European Journal of Operational Research xxx (2013) xxx–xxx

Contents lists available at SciVerse ScienceDirect

European Journal of Operational Research

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

Interfaces with Other Disciplines

About negative efficiencies in Cross Evaluation BCC input oriented models

João Carlos C.B. Soares de Mello a,⇑, Lidia Angulo Meza b, Juliana Quintanilha da Silveira a,
Eliane Gonçalves Gomes c
a
Universidade Federal Fluminense, Rua Passo da Pátria 156, 24210-240 Niterói, RJ, Brazil
b
Universidade Federal Fluminense, Av. dos Trabalhadores 420, 27255-125 Volta Redonda, RJ, Brazil
c
Brazilian Agricultural Research Corporation (Embrapa), Parque Estação Biológica, Av. W3 Norte final, 70770-901, Brasília, DF, Brazil

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

Article history: It will be shown in this paper that the input oriented DEA BCC model can generate negative efficiencies
Received 10 June 2011 that are usually hidden in the model. The impact of these negative efficiencies becomes obvious when
Accepted 12 February 2013 using input oriented Cross Evaluation models. With the help of an example with one input and one
Available online xxxx
output, the conditions for the possible occurrence of negative efficiencies will be shown. Furthermore,
we will show that a small intuitive change in the BCC multipliers model, previously presented in other
Keywords: papers, corrects this situation. We show why this change is used and compared it with an alternative for-
Data envelopment analysis
mulation, which avoid negative efficiencies, namely the Non-Decreasing Returns to Scale (NDRS) model.
Negative efficiencies
BCC model
We also show that the formulation studied in this paper is less restrictive than the NDRS model. The
Cross Evaluation study of this variation in the DEA BCC model will be complemented with the formulation of the dual
envelope model. This model changes the original frontier. Using the concept of non-observed DMUs,
those variations can be graphically analyzed. We have also carried out some algebraic studies concerning
benchmarks, multipliers and returns to scale.
Ó 2013 Elsevier B.V. All rights reserved.

1. Introduction makes the Production Possibility Sets to be convex. As the envelope

formulation has one equality type constraint, its dual, the multipli-
The BCC DEA model was proposed by Banker et al. (1984) and ers formulation, has one unrestricted (free) variable. This means
introduces the hypothesis of variable returns to scale. Depending we cannot ensure the non-negativity of the efficiencies when the
on whether the right orientation is chosen, the invariability of multipliers of a DMU are used to evaluate other DMUs. When
some translations, which this model includes, allows negative data working with the classic DEA models this is not a problem. How-
to be used (Pastor, 1996; Thrall, 1996). ever, in Cross Evaluation models (Sexton et al., 1986; Doyle and
Although the BCC model with the adequate translation can use Green, 1994) negative efficiencies can be generated by BCC input
negative data, the efficiencies obtained by DEA models have always oriented models.
been taken as non-negative. In fact, the input or output oriented To correct this situation, this paper revisits a small intuitive
CCR model and the output oriented BCC model do not generate change to the BCC multiplier model previously used by Angu-
negative efficiencies. The latter model includes a greater or equal lo-Meza et al. (2004) and Wu et al. (2009). These authors have
type constraint in the multipliers formulation, which ensures just used the change with no further studies. We shall interpret
non-negativity for all efficiencies, as demonstrated in Section 3. and study in this paper the consequences of this change. It will
For all classic CCR or BCC models efficiency is the objective be shown that Angulo-Meza et al. (2004) and Wu et al. (2009)
function result. In the specific case of the radial input oriented change generates a major alteration in the dual envelope model,
model the efficiency lies in the interval [0, 1]. causing a frontier displacement. Therefore, it also changes some
Now, in the case of the DEA-BCC model, the multipliers benchmarks, multipliers and assumptions of returns to scale. We
formulation was initially the dual of the envelope formulation. will also show that the modified BCC model brings further
The envelope formulation has an equality type constraint that change: the efficiency of some DMUs will depend on both effi-
cient and inefficient DMUs. We shall further go onto show the
conditions for the existence of the negative efficiencies. Finally
⇑ Corresponding author. Tel.: +55 21 26080844. we shall compare the BCC modified model with the Non-
E-mail addresses: [email protected] (J.C.C.B.), [email protected] Decreasing Returns to Scale (NDRS) model (Charnes et al.,
(L. Angulo Meza), [email protected] (J.Q. da Silveira), [email protected] 1994; Cooper et al., 2007).
(E.G. Gomes).

0377-2217/$ - see front matter Ó 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ejor.2013.02.020

Please cite this article in press as: Soares de Mello, J.C.C.B., et al. About negative efficiencies in Cross Evaluation BCC input oriented models. European Journal
of Operational Research (2013), http://dx.doi.org/10.1016/j.ejor.2013.02.020
2 J.C.C.B. Soares de Mello et al. / European Journal of Operational Research xxx (2013) xxx–xxx

2. Existence of implicit negative efficiencies in the DEA BCC negative efficiencies have yet to be studied and are the main object
model of this study.
In the case of just one output, r = 1, negative efficiencies may
The DEA-BCC model was proposed for dealing with situations appear in the restrictions of model (1) for a specific DMU j when
whereby the proportionality between inputs and outputs is not l1o y1j þ lo < 0, i.e., y1j < lo =l1o .
constant along the efficient frontier. This generates a variable re- Table 1 shows both data and results of a numerical example
turns to scale (VRS) frontier. The BCC model was originally devel- with one input and one output, illustrating the situation previously
oped by adding a convexity restriction to the CCR (Charnes et al., described. The multipliers were just the first obtained by the SIAD
1978) model envelope formulation. The frontier is piecewise linear software (Angulo-Meza et al., 2005a,b). Notice that in Table 1, DMU
and takes into account the different production scales, like increas- A lo , DMU F lf , and DMU G lg are negative.
ing, constant and decreasing returns to scale. The input oriented Table 2 shows DMUs’ efficiency indexes calculated by using
BCC multipliers model – the dual of the envelope model, is pre- other DMU multipliers, an approach similar to Cross Evaluation
sented in (1). (Sexton et al., 1986; Doyle and Green, 1994). In this table, the val-
Ps ues of cell i/j (i, j = A, . . . , G) are DMU j efficiencies as evaluated by
l yro þ lo DMU i.
Max r¼1
Pmro
mio xio Table 2 shows that when using DMU A’s output multiplier,
Ps i¼1 lA = 0.375, and the independent term, lA ¼ 2:750, three nega-
r¼1 ro rj þ lo
l y ð1Þ
S:t: Pm 6 1; j ¼ 1; . . . ; n tive efficiencies are obtained. This happens for DMUs B, C and E.
i¼1 mio xij
The outputs of these DMUs are less than 7.333 (result of
v io ; lro P 0; lo 2 R lA =lA ). These values are meaningless in the classic theory of effi-
ciency. Besides, these implicit negative efficiencies are the main
As well known in DEA literature this model can be linearized as in
reason to use of Cross Evaluation almost exclusively with constant
model (2).
returns to scale (CRS) DEA models.
X
s
Max lro yrj þ lo 3. Study of the BCC model with non-negativity constraints
r¼1
Xm
In order to avoid the negative efficiencies we shall remember
S:t: mio xij ¼ 1
i¼1 ð2Þ that they are caused only by the possible negativity of the indepen-
X
s X
m dent term lo . Therefore, a straightforward approach would be to
lro yrj þ lo  mio xij 6 0; j ¼ 1; . . . ; n restrain lo to be non-negative, i.e., lo P 0. A model with this
r¼1 i¼1 restriction is the Non-Decreasing Returns to Scale (NDRS) model
v io ; lro P 0; lo 2 R (Charnes et al., 1994; Cooper et al., 2007). However, we shall note
that not all negative values for lo lead to negative efficiencies. In
In this and similar models, DMU o is the DMU being evaluated, i.e., fact, negative efficiencies only appear if the absolute value of neg-
the observed DMU; lro are the multipliers of the outputs, vio are the ative lo is larger than the weighted sum of the output, i.e., nega-
P
multipliers of the inputs, and lro and vio are the model decision vari- tive efficiencies appear if and only if lo <  sr¼1 lro yrj . So, we
ables, yrj is the output r of DMU j, and xij is the input i of DMU j, with must only avoid this condition to happen. We must restrain
P P
j = 1, . . . , n. The variable lo indicates where the observed DMU finds lo to be not less  sr¼1 lro yrj i.e., lo P  sr¼1 lro yrj , or
Ps
itself: whether in increasing, constant or decreasing returns to scale lo þ r¼1 lro yrj P 0. With this additional constraint model (2)
frontier regions. becomes model (3).
It should be noted in model (1), that when DMU o multipliers
P X
s
are used to evaluate other DMUs, the expression sr¼1 lro yrj þ lo Max lro yrj þ lo
may be negative, if lo is negative enough. This means that the r¼1
DMU j efficiency can be negative if and when evaluated by DMU Xm

o’s multipliers. This may occur when the DMU o is operating under S:t: mio xij ¼ 1
decreasing returns to scale. However, these negative efficiencies do i¼1

not appear in DEA results except when Cross Evaluation models are Xs X
m
ð3Þ
lro yrj þ lo  mio xij 6 0; j ¼ 1; . . . ; n
used. Even in this case, the negative efficiencies exist only in the r¼1 i¼1
input oriented model. X
s
These negative efficiencies were previously detected by Soares lro yrj þ lo P 0; j ¼ 1; . . . ; n
de Mello et al. (2002), when using a smoothed DEA frontier, gener- r¼1

alized by Nacif et al. (2009), as they carried out an input oriented v io ; lro P 0; lo 2 R
Cross Evaluation with variable returns to scale. That work did not
study the negative efficiencies themselves. To perform Cross Eval-
uation, Soares de Mello et al. (2002) have used an unsophisticated
approach: DMUs that generated negative efficiencies in the Cross Table 1
Evaluation model were not taken into account. Appa et al. (2006) Numerical data example – 1 input and 1 output.

have also noticed the negative efficiencies problem in BCC Cross DMU Input Output Multipliers Efficiency
Evaluation models. v l l
Angulo-Meza et al. (2004) and Wu et al. (2009) have indepen-
A 4 10 0.250 0.375 2.750 1.000
dently approached how to deal with negative efficiencies in the in- B 1 5 1.000 0.000 1.000 1.000
put oriented Cross Evaluation BCC model. Without further C 2 7 0.500 0.000 0.500 0.500
analyses, those authors limited themselves to add a set of con- D 1 8 1.000 0.125 0.000 1.000
straints in the multipliers model to avoid negative efficiencies. E 6 6 0.167 0.000 0.167 0.167
F 5 10 0.200 0.300 2.200 0.800
The required interpretation and the consequences of the modified
G 3 9 0.333 0.500 3.667 0.833
BCC model as well as the conditions for the existence of those

Please cite this article in press as: Soares de Mello, J.C.C.B., et al. About negative efficiencies in Cross Evaluation BCC input oriented models. European Journal
of Operational Research (2013), http://dx.doi.org/10.1016/j.ejor.2013.02.020
 J.C.C.B. Soares de Mello et al. / European Journal of Operational Research xxx (2013) xxx–xxx 3

Table 2 Table 3
Cross efficiencies for the numerical example. Numerical example for the modified DEA-BCC model.

A B C D E F G DMU Input Output Multipliers Efficiency

A 1.0000 3.5000 0.2500 1.0000 0.3333 0.8000 0.8333 V l l
B 0.2500 1.0000 0.5000 1.0000 0.1667 0.2000 0.3333
A 4 10 0.250 0.083333 0.416667 0.416667
C 0.2500 1.0000 0.5000 1.0000 0.1667 0.2000 0.3333
B 1 5 1.000 0.000 1.000 1.000
D 0.3125 0.6250 0.4375 1.0000 0.1250 0.2500 0.3750
C 2 7 0.500 0.000 0.500 0.500
E 0.2500 1.0000 0.5000 1.0000 0.1667 0.2000 0.3333
D 1 8 1.000 0.000 1.000 1.000
F 1.0000 3.5000 0.2500 1.0000 0.3333 0.8000 0.8333
E 6 6 0.1666667 0.000 0.1666667 0.1666667
G 1.0008 3.5045 0.2508 1.0000 0.3338 0.8006 0.8338
F 5 10 0.2 0.066667 0.333333 0.333333
G 3 9 0.333333 0.111111 0.555556 0.444444

P
In model (3), each restriction sr¼1 lro yrj þ lo P 0 would be called a
non-negativity restriction. This model is the linearized form of the
nonlinear problem postulated by Wu et al. (2009) and Angulo-Meza Table 4
et al. (2004), presented in model (4). Cross efficiencies for the modified DEA BCC model.

Ps A B C D E F G
l yrj þ lo
r¼1
Max Pmro A 0.4167 0.0000 0.3333 1.0000 0.0556 0.3333 0.4445
i¼1 mio xij B 0.2500 1.0000 0.5000 1.0000 0.1667 0.2000 0.3333
Ps
l yrj þ lo ð4Þ C 0.2500 1.0000 0.5000 1.0000 0.1667 0.2000 0.3333
S:t: 0 6 r¼1 Pmro 6 1; j ¼ 1; . . . ; n D 0.2500 1.0000 0.5000 1.0000 0.1667 0.2000 0.3333
i¼1 mio xij E 0.2500 1.0000 0.5000 1.0000 0.1667 0.2000 0.3333
v io ; lro P 0; lo 2 R F 0.4167 0.0000 0.3333 1.0000 0.0556 0.3333 0.4444
G 0.4167 0.0000 0.3333 1.0000 0.0556 0.3333 0.4444
ek 0.3214 0.5714 0.4286 1.0000 0.1190 0.2571 0.3810
We point out that all the denominators in model (4) are positive,
due to the fact that we are dealing only with non-negative inputs
and outputs. As the NDRS imposes l to be greater or equal to zero
and the non-negative efficiency constraint model is less restrictive was used by Lins et al. (2003) when performing a model similar
than the NDRS model, as it only imposes l to be greater or equal to Cross Evaluation, output oriented, to analyse Olympic results.
than minus the outputs weighted sum.For the output oriented Table 3 shows the results of model (3) using Table 1 data. The
BCC model there is also a free variable v o , and the nonlinear form results show that the modified input oriented BCC model has de-
of this model is presented in (5). creased DMU A, F and G efficiencies as they are the only ones that
Pm generated negative efficiencies when used to evaluate other DMUs.
i¼1 mio xij þ v o For instance, the term lA =lA of DMU A is equal to 5.00, being the
Min Ps
r¼1 lro yrj lowest output value in the set of DMUs.
Pm
i¼1 mio xij þ v o ð5Þ Once again we have used the first multipliers obtained with
S:t: Ps P 1; j ¼ 1; . . . ; n model (3) to calculate the Cross Evaluation Matrix for all DMUs.
r¼1 lro yrj
The results are shown in Table 4.
v io ; lro P 0; v o 2 R We draw the readers’ attention to the fact that all evaluations
To avoid negative efficiencies as in model (4), we should add the carried out by all DMUs generate efficiencies between 0 and 1. It
constraint in (6). should be noted that DMU B has zero efficiency, when evaluated
by A, F and G DMUs. In other words, the non-negative restriction
X
m is an active restriction. We may point out that despite the fact
mio xij þ v o P 0 ð6Þ negative efficiencies are unacceptable in standard DEA model,
i¼1
nil value efficiencies may occur in those models but only when
The restriction in model (5) can be rewritten as all outputs are zero. If we want to avoid also null efficiencies
it is sufficient to replace zero by a non-Archimedean number,
X
m X
s
e, in the non-negativity restriction in model (3), as is done in
mio xij þ v o P lro yrj ð7Þ
the non-Archimedean DEA models presented by Charnes et al.
i¼1 r¼1
(1985).
P
And, as lro ; yrj P 0, then sr¼1 lro yrj P 0. So restriction (7) becomes In some cases, the modified DEA BCC model will not change any
restriction (6), i.e., restriction (6) is already implicitly included in efficiency index. This is caused by the multiplicity of the optimum
model (5). Therefore, in output oriented BCC models there are no multipliers set. This situation is exemplified in Tables 5 and 6, with
implicitly negative efficiencies, even in Cross Evaluation. This fact a data set of two inputs and one output.

Table 5
Data and BCC results for a three-dimensional numerical example.

DMU Input1 Input2 Output1 Multipliers Efficiency

v1 v2 l lo
A 0.489 0.637 0.607 0.300 1.340 1.627 0.012 1.000
B 1.000 1.000 1.000 1.000 0.000 1.300 0.300 1.000
C 0.019 0.190 0.010 30.702 2.193 0.000 1.000 1.000
D 0.032 0.008 0.005 30.702 2.193 0.000 1.000 1.000
E 0.096 0.052 0.032 0.000 19.231 19.173 0.058 0.672
F 0.053 0.035 0.007 17.754 1.687 15.240 0.505 0.612
G 0.898 0.164 0.115 0.000 6.098 6.079 0.018 0.717

Please cite this article in press as: Soares de Mello, J.C.C.B., et al. About negative efficiencies in Cross Evaluation BCC input oriented models. European Journal
of Operational Research (2013), http://dx.doi.org/10.1016/j.ejor.2013.02.020
4 J.C.C.B. Soares de Mello et al. / European Journal of Operational Research xxx (2013) xxx–xxx

Table 6
Data and modified BCC results for a three-dimensional numerical example.

DMU Input1 Input2 Output1 Multipliers Efficiency

v1 v2 l lo
A 0.489 0.637 0.607 0.300 1.340 1.627 0.012 1.000
B 1.000 1.000 1.000 0.183 0.817 0.993 0.007 1.000
C 0.019 0.190 0.010 26.989 2.564 23.168 0.768 1.000
D 0.032 0.008 0.005 30.525 2.900 26.203 0.869 1.000
E 0.096 0.052 0.032 0.000 19.231 19.173 0.058 0.672
F 0.053 0.035 0.007 17.754 1.687 15.240 0.505 0.612
G 0.898 0.164 0.115 0.000 6.098 6.079 0.018 0.717

These tables show that the classic BCC model (Table 5) and the only to the DEA BCC model; the red discontinuous line represents
modified BCC model (Table 6) provide the same efficiencies in- the frontiers of the modified DEA BCC model, and the blue line rep-
dexes for all DMUs even with a different set of multipliers. resents the NDRS frontier. The modified DEA BCC frontier displace-
Soares de Mello et al. (2002) have used a different approach to ment from the original DEA BCC frontier is clearly seen. The
deal with negative efficiencies. In their paper they did not allow modified DEA BCC frontier is closer than the NDRS one to the ori-
the DMUs generating negative efficiencies to evaluate other DMUs. ginal frontier.
In Table 2, rows corresponding to DMU A, F and G are not taken Besides the changes in the efficient frontier, other changes may
into account when calculating the average cross efficiency index, occur when introducing the non-negativity constraint. One of
ek for all and each DMU. Table 7 compares the rankings obtained those changes is in the scale elasticity. As the modified model is
by both approaches. In this case, rankings are the same. closer to CCR model than the original BCC model, scale elasticity
presents less variation than in the BCC model. Since Forsund and
Hjalmarsson (2004) and Cooper et al. (2006) shown that the scale
4. Geometrical interpretation of the additional constraint
elasticity is a function of lo , we will analyze that variable instead
of analyzing the scale elasticity itself. As mentioned before, these
Discussing the effects of the new constraint on the efficient
values of lo , as well as the values of all multipliers, are note un-
frontier is important. Thanassoulis and Allen (1988) have shown
iquely determined. So, instead of a single value of lo , we deter-
that multiplier restrictions can be replaced by one or more unob-
mine its range both in the BCC and in the modified BCC models.
served or artificial DMUs, i.e., DMUs that do not exist in the original
To determine the range of lo , for each model, we have used the
data set. As the restrictions of non-negativity, one for each DMU,
efficiency indexes obtained by each model (Table 1 for the classic
are in fact multiplier restrictions, they can be replaced by unob-
BCC model and Table 3 for the modified BCC model). Then, we have
served DMUs. So, the efficient frontier in the model with non-neg-
obtained the maximum value of lo for each DMU with a new
ativity constraints may also be dependent on the inefficient DMUs.
model whose goal is to maximize the value of lo , with the con-
The only inefficient DMUs that can change the frontier have, at
straints in the previous model and maintaining the efficiency index
least, one cross efficiency with negative value in the classic input
(a new restriction) found previously. The same was done to obtain
oriented BCC model. Models with non-observed DMUs have been
the minimum value of lo . The formulation for these BCC model is
used by Thanassoulis et al. (2011), Jahanshahloo and Soleimani-
presented in (8). This was done for both models. Table 8 shows the
Damaneh (2005), Sowlati and Paradi (2004), Diallo et al. (2007),
results for the lo range in both models.
Figueiredo and Soares de Mello (2009), Gonçalves et al. (2004),
among others.
DMU A is the only efficient DMU, the efficiency value of which
Max ðor MinÞ lo
was changed by model (3) as shown by the results of the numerical X
m
S:t: mio xij ¼ 1
two-dimensional example (Table 1). Thus, a single non-observed
i¼1
DMU is enough to modify the model frontier. The values of this Xs X
m
non-observed DMU and that of DMU A outputs are the same. The lro yrj þ lo  mio xij 6 0;
non-observed DMU input is obtained by multiplying DMU A effi- r¼1 i¼1 ð8Þ
ciency in the modified model by the actual input of DMU A. The j ¼ 1; . . . ; n; and j – 0
non-observed DMU output and input are respectively 10 and X s Xm
1.6667 for our numerical example. lro yro þ lo  Effo mio xi0 ¼ 0
Fig. 1 shows the frontiers of the DEA BCC, the modified DEA BCC r¼1 i¼1

and the NDRS models. In this figure, the black line represents the v io ; lro P 0; lo 2 R
three models common frontier; the continuous line in red belongs
All DMUs shown in this table that have changed their efficiency
in the modified BCC model have new values for the lo range. On
Table 7 top of it, DMU D, that did not change its efficiency in the modified
Average cross efficiency indexes and ranking for both approaches. BCC model has also a changed value for the lo lower bound. This
Ranking Approaches happens because this DMU is located at the point where the two
frontiers separate from each other.
Non-negativity constraint Soares de Mello et al. (2002)
D 1.000 1.000
B 0.5714 0.9063
C 0.4286 0.4844 5. Interpretation of the new constraint using the envelope
G 0.3810 0.3438 model
A 0.3214 0.2656
F 0.2571 0.2125
The geometric representation in Fig. 1 is only valid for one input
E 0.1190 0.1563
and one output. The multidimensional case can only be interpreted

Please cite this article in press as: Soares de Mello, J.C.C.B., et al. About negative efficiencies in Cross Evaluation BCC input oriented models. European Journal
of Operational Research (2013), http://dx.doi.org/10.1016/j.ejor.2013.02.020
 J.C.C.B. Soares de Mello et al. / European Journal of Operational Research xxx (2013) xxx–xxx 5

Fig. 1. The BCC, the modified BCC and the NDRS frontiers.

Table 8 Table 9
Ranges for lo in classic BCC and modified BCC models. Values of k and k0 .

DMU Classic BCC Modified BCC DMU k k0

lo lo
A kD = 1.666667 k0B = 0.666667
Max Min Max Min B kB = 1 –
C kB = 0.333333, kD = 0.6666667 –
A 2.7500 Unbound 0.4167 0.4167
D kD = 1 –
B 1.0000 1.0000 1.0000 1.0000
E kB = 666667, kD = 0.3333333 –
C 0.5000 0.5000 0.5000 0.5000
D 1.0000 11.0000 1.0000 1.6667
F kD = 1.666667 k0B = 0.666667
E 0.1667 0.1667 0.1667 0.1667 G kD = 1.333333 k0B = 0.333333
F 2.2000 Unbound 0.3333 0.3333
G 3.6640 3.6640 0.5556 0.5556

the unobserved DMU is not a real DMU, it was not included in

the data set, and consequently, it cannot be a benchmark for any
inefficient DMU. The unobserved DMU is only a tool to enable us
using the model (3) dual, i.e., the envelope formulation for the
modified BCC model. This formulation is presented in model (9). to visualize the new shape of the efficient frontier.
For the DMUs that have DMU A as one of the benchmarks in the
Min ho classic BCC model the new benchmark is DMU D alone. For the
Xn X
n other DMUs there is no change in the benchmarks’ set.
S:t: yrj kj  yrj k0j P yro ; 8r From the complementary slack theorem, we know that k0 times
j¼1 j¼1 the corresponding constraints slack must be null. Consequently,
X
n the necessary, but not sufficient, condition for k0 to be other than
xij kj 6 hxio ; 8i ð9Þ zero, is for the slack in the corresponding additional restriction of
j¼1
model (9) to be zero. This happens when the corresponding addi-
X
n X
n
tional restriction is active. This means the DMU corresponding to
kj  k0j ¼ 1; 8j
j¼1 j¼1
the active restriction will have negative efficiency using DMU o
multipliers.
kj ; k0j P 0; 8j
In the numerical example showed in Table 3, DMU A efficiency
has been changed owing to the additional constraints, and this
Model (3) additional constraints, one for each DMU, generate the
change was caused by the non-negativity constraint relative to
same number of additional decision variables in the dual model
DMU B. So, all DMU A k0 , except k0B , are null. Similar considerations
(9). These variables are called k0 . It should be noted that when the
can be made to DMUs F and G.
sum of all k0 is not null, then the sum of k will not be unitary. There-
fore, in this case, there will be no guarantee of convexity in the
modified model. 6. Final comments
Model (9) also determines benchmarks for the DMUs’ set. Table
9, shows the values of k and k0 . All benchmarks are real DMUs. This paper has shown that the input oriented DEA BCC model
Although a geometrical observation of Fig. 1 may lead to the con- can generate negative efficiencies when using Cross Evaluation
clusion that an unobserved DMU is a benchmark for G, we can models. The one input and one output example has defined the
see in Table 9, the unobserved DMU is not a benchmark. In fact, one condition for the possible occurrence of negative efficiencies.

Please cite this article in press as: Soares de Mello, J.C.C.B., et al. About negative efficiencies in Cross Evaluation BCC input oriented models. European Journal
of Operational Research (2013), http://dx.doi.org/10.1016/j.ejor.2013.02.020
6 J.C.C.B. Soares de Mello et al. / European Journal of Operational Research xxx (2013) xxx–xxx

Negative efficiencies are avoided with the non-negative con- References

straint presented by Angulo-Meza et al. (2004) and Wu et al.
(2009). We have shown that this constraint is less restrictive than Angulo-Meza, L., Soares de Mello, J.C.C.B., Gomes, E.G., Biondi Neto, L., 2004.
Eficiências negativas em modelos DEA-BCC: como surgem e como evitá-las. In:
some alternative approaches, namely the NDRS model. Paper read at VII Simpósio de Pesquisa Operacional e Logística da Marinha –
As the model with the non-negativity constrain is closer to the SPOLM, at Niterói.
original BCC model, a Cross Evaluation with an input oriented BCC Angulo-Meza, L., Biondi Neto, L., Soares de Mello, J.C.C.B., Gomes, E.G., 2005a. ISYDS
– integrated system for decision support (SIAD Sistema Integrado de Apoio a
model should use this restriction instead of the NDRS model. Of Decisão): a software package for data envelopment analysis model. Pesquisa
course, the average cross efficiency scores are higher using the Operacional 25 (3), 493–503.
non-negativity restriction and lower using the NDRS. We shall Angulo-Meza, L., Biondi Neto, L., Soares de Mello, J.C.C.B., Gomes, E.G., Coelho,
P.H.G., 2005b. Free software for decision analysis a software package for data
point out that the model with the non-negativity restriction is
envelopment models. In: Proceedings of the 7th International Conference on
not a pure BCC model, and so further research is needed in the mat- Enterprise Information Systems, at Miami Paper Read at ICEIS 2005.
ter of BCC Cross Evaluation. For instance, using the virtual peer ap- Appa, G., Argyris, N., Williams, H.P., 2006. A methodology for cross-evaluation in
proach described in Rödder and Reucher (2012). DEA. In: Working Paper: LSEOR 06.81. London School of Economics and Political
Science.
Also, the inclusion of a new set of constraints in the BCC multi- Banker, R.D., Charnes, A., Cooper, W.W., 1984. Some models for estimating technical
pliers model generates an important modification in the dual scale inefficiencies in data envelopment analysis. Management Science 30 (9),
(envelope) model. The new dual model, with a new set of variables, 1078–1092.
Charnes, A., Cooper, W.W., Rhodes, E., 1978. Measuring the efficiency of decision-
can change the efficient frontier. making units. European Journal of Operational Research 2, 429–444.
This paper includes two numerical examples. The first one Charnes, A., Cooper, W.W., Golany, B.E., Seiford, L., Stutz, J., 1985. Foundations of
has shown that the modified model causes a change in the effi- data envelopment analysis for Pareto–Koopmans efficient empirical production
functions. Journal of Econometrics 30, 91–107.
cient frontier. These changes have been analyzed through the Charnes, A., Cooper, W.W., Lewin, A.Y., Seoiford, L.M., 1994. Data Envelopment
introduction of a non-observed DMU. In this first example, the Analysis: Theory, Methodology and Applications. Kluwer Academic Publishers.
DMU causing negative efficiencies in other DMUs has had an Cooper, W.W., Seiford, L., Tone, K., 2006. Introduction to Data Envelopment Analysis
and its Uses: With DEA-Solver Software and References. Springer Science, USA.
efficiency decrease in the modified model. This decrease did Cooper, W.W., Seiford, L.M., Tone, K., 2007. A Comprehensive Text with Models,
not occur in the second example (two inputs and one output). Applications, References and DEA-Solver Software. Springer, New York.
This has happened on account of the optimum multiplier Diallo, M., Pereira de Souza, M.V., Guedes, L.E., Castro Souza, R., 2007. Artificial
DMUs and contingent weight restrictions for the analysis of Brazilian retail
multiplicity.
banks efficiency. In: Operations Research Proceedings 2007. Springer, pp. 97–
We have also analyzed the algebraic characteristics of the mod- 102.
ified BCC model, and have empirically observed the relationship Doyle, J., Green, R.H., 1994. Efficiency and cross-efficiency in DEA derivations,
between k and k0 . In all cases, the benchmarks were always real meanings and uses. Journal of the Operational Research Society 45, 567–578.
Figueiredo, D.S.D., Soares de Mello, J.C.C.B., 2009. Índice híbrido de eficácia e
DMUs. eficiência para lojas de varejo. Gestão & Produção 16 (2), 286–300.
Possible changes in the efficient frontier, and consequently in Forsund, F.R., Hjalmarsson, L., 2004. Calculating scale elasticity in DEA models.
the efficiency indexes, may occur only in DMUs that generate Journal of the Operational Research Society 55, 1023–1038.
Gonçalves, D.A., Soares de Mello, J.C.C.B., Angulo-Meza, L., Lins, M.P.E., 2004. When a
negative efficiencies when analyzing other DMUs. However, value judgment using the technique of unobserved DMUs can be changed by
changes in the range of multipliers and in the scale factor may weight restrictions. Relatórios de Pesquisa em Engenharia de Produção. 4.
occur in DMUs whose efficiency is not affected by the new con- Jahanshahloo, G.R., Soleimani-Damaneh, M., 2005. A note on simulating weights
restrictions in DEA: an improvement of Thanassoulis and Allen’s method.
straint. This happens to DMUs located in the points where the Computers and Operations Research 32 (4), 1037–1044.
BCC and the modified BCC frontiers separate from each other. Lins, M.P.E., Gomes, E.G., Soares de Mello, J.C.C.B., Soares de Mello, A.J.R., 2003.
It is obvious that the model presented in this paper is a new Olympic ranking based on a zero sum gains DEA model. European Journal of
Operational Research 148, 312–322.
DEA model different from the BCC model and with different Nacif, F.B., Soares de Mello, J.C.C.B., Angulo-Meza, L., 2009. Choosing weights in
assumptions. One of its many characteristics is that the location optimal solutions for DEA-BCC models by means of a n-dimensional smooth
of the unobserved DMU depends on some efficient and ineffi- frontier. Pesquisa Operacional 29 (3), 623–642.
Pastor, J.T., 1996. Translation invariance in data envelopment analysis: a
cient DMUs. In other words, contrary to standard DEA model,
generalization. Annals of Operations Research 66, 93–102.
in the model presented here the frontier shape depends also Rödder, W., Reucher, E., 2012. Advanced X-efficiencies for CCR- and BCC-models –
on some inefficient DMUs. towards peer-based DEA controlling. European Journal of Operational Research
It should be noted that the negative efficiencies mentioned in 219 (2), 467–476.
Sexton, T.R., Silkman, R.H., Logan, A.J., 1986. Data envelopment analysis: critique
this work apply neither to the CCR model nor to the output ori- and extensions. In: Silkman, H. (Ed.), Measuring Efficiency: An Assessment of
ented BCC model. The latter includes a greater or equal type con- Data Envelopment Analysis. Jossey-Bass, San Francisco, pp. 73–105.
straint in the multipliers model, ensuring the non-negativity of Soares de Mello, J.C.C.B., Lins, M.P.E., Gomes, E.G., 2002. Construction of a smoothed
DEA frontier. Pesquisa Operacional 28 (2), 183–201.
efficiency measures, as demonstrated in Section 3. Sowlati, T., Paradi, J.C., 2004. Establishing the ‘‘practical frontier’’ in data
It is important to observe the model presented here solves one envelopment analysis. Omega 32 (4), 261–272.
of the problems of using Cross Evaluation with the variable returns Thanassoulis, E., Allen, R., 1988. Simulating weight restrictions in data envelopment
analysis by means of unobserved DMUs. Management Science 44 (4), 586–594.
to scale models. Other problems may occur and will be studied in Thanassoulis, E., Kortelainen, M., Allen, R., 2011. Improving envelopment in data
future works. Also, other approaches may exist to deal with nega- envelopment analysis under variable returns to scale. European Journal of
tive efficiencies in Cross Evaluation. Operational Research 218 (1), 175–185.
Thrall, R.M., 1996. The lack of invariance of optimal dual solutions under
We may point out that this study has been developed for non- translation. Annals of Operations Research 66, 103–108.
negative data. Further studies are required regarding the presence Wu, J., Liang, L., Chen, Y., 2009. DEA game cross-efficiency approach to Olympic
of negative data in the modified BCC model with non-negativity rankings. Omega 37 (4), 909–918.
constraints.

Please cite this article in press as: Soares de Mello, J.C.C.B., et al. About negative efficiencies in Cross Evaluation BCC input oriented models. European Journal
of Operational Research (2013), http://dx.doi.org/10.1016/j.ejor.2013.02.020