Data Envelopment Analysis (DEA) : Jean-Marc Huguenin

Jean-Marc Huguenin
Data Envelopment Analysis (DEA)

A pedagogical guide for decision makers
in the public sector
IDHEAP – Cahier 276/2012

Chair of Public finance
 2012 IDHEAP, Lausanne

ISBN 978-2-940390-54-0
CONTENTS
LIST OF TABLES
.......................................................7
..................................................................................32
.....................................................................40
...................................................66
LIST OF FIGURES
.................. 10
.................................... 12
............................................................. 14
............................................................................................................... 16
.....................................................................................18
.....................................................................................19
...............................................................................................21
................................22
...........................................33
................34
...........................................................................................36
.........................38
................................................................................................39
.....................................................................................42
..................44
..................45
..................46
........................47
...............53
............................................58
...................................................................................65
...................................................................................68
........................69
1 INTRODUCTION
This guide introduces Data Envelopment Analysis (DEA), a
performance measurement technique, in such a way as to be appropriate
to decision makers with little or no background in economics and
operational research. The use of mathematics is kept to a minimum. This
guide therefore adopts a strong practical approach in order to allow
decision makers to conduct their own efficiency analysis and to easily
interpret results.
DEA helps decision makers for the following reasons:
- By calculating an efficiency score, it indicates if a firm is
efficient or has capacity for improvement.
- By setting target values for input and output, it calculates how
much input must be decreased or output increased in order to
become efficient.
- By identifying the nature of returns to scale, it indicates if a
firm has to decrease or increase its scale (or size) in order to
minimize the average cost.
- By identifying a set of benchmarks, it specifies which other
firms’ processes need to be analysed in order to improve its
own practices.
After this introduction, Chapter 2 presents the essentials about DEA,
alongside a case study to intuitively understand its application.
Chapter 3 introduces Win4DEAP, a software package that conducts
efficiency analysis based on DEA methodology. Chapter 4 is dedicated
to more demanding readers interested in the methodical background of
DEA. Four advanced topics of DEA (adjustment to the environment;
preferences; sensitivity analysis; time series data) are presented in
Chapter 5. Finally, Chapter 6 shows how to program the Solver in
Microsoft Excel © in order to run a basic DEA efficiency analysis.
2 BASICS OF DEA
DEA is used to measure the performance of firms or entities (called

Decision-Making Units –DMUs–) which convert multiple inputs into
multiple outputs. Firm efficiency is defined as the ratio of the sum of its
weighted outputs to the sum of its weighted inputs (Thanassoulis et al.,
2008, p. 264). DEA is suitable for the use of both private sector firms
and public sector organizations (and even for entities such as regions,
countries, etc.)1. It was formulated in Charnes et al. (1978, 1981) in
order to evaluate a US federal government program in the education
system called ‘Program Follow Through’. The use of DEA then spread
to other public organizations (hospitals, aged-care facilities, social
service units, unemployment offices, police forces, army units, prisons,
waste management services, power plants, public transportation
companies, forestry companies, libraries, museums, theatres, etc.) and to
the private sector (banks, insurance companies, retail stores, etc.).
Each firm’s efficiency score is calculated relative to an efficiency
frontier. Firms located on the efficiency frontier have an efficiency score
of 1 (or 100%). Firms operating beneath the frontier have an efficiency
score inferior to 1 (or 100%) and hence have the capacity to improve
future performance. Note that no firm can be located above the
efficiency frontier because they cannot have an efficiency score greater
than 100%. Firms located on the frontier serve as benchmarks –or
peers– to inefficient firms. These benchmarks (i.e. real firms with real
data) are associated with best practices. DEA is therefore a powerful
benchmarking technique.
1
To keep it simple and to make this pedagogical guide more understandable, the term
‘firm’ is used in a generic way.
To better understand the mechanics behind DEA, this section develops a
simple practical case study. It includes only one input and one output,
although DEA can handle multiple inputs and multiple outputs.
Five register offices (A to E) produce one output (total number of
documents, such as marriage or birth certificates) with one input
(number of full-time equivalent public servants)2. The data are listed in
Table 1. For example, two public servants work in Register Office A.
They produce one document (during a certain period of time).
Input Output
Register Office
Public servants (x) Documents (y)
A 2 1
B 3 4
C 5 5
D 4 3
E 6 7
a) Case study 1 – Two basic DEA models

Two basic models are used in DEA, leading to the identification of two
different frontiers:
- The first model assumes constant returns to scale technology
(CRS model). This is appropriate when all firms are operating
at an optimal scale. However, note that this is quite an
ambitious assumption. To operate at an optimal scale, firms
should evolve in a perfectly competitive environment, which is
seldom the case. The CRS model calculates an efficiency score
called constant returns to scale technical efficiency (CRSTE).
- The second model assumes variable returns to scale
technology (VRS model). This is appropriate when firms are
2 Note that DEA can handle more outputs and inputs. In order to represent this example
in a two-dimensional graph, we consider a total of two outputs and inputs of two (one
output, one input; no variable representing the quality of the variables).
not operating at an optimal scale. This is usually the case when
firms face imperfect competition, government regulations, etc.
The VRS model calculates an efficiency score called variable
returns to scale technical efficiency (VRSTE).
Comparison between the two models reveals the source of inefficiency.
Constant returns to scale technical efficiency corresponds to the global
measure of firm performance. It is composed by a ‘pure’ technical
efficiency measure (captured by the variable returns to scale technical
efficiency score) and a scale efficiency measure (SE). Section 4.2
demonstrates how these three notions (CRSTE, VRSTE and SE) relate
to each other.
b) Case study 1 – Input or output orientation

A DEA model can be input or output oriented:
- In an input orientation, DEA minimizes input for a given level
of output; in other words, it indicates how much a firm can
decrease its input for a given level of output.
- In an output orientation, DEA maximizes output for a given
level of input; in other words, it indicates how much a firm can
increase its output for a given level of input.
The efficiency frontier will be different in a CRS or a VRS model (see
Section 4.2). However, within each model, the frontier will not be
affected by an input or an output orientation. For example, the efficiency
frontier under VRS will be exactly the same in an input or an output
orientation. Firms located on the frontier in an input orientation will also
be on the frontier in an output orientation.
In a CRS model, technical efficiency scores have the same values in an
input or an output orientation. But these values will be different
according to the model’s orientation when VRS is assumed. However,
Coelli and Perelman (1996, 1999) note that, in many instances, the
choice of orientation has only a minor influence upon the technical
efficiency scores calculated in a VRS model.
Choosing between an input or an output orientation
The model’s orientation should be selected according to which variables
(inputs or outputs) the decision maker has most control over. For
example, a school principal will probably have more control over his
teaching staff (input) than over the number of pupils (output). An input
orientation will be more appropriate in this case.
In the public sector, but sometimes also in the private, a given level of
input can be granted and secured to a firm. In this case, the decision
maker may want to maximize the output (and therefore choose an output
orientation). Alternatively, if the decision maker’s task is to produce a
given level of output (e.g. a quota) with the minimum input, he will opt
for an input orientation.
If the decision maker is not facing any constraints and has control of
both input and output, the model’s orientation will depend on his
objectives. Does he need to cut costs (input orientation) or does he want
to maximize production (output orientation)?
c) Case study 1 – CRS efficient frontier

Figure 1 represents the efficient frontier assuming constant returns to
scale technology (CRS efficient frontier). The CRS efficient frontier
starts at the origin and runs through Register Office B. Register Office B
happens to be the observation with the steepest slope, or the highest
productivity ratio, among all register offices (4 / 3 = 1.33, meaning that
one public servant produces 1.33 documents). Register Office B is on
the frontier; it is 100% efficient. Register Offices A, C, D and E are
beneath the frontier. Their respective efficiency scores are less than
100%. DEA assumes that the production possibility set is bounded by
the frontier. This actually implies that DEA calculates relative and not
absolute efficiency scores. Although firms on the efficient frontier are
granted a 100% efficiency score, it is likely that they could further
improve their productivity.
CRS efficient frontier
8
7 E
6
Documents
5 C
4 B
3 D
ACRS-O
2
Production possibility set
1 A
S ACRS-I
T
0
0 1 2 3 4 5 6 7
Public servants
Figure 1 also illustrates how DEA measures efficiency scores. The

example of Register Office A is described below:
- In an input orientation, A’s efficiency score is equal to the
distance SACRS-I divided by the distance SA. ACRS-I is the
projection of point A on the efficient frontier (assuming
constant returns to scale –CRS– and an input orientation –I).
Note that one can easily calculate efficiency scores using a
ruler and measuring the distances on the graph. A’s score is
37.5%. This means that Register Office A could reduce the
number of public servants employed (input) by 62.5%
(100 - 37.5) and still be able to produce the same number of
documents (one).
- In an output orientation, A’s efficiency score is equal to the
distance TA divided by the distance TACRS-O. ACRS-O is the
projection of point A on the efficient frontier (assuming
constant returns to scale –CRS– and an output orientation –O–).
A’s score is 37.5%, as in an input orientation3. This means that
Register Office A could increase its production of documents
(output) by 62.5% (100 - 37.5) whilst holding the number of
public servants constant at two.
d) Case study 1 – VRS efficient frontier

Figure 2 represents the efficient frontier assuming variable returns to
scale technology (VRS efficient frontier). The VRS efficient frontier is
formed by enveloping all the observations. Register Offices A, B and E
are on the frontier. They are 100% efficient. Register Offices C and D
are beneath the frontier. Their respective efficiency scores are inferior to
100%. DEA assumes that the production possibility set is bounded by
the frontier. Again, this implies that DEA calculates relative and not
absolute efficiency scores. Although firms on the efficient frontier are
granted a 100% efficiency score, it is likely that they could further
improve their productivity.
3 Note that the efficiency scores in a CRS model are always the same for an input or an
output orientation.
8
7
VRS efficient frontier E
6
Documents
5 C
DVRS-O
4 B
3 D
U DVRS-I
2
Production possibility set
1 A
0 V
0 1 2 3 4 5 6 7
Public servants
Figure 2 also illustrates how DEA measures efficiency scores. The

example of Register office D is described below:
- In an input orientation, D’s efficiency score is equal to the
distance UDVRS-I divided by the distance UD. DVRS-I is the
projection of point D on the efficient frontier (assuming
variable returns to scale –VRS– and an input orientation –I–).
Note that one can easily calculate efficiency scores using a
ruler and measuring the distances on the graph. D’s score is
66.7%. This means that Register Office D could reduce the
number of public servants employed (input) by 33.3%
(100 - 66.7) and still be able to produce the same number of
documents (three).
- In an output orientation, D’s efficiency score is equal to the
distance VD divided by the distance VDVRS-O. DVRS-O is the
projection of point D on the efficient frontier (assuming
variable returns to scale –VRS– and an output orientation –O–).
D’s score is 60%4. This means that Register Office D could
increase its production of documents (output) by 40%
(100 - 60) whilst holding the number of public servants
constant at four.
How to interpret efficiency scores according to the DEA model’s

output or input orientation
Register Office C has an efficiency score of 75% in the CRS model. It
will get the same efficiency score in an output or in an input-oriented
model under the constant returns to scale assumption. However:
- In the input-oriented model, the capacity to improve input (i.e. a
reduction) by 25% (100 - 75) is calculated using the original input
value of 5 public servants. The DEA model calculates a projected
value of 3.75. The 25% improvement is then calculated according to
the original value: [((5 - 3.75) / 5) x 100] = 25. From a practical point
of view, the capacity to improve input by 25% means that the Register
Office should reduce all of its inputs by 25% in order to become
efficient.
- In the output-oriented model, the capacity to improve output (i.e. an
augmentation) by 25% (100 - 75) is calculated using the projected
output value. Register Office C has an original output value of
5 documents. The DEA model calculates a projected value of
6.67 documents. The 25% improvement is calculated according to the
projected value: [((6.67 - 5) / 6.667) x 100] = 25. From a practical
point of view, the capacity to improve output by 25% means that the
Register Office should augment all of its outputs by 25% in order to
become efficient.
4 Note that the efficiency scores in a VRS model are different for an input or an output
orientation.
e) Case study 1 – CRS, VRS and scale efficiency
Figure 3 represents the CRS and the VRS efficient frontiers on the same
graph. Register Office B is CRS and VRS efficient, as it is located on
both frontiers. Register Offices A and E are efficient under the variable
returns to scale assumption but inefficient under the constant returns to
scale assumption. Finally, Register Offices D and C are both CRS and
VRS inefficient; they are located neither on the CRS nor on the VRS
frontiers.

8
7
E
6 VRS efficient frontier

Documents
5 C
4 B
DCRS-I
3 D
U DVRS-I
2
ACRS-I
1 A
0
0 1 2 3 4 5 6 7
Public servants
The gap observed between the CRS and the VRS frontiers is due to a
problem of scale. For example, Register Office A is VRS efficient. To
become CRS efficient, Register Office A should modify its scale (or
size). Only by operating at point ACRS-I would Register Office A be as
productive as Register Office B, which is the only CRS efficient
Register Office.
Some Register Offices (D and C) are not even located on the VRS
frontier. These Register Offices not only have a scale problem but are
also poorly managed. For example, Register Office D should move to
point DVRS-I located on the VRS frontier in order to become VRS
efficient (i.e. to eliminate the inefficiency attributable to poor
management). Furthermore, Register Office D should move from point
DVRS-I to point DCRS-I located on the CRS frontier in order to become
CRS efficient (i.e. to eliminate the inefficiency attributable to a problem
of scale).
As a result, the CRS efficiency (also called ‘total’ efficiency) can be
decomposed into two components: the VRS efficiency (also called
‘pure’ efficiency) and the scale efficiency. The following ratios
represent these three types of efficiency for Register Office D (input
orientation).
Technical efficiency of Technical efficiency of Scale efficiency of D
D under CRS D under VRS
UDCRS − I UDVRS − I UDCRS − I
TECRS = = 56.3% TEVRS = = 66.7% SE = = 84.4%
UD UD UDVRS − I
f) Case study 1 – Nature of returns to scale

The nature of returns to scale of register offices not located on the CRS
frontier (in other words, scale inefficient) has to be identified. Figure 4
represents the CRS efficient points ACRS-I and ECRS-I of Register Offices
A and E (which are CRS inefficient but VRS efficient). It also represents
the CRS efficient points DCRS-I and CCRS-I and the VRS efficient points
DVRS-I and CVRS-I of Register Offices D and C (which are CRS and VRS
inefficient).
8
ECRS-I
7
E
VRS efficient frontier
6
DRS
Documents
CCRS-I
5 C
CVRS-I
4 B
DCRS-I
3 D
DVRS-I
2 IRS
ACRS-I
1 A
0
0 1 2 3 4 5 6 7
Public servants
To identify the nature of returns to scale, one has to focus on the slope
of the VRS efficient points A, DVRS-I, B, CVRS-I and E (or productivity).
Three situations can occur:
- A register office is located both on the CRS and the VRS
efficient frontiers (such as point B). Register Office B has the
highest productivity of all VRS efficient points (4 / 3 = 1.33). It
is facing constant returns to scale. Such a firm reaches its
optimal size (or efficient scale)5. It is operating at a point where
the scale (or size) has no impact on productivity. This situation
occurs when the average inputs consumption is minimized and
does not vary with output. In a situation of constant returns to
5 In the economic context, a firm operates at the optimal size (or efficient scale) when it
minimizes its average cost. In the context of DEA, we can measure efficiency in
physical or in monetary terms. Because cost and price information is not always
available or appropriate, the use of technical efficiency is often preferred. As this latter
measure is based on physical terms, we prefer to use the expression of average inputs
consumption instead of average cost.
scale, an increase in output of 1 percent requires a
proportionate increase in input (i.e. 1 percent).
- A register office (or the projected point of a register office) is
located at a point where the scale (or the size) has a positive
impact on productivity. Points A and DVRS-I are in such a
position (see Figure 5). The productivity of A (1 / 2 = 0.5) is
inferior to the productivity of DVRS - I (3 / 2.67 = 1.12). The ratio
of productivity is increasing with the scale. This situation
occurs until point B, which has a productivity of 1.33. Register
Offices A and D are therefore facing increasing returns to scale
(IRS) –or economies of scale–. In this situation, the average
inputs consumption declines whilst output rises. Register
Offices A and D have not yet reached their optimal size (or
efficient scale). To improve their scale efficiency, they have to
expand their output. In a situation of economies of scale, a
variation in output of 1 percent results in a variation in input of
less than 1 percent. Hence, an increase in output results in a
reduction of the average inputs consumption.
8
7
E
6
Documents
5 C
CVRS-I
4 B
3 D
DVRS-I
2
1 A
0
0 1 2 2.67 3 4 5 6 7
Public servants
- A register office (or the projected point of a register office) is

located at a point where the scale (or the size) has a negative
impact on productivity. Points CVRS-I and E are in such a
position (see Figure 6). The productivity of CVRS-I (5 / 4 = 1.25)
is superior to the productivity of E (7 / 6 = 1.17). The ratio of
productivity is decreasing with the scale. This situation occurs
from point B, which has a productivity of 1.33. Register
Offices C and E are therefore facing decreasing returns to scale
(DRS) –or diseconomies of scale–. In this situation, the average
inputs consumption rises whilst output rises. Register Offices C
and E have exceeded their optimal size (or efficient scale). To
improve their scale efficiency, they have to reduce their output.
In a situation of diseconomies of scale, a variation in output of
1 percent results in a variation in input of more than 1 percent.
Hence, a decrease in output results in a reduction of the average
inputs consumption.
8
7
E
6
Documents 5 C
CVRS-I
4 B
3 D
DVRS-I
2
1 A
0
0 1 2 3 4 5 6 7
Public servants
The specific cases of the five Register offices are described below (see
Figure 4):
- Register Office A is located on the VRS frontier but not on the
CRS frontier. Its inefficiency is due to an inappropriate scale. A
is facing increasing returns to scale. A variation in output of
1 percent results in a variation in input of less than 1 percent.
- Register Office D is neither located on the CRS nor on the VRS
frontier. Its inefficiency is due to an inappropriate scale and to
poor management. D is facing increasing returns to scale. A
less than 1 percent.
- Register Office B is located both on the CRS and on the VRS
frontier. It has no inefficiency at all. B is facing constant
returns to scale. A variation in output of 1 percent results in a
variation in input of 1 percent.
- Register Office C is neither located on the CRS nor on the VRS
frontier. Its inefficiency is due to an inappropriate scale and to
poor management. C is facing decreasing returns to scale. A
more than 1 percent.
- Register office E is located on the VRS frontier (but not on the
CRS frontier). Its inefficiency is due to an inappropriate scale.
E is evolving in a situation of decreasing returns to scale. A
more than 1 percent.
g) Case study 1 – Peers (or benchmarks)

DEA identifies, for each inefficient firm, the closest efficient firms
located on the frontier. These efficient firms are called peers or
benchmarks. If inefficient firms want to improve their performance, they
have to look at the best practices developed by their respective peers.
Under the CRS assumption, Register Office B is the only firm located
on the efficient frontier. Hence it is identified as the peer for all other
inefficient register offices.
Figure 7 illustrates the peers under the VRS assumption. Three Register
Offices (A, B and E) are located on the efficient frontier. Two Register
Offices (C and D) are inefficient. Register Office C has two assigned
peers: B and E. CVRS-I, the projected point of C on the VRS frontier, lies
between these two benchmarks. Register Office D also has two assigned
peers: A and B. DVRS-I, the projected point of D on the VRS frontier, lies
between these two benchmarks.
8
7
E
6
Documents 5 C
CVRS-I
4 B
3 D
DVRS-I
2
1 A
0
0 1 2 3 4 5 6 7
Public servants
h) Case study 1 – Slacks

Particular positions located on the frontier are inefficient. Assume there
is an additional register office in our sample, F. It produces
0.5 document with two public servants. Figure 8 illustrates the efficient
frontier under VRS. Register Office F is not located on the frontier. In
order to become efficient, it has first to move to point FVRS-I without slacks.
At this location, Register Office F should have an efficiency score of
100%, as it is located on the frontier. But Register Office A, next to him
on the frontier, is also 100% efficient. The difference between F and A
is striking. With the same number of inputs (two public servants), F
produces 0.5 document and A produces one document (i.e. 0.5 more
than F). Therefore point FVRS-I without slacks cannot be considered as 100%
efficient, because it produces less output with the same amount of input
than another register office (A). To get a 100% efficiency score, point
FVRS-I without slacks has to move further up to point A. This additional
improvement needed for a firm to become efficient is called a slack.
Indeed, every point located on the sections of the frontier which run
parallel to either the x or the y axes has to be adjusted for slacks. DEA is
designed to take slacks into account.
7
E
6 VRS efficient frontier
Documents
5
C
4
B
3
D
1
Slacks A
F
FVRS-I without slacks
0
0 1 2 3 4 5 6 7
Public servants
DEA allows multiple outputs and multiple inputs to be taken into

account. For example, a shirt company uses machines, workers and
tissue (three inputs) in order to produce T-shirts, pants and underwear
(three outputs). DEA can account for all of these variables and even
more. As a result, DEA goes far beyond the calculation of single
productivity ratios such as, for example, the number of T-shirts
produced per worker (one output divided by one input).
However, the total number of outputs and inputs being considered is not
limitless from a practical point of view. It depends on the number of
firms in the data set. If the number of firms is smaller than, roughly
speaking, three times the sum of the total number of inputs and outputs,
it is highly probable that several firms, if not all, will obtain a 100%
score6. For example, a dataset containing 21 shirt companies allows a
total of seven outputs and inputs to be dealt with (21 divided by 3). As
Cooper et al. (2006, p. 106) point out, “if the number of DMUs (n) is
less than the combined number of inputs and outputs (m + s), a large
portion of the DMUs will be identified as efficient and efficiency
discrimination among DMU is questionable due to an inadequate
number of degrees of freedom. (…). Hence, it is desirable that n exceeds
m + s by several times. A rough rule of numbs in the envelopment
model is to choose n (= the number of DMUs) equal to or greater than
max {m x s, 3 x (m + s)}. ”
DEA measures firm efficiency based on multiple outputs and multiple
inputs. If Shirt Company A produces a lot of T-shirts but only a few
pants and underwear, DEA will automatically attribute a high weighting
to the T-shirts variable in order to emphasize this strength. As a result,
DEA ‘automatically’ optimizes the weighting of each variable in order
to present each firm in the best possible light.
The particularity of DEA is that weights assigned to outputs and inputs
are not decided by users. Moreover, it does not use a common set of
weights for all firms. Instead, a different set of weights is calculated by a
linear optimization procedure.
Unfortunately, DEA does not work with negative or zero values for
inputs and outputs. However, zero values can be substituted with very
low values such as 0.01.
It is also noted that each DMU must have the same number of inputs and
outputs in order to be compared, otherwise DEA cannot be applied.
A distinction has to be made between variables which are under the
control of management (discretionary variables) and variables which are
not (non-discretionary or environmental variables). Ideally, a DEA
model will exclusively include discretionary variables although some
6 The higher the number of inputs and outputs that are taken into consideration for a
given number of firms, the more probable it is that each firm will be the best producer
of at least one of the outputs. Therefore, all firms could obtain a 100% efficiency score.
DEA models can also accommodate non-discretionary. In a second step,
efficiency scores can be adjusted to account for environmental variables
(i.e. such variables influence the efficiency of a firm but are not a
traditional input and are not under the control of the manager).
Moreover, variables should reflect both quantitative and qualitative
characteristics of firms’ resources and services. Although it may not be
easy to identify and to convert qualitative characteristics into numbers, it
is desirable to include such variables in the model in order to
appropriately benchmark firms.
The notion of efficiency refers to an optimal situation; the maximum

output for a given level of input or the minimum input for a given level
of output. Subject to data availability, several types of efficiency can be
measured:
- Technical efficiency, in which both outputs and inputs are
measured in physical terms7.
- Cost efficiency: identical to technical efficiency, except that
cost (or price) information about inputs is added to the model.
- Revenue efficiency: identical to technical efficiency, except
that price information about outputs is added to the model.
- Profit efficiency: identical to technical efficiency, except that
cost information about inputs and price information about
outputs are added to the model.
7 This pedagogical guide will focus on the measurement of technical efficiency for two
main reasons: first, firms in the public sector are often not responsible for the age
pyramid of their employees; therefore taking into account the wages of the employees
(which often grow higher alongside seniority) would unfairly alter efficiency of a firm
with a greater proportion of senior employees; second, firms in the public sector do not
often produce commercial goods or services with a set price.
Technical efficiency is a global measure of firm performance. However,
it does not indicate the source of inefficiency. This source could be
twofold:
- First, the firm could be poorly managed and operated.
- Second, it could be penalised for not operating at the right
scale.
Technical efficiency can be decomposed into a ‘pure’ technical
efficiency measure and a scale efficiency measure to reflect these two
sources of inefficiency8.
DEA is a benchmarking technique. The efficiency scores provide

information about a firm’s capacity to improve output or input. In this
sense, DEA offers strong support to decision making. To conduct an
efficiency analysis and to interpret results often raises practical
questions. The following list of frequently asked questions offers some
advice.
- Is it advisable to involve the managers of the firms to be
benchmarked in the efficiency analysis from the beginning
of the process?
Yes, it is, and for two main reasons. First, managers know the
processes of their firms and the data available. Therefore they
are the right persons to pertinently identify which inputs and
outputs have to be integrated into the analysis. Second,
managers involved from the beginning of the process are more
likely to accept the results of the analysis (rather than to reject
them) if they have been involved in the process.
8 The firm’s management team will definitely be held responsible for the ‘pure’ technical
efficiency score. In a situation where it does not have the discretionary power to modify
the firm’s size, it will likely not be accountable for the scale efficiency score. However,
especially in the private sector, one has the choice of the scale at which it operates: the
management team can easily downsize the firm and, with some efforts, upsize it also.
- How should one respond to managers who claim that their
firms are different from others, and therefore cannot be
compared to them?
Sometimes, inefficiencies can be explained by indisputable
environmental variables. But sometimes they cannot. Managers
often justify the low efficiency scores of their firms by arguing
that their situations are different compared to the situations of
the other firms. They claim to be a ‘special case’ (and therefore
it is acceptable to be inefficient). Actually, the majority of
firms could possibly claim to be different as most possess a
specificity that others do not have. However, it is likely that the
difference of one firm will be compensated by the difference of
another. More generally, it is up to the managers to prove that
they really face a hostile environment. If they cannot prove it,
management measures have to be taken to improve efficiency.
- Assume that a firm obtains an efficiency score of 86.3%.
Does this number have to be strictly applied?
Not really, it should be interpreted more as an order of
magnitude. This order of magnitude informs managers that they
have to increase their outputs or to decrease their inputs in
order to become more efficient. But one should not focus too
strictly on the capacity for 13.7% improvement. Such a number
could be interpreted by practitioners as too ‘accurate’ and may
offend their sensibilities. Therefore it is better to consider
efficiency scores more as more of an objective basis to hold an
open discussion about the way to improve firm efficiency
rather than a number to be strictly applied.
- A firm faces increasing returns to scale. It has economies of
scale. What does that concretely mean from a managerial
point of view?
Such a firm has not yet reached its optimal size. In order to
reduce its average cost (or its average inputs consumption), it
has to increase its size. Practically, this could be done either by
internal growth (i.e. producing more output) or by merging
with another firm which is also facing increasing returns to
scale. If, for some reason, managers cannot influence the scale
of a firm, they should not be held accountable for this source of
inefficiency.
- A firm faces decreasing returns to scale. It has
diseconomies of scale. What does that concretely mean from
a managerial point of view?
Such a firm is already oversized, having exceeded its optimal
size. In order to reduce its average cost (or its average inputs
consumption), it has to decrease its size. Practically, this could
be done either by internal decay (i.e. producing less output) or
by splitting the firm into two separate businesses. Note that
some of the production could be transferred to a firm facing
increasing returns to scale. If, for some reason, managers
cannot influence the scale of a firm, they should not be held
accountable for this source of inefficiency.
- Is efficiency the only criteria to assess a firm’s
performance?
Not necessarily. Basically, the assessment of a firm’s
performance will depend on the management objectives. Other
criteria such as effectiveness or equity are often considered
alongside efficiency. If this is the case, the overall performance
should be balanced with the various criteria.
- One firm obtains a score of 100% but all the others in the
dataset obtain much lower scores (for example, starting at
40% or lower). Is this realistic?
It could be realistic, but the gap appears to be important. In
such a case, data have to be carefully checked, and especially
data of the efficient firm. If a data problem is not identified,
such results mean that the efficient firm is likely to have
completely different processes than the other firms. It should
therefore be absolutely presented as a best practice model.
However, even if they are realistic, such results are likely to be
rejected by managers whose firms have low efficiency scores.
These managers are likely to be discouraged because it is
obviously unrealistic for them to improve their firm’s
efficiency by 60% (or more) in the short run. Therefore it is
better to exclude the efficient firm from the sample and to run a
new model.
- Almost all the firms obtain an efficiency score of 100%.
Does that mean that all of them are really efficient?
Yes, it could mean that all the firms are efficient. Such results
would be great! But they are unlikely. Here, the total number of
inputs and outputs is probably too high compared to the
number of firms in the dataset. In this case, one variable has to
be excluded and a new model has to be run. If the number of
firms obtaining a 100% score decreases, it indicates that the
number of variables was too high compared to the number of
firms. If not, all the firms are just efficient and must be
congratulated.
- The model does not show any results. What does that
mean?
Data has to be checked. This could happen when data with a
value of zero are in the set. Zeros have to be substituted by a
very small number (0.01).
Exercise 1
The following multiple choice questions test one’s knowledge on the
basics of DEA. Only one answer is correct. Solutions are listed at the
end of this exercise.
1. What is the main purpose of DEA?
a) DEA measures firms’ effectiveness
b) DEA measures firms’ efficiency
c) DEA measures firms’ profit
d) DEA measures firms’ productivity
2. A dataset includes information about input quantity, input cost and
output quantity. Which type of efficiency cannot be measured?
a) Technical efficiency
b) Cost efficiency
c) Revenue efficiency
d) Scale efficiency
3. ‘Pure’ technical efficiency reflects:
a) A global measure of firm performance
b) The efficiency of a firm operating at an incorrect scale
c) A measure of profit efficiency
d) The efficiency of a poorly managed firm
4. Firm A is inefficient. Who is its peer(s)?
a) One or several firms whose efficiency scores are worse than
firm A’s efficiency
b) One or several firms whose efficiency scores are better than
firm A’s efficiency, but which are not located on the
efficiency frontier
c) Any firm located on the efficiency frontier
d) One or several specific firms (i.e. a subgroup of efficient
DMUs) located on the efficiency frontier
5. A firm is producing laptops. Which input reflects quality?
a) The number of FTE employees with a Master’s degree
b) Total number of FTE employees, disregarding their
educational background
c) Total number of square meters of the factory
d) Energy consumption
6. A firm has diseconomies of scale. How can the management team
improve its efficiency?
a) By merging with another firm
b) By producing more output
c) By producing less output
d) By producing the same amount of output
7. A manager plans to measure efficiency using three inputs and two
outputs. What is the minimum number of firms that should be
included in the dataset?
a) 10
b) 6
c) 15
d) It does not matter
Correct answers: 1b; 2c; 3d; 4d; 5a; 6c; 7c.

3 DEA SOFTWARE
The user-friendly software packages of DEA incorporate intuitive

graphical user interfaces and automatic calculation of efficiency scores.
Some of them are compatible with Microsoft Excel ©. For a survey of
DEA software packages, one can refer to Barr (2004). Today, several
software packages have been developed:
- Free packages include DEAP (Timothy Coelli, Coelli
Economic Consulting Services) and Win4DEAP (Michel
Deslierres, University of Moncton), Benchmarking package in
R (Peter Bogetoft, Copenhagen Business School, and Lars Otto,
University of Copenhagen), Efficiency Measurement System
(Holger Scheel, University of Dortmund) or DEA Solver Online
(Andreas Kleine and Günter Winterholer, University of
Hohenheim).
- Commercial packages include DEAFrontier9 (Joe Zhu,
Worcester Polytechnic Institute), DEA-Solver PRO10 (Saitech,
Inc.), PIM-DEA (Ali Emrouznejad, Aston Business School) or
Frontier Analyst (Banxia Software Ltd).
This section focuses on the ‘twin’ DEA software packages
DEAP/Win4DEAP11. These packages centre on the basics of DEA, are
simple to use and are stable over time. They are freely available12 and
come with data files as examples. As Win4DEAP is the Windows based
interface of DEAP (which is a DOS program), the current section refers
only to Win4DEAP. All screenshots and icons presented in this section
9 Zhu (2003) includes an earlier version of DEAFrontier, DEA Excel Solver, on a CD-
ROM. This software works only under Excel © 97, 2000 and 2003. It deals with an
unlimited number of DMUs and is available at little cost.
10 Cooper et al. (2006) include a CD-ROM with a DEA-Solver version limited at 50
DMUs. It is available at little cost.
11 As DEAP is a DOS program, a user friendly Windows interface has been developed for
it (Win4DEAP). These ‘twin’ software packages have to be both downloaded and
extracted to the same folder. Win4DEAP cannot work without DEAP.
12 DEAP Version 2.1: http://www.uq.edu.au/economics/cepa/deap.htm
Win4DEAP Version 1.1.3: http://www8.umoncton.ca/umcm-
deslierres_michel/dea/install.html
and coming from DEAP or Win4DEAP are reproduced by permission of
Timothy Coelli and Michel Deslierres.
The use of Win4DEAP is illustrated by a case study including a sample

of 15 primary schools (see Table 2 below).
The data used in this case study are fictitious (but are very similar to real
ones). 15 schools produce one output (number of pupils) with three
inputs (number of full-time equivalent teachers, number of full-time
administrative staff and number of computers –used as a proxy for
technology investment). For example, School # 8 educates 512 pupils
with 28.6 teachers, 1.3 administrative staff and 26 computers.
Input Output
School
FTE teachers FTE adm. staff Computers Pupils
1 40.2 2.0 37.0 602.0
2 18.1 1.1 17.0 269.0
3 42.5 2.1 41.0 648.0
4 11.0 0.8 10.0 188.0
5 24.8 1.3 22.0 420.0
6 21.1 1.3 19.0 374.0
7 13.5 1.0 13.0 247.0
8 28.6 1.3 26.0 512.0
9 23.5 1.3 22.0 411.0
10 15.9 1.0 15.0 285.0
11 23.2 1.3 22.0 397.0
12 26.0 1.4 25.0 466.0
13 11.1 0.8 11.0 198.0
14 28.8 1.6 26.0 530.0
15 19.7 1.3 18.0 357.0
a) Case study 2 – Building a spreadsheet in Win4DEAP
Win4DEAP is launched by clicking the MD icon ( ). Firms (called
decision-making units or DMUs) are listed in the rows and variables
(outputs and inputs) in the columns. The opening spreadsheet contains
one decision-making unit (DMU1), one output (OUT1) and one input
(IN1) by default (see Figure 9).
To edit and name firms, outputs and inputs, the user has to click the
DMU1 ( ), OUT1 ( ) and IN1 ( ) icons, respectively. The
window reproduced in Figure 10 allows the user to (1) assign a long
name and a label (maximum of eight characters) to any variable and (2)
select the nature of the variables (either ‘input’ or ‘output’). Finally, the
user has to select the ‘with price’ option if he intends to measure cost,
revenue or profit efficiency (i.e. a ‘price’ column will be added to the
selected variable in the spreadsheet).
The icons enable the user to add firms (DMUs). The icons
enable the user to add variables (inputs or outputs). The icons
and are used to delete any existing DMUs or variables. Finally,
the following icons allow the user to reverse the order of
appearance of DMUs (rows) or variables (columns).
How to import Microsoft Excel © data into Win4DEAP
Note that data can be imported from an Excel © file into Win4DEAP by
following these steps:
- Save the Microsoft Excel © data (only numbers, no names of DMUs
or variables should be included) into the CSV format (Comma
delimited).
- In Win4DEAP, first select the ‘File’ menu, then the ‘Import’ option
and finally the ‘New data set’ application.
- Select the CSV file and open it.
- The data is now presented in the Win4DEAP spreadsheet, which
still has to be configured (DMUs and variables have to be named
and variables must be defined as inputs or outputs).
Exercise 2
The objective of this exercise is to correctly calibrate a spreadsheet in
Win4DEAP.
Tasks
a) Prepare a spreadsheet in Win4DEAP including 15 DMUs, 3 inputs
and one output. Name the DMUs ‘School 1’ to ‘School 15’. The
first input is ‘FTE teachers’, the second ‘FTE administrative staff’
and the third ‘Number of computers’. The output corresponds to the
number of pupils.
b) Feed the data appearing in Table 2 into the spreadsheet.
c) Save the file, preferably into the same folder containing
DEAP/Win4DEAP (menu ‘File’, option ‘Save as’).
Answer: The spreadsheet should be similar to the one represented in

Figure 11.
b) Case study 2 – Running a DEA model
To run a DEA model, the user has to click the ‘lightning’ icon ( ).
The window represented in Figure 12 will then appear. This window
allows a calibration of the model following steps 1 to 4 described below:
1. Select an input or an output orientation (Orientation box).
2. Select the assumption about returns to scale (Returns to scale box).
By ticking ‘constant’, one assumes constant returns to scale (CRS);
by ticking ‘variable’, one assumes variable returns to scale (VRS).
If one cannot be certain about the fact that firms are operating at an
optimal scale, running a VRS model is recommended.
3. Select a model (Calculate box). Three main options are available:
- To calculate technical efficiency (TE) or technical (CRS), ‘pure’
(VRS) and scale efficiency (SE), tick ‘DEA (multi-stage)’.
Options ‘DEA (1-stage)’, ‘DEA (2-stage)’ and ‘DEA (multi-
stage)’ correspond to different treatments of slacks. Following
Coelli (1998), the multi-stage treatment is recommended.
- To calculate cost, revenue or profit efficiency, tick ‘DEA-COST’.
For this option, cost and/or price information about variables
must be available and added to the spreadsheet.
- To calculate technical and scale efficiency when panel data are
available, tick ‘MALMQUIST’. See Section 5.4 to learn more
about this.
4. Choose how to display the results: only summarized or reported
firm by firm (Report box).
5. Click ‘Execute’ to run the model.
Exercise 3
The objective of this exercise is to run a DEA model in Win4DEAP
based on the schools case study. The following information is available:
- Schools are confronted with budget restrictions;
- The school system is heavily regulated;
- An obligatory school by school report is expected.
Prerequisites
Exercise 2
Tasks
a) Open the schools data spreadsheet (i.e. the calibrated spreadsheet in
exercise 2)
b) Calibrate the model
c) Execute the model
Answer: The model should be similar to the one represented in

Figure 13.
c) Case study 2 – Interpreting results
After executing the selected model, a short notice appears with
information about Timothy Coelli, the developer of DEAP. Results are
displayed after closing this window. It is recommendable for first time
users to take some time navigating through the results file in order to
become familiar with it. Some results tables are commented on in this
section. Table 3 contains a list of abbreviations with the main acronyms
used in the results file.
Acronym Full name

DEA Data Envelopment Analysis
CRS Constant Returns to Scale
VRS Variable Returns to Scale
TE Technical Efficiency
CRSTE Constant Returns to Scale Technical Efficiency
VRSTE Variable Returns to Scale Technical Efficiency
SE Scale Efficiency
IRS Increasing Returns to Scale
DRS Decreasing Returns to Scale
Figure 14 represents the first table to be commented on. It is an extract

of the results file and features an efficiency summary. The first column
contains the 15 schools (listed 1 to 15). The second one displays the
constant returns to scale technical efficiency scores (CRSTE)13. This
‘total’ efficiency score is decomposed into a ‘pure’ technical efficiency
measure (variable returns to scale technical efficiency –VRSTE– in the
third column) and a scale efficiency measure (scale efficiency –SE– in
the fourth column). The last column indicates the nature of returns to
scale (IRS, DRS or a dash):
13 Note that if you had run a constant returns to scale model instead of a variable returns
to scale one, you would have obtained only one type of efficiency score in your results
file (technical efficiency –TE). Technical efficiency scores are strictly equal to constant
returns to scale technical efficiency scores obtained in the CRSTE column of your
variable returns to scale model.
- Firms associated with IRS are facing increasing returns to scale
(economies of scale).
- Firms associated with DRS are facing decreasing returns to
scale (diseconomies of scale).
- Firms associated with a dash are facing constant returns to
scale; they are operating at an optimal scale.
On average, schools efficiency scores are:
- 94% for CRSTE; overall, schools could reduce their inputs by
6% whilst educating the same number of pupils.
- 97.5% for VRSTE; a better school organization would be able
to reduce input consumption by 2.5%.
- 96.4% for SE; in adjusting their scale, schools could reduce
their inputs by 3.6%.
All subsequent tables displayed in the results file refer to the VRSTE
scores. These tables contain the following information:
- The number of the DMU under review (‘Results for firm’).
- The technical efficiency score (‘Technical efficiency’),
corresponding to the VRSTE when a VRS model has been run
or to the CRSTE when a CRS model has been run.
- The scale efficiency score (‘Scale efficiency’); note that the SE
is mentioned only when a VRS model has been run.
- The lines of the matrix represent the outputs and the inputs of
the model (‘output 1’, ‘output 2’, etc., ‘input 1’, ‘input 2’, etc.).
- The first column of the matrix recalls the original values of the
variables’ outputs and inputs (‘original values’).
- The second column of the matrix represents the movement an
inefficient DMU has to take in order to be located on the
frontier (‘radial movement’).
- The third column of the matrix is the additional movement a
DMU located on a segment of the frontier running parallel to
the axis has to take in order to become efficient (‘slack
movement’).
- The fourth column of the matrix lists the values of the variables
which enable the DMU to be efficient (‘projected value’); these
projected values take into account both the radial and the slack
movements.
- Finally, the listing of peers is mentioned. Each peer is
identified by a number and has an associated weight (‘lambda
weight’) representing the relative importance of the peer.
As illustrations, three individual school tables are specifically
commented on below: School # 1 (Figure 15), # 2 (Figure 16) and # 3
(Figure 17).
School # 1 (Figure 15) has a ‘pure’ efficiency score of 95.1% and a scale
efficiency score of 86.9%. It is facing decreasing returns to scale (DRS).
By improving the operation of the school, 4.9% (100 - 95.1) of inputs
could be saved. By adjusting the school to its optimal size, 13.1%
(100 - 86.9) of inputs could be saved.
The ‘original value’ column contains the original values of the school’s
variables: School # 1 educates 602 pupils with 40.2 teachers, 2
administrative staff and 37 computers. However, School # 1 could
‘produce’ the same quantity of output with fewer inputs: 37.186 teachers
instead of 40.2; 1.902 administrative staff instead of 2;
35.185 computers instead of 37 (see the ‘projected value’ column). The
decreases in inputs 2 and 3 are equal to 4.9 % of the original values:
(- 0.098 / 2) x 100 for input 2 and (- 1.815 / 37) x 100 for input 314. The
case of input 1 is slightly different: to become efficient, it has to
14 In a VRS model, the improvement in variables (decrease in inputs or increase in

outputs) is calculated according to the VRS technical efficiency score (only). In a CRS
model, it is calculated according to the CRS technical efficiency score, or TE score,
including not only the pure efficiency but also the scale efficiency.
decrease not only by 4.9% (minus 1.972 from the ‘radial movement’
column) but also by an additional 1.042 (from the ‘slack movement’
column). Overall School # 1 has to decrease its first input by
minus 3.014 [(- 1.972) + (- 1.042)] to become efficient. This represents
7.5 % [(- 3.014 / 40.2) x 100].
To improve its efficiency, School # 1 has to analyse the practice of
Schools # 3, # 14 and # 8, which are identified as its peers. To be a peer
(or a benchmark), a firm must have a ‘pure’ efficiency score of 100%.
The lambda weight associated with each peer corresponds to its relative
importance among the peer group. Ideally, School # 1 should analyse
best practice from a composite school formed by 61.2% of School # 3,
37.3% of School # 14 and 1.4% of School # 8. As such a ‘virtual’ school
does not exist. School # 1 should concentrate its best practice analysis
on the peer associated with the highest lambda value (i.e. School # 3).
School # 2 (Figure 16) has a ‘pure’ efficiency score of 83.8% and a scale
efficiency score of 96.4%. It is facing increasing returns to scale (IRS).
By improving the operation of the school, 16.2% (100 - 83.8) of inputs
could be saved. By adjusting the school to its optimal size, 3.6%
variables: School # 2 educates 269 pupils with 18.1 teachers,
1.1 administrative staff and 17 computers. However, School # 2 could
‘produce’ the same quantity of output with fewer inputs: 15.163 teachers
instead of 18.1, 0.922 administrative staff instead of 1.1;
14.242 computers instead of 17 (see the ‘projected value’ column). The
decreases in inputs 1, 2 and 3 are equal to 16.2% of the original values
(‘radial movement’ column). No slack movement is identified.
To improve its efficiency, School # 2 has to refer to Schools # 13, # 4,
# 14 and # 8, which are identified as its peers.
School # 3 (Figure 17) has a ‘pure’ efficiency score of 100% and a scale
efficiency score of 84.2%. It is facing decreasing returns to scale (DRS).
This school is well managed. It cannot improve its ‘pure’ efficiency. The
only capacity for improvement lies in a scale adjustment: 15.8%
variables: School # 3 educates 648 pupils with 42.5 teachers,
2.1 administrative staff and 41 computers. These values are equal to the
projected ones (‘pure’ efficiency = 100%).
As School # 3 is purely efficient, it acts as its own peer.
Exercise 4
The objective of this exercise is to interpret DEA results. Figure 18
displays results for one of the 15 schools. It has been truncated in order
to hide the VRS technical efficiency score.
Tasks
Answer the following questions:
a) The variable returns to scale technical efficiency score has been
removed from the table. Find a way to calculate it.
Answer: For example, input 1 can be reduced by 0.864 (radial
movement). This represents 3.7% [(- 0.864 / 23.5) x 100]. Therefore
VRSTE corresponds to 96.3% (100 - 3.7). You would have obtained
the same result if you based your calculations on input 2. If you
based your calculations on input 3, be careful not to take into
account the slack movement.
b) Assume that the ‘pure’ efficiency score is equal to 96.3%. What is
the main feature in need for improvement: the school’s management
or the school’s scale?
Answer: The school’s management has the capacity to improve
efficiency by 3.7% (100 - 96.3). Modifying the school’s scale could
improve efficiency by 1.3% (100 - 98.7). Therefore the main feature
in need of improvement is school management.
c) Assume that the school has only time to analyze best practice from
one of its peers. Which one should it select?
Answer: School # 14. Among the three peers listed (13, 14 and 8),
School # 14 is associated with the highest weight (41.7%).
d) How much must the school reduce input 3 in order to be located on
the efficiency frontier?
Answer: - 1.184 [(- 0.809) + (- 0.375)].
4 DEA IN THE BLACK BOX
This section describes the two principal DEA models: the constant
returns to scale model (Charnes et al., 1978) and the variable returns to
scale model (Banker et al., 1984). DEA is based on the earlier work of
Dantzig (1951) and Farrell (1957), whose approach adopted an input
orientation. Zhu and Cook (2008), Cooper et al. (2007) or Coelli et al.
(2005) provide a comprehensive treatment of the methodology. By
2007, Emrouznejad et al. (2008) identified more than 4000 research
articles about DEA published in scientific journals or books.
DEA is a non-parametric method. Unlike parametric methods (such as
ordinary least square, maximum likelihood estimation or stochastic
frontier analysis), inputs and outputs are used to compute, using linear
programming methods, a hull to represent the efficiency frontier. As a
result, a non-parametric method does not require specification of a
functional form.
Charnes et al. (1978) propose a model assuming constant returns to

scale (CRS model)15. It is appropriate when all firms operate at the
optimal scale. Efficiency is defined by Charnes et al. (1978, p. 430) as
“the maximum of a ratio of weighted outputs to weighted inputs subject
that the similar ratios for every DMU be less or equal to unity”. The
following notation is adopted, as in Johnes (2004):
s
∑u y
r =1
r rk
TEk = m (1)
∑v x
i =1
i ik
15 This model is also known as the Charnes, Cooper & Rhodes model (CCR model).
Where:
TEk is the technical efficiency of firm k using m inputs to produce s
outputs;
yrk is the quantity of output r produced by firm k;
xik is the quantity of input i consumed by firm k;
ur is the weight of output r;
vi is the weight of input i.
n is the number of firms to be evaluated
s is the number of outputs
m is the number of inputs
The technical efficiency of firm k is maximized under two constraints.
First, the weights applied to outputs and inputs of firm k cannot generate
an efficiency score greater than 1 when applied to each firm in the
dataset (equation # 3). Second, the weights on the outputs and on the
inputs are strictly positive (equation # 4). The following linear
programming problem has to be solved for each firm:
s
∑u y
r =1
r rk
Maximize m (2)
∑v x
i =1
i ik
∑u
r =1
r yrj
Subject to m
≤1 j = 1, K , n (3)
∑v x
i =1
i ij
u r , vi > 0 ∀r = 1, K , s; i = 1, K , m (4)
This linear programming problem can be dealt following two different

approaches. In the first one, the weighted sums of outputs are
maximized holding inputs constant (output-oriented model). In the
second one, the weighted sums of inputs are minimized holding outputs
constant (input-oriented model)16. The primal equations for each model,
known as the multiplier form, are given below:
CRS output-oriented model CRS input-oriented model

Primal equation Primal equation
m s
Minimize ∑v x
i =1
i ik (5) Maximize ∑u y
r =1
r rk (9)
Subject to Subject to
m s m s
∑v x − ∑u
i =1
i ij
r =1
r y rj ≥ 0 j = 1, K , n (6) ∑v x − ∑u
i =1
i ij
r =1
r yrj ≥ 0 j = 1, K , n (10)
s m
∑u
r =1
r yrk = 1 (7) ∑v x
i =1
i ik =1 (11)
u r , vi > 0 ∀r = 1, K , s; i = 1, K , m (8) u r , vi > 0 ∀r = 1, K , s; i = 1, K , m (12)
Using the duality in linear programming, an equivalent form, known as

the envelopment form, can be derived from this problem. It is often
preferable to solve the computation using the envelopment form because
it contains only s+m constraints rather than n+1 constraints in the
multiplier form.

Dual equation Dual equation
Maximize φk (13) Minimize θ k (17)
n n
φk y rk − ∑ λ j yrj ≤ 0 r = 1, K , s (14) y rk − ∑ λ j y rj ≤ 0 r = 1, K , s (18)
j =1 j =1
n n
xik − ∑ λ j xij ≥ 0 i = 1, K , m (15) θ k xik − ∑ λ j xij ≥ 0 i = 1, K , m (19)
j =1 j =1
λ j ≥ 0 ∀j = 1, K , n (16) λ j ≥ 0 ∀j = 1, K , n (20)
Where:
1 and θk represent the technical efficiency of firm k
φk
λj represents the associated weighting of outputs and inputs of firm j
16 Note that the input and output orientations refer to the dual equations of each model
(and not to the primal ones).
Every firm located on the sections’ envelope running parallel to the axes
has to be adjusted for output and input slacks. However, the preceding
formulation does not integrate the role of slacks in measuring efficiency.
Considering output slacks, sr, and input slacks, si, the above equations
become:
Dual equation with slacks Dual equation with slacks
s m s m
Maximize φk + ε ∑ sr + ε ∑ si (21) Minimize θ k − ε ∑ sr − ε ∑ si (25)
r =1 i =1 r =1 i =1
n n
φk yrk − ∑ λ j yrj + sr = 0 r = 1, K , s (22) y rk − ∑ λ j y rj + sr = 0 r = 1, K , s (26)
j =1 j =1
n n
xik − ∑ λ j xij − si = 0 i = 1, K , m (23) θ k xik − ∑ λ j xij − si = 0 i = 1, K , m (27)
j =1 j =1
λ j , sr , si ≥ 0 ∀j = 1, K , n; r = 1, K , s; i = 1,..., m λ j , sr , si ≥ 0 ∀j = 1, K , n; r = 1, K , s; i = 1, K , m
(24) (28)
Here, ε is a non-Archimedean value defined to be smaller than any

positive real number. ε is greater than 0. The firm k is efficient only if:
- the efficiency score TEk =  1  = 1 (or TEk = θ k = 1 );

 φk 
- and the slacks sr , si = 0 , ∀ r = 1, K s and i = 1, K , m .
For an in-depth analysis on the treatment of slacks, and especially the

multi-stage methodology, see Coelli (1998).
Banker et al. (1984) propose a model assuming variable returns to scale
(VRS model)17. It is appropriate when all firms do not operate at optimal
scale. As Coelli et al. (2005, p. 172) point out, “the use of the CRS
specification when not all firms are operating at the optimal scale,
results in measures of TE that are confounded by scale efficiencies (SE).
The use of the VRS specification permits the calculation of TE devoid
of these SE effects”. The CRS model can be modified by relaxing the
constant returns to scale assumption. A measure of return to scale for
firm k is added in the primal equation (or the convexity constraint
n
∑λ
j =1
j = 1 in the dual equations).
Figure 19 represents the CRS efficiency frontier (the dashed line) and
the VRS efficiency frontier (the solid line) on the same graph to
illustrate a simple example with one output and one input. Only one
firm, B, is located on both frontiers. A and C are 100% efficient under
the VRS assumption, but inefficient under the CRS assumption. D and E
are inefficient under both specifications.
17 This model is also known as the Banker, Charnes & Cooper model (BCC model).
Output C’ C
W
E’’ E’ E
V
B
U
D’’ D’ D
T
S A
A’
0
Input
The specific situation of firm D is commented on in detail below:

- Firm D is inefficient under VRS and CRS. In order to become
VRS efficient, it has to move to point D’. The input-oriented
VRS technical inefficiency of point D is the distance DD’. In
order to become CRS efficient, firm D has to move further
toward point D’’. The input-oriented CRS technical
inefficiency of point D is the distance DD’’. The distance
between D’ and D’’ corresponds to scale inefficiency. The ratio
efficiency measures, bounded by zero and one, are as follows:
Technical efficiency Technical efficiency Scale efficiency
of D under CRS of D under VRS of D
TD ' ' TD' TD ' '
TECRS = TEVRS = SE =
TD TD TD '
Exercise 5
The objective of this exercise is (1) to comment in detail on the situation
of firms E, A, B and C represented in Figure 19 (as above for firm D)
and (2) to provide ratios of TECRS, TEVRS and SE for firms E, A, B and
C.
Answer for firm E:

Firm E is inefficient under VRS and CRS. In order to become CRS
efficient, it has to move toward point E’’. The input-oriented CRS
technical inefficiency of point E is the distance EE’’. In order to become
VRS efficient, it has to move to point E’. The input-oriented VRS
technical inefficiency of point E is the distance EE’. The difference
between these two distances, i.e. the distance E’E’’, corresponds to
scale inefficiency. The ratio efficiency measures, bounded by zero and
one, are as follows:
of E under CRS of E under VRS of E
VE ' ' VE ' VE ' '
TECRS = TEVRS = SE =
VE VE VE '
Answer for firm A:

Firm A is efficient under VRS but inefficient under CRS. In order to
become CRS efficient, it has to move toward point A’. The input-oriented
CRS technical inefficiency of point A is the distance AA’; this also
corresponds to scale inefficiency. The ratio efficiency measures,
bounded by zero and one, are as follows:
of A under CRS of A under VRS of A
SA' SA SA'
TECRS = TEVRS = =1 SE =
SA SA SA
Answer for firm B:
Firm B is efficient both under VRS and CRS. It is operating at the
optimal scale. The ratio efficiency measures, bounded by zero and one,
are as follows:

of B under CRS of B under VRS of B
UB UB UB
TECRS = =1 TEVRS = =1 SE = =1
UB UB UB
Answer for firm C:

Firm C is efficient under VRS but inefficient under CRS. In order to
become CRS efficient, it has to move toward point C’. The input-
oriented CRS technical inefficiency of point C is the distance CC’; this
also corresponds to scale inefficiency. The ratio efficiency measures,
bounded by zero and one, are as follows:

of C under CRS of C under VRS of C
WC ' WC WC '
TECRS = TEVRS = =1 SE =
WC WC WC
Knowing TE under CRS and TE under VRS, the scale efficiency is

easily calculated. As TEk ,CRS = TEk ,VRS × SE k , the scale efficiency is
obtained through the division of TE under CRS by TE under VRS:
TE k ,CRS
SEk = .
TEk ,VRS
The linear programming problem to be solved under VRS includes a
measure of returns to scale on the variables axis, ck, for the firm k. The
primal equations are as follows:
VRS output-oriented model VRS input-oriented model

Primal equation Primal equation
m s
Minimize ∑v x
i =1
i ik − ck (29) Maximize ∑u
r =1
r yrk + ck (33)
m s m s
∑v x − ∑u
i =1
i ij
r =1
r yrj − ck ≥ 0 j = 1, K , n (30) ∑v x − ∑u
i =1
i ij
r =1
r yrj − ck ≥ 0 j = 1, K , n (34)
s m
∑u
r =1
r yrk = 1 (31) ∑v x
i =1
i ik =1 (35)
u r , vi > 0 ∀r = 1, K , s; i = 1, K , m (32) u r , vi > 0 ∀r = 1, K , s; i = 1, K , m (36)
The dual linear programming models are presented hereafter.

Dual equation Dual equation
Maximize φk (37) Minimize θ k (42)
n n
φk y rk − ∑ λ j yrj ≤ 0 r = 1, K , s (38) y rk − ∑ λ j y rj ≤ 0 r = 1, K , s (43)
j =1 j =1
n n
xik − ∑ λ j xij ≥ 0 i = 1, K , m (39) θ k xik − ∑ λ j xij ≥ 0 i = 1, K , m (44)
j =1 j =1
n n
∑λ
j =1
j =1 (40) ∑λ
j =1
j =1 (45)
λ j ≥ 0 ∀j = 1, K , n (41) λ j ≥ 0 ∀j = 1, K , n (46)
When slacks are added into the model, the dual linear programming
equations become:

Dual equation with slacks Dual equation with slacks
s m s m
Maximize φk + ε ∑ sr + ε ∑ si (47) Minimize θ k − ε ∑ sr − ε ∑ si (52)
r =1 i =1 r =1 i =1
n n
φk yrk − ∑ λ j yrj + sr = 0 r = 1, K , s (48) y rk − ∑ λ j y rj + sr = 0 r = 1, K , s (53)
j =1 j =1
n n
xik − ∑ λ j xij − si = 0 i = 1, K , m (49) θ k xik − ∑ λ j xij − si = 0 i = 1, K , m (54)
j =1 j =1
n n
∑λ
j =1
j =1 (50) ∑λ
j =1
j =1 (55)
λ j , sr , si ≥ 0 ∀j = 1, K , n; r = 1, K , s; i = 1,..., m λ j , sr , si ≥ 0 ∀j = 1, K , n; r = 1, K , s; i = 1, K , m
(51) (56)
A further step has to be taken in order to identify the nature of the

returns to scale. This relates to another model, the non-increasing returns
to scale model (NIRS), derived from the VRS model in which the
n n
∑λ
j =1
j = 1 restriction is substituted by the ∑λ
j =1
j ≤ 1 constraint
(Coelli et al., 2005). In Figure 20, the NIRS efficiency frontier has been
added (the dotted line). It corresponds to the CRS frontier from the
origin to point B followed by the VRS frontier from point B. The nature
of the scale inefficiencies for each firm can be determined by comparing
technical efficiency scores under NIRS and VRS. If
NIRS TE ≠ VRS TE (as for firms A and D), increasing returns to scale
apply. If NIRS TE = VRS TE (but ≠ CRS TE) (as for firms E and C),
decreasing returns to scale apply. Finally, if
NIRS TE = VRS TE = CRS TE, as for firm B, constant returns to scale
apply.
Output C’ C
W
E’’ E’ E
V
B
U
D’’ D’ D
T
S A
A’
NIRS efficient frontier
0
Input
5 EXTENSIONS OF DEA
In this section, a selection of four extensions of DEA is shortly
introduced: adjusting for the environment, preferences (weight
restrictions), sensitivity analysis and time series data. For a broader
overview of the major developments in DEA, see Cook and Seiford
(2008). For an up-to-date review of DEA, readers will refer to Cooper et
al. (2011).
Environmental variables influence the efficiency of firms but are not

under the control of the management team. In DEA, several methods
accommodate such variables. Those include the Charnes et al. (1981)
approach, the categorical model (Banker & Morey, 1986a) or the non-
discretionary variable model derived by Banker and Morey (1986b)
(which indeed includes the environmental variable directly into the DEA
model).
The most convincing of these methods, however, is the two-stage
method, the advantages of which are described in Coelli et al. (2005,
pp. 194-195) or in Pastor (2002, p. 899). The two-stage method
combines a DEA model and a regression analysis. In the first stage, a
traditional DEA model is conducted. This model includes only
discretionary inputs and outputs. In the second stage, the efficiency
scores are regressed against the environmental (i.e. non-discretionary or
exogenous) variables. Tobit regression is often used in the second stage.
However, recent studies have shown that ordinary least squares
regression is sufficient to model the efficiency scores (Hoff, 2007) or
even more appropriate than Tobit (McDonald, 2009).
The coefficients of the environmental variables, estimated by the
regression, are used to model the efficiency scores to correspond to an
identical condition of environment (e.g. usually the average condition).
Simar and Wilson (2007, p. 32) provide a selection of studies using the
two-stage method. Among those are applications in education
(Chakraborty et al., 2001, McMillan & Datta, 1998, McCarty &
Yaisawarng, 1993), hospitals (Burgess & Wilson, 1998), defence
(Barros, 2004), police (Carrington et al., 1997), farming (Binam et al.,
2003) or banking (O’Donnell & van der Westhuizen, 2002). Sueyoshi et
al. (2010) and Sibiano and Agasisti (2012) provide more recent
applications in the sector of manufacturing sector and education.
For different reasons (e.g. the weights assigned to variables by DEA are
considered unrealistic for some firms; the management team may wish
to give priority to certain variables; etc.), preferences about the relative
importance of individual inputs and outputs can be set by the decision
maker. This is done by placing weight restrictions onto outputs and
inputs (also called multiplier restrictions). Cooper et al. (2011) and
Thanassoulis et al. (2004) provide a review of models regarding the use
of weights restrictions. An earlier review can be found in Allen et al.
(1997). Generally, the imposition of weight restrictions worsens
efficiency scores. Three main approaches are identified to accommodate
preferences:
- Dyson and Thanassoulis (1988) propose an approach which
imposes absolute upper and lower bounds on input and output
weights. This technique is applied in Roll et al. (1991) to
highway maintenance units or in Liu (2009) to garbage
clearance units.
- Charnes et al. (1990) develop the cone-ratio method. This
approach imposes a set of linear restrictions that define a
convex cone, corresponding to an ‘admissible’ region of
realistic weight restrictions. See Brockett et al. (1997) for an
application to banks.
- Thompson et al. (1986, 1990) propose the assurance region
method. This approach is actually a special case of the cone
ratio. It imposes constraints on the relative magnitude of the
weights. For example, a constraint on the ratio of weights for
input 1 and input 2 can be included, such as the following:
ν2
L1, 2 ≤ ≤ U1, 2 , where L1,2 and U1,2 are lower and upper
ν1
bounds for the ratio of the weight of input 2 (ν 2 ) to the weight
of input 1 (ν 1 ). As a result, the assurance region method limits
the ‘region’ of weights to a restricted area by prohibiting large
differences in the value of those weights. An application of this
model is provided by Sarica and Or (2007) in the assessment of
power plants.
Cooper et al. (2006, p. 271) mention that the term ‘sensitivity’

corresponds to stability or robustness. For Zhu (2003, p. 217), “the
calculated frontiers of DEA models are stable if the frontier DMUs that
determine the DEA frontier remain on the frontier after particular data
perturbations are made”. Sensitivity analysis aims to identify the impact
on firm efficiency when certain parameters are modified in the model.
The first way to test the sensitivity of DEA results consists in adding or
extracting firms to DEA models. Dusansky and Wilson (1994, 1995)
and Wilson (1993, 1995) provide different approaches to deal with this
concern. The approach of Pastor et al. (1999) allows users to identify the
observations which considerably affect the efficiency of the remaining
firms. It also determines the statistical significance of efficiency
variations which are due to the inclusion of a given firm in the sample.
Another way to test the sensitivity of DEA results consists in modifying
the values of outputs and inputs. They focus on the maximum data
variations a given firm can endure, whilst maintaining its efficiency
status. Approaches include data perturbation of:
- a single variable of an efficient firm (Charnes et al., 1985), data
of other firms remaining fixed;
- simultaneous proportional data perturbation of all outputs and
inputs of an efficient firm (Charnes & Neralic, 1992), data of
other firms remaining fixed;
- simultaneous data perturbation of an efficient firm in a situation
where outputs and inputs can be modified individually (Seiford
& Zhu, 1998a, or Neralic & Wendell, 2004), data of other firms
remaining fixed;
- simultaneous proportional data perturbation of all outputs and
inputs of all firms (Seiford & Zhu, 1998b).
For further review of sensitivity analysis, readers can refer to Zhu
(2001).
In DEA, panel data are considered using two methods: window analysis
and the Malmquist index.
Window analysis, introduced by Charnes et al. (1985), examines the
changes in the efficiency scores of a set of firms over time. A ‘window’
of time periods is chosen for each firm. The same firm is treated as if it
represented a different firm in every time period. In this sense, window
analysis can also be considered as a sensitivity analysis method. For
instance, a model including n firms with annual data and a chosen
‘window’ of t years will result in n x t units to be evaluated. For each
firm, t different efficiency scores will be measured. The ‘window’ is
then shifted by one period (one year in our example) and the efficiency
analysis is repeated. Yue (1992) provides a didactical application of
window analysis. Other applications include Yang and Chang (2009),
Avkiran (2004) or Webb (2003).
The Malmquist total factor productivity index was first introduced by
Malmquist (1953) before being further developed in the frame of DEA.
It is used to measure the change in productivity over time. The
Malmquist index decomposes this productivity change into two
components:
- The first one is called ‘catch-up’. This captures the change in
technical efficiency over time.
- The second one is called ‘frontier-shift’. This captures the
change in technology which occurs over time (i.e. the
movement of efficiency frontiers over time).
Readers will refer to Färe et al. (2011) and Tone (2004) for actual
reviews. Applications of the Malmquist index can be found in Coelli and
Prasada Rao (2005) and Behera et al. (2011).
6 DEA WITH MICROSOFT EXCEL © SOLVER
Excel © Solver is a tool used to find the best way to do something, in

other words to optimize an objective. Instructions on loading
Excel © Solver are easily found on the Internet18.
Excel © Solver allows users to solve optimization problems. An
optimization model is composed of three elements: the target cell, the
changing cells and the constraints. These three elements correspond to
the parameters to be defined in Excel © Solver (see Figure 21).
- The target cell (‘Set objective’) corresponds to the objective. It
has to be either minimized or maximized.
- The changing variable cells are the cells which can be altered in
order to optimize the target cell.
- The constraints (one or several) correspond to restrictions
placed on the changing cells.
18 In Microsoft Excel © 2010, the Solver has to be loaded by clicking the File button, then
the Excel Options and finally the Add-Ins button. In the Manage box, Excel Add-ins has
to be selected before clicking the Go button. In the Add-Ins box, the Solver Add-in has
to be selected. Finally, the OK button has to be clicked. Once the Solver is loaded, it is
located in the Analysis group on the Data tab.
Consider five register offices (A to E) producing two outputs (birth and
marriage certificates) with one input (full-time equivalent public
servant). The data are listed in Table 4. For example, one full-time
equivalent (FTE) public servant works in Register Office A. He
produces one birth and six marriage certificates during a certain period
of time.
Register Input Output
Office Public servant (x) Birth (y1) Marriage (y2)
A 1 1 6
B 1 3 8
C 1 4 3
D 1 5 6
E 1 6 2
The use of Excel © Solver is illustrated with the following CRS model.
CRS input-oriented model

Primal equation
s
Maximize ∑u
r =1
r yrk
m s
Subject to
∑v x − ∑u y
i =1
i ij
r =1
r rj ≥0 j = 1, K , n
m
∑v x
i =1
i ik =1
u r , vi > 0 ∀r = 1, K , s; i = 1, K , m
In this model, the objective is to maximize the weighted sum of outputs

of firm k. Two constraints have to be considered. First, the weighted
sum of inputs minus the weighted sum of outputs of firm j has to be
greater than or equal to zero. Second, the weighted sum of inputs of firm
k has to be equal to one.
Users have to prepare an Excel © spreadsheet, such as the one appearing
in Figure 22. This is divided into two parts:
- The first part comprises rows 2 and 3. This section enables
users to successively calculate the efficiency of the five register
offices (one at a time). To do this, data of each register office
have to be entered successively in cells B2 to D2 (dark grey
cells). Figure 22 already contains data on Register office C. The
two outputs and one input of Register Office C are assigned
weights in cells B3 to D3 (light grey cells). A value of one has
been assigned to all of them in the spreadsheet. These values
will be precisely modified by Excel © Solver in order to
maximize the register offices’ efficiency scores. Cell E2
contains the weighted sum of outputs for Register Office C.
The formula associated with cell E2 is
(B2*$B$3) + (C2*$C$3). Cell F2 contains the weighted sum of
the input for Register Office C. The formula associated with
cell F2 is (D2*$D$3). Finally, cell G2 contains the efficiency
score of Register office C as a percentage (light grey cell). The
formula associated with cell G2 is (E2/F2)*100. Note that the
score of 700% appearing in the spreadsheet is calculated using
weighted values of 1 and without any constraints. In other
words, this score has not yet been optimized under varying
constraints.
- The second part comprises rows 6 to 10. It contains the data for
register offices A to E (output 1 = column B,
output 2 = column C, input = column D, weighted sum of
outputs = column E, weighted sum of the input = column F).
The same formulae as above apply to the weighted sums of
outputs and the input. An additional column, G, is added in the
spreadsheet. It is a working column which will be used by
Excel © Solver. Column G contains the weighted sum of the
input minus the weighted sum of outputs to adequately reflect
m s
the ∑v x − ∑u y
i =1
i ij
r =1
r rj ≥0 constraint. The formula
associated with cell G6 is F6 - D6, the formula associated with

cell G7 is F7 - D7, etc.
Once the spreadsheet is ready, the parameters of Excel © Solver have to
be specified in the following way:
- The objective is to maximize the weighted sum of outputs of
s
Register Office k ( ∑u
r =1
r yrk ). In the objective parameter, cell
$E$2 has to be specified. The Max option has to be ticked.

- To optimize the objective, the changing variable cells have to
be specified. They correspond to the weights associated with
outputs and inputs. In the changing variable cells parameter,
cells $B$3 to $D$3 ($B3:$D$3) have to be specified.
- Finally, the restrictions placed on the changing cells have to be
introduced as constraints. A constraint is added by clicking the
Add button. In the Add Constraint box, three parameters have
to be specified: the cell reference, the sign of the constraint (<=,
= or >=) and the value of the constraint. The first constraint of
m s
the CRS model ( ∑v x − ∑u y
i =1
i ij
r =1
r rj ≥ 0 ) is therefore
specified as follows: $G$6:$G$10>=0 (where the cell reference

is $G$6:$G$10, the sign is >= and the constraint is 0). The
m
second constraint ( ∑v x
i =1
i ij = 1 ) is specified as follows:
$F$2=1 (where the cell reference is $F$2, the sign is = and the
constraint is 1). Note that this constraint means that the given
level of input is kept constant.
Figure 23 represents the Solver Parameters defined above.
Finally, a Simplex LP solving method has to be selected and the ‘Make

Unconstrained Variables Non-Negative box’ has to be ticked. This
indicates that a linear model with non-negative variables is appropriate
(and therefore the third and last ‘constraint’ u r , vi > 0 is taken into
account).
The Solve button should be clicked in order to execute Excel © Solver.
Excel © Solver will search every feasible solution to determine the
solution which has the best target cell value. Register Office C obtains
an efficiency score of 73.08% (cell G2). This score is obtained using
weights of 0.15, 0.04 and 1 assigned to output 1, output 2 and input 1,
respectively (cells B3, C3 and D3). A Solver Results box appears after
solving the model. Before solving the model again for the other register
offices, ‘Restore Originals Values’ has to be ticked before clicking the
OK button.
To measure the efficiency of Register Office A (for example), it is
necessary to replace the values of cells B3 to D3 (which currently refer
to Register Office C) with the values of cells B6 to D6 (which refer to
Register Office A). Solving the model will calculate an efficiency score
of 75% for Register Office A.
Exercise 6
The objective of this exercise consists in programming the following
CRS model using Microsoft Excel © Solver. The same data as above
(see Table 4) have to be used. Note that this CRS model is equivalent to
the one developed above. Instead of maximizing the weighted sum of
outputs, however, it minimizes the weighted sum of the input.
m
Minimize ∑v x
i =1
i ik
m s
Subject to
∑ vi xij − ∑ ur yrj ≥ 0
i =1 r =1
j = 1, K , n
s
∑u y
r =1
r rk =1
u r , vi > 0 ∀r = 1, K , s; i = 1, K , m
Tasks
a) Prepare an Excel © spreadsheet adapted to the use of
Excel © Solver.
b) Which cell has to be optimized?
Answer: The target cell to be optimized is $F$2
c) In this case, do you maximize output or minimize input?
Answer: Input minimization (the Min option has to be ticked in the
Solver)
d) Which equation of the CRS model is optimized by the Solver?
m
Answer: ∑v xi =1
i ik . This equation corresponds to the minimization
of the weighted sum of the input.
e) Which variables can be changed in the optimization process?

Answer: The weights associated with output 1, output 2 and input 3
can be changed (cells B3, C3 and D3)
f) Which constraint in the Excel © Solver corresponds to the

m s
equation ∑ vi xij − ∑ ur yrj ≥ 0 in the CRS model?
i =1 r =1
Answer: $G$6:$G$10>=0
g) Which constraint in the Excel © Solver corresponds to the equation

s
∑u y
r =1
r rk = 1 in the CRS model?
Answer: $E$2=1. This constraint means that the weighted sum of

outputs must equal one (e.g. the given levels of outputs are kept
constant).
h) Solve the CRS model for each register office. What are the
efficiency scores?
Answer: A=75, B=100, C=73.08, D=100 and E=100
REFERENCES
Allen, R., Athanassopoulos, A., Dyson, R. G. & Thanassoulis, E.
(1997). Weights restrictions and value judgements in Data Envelopment
Analysis: Evolution, development and future directions. Annals of
Operations Research, 73(1), 13-34.
Avkiran, N. K. (2004). Decomposing technical efficiency and window
analysis. Studies in Economics and Finance, 22(1), 61-91.
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), 1078-1092.
Banker, R. D. & Morey, R. C. (1986a). Efficiency Analysis for
Exogenously Fixed Inputs and Outputs. Operations Research, 34(4),
513-521.
Banker, R. D. & Morey, R. C. (1986b). The Use of Categorical
Variables in Data Envelopment Analysis. Management Science, 32(12),
1613-1627.
Barr, R. (2003). DEA Software Tools and Technology: A State-of-the-
Art survey. In W. W. Cooper, L. M. Seiford & J. Zhu (Eds.), Handbook
on Data Envelopment Analysis (pp. 539-566). Boston: Kluwer
Academic Publishers.
Barros, C. P. (2004). Measuring performance in defence-sector
companies in a small NATO member country. Journal of Economic
Studies, 31(2), 112-128.
Behera, S. K., Faroquie, J. A. & Dash, A. P. (2011). Productivity change
of coal-fired thermal power plants in India: a Malmquist index approach.
IMA Journal of Management Mathematics, 22(4), 387-400.
Binam, J. N., Sylla, K., Diarra, I. & Nyambi, G. (2003). Factors
affecting technical efficiency among coffee farmers in Côte d’Ivoire:
evidence from the centre west region. R&D Management, 15(1), 66-76.
Bogethoft, P. & Otto, L. (2010). Benchmarking with DEA, SFA, and R.
New York: Springer.
Brockett, P. L., Charnes, A., Cooper, W. W., Huang, Z. M. & Sun, D. B.
(1991). Data Transformations in DEA Cone Ratio Envelopment
Approaches for Monitoring Bank Performance. Journal of Operational
Research, 98(2), 250-268.
Burgess, J. F. & Wilson, P. W. (1998). Variation in inefficiency among
US hospitals. Canadian Journal of Operational Research and
Information Processing (INFOR), 36(3), 84-102.
Carrington, R., Puthucheary, N., Rose, D. & Yaisawarng, S. (1997).
Performance measurement in government service provision: the case of
police services in New South Wales. Journal of Productivity Analysis,
8(4), 415-430.
Chakraborty, K., Biswas, B. & Lewis, W. C. (2001). Measurement of
technical efficiency in public education: a stochastic and nonstochastic
production approach. Southern Economic Journal, 67(4), 889-905.
Charnes, A., Clarke, C., Cooper, W. W. & Golany, B. (1984). A
development study of DEA in measuring the effect of maintenance units
in th U.S. Air Force. Annals of Operations Research, 2(1), 95-112.
Charnes, A., Cooper, W. W., Huang, Z. M. & Sun, D. B. (1990).
Polyhedral cone-ratio DEA models with an illustrative application to
large commercial banks. Journal of Econometrics, 46(1-2), 73-91.
Charnes, A, Cooper, W. W. & Rhodes E. L. (1978). Measuring the
efficiency of decision making units. European Journal of Operational
Research, 2(6), 429-444.
Charnes, A, Cooper, W. W. & Rhodes E. L. (1981). Evaluating program
and managerial efficiency: An application of DEA to program follow
through. Management Science, 27(6), 668-697.
Charnes, A., Haag, S., Jaska, P. & Semple, J. (1992). Sensitivity of
efficiency calculations in the additive model of data envelopment
analysis. International Journal of System Sciences, 23(5), 789-798.
Charnes, A. & Neralic, L. (1992). Sensitivity analysis in data
envelopment analysis. Glasnik Matematicki, 27(47), 191-201.
Coelli, T. J. (1996). A Guide to DEAP Version 2.1: A Data Envelopment
Analysis (Computer) Program (CEPA Working Paper 96/08). Brisbane:
Centre for Efficiency and Productivity Analysis, University of
Queensland.
Coelli, T. J. (1998). A Multi-stage Methodology for the Solution of
Oriented DEA Models. Operations Research Letters, 23(3-5), 143-149.
Coelli, T. J. & Perelman, S. (1996). Efficiency measurement, Multiple-
output Technologies and Distance Functions: With Application to
European Railways (CREPP Working Paper 96/05). Liège: Centre de
Recherche en Economie Publique et de la Population, University of
Liège.
Coelli, T. J. & Perelman, S. (1999). A Comparison of Parametric and
Non-Parametric Distance Functions: With Application to European
Railways. European Journal of Operational Research, 117(2), 326-339.
Coelli, T. J. & Prasada Rao, D. S. (2005). Total factor productivity
growth in agriculture: a Malmquist index analysis of 93 countries, 1980-
2000. Agricultural Economics, 32(s1), 115-134.
Coelli, T. J., Prasado Rao, D. S., O'Donnel, C. J. & Battese, G. E.
(2005). An Introduction to Efficiency and Productivity Analysis. New
York: Springer.
Cook, W. D. & Seiford, L. M. (2008). Data envelopment analysis
(DEA) – Thirty years on. European Journal of Operational Research,
192(1), 1-17.
Cooper, W. W., Ruiz J. L. & Sirvent, I. (2011). Choices and Uses of
DEA Weights. In W. W. Cooper, L. M. Seiford & J. Zhu (Eds.),
Handbook on Data Envelopment Analysis (pp. 93-126). New York:
Springer.
Cooper, W. W., Seiford, L. M. & Tone, K. (2006). Introduction to Data
Envelopment Analysis and Its Uses. New York: Springer.
Cooper, W. W., Seiford, L. M. & Tone, K. (2007). Data Envelopment
Analysis: A comprehensive Text with Models, Applications, References
and DEA-Solver Software. New York: Springer.
Cooper, W. W., Seiford, L. M. & Zhu, J. (2004). Handbook on Data
Envelopment Analysis (1st ed.). Boston: Kluwer Academic Publishers.
Cooper, W. W., Seiford, L. M. & Zhu, J. (2011). Handbook on Data
Envelopment Analysis (2nd ed.). New York: Springer.
Dantzig, G. B. (1951). Maximization of a linear function of variables
subject to linear inequalities. In T. C. Koopmans (Ed.), Activity Analysis
of Production and Allocation (pp. 339-347). New York: John Wiley and
Sons.
Dusansky, R. & Wilson, P. W. (1994). Measuring efficiency in the care
of developmentally disabled. Review of Economics and Statistics, 76(2),
340-345.
Dusansky, R. & Wilson, P. W. (1995). On the relative efficiency of
alternative modesof producing public sector output: The case of
developmentally disabled. European Journal of Operational Research,
80(3), 608-628.
Dyson, R. G. & Thanassoulis, E. (1988). Reducing weight flexibility in
DEA. Journal of the Operational Research Society, 39(6), 563-576.
Emrouznejad, A., Parker, B. R. & Tavares G. (2008). Evaluation of
research in efficiency and productivity: A survey and analysis of the first
30 years of scholarly literature in DEA. Socio-Economic Planning
Sciences, 42(3), 151-157.
Färe, R., Grosskopf, S. & Margaritis, D. (2011). Malmquist Productivity
Indexes and DEA. In W. W. Cooper, L. M. Seiford & J. Zhu (Eds.),
Handbook on Data Envelopment Analysis (pp. 127-150). New York:
Springer.
Farrell, M. J. (1957). The measurement of productive efficiency.
Journal of Royal Statistical Society, 120(3), 253-281.
Hoff, A. (2007). Second stage DEA: Comparison of approaches for
modelling the DEA score. European Journal of Operational Research,
181(1), 425-435.
Johnes, J. (2004). Efficiency measurement. In G. Johnes & J. Johnes
(Eds.), International Handbook on the Economics of Education
(pp. 613-742). Cheltenham: Edward Elgar Publishing.
Liu, C.-C. (2009). A study of optimal weights restriction in Data
Envelopment Analysis. Applied Economics, 41(14), 1785-1790.
McCarty, T. A. & Yaisawarng, S. (1993). Technical efficiency in New
Jersey school districts. In H. O. Fried, C. A. K Lovell & S. S. Schmidt
(Eds.), The Measurement of Productive Efficiency: Techniques and
Applications (pp. 271-287). Oxford University Press: New York.
McDonald, J. (2009). Using least squares and tobit in second stage DEA
efficiency analyses. European Journal of Operational Research, 197(2),
792-798.
McMillan, M. L. & Datta, D. (1998). The relative efficiencies of
Canadian universitites: a DEA perspective. Canadian Public Policy-
Analyse de Politiques, 24(4), 485-511.
Malmquist, S. (1953). Index numbers and indifference surfaces.
Trabajos de Estatistica, 4(2), 209-242.
Neralic, L. (2004). Preservation of efficiency and inefficiency
classification in data envelopment analysis. Mathematical
Communications, 9(1), 51-62.
Neralic, L. & Wendell, R. E. (2004). Sensitivity in data envelopment
analysis using an approximate inverse matrix. Journal of the
Operational Research Society, 55(11), 1187-1193.
O’Donnell, C. J. & van der Westhuizen, G. (2002). Regional
comparisons of banking performance in SouthAfrica. South African
Journal of Economics,70(3), 485-518.
Pastor, J. M. (2002). Credit risk and efficiency in the European banking
system: A three-stage analysis. Applied Financial Economics, 12(12),
895-911.
Pastor, J. M., Ruiz, J. L. & Sirvent, I. (1999). A statistical test for
detecting influential observations in DEA. European Journal of
Operational Research, 115(3), 542-554.
Rhodes, E. L. (1978). Data envelopment analysis and approaches for
measuring the efficiency of decision making units with an application to
program follow through in U.S. education (Doctoral dissertation).
School of Urban and Public Affairs, Carnegie-Mellon University.
Roll, Y., Cook, W. D. & Golany B. (1991). Controlling Factor Weights
in Data Envelopment Analysis. IEE Transactions, 23(1), 2-9.
Sarica, K. & Or, I. (2007). Efficiency assessment of Turkish power
plants using data envelopment analysis. Energy, 32(8), 1484-1499.
Scheel, H. (2000). EMS: Efficiency Measurement System User's
Manual. Retrieved from Technische Universität Dortmund, Wirtschafts-
und Sozialwissenschaftliche Fakultät Web site:
http://www.wiso.tu-
dortmund.de/wiso/de/fakultaet/personen/institut/or/EXT-HOSC.html
Seiford, L. M. & Zhu, J. (1998a). Stability regions for maintaining
efficiency in data envelopment analysis. European Journal of
Operational Research, 108(1), 127-139.
Seiford, L. M. & Zhu, J. (1998b). Sensitivity analysis of DEA models
for simultaneous changes in all the data. Journal of the Operational
Research Society, 49(10), 1060-1071.
Sibiano, P. & Agasisti, T. (2012). Efficiency and heterogeneity of public
spending in education among Italian regions. Journal of Public Affairs,
doi:10.1002/pa. 1404.
Simar, L. & Wilson P. W. (2007). Estimation and inference in two-
stage, semi-parametric models of production processes. Journal of
Econometrics, 136(1), 31-64.
Sueyoshi, T., Goto, M. & Omi, Y. (2010). Corporate governance and
firm performance: Evidence from Japanese manufacturing industries
after the lost decade. European Journal of Operational Research,
203(3), 724-736.
Thanassoulis, E, Portela, M. C. S. & Allen, R. (2004). Incorporating
value judgements in DEA. In W. W. Cooper, L. M. Seiford & J. Zhu
(Eds.), Handbook on Data Envelopment Analysis (pp. 99-138). Boston:
Kluwer Academic Publishers.
Thanassoulis, E., Portela, M. C. S., & Despic, O. (2008). Data
Envelopment Analysis: The Mathematical Programming Approach to
Efficiency Analysis. In H. O. Fried, C. A. Lovell, & S. S. Schmidt, The
Measurement of Productive Efficiency and Productivity Growth
(pp. 251-420). New York: Oxford University Press.
Thomson, R. G., Langemeier, L. N., Lee, C.-T., Lee, E. & Thrall, R. M.
(1990). The role of multiplier bounds in efficiency analysis with
application to Kansas farming. Journal of Econometrics, 46(1-2), 93-
108.
Thomson, R. G., Singleton Jr., F. D., Thrall, R. M. & Smith, B. A.
(1986). Comparative site evaluations for locating a high-energy physics
lab in Texas. Interfaces, 16(6), 35-49.
Tone, K. (2004). Malmquist productivity index – Efficiency change over
time. In W. W. Cooper, L. M. Seiford & J. Zhu (Eds.), Handbook on
Data Envelopment Analysis (pp. 203-227). Boston: Kluwer Academic
Publishers.
Webb, R. (2003). Level of efficiency in UK retail banks: a DEA window
analysis. International Journal of the Economics of Business, 10(3),
305-322.
Wilson, P. W. (1993). Detecting outliers in deterministic non-parametric
frontier models with multiple outputs. Journal of Business and
Economic Statistics, 11(3), 319-323.
Wilson, P. W. (1995). Detecting influential observations in data
envelopment analysis. Journal of Productivity Analysis, 6(1), 27-45.
Yang, H.-H. & Chang, C.-Y. (2009). Using DEA window analysis to
measure efficiencies of Taiwan's integrated telecommunication firms.
Telecommunications Policy, 33(1-2), 98-108.
Yue, P. (1992). Data Envelopment Analysis and Commercial Bank
Performance: A Primer with Applications to Missouri Banks. Federal
Reserve Bank of St Louis Review, 74(1), 31-45.
Zhu, J. (2001). Super-efficiency and DEA sensitivity analysis. European
Journal of Operational Research, 129(2), 443-455.
Zhu, J. (2003). Quantitative Models for Performance Evaluation and
Benchmarking. New York: Springer.
Zhu, J. & Cook, W. D. (2008). Data Envelopment Analysis: Modeling
Operational Processes and Measuring Productivity. Seattle:
CreateSpace.

Data Envelopment Analysis (DEA) : Jean-Marc Huguenin

Uploaded by

Data Envelopment Analysis (DEA) : Jean-Marc Huguenin

Uploaded by

Jean-Marc Huguenin

Data Envelopment Analysis (DEA)

IDHEAP – Cahier 276/2012

 2012 IDHEAP, Lausanne

DEA is used to measure the performance of firms or entities (called

a) Case study 1 – Two basic DEA models

b) Case study 1 – Input or output orientation

c) Case study 1 – CRS efficient frontier

Figure 1 also illustrates how DEA measures efficiency scores. The

d) Case study 1 – VRS efficient frontier

Figure 2 also illustrates how DEA measures efficiency scores. The

How to interpret efficiency scores according to the DEA model’s

CRS efficient frontier

6 VRS efficient frontier

f) Case study 1 – Nature of returns to scale

- A register office (or the projected point of a register office) is

g) Case study 1 – Peers (or benchmarks)

h) Case study 1 – Slacks

DEA allows multiple outputs and multiple inputs to be taken into

The notion of efficiency refers to an optimal situation; the maximum

DEA is a benchmarking technique. The efficiency scores provide

Correct answers: 1b; 2c; 3d; 4d; 5a; 6c; 7c.

The user-friendly software packages of DEA incorporate intuitive

The use of Win4DEAP is illustrated by a case study including a sample

Answer: The spreadsheet should be similar to the one represented in

Answer: The model should be similar to the one represented in

Acronym Full name

Figure 14 represents the first table to be commented on. It is an extract

14 In a VRS model, the improvement in variables (decrease in inputs or increase in

Charnes et al. (1978) propose a model assuming constant returns to

This linear programming problem can be dealt following two different

CRS output-oriented model CRS input-oriented model

u r , vi > 0 ∀r = 1, K , s; i = 1, K , m (8) u r , vi > 0 ∀r = 1, K , s; i = 1, K , m (12)

Using the duality in linear programming, an equivalent form, known as

CRS output-oriented model CRS input-oriented model

Here, ε is a non-Archimedean value defined to be smaller than any

- the efficiency score TEk =  1  = 1 (or TEk = θ k = 1 );

For an in-depth analysis on the treatment of slacks, and especially the

The specific situation of firm D is commented on in detail below:

Answer for firm E:

Answer for firm A:

Technical efficiency Technical efficiency Scale efficiency

Answer for firm C:

Technical efficiency Technical efficiency Scale efficiency

Knowing TE under CRS and TE under VRS, the scale efficiency is

VRS output-oriented model VRS input-oriented model

u r , vi > 0 ∀r = 1, K , s; i = 1, K , m (32) u r , vi > 0 ∀r = 1, K , s; i = 1, K , m (36)

The dual linear programming models are presented hereafter.

VRS output-oriented model VRS input-oriented model

VRS output-oriented model VRS input-oriented model

A further step has to be taken in order to identify the nature of the

NIRS efficient frontier

Environmental variables influence the efficiency of firms but are not

Cooper et al. (2006, p. 271) mention that the term ‘sensitivity’

Excel © Solver is a tool used to find the best way to do something, in

CRS input-oriented model

In this model, the objective is to maximize the weighted sum of outputs

associated with cell G6 is F6 - D6, the formula associated with

$E$2 has to be specified. The Max option has to be ticked.

specified as follows: $G$6:$G$10>=0 (where the cell reference

Finally, a Simplex LP solving method has to be selected and the ‘Make

of the weighted sum of the input.

e) Which variables can be changed in the optimization process?

f) Which constraint in the Excel © Solver corresponds to the

g) Which constraint in the Excel © Solver corresponds to the equation

Answer: $E$2=1. This constraint means that the weighted sum of

You might also like