You are on page 1of 5

2019 8th International Conference on Modeling Simulation and Applied Optimization (ICMSAO)

Solving DEA models in spreadsheets and modeling


languages
Josef Jablonsky
Department of Econometrics
University of Economics, Prague
13067 Praha 3, Czech Republic
email: jablon@vse.cz
ORCID: 0000-0003-0606-354X

Abstract—Data envelopment analysis (DEA) models are Maximize


tools for efficiency and performance evaluation of a set of r
decision-making units (DMUs). They are formulated as linear
programming problems that have to be solved for each DMU of
 uk yqk
k 1
the given set. The paper discusses problems with solving DEA
q  m
models and presents an original system written in LINGO  v j xqj
modeling language and using MS Excel as the user's interface. j 1
This system covers main DEA models including multiplier and
envelopment models, models for ranking of DMUs, network subject to (2)
models, etc. Depending on the version of the LINGO system the r
DEA solver can analyze problems from 200 until the unlimited
number of DMUs.
 uk yik
k 1
m
 1, i  1,..., n,
Keywords—data envelopment analysis, modeling languages,
LINGO, MS Excel, solver
 v j xij
j 1

v j   , uk   ,
I. INTRODUCTION
The history of data envelopment analysis (DEA) models where  is an infinitesimal constant that ensures strict
started in 1978 when Charnes, Cooper and Rhodes published positivity of the weights. Model (2) is not linear but can be
their pioneering work [1]. In general, they define efficiency as easily transformed into a linear program using Charnes-
the level of transformation of inputs into outputs, and the Cooper transformation. The linearized version of input-
efficiency score of the i-th decision-making unit (DMUi, oriented CCR model is as follows:
i = 1,…,n) as the weighted sum of its outputs divided by the
weighted sum of inputs as shown in (1): Maximize
r

r
 uk yqk ,
k 1
 uk yik subject to (3)
k 1
 i    m
m
 v j xij  v j xqj  1,
j 1
j 1
r m

where xij, j = 1,…, m, is the non-negative value of the j-th input


 uk yik   v j xij  0, i  1,..., n,
k 1 j 1
for the DMUi, yik, k = 1,…, r, is the non-negative value of the
k-th output for the same unit, and vj, j = 1,…, m, and uk, v j   , uk   .
k = 1,…, r, are the weights of the inputs and outputs,
respectively. In this way, the efficiency score i of the DMUi Model (3) is often denoted as the multiplier version of
is a measure of the transformation of m inputs into r outputs. CCR input-oriented model. The dual problem to (3) is the
envelopment CCR input-oriented model:
In DEA models, the efficiency score (1) is not considered
absolutely but relatively with respect to other units of the set. Minimize
The model introduced in [1] maximizes the efficiency score of  m r 
the unit under evaluation under the assumption that the q    s j   sk 
efficiency scores of all other units are lower or equal than 1.  j 1 
 k 1 
This model (often denoted as CCR model) looks as follows: subject to (4)
n
 xij λi  s j  θq xqj , j  1,...,m,
i 1
n
 yik λi  sk  yqk , k  1,...,r,
i 1
The research is supported by the Grant Agency of the Czech Republic,
project no. 19-08985S, Models for efficiency and performance evaluation
in non-homogeneous economic environment.

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

978-1-5386-7684-4/19/$31.00 ©2019 IEEE


2019 8th International Conference on Modeling Simulation and Applied Optimization (ICMSAO)

sk  0, k  1,..., r , solver does not depend on the number of DMU as in


the previous two cases.
s j  0, j  1,..., m,
 DEA Frontier Excel add-in (www.deafrontier.net) is
i  0, i  1,..., n. an MS Excel add-in application created by J. Zhu.
This solver contains all models included in the
Output-oriented versions (multiplier and envelopment) of author's book [4], i.e. the number of models if high and
the CCR model can be derived in a similar way. Since 1978, is still being extended. A free version of this software
many modifications of the CCR model have been formulated. is limited by 20 DMUs and a selection of available
Banker, Charnes and Cooper [2] extended the formulations models.
above by different assumptions about returns to scale. Models
 MaxDEA Ultra (Basic) version 7 (www.maxdea.cn)
(3) and (4) assume constant returns to scale, i.e. define the
includes probably the highest number of DEA models
conic envelope of the data set. In order to consider variable
(Ultra version) at all. It is an exciting option for all
(VRS), non-increasing (NIRS) or non-decreasing (NDRS)
individuals interested in efficiency and performance
returns to scale model (4) have to be extended by one of the
evaluation. Basic version allows solving problems
following constraints:
with the unlimited number of DMUs using traditional
n models (3) and (4) and several of their modifications,
VRS  i  1, and it is free.
i 1
n Except for those rather commercial software tools, one
NIRS  i  1, (5) can find several free solvers with different capabilities and
i 1 user's comfort. Some time ago we created a solver that was
n designed as an add-in application for MS Excel and used
NDRS  i  1. in/built Excel optimization tool for solving linear programs
i 1
within DEA models. More information about this system can
be found in [5]. This tool was and still is downloadable from
The multiplier form (3) of the CCR model must be modified author's web pages, but it has several disadvantages. It is
accordingly. The units under evaluation are identified as limited to problems up to 200 DMUs which is given by limits
efficient if their efficiency scores (objective functions of of standard solver in MS Excel, and it is quite slow. Another
models (3) or (4)) equal to 1, lower values indicate problem is a quite difficult extension of the system by other
inefficiency. models. That is why we have decided to create a new system
Next Section of the paper informs about several most that will cover a higher number of models and will be
important commercial and academic solvers for DEA models. available to interested users easily even for a higher number
Section 3 describes in detail the system for solving DEA of DMUs. This new system uses MS Excel similarly that the
models created by the authors. Its properties and advantages previous one but uses LINGO solver as an external linear
are shortly discussed in the final Section of the paper. programming solver.

II. SOLVING DEA MODELS III. MS EXCEL/LINGO DEA SOLVER


In order to verify the level of efficiency, the presented LINGO is a modeling language and system with several
DEA models have to be solved n times, where n is the number favorable properties concerning our goals. See Schrage [6]
of DMUs. That is why decision makers interested in using for more details. Among its main advantages belong:
DEA models need appropriate software tools. There are  The model written in LINGO modeling language is
several commercial DEA solvers but they are not free, and general and can be used for any data sets of any sizes.
their availability is limited. Among the most known DEA
solvers belong:  The model can read data sets from MS Excel sheet and
export its results back MS Excel.
 Frontier Analyst version 4 (www.banxia.com) which
is one of the first commercial solvers for DEA models.  LINGO commands allow repeating of calculations of
It is oriented rather on managerial use and not suitable the same model with different parameters which is the
for research because the system covers a quite limited case of DEA models.
number of models. The top version allows solving
problems with 20000 DMUs and 32 variables (inputs  LINGO models can be easily modified which is
and outputs). especially crucial for DEA models that often differ not
significantly in their mathematical formulation.
 Performance Improvement Management (PIM) DEA
software (www.deasoftware.co.uk) is a better option  LINGO models can be placed directly to a hidden
for both managerial and research purposes than the Excel sheet and run using a simple VBA procedure.
previous one. It is cheaper and includes a much higher This property allows building a library of models that
number of DEA models of various categories. can be called from MS Excel menu.

 DEA Solver Pro 12.1 (www.saitech-inc.com) is the  LINGO linear solver has better properties (is faster)
tool that is based on the textbook Cooper, Seiford and than in-built Excel solver.
Tone [3] and supports all models included in this For illustration purposes, we will show how close the
book. It means the number of models is high but newer model written in LINGO language is to its mathematical
versions of models are not included. The price of this
2019 8th International Conference on Modeling Simulation and Applied Optimization (ICMSAO)

formulation. We will use the envelopment form of CCR B. Multiplier models


model (4):
The same models as in the previous case but in their
MODEL: multiplier form. The user can choose input or output
SETS: orientation of the model and one of four assumptions for
INP/@OLE('datadea.xlsx')/:SMIN; returns to scale. Output information given by this group of
OUT/@OLE('datadea.xlsx')/:SPLUS; models are:
DMU/@OLE('datadea.xlsx')/:LAMBDA, ESCORE;
MATX(DMU,INP): X;  Efficiency scores,
MATY(DMU,OUT): Y;
ENDSETS  Target values for inefficient units (how to improve
their input/outputs to reach the efficient frontier), and
DATA:
X, Y = @OLE('datadea.xlsx');  Multipliers for inputs and outputs, i.e. optimal values
EPS = 10E-8; of variables vj, j = 1,…, m, and uk, k = 1,…, r.
ENDDATA
C. Slack-based measure (SBM) models
MIN = THETA - EPS*(@SUM(INP:SMIN) +
@SUM(OUT:SPLUS)); SBM models measure the distance of inefficient units from
@FOR(INP(J): @SUM(DMU(I): X(I,J)*LAMBDA(I)) the efficient frontier using slack and surplus variables
+ SMIN(J) = THETA*X(Q,J)); s j , j  1,..., m , and sk , k  1,..., r. This group of models con-
@FOR(OUT(K): @SUM(DMU(I): Y(I,K)*LAMBDA(I)) -
SPLUS(K) = Y(Q,K));
tains the following three modules in our application:
 Additive model with different assumptions about
CALC: returns to scale. These models are identical to
@FOR(DMU(P): Q = P;
formulation (4) after removing radial variable q and
@SOLVE();
ESCORE(P) = THETA;); an infinitesimal constant  from the objective function.
ENDCALC  Weighted additive model is identical to the previous
case but allows adding weights of the slacks/surpluses
DATA:
to the objective function.
@OLE('datadea.xlsx') = ESCORE;
ENDDATA  SBM model introduced in Tone [7]. Recently, this
END model belongs among very popular DEA models in
This LINGO model reads the names of inputs, outputs and various applications. Instead of the sum of slacks and
DMUs from the ranges of the file datadea.xlsx INP, OUT, surpluses, this model minimizes the following
and DMU respectively. The matrices of inputs and outputs expression:
are read from ranges X and Y. The results (in this case
1 m 
efficiency scores of all DMUs) are exported to the range
ESCORE of the same file.
1
m j 1

 s j / xqj 
q  (6)
For the case of variable returns to scale the LINGO model 1 r
can be easily modified by adding the row 
1   sk / yqk
r k 1

@SUM(DMU: LAMBDA) = 1;
Constraints of the model are the same as in the previous
to the body of the model. two cases. The objective function is not linear, but as
The system we created consists of several groups of DEA proposed in [7] it can be linearized, and our solver uses
models that will be described briefly below. this linearized version.
The DMUs under evaluation are efficient according to the
A. Envelopment models SBM models if all slack and surplus variables are 0, i.e. the
This group of models contains traditional input-oriented objective function of additive models is 0 and Tone's SBM
model (4) and its output-oriented modification. Both models model is 1. Otherwise, the DMUs are inefficient.
can be extended by assumptions about returns to scales (VRS, This module of the system returns the output information
NIRS, NDRS) by adding one of the constraints (5). The in the same structure as envelopment models.
following is the main information offered by this group of
models for all DMUs: D. Models for ranking of efficient DMUs
 Efficiency scores, Even though the main aim of DEA models is not to rank
 Target values for inefficient units (how to improve the DMUs, decision makers often require a ranking of the
their input/outputs to reach the efficient frontier), and DMus within the given set. All models of first three groups
split the DMUs into two classes – efficient and inefficient. The
 Non-zero weights of DMUs (peers) – optimal values inefficient ones can be ranked using their efficiency score
of i, i = 1,…, n, variables. given by the DEA model applied, but the efficient ones cannot
be raked using this measure because their efficiency scores are
identical. To distinguish among the efficient units, various
models on various methodological principles have been
2019 8th International Conference on Modeling Simulation and Applied Optimization (ICMSAO)

proposed in the past. Our application contains three of them the module that allows solving these models are identical to
that belong probably the most often used models of this results of the module with envelopment models.
category:
G. Network models
 Super-efficiency model introduced by Andersen and
The last module of the application contains the possibility
Petersen in [8]. This model is identical to model (4),
to evaluate the efficiency of two-stage serial production
only removes the unit under evaluation from the set of
processes. In such processes, the outputs of the first stage enter
DMUs, i.e. the weight of this unit q = 0. In this way, the second stage as inputs. The outputs of the second stage are
the efficiency score of the inefficient units remains considered as final outputs of the production process. Several
unchanged, but the score for efficient units is greater models for two-stage processes have been proposed in the
than 1 which allows their ranking. past. Our application includes two of them. They are one of
the first models of this category. They are:
1 m *
 x j / xqj
m j 1  Kao and Hwang model introduced in [12]. It is input or
Minimize  qSBM  , (7) output oriented, and its result is the overall efficiency
1 r * score over both stages.
 yk / yqk
r k 1  Chen model formulated in [13] that calculates
n efficiency in both stages – the first one with input
subject to  xij i  s j  x*j , i = 1,2,...,n, orientation, the second one with output orientation.
i 1,  q The application returns efficiency score of both stages
n given by the model.
 yij i  sk  yk* , k = 1,2,...,r,
Working with the application is very easy – everything is
i 1,  q
managed in MS Excel environment. Users need not know
x*j  xqj , j = 1,2,...,n, anything about LINGO or other modeling languages. The only
yk*  yqk , k = 1,2,...,r, need is to prepare the data set in an Excel file. There are no
special requirements on the data set – only need is to place the
i ≥ 0, i = 1,2,…,n, i ≠ q. matrices with inputs and outputs in continuous ranges in any
sheet. The results for each applied model are stored in
Model (7) is not linear but can be easily transformed separated Excel sheets.
into a linear program that is used in our application.
For the units that are not efficient in Tone’s SBM IV. CONCLUSIONS
model returns model (7) objective function equal to 1,
the SBM efficient units have in this model super- MS Excel/LINGO DEA solver is a user-friendly
efficiency score greater than 1. application. There are no special requirements on decision-
makers skills except basic knowledge of working with
 Super-efficiency model formulated by Jablonsky in spreadsheets. The following are its main features and
[10]. This model measures the distance of the unit advantages:
under evaluation from the new efficient frontier after
its removing using undesirable slack and surplus  The application includes the most often used DEA
variable and goal programming methodology. models.
Mathematical formulation of this model can be found  Using the application does not need any knowledge
in [10] or [11]. about DEA or modeling languages.
All three super-efficiency models return super-efficiency
 Extension of the system by other DEA models is
scores that can be used for complete ranking of efficient
straightforward even though it cannot be done directly
units.
by users of the system.
E. Models with undesirable variables  Theoretically, the system can solve problems with an
unlimited number of DMUs. Its capacity is given by
In many cases, the efficiency evaluation model contains the version of the installed LINGO solver that is free
undesirable variables, especially outputs. In this case, the data for academic purposes in its top release.
set for undesirable measures must be modified (e.g. by
subtraction from the highest values of undesirable variables  The system is fast – solving problems with 1000
within the set). Then, the traditional model (4) with four DMUs takes just a few seconds.
possible assumptions about returns to scale is applied. The
results that are offered by this module are identical to the  The DEA solver is free and can be downloaded from
results returned by envelopment models. the author’s web pages.

F. Models with uncontrollable variables REFERENCES


[1] A. Charnes, W. W. Cooper, and E. Rhodes, “Measuring the efficiency
The inputs or outputs of the model cannot always be of decision making units”, Eur J Oper Res, vol. 2, pp. 429–444, 1978.
directly controllable by decision-maker. In this case, the
[2] R. D. Banker, A. Charnes, and W. W. Cooper, “Some models for
uncontrollable variables must be fixed, and model (4) must be estimating technical and scale inefficiencies in data envelopment
modified accordingly. More information about this class of analysis”, Manage Sci, vol. 30, pp. 10781092, 1984.
DEA models can be found in [4], [5] or [11]. The results of [3] W. W. Cooper, L. M. Seiford, K. Tone, Data envelopment analysis:
A comprehensive text with models, applications, references and DEA-
solver software, Berlin, Springer, 2006.
2019 8th International Conference on Modeling Simulation and Applied Optimization (ICMSAO)

[4] J. Zhu, Quantitative models for performance evaluation and bench- [9] K. Tone, “A slack-based measure of super-efficiency in data envelop-
marking: Data envelopment analysis with spreadsheets. Boston, ment analysis”, Eur J Oper Res, vol. 143, pp. 3241, 2002.
Springer, 2009. [10] J. Jablonsky, “Multicriteria approaches for ranking of efficient units in
[5] J. Jablonský, “MS Excel based software support tools for decision DEA models”, Cent Eur J Oper Res, vol. 20, pp. 435–449, 2012.
problems with multiple criteria”, In: 17th International Conference [11] M. Dlouhy, J. Jablonsky, P. Zykova, Data envelopment analysis (in
Enterprise and Competitive Environment 2014. Book Series: Procedia Czech). Praha, Professional Publishing, 2018.
Economics and Finance, vol. 12, pp. 251258, Amsterdam, Elsevier,
2014. [12] C. Kao, and S. N. Hwang, “Efficiency decomposition in two-stage data
envelopment analysis: An application to non-life insurance companies
[6] L. Schrage, Optimization modeling with LINGO. Chicago, Lindo in Taiwan”, Eur J Oper Res, vol. 185, pp. 418–429, 2008.
Systems Inc., Chicago, 1999.
[13] Y. Chen, L. Liang, and J. Zhu, “Equivalence in two-stage DEA
[7] K. Tone, “A slack-based measure of efficiency in data envelopment approaches”, Eur J Oper Res, vol. 193, pp. 600–604, 2009.
analysis”, Eur J Oper Res, vol. 130, pp. 498509, 2001.
.
[8] P. Andersen, and N. C. Petersen, “A procedure for ranking efficient
units in data envelopment analysis”, Manag Sci, vol. 39, pp. 1261
1264, 1993.

You might also like