SUPER EFFICIENCY DATA DATA ENVELOPMENT ANALYSIS MODELS1

JOSEF JABLONSKÝ
UNIVERSITY OF ECONOMICS PRAGUE, CZECH REPUBLIC

Abstract
Data Envelopment Analysis (DEA) offers several models for comparison the
relative efficiency of decision making units (DMU) described by multiple inputs and
outputs. In most DEA models the best DMUs reach the efficient score 100%. Typically
the number of units with 100% efficiency can be high (depending on the number of units
and/or inputs and outputs used in the model). The necessity to classify the efficient units
leads to the several definition of super efficiency. The efficient units in super efficient
models can reach the efficient score higher than 100%. In this way they can be simply
classified. The paper presents two super efficiency models and discusses their results on a
numerical example.

1. Basic DEA models

Data envelopment analysis (DEA) is a tool for measuring the relative efficiency and
comparison of decision making units (DMU). The DMUs are usually characterised by several
inputs that are spent for production of several outputs. Let us consider set E of n decision
making units E = {DMU1, DMU2, …, DMUn}. Each of the units produces r outputs and for
their production spent m inputs. Let us denote Xj = {xij, i=1,2,…,m} the set of inputs and Yj =
{yij, i=1,2,…,r} the set of outputs for the DMUj. Then X is the (m,n) matrix of inputs and Y
the (r,n) matrix of outputs.
The basic principle of the DEA in evaluation of efficiency of the DMUq, q∈{1,2.,,,.n}
consists in looking for a virtual unit with inputs and outputs defined as the weighted sum of
inputs and outputs of the other units in the decision set - Xλ a Yλ, where λ=(λ1, λ2,…, λn),
λ>0 is the vector of weights of the DMUs. The virtual unit should be better (or at least not
worse) than the analysed unit DMUq. The problem of looking for a virtual unit can generally
be formulated as standard linear programming problem:
minimise z=θ,

1
The paper is supported by the Grant Agency of Czech Republic – grant no. 402/01/0133 and corresponds to the
research program of the Faculty of Informatics and Statistics no. CEZ:J18/98:311401001.

112
subject to Yλ ≥ Yq , (1)
Xλ ≤ θXq ,
λ ≥ 0.
The DMUq is to be considered as efficient if the virtual unit is identical with analysed
unit (does not exist the virtual unit with better inputs and outputs). In this case Yλ = Yq, Xλ =
Xq and minimum value of z = θ = 1. Otherwise the DMUq is not efficient and minimum value
of θ<1 can be interpreted as the need of proportional reduction of inputs in order to reach the
efficient frontier. The presented model is so called input oriented model because its objective
is to find reduction of inputs in order to reach the efficiency. Analogously can be formulated
output oriented model.
Model (1) presents just the basic philosophy of the DEA models. The input oriented
form of the Charnes, Cooper, Rhodes model (CCR-I) is formulated as follows:
minimise z = θ − ε(eTs+ + eTs−) ,
subject to Yλ − s+ = Yq , (2)
Xλ + s− = θXq ,
λ, s+, s− ≥ 0,
where eT = (1,1,…,1) and ε is a infinitesimal constant (usually 10-6 or 10-8). Presented
formulations (1) and (2) are very close each other. The variables s+, s− are just slack variables
expressing the difference between virtual inputs/outputs and appropriate inputs/outputs of the
DMUq. Obviously, the virtual inputs/outputs can be computed with optimal values of
variables of the model (2) as follows:
X’q = Xqθ* − s− ,
Y’q = Yq + s+ .
2. Super efficiency models
The efficiency score in the standard CCR input oriented model is limited to unity
(100%). Nevertheless, the number of efficient units identified by the CCR model and reaching
the maximum efficiency score 100% can be relatively high and especially in problems with a
small number of decision units the efficient set can contain almost all the units. In such cases
it is very important to have a tool for a diversification and classification of efficient units.
That is why several DEA models for classification of efficient units were formulated. In these
models the efficient scores of inefficient units remain lower than 100% but the efficient score
for efficient units can be higher than 100%. Then, the efficient score can be taken as a basis

113
for a complete ranking of efficient units. The DEA models that relax the condition for unity
efficiency are called super efficiency models.
The first super efficiency DEA model was formulated by Andersen and Petersen in
[1]. Its input oriented formulation (3) is very close to the standard input oriented formulation
of CCR model (2). In this model the weight λq of the evaluated unit DMUq is equated to zero.
This cannot influence the efficient score of the inefficient units but the efficient score of the
efficient units is not limited by unity in this case. The input oriented formulation of the
Andersen and Petersen model is as follows:
minimise z=θ,
n
subject to ∑ x ijλ j + s i− = θx iq , i = 1,2,...,m, (3)
j=1, ≠ q
n
∑ y ij λ j − s i+ = y iq , i = 1,2,...,r,
j=1, ≠ q

λ, s , s− ≥ 0.
+

Tone in [6] proposes a slack based measure of efficiency (SBM model) that is basis for
his formulation of the super efficiency model presented in [7]. The SBM model is formulated
as follows:
1 m −
1− ∑ s i / x iq
m i =1
minimise ρ= ,
1 r
1 + ∑ s i+ / y iq
r i =1
n
subject to ∑ x ij λ j + s i− = x iq , i = 1,2,...,m, (4)
j=1
n
∑ y ij λ j − s i+ = y iq , i = 1,2,...,r,
j=1

λ, s+, s− ≥ 0.
The formulation shows that the SBM model is non-radial and deals directly with all
the slacks variables. The model returns efficiency score between 0 and 1 and is equal to 1 if
and only if the DMUq is on the efficient frontier without any slacks. It is possible to prove that
the efficient score of the SBM model is always lower or equal than the efficient score of the
appropriate CCR input oriented model. The formulation of the model (4) with fractional
objective function can be simply transformed into a standard problem with linear objective
function.

114
 The super efficiency SBM model removes the evaluated unit DMUq from the set of
units (like Andersen and Petersen model) and looks for a DMU* with inputs x* and outputs
y* being SBM (and CCR) efficient after this removal. It is clear that all the inputs of the unit
DMU* have to be higher or equal than the inputs of the unit DMUq and all the outputs will be
lower or equal comparing to outputs of DMUq. The super efficiency is measured as a distance
of the inputs/outputs of both the units. As a distance measure in the mathematical formulation
of the super SBM model below the index δ is used:

1 m *
∑ x i / x iq
m i =1
minimise δ= , (5)
1 r *
∑ y i / y iq
r i =1
n
subject to ∑ x ij λ j + s i− = x iq , i = 1,2,...,m,
j=1, ≠ q
n
∑ y ij λ j − s i+ = y iq , i = 1,2,...,r, (6)
j=1, ≠ q

x *i ≥ x iq , i = 1,2,...,m,
y *i ≤ y iq , i = 1,2,...,r,
λ, s+, s− , y *i ≥ 0.
The numerator in the ratio (5) can be interpreted as a distance of both the units in the
input space and an average reduction rate of inputs of DMU* to inputs of DMUq. The same
holds for the output space in the denominator of the ratio (5). The input oriented formulation
of the super SBM model is given as follows:

1 m *
minimise δ = ∑ x i / x iq , (7)
m i =1
n
subject to ∑ x ij λ j + s i− = x iq , i = 1,2,...,m,
j=1, ≠ q
n
∑ y ij λ j + s i− = y iq , i = 1,2,...,r, (8)
j=1, ≠ q

x *i ≥ x iq , i = 1,2,...,m,
y *i = y iq , i = 1,2,...,r,
+ −
λ, s , s ≥ 0.
The formulation of the previous model is derived from the model (5)-(6) by setting the
denumerator equal to 1, i.e. yi* = y iq.

115
 The super SBM model (7)-(8) gives optimal objective values greater or equal 1. The
optimal efficient score is greater than 1 for efficient DMUs – higher value is assigned to more
efficient units. All the SBM inefficient units reach in the super SBM model optimal score 1. It
means that this model cannot be used for classification of inefficient units. The SBM models
have to be used in two steps. The first step is applied to the entire set of units in order to
identify efficient units and classify inefficient units. The second step is the computation of the
super efficiency scores by means of the model (7) and (8).
3. A numerical example
The basic features of the super efficiency models will be illustrated on the numerical
example. This example is based on a real data set originating from the survey among the
Central European firms. Table 1 contains a data set of 12 firms of the meat processing
industry from Czech Republic (CZ), Germany (GW) and Hungary (H). Each firm is described
by 6 characteristics: turnover, market share, fixed costs, # of workers, floor space and
investments (all the financial characteristics are in millions of Euro). The first and second
characteristics are taken as outputs of the model, the remaining ones are inputs.

Market Fixed # of Floor

Turnover share costs workers space Invest.
CZ1 1.486 40 0.8 39 2500 0.01
CZ2 3.143 5 0.023 30 960 0.057
CZ3 3.871 70 0.949 73 800 0.266
CZ4 2.857 5 1 52 500 0.057
GW1 7.669 40 6.136 26 2500 0.102
GW2 3.835 5 1.023 18 1000 0.102
GW3 30.678 40 4.001 100 2700 0.511
GW4 5.466 15 2.572 65 2900 0.378
H1 1.753 10 0.253 20 1300 0.078
H2 1.851 10 0.07 19 400 0.148
H3 3.409 10 0.031 27 920 0.249
H4 1.648 12 0.105 27 4601 0.23
Tabulka 1 – The data set.
Table 2 contains the efficiency scores computed by:
‰ standard CCR input oriented model (2),
‰ SBM model (4),
‰ input oriented Andersen and Petersen model (3) and
‰ input oriented super SBM model (7)-(8).

116
 Andersen Super
CCR SBM Petersen SBM
CZ1 1.000 1.000 10.200 4.505
CZ2 1.000 1.000 3.588 1.818
CZ3 1.000 1.000 3.634 2.095
CZ4 0.798 0.502 0.798 xxx
GW1 1.000 1.000 1.997 1.288
GW2 0.694 0.537 0.694 xxx
GW3 1.000 1.000 2.643 1.945
GW4 0.371 0.294 0.371 xxx
H1 0.803 0.644 0.803 xxx
H2 1.000 1.000 1.234 1.115
H3 1.000 1.000 1.733 1.309
H4 0.829 0.612 0.829 xxx
Tabulka 2 – Results of DEA models.

On the example of the DMU1 we will explain the difference between both the super
efficiency models. The optimal efficiency score for DMU1 in the Andersen and Petersen
model is 10.2. This value is the absolute expansion rate for input values of this unit, i.e. if all
the inputs of the projected unit are expanded 10.2 times the unit still becomes CCR efficient
and the projected point has the following reduced inputs x* = (8.16, 397.8, 25500, 0.102) with
the identical outputs. The projected inputs for the point on the efficient frontier result directly
from the optimisation model (7)-(8). For DMU1 they are x* = (0.8, 40.99, 2500, 0.1497). The
optimal value of δ* = 4.505 is the average distance of the vector of projected inputs and the
vector of inputs x1 = (0.8, 39, 2500, 0.01), i.e.

0.8 / 0.8 + 40.99 / 39 + 2500 / 2500 + 0.1497 / 0.01

= 4.505.
4

It is clear that the projected point given by the super SBM model is much closer (in
sense of the selected metric) to the efficient frontier. It can be taken as advantage of the super
SBM model because it leads always to lower super efficiency scores comparing to the
Andersen and Petersen model. These lower super efficiency scores can be better explained
and interpreted. The ranking of units by both the super efficiency models cannot be identical.
The results in Table 2 show some rank reversals of efficient units in Andersen and Petersen
and super SBM model.

117
References
1. Andersen,P., Petersen,N.C.: A procedure for ranking efficient units in data envelopment
analysis. Management Science 39 (1993), pp.1261-1264.
2. Charnes,A., Cooper,W.W., Lewin,A., Seiford,L.: Data Envelopment Analysis: Theory,
Methodology and Applications. Kluwer Publ., Boston 1994.
3. Cooper, W.W., Seiford, L.M, Tone, K.: Data Envelopment Analysis. Kluwer Publ.,
Boston 2000.
4. Jablonský,J.: Implementation of Data Envelopment Analysis in Lingo Modelling System.
Acta Oeconomia Pragensia 4, 7 (1999), pp.75-84.
5. Jablonský,J.: Comparison of efficiency of Czech industrial branches by data envelopment
analysis. In: CD Proceedings of the APORS 2000 Conference, Singapore 2000, p. 8.
6. Tone,K.: A slack-based measure of efficiency in data envelopment analysis. European
Journal of Operational Research 130 (2001), pp.498-509.
7. Tone,K.: A slack-based measure of super-efficiency in data envelopment analysis.
Submitted to EJOR, p.21.

Josef Jablonský
Department of Econometrics, Faculty of Informatics and Statistics, University of Economics
Prague, W. Churchill Sq. 4, 130 67 Praha 3, e-mail: jablon@vse.cz, www: nb.vse.cz/~jablon

118