Omega 91 (2020) 102022

Contents lists available at ScienceDirect

Omega
journal homepage: www.elsevier.com/locate/omega

Additive slacks-based measure: Computational strategy and extension

to network DEA
Kun Chen a, Joe Zhu a,b,∗
a
College of Auditing and Evaluation, Nanjing Audit University, Nanjing, Jiangsu Province 211815, China
b
Foisie Business School, Worcester Polytechnic Institute, Worcester, MA 01609, USA

a r t i c l e i n f o a b s t r a c t

Article history: The current paper visits a set of data envelopment analysis (DEA) models that identify inefficiency by
Received 14 August 2018 optimizing input and output slacks. These slacks are aggregated either in an additive or ratio form. Only
Accepted 27 December 2018
the ratio slacks-based DEA models can be solved as a linear program and generate a DEA score between
zero and unity. The additive slacks-based model can be equivalent to the Russell graph measure and
Keywords: converted into a second order cone programming (SOCP) problem whose solving procedure has become a
Data envelopment analysis (DEA) mature technology. As such, the additive slacks-based model can also yield a DEA score between zero and
Additive slacks-based measure unity. This study shows that the additive slacks-based model can be applied to modelling network DEA
Russell measure where the internal structures of decision making units (DMUs) are of interest. The additive slacks-based
Second order cone programming
network DEA can be solved using SOCP technique and adapted to the preference of the decision maker by
Network DEA
choosing the weights for aggregating individual components in the network structures. It is shown that
the additive slacks-based approach can yield divisional efficiencies of Pareto optimal equivalences to be
selected by the decision maker when compared to the existing ratio slacks-based measure. An example
and solving codes are provided in the current study.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction tween zero and unity and satisfies the property of units invariance.
However, Green et al.’s [22] ASBM model is a nonlinear program-
Data envelopment analysis (DEA) is a data-oriented tool for ef- ming problem. Their model is at the mercy of non-linear program-
ficiency analysis, performance evaluation and benchmarking. DEA ing software packages and does not guarantee that the global op-
identifies a set of best practices from a given set of decision mak- timum is obtained.
ing units (DMUs) where multiple performance measures classified Using the additive model concept, Färe et al. [21] develop a Rus-
as inputs and outputs are present. DEA provides a composite in- sell graph measure (RGM) which simultaneously minimizes the in-
dex measure that is equivalent to the ratio of weighted outputs put efficiency measure and maximizes the output inefficiency mea-
to weighted input without the need for specifying the weights. sure. Note that the original Russell measure proposed by Färe and
Such a composite index can be viewed as an efficiency measure Lovell [20] is an input-oriented measure. The RGM is non-oriented
or a benchmarking index depending on the actual application it- as in the additive DEA model. However, as noted in Cooper et al.
self (Cook et al. [12]). [13] and Cooper et al. [15], the RGM model is difficult to compute
Note that there are several DEA models that are based upon due to its nonlinear objective function. They therefore develop an
so-called DEA input and output slacks. For example, Charnes et al. enhanced Russell graph measure (ERGM) where the DEA score or
[4] develop the first additive DEA model where the DEA objec- the objective function is defined as a ratio of weighted input re-
tive function is defined as the summation of all input and output ductions to weighted output increases. While the objective func-
slacks. While the additive model can identify the best practices or tion is non-linear in ERGM, ERGM can be converted into a lin-
efficient DMUs, it fails to produce a comparable DEA score or com- ear programming problem by a technique similar to the Charnes–
posite index. To overcome this limitation of the additive model, Cooper transformation (Charnes and Cooper [3]).
Green et al. [22] modify the additive DEA model into an additive On the other hand, the slacks-based measure (SBM) proposed
slacks-based measure (ASBM) that yields a DEA efficiency score be- by Tone [34] is defined in a ratio form of aggregated output slacks
to input slacks. Cooper et al. [14] demonstrate that ERGM is equiv-
alent to Tone [34]’s SBM. While the SBM measure is defined in a
∗
Corresponding author. non-linear ratio form so that DEA efficiency scores are in-between
E-mail addresses: [email protected] (K. Chen), [email protected] (J. Zhu).

https://doi.org/10.1016/j.omega.2018.12.011
0305-0483/© 2018 Elsevier Ltd. All rights reserved.
2 K. Chen and J. Zhu / Omega 91 (2020) 102022

0 and 1, it can be converted into a linear programming problem tained. This advantage from the perspective of multi-objective pro-
using the Charnes–Cooper transformation. gramming helps internal evaluation of network DEA using ASBM to
Due to its ability of producing a comparable DEA score, the SBM match the preferences of decision makers.
measure of Tone [34] has been used to develop network DEA mod- The remainder of the article is organized as follows. In
els where the internal network structures of DMUs are studied Section 2, we present the ASBM by Green et al. [22]. We show that
(Chen et al. [9], Tone and Tsutsui [35]). Network DEA can be re- the ASBM is equivalent to the RGM and can be solved directly via
garded as a generalization of conventional DEA in modelling DMUs SOCP. In Section 3, we apply the ASBM in evaluating two-stage net-
with network structures. See, for example, Cook et al. [10,11], Chen work structures. For external evaluation, the ASBM is a valid alter-
et al. [7], Kao [25], An et al. [1], and Li et al. [27]. native to the SBM. For internal evaluation, the ASBM outperforms
In the current study, we re-visit the ASBM developed by Green the SBM with respect to providing more optimal solutions of di-
et al. [22]. We first present the ASBM in a form that captures the visional efficiencies and adapting to the preference of the decision
efficiency directly. We show that the ASBM is equivalent to the maker due to its flexibility of assigning combining weights of link-
RGM. Consequently, ASBM can be converted into a second order ing divisions. Extension to multiple stage network structures and
cone programming (SOCP) problem. Note that SOCP is a special illustrative example are provided in Sections 4 and 5, respectively.
form of convex optimization where local optimum ensures to be Section 6 concludes.
the global optimum and can be solved as easily as linear program-
ming problem via interior point algorithm (Lobo et al. [30]). Nowa- 2. Additive slacks-based measure
days, SOCP has become a mature technology and a number of com-
puter solvers guarantee its global optimal solution. In DEA, non- Suppose there are n DMUs to be evaluated. Each DMUj has
linearity has often been found either in formulating additive effi- m inputs denoted as x1j , x2j , , xmj and s outputs denoted as
n n
y1j , y2j , , ysj . Let s+ j=1 yr j λ j − yr0 and si = xi0 − j=1 xi j λ j
−
r =
ciency measures, or in studying more complex relationships among
inputs and outputs such as network structures (Cook et al. [10]). denote the slack variables for outputs and inputs, respectively.
s + m −
r=1 sr +
One recent attempt in solving these nonlinear DEA problems is In the additive model of Charnes et al. [4], i=1 si is
to reformulate the non-linear DEA models as SOCP problems by used as the objective function. Obviously, the additive model can
identifying second order cone structures in disguise and to con- identify the efficient DMUs, but fails to provide a DEA score be-
vert them into useful and solvable DEA models (see, e.g, Sueyoshi cause the objective function is not commensurate. One can use
and Sekitani [33], Chen and Zhu [5], Chen and Zhu [6], Halická and s s+ m s−i
s+m ( r=1 y + i=1 x ) as an objective function. However, since
1 r
Trnovská [23]). In fact, Sueyoshi and Sekitani [33] show that SOCP r0 i0
s+
can be used to solve the RGM. Unlike the SOCP procedure for solv- s+
r is likely to be greater than yr0 , there is no guarantee that yr0
r

ing the RGM developed by Sueyoshi and Sekitani [33], the current will lie on [0, 1]. In response, Green et al. [22] develop the follow-
study presents a different SOCP procedure for the ASBM without ing measures of efficiency and inefficiency. For output r of DMU0 ,
altering the production possibility set and DEA constraints. As a the measures are
result of the current study, the ASBM becomes an alternative to  efficiency = yr0 = yr0
≤1 (1)
the SBM models developed by Tone [34]. We now have both the yr p yr0 + s+
r
ratio and additive forms of SBM along with the ERGM where the
global optimum is guaranteed.  inefficiency = 1 − efficiency = s+
r
≤1 (2)
The current paper further shows that the ASBM can be applied yr0 + s+
r
to model network DEA. The resulting ASBM-based network DEA n
models can all be solved via SOCP. According to Kao [26], there where yr p = j=1 yr j λ j . For input i of DMU0 , the measures are
are two types of network performance evaluations. One is exter- −
 efficiency = xip = xi0 − si ≤ 1 (3)
nal evaluation where the overall efficiency defined by outsiders
xi0 xi0
who do not concern the performance of intermediates which are
supposed to be invisible to them. The other is internal evaluation −
 inefficiency = 1 − efficiency = si ≤ 1 (4)
where performance of intermediates are included into the overall
xi0
efficiency which is defined by internal managers.

In order to convert the non-linear network DEA model into a where xip = nj=1 xi j λ j . Green et al. [22] suggest the following in-
linear model, the SBM-based network DEA model requires a set of efficiency measure model by summing up inefficiency measures
special unknown weights to be given posterior to the computation. (2) and (4) in the objective function.
These special endogenous unknown weights are functions of slack  
variables and may vary from one DMU to another. Magnitudes of 1 
s
s+ 
m
s−
max s+m
r
yr0 +s+
+ i
xi0
r
those special weights become known only after the SBM network r=1 i=1
DEA is solved. The resulting weights may violate the preference 
n
s.t yr j λ j − s+
r = yr0 , ∀r
of the decision maker. For example, it is likely that the resulting j=1
weight of the first division is greater than that of the second di- 
n (5)
xi j λ j + s− = xi0 , ∀i
vision whereas the decision maker thinks the second division is j=1
i

more important. In other words, these special weight functions do n

not necessarily reflect the preferences over individual components λ j = 1,
j=1
in a network structure. Obviously, as the number of divisions in-
λ j , s+r , s−i ≥ 0, ∀ j, r, i
creases in the network structure, it is less likely that the order of
special combining weights will match the preference of decision Let E0 be the optimal objective value obtained from model (5).
makers. Then, 1 − E 0 is reported as the efficiency score (Green et al. [22]).
The current study shows that by applying the ASBM under the Note that 1 − E 0 is guaranteed to lie on [0, 1] since the range
network DEA, there is no need to specify the weights for combin- of objective function of model (5) is ensured to be between zero
ing divisional efficiencies as in Tone and Tsutsui [35] and Kao [26]. and unity. However, model (5) is a nonlinear optimization problem
The ASBM-based network DEA can relax such a weighting require- and its computational strategy is not known. Although Green et al.
ment so that all Pareto optimal divisional efficiencies can be ob- [22] has mentioned to use some optimization software packages
 K. Chen and J. Zhu / Omega 91 (2020) 102022 3

to solve model (5), whether the global optimal solution can be ob- As shown in Fig. 1, the introduction of ASBM, i.e. model (6), en-
tained has not been discussed in detail. Further, the relationship ables us to visualize its relationships to other existing for efficiency
between model (5) and other efficiency measures such as RGM is measures, namely RGM, ERGM, and SBM.
also unknown.
Hence, it is useful to reconstruct an efficiency or performance 3. Extension to two-stage network structures
measure that has a transparent computational strategy. We will
show that such a measure can be equivalent to existing well- 3.1. Production possibility set
known efficiency measures in the literature. In addition, this new
efficiency measure can be extended to network DEA. Specifically, We consider a general two-stage network structure as shown
instead of combining (2) and (4) to derive an inefficiency measure, in Fig. 2 which has been widely studied in the DEA literature
we propose to add up (1) and (3) to establish the ASBM as follows. (Liang et al. [28], Lim and Zhu [29], Chen and Zhu [5], Kao [26]).
Each DMUj ( j = 1, 2, · · · , n ) has M inputs xij , (i = 1, 2, · · · , M ) to
 

s 
m
xi0 −s−
the first stage and P outputs y1p j ( p = 1, 2, · · · , P ) that leave the
1 yr0
min s+m yr0 +s+
+ xi0
i
system. In addition to these P outputs, stage 1 has D outputszdj
r
r=1 i=1

n (d = 1, 2, · · · , D ) called intermediate measures that become in-
s.t yr j λ j − s+
r = yr0 , ∀r puts to the second stage. The second stage has its own inputs
j=1

n x2h j (h = 1, 2, · · · , H ). The outputs from the second stage are yrj
(6)
xi j λ j + s−
i
= xi0 , ∀i ( r = 1, 2, · · · , S ).
j=1
Färe and Grosskopf [19] suggest that the production possibility
n
λ j = 1, set (PPS) of network system is the aggregation of PPS of individual
j=1 divisions. Thus, based upon Tone and Tsutsui [35] and Kao [26],
λ j , s+r , s−i ≥ 0, ∀ j, r, i PPS of general two-stage network shown in Fig. 2 can be defined
as follows.
Let E0 ∗ be the optimal objective value obtained from model (6). ⎧
It is straightforward to show that models (5) and (6) will locate ⎪
⎨
the same optimal solution and E 0∗ = 1 − E 0 . Moreover, model (6) is T = x, x2 , y, y1 , z
equivalent to the RGM developed in Färe et al. [21]. ⎪
⎩
The RGM model can be presented as follows
  
n 
n 
n ⎫

m 
s xi j λ1j ≤ xi , x2h j λ2j ≤ xh , ∀h, yr j λ2j ≥ yr , ∀r, ⎪
⎬
min 1
m+s
θi + φr
1
j=1 j=1 j=1
i=1 r=1
n 
n 
n

n
y1p j λ ≥ y p , ∀ p,
1
zdj λ ≥ zd , ∀d,
1
zdj λ2j ≤ zd , ∀d ⎪
⎭
s.t − x i j λ j + θi x i 0 ≥ 0 , ∀i j=1
j
j=1
j
j=1
j=1

n (7) (9)
yr j λ j − φr yr0 ≥ 0, ∀r
j=1 Clearly, in the general two-stage network system, xij , x2h j , y1p j ,
n
and yrj are visible to outsiders. Then, based on the PPS (9), we have
λ j = 1,
j=1 the following constraints after slacks are introduced:
λ j , θi , φr ≥ 0, ∀ j, i, r 
n
xi j λ1j + s−
i
= xi0 , ∀i
Based upon Halická and Trnovská [23], let θi xi0 = and xi0 − s−
i
j=1
φr yr0 = yr0 + s+r , then model (7) is equivalent to the following n
x2h j λ2j + s−
h
= x2h0 , ∀h
model. j=1
     n
1 
m
xi0 −s− 
s
yr0
yr j λ2j − s+
r = yr0 , ∀r
min m+s xi0
i
+ yr0 +s+ j=1 (10)

r
i=1 r=1 n

n y1p j λ −
1
s+p = y1p0 , ∀p
s.t − xi j λ j + xi0 − s−
i
≥ 0, ∀i j=1
j

j=1 n 
n

n (8) λ
1
= 1, 2
λ = 1,
yr j λ j − yr0 − s+
r ≥ 0, ∀r j=1
j
j=1
j

j=1 − − + +
si , sh , sr , s p , j ,
1
λ λ2j ≥ 0
n
λ j = 1, Note that (10) is the union of constraints of slack variables
j=1
λ j , s+r , s−i ≥ 0, ∀ j, r, i for each stage without considering the connection in-between the
stages. In this sense, (10) can be viewed as the constraints for solv-

Further, for constraint − nj=1 xi j λ j + xi0 − s− ≥ 0, we can prove ing a variable returns to scale (VRS) SBM model for each stage su-
 i
that we will always have − nj=1 xi j λ j + xi0 − s− = 0 at optimality. perlatively.
 i
Further, zdj are intermediates of the two-stage network struc-
Suppose − nj=1 xi j λ j + xi0 − s− i
> 0 at optimality, we can achieve
a smaller objective value by increasing the value of s− such ture and are invisible to the outsiders. Two major views have been
  i
established to impose constraints on zdj . The first type can be
that − nj=1 xi j λ j + xi0 − s− i
= 0. Thus, − nj=1 xi j λ j + xi0 − s−
i
>0
could not hold at optimality. Following this argument, we obtain viewed as cooperation on price and is defined as follows (Chen
n et al. [8])
j=1 yr j λ j − yr0 − sr = 0 at optimality. Thus, model (8) is equiva-
+

lent to model (6). 

n 
n

Note that the RGM model (7) can be solved via SOCP or λ1j zdj ≥ λ2j zdj , ∀d (11)
semidefinite programming (Sueyoshi and Sekitani [33], Halická and j=1 j=1

Trnovská [23]). Given the equivalence between the ASBM and RGM, Note that Kao [26] terms (11) as relational type. However, since
the ASBM can also be solved via SOCP or semidefinite program- zdj shares the same virtual price in multiplier form of (11, 11) can
ming. be viewed as a cooperation on price.
4 K. Chen and J. Zhu / Omega 91 (2020) 102022

Fig. 1. The relationship among four efficiency measures.

We now define the corresponding two-stage network DEA prob-

lem using ASBM as follows.
1
min
 S+P+M+H 

S
yr0 P y1p0 
M
xi0 −s− 
H
x2h0 −s−
yr0 +s+
+ y1p0 +s+
+ xi0
i
+ x2h0
h
(14)
r p
r=1 p=1 i=1 h=1
s.t Constraint Sets (10 ) and (12 )

Note when the SBM is extended into two-stage structure, cor-

Fig. 2. General two-stage network structure. responding DEA problem can be defined as follows (Kao [26]).
 

M s− s− 
H
1
1− M+ i +h
H xi0 x2
i=1
The second type of constraints on zdj can be viewed as a coop- min  h=1 h0

S
s+ P s+
p (15)
eration on quantity since the quantity being produced by the first 1+ S+1 P y +
r
y1
r=1 r0 p=1 p0
division is exactly the quantity needed in the second division. s.t Constraint Sets (10 ) and (12 )

n 
n
λ1j zdj = λ2j zdj , ∀d (12) Based upon the fact that the harmonic mean is always no
j=1 j=1
greater than the arithmetic mean, we know that efficiency score
derived by (15) is always less than the efficiency score obtained
Note that (12) is simply a special case of (11) and PPS with by (14). Both (14) and (15) identify the same efficient DMUs, since
(12) is smaller than PPS with (11). However, the choice of (12) or both of them are slacks-based measures.
(11) does not affect the computational complexity of the corre- Note that the nonlinear objective function of model
sponding DEA model since both of them are linear. 1 yr0 1 y1p0
(14) can be rewritten as S+P+M+H yr0 +s+ + S+P+M+H y1 +s+ +
r p0 p

3.2. External evaluation with ASBM M xi0 −s− H x2h0 −s−

S+P+M+H ( i=1 ). Then, for each term, we
1
xi0
i
+ h=1
h
x2
h0
Kao [26] points out that external DEA network evaluation is to introduce its upper bounds as ξs1 ,
ξ p2 , and ξ 3 respectively. Con-
evaluate the system efficiency from the perspective of outsiders. sequently, by an epigraph transformation which replaces the
In this case, overall efficiency of corresponding DEA problems does nonlinear objective function of model (14) with sum of those up-
not include slack variables of intermediates which are only observ- per bounds, model (14) is equivalent to the following optimization
able by the insiders. The overall efficiency index using ASBM based model.
on xij , x2h j , y1p j , and yrj in general two-stage network system can be 
S 
P
defined as follows. min ξs1 + ξ p2 + ξ 3
r=1 p=1
1 s.t 1 yr0
S+P+M+H yr0 +s+
≤ ξs1 , ∀s
r
S+P+M+H
  1 y1p0
≤ξ p,
2
∀p (16)
S+P+M+H y p0 +s+
 
1

S
yr0 P
y1p0 M
xi0 − s− 
H
x2h0 − s− p
i h
+ + + + (13)  
M H
xi0 −s− x2h0 −s−
yr0 + sr y + sp
1
p=1 p0
+
i=1
xi0 x2h0 1
S+P+M+H xi0
i
+ x2h0
h
≤ ξ3
r=1 h=1
i=1 h=1
n n Constraint Sets (10 ) and (12 )
j=1 λ j zdj > j=1 λ j zdj , then
Note that if optimality occurs at 1 2
n 
λ1 z − z , = − ( z
j=1 j dj d0 d0 − n
λ 2 z ) and there would be mul-
j=1 j dj Evidently, model (16) is a quadratic optimization problem
tiple projections for zd0 . Therefore, for the benefit of locating the which can be convex or nonconvex. We next show that model
unique frontier projection for intermediates, the current paper only (16) can be converted into a SOCP problem whose global opti-
considers the case of cooperation on quantity for both external and mal solution is ensured and can be obtained by solvers such as
internal DEA network evaluations. Nevertheless, if application con- CVX in Matlab since SOCP is a special form of convex optimiza-
n n
j=1 λ j zdj > j=1 λ j zdj and + ≤ ξs
yr0
text allows that optimality occurs at 1 2 tion (Boyd and Vandenberghe [2]). Obviously, S+P+1M+H y + 1
r0 sr
√
r )ξs ≥ ( yr0 ) . Then, ac-
is equivalent to (S + P + M + H )(yr0 + s+
multiple projections for zd0 are acceptable, the following conclu- 1 2

sions and solving procedure can still be applied when nj=1 λ1j zdj = cording to the transformation utilized in Chen and Zhu [5] which
n  
j=1 λ j zdj is replaced with j=1 λ j zdj ≥ j=1 λ j zdj .
2 n 1 n 2 involves some basic algebraic manipulations, we know that
 K. Chen and J. Zhu / Omega 91 (2020) 102022 5

√ n n n
r )ξs ≥ ( yr0 ) can be converted into
(S + P + M + H )(yr0 + s+ λ1j zdj = λ2j zdj (Kao [26]). For zd0 = λ1j zdj =
1 2 or zd0 = j=1 j=1 j=1
 n
λ j zdj , the efficiency of zd0 will be greater than unity. For
√
1 2 j=1
2
2 n n
yr0 + (S + P + M + H ) yr0 + s+r − ξs1 zd0 = j=1 λ j zdj =
1
j=1 λ j zdj , the efficiency of zd0 will be unity
2
2 n 
since td = td = j=1 λ j zdj − zd0 = zd0 − nj=1 λ2j zdj = 0.
+ − 1
1
≤ (S + P + M + H ) yr0 + s+r + ξs1 (17) Then, using the concept of ASBM, divisional efficiencies can be
2
defined as follows.
Note that the equivalence of (S + P + M + H )(yr0 + s+r ) ξs ≥
1
√  
( yr0 )2 and (17) can be verified by squaring both the left hand 1 P
y1p0 M
xi0 − s− 
D
zd0
i
side and the right hand side of (17) and expanding those square E1 = +
+ + +
P+M+D y 1 + s p x i 0 z d0 + td
terms out. p=1 p0 i=1 d=1
y1p0
Similarly, 1
S+P+M+H y1 +s+ ≤ ξ p2 is equivalent to (23)
p0 p

 2 1 2
y1p0 + (S + P + M + H ) y1p0 + s+p − ξ p2  
2 1 S
yr0 
H
x2h0 − s− 
D
zd0 − td−
h
E2 = + + +
1 S+H+D yr0 + sr 2
xh0 zd0
≤ (S + P + M + H ) y1p0 + s+p + ξ p2 (18) r=1 h=1 d=1
2
(24)
Note that models (17) and (18) are second order cone con-
straints (Boyd and Vandenberghe [2]). Thus, model (16) can be fur- In this manner, external DEA network evaluation can be viewed
ther converted into the following SOCP problem. as the process of efficiency decomposition where overall efficiency

S 
P is computed first and divisional efficiencies are derived afterwards.
min ξs1 + ξ p2 + ξ 3
r=1 p=1 3.3. Internal evaluation with ASBM
 √ 
 yr0 
s.t 
 1 (S + P + M + H )(yr0 + s+r ) − ξs1


Kao [26] also introduces a concept of internal DEA network
2 2 evaluation where the overall efficiency includes efficiencies of in-
1 termediates. In fact, internal evaluation is a process of efficiency
≤ (S + P + M + H ) yr0 + s+r + ξs1 , ∀s aggregation which can be viewed as a reverse process of external
2
   evaluation. Unlike external evaluation where the overall efficiency
 
 y1p0  and divisional efficiencies are computed in a successive manner,
1  internal evaluation is able to obtain overall efficiency and divi-
 (S + P + M + H ) y + s p − ξ p 
1 + 2
2 p0 2 sional efficiencies simultaneously.
1 For internal evaluation, the overall efficiency has to be defined
≤ (S + P + M + H ) y1p0 + s+p + ξ p2 , ∀p as the weighted average of divisional efficiencies. For example, un-
2
  der SBM, Kao [26] define the following divisional efficiencies
1 M
xi0 − s− 
H
x2h0 − s−
i
+ h
≤ ξ3 
M
s−
S+P+M+H xi0 x2h0 1− 1
M
i
xi0
i=1 h=1
i=1
E1 =   (25)
Constraint Sets (10 ) and (12 ) (19) 1 
P
s+ 
D
td+
1+ P+D
p
y1p0
+ zd0
where 2 is a standard notation of l2 norm. Compared with the p=1 d=1
transformation method developed by Sueyoshi and Sekitani [33] to  
derive their SOCP model, our method keeps the constraints cor- 1 
H
s− 
D
td−
1− H+D
h
x2h0
+ zd0
responding to the production possibility set, i.e. constraint sets h=1 d=1
(10) and (12), unchanged and reusable. For network DEA which E2 = (26)
1 
S
s+
has complicated structures with respect to production possibility 1+ S
r
yr0
r=1
set, our method can also avoid notation troubles.
After solving model (19), we obtain the overall efficiency. In or- Therefore, under SBM, if we want to solve the resulting SBM-
η
der to compute the divisional efficiency, we need to examine effi- based network DEA model, we have to set weights w(1 ) = η +1η
1 2
ciency of intermediate variables. According to Kao [26], slack val- η
and w(2 ) = η +2η , where η1 and η2 are denominator of (25) and
ues of intermediates variables are defined as follows. 1 2
(26), respectively to aggregate (25) and (26) into the following

n 
n
overall efficiency
td+ = λ1j zdj − zd0 , td− = zd0 − λ2j zdj , ∀d (20)   
j=1 j=1
1 1 
M
s− 1 
H
s− 
D
td−
1− 2 M
i
xi0
+ H+D
h
x2h0
+ zd0
Thus, efficiency of zd0 as outputs of the first division can be i=1 h=1 d=1
defined as follows. E=     (27)
1 1 
P
s+ 
D
td+ 1 
S
s+
zd0 1+ 2 P+D
p
y1p0
+ zd0
+ S
r
yr0
(21) p=1 r=1
zd0 + td+ d=1

Similarly, the efficiency of zd0 as inputs of the second division Note that these weights w(1)
and w(2) are in fact functions of
can be defined as follows. slacks. We do not know their values before the models are calcu-
lated. In other words, when SBM is applied in the network DEA, its
zd0 − td−
(22) limitation in the case of internal evaluation is that this special set
zd0 of weights with respect to divisional efficiencies has to be assigned
Note that in the case of cooperation on quantity, i.e. (12), in order to ensure that the derived overall efficiency remains in a
 
the position of zd0 can be either zd0 = nj=1 λ1j zdj = nj=1 λ2j zdj format of SBM that can be converted into a linear program.
6 K. Chen and J. Zhu / Omega 91 (2020) 102022

Fig. 3. General multi-stage series network structure.

Note also that after the model is solved, the resulting values 
n
η η
on the weights w(1 ) = η +1η and w(2 ) = η +2η do not necessarily td+ = λ1j zdj − zd0 , ∀d
1 2 1 2
j=1
reflect the relative importance of the divisional efficiencies. It is
likely that the special weight of the first division is greater than 
n

that of the second division whereas decision makers think the sec- td− = zd0 − λ2j zdj , ∀d
ond division is more important. j=1

In addition, for DMUs which are close to be efficient, s+ +

p , td , Constraint Sets (10 ) and (12 ) (29)
and s+ r are small numbers with respect to one. Thus, w
(1) and w(2)
1 Second, model (29) is equivalent to the following SOCP prob-
would be close to 2 . In other words, in such a case, we cannot
lem.
consider cases where w(1) is much greater or much smaller than
w(2) . 
P 
D 
S

We next show that if the internal network evaluation is based

min ξ p1 + ξd1 + ξr2 + ξ 3
p=1 d=1 r=1
upon the ASBM, there is no need to specify such weights. Based   
upon the divisional efficiencies defined in (23) and (24), we have  
 y1p0 w(1) 
the following ASBM-based network DEA model for internal net- s.t  
 1 (P + M + D) y1 + s+p − ξ p1 
work evaluation 2 p0 2
  1

P y1p0 
M
xi0 −s− 
D ≤ (P + M + D) y1p0 + s+p + ξ p1 , ∀p
min w(1) P+M+
1
y1p0 +s+
+ i
+ zd0
zd0 +td+ 2
D
p=1 
p
i=1
xi0
d=1    
  
 zd0 w(1)

+ 1 − w (1 ) 1
S
yr0
H
x2h0 −s−
D
zd0 −td−  
yr0 +s+
+ h
+ 1 
S+H+D
r=1
r
h=1
x2h0
d=1
zd0
 2 (P + M + D) zd0 + td+ − ξd1 

n (28) 2
s.t td+ = λ 1
z
j dj
− zd0 , ∀d 1
j=1 ≤ (P + M + D) zd0 + td+ + ξd1 , ∀d

n 2
td− = zd0 − λ2j zdj , ∀d   
 
j=1  yr0 1 − w(1) 
Constraint Sets (10 ) and (12 ) 1 
 (S + H + D)(yr0 + s+r ) − ξr2 
2 2
where w(1) and w(2 ) (= 1 − w(1 ) ) are user-specified weights or ex- 1
ogenous weights satisfying 0 < w(1) < 1. Unlike the SBM-based net- ≤ (S + H + D) yr0 + s+r + ξr2 , ∀r
2
work DEA, in model (28), once the weights are specified, they be-
come constants. w (1 ) 
M
xi0 − s− 1 − w (1 )
i
+
Based upon the similar transformation from model (16) to P+M+D xi0 S+H+D
i=1
model (19), model (28) can be converted into a SOCP problem as
 
follows. First, by an epigraph transformation, we have 
H
x2h0 − s− 
D
zd0 − td−
× h
+ ≤ ξ3

P 
D 
S x2h0 zd0
h=1 d=1
min ξ p1 + ξd1 + ξr2 + ξ 3
p=1 d=1 r=1 
n
td+ = λ1j zdj − zd0 , ∀d
w (1 ) y1p0
≤ ξ p1 , ∀p j=1
s.t
P + M + D y1p0 + s+p 
n
td− = zd0 − λ2j zdj , ∀d
w (1 ) zd0
≤ ξd1 , ∀d j=1
P + M + D zd0 + td+ Constraint Sets (10 ) and (12 ) (30)
1 − w (1 ) yr0 One would argue that if we use our ASBM-based network DEA
≤ ξr2 , ∀r
S + H + D yr0 + s+
r model (28), we have to choose the constant weights for aggre-
gating the divisional efficiencies. Note the fact that network DEA
w (1 ) 
M
xi0 − s− has been regarded as a problem of multi-objective optimization
i
P+M+D xi0 (Despotis et al. [17], Li et al. [27]). The key spirit of multi-objective
i=1
  optimization is to obtain multiple Pareto optimal solutions to be
1 − w(1)  x2h0 − sh 
−
H D
zd0 − td− chosen by decision makers and evaluators according to their own
+ + ≤ ξ3 personal preferences (Miettinen [31]). Because the ASBM-based
S+H+D x2h0 zd0
h=1 d=1 model allows a user to specify constant weights, the user can lo-
 K. Chen and J. Zhu / Omega 91 (2020) 102022 7

cate different divisional efficiencies of Pareto optimal equivalence For external network evaluation, the ASBM-based overall effi-
by varying combining weight w(1) from zero to unity. However, in ciency can be defined as follows.
the case of SBM-based network DEA, only one pair of divisional
efficiencies can be found since combining weights are modeled as 1
endogenous variables. Given the fact that w(1) and w(2) are func- 
Q 
Q
O (q ) + I (q )
tions of the decision variables, these weights introduce implicit q=1 q=1
constraints that change the feasible region of SBM-based network ⎛ ⎛ (q ) ⎞⎞
DEA. The resulting SBM-based divisional efficiencies may not be 
Q |
O |
y(p0
q) |
I (q ) | ( q )
x − s ( q )−
one of the Pareto solutions.
⎝ ⎝ + h0 h ⎠⎠ (34)
q=1 p=1 y(p0
q)
+ s(pq )+ h=1 xh(q0)
One may argue that it is likely that we do not know the pref-
erence over the divisional efficiencies. Under such a case, we rec-
ommend that using model (28) to locate multiple Pareto optimal where |O(q) | and |I(q) | are cardinality of O(q) and I(q) , respectively.
divisional solutions by varying weight w1 from zero to unity (see The corresponding multi-stage network DEA problem using
Appendix I). ASBM is defined as follows.
We finally note that the objective functions of SBM, i.e. (25) and 1
(26), are essentially quasi-convex functions. Thus, it is not clear min

Q 
Q
whether SBM will help to find all divisional efficiencies of Pareto O (q ) + I (q )
optimal equivalence in the case of two-stage network DEA even q=1 q=1
the restriction of combining weights is relaxed. Also, the weighted
⎛ ⎛ ⎞⎞

Q |
O (q ) |
y(p0
q) |
I (q ) | ( q )
x − s ( q )−
sum of quasi-convex functions is a non-convex function. Therefore, ⎝ ⎝ + h0 h ⎠⎠
internal evaluation with exogenous combining weights based on
q=1 p=1 y(p0
q)
+ s(pq )+ h=1 xh(q0)
SBM is a NP-hard optimization problem. In such a case, although
convex relaxation model can be constructed to find the global op- s.t Constraint Sets (31 ) and (33 ) (35)
timal solution via branch and bound algorithm (Jiao and Liu [24]),
the solving process is more laborious than solving a SOCP problem. Likewise, we have that the DEA scores obtained by model
(35) are greater than those derived by SBM. In addition, both of
4. Extension to general multiple stage network structures them will identify the same efficient DMUs. Following the trans-
formation method in the two-stage network system, model (35) is
We consider the multi-stage series network structure shown in equivalent to the following SOCP problem.
Fig. 3 where each division only provides intermediates to another ⎛ ⎛ (q ) ⎞⎞
single division (Despotis et al. [18]). This network structure is also 
Q |
O | |
I (q ) |
1
called distinctive case of general network structure by Kao [26]. min ⎝ ⎝ ξ p(q) + ξh(q) ⎠⎠

Q 
Q
In Fig. 3, x(q) , y(q) , z(q) (q = 1, · · · , Q ) denote the inputs, outputs, O (q ) + I (q ) q=1 p=1 h=1
q=1 q=1
and intermediates of division q, respectively. Let I(q) , O(q) , M(q) be
the sets of the indices of x(q) , y(q) , z(q) respectively. Then, for in-   
 
 y(p0
q)
 1 (q ) ( q )+ (q )
puts and outputs, equality constraints with slack variables can be s.t   ≤ y p0 + s p + ξ p ,
defined as follows.  1 y(q) + s(pq)+ − ξ p(q)  2
2 p0 2

n
xh(qj) λ(jq ) + sh(q )− = xh(q0) , h∈ I (q ) p ∈ O (q ) , q = 1, · · · Q
j=1
n xh(q0) − sh(q )−
y(pqj) λ(jq ) − s(pq )+ = y(p0
q)
, p ∈ O (q ) ≤ ξh(q ) , h ∈ I (q ) , q = 1, · · · Q
j=1 (31) xh(q0)
n
(q )
λ j
= 1, Constraint Sets (31 ) and (33 ) (36)
j=1
sh(q )− , s(pq )+ , (jq ) λ ≥0 Further, divisional efficiencies can be computed after model
(36) is solved. Specially, efficiency of the first division denoted as
For intermediates in the case of cooperation on price, we have
E(1) and efficiency of the last division denoted as E(Q) can be ob-

n 
n tained as follows.
λ(jq) zdj ≥ λ(jq+1) zdj , d ∈ M (q ) (32)
1
j=1 j=1
E (1 ) =
O (1 ) + I (1 ) + M (1 )
For intermediates in the case of cooperation on quantity, we ⎛ ⎞
have |
O (1 ) |
y(p0
1) |
I (1 ) | ( 1 )
x − s ( 1 )−
|M
|
(1 )
(1 )
zd0

n 
n ⎝ + i0 i
+ ⎠ (37)
λ(jq) zdj = λ(jq+1) zdj , d ∈ M (q ) (33) p=1 y(p0
1)
+ s(p1)+ i=1 xi(01) d=1
(1 )
zd0 + td(1)+
j=1 j=1

The efficiency for inputs and outputs for division q can be 1

x
(q ) −s(q )− y
(q ) E (Q ) =
defined as h0 h
(q )
and (q ) +s(q )+
p0
, respectively. Similarly, the effi- O (Q ) + I (Q ) + M (Q )
x
h0
y
p0 p ⎛ (Q ) ⎞
ciency of intermediates as inputs of division q + 1 can be defined |
O |
y(p0
Q) |
I (Q ) | ( Q ) |M(Q ) | (Q ) (Q )−
xi0 − si(Q )−  zd0 − td
(q ) −t (q )−
as
z
d0 d and the efficiency of intermediates as outputs of divi- ⎝ + + ⎠ (38)
z
(q )
d0 p=1 y(p0
Q)
+ s(pQ )+ i=1 xi(0Q ) d=1
(Q )
zd0
(q )
( q )− (q ) n
λ(jq) zdj
z
sion q can be defined as d0
(q ) +t (q )+
, where td = zd0 − j=1
z
d0 d For divisions between the first division and the last division, di-
( q )+ n
and td = j=1 λ(jq) zdj − zd0
(q )
. visional efficiency denoted as E(q) , 1 < q < Q can be computed as
8 K. Chen and J. Zhu / Omega 91 (2020) 102022

follows.

|O
(q )
| y(p0
q) |
I ( q ) | ( q ) ( q )−
x −s
E (q ) = 1
+ i0 i
+
| | | | |
O(q ) + I (q ) + M (q ) + M (q−1) | | | p=1
(q )
y p0 +s p(q )+
i=1
xi(0q )
 (39)
|M
(q−1 )
| z(q−1) −t (q−1)− |M(q) | (q )
zd0
d0 d
(q−1 ) + (q )
zd0 zd0 +td(q )+
d=1 d=1

For internal network evaluation, we define a weighted average

of E(1) , E(q) , 1 < q < Q, and E(Q) is as follows.
Fig. 4. Three-stage series network structure of electricity industry chain.

Q−1
min w (1 ) E (1 ) + w (q ) E (q ) + w (Q ) E (Q )
q=2
n
s.t td(q )− = zd0
(q )
− λ(jq) zdj , ∀d, q [26]) as shown in Fig. 4. x(1) , x(2) , and x(3) represent labor for gen-
j=1 (40) eration, labor for transmission, and labor for distribution respec-

n
tively. y(1) and y(2) refer to electric power sold to large customers
td(q )+ = λ(jq) zdj − zd0
(q )
, ∀d, q
j=1 and electric power sold to small customers respectively. z(1) and
Constraint Sets (31 ) and (33 ) z(2) stand for electric power generated and sent respectively. For
each division, there are only one input, one output and one inter-
where w(1) ,w(q) , and w(Q) are exogenous weights satisfying w(1 ) + mediate. There are ten DMUs to be evaluated.
Q−1 (q )
q=2
w + w(Q ) = 1. Similarly, model (40) is equivalent to the For DMU0 under evaluation, the constraints for the input and
following SOCP problem. output excluding intermediate are as follows.
 
|O
(1 )
| |
I (1 ) | |M(1) |
(1 ) (1 ) (1 )
min w(1) 1
ξp + ξi + ξd 
10 
10
|O(1) |+|I(1) |+|M(1) | p=1 i=1 d=1 λ(j1) x(j1) + s(1)− = x0(1) , λ(j2) x(j2) + s(2)− = x0(2) ,
j=1 j=1

Q−1

w (q ) 1 10
+ (3 ) (3 )
λj xj + s ( 3 )− = x0 ,(3 )
|
| | | | | | |
O(q ) + I (q ) + M (q ) + M (q−1)
 q=2  j=1
|| O (q ) || |  |
I (q )
M (q−1 )
|M(q) | 10 
10
× (q )
ξp + (q )
+ ξi + ξd (q−1 )
+ ξd (q )
λ(j2) y(j2) − s(2)+ = y0(2) , λ(j3) y(j3) − s(3)+ = y0(3) ,
p=1 i=1 d=1 d=1 j=1 j=1
 (Q )  10
|O | |I(Q )
| |M (Q )
| (k )
λ j = 1, k = 1, 2, 3
( ) ( ) ( )
+w ( Q ) ( Q ) 1
ξ + Q
ξi +Q
ξd Q
|O |+|I(Q ) |+|M(Q ) | p=1 p j=1

  
i=1 d=1
s( 1 )− , s(2)− , s(3)− , s(2)+ , s(3)+ , λ(jk ) ≥ 0. k = 1, 2, 3 j = 1, · · · , 10.
 
 y(p0
1)

s.t   ≤ 12 y(p01) + s(p1)+ + ξ p(1) (42)
 1 y(1) + s(p1)+ − ξ p(1) 
 2 p0  2
 (1 )  As discussed above, we only consider the case of cooperation
 zd0  (1 )
 ( 1 ) ( 1 )+  ≤ 12 zd0 + td(1)+ + ξd(1) ; on quantity and let intermediate satisfy the following conditions
1 z + t − ξd(1 ) 
2 d0 d 2
xi(01) −si(1)−
≤ ξ (1 )    
10 10 10 10
(1 )
 xi0  i  (41) λ(j1) z(j1) = λ(j2) z(j1) , λ(j2) z(j2) = λ(j3) z(j2) (43)
 
 y (q )  (q ) ( q )+ (q )
j=1 j=1 j=1 j=1
 (q) (qp0  ≤ 12 y p0 + s p + ξ p
 1 y + s )+ − ξ p(q) 
 2 p0  p 2 Note that in such a case, super-efficiencies are likely to occur,
 
 z (q )  (q ) ( q )+ (q ) i.e. divisional efficiencies can be greater than one. If we want di-
 (q) (qd0  ≤ 12 zd0 + td + ξd
 1 z + t )+ − ξ ( q )  visional efficiencies to be between zero and unity, we can replace
10 (1 ) (1 )  10 (2 ) (2 )
2 d0 d d
(q−1 )
2 (43) with j=1 λ j z j = z0(1 ) = 10 (2 ) (1 )
j=1 λ j z j and j=1 λ j z j =
xi(0q ) −si(q )− −td(q−1)−
≤ ξi(q ) ; ≤ ξd(q−1) ; 
zd0
(q ) (q−1 ) q = 2, · · · Q − 1 (2 ) (3 ) (2 )
z0 = j=1 λ j z j .
10
 xi0  zd0

 (Q ) 
 y p0 
 ( Q ) ( Q )+  ≤ 1
y(p0
Q)
+ s(pQ )+ + ξ p(Q ) ;
 1 y + s p − ξ p(Q )  2
2 p0 2
(Q )
xi(0Q ) −si(Q )− −td(Q )−
≤ ξi(Q ) ; ≤ ξd(Q )
zd0
(Q ) (Q ) 5.1. External evaluation
xi0 zd0

n
td(q )− = zd0
(q )
− λ(jq) zdj , ∀d, q For external evaluation, the corresponding network DEA model
j=1
based on ASBM is as follows.

n
td(q )+ = λ(jq) zdj − zd0
(q )
, ∀d, q
j=1  
x0(1) −s(1)− x0(2) −s(2)− x0(3) −s(3)− y0(2) y0(3)
Constraint Sets (31 ) and (33 ) min 1
5 (1 ) + (2 ) + (3 ) + (2 ) + (3 )
x0 x0 x0 y0 +s(2)+ y0 +s(3)+

s.t Constraint Sets (42 ) and (43 )

5. An example
(44)
We consider a three-stage network model of the operations of
electricity industry chain consisting of power generation, power Then, the corresponding SOCP model equivalent to (44) is as
transmission, and power distribution (Tone and Tsutsui [35]; Kao follows.
 K. Chen and J. Zhu / Omega 91 (2020) 102022 9

Table 1 The comparison between values of slack variables in Table 2 of

Overall and divisional efficiencies of external evaluation
the current paper and Table 3 of Kao [26]’s paper is also made.
based on ASBM.
We find that except that the absolute values of t (2 )+ and t (2 )− for
Company E E(1) E(2) E(3) A and B in Table 2 are larger than those obtained by Kao [26], all
A 0.6196 0.7262 0.8127 0.8037 other slack values for other DMUs are the same.
B 0.5514 0.5357 0.7624 0.9025 As a result, ASBM is a valid alternative to SBM in the case of
C 0.9838 1.0 0 0 0 1.0 0 0 0 0.9729 external evaluation due to the consistency and similarities detected
D 0.8593 0.6484 1.0 0 0 0 1.0 0 0 0
in the above discussion.
E 0.7203 0.6317 1.0 0 0 0 0.7793
F 0.6570 0.7809 0.7901 0.8613
G 0.7828 0.7428 0.9108 0.9505
H 0.9843 0.9608 1.0 0 0 0 1.0 0 0 0
I 0.9150 1.0 0 0 0 1.0 0 0 0 0.8583
J 0.6347 0.6050 0.7763 0.9572 5.2. Internal evaluation

Under the internal evaluation, the key components of linking

min ξ1 + ξ2 + ξ3 overall efficiency and divisional efficiencies are the combin-
 ( 1 ) ( 1 )− (2 )
 ing weights. When SBM is applied to internal evaluation, the
x0 −s x0 −s(2)− x0(3) −s(3)−
s.t 1
(1 ) + ≤ ξ1
(2 ) + (3 ) combining weights are computed as w(1 ) = P1 /(P1 + P2 + P3 ),
 5 x0
 x0  x0 w(2 ) = P2 /(P1 + P2 + P3 ), w(3 ) = P3 /(P1 + P2 + P3 ),
  and where
 y0(2)  P1 = 1 + t (1 )+ /z0(1 ) , P2 = 1 + 12 (s(2 )+ /y0(2 ) + t (2 )+ /z0(2 ) ),
 (2 )  ≤ 12 y0(2) + s(2)+ + 5ξ 2 and
 1 y + s ( 2 )+ − 5 ξ 2  (45)
P3 = 1 + s(3 )+ /y0(3 ) for mathematical convenience in terms of
2 0  2
 (3 )  converting the objective function from sum of linear fractional to
 y0 
 (3 )  ≤ 12 y0(3) + s(3)+ + 5ξ 3 a single linear fractional. There are three findings with respect
 1 y + s ( 3 )+ − 5 ξ 3  to combining weights derived by SBM as shown in Table 4 of
2 0 2
Constraint Sets (42 ) and (43 ) Kao [26]. First, they are heterogeneous on the level of DMUs.
Namely, those weights are distinct for each DMU under evaluation.
Based on the solution of model (45), slack variables of interme- For instance, for company A, the combining weights are (0.3197,
diates can be computed as follows. 0.3384, 0.3419) whereas they become (0.3314, 0.3756, 0.2930) for

10 
10 company B. Second, combining weights are fixed for each DMU. In
t ( 1 )+ = λ(j1) z(j1) − z0(1) , t (1)− = z0(1) − λ(j2) z(j1) , other words, combining weights cannot be modified unless DMU
j=1 j=1
(46) is changed. Third, weights of all DMUs are close to 13 .
10
(2 ) (2 ) (2 ) (2 ) 10
t ( 2 )+ = λ j z j − z0 , t ( 2 )− = z0 − λ(j3) z(j2) From the perspective of mathematical optimization, network
j=1 j=1 DEA is a problem of multi-objective optimization since each stage
in the network structure aims to maximize its own performance.
Further, divisional efficiencies can be obtained as follows.
  Consequently, the decision maker who uses network DEA as a tool
x0(1) −s(1)− z0(1) of performance evaluation should be able to decide the divisional
E (1 ) = 1
2 (1 ) + (1 )
+t (1)+
 x0 z0
 combining weights. However, the order of the combining weighs
x0(2) −s(2)− y0(2) z0(1) −t (1)− z0(2)
E (2 ) = 1
4 (2 ) + (2 ) + (1 ) + (2 ) (47) derived by the SBM is very likely to violate the decision maker’s
+s(2)+ +t (2)+
 x0 y0 z0
 z0 preference. For instance, for company A, the preference of the de-
x0(3) −s(3)− y0(3) z0(2) −t (2)− cision maker might be w(1) > w(2) > w(3) which is inconsistent with
E (3 ) = 1
+ +
3 x0(3) y0(3) +s(3)+ z0(2)
the order of (0.3197, 0.3384, 0.3419) derived by the SBM. In addi-
CVX in Matlab is adopted to solve the above model of exter- tion, the decision maker may want to set the weights to be fixed
nal evaluation (Appendix II). The results of overall efficiency and across all DMUs. Internal evaluation based on SBM may not be able
divisional efficiencies of external evaluation are summarized in to deal with such a case based upon the above findings. Moreover,
Table 1. The results of slack variables are presented in Table 2. practical cases where one of the divisions is extremely crucial in
We compare the results in Table 1 of the current paper and the view of the decision maker cannot be dealt with by the SBM-
those presented in Table 4 of Kao [26] where SBM is applied. based network DEA due to SBM’s way of specifying the combining
Note that in Kao [26] where different types of production possi- weights.
bility set are concerned, the corresponding DEA model is termed In comparison, ASBM can always adapt to the preference of the
as cooperative-discretionary type. Recall that the efficiency score decision maker since the requirement of the divisional combining
of ASBM proves to be larger than that of SBM in theory. In addi- weights is flexible in ASBM. Based upon ASBM, internal evaluation
tion, both ASBM and SBM are supposed to identify the same ef- of this particular case can be modeled as follows.
ficient DMUs. Our results confirm that overall efficiency scores in
 
Table 1 are greater than those in Table 4 in Kao [26], and the effi- x0(1) −s(1)− z0(1)
min w(1) 12 (1 ) + (1 ) +
cient stages of each DMU are identified as the same in these two +t (1)+
 x0 z0

Tables. x0(2) −s(2)− y0(2) z0(1) −t (1)− z0(2)
w(2) 14 (2 ) + (2 ) + (1 ) + (2 ) +
Further, similarities between results derived by ASBM and those +s(2)+ +t (2)+
 (3 )
x0 y0
(3 ) (2 )
z0
 z0
by SBM are also found. In particular, the finding suggests that al- x0 −s(3)− y0 z0 −t (2)−
w(3) 13 + +
though efficiency cores obtained in Table 1 are different from those x0(3) y0(3) +s(3)+ z0(2)
(48)

10 
10
obtained by SBM (Kao [26]), the rankings of the overall efficiency s.t t ( 1 )+ = λ(j1) z(j1) − z0(1) , t (1)− = z0(1) − λ(j2) z(j1) ,
and divisional efficiencies derived by these two methods are the j=1 j=1
same. Further, efficient stages with unity efficiency score are also 
10 
10

the same. Finally, under external evaluation, overall efficiency is t ( 2 )+ = λ(j2) z(j2) − z0(2) , t ( 2 )− = z0 (2 )
− λ(j3) z(j2)
j=1 j=1
likely to be lower than all divisional efficiencies (Kao [26]). Such Constraint Sets (42 ) and (43 )
a phenomenon has also been found in Table 1 where the overall
efficiency of company A is less than all its stage efficiencies.
10 K. Chen and J. Zhu / Omega 91 (2020) 102022

Table 2
Slack variables of external evaluation based on ASBM.

Company s ( 1 )− s ( 2 )− s ( 3 )− s ( 2 )+ s ( 3 )+ t ( 1 )+ t ( 1 )− t ( 2 )+ t ( 2 )−

A 0.5170 0.1125 0.5598 0.1873 0.0467 −0.0580 0.0580 0.0415 −0.0415

B 0.9120 0.0836 0.1454 0.3536 0.0305 0.1580 −0.1580 0.0340 −0.0340
C 0.0 0 0 0 0.0 0 0 0 0.0391 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0
D 1.0429 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0
E 1.1726 0.0 0 0 0 0.6688 0.0 0 0 0 0.0150 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0
F 0.4690 0.0807 0.2218 0.1120 0.0 0 0 0 −0.1300 0.1300 −0.0025 0.0025
G 0.1300 0.0195 0.0500 0.3412 0.0 0 0 0 0.1890 −0.1890 −0.0031 0.0031
H 0.0320 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0
I 0.0 0 0 0 0.0 0 0 0 0.2897 0.0 0 0 0 0.0022 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0
J 0.9010 0.1040 0.0481 0.1420 0.0182 0.0440 −0.0440 0.0200 −0.0200

Table 3 (49)
Internal evaluation based on ASBM.
When it comes to decide divisional combining weights accord-
Company E E(1) E(2) E(3)
ing to the preference of the decision maker, there are usually two
A 0.7811 0.7262 0.8245 0.7894 cases to be considered. One is when the decision maker has an ex-
B 0.6490 0.5357 0.7682 0.8902 plicit preference and the other is when the decision maker does
C 0.9946 1.0 0 0 0 1.0 0 0 0 0.9729
D 0.8242 0.6484 1.0 0 0 0 1.0 0 0 0
not. If the decision maker does have an explicit preference, then
E 0.7717 0.6317 1.0 0 0 0 0.7793 the preference can be computed as weights via Multi-Criteria De-
F 0.8658 0.8163 0.7857 0.8613 cision Making (MCDM) methods such as Delphi method (Dalkey
G 0.8021 0.7428 0.9108 0.9505 and Helmer [16]) and Analytic Hierarchy Process (AHP) (Saaty [38];
H 0.9804 0.9608 1.0 0 0 0 1.0 0 0 0
Saaty [39]). Suppose that all DMUs share the same combining
I 0.9717 1.0 0 0 0 1.0 0 0 0 0.8583
J 0.7198 0.6050 0.7763 0.9572 weights and the preference of the decision maker is derived as
w(1) > w(2) > w(3) via MCDM methods and is valid for all DMUs.
For the purpose of illustration, we let w(1 ) = 0.5, w(2 ) = 0.3 and
w(3 ) = 0.2. The CVX codes in Matlab for internal evaluation are
Consequently, the equivalent SOCP model is as follows. presented in Appendix II. The results of overall efficiency and divi-
sional efficiencies are summarized in Table 3. The results of slack
min ξ1 + ξ2 + ξ3 + ξ4 + ξ5
   variables are presented in Table 4. Note that SBM-based network
  DEA cannot yield the results similar to those in Tables 3 and 4,
 z0(1) 
s.t   ≤ 12 w2(1) z0(1) + t (1)+ + ξ 1 because the same combining weights w(1 ) = 0.5, w(2 ) = 0.3 and
 1 2(1) z(1) + t (1)+ − ξ 1  w(3 ) = 0.2 have to be applied to all DMUs.
2 w 0 2
  Comparisons between the results of Table 3 and those obtained
 y0(2) 
  ≤ 12 w4(2) y0(2) + s(2)+ + ξ 2 by SBM are made. We find that for DMU F in Table 3, the efficiency
 1 4(2) y(2) + s(2)+ − ξ 2  of the first stage is larger than that of the second stage whereas
2 w 0 2
  their order is switched in Table 4 in Kao [26]. Consequently, the
 z0(2) 
  ≤ 12 w4(2) z0(2) + t (2)+ + ξ 3 corresponding results of slack variables shown in Table 4 and those
 1 4(2) z(2) + t (2)+ − ξ 3  presented in Table 3 in Kao [26] are very different. Due to these
2 w 0  2
  differences, it is necessary to adopt ASBM when divisional com-
 y0(3) 
  ≤ 12 w3(3) y0(3) + s(3)+ + ξ 4 bining weights are set by the decision maker.
 1 3(3) y(3) + s(3)+ − ξ 4  Moreover, since divisional combining weights of the ASBM-
2 w
(1 )
0  ( 2 ) ( 2 )− 2 ( 1 ) ( 1 )− 
( 1 )−
x0 −s x0 −s z0 −t based model can be assigned according to the decision maker’s
w(1) 12 + w(2) 14 +
x0(1)  x0(2)  z0(1) preference, the ASBM-based model can also allow the case where
(3 )
x0 −s(3)− z0(2) −t (2)−
+w(3) 13 + ≤ ξ5 each DMU under evaluation is assigned by a set of different
x0(3) z0(2)
weights. In other words, ASBM can also deal with the case of

10 
10
t ( 1 )+ = λ(j1) z(j1) − z0(1) , t (1)− = z0(1) − λ(j2) z(j1) , heterogeneous combining weights where weights can be different
j=1 j=1 across DMUs.
10 10
If the decision maker does not have an explicit preference over
t ( 2 )+ = λ(j2) z(j2) − z0(2) , t (2)− = z0(2) − λ(j3) z(j2)
j=1 j=1 the divisions in a network structure, we can locate all divisional ef-
Constraint Sets (42 ) and (43 ) ficiencies of Pareto optimal equivalence by varying each divisional

Table 4
Slack variables of internal evaluation based on ASBM.

Company s ( 1 )− s ( 2 )− s ( 3 )− s ( 2 )+ s ( 3 )+ t ( 1 )+ t ( 1 )− t ( 2 )+ t ( 2 )−

A 0.5170 0.1422 0.5880 0.1486 0.0 0 0 0 −0.0580 0.0580 −0.0074 0.0074

B 0.9120 0.0870 0.1541 0.3730 0.0162 0.1580 −0.1580 0.0190 −0.0190
C 0.0 0 0 0 0.0 0 0 0 0.0391 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0
D 1.0429 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0
E 1.1726 0.0 0 0 0 0.6688 0.0 0 0 0 0.0150 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0
F 0.3896 0.0825 0.2218 0.1386 0.0 0 0 0 −0.1080 0.1080 −0.0025 0.0025
G 0.1300 0.0195 0.0500 0.3412 0.0 0 0 0 0.1890 −0.1890 −0.0031 0.0031
H 0.0320 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0
I 0.0 0 0 0 0.0 0 0 0 0.2897 0.0 0 0 0 0.0022 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0 0.0 0 0 0
J 0.9010 0.1040 0.0481 0.1420 0.0182 0.0440 −0.0440 0.0200 −0.0200
 K. Chen and J. Zhu / Omega 91 (2020) 102022 11

Table 5 ratio-based slacks measures along with the Russell measure to be

Pareto equivalent solution in internal evaluation based
used in developing DEA network models.
on ASBM.

Company E E(1) E(2) E(3) Acknowledgments

A 0.8008 0.7262 0.8245 0.7894
B 0.7269 0.5357 0.7682 0.8902 The authors are grateful to the comments and suggestions by
C 0.9892 1.0 0 0 0 1.0 0 0 0 0.9729 three anonymous reviewers. This research is supported by National
D 0.9121 0.6484 1.0 0 0 0 1.0 0 0 0
Natural Science Foundation of China (No. 71828101) .
E 0.8196 0.6317 1.0 0 0 0 0.7793
F 0.8639 0.9508 0.7779 0.8613
G 0.8465 0.7428 0.9108 0.9505 Supplementary material
H 0.9902 0.9608 1.0 0 0 0 1.0 0 0 0
I 0.9433 1.0 0 0 0 1.0 0 0 0 0.8583
Supplementary material associated with this article can be
J 0.7976 0.6050 0.7763 0.9572
found, in the online version, at doi:10.1016/j.omega.2018.12.011.

Appendix I. Locating Pareto solutions

combining weight from zero to unity. Compared with SBM-based
network DEA, our ASBM-based network DEA model is able to per- Model (28) can be viewed as the weighting method of the fol-
form such a task. For example, if we set w(1 ) = 0.25, w(2 ) = 0.35 lowing multi-objective optimization problem.
and w(3 ) = 0.40 (w(1) < w(2) < w(3) ), we obtain the results shown in  
1 
P y1p0 
M
xi0 −s− 
D
zd0
Table 5. min P+M+D y1p0 +s+
+ xi0
i
+ zd0 +td+
 p=1 
p
Note that we have located another Pareto optimal stage ef- i=1 d=1
ficiency for company F by using a different set of combining 1 
S
yr0 
H
x2h0 −s− D
zd0 −td−
min S+H+D yr0 +s+
+ x2h0
h
+ zd0
weights. For example, divisional efficiencies for company F are now r=1
r
h=1 d=1
(0.9508, 0.7779, 0.8613) which is a Pareto equivalence to (0.8163, 
n (50)
0.7857, 0.8613) shown in Table 3. Values of slack variables are
s.t td+ = λ1j zdj − zd0 , ∀d
j=1
changed accordingly. s(1 )− decreases from 0.3896 in Table 4 to 
n
0.1014. s(2 )+ increases from 0.1386 in Table 4 to 0.2169. The ab- td− = zd0 − λ2j zdj , ∀d
solute value of t (1 )+ and t (1 )− decreases from 0.1080 in Table 4 to
j=1
Constraint Sets (10 ) and (12 )
0.0281. Similarly, additional optimal divisional efficiencies of Pareto
equivalence can be found by considering other orders of w(1) , w(2) , According to Miettinen [31], the sufficient and necessary condi-
and w(3) or increasing the precision level of combining weights ac- tion for weighting method to find all Pareto optimal solutions are
cording to the application context. that the multi-objective optimization problem is convex, i.e. all the
Therefore, ASBM is more than a valid alternative to SBM in the objective functions and the feasible region are convex.
case of internal evaluation because ASBM is more flexible in terms y1p0 zd0
Evidently,
y1p0 +s+
, ∀ p and zd0 +t +
, ∀d are convex func-
of dealing with combining weights of network DEA model. Such a p d

merit can help network DEA to match the preference of a decision P y1p0 M xi0 −s− D zd0
P+M+D ( p=1 y1 +s+ )
1
tions. Thus, + i=1 xi0
i
+ d=1 z +t +
maker. p0 p d0 d
is a convex function since nonnegative weighted sum
of convex function is still a convex function. Similarly,
S yr0  x2h0 −s−  zd0 −t −
S+H+D ( r=1 yr0 +s+ + H d=1 zd0 ) is also a con-
+ D
6. Conclusions 1 h d
h=1 x2
r h0 
As a DEA model, the ASBM is less known by researchers and vex function. Constraint sets td+ = nj=1 λ1j zdj − zd0 , td− =

practitioners than SBM which has been extensively studied and ex- zd0 − nj=1 λ2j zdj , (10), and (12) are convex since all of them
tended into network DEA (see, e.g., Tone and Tsutsui [36,37], Yu are linear. Consequently, model (50) is a convex multi-objective
[40], Morita et al. [32]). In this paper, we find that the ASBM is optimization problem and model (28) is ensured to find all Pareto
equivalent to the RGM. Further, the ASBM can be transformed into optimal solutions.
a problem of SOCP whose solving method is as mature as linear
programming problems. In this sense, both ASBM and SBM are eas- Appendix II. Computing codes
ily solvable DEA models.
The current paper applies the ASBM into two-stage and multi- External network DEA
stage network structures. We show that the ASBM-based network
DEA can be solved via SOCP. For the internal DEA network evalua- % input the data
tion, the advantage of using ASBM lies in its flexibility in assigning D = [0.838 0.277 0.962 0.879 0.337 0.894 0.362;
combining weights of linking divisions. In the case of SBM, com- 1.233 0.132 0.443 0.538 0.180 0.678 0.188;
bining weights have to be specified as functions of the slack vari- 0.321 0.045 0.482 0.911 0.198 0.836 0.207;
ables for each DMU. This property makes SBM unusable when the 1.483 0.111 0.467 0.570 0.491 0.869 0.516;
combining weights are required to be the same across all DMUs 1.592 0.208 1.073 1.086 0.372 0.693 0.407;
and adapt to the preferences of the decision maker. When the 0.790 0.139 0.545 0.722 0.253 0.966 0.269;
ASBM is applied to internal network evaluation, multiple solutions 0.451 0.075 0.366 0.509 0.241 0.647 0.257;
of divisional efficiencies of Pareto equivalence can be obtained. 0.408 0.074 0.229 0.619 0.097 0.756 0.103;
In summary, ASBM is a computable and useful DEA model 1.864 0.061 0.691 1.023 0.380 1.191 0.402;
that can be used in both conventional and network DEA mod- 1.222 0.149 0.337 0.769 0.178 0.792 0.187]
eling. It deserves new attention for the purpose of DEA devel- % specify the index of DMU under evaluation
opment and applications. Since the ASBM is equivalent to the t=1
RGM, RGM can also be applied to solve network DEA models (see % specify the length of the vector of upper bounds
Appendixes III and IV). As a result, we have both additive and n = 3 cvx_begin
12 K. Chen and J. Zhu / Omega 91 (2020) 102022

% L1,L2,L3 refer to lambda1, lambda2, and lambda 3 respectively D(t,1) + w2/4 ∗ ((D(t,2)-SX(2))/D(t,2) + (D(t,6)-t1m)/D(t,6)) +
% SX refers to slack variables for inputs w3/3 ∗ ((D(t,3)-SX(3))/D(t,3) + (D(t,7)-t2m)/D(t,7)) < = U(5) norm
% SY2 refers to slack variable for output of the second stage (CZ1, 2) < = (D(t,6) + t1p + 2∗ U(1))/2 norm(CY2, 2) < = (D(t,4) +
% SY3 refers to slack variable for output of the third stage SY2 + 4∗ U(2))/2 norm(CZ2, 2) < = (D(t,7) + t1p + 4∗ U(3))/2
% U refers to upper bounds in the objective function norm(CY3, 2) < = (D(t,5) + SY3 + 3∗ U(4))/2
% C2 refers to the LHS vector of the cone structure of the output t1p = = (transpose(D(:,6)) ∗ L1) - D(t,6)
of the second stage t1m = = D(t,6) - (transpose(D(:,6)) ∗ L2)
% C3 refers to the LHS vector of the cone structure of the output t2p = = (transpose(D(:,7)) ∗ L2) - D(t,7)
of the third stage t2m = = D(t,7) - (transpose(D(:,7)) ∗ L3) transpose(D(:,1))
% LHS refers to left hand side variables L1(10) L2(10) L3(10) ∗ L1 + SX(1) = = D(t,1) transpose(D(:,2)) ∗ L2 + SX(2) = = D(t,2)

SX(3) SY2(1) SY3(1) U(3) C2(2) C3(2) transpose(D(:,3)) ∗ L3 + SX(3) = = D(t,3) transpose(D(:,4)) ∗
C2(1) = = sqrt(D(t,4)); C2(2) = = (D(t,4) + SY2 - 5∗ U(2))/2; L2 - SY2 = = D(t,4) transpose(D(:,5)) ∗ L3 - SY3 = = D(t,5)
C3(1) = = sqrt(D(t,5)); C3(2) = = (D(t,5) + SY3 - 5∗ U(3))/2; mini- transpose(ones(10,1)) ∗ L1 = = 1 transpose(ones(10,1)) ∗

mize sum(U) subject to L2 = = 1 transpose(ones(10,1)) ∗ L3 = = 1 transpose(D(:,6))

1/5 ∗ ((−1./D(t,1:3) ∗ SX) + 3) < = U(1) norm(C2, 2) < = ∗ L1 = = transpose(D(:,6)) ∗ L2 transpose(D(:,7)) ∗

(D(t,4) + SY2 + 5∗ U(2))/2 norm(C3, 2) < = (D(t,5) + SY3 + 5∗ U(3))/2 ∗

L2 = = transpose(D(:,7)) L3
transpose(D(:,1)) ∗ L1 + SX(1) = = D(t,1) transpose(D(:,2)) ∗ L1 > = 0; L2 > = 0; L3 > = 0; SX > = 0
L2 + SX(2) = = D(t,2) transpose(D(:,3)) ∗ L3 + SX(3) = = D(t,3) SY2 > = 0; SY3 > = 0; U > = 0 cvx_end
transpose(D(:,4)) ∗ L2 - SY2 = = D(t,4) transpose(D(:,5)) ∗ % compute divisional efficiencies
L3 - SY3 = = D(t,5) transpose(ones(10,1)) ∗ L1 = = 1 trans- E1 = 1/2 ∗ ((D(t,1)-SX(1))/D(t,1) + D(t,6)/(D(t,6) + t1p))
pose(ones(10,1)) ∗ L2 = = 1 transpose(ones(10,1)) ∗ L3 = = 1 trans- E2 = 1/4 ∗ ((D(t,2)-SX(2))/D(t,2) + D(t,4)/(D(t,4) + SY2) + (D(t,6)-
pose(D(:,6)) ∗ L1 = = transpose(D(:,6)) ∗ L2 transpose(D(:,7)) ∗ t1m)/D(t,6) + D(t,7)/(D(t,7) + t2p))
L2 = = transpose(D(:,7)) ∗ L3 E3 = 1/3 ∗ ((D(t,3)-SX(3))/D(t,3) + D(t,5)/(D(t,5) + SY3) + (D(t,7)-
L1 > = 0; L2 > = 0; L3 > = 0; SX > = 0 t2m)/D(t,7))
SY2 > = 0; SY3 > = 0; U > = 0 cvx_end
% compute the slack variables of intermediates
t1p = (transpose(D(:,6)) ∗ L1) - D(t,6)
t1m = D(t,6) - (transpose(D(:,6)) ∗ L2)
t2p = (transpose(D(:,7)) ∗ L2) - D(t,7)
t2m = D(t,7) - (transpose(D(:,7)) ∗ L3)
Appendix III. The equivalence of ASBM and RGM in two-stage
% compute divisional efficiencies
external network DEA evaluation and SOCP conversion
E1 = 1/2 ∗ ((D(t,1)-SX(1))/D(t,1) + D(t,6)/(D(t,6) + t1p))
E2 = 1/4 ∗ ((D(t,2)-SX(2))/D(t,2) + D(t,4)/(D(t,4) + SY2) + (D(t,6)-
Let θi xi0 = xi0 − s− , θh x2h0 = x2h0 − s− , φr yr0 = yr0 + s+
r , and
t1m)/D(t,6) + D(t,7)/(D(t,7) + t2p)) i h

E3 = 1/3 ∗ ((D(t,3)-SX(3))/D(t,3) + D(t,5)/(D(t,5) + SY3) + (D(t,7)- φ p y1p0 = y1p0 + s+p . Then, model (14) becomes
t2m)/D(t,7))
 
Internal network DEA 
S 
P 
M 
H
min 1
S+P+M+H
1
φr +
1
φp + θi + θh
r=1 p=1 i=1 h=1
% specify the index of DMU under evaluation 
n n

t=1 s.t. xi j λ1j = θi xi0 , ∀i; x2h j λ2j = θh x2h0 , ∀h

j=1 j=1
% specify the length of the vector of upper bounds 
n 
n (51)
n=5 yr j λ2j = φr yr0 , ∀r; y1p j λ1j = φ p y1p0 , ∀ p
j=1 j=1
% set combining weights n 
n 
n 
n
w1 = 0.5; w2 = 0.3; w3 = 0.2 cvx_begin λ1j zdj = λ2j zdj , ∀d; λ1j = 1, λ2j = 1,
% L1,L2,L3 refer to lambda1, lambda2, and lambda 3 respectively j=1 j=1 j=1 j=1

% SX refers to slack variables for inputs φr , φ p , θi , θh , λ1j , λ2j ≥ 0

% SY2 refers to slack variable for output of the second stage
% SY3 refers to slack variable for output of the third stage
% U refers to upper bounds in the objective function Further, since optimality always occur at equality with respect
% CY2 refers to the LHS vector of the cone structure of the out- to input and output constraints, model (51) is the same as the fol-
put of the second stage lowing model in RGM format
% CY3 refers to the LHS vector of the cone structure of the out-
put of the third stage  
% CZ1 refers to the LHS vector of the cone structure of the first 
S 
P 
M 
H
min 1
S+P+M+H
1
φr +
1
φp + θi + θh
intermediate outputs r=1 p=1 i=1 h=1
% CZ2 refers to the LHS vector of the cone structure of the sec- 
n n
s.t. xi j λ1j ≤ θi xi0 , ∀i; x2h j λ2j ≤ θh x2h0 , ∀h
ond intermediate outputs j=1 j=1
% t1p t1m t2p t2m are slack variables of intermediates variables 
n 
n (52)
L1(10) L2(10) L3(10) SX(3) SY2(1) SY3(1) U(5) CY2(2) CY3(2) CZ1(2) yr j λ2j ≥ φr yr0 , ∀r; y1p j λ1j ≥ φ p y1p0 , ∀ p
j=1 j=1
CZ2(2) t1p(1) t1m(1) t2p(1) t2m(1) n 
n 
n 
n
CZ1(1) = = sqrt(D(t,6) ∗ w1); CZ1(2) = = (D(t,6) + t1p - 2∗ U(1))/2; λ1j zdj = λ2j zdj , ∀d; λ1j = 1, λ2j = 1,
CY2(1) = = sqrt(D(t,4) ∗ w2); CY2(2) = = (D(t,4) + SY2 - 4∗ U(2))/2; j=1 j=1 j=1 j=1

CZ2(1) = = sqrt(D(t,7) ∗ w2); CZ2(2) = = (D(t,7) + t1p - 4∗ U(3))/2;

φr , φ p , θi , θh , λ1j , λ2j ≥ 0
CY3(1) = = sqrt(D(t,5) ∗ w3); CY3(2) = = (D(t,5) + SY3 -
3 U(4))/2; minimize sum(U) subject to w1/2 ∗ (D(t,1)-SX(1))/
∗
 K. Chen and J. Zhu / Omega 91 (2020) 102022 13

 
Consequently, model (52) can be written as a SOCP problem as 
P 
D S
follows. min w(1) P+M+
1
D
ξp +
ξd + 1 −w(1) S+H+ 1
D
ξr + ξ
   p=1 d=1  r=1
1 
S 
P  1   1 
min ξr + ξp +ξ s.t   
 1 (φr − ξr ) ≤ 2 (φr + ξr ), ∀r;  1 φ + − ξd 
1 
S+P+M+H 2 2 d
r=1 p=1 2 2
    ≤ 12 φd+ + ξd , ∀d
 1   1   
s.t.   ≤ 1 (φr + ξr ), ∀r;    1 
 1 (φr − ξr )  φ − ξ p )2  
 1 (φ p − ξ p ) ≤ 2 (φ p + ξ p ), ∀ p
1
2( p
2 1
2 2
2 2  
1 M 
H D
≤ (φ p + ξ p ), ∀ p ( 1)
w P+M+D 1
θi + 1 −w S+H+D
( 1) 1
θh + θd ≤ ξ−
2 i=1 h=1
  
n 
n
d=1

1 
M 
H
xi j λ1j ≤ θi xi0 , ∀i; x2h j λ2j ≤ θh x2h0 , ∀h;
θi + θh ≤ ξ , ∀i, h j=1 j=1 (55)
S+P+M+H 
n
i=1 h=1
λ2
z
j dj
≤θ −
z ,
d d0
∀d
j=1

n 
n
xi j λ1j ≤ θi xi0 , ∀i; x2h j λ2j ≤ θh x2h0 , ∀h 
n 
n
yr j λ2j ≥ φr yr0 , ∀r; y1p j λ1j ≥ φ p y1p0 , ∀ p;
j=1 j=1 j=1 j=1

n

n 
n λ1j zdj ≥ φd+ zd0 , ∀d
yr j λ ≥ φr yr0 , ∀r;
2
j y1p j λ ≥φ
1
j
1
p y p0 , ∀p j=1

n 
n
j=1 j=1 λ1j zdj = λ2j zdj , ∀d
j=1 j=1

n 
n 
n 
n
n n
λ1j zdj = λ2j zdj , ∀d; λ1j = 1, λ2j = 1, λ1j = 1, λ2j = 1,
j=1 j=1 j=1 j=1 j=1 j=1
φr , φ p , θi , θh , φd+ , θd− , λ1j , λ2j ≥ 0
φr , φ p , θi , θh , λ1j , λ2j ≥ 0 (53)

Appendix IV. SOCP conversion RGM-based internal network References

DEA evaluation
[1] An Q, Yan H, Wu J, Liang L. Internal resource waste and centralization degree
Let θi xi0 = xi0 − s−
i
, θh x2h0 = x2h0 − s−
h
, φr yr0 = yr0 + s+
r , φ p y p0 =
1 in two-stage systems: an efficiency analysis. Omega 2016;61:89–99.
[2] Boyd S, Vandenberghe L. Convex optimization. Cambridge University Press;
p , θd zd0 = zd0 − td and φd zd0 = zd0 + td . Based on the fact
y1p0 + s+ − − + +
2004.
[3] Charnes A, Cooper W, W. Programming with linear fractional functionals. Naval
that optimality will always occur at equality constraints of produc-
Res Logist Quart 1962;9:181–5.
tion possibility set as shown in Appendix III. Then, model (28) be- [4] Charnes A, Cooper WW, Golany B, Seiford L, Stutz J. Foundations of data envel-
comes opment analysis for Pareto-Koopmans efficient empirical production functions.
  J Econometrics 1985;30(1):91–07.
1 
P
1 
M 
D
1 [5] Chen K, Zhu J. Second order cone programming approach to network data en-
min w (1 ) + θi + velopment analysis. Eur J Oper Res 2017;262(1):231–8.
P+M+D φp φd+ p=1 i=1 d=1 [6] Chen K, Zhu J. Scale Efficiency in two-stage network DEA. J Oper Res Soc 2018
  doi.org/10.1080/01605682.2017.1421850.
[7] Chen Y, Cook WD, Li N, Zhu J. Additive efficiency decomposition in two-stage
1 
S
1 
H 
D
(1 )
+ 1−w + θh + θd− DEA. Eur J Oper Res 2009;196(3):1170–6.
S+H+D φr r=1 h=1 d=1
[8] Chen Y, Cook WD, Zhu J. Deriving the DEA frontier for two-stage processes.
Eur J Oper Res 2010;202(1):138–42.
[9] Chen Y, Li Y, Liang L, Salo A, Wu H. Frontier projection and efficiency de-

n 
n
composition in two-stage process with slacks-based measure. Eur J Oper Res
s.t xi j λ1j ≤ θi xi0 , ∀i; x2h j λ2j ≤ θh x2h0 , ∀h; 2016;250(2):543–54.
j=1 j=1 [10] Cook WD, Liang L, Zhu J. Measuring performance of two-stage network struc-
tures by DEA: a review and future perspective. Omega 2010;38(6):423–30.

n [11] Cook WD, Zhu J, Bi G, Yang F. Network DEA: additive efficiency decomposition.
λ2j zdj ≤ θd− zd0 , ∀d Eur J Oper Res 2010;207(2):1122–9.
[12] Cook WD, Tone K, Zhu J. Data envelopment analysis: prior to choosing a model.
j=1
Omega 2014;44(2):1–4.
[13] Cooper WW, Park KS, Pastor JT. RAM: A range adjusted measure of inefficiency

n 
n
for use with additive models and relations to other models and measures in
yr j λ2j ≥ φr yr0 , ∀r; y1p j λ1j ≥ φ p y1p0 , ∀ p; DEA. J Product Anal 1999;11(1):5–42.
j=1 j=1 [14] Cooper WW, Seiford LM, Tone K. Introduction to data envelopment analysis
and its uses. Springer Science; 2006.

n [15] Cooper WW, Huang Z, Li SX, Parker BR, Pastor JT. Efficiency aggregation with
λ1j zdj ≥ φd+ zd0 , ∀d enhanced Russell measure in data envelopment analysis. Socioecon Plann Sci
j=1 2007;41(1):1–21.
[16] Dalkey N, Helmer O. An experimental application of the Delphi method to the

n 
n user of experts. Manag Sci 1963;9(3):458–67.
λ1j zdj = λ2j zdj , ∀d [17] Despotis DK, Koronakos G, Sotiros D. The “weak-link” approach to network
DEA for two stage processes. Eur J Oper Res 2016;254(2):481–92.
j=1 j=1 [18] Despotis DK, Sotiros D, Koronakos G. A network DEA approach for series mul-
ti-stage processes. Omega 2016;61:35–48.

n 
n
[19] Färe R, Grosskopf S. Network DEA. Socioecon Plann Sci 20 0 0;34(1):35–49.
λ1j = 1, λ2j = 1, [20] Färe R, Lovell CAK. Measuring the technical efficiency of production. J Econom
j=1 j=1 Theory 1978;19(1):150–62.
[21] Färe R, Grosskopf S, Lovell. The measurement of efficiency of production.
φr , φ p , θi , θh , φd+ , θd− , λ1j , λ2j ≥ 0 (54) Boston: Kluwer-Nijhoff; 1985.
[22] Green RH, Cook WD, Doyle J. A note on the additive data envelopment analysis
Consequently, model (54) is equivalent to the following SOCP model. J Oper Res Soc 1997;48(4):446–8.
problem. [23] Halická M, Trnovská M. The Russell measure model: computational aspects,
duality, and profit efficiency. Eur J Oper Res 2018;268(1):386–97.
14 K. Chen and J. Zhu / Omega 91 (2020) 102022

[24] Jiao H-W, Liu S-Y. A practicable branch and bound algorithm for sum of linear [32] Morita H, Hirokawa K, Zhu J. A slack-based measure of efficiency in contex-
ratios problem. Eur J Oper Res 2015;243(3):723–30. t-dependent data envelopment analysis. Omega 2005;33(4):357–62.
[25] Kao C. Network data envelopment analysis: foundations and extensions. [33] Sueyoshi T, Sekitani K. Computational strategy for Russell measure in DEA:
Springer; 2016. Second-order cone programming. Eur J Oper Res 2007;180(1):459–71.
[26] Kao C. A classification of slacks-based efficiency measures in network data en- [34] Tone K. A slacks based measure of efficiency in data envelopment analysis. Eur
velopment analysis with an analysis of the properties possessed. Eur J Oper J Oper Res 2001;130(3):498–509.
Res 2018;270(3):1109–21. [35] Tone K, Tsutsui M. Network DEA: a slacks-based measure approach. Eur J Oper
[27] Li H, Chen C, Cook WD, Zhang J, Zhu J. Two-stage network DEA: who is the Res 2009;197(1):243–52.
leader? Omega 2018;74:15–19. [36] Tone K, Tsutsui M. Dynamic DEA: a slacks-based measure approach. Omega
[28] Liang L, Yang F, Cook WD, Zhu J. DEA models for supply chain efficiency eval- 2010;38(3):145–56.
uation. Ann Oper Res 2006;145(1):35–49. [37] Tone K, Tsutsui M. Dynamic DEA with network structure: a slacks-based mea-
[29] Lim S, Zhu J. Integrated data envelopment analysis: Global vs local optimum. sure approach. Omega 2014;42(1):124–31.
Eur J Oper Res 2013;229(1):276–8. [38] Saaty TL. A scaling method for priorities in hierarchical structures. J Math
[30] Lobo MS, Vandenberghe L, Boyd S, Lebret H. Applications second-order cone Psych 1977;15(3):234–81.
programming. Linear Algebra Appl 1998;284(1-3):193–228. [39] Saaty TL. Decision-making with the AHP: why is the principal eigenvector nec-
[31] Miettinen K. Nonlinear multiobjective optimization. Kluwer Academic Publish- essary. Eur J Oper Res 2003;145(1):85–91.
ers; 1998. [40] Yu M-M. Assessment of airport performance using the SBM-NDEA model.
Omega 2010;38(6):440–52.