Data Envelopment Analysis

Data Envelopment Analysis (DEA) is an increasingly popular management tool. This write-up is
an introduction to Data Envelopment Analysis (DEA) for those who are unfamiliar with the
technique. For a more in-depth discussion of DEA the interested reader is referred to !eiford
and Thrall "#$$%& or the seminal wor' (y )harnes )ooper and *hodes "#$+,&.
DEA is commonly used to evaluate the efficiency of a num(er of producers. A typical statistical
approach is characteri-ed as a central tendency approach and it evaluates producers relative to an
average producer .n contrast DEA compares each producer with only the /(est/ producers. 0y
the way in the DEA literature a producer is usually referred to as a decision ma'ing unit or
D12. DEA is not always the right tool for a pro(lem (ut is appropriate in certain cases. (!ee
!trengths and 3imitations of DEA.)
.n DEA there are a num(er of producers. The production process for each producer is to ta'e a
set of inputs and produce a set of outputs. Each producer has a varying level of inputs and gives a
varying level of outputs. For instance consider a set of (an's. Each (an' has a certain num(er of
tellers a certain square footage of space and a certain num(er of managers (the inputs). There
are a num(er of measures of the output of a (an' including num(er of chec's cashed num(er of
loan applications processed and so on (the outputs). DEA attempts to determine which of the
(an's are most efficient and to point out specific inefficiencies of the other (an's.
A fundamental assumption (ehind this method is that if a given producer A is capa(le of
producing 4(A) units of output with 5(A) inputs then other producers should also (e a(le to do
the same if they were to operate efficiently. !imilarly if producer 0 is capa(le of producing 4(0)
units of output with 5(0) inputs then other producers should also (e capa(le of the same
production schedule. 6roducers A 0 and others can then (e com(ined to form a composite
producer with composite inputs and composite outputs. !ince this composite producer does not
necessarily e7ist it is typically called a virtual producer.
The heart of the analysis lies in finding the /(est/ virtual producer for each real producer. .f the
virtual producer is (etter than the original producer (y either ma'ing more output with the same
input or ma'ing the same output with less input then the original producer is inefficient. The
su(tleties of DEA are introduced in the various ways that producers A and 0 can (e scaled up or
down and com(ined.
Numerical Example
To illustrate how DEA wor's let8s ta'e an e7ample of three (an's. Each (an' has e7actly #%
tellers (the only input) and we measure a (an' (ased on two outputs9 )hec's cashed and 3oan
applications. The data for these (an's is as follows9
0an' A9 #% tellers #%%% chec's :% loan applications
0an' 09 #% tellers ;%% chec's <% loan applications
0an' )9 #% tellers :%% chec's #<% loan applications
=ow the 'ey to DEA is to determine whether we can create a virtual (an' that is (etter than one
or more of the real (an's. Any such dominated (an' will (e an inefficient (an'.
)onsider trying to create a virtual (an' that is (etter than 0an' A. !uch a (an' would use no
more inputs than A (#% tellers) and produce at least as much output (#%%% chec's and :% loans).
)learly no com(ination of (an's 0 and ) can possi(ly do that. 0an' A is therefore deemed to (e
efficient. 0an' ) is in the same situation.
>owever consider (an' 0. .f we ta'e half of 0an' A and com(ine it with half of 0an' ) then
we create a (an' that processes ?%% chec's and ,< loan applications with @ust #% tellers. This
dominates 0 (we would much rather to have the virtual (an' we created than (an' 0). 0an' 0 is
therefore inefficient.
Another way to see this is that we can scale down the inputs to 0 (the tellers) and still have at
least as much output. .f we assume that inputs are linearly scala(le then we estimate that we can
get (y with ?.A tellers. Be do that (y ta'ing .A; times (an' A plus .:$ times (an' 0. The result
uses ?.A tellers and produces at least as much as (an' 0 does. Be say that (an' 08s efficiency
rating is .?A. 0an's A and ) have an efficiency rating of #.
Graphical Example
The single input two-output or two input-one output pro(lems are easy to analy-e graphically.
The previous numerical e7ample is now solved graphically. (An assumption of constant returns
to scale is made and e7plained in detail later.) The analysis of the efficiency for (an' 0 loo's li'e
the following9
.f it is assumed that conve7 com(inations of (an's are allowed then the line segment connecting
(an's A and ) shows the possi(ilities of virtual outputs that can (e formed from these two (an's.
!imilar segments can (e drawn (etween A and 0 along with 0 and ). !ince the segment A) lies
(eyond the segments A0 and 0) this means that a conve7 com(ination of A and ) will create
the most outputs for a given set of inputs.
This line is called the efficiency frontier. The efficiency frontier defines the ma7imum
com(inations of outputs that can (e produced for a given set of inputs.
!ince (an' 0 lies (elow the efficiency frontier it is inefficient. .ts efficiency can (e determined
(y comparing it to a virtual (an' formed from (an' A and (an' ). The virtual player called C is
appro7imately <;D of (an' A and ;?D of (an' ). (This can (e determined (y an application of
the lever law. 6ull out a ruler and measure the lengths of AC )C and A). The percentage of
(an' ) is then ACEA) and the percentage of (an' A is )CEA).)
The efficiency of (an' 0 is then calculated (y finding the fraction of inputs that (an' C would
need to produce as many outputs as (an' 0. This is easily calculated (y loo'ing at the line from
the origin F to C. The efficiency of player 0 is F0EFC which is appro7imately ?AD. This
figure also shows that (an's A and ) are efficient since they lie on the efficiency frontier. .n
other words any virtual (an' formed for analy-ing (an's A and ) will lie on (an's A and )
respectively. Therefore since the efficiency is calculated as the ratio of FAEFC or FAEFC (an's
A and ) will have efficiency scores equal to #.%.
The graphical method is useful in this simple two dimensional e7ample (ut gets much harder in
higher dimensions. The normal method of evaluating the efficiency of (an' 0 is (y using a linear
programming formulation of DEA.
!ince this pro(lem uses a constant input value of #% for all of the (an's it avoids the
complications caused (y allowing different returns to scale. *eturns to scale refers to increasing
or decreasing efficiency (ased on si-e. For e7ample a manufacturer can achieve certain
economies of scale (y producing a thousand circuit (oards at a time rather than one at a time - it
might (e only #%% times as hard as producing one at a time. This is an e7ample of increasing
returns to scale (.*!.)
Fn the other hand the manufacturer might find it more than a trillion times as difficult to
produce a trillion circuit (oards at a time though (ecause of storage pro(lems and limits on the
worldwide copper supply. This range of production illustrates decreasing returns to scale (D*!.)
)om(ining the two e7treme ranges would necessitate varia(le returns to scale (C*!.)
)onstant *eturns to !cale ()*!) means that the producers are a(le to linearly scale the inputs
and outputs without increasing or decreasing efficiency. This is a significant assumption. The
assumption of )*! may (e valid over limited ranges (ut its use must (e @ustified. As an aside
)*! tends to lower the efficiency scores while C*! tends to raise efficiency scores.
Using Linear Programming
Data Envelopment Analysis is a linear programming procedure for a frontier analysis of inputs
and outputs. DEA assigns a score of # to a unit only when comparisons with other relevant units
do not provide evidence of inefficiency in the use of any input or output. DEA assigns an
efficiency score less than one to (relatively) inefficient units. A score less than one means that a
linear com(ination of other units from the sample could produce the same vector of outputs (y
using a smaller vector of inputs. The score reflects the radial distance from the estimated
production frontier to the D12 under consideration.
There are a num(er of equivalent formulations for DEA. The most direct formulation of the
e7ample a(ove is as follows9
3et (e the vector of inputs into D12 i. 3et (e the corresponding vector of outputs. 3et
(e the inputs into a D12 for which we want to determine its efficiency and (e the outputs. !o
the X8s and the Y8s are the data. The measure of efficiency for is given (y the following
linear program9
where is the weight given to D12 i in its efforts to dominate D12 % and is the efficiency of
D12 %. !o the 8s and are the varia(les. !ince D12 % appears on the left hand side of the
equations as well the optimal cannot possi(ly (e more than #. Bhen we solve this linear
program we get a num(er of things9
#. The efficiency of D12 % ( ) with meaning that the unit is efficient.
:. The unit8s GGcompara(les88 (those D12 with non-ero ).
A. The GGgoal88 inputs (the difference (etween and )
;. Alternatively we can 'eep inputs fi7ed and get goal outputs ( )
DEA assumes that the inputs and outputs have (een correctly identified. 2sually as the num(er
of inputs and outputs increase more D12s tend to get an efficiency rating of # as they (ecome
too speciali-ed to (e evaluated with respect to other units. Fn the other hand if there are too few
inputs and outputs more D12s tend to (e compara(le. .n any study it is important to focus on
correctly specifying inputs and outputs.
The linear programs for evaluating the A D12s are given (y9
LP for evaluating DMU 1:
min THETA
st
5L1+8L2+7L3 - 5THETA <= 0
14L1+15L2+12L3 - 14THETA <= 0
9L1+5L2+4L3 >= 9
4L1+7L2+9L3 >= 4
16L1+10L2+13L3 >= 16
L1, L2, L3 >= 0
LP for evaluating DMU :
min THETA
st
5L1+8L2+7L3 - 8THETA <= 0
14L1+15L2+12L3 - 15THETA <= 0
9L1+5L2+4L3 >= 5
4L1+7L2+9L3 >= 7
16L1+10L2+13L3 >= 10
L1, L2, L3 >= 0
LP for evaluating DMU !:
min THETA
st
5L1+8L2+7L3 - 7THETA <= 0
14L1+15L2+12L3 - 12THETA <= 0
9L1+5L2+4L3 >= 4
4L1+7L2+9L3 >= 9
16L1+10L2+13L3 >= 13
L1, L2, L3 >= 0
The solution to each of these is as follows9
DMU 1"
Adjusta!" #"!!s
$ina! %"du&"d 'j"&ti(" A!!)*a!" A!!)*a!"
#"!! +am" ,a!u" #)st #)"--i&i"n .n&/"as" 0"&/"as"
12110 t3"ta 1 0 1 1E+30 1
12111 L1 1 0 0 0492857142 04619047619
12112 L2 0 0424285714 0 1E+30 04242857143
12113 L3 0 0 0 0436710963 04412698413

#)nst/aints
$ina! 53ad)* #)nst/aint A!!)*a!" A!!)*a!"
#"!! +am" ,a!u" 6/i&" %4H4 5id" .n&/"as" 0"&/"as"
12116 .+1 -04103473 0 0 1E+30 0
12117 .+2 -04289724 -0407142857 0 0 1E+30
12118 '7T1 9 04085714286 9 0 0
12119 '7T2 4 04057142857 4 0 0
12120 '7T3 16 0 16 0 1E+30
DMU "
Adjusta!" #"!!s
$ina! %"du&"d 'j"&ti(" A!!)*a!" A!!)*a!"
#"!! +am" ,a!u" #)st #)"--i&i"n .n&/"as" 0"&/"as"
12110 t3"ta 04773333333 0 1 1E+30 1
12111 L1 04261538462 0 0 04866666667 04577777778
12112 L2 0 0422666667 0 1E+30 04226666667
12113 L3 04661538462 0 0 04342635659 04385185185

#)nst/aints
$ina! 53ad)* #)nst/aint A!!)*a!" A!!)*a!"
#"!! +am" ,a!u" 6/i&" %4H4 5id" .n&/"as" 0"&/"as"
12116 .+1 -0424820512 0 0 1E+30 04248205128
12117 .+2 -04452651 -040666667 0 0446538461 1E+30
12118 '7T1 5 0408 5 10475 04655826558
12119 '7T2 7 040533333 7 1405676855 3441509434
12120 '7T3 12478461538 0 10 2478461538 1E+30
DMU !"
Adjusta!" #"!!s
$ina! %"du&"d 'j"&ti(" A!!)*a!" A!!)*a!"
#"!! +am" ,a!u" #)st #)"--i&i"n .n&/"as" 0"&/"as"
12110 t3"ta 1 0 1 1E+30 1
12111 L1 0 0 0 1408333333 04722222222
12112 L2 0 04283333333 0 1E+30 04283333333
12113 L3 1 0 0 0442829457 04481481481

#)nst/aints
$ina! 53ad)* #)nst/aint A!!)*a!" A!!)*a!"
#"!! +am" ,a!u" 6/i&" %4H4 5id" .n&/"as" 0"&/"as"
12116 .+1 -04559375 0 0 1E+30 0
12117 .+2 -04741096 -0408333333 0 0 1E+30
12118 '7T1 4 041 4 16425 0
12119 '7T2 9 04066666667 9 0 0
12120 '7T3 13 0 13 0 1E+30
=ote that D12s # and A are overall efficient and D12 : is inefficient with an efficiency rating
of %.++AAAA.
>ence the efficient levels of inputs and outputs for D12 : are given (y9
Efficient levels of .nputs9
Efficient levels of Futputs9
=ote that the outputs are at least as much as the outputs currently produced (y D12 : and inputs
are at most as (ig as the %.++AAAA times the inputs of D12 :. This can (e used in two different
ways9 The inefficient D12 should target to cut down inputs to equal at most the efficient levels.
Alternatively an equivalent statement can (e made (y finding a set of efficient levels of inputs
and outputs (y dividing the levels o(tained (y the efficiency of D12 :. This focus can then (e
used to set targets primarily for outputs rather than reduction of inputs.
!A! E7ample.
Alternate #ormulation
There is another pro(a(ly more common formulation that provides the same information. Be
can thin' of DEA as providing a price on each of the inputs and a value for each of the outputs.
The efficiency of a D12 is simply the ratio of the inputs to the outputs and is constrained to (e
no more than #. The prices and values have nothing to do with real prices and values9 they are an
artificial construct. The goal is to find a set of prices and values that puts the target D12 in the
(est possi(le light. The goal then is to
>ere the varia(les are the u8s and the v8s. They are vectors of prices and values respectively.
This fractional program can (e equivalently stated as the following linear programming pro(lem
(where Y and X are matrices with columns and respectively).
Be denote this linear program (y (D). 3et us compare it with the one introduced earlier which
we denote (y (6)9
To fi7 ideas let us write out e7plicitely these two formulations for D12 : say in our e7ample.
#ormulation $P% for DMU :
min THETA
st
-5 L1 - 8 L2 - 7 L3 + 8 THETA >= 0
-14L1 -15 L2 -12 L3 + 15THETA >= 0
9 L1 + 5 L2 + 4 L3 >= 5
4 L1 + 7 L2 + 9 L3 >= 7
16L1 +10 L2 +13 L3 >= 10
L1>=0, L2>=0, L3>=0
#ormulation $D% for DMU :
ma8 5 71 + 7 72 + 10 73
st
- 5 ,1 - 14,2 + 9 71 + 4 72 + 16 73 <= 0
- 8 ,1 - 15,2 + 5 71 + 7 72 + 12 73 <= 0
- 7 ,1 - 12,2 + 4 71 + 9 72 + 13 73 <= 0
8 ,1 + 15,2 = 1
,1>=0, ,2>=0, 71>=0, 72>=0, 73>=0
Formulations (6) and (D) are dual linear programsH These two formulations actually give the
same information. 4ou can read the solution to one from the shadow prices of the other. Be will
not discuss linear programming duality in this course. 4ou can learn a(out it in some of the F*
electives.
Applications
The simple (an' e7ample descri(ed earlier may not convey the full view on the usefulness of
DEA. .t is most useful when a comparison is sought against /(est practices/ where the analyst
doesn8t want the frequency of poorly run operations to affect the analysis. DEA has (een applied
in many situations such as9 health care (hospitals doctors) education (schools universities)
(an's manufacturing (enchmar'ing management evaluation fast food restaurants and retail
stores.
The analy-ed data sets vary in si-e. !ome analysts wor' on pro(lems with as few as #< or :%
D12s while others are tac'ling pro(lems with over #%%%% D12s.
&trengths an' Limitations of DEA
As the earlier list of applications suggests DEA can (e a powerful tool when used wisely. A few
of the characteristics that ma'e it powerful are9
DEA can handle multiple input and multiple output models.
.t doesn8t require an assumption of a functional form relating inputs to outputs.
D12s are directly compared against a peer or com(ination of peers.
.nputs and outputs can have very different units. For e7ample 5# could (e in units of
lives saved and 5: could (e in units of dollars without requiring an a priori tradeoff
(etween the two.
The same characteristics that ma'e DEA a powerful tool can also create pro(lems. An analyst
should 'eep these limitations in mind when choosing whether or not to use DEA.
!ince DEA is an e7treme point technique noise (even symmetrical noise with -ero mean)
such as measurement error can cause significant pro(lems.
DEA is good at estimating /relative/ efficiency of a D12 (ut it converges very slowly to
/a(solute/ efficiency. .n other words it can tell you how well you are doing compared to
your peers (ut not compared to a /theoretical ma7imum./
!ince DEA is a nonparametric technique statistical hypothesis tests are difficult and are
the focus of ongoing research.
!ince a standard formulation of DEA creates a separate linear program for each D12
large pro(lems can (e computationally intensive.
(eferences
DEA has (ecome a popular su(@ect since it was first descri(ed in #$+,. There have (een
hundreds of papers and technical reports pu(lished along with a few (oo's. Technical articles
a(out DEA have (een pu(lished in a wide variety of places ma'ing it hard to find a good starting
point. >ere are a few suggestions as to starting points in the literature.
#. )harnes A. B.B. )ooper and E. *hodes. /1easuring the efficiency of decision ma'ing
units./ European Iournal of Fperations *esearch (#$+,)9 ;:$-;;.
:. 0an'er *.D. A. )harnes and B.B. )ooper. /!ome models for estimating technical and
scale inefficiencies in data envelopment analysis./ 1anagement !cience A% (#$,;)9 #%+,-
$:.
A. Dyson *.J. and E. Thanassoulis. /*educing weight fle7i(ility in data envelopment
analysis./ Iournal of the Fperational *esearch !ociety A$ (#$,,)9 <?A-+?.
;. !eiford 3.1. and *.1. Thrall. /*ecent developments in DEA9 the mathematical
programming approach to frontier analysis./ Iournal of Econometrics ;? (#$$%)9 +-A,.
<. Ali A... B.D. )oo' and 3.1. !eiford. /!trict vs. wea' ordinal relations for multipliers
in data envelopment analysis./ 1anagement !cience A+ (#$$#)9 +AA-,.
?. Andersen 6. and =.). 6etersen. /A procedure for ran'ing efficient units in data
envelopment analysis./ 1anagement !cience A$ (#$$A)9 #:?#-;.
+. 0an'er *.D. /1a7imum li'elihood consistency and data envelopment analysis9 a
statistical foundation./ 1anagement !cience A$ (#$$A)9 #:?<-+A.
The first paper was the original paper descri(ing DEA and results in the a((reviation ))* for
the (asic constant returns-to-scale model. The !eiford and Thrall paper is a good overview of the
literature. The other papers all introduce important new concepts. This list of references is
certainly incomplete.
A good source covering the field of productivity analysis is The 1easurement of 6roductive
Efficiency edited (y Fried 3ovell and !chmidt #$$A from F7ford 2niversity 6ress. There is
also a recent (oo' from Kluwer 6u(lishers Data Envelopment Analysis9 Theory 1ethodology
and Applications (y )harnes )ooper 3ewin and !eiford.
To stay more current on the topics some of the most important DEA articles appear in
1anagement !cience The Iournal of 6roductivity Analysis The Iournal of the Fperational
*esearch !ociety and The European Iournal of Fperational *esearch. The latter @ust pu(lished a
special issue /6roductivity Analysis9 6arametric and =on-6arametric Approaches/ edited (y
3ewin and 3ovell which has several important papers.
Volume 15, Issue 1, 2002
Estimation of technical efciency: a review of some of the
stochastic frontier and DEA software
In!s "errero
Department of Applied Economics, University of Huelva, and
#ean $ascoe
CEMARE, Department of Economics, University of Portsmouth
The level of technical efficiency of a particular firm is characterised (y the relationship (etween
o(served production and some ideal or potential production (Jreene #$$A). The measurement of
firm specific technical efficiency is (ased upon deviations of o(served output from the (est
production or efficient production frontier. .f a firm8s actual production point lies on the frontier
it is perfectly efficient. .f it lies (elow the frontier then it is technically inefficient with the ratio
of the actual to potential production defining the level of efficiency of the individual firm.
Farrell8s (#$<+) definition of technical efficiency led to the development of methods for
estimating the relative technical efficiencies of firms. The common feature of these estimation
techniques is that information is e7tracted from e7treme o(servations from a (ody of data to
determine the (est practice production frontier (3ewin and 3ovell #$$%). From this the relative
measure of technical efficiency for the individual firm can (e derived. Despite this similarity the
approaches for estimating technical efficiency can (e generally categorised under the distinctly
opposing techniques of parametric and non-parametric methods (!eiford and Thrall #$$%).
!tochastic estimations incorporate a measure of random error. This involves the estimation of a
stochastic production frontier where the output of a firm is a function of a set of inputs
inefficiency and random error. An often quoted disadvantage of the technique however is that
they impose an e7plicit functional form and distri(ution assumption on the data. .n contrast the
linear programming technique of data envelopment analysis (DEA) does not impose any
assumptions a(out functional form hence it is less prone to mis-specification. Further DEA is a
non-parametric approach so does not ta'e into account random error. >ence it is not
su(sequently su(@ect to the pro(lems of assuming an underlying distri(ution a(out the error term.
>owever since DEA cannot ta'e account of such statistical noise the efficiency estimates may
(e (iased if the production process is largely characterised (y stochastic elements.
The purpose in this paper is to present a range of methods for estimating technical efficiency and
to review the software that is availa(le to do so.
Economic, allocative and technical efciency
Technical efficiency is @ust one component of overall economic efficiency. >owever in order to
(e economically efficient a firm must first (e technically efficient. 6rofit ma7imisation requires
a firm to produce the ma7imum output given the level of inputs employed (i.e. (e technically
efficient) use the right mi7 of inputs in light of the relative price of each input (i.e. (e input
allocative efficient) and produce the right mi7 of outputs given the set of prices (i.e. (e output
allocative efficient) (Kum(ha'er and 3ovell :%%%). These concepts can (e illustrated graphically
using a simple e7ample of a two input (x# x:)-two output (y# y:) production process (Figure #).
Efficiency can (e considered in terms of the optimal com(ination of inputs to achieve a given
level of output (an input-orientation) or the optimal output that could (e produced given a set of
inputs (an output-orientation).
.n Figure #(a) the firm is producing a given level of output (y#
L
y:
L
) using an input com(ination
defined (y point A. The same level of output could have (een produced (y radially contracting
the use of (oth inputs (ac' to point 0 which lies on the isoquant associated with the minimum
level of inputs required to produce (y#
L
y:
L
) (i.e. .so(y#
L
y:
L
)). The input-oriented level of
technical efficiency (TE.(yx)) is defined (y %0E%A. >owever the least-cost com(ination of
inputs that produces (y#
L
y:
L
) is given (y point ) (i.e. the point where the marginal rate of
technical su(stitution is equal to the input price ratio w:Ew#). To achieve the same level of cost
(i.e. e7penditure on inputs) the inputs would need to (e further contracted to point D. The cost
efficiency ()E(y7w)) is therefore defined (y %DE%A. The input allocative efficiency
(AE.(yww)) is su(sequently given (y )E(y7w)ETE.(y7) or %DE%0 in Figure #.#(a)
(Kum(ha'er and 3ovell :%%%).
The production possi(ility frontier for a given set of inputs is illustrated in Figure #(() (i.e. an
output-orientation). .f the inputs employed (y the firm were used efficiently the output of the
firm producing at point A can (e e7panded radially to point 0. >ence the output oriented
measure of technical efficiency (TEF(y7)) can (e given (y %AE%0. This is only equivalent to the
input-oriented measure of technical efficiency under conditions of constant returns to scale.
Bhile point 0 is technically efficient in the sense that it lies on the production possi(ility
frontier a higher revenue could (e achieved (y producing at point ) (the point where the
marginal rate of transformation is equal to the price ratio p:Ep#). .n this case more of y# should (e
produced and less of y: in order to ma7imise revenue. To achieve the same level of revenue as at
point ) while maintaining the same input and output com(ination output of the firm would need
to (e e7panded to point D. >ence the revenue efficiency (*E(y7p)) is given (y %AE%D. Futput
allocative efficiency (AEF(yww)) is given (y *E(y7w)ETE.(y7) or %0E%D in Figure #(()
(Kum(ha'er and 3ovell :%%%).
#igure 1: .nput (a) and output (() oriented efficiency measures
#tochastic %roduction frontier software
!tochastic frontiers can (e estimated using a different range of multi-purpose econometric
software which can (e adapted for the desired estimation. This software includes 'nown
statistical pac'ages such as 3.1DE6 T!6 !ha-am JA2!! !A! etc. The two most commonly
used pac'ages for estimating of stochastic production frontiers and inefficiency are F*F=T.E*
;.# ()oelli #$$?a) and 3.1DE6 (Jreene #$$<). A recent review of (oth pac'ages is provided (y
!ena (#$$$).
F*F=T.E* ;.# is a single purpose pac'age specifically designed for the estimation of stochastic
production frontiers (and nothing else) while 3.1DE6 is a more general pac'age designed for a
range of non-standard (i.e. non-F3!) econometric estimation. An advantage of the former model
(F*F=T.E*) is that estimates of efficiency are produced as a direct output from the pac'age.
The user is a(le to specify the distri(utional assumptions for the estimation of the inefficiency
term in a program control file. .n 3.1DE6 the pac'age estimates a one-sided distri(ution (ut
the separation of the inefficiency term from the random error component requires additional
programming.
F*F=T.E* is a(le to accommodate a wider range of assumptions a(out the error distri(ution
term than 3.1DE6 (Ta(le #) although it is una(le to model e7ponential distri(utions. =either
pac'age can include gamma distri(utions. Fnly F*F=T.E* is a(le to estimate an inefficiency
model as a one-step process. An inefficiency model can (e estimated in a two-stage process
using 3.1DE6. >owever this may create (ias as the distri(ution of the inefficiency estimates is
pre-determined through the distri(utional assumptions used in its generation.
)a*le 1" Distri*utional assumptions allo+e' *y the soft+are
Distri&ution
'I(D
E$
)*+,-I
E*
Time invariant frm specifc
ineciency
! Half"normal distri#ution $es $es
! Truncated normal distri#ution $es $es
! E%ponential distri#ution $es &o
Time variant frm specifc
ineciency
! Half"normal distri#ution &o $es
! Truncated normal distri#ution &o $es
'ne step ineciency model &o $es
!ource9 !ena (#$$$)
)*+,-IE* .1
The most commonly used pac'age for estimation of stochastic production frontiers in the
literature is F*F=T.E* ;.# ()oelli #$$?a). This incorporates the ma7imum li'elihood
estimation of the parameters. The estimation process consists on three main steps. At the first
step F3! is applied to estimate the production function. This provides un(iased estimators for
the 8s (e7cept for the intercept term and the variance estimate). The F3! estimates are used as
starting values to estimate the final 13 model. First the value of the li'elihood function is
estimated for different values of (etween % and # given the values for the 8s derived in the
F3!. Finally an interative Davidon-Fletcher-6owell algorithm calculates the final parameter
estimates using the values of the 8s from the F3! and the value of from the intermediate step
as starting values.
F*F=T.E* ;.# has (een created specifically for the estimation of production frontiers. As such
it is a relatively easy tool to use in estimating stochastic frontier models. .t is fle7i(le in the way
that it can (e used to estimate (oth production and cost functions can estimate (oth time-varying
and invariant efficiencies or when panel data is availa(le and it can (e used when the functional
form have the dependent varia(le (oth in logged or in original units.
F*F=T.E* solves two general models. The error components model can (e formulated as
Yit M Xit N (Vit - Uit)
Bhere Yit is the (logged) output o(tained (y the i-th firm in the t-th time periodO Xit is a ('7#)
vector of (transformation of the) input quantities of the i-th firm in the t-th time periodO is a
('7#) vector of un'nown parametersO and Vit are assumed to (e iid =(% v
:
) random errors and
Uit M Ui e7p (-(t-T)) where Ui are assumed to (e iid as truncations at -ero of the =(i u
:
).
This is the 0attese and )oelli (#$$:) model. >owever some other models can (e summari-ed as
special cases of this one and can also (e solved using F*F=T.E*. !etting M% the time
invariant model of 0attese )oelli and )ol(y (#$,$) is o(tained. The 0attese and )oelli (#$,,)
model results from the previous one for the particular case of pro(lems in which (alanced data is
availa(le. .f we add M% to the aforementioned assumptions the 6itt and 3ee (#$,#) model
results. And if we finally set TM# in the 6itt and 3ee model we o(tain the original cross-sectional
data model of Aigner 3ovell and !chmidt (#$++).
.f P% the inefficiency term Uit is always decreasing with time whereas Q% implies that Uit is
always increasing with time. That could (e one of the main pro(lems when using this model
technical efficiency is forced to (e a monotonous function of time.
The second model included in the F*F=T.E* pac'age is the Technical Efficiency (TE) effects
model (0attese and )oelli #$$<). .t can (e e7pressed as Yit M Xit N (Vit - Uit) where Yit Xit and
Vit are as defined earlier and Uit R=(mit u
:
) where mit M Zit Zit is the vector of firm-specific
varia(les which may influence the firms8 efficiency. F*F=T.E* offers also the solution of the
model of !tevenson (#$,%) which is a particular case of the previous model that can (e o(tained
for the cases in which T is equal to # (for cross-sectional data).
There are two approaches to estimating the inefficiency models. These may (e estimated with
either a one step or a two step process. For the two-step procedure the production frontier is first
estimated and the technical efficiency of each firm is derived. These are su(sequently regressed
against a set of varia(les Zit which are hypothesised to influence the firms8 efficiency. A pro(lem
with the two-stage procedure is the inconsistency in the assumptions a(out the distri(ution of the
inefficiencies. .n the first stage the inefficiencies are assumed to (e independently and
identically distri(uted (iid) in order to estimate their values. >owever in the second stage the
estimated inefficiencies are assumed to (e a function of a num(er of firm specific factors and
hence are not identically distri(uted unless all the coefficients of the factors are simultaneously
equal to -ero ()oelli *ao and 0attese #$$,). F*F=T.E* uses the ideas of Kum(ha'ar Jhosh
and 1cJuc'in (#$$#) and *eifschneider and !tevenson (#$$#) and estimates all of the
parameters in one step to overcome this inconsistency. The inefficiency effects are defined as a
function of the firm specific factors (as in the two-stage approach) (ut they are then incorporated
directly into the 13E. This is something that should (e ta'en into consideration when
programming in some of the general statistical pac'ages.
F*F=T.E* offers a wide variety of tests on the different functional forms of the models that can
(e conducted easily (y placing restrictions on the models and testing the significance of the
restrictions using the li'elihood ratio test. The F*F=T.E* program is easy to use. A (rief
instruction file and a data file have to (e created. The e7ecuta(le file and the start-up file can (e
downloaded from the .nternet free of charge at the )E6A (2niversity of =ew England) we(
page9 http9EEwww.uq.edu.auEeconomicsEcepaEfrontier.htm.
+ther %ac/a0es
3.1DE6 is a multi-purpose econometric pac'age that can (e programmed to estimate several
stochastic frontier models using some of the stochastic frontier estimation routines that are
already implemented in the pac'age. The estimation process can (e carried out when the
inefficiency varia(le is assumed to have the half-normal and truncated normal distri(ution (ut
also when the e7ponential distri(ution is assumed. >ence when there seems to (e strong
evidence to assume an e7ponential distri(ution 3.1DE6 may (e a (etter option than
F*F=T.E*. Among other features 3.1DE6 offers the possi(ility of including random effects
for panel data and estimates of one-sided distur(ance components. The pac'age includes the
optimisation of non-linear regression with two-step least squares and user defined ma7imum
li'elihood estimation. There is a 3.1DE6 Be( page at http9EEwww.limdep.comE
The fle7i(ility feature of JA2!! allows the possi(ility to (e adapted for the estimation of
production frontiers. JA2!! includes some non-linear optimisation routines which can (e used
for this purpose. 0eing a very general statistical pac'age also has some disadvantages and the
programming structure is not as easy as some of the other software pac'ages. The JA2!!
program can (e ordered at http9EEwww.aptech.comEprod.html. There is also a Be( page that
provides answers to questions frequently as'ed a(out the program9
http9EEwww.indiana.eduERstatmathEgaussEinde7.html
The T!6 pac'age can (e specifically adapted for the estimation of the most (asic models of
production frontiers using the 13 user programmed li'elihood functions. The program also
includes non-linear three-stage least squares. The program can (e ordered and more information
can (e o(tained at9 http9EEwww.tspintl.comEproductsEtsp
The statistical program !ha-am can also (e adapted for the estimation of production frontiers. .t
allows the optimisation of non-linear functions either (y 1a7imum 3i'elihood or least squares
and it includes some hypothesis testing. .nformation a(out the program can (e found at9
http9EEsha-am.econ.u(c.caE
DEA software
1ost of the general-purpose mathematical optimisation software can (e adapted to solve Data
Envelopment Analysis pro(lems. E7amples of program code for DEA models have already (een
pu(lished and are readily adapta(le (e.g. Flesen and 6etersen (#$$<) present the JA1! code for
a DEA model that can (e adapted to suit most analyses). Emrou-ne@ad (:%%%) developed a !A!
program for different DEA models including different options as inputEoutput orientation and
)*!EC*!. The Emrou-ne@ad (:%%%) program and comments can (e downloaded from9
http9EEwww.dea-one.comEsoftwareEinde7.htmSsasdea.
These general programs offer the possi(ility of a wide range of applications using non-specialist
DEA software. >owever there are several DEA-specific programs that provide a variety of
interesting facilities. !even of the most common ones are listed (elow including the sources
where further information can (e found9
() EM* ()+ ,Eciency measurement system- #y H) *cheel) University of
Dortmund) .ermany) This can #e do/nloaded from the net)
http011///)/iso)uni"dortmund)de1
2) DEAP 2)( ,Coelli, (334#-) Centre for Eciency and Productivity Analysis
,CEPA-) University of &e/ En5land, &*6, Australia) 7t is no/ free and can also
#e do/nloaded from the net, as /ell as its manual from the CEPA 6e# pa5es
http011///)u8)edu)au1economics1cepa1deap)htm
+) 9rontier Analyst, version +) Professional Edition) :an%ia *oft/are ;td)
.las5o/) *cotland) http011///)#an%ia)com1
<) 6AR67C= DEA soft/are) Aston :usiness *chool) University of Aston)
:irmin5ham, En5land, U=) http011///)/ar/ic>)ac)u>1?#srlu1
@) 'n9ront 2) Economic Measurement and Auality 7 ;und A: ;und) */eden)
http011///)em8)se1soft/are)html
4) DEA"*olver Professional 2)B) *aitech 7nc) &e/ Cersey) U)*)A) This is the most
recent pac>a5e) 7ts manual is the Cooper, *eiford and Tone ,2BBB- #oo> ,Data
Envelopment Analysis) =lu/er Academic Pu#lishers-) http011///)saitech"
inc)com1
D) 7DEA* 4)() Professional version) Consultin5 7nc), Amherst, MA, U*A)
http011///)ideas2BBB)com1
$ac/a0e )acilities
The 'ey features of the different software pac'ages have (een summarised in Ta(le :.
)a*le " ,ey features of the pac-ages
1indo
w
analys
is
(alm
2
inde3
1ei0
ht
restr
In%ut
4
out%
ut
orien
t
A5$
#u%e
r6
Efc
E3o
0
)i3e
d
V&le
s
7ate0
o6
rical
v&les
(ul
ti
#ta
0e
DEA
7*
#
V*
#
,I*
#
Addi
t
(od
el
*-
#
inf
o
E(#
E E E E E E E E E
DEA$
E E E E E
8A,9IA
E E E E E
1A*1I
7:
,F- ,F- E E E E E E E
+n)ron
t 2
,F- E E ,F- E E E E
DEA
#olver
E E E E E E E E E
IDEA#
,F- E E E E E E
(L) This can (e carried out (y doing some modifications to the program
E1! and DEA-solver appear to offer the most options. Ff the software e7amined only E1! and
DEA-!olver can (e used to carry out window analysis when time series data are availa(le
although if this is the case DEA6 and FnFront can (e used for the calculation of 1almquist
indices. FnFront: also includes many different options some of them not offered (y any other
pac'age such as the a(ility to carry out input and output congestion measures and )apacity
2tili-ation measures. Barwic' software can (e adapted to compute 1almquist indices (ut only
(y running the DF! version in (atch mode to compute efficiencies. These efficiency scores are
needed later in order to compute the indicesO similarly it can (e adapted to carry out window
analysis.
The additive model can (e estimated using E1! and Barwic' software and the !lac's-0ased-
1easure (!01) model using either E1! or DEA-solver.
All the programs under study - e7cept DEA6 - allow the possi(ility of choosing (etween an input
or output orientation the assumption of either constant or varia(le returns to scale the possi(ility
of including non discretionary varia(les (i.e. e7ogenously fi7ed varia(les) and offer some scale
information. This latter possi(ility has recently (een added in the new version of FnFront under
the name of vector analysis.
0an7ia lac's some of the options that other programs offer such as window analysis (ut it
presents the advantage of offering many graphs relating efficiency to the different inputs and
outputs. .t also allows the introduction andEor elimination of some of the decision-ma'ing units
(D12s). .t is a very user-friendly pac'age and very easy to wor' with even for those who do not
have a wide 'nowledge of DEA. A thorough userTs guide is also provided. Fne-day wor'shops
introducing some of the theory of DEA its practical applications and how to handle the program
ta'e place regularly in 3ondon. Details regarding contents and cost information can (e found at9
http9EEwww.(an7ia.comEtrainingE(7fanalyst.html
*egarding the use of categorical varia(les only .DEA! and DEA-!olver offer their inclusion as
an option in the different models. The possi(ility of including wea' or strong disposa(ility of
inputs has (een added in the new version of FnFront.
E1! Barwic' software and the new version of FnFront are the only pac'ages that compute
Andersen and 6etersen super-efficiency scores. E1! and DEA-solver include the possi(ility of
wor'ing with non-conve7 technologies and to compare different programs8 efficiencies.
E1! and DEA6 offer a good on-line guide for users. 0an7ia and FnFront provide a reference
guide to users together with the program. DEA-!olver program @ust includes a very (rief
reference guide in the pac'age while more detailed e7planations a(out the options can (e found
in the associated reference (oo' ()ooper et al. :%%%).
+%eratin0 #ystem and $7 *e2uirements
1ost of the software pac'ages run under Bindows though some of them li'e FnFront and
Barwic' DEA offer (oth DF! and Bindows versions. There are some others li'e 6.F=EE*
(not included in the a(ove ta(le) that run under 2ni7 (for additional information contact
(arrUseas.smu.edu). DEA6 is the only pac'age among those presented here that does not offer a
Bindows version. .t is (ased on 1!-DF! and the input files have to (e te7t files. .t also lac's
some of the options that other programs offer. .n spite of these disadvantages DEA6 presents
many desira(le technical characteristics. Foremost of these is that DEA6 allows a multi-step
procedure to (e used in solving the DEA models. This can (e either a one-step two-step or a
multi-stage procedure. This is important as using a one-step procedure can produce insta(ilities.
1any of the other programs such as the Barwic' !oftware avoid insta(ilities using a two-step
procedure (ut wrong peers can (e o(tained for some of the inefficient units. 0oth pro(lems can
(e overcome (y the use of the DEA6 multi-step procedure. DEA6 does not requires a large
memory 6) and (eing currently free of charge is another good reason for using it.
The operating !ystem used (y all programs with the e7ception of DEA6 is Bindows $< or
later
(note)
and =T. E1! accepts E7cel$+ or te7t data files. Barwic' and FnFront require input in
the form of te7t files (although this restriction can (e lessened using the Bindows clip(oard).
The new version of 0an7ia offers a direct data import from E7cel and !6!! and the reports
produced there can (e directly e7ported to *TF (as Bord) E7cel and html files. The DEA-solver
is (ased in E7cel $+. Access D(ase E7cel and Te7t files are all accepted files for .DEA!.
*equirements regarding the capacity of the computer are highly related to how sophisticated the
software is. The most demanding are 0an7ia and FnFront. 0an7ia software requires <%10 of
hard dis' space A: 10 of *A1 and they strongly recommend using a 6entium 6). FnFront
requires at least #? 10 of *A1 and ,1( of dis' space. DEA6 only requires ; 10 of *A1 and
it can even (e run using a A,? 6). >owever it is not as user-friendly as the others.
#oftware $rices
The prices in :%%# (e7cluding CAT shipping and handling) for the software under analysis are
presented in Ta(le A.
)a*le !" Price for main DEA soft+are. //1 $exclu'es 0A)%
Academic
%rice
7ommercial
%rice
Additional
licence
;academic<
Additional
licence
;commercial<
E(# 1=
9ree 9ree 9ree 9ree
DEA$ 21
9ree 9ree 9ree 9ree
8A,9IA =
G(3@"@3@ G+3@"2+3@ +@H discount +@H discount
1A*1I7:
G2BB1@BB G2BB1@BB " "
+n)ront2
ID@B I(D@B I<@B I3BB
DEAsolver
20
IJBB I(4BB I<BB IJBB
IDEA#
I(<B1(<BB I2JB12JBB " "
The prices are for individual licences either for academic or (usiness purposes. There are some
other prices for )ampus licences educational purposes etc. !ome of the software presented
here such as 0an7ia Barwic' or .DEA! can (e o(tained at different prices if the num(er of
D12s under study is limited. 0an7ia software offers a wide variety of prices depending on the
ma7imum num(er of units it analyses. The lower level is for +< units (V#$<) then :<% (V:$<)
<%% units (VA$<) #<%% units (V;$<) and :<%% units (V<$<). The same range can (e (ought for
commercial purposes and prices vary from VA$< to V:A$<. Barwic' software offers a limited
version for analysis underta'ing less than <% units (V:%%) and a full version for an unlimited
num(er of D12s (V<%%). The same applies for the DF! Barwic' version (V#<%EA%%). .DEA!
also offers two versions the standard version which allows analyses of up to <%% D12s :%
varia(les and :% price ratios and the professional version for analyses of up to :%%% D12s :<
varia(les and #:% price ratios. The rest of the programs can (e used for any num(er of units
(within the capacity of the memory of the computer). 0an7ia has recently launch some e7tra
pac'ages at different prices9 )ross efficiency module (V$<) <%% unit three-month-trial licence
programme (V:$<) !upport and maintenance for a year (V::$) and the upgraded from the
previous version is offered (y o(taining the )D and new manual (V#?$) or (y o(taining an
electronic copy (V#:$ with no physical products).
Ac/nowled0ement
This review was underta'en as part of the E2 funded pro@ect 8Technical efficiency in E2
fisheries9 implications for monitoring and management through effort controls8 (W3K<-)T#$$$-
%#:$<).
,ote
Barwic' software also allows for Bindows A.#
*eferences
Aigner D. ). A. K. 3ovell and 6. !chmidt (#$++). Formulation and estimation of stochastic
frontier production function models. Journal of Econometrics ?9 :#-A+.
0an7ia !oftware (#$$,). 0an7ia Frontier Analyst. 2ser8s Juide. 6rofessional Edition. Frontier
Analyst.
0attese E. and T. I. )oelli (#$,,). 6rediction of firm level technical inefficiencies with a
generalised frontier production function and panel data Journal of Econometrics A, A,+-A$$.
0attese J. E. T. I. )oelli and T. ). )ol(y (#$,$) Estimation of Frontier 6roduction Functions
and the Efficiencies of .ndian Farms 2sing 6anel Data from .)*.!AT8s Cillage 3evel !tudies.
Journal of Quantitative Economics <(:)9 A:+-;,
0attese J. E. and T. I. )oelli (#$$:). Frontier 6roduction Function technical efficiency and
panel data9 with application to 6addy Farmers in .ndia. Journal of Productivity Analysis A9 #<A-
#?$.
0attese J. E. and T. I. )oelli (#$$<). A model for technical inefficiency effects in a stochastic
frontier production function for panel data. Empirical Economics :%9 A:<-AA:.
)oelli T. (#$$?a). A guide to F*F=T.E* version ;.#9 a computer program for frontier
production function estimation. )E6A Bor'ing 6aper $?E%+ Department of Econometrics
2niversity of =ew England Armidale Australia.
)oelli T. I. (#$$?(). 8A guide to DEA6 version :.#9 A Data Envelopment Analysis ()omputer)
program8. )E6A Bor'ing 6apers =o.,E$?. .!!= #A:+-;A<5. .!0= # ,?A,$ ;$?$.
)oelli T. I. D. !. 6. *ao and J. E. 0attese (#$$,). An ntroduction to Efficiency and
Productivity Analysis Kluwer Academic 6u(lishers9 2!A.
)ooper B. B. !eiford 3. 1. and Tone K. (:%%%). !ata Envelopment Analysis. Kluwer
Academic 6u(lishers. .!0= %+$:A,?$A%.
Emrou-ne@ad A. (:%%%). An E7tension to !A!EF* for Decision !ystem !upport 6roceeding of
the :< !A! 2ser Jroup .nternational )onference April :%%% .ndianapolis pp#;<?-#;?#.
FXre *. and Jross'opf !. (#$$,). 8*eference Juide to FnFront8. Economic 1easurement and
Wuality . 3und A0 3und. !weden.
Farrell 1. I. (#$<+). The measurement of productive efficiency Journal of t"e #oyal $tatistical
$ociety #:% :<:-$%.
Jreene B. (#$$<). %&!EP 'Version ()* User+s &anual and #eference ,uide Econometric
!oftware .nc. =ew 4or'.
Jreene B. >. (#$$A). /Frontier 6roduction Functions/ E)-$A-:%. !tern !chool of 0usiness
=ew 4or' 2niversity.
>ollingsworth 0. (#$$+). 8A review of Data Envelopment Analysis !oftware8. T"e Economic
Journal #$$+ #:?,-#:+%
>ollingsworth 0. (#$$$). Data Envelopment Analysis and 6roductivity Analysis9 a review of
the options. T"e Economic Journal ;<,-;?:.
Kum(ha'er !. ). and 3ovell ). A. K. (:%%%). $toc"astic -rontier Analysis. )am(ridge9
)am(ridge 2niversity 6ress.
Kum(ha'ar !. !. Jhosh and I. 1cJuc'in. (#$$#). A generali-ed 6roduction Frontier Approach
for Estimating Determinants of .nefficiency in 2.!. Dairy Farms. Journal of .usiness and
Economic $tatistics $9:+$-:,?.
3ewin A. 4. and ). A. K. 3ovell (#$$%). Editors .ntroduction Journal of Econometrics ;? A-<.
Flesen and 6etersen (#$$<). A presentation of JA1! for DEA. /omputers 0 1perations
#esearc" :A(;) pp A:A-AA$
6itt 1. 1. and 3. F. 3ee. (#$,#). The measurement and sources of technical inefficiency in the
.ndonesian weaving industry. Journal of !evelopment Economics $9;A-?;.
*eifschneider D. and *. !tevenson. (#$$#). !ystematic departures from the frontier9 A
framewor' for the analysis of firm inefficiency. nternational Economic #eview A: (A)9+#<-+:A.
!cheel >. (:%%%). E1!9 Efficiency 1easurement !ystem 2serTs 1anual. Cersion #.A
!eiford 3. 1. and *. 1. Thrall (#$$%) *ecent developments in DEA9 the mathematical
programming approach to frontier analysis Journal of Econometrics ;? +-A,.
!ena C. (#$$$). !tochastic frontier estimation9 a review of the software options. Journal of
Applied Econometrics #;9 <+$-<,?.
!tevenson *. E. (#$,%). 3i'elihood function for generali-ed stochastic frontier estimation
Journal of Econometrics #A <+-??.