1

Introduction

Data Envelopment Analysis (DEA) is a methodology based upon

an interesting application of linear programming. It was originally
developed for performance measurement. It has been successfully
employed for assessing the relative performance of a set of firms
that use a variety of identical inputs to produce a variety of iden-
tical outputs. The principles of DEA date back to Farrel (1957).
The recent series of discussions on this topic started with the
article by Charnes et al. (1978). A good introduction to DEA is
available in Norman and Stoker (1991). Cooper et al. (2000)
provide recent and comprehensive material on DEA.

1.1 Decision-making Units

Data Envelopment Analysis is a linear programming-based

technique for measuring the performance efficiency of organiza-
tional units which are termed Decision-Making Units (DMUs).
This technique aims to measure how efficiently a DMU uses the
resources available to generate a set of outputs (Charnes et al.
1978). Decision-making units can include manufacturing units,
departments of big organizations such as universities, schools,
bank branches, hospitals, power plants, police stations, tax offices,
prisons, defence bases, a set of firms or even practising individuals
26 An Introduction to Data Envelopment Analysis

such as medical practitioners. As we shall see later in this book,

DEA has been successfully applied to measure the performance
efficiency of all these kinds of DMUs.
Most of these DMUs are non-profit organizations, where the
measurement of performance efficiency is difficult.1 Note that
the efficiency of commercial organizations can be assessed easily
by their yearly profits, or their stock market indices. However,
such measurable factors are not applicable to non-profit organiza-
tions. The problem is complicated by the fact that the DMUs
consume a variety of identical inputs and produce a variety of
identical outputs. For example, schools can have a variety of in-
puts, which are the same for each school—quality of students,
teachers, grants, etc. They have a variety of identical outputs—
number of students passing the final year, average grade obtained
by the students in their final year, etc.
The performance of DMUs is assessed in DEA using the con-
cept of efficiency or productivity, which is the ratio of total outputs
to total inputs. Efficiencies estimated using DEA are relative,
that is, relative to the best performing DMU (or DMUs if there
is more than one best-performing DMUs). The best-performing
DMU is assigned an efficiency score of unity or 100 per cent, and
the performance of other DMUs vary, between 0 and 100 per
cent relative to this best performance.

1.2 Basic Concepts of Efficiency Measurement

As mentioned earlier, the basic efficiency measure used in DEA

is the ratio of total outputs to total inputs.

Output
Efficiency = (1.1)
Input

1
Over the past years, DEA has been applied to profit-making organiza-
tions also. This is partly because profit per se is not a good indication of the
potential for improvement within an organization, and because other factors
are necessary for a holistic assessment of performance.
 Introduction 27

1.2.1 Case of Single Input and Single Output

Let us consider the performance of four firms on the basis of one
input measure (capital) and one output measure (value added)2
as listed in Table 1.1.

Table 1.1 Performance of four firms

Firm Input Output
(Fixed + Working (Value Added)
Capital) ($million) ($million)
A 8.6 1.8
B 2.2 0.2
C 15.6 2.8
D 31.6 4.1

Note that the firms consume differing quantities of inputs and

produce differing levels of outputs. Given the data, how can we
compare their performance?
The easiest approach is to use ratios. As there is only one input
and only one output, ratio computation is simple. Details are
shown in Table 1.2.

Output Value Added

Efficiency = =
Input Capital Employed

Table 1.2 Comparison of performance of the firms

Firm Capital Employed Value Added Value Added per
($million) ($million) Capital Employed
A 8.6 1.8 0.209
B 2.2 0.2 0.091
C 15.6 2.8 0.179
D 31.6 4.1 0.130

Firm A has the highest value added per unit of capital employed,
while Firm B has the lowest. As Firm A has the highest ratio, we
can compare the performance of other firms relative to that of

2
The term ‘value added’ is an economic term, and is the money value of
all intermediate inputs in a firm subtracted from the output.
28 An Introduction to Data Envelopment Analysis

Firm A. Setting the performance efficiency of Firm A as 100 per

cent, we can calculate the relative efficiencies of the other firms, as
shown in Table 1.3.

Table 1.3 Relative efficiencies of firms

Firm Value Added per Relative Efficiency
Capital Employed (%)
A 0.209 100.0
B 0.091 43.4
C 0.179 85.8
D 0.130 62.0

A fundamental assumption behind the computation of relative

efficiency is that if a given firm, A, is capable of producing Y(A)
units of output using X(A) of inputs, then other firms should
also be able to do the same if they were to operate efficiently.
We can set Performance Targets for inefficient firms to enable
them to reach 100 per cent relative efficiency in comparison with
Firm A, the most efficient. Firm A has operated in an environment
similar to the others and hence using its performance as a bench-
mark is realistic. Input Target for Firm B is the amount of capital
employed that will enable the firm to have the same ratio of value
added to capital employed as Firm A.

Input Target = Actual Input ´ Relative Efficiency/100 (1.2)

For Firm B,
Input Target = 2.2 ´ 0.434 = 0.955

This means that if Firm B operates using $0.955 million as

input, and produces $0.2 million as value added output, then it
will be considered as efficient as Firm A.
For inefficient firms, input target will be less than Actual Input.
The difference between actual input and input target is Input
Slack.
For Firm B,

Input Slack = Actual Input – Input Target (1.3)

= 2.2 – 0.955
= 1.245
 Introduction 29

Input Slack can also be expressed as a percentage.

Input Slack
Input Slack Percentage = × 100 (1.4)
Actual Input
For Firm B,
1.245
Input Slack Percentage = × 100 = 56.6
2.2

Thus, if Firm B has to be as efficient as Firm A, it should pro-

duce the same output using 57 per cent less inputs.
Input targets and slacks for other firms can be computed
similarly.
Using a similar logic, we can compute Output Targets and Output
Slacks.
Actual Output
Output Target = (1.5)
Relative Efficiency/ 100
Output Slack = Output Target – Actual Output (1.6)
Output Slack
Output Slack Percentage = × 100 (1.7)
Actual Output
For Firm B,
Actual Output 0.2
Output Target = = = $0.46 billion
Relative Efficiency/ 100 0.434
Output Slack = Output Target – Actual Output
= 0.46 – 0.2
= $0.26 billion

Output Slack
Output Slack Percentage = × 100
Actual Output
0.26
= × 100
0.2
= 130
30 An Introduction to Data Envelopment Analysis

Thus, if firm B is to achieve the same efficiency as Firm A, it

should increase its output by $0.26 million or by 130 per cent
for the same level of input (capital).

1.2.2 Case of Single Output and Two Inputs

In practice, no firm consumes only one input to produce a single
output. Let us now add one more input—number of employees,
as listed in Table 1.4.

Table 1.4 An additional input for four firms

Firm No. of Employees (’000)
A 1.8
B 1.7
C 2.6
D 12.3

How do we compare the performance of the firms now? We can

still use ratios, but we now have two ratios. Calculations are shown
in Table 1.5. Throughout this textbook, whenever this example
is considered, value added and capital employed are always meas-
ured in million dollars, and employees in thousands.

Table 1.5 Comparison of performance of firms

with one output and two inputs
Firm Value Added per Value Added per
Capital Employed Employees
A 0.209 1.000
B 0.091 0.118
C 0.179 1.077
D 0.130 0.333

Note that Firm A has produced the largest value added per
unit of capital employed, while Firm C has produced the maximum
value added per employee. As we do not know which ratio is more
crucial, we cannot say Firm A is more efficient than Firm C or
otherwise. What we can conclude is that Firms B and D are not
as efficient as Firms A and C, because their ratios of output to in-
put (i.e., value added to capital or value added to employees) are
lower.
 Introduction 31

1.3 Graphical Description—Frontier Analysis

One way to tackle the problem of interpreting different ratios, for

problems involving two inputs and one output, is through graph-
ical analysis. Let us plot the two ratios for each firm as shown in
Figure 1.1.

Figure 1.1 Frontier analysis in DEA

As we have two inputs and one output, we have used ratios of

input to output as axes in Figure 1.1. Therefore, the values used
for the two axes are the reciprocals of the values reported in Table
1.5. The reciprocals are shown in Table 1.6. The firms that are
more efficient consume lower levels of input per unit of output
and hence lie closer to the origin. As discussed previously, Firms
A and C are more efficient than the others. Given these ratios,
one can draw an Efficiency Frontier, as the line joining the more
efficient firms and the vertical and horizontal lines connecting
them to the two axes. The efficiency frontier is indicated in Figure
1.1. It represents a standard of performance that the firms not
on the frontier should try to achieve. Firms on the frontier (Firms
A and C here) are considered 100 per cent efficient.
32 An Introduction to Data Envelopment Analysis

Table 1.6 Performance of firms with one output and

two inputs (input/output ratios)
Firm Capital Employed/ Employees/
Value Added Value Added
A 4.778 1.000
B 11.000 8.500
C 5.571 0.929
D 7.707 3.000

Such an analysis using the efficiency frontier is often termed

Frontier Analysis (Farrel 1957).3 This efficiency frontier forms
the basis of efficiency measurement. The efficiency frontier
envelops the available data. Hence, the term Data Envelopment
Analysis. In the DEA literature, Firms A and C are called efficient
firms while Firms B and D that do not lie on the efficiency frontier
are called inefficient firms.
Firms A and C lie on the efficiency frontier, and hence are the
most efficient. Note that this does not mean that their performance
cannot be improved. It may or may not be possible. The available
data does not give any idea regarding the extent to which their
performance can be improved. These are the best firms with regard
to the data we have. As no other firm shows better performance,
we should assume that their performance is the best achievable.
We rate the performance of all other firms in relation to this best
achieved performance. Thus, we consider relative efficiencies, not
absolute efficiencies.

1.3.1 Estimating Efficiencies of Inefficient Firms

Consider Firm B. The firm is inefficient because it does not lie
on the frontier. Can we make a quantitative estimate of its effi-
ciency, in relation to the performance of the best firm lying on the
frontier? Firm B uses $2.2 million of capital and 1,700 employees.
3
The economic treatment of Frontier Analysis is normally explained using
the concept of production possibility frontier (Samuelson and Nordhaus 1989;
Thanassoulis 1999). A production possibility set for the example in this sec-
tion can be constructed by considering all possible combinations of the two
inputs (capital employed and employees) that will result in the same, given
level of output (value added), and the production possibility frontier is the
frontier enveloping the production possibility set.
 Introduction 33

Let the firm’s output be increased keeping the ratio of inputs

(i.e., capital to employees) unchanged. In Figure 1.1, the firm
will move along the line joining the origin with Firm B (dotted
line OB in the figure). Obviously, the best performance possible
for Firm B (retaining the same ratio of inputs) occurs at the point
E, the intersection of the line OB and the efficiency frontier. The
co-ordinates of the point E can be determined using the principles
of analytical geometry.
Note that the best possible performance cannot go below the
line AG as shown in Figure 1.1 because the line defines the best
achieved performance. This is exemplified by Firm A in this case.
Thus, Firm A defines the best achievable performance of Firm B.
Hence Firm A is called the peer for Firm B. Peers are always ef-
ficient firms. Inefficient firms can try to emulate their peers in
order to improve their efficiency.
Given this best possible performance, one can measure the effi-
ciency of Firm B as the ratio of distance OE to the distance OB.

Relative Efficiency of Firm B

= (1.8)

OE
=
OB
3.69 2 + 4.782
=
8.52 + 112
= 0.4344

Hence, the relative efficiency of Firm B is 43.44 per cent. Simi-

larly, the relative efficiency of Firm D can be estimated.

1.3.2 Performance Targets and Slacks

for Inefficient Firms
Similar to the earlier exercise, we can estimate input and output
targets. Let us again consider Firm B. It can move up to the effi-
ciency frontier in at least three ways.
34 An Introduction to Data Envelopment Analysis

(a) If Firm B reduces the number of employees, with capital

and value added being unchanged it moves to point F,
where the ratio of employees to value added is 0.9286.
As value added for Firm B is $0.2 million,
Input Target for Employees = 0.9286 ´ 0.2 = 0.1857
Input Slack for Employees = Actual Input – Input Target
= 1.7 – 0.1857
= 1.5143
Input Slack
Input Slack Percentage for Employees = × 100
Actual In put
1.5143
= × 100
1.7
= 89.1
(b) If Firm B reduces only its capital employed, keeping
employees and value added unchanged, to move to point
G. At this point, the ratio of capital employed to value
added is 4.78. As value added for Firm B is $0.2 million,4
Input Target for
Capital Employed = 4.78 ´ 0.2 = 0.956
Input Slack for
Capital Employed = Actual Input – Input Target
= 2.000 – 0.956
= 1.244
Input Slack Percentage Input Slack
for Capital Employed = × 100
Actual In put
1.244
= × 100
2.200
= 56.5
(c) If Firm B reduces both capital and employees in the same
ratio, or increases value added (keeping the same ratio
of inputs), it moves to point E. Targets can be calculated
accordingly.
4
These calculations are the same as those carried out in the previous
section, except for round-off errors.
 Introduction 35

2.2 1.7
Output Target for Value Added = ≡ = $0.46 billion
4.78 3.69
Output Slack = Output Target – Actual Output
= 0.46 – 0.2
= $0.26 billion
Output Slack
Output Slack Percentage = × 100
Actual Output
0.26
= × 100
0.2
= 130

1.4 Strongly and Weakly Efficient DMUs

In the previous section, it was claimed that the DMUs lying on

the efficiency frontier are efficient. However, further distinction
among these efficient units is possible. Consider the firm cor-
responding to the point E in Figure 1.1. This firm is considered
efficient because it lies in the efficiency frontier, but is weakly
efficient as it has a positive slack in one of its inputs (thousand
employees). Firm A is strongly efficient as it has no slack. Firm C
is also strongly efficient.

1.5 Exercises

1. Consider only one input (capital employed) and only one output
(value added). Estimate the relative efficiencies, input targets,
input slacks, output targets, and output slacks for four firms
A, B, C and D. Use the data given in section 1.2.
2. Consider two inputs (capital employed and number of em-
ployees) and one output (value added). Estimate the relative
efficiencies, input targets, input slacks, output targets, and
output slacks for the four firms A, B, C and D. Use the data
given in section 1.2.
3. The table below gives the input parameters for 19 schools.
A, B, C, etc., designate the schools. Plot them on a graph and
36 An Introduction to Data Envelopment Analysis

identify the most efficient schools. Draw the efficiency frontier.

Which is the most inefficient school and why? Also, use your
own knowledge in analytical geometry or any other idea to
compute the efficiency of the following schools: B, H, Q, R
and S.

School No. of Non-teaching No. of Teaching

Staff/Proportion of Staff/Proportion of
Students Passing in Students Passing in
the Final School Exam. the Final School Exam.
A 9.443000 252.8089
B 96.45864 202.9064
C 152.2905 175.5855
D 71.88695 150.1708
F 87.27089 144.5602
G 68.9087 139.9097
H 35.56934 121.2792
I 92.76387 100.1946
J 166.9809 94.35873
K 18.09131 93.10026
L 99.06741 91.41724
M 38.35395 90.25016
N 11.41596 83.5886
O 60.25421 79.99854
P 122.7965 74.52081
Q 30.28074 57.25475
R 82.86489 55.91398
S 27.82466 54.57391
T 50.35362 50.12258

4. Some output parameters for hospitals designated as A, B, C,

etc., are tabulated below. Plot them on a graph and identify
the most efficient hospitals. Draw the efficiency frontier. What
is the most inefficient hospital and why? Also, use your own
knowledge in analytical geometry or any other idea to compute
the efficiency of the following hospitals: B, C, H and R.
 Introduction 37

Hospital Average Number of Average Number of

Medical Surgical Intensive Medical Surgical Acute
Care Discharges/Average Discharges/Average
Staffed Beds Staffed Beds
A 0.944307 25.28089
B 9.645874 20.29064
C 15.2290500 17.55855
D 7.188695 15.01708
F 8.727089 14.45602
G 6.890870 13.99097
H 3.556934 12.12792
I 9.276387 10.01946
J 16.6980900 009.435873
K 1.809131 009.310026
L 9.906741 009.141724
M 3.835395 009.025016
N 1.141596 08.35886
O 6.025421 007.999854
P 12.2796500 007.452081
Q 3.028074 005.725475
R 8.286489 005.591398
S 2.782466 005.457391
T 5.035362 005.012258
 2

Mathematical Programming
Aspects of DEA

As discussed in the first chapter, performance evaluation for the

case of two inputs and one output was more complicated than in
the case of single input–output. Graphical analysis was used for
analysing this case. However, graphical models cannot be used if
we consider a greater number of inputs and outputs. Hence, a
general mathematical formulation is needed to handle the case of
multiple inputs and multiple outputs.
Note that the techniques of frontier analysis has been described
by Farrel in 1957, but a mathematical framework to handle fron-
tier analysis could be established only after 20 years. This math-
ematical formulation was provided by Charnes et al. (1978). This
seminal paper provided the fundamentals of the mathematical
aspects of frontier analysis. The authors also coined the term Data
Envelopment Analysis.

2.1 Mathematical Formulation

Let us use x and y to represent inputs and outputs, respectively.
Let the subscripts i and j to represent particular inputs and out-
puts respectively. Thus xi represents the ith input, and yj represent
the jth output of a decision-making unit. Let the total number of
inputs and outputs be represented by I and J respectively, where
I, J > 0.
 Mathematical Programming Aspects of DEA 39

In DEA, multiple inputs and outputs are linearly aggregated

using weights. Thus, the virtual input of a firm is obtained as the
linear weighted sum of all its inputs.

I
Virtual Input = ∑ ui xi , (2.1)
i=1

where ui is the weight assigned to input xi during the aggregation,

and ui ³ 0.
Similarly, the virtual output of a firm is obtained as the linear
weighted sum of all its outputs.

J
Virtual Output = ∑ v j y j , (2.2)
j =1

where vj is the weight assigned to output yj during the aggregation.

Also vj ³ 0.
Given these virtual inputs and outputs, the Efficiency of the
DMU in converting the inputs to outputs can be defined as the
ratio of outputs to inputs.
J

Virtual Output
∑
j 1
=
vj yj
Efficiency = = (2.3)
Virtual Input I
∑
i 1
=
ui xi

Obviously, the most important issue at this stage is the assess-

ment of weights. This is a tricky issue as there is no unique set of
weights.
For example, a school that has a good reputation of teaching
humanities will like to attach higher weights to its humanities’
output. A school that has a higher percentage of socially weaker
groups in its students would like to emphasize this fact, assigning
a greater weight to this input category. Thus, the weights assigned
should be flexible and reflect the requirement (performance) of
the individual DMUs.
This issue of assigning weights is tackled in DEA by assigning
a unique set of weights for each DMU. The weights for a DMU
40 An Introduction to Data Envelopment Analysis

are determined, using mathematical programming, as those

weights which will maximize its efficiency subject to the condition
that the efficiencies of other DMUs (calculated using the same
set of weights) is restricted to values between 0 and 1. The DMU
for which the efficiency is maximized is normally termed as the
reference or base DMU or the DMU under the assessment.

2.1.1 Fractional DEA Programs

Let there be N DMUs whose efficiencies have to be compared.
Let us take one of the DMUs, say the mth DMU, and maximize
its efficiency according to the formula given above. Here the mth
DMU is the reference DMU.
The mathematical program now is,

J
∑
j 1
=
v jm y jm
max Em = I
∑
i 1
=
uim xim

subject to
J
∑
j 1
=
v jm y jn
0≤ I ≤ 1; n = 1, 2, K, N
∑ uim xin
i=1
v jm, uim ≥ 0; i = 1, 2, K, I ; j = 1, 2, K, J (2.4)

where
Em is the efficiency of the mth DMU,
yjm is jth output of the mth DMU,
vjm is the weight of that output,
xim is ith input of the mth DMU,
uim is the weight of that input, and
yjn and xin are jth output and ith input, respectively, of the nth
DMU, n = 1, 2, …, N.
Note that here n includes m.
Consider the Firms A, B, C and D discussed in the first chapter.
Let vVA, A be the weight associated with the only output (value
 Mathematical Programming Aspects of DEA 41

added, VA) when Firm A is the reference DMU. The first subscript
denotes the output, while the second subscript denotes the refer-
ence DMU. Using a similar notation, let uCAP, A and uEMP, A represent
the weights of the two inputs, capital employed (CAP) and the
number of employees (EMP), respectively. Thus, in DEA, the ef-
ficiency of Firm A, denoted as EA, is defined as follows.

1.8vVA, A
EA = (2.5)
8.6uCAP, A + 1 .8uEMP, A
This efficiency is maximized subject to the following conditions.

1 .8vVA, A
max E A =
8.6uCAP, A + 1.8uEMP, A
subject to
1 .8vVA, A
0 ≤ EA = ≤1
8.6uCAP, A + 1 .8uEMP, A
0.2vVA A,
0 ≤ EB = ≤1
2.2uCAP, A + 1 .7uEMP, A
2.8vVA, A
0 ≤ EC = ≤1
15.6uCAP, A + 2.6uEMP, A
4.1vVA, A
0 ≤ ED = ≤1
31.6uCAP, A + 12.3uEMP, A
vVA A , uCAP A , uEMP A ≥ 0 (2.6)
, , ,

The above mathematical program, when solved, will give the

values of weights u and v that will maximize the efficiency of
Firm A. If the efficiency is unity, then the firm is said to be effi-
cient, and will lie on the frontier. Otherwise, the firm is said to be
relatively inefficient.
Note that the above mathematical program gives the efficiency
of only one firm (the reference firm—Firm A here). To get the
42 An Introduction to Data Envelopment Analysis

efficiency scores of the Firms B, C and D, more such mathematical

programs have to be solved, considering each of them as the refer-
ence firm.
For example, to obtain the efficiency of Firm B, the following
mathematical program is used.

0.2vVA, B
max EB =
2.2uCAP, B + 1 .7uEMP, B
subject to
1.8vVA, B
0 ≤ EA = ≤1
8.6uCAP, B + 1 .8uEMP, B
0.2vVA B,
0 ≤ EB = ≤1
2.2uCAP, B + 1.7uEMP, B
2.8vVA B ,
0 ≤ EC = ≤1
15.6uCAP, B + 2.6uEMP, B
4.1vVA, B
0 ≤ ED = ≤1
31.6uCAP, B + 12.3uEMP, B
vVA B , uCAP B , uEMP B ≥ 0
, , , (2.7)

2.1.2 Output Maximization and Input

Minimization DEA Programs
Note that these mathematical programs are fractional programs.
It is generally difficult to solve fractional programs. If they are
converted to simpler formulations, such as the linear programming
(LP) formats, then they can be solved easily. The simpliest way to
convert these fractional programs to linear programs is to normal-
ize either the numerator or the denominator of the fractional pro-
gramming objective function!
Let us first normalize the denominator of the objective function
of the fractional program that estimates the efficiency of Firm A.
We obtain the following linear program (LP) for maximizing the
efficiency of Firm A.
 Mathematical Programming Aspects of DEA 43

max 1 .8vVA, A
subject to
8.6uCAP, A + 1 .8uEMP, A = 1

1.8vVA, A –  8.6uCAP, A + 1 .8uEMP, A  ≤ 0

 
 
0.2vVA, A −  2.2uCAP, A + 1.7uEMP, A  ≤ 0
 
2.8vVA, A −  15.6uCAP, A + 2.6uEMP, A  ≤ 0

 
4.1vVA, A −  31.6uCAP, A + 12.3uEMP, A  ≤ 0

 
vVA A , uCAP A , uEMP A ≥ 0 (2.8)
, , ,

The weighted sum of inputs is constrained to be unity in this

linear program. As the objective function is the weighted sum of
outputs that has to be maximized, this formulation is referred to
as the Output Maximization DEA program.
An analogous LP formulation is possible by minimizing the
weighted sum of inputs, setting the weighted sum of outputs equal
to unity. That is the Input Minimization DEA program.
The following is the Input Minimization DEA program for
Firm A.
min 8.6u′ + 1 .8u′
CAP, A EMP, A
subject to
1.8vVA
′
A
,
=1

 
1.8vVA
′
,A
−  8.6uCAP
′
,A
+ 1 .8u′EMP A  ≤ 0
,
 
 
0.2vVA
′
,A 
−  2.2uCAP′
,A
+ 1 .7u′EMP A  ≤ 0
, 

 
2.8vVA
′
,A 
−  15.6u′CAP A + 2.6u′EMP A  ≤ 0
, , 

4.1v′VA, A −  31.6uCAP′
,A , 

+ 12.3u′EMP A  ≤ 0

vVA
′ , u′
A CAP A EMP A
, u′ ≥0 (2.9)
, , ,
44 An Introduction to Data Envelopment Analysis

Because of the nature of the formulations, the optimal objective

function value of the input minimization DEA program for Firm
A will be the reciprocal of the optimal objective function value
(i.e., the value of efficiency) of the output maximization DEA
program for Firm A.
These were the original models introduced by Charnes et al. in
1978. Immediately after, the authors made a minor modification
(Charnes et al. 1979). In a conventional LP, the decision variables
are non-negative—they can be either zero or positive. However,
the authors chose to define the decision variables of the DEA
programs (i.e., the weights) to be strictly positive. They replaced
the non-negativity constraints,

vVA A , uCAP A , uEMP A ≥ 0

, , ,

by the strict positivity constraints

vVA A , uCAP A , uEMP A > 0.

, , ,

This modification restricted the input and output weights such

that
vVA, A, uCAP, A, uEMP, A > ε ,

where e is an infinitesimal or non-Archimedean constant, usually

of the order of 10–5 or 10–6.
It must be emphasized that this non-Archimedean infinitesimal
is not a number, and hence in principle cannot be approximated
by any finite valued number. However, standard LP packages re-
quire that this infinitesimal be represented in the form of a small
number.
The non-Archimedean infinitesimals (i.e., es) were introduced
because, under certain circumstances, the earlier model implied
unit efficiency ratings for DMUs with non-zero slack variables
such that further improvements in performance remained feasible.
We shall discuss the use of these non-Archimedean infinitesimals
in greater detail later in this chapter.
The models developed so far are called the CCR (Charnes,
Cooper and Rhodes) models in the DEA literature.
 Mathematical Programming Aspects of DEA 45

2.1.3 General Form of CCR DEA Models1

A general output maximization CCR DEA model can be repre-
sented as follows.
J
max z = ∑ v jm y jm
j =1
subject to
I
∑
i 1
=
uim xim = 1
J I
∑
j 1
=
v jm y jn − ∑ uim xin ≤ 0 ;
i 1
=
n = 1, 2, K, N

v jm , uim ≥ ε ; i = 1, 2, K, I ; j = 1, 2, K, J (2.10)

This program can be represented in matrix form as shown below.

max z = V mT Ym
subject to
U mT X m = 1
V mT Y − U mT X ≤ 0
V mT , U mT > ε (2.11)

where X is the matrix of inputs and Y is the matrix of outputs.

Similarly, a general input minimization CCR DEA model can
be represented as follows.
I
min z′ = ∑ u′im xim
i=1
subject to
J
∑
j 1
=
v′jm y jm = 1
J I
∑
j 1
=
v′jm y jn − ∑ uim
i 1
=
′ x ≤ 0;
in
n = 1, 2, K, N

v′jm , u′im ≥ ε ; i = 1, 2, K, I ; j = 1, 2, K, J (2.12)

1
This section requires knowledge of matrix algebra, and can be skipped
without any loss of continuity.
46 An Introduction to Data Envelopment Analysis

The program can be represented in matrix form as shown below.

min z′ = U m′ T X m
subject to
V m′T Ym = 1
V m′T Y – U m′T X ≤ 0
V m′T, U m′T > ε (2.13)

2.2 Exercises

For the exercises below, use e = 10–6 whenever needed.

1. Write fractional programs for estimating the efficiencies of all
the four firms discussed in Chapter 1.
2. Write the output maximization and input minimization CCR
DEA programs for all the four firms. Consider capital employed
as the only input variable and value added as the output vari-
able.
3. Solve the CCR DEA programs listed in this section using a
suitable LP software package. List the efficiencies and the
weights for all the models.
4. Repeat exercises 2 and 3 considering capital employed and
number of employees as the two input variables, and value
added as the output variable.
5. Compute the efficiencies of the schools in the table in page 47
using an LP package. Use both output maximization and input
minimization CCR DEA programs. Make use of the same data
as in the previous exercise. Since there is no separate output
variable, assume a dummy output of 1 for all schools.
6. Compute the efficiencies of the following hospitals using an
LP package. Use both output maximization and input mini-
mization CCR DEA programs. The data to be used is the
same as in the previous exercise. As there are no separate input
variables, assume a dummy input of 1 for all the hospitals.2
2
The implication of inclusion of dummy inputs/outputs in a DEA
program is discussed in Section 6.2.2.
 Mathematical Programming Aspects of DEA 47

Compare the results of the LPs to the results of graphical cal-

culations in Exercise 1.5.

School No. of Non-teaching No. of Teaching

Staff/Proportion of Staff/Proportion of
Students Passing in Students Passing in
the Final School Exam. the Final School Exam.
A 9.443067 252.8089
B 96.4586400 202.9064
C 152.29050000 175.5855
D 71.8869500 150.1708
F 87.2708900 144.5602
G 68.9087000 139.9097
H 35.5693400 121.2792
I 92.7638700 100.1946
J 166.98090000 0094.35873
K 18.0913100 0093.10026
L 99.0674100 0091.41724
M 38.3539500 0090.25016
N 11.4159600 083.5886
O 60.2542100 0079.99854
P 122.79650000 0074.52081
Q 30.2807400 0057.25475
R 82.8648900 0055.91398
S 27.8246600 0054.57391
T 50.3536200 0050.12258

Hospital Average Number of Average Number of

Medical Surgical Intensive Medical Surgical Acute
Care Discharges/Average Discharges/Average
Staffed Beds Staffed Beds
A 0.944307 25.28089
B 9.645874 20.29064
C 15.2290500 17.55855
D 7.188695 15.01708
F 8.727089 14.45602
G 6.890870 13.99097

(Table contd.)
48 An Introduction to Data Envelopment Analysis

(Table contd.)

Hospital Average Number of Average Number of

Medical Surgical Intensive Medical Surgical Acute
Care Discharges/Average Discharges/Average
Staffed Beds Staffed Beds
H 3.556934 12.12792
I 9.276387 10.01946
J 16.6980900 009.435873
K 1.809131 009.310026
L 9.906741 009.141724
M 3.835395 009.025016
N 1.141596 08.35886
O 6.025421 007.999854
P 12.2796500 007.452081
Q 3.028074 005.725475
R 8.286489 005.591398
S 2.782466 005.457391
T 5.035362 005.012258

2.3 Dual DEA Models

The basic theory of linear programming states that every linear

programming problem (usually called the primal problem) has
another closely related linear program, called its dual. Thus, all
the linear programming problems developed in Section 2.1 have
duals. These duals play a very important role in DEA.
Consider the output maximizing DEA program for Firm A.
Let us call this as the primal problem and write its dual.
It is possible to write the dual of any linear programming prob-
lem using certain rules. These rules are available in textbooks
on linear programming, such as Taha (1997). Following the
definitions and rules described in Taha (1997), let us first write
the primal problem in standard form as follows.3

3
Please note that the standard form of the primal model is written only
in case of the procedure suggested by Taha (1997) for writing the dual of a
linear program. Another procedure is to skip this step, and directly write
the dual (2.15).
 Mathematical Programming Aspects of DEA 49

max 1 .8vVA, A
subject to
8.6uCAP, A + 1 .8uEMP, A = 1

1.8vVA, A −  8.6uCAP, A + 1.8uEMP, A  + pA = 0

 
0.2vVA, A −  2.2uCAP, A + 1 .7uEMP, A  + pB = 0
 
2.8vVA, A − 15.6uCAP, A + 2.6uEMP, A  + pC = 0

 
4.1vVA, A −  31.6uCAP, A + 12.3uEMP, A  + pD = 0

 
vVA A , uCAP A , uEMP A , pA , pB , pC , pD ≥ 0 (2.14)
, , ,

Let q be the dual variable corresponding to the equality con-

straint that normalizes the weighted sum of inputs. Let l be the
dual variable corresponding to the other inequality constraints of
the primal. The dual can be written as follows.
min θ A
such that
1.8λ AA + 0.2λ BA + 2.8λCA + 4.1λ DA ≥ 1.8
8.6θ A − 8.6λ AA − 2.2λ BA − 15.6λCA − 31.6λ DA ≥ 0
1.8θ A − 1.8λ AA − 1.7λ BA − 2.6λCA − 12.3λ DA ≥ 0
λ AA , λ BA , λCA , λ DA ≥ 0
qA unrestricted (2.15)

Similar to the notations of the primal, the first subscript of the

dual variables refers to all the DMUs, while the second denotes
the reference DMU. For convenience, the comma is omitted.
For completeness, the primal (with e constraints) in standard
form is also given below.4

4
Please note that the above standard form of the primal is written only
in case of the procedure suggested by Taha (1997) for writing the dual of a
linear program. Another procedure is to skip this step.
50 An Introduction to Data Envelopment Analysis

max 1 .8vVA, A
subject to
8.6uCAP, A + 1 .8uEMP, A = 1

1.8vVA, A −  8.6uCAP, A + 1.8uEMP, A  + pA = 0

 
0.2vVA, A −  2.2uCAP, A + 1 .7uEMP, A  + pB = 0

 
2.8vVA, A − 15.6uCAP, A + 2.6uEMP, A  + pC = 0

 
4.1vVA, A −  31.6uCAP, A + 12.3uEMP, A  + pD = 0
 
vVA A − pVA = ε
,
uCAP A − pCAP = ε
,
uEMP A − pEMP = ε
,
pVA , pCAP , pEMP , pA , pB , pC , pD ≥ 0 (2.16)

Let s and t represent the dual variables corresponding to the e

constraints of inputs and outputs, respectively. The dual can be
written as follows.

min θ A − ε  tVA, A + sCAP, A + s EMP, A 

 
such that
1.8λ AA + 0.2λ BA + 2.8λCA + 4.1λ DA − tVA, A = 1 .8
8.6θ A − 8.6λ AA − 2.2λ BA − 15.6λCA − 31.6λ DA − sCAP, A = 0
1.8θ A − 1.8λ AA − 1.7λ BA − 2.6λCA − 12.3λ DA − s EMP A = 0,
λ AA , λ BA , λCA , λ DA , tVA A , sCAP A , s EMP ≥0
, , ,A
qA unrestricted (2.17)

We shall mostly use the former version of the dual (without the
e constraints in the primal). However, our conclusions can be
generalized for the other dual too.
 Mathematical Programming Aspects of DEA 51

2.3.1 Comparing Primal and Dual

Using the basic theory of linear programming, the following ob-
servations can be made.

(a) As the optimal values of primal and dual objective func-

tions are equal, qA represents the efficiency of Firm A.
(b) The number of constraints of the primal depends upon
the number of DMUs, while the number of constraints of
the dual depends upon the number of inputs and outputs.
(c) The computational efficiency of LP codes depends to a
greater extent upon the number of constraints than on
the number of variables. In a typical DEA exercise, about
5 inputs and 5 outputs are considered, while the number
of units being compared is much larger (of the order of
hundreds or even thousands). Hence, the dual formu-
lation is computationally more efficient than the primal.

2.3.2 Interpreting the Dual

Though we have derived the dual mathematically, rather than in-
tuitively, from the primal, we can interpret the dual (2.15) intui-
tively. First of all, let us observe that while the primal provided
optimal weights to inputs and outputs, the dual provides weights
to the DMUs (l).
The first dual constraint is the following.

1.8λ AA + 0.2λ BA + 2.8λCA + 4.1λ DA ≥ 1.8

The left-hand side (LHS) of this constraint is the weighted

sum of the outputs of all the firms; the right-hand side (RHS) is
the output of the reference firm (A here). This constraint states
that the dual variables l should be chosen such that the weighted
combination of all the outputs of all the firms should be at least
equal to the output of the reference firm. As we shall see later, if
the firm is efficient, the strict equality will hold, with no slack in
the constraint. We shall also see later that for an inefficient firm,
the weights are actually the weights to be assigned to their peers.
Let us now consider the second constraint of the dual.

8.6λ AA − 2.2λ BA − 15.6λCA − 31.6λ DA ≤ 8.6θ A

52 An Introduction to Data Envelopment Analysis

This constraint corresponds to the capital input. It says that

the weighted combination of the capital inputs of all the firms
cannot be more than the capital input for the reference firm multi-
plied by its efficiency. This can be proved using the complementary
slackness conditions of linear programming (see Taha 1997) that,
at optimality, for an efficient firm, q * is unity; the variable q *
refers to optimal values. The constraint becomes strict equality
with zero slacks (Cooper et al. 2000).
We know that Firm A is an efficient firm. We already know that
q A* is equal to unity, and the optimal values of slack variables of
inputs and outputs are zero. Therefore, the constraints in Equ-
ation (2.15) are reduced to,
θ A* = 1
and
1.8λ*AA = 1.8
8.6λ*AA = 8.6θ A*
1.8λ*AA = 1.8θ A* (2.18)
Thus,
l*AA = 1
in this case. The other ls are equal to zero.
Let us now consider Firm B. The dual DEA program for this
firm is the following.
min θ B
such that
1.8λ AB + 0.2λ BB + 2.8λCB + 4.1λ DB ≥ 0.2
8.6λ AB + 2.2λ BB + 15.6λCB + 31.6λ DB ≤ 2.2θ B
1.8λ AB + 1.7λ BB + 2.6λCB + 12.3λ DB ≥ 1 .7θ B
λ AB , λ BB , λCB , λ DB ≥ 0
θ B unrestricted (2.19)

We know that this firm is inefficient, and that firm A is its

peer. If the LP given above is solved for firm B, the following
solution is obtained. (For details of the notations of points A,
B, …, E, please see Figure 1.1.)
 Mathematical Programming Aspects of DEA 53

OE
θ B* = = 0.434
OB
and
1.8λ*AB = 0.2
8.6λ*AB = 2.2θ B*
1.8λ*AB < 1.7θ B*
Note that
0.2 1
l*AB = =
1.8 9
is positive in this case as Firm A is the only peer for Firm B.
Other ls are zero in value. The value of qB can also be obtained
from the above equalities.
Thus, mathematically, peers for an inefficient DMU can be
identified by solving the dual DEA program. Peers are those effi-
cient DMUs that have positive ls in the optimal solution of the
dual DEA program.
Input target for capital input can be obtained as

2.2qB = 0.95

Note that the constraint representing capital input,

8.6l*AB = 2.2q B*

is satisfied in equation form; however, there is a slack in the

constraint representing employees input

1.8l*AB < 1.7q B*

This is because point E (in Figure 1.1), which represents the best
achievable performance for Firm B, corresponds to the same capital
input as Firm A but has a higher input of employees. This slack
means that though the point E is efficient, there is still scope for
reducing the employees input.
In DEA, a DMU is considered efficient if and only if q * = 1
and all slacks are zero; otherwise it is inefficient. However, a DMU
may be inefficient in various ways. If q * = 1 at the optimal solution
54 An Introduction to Data Envelopment Analysis

but some of the slack variables do not equal zero, then there exists
a combination of other units which does not dominate the current
output vector of the reference DMU, but uses less resources. As
discussed in Chapter 1, these DMUs may be termed weakly effi-
cient. Thus, the firm represented by point E in Figure 1.1 is
weakly efficient, while Firms A and B are strongly efficient firms.
The value of the slack at point E is,
* * *
sEMP, B = 1.7q B – 1.8 lAB = 0.54

We can now understand why the LHS of the constraints corres-

ponding to inputs are multiplied by the efficiency score q, as can
be derived from Equation 2.15: they represent the best achievable
performance. For example, for Firm B, 2.2q B* represents the best
achievable value (i.e., input target) for capital employed, while
1.7q B* represents the best achievable value for employees. Together,
these two values represent point E.
Weakly efficient firms can be distinguished from strongly effi-
cient ones using e (infinitesimal) constraints in the dual formula-
tion (2.17). The dual objective function for a firm corresponding
to point E is
min θ E − ε s EMP, E ,
*
which is less than 1 as e and sEMP, B are positive.
Let us now consider four firms P, Q, R and T that produce the
same level of a single output Y, from two inputs X1 and X2, shown
in Figure 2.1.
Firms P and R are efficient. They represent the best practice.
This implies that no other firm or no linear combination5 of firms
can be identified which produces the same level of output for less
than either or both the inputs.
As we have seen earlier, the dual DEA program for these two
firms will indicate zero slacks and unit efficiency. For example,
the dual DEA optimal results for Firm P are the following, where
5
A linear combination of a set of variables is one in which each variable
is multiplied by a coefficient and the products summed. For example, it is
possible to create a linear combination (Y) of variables X1, X2, and X3 as
Y = 2X1 + 100X2 + 0.5X3. If the formula for the new variable contains
functions such as square roots or logarithms, then the new variable is not a
linear combination of the other variables.
 Mathematical Programming Aspects of DEA 55

Figure 2.1 Efficiency frontier of four firms consuming two inputs

lP* = q P* = 1. For convenience, the subscript representing the refer-

ence firm is suppressed.
θ P* = 1
and
X 1*P λ*P = X1*Pθ P*
X 2*P λ*P = X 2*Pθ P*
YP λ*P = YP (2.20)

Note that both the efficient firms (Firms P and R) are peers for
the inefficient Firm Q. Hence, lP* and lR* are positive. For this
firm, the following equalities hold as obtained from the optimal
DEA dual results (2.21).
OU
θ Q* = ≤1
OQ
and
X 1*P λ*P + X1*Rλ*R = X1*Qθ Q*
X 2*P λ*P + X 2*R λ*R = X 2*Qθ Q*
YP λ*P + YRλ*R = YQ (2.21)
56 An Introduction to Data Envelopment Analysis

Note that RHSs of the above equations represent the best achiev-
able target performance (corresponding to point U) for Firm B.
*
Also, (X 1Qq Q* = X 1U
* *
; X 2Qq Q* = X 2U
*
), the hypothetical firm U is
actually a linear combination of the best practice firms P and R,
where the weights for the linear combination are the ls obtained
from the dual DEA program.
For completeness, the dual DEA program corresponding to
input minimizing multiplier model for Firm A is given below.

min φ A
such that
1.8µ AA + 0.2µ BA + 2.8µ CA + 4.1 µ DA ≥ 1.8φ A
8.6µ AA − 2.2µ BA − 15.6µ CA − 31.6µ DA ≤ 8.6
1.8µ AA − 1.7µ BA − 2.6µ CA − 12.3µ DA ≥ 1 .8
µ AA , µ BA , µ CA , µ DA ≥ 0
φ A unrestricted (2.22)

Before closing this section, it is important to compare the op-

timal values of the objective functions qA and fA. Note that qA is
the dual objective function for Firm A corresponding to the output
maximizing primal, and fA is the dual objective function for Firm
A corresponding to the input minimizing primal. At the optimal
solution, the optimal objective function values of the primal and
the dual are equal. Hence, by comparing the optimal objective
function values of the output maximizing and the input minimiz-
ing primals, it is clear that the optimal value of qA is the reciprocal
of the optimal value of fA.

2.3.3 Two-stage Optimization Procedure

We have seen in Section 2.3 that the use of infinitesimals (e) will
distinguish weakly efficient DMUs from strongly efficient ones.
As discussed in that section, the dual formulation (with e con-
straints) to obtain the efficiency of Firm A is the following (Equ-
ation 2.17, Section 2.3).
 Mathematical Programming Aspects of DEA 57

min θ A − ε  tVA, A + sCAP, A + s EMP, A 

 
such that
1.8λ AA + 0.2λ BA + 2.8λCA + 4.1λ DA − tVA, A = 1 .8
8.6θ A − 8.6λ AA − 2.2λ BA − 15.6λCA − 31.6λ DA − sCAP, A = 0
1.8θ A − 1.8λ AA − 1.7λ BA − 2.6λCA − 12.3λ DA − s EMP A = 0
,
λ AA , λ BA , λCA , λ DA , tVA A , sCAP A , s EMP A ≥ 0
, , ,
θ A unrestricted

However, the numeric values for infinitesimals (e’s) in actual

computations should be chosen to be much smaller than other
input and output values so that they will not affect optimization.
This is often troublesome.6 A two-stage optimization procedure
(Ali and Seiford 1993; Joro et al. 1998) has been suggested to
avoid using the infinitesimals (e) in DEA computations.
The two-stage optimization procedure corresponding to the dual
formulation (Equation 2.23) for Firm A can be written as follows.
 
lex min θ A ,  tVA, A + sCAP, A + s EMP, A 
  
such that
1.8λ AA + 0.2λ BA + 2.8λCA + 4.1λ DA − tVA, A = 1 .8
8.6θ A − 8.6λ AA − 2.2λ BA − 15.6λCA − 31.6λ DA − sCAP, A = 0
1.8θ A − 1.8λ AA − 1.7λ BA − 2.6λCA − 12.3λ DA − s EMP, A = 0
λ AA , λ BA , λCA , λ DA , tVA A , sCAP A , s EMP A ≥ 0
, , ,
θA unrestricted (2.23)

Here ‘lex min’ means that the objective function q is mini-

mized first (Stage 1); if the solution is not unique, the second
objective function is minimized lexicographically, subject to the
additional constraint, that q ³ q * (where q * is the optimal value
of q ) (Stage 2). In case the optimal solution of the first objective
function is unique, the second optimization is not needed.
6
See Chapter 5 for further discussion on the computational features of
DEA.
58 An Introduction to Data Envelopment Analysis

The optimization models for Stages 1 and 2 are given below.

Stage 1
min θ A
such that
1.8λ AA + 0.2λ BA + 2.8λCA + 4.1λ DA − tVA, A = 1 .8
8.6θ A − 8.6λ AA − 2.2λ BA − 15.6λCA − 31.6λ DA − sCAP, A = 0
1.8θ A − 1.8λ AA − 1.7λ BA − 2.6λCA − 12.3λ DA − s EMP, A = 0
λ AA , λ BA , λCA , λ DA , tVA A , sCAP A , s EMP A ≥ 0
, , ,
θA unrestricted (2.24)

Note that this Stage 1 model is the same as the DEA model
with no e constraints.

Stage 2

min  tVA, A + sCAP, A + s EMP, A 

 
such that
1.8λ AA + 0.2λ BA + 2.8λCA + 4.1λ DA − tVA, A = 1 .8
8.6θ A − 8.6λ AA − 2.2λ BA − 15.6λCA − 31.6λ DA − sCAP, A = 0
1.8θ A − 1.8λ AA − 1.7λ BA − 2.6λCA − 12.3λ DA − s EMP, A = 0
θA ≥ θ*
λ AA , λ BA , λCA , λ DA , tVA A , sCAP A , s EMP A ≥ 0
, , ,
θ A unrestricted (2.25)

where q * is the optimal value of q A in stage 1.

2.4 Multiplier and Envelopment DEA Programs

We know that the dual of a dual is primal. Hence, the terms

primal DEA program and dual DEA program are relative. New
terms are being increasingly used to represent DEA formulations.
 Mathematical Programming Aspects of DEA 59

The DEA programs involving weights of inputs and outputs

(u and v) are called Multiplier DEA Programs. Those involving
weights of firms (q and l) are called Envelopment DEA Programs.
Output maximizing and input minimizing multiplier versions
of DEA programs have been discussed in Section 2.1. A general
envelopment DEA program corresponding to the output maximizing
multiplier model can be written as follows.

min θ m
such that
N
∑
n 1
=
y jnλ n ≥ y jm ; j = 1, 2, K, J
N
∑
n 1
=
xinλn ≤ θ m xim ; i = 1, 2, K, I

λ n ≥ 0;n = 1, 2, K, N
θ m unrestricted (free) (2.26)

This model can also be succinctly represented using matrix

notation as follows.
min θm
,
θ λ

such that
Yλ ≥ Ym
Xλ ≤ θ X m
λ ≥ 0; θ m free (2.27)

Similarly, the envelopment DEA program corresponding to the

input minimizing multiplier model is given below in its general
matrix form.
max φ m
φ, µ

such that
Yµ ≥ φ mYm
Xµ ≤ X m
µ ≥ 0; φ m free (2.28)
60 An Introduction to Data Envelopment Analysis

2.5 Input and Output Oriented Envelopment

DEA Programs

Let us study the two envelopment versions, one involving q and

the other involving f. The version involving q aims to produce
the observed outputs with minimum inputs. That is why inputs
are multiplied by efficiency, according to its constraint rules. Be-
cause of this characteristic, this version is often referred to as an
input oriented envelopment DEA program. The other version in-
volving f is referred to as an output oriented envelopment DEA
program as it aims to maximize output production, subject to
the given resource level.
Note that the dual of the output maximizing multiplier program
is the input oriented envelopment program. Similarly, the dual
of the input minimizing multiplier program is the output oriented
envelopment program.
Let us examine the behaviour of the input and output oriented
envelopment DEA programs more closely, using data regarding
the Firms A, B, C, and D. For simplicity, let us consider only one
input (capital employed) and only one output (value added).
When the values are plotted in a graph (see Figure 2.2), we
observe that Firm A has the maximum output (value added) for a
given input (capital). Hence, Firm A is the most efficient, and
acts as a peer for all other (inefficient) firms. If we can draw a line

Figure 2.2 Behaviour of input and output oriented envelopment

DEA programs
 Mathematical Programming Aspects of DEA 61

joining the origin to A, all other firms are found to the left of this
line.
The input oriented envelopment program used to estimate the
efficiency of Firm B is the following.
min θ B
such that
1.8λ AB + 0.2λ BB + 2.8λCB + 4.1λ DB ≥ 0.2 (for value added)
8.6λ AB + 2.2λ BB + 15.6λCB + 31.6λ DB ≤ 2.2θ B (for capital)
λ AB , λ BB , λCB , λ DB ≥ 0
qB unrestricted (2.29)
Consider its solution. Since Firm A is its peer, zero weights are
assigned to Firms C and D (i.e., no other firm influences the per-
formance of Firm B). Hence,
λ*AB ≠ 0 and λ*BB = λ*CB = λ*DB = 0
0.2 1
λ*AB = =
1 .8 9
8.6 × λ*AB
θ B* = = 0.434
2.2
Given q *B, the best achievable capital input for Firm B is

2.2 ´ 0.434 = 0.96.

The point is shown by B1 in Figure 2.2. Thus, the constraint on

capital input gives the target for capital input. Therefore, the
input oriented envelopment DEA program projects the Firm B
from right to left, as shown by the arrows in Figure 2.2.
Let us study the behaviour of the output oriented envelopment
program for the Firm B.
max φ B
such that
8.6µ AB + 2.2µ BB + 15.6µ CB + 31.6µ DB ≤ 2.2 (for capital)
1.8µ AB + 0.2µ BB + 2.8µ CB + 4.1 µ DB ≤ 0.2φ B (for value added)
µ AB , µ BB , µ CB , µ DB ≥ 0
fB unrestricted (2.30)
62 An Introduction to Data Envelopment Analysis

As before, we have the following results.

µ *AB ≠ 0 and µ BB
* = µ* = µ* = 0
CB DB
2.2
µ *AB = = 0.256
8.6
1 .8 × µ *AB
φ B* = = 2.30
0.2
The best achievable performance, using the constraint on out-
put, is 0.2 ´ f B* = 0.46. This corresponds to the point B0 in Fig-
ure 2.2. Thus, the output oriented envelopment DEA program
projects the Firm B from bottom to top.
The projections can be seen more clearly for the Firm D. The
calculations corresponding to this firm are left as an exercise for
students.
Further information on this topic can be obtained from Seiford
and Thrall (1990).

2.6 Relationships among Different

DEA Formulations7

So far, we have studied four different DEA programs.

(a) Output maximizing multiplier program

(b) Input minimizing multiplier program
(c) Input oriented envelopment program, and
(d) Output oriented envelopment program.

The general matrix representations of the four programs are

given below.

7
This section requires knowledge of matrix algebra, and can be skipped
without any loss of continuity.
 Mathematical Programming Aspects of DEA 63

Multiplier Versions Envelopment Versions

Output Maximizing Input Oriented
max z = VmT Ym min θm
U, V θ ,λ

such that such that

UmT Xm = 1 Yλ ≥ Ym
VmTY − UmT X ≤ 0 Xλ ≤ θ m Xm
(
VmT , UmT ≥ 0 or VmT , UmT ≥ ε ) λ ≥ 0; θ m free

Input Minimizing Output Oriented

min z′ = Um′T Xm max φm
U′, V′ φ, µ

such that such that

Vm′ TYm = 1 Yµ ≥ φ mYm
Vm′ TY − Um′T X ≤ 0 Xµ ≤ Xm
(
Vm′ T, Um′T ≥ 0 or Vm′ T, Um′T ≥ ε ) µ ≥ 0; φ m free

Using the theory of duality, we can make certain observations

regarding the relationships among the optimal values of these
programs. See Seiford and Thrall (1990) for rigorous treatments
and proofs of the relationships. In the next few paragraphs, we
present the relationships in a more qualitative manner.
It has been explained at the beginning of this chapter (Sec-
tion 2.1) that
1
z* = .
z′ *
This is because z and z′ represent the denominator and numerator
of the equation defining efficiency.
Consider the first constraint of the output maximizing multi-
plier version of the program. Divide both sides of this equation
by z*. Note that 0 £ z* £ 1, because efficiency is restricted to be
between zero and unity in the corresponding multiplier model.
Also, assume that z* £ 0. We will not use the asterisk in further
discussions, but deal only with optimal values.
64 An Introduction to Data Envelopment Analysis

U mT X m = 1
UT X 1
⇒ m m =
z z
 U mT 
⇒  X = 1
z  m z
 
 U mT 
⇒   X = z′ (2.31)
 m
 z 

Comparing the last equation with the objective function of the

input minimizing multiplier version, we get,

1 *
U ′* =  * U (2.32)
z 

Using a similar argument, we can show

1 *
V ′′* = * V (2.33)
z 

Let (q *, l*) and (f *, m *) be the optimal solutions for the input

oriented and output oriented DEA envelopment programs respec-
tively (with the subscript representing the reference firm m sup-
pressed). From analogies with their multiplier versions, we have
1
φ* = * (2.34)
θ

Divide the two constraints of the input oriented envelopment

version of the program by q m* . Note that 0 £ q m* £ 1 by definition.
Further, assume that q m* ¹ 0. As before, though asterisks are not
used below, let us deal with optimal values only. We have,

Yλ Y  λ 
≥ m ⇒ Y   ≥ φ mYm
 (2.35)
θm θm θm 
  θ  
X  λ  ≤ m X m ⇒ X  λ  ≤ X m (2.36)
θm  θm θm 
 Mathematical Programming Aspects of DEA 65

Comparing the above two equations with the two constraints

of output oriented envelopment version of the program, we have

 1 
µ * =  * λ* (2.37)
θ 

We can easily verify this using the numerical calculations we

have carried out for Firm B.

1 1
= = 2.3 = φ *
θ * 0.434
 1  *  1   1  0.256
 * λ =  ×  = = µ*
θ
   0.434 9
  

2.7 Exercises

Some of the questions in this exercise require knowledge of matrix

algebra.

1. Write short notes on peers.

2. Say Yes or No with reasons.

(a) The input oriented envelopment DEA problem is the

dual of the input minimizing multiplier DEA problem.
(b) The objective function of an input oriented DEA pro-
gram can take any value.
1
(c) θ * * = 1
φ
(d) u′ * = z′ *u*
(e) Output targets for an inefficient DMU can be calculated
using the input oriented envelopment program.

3. State the errors in the general envelopment DEA program for

a reference firm ‘m’ shown below.
66 An Introduction to Data Envelopment Analysis

min ω m
such that
Yγ ≤ ωYm
Xγ ≤ X m
ω, γ ≥ 0

4. Interpret the input constraints in an input oriented envelop-

ment DEA program.
5. Compare and contrast the envelopment and multiplier DEA
programs.
6. Explain input and output targets and slacks in DEA using a
graph for the case of two outputs and one input.
7. Write the input minimizing multiplier DEA program using
the following data.

A B C D E
Input I 1 2.5 2 6 10
Output O 5 12.50 11 40 80

Estimate the efficiency of the Firm B for the above data (using
hand calculations). Identify its peer in case B is an inefficient
firm. Using your results, calculate the optimal objective function
value and other parameters for the output oriented DEA program
for the same firm.
 3

Economies of Scale

In economics, the concept of production function specifies the output

in an industry for all combinations of inputs. A production func-
tion can be depicted on a two-dimensional graph as shown in Fig-
ure 3.1. To facilitate depiction on a two-dimensional graph, let
us aggregate inputs and outputs, i.e., assume that all the inputs
are aggregated into one input, and all outputs are aggregated into
one output.
Suppose that a firm consumes inputs amounting to X1, and
produces Y1 amounts of output. In automated operations, it is
possible to consume a certain amount of inputs, and produce
more than a proportional amount of output. For example, consider
a manufacturer producing shock absorbers. If only a few shock
absorbers need to be produced, he may prefer to do so manually.
But, if he needs to produce a large amount of shock absorbers, he
may prefer automate his process. Hence he will be able to produce
a larger amount of output in proportion to the inputs. Therefore,
he can consume a larger input X2, and can produce the output Y2,
which is more than a proportional increase in output, i.e.,

Y2 X 2
> . (3.1)
Y1 X1

This concept is termed Economy of Scale. Actually, the manu-

facturer is operating under Increasing Returns to Scale as his returns
(profits) will increase if he increases his production.
68 An Introduction to Data Envelopment Analysis

Figure 3.1 Production functions

One can define Increasing Returns to Scale (IRS) as a property

of a production function such that changing all inputs by the
same proportion changes the output by a greater extent than the
proportional value.
However, beyond a limit, IRS does not hold. If the manufacturer
needs to produce billions of shock absorbers, he might find it dif-
ficult to produce that amount because of storage problems and
limits on the supply of raw materials. In this case, he is said to be
operating under Decreasing Returns to Scale (DRS).
Combining the two extremes (IRS and DRS) would necessitate
Variable Returns to Scale (VRS). This property signifies that in a
production process, the operations will follow IRS or DRS (or
CRS—see discussion below) for different ranges of output. The
same concept can be extended to areas other than production
processes, such as schools, banks, hospitals, and other categories
of DMUs.
 Economies of Scale 69

Note that the IRS changes to DRS at a particular level of produc-

tion, represented by (X2, Y2) in Figure 3.1. At this point a DMU
is said to be operating at its Most Productive Scale Size (MPSS),
because it enjoys the maximum possible economy of scale.
Another variant of economies of scale is Constant Returns to
Scale (CRS). This property signifies that the manufacturer is able
to scale the inputs and outputs linearly without increasing or
decreasing efficiency. In such a case, he is able to obtain of output
Y1′ by consuming X1 of input, Y2 by consuming X2 and, Y3′ by con-
suming X3. This is a significant assumption because may be valid
over limited ranges. Hence, if the CRS assumption is employed
for a particular case, its use must be justified showing evidences
for the existence of CRS.

3.1 Returns to Scale and DEA

It is important to note that the DEA models discussed so far as-

sume that the operations follow constant returns to scale. This
represented one of the most limiting factors for the applicability
of DEA, at least in the early years. DEA has not received wide-
spread attention for the analysis of production processes because
of this limitation. Many economists viewed this assumption as
over-restrictive and preferred alternative statistical procedures in
spite of the advantages offered by DEA.
Modifications on DEA to handle VRS categories were first de-
scribed in 1984, when Banker et al. (1984) came up with a simple
yet remarkable modification to the CCR DEA models in order to
handle variable returns to scale. This modification was suggested
by comparing some previous studies on production functions. We
will not report the previous studies, and hence will not provide
rigorous proofs of the modification, but we shall certainly study
the effect of modification intuitively. Those interested in rigorous
proof should consult Banker et al. (ibid.).
Before discussing returns to scale properties in the context of
DEA, let us complete estimating the parameters (q and l) for
the Firms C and D (for the case of one input—capital employed,
CAP and one output—value added, VA). Students are advised to
check these computations.
70 An Introduction to Data Envelopment Analysis

for Firm C,
l*AC = 1.56 and q *C = 0.86
for Firm D,
l*AD = 2.28 and q *D = 0.62
Also, as previously calculated,
for Firm A,
l*AA = 1 and q A* = 1
for Firm B,
l*AB = 1/9 and q B* = 0.434.
Observe that,
l*AA = 1; l*AB < 1; l*AC > 1 and l*AD > 1.

We can identify a relationship between the values of l and the

scales of operation of the firms. The performances of the four
Firms A, B, C, and D are plotted in Figure 3.2. Note that the
scale of operation of Firm B is smaller, and that of Firms C and D
is larger, as compared to the efficient Firm A. Accordingly, the
weights of lA for these firms differ. The target for Firm B is scaled

Figure 3.2 Scale of operations of Firms A, B, C, and D

 Economies of Scale 71

down as depicted by the relation l*AB < 1, while the targets for
Firms C and D are scaled up as depicted by the relationships
l*AC > 1 and l*AD > 1.
It has already been shown in previous chapter that, when only
capital employed is considered as input and value added as output,
Firm A displays the highest ratio of input to output. Hence, Firm
A is the most efficient and is considered to be operating at the
most productive scale size. Firms operating at lower scale sizes
(such as Firm B) are said to be operating under IRS because they
can achieve greater economies of scale if they increase their volume
of operation. Note that as observed earlier,

l*AB < 1.

Firms operating at higher scales sizes (such as Firms C and D)

are said to be operating under DRS. Again, as observed earlier,
note that
l*AC > 1 and l*AD > 1
and also l*AA = 1.

In other words, a useful test of returns to scale properties of

DMUs can be obtained by observing the corresponding values
of l*.

(a) If l*bp < 1 Þ Increasing Returns to Scale, IRS

(b) If l*bp > 1 Þ Decreasing Returns to Scale, DRS

where bp denotes the best practice DMU. Note that l*bp = 1 for the
best practice DMU.
We have considered one input (capital employed, CAP) and only
one output (value added, VA). But, in practice, we may need to
consider a greater number of inputs and outputs. In such cases,
the foregoing conditions will be modified as follows (Ganley and
Cubbin 1992).
N
∑
n 1
=
λ n < 1 ⇒ Increasing Returns to Scale, IRS

N
∑
n 1
=
λ n > 1 ⇒ Decreasing Returns to Scale, DRS
72 An Introduction to Data Envelopment Analysis

Note that, for efficient firms,

N
∑
n 1
=
λ n = 1.

Note also that some firms that are not efficient in the models
considered so far may become efficient if we assume variable re-
turns to scale relaxing the assumption of CRS.
Rating firms as efficient/inefficient depends upon the con-
straints imposed on a CCA DEA program. Suppose we force the
N
condition ∑
n 1
=
λ n = 1 in the CCR DEA program. The introduction

of this additional constraint ensures that firms operating at

different scales are recognized as efficient. Therefore, the envelop-
ment is formed by the multiple convex linear combinations of
N
best practice (incorporating VRS). The constraint ∑
n 1
=
λn = 1 is

termed the convexity constraint in the mathematics literature.

The VRS frontier for the four firms is shown in Figure 3.3.

Figure 3.3 CRS and VRS frontiers for the Firms A, B, C, and D
 Economies of Scale 73

3.1.1 Variable Returns to Scale Envelopment

DEA Programs
Thus, the DEA envelopment program for considering variable
returns to scale is the following.

min θm
,
θ λ

such that
Yλ ≥ Ym
Xλ ≤ θ X m
N
∑
n 1
=
λn = 1 (or, eT λ = 1, where e is a unit vector)

λ ≥ 0; θ m free (3.2)

As mentioned earlier, this modification was first suggested

by Banker et al. (1984). Hence, the foregoing DEA model is
termed the BCC (Banker, Charnes and Cooper) model. In general,
DEA programs incorporating the additional convexity constraint
to take into account variable returns to scale are called BCC DEA
models or VRS DEA models. In contrast, CCR DEA models are
also called CRS DEA models.
According to Figure 3.3, all the four firms have been recognized
as efficient. They were considered inefficient by the CCR program
because of their differences in scale sizes.
Consider another Firm E using CAP = 7.5 to produce VA = 1.
It will certainly be designated as inefficient because Firms B and
A operate more efficiently, though they are smaller and larger in
size, respectively, as compared to E. The VRS efficiency of Firm
E is 0.72, with

l*AE = 0.5 = l*BE.

Note that the input oriented as well as output oriented envelop-

ment models will project E on to the Facet AB.
Consider another Firm F. As can be observed from Figure 3.3,
it is inefficient, but it is projected onto different facets depending
upon which orientation model is used. The input oriented model
74 An Introduction to Data Envelopment Analysis

projects Firm F on the Facet AB on to point G, while the output

oriented model projects it at the Facet AC on to point H. Of
course, without the convexity constraint
N
∑
n 1
λn = 1 ,
=

the CCR envelopment models will project the firm onto the points
M and L respectively.

3.1.2 Non-increasing and Non-decreasing Returns to

Scale Envelopment DEA Programs
N
We now know that appending the constraint ∑
n 1
=
λn = 1 has the

effect of introducing VRS into the model. Not appending such a

constraint has the effect of introducing CRS.
N
What is the effect of introducing the constraints, ∑
n 1
=
λn ≤ 1 or
N
∑
n 1
=
λn ≥ 1 ?
N
Suppose, we add the constraint ∑
n 1
λn ≤ 1 . Firm B, which is at
= N
IRS, will be chosen as efficient only if ∑ λn = 1 is forced. But,
N n 1 =

∑
n 1
=
λn ≤ 1 does not force this. Without any convexity constraint,
N
Firm B has the constraint ∑ λn < 1 . This requirement is allowed
N n 1=

by the condition ∑ λn ≤ 1 . Hence, Firm B will not be considered

n 1=
efficient.
N
In contrast, for Firms C and D, ∑n 1
=
λn > 1, which is not allowed
N N
by the condition ∑
n 1=
λ n ≤ 1 . Hence, the condition ∑
n 1
=
λn = 1 is

forced, in their case. Accordingly, they will be considered efficient

by the model.
 Economies of Scale 75

N
Thus, adding the constraint, ∑
n 1
=
λn ≤ 1 , has the effect of forcing

CRS up to Firm A, and VRS beyond it. Firm B, which is operating

under IRS, will be considered inefficient, while Firms C and D,
which are operating under DRS, will be considered efficient. Thus
the resulting model will be said to be capturing Non-Increasing
Returns to Scale (NIRS). The NIRS frontier is shown in Fig-
ure 3.4.

Figure 3.4 NIRS DEA frontier

Using a similar argument, we can prove that the condition

N
∑
n 1
=
λn ≥ 1 will capture Non-Decreasing Returns to Scale (NDRS).

If this constraint is introduced, Firm B will be considered efficient,

while Firms C and D will be considered inefficient. The NDRS
frontier is shown in Figure 3.5.
88 An Introduction to Data Envelopment Analysis

A data set of Japanese electric power generation companies

Name Generation Fuel Number Total
(DMU) Capacity Consumption of Generation
(MW) (10 9 kcal) Employees (GWh)
Chubu 18,075 1,53,335 2,472 69,300
Chugoku 6,406 56,194 1,176 25,717
Dengen-kaihatsu 4,655 65,728 825 30,071
Fukui 250 754 59 326
Fukuyama 699 9,704 187 4,339
Hokkaido 3,060 28,850 659 13,120
Hokuriku 2,662 17,127 664 7,648
Jyouban 1,625 15,758 206 6,940
Kansai 18,581 1,14,707 3,132 50,846
Kashima 1,400 12,380 141 5,407
Kimitsu 950 12,745 177 5,533
Kyushu 9,321 57,229 1,480 27,448
Mizushima 543 6,946 106 3,107
Okinawa 1,290 8,134 477 4,018
Ooita 500 7,837 85 3,351
Sakai 150 1,530 63 594
Sakata 700 9,468 86 4,229
Shikoku 3,171 29,611 554 13,113
Sumitomo 463 2,456 97 990
Tobata 781 10,446 104 4,488
Tohoku 7,868 73,705 107 34,228
Tokya 29,254 2,80,478 3,751 1,27,538
Tomakomai 250 1,817 55 790
Toyama 500 6,192 84 2,703
Wakayama 306 4,541 112 1,921

7. The performance of some selected international airlines is given

in the following table. Operating costs, nonflight assets and
nonpassenger revenue are measured in appropriate money units
(US$).
Write all the eight DEA programs (Envelopment: input
oriented and output oriented, Multiplier: output maximizing
and input minimizing, and Returns to scale: constant and
variable) for estimating the efficiency of Airline 4.
Solve these programs using an LP software. Tabulate the
results. Is Airline 4 efficient? If not, which airlines would
you recommend that Airline 4 consider emulating to improve
the efficiency of its operation? What are the CRS, VRS and
 Economies of Scale 89

A data set for selected international airlines

Inputs Outputs
Airline Available Operating Non- Revenue Non-
(DMU) ton km Cost flight Passenger passenger
Assets (km) Revenue
1 10,884 6,730 3,934 26,677 7,688
2 4,603 3,457 2,360 22,112 969
3 12,097 6,779 6,474 52,363 2,001
4 6,587 3,341 3,581 26,504 1,297
5 5,723 3,239 2,003 26,677 697
6 24,099 9,560 6,267 1,24,055 1,266
7 22,793 9,874 4,145 1,22,528 1,404
8 19,080 8,032 3,272 95,011 572
9 5,654 1,878 1,916 19,277 972
10 12,559 8,098 3,310 41,925 3,398
11 5,728 2,481 2,254 27,754 982
12 4,715 1,792 2,485 31,332 543
13 13,565 7,499 3,213 64,734 1,563
14 5,183 1,880 783 23,604 513
15 5,895 4,225 4,557 33,081 539

scale efficiencies of this airline? What is its most productive

scale size?
Which airlines are efficient according to this data set? For
each, prepare a table showing the inefficient airlines for which
the efficient airline is a peer.
8. The data, choice of inputs and choice of output in this exercise
are based on the article, Shafer and Byrd (2000). The efficiency
of organizational investments in information technology firms
need to be studied. The three inputs related to investments in
IT exercise are: information system (IS) budget as a percentage
of sales, an organization’s total processor value as a percentage
of sales, and the percentage of the IS budget allocated to train-
ing. Because of a wide range of companies (in terms of sales,
profits, number of employees, and industry) all three inputs
were normalized to facilitate comparisons. Specifically, IS
budget and an organization’s total processor value were taken
as a percentage of sales to normalize the differences in company
size. Similarly, training expenditures were taken as a percentage
of the total IS budget. Normalized measures are used to repre-
sent two outputs. Data on inputs and outputs for 36 of the
209 companies considered in the article are shown in the table
below.