v) = uv can be obtained
from the Gaussian and Gumbel copula for q = 0 and q = 1, respectively, and as the limit of the Clayton copula when q tends to zero.
Given our distributional assumptions and denoting the probability densities for the fuel and labor composite error terms as fitf
and fitl , respectively, and their respective cumulative distributions
as by F fit and F lit , it is straightforward to write the log likelihood
for a given copula function as20

19
As noted by Karagiannis et al. (2002), in this stochastic framework, the
autonomous changes in inefciency can be isolated from those of technical change.
20
Note that we assume that the dependence structure remains the same across i
and t so that the copula function is not indexed by i nor by t. It is possible to model the
dependence parameters in the copula function in such a way that they show variation
across time and rms but we will assume that they are constant.

849

M. Genius et al. / European Journal of Operational Research 220 (2012) 844852

i X X 

XX h  f f
LnB
ln ck F it eit ; F lit elit
ln fitf efit
i

lnfitl elit
where ck F 1 ; F 2 @

Table 1
Parameter estimates of factor demand equations for US electric utilities.

20

2 k

C F 1 ;F 2
@F 1 @F 2

and k = G, C, B for the three alternative

copula specications discussed previously.

After estimating the underlying system of derived demand
equations, the estimates of the different components integrating
the productivity decomposition in (14) can be derived. For instance, the dual and the primal rates of technical change are related to each other as follows21

C tj bjd 2bjdd Dt

21

The hypothesis of zero technical change can be tested by imposing

the following restriction: bjd bjdd 08j. If the hypothesis of zero
technical change cannot be rejected, the fourth term in (14) is zero,
and technical change has no impact on productivity changes. In
addition, Hicks-neutral technical change (i.e., passage of time
affects equally both variable inputs) can be statistically tested by
imposing the restriction that bfd bld ^ bfdd bldd .22 Then, factor-specic returns to scale are given by

eCy
j

@ ln xjit
bjy
@ ln yit

Parameter

b0
by
bz
bd
bdd
d0
dd
ddd
dD
dz

r
c
q
Ln(h)

Fuel Input

Labor Input

Estimate

S.E.

Estimate

S.E

6.6499
0.8036
0.3668
0.0246
0.0012
3.7351
0.0082
0.0020
0.2969
0.2828
0.3121
0.6712
1.0663
412.391

(0.4572)*
(0.0282)*
(0.0436)*
(0.0124)**
(0.0019)
(0.5413)*
(0.0044)**
(0.0034)
(0.0555)*
(0.0388)*
(0.0184)*
(0.0650)**
(0.0258)*

6.8054
0.5476
0.4026
0.0429
0.0051
3.2524
0.1961
0.0104
0.8378
0.2797
0.6178
0.7521

(0.4230)*
(0.0434)*
(0.0540)*
(0.0216)**
(0.0027)**
(1.8432)**
(0.1043)**
(0.0112)
(0.4102)**
(0.1788)
(0.1338)*
(0.1037)**

Where y stands for output, z for capital, d for time and D for the regulation dummy.
The critical value for a 5% signicance level is equal to 2.71 and is obtained from
Kodde and Palm (1986, Table 1).
*
Statistical signicance at the 1% level.
**
Statistical signicance at the 5% level.

The LR test for H0: q = 1 versus q > 1 is 9.05 and it follows a mixed chi-squared
distribution.

22

The hypothesis of constant returns to scale is examined imposing

the restriction that bjy 18j. If it cannot be rejected the nal term
in (14) vanishes. Finally, sufcient and necessary condition for
homotheticity of the production structure implies the restriction
that bfy bly .
Next, given the conditional density of ujit and the conditional
mean of the inefciency terms the components of the technical
efciency changes effect in (14) are computed as

@ ln TEjit @TEjit 1
1
djd 2djdd Dt njit j

@t
@t TEjit
TEit

23

For the calculation of the sub-equilibrium effects in (14) we

need the shadow value of the quasi-xed input. The price of z at
long-run equilibrium is

C z c
,
rd

where C z @Cy;w;z;t
, c is the user cost
@z

of capital, r is the discount rate, and d is the depreciation rate of

z. The long-run optimal value of z is determined by the rst order
z c
condition: kw; y; z; tF z  Crd
, where kw; y; z; t is the short-run
marginal product and
F
(

)
is
the marginal physical product of z.
 z

z c
This leads to z h Crd
; y; w; t , which suggests that the estimation of variable input factors is linked to the determination of z
via the components of Cz. However, in the Leontief framework, input demands are independent of input prices.

and
4.2. Data and estimation

@ ln TEjit
@TEjit 1
1

djz njit j
@ ln zit
@ ln zit TEj
TE
it
it

24

where

njit



r2v j
1 2
j
~
~
Ukjit 1
exp

l

r
it
2 j
r2j
"

 Ugjit

#
ukjit
1
ugjit  Ugjit

r~ j
Ukj
it

and

kjit

l~ jit j l~ jit ~ ~ j r2v j ljit r2uj ejit ~ 2 r2uj r2v j

; g
 rj ; lit
; rj
:
r~ j it r~ j
r2j
r2j

The hypothesis that factor-specic technical inefciency is time

invariant is tested by imposing the restrictions that djd djdd 08j.
Similarly the hypothesis that the autonomous rate of change in
technical inefciency is common across factors of production implies the restrictions dfd dld ^ dfdd dldd to the model in (19).
21
According to Frsund (1996) and Atkinson and Cornwell (1998), the rate of
technical change should be evaluated at the frontier and therefore the marginal effect
of time in the one-sided error term is not included in (21). Further, Lindebo et al.
(2007) analysing a decomposition of the economic capacity into technical and
economic factors, thus highlighting the factors that need attention in any management scheme that aims to improve efciency.
22
Conventional LR-test can be used to statistically examine the aforementioned
hypothesis.

The data are those employed in Rungsuriyawiboon and

Stefanou (2007) and their construction are described in greater detail therein. As an overview, the output variable is represented by
net steam electric power generation in megawatt-hour, which is
dened as the amount of power produced using fossil-fuel red
boilers to produce steam for turbine generators during a given
period of time. The nal data set is a balanced panel of 72 electric
utilities for the years 19861999. Among these electric utilities,
there are 45 electric utilities having all plants located in states
within deregulation acts and 27 electric utilities having all plants
located in states without the deregulation acts.23
The log-likelihood is maximized under the three different copula scenarios but the estimation results for the Gumbel copula only
are presented in Table 1. The choice of the Gumbel copula is based
on the application of the Akaike Information Criteria which favors
it and of two model selection tests for non-nested models based on
the KLIC: the Vuong (1989) test which has an asymptotic normal
distribution and the recent distribution-free test of Clarke (2007)
which is based on the binomial distribution with parameters
23
Among the twenty-seven electric utilities located in states without deregulation
plan, seven electric utilities, i.e., Empire District Electric, Interstate Power, Kentucky
Utilities, Union Electric, UtiliCorp United, Wisconsin Power and Light, and Wisconsin
Public Service served states that passed deregulation acts according to the Energy
Information Administration (1996). Financial Statistics of US Major US IOUs (1996).
However, the data used for these utilities was utility data in that state without
deregulation acts only.

850

M. Genius et al. / European Journal of Operational Research 220 (2012) 844852

Table 2
Model specication tests.
LR-statistic

Critical value (a = 0.05)

Technical efciency:
Technical efciency (i.e., cj = 0 "j)

33.43

Technical efciency in fuel input (i.e., cf = 0)

21.55

v22 5:14
v21 2:71
v21 2:71
v24 9:48
v22 5:99

Hypotheses

Technical efciency in labor input (i.e., cl = 0)

19.62

Time invariant inefciency i.e., djd djdd 08j

16.72

Neutral time varying inefciency i.e., dfd dld ^ dfdd dldd

15.68

Structure of production:
36.72

Homotheticity i.e.,bfy bly

CRTS i.e.,

bjy

v21 3:84
v22 5:99
v21 3:84
v21 3:84

54.98

18j

CRTS in fuel input i.e., bfy 1

28.72

CRTS in labor input i.e., bly 1

23.54

Technical change:
15.62

Hicks neutral TC i.e., bfd bld ^ bfdd bldd

Zero TC i.e.,

bjd

bjdd

v22 5:99
v24 9:48
v22 5:99
v22 5:99

19.74

08j

Zero TC in fuel input i.e., bfd bfdd 0

6.02

Zero TC in fuel input i.e., bld bldd 0

12.35

Regulation:
12.34

Absence of regulation effect i.e., djD 0 8j

v22 5:99
v21 3:84

9.41

Neutral regulation effect i.e., dfD dlD

Note: When the null hypothesis involves the restriction of c = 0 (rst three hypotheses) then the LR-test statistic follows a mixed chi-squared distribution, the critical values
of which are obtained from Kodde and Palm (1986, Table 1). These rst three critical values are for the Wald statistic of the same null hypothesis, where the likelihood ratio is
less than the Wald statistic. If likelihood ratio exceeds the critical value of Wald statistic, then so does the Wald test. Consequently, we still reject the null in the rst three
hypotheses.

85

Technical Efficiency (%)

Labour
80

Russell TE Index
75

Fuel
70

65

60
1986

1987

1988

1989

1990

1991

1992

1993

1994

1995

1996

Years
Fig. 2. Time development of technical efciency estimates for US electric utilities.

N = 792 and p = 0.5. When comparing (a) the Gumbel versus the
Gaussian and (b) the Gumbel versus the Clayton, the Vuong test statistic is equal to 1.68 with a p-value of 0.09 for case (a) and 1.61
with p-value 0.11 for case (b). The Clarke test statistic is 442 with
p-value 0.0005 for case (a) and 429 with p-value 0.009 for case (b).
Therefore, both Clarkes tests indicate that the Gumbel model is
more appropriate than the two competing choices while the Vuong
test, which is known to have low power, does not lead to a strong
rejection that the Gumbel and the two competing models are
equivalent. Turning to estimation results in Table 1, we note that
although dependence is not very strong (the estimate of q is
1.06), the likelihood ratio test rejects the null hypothesis of independence at the 5% level. The coefcient estimates are statistically
signicant for both variable input demand equations, except for

the second-order term for time. The presence of the regulatory

dummy leads to declining mean fuel and labor input efciency levr2

els. Similarly, cj rUj2 reects importance of the one-sided error

j
which conrms the presence of inefciency in the use of both fuel
and labor.
A set of hypotheses concerning the presence of technical
inefciency, the production structure and the characterization of
technical change are evaluated in Table 2. The null hypotheses concerning overall technical efciency, technical efciency for fuel and
then technical efciency for labor are all rejected soundly. Further,
technical efciency is found to be non-neutral and time varying.
The assumption of homothetic production is rejected as well as
the presence of constant returns to scale over all inputs and for

M. Genius et al. / European Journal of Operational Research 220 (2012) 844852

Table 3
Partial factor productivity growth and decomposition results for US electric utilities
(as percentages).
19861996

19861991

19921996

Fuel input
PFP growth
Technical change
Scale effect
Capacity utilization
TE change
Autonomous
Capital

0.4233
0.1560
0.2187
0.1382
0.1868
0.0802
0.1065

0.5090
0.1625
0.4929
0.1686
0.0221
0.0009
0.0230

0.3376
0.1495
0.0555
0.1078
0.3514
0.1613
0.1901

Labor input
PFP growth
Technical change
Scale effect
Capacity utilization
TE change
Autonomous
Capital

1.1115
0.0731
0.5288
0.1517
0.8075
0.7021
0.1054

2.1336
0.0735
1.1355
0.2549
1.3265
1.3233
0.0033

0.0894
0.0727
0.0778
0.0485
0.2884
0.0809
0.2075

the fuel and labor inputs separately. When testing for technical
change, we nd that Hicks neutrality is rejected but perceptible
technical change is present jointly and separately for the variable
inputs. The presence of a regulation effect is not rejected and this
regulation effect has a positive but differential impact on the
variable inputs. When evaluating the mean percentage change in
variable input use given the presence of a regulation effect, we nd
that fuel use increases by 15.22% while labor use increases much
slower at 3.5% reecting the relative importance of the fuel input
in terms of cost share. The results are discussed in the context of
two distinct periods: 19861991 and 19921996 in an attempt
to pick up a deregulation anticipation effect on the part of rms
to assess if their production decisions reect this potential change.
4.3. Technical efciency
Fig. 2 presents a frequency distribution of technical efciency
measures for both the Kopp (single factor) and Russell technical
efciency measures. For the last nine years of the sample period,
the Russell measure is bracketed by the factor-specic measures
and the trajectory of efciency levels indicates a gradual increase
in all efciency measures over the period. In comparison to recent
studies addressing technical efciency for panels of US electric utilities with non-Leontief specications, Knittel (2002) nds technical
efciency for the CobbDouglas specication for coal- and gas-red
plants to average 80% and 94%, respectively, with Hiebert (2002), in
contrast, nding fairly high average technical efciency of 87.9%
and 80.5% for coal- and gas-red plants, respectively, using the
more exible translog specication. Atkinson and Primont (2002)
employ a panel of privately-owned electric utilities engaged in
steam electric generation for the period 19611997. Both dual
and distance functions are estimated with a exible functional form
specication of the non-Leontief variety with an average technical
efciency levels of 0.7154 and 0.6675 using the cost and distance
functions, respectively. Granderson and Linvill (1997) nd a sample
of US electric utilities between 1977 and 9187 present technical
efciency score ranging from 0.55 to 0.87 using the data envelope
approach. In contrast, Rungsuriyawiboon and Stefanou (2007) nd
the technical efciency scores of variable inputs average of 76.7%
using a dynamic adjustment specication with a exible functional
form specication. The Russell aggregate TE index estimated here
averages higher generally than these studies at 90.3%.
4.4. Productivity growth
Rungsuriyawiboon and Stefanou (2007) estimate efciency under dynamic adjustment for these electric utility rms and nd

851

that the estimated capital adjustment rate is nearly equal to the

depreciate rate (3%).24 Since this industry is at a long-run equilibrium position and a Leontief technology is maintained, the shadow
value of capital is constant and a proxy for the optimal ^z is generated
by regressing gross investment against the arguments (y, z, t).25
Table 3 reports the partial productivity growth measures.26 Fuel
productivity averages 0.42% with the earlier period growing marginally faster than the later period. The contribution of technical change
accounts for more than a third of this growth and is fairly consistent
in its contribution over the entire period which can be characterized
as modest. The most signicant change over the two sub-periods is
attributed to the technical efciency change effect, which accounts
for nearly 45% of the fuel productivity growth over the entire period.
The later period reects the impact of efciency gains in fuel use
with the capital adjustment contribution marginally outweighing
the autonomous technical efciency change contribution. The scale
effect presents an opposite pattern being a signicant contributor
to fuel productivity growth in the early period and then being a negative, albeit marginal, contributor in the later period. These results
suggest that fuel use decisions were targeted for efciency gains in
the later period as the prospect of deregulation loomed large.27
Labor productivity is growing nearly three times faster than fuel
productivity over the entire period with most of that growth taking
place in the early period. Technological regress is present for labor
but quite minor. The technical efciency change contribution is
even more dramatic in this case accounting for 73% of labor productivity growth over the entire period. Contrary to the fuel productivity growth pattern over time, the labor productivity gains
from technical efciency changes in the earlier period dominate
the later period gains with the capital adjustment contribution to
the efciency change being the dominating factor with a similar
magnitude to that of the fuel productivity growth case. Similar to
the fuel productivity growth case, the scale effect presents an
opposite pattern being a signicant contributor to labor productivity growth in the early period and then being a negative, albeit
marginal, contributor in the later period. These results suggest that
labor use decisions were targeted for efciency gains in the earlier
period and can reect the relative importance of managing for fuel
productivity gains over labor productivity gains as the prospect of
deregulation loomed large.
The capital adjustment effect is nearly the same for both factors
by retarding each factor productivity growth by a similar magnitude over the entire period with the earlier period presenting the
stronger impact and the later period presenting the most trivial
constitution to overall factor productivity growth. This suggests
that these rms have made adjustments to the point that the
long-run equilibrium capital stock is being maintained.
24
Thermal conversion efciency is used typically to measure the performance of
generating plants. The report of EIA indicates that the standard deviation of an
average plant efciency of steam electric power generating plants measured by
thermal conversion efciency is very low for each plant which supports the
estimation results in Rungsuriyawiboon and Stefanou (2007) that these rms are
technically efcient in capital.
25
When testing for the presence of a Leontief technology in the use of capital using
xed effects, we cannot reject the null hypothesis that gross investment depends on
(y, z, t) and this estimation is used to generate the predicted ^z.
26
On productivity growth coinciding with our study period of 19871996, Atkinson
and Primont (2002) nd total factor productivity growth of 3.48% and 4.45% for the
cost and distance functions estimation, respectively. For their entire study period of
19611996, they nd negligible growth of 0.27% and 0.67%, for the cost and distance
functions, respectively. They report the productivity change and its components for
each year and we present the simple average for 19871996 period.
27
The studies of Rungsuriyawiboon and Stefanou (2007) and Rungsuriyawiboon
and Stefanou (2008) undertake the econometric specication and estimation of
dynamic efciency and productivity growth measures. They nd that total factor
productivity of the production of the US electricity industry grew at 2.26% per annum,
with technical change having the dominant contribution to the growth while the
combined scale effects had the least contribution.

852

M. Genius et al. / European Journal of Operational Research 220 (2012) 844852

5. Concluding comments
The measurement of productivity and technical efciency is
problematic in the presence of factor non-substitution, as is the
case for a Leontief technology. Radial measurements of efciency
are not adequate in this context as they can fail to recognize inefciencies associated with a subset of inputs, while non-radial measures overcome this limitation. With a view toward generalizing
the econometric measurement of factor demands in this setting,
the Leontief technology specication is merged with the copula
estimation of cross equation dependences to account for technical
efciency in the estimation of fuel and labor demand. The decomposition of partial factor productivity measures is developed that
allows for scale effects, technical change, efciency change and
the impact of capital utilization. The tools developed have been applied to data related to the large, fossil red steam electric generating utility facilities in the U.S.
Our results indicate that labor productivity grew nearly three
times faster than fuel productivity over the entire period with most
of that growth taking place in the early period. The contribution of
technical efciency improvement is more dramatic for labor productivity growth. When we partition the 19861996 period into
two sub-periods, we nd that the labor productivity gains from
technical efciency changes in the earlier period dominate the later period gains with the capital adjustment contribution to the
efciency change being the dominating factor with a similar magnitude to that of the fuel productivity growth case.
Acknowledgements
The authors acknowledge the nancial support of a Marie Curie
Transfer of Knowledge Fellowship of the European Communitys
Sixth Framework Programme under Contract Number MTKD014288.
References
Athanassopoulos, A.D., Lambroukos, N., Seiford, L., 1999. Data envelopment scenario
analysis for setting targets to electricity generating plants. European Journal of
Operational Research 115, 413428.
Atkinson, S.E., Cornwell, C., 1998. Estimating radial measures of productivity
growth: frontier versus non-frontier approaches. Journal of Productivity
Analysis 10, 3546.
Atkinson, S.E., Primont, D., 2002. Stochastic estimation of rm technology,
inefciency, and productivity growth using shadow cost and distance
functions. Journal of Econometrics 108, 203225.
Barnum, D.T., Gleason, J.M., 2011. Measuring efciency under xed proportions
technologies. Journal of Productivity Analysis 35, 243262.
Battese, G.E., Coelli, T.J., 1995. A model for technical inefciency effects in a
stochastic frontier production function for panel data. Empirical Economics 20,
325332.
Bauer, P.W., 1990. Decomposing TFP growth in the presence of cost inefciency,
nonconstant returns to scale, and technological progress. Journal of Productivity
Analysis 1, 287299.
Bravo-Ureta, B., Sols, D., Moreira Lpez, V.H., Maripani, J.F., Thiam, A., Rivas, T.,
2007. Technical efciency in farming: a meta-regression analysis. Journal of
Productivity Analysis 27, 5272.
Buccola, S.T., Sil, J., 1996. Productivity measurement in the agricultural marketing
sector. American Journal of Agricultural Economics 78, 13661371.
Cherchye, L., Van Puyenbroek, T., 2009. Semi-radial technical efciency
measurement. European Journal of Operational Research 193, 616625.
Clarke, K.A., 2007. A simple distribution-free test for nonnested model selection.
Political Analysis 15, 347363.
Diewert, W.E., Nakamura, A.O., 1993. Essays in Index Number Theory, vol. 1.
Elsevier, Amsterdam.
Energy Information Administration, 1996. Financial Statistics of Major US Investor
Owned
Electric
Utilities.
<http://www.eia.gov/cneaf/electricity/invest/
invest_sum.html>.
Fre, R., Lovell, C.A.K., 1978. Measuring the technical efciency of production.
Journal of Economic Theory 19, 150162.
Fre, R., Grosskopf, S., Margaritis, D., 2008. Efciency and productivity: Malmquist
and more. In: Fried, H.O., Lovell, C.A.K., Schmidt, S.C. (Eds.), The Measurement of
Productive Efciency and Productivity Growth. Oxford University Press, New
York, pp. 522622.

Farrell, M.J., 1957. The measurement of productive efciency. Journal of the Royal
Statistical Society A 120, 253281.
Frsund, F.R., 1996. On the calculation of the scale elasticity in DEA models. Journal
of Productivity Analysis 7, 283302.
Fried, H.O., Lovell, C.A.K., Schmidt, S.C., 2008. Efciency and productivity. In: Fried,
H.O., Lovell, C.A.K., Schmidt, S.C. (Eds.), The Measurement of Productive
Efciency and Productivity Growth. Oxford University Press, New York, pp. 3
91.
Granderson, G., Linvill, C., 1997. Nonparametric measurement of productive
efciency and in the presence of regulation. European Journal of Operational
Research 110, 643657.
Granderson, G., Linvill, C., 2002. Regulation, efciency, and Granger causality.
International Journal of Industrial Organization 20, 12251245.
Guan, Z., Oude Lansink, A., Van Ittersum, M., Wossink, A., 2006. Integrating
agronomic principles into production function specication: a dichotomy of
growth inputs and facilitating inputs. American Journal of Agricultural
Economics 88, 203214.
Haldi, J., Whitcomb, D., 1967. Economies of scale in industrial plants. Journal of
Political Economy 75, 373385.
Hiebert, L.D., 2002. The determinants of the cost efciency of electric generating
plants: a stochastic frontier approach. Southern Economic Journal 68, 935946.
Holloway, G., Paris, Q., 2002. Production efciency in the von Liebig model.
American Journal of Agricultural Economics 84, 12711278.
Holvad, T., Hougaard, J.L., Kronborg, D., Kvist, H.K., 2004. Measuring inefciency in
the Norwegian bus industry using multi-directional efciency analysis.
Transportation 31, 349369.
Hulten, C.R., 1973. Divisia index numbers. Econometrica 41, 10171026.
Karagiannis, G., Midmore, P., Tzouvelekas, V., 2002. Time-varying technical
inefciency, rm-specic indices of technical change, and panel data: an
application to the England and Wales dairy sector. Journal of Productivity
Analysis 18, 2338.
Knittel, C.R., 2002. Alternative regulatory methods and rm efciency: stochastic
frontier evident from US electric industry. Review of Economics and Statistics
84, 530540.
Kodde, D.A., Palm, F., 1986. Wald criteria for jointly testing equality and inequality
restrictions. Econometrica 54, 12431248.
Komiya, R., 1962. Technological progress and production function in the United
States steam power industry. Review of Economics and Statistics 44, 156166.
Kopp, R.J., 1981. The measurement of productive efciency: a reconsideration.
Quarterly Journal of Economics 96, 477503.
Lau, L., Tamura, S., 1972. Economies of scale, technical progress, and the
nonhomothetic Leontief production function: an application to the Japanese
petrochemical processing industry. Journal of Political Economy 80, 11671187.
Lindebo, E., Hoff, A., Vestergaard, N., 2007. Revenue-based capacity utilisation
measures and decomposition: the case of Danish North Sea trawlers. European
Journal of Operational Research 180, 215227.
Lovell, C.A.K., 1996. Applying efciency measurement techniques to the measurement
of productivity change. Journal of Productivity Analysis 7, 329340.
Luh, Y.H., Stefanou, S.E., 1991. Productivity growth in US agriculture under dynamic
adjustment. American Journal of Agricultural Economics 73, 11161125.
Morrison, C.J., 1992. Unraveling the productivity growth in the US, Canada and
Japan: the effects of sub-equilibrium, scale economies and mark-ups. Review of
Economics and Statistics 74, 381393.
Nakamura, S., 1990. A non-homothetic generalized Leontief cost function based on
pooled data. Review of Economics and Statistics 72, 649656.
Nelsen, R.B., 1999. An Introduction to Copulas. Springer-Verlag, New York.
Ozaki, I., 1970. Economies of scale and input-output coefcients. In: Carter, A.,
Brody, A. (Eds.), Applications of InputOutput Analysis. North-Holland,
Amsterdam, pp. 280302.
Paris, Q., 1992. The von Liebig hypothesis. American Journal of Agricultural
Economics 74, 10191028.
Portela, M., Thanassoulis, E.T., 2005. Protability of a sample of Portuguese bank
branches and its decomposition into technical and allocative components.
European Journal of Operational Research 162, 850866.
Reifschneider, D., Stevenson, R., 1991. Systematic departures from the frontier: a
framework for the analysis of rm inefciency. International Economic Review
32, 715723.
Rungsuriyawiboon, S., Stefanou, S.E., 2007. Dynamic efciency estimation: an
application to US electric utilities. Journal of Business and Economic Statistics
25, 226238.
Rungsuriyawiboon, S., Stefanou, S.E., 2008. Decomposition of total factor
productivity growth in the US electricity industry under dynamic adjustment.
Journal of Productivity Analysis 30, 177190.
Russell, R.R., 1985. Measures of technical efciency. Journal of Economic Theory 35,
109126.
Russell, R.R., 1987. On the axiomatic approach to the measurement of technical
efciency. In: Eichorn, W. (Ed.), Measurement in Economics: Theory and
Applications of Economic Indices. Physica-Verlag, Heidelberg, pp. 790.
Silberberg, E., 1972. Duality and the many consumers surpluses. American
Economic Review 62, 942952.
Sklar, A., 1959. Fonctions de repartition a n dimensions et Leurs Marges. Publications
de lInstitut de Statistique de LUniversit de Paris, Paris 8, 229231.
Trivedi, P.K., Zimmer, D.M., 2007. Copula modeling: an introduction for
practitioners. Foundations and Trends in Econometrics 1, 1111.
Vuong, Q.H., 1989. Likelihood ratio tests for model selection and non-nested
hypothesis. Econometrica 57, 307333.