International Conference on Knowledge Engineering and Decision Support

DATA ENVELOPMENT ANALYSIS (DEA): A TOOL FOR MEASURING

EFFICIENCY AND PERFORMANCE
Jorge A. Santos
Phd student at: Departamento de Matemtica Universidade de vora
vora Portugal
[email protected]

Jos H. Dul
School of Business Administration University of Mississippi
University, MS 38677
[email protected]
Abstract This paper introduces Data Envelopment
Analysis and highlights its potentialities to evaluate the
performance of organisations, it describes too the concept
of superefficiency an extension of DEA which increases
its performance.

The present paper is structured as follows. The next

section describes the development and fields of
application of the technique, while section III
introduces the DEA models followed by a numerical
example and a final section about superefficiency
evaluation, an extension of DEA. The readership not
familiar with DEA, a may find the brief introduction to
the method presented below usefull, but for those who
wish to follow the matter further there is a good review
of DEA in Boussofiane et al [4].

Keywords: Data Envelopment Analysis, Efficiency

Evaluation.
I. INTRODUCTION
Data Envelopment Analysis (DEA) is a
mathematical programming based technique to evaluate
the relative performance of organisations. While the
main applications have been in the evaluation of notfor-profit organisations, the technique can be
successfully applied to other organisations, as a recent
evaluation of Russian industry has demonstrated [1].
With this paper we have two objectives in mind. The
first one is to present DEA-Data Envelopment Analysis,
a technique which may have useful applications in
many evaluation contexts, namely when assessing not
for profit organizations. In addition to allowing the
ranking of the organizations traditionally termed
decision-making units, DEA also creates the conditions
to improve performance through target setting and rolemodel identification.
The second objective is to describe briefly the
technique of deleted domain, also known as
superefficiency.
DEA is suited for this type of evaluation because it
enables results to be compared making allowances for
factors [2]. DEA makes it possible to identify efficient
and inefficient units in a framework where results are
considered in their particular context. In addition, DEA
also provides information that enables the comparison
of each inefficient unit with its "peer group", that is a
group of efficient units that are identical with the units
under analysis. These role-model units can then be
studied in order to identify the success factors which
other comparable units can attempt to follow.
Thanassoulis et al [3] argue that DEA is preferable to
other methods, such as regression analysis, which also
make it possible to contextualize results.

II. HISTORY AND APPLICATIONS OF DEA

DEA is a mathematical programming technique
presented in 1978 by Charnes, Cooper and Rhodes [5],
although its roots may be found as early as 1957 in
Farrel`s seminal work [6]. This technique is usually
introduced as a non-parametric one, but in fact it rests
on the assumption of linearity [7] and for the original
models even in the more stringent assumption of
proportionality.
Its application has been focused mainly on the
efficiency assessment of not-for-profit organizations,
since these cannot be evaluated on the basis of
traditional economic and financial indicators used for
commercial companies.
The first application of DEA was in the field of
Education, in the analysis of the Program Follow
Through, conducted in the USA, in the late seventies
[8]. Since then it has been used to assess efficiency in
areas such as health [9, 10], prisons [11], courts [12],
universities and many other not-for-profit sectors.
Nowadays DEA can be seen to have spread to other
fields such as Transit [13], Mining [14], Air
Transportation [15], and even Banking [16].
However, many applications belong to the education
field and range from primary education [17, 18], to
secondary [19, 20, 21] and university levels [22].
In Data Envelopment Analysis the organizational
units to be assessed should be relatively homogeneous
and were originally termed Decision Making Units. As
the whole technique is based on comparison of each
DMU with all the remaining ones a considerable large

423

International Conference on Knowledge Engineering and Decision Support

set of units is necessary for the assessment to be

meaningful. We will assume that each DMU produces
N outputs by means of M inputs.

This Fractional Linear Program can be solved by

means of the Charnes and Cooper transformation [25]
which yields the following Linear Program:

III. DEA FORMULATIONS

In DEA, efficiency (hj) of a specific decision
making unit (DMU j) is defined as the ratio between a
weighted sum of its N outputs Ynj and a weighted sum
of its M inputs Xmj, a natural extension of the concept
of efficiency used in the fields of physics and
engineering [23]:

Max

s.t.

m =1 mj '

mj '

j'

nj ' nj '

DMU
1
2
3
4
5
6
7
8
9
10
11
12

(2)

x
m =1 mj '

mj '

s.t. mj > 0
nj > 0

m=1...M
n=1...N

(3)
(4)

j=1...J

(5)

v y
n =1
M

nj ' nj '

x
m=1 mj '

mj '

x
m =1 mj '

mj '

=1

x
m =1 mj '

mj

j=1...J

m=1...M
Eq. 1.1
n=1...N

(9)
(10)
(11)

Table I- Outputs normalised by Input X1

v y
n =1
M

IV. NUMERICAL EXAMPLE

To illustrate the Data Envelopment Analysis
technique, an example is introduced in Table I, with 12
DMUs producing two Outputs Y1 and Y2 from a single
Input X1, under the assumption of constant returns to
scale, which simply means that if one doubles the Inputs
of any unit it would be expected that its Outputs would
also double. In algebraic form this can be stated as: if xj
yields Outputs yj than Inputs kxj should produce
Outputs kyj.

nj ' nj

(8)

The problem above is known as the multiplier

problem, since its unknowns are the weights, which are
usually lower bounded by a small quantity (Eq: 10-11)
so that all Inputs and Outputs are considered in the
evaluation [24], even if with a minor weight , set in all
the following formulations equal to 10-6.
The dual of this problem, which we shall call the
envelopment problem, provides important information
about economies that could be achieved in all the
inputs; it also indicates which efficient units the
inefficient unit being assessed should emulate. Those
efficient units are usually referred to as the reference set
or peer group of the unit under evaluation.

When assessing a set of J organisations, where Xmj

th
th
stands for the m input of the j DMU, with a similar
meaning for Ynj , the weights mj and nj, in Eq (1),
are chosen for each j DMU under evaluation as those
that maximize its efficiency as defined by hj. Several
constraints have to be added to the maximization
problem:

The strict positivity [24] of the weights

mj ,nj (also known as virtual multipliers).

For scaling purposes, all J DMUs under

analysis must have efficiencies not exceeding an agreed
value, typically one or 100%, as is usual in engineering
definitions of efficiency.

A third kind of restriction has to be included

since otherwise this linear fractional program would
yield an infinite number of solutions. In fact, if a set of
weights mj ,nj returns the optimal solution, so
would kmj , knj. Making the denominator, in Eq
(1), equal to one or 100%, circumvents this situation.
So, we have to solve the following maximization
problem for each one of the J DMUs under analysis:

Max

=1

mj > 0
nj > 0

(1)

mj '

v y
n =1

nj ' nj '

n=1
M

x
m =1 mj '
N

v y

h=

(7)

nj ' nj '

n =1

'

h j' = v y

(6)

424

x1
4.0
5.0
6.0
10.0
11.0
8.0
9.0
5.0
5.5
8.0
10.0
8.0

y1
2.0
1.0
6.0
8.0
7.0
6.0
7.0
3.0
4.4
4.0
2.0
1.0

y2
28.0
22.5
12.0
60.0
16.5
12.0
6.0
30.0
5.5
72.0
20.0
4.0

y1/x1
0.500
0.200
1.000
0.800
0.636
0.750
0.778
0.600
0.800
0.500
0.200
0.125

y2/x1
7.000
4.500
2.000
6.000
1.500
1.500
0.667
6.000
1.000
9.000
2.000
0.500

International Conference on Knowledge Engineering and Decision Support

This related problem can be written as follows:

In this simple example we can normalise the Outputs

by the only Input and plot them in the plane. From
Figure 1 it is easy to understand the reason for naming
this technique Data Envelopment Analysis; in fact each
DMU is analysed against the envelope of the most
efficient units. For instance, the efficiency of DMU 8 is
0.857 (see Table II). This means that it could reduce its
input to 85.7% of its current value reaching its target,
Ci 8 (where Ci stands for Composite unit under
minimisation of inputs), which is the same as for
DMU 11 except that, for the later, a reduction to 28.6%
of X1s current level of inputs would be necessary for
this DMU to become efficient, since its efficiency is
only 0.286.
9
8
7
6
5
4
3
2
1
0

Ci2

s.t.

12
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Ci9
0.9

y1/x

Figure 1- Efficiency frontier and radial projections for

inefficient units.

Althought the results could be obtained graphically

we present in table II the results obtained by a standard
linear optimisation software
Table II- Results for the multiplier problem

DMU
1
2
3
4
5
6
7
8
9
10
11
12

1
0.250
0.200
0.167
0.100
0.091
0.125
0.111
0.200
0.182
0.125
0.100
0.125

1
0.179
0.000
0.152
0.091
0.083
0.114
0.111
0.143
0.182
0.089
0.071
0.114

ynj Snj+ ' = ynj

n=1...N

(14)

j=1...J

(15)

+
nj '

n=1...N

(16)

mj ' 0

m=1...M

(17)

S 0

6 9
7

(13)

j 0

x + Smj ' = Zj xmj m=1...M

j mj

j =1

2
11

Ci8;Ci11

(12)

j =1

10
1

+
S
+
Smj '

nj '
m =1
n =1

min: Zj -

Efficiency
2
0.018
0.857
0.022
0.500
0.008
1.000
0.005
1.000
0.004
0.647
0.006
0.750
0.000
0.778
0.014
0.857
0.000
0.800
0.009
1.000
0.007
0.286
0.006
0.136

This formulation can be interpreted as follows: given

DMUj find the Composite unit which has no smaller
outputs than this one and whose inputs are smaller than
those of DMUj scaled down by a factor Zj as little as
possible. This is why this formulation is known as Input
minimisation; since we are minimising Zj, we are
seeking the minimal inputs that, based on best achieved
performance, could still produce the same amount of
Outputs as DMUj is currently doing. These Composite
units are linear combinations of efficient units which lie
on the efficiency frontier. These efficient DMUs are
known as the peer group, the role model units that
inefficient DMUj should try to emulate.
However, some DMUs cannot be expressed as a
linear combination (with positive weights) of the
efficient units. In our example these are DMUs 2, 7 and
9 (see Figure 1). The goal for these inefficient units is
an equally proportional contraction of their Input and
the increase in Output 1 or 2; S1j+ or S2j+ , known as
Slacks, their name in standard Linear Programming
terminology.
The Input minimisation for the above envelopment
problem would return the values presented in Table III
(all the absent s are null).
Table III- Dual Variables(s), Slacks and Efficiency

DMU
1
2
3
4
5
6
7
8
9
10
11
12

We minimised the consumption of X1, the only

Input, so that DMU 8 (or 11) become efficient. This
contrasts clearly with Eq (2) where we have a
maximisation problem. The reason for this is simply
that we have been using the dual of our original
problem (for a generic introduction to duality in Linear
Programming see [26] or [27]).

425

10

0.0952 0.3095
0.3125
1
1
1.0909 0.0568
1
1.1667
0.2857 0.1786
0.7333
1
0.1905 0.1190
0.1061 0.0455

S1+ S2+ Efficiency

0.857
0.25
0.500
1.000
1.000
0.647
0.750
8.00
0.778
0.857
3.30
0.800
1.000
0.286
0.136

International Conference on Knowledge Engineering and Decision Support

This extension to DEA was first suggested by

Andersen and Petersen [30] and its use is strongly
recommended by the authors as a consequence of its
simplicity and usefulness.
By using superefficiency, it is possible to rank all
units, even the efficient ones, that by standard DEA
techniques would all be rated as equal - their efficiency
having reached the top value of 100%.
To exemplify the benefits of using superefficiency let
us look at two examples of the CCR model represented
in Figures 2 and 3. While DMU C in Figure 2 is clearly
efficient and robust, since it would still continue to be
efficient if the output were to decrease until OC/OCr,
the same does not happen with the case represented in
Figure 3 where a small variation in OC'/OC'r could
bring it under 100% which would mean that DMU C'
would become inefficient.

In this model DMUs 1 and 8 can be thought of as

combinations of efficient DMUs 4 and 10. Or, in other
words, the later being the peers with which the former
should be compared. This reflects the basic idea of DEA
- each unit is evaluated by comparison with other
similar ones; if other DMUs with similar inputs can
achieve higher outputs so should the one being
evaluated. Note that inputs and outputs should include
all relevant variables. For a discussion of the choice of
variables see Norman and Stoker [28].
As a whole the interpretation of the DEA technique
is straightforward and can be put in the following terms:

Multiplier problem
Evaluate each DMU with the set of weights which
maximises its efficiency, provided that all other DMUs,
rated with that set of weights, have efficiency not
greater than unity.

Envelopment problem
Find the smallest proportion of Inputs that would
bring the current DMUj to the enveloping surface of all
DMUs.
The model presented above, named CCR after the
authors: Charnes Cooper and Rhodes [23], assumes
constant returns to scale. This means that when the
input of an efficient unit is multiplied by a given factor
its output level is also multiplied by the same factor. In
many situations this is not the case. The scale of
operations may have an impact on the outputs creating
"economies" or "diseconomies" of scale. The BCC
model, developed by Banker, Charnes and Cooper [29]
can deal with variable returns to scale. The example
presented above assumes constant returns to scale.
Great care should be taken when applying DEA,
since it is an extreme point technique, very sensitive to
errors on outliers. Since DEA permits substitution
between Outputs or Inputs, these should be very
carefully chosen and kept to as few as possible since
otherwise units deemed as efficient are merely
comparing themselves with themselves. In extreme
cases one can be faced with the situation of all units
being efficient, especially when the ratio: [number of
variables]/[number of units] increases. This weakness is
overcome by the judicious selection of Inputs and
Outputs. In addition many applications also limit the
range of variation of the weights in order to overcome
this problem as will be discussed later.

Out2
inp
7

Cr

5
4
3

2
1

8 Out1
inp

Figure 2- Example of a DMU, efficient and robust

This ratio (superefficiency) is in fact a measure of the

robustness of the efficient units that can be used to rank
these units.
Out2'
inp'
7

A'
C'

C'r

5
4
3

B'

V. SUPEREFFICIENCY
We arrive at the concept of superefficiency by
allowing the efficiency of the DMU being assessed, to
be greater than unity. This is achieved by removing the
corresponding constraint from the set of J constraints in
Eq (9) Superefficiency only affects units deemed as
efficient as the removed constraint is not binding for the
inefficient units since their efficiency is, by definition,
less than unity.

8 Out1'
inp'

Figure 3- Example of a DMU, efficient but not robust

For the example presented in the previous section the

superefficiency for the 3 efficient DMUs would be as
presented in table IV.

426

International Conference on Knowledge Engineering and Decision Support

Table IV- Superefficiency scores for the efficient units

Unit

Superefficienc
y
Unit 4
107.50%
Unit 3
125.00%
Unit 10
128.57%
Units 3 and 10 are efficient and robust while any
small increase in the Input or decrease in the Outputs of
Unit 4 may make it inefficient.
An important additional benefit from this extension
to the DEA model is that the set of weights, is uniquely,
determined for the efficient units in all practical
applications [21].

[10]

[11]

[12]

[13]
REFERENCES
[1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

Paterson, I., Data envelopment analysis of

enterprises in the Russian metallurgy sector, IHS
(Institute of Advanced Studies, Vienna), Working
Paper.
Thanassoulis, E. and P. Dunstan, (1994), Guiding
schools to improved performance using data
envelopment analysis: An illustration with data
from a local education authority, Journal of the
Operational Research Society, 45, (11), 12471262.
Thanassoulis, E., (1993), Comparison of
regression analysis and data envelopment analysis
as alternative methods for performance
assessments, Journal of the Operational
Research Society , 44, (11), 1129-44.
Boussofiane, A., R. G. Dyson, Thanassoulis, E.,
(1991), Applied data envelopment analysis,
European Journal of Operational Research, 52,
(1), 1-15.
Charnes, A., Cooper, W. W., Rhodes, E., (1978),
Measuring the efficiency of decision making
units, European Journal of Operations Research,
2, (6), 429-44.
Farrell, M. J., (1957), The measurement of
productive efficiency, Journal of Royal
Statistical Society A, 120, 253-281.
Chang, K.-P. and Y.-Y. Guh, (1991), Linear
production functions and the data envelopment
analysis, European Journal of Operational
Research, 52, (2), 215-23
Rhodes, E., (1978), Data envelopment analysis
and related approaches for measuring the
efficiency of decision-making units with an
application to program follow through in U.S.
education, unpublished doctoral thesis, School of
Urban and Public Affairs, Carnegie - Mellon
University.
Nunamaker, T. R. and A. Y. Lewin, (1983),
Measuring routine nursing service efficiency: A

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]

427

comparison of cost per patient day and data

envelopment analysis models, Health Services
Research, 18, (2 (Part 1)), 183-208.
Thanassoulis, E., A. Boussofiane, R. G.
Dyson,(1995), Exploring output quality targets in
the provision of prenatal care in England using
data envelopment analysis, European Journal of
Operational Research, 80, (3), 588-607.
Ganley, J. A., Cubbin, J. S., (1992), Public Sector
Efficiency Measurement - Applications of Data
Envelopment Analysis, North - Holland, 57-80.
Lewin, A. Y. and R. C. Morey, Cook, T.J., (1982),
Evaluating the Administrative Efficiency of
Courts, Omega, 10 ,(4), 401-11.
Chu, X. and G. J. Fielding, (1992), Measuring
transit performance using data envelopment
analysis, Transportation Research Part A (Policy
and Practice), 26A, (3), 223-30.
Thompson, R. G., Dharmapala, P. S., Thrall, R.,
(1995), Sensitivity analysis of efficiency measures
with applications to Kansas farming and Illinois coal
mining, Data Envelopment Analysis: Theory,
Methodology and Applications,. A. Charnes, W. W.
Cooper, A. Lewin and L. M. Seiford. Boston,
Kluwer Academic Publishers, 89-94.
Schefczyk, M., (1993), Operational performance
of airlines: an extension of traditional
measurement paradigms, Strategic Management
Journal, 14, 301-317.
Vassiloglou, M. and D. Giokas, (1990), A study
of the relative efficiency of bank branches: an
application of data envelopment analysis,
Journal of the Operational Research Society , 41,
( 7), 591-7.
Bessent, A., Bessent, W., Kennington, J., e Regan,
B., (1982), An application of Mathematical
Programming to assess productivity in the
Houston
independent
school
district,
Management Science, Vol. 28, No. 12, 1355-67.
Charnes, A., Cooper, W.W., Rhodes, E., (1981),
Evaluating program and managerial efficiency:
an application of Data Envelopment Analysis to
Program Follow Through, Management Science,
27, (6), 668-97.
Charnes, A., Cooper, W.W., e Rhodes, E., (1978),
Measuring the efficiency of decision making
units, European Journal of Operational
Research, 2, (6), 429-44.
Castro, R., (1993), O mtodo DEA - Aplicao
avaliao da eficincia comparativa das escolas
secundrias do distrito do Porto, unpublished
Msc. thesis, Faculdade de Economia, University
of Oporto.
Santos, J., (1994), Ordenao de unidades eficientes
por tcnicas de data envelopment analysis,

International Conference on Knowledge Engineering and Decision Support

[22]
[23]

[24]

[25]

[26]

unpublished Msc. thesis, Instituto Superior Tcnico,

Technical University of Lisbon.
Beasley, J.E., (1990), Comparing university
departments, Omega, 18, 171-83.
Charnes, A., Cooper, W.W., Rhodes, E., (1978),
Measuring the efficiency of decision making
units, European Journal of Operational
Research, 2, (6), 429-44.
Charnes, A., Cooper, W. W., Rhodes, E., (1979), Short
communication: measuring the efficiency of decision
making units, European Journal of Operational
Research, 3, (4), 339.
Charnes,
A.,
Cooper,
W.W.,
(1962),
Programming with linear fractional functionals,
Naval Research Logistics Quarterly, 9, 181-185.
Gass, S. I. & Harris, C. M., Encyclopedia of
Operations Research and Management Science,
1996

[27] Frederick S. Hillier, Gerald J. Lieberman,

Introduction to Operations Research, Mc Graw
Hill, 1990.
[28] Norman, M., Stoker ,B., Data Envelopment
Analysis: the Assessment of Performance, John
Wiley, Chichester, 1991.
[29] Banker, R. D., Charnes, A., Cooper, W. W.,
(1984), Some models for estimating technical
and scale inefficiencies in data envelopment
analysis, Management Science, 30, (9), 10781092.
[30] Andersen, P., Petersen, N.C., (1993), A
procedure for ranking efficient units in data
envelopment analysis, Management Science, 39
,(10), 1261-4.

428