Applied Mathematics and Computation 160 (2005) 363–378

www.elsevier.com/locate/amc

A mathematical model for dynamic

efficiency using data envelopment analysis
a,* b
Ali Emrouznejad , Emmanuel Thanassoulis
a
Statistics and Operational Research Group, Priority Street, Coventry University, Coventry,
CV1 5FB, UK
b
Aston Business School, University of Aston, Aston Triangle, Birmingham B4 7ET, UK

Abstract
In this paper we propose a data envelopment analysis (DEA) based method for
assessing the comparative efficiencies of units operating production processes where
input–output levels are inter-temporally dependent. One cause of inter-temporal
dependence between input and output levels is capital stock which influences output
levels over many production periods. Such units cannot be assessed by traditional or
ÔstaticÕ DEA which assumes input–output correspondences are contemporaneous in the
sense that the output levels observed in a time period are the product solely of the input
levels observed during that same period. The method developed in the paper overcomes
the problem of inter-temporal input–output dependence by using input–output ÔpathsÕ
mapped out by operating units over time as the basis of assessing them. As an appli-
cation we compare the results of the dynamic and static model for a set of UK uni-
versities. The paper is suggested that dynamic model capture the efficiency better than
static model.

2003 Elsevier Inc. All rights reserved.

Keywords: Data envelopment analysis; Dynamic efficiency; Performance measurement; Higher

education

*
Corresponding author.
E-mail addresses: a.emrouznejad@coventry.ac.uk (A. Emrouznejad), e.thanassoulis@
aston.ac.uk (E. Thanassoulis).
URL: http://www.deazone.com.

0096-3003/$ - see front matter
2003 Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2003.09.026
364 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378

1. Introduction

This paper addresses the issue of assessing the comparative performance of

decision making units (DMUs) whose operations have not ceased at the time of
assessment and where output levels over a given period of time depend at least
in part on prior resources. A typical case of this kind is that where DMUs
secure their outputs using resources which include capital stock. Such stock,
which may occasionally be upgraded, affects output levels over a continuous
time interval which may span several assessment periods. In such cases tradi-
tional or ÔstaticÕ approaches to assessing performance break down because they
implicitly assume that there is correspondence between coincident input–output
levels. We distinguish between correspondence and coincidence of input–out-
put levels as follows:

• coincident input–output levels are those observed during the same time
period;
• corresponding input–output levels are those where the output levels are
caused exclusively by the input levels

where correspondence of coincident input–output levels does not hold we have

inter-temporal input–output dependence. This paper develops a method for
assessing the comparative performance of DMUs where we have panel data
and inter-temporal input–output dependence.
The paper is layed out as follows. Section 2 contains a taxonomy of
inter-temporal input–output dependencies and a discussion of their causes.
Section 3 develops the concept of an input–output path for capturing input–
output correspondences in the presence of inter-temporal effects. Sections 4
and 5 define a production possibility set (PPS) of input–output paths. Sec-
tion 6 develops a model for assessing the efficiency of an input–output
path and Section 7 illustrates the method developed, using a case of effi-
ciency measurement of higher education. Section 8 concludes the
paper.

2. A classification of multi-period production process

Depending on the duration of the life of operating units and the nature of
any inter-temporal dependence of input–output levels we can discern three
types of production process:

• single period;
• multi-period without inter-temporal input–output dependence;
• multi-period with inter-temporal input–output dependence.
 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378 365

2.1. Single period production processes

In such production processes clearly the issue of inter-temporal dependence

of input–output levels does not arise. We have correspondence between the
coincident input and output levels of each DMU and efficiency can be assessed
by the DEA models developed by Charnes et al. [6] or their extensions ([5] and
Charnes et al. (1999), see also [12]).

2.2. Multi-period production processes without inter-temporal input–output

dependence

In these production processes we continue to have contemporaneous cor-

respondence of input–output levels. Thus we can again use the DEA models
developed by Charnes et al. [6] or their extensions to assess the DMUs con-
cerned. However, DMUs are now in existence over several time-periods and
issues arise as to their performance over time rather than just at each specific
point in time. In essence, in multi-period production processes we can assess
performance in two contexts: cross-sectionally and diachronically.
Let us consider the case where a set of DMUs have been observed in time
periods t ¼ 1; 2; . . . ; T . The cross-sectional performance of a DMU relates to a
single time period t and it is assessed relative to best observed practice in that
time period. Cross-sectional efficiencies for each time period t can be estimated
using the standard DEA models.
Cross-sectional efficiency offers a snap-shot of the performance of a unit in
the time period concerned. It fails to identify the progress or regress over time
either of the efficient boundary itself or of a given operating unit. This point is
illustrated in Fig. 1.
Fig. 1 shows the efficient boundary Lt and Ltþ1 for producing a unit of
output in periods t and t þ 1 respectively. Two inputs are used in the pro-
duction process. Let us assume a production unit operates at A in time period t
and at B in time period t þ 1. Clearly the unit is more efficient in time period
t þ 1 in that it secures a unit of output using much lower input levels. Nev-
ertheless, the cross-sectional efficiency of the unit in time period t is OD/OA
and in time period t þ 1 OC/OB. Since OC/OB < OD/OA the cross-sectional
efficiencies of the unit convey the incorrect impression that its performance
deteriorates over time.
The problem is addressed using diachronic performance measurement. One
approach frequently used for measuring productivity change over time is that
developed by F€ are et al. [9] using a Malmquist index. F€are et al. [9] decompose
the total productivity change of a unit into that attributable to the ÔshiftÕ in
the efficient boundary between period t and t þ 1 and that attributable to the
Ôcatch-upÕ of the unitÕs efficiency. The catch-up factor reflects the change in the
cross-sectional efficiency of an operating unit as we move from time period t to
366 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378

Input2

A
D
B
Lt
C

Lt+1

Input1

Fig. 1. Cross-sectional efficiency does not reflect diachronic productivity changes.

time period t þ 1. The boundary shift term reflects the movement in the effi-
cient boundary from time period t to time period t þ 1 in terms of how much
more (less) input is needed to secure a given level of output, under efficient
operation. For more details of the approach (see [9]). Alternative approaches
for measuring progress and regress of production over time have been devel-
oped by Tulkens and Eeckaut [13].

2.3. Multi-period production processes with inter-temporal input–output depen-

dence

This is the case examined in this paper. The production units operate over a
continuing sequence of time periods and we do not have correspondence of
coincident input–output levels. Three particular causes of inter-temporal
input–output dependence are those of capital stock, lagged output and capital
output. These causes are elaborated below.

2.3.1. Capital stock

Capital stock, such as robots in car plants, enhance productivity. The
productive life of capital stock spans in general many time units such as years
or quarters typically used for recording coincident input–output data. Inter-
temporal dependence of input–output levels is caused by changes in the level of
capital stock, such as those due to capital investments. Asset acquisition does
not generally lead to an instantaneous rise in productivity and may indeed
 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378 367

initially lead to its drop. This is because of the ÔadjustmentÕ and ÔdisruptionÕ
processes generally associated with asset acquisition. The adjustment process is
typically referred to as the Ôlearning curveÕ as units need to learn how to use
new assets acquired [11]. Asset acquisition can also entail disruption due to the
need to integrate the new with existing assets. The duration and timing of the
adjustment and disruption effects will generally differ from DMU to DMU
depending on their asset acquisition activities.

2.3.2. Lagged output

In some production situations output can lag input in a way which makes it
difficult to establish correspondence between input and output. One case in
point is that of promotion of sales. Consider for example sales teams pro-
moting financial products such as personal insurance, pension plans etc. Over
some given assessment period a team may use the bulk of its time to make a
wide range of introductory contacts with potential clients, hold explanatory
workshops on the products for sale etc. Actual recorded sales may be low
during such a period. However, the team may have been successful in building
up goodwill among potential clients which will manifest itself in increased sales
over future periods. Thus in essence there is a lag between sales effort and
actual sales. Such a lag may span several assessment periods which makes it
difficult to establish correspondence between input (time devoted to promoting
sales) and output (sales achieved) within a given assessment period.

2.3.3. Capital output

In certain production contexts it is possible for intermediate or capital output
to be created which is not directly measurable but can enhance productivity in
subsequent periods. An example of intermediate or capital output is that of
research. Typically research output is measured by the number of research
papers or reports published, research grants obtained and so on. (We are
ignoring here the important but difficult issue of the quality of the research
output.) A research team may generate intermediate output in the form of
research ideas and provisional research results which are incomplete for pub-
lication. Such intermediate output is in effect Ôwork––in progressÕ and cannot
be captured by the usual research output measures. Yet it may have important
implications for a teamÕs productivity in subsequent periods.
Capital output whose generation and/or impact spans several assessment
periods distorts the correspondence of input–output levels within any given
assessment period. Inter-temporal dependence of input–output levels was ad-
dressed by Sengupta [11]. However, Sengupta addressed only the limited case
of the two period production process with capital inputs which have inter-
temporal effects on output. The approach developed in this paper differs from
that of Sengupta and considers inter-temporal input–output dependencies in
the general multi-period case.
368 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378

3. Capturing inter-temporal input–output correspondence using input–output

paths

Correspondence of coincident input–output levels is at the heart of the def-

inition of production possibility sets (PPS) used to assess the comparative
performance of DMUs. See for example [2, p. 1081;13, p. 475] for the definition
of the PPS in the DEA and free disposal hull (FDH) approaches to measuring
efficiency respectively. Input–output correspondences are fundamental to
performance measurement since what we measure is useful output secured
against the resources (inputs) used for its procurement. This fundamental
requirement of input–output correspondence does not alter in the presence of
inter-temporal input–output dependence. Thus, we need a method of capturing
inter-temporal input–output correspondence.
We can attempt to capture inter-temporal input–output correspondences
through the use of paths of coincident input-output levels as follows. Consider
a DMU j which came into existence n þ T time periods ago, it has been in
existence up to the current i.e. the ðn þ T Þth period and it is expected to con-
tinue in existence after the end of the current period. Let us further assume that
we have observed input–output coincidences ðxtj ; yjt Þ where xtj ¼ ðxt1j ; xt2j ; . . . ; xtmj Þ
are the input and yjt ¼ ðy1j t
; y2jt
; . . . ; ysjt Þ the output levels observed in time period
t at DMU j. Finally let the final T periods ending up at the current i.e.
ðn þ T Þth period be referred to as the assessment window.
We define the sequence ðxtj ; yjt Þ t ¼ 1; . . . ; T as the assessment path of DMU j
and denote it ðxj1;2;...;T ; yj1;2;...;T Þ.
The concept of the assessment path of DMU j is illustrated graphically in
Fig. 2.
In the case of a stock input (e.g. capital) the levels within the path reflect its
variation over time such as might be caused by occasional investment activity.
The shorter the periods into which the assessment window is subdivided the
more accurate the reflection of the underlying continuous variable of stock
input. In the case of a flow input (e.g. recurrent operating expenditure) the
levels within the path reflect the resource used up during each period in the
procurement of outputs.

t =1 t = n t = n+1 t = n +T
.... ....

Assessment window t = n + 1 to t = n + T

Fig. 2. The assessment path of DMU, j is the sequence of its input–output.

 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378 369

Let us now consider an assessment window covering the entire life of the
DMU. The assessment path can be said to capture the input–output corre-
spondence represented by the DMU. This is because all inputs used by the
DMU are reflected in the assessment path as are the corresponding outputs
procured, irrespective of the time lag between inputs and corresponding outputs.
The concept of a path covering the entire life of a DMU is useful for seeing
how input–output correspondences under inter-temporal effects can be cap-
tured in a path. However, a path covering the entire life of a DMU is not very
practical. In most situations the DMUs are expected to continue in existence
long into the future and what management usually wants is to measure per-
formance over a ÔsensibleÕ length of time leading up to the present. In view of
this we need to restrict our attention to assessment paths which cover a part of
the life of a DMU. The path covering the last T periods of a DMUÕs life, e.g.
from t ¼ n þ 1 to t ¼ n þ T in Fig. 2, is the type of assessment path which can
be constructed in practice. We shall focus in the remainder of this paper on
paths of this type.
It is evident that the longer an assessment path is the better it will reflect the
input–output correspondences mapped out by a DMU. This because most
lagged and capital output effects are likely to relate to inputs within the path
while any adjustment periods will represent a short proportion of the time
covered by the window. The length of the assessment window to be used is a
matter of judgement formed in the light of output lag, adjustment periods and
capital output effects likely to apply to the situation modelled. How many
assessment periods are used within the assessment window is also an issue
which needs to be addressed and to which we shall return after presenting the
assessment method to be used.

4. Defining a production possibility set using assessment paths

Let ðxj1;2;...;T ; yj1;2;...;T Þ denote the assessment path of DMU j ¼ 1; . . . ; N . The

production possibility set PPS is denoted P and is defined as follows:

P ¼ fðx1;2;...;T ; y 1;2;...;T Þjx1;2;...;T can produce y 1;2;...;T g:

Following the construction of the PPS in DEA elsewhere, e.g. [2, p. 1081] apart
from convexity, we will assume P has the following properties:

(i) P is non-empty. All observed paths fðxj1;2;...;T ; yj1;2;...;T Þ, j ¼ 1; 2; . . . ; N g 2 P .

(ii) Strong disposability of input. If ðxj1;2;...;T ; yj1;2;...;T Þ 2 P and x1;2;...;T P xj1;2;...;T ,
then ðx1;2;...;T ; yj1;2;...;T Þ 2 P where x1;2;...;T P xj1;2;...;T means xt P xtj for
t ¼ 1; 2; . . . ; T , and xt P xtj means that at least one element of xt is greater
than the corresponding element of xtj .
370 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378

(iii) Strong disposability of output. If ðxj1;2;...;T ; yj1;2;...;T Þ 2 P and y1;2;...;T 6 yj1;2;...;T ,

then ðxj1;2;...;T ; y1;2;...;T Þ 2 P .
(iv) No output can be produced without some input (No free lunch).
ðxj1;2;...;T ; 0Þ 2 P ; but if yj1;2;...;T P 0 then ð0; yj1;2;...;T Þ 62 P .
(v) Constant returns to scale. If ðxj1;2;...;T ; yj1;2;...;T Þ 2 P then for each positive real
value k > 0 we have ðkxj1;2;...;T ; kyj1;2;...;T Þ 2 P .
(vi) Minimum extrapolation. P is the smallest closed and convex set satisfying
(i)–(v).

A PPS P which satisfies the above postulates can be constructed from the
observed assessment paths ðxj1;2;...;T ; yj1;2;...;T Þ, j ¼ 1; . . . ; N as follows:
(
X 1;2;...;T
P ¼ ðx1;2;...;T ; y 1;2;...;T Þjx1;2;...;T P kj x j ; y 1;2;...;T
j
)
X
6 kj yj1;2;...;T ; kj > 0 : ð1Þ
j

5. Capturing initial and terminal stock of capital within the PPS

There is, however, a further aspect, which is important from a capital theory
viewpoint. As noted earlier, capital is viewed here as stock. Once, a capital input
is implemented, it produces a flow of outputs in future periods (see Fig. 3).
It is clear from this figure that there is lagged production of output from
changes in capital taking place at some point in time. So long as lagged outputs
are within the assessment window used we are not concerned about their
timing. However, where lagged output due to changes in capital stock made
within the assessment window, falls outside of it, and also where output within
the assessment window is the result of changes in capital prior to the assess-
ment window, then lagged output of this kind needs to be reflected in the
dynamic assessment. Thus one distinction between static and dynamic PPS is
that the definition reflects initial and terminal conditions of capital stock.
Whereas, the static PPS does not require these two additional conditions.
Now focus upon DMU-paths from the point of view of terminal stock in
each assessment path. To clarify this issue, consider two feasible paths P and

Periods
Period of use
Period of
installation

Fig. 3. The flow of output from capital.

 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378 371

P 0 , both of a finite duration and length t ¼ 1; . . . ; s. Assume that they start with
the same level of capital stock in the first period and they provide identical
output streams y 1;...;s ¼ y 01;...;s but that the terminal capital stocks differ, with
K s >6¼ K 0s . Thus path P provides more terminal capital stock than P 0 which
can contribute to future outputs. Clearly, then, this capability of path P should
be reflected in its assessment. However

• if period s is literally the end of the life of the DMU-path, then terminal cap-
ital stock of path P cannot be used to produce output in future and can be
ignored;
• on the other hand if DMU-path P survives after period s then having more
of terminal capital stock than DMU P 0 will enable higher future output at
DMU-path P .

Let us assume that stock of capital input at period s can be used for producing
output in future. To take into account this assumption in the PPS, terminal
capital stock must be treated as another output.
Similar discussion can be made for initial capital stock in the assessment
window. If, in time horizon t ¼ 1; 2; . . . ; s, the initial capital stocks of DMUs at
t ¼ 0 are not identical the PPS should take into account the difference between
those DMUs which start with a large and those which start with a small
quantity of capital stock at the beginning of the process under consideration.
Initial capital stock should be reflected in the PPS as another input, as it can be
converted to output within the assessment period.
Let us consider a window of periods t ¼ s; s þ 1; . . . ; s þ T . Assume that the
set of inputs, I ¼ f1; . . . ; mg, can be divided into two sub-sets of period-specific
inputs and capital-inputs, respectively I1 and I2 such that

I1 & I2
I; I1 [ I2 ¼ I and I1 \ I2 ¼ £:

Then the set of inputs is:

period-specific input paths: xs;sþ1;...;sþT ;

changes in stock input paths: zs;sþ1;...;sþT ;
Initial-stock inputs: Z s1 .

The set of outputs is:

output-paths: y s;sþ1;...;sþT ;
terminal-stock inputs as outputs: Z sþT :

For example in case of capital the changes in stock inputs will be reflected by
investment.
372 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378

This raises the issue of how to estimate the level of initial and terminal stock
of inputs. The details of how to estimate such values are not directly addressed
in this paper. However one possibility is to reflect stock input by means of
converting it to a capital value which takes into account the age and productive
capabilities of the stock. Depreciation is of use as a means of reflecting in
monetary terms the age of stock of capital.
The PPS within the assessment window s; . . . ; s þ T can be now stated as
follows:


P ¼ ðxs;...;sþT ; zs;...;sþT ; y s;...;sþT Þj:
X
xti P kj xtij ; 8t ¼ s; . . . ; s þ T and i 2 I1
j
X
zti P kj ztij ; 8t ¼ s; . . . ; s þ T and i 2 I2
j
X
yt 6 kj yjt ; 8t ¼ s; . . . ; s þ T
j
X
Zis1 P kj Zijs1 ; 8i 2 I2
j

X 
ZisþT 6 kj ZijsþT ; 8i 2 I2 kj 2 Rþ 8j : ð2Þ
j

Note that if it is assumed that there is only one period, then this PPS will
collapse to the static PPS as described in Banker et al. [2].
The next section uses P as defined in (2) to measure Ôthe relative efficiencyÕ of
the assessment path of a DMU.

6. Measuring the comparative efficiency of an assessment path

We begin by extending the definition of Pareto efficiency to assessment

paths. Drawing from Charnes et al. [6, p. 433], Abel et al. [1] and Burmeister [4]
we define a Pareto efficient path as follows:
The assessment path ðxs;...;sþT
j0
s;...;sþT
; yj0 Þ of DMU j0 is said to be Pareto effi-
cient if no other assessment path or combination of assessment paths yields
more than the path of DMU j0 on at least one output in at least one period t,
t ¼ s; . . . ; s þ T , without producing less in some other output in some period or
requiring more of at least one input in at least one period.
We shall refer to an assessment path which is Pareto efficient in the fore-
going sense as a dynamically efficient path. The following linear programming
s;...;sþT s;...;sþT
model can be used to ascertain whether the assessment path ðxj0 ; yj0 Þ of
DMU j0 is dynamically efficient.
 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378 373

MODEL M1
PsþT X
sþT X sþT X
X sþT X
X s
at
Min a¼ t¼s
e Sit þ dt
i þ Srtþ
T t¼s i2I1 t¼s i2I2 t¼s r¼1
!
X X
þ c
i þ cþ
i ;
i2I2 i2I2
X
N
s:t: C1 : kj xtij ¼ at xtij0  Sit ; i 2 I1 ; t ¼ s; . . . ; s þ T ;
j¼1
X
N
C2 : kj ztij ¼ at ztij0  dt
i ; i 2 I2 ; t ¼ s; . . . ; s þ T ;
j¼1
X
N
C3 : kj yrjt ¼ yrjt 0 þ Srtþ ; r ¼ 1; . . . ; s; t ¼ s; . . . ; s þ T ;
j¼1
X
N
C4 : kj ZijsþT ¼ ZijsþT
0
þ cþ
i ; i 2 I2 ;
j
X
N
C5 : kj Zijs1 ¼ Zijs1
0
 c
i ; i 2 I2 ;
j¼1

kj P 0; 8j; Sit P 0; dt i P 0 ð8t; 8i 2 I1 Þ;

Srtþ P 0 ð8r; 8tÞ; cþ
i P 0; c
i P 0 ð8i 2 I2 Þ;

where I1  f1; . . . ; mg are flow inputs, I2  f1; . . . ; mg are those inputs that
their end-stock will be converted, directly or indirectly, into more output some
type at some future period. Zijs1 is the initial-stock of capital of type i for DMU
j; i 2 I2 , ZijsþT is the end-stock capital of type i for DMU j; i 2 I2 .
Model (M1) modifies static DEA essentially by assessing the DMU over
horizon simultaneously and by adding constraint sets C4 and C5. Thus the
model measures the extent to which inputs, both flow and stock, can be re-
duced further, given the initial and terminal stock input of the unit and given
its output levels during the assessment periods.
It is important to notice that this model distinguishes between investment
and capital stock. The model will treat Zjs1 and any investment within the
assessment periods as input but ZjsþT as output since end-capital stock is
capable to produce further output in future. In this view a high level of capital
stock at the end would not be seen as inefficiency, The model will balance
between investment and capital stock.
In this model, constraints sets C1–C5 as follows:

)C1 are period-specific input constraints,

)C2 are stock-change input constraints,
374 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378

)C3 are output constraints and,

)C4 are end-stock constraints,
)C5 are initial-stock constraints.

Set C4 treats terminal capital stock as an output and that is why constraint
sets C3 and C4 are essentially the same. Constraint C5 treats initial stock of
capital as an exogenously fixed input. Thus the model measures the extent to
which inputs, both flow and stock, can be reduced further, given the initial and
terminal stock input of the unit and given its output levels during the assess-
ment window.
An optimal solution to model (M1) specifies a production point
ðxti ; i ¼ 1 . . . m; yrt ; r ¼ 1 . . . s; t ¼ s; . . . ; s þ T Þ within the PPS, where

X
N
xti ¼ kj xtij ¼ /t xtij0  Sit i 2 I1 ; t ¼ s; . . . ; s þ T ;
j

X
N
zti ¼ kj ztij ¼ /t ztij0  dt
i i 2 I2 ; t ¼ s; . . . ; s þ T ; ð3Þ
j

X
N
yrt ¼ kj yijt ¼ yijt 0 þ Srtþ r ¼ 1; . . . ; s; t ¼ s; . . . ; s þ T :
j

The superscript * denotes the optimal value of the corresponding variable in

model (M1).
The assessment path specified in (3) is dynamically efficient in line with our
earlier definition. By virtue of model (M1) it is the case that there exists no
point within the PPS which offers a reduction in one of the input levels without
either a consequent rise in some other input level or a reduction in the level of
at least one of the outputs in some time period t. By implication,when at the
optimal solution to model (M1) we have a ¼ 1, Sit ¼ 08i; t, dt
i ¼ 0 8i; t and
Srtþ ¼ 0 8r; t the assessment path of DMU j0 is dynamically efficient.
Where the assessment path of DMU j0 is not dynamically efficient, the value
of a0 can be seen as a measure of its dynamic (input) efficiency. Specifically, a0
measures the average proportion to which the observed input levels of DMU j0
can be lowered without detriment to any one of its output levels in any time
period while maintaining its input mix in each period of the assessment
window.
Each component a0 measures the extent to which the input levels in period t
can be lowered radially under efficient production. The cross-sectionally radial
measure of efficiency used is consistent with the basic notion in DEA of not
imposing a prior value system over input–output levels within each time
period.
 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378 375

The measure of efficiency yielded by model (M1) will not alter if periods of
the assessment window are aggregated or subdivided, provided the input–
output levels in the new periods are obtainable by simply scaling the input–
output levels of the original periods. This can be readily seen by noting that
such scaling of input–output levels merely generates redundant constraints
within model (M1) in going from the original to the new periods within the
assessment window. In the more general case, however, where the subdivision
or aggregation of the original periods does not preserve the mix of the original
input–output levels the efficiency measure a0 will be assessment window sub-
division variant. This is as should be since what we aim to do is to assess
DMUs by charting their resource use and output creation over time during the
assessment window. Where the pattern of resource use and output creation
differs over time then so do the underlying efficiencies. An important question
is then which assessment window subdivision yields the more reliable efficiency
measure? The answer is the sub-division with the largest number of assessment
periods. The larger the number of assessment periods used the more accurate
the reflection of the true path of input–output levels mapped out by each
DMU. Obviously a balance has to be struct between reflecting accurately the
input–output path of each DMU and the number of assessment periods used.
The assessment paths corresponding to positive k values at the optimal
solution to model (M1) will be referred to as the referent assessment paths or
efficient peer paths of DMU j0 . The dynamic efficiency rating of the assessment
path of DMU j0 is with reference to these input–output paths. Note that when
T ¼ 1 model (M1) collapses to model (9) in Charnes et al. [6, p. 433] used to
measure efficiency in the single production period context.
The next section illustrates the assessment of the dynamic efficiencies of 15
UK universities and contrasts the results obtained with that would be obtained
if a ÔstaticÕ single period assessment model was used in each time period.

7. An application of dynamic DEA in higher education

In this section we demonstrate a dynamic DEA model that could be used for
evaluating efficiency in higher education.

7.1. Input output variables

The determination of input output variables is difficult in an educational

organisation and in particular in university assessments. The main products of
a university are its teaching and research outcomes. Therefore in order to assess
universities on their responsibility of delivering knowledge it is necessary to
identify input output variables pertaining to this function.
376 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378

The inputs should represent all the resources used and the outputs the
corresponding activity levels of the research and teaching as main objectives of
the universities.Therefore the inputs that we selected are:

X ¼ Total recurrent grants.

C ¼ Total capital grants.

On the output side we included.

Y1 ¼ Income from research grants and contracts.

Y2 ¼ Number of PhDs awarded.
Y3 ¼ Number of other postgraduate degrees awarded, not including PhDs.
Y4 ¼ Number of undergraduate degrees awarded.

The assessment periods we are examining in this section are the academic
years from 1994 to 1998. The data are derived from the publication of Higher
Education Statistics Agency (see [10]). Fifteen Institutions are included in the
analysis. Following Beasley [3] we decided to impose the weight restriction for
number of qualification awarded. The suggestion is that the weight associated
with a PhD is at least 25% greater than the weight associated with a taught
postgraduate and a weight associated with a taught postgraduate is at least
25% greater than the weight associated with an undergraduate student.
However user can set up different weight restriction and run the model again.
Table 1 shows the average dynamic efficiency scores obtained and static
DEA model.

Table 1
Dynamic and static efficiency
University Average static DEA over three Dynamic efficiency
academic years 1995 to 1998 score
Cranfield University 100 100
Imperial College 74 83
Keele University 100 100
London Business School 100 100
London School of Economics 82 100
The London Institute 100 100
University College London 100 100
University of Bath 50 53
University of Birmingham 70 62
University of Cambridge 90 86
University of Durham 56 53
University of London 91 100
University of Oxford 93 77
University of Warwick 73 72
University of York 57 62
 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378 377

Despite the overall agreement between the static and dynamic DEA the
two approaches disagree substantially in some institutions. The main reason
that dynamic DEA gives different scores to these institutions is that, firstly,
dynamic DEA assesses the institutions by examining them over three aca-
demic years simultaneously and secondly, the variation of capital input af-
fects much more the dynamic than the static efficiency. We, would argue that
the two approaches complement each other rather to replace one another.
Each gives a different insight to the efficiency of organisations like univer-
sities.

8. Conclusions

This paper has developed a method for assessing the performance of DMUs
when their input–output levels are characterised by inter-temporal dependen-
cies. Such dependencies typically arise when capital stock is used in production
and its level varies over time. Inter-temporal dependencies of input–output
levels can, however, also be due to other causes such as lags in output or the
creation of capital or latent output (e.g. generation of product awareness or
goodwill within a client base).
Static, single period assessment fails in the presence of inter-temporal input–
output dependencies because we do not have (causal) correspondence between
coincident input–output levels, a fundamental assumption in DEA models
developed todate. The model developed in this paper overcomes the problem of
lack of correspondence between coincident input–output levels by using
dependencies within the path of input–output levels generated by a DMU. Such
paths cannot capture any impact input–output levels inside the assessment
window have outside. However, the length of the window used can be such that
the paths give a good approximation to the correspondence of input–output
levels generated by each DMU as the vast majority of inter-temporal effects
will be reflected within the assessment window.
The model developed follows closely the basic DEA model introduced by
Charnes et al. [6] for assessing the comparative performance of DMUs. The
novelty in the model developed is the use of input–output paths in place of
the single period input–output levels. An application of efficiency measure-
ment in higher education has demonstrated how static efficiencies can be very
volatile in the presence of inter-temporal effects, while the dynamic model
captures better such effects. Further research can improve the model devel-
oped. One aspect in need of investigation is how to decide on the length of
the assessment window to use and on its subdivision into periods in order to
get as accurate a measure of the performance of the units being assessed as
possible.
378 A. Emrouznejad, E. Thanassoulis / Appl. Math. Comput. 160 (2005) 363–378

References

[1] A.B. Abel, N.G. Mankiw, L.H. Summer, R.J. Zeckhauser, Assessing dynamic efficiency:
theory and evidence, Review of Economics Studies 56 (1989) 1–20.
[2] R.D. Banker, A. Charnes, W.W. Cooper, Some models for estimating technical and scale
inefficiency in data envelopment analysis, Management Science 30 (9) (1984) 1078–1092.
[3] J.E. Beasley, Comparing university departments, OMEGA International Journal of Manage-
ment Science 18 (2) (1989) 171–183.
[4] E. Burmeister, Capital Theory and Dynamics, Cambridge University Press, Cambridge, 1980.
[5] A. Charnes, W.W. Cooper, Y. Arie, M.L. Seiford (Eds.), Data Envelopment Analysis: Theory,
Methodology and Application, Kluwer Academic Publishers, Boston, Dordrecht, London,
1994.
[6] A. Charnes, W.W. Cooper, E. Rhodes, Measuring the efficiency of decision making units,
European Journal of Operational Research 2 (6) (1978) 429–444.
[9] R. F€are, Sh. Grosskopf, B. Lindgren, P. Roos, Productivity changes in swedish pharmacies
1980–1989: a non-parametric malmquist approach, Journal of Productivity Analysis 3 (1992)
85–101.
[10] HESA Publications, 1994–1999. Resources of Higher Education Institutions, Available from
<www.hesa.ac.uk>.
[11] K. Sengupta, Dynamics of Data Envelopment Analysis: Theory of System Efficiency, Kluwer
Academic Publishers, Dordrecht, London, 1995.
[12] E. Thanassoulis, Introduction to the Theory and Application of Data Envelopment Analysis:
A Foundation Text with Integrated Software, Kluwer Academic Publishers, Boston, 2001.
[13] H. Tulkens, P.V. Eeckant, Nonparametric efficiency: progress and regress measures for panel
data. Methodological aspects, European Journal of Operational Research 80 (3) (1995) 474–
499.