Professional Documents
Culture Documents
DOI 10.1007/s10100-015-0401-z
ORIGINAL PAPER
Josef Jablonsky
1 Introduction
J. Jablonsky (B)
University of Economics Prague, W. Churchill Sq. 4, 13067 Praha 3, Czech Republic
e-mail: jablon@vse.cz
URL: http://webhosting.vse.cz/jablon
123
J. Jablonsky
outputs. This approach introduced in (Charnes et al. 1978) is based on solving linear
programming problems for each of the DMUs under evaluation. Conventional DEA
models analyse relative technical efficiency of the set of n DMUs that are characterized
by m inputs and r outputs in one period. The efficiency score θq of the DMUq is
defined as the weighted sum of outputs divided by the weighted sum of inputs as
follows: r
u k ykq
θq = k=1 m , (1)
i=1 vi x iq
123
Efficiency analysis in multi-period systems
Conventional DEA models can be formulated either in their primal or dual form - see
e.g. (Zhu 2003). In this section we will use the dual form that has several advantages
with respect to formulation of multi-period models. Let us suppose that the DMUs
are described by the same set of inputs and outputs in T consecutive time periods
t = 1, 2, . . ., T , and assume that xit j , i = 1, 2, . . ., m, j = 1, 2, . . ., n, and ykt j , k =
1, 2, . . ., r, j = 1, 2, . . ., n are the values of the i-th input and the k-th output in the
t-th period of the DMU j. The first phase of the output oriented PP model with constant
(variable) returns to scale assumption that evaluates the aggregative efficiency of the
DMUq is formulated as follows:
Maximize θq
n
subject to xit j λtj ≤ xiq
t
, i = 1, 2, . . . , m, t = 1, 2, . . . , T,
j=1
n
yit j λtj ≥ θq yiq
t
, k = 1, 2, . . . , r, t = 1, 2, . . . , T, (2)
j=1
⎛ ⎞
n
⎝ λtj = 1, t = 1, 2, . . . , T ⎠ ,
j=1
λtj ≥ 0, j = 1, 2, . . . , n, t = 1, 2, . . . , T,
T
m
T
r
Maximize sit− /xiq
t
+ skt+ /yiq
t
,
t=1 i=1 t=1 k=1
n
subject to xit j λtj + sit− = xiq
t
, i = 1, 2, . . . , m, t = 1, 2, . . . , T,
j=1
123
J. Jablonsky
n
yit j λtj − skt+ = θq∗ yiq
t
, k = 1, 2, . . . , r, t = 1, 2, . . . , T,
j=1
⎛ ⎞
n
⎝ λtj = 1, t = 1, 2, . . . , T ⎠ , (3)
j=1
sit− ≥ 0, i = 1, 2, . . . , m, t = 1, 2, . . . , T,
skt+ ≥ 0, k = 1, 2, . . . , r, t = 1, 2, . . . , T,
λtj ≥ 0, j = 1, 2, . . . , n, t = 1, 2, . . . , T,
where sit− and skt+ are slack/surplus variables belonging to constraints of Model (2).
According to the results of both stages the DMUs can be classified as follows:
• (Fully) efficient DMUs—the optimal objective function value of Model (2) θq∗ = 1
and matrices of slack and surplus variables are zero matrices.
• Weakly efficient DMUs—the optimal objective function value of Model (2) θq∗ = 1
and at least one element of matrices of slack/surplus variables is positive.
• Inefficient units—the optimal objective function value of Model (2) θq∗ > 1 (for
output oriented models).
The second phase of the PP model is slightly modified comparing to its original
formulation. The original formulation contains in its objective function a simple sum
of slack and surplus variables which seems to be meaningless due to possible high
differences in input/output values. That is why a sum of relative slack and surplus
variables is used in objective function of Model (3).
In order to evaluate the DMUs, Model (2) must be solved in the first phase and
depending on the optimal objective function value the second phase follows. This phase
must be applied when the optimal objective function value of the first phase is θq∗ = 1,
i.e. the DMU under evaluation is recognized as fully or weakly efficient. In this case the
second phase allows determination whether this unit is fully or weakly efficient. The
inefficient or weakly efficient DMUs can be ranked according to the optimal objective
function values from the first or second phase respectively. Nevertheless, the efficient
units cannot be ranked at all using the PP model as proposed in (Park and Park 2009).
In single period systems, this problem is solved by using of super-efficiency models
or any other models that allow ranking of efficient DMUs. Information about the most
important models of this class can be found e.g. in (Jablonsky 2007).
First super-efficiency model was introduced by Andersen and Petersen (1993).
Their model is based on the idea to remove the unit under evaluation from the set of
units and then measure its distance from the new efficient frontier. The super-efficiency
score of an originally efficient DMU is lower than one (taking into account the output
oriented model) and can be explained as a rate for possible worsening of outputs in
order to keep the efficiency status. A similar approach as in Andersen and Petersen
model can be used for multi-period models too. For these purposes Model (2) can be
extended by additional constraints that ensure zero values for all weights of the DMU
under evaluation, i.e.
λqt = 0, t = 1, 2, . . . , T, (4)
123
Efficiency analysis in multi-period systems
where θqt is the efficiency score in the period t and ψq is the worse efficiency score of
the of the DMUq , and ε is an infinitesimal constant. The objective function of Model
123
J. Jablonsky
(5) consists of two parts. The first part is the maximization of the efficiency scores
in the particular periods and the second one is the minimization of the maximum
efficiency score. The model returns optimal value ψq∗ which is the worse efficiency
score (or super-efficiency score for DMUs that are efficient in all particular periods)
of the unit under evaluation over all periods.
Average efficiency score ωq of the unit under evaluation can be given using the
following Model (6):
T
Maximize ωq = θqt /T,
t=1
n
subject to xit j λtj ≤ xiq
t
, i = 1, 2, . . . , m, t = 1, 2, . . . , T, (6)
j=1
n
yit j λtj ≥ θqt yiq
t
, k = 1, 2, . . . , r, t = 1, 2, . . . , T,
j=1
λtj ≥ 0, j = 1, 2, . . . , n, t = 1, 2, . . . , T,
The objective function of Model (6) returns average efficiency score of the unit under
evaluation over all periods. The optimum value of this model ωq∗ > 1 indicates ineffi-
ciency at least in one period, ωq∗ = 1 means that the unit under evaluation is globally
efficient, i.e. it is efficient in all considered periods. It is possible to discriminate among
these units using a super-efficiency model. Andersen and Petersen (1993) model is one
of suitable alternatives. For this purpose Model (6) has to be extended by constraint (4).
Then, the objective function of this extended model measures average super-efficiency
of the DMUq .
SBM models are quite popular group of models that allow evaluation of efficiency
independently on the orientation of the model. Probably the most often applied and
discussed is the Tone’s model in its standard and super-efficiency versions—see (Tone
2002)—but this model is not suitable for its modification for multi-period analysis.
We propose for this purpose a standard SBM model that measures inefficiencies using
relative slacks, i.e. negative relative deviations in the input space and positive ones
in the output space. This model can be easily written for multi-period system as
follows:
m
T
r
t− t t+ t
Minimize υ= s1i /xiq /m + s2k /ykq /r /T,
t=1 i=1 k=1
n
t−
subject to xit j λtj + s1i = xiq
t
, i = 1, 2, . . . , m, t = 1, 2, . . . , T,
j=1
123
Efficiency analysis in multi-period systems
n
t+
ykt j λtj − s2i = ykq
t
, k = 1, 2, . . . , r, t = 1, 2, . . . , T, (7)
j=1
λtj ≥ 0, j = 1, 2, . . . , n,
t− t+
s1i ≥ 0, s2k ≥ 0, i = 1, 2, . . . , m, k = 1, 2, . . . , r, t = 1, 2, . . . , T,
t− t+
where s1i , s2k are negative and positive slacks in the period t of the i-th input and
the k-th output respectively. The other symbols have the same meaning as in Model
(6). The objective function of Model (7) consists of two parts. The first part measures
average inefficiency in input space and the second one the average inefficiency in the
output space. The total objective function value is the average of both inefficiencies
over all time periods.
The optimal objective function of Model (7) υ ∗ = 0 if the DMU under evaluation
is efficient in all periods, otherwise the DMU is inefficient at least in one period. The
units can be easily ranked according to the average inefficiency over all periods. The
units that are efficient in all periods can be ranked according to their average super-
efficiency measure that can be given using a modified version of the model proposed
in (Jablonsky 2012). Here, the super-efficiency is measured using undesirable slacks,
i.e. using positive deviations in the input space and negative ones in the output space.
The model is formulated in the following way:
m
T
r
t+ t t− t
Minimize ρ =1+ s1i /xiq /m + s2k /ykq /r /T,
t=1 i=1 k=1
n
t− t+
subject to xit j λtj + s1i − s1i = xiq
t
, i = 1, 2, . . . , m, t = 1, 2, . . . , T,
j=1
n
t− t+
ykt j λtj + s2i − s2i = ykq
t
, k = 1, 2, . . . , r, t = 1, 2, . . . , T, (8)
j=1
λtj ≥ 0, j = 1, 2, . . . , n, λqt = 0, t = 1, 2, . . . , T,
t− t+ t− t+
s1i ≥ 0, s1i ≥ 0, s2k ≥ 0, s2k ≥ 0,
i = 1, 2, . . . , m, k = 1, 2, . . . , r, t = 1, 2, . . . , T,
t+ t−
where s1i , s2k are positive and negative slacks in the period t of the i-th input and
the k-th output respectively and other symbols correspond to Model (7). It is possible
to prove that the DMUs inefficient in all periods have the efficiency score given by
Model (8) ρ ∗ = 1 and the units that are efficient in at least one period ρ ∗ > 1 (higher
values correspond to more efficient units). That is why the inefficient DMUs can be
ranked according to the efficiency scores given by Model (7) and the units identified
as efficient by this model can be discriminated according to the results of Model (8).
As stated previously Model (2) and Model (3) evaluate the DMUs over the
given time periods according to the best conditions of the DMU under evalua-
tion which can lead to very unsatisfactory results, e.g. the DMU highly efficient in
the first period and inefficient in all other periods can easily be rated as the best
123
J. Jablonsky
DMU according to the PP Model (2). In order to analyze deeper the given multi-
period problem it is useful to derive the efficiency score based on the “worse”
period of the unit under evaluation, and/or calculate the average efficiency score.
This can be given using Model (5) and Model (6) that are formulated in Sect. 2.
A mutual comparison of the “best”, “worse”, and average efficiencies over all peri-
ods can contribute to better understanding the problem solved. An additional pos-
sibility for analysis of multi-period systems consists in application of the SBM
Model (7) and, if necessary, its super-efficiency version which is Model (8). They
are based on measuring the efficiency using slacks and using average efficiency
score instead of maximum efficiency score which usually leads to more appropriate
results.
123
Efficiency analysis in multi-period systems
123
J. Jablonsky
123
Efficiency analysis in multi-period systems
scores are 0 (see Tables 2, 3). In order to discriminate among these globally efficient
DMUs we propose to use super-efficiency Model (8). The results of this model for
both cases, three and five-output DEA model, are presented in Table 4.
As it is evident from Table 4, the SBM super-efficiency measures do not differ signif-
icantly. They lead to the rankings that correspond more or less to the results of other
models presented in Tables 2 or 3, i.e. the most efficient faculty according to the model
with three outputs is FSV UK, which corresponds to national expectations, and the
most efficient unit according to the model with five outputs is FIS VSE.
The comparison of results given by the original PP model and its super-efficiency
modification is quite clear. The original procedure cannot rank fully efficient DMUs
Table 4 Results of
Faculty Three outputs Rank Five outputs Rank
super-efficiency Model (8)
FSV UK 1.1092 2 1.1230 3
FME Zlin 1.1284 1 1.1297 2
FIS VSE Inefficient xx 1.1328 1
NH VSE Inefficient xx 1.1126 4
123
J. Jablonsky
and, as it is shown, among the 19 faculties 6 are efficient (10 of them in case of the
second set of outputs), i.e. 6 or even 10 units cannot be ranked by using the original
model. The modified procedure allows ranking of all units easily.
Another question is a discussion about the contribution of the presented PP model
and its modification. As it is clear the efficiency (super-efficiency) score of this multi-
period model is given as the best (minimum in case of output oriented model) efficiency
score of particular periods. That is why we propose other two modifications of the PP
model that measure the efficiency according to the worse and average periods of the
unit under evaluation. The advantage of the PP model and its modifications consists
in a possibility to get the multi-period efficiency score by solving one optimization
problem for each DMU.
The results of models presented in Tables 2, 3 and 4 can be discussed from two
different points of view. The first one is a comparison of results given by different multi-
period models, i.e. the results separately within Table 2 on one hand and Table 3 on
the other hand. The second one is discussion about differences in efficiency evaluation
depending on two different sets of outputs used.
The results in both tables show that there is a significant difference in evaluation of
DMUs when they are evaluated according to their best or worst period, or according
to the average efficiency score from all periods. E.g. ESF MU (the fourth faculty in the
list) has its highest efficiency score 1.0405 (Table 2), which indicates efficiency at least
in one period, and quite high ranking (6th). Nevertheless, its worse efficiency score is
0.6151, which is one of the worse results among all faculties (17th). Probably more
correct information is given by average efficiency score—0.7751 and the 12th rank.
The SBM efficiency score leads in this case to similar results as average efficiency.
Comparison of results in Tables 2 and 3 are quite contradictory. They show that it
is very important to pay attention to a proper selection of input and output variables
of the model. The models with three outputs contain only one research output—the
number of RIV points. This number is given directly as the weighted sum of the
numbers of publications in particular categories (each category has its weight reflect-
ing its importance). The model with five outputs take into account three research
outputs—books, Jimp and CPCI proceedings (weighting of these three categories
is the main issue for definition of RIV points). The results in Table 2 show that
all DMUs that are globally efficient in first group of models remain globally effi-
cient in the second group as well. Nevertheless, the opposite relation does not hold.
Some of highly inefficient units in models with three outputs are efficient in models
with five outputs. An extreme situation occurs for FIS VSE. This faculty is highly
inefficient in the first group of models—its best efficiency score over all periods is
only 0.8236 (rank 16th). There is a little better result for this unit when the model
for the worse or average efficiency is used. In the contrary, FIS VSE is globally
efficient in the second group of models and its super-efficiency score is the best
and it is on the highest rank among all faculties. This difference can be explained
quite easily by analysis of the source data set presented in Table 1. FIS VSE has
the highest number of publication in CPCI proceedings in 2012, it is much higher
than the other faculties have. This fact causes that the faculty is globally efficient,
i.e. is efficient in all periods considered. This example demonstrates very well that
123
Efficiency analysis in multi-period systems
it is necessary to build the model carefully and explain its results in the context of
reality.
5 Conclusions
Analysis of efficiency of DMUs within multiple periods is an important task and many
various models were proposed for these purposes in the past. The paper presents the
PP model (2), which is one of the latest contributions in the area of multi-period DEA
models. Except the conventional PP model several its modifications are formulated.
The main motivation for formulation of these models consists in disadvantages of the
PP model that were discussed mainly in Sect. 2 of the paper. A super-efficiency PP
model and models that are, in the contrary to the PP model, based on the efficiency in
the “worse” period and average efficiency over all periods of the unit under evaluation
were formulated and tested on the real data set. In addition to the PP model and its
modifications an original SBM and super-SBM multi-period model has been proposed
as an alternative approach to conventional PP models.
The advantage of the proposed super-efficiency multi-period DEA model consists
in its computational efficiency because it is necessary to solve only one optimization
problem for each DMU in order to get its aggregative (super-) efficiency score. The
disadvantage is given by the fact that the final results depend on the efficiencies of
the DMUs in particular periods without any interconnections among periods. Another
problem consists in a possible infeasibility of Model (2) and its modifications with
additional constraints (4) under the assumption of variable returns to scale. This prob-
lem is widely discussed by many researchers, see e.g. Lee and Zhu (2012). A future
research in this field can be focused on analysis of multi-period production systems
with interconnections among the inputs and/or outputs in particular periods, and in
analysis of multi-period systems under the assumption of uncertainty in data set in all
or at least some periods.
Acknowledgments The research is supported by the Grant Agency of the Czech Republic, Project No.
P403/12/1387.
References
Andersen P, Petersen NC (1993) A procedure for ranking efficient units in data envelopment analysis.
Manag Sci 39(10):1261–1264
Beasley JE (1995) Determining teaching and research efficiencies. J Oper Res Soc 46(4):441–452
Begicevic N, Divjak B, Hunjak T (2010) Decision-making on prioritization of projects in higher education
institutions using analytic network process approach. CEJOR 18(3):341–364
Charnes A, Cooper WW, Rhodes E (1978) Measuring the efficiency of decision making units. Eur J Oper
Res 2(6):429–444
Fare R, Grosskopf S (1996) Intertemporal production frontiers with dynamic DEA. Kluwer, Boston
Jablonsky J (2007) Measuring the efficiency of production units by AHP models. Math Comput Model
46(7–8):1091–1098
Jablonsky J (2012) Multicriteria approaches for ranking of efficient units in DEA models. CEJOR 20(3):435–
449
Lee HS, Zhu J (2012) Super-efficiency infeasibility and zero data in DEA. Eur J Oper Res 216(2):429–433
Park KS, Park K (2009) Measurement of multiperiod aggregative efficiency. Eur J Oper Res 193(2):567–580
123
J. Jablonsky
Sueyoshi T, Sekitani K (2005) Returns to scale in dynamic DEA. Eur J Oper Res 161(2):536–544
Thanassoulis E, Kortelainen M, Johnes G, Johnes J (2011) Costs and efficiency of higher education insti-
tutions in England: a DEA analysis. J Oper Res Soc 62(7):1282–1297
Tone K (2002) A slack-based measure of super-efficiency in data envelopment analysis. Eur J Oper Res
143(1):32–41
Zhu J (2003) Quantitative models for performance evaluation and benchmarking: Data envelopment analysis
with spreadsheets and DEA Excel Solver. Kluwer, Boston
123