Chencp

International Journal of Information Technology & Decision Making
Vol. 14 (2015)
°c World Scienti¯c Publishing Company
DOI: 10.1142/S0219622015500091
Nonradial Directional Performance Measurement

with Undesirable Outputs: An Application to OECD
and Non-OECD Countries
Po-Chi Chen*, Ming-Miin Yu†,**, Ching-Cheng Chang‡, Shih-Hsun Hsu§

and Shunsuke Managi¶,||
Int. J. Info. Tech. Dec. Mak. Downloaded from www.worldscientific.com
by MONASH UNIVERSITY on 05/04/15. For personal use only.
*Department of International Business, Chung Hua University, Hsinchu, Taiwan

†
Department of Transportation Science, National Taiwan Ocean University
2 Pei-Ning Road, 20224, Keelung, Taiwan
‡
Institute of Economics, Academia Sinica, Taipei, Taiwan
§Department of Agricultural Economics, National Taiwan University, Taipei, Taiwan
¶
Graduate School of Environmental Studies, Tohoku University, Sendai, Japan
||QueenslandUniversity of Technology
Brisbane, Australia
**yumm@mail.ntou.edu.tw
Published 16 April 2015
The objective of this paper is to provide a more comprehensive e±ciency measure to estimate
the performance of OECD and non-OECD countries. A Russell directional distance function
that appropriately credits the decision-making unit not only for increase in desirable outputs
but also for the decrease of undesirable outputs is derived from the proposed weighted Russell
directional distance model. The method was applied to a panel of 116 countries from 1992 to
2010. This framework also decomposes the comprehensive e±ciency measure into individual
input/output components' ine±ciency scores that are useful for policy making. The results
reveal that the OECD countries perform better than the non-OECD countries in overall, goods,
labor and capital e±ciencies, but worse in bad and energy e±ciencies.
Keywords: Data envelopment analysis; directional distance function; undesirable output;

Russell measure; slacks-based model.
1. Introduction
It has long been recognized in the technical e±ciency literature that data envelop-
ment analysis (DEA) is particularly adept at computing multiple input and output
production correspondences.1 Generally, there are two types of measures in DEA,
namely radial and nonradial.2 Radial measures are represented by Charnes–Cooper–
Rhodes (CCR)3 and Banker–Charnes–Cooper (BCC)4 models. However, since radial
measures of e±ciency overestimate technical e±ciency when there are nonzero slacks
1
2 P. C. Chen et al.
in the constraints de¯ning the piecewise linear technology,5 recently research has
sought to construct alternative nonradial e±ciency measures that account for slacks.
In the literature of the nonradial models, the proposed methods can be roughly
divided into the following three groups: (1) The Russell measure, which was ¯rst
presented by Färe and Knox Lovell6 with an input-oriented form. Nevertheless, it
only accounts for all the slacks of inputs, but fails to consider the ine±ciencies
associated with outputs. It was later extended by Färe et al.7 in a nonlinear form that
they refer to as the \Russell graph measure", which combines the input and output
Russell measures in an additive way and accounts for all the input slacks as well as
the output slacks. Pastor et al.8 then further revised it to a new measure called the
\enhanced Russell graph measure" which in turn combines input and output Russell
measures in a ratio form. (2) The additive model, which was developed by Charnes
et al.9 and also accounts for all sources of ine±ciency both in inputs and outputs.
However, it does not directly provide an e±ciency measure.8 (3) The slacks-based
measure model, which was proposed by Tone10 with the objective of maximizing all
the input and output slacks in fractional programming form. Cooper et al.11 showed
that the slacks-based measure is equivalent to the enhanced Russell graph measure.
Recently, Fukuyama and Weber5 introduced the directional distance function
technology into the slacks-based measure to develop a generalized measure of technical
ine±ciency which also accounts for all slacks in input and output constraints. This new
measure was referred to as the directional slacks-based ine±ciency (SBI) measure.
This is shown to yield the same information on performance as Tone's slacks-based
measure of e±ciency when the directional vectors for inputs and outputs are chosen to
equal the actual input and output vector, and can also be thought of as a generalization
of the original Russell measure of e±ciency. On the other hand, Färe and Grosskopf12
also proposed a generalization of the slacks-based measure based on the directional
distance function. The optimization problem of this measure is based on the sum of the
directional distance function, and can tell how much excess inputs have been employed
and how much outputs short of an e±cient level have been produced.
While these previous studies have made positive contributions, the undesirable
outputs are ignored in almost all these measures. In practice, there are some cases in
which both outputs which are desirable (goods) and undesirable (bads, such as air
pollution, waste or bad loans) are produced jointly.a Environmental e±ciency analysis
is thus one of the most studied topics in the last two decades. Environmental e±ciency
focuses on e±ciency levels across many decision-making units (DMUs). However,
temporal changes in e±ciency and shifts in production technology are also important.
Hoang and Coelli13 proposed a nutrient-orientated environmental e±ciency to con-
struct a nutrient total factor productivity index.b Zhang and Choi14 extended the total
productivity index by incorporating group heterogeneity and a nonradial directional
a As Smith25 pointed out, undesirable outputs may also appear in health care (e.g., complications of
medical operations) and business (e.g., tax payments) applications.
b Fujii et al.26 examine technical e±ciency and productivity growth in the Indian banking sector over the
period from 2004 to 2011 by using a weighted Russell directional distance model.
Nonradial Directional Performance Measurement with Undesirable Outputs 3
distance function measure to analyze the changes in the total-factor CO2 emission
performance of 93 fossil fuel power plants in China for the 2005–2010 period.
Early studies adopted various methods and data sets to conduct environmental
e±ciency analysis. There are four representative frontier models
the directional
distance function,15 the hyperbolic model,16 the Seiford and Zhu model,17 and the
undesirable output as input model18
to model undesirable outputs. Those can be
categorized into two major approaches: one assuming that bad outputs are strongly
disposable inputs (the Seiford and Zhu model and the undesirable output as input
model, for example), and the other assuming bad outputs are weakly disposable
outputs (the hyperbolic model and the directional distance function model). Both
approaches have their own problems in terms of yielding unacceptable implications
concerning tradeo®s among inputs, goods and bads. The undesirable output as input
model has been criticized for not accurately representing the production process,
because undesirable outputs are modeled as inputs.17 Sahoo et al.19 argued that the
ine±ciency score may depend on the choice of translation vector in the Seiford and
Zhu model. Both the hyperbolic model and the directional distance function model
have identical production sets but di®er in their ine±ciency indexes. Chen and
Delmas20 showed that both hyperbolic and directional distance function models are
not monotonic in undesirable outputs, point out some serious drawbacks in existing
frontier models, and propose a new eco-ine±ciency frontier model that avoids
identifying eco-ine±cient ¯rms as eco-e±cient. Murty et al.21 showed that the correct
trade-o®s in production can be best captured if a production technology is modeled
as an intersection of an intended-production technology of the ¯rm and nature's
residual-generation set. Leleu22 modeled undesirable outputs with a meaningful
primal/dual economic interpretation, by introducing a hybrid model to ensure the
jointness of desirable and undesirable outputs while constraining shadow prices of
undesirable outputs as an expected sign. As a whole, the ine±ciency score and
benchmark target depend on di®erent frontier assumptions and di®erent objective
function settings.
To our knowledge, most previous studies did not consider all the ine±ciency
sources of inputs, desirable and undesirable outputs simultaneously in a uni¯ed
model. For example, both Refs. 23 and 24 had extended the slacks-based measure
and Russell measure to incorporate undesirable outputs. Nevertheless, the models
established in the former one do not really account for all the ine±ciency sources of
undesirable outputs. The models constructed in the latter study only focus on
measuring the performance of undesirable outputs, without taking ine±ciency
sources of inputs and desirable outputs into consideration in the objective function
(referred to as a Russell environmental performance index in Ref. 24). It is therefore
reasonable to consider not only all the ine±ciency sources of inputs and desirable
outputs, but also all the ine±ciency sources of undesirable outputs when we evaluate
the performance of a DMU.
As presented by Zhang and Choi,27 the directional distance function has been
attracting positive attention in the ¯eld of environmental modeling. Zhang choi27
4 P. C. Chen et al.
surveyed various directional distance functions employed in energy and environ-

mental studies, and found that the radial e±ciency measures and output-oriented
model are the most widely used speci¯cations. However, the nonradial directional
distance model (DDF) extensions can provide promising insights in future studies.
Therefore, the purpose of this study is to extend the directional Russell measure of
ine±ciency proposed by Chung et al.5 into the cases where undesirable outputs exist.
We refer to the proposed model as the weighted Russell directional distance model
(WRDDM).c Compared to the original slacks-based measure and enhanced Russell
graph measure, which combine input and output e±ciency measures in a nonlinear
fractional form, our directional distance function-based measure is evaluated in
linear form, and hence possesses the attractive advantages of easy computation and
easy extension of incorporating the additional undesirable outputs into the pro-
gramming problems. Zhou et al.28 and Zhang et al.29 had de¯ned a directional dis-
tance model similar to ours, but did not discuss the relationships of the proposed
model with the slacks-based measure and Russell measure as well as the uniqueness
problems of nonradial measures. Thus, the proposed WRDDM model not only
considers the Russell measure to estimate each input, output and undesirable out-
put's ine±ciency but also imposes both a strong complementary slackness condition
on the WRDDM model to deal with an occurrence of multiple projections and
weights on the source of ine±ciencies. These are the bene¯ts of the proposed model.
The remainder of this paper is organized as follows. The next section presents the
model. For comparison purpose, we ¯rst illustrate the traditional directional distance
function model (TDDFM) problem, which credits a producer for simultaneously
increasing production of the good output, reducing production of bad outputs and
contracting employment of inputs. Then, the extended models follow. In addition,
the slacks-based measure and Russell measure, which correspond to our WRDDM,
are also presented. Section 3 demonstrates using a case study of panel data of 116
countries. Our measure obtained is easily decomposed into separate measures of
input, and desirable and undesirable output e±ciencies, and can help us to shed more
light on the sources of ine±ciencies. In the case where the environmental issues are
evaluated, it provides us an integrated measure of economic and environmental
performance. The ¯nal section concludes.
2. Preliminary
2.1. The traditional directional distance function model
Let inputs be denoted by x 2 R N þ , good outputs by y 2 R þ , and bad or undesirable
M
outputs by b 2 R Jþ , where R þ represents the non-negative Euclidean *-orthant. The
c The concept of weighted Russell measure is not new. Ruggiero and Bretschneider30 had developed a
weighted Russell measure to evaluate technical e±ciency. However, their measure is evaluated in distance
function rather than in directional distance function and did not consider undesirable outputs, so it is
di®erent from our measure.
directional distance function seeking to increase the desirable outputs and decrease
the undesirable outputs and inputs directionally can be de¯ned by the following
formulation:
Dðx; y; b; gÞ ¼ supf : ðx g x ; y þ g y ; b g b Þ 2 T g; ð1Þ
where the nonzero vector g ¼ ðg x ; g y ; g b Þ determines the \directions" in which
inputs, desirable outputs and undesirable outputs are scaled, and the technology
reference set T ¼ fðx; y; bÞ : x can produce ðy; bÞg satis¯es the assumptions of con-
stant returns to scale (CRS), strong disposability of inputs, and weak disposability of
both desirable outputs and undesirable outputs.
Suppose there are k ¼ 1; . . . ; K DMUs in the data set. Each DMU uses input
x ¼ ðx 1k ; x 2k ; . . . ; x Nk Þ 2 R N
k
þ to jointly produce desirable outputs y ¼ ðy 1 ; y 2 ; . . . ;
k k k
y M Þ 2 R þ and undesirable outputs b ¼ ðb 1 ; b 2 ; . . . ; b J Þ 2 R þ . The DEA piecewise

k M k k k k J
reference technology can be constructed as follows:

X
K
T ¼ fðx; y; bÞ : z k y mk y m ; m ¼ 1; . . . ; M ;
k¼1
XK
z k b jk ¼ b j ; j ¼ 1; . . . ; J ;
ð2Þ
k¼1
XK
z k x nk x n ; n ¼ 1; . . . ; N ;
k¼1
z k 0; k ¼ 1; . . . ; K g;
where z k are the intensity variables to shrink or expand the individual observed
activities of DMU k for the purpose of constructing convex combinations of the
observed inputs and outputs.
Relative to the CRS reference technology T constructed in (2), traditionally, for
each DMU k 0 ¼ 1; . . . ; K , the directional distance function can be obtained by
solving the following linear programming problem:
0 0 0 0
Dðx k ; y k ; bk ; gÞ ¼ max k
XK
0
s:t: z k y mk y mk 0 þ k g y m ; m ¼ 1; . . . ; M ;
k¼1
X
K
0
z k b jk ¼ b jk 0 k g b j ; j ¼ 1; . . . ; J ; ð3Þ
k¼1
XK
0
z k x nk x nk 0 k g x n ; n ¼ 1; . . . ; N ;
k¼1
z k 0; k ¼ 1; . . . ; K ;
k0
where measures the maximum expansion of desirable outputs and contraction of
undesirable outputs and inputs that remain technically feasible and can serve as a
0
measure of technical ine±ciency. If k ¼ 0, then DMU k 0 operates on the frontier of
0
T with technical e±ciency. If k > 0, then DMU k 0 operates inside the frontier of T .
6 P. C. Chen et al.
Other than being the generalization of the Shephard's distance functions,d one of
the important characteristics of the directional distance function is that the direction
in which performance is scaled can be speci¯ed °exibly to accommodate di®erent
0 0 0
analytical purposes. For example, if we set g ¼ ðg x ; g y ; g b Þ ¼ ðx k ; y k ; bk Þ, i.e.,
0
the direction is chosen based on the observed data, represents the potential pro-
k
portionate change in goods, bads and inputs. If instead we take

g ¼ ðg x ; g y ; g b Þ ¼ ð1; 1; 1Þ, then we can interpret the solution value as the net
improvement in performance in terms of feasible increase in goods outputs and feasible
decreases in bad outputs and inputs.31 On the other hand, setting g ¼ ð0; g y ; g b Þ, we
get the directional output (including goods and bads) distance function or environ-
mental directional output distance function as referred to by Färe et al.32
However, as mentioned previously, the measure of this approach fails to consider

the ine±ciencies associated with nonzero slacks, and would have the problem of
incorrectly regarding some evaluated DMUs as e±cient units.
2.2. The weighted directional distance model

The e±ciency measurement constructed in (3) expands all desirable outputs and con-
tracts all inputs and undesirable outputs by the same rate of . However, there is no
guarantee that the proportional contraction (expansion) rate for input items (Þ, de-
sirable output items (Þ and undesirable output items (Þ must be the same in practice.
We believe this is a strong assumption of the model. Thus, the formulation of (3) can be
generalized to accommodate di®erent expansion and contraction scales as follows:
0 0 0 0 0 0 0
Dw ðx k ; y k ; bk ; gÞ ¼ max w y k þ w b k þ w x k ¼ kw
XK
0
s:t: z k y mk y mk 0 þ k g y m ; m ¼ 1; . . . ; M ;
k¼1
X
K
0
z k b jk ¼ b jk 0 k g b j ; j ¼ 1; . . . ; J ; ð4Þ
k¼1
X
K
0
z k x nk x nk 0 k g x n ; n ¼ 1; . . . ; N ;
k¼1
z k 0; k ¼ 1; . . . ; K :
It is required that the directional vectors have same units of measurement as the
vectors of the observed data, so that it allows the , and to be added.e The
0 0 0 0
measure Dw ðx k ; y k ; bk ; g x n ; g y m ; g b j Þ given in (4) is maximized hyperbolically kw
0 0 0 k0 k0 k0
¼ w y þ w b þ w x by comparing the observed ðx n ; y m ; b j Þ with the frontier
k k k
0 0 0
(ðx nk 0 k g x n Þ,ðy mk 0 þ k g y m Þ, ðb jk 0 k g b j ÞÞ. The weighted directional distance
d Details of the relationship between directional distance functions and Shephard distance functions can be
found in Refs.15 and 33.
e This requirement is the same as what Ref. 5 (p. 276) point out, that the directional vectors have the same
units of measurement as the vectors of the input and output slacks, and when g ¼ 1, the role of the
directional vectors is also similar to that of the e l in Ref. 12.
function gives the expansion in good outputs and contraction in bad outputs and
0 0 0
inputs simultaneously. When Dw ðx k ; y k ; bk ; g x n ; g y m ; g b j Þ ¼ 0, DMU k 0 is tech-
nically e±cient since no additional improvements in good outputs, bad outputs and
0 0 0
inputs are feasible. Dw ðx k ; y k ; bk ; g x n ; g y m ; g b j Þ > 0 indicates technical ine±-
ciency. The coe±cients w y , w b and w x are associated with the priorities or managerial
preferences given to the outputs (goods and bads) and inputs, and their sum is
normalized to unity. The improvements for desirable outputs, undesirable outputs
0 0 0
and inputs can be measured by k , k and k , respectively, and then used to
k0
calculate the weighted ine±ciency score w .
0 0 0
Note that if we set k ¼ k ¼ k , then model (4) degenerates to model (3). In
addition, in the one input, one desirable output and one undesirable output case,
formulation (4) is able to account for all the slacks. However, in the multiple inputs,
desirable outputs and undesirable outputs case, it may still fail to identify all the
nonzero slacks associated with the input and output constraints.
3. The Model
3.1. The weighted russell directional distance model
Inspired by the equivalence of the enhanced Russell graph measure and slacks-based
measure, we follow an idea similar to that of the enhanced Russell graph measure to
generalize the traditional directional distance function and develop another similar
measure called the WRDDM. It is important to note that the WRDDM is a measure
closely related to the enhanced Russell graph measure, while the enhanced Russell
graph measure and slacks-based measure are special cases of the WRDDM measure.f
The proposed programming model is:
0 0 0
DR ðx k ; y k ; bk ; gÞ
! ! !
0
X
M
0
X
J
0
X
N
0
¼ kR ¼ max w y $ ym km þ wb $ bj kj þ wx $ xn kn
m¼1 j¼1 n¼1
X
K
k0
s:t: z k y mk y mk 0 þ g y m ;
m m ¼ 1; . . . ; M ;
k¼1 ð5Þ
XK
0
z k b jk ¼ b jk 0 kj g b j ; j ¼ 1; . . . ; J ;
k¼1
XK
0
z k x nk x nk 0 kn g x n ; n ¼ 1; . . . ; N ;
k¼1
z k 0; k ¼ 1; . . . ; K ;
f The di®erence between the enhanced Russell graph measure and the WRDDM exists in their objective
functions and constraints. The objective function of the enhanced Russell graph measure is speci¯ed for
calculating e±ciency measures, while those respective variables in the WRDDM are ine±ciency measures.
In addition, WRDDM has an additive form objective function, while the enhanced Russell graph measure
is in ratio form and the enhanced Russell graph measure does not consider undesirable outputs (cf. Ref. 8
for more details).
8 P. C. Chen et al.
0 0 0
where km , kj and kn are the individual ine±ciency measures for each desirable
output y m , each undesirable output b j and each input x n . In other words, this
speci¯cation allows for not only the technical ine±ciency associated with desirable
output, undesirable output and input to be di®erent as like model (4), but also allows
the technical ine±ciency among each of the desirable outputs, the undesirable
0 0
outputs and the inputs to be di®erent ( km ðm ¼ 1; . . . ; M Þ 6¼ kj ðj ¼ 1; . . . ; J Þ 6¼
0
kn ðn ¼ 1; . . . ; N ÞÞ. This makes sense, since, for example, the ine±ciency of
the input uses of a ¯rm could be more from the ine±cient use of labor (using too
many workers or there is labor congestion) but less from capital. On the other
hand, a producer may produce several products at the same time (e.g., crops
and livestock production of farmers or loans and securities investment production
of banks), but with di®erent production ability, and hence the production e±-
ciency for di®erent products would be di®erent. Therefore, one of the advantages
of this model is that it can help us to identify the place where we need to improve
most.
Once again, the directional vectors are required to have the same units of
0 0
measurement as the vectors of the observed data, so that it allows the km , kj and
k0
n to be added. If the coe±cients w y , w b and w x denote the given priorities asso-
ciated with the outputs (goods and bads) and inputs, and their sum is normalized
to unity, and the ine±ciencies of each corresponding input (output) are also
speci¯ed to allow assigning di®erent priorities to each of them, and their sums are
PM y PJ PN
m¼1 $ m ¼ 1, j¼1 $ j ¼ 1 and n¼1 $ n ¼ 1, then this is
b x
assumed to be one:
g
similar to what Liu and Tone34 did. However, it is noted that if the direction
vectors do not have the same units of measurement as the vectors of the observed
data, we can alternatively set the weights $ ym , $ bj and $ xn as values which can
normalize the direction vectors, such as the sample standard deviations of the
inputs and outputs. We will say more about this below when we discuss the unit
invariant property of WRDDM.
By means of the following change of variables,h the Russell directional-type in-
e±ciency measures can be changed to the slacks-based ones.
þ
0 s mk 0 0 s jk 0 0 s nk 0
km ¼ ; kj ¼ ; kn ¼ ;
gym gbj gxn
þ
where s mk 0 ; s jk 0 ; s nk 0 are the desirable and undesirable outputs and inputs slacks,
respectively which cause the ine±ciency for the evaluated unit k 0 . Then, the model
g That is, if we alternatively specify the objective function in (5) as 1 ð M $ y k 0 þ

P
PJ 0 P 0
3 m¼1 m m
j¼1 $ j j þ n¼1 $ n n Þ, the corresponding weighted slacks-based measure for the WRDDM will be
b k N x k
close to the weighted slacks-based measure presented by Liu and Tone,34 in which only inputs and outputs
are considered. Therefore, Liu and Tone's formulation is the special case of our corresponding weighted
slacks-based measure which will be introduced in Sec. 3.2. þ
h In the enhanced Russell graph measure, if considering bads, k 0 ¼ s mk0 0 , k 0 ¼ s jk 00 and k 0 ¼ s nk 00 .
m yk j bk n xk
m j n
can be re-expressed as follows:

0 0 0
DR ðx k ; y k ; bk ; gÞ
! ! !
0
XM þ
$ ym s mk 0
XJ
$ bj s jk 0
XN
$ xn s nk 0
¼ kR ¼ max w y þ wb þ wx
m¼1
g y m j¼1
gbj n¼1
gxn
X
K
þ
s:t: z k y mk ¼ y mk 0 þ s mk 0; m ¼ 1; . . . ; M ;
k¼1
XK ð6Þ

z k b jk ¼ b jk 0 s jk 0; j ¼ 1; . . . ; J ;
k¼1
X
K

z k x nk ¼ x nk 0 s nk 0; n ¼ 1; . . . ; N ;
k¼1
þ
0 ; s jk 0 ; s nk 0 0 8 m; j; n;
s mk
z k 0; k ¼ 1; . . . ; K:
Because the slacks for each variable are allowed to be di®erent, the objective in (6)
can help us to re°ect all ine±ciencies by calculating the maximum expansion of all
desirable outputs and contraction of all undesirable outputs and inputs that the
0
model can identify. Thus, Rk provides us an aggregate or overall ine±ciency measure
PM y k0
of performance in a nonradial manner in which the component of m¼1 $ m m ¼
P M $ m s mk 0
y þ
0
m¼1 g y ¼ k in the objective of (5) or (6) corresponds to the average desirable
m PJ b k0
P J $ bj s jk 0 k 0 and
output mix ine±ciencies and similarly, the j¼1 $ j j ¼ j¼1 g b ¼
PN x k0 P x
$ n s nk 0 0
j
n¼1 $ n n ¼ ¼ k correspond to the average undesirable output and

N
n¼1 g x
n
input mix ine±ciencies, respectively.
Let an optimal solution of model (6) be z k 0 , s mk þ
0 , s jk 0 and s nk 0 ; the WRDDM
0
overall ine±ciency is the weighted average of the desirable output ine±ciency, k ,
k 0 0
undesirable output ine±ciency, , and input ine±ciency, . We de¯ne the k
0
WRDDM overall ine±ciency measure kR by
0 0
kR ¼ w y
0
k þ w x k ðk ¼ 1; . . . ; K Þ;
k þ wb
0
ð7Þ
where
0
X
M
$ ym s þ 0
k ¼
mk
;
m¼1
gym
X
J
$ bj s jk 0
k 0 ¼
;
j¼1
gbj
0
X
N
$ xn s 0
k ¼ nk
:
n¼1
gxn
The optimal solution to this linear program (6) is an ine±ciency score which
10 P. C. Chen et al.
measures the largest rectilinear distance from the observation being evaluated to the
e±cient production frontier. Therefore, we have the following de¯nitions:
De¯nition 3.1 (WRDDM desirable output e±cient). If all optimal solutions
0
k ¼ 0, DMU k 0 is called desirable output e±cient. This implies that
of (6) satisfy
þ
the optimal slacks for the desirable outputs in (6) are all zero, i.e., s mk 0 ¼ 0ð8 mÞ.
De¯nition 3.2 (WRDDM undesirable output e±cient). If all optimal

k 0 ¼ 0, DMU k 0 is called undesirable output e±cient.
solutions of (6) satisfy
This implies that the optimal slacks for the undesirable outputs in (6) are all zero,

i.e., s jk 0 ¼ 0ð8 jÞ.
De¯nition 3.3 (WRDDM input e±cient). If all optimal solutions of (6) satisfy
0
k ¼ 0, DMU k 0 is called input e±cient. This implies that the optimal slacks for the

inputs in (6) are all zero, i.e., s nk 0 ¼ 0ð8 nÞ.
De¯nition 3.4 (WRDDM overall e±cient). If all optimal solutions of (6)

0
k 0 ¼ 0 and k 0 ¼ 0, DMU k 0 is called WRDDM
k ¼ 0,
simultaneously satisfy
overall e±cient. This implies that the optimal slacks for the desirable outputs,
þ
undesirable outputs and inputs in (6) are all zero, i.e., s mk 0 ¼ s jk 0 ¼ s nk 0 ¼ 0ð8m; j; nÞ.
If the WRDDM ine±ciency measure is zero, then the DMU is fully e±cient. The
ine±ciency measure of the WRDDM has the following properties:
Theorem 3.1. DMU k 0 is WRDDM overall e±cient, if and only if it is WRDDM
e±cient for all e±cient in desirable output, undesirable output and input.
Proof. From the equality (7), this theorem holds.

Theorem 3.2. The projected DMU o on the frontier is WRDDM overall e±cient.
Proof. See Appendix A.

Theorem 3.3. If, for any two DMUs k 0 and k 0 0 , their ine±ciencies simultaneously
0
k
satisfy all of the three inequalities
00
k 0
k ; k 0 0 and k 0 k 0 0 and then
0 0 0
it holds that kR kR .
Proof. From the equality (7), this theorem holds.

To solve the model, we need to set the directional vector. As mentioned previ-
ously, one of the important characteristics of the directional distance function is that
the direction in which performance is scaled can be speci¯ed °exibly to accommodate
di®erent analytical purposes. The directional vectors in the literature are commonly
the vector based on the observed data and the unit vector. Therefore, we can de¯ne
the following two WRDDM models:
De¯nition 3.5. If the directional vector g ¼ ðg x ; g y ; g b Þ is set to be g ¼
ðx; y; bÞ, then the WRDDM is called the observation directional WRDDM.
De¯nition 3.6. If the directional vector g ¼ ðg x ; g y ; g b Þ is set to be unit

direction, g ¼ ð1; 1; 1Þ, then the WRDDM is called the unit directional WRDDM.
0 0 0
It is noted in this case, that the km ; kj and kn in the objective function in model
þ k0
(5) are equivalent to s mk 0 , s jk 0 and s nk 0 without units, respectively, and R is the
weighted average of those weighted sums of respective slacks. When all the indi-
vidual weights, $ ym ; $ bj and $ xi are set to be one, then model (5) degenerates to
Ref. 12 model (5).
Because the vectors of these two WRDDMs are associated with same units of the
observed data, it can be easily veri¯ed that they both possess the desirable properties
of units invariance and translation invariance as described in Ref. 5. However, if the
directional vectors do not have the same units of measurement as the vectors of the
observed data, we can specify the following inverse variance WRDDM to have these
two desirable properties.

De¯nition 3.7. If the directional vectors do not have the same units of
measurement as the vectors of the observed data, and if the weights corresponding
to each of the inputs, and desirable and undesirable outputs are the reciprocal of
sample standard deviations, then the WRDDM is called the inverse variance
WRDDM.i
Theorem 3.4. The inverse variance WRDDM is units invariant.

Proof. See Appendix B.
3.2. Relationship between WRDDM, slacks-based measure model

and enhanced russell graph measure model
Fukuyama and Weber5 showed that their directional SBI measure yields the same
information on performance as Tone's slacks-based measure of e±ciency when the
directional vectors for inputs and outputs are chosen to equal the actual input and
output vectors. That is, the slacks-based measure is a special case of SBI. Besides,
as mentioned, Cooper et al.11 showed that the slacks-based measure is equivalent
to an enhanced Russell graph measure. Therefore, it is natural for us to develop
the corresponding weighted slacks-based measure and enhanced Russell graph
measure models to extend this relationship to the situation where undesirable
outputs exist.
0 0 0
Proposition 3.1. By setting g ¼ ðg x ; g y ; g b Þ ¼ ðx k ; y k ; bk Þ, the correspond-
ing weighted slacks-based measure and enhanced Russell graph measure yield
information on performance equivalent to the WRDDM.
Proof. See Appendix C.
i Lovell and Pastor36 used the same way to propose a unit invariant weighted additive DEA model.
3.3. Uniqueness problems and strong complementary

slackness condition
Unique determination of the optimum value is one of the desirable properties for
e±ciency comparison. However, as Sueyoshi and Sekitani35 pointed out, many DEA
models su®er from an occurrence of multiple projections, especially for nonradial
measures. Mathematically, although the overall ine±ciency R calculated from the
þ
WRDDM is uniquely determined, the corresponding slacks s m and/or s j and/or s n
may have multiple optima and hence solutions of b and x may also be multiple.
y,
If the obtained slacks have multiple optima, it means that the projections to the
e±cient status described in Theorem 3.2 cannot be uniquely determined. We
therefore have the following proposition:
Proposition 3.2. The WRDDM violates the property of unique projection for
e±ciency comparison.
We give an example to show that the WRDDM ine±ciency scores may not be
uniquely determined for e±ciency comparison in Appendix D.
To overcome this di±culty, Sueyoshi and Sekitani35 theoretically discussed use of
primal/dual DEA combined with a Strong Complementary Slackness Condition
(SCSC) to uniquely identify a single reference set onto which an ine±cient DMU is
projected. In order to describe the proposed WRDDM model with SCSC, we need to
consider the dual of model (5).
It is as follows:
X
M
0 0
X
N
0 0
X
J
0 0
min ym þ
k k
um v nk x nk þ kj b kj
m¼1 n¼1 j¼1
X
M
0
X
N
0
X
J
0
s:t ym þ
k k
um v nk x nk þ kj b kj 0k ¼ 1; . . . ; K ;
m¼1 n¼1 j¼1
k0 y ð8Þ
u m gym wy $ m m ¼ 1; . . . ; M ;
0
kj g b j w b $ bj j ¼ 1; . . . ; J ;
0
v nk g x n w x $ xn n ¼ 1; . . . ; N ;
0 0 0
k
um ; v nk 0; kj free;
0 0 0
k
where u m ; v nk and kj are multipliers for desirable outputs, inputs and undesirable
outputs, respectively. It is worth mentioning that the constraints in (8) show that
k0 0 0
each u m g y m , kj g b j and v nk g x n should not be less than w y $ ym , w b $ bj and w x $ xn ,
respectively.
Following Ref. 38, Models (5) and (8) along with SCSC are asj:
max
j Because all the slacks have been accounted for in the WRDDM, we do not need to consider slack variables
in this model.
s.t. all constraints in primal (5) and dual (8) problems, and
! ! !
X
M
y k0
XJ
b k0
XN
x k0
wy $ m m þ wb $ j j þ wx $n n
m¼1 j¼1 n¼1
X
M
0 0
X
N
0 0
X
J
0 0
¼ ym þ
k k
um v nk x nk þ kj b kj
m¼1 n¼1 j¼1
!
X
M XN X
J
k0 k 0 0
um ym þ v nk x nk þ kj b kj þ zk k ¼ 1; . . . ; K;
m¼1 n¼1 j¼1
0 0 0
ym þ
k k
um km w y $ ym þ m ¼ 1; . . . ; M ; ð9Þ
k0 k0 k0
j b j þ j w b $ bj þ j ¼ 1; . . . ; J ;
0 0 0
v nk x nk þ kn w x $ xn þ n ¼ 1; . . . ; N ;
0
k
um m ¼ 1; . . . ; M ;
0
v nk n ¼ 1; . . . ; N ;
0
kj 6¼ 0 j ¼ 1; . . . ; J ;
0:
3.4. The WRDDM with variable returns to scale technology

For standard DEA with only inputs and desirable outputs, if the e±ciency is eval-
uated with variable returns to scale (VRS) technology, one only need to impose the
P
additional constraint of K k¼1 z k ¼ 1 in the CRS DEA formulation. However, when
the undesirable outputs are considered, Refs. 39 and 40 had showed that, in order to
correctly model a convex technology exhibiting weakly disposability of good and bad
P
outputs, only imposing the constraint of K k¼1 z k ¼ 1 based on technology in Ref. 41
on the formulation of (2) is not su±cient. Kuosmmanen39 thus provided an alter-
native VRS production possibility set as follows:
(
X
K
T v ¼ ðx; y; bÞ : k z k y mk y m ; m ¼ 1; . . . ; M ;
k¼1
X
K
k z k b jk ¼ b j ; j ¼ 1; . . . ; J ;
k¼1
XK
z k x nk x n ; n ¼ 1; . . . ; N ;
k¼1
XK
z k ¼ 1;
k¼1
0 k 1; k ¼ 1; . . . ; K ;
)
z k 0; k ¼ 1; . . . ; K ;
ð10Þ
where k is the proportion by which the good and bad outputs are reduced, and is
referred to as the abatement factor of DMU k. It is noted that this abatement factor
is allowed to be nonuniform across DMUs. Kuosmanen and Podinovski40 proved that
the use of multiple abatement factors is the only correct technology suitable for the
implementation of the weak disposability axiom. We thus follow Kuosmanen's
technology to construct our VRS WRDDM. Nevertheless, this technology set is
nonlinear and can be linearized as followsk:
(
X
K
T v ¼ ðx; y; bÞ : k y mk y m ; m ¼ 1; . . . ; M ;
k¼1
X
K
k b jk ¼ b j ; j ¼ 1; . . . ; J ;
k¼1
XK
ð k þ k Þx nk x n ; n ¼ 1; . . . ; N ;
k¼1
XK
ð k þ k Þ ¼ 1;
k¼1 )
k ; k 0; k ¼ 1; . . . ; K :
ð11Þ
Relative to Kuosmanen's technology formulated in (11), the programming model
of the VRS WRDDM is:
0 0 0
v ðx ; y ; b ; gÞ
DR k k k
! ! !
0
X
M
0
X
J
0
X
N
0
¼ kR v ¼ max w y $ ym km þ wb $ bj kj þ wx $ xn kn
m¼1 j¼1 n¼1
X
K
0
s:t: k y mk y mk 0 þ km g y m ; m ¼ 1; . . . ; M ;
k¼1
XK
0
k b jk ¼ b jk 0 kj g b j ; j ¼ 1; . . . ; J ; ð12Þ
k¼1
X
K
0
ð k þ k Þx nk x nk 0 kn g x n ; n ¼ 1; . . . ; N ;
k¼1
XK
ð k þ k Þ ¼ 1;
k¼1
k ; k 0; k ¼ 1; . . . ; K:
k0
Similarly, R v provides the overall ine±ciency measure of performance under VRS
PM y k0 0 PJ b k0 k 0 and
technology, and its components, m¼1 $ m m ¼
kv , j¼1 $ j j ¼ v
PN x k0 0
n¼1 $ n n ¼
kv , in the objective of (10) correspond to the VRS weighted average
of the desirable output ine±ciency, undesirable output ine±ciency, and input
k For details, please see Refs. 39 and 40.
ine±ciency, respectively. This VRS WRDDM measure also enjoys the property of
unit invariance but not property of translation invariance.
4. Empirical Application
Increasing environmental performance has long been a policy objective in most
countries. As concern over global warming mounts, there have been several studies
carrying out cross-country performance comparisons or time-series productivity
growth analysis taking into account carbon dioxide (CO 2 Þ emission by using the
method of DEA, such as Refs. 23, 34 and 42–46.l However, no matter which model
these papers adopted, few of them gauge performance in terms of increased good
outputs and decreased bad outputs and inputs simultaneously and account for all the
slacks which the model can identify. We apply a dataset consisting of 116 countries
from 1992 to 2010 to illustrate our WRDDM. The sample period covers 19 years and
allows us to compare the trends of ine±ciency over di®erent types of classi¯cations.
4.1. Data and variables description

Our measure of aggregate outputs is gross domestic product (GDP) as the desirable
output, and CO2 emission as the undesirable output in this study. Regarding inputs,
other than capital and employment by Refs. 42 and 43, energy use is included to
re°ect the fact that CO2 emissions are directly related to the use of energy, and
energy use plays an important role in interacting with production productivity.47,48
That is, we assume that each country employs labor, capital and energy to produce
aggregate outputs including goods and services associated with the undesirable by-
product, CO2. This variable speci¯cation is the same as Refs. 45, 46 and 49–52.
As for the sources of these data, energy use and CO2 are compiled from World
Development Indicators (WDI, World Bank), while GDP, employment and capital
are provided by Penn World Tables (PWT 8.0). GDP and capital are both expressed
in 2005 international prices.m
Table 1 displays the summary statistics of all the input and output variables used
in the study as well as the CO2 intensities in terms of CO2 emissions per GDP and
CO2 factor in terms of CO2 emissions per energy consumption according to groups
classi¯ed by income levels following the World Bank classi¯cation and OECD/non-
OECD countries. It can be seen that the economies of the low income countries have
the lowest average levels for all the inputs and outputs variables and both the CO2
intensities and factor. As for OECD/non-OECD countries, the same holds, with all
the average levels of the inputs, outputs and CO2 intensities and factor lower for non-
OECD countries than for OECD countries. This means that the non-OECD coun-
tries employ fewer inputs to produce fewer outputs with lower CO2 intensities in
terms of CO2 emissions per GDP and CO2 factor in terms of CO2 emissions per
energy consumption.
l For more related literature, please see Ref. 27.
m GDP, employment and capital are from the data of rgdpna, emp and rkna in the PWT 8.0, respectively.
Table 1. Summary statistic (1992–2010).
GDP Labor Capital Energy CO 2 kt CO 2 /energy CO 2 /GDP

millions millions millions kt of oil (kt/kt of oil (kt/million)
2005 US$ 2005 US$ equivalent equivalent)
116 countries
mean 428,451.3 21.464 1,346,770.5 83,384.3 208,841.1 2.157 0.454
SD 1,276,980.3 77.818 4,064,285.2 260,818.9 712,316.3 0.920 0.321
max 13,144,400.0 781.377 40,615,356.0 2,516,731.2 8,286,892.0 5.617 2.599
min 2977.8 0.110 7952.7 664.9 707.7 0.087 0.063
High income countries (43)
mean 444,871.4 23.185 1,375,818.6 90,187.0 226,857.1 2.181 0.459
SD 1,394,330.8 79.733 4,329,602.2 283,941.1 749,816.1 0.919 0.314
max 12,992,000.0 766.950 40,615,356.0 2,307,819.2 6,791,804.7 5.257 2.123
min 3622.4 0.114 7952.7 664.9 953.4 0.091 0.063

Upper middle income countries (35)

mean 426,265.7 22.288 1,350,530.2 82,399.3 207,939.4 2.170 0.465
SD 1,157,161.3 85.660 3,732,286.9 247,172.4 700,471.2 0.921 0.334
max 12,564,300.0 772.350 37,569,048.0 2,318,861.1 7,035,443.9 5.617 2.599
min 2977.8 0.137 8775.1 676.1 707.7 0.111 0.081
Lower middle income countries (26)
mean 447,289.7 19.159 1,411,822.7 84,373.8 211,108.7 2.134 0.440
SD 1,425,005.5 67.414 4,585,021.3 281,672.9 778,584.2 0.910 0.304
max 13,144,400.0 781.377 40,057,600.0 2,516,731.2 8,286,892.0 5.281 2.312
min 3970.6 0.122 8966.9 719.6 777.4 0.087 0.100
Low income countries (12)
mean 335,170.6 17.889 1,090,768.7 59,736.5 141,999.9 2.085 0.436
SD 706,648.0 67.418 2,538,246.5 136,458.2 381,723.7 0.945 0.344
max 4,251,395.9 717.395 17,282,742.0 1,161,353.1 3,469,510.0 5.343 2.416
min 3392.7 0.110 9148.0 691.3 821.4 0.092 0.095
OECD countries (34)
mean 442,565.2 23.6 1,383,553.7 89,717.2 225,559.1 2.165 0.456
SD 1,378,797.4 81.1 4,330,339.9 277,836.8 742,417.7 0.913 0.312
max 12,992,000.0 772.3 40,615,356.0 2,307,819.2 7,035,443.9 5.421 2.123
min 3957.2 0.1 7952.7 664.9 986.4 0.091 0.074
Non-OECD countries (82)
mean 422,599.2 20.6 1,331,518.9 80,758.4 201,909.2 2.154 0.453
SD 1,232,717.5 76.4 3,950,064.8 253,474.3 699,582.9 0.923 0.325
max 13,144,400.0 781.4 40,057,600.0 2,516,731.2 8,286,892.0 5.617 2.599
min 2977.8 0.1 8304.1 676.1 707.7 0.087 0.063
Note: 1. We use the classi¯cation drawn from the World Bank in which the economies are divided
according to 2011 GNI per capita, calculated using the World Bank Atlas method. The groups are: low
income, $1005 or less; lower middle income, $1006–$3975; upper middle income, $3976–$12,275; and high
income, $12,276 or more.
2. The value inside the parentheses after each category is the number of countries.
4.2. Di®erences between the TDDFM and WRDDM models

Table 2 reports the results of average overall ine±ciency scores of all countries
calculated by the TDDFM, individual factors' ine±ciency scores of all countries
calculated by the WRDDM, and the average overall ine±ciency di®erences (BIAS)
Table 2. Aggregate (overall) technical ine±ciency.
1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 AVG. 92-10 AVG. 01-10
TDDFM
overall 0.194 0.214 0.213 0.233 0.223 0.232 0.233 0.222 0.235 0.224 0.230 0.214 0.217 0.207 0.199 0.193 0.188 0.182 0.192 0.213 0.205
WRDDM
overall 0.350 0.363 0.372 0.378 0.375 0.363 0.362 0.334 0.351 0.332 0.358 0.321 0.321 0.312 0.305 0.302 0.305 0.301 0.309 0.338 0.317
goods 0.415 0.508 0.461 0.535 0.465 0.458 0.463 0.404 0.462 0.393 0.417 0.334 0.339 0.278 0.260 0.262 0.243 0.225 0.213 0.376 0.297
bads 0.383 0.342 0.395 0.359 0.381 0.361 0.378 0.343 0.349 0.332 0.379 0.357 0.326 0.340 0.350 0.354 0.366 0.365 0.400 0.361 0.357
inputs 0.252 0.240 0.259 0.240 0.278 0.269 0.245 0.254 0.241 0.271 0.278 0.272 0.299 0.316 0.306 0.290 0.304 0.311 0.314 0.276 0.296
labor 0.428 0.366 0.444 0.354 0.474 0.468 0.445 0.460 0.462 0.471 0.471 0.479 0.495 0.536 0.478 0.449 0.426 0.450 0.441 0.452 0.470
capital 0.092 0.150 0.100 0.126 0.106 0.093 0.073 0.091 0.060 0.120 0.104 0.080 0.120 0.109 0.047 0.037 0.025 0.026 0.029 0.084 0.070
energy 0.237 0.205 0.232 0.239 0.254 0.246 0.216 0.212 0.201 0.221 0.258 0.257 0.284 0.303 0.393 0.384 0.463 0.458 0.473 0.291 0.349
BIAS 0.156 0.149 0.158 0.145 0.152 0.131 0.129 0.112 0.116 0.108 0.128 0.107 0.104 0.105 0.106 0.109 0.117 0.119 0.118 0.125 0.112
Note: TDDFM: traditional directional distance function results, formulation (3). WRDDM: weighted Russell directional distance function results, formulation (6) in
which weights are set equally as 1/3. BIAS ¼ WRDDM (overall scores)–TDDFM (overall scores).
Nonradial Directional Performance Measurement with Undesirable Outputs
17
between these two models. It is important to observe that requiring g ¼

ðg x ; g y ; g b Þ having a value of one for each type of input and output implies that
the directional vector is in the units of the observed data. The resulting values of the
ine±ciency measures are interpreted as the number of units of each type of input
(output) that can be decreased (increased) for each observation.12 Thus, models are
0 0 0
solved by setting g ¼ ðg x ; g y ; g b Þ ¼ðx k ; y k ; bk Þ, since we would like to observe
how much the percentage needed to be improved, and all the weights in the
WRDDM are speci¯ed to be equal (i.e., w y ¼ w b ¼ w x ¼1/3, $ ym ¼ 1=M , $ bj ¼ 1=J
and $ xn ¼ 1=N ). It should be noted that a uniform weight scheme is used in this
paper because we assume that the improvement of inputs and both the desirable and
undesirable outputs are of equal importance; however, their relative importance is
not always uniform.n For example, if the problem of global warming becomes worse,
we may wish to attach a larger weight on the improvement of bads. In addition, the
aggregate overall technical ine±ciency computed from the models of the traditional
directional distance function and weighted Russell directional distance function are
evaluated at the same input bundles transformed to outputs of 116 countries in the
same period. The model (6) with CRS technology is used because,o given that the
analysis involves the use of aggregate country-level data, it does not appear to be
sensible to consider a VRS technology.53 Zhou et al.50 also pointed out that the CRS
environmental DEA technology is the most commonly adopted practice in the lit-
erature.
As can be found in Table 2, in 1992–2010, the TDDFM shows slightly larger
di®erences than it does in 2001–2010 (0.213 versus 0.205). The same di®erence
pattern is also found in the respective periods of the WRDDM (0.338 in 1992–2010
and 0.317 in 2001–2010). The average overall ine±ciency score estimated by the
TDDFM is lower than those by the WRDDM, thus the BIAS scores during the study
period are all positive. It is an explicit relationship derived from the TDDFM and
WRDDM models that ine±ciency scores estimated by the latter are no less than the
former due to the lack of accounting for nonzero slacks in the former one. It is also
found that all ine±ciency scores are broadly smooth over the period 1992–2010. This
may imply that these sample countries did not have large e±ciency changes during
this 1992–2010 period, even though the average ine±ciency scores of these countries
showed slight improvement in 2001–2010.
A Wilcoxon test is used to test the hypothesis that the two vectors of means of
overall ine±ciency score were equal. We ¯nd there are highly signi¯cant di®erences
(z-score ¼ 3:823, p-value < 0:001) between the results of these two models,
n The WRDDM model can be applied to any case wherein the weights of inputs, of desirable or undesirable
outputs are known or can be obtained exogenously, for example, the cost share for inputs, revenue share for
outputs or weights based on decision makers' preference. Tsai and Mar Molinero54 adopted the proportions
of individual activities' current operating expenditures in relation to total expenditures as the initial
weights for each activity. Tsutsui and Goto55 used the cost share of each input as the input weights.
o Picazo-Tadeo et al.56 provided a detailed justi¯cation of the CRS assumption in the framework of eco-
e±ciency measurement and indicated that the size of production activity generally matters little in the
assessment of eco-e±ciency.
implying that the null hypothesis of equal overall ine±ciency score has to be rejected.
Therefore, this con¯rms that the TDDFM and the WRDDM model di®er in their
ability to measure comprehensive e±ciencies for production of the sample data.
In addition to signi¯cant di®erence in overall ine±ciency score, the WRDDM
provides more information on how to get priority improvement in those ine±ciency
factors which a country needs to focus on. This is what the TDDFM was unable to
do. For example, the fact that the most ine±cient factor on average (0.357) is bads
during the period 2001–2010 implies that policies to improve e±ciency in bads are
needed. Moreover, input factor ine±ciencies also have been assessed for each country
around the world in our sample in three dimensions, namely, directional reduction of
three inputs considered and speci¯c reductions of labor, capital and energy, respec-
tively. Labor ine±ciency is the most ine±cient one during the period 2001–2010 with
an annual average ine±ciency of 0.470, while the capital ine±ciency (0.0070) is the
less ine±cient input factor. In other words, improvements in labor ine±ciency con-
tribute 52.9% to the input e±ciency improvement in the sample data during
2001–2010, while improvements in the capital ine±ciency contribute just 7.9%. The
ine±ciency pro¯le observed in 1992–2000 is similar, with more ine±ciency in goods
and less ine±ciency in energy factor. Note that a practically useful advantage of
the WRDDM illustrated here provides further insight into the relative levels of
individual factor e±ciencies among countries.
4.3. Comparison of input with output e±ciencies

We now turn to the computed overall ine±ciency and its components separated by
OECD and non-OECD countries (see summary in Table 3). Our e±ciency modeling
not only describes the input side of the production technology, but also includes
desirable and undesirable output e®ects, which are present on the output side of
production. Especially for the bad outputs, which are common in production, they
are expected to in°uence the provision of a country's desirable output. Thus, these
undesirable by-products can have an evident e®ect on the output e±ciency for
countries at a point where joint-production appears in production. Table 3 also
presents the decomposition of input ine±ciency scores of OECD and non-OECD
countries into labor, capital and energy parts. For instance, the 0.301 value (the
value of overall ine±ciency, which is the weighted sum of corresponding ¯gures of
inputs, goods and bads) computed for the OECD countries in 1992 on average, shows
the factor by which the good output can be expanded by more than 19.5% times
compared to its current level while simultaneously cutting 48.5% of the pollution
emissions and 22.4% of resource use (34.5% labor, 11.4% capital and 21.3% energy)
and still remaining in the feasible production set. A similar interpretation applies to
the values of the non-OECD countries in each component.
As regards the target setting by countries as to which inputs or bads to reduce or
which outputs to increase, one can obtain by the results how a country can attain its
full technical e±ciency. For the individual countries, the WRDDM results allow each
20
P. C. Chen et al.
Table 3. Components of the overall ine±ciency by OECD and non-OECD countries.
1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 AVG. AVG.
92-10 01-10
OECD countries
overall 0.301 0.298 0.325 0.302 0.326 0.324 0.308 0.281 0.273 0.281 0.295 0.275 0.275 0.278 0.259 0.250 0.257 0.261 0.266 0.286 0.270
goods 0.195 0.192 0.217 0.209 0.250 0.262 0.236 0.202 0.248 0.217 0.251 0.201 0.175 0.153 0.149 0.107 0.052 0.041 0.053 0.179 0.140
bads 0.485 0.481 0.522 0.456 0.489 0.476 0.455 0.401 0.364 0.391 0.406 0.386 0.398 0.410 0.390 0.411 0.450 0.464 0.472 0.437 0.418
inputs 0.224 0.221 0.235 0.240 0.238 0.234 0.232 0.241 0.207 0.235 0.226 0.240 0.251 0.271 0.239 0.233 0.269 0.278 0.273 0.241 0.251
labor 0.345 0.330 0.346 0.340 0.336 0.331 0.338 0.358 0.305 0.348 0.331 0.356 0.358 0.381 0.280 0.272 0.270 0.289 0.272 0.326 0.316
capital 0.114 0.126 0.138 0.140 0.138 0.146 0.149 0.163 0.129 0.161 0.127 0.158 0.160 0.176 0.093 0.073 0.045 0.045 0.033 0.122 0.107
energy 0.213 0.206 0.221 0.239 0.241 0.226 0.210 0.202 0.186 0.197 0.221 0.205 0.235 0.255 0.345 0.353 0.492 0.500 0.513 0.277 0.332
Non-OECD countries
overall 0.370 0.390 0.391 0.410 0.395 0.379 0.384 0.356 0.383 0.353 0.384 0.340 0.341 0.326 0.324 0.323 0.324 0.317 0.327 0.359 0.336
goods 0.506 0.638 0.562 0.670 0.554 0.540 0.557 0.488 0.550 0.466 0.486 0.390 0.407 0.330 0.306 0.326 0.322 0.302 0.280 0.457 0.362
bads 0.341 0.284 0.342 0.319 0.337 0.313 0.346 0.320 0.342 0.308 0.368 0.345 0.296 0.311 0.334 0.330 0.332 0.324 0.371 0.330 0.332
inputs 0.264 0.248 0.269 0.240 0.294 0.284 0.250 0.260 0.256 0.286 0.299 0.285 0.320 0.335 0.334 0.314 0.319 0.325 0.331 0.290 0.315
labor 0.462 0.380 0.485 0.360 0.531 0.525 0.490 0.502 0.527 0.522 0.529 0.530 0.551 0.600 0.560 0.522 0.490 0.516 0.511 0.505 0.533
capital 0.083 0.161 0.085 0.121 0.092 0.072 0.042 0.061 0.032 0.104 0.094 0.047 0.104 0.081 0.027 0.021 0.017 0.018 0.027 0.068 0.054
energy 0.247 0.204 0.236 0.239 0.260 0.254 0.218 0.217 0.208 0.231 0.273 0.279 0.304 0.323 0.413 0.397 0.451 0.440 0.456 0.297 0.357
country to ¯nd the area in which the greatest improvement is needed. For example,
in 2010, Estonia was the most ine±cient country among the OECD countries, with
an overall ine±ciency score of 0.522. Its factors' ine±ciency scores for inputs, goods
and bads are 0.319 (0.427 for labor, 0.000 for capital and 0.531 for energy), 0.646 and
0.600, respectively. These suggest that expanding good output production should be
the top priority for Estonia to improve its performance. In contrast, a non-OECD
country, Mozambique, has an overall ine±ciency score of 0.220 with 0.519 (0.861 for
labor, 0.000 for capital and 0.695 for energy), 0.000 and 0.142 scores for inputs, goods
and bads components, respectively. These suggest that the top performance im-
provement priority for Mozambique is to cut its resource use, especially for labor.p
This provides additional guidance to the country as to how to improve e±ciency, but
also provides the basis for individual factor ine±ciency scores. Furthermore, it is
apparent that the rates of improvements in individual input-use, individual bad-

reduction and individual output-production are di®erent, implying that a country
need not reduce its inputs and bads, and increase its outputs equiproportionately so
as to improve its technical e±ciency.
The degree of OECD's overall ine±ciency was generally lower than non-OECD
throughout the period, which is con¯rmed by the Mann–Whitney U test with
p-value < 0:001. This implies that OECD's comprehensive e±ciency is generally
better than non-OECD countries. It is hence interesting to investigate whether all
components of overall ine±ciency have the same pattern. Therefore, the rank order
correlation coe±cients for both OECD and non-OECD countries are also computed
for each component. The results are displayed in Table 4.
On the output side, the means of the rank order correlation coe±cients for both
categories of countries are signi¯cantly unequal for the goods ine±ciency (OECD 10.0
and non-OECD 29.0 with p-value < 0:001). Therefore, we accept that the OECD
countries' goods e±ciencies are better than non-OECD countries'. However, this is not
Table 4. Rank order correlation coe±cient of the OECD and non-OECD

countries: Mann–Whitney test.
Component Rank order correlation coe±cient U p-value
OECD non-OECD
overall 10.610 28.390 11.500 <0.001*

goods 10.000 29.000 0.000 <0.001*
bads 28.890 10.110 2.000 <0.001*
Inputs 11.710 27.290 32.500 <0.001*
labor 10.110 28.890 2.000 <0.001*
capital 25.660 13.340 63.500 0.001*
energy 16.870 22.130 130.500 0.144
Note: *Signi¯cant at 0.01 level.
p For lack of space, the detail results on the improvements in input-using, bads-reducing and output-
producing by the individual countries are not reported here. The results of individual countries are
available upon request.
the case for bads ine±ciency (OECD 28.89 and non-OECD 10.11 with p-value < 0:001).
The non-OECD countries are better than the OECD countries in bads e±ciency.
On the input side, a similar result is found for the labor component; the mean of the
rank order correlation coe±cients of the OECD countries is 10.11, and that of the non-
OECD countries 28.89. Again, the labor e±ciency of OECD countries is signi¯cantly
better than the non-OECD countries' (p-value < 0:001). As for the capital component,
the performance of the OECD countries is worse than that of the non-OECD coun-
tries, since the Mann–Whitney test indicates that their di®erences are signi¯cant.
However, a di®erent result is found for the energy e±ciency for the OECD and the
non-OECD countries; the mean of the rank order correlation coe±cients of the OECD
countries is 16.87, and that of the non-OECD countries is 22.13. This implies that the
OECD countries are slightly better than the non-OECD countries in energy e±ciency,
although the result of the Mann–Whitney test is not signi¯cant.

One interesting discrepancy in ine±ciency between the OECD and non-OECD
countries is the annual trend pattern of e±ciency variation. Figure 1 shows the
relative di®erence in levels and the relative comparability in temporal patterns for
the overall ine±ciency scores and their components during the sample period for the
OECD and non-OECD countries. The average level of overall ine±ciency scores over
the 19-year period was calculated to be 0.286 and 0.359 for OECD and non-OECD
countries, respectively. Ine±ciency was at its lowest in 2007 for OECD countries and
in 2009 for non-OECD countries, when the level of ine±ciency was 0.250 and 0.317,
respectively. These are followed by 0.257 in 2008 for OECD countries and 0.323 in
2007 for non-OECD countries. Figure 1(a) shows there is a relatively wide gap in the
technical e±ciencies between the OECD and non-OECD countries of about 10%
persistently during the sample years.
Figures 1(a)–1(g) show the temporal patterns of the overall ine±ciency and its
decompositions. On average, Figs. 1(b) and 1(e) indicate that the goods ine±ciency and
labor ine±ciency components for OECD countries are considerably below those for their
non-OECD counterparts at the beginning of the study period, and that this pattern was
rather stable over the sample period. The goods (labor) ine±ciency averages to 0.179
(0.326) over the 19-year period for the OECD countries, which is smaller than 0.457
(0.505) for the non-OECD countries. Oppositely, Figs. 1(c) and 1(f) show OECD
countries had annual bads and capital ine±ciencies which were above those for the non-
OECD countries except for 1993, 0.437 versus 0.330 and 0.122 versus 0.068 of capital.
Fig. 1(g), however, indicates that the di®erences of the annual ine±ciencies of energy
between the OECD countries and non-OECD countries are not consistent over time.
Over 1992–2000, the OECD countries and non-OECD countries have almost the same
performance in energy e±ciency. Then the OECD countries perform better than non-
OECD countries over 2001–2007 and but worse afterwards. Additionally, it is worth
noting that the ine±ciencies for both are increasing quickly after 2002.
Overall, capital and bads e±ciencies of non-OECD countries are better than
OECD countries' during the study period. For energy, these two kinds of countries
perform closely. In particular, the energy ine±ciencies of OECD countries become
(a) Average overall ine±ciency scores: (b) Average goods ine±ciency scores:
OECD V.S. nonOECD countries. OECD V.S. nonOECD countries.
(c) Average bads ine±ciency scores: (d) Average inputs ine±ciency scores:
(e) Average labor ine±ciency scores: (f) Average capital ine±ciency scores:
(g) Average energy ine±ciency scores:

OECD V.S. nonOECD countries.
Fig. 1. Temporal patterns of the overall ine±ciency and its decompositions during 1992–2010, OECD
and non-OECD countries.
larger than those of non-OECD countries after 2008. The results show that the bads
ine±ciencies for the OECD countries are generally larger than those of the non-
OECD countries. The other results are opposite, i.e., the OECD countries generally
perform better than the non-OECD countries in overall, goods and one (labor) of the
three input items' e±ciencies. In particular, the labor ine±ciency of the OECD
countries is much smaller than that of the non-OECD countries after 1996. The high
ine±ciency scores of bads and energy ine±ciencies for the OECD countries show the
importance of policies that focus on improving pollution abatement, and also dem-
onstrate that the energy use of the OECD countries is not e±cient. The results imply
that the advantage of the OECD countries' higher technology level results in lower
labor and goods ine±ciencies, while the OECD countries' people produce too much
CO2 in their daily lives. More importantly, the increasing ine±ciencies of energy
indicate that the energy resources wasted by both groups of countries are growing.
Therefore, research and development policies focused on emission-saving technolo-
gies might not be enough if best practices are not later applied in full to production
processes and generally to the management of air pollutants by OECD countries,
allowing them to approach the environmental technological frontier.57
Based on this discussion, we conclude that membership in the OECD is one of the
most important factors in°uencing the components' ine±ciencies. A possible ex-
planation for this variability is that the income levels for these two groups of
countries are di®erent. Therefore, in order to capture more insights, we further
separate all of the countries into four income levels and investigate various compo-
nents' ine±ciencies in all income level countries.
4.4. Di®erences in income levels

Table 5 shows that the overall ine±ciency in the high income countries is the second
lowest and only higher than that in low income countries. Because there are only 12
countries in low income countries group, this implies the high income countries are in
general more e±cient than those at the other income levels and have limited room for
improvement in their e±ciencies. This result is similar to that of Refs. 51 and 52. To
some extent, this may be because higher income countries are more likely to employ
advanced technology with more e±cient output production which is a potential
source of their competitive advantage. Similar to Ref. 51, the upper middle income
countries perform worse than the lower middle and low income countries because of
higher CO2 emissions and lower good output production in these countries.
As can be seen in Table 5, the e®ect of outputs on overall ine±ciency is generally
more serious than that of inputs (except labor) of each income level country. Also,
the bads' impact on ine±ciency is more demanding than that of the goods for high
income countries, whereas it is opposite for the other countries. This indicates that
improving e±ciency of bads is more e®ective than improving e±ciency of goods and
inputs for increasing overall e±ciency for the high income countries, while for the
other countries, improving e±ciency of goods and labor are more important. In
Table 5. Components of the overall ine±ciency by income levels.
1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 AVG. AVG.
92-10 01-10
116 countries
overall 0.350 0.363 0.372 0.378 0.375 0.363 0.362 0.334 0.351 0.332 0.358 0.321 0.321 0.312 0.305 0.302 0.305 0.301 0.309 0.338 0.317
goods 0.415 0.508 0.461 0.535 0.465 0.458 0.463 0.404 0.462 0.393 0.417 0.334 0.339 0.278 0.260 0.262 0.243 0.225 0.213 0.376 0.297
bads 0.383 0.342 0.395 0.359 0.381 0.361 0.378 0.343 0.349 0.332 0.379 0.357 0.326 0.340 0.350 0.354 0.366 0.365 0.400 0.361 0.357
inputs 0.252 0.240 0.259 0.240 0.278 0.269 0.245 0.254 0.241 0.271 0.278 0.272 0.299 0.316 0.306 0.290 0.304 0.311 0.314 0.276 0.296
labor 0.428 0.366 0.444 0.354 0.474 0.468 0.445 0.460 0.462 0.471 0.471 0.479 0.495 0.536 0.478 0.449 0.426 0.450 0.441 0.452 0.470
capital 0.092 0.150 0.100 0.126 0.106 0.093 0.073 0.091 0.060 0.120 0.104 0.080 0.120 0.109 0.047 0.037 0.025 0.026 0.029 0.084 0.070
energy 0.237 0.205 0.232 0.239 0.254 0.246 0.216 0.212 0.201 0.221 0.258 0.257 0.284 0.303 0.393 0.384 0.463 0.458 0.473 0.291 0.349
High income countries
overall 0.309 0.304 0.328 0.302 0.325 0.321 0.306 0.283 0.279 0.283 0.295 0.274 0.264 0.266 0.253 0.244 0.248 0.258 0.263 0.285 0.265
goods 0.235 0.246 0.276 0.263 0.300 0.301 0.266 0.227 0.263 0.244 0.269 0.206 0.189 0.163 0.177 0.153 0.103 0.104 0.119 0.216 0.173
bads 0.486 0.468 0.504 0.435 0.468 0.457 0.443 0.405 0.387 0.395 0.413 0.396 0.382 0.397 0.376 0.388 0.419 0.430 0.436 0.426 0.403
inputs 0.207 0.198 0.205 0.208 0.207 0.205 0.210 0.216 0.185 0.210 0.203 0.220 0.221 0.239 0.205 0.192 0.224 0.242 0.235 0.212 0.219
labor 0.302 0.280 0.282 0.283 0.273 0.271 0.282 0.293 0.252 0.284 0.268 0.293 0.284 0.304 0.200 0.194 0.198 0.215 0.201 0.261 0.244
capital 0.114 0.113 0.117 0.119 0.119 0.129 0.140 0.156 0.115 0.148 0.120 0.149 0.152 0.169 0.076 0.060 0.038 0.049 0.039 0.112 0.100
energy 0.205 0.201 0.217 0.221 0.230 0.215 0.207 0.200 0.189 0.199 0.223 0.219 0.226 0.244 0.338 0.322 0.437 0.461 0.466 0.264 0.313
Upper middle income countries
overall 0.407 0.420 0.417 0.434 0.419 0.418 0.400 0.381 0.417 0.368 0.424 0.373 0.389 0.382 0.374 0.368 0.359 0.365 0.365 0.394 0.377
goods 0.579 0.699 0.598 0.681 0.608 0.627 0.594 0.550 0.690 0.539 0.628 0.528 0.545 0.484 0.412 0.368 0.319 0.319 0.316 0.531 0.446
bads 0.401 0.353 0.426 0.431 0.403 0.380 0.400 0.362 0.349 0.305 0.361 0.343 0.326 0.349 0.386 0.407 0.433 0.432 0.445 0.384 0.379
inputs 0.242 0.207 0.228 0.190 0.247 0.247 0.207 0.231 0.212 0.260 0.284 0.249 0.297 0.312 0.323 0.327 0.324 0.344 0.333 0.267 0.305
labor 0.438 0.348 0.420 0.252 0.468 0.470 0.439 0.485 0.461 0.539 0.592 0.514 0.562 0.589 0.580 0.557 0.497 0.548 0.523 0.488 0.550
capital 0.093 0.082 0.081 0.125 0.048 0.052 0.059 0.073 0.050 0.114 0.104 0.064 0.096 0.065 0.029 0.026 0.016 0.021 0.013 0.064 0.055
energy 0.194 0.191 0.182 0.193 0.224 0.220 0.125 0.135 0.124 0.126 0.157 0.170 0.232 0.284 0.360 0.398 0.457 0.462 0.464 0.247 0.311
25
26
Table 5. (Continued )
P. C. Chen et al.
1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 AVG. AVG.
92-10 01-10
Lower middle income countries

overall 0.368 0.394 0.420 0.455 0.417 0.396 0.400 0.359 0.385 0.375 0.377 0.341 0.355 0.326 0.321 0.331 0.348 0.315 0.322 0.369 0.341
goods 0.538 0.661 0.628 0.818 0.572 0.573 0.581 0.469 0.490 0.515 0.450 0.328 0.394 0.271 0.242 0.340 0.398 0.300 0.215 0.462 0.345
bads 0.270 0.220 0.307 0.266 0.325 0.291 0.327 0.294 0.345 0.295 0.364 0.371 0.303 0.325 0.326 0.303 0.281 0.278 0.361 0.308 0.321
inputs 0.297 0.301 0.326 0.282 0.353 0.326 0.292 0.316 0.319 0.314 0.317 0.323 0.368 0.382 0.394 0.350 0.365 0.368 0.391 0.336 0.357
labor 0.555 0.471 0.661 0.498 0.727 0.701 0.633 0.668 0.694 0.620 0.604 0.642 0.680 0.757 0.719 0.622 0.605 0.636 0.623 0.638 0.651
capital 0.048 0.173 0.059 0.102 0.051 0.036 0.010 0.020 0.012 0.074 0.051 0.023 0.091 0.058 0.018 0.019 0.013 0.001 0.046 0.048 0.039
energy 0.287 0.261 0.259 0.247 0.281 0.241 0.234 0.259 0.252 0.248 0.297 0.305 0.335 0.332 0.447 0.410 0.476 0.467 0.504 0.323 0.382
Lower income countries
overall 0.289 0.343 0.287 0.320 0.331 0.278 0.367 0.323 0.342 0.310 0.348 0.296 0.256 0.237 0.261 0.254 0.253 0.232 0.283 0.295 0.273
goods 0.311 0.556 0.365 0.471 0.404 0.283 0.534 0.470 0.447 0.239 0.262 0.245 0.157 0.110 0.154 0.173 0.187 0.226 0.248 0.308 0.200
bads 0.209 0.119 0.104 0.080 0.129 0.108 0.190 0.175 0.218 0.264 0.345 0.230 0.172 0.145 0.203 0.186 0.168 0.130 0.228 0.179 0.207
inputs 0.348 0.354 0.393 0.409 0.459 0.441 0.377 0.325 0.361 0.426 0.438 0.413 0.440 0.458 0.427 0.403 0.405 0.340 0.373 0.399 0.412
labor 0.572 0.495 0.623 0.599 0.661 0.669 0.642 0.532 0.719 0.618 0.560 0.692 0.649 0.731 0.656 0.673 0.646 0.602 0.665 0.632 0.649
capital 0.104 0.434 0.184 0.210 0.343 0.212 0.015 0.062 0.000 0.141 0.158 0.000 0.140 0.131 0.055 0.021 0.030 0.007 0.000 0.118 0.068
energy 0.367 0.132 0.371 0.417 0.373 0.444 0.475 0.380 0.363 0.519 0.597 0.547 0.530 0.512 0.568 0.515 0.540 0.412 0.454 0.448 0.519
addition, we can ¯nd that the ine±ciency scores of bads increase from 0.179 to 0.207
on average for the low income countries, to 0.384 and 0.379 for the upper middle
income countries, then further increase to 0.426 and 0.403 for the high income
countries over 1992–2010 and 2001–2010, respectively. In low income countries,
improving e±ciency of labor is more e®ective for increasing overall e±ciency. If we
regard the ine±ciency score of bads as the environmental performance measure, this
tendency suggests that there does not exist an inverse-U-shaped environmental
Kuznets curve relationship in our sample data. This is because the performance
deteriorates along with the increase of the income. Chiu et al.51 also found that
higher average per capita annual income is accompanied by higher CO2 emissions.
More evidence is of course needed to support the existence of this relationship in our
case. However, that is beyond the scope of our study.

In contrast to the impact of outputs on overall ine±ciency, the relative magnitude

of the impact of each input item on overall ine±ciency between each income level is
shown to have dissimilar patterns. For example, the labor ine±ciency of the lower
middle income countries is greater than that of the other income level countries,
while the lower middle income countries perform the best in capital e±ciency. When
we compare energy e±ciency, apart from the high income countries, there seems to
be a tendency for lower income countries to have higher energy ine±ciency. That is,
the low income countries have the highest energy ine±ciency. Moreover, if we focus
on the performance of the upper and lower middle income countries, we ¯nd that
even though they have the lower capital ine±ciency than that of high income
countries, its positive impact on the overall ine±ciency is o®set by both the labor and
energy ine±ciency. This result indicates that ine±ciencies of these two income level
countries are more related to poor technical management of production factors, i.e.,
overuse of labor and energy, and improving the labor and energy e±ciency for these
countries is important. Overall, the labor and energy ine±ciency scores are relatively
high in the input ine±ciency scores on average. The capital ine±ciency scores are
marginally lower than the labor ine±ciency scores, and thus capital e±ciency con-
tributes to a very large extent to the input utilization performance.
The WRDDM analysis o®ers both the aggregate ine±ciency measure and its
components. Therefore our basic hypothesis is that the aggregate ine±ciency mea-
sure and its components themselves are not di®erent among these groups of countries
by income level. The ¯rst analysis concentrates on testing whether the overall inef-
¯ciency of the di®erent income level countries is not di®erent from others. The second
hypothesis addresses whether the in°uence of each input/output ine±ciency on
overall ine±ciency is not di®erent at a given income level countries.
Hypothesis 1. The overall ine±ciencies of di®erent income level countries are not
di®erent from each other.
This hypothesis assumes that di®erences of overall ine±ciency between di®erent

income level countries are not obvious, whereas we expect the overall ine±ciencies to
Table 6. Kruskal wallis test: Hypothesis 1.
Income level Rank order correlation coe±cient p-value
High income level 18.470 <0.001

Upper middle income level 61.240 –
Lower middle income level 51.260 –
Low income level 23.030 –
vary distinctly between di®erent income level countries. A Kruskal Wallis test was
used to test this hypothesis, and the results displayed in Table 6 show a signi¯cant
di®erence among di®erent income level countries (p-value < 0:001), implying that
the null hypothesis of equal overall ine±ciency has to be rejected. This ¯nding
con¯rms our expectation that the overall ine±ciency was signi¯cantly di®erent
among di®erent income level countries and implies that the overall performance of a
country would be a®ected by its income level.
Hypothesis 2. The in°uence of all components' ine±ciencies on overall ine±ciency

is the same in a given income level group of countries.
This hypothesis concerns the impacts of the various input and output items on the
overall ine±ciency within each of the income levels of countries. A Kendall's W
statistic is used to measure the extent to which the ranking of all the input and
output items' ine±ciencies resemble the ranking of each other items'. The results for
each of income levels' countries are displayed in Table 7. As can be seen in the ¯rst
row of Table 7, the rank order correlation coe±cients of each component ine±ciency
score are signi¯cantly di®erent from each other (goods 3.32, bads 5.84, labor 4.32,
capital 1.05 and energy 3.58), resulting in a very small p-value of the Kendall's W
test. This implies that the null hypothesis of equal ine±ciency scores for the high
income countries has to be rejected. For further exploration, the same outcome can
be found for the remainder of the three income levels. These results provide evidence
to show that di®erent components' ine±ciencies lead to di®erent degrees of impact
on the overall ine±ciency and, hence, indicate that separating overall ine±ciency
into di®erent parts is important to provide management insight for the policy de-
cision making of each country.
Table 7. Kendal's W test: Hypothesis 2.
Income level Rank order correlation coe±cient p-value

Goods Bads Inputs Labor Capital Energy
High income level 3.320 5.840 2.890 4.320 1.050 3.580 <0.001
Upper middle income level 5.030 4.210 2.840 5.240 1.000 2.680 <0.001
Lower middle income level 4.320 2.950 3.790 5.890 1.000 3.050 <0.001
Low income level 3.260 2.050 3.840 5.890 1.420 4.530 <0.001
Fig. 2. Comparison of average goods, bads and input ine±ciencies during 1992–2010.
To elaborate more details on each component of the production performance of the

countries, we plot in Fig. 2 the mean value of the goods, bads and input ine±ciency
indices computed over the 116 countries for the period 1992–2010. The mean goods
ine±ciency index shows a progressive improvement in good output production e±-
ciency from 1995. Despite the di®erences in overall means, we observe that the changes
in the bads and input ine±ciency are relatively stable during the study period. The
decomposition of overall ine±ciency index into the goods, bads and input ine±ciency
indices is also successful in capturing each component of the overall ine±ciency of
countries. The result shows that the bads and input e±ciencies are better than goods
e±ciency globally before 2004, while it appear to be opposite after that.
5. Conclusions
As environmental concerns become increasingly pronounced in relation to global
commons, environmental issues are being treated more and more as international
matters. The accurate assessment of environmental conditions is essential. In
this study, the WRDDM model is proposed and applied to a 116-country data
set from 1992 to 2010 to examine the input, desirable output, and undesirable output
items' e±ciencies in order to comprehensively gauge overall production e±ciency.
From the performance evaluation aspect, one of the advantages of the WRDDM over
the traditional directional distance function model is that it directly incorporates
weights to consider the appropriate relationship among input and output items, while
the traditional model weights them equally. In addition, the proposed model accounts for
all the slacks for the inputs, desirable outputs and undesirable outputs, and thus is able to
provide more accurate performance evaluation results. From the application aspect, this
model allows a DMU to be able to adjust all the amounts of inputs, desirable outputs and
undesirable outputs (subject only to the constraints imposed by the production tech-
nology) and hence the technical e±ciencies for each input, desirable output and unde-
sirable output may be rated di®erently and can help us to identify which resource uses or
production of outputs (including goods and bads) need to be improved most.
The empirical results reveal that the separation of overall ine±ciency into dif-
ferent parts is important to provide management insight for the policy decision
making of each country. For example, it is shown that the OECD countries perform
better than the non-OECD countries in overall, goods and one out of three input
items' e±ciencies. On the other hand, the bads ine±ciency for the OECD countries is
larger than those of the non-OECD countries. In particular, it is the largest ine±-
ciency source for the OECD countries. These results imply that the abatement of the
bad output of OECD countries would be more important than that of the non-
OECD countries. While this result does not deny the importance of e®orts to
improve e±ciency in the other components, the most e®ective way to improve the
overall production e±ciency of the non-OECD countries is to focus on the im-
provement in the goods and labor e±ciencies. Accordingly, for OECD countries,
policy measures aimed at enhancing environmental e±ciency are recommended to do

more e®ort on controlling CO2 emission to encourage catching-up to approach the
environmental technological frontier. For the non-OECD countries, more e±cient
labor usage is needed to expand their good outputs. Promoting labor skills and
managerial capabilities might be welcome means of improving their performances.
Future works in this burgeoning ¯eld of research should include: Firstly, one
important quali¯cation of using the weighted Russell directional distance model is
the necessity of determining the weight schemes prior to measurement. The
weighting schemes based on decision makers' preferences, expert experiences or a
multiple criteria decision-making (MCDM) approach (For example, the analytic
hierarchy process to derive a reliable priority vector from a pairwise comparison
matrix,59 or an MCDM-based approach to rank a selection of the weight
schemes,60,61 etc.) may be alternatives for further investigation. Secondly, modeling
pollution-generating technologies as a composition of an intended-production tech-
nology and a residual-generation technology as proposed by Murty et al.21 to best
capture correct trade-o®s in production is an essential work to be resolved in the
future studies if the abatement e®orts data are available. Thirdly, further study is
needed to examine technical e±ciency and productivity growth in the OECD and
non-OECD countries over the study periods. Lastly, there is need for extending the
present paper to develop a meta-frontier weighted Russell directional distance
function by considering the group heterogeneity of all countries in the world.
Appendix A
Theoerm 3.2. The projected DMU o on the frontier is WRDDM overall e±cient.
Proof. Let an optimal solution of model (6) be z o , s mo
þ
, s jo
and s no . We de¯ne the
projection of DMU o as follows:
X
K
y~ mo ¼ z k y mk ¼ y mo þ s mo
þ
ðm ¼ 1; . . . ; M Þ;
k¼1
X
K
b~jo ¼ z k b jk ¼ b jo s jo

ðj ¼ 1; . . . ; J Þ;
k¼1
X
K
x~ no ¼ z k x nk ¼ x no s no

ðn ¼ 1; . . . ; N Þ:
k¼1
Then, re-estimate the overall e±ciency of the projected DMU. Let an optimal
solution of the projected DMU be zao , saþ a a
mo , s jo and s no . We have:
X
K
y~ mo ¼ zak y mk saþ
mo ðm ¼ 1; . . . ; M Þ;
k¼1
X
K
b~jo ¼ zak b jk þ sa
jo ðj ¼ 1; . . . ; J Þ;
k¼1
X
K
x~ no ¼ zak x nk þ sa
no ðn ¼ 1; . . . ; N Þ:
k¼1
þ
Replacing y~ mo ¼ y mo þ s mo (m ¼ 1; . . . ; M Þ, b~jo ¼ b jo s jo ðj ¼ 1; . . . ; J Þ, and

x~ no ¼ x no s no ðn ¼ 1; . . . ; N Þ, we have:
X
K
y mo ¼ zak y mk saþ þ
mo s mo ðm ¼ 1; . . . ; M Þ;
k¼1
X
K
b jo ¼ zak b jk þ sa
jo þ s jo ðj ¼ 1; . . . ; J Þ;
k¼1
X
K
x no ¼ zak x nk þ sa
no þ s no ðn ¼ 1; . . . ; N Þ:
k¼1
Corresponding to this expression we have the overall ine±ciency,

o
~ o þ w b~ þ w x ~o ðk ¼ 1; . . . ; K Þ;
R ¼ wy
~ k
where
XM
$ ym ðsaþ þ
mo þ s mo Þ
~ o ¼
;
m¼1
gym
o
XJ
$ bj ðsa
jo þ s jo Þ
~ ¼ ;
j¼1
gbj
XN
$ xn ðsa
no þ s no Þ
~o ¼ :
n¼1
gxn
If any element of fsa a aþ
no g; fs jo g and fs yo g is positive, then it holds that R >
~ o o
R . This
a a
contradicts the optimality of R . Thus, we have s no ¼ 0; s jo ¼ 0 and
o
saþ
mo ¼ 0ð8 n; j; mÞ. Hence, the projected DMU is overall-e±cient.
Appendix B
Theorem 3.4. The inverse variance WRDDM is units invariant.
Proof. Let the sample standard deviations of the desirable output y m , undesirable
output b j and input x n be ym , bj and xn , respectively and the corresponding weights
of their slacks be 1y , 1b and 1xn .
m j
Consider rescaling desirable output y m , undesirable output b j and input x n by

multiplying by the scalar m , j and n , respectively. Then the rescale sample
standard deviations of the desirable output m y m , undesirable output j b j and input
n xnbecome m ym , j bj and n xn and the weights, $ ym , $ bj and $ xn in the objective
function in model (6) become m1 y , 1 b and n1 x , respectively. This implies that
m m j n
those rescale slacks corresponding to desirable output y m , undesirable output b j and

input x n are normalized by their standard deviations. Thus, the value of the objective
function (6) is not di®erent from the original oneq Furthermore, the translated
restrictions are equivalent to the original restrictions. The value of the objective is
also not a®ected, thus, the standard deviation of WRDDM is units invariant.
Appendix C
0 0 0
Proposition 3.1.By settingg ¼ ðg x ; g y ; g b Þ ¼ ðx k ; y k ; bk Þ, the corresponding
weighted slacks-based measure and enhanced Russell graph measure yield informa-
tion on performance equivalent to the WRDDM.
Proof. See Appendix C
0 0 0
When g ¼ ðg x ; g y ; g b Þ ¼ ðx k ; y k ; bk Þ, the programming problem of (6) can
be re-expressed as follows:
0 0 0 0
DR ðx k ; y k ; bk ; gÞ ¼ Rk
!
XN
$ xn s nk 0
XM þ
$ ym s mk 0
XJ
$ bj s jk 0
¼ max w x þ wy þ wb
n¼1
x nk 0 m¼1
y mk 0 j¼1
b jk 0
q That is
XM þ
$ ym s mk 0
XJ
$ bj s jk
0 XN
$ xn s nk 0
wy þ wb þ wx
m¼1
g y m j¼1
g b j n¼1
g x n
X M
1 m s mk þ
0
XJ
1
m s jk 0
X N
1
n s nk 0
¼ wy þ wb þ wx
m¼1 m m
y
gym j¼1 m j
b gbj n¼1 n
x
n
gxn
X M y þ
$ m s mk 0 X J
$ b
s
j jk 0
X N
$ xn s nk 0
¼ wy þ wb þ wx :
m¼1
gym j¼1
gbj n¼1
gxn
X
K
þ
s:t: z k y mk ¼ y mk 0 þ s mk 0 ; m ¼ 1; . . . ; M ;
k¼1
X
K

z k b jk ¼ b jk 0 s jk 0; j ¼ 1; . . . ; J ;
k¼1
XK

z k x nk ¼ x nk 0 s nk 0 ; n ¼ 1; . . . ; N ;
k¼1
þ
s mk 0 ; s jk 0 ; s nk 0 0 8 m; j; n;
z k 0; k ¼ 1; . . . ; K :
ðC:1Þ
Since the value of the objective in (C.1) is greater than or equal to zero, we have
X
N
s nk 0
XM
sþ 0 XJ
s jk 0
wx $ xn þ wy $ ym mk þ w b $ bj 0
x nk 0 y mk 0 b jk 0
n¼1 m¼1 j¼1
! !
XM
s þ
0
XN
s
0
X
J
s jk 0
() 2 þ w y $ ym mk 1 w x $ xn nk þ 1 wb $ bj
m¼1
y mk 0 n¼1
x nk 0 j¼1
b jk 0
PN
P s0

s nk
1 wx n¼1 $ n
x 0
x nk 0 þ 1 w b Jj¼1 $ bj b jkjk 0
() PM 1: ðC:2Þ
sþ
2 þ wy m¼1 $ ym y mk
mk 0
0
Therefore, the corresponding weighted slacks-based measure which is equivalent to

our WRDDM would be
P
PJ
x s nk 0 b s jk 0
1 wx N n¼1 $ n x nk 0 þ 1 w b j¼1 $ j b jk 0
0
min k ¼ PM þ
y s mk 0
2 þ w y m¼1 $ m y mk 0
X
K
þ
s:t: z k y mk ¼ y mk 0 þ s mk 0; m ¼ 1; . . . ; M ;
k¼1
X
K

z k b jk ¼ b jk 0 s jk 0; j ¼ 1; . . . ; J ; ðC:3Þ
k¼1
X
K

z k x nk ¼ x nk 0 s nk 0; n ¼ 1; . . . ; N ;
k¼1
þ
s mk 0 ; s jk 0 ; s nk 0 0 8 m; j; n;
z k 0; k ¼ 1; . . . ; K :
It is noted that the objective of (6) is to maximize the weighted sum of associated
slacks, while the input slacks in the numerator and output slacks in the denominator
of (C.3) are minus and positive, respectively, therefore, (C.3) become a minimization
0
scheme. As stated in Eq. (C.2), the value of k in the objective function of formu-
lation (C.3) is less than or equal to one, and it provides us an aggregate measure of
0
e±ciency performance in which the numerator of k corresponds to the sum of the
mean reduction rate of inputs and undesirable output; and similarly, the denomi-
nator corresponds to the mean expansion rate of desirable output plus one. In ad-
0
dition, a larger value of k indicates that k 0 performs better. If there are no slacks in
0
all outputs, then k will attain the value of unity, indicating that k 0 is a technically
e±cient ¯rm.
Furthermore, if we set each weight to be equal (i.e., w y ¼ w b ¼ w x Þ in (C.1), the
corresponding objective function in (C.2) will become
X
N
s nk XM
sþ 0 XJ
s jk
0 0
wx $ xn þ wy $ ym mk þ w b $ bj 0
n¼1
x nk 0 m¼1
y mk 0
j¼1
b jk 0
X
N
s nk 0
X
M
sþ 0 X J
s jk 0
() $ xn þ $ ym mk þ $ bj 0
n¼1
x nk 0 m¼1 y mk 0 j¼1
b jk 0
! !
X
M þ XN XJ
s mk 0 x s nk 0 b s jk 0
2þ $ ym 1 $n þ 1 $j
m¼1
y mk 0 n¼1
x nk 0 j¼1
b jk 0
P
PJ
x s nk 0 b s jk 0
1 N n¼1 $ n x nk 0 þ 1 j¼1 $ j b jk 0
() P þ 1:
y s mk 0
2þ M m¼1 $ m y mk 0
The objective function in (C.3) becomes

PN
PJ
x s nk 0 b s jk 0
0
1 n¼1 $ n x 0
þ 1 j¼1 $ j b jk 0
min k ¼ : ðC:4Þ
nk
PM þ
y s mk 0
2 þ m¼1 $ m y mk 0
Then, the weighted slacks-based measure can further be shown to be equivalent to

the following enhanced Russell graph measure by means of the following change of
variables:
þ
y y mk 0 þ s mk 0 sþ 0
mk 0 ¼ ¼ 1 þ mk ;
y mk 0 y mk 0

b jk 0 s jk 0 s jk 0
b
jk 0 ¼ ¼1 ; ðC:5Þ
b jk 0 b jk 0

x nk 0 s nk 0 s 0
x
nk 0 ¼ ¼ 1 nk :
b nk 0 x nk 0
Table A.1. Numerical test of uniqueness.
DMU Good Bad Input 1 Input 2 R b

y x 1 2 max
y min
y max b max x min x max 1 min 1 max 2 Min 2
b min
1 2 1 1 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2 2 1 2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3 2 1 2 2 0.083 0 0 0.250 0 0.500 0 0 0 0 0.250 0.250 0.500 0 0.500 0
4 2 2 4 4 0.417 1 0 0.250 0 0.500 1 1 0 0 0.250 0.250 0.500 0 0.500 0
5 2 1 6 6 0.250 0 0 0.750 0.667 0.833 0 0 0 0 0.750 0.750 0.833 0.667 0.833 0.667
35
PM y PJ
m¼1 $ m ¼ 1, j¼1 $ j ¼ 1 and
b
We derive the alternative expression. Since
PN
n¼1 $ n ¼ 1, the formulation (C.5) can be easily re-expressed as follows:
x
PN PJ
n¼1 $ n nk 0 þ j¼1 $ j jk 0
x x b b
0
min k ¼ PM y y
1 þ m¼1 $ m mk 0
X
K
y
s:t: z k y mk mk 0 y mk 0 m ¼ 1; . . . ; M ;
k¼1
X
K
z k b jk ¼ b
jk 0 b jk 0 j ¼ 1; . . . ; J ; ðC:6Þ
k¼1
X
K
z k x nk x
nk 0 x nk 0 n ¼ 1; . . . ; N ;
k¼1
y
mk 0 1; b
jk 0 1; x
nk 0 1 8 m; j; n
zk 0 k ¼ 1; . . . ; K :
The objective function in (C.6) gives
PN PJ
n¼1 $ n nk 0 þ j¼1 $ j
x x b b
jk 0
0 PM
1 þ m¼1 $ ym ymk 0
! !1
X
N X
J X
M
y
¼ $ xn x
nk 0 þ $ bj b
jk 0 1þ $ ym mk 0
n¼1 j¼1 m¼1
¼ ð k 0 þ k 0 Þð1 þ k 0 Þ1 1:
x b y
ðC:7Þ
This expression (C.7) is interpreted as the ratios of the weighted average of

input and undesirable output e±ciency scores to the average desirable output
e±ciency scores that represent the e±ciency measure for each of the DMU k
(k ¼ 1; . . . ; K ). This measure is also able to provide us an overall e±ciency mea-
sure which simultaneously considers resource utilization, good output production
0
and bad output reduction. Moreover, since k contains the performance infor-
mation about each y m , b j and x n of DMU k 0 , we can easily ¯nd the percentage by
which each y m , b j and x n ought to improve from this programming problem, and
hence it allows us to give a more comprehensive picture on the performance of the
DMU. Nevertheless, the formulation (C.6) is nonlinear and its solution is not
easily obtained. We may follow the methods developed by Pastor et al.8 or Ray
and Jeon58 to transform it into a linear programming problem when we execute
the solution calculation process.
Appendix D
In this example, we generate the data for ¯ve DMUs with one good output, one bad
output and two inputs. We use model (5) to calculate each DMU's overall
ine±ciency score, and the ine±ciency scores of goods, bads and inputs. Because there
are two inputs, we also have the ine±ciency scores for each input. In order to ¯nd its
range of variation, we follow the method in Ref. 37 to adopt two-step procedures to
solve max (min) y, b and x , while keeping R at the optimum respectively in the
¯rst step; then in turn to solve max (min) of each m , j and n , while keeping R as
well as b or x (depending on whether we solve the bounds for m , j or n ,
y,
respectively) at the optimum in the second step.
Accordingly, it can be found the projected desirable output and undesirable
output are uniquely determined, while the projected inputs are not, since their max
and min values are not equal, respectively. Tone and Tsutsui37 indicated that if the
relative gap (max–min)/max is comparatively small (e.g., < 0:5%), we can practi-
cally neglect the multiple optima problem. In this example, it is shown that the
relative gaps for the two inputs of DMU 3, DMU 4 and DMU 5 are too large (100%,
100% and 19.9%, respectively) to be neglected.
References
1. L. M. Seiford and J. Zhu, An investigation of returns to scale in data envelopment
analysis, Omega 27 (1999) 1–11.
2. K. Tone, Variations on the theme of slacks-based measure of e±ciency in DEA, European
Journal of Operational Research 200 (2010) 901–907.
3. A. Charnes, W. W. Cooper and E. Rhodes, Measuring the e±ciency of decision making
units, European Journal of Operational Research 2 (1978) 429–444.
4. R. D. Banker, A. Charnes and W. W. Cooper, Some models for estimating technical and scale
ine±ciencies in data envelopment analysis, Management Science 30(9) (1984) 1078–1092.
5. H. Fukuyama and W. L. Weber, A directional slacks-based measure of technical ine±-
ciency, Socio-Economic Planing Sciences 43 (2009) 274–287.
6. R. Färe and C. A. Knox Lovell, Measuring the technical e±ciency of production, Journal
of Economic Theory 19(1) (1978) 150–162.
7. R. Färe, S. Grosskopf and C. A. Knox Lovell, The Measurement of E±ciency of Pro-
duction (Kluwer Academic Publishers, Boston, 1985).
8. J. T. Pastor, J. L. Ruiz and I. Sirvent, An enhanced DEA Russell graph e±ciency
measure, European Journal of Operational Research 115 (1999) 596–607.
9. A. Charnes, W. W. Cooper, B. Golany, L. M. Seiford and J. Stutz, Foundations of data
envelopment analysis for Pareto-Koopmans e±cient empirical production functions,
Journal of Econometrics 30 (1985) 91–107.
10. K. Tone, A slacks-based measure of e±ciency in data envelopment analysis, European
11. W. W. Cooper, L. M. Seiford and K. Tone, Some models and measures for evaluating
performance with DEA: Past accomplishments and future prospects, Journal of Pro-
ductivity Analysis 28 (2007) 151–163.
12. R. Färe and S. Grosskopf, Directional distance functions and slacks-based measures of
e±ciency, European Journal of Operational Research 200 (2010) 320–322.
13. V. N. Hoang and T. Coelli, Measurement of agricultural total factor productivity growth
incorporating environmental factors: A nutrients balance approach, Journal of Envi-
ronmental Economics and Management 62(3) (2011) 462–474.
14. N. Zhang and Y. Choi, Total-factor carbon emission performance of fossil fuel power
plants in China: A metafrontier non-radial Malmquist index analysis, Energy Economics
40 (2013) 549–559.
15. Y. H. Chung, R. Färe and S. Grosskopf, Productivity and undesirable outputs: A di-
rectional distance function approach, Journal of Environmental Management 51 (1997)
229–240.
16. R. Färe, S. Grosskopf, C. A. K. Lovell and C. Pasurka, Multilateral productivity com-
parisons when some outputs are undesirable: A nonparametric approach, Review of
Economics and Statistics 71 (1989) 90–98.
17. L. M. Seiford and J. Zhu, Modeling undesirable outputs in e±ciency evaluation, European
18. A. Hailu and T. S. Veeman, Nonparametric productivity analysis with undesirable out-
puts: An application to the Canadian pulp and paper industry, American Journal of
Agricultural Economics 83 (2001) 605–616.

19. B. K. Sahoo, M. Luptacik and B. Mahlberg, Alternative measures of environmental
technology structure in DEA: An application, European Journal of Operational Research

215 (2011) 750–762.
20. C. M. Chen and M. Delmas, Measuring eco-ine±ciency: A new frontier approach,
Operations Research 60(5) (2012) 1064–1079.
21. S. Murty, R. R. Russell and S. B. Levkov, On modeling pollution-generating technologies,
Journal of Environmental Economics and Management 64 (2012) 117–135.
22. H. Leleu, Shadow pricing of undesirable outputs in nonparametric analysis, European
23. P. B. Zhou, W. Ang and K. L. Poh, Slacks-based e±ciency measures for modeling envi-
ronmental performance, Ecological Economics 60 (2006) 111–118.
24. P. Zhou, K. L. Poh and B. W. A. Ang, Non-radial DEA approach to measuring envi-
ronmental performance, European Journal of Operational Research 178 (2007) 1–9.
25. P. Smith, Data envelopment analysis applied to ¯nancial statements, Omega 18 (1990)
131–138.
26. H. Fujii, S. Managi and R. Matousek, Indian bank e±ciency and productivity changes
with undesirable outputs: A disaggregated approach, Journal of Banking and Finance 38
(2014) 41–50.
27. N. Zhang and Y. Choi, A note on the evolution of directional distance function and its
development in energy and environmental studies 1997–2013, Renewable and Sustainable
Energy Reviews 3(3) (2014) 50–59.
28. P. Zhou, B. W. Ang and H. Wang, Energy and CO2 emission performance in electricity
generation: A non-radial directional distance function approach, European Journal of
Opererational Research 221 (2012) 625–635.
29. N. Zhang, P. Zhou and Y. Choi, Energy e±ciency, CO2 emission performance and
technology gaps in fossil fuel electricity generation in Korea: A meta-frontier non-radial
directional distance function analysis, Energy Policy 56 (2013) 653–662.
30. J. Ruggiero and S. Bretschneider, The weighted Russell measure of technical e±ciency,
European Journal of Operational Research 108 (1998) 438–451.
31. R. Färe and S. Grosskopf, Modeling undesirable factors in e±ciency evaluation: Com-
ment, European Journal of Operational Research 157 (2004) 242–245.
32. R. Färe, S. Grosskopf and C. A. Pasurka, Environmental production functions and en-
vironmental directional distance functions, Energy 32 (2007) 1055–1066.
33. R. Färe and S. Grosskopf, Network DEA, Socio-Economic Planing Science 34 (2000) 35–49.
34. J. Liu and K. Tone, A multistage method to measure e±ciency and its application to
Japanese banking, Socio-Economic Planning Sciences 42 (2008) 75–91.
35. T. Sueyoshi and K. Sekitani, An occurrence of multiple projections in DEA-based mea-

surement of technical e±ciency: Theoretical comparison among DEA models from de-
sirable properties, European Journal of Operational Research 196 (2009) 764–794.
36. C. A. K. Lovell and J. T. Pastor, Units invariant and translation invariant DEA models,
Operations Research Letters 18 (1995) 147–151.
37. K. Tone and M. Tsutsui, Dynamic DEA: A slacks-based measure approach, Omega 38
(2010) 145–156.
38. T. Sueyoshi and M. Goto, Measurement of a linkage among environmental, operational,
and ¯nancial performance in Japanese manufacturing ¯rms: A use of data envelopment
analysis with strong complementary slackness condition. European Journal of Opera-
tional Research 207(3) (2010) 1742–1753.
39. T. Kuosmanen, Weak disposability in nonparametric production analysis with undesir-
able outputs, American Journal of Agricultural Economics 87(4) (2005) 1077–1082.
40. T. Kuosmanen and V. Podinovski, Weak disposability in nonparametric production

analysis: Reply to Färe and Grosskopf, American Journal of Agricultural Economics
91(2) (2009) 539–545.

41. R. W. Shephard, Indirect Production Functions, Mathematical Systems in Economics,
Vol. 10 (Anton Hain, Meisenheim am Glan, 1974).
42. O. Zaim and F. Taskin, Environmental e±ciency in carbon dioxide emissions in the OECD:
A non-parametric approach, Journal of Environmental Management 58 (2000) 95–107.
43. F. Taskin and O. Zaim, The role of international trade on environmental e±ciency: A
DEA approach, Economic Modeling 18 (2001) 1–17.
44. R. Färe, S. Grosskopf and F. Hernandez-Sancho, Environmental performance: An index
number approach, Resource and Energy Economics 26 (2004) 343–352.
45. B. M. Jeon and R. C. Sickles, The role of environmental factors in growth accounting,
Journal of Applied Econometrics 19 (2004) 567–591.
46. S. Kumar, Environmentally sensitive productivity growth: A global analysis using
Malmquist-Luenberger index, Ecological Economics 56 (2006) 280–293.
47. E. R. Berndt, Energy use, technical progress and productivity growth: A survey of eco-
nomic issues, Journal of Productivity Analysis 2(1) (1990) 67–83.
48. L. R. Murillo-Zamorano, The role of energy in productivity growth: A controversial issue?
Energy Journal 26(2) (2005) 69–88.
49. S. Kumar and M. Khanna, Measurement of environmental e±ciency and productivity:
Across-country analysis, Environment and Development Economics 14 (2009) 473–495.
50. P. Zhou, B. W. Ang and J. Y. Han, Total factor carbon emission performance: A
Malmquist index analysis, Energy Economics 32 (2010) 194–201.
51. C. R. Chiu, J. L. Liou, P. I. Wu and C. L. Fang, Decomposition of the environmental
ine±ciency of the meta-frontier with undesirable output, Energy Economics 34 (2012)
1392–1399.
52. Y. Y. Lin, P. Y. Chen and C. C. Chen, Measuring the environmental e±ciency of
countries: A directional distance function metafrontier approach, Journal of Environ-
mental Management 119 (2013) 134–142.
53. T. J. Coelli and S. Prasada Rao, Total factor productivity growth in agriculture: A
Malmquist index analysis of 93 countries, 1980–2000, Agricultural Economics 32 (2005)
115–134.
54. P. F. Tsai and C. Mar Molinero, A variable returns to scale data envelopment analysis
model for the joint determination of e±ciencies with an example of the UK health service,
European Journal of Operational Research 141 (2002) 21–38.
55. M. Tsutsui and M. Goto, A multi-division e±ciency evaluation of U.S. electric power
companies using a weighted slacks-based measure, Socio-Economic Planning Sciences
43(3) (2009) 201–208.
56. A. J. Picazo-Tadeo, M. Beltran-Esteve and J. A. G
omez-Lim on, Assessing eco-e±ciency
with directional distance functions, European Journal of Operational Research 220
(2012) 798–809.
57. A. J. Picazo-Tadeo, J. Castillo-Gimenez and M. Beltr an-Esteve, An intertemporal ap-
proach to measuring environmental performance with directional distance functions:
Greenhouse gas emissions in the European union, Ecological Economics 100 (2014) 173–
182.
58. S. C. Ray and Y. Jeon, Reputation and e±ciency: A non-parametric assessment of
America's top-rated MBA programs, European Journal of Operational Research 189
(2008) 245–268.
59. G. Kou and C. Lin, A cosine maximization method for the priority vector derivation in
AHP, European Journal of Operational Research 235 (2014) 225–232.
60. G. Kou, Y. Lu, Y. Peng and Y. Shi, Evaluation of classi¯cation algorithms using MCDM
and rank correlation, International Journal of Information Technology and Decision
Making 11(1) (2012) 197–225.
61. G. Kou and Y. Peng, Evaluation of clustering algorithms for ¯nancial risk analysis using
MCDM methods, Information Sciences 275 (2014) 1–12.

Chencp

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chencp

Uploaded by

Copyright:

Available Formats

International Journal of Information Technology & Decision Making

Nonradial Directional Performance Measurement

Po-Chi Chen*, Ming-Miin Yu†,**, Ching-Cheng Chang‡, Shih-Hsun Hsu§

*Department of International Business, Chung Hua University, Hsinchu, Taiwan

Published 16 April 2015

Keywords: Data envelopment analysis; directional distance function; undesirable output;

surveyed various directional distance functions employed in energy and environ-

outputs by b 2 R Jþ , where R þ represents the non-negative Euclidean *-orthant. The

y M Þ 2 R þ and undesirable outputs b ¼ ðb 1 ; b 2 ; . . . ; b J Þ 2 R þ . The DEA piecewise

reference technology can be constructed as follows:

portionate change in goods, bads and inputs. If instead we take

However, as mentioned previously, the measure of this approach fails to consider

2.2. The weighted directional distance model

g That is, if we alternatively specify the objective function in (5) as 1 ð M $ y k 0 þ

can be re-expressed as follows:

n¼1 $ n n ¼ ¼ k correspond to the average undesirable output and

De¯nition 3.2 (WRDDM undesirable output e±cient). If all optimal

De¯nition 3.4 (WRDDM overall e±cient). If all optimal solutions of (6)

Proof. From the equality (7), this theorem holds.

Proof. See Appendix A.

Proof. From the equality (7), this theorem holds.

De¯nition 3.6. If the directional vector g ¼ ðg x ; g y ; g b Þ is set to be unit

two desirable properties.

Theorem 3.4. The inverse variance WRDDM is units invariant.

3.2. Relationship between WRDDM, slacks-based measure model

Proof. See Appendix C.

3.3. Uniqueness problems and strong complementary

3.4. The WRDDM with variable returns to scale technology

4.1. Data and variables description

Table 1. Summary statistic (1992–2010).

GDP Labor Capital Energy CO 2 kt CO 2 /energy CO 2 /GDP

min 3622.4 0.114 7952.7 664.9 953.4 0.091 0.063

Upper middle income countries (35)

4.2. Di®erences between the TDDFM and WRDDM models

Table 2. Aggregate (overall) technical ine±ciency.

between these two models. It is important to observe that requiring g ¼

4.3. Comparison of input with output e±ciencies

Table 3. Components of the overall ine±ciency by OECD and non-OECD countries.

apparent that the rates of improvements in individual input-use, individual bad-

Table 4. Rank order correlation coe±cient of the OECD and non-OECD

Component Rank order correlation coe±cient U p-value

overall 10.610 28.390 11.500 <0.001*

Note: *Signi¯cant at 0.01 level.

although the result of the Mann–Whitney test is not signi¯cant.

(g) Average energy ine±ciency scores:

4.4. Di®erences in income levels

Table 5. Components of the overall ine±ciency by income levels.

Lower middle income countries

case. However, that is beyond the scope of our study.

In contrast to the impact of outputs on overall ine±ciency, the relative magnitude

This hypothesis assumes that di®erences of overall ine±ciency between di®erent

Table 6. Kruskal wallis test: Hypothesis 1.

Income level Rank order correlation coe±cient p-value

High income level 18.470 <0.001

Hypothesis 2. The in°uence of all components' ine±ciencies on overall ine±ciency

Table 7. Kendal's W test: Hypothesis 2.

Income level Rank order correlation coe±cient p-value

To elaborate more details on each component of the production performance of the

policy measures aimed at enhancing environmental e±ciency are recommended to do

Corresponding to this expression we have the overall ine±ciency,

Consider rescaling desirable output y m , undesirable output b j and input x n by

those rescale slacks corresponding to desirable output y m , undesirable output b j and

Therefore, the corresponding weighted slacks-based measure which is equivalent to