Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Manual
http://www.MaxDEA.cn
1 Overview .................................................................................................1
Radial ........................................................................................................... 12
i
Hybrid distance: EBM ................................................................................. 17
Orientation........................................................................................................ 18
ii
Preference (weighted) Model .......................................................................... 28
Self-benchmarking ....................................................................................... 30
Cross-benchmarking .................................................................................... 31
Downward-benchmarking............................................................................ 31
Upward-benchmarking ................................................................................ 32
Lower-adjacent-benchmarking .................................................................... 33
Upper-adjacent-benchmarking ..................................................................... 34
Window-benchmarking ................................................................................ 34
Score ............................................................................................................ 48
iii
Input Non-radial Inefficiency(for Hybrid model) ........................................ 49
Σλ ................................................................................................................. 49
Original ........................................................................................................ 50
Projection ..................................................................................................... 50
α, β and γ ...................................................................................................... 50
iv
Variable-benchmark Model ............................................................................ 54
Weight .......................................................................................................... 57
v
Efficiency(t) ................................................................................................. 62
Bootstrapped Scores..................................................................................... 68
Smoothed Bootstrap......................................................................................... 69
Simplify Results................................................................................................ 69
vi
Results from Bootstrapping DEA scores ........................................................ 69
Score(Original) ............................................................................................ 70
Bias .............................................................................................................. 70
Mean ............................................................................................................ 70
Median ......................................................................................................... 70
SD ................................................................................................................ 70
CI_LowerBound .......................................................................................... 70
CI_UpperBound ........................................................................................... 70
2) RTS = Scale Efficiency or Scale Effect (CRS Score / VRS Score) ......... 73
8 Options ..................................................................................................75
Slack Computation........................................................................................... 75
Big M ................................................................................................................. 76
No optima .......................................................................................................... 76
Export LP ......................................................................................................... 77
vii
9 Browse and Export Results .................................................................78
12 Appendix .............................................................................................85
12.1 How to get the decomposition of the Malmquist Index in the literature
MaxDEA? ......................................................................................................... 91
viii
1 Overview
MaxDEA Pro is an easy-to-use but powerful and professional DEA software. It has the
most extensive range of the up-to-date DEA models.
1
DEA linear programming can be exported to text files with 3 formats: matrix
format (*.txt), lp format (*.lp) and mps format (*.mps).
Most important of all, MaxDEA Pro provides all the possible combinations of
up-to-date DEA models. To use a combination of multiple DEA models, just
choose all the relevant options. For example, Network-Malmquist model with
weakly disposable bad outputs can be achieved by choosing the settings for
Network, Undesirable outputs, Weak disposability and Malmquist, at the same
time.
2
DEA Models available in MaxDEA Pro
MaxDEA Pro has the most comprehensive DEA models and all their possible
combinations.
1) Distance to measure efficiency
Radial
Maximum distance to frontier:Slack based measure (SBM)
Modified SBM
Minimum distance to weak efficient frontier
Minimum distance to strong efficient frontier
Directional distance function,
Range directional model (RDM)
Hybrid: mixture of radial and SBM measure (EBM)
Hybrid: mixture of radial and non-radial fields
Cost /Revenue / Profit / Revenue-cost ratio
2) Orientation to measure efficiency
Input-oriented
Output-oriented
Non-oriented
Input-oriented (modified)
Output-oriented (modified)
Non-oriented (input-prioritized)
Non-oriented (output-prioritized)
Non-oriented (generalized priority)
3) RTS to measure efficiency
Constance returns to scale (CRS)
Variable returns to scale (VRS)
Non-increasing returns to scale (NIRS)
Non-decreasing returns to scale (NDRS)
3
Generalized returns to scale (GRS)
4) FDH model
5) Window model
6) Malmquist model
Adjacent Malmquist
Fixed Malmquist
Global Malmquist
Sequential Malmquist
Window-Malmquist (Adjacent)
Window-Malmquist (Fixed)
7) Network model
8) Parallel model
9) Dynamic model
10) Context-dependent model
11) Super-efficiency model
12) Cross efficiency model.
13) Undesirable output model
14) Nondiscretionary input/output model (non-controllable model, measure
specific model)
15) Bounded input/output model
16) Preference (weighted) model
17) Restricted projection model
18) Weak disposability model
19) Restricted multiplier model (assurance region model, trade-offs between
inputs and outputs)
20) Cluster model
21) Customized reference ret model
Variable-benchmark model
Fixed-benchmark model
Minimum efficiency model
4
22) Bootstrap
Bootstrap of DEA Score
Bootstrap of Malmquist Index
5
2 A Quick Guide
System Requirements
MaxDEA runs under Windows systems in any language, including:
Windows XP
Windows 2003
Windows Vista
Windows 7
Windows 8
MaxDEA is developed with VBA for Access, so Microsoft Access is required.
If the program file (MaxDEA.mdb) cannot open, it indicates that Microsoft Office
Access is not installed in your computer, and you must install MS Access 2003 or
higher version (professional edition or runtime edition) first.
Access 2013 Runtime can be downloaded free at Microsoft website:
http://www.microsoft.com/en-us/download/details.aspx?id=39358
(32-bit & 64-bit)
Access 2010 Runtime can be downloaded free at Microsoft website:
http://www.microsoft.com/en-us/download/details.aspx?id=10910
(32-bit & 64-bit)
Access 2007 Runtime can be downloaded free at Microsoft website:
http://www.microsoft.com/en-us/download/details.aspx?id=4438
(32-bit)
If your Office is 32-bit, please download the 32-bit MaxDEA; if your Office is
64-bit, please download the 64-bit MaxDEA at www.maxdea.cn.
6
Prepare Data
MaxDEA Pro can import data from Excel, Access, dBase and comma delimited
text files. Preparing your data in Excel is the most convenient way.
Data format
The first and only the first row must contain field (column, variable) names. The
field names can be anything you want, but there must not contain spaces left to
the column names. For example, column names such as “ Input1” are not
permitted.
Table 2-1(A) An example of correct data format
Company Capital Labor Product
A 4323 875 93608
B 2295 469 225559
C 6379 1286 327068
D 6644 1339 201354
E 1436 297 188926
F 6281 1266 413738
G 7459 1502 114022
7
DMU Name
DMU Name is the identifier for each DMU, so it must be unique. The name for
DMUs can be characters or numbers. The DEA results are sorted by DMU name,
so if you expect the results to be displayed by the original order of the dataset,
your dataset should be sorted by DMU name. If you use numbers as DMU name,
such as 1,2,…,100, MaxDEA will automatically format the numbers by adding
leading zeros, like 001, 002,…,100, so that the results are displayed in numerical
order. The number of leading zeros added depends on the length of the maximum
number.
Panel Data
For panel data, there must be an additional column indicating the time of the data.
DMU Name must be unique within each period. Panel data are prepared for
Malmquist, Window, and Dynamic models. If you try to run a cross-sectional
DEA model using panel data, there will be an error message indicating that the
DMU names are not unique. Panel data can be sorted by DMU name or by period,
but not necessary.
Panel data for Malmquist model and Window model are not necessary to be
balanced, i.e., missing values at some periods are permitted. But panel data for
Dynamic model must be balanced.
8
Table 2-2(B) A second example of correct panel data format (balanced)
Period Company Capital Labor Product
1 A 4323 875 93608
2 B 1436 297 188926
1 B 2295 469 225559
3 A 7459 1502 114022
3 B 4464 903 212444
1 C 6379 1286 327068
2 A 6644 1339 201354
2 C 6281 1266 413738
3 C 4524 915 462677
Negative Data
MaxDEA Pro permits negative values in inputs or outputs. The default method of
dealing with negative values is the variant radial measure (VRM) proposed by
Cheng, Zervopoulos, and Qian (2013). This method of dealing with negative
9
values in the VRM is also applied to other types of distances, such as the
non-radial (SBM) model.
After data is imported, the “Data Define” window will open automatically.
Defining data is to tell MaxDEA Pro which column is the DMU name, which
columns are inputs, and which columns are outputs. At least the fields for DMU
Name, Inputs and Outputs must be defined.
Missing values are not permitted in MaxDEA. If “DMU Name” has missing
values, there will be an error message “There is at least one DMU whose ‘DMU
Name’ is empty” or “Invalid use of Null”. If other fields (such as Input, Outputs,
Period and Cluster) have missing values, there will be an error message “... must
be numeric” or “... must be integer”.
Table 2-3 Missing values not permitted
DMU Input1 Input2 Output
A 4323 875 93608
B 2295 469 225559
C 6379 327068
D 6644 1339 201354
1436 297 188926
F 6281 1266 413738
G 7459 1502 114022
If you want to eliminate some inputs or outputs from the model, just deactivate
10
them. Only those active inputs/outputs will be included in your DEA model.
The imported data and their definitions are permanently saved, so after MaxDEA
Pro is closed, they will not be lost.
Run Model
There are two types of DEA models. One is called multiplier model, i.e., the
primal model, and the other is called envelopment model, i.e., the dual model.
The menu “Run Envelopment Model” is for envelopment models, “Run
Multiplier Model” is for multiplier models, and “Express to Basic Models” is an
express way to basic models (CCR and BCC), which is easy to use for new DEA
leaners.
The settings (options) for the DEA model are permanently saved, so after
MaxDEA Pro is closed, they will not be lost. If you want to save a backup of your
DEA model, just copy the “MaxDEA X.mdb” file and rename it as you want, such
as CCR.mdb, BCC.mdb.
11
3 Envelopment DEA model
Distance (method of measuring efficiency)
Including radial measure, non-radial measure, hybrid measure and financial measure
Radial
It measures the necessary proportional improvements of relevant factors (inputs /
outputs) for the evaluated DMU to reach the frontier.
Ref. to (Charnes, Cooper, & Rhodes, 1978; Banker, Charnes, & Cooper, 1984)
for CCR model and BCC model.
Maximum distance to frontier: SBM
The SBM model maximizes the average improvements of relevant factors (inputs
/ outputs) for the evaluate DMU to reach the frontier(Tone, 2001 ). The target
(projected point) determined by this method is a strong efficient point on the
frontier which is the farthest to the evaluated DMU. This method seems
unreasonable in that for an inefficient unit much more efforts may be needed to
reach the much farther target.
12
Modified SBM
A modified SBM model that can deal with negative data (Sharp, Meng, & Liu,
2007).
Minimum distance to weak efficient frontier
A closer target means less effort for an inefficient unit to reach the efficient frontier.
The minimum distance to weak efficient frontier finds a closest target located on
the weak efficient frontier.
Ref. to (Charnes, Roussea, & Semple, 1996; Briec, 1999)
Minimum distance to strong efficient frontier
The minimum distance to strong efficient frontier determines a closest target
located on the strong efficient frontier.
Ref. to (Aparicio, Ruiz, & Sirvent, 2007; Jahanshahloo, Vakili, & Zarepisheh,
2012)
Directional Distance Function
Directional distance function is a generalized form of radial model (Chambers,
Chung, & Färe, 1996; Chung, Färe, & Grosskopf, 1997). MaxDEA Pro provides
a further generalized form of the directional distance function. Here, α, β and γ
are not necessarily have same values, i.e. the proportion of input decrease is not
necessarily equal to that of output increase, and good output increase is not
13
necessary equal to bad output decrease. But, you can set α=β=γ.
The directional distance function model in MaxDEA Pro is expressed as
For DMU0
m
1 m1 wi g xio / xio
i 1
Efficiency score = min sd su
1 od 1
sd wd g ydo / ydo ou
d 1
1
su w g
u 1
u yuo / yuo
s.t. X g x xo
Y d g dy yod
Y u g uy you
u
0
Note that g y .
If the bad outputs are set weak disposable, the directional distance function model
need combine with the “Weak disposability” model
Y u g uy you
.
14
m sd su
wi m,
i 1
wd sd ,
d 1
w
u 1
u su
;
od + ou = 1.
In MaxDEA, the user can set the weights freely, for example, suppose there are 2
inputs and 2 outputs in the model, you can set weight for input1 = 1, weight for
input1 = 2, weight for output1 = 1, weight for output2 = 2, od =1, ou=1, and
MaxDEA will calculate the weights automatically.
This method to compute the efficiency score is defined by the authors of
MaxDEA Pro, with the following features:
1) The efficiency scores computed with this method are independent of the
length of the direction vector, although the values of α and β are dependent of the
length of the direction vector. For example, the computed efficiency scores are the
same if the direction vector (1, 1, 1, 1) or (2, 2, 2, 2) is used.
2) It keeps consistent with the measure used in radial model. If the directional
distance function model defined is equivalent to radial model, i.e., using x0 and y0
as the direction vector, the efficiency scores are the same.
3) The inputs and outputs can be attached with user-defined weights to indicate
their relative importance.
4) Overall weights of good outputs and bad outputs can be defined.
If you want to compute the efficiency scores using your own formula, your can
do it using the values of α, β and γ provided in the results.
There are six types of direction vectors for the directional distance function in
MaxDEA.
15
1) The values of the evaluated DMU, i.e., using x0 and y0 as the direction
vector;
2) The mean of all DMUs;
3) Vector (1, 1, 1…);
4) Gap between the evaluated value and best value, RDM (Portela, Thanassoulis,
& Simpson, 2004)
5) A self-defined vector same to all DMUs;
6) Self-defined vectors different to different DMUs.
The directional distance function model will be equivalent to the radial model with
the first type of direction vector, i.e. gx = x0, and gy = y0.
Note: Importance differences between MaxDEA Pro 5 and 6
In MaxDEA Pro 5, the users cannot assign negative values to the direction vector.
The values of the direction vector for bad outputs will be changed into negative
values automatically.
To make the direction vector more flexible, in MaxDEA Pro 6, the users are
permitted to assign both positive and negative values to the direction vector. But,
the users need to note that
With the non-user-defined direction vector types, i.e.,
1) Value of the evaluted DMU,
2) Mean of All DMUs,
16
3) Vector (1, 1, ……, 1),
4) Gap between the evaluated value and best value (RDM: Portelal et al,
2004), and
7) Direction Vector Scanning,
MaxDEA Pro 6 will automatically change positive values into negative values for
bad outputs, just as MaxDEA Pro 5 does.
With the user-defined direction vector types, i.e.,
5) Customized (same for all DMUs), and
6) Customized (DMU specific),
MaxDEA Pro 6 will use the user-defined (positive or negative) direction vector
values for bad outputs, which means that if the user defines positive values for
bad outputs, MaxDEA Pro 6 will not change them into negative values.
See Weak Disposability Model
Hybrid distance: EBM
EBM is a hybrid distance prosed by Tone and Tsutsui (2010b).
See 12.4 An epsilon-based measure of efficiency in DEA - An alternative method
for the affinity index.
Hybrid distance(Radial and SBM fields)
This is another type of hybrid model. It is a mixture of radial model and Non-radial
model, and it is a generalized form of radial and non-radial models(Cooper,
Seiford, & Tone, 2007). If Hybrid model is selected, you should define the
distances of inputs/outputs. Click the “Define” button right to the Hybrid check
box, the distance definition form will open.
If all the inputs/outputs are set to be radial in Hybrid model, it is equivalent to
radial model, and if all set to be non-radial, it is equivalent to SBM model.
Cost, Revenue, Profit, and Revenue/Cost Ratio Models
These models deal with financing aspects of DEA models in the case that prices of
inputs/outputs are known. It is related to the topic of allocative efficiency(Cooper,
Seiford, & Tone, 2007).
There are two types models, and MaxDEA names them as “Type I “and “Type II”
17
respectively. The difference between type I and type II is that type I models use the
original inputs/outputs values in constraints, while type II models use cost/revenue
values of inputs/outputs in constraints. (Type I is traditional and commonly used.)
To run the above models, price information must be set first using the “Define”
button on the right side. Cost model needs input prices, Revenue model needs
output prices, and Profit and Revenue/Cost Ratio models need both input and
output prices. The price fields must be kept “Not define” at the stage of “Define
Data”.
Note that efficiency score of profit model might be negative.
Orientation
Five types of RTS (returns to scale) are available with MaxDEA: constant(CRS),
variable(VRS), non-increasing(NIRS), non-decreasing(NDRS) and
generalized(GRS). If GRS is selected, you should also set the lower bound (L) and
upper bound (U) of Σλ. GRS is a generalized form of the other four types of RTS,
with the following relationship (Cooper, Seiford, & Tone, 2007):
Table 3-1 Relationship between GRS and other RTS
GRS
L U
CRS 0
VRS 1 1
NIRS 0 1
NDRS 1
The option “Scale Efficiency or Scale Effect (CRS Score / VRS Score)” is a special
case in which both CRS and VRS models will be solved, and the scale efficiency or
scale effect will be computed.
In Radial model, Scale Efficiency = CRS efficiency / VRS efficiency.
19
In non-radial models, Scale Effect = CRS efficiency / VRS efficiency.
RTS Estimation is also provided with this option.
For Malmquist model, the relationship between scale factors are as follows
Scale Effect on Malmquist Index = Malmquist Index (CRS) / Malmquist
Index (VRS)
Scale Effect on Efficiency Change = Efficiency Change (CRS) / Efficiency
Change (VRS)
In Radial model, Scale Effect on Efficiency Change is “Scale Efficiency
Change”.
Scale Effect on Technological Change = Technological Change (CRS) /
Technological Change (VRS)
Scale Effect on Malmquist Index = Scale Effect on Efficiency Change *
Scale Effect on Technological Change
Note that when “Scale Efficiency or Scale Effect (CRS Score / VRS Score)” is selected,
the result window only reports the scale efficiency scores or scale effect scores, but all
the detailed results under CRS and VRS can be obtained through the Menu “Tools –
Browse Results”.
CCR model
See Distance-Radial
BCC model
See Distance-Radial
SBM model
See Distance- Maximum distance to frontier: SBM
20
Modified SBM
FDH Model
The purpose of Free Disposal Hull (FDH) model is to ensure that efficiency
evaluation is based on only actually observed performances. The input-oriented
FDH model can be expressed using the following mixed integer LP,
min
st x0 X s 0
Y y0 s 0
s , s 0
21
e 1, {0,1}
Note that the FDH model is located under “Returns to scale (RTS)”.
Ref. to (Deprins, Simar, & Tulkens, 1984; Tulkens, 1993)
Advanced models
Super-efficiency model
The difference between Super-efficiency model and standard efficiency model is
that in super models the DMU0 (the DMU evaluated) is eliminated from the
reference set (indicated by j≠0 in the LP).
Ref. to (Andersen & Petersen, 1993; Tone, 2002).
The Super-efficiency score can be greater than 1. In some cases, the LP for some
DMUs will be infeasible(Seiford & Zhu, 1999). In such cases, you can decide
whether the program returns 1 as the score through the option “No optima”.
22
Range Directional model (RDM)
Window Model
The window model deals with panel data. Window width must be set for window
models.
Panel data for Malmquist model and Window model are not necessary to be
balnced, i.e., missing values at some periods are permitted.
If you want to get the DEA results for each year independently, you should set
window width to 1. It is equivalent to dividing the panel data into many datasets
(each dataset contains one year’s data) and analyzing each dataset one by one.
1
min
wg wb
1
wg wb
st (1 ) x0 X s 0
Y g (1 ) y 0g s g 0
Y b (1 ) y b0 s b 0
Y b (1 ) y b0 0
,
the weak disposability model should be applied at the same time.
The SBM model with undesirable outputs is expressed as
m
1 1
m s
i / xio
min i 1
s
1 1s sr / yro
r 1
st x0 X s 0
Y g y 0g s g 0
Y b y b0 s b 0
, s , s 0
st x0s X s s s 0
x0w X w 0
Y s y 0s s s 0
Y w y 0w 0
27
disposability.
Note: Weak Disposability is not applicable in non-radial (SBM) model.
w
i 1
i m
s
w
r 1
r s
Context–dependent model
Ref. to Seiford and Zhu (2003)
Cluster Model
Cluster Model deals with the situation that the DMUs are categorized according to
their characteristics. There are 7 types of cluster models according to the relationship
between the clusters evaluated and the clusters as benchmarks.
28
Data format for Cluster Model
For cluster data, there must be an additional column indicating the clusters. Each
DMU belongs to only one cluster, and a DMU should not belong to different
clusters.
The following are examples of cluster data:
Table 3-2(A) Example data with clusters
DMU Cluster Input1 Input2 Output1 Output2
A 1 4323 875 93608 187196
B 1 2295 469 225559 451099
C 1 6379 1286 327068 654116
D 1 6644 1339 201354 402688
E 1 1436 297 188926 377833
F 1 6281 1266 413738 827456
G 2 7459 1502 114022 228024
H 2 4464 903 212444 424867
I 2 4524 915 462677 925334
Cluster data needn’t be balanced. The number of DMUs in each cluster can be
different.
Table 3-2(B) Example panel data with clusters
Period DMU Cluster Input1 Input2 Output1 Output2
1 A 1 4323 875 93608 187196
1 B 1 6644 1339 201354 402688
1 C 1 7459 1502 114022 228024
1 D 1 2295 469 225559 451099
1 E 1 1436 297 188926 377833
1 F 1 4464 903 212444 424867
1 G 2 6379 1286 327068 654116
1 H 2 6281 1266 413738 827456
1 I 2 4524 915 462677 925334
2 A 1 4341 944 93656 187279
2 B 1 6689 1374 201442 402767
2 C 1 7559 1511 114107 228101
2 D 1 2362 530 225604 451107
2 E 1 1513 381 188937 377886
2 F 1 4544 937 212446 424869
2 G 2 6406 1361 327102 654188
29
2 H 2 6369 1327 413797 827479
2 I 2 4563 942 462739 925393
Note: For panel data, the same DMU in different periods must belong to the
same cluster. For example, DMU A belongs to cluster 1 in period 1, and it must
belong to cluster 1 in period 2 and other periods.
The cluster field must be integer numbers, such as
1, 2, 3……
But they needn’t to be continuous. The following time series are permitted:
1, 2, 5, 8……
2001, 2005, 2009……
200101, 200102, 200302……
Self-benchmarking
In self-benchmarking model, each DMU is evaluated within the cluster it
belongs to. In effect it is a batch mode of evaluating the DMUs many times,
analyzing one cluster each time.
For example, if all the DMUs are categorized into 3 clusters, the results of
self-benchmarking model are the same as the following models. Firstly run the
model using the data of cluster 1 only, secondly cluster 2 only, and lastly cluster
3 only.
Self-benchmarking
Cluster Evaluated Cluster as Benchmark
Cluster 3 Cluster 3
Cluster 2 Cluster 2
Cluster 1 Cluster 1
30
Cross-benchmarking
In cross-benchmarking model, each DMU is evaluated successively with every
cluster as the reference set separately.
For example, if all the DMUs are categorized into 3 clusters,
1) Firstly all DMUs are evaluated using the DMUs in cluster 1 as reference set,
2) Secondly all DMUs are evaluated using the DMUs in cluster 2 as reference
set, and
3) Lastly all DMUs are evaluated using the DMUs in cluster 3 as reference set.
Cross-benchmarking
Cluster Evaluated Cluster as Benchmark
Cluster 3 Cluster 3
Cluster 2 Cluster 2
Cluster 1 Cluster 1
Downward-benchmarking
In downward-benchmarking, the DMUs in a cluster are evaluated with the
DMUs in its own cluster and those in the cluster with a lower cluster number
(cluster ID) as the reference set.
For example, if all the DMUs are categorized into 3 clusters: clusters 1, 2 and 3,
1) the DMUs in cluster 1 are evaluated with the DMUs in cluster 1 only as the
reference set;
2) the DMUs in cluster 2 are evaluated with the DMUs in cluster 1 and cluster 2
as the reference set; and
3) the DMUs in cluster 3 are evaluated with the DMUs in cluster 1, cluster 2,
and cluster 3 (all the DMUs) as the reference set.
31
Downward -benchmarking
Cluster Evaluated Cluster as Benchmark
Cluster 3 Cluster 3
Cluster 2 Cluster 2
Cluster 1 Cluster 1
Upward-benchmarking
Contrary to downward-benchmarking, in upward-benchmarking, the DMUs in a
cluster are evaluated with the DMUs in its own cluster and those in the cluster
with a upper cluster number (cluster ID) as the reference set.
For example, if all the DMUs are categorized into 3 clusters: clusters 1, 2 and 3,
1) the DMUs in cluster 1 are evaluated with the DMUs in cluster 1, cluster 2,
and cluster 3 (all the DMUs) as the reference set;
2) the DMUs in cluster 2 are evaluated with the DMUs in cluster 2 and cluster 3
as the reference set; and
3) the DMUs in cluster 3 are evaluated with the DMUs in cluster 3 only as the
reference set.
32
Upward-benchmarking
Cluster Evaluated Cluster as Benchmark
Cluster 3 Cluster 3
Cluster 2 Cluster 2
Cluster 1 Cluster 1
Lower-adjacent-benchmarking
In lower-adjacent-benchmarking, the DMUs in a cluster are evaluated with the
DMUs in the cluster with a lower cluster number (cluster ID) as the reference
set.
For example, if all the DMUs are categorized into 3 clusters: clusters 1, 2 and 3,
1) the DMUs in cluster 1 are not evaluated;
2) the DMUs in cluster 2 are evaluated with the DMUs in cluster 1 as the
reference set; and
3) the DMUs in cluster 3 are evaluated with the DMUs in and cluster 2 as the
reference set.
Downward -benchmarking
Cluster Evaluated Cluster as Benchmark
Cluster 3 Cluster 3
Cluster 2 Cluster 2
Cluster 1 Cluster 1
33
Upper-adjacent-benchmarking
Contrary to lower-adjacent-benchmarking, in upper-adjacent-benchmarking, the
DMUs in a cluster are evaluated with the DMUs in the cluster with a upper
cluster number (cluster ID) as the reference set.
For example, if all the DMUs are categorized into 3 clusters: clusters 1, 2 and 3,
1) the DMUs in cluster 1 are evaluated with the DMUs in cluster 2 as the
reference set;
2) the DMUs in cluster 2 are evaluated with the DMUs in cluster 3 as the
reference set; and
3) the DMUs in cluster 3 are not evaluated.
Upward-benchmarking
Cluster Evaluated Cluster as Benchmark
Cluster 3 Cluster 3
Cluster 2 Cluster 2
Cluster 1 Cluster 1
Window-benchmarking
In window-benchmarking, the DMUs in a cluster are evaluated with the DMUs
in a cluster-window including its own cluster as the reference set. By default
(offset =0), the cluster-window is composed of its own cluster and a number of
adjacent clusters. The number of the adjacent clusters included in the window is
width – 1.
Suppose all the DMUs are categorized into 5 clusters: clusters 1, 2, 3, 4 and 5.
If window width = 2, offset = 0,
1) the DMUs in cluster 1 are evaluated with the DMUs in cluster 1 only as the
reference set, and its referred window is not full width;
2) the DMUs in cluster 2 are evaluated with the DMUs in cluster 1 and cluster 2
as the reference set;
34
3) the DMUs in cluster 3 are evaluated with the DMUs in cluster 2 and cluster 3
as the reference set;
4) the DMUs in cluster 4 are evaluated with the DMUs in cluster 3 and cluster 4
as the reference set;
5) the DMUs in cluster 5 are evaluated with the DMUs in cluster 4 and cluster 5
as the reference set.
Downward –benchmarking
Width = 2, Offset = 0
Cluster Evaluated Cluster as Benchmark
Cluster 5 Cluster 5
Cluster 4 Cluster 4
Cluster 3 Cluster 3
Cluster 2 Cluster 2
Cluster 1 Cluster 1
35
Downward –benchmarking
Width = 2, Offset = 1
Cluster Evaluated Cluster as Benchmark
Cluster 5 Cluster 5
Cluster 4 Cluster 4
Cluster 3 Cluster 3
Cluster 2 Cluster 2
Cluster 1 Cluster 1
36
Downward –benchmarking
Width = 3, Offset = 0
Cluster Evaluated Cluster as Benchmark
Cluster 5 Cluster 5
Cluster 4 Cluster 4
Cluster 3 Cluster 3
Cluster 2 Cluster 2
Cluster 1 Cluster 1
37
Downward –benchmarking
Width = 3, Offset = 1
Cluster Evaluated Cluster as Benchmark
Cluster 5 Cluster 5
Cluster 4 Cluster 4
Cluster 3 Cluster 3
Cluster 2 Cluster 2
Cluster 1 Cluster 1
38
Downward –benchmarking
Width = 3, Offset = 2
Cluster Evaluated Cluster as Benchmark
Cluster 5 Cluster 5
Cluster 4 Cluster 4
Cluster 3 Cluster 3
Cluster 2 Cluster 2
Cluster 1 Cluster 1
Note: in the stage of “Define Data”, the field indicating cluster ID should be
defined as “Cluster”.
Variable-benchmark Model
See Customized Benchmarking
Customized Benchmarking
In customized benchmarking model, you can customize the reference DMU set
for calculating efficiency scores. Take the input-oriented CCR model as an
example,
min
st x0 X r s 0
Y r y0 s 0
, s , s 0
The superscript r indicates the DMUs included in the customized reference DMU
set R.
39
The above model is also called variable-benchmark model (Zhu, 2009),
because any of the DMUs in R might be a benchmark, but it is not necessary for
all the DMUs in R to be the benchmarks.
To set a variable-benchmark model, open the menu “Run Envelopment Model”,
check the option “Variable-benchmark”, and click the “Define” button to select
variable benchmarks.
40
An example of Network DEA Model (net-shape)
Input 1
Node 1 Output 1
Intermediate 1 Intermediate 2
Input 2 Input 3
Output 2 Output 3
41
An example of Network DEA Model (chain-shape)
Input 1
Node 1
Intermediate 1
Input 2 Input 3
Output 2 Output 3
42
An example of Network DEA Model (chain-shape)
Input
Node 1
Intermediate 1
Node 2
Intermediate 2
Node 3
Output
43
MaxDEA Pro is a powerful tool for Network DEA. It provides the most extensive
options for Network DEA models.
There are 2 options for Network DEA. One is the type of intermediates. There are 4
types of intermediates:
1) Free
2) Fixed
3) Non-increasing
4) Non-decreasing.
If you want the intermediates to be different types, you should choose the “Free” type
and at the same time apply the “Bounded Inputs/Outputs” model. In the Bounded
settings, you may set the lower and upper bounds for the intermediates.
See Bounded Inputs/outputs Model.
The other option is “Use Identical Lambda”. It means that the lambdas for all nodes
are the same.
To run a Network DEA model, you should
1) At the stage of “Define Data”, define the direct inputs/outputs just as usual, but
define indirect inputs/outputs (intermediate inputs/outputs) as “Intermediate”;
2) At the stage of “Run Model”, check the box “Network”, and
a) Decide whether different nodes use same or different intensity vector λ
through the check box “Use Identical Lambda”;
b) Decide whether the intermediate value is free or fixed through the check box
“Fixed Intermediate Value”;
c) Define the Sub-processes of the Network through the “Define” button;
i. Firstly define the nodes for the network, including node name and node
weight;
ii. Secondly, assign the direct inputs, outputs and intermediate to
corresponding nodes. Each input or output must be assigned to one and
only one node, and each intermediate must be assigned to one node as
input and another node as output.
d) Run the model.
In MaxDEA Pro, it is very convenient to compare the results between the traditional
model (so-called “Black box” model) and the Network model. Just uncheck
“Network” with other options unchanged, and it will be the traditional model.
44
There are three tables for the results of the Network models: the first is a summary
just as usual, the second is overall efficiency and slacks and projections, and the last is
efficiency scores of the nodes and the intensity vectors.
Dynamic Model
Dynamic model deals with panel data, and take into consideration the links between
periods. Panel data for Dynamic model must be balanced, i.e., each DMU must have
one and only one observation in each period.
Table 3-3(A) is a panel dataset with one input, one output and one intermediate with
three periods for dynamic model. Note that in MaxDEA Pro, the intermediate in
period 1 is treated as the link between period 1 and 2, the intermediate in period 2 as
45
the link between period 2 and 3, and the intermediate in period 3 is not included in the
dynamic model.
Essentially, dynamic model is a special case of network model, so all dynamic models
can be equivalently achieved through network model. To do so, the panel data (long
data) must be transformed into wide data for network model. Table 3-3(B) is the
equivalent dataset to Table 3-3(A) for the same dynamic model achieved though
network model, which treats the periods as the nodes.
With MaxDEA, more complex dynamic model , such as dynamic model with different
intermediates, inputs and outputs in each period, can be developed through network
model.
46
Table 3-3(A) Panel data for dynamic model (example)
Period DMU Input Output Intermediate
1 A 4323 93608 875
1 B 2295 225559 469
1 C 6379 327068 1286
2 A 6644 201354 1339
2 B 1436 188926 297
2 C 6281 413738 1266
3 A 7459 114022 1502
3 B 4464 212444 903
3 C 4524 462677 915
Intermediate (1)
Intermediate (2)
Table 3-3(B) The equivalent dataset to table 3-3(A) for the same dynamic model
achieved though network model
47
Results for Envelopment Model
Score
Please note that MaxDEA uses θ as efficiency score in input-oriented model, and
uses 1/φ (the reciprocal of φ) as efficiency score in output-oriented model.
For non-oriented models,
1 Input Inefficiency
Efficiency Score = 1 Output Inefficiency
The score may be technical efficiency, pure technical efficiency, cost, revenue, or
profit efficiency. Its meaning depends on the specific model you develop.
Scale Efficiency
In Radial model, Scale Efficiency = CRS efficiency / VRS efficiency.
In non-radial models, Scale Effect = CRS efficiency / VRS efficiency.
Allocative Efficiency
Allocative Efficiency is not provide directly in MaxDEA. It can be calculated as
follows
Allocative Efficiency = Cost (or Revenue or Profit) Efficiency / Technical
Efficiency
Input Inefficiency
= Input Radial Inefficiency + Input Non-radial Inefficiency
48
Input Radial Inefficiency (for Hybrid model)
mm1 1
ss1 ( 1)
49
RTS Estimation
See Returns to scale (RTS)
Original
The original value of inputs/outputs.
Proportionate Movement (Radial Movement)
Only available for radial, directional distance and hybrid models. It is the radial
part of improvement of inputs/outputs, the proportional decrease of inputs or the
proportional increase of outputs. Positive values mean increase, and negative
values mean decrease.
Slack Movement
It’s absolute value is the s- (input slack) or s+ (output slack) in the LP equations.
Positive values indicate increase, and negative values indicate decrease.
Projection
It is the efficient target.
For radial models,
Strong Efficient Projection =Original + Radial Movement + Slack Movement
Weak Efficient Projection = Original + Radial Movement
If strong efficient projection is wanted, it is suggested that “Two Stage” method be
used to compute the slacks (in “Options – Slack Computation”), because the
slacks from the first stage may be incomplete.
For SBM models, Projection =Original + Slack Movement
α, β and γ
The values of α, β and γ in the solution.
RHS Lower Bound and RHS Upper Bound
See Dual Prices and Sensitivity Analysis
Dual Prices and Sensitivity Analysis
The dual price (also named as shadow price, dual value) of an input (output)
indicate that the objective function will change with the value of the dual price if
the value of the right-hand side (RHS) for the input (output) is changed with 1
unit. Note that the sign indicates whether the objective function will increase or
50
decrease.
The dual price remains constant as long as the RHS for the input (output) stays
within the lower/upper bounds. If there is no dual price, or no lower/upper limit,
then these values are positive or negative infinity1.
The absolute values of the dual prices correspond to the values of the weights of
inputs (outputs) in the dual model (Multiplier Model). Note that there often exist
multiple optimal solutions for multiplier model, so the absolute values of the dual
prices got from envelopment model might not be equal to the values of the
weights got from corresponding multiplier model.
1
1E+30 indicates +finite, and -1E+30 indicates -finite.
51
4 Multiplier DEA Model
Scale Elasticity
A quantitative estimate of RTS.
Ref. to Banker and Thrall (1992).
52
and type II respectively. The restricted multiplier model is to add constraints on the ratios
or proportions of weights for inputs/outputs. It is also called Assurance Region model in
literature.
Type I restricted multiplier model is to add constraints on the ratios of weights for
inputs/outputs, such as
2
L1 U1
1
3
L2 U2
1
while Type II restricted multiplier model is to add constraints on the proportion of a (some
or all) weighted (virtual) input/output over the total weighted (virtual) inputs/outputs, for
example, suppose there are two inputs and two outputs, we can restrict their virtual
proportions as follows,
1 x1
0.1 0.3
1 x1 2 x2 ,
1 y1
0.1 0.3
1 y1 2 y2 ,
2 y2
0.5 0.8
1 y1 2 y2 .
Please note that in type II restricted multiplier model, (1) the sum of the lower bounds of
weight proportions must be less than or equal to 1, and (2) the sum of the upper bounds
must be greater than or equal to 1 in case that all the weight proportions of inputs/outputs
are restricted such as the above example, in which all the weight proportions of outputs are
restricted.
To run the restricted multiplier model, set the restrictions first by a click on the “Define”
button on the right side.
If you want to eliminate one piece of restriction, (1)click the arrow on the left, and (2)
press the “Delete” key or use the “delete record button” on the toolbar. You can also
53
disable the restriction through the active check box.
Parallel model
The parallel DEA model takes the operation of individual components of the system into
account in calculating the efficiency.
The data for parallel system must have two columns indicating the DMU and Sub-DMU
respectively.
Ref. to Kao (2009)
Table 4-1 An example for parallel data format
DMU Sub-DMU Capital Labor Product
A a 4323 875 93608
A b 2295 469 225559
A c 6379 1286 327068
B d 6644 1339 201354
B e 1436 297 188926
C f 6281 1266 413738
C g 7459 1502 114022
Variable-benchmark Model
See Customized Benchmarking
Fixed-benchmark Model
See Customized Benchmarking
54
Minimum Efficiency Model
See Customized Benchmarking
Customized Benchmarking
In customized benchmarking model, you can customize the reference DMU set for
calculating efficiency scores. Take the input-oriented CCR model as an example,
max ' y0
st ' x0 1
'Y r 'X r 0
, 0 (or , )
The superscript r indicates the DMUs included in the customized reference DMU set.
The above model is also called variable-benchmark model (Zhu, 2009), because any of
the DMUs in R might be the benchmark for a DMU evaluation, but it is not necessary for
all the DMUs in R to be the benchmarks.
If you want one or some DMUs in R to be fixed benchmarks, you can develop a
fixed-benchmark model (Zhu, 2009) as follows, also take the input-oriented CCR
model as an example,
max ' y0
st ' x0 1
'Y f 'X f 0
'Y n 'X n 0
, 0 (or , )
55
max ' y0
,
to
min ' y0
,
The model will change from a maximum efficiency fixed-benchmark model to a
minimum efficiency fixed-benchmark model (Zhu, 2009). Such a change for
output-oriented models is vice versa.
Variable-benchmarking can be applied to both envelopment models and multiplier
models, but fixed-benchmarking can be applied to multiplier models only.
To set a variable-benchmark model, open the menu “Run Multiplier Model”, check the
option “Variable-benchmark”, and click the “Define” button to select variable
benchmarks.
To set a fixed-benchmark model, check the option “Fixed-benchmark”, choose
maximum efficiency or minimum efficiency, and click the “Define” button to select both
variable benchmarks and fixed benchmarks. The selection of variable benchmarks is
optional.
Note: Be cautious about setting the fixed benchmarks. If the fixed benchmarks are not
properly chosen, it will result in infeasible LP.
56
Results for Multiplier Model
Weight
The optimal weight (multiplier) of the inputs (outputs), i.e., the optimal solution of ν
and μ.
Weighted Value
The product of an input (output) and its multiplier, i.e., ν’x0 or μ’y0.
Coefficient Lower Bound and Coefficient Upper Bound
See Sensitivity of Objective Function
Sensitivity of Objective Function
It provides the lower and upper bounds of a coefficient in the objective without causing
any of the optimal values of the decision variables (weights of inputs and outputs) to
change.
Dual Solution
It provides the benchmarks with the values of lambda and the values of projection.
57
5 Productivity Analysis (Malmquist Model)
The Malmquist model deals with panel data. Panel data for Malmquist model and
Window model are not necessary to be balanced, i.e., missing values at some
periods are permitted. If the values for some DMUs in one or more periods are
missing, the Malmquist indices for the relevant periods will not be computed, but
the Malmquist indices for the periods that have complete data will be computed.
In addition, the DMUs with missing values may serve as benchmarks for other
DMUs at the periods that they have data.
The Malmquist productivity index evaluates the total factor productivity change of
a DMU between two periods. It is defined as the product of efficiency change
(catch-up) and technological change (frontier-shift). The efficiency change
reflects to what extent a DMU improves or worsens its efficiency, while
technological change reflects the change of the efficiency frontiers between two
58
periods.
MaxDEA Pro provides 6 types of Malmquist Indices.
Adjacent Malmquist
Mean of 2 TFP Indices
The traditional method of Malmquist index developed by Färe, Grosskopf,
Lindgren, and Roos (1992)
Single TFP Index
A method of Malmquist index that overcome the infeasibility problem proposed
by Cheng and Qian (2012)
Fixed Malmquist
A method of Malmquist index that uses a fixed period as reference proposed by
Berg, Forsund, and Jansen (1992)
Global Malmquist
A method of Malmquist index that uses all periods as reference proposed by
Pastor and Lovell (2005)
Sequential Malmquist
Mean of 2 TFP Indices
A method of Malmquist index that uses all previous periods as reference. Ref. to
(Tulkens & Vanden Eeckaut, 1995; Shestalova, 2003)
Single TFP Index
A method of Sequential Malmquist index that overcome the infeasibility problem
using the idea by Cheng and Qian (2012).
Window-Malmquist (adjacent)
Mean of 2 TFP Indices
A method of Malmquist index that uses period window as reference. Ref. to
(Sueyoshi & Aoki, 2001; Asmild, Paradi, Aggarwall, & Schaffnit, 2004)
Single TFP Index
A method of Window-Malmquist index that overcome the infeasibility problem
using the idea by Cheng and Qian (2012).
See Cluster Model, window-benchmarking for the meaning of Width and Offset.
59
Window-Malmquist (fixed)
A method of Window-Malmquist index that uses fixed period window as
reference. The idea is similar with that in Fixed Malmquist proposed by Berg,
Forsund, and Jansen (1992)
See Cluster Model, window-benchmarking for the meaning of Width and Offset.
61
Results For Malmquist Models
62
efficiency(1+1) is the efficiency in period 1 calculated using the input/output
values in period 2 as the reference set.
Efficiency_Fixed (t) for adjacent reference
It is the efficiency calculated with a user-defined fixed period as the reference set.
For example, if the reference period is 1, the Efficiency_Fixed (2) is the efficiency
in period 2 calculated using the input/output values in period 1 as the reference set.
Efficiency_Global (t) for global reference
It is the efficiency calculated with all periods as a whole as the reference set. For
example, the Efficiency_Global (2) is the efficiency in period 2 calculated using
the input/output values in period 1 and period 2 as the reference set.
Multiplicative measurement of productivity change
The multiplicative method measures the relative change of productivity. It is the
commonly used method.
For example,
Efficiency 2
Efficiency Change (1 TO 2)
Efficiency 1
2) Fixed reference:
Efficiency_Fixed t of the ' to ' period
Fixed Malmquist Index
Efficiency_Fixed t of the ' from ' period
63
Efficiency_Fixed 2
Fixed Malmquist Index (1 TO 2)
Efficiency_Fixed 1
3) Global reference:
Efficiency_Global t of the ' to ' period
Global Malmquist Index
Efficiency_Global t of the ' from ' period
Efficiency_Global 2
Global Malmquist Index (1 TO 2)
Efficiency_Global 1
Malmquist Index
Technological Change
Efficiency Change
Note: Malmquist Indices, Efficiency Change and Technological Change
with fixed or global reference are circular. For example,
Fixed MI (1 TO 2) * Fixed MI (2 TO 3) = Fixed MI (1 TO 3).
In addition, Global Malmquist model with non-CRS technology doesn’t
suffer from the infeasibility problem.
See 12.1 How to get the decomposition of the Malmquist Index in the literature
using MaxDEA Pro?
65
Additive measurement of productivity change
The additive method measures the absolute change of productivity.
For example,
Efficiency Change Efficiency t of the ' to ' period Efficiency t of the ' from ' period
2) Fixed reference:
Fixed Malmquist Index
Efficiency_Fixed t of the ' to ' period Efficiency_Fixed t of the ' from ' period
3) Global reference:
Global Malmquist Index Efficiency_Global t of the ' to ' period Efficiency_Global t of the ' from ' period
66
Fixed MI (1 TO 2) + Fixed MI (2 TO 3) = Fixed MI (1 TO 3).
In addition, Global Malmquist model with non-CRS technology doesn’t
suffer from the infeasibility problem.
67
7 Bootstrapping DEA Scores and Malmquist Indices
Method of Bootstrapping
MaxDEA Pro provides two types of bootstrap method: one is that proposed by
Simar and Wilson (1998), and the other is by Simar and Wilson (1999).
68
Original Score / (1+ Bootstrapped Bias)
The confidence intervals are calculated using the bias-adjusted scores. The scores
are adjusted by
Adjust Score = Original Score / (1+ Bootstrapped Bias).
Smoothed Bootstrap
Whether the smoothed bootstrapping is used.
Avoid Over-Smooth
It is useful when some DMUs have very small efficiency scores. It can avoid
negative bootstrapped scores.
Number of replications
Bootstrapping is a time-consuming analysis. The running time is approximately
the product of time for a single DEA or Malmquist model and the number of
replications.
Simplify Results
Some bootstrapping details are compressed.
69
Score(Original)
The original efficiency score.
Bias
The average (mean) bias of bootstrapped scores..
Mean
The mean of the bias-corrected bootstrapped scores.
Median
The median of the bias-corrected bootstrapped scores.
SD
70
Table 7-1: Result tables for bootstrapping DEA scores
Table Name Contents
Tbl_Results_Summary the summary of the DEA model
Tbl_Results of Envelopment Model the original bootstrapped scores
(Envelopment model or dual solution of
Multiplier model )
Tbl_Results_Weights the original bootstrapped scores
(Multiplier model or dual values of
Envelopment model )
71
Table 7-2: Result tables for bootstrapping Malmquist indices
(RTS=CRS, VRS, NIRS, NDRS, GRS)
Table Name Contents
Tbl_Results_Summary the summary of the DEA model
Tbl_Results of Malmquist Model the original bootstrapped Malmquist
indices
Tbl_Results of Malmquist Model_Adjusted the bias-corrected bootstrapped Malmquist
indices
72
2) RTS = Scale Efficiency or Scale Effect (CRS Score / VRS Score)
The result window “Bootstrap” reports the summary of the bootstrapping: the CIs
for
Scale Effect on Efficiency Change(SEEC),
Scale Effect on Technological Change(SETC) and
Scale Effect on Malmquist Index(SEMI).
All the detailed bootstrapped results, including the original and adjusted
bootstrapped Malmquist indices under CRS and VRS, the bias-corrected mean,
median, lower bound and upper bound for scale effect, can be obtained from the
result tables in MaxDEA Pro.
73
Table 7-3: Result tables for bootstrapping scale effect on Malmquist indices
(RTS= “Scale Efficiency or Scale Effect (CRS Score / VRS Score)”)
Table Name Contents
Tbl_Results_Summary the summary of the DEA model
Tbl_Results of Malmquist Model_CRS the original bootstrapped Malmquist
indices (RTS=CRS)
Tbl_Results of Malmquist Model_CRS_Adjusted the bias-corrected bootstrapped Malmquist
indices (RTS=CRS)
Tbl_Results of Bootstrap_UpperBound the upper bound of the CIs for scale effect
Tbl_Results of Bootstrap_CI the CIs for scale effect
74
8 Options
Result Decimals
It sets the decimals of results.
Slack Computation
You can choose a second stage calculation for slacks in radial models by using this
option.
Take CCR model as an example,
A two stage model carries out the LP with the following objective
min (es es )
,
75
2nd stage: max es es ,
while one stage model only carries out the 1st stage model.
Big M
The big M is used in the mixed integer linear programming of Minimum Distance
to Strong Efficient Frontier. There are three types of methods to set a big M: The
first type lets MaxDEA automatically find a proper Big M for each DMU; the
second lets MaxDEA automatically find a proper Big M for the 1st DMU and apply
it to all the DMUs; and the last lets the user set the value. The first type is
recommended.
No optima
Using this option, you can decide whether the score will be set to be 1 in case of no
optima. The problem of no optima often takes place in Super-efficiency,
Malmquist, and variable-benchmark and fixed-benchmark models.
Note: No optima may be infeasible, unbounded, degenerative or numerical
failure.
Weight Computation
Using the two stage method in Multiplier Model, the weights will be re-computed
in the second stage, with the objective of minimizing (or maximizing) the average
efficiency of other DMUs. This method is especially useful for computing cross
efficiencies since there usually exist multiple optimal solutions in multiplier
models.
76
Specify the value of epsilon
The epsilon in Multiplier Model is set to be 0 in default. Be cautious to change this
parameter. If it is not properly set, infeasible LP may occur.
Export LP
Linear programming equations of DEA models can be exported to text files with
three formats: matrix format (*.txt), mps format (*.mps) and lp format(*.lp). It is
very useful to help DEA users to develop their own DEA models using a general
LP software.
LP equations are the core of DEA models, exporting LP equations is also helpful to
verify the results of DEA models.
The user can choose to export LP for the first DMU only, or for all the DMUs, or
not to export LP. In addition to LP equations, value of the objective function and
values of all variables in the optimal solution are also exported. (See details in
MaxDEA Linear Programming Manual.)
77
9 Browse and Export Results
Browse results
After running your model, the results will show automatically in the result form.
You can also browse the results in the result tables saved in MaxDEA Pro. It is
more convenient to copy or export the results.
Export results to Excel or other database
The results can be exported to Excel and other types of databases, such as Access,
dBase. Other types of databases are useful when the results exceed the maximum
rows in Excel.
Note: Export results to other types of database may require that MS Access be a
full installation.
78
10 How to buy and register MaxDEA Pro
MaxDEA Basic can be downloaded at http://www.MaxDEA.cn free. It is fully
functional with basic DEA models. MaxDEA Basic will be upgraded to MaxDEA Pro
after registration.
2) Open MaxDEA Basic, click the menu “Tools – Register and Upgrade to MaxDEA
3) You will receive the registration code by email in days (within 1 day in most
cases);
4) Fill in the “Registration Code ” sent to you, click the button “Register”.
2) Remove all the removable disks (USB disks) from your computer temporarily;
3) Plug the flash disk you want to use as the USB key for MaxDEA into your
computer;
4) Open MaxDEA Basic (the MaxDEA file can be placed anywhere, not necessarily in
the USB disk), click the Menu “Tools – Register and Upgrade to MaxDEA Pro”,
remember to check the option box "Portable licence", and send the "Machine
Code" to MaxDEA@qq.com;
79
5) You will receive the "Registration Code" by email in days (within 1 day in most
cases);
Note:
1) The portable licence holder can use MaxDEA on any computer the USB key is
plugged into.
2) The flash disk used for USB key is prepared by the user. Please ensure that the flash
3) The "Machine Code" for portable licence should begin with the letter "U", such as
"U60A-U123456";
4) Every time MaxDEA is used, plug in the USB key first, and do not remove it until
MaxDEA is closed.
80
11 Frequently asked questions (FAQ)
1) I cannot find the “setup” file in the downloaded zip file, why?
MaxDEA doesn’t need installation. The only file needed to run MaxDEA is the
mdb file (such as MaxDEA 4.mdb). All the data needed to develop a DEA model,
including dataset and model options, are permanently stored in this file, which
means that after closing and reopening the file, the database and model options are
still there unchanged. MaxDEA provides extreme convenience to develop DEA
models and keep backups of the models. If you want to keep a backup of a DEA
model, just rename it as you want, such as Model_CCR.mdb, it is a full backup of
the DEA model.
2) Why I cannot open the program file (mdb file), why?
MaxDEA is developed with VBA for Access, so Microsoft Access is required for
running MaxDEA.
If you cannot open the program file (the file with extension “mdb” cannot be
recognized), it indicates that Microsoft Office Access is not installed in your
computer, and you must install MS Access 2003, 2007 or 2010 (Professional OR
Runtime) first. Access Runtime can be downloaded free at Microsoft website.
3) After registration, Can I use MaxDEA on multiple computers?
MaxDEA has two types of licences.
One is fixed licence, which is bonded to a computer and can be used in this
computer only. The licence will not be affected by reinstalling system or
formatting hard disk.
The other is portable licence. This licence type uses a flash disk as the USB key.
The license holder may work with MaxDEA on any computer the USB key is
plugged into. The flash disk is prepared by the user.
4) How to update MaxDEA to a higher version? Is update free?
Update to a minor higher version within the same major version is free of charge.
For instance, MaxDEA Pro 5.0 can be updated to 5.x free.
81
MaxDEA update is very convenient. Just download an updated version of
MaxDEA Basic from http://www.MaxDEA.cn, and register it with your
registration code.
5) How to backup my DEA models?
MaxDEA provides extreme convenience to develop DEA models and keep
backups of the models. If you want to keep a backup of a DEA model, just rename
it as you want, such as Model_CCR.mdb, it is a full backup of the DEA model.
6) When I reopen MaxDEA, the data and model options are missing, why?
If you open MaxDEA directly in the zip file, the data may be lost after the program
file is closed. It is strongly suggested that you unzip the file (MaxDEA.zip) first.
7) Can MaxDEA run multiple models at the same time? If yes, how to do it?
Multiple models can be run at the same time. Because the only files needed for
running the program is MaxDEA.mdb, you can rename or copy this file freely.
Each copy of this file contains one DEA model with all its data and model options
saved in the file. And you can open and run multiple files simultaneously.
8) How to deal with zeros in inputs or outputs in MaxDEA?
There are many methods to deal with zero data in the literature. One easy approach
is replacing zeros with a small number, such as 0.01. Please note that a too small
number might cause a numerical problem in linear programming. Suppose, the
value of input1 of DMU1 is 100000000, and the value of input1 of DMU2 is 0. If
you replace the zero with 0.0000001, it will cause a error in solving the linear
program.
9) What is the largest number of DMUs that MaxDEA can deal with?
Theoretically the number of DMUs that MaxDEA can deal with is unlimited. We
have tested with a dataset with 20000 DMUs. Please note that the computing time
will increase dramatically when the number of DMUs increases.
10) MaxDEA terminates with a message “Inputs must be numeric”
Please check your data. There may be spaces in the data, which will be treated as
strings. This problems often take place when you copy the data from a pdf file.
Remove the spaces from the data, or do it in your original dataset and import the
82
data again.
11) When I import data from Excel, MaxDEA terminates with a message “The
search key was not found in any record ”
The first row of the Excel sheet contains column names, and there must not be
spaces left to the column names. For example, column names such as “ Input1”
will result in an error. But spaces in the middle of column names or right to the
column names, such as “Input1 ” and “Input 1”, are permitted, but not
recommended.
This problems often take place when you copy the data from a pdf file.
Remove the spaces from your data in Excel and import the data again.
In addition, this problem may also result from non-uniqueness of the column
names. The column name must be unique for each column.
12) MaxDEA terminates with a message “There is no object in this control ” or
“The progress bar cannot be correctly displayed. You can either close the
progress bar in MaxDEA Options or fix this problem in your Windows
system…..
MaxDEA uses an ActiveX control “Microsoft Progress Bar” to display the
progress during the model is running. This error indicates that the ActiveX control
library “Microsoft Common Controls” is not registered or the registration is
deleted by mistake. You can either close the progress bar in MaxDEA or fix this
problem in your Windows system as follows.
A) Re-register the ActiveX control library in the Windows system.
Find the file “cmd.exe” in the folder “C:\Windows\System32”, double-click it
to open a command window. If you use Windows Vista, 7 or 8, right-click it
and run as administrator.
In the command window, run
regsvr32.exe C:\Windows\System32\mscomctl.ocx
If you use the 64-bit Windows, run
regsvr32.exe C:\Windows\SysWOW64\mscomctl.ocx
B) Close the Progress Bar in MaxDEA Options.
83
84
12 Appendix
Using MaxDEA Pro, a most comprehensive decomposition of the Malmquist Index can
be achieved.
A. Composition of FGLR
Färe R, Grosskopf S, Lindgren B, Roos P. Productivity changes in Swedish
pharamacies 1980–1989: A non-parametric Malmquist approach. J Prod Anal.
1992; 3(1-2): 85-101.
“By allowing for inefficiencies, the productivity index can be decomposed into two
components, one measuring change in efficiency and the other measuring
technical change or equivalently change in the frontier technology.”
MI
EC TC
In MaxDEA Pro, set the “RTS” to “Constant (CRS)”, and the Malmquist Index
and its decomposition are those according to FGLR.
In the results of MaxDEA, change in efficiency or efficiency change (EC) are
named as “Efficiency Change (t to t+1)”; technical change (TC) are named as
“Technological Change (t to t+1)”
The Malmquist Index (MI) =EC*TC
B. Composition of FGNZ
Fare R, Grosskopf S, Norris M, Zhang Z. Productivity Growth, Technical Progress,
and Efficiency Change in Industrialized Countries. Am Econ Rev. 1994; 84(1):
66-83.
“We use an enhanced decomposition of the Malmquist index developed in Rolf Fare
et al. (1994). This enhanced decomposition takes the efficiency-change component
85
calculated relative to the constant-returnsto- scale technology and decomposes it
into a pure efficiency-change component (calculated relative to the
variable-returns technologies) and a residual scale component which captures
changes in the deviation between the variable-returns and constant returns-
to-scale technology.”
The composition of FGNZ is a further decomposition based on FGLR as follows.
MI
EC TC
PEC SEC
86
MI
In the figure, PTC is the technical change factor (TECHCH ) by Ray and Desli.
In MaxDEA Pro, three steps are needed to get the above compositions:
1) set the “RTS” to “Constant (CRS)”, you can get the MI;
2) set the “RTS” to “Variable (VRS)”, the “Efficiency Change (t to t+1)” in the
results is PEC and the “Technological Change (t to t+1)” in the results is PTC;
3) set the “RTS” to “Scale Efficiency or Scale Effect (CRS Score / VRS Score)”,
the “Scale Effect on Malmquist (t to t+1)” in the results is SCH.
MI = PEC * PTC * SCH
D. Composition of Zofio
Zofio, J.L., Malmquist productivity index decompositions: a unifying framework.
Applied Economics, 2007. 39(18): 2371-2387.
Zofio provides a more complete decomposition of Malmquist Index as follows.
It can be seen as a further composition on the basis of FGNZ
MI
EC TC
In MaxDEA Pro, three steps are needed to get the above compositions:
1) set the “RTS” to “Constant (CRS)”, you can get the EC, TC and MI;;
2) set the “RTS” to “Variable (VRS)”, the “Efficiency Change (t to t+1)” in the
results is PEC and the “Technological Change (t to t+1)” in the results is PTC;
3) set the “RTS” to “Scale Efficiency or Scale Effect (CRS Score / VRS Score)”,
the “Scale Effect on Efficiency Change (t to t+1)” in the results is SEC, and the
“Scale Effect on Technological Change (t to t+1)” in the results is STC.
It can also be seen as a further composition on the basis of Ray and Desli.
87
MI
SEC STC
88
12.2 How to get the Luenberger productivity index and
90
12.3 How to calculate MetaFrontier Malmquist and its components
using MaxDEA?
Cheng Gang
The MetaFrontier Malmquist proposed by Oh and Lee (2010) is essentially the same as
the Global Malmquist proposed by Pastor and Lovell (2005). Their difference is that
Oh and Lee (2010) add a "Group" factor to the Global Malmquist. After introducing
The MetaFrontier Malmquist and its components can be calculated in a batch mode as
follows.
91
12.4 An epsilon-based measure of efficiency in DEA - An alternative
1.Introduction
Data envelopment analysis (DEA) is a linear programming methodology to
evaluate the technical efficiency for each member of a set of peer decision making units
(DMUs) with multiple inputs and multiple outputs. There are mainly two types of
approaches to measuring technical efficiency in DEA: radial and non-radial. The radial
measure was first introduced by the CCR model developed by Charnes, Cooper, and
Rhodes (1978). The non-radial model is represented by the widely used slacks-based
measure (SBM model) developed by Tone (2001). On the basis of these two types of
measures, Tone and Tsutsui (2010b) introduced a hybrid model (epsilon-based measure,
EBM), which combines both radial and non-radial measures in a unified framework. In
the EBM model, there are two parameters that need to be determined: a scalar named as
epsilon indicating the relative importance of the non-radial measure over the radial
measure, and a vector indicating the weights of inputs or outputs. To compute these two
parameters, Tone and Tsutsui (2010b) constructed an affinity matrix and then the
largest eigenvalue of the defined affinity matrix was used to compute the epsilon, and
its associated nonnegative eigenvector was used to compute the weights of inputs or
outputs. However, we find their method to construct the affinity matrix may be
questionable. Their method may not reflect the true degree of scattered distribution of
inputs or outputs. Based on Tone & Tsutsui’s idea, we introduced an alternative method
of constructing the affinity matrix, which overcomes the drawback of their method.
2.An epsilon-based measure (EBM) of efficiency
The input-oriented EBM model under constant returns to scale technology is used
for demonstration. It can be easily extended to out-orientation and variable returns to
scale. The EBM model is defined as follows:
92
m
wi si
* min x
i 1 x0
subject to x0 X s 0,
Y y0 ,
0, s 0.
m
i 1
wi 1 (wi 0 i) , and εx is a key parameter which indicates the relative
importance of the non-radial slacks over the radial θ. Parameters εx and wi must be
supplied prior to the efficiency measurement, and they should be units-invariant values.
To determine the two parameters, Tone and Tsutsui (2010b) constructed an affinity
index between two input vectors instead of the Pearson’s correlation coefficient. The
affinity index S(a,b) between the vector a and vector b has the following properties.
(P1) Identical: S(a,a) = 1,
(P2) Symmetric: S(a,b) = S(b,a),
(P3) Units-invariant : S(ta,b) = S(a,b) (t > 0), and
(P4) 1≥S(a,b)≥0.
The ‘‘affinity index” S(a,b) was defined as follows:
S(a,b)=1-2D(a,b)
D(a,b) is a diversity index, which indicates the degree of scattered distribution, and
was defined as follows:
n | cj c |
j 1 (if cmax cmin )
D (a, b) n(cmax cmin ) ,
0 (if cmax cmin )
n
where c j ln a jj , c ln a jj , cmax max(c j ) , and cmin min(c j ) .
b 1 b
n j 1
The largest value of eigenvalue ρ of the affinity matrix S was used to compute the
epsilon, and its associated nonnegative eigenvector wx was used to compute the weights,
as follows.
m max( x )
x , and
m 1
93
wx
w m
.
wi 1
xi
As discussed in the paper by Tone and Tsutsui (2010b), in the narrow range case,
the affinity index should approximate to 1; and in the widely scattered case, the affinity
index should approximate to 0. In their paper, they give an widely scatted example
(example 2 in their paper), the computed affinity index between x1 and x2 is 0.
However their example is too simple to illustrate the rationale of the index, it has only 2
SBM-efficient DMUs. When we apply their method to a completely scattered case
which has more than 2 efficient DMUS, we get questionable results.
The following example data have two inputs (x1 and x2) and one output (Table1).
For the 5 efficient DMUs (A-E), a decrease in x1 is always accompanied by an equal
amount of increase in x2, which shows the complete substitutability between the two
inputs. The complete negative linear correlation exists between x1 and x2 in the frontier
(Figure 1).
Table 1 Example data
DMU x1 x2 y
A 1 5 1
B 2 4 1
C 3 3 1
D 4 2 1
E 5 1 1
F 3 4 1
G 4 3 1
94
A
5
B F
4
C G
3
x2
D
2
E
1
0
0 1 2 3 4 5
x1
DMU x1 x2
x1 0 0.286
x2 0.286 0
DMU x1 x2
x1 1 0.428
x2 0.428 1
DMU x1 x2
x1 1 0
x2 0 1
The largest eigenvalue and eigenvector of the new affinity matrix are max(ρx) = 1.
m max( x ) 2 1
So we have x 1 . This result conforms to the rationale of the
m 1 2 1
affinity index defined by Tone and Tsutsui (2010b).
96
12.5 Overcoming the infeasibility of the super-efficiency DEA model:
1. Introduction
However the SE model may not have feasible solutions under the assumption of variable,
non-increasing or non-decreasing returns to scale (VRS, NIRS, NDRS). Many efforts have
been made to address this problem. Seiford and Zhu (1999) provided necessary and sufficient
conditions for infeasibility of the input-oriented and output-oriented SE-VRS DEA models.
Lovell and Rouse (2003) (Lovell & Rouse, 2003) developed an equivalent standard DEA model
to provide SE-VRS scores by scaling up the inputs (scaling down the outputs) of a DMU under
evaluation. According to the equivalent standard DEA model, the SE score for an efficient unit
that lacks feasible solutions in the standard SE model is equal to the user-defined scaling factor.
Chen (2005)(Chen, 2005) suggests that the input-oriented model be used when the
output-oriented model is infeasible, and vice versa, however both the input-oriented and the
output-oriented SE models may simultaneously be infeasible for an efficient unit. Cook et al.
(2009)(Cook, Liang, Zha, & Zhu, 2009) developed a novel approach which can always get a
feasible solution for the SE-VRS model. However, it is proved in following section of this study
97
that the results yielded by applying Cook et al.’s formula are not flawless.
In this paper, on the basis of Cook et al. (2009)’s idea, we develop a DEA model with
generalized orientation to resolve the infeasibility problem in the SE-VRS model. In addition,
we discuss several special cases derived from the generalized model, among which there are
two cases that are equivalents of the traditional input-oriented and output-oriented models, and
two other cases (named as the modified input-oriented model and the modified output-oriented
model) that provide solutions to the problem of infeasibility. It is shown that the new approach
yields standard efficiency scores and SE scores that are identical to those determined by its
counterpart traditional models, and it also yields optimal solutions and scores for efficient
DMUs that are regarded as infeasible by the traditional models. This paper presents the new
approach for the SE-VRS model only, but it can also applied to SE-NIRS and SE-NDRS
models.
This paper unfolds as follows. In Section 2, we identify the problems associated with the
traditional SE-VRS models and analyze the properties of the new (general-oriented) model. In
Section 3 we provide the method of soling the new model. Section 4 provides a comparison
between the general-oriented model and Cook et al.’s approach. The new model is applied to a
real dataset as an illustrative example in Section 5. Finally, Section 6 concludes.
The input-oriented VRS model for the evaluated DMUk can be formulated as(Banker,
Charnes, & Cooper, 1984):
min
s.t. x j ij
xik , i 1, 2,..., m
j 1
y j rj
yrk , r 1, 2,..., s
j 1
j
1
j 1
j 0, j 1, 2, ..., n
(1)
For an efficient DMUk, the SE-VRS model is(Andersen & Petersen, 1993):
min
98
n
s.t. x j ij
xik , i 1, 2,..., m
j 1
jk
y j rj
yrk , r 1, 2,..., s
j 1
jk
j
1
j 1
jk
j 0, j 1, 2,..., n ( j k ) (2)
In the input-oriented SE-VRS model, infeasibility occurs when an efficient DMUk cannot reach
the frontier formed by the remaining DMUs through increasing inputs. Since the constraint for
inputs in (2) is always feasible, the necessary and sufficient condition for infeasibility is that the
n
A sufficient but not necessary condition for infeasibility of the input-oriented SE-VRS models
is that there exists at least one output that has a value for the evaluated DMUk greater than the
values of any other DMUs.
n n n
n n n
s.t . x j ij
xik , i 1, 2,..., m
j 1
y j rj
yrk , r 1, 2,..., s
j 1
99
n
j
1
j 1
j 0, j 1, 2, ..., n
(3)
s.t . x j ij
xik , i 1, 2,..., m
j 1
jk
y j rj
yrk , r 1, 2,..., s
j 1
j k
j
1
j 1
jk
j 0, j 1, 2,..., n ( j k )
(4)
In the output-oriented SE-VRS model, infeasibility occurs when an efficient DMUk cannot
reach the frontier formed by the rest of the DMUs through decreasing outputs. Since the
constraint for outputs in (4) is always feasible, the necessary and sufficient condition for
n
A sufficient but not necessary condition for infeasibility of the output-oriented SE-VRS models
is that there exists at least one input that has a value for the evaluated DMUk smaller than the
values of any other DMUs.
n n
n n n n
infeasible.
1
min
1
s.t. x j ij
(1 ) xik , (i 1, 2,..., m)
j 1
y j rj
(1 ) yrk , ( r 1, 2,..., s )
j 1
j
1
j 1
j 0, j 1, 2, ..., n
(5)
where both input-decreasing and output-increasing movements are permitted for an inefficient
DMU to reach the frontier, and the efficiency score is defined as (1-α*)/(1+β*), with the
numerator (1-α*) indicating the degree of input shrinkage and the denominator (1+β*) denoting
the output expansion.
1
min
1
s.t. x j ij
(1 ) xik , (i 1, 2,..., m)
j 1
jk
y j rj
(1 ) yrk , ( r 1, 2,..., s )
j 1
jk
j
1
j 1
jk
0, 0, j 0, j 1, 2, ..., n ( j k )
(6)
101
where both input increasing (α ≤ 0) and output decreasing (β ≤ 0) are allowed for an efficient
DMU to reach the frontier formed by the rest of the DMUs, and the SE score is defined as the
ratio: (1-α*)/(1+β*), in which the numerator (1-α*) indicates the degree of input expansion and
the denominator (1+β*) indicates the output shrinkage.
In model (6), it is possible that an efficient DMU increases its inputs and decreases its outputs
simultaneously to reach the frontier determined by the rest of the DMUs. Consequently, model
(6) is always feasible.
Now we extend models (1), (3) and (5) to a generalized form, for DMUk
1 w
I
min
1 w
O
s.t . x j ij
(1 f ( w ) * ) xik , i 1, 2,..., m
I
j 1
y j rj
(1 f ( w ) * ) yrk , r 1, 2,..., s
O
j 1
j
1
j 1
j 0, j 1, 2, ..., n
wI +wO 0
0, if w 0
f ( w) (7)
1, if w 0
where w I and wO are user-defined non-negative numbers, with at least one of them being
positive. The superscripts I and O denote that the user-defined values are attached to the
input-orientation or the output-orientation respectively. The efficiency score is defined by the
ratio: (1-α)/(1+β).
102
If DMUk is efficient in model (7), its generalized SE-VRS model is expressed as:
1 w
I
min
1 w
O
s.t. x j ij
(1 f ( w ) * ) xik , i 1, 2,..., m
I
j 1
j k
y j rj
(1 f ( w
O
) * ) yrk , r 1, 2,..., s
j 1
jk
j
1
j 1
jk
0, 0, j 0, j 1, 2,..., n ( j k )
wI +wO 0
0, if w 0
f ( w)
1, if w 0 (8)
In model (7) and model (8), w I and wO denote the priority of the orientation. The generalized
SE model (8) is always feasible when both w I and wO have non-zero values. The proof of the
feasibility statement for model (8) is based on the proof of universal feasibility for model (6).
Table 1 lists seven special cases of the generalized model and their definitions for (super-)
efficiency score. Case 1 is the equivalent of the traditional input-oriented model, and case 2 is
the modified input-oriented model which overcomes the problem of infeasibility while keeping
input-orientation. Case 3 is the equivalent of the traditional output-oriented model, and case 4 is
the modified output-oriented model which overcomes the problem of infeasibility while
keeping output-orientation. Case 5 to case 7 are non-orientated models. The aim of this paper is
to address the infeasibility of SE-VRS models acknowledging that non-oriented SE-VRS
models do not experience infeasibility problem. In this context, we explain only cases 1 to 4 in
detail.
Case 1: The generalized model is equivalent to the traditional input-oriented model, and the
(super-) efficiency score 1 – α* in the generalized model is equal to θ* in the traditional
input-oriented model.
103
Proof: In case 1, w I 1 and wO 0 , the standard efficiency model can be expressed as:
min (1 )
s.t. x j ij
(1 ) xik , i 1, 2,..., m
j 1
y j rj
yrk , r 1, 2,..., s
j 1
j
1
j 1
j 0, j 1, 2, ..., n
(9)
If (1-α) is replaced by θ, model (9) becomes the traditional input-oriented VRS model (1).
min (1 )
s.t. x j ij
(1 ) xik , i 1, 2,..., m
j 1
jk
y j rj
yrk , r 1, 2,..., s
j 1
jk
j
1
j 1
jk
0, 0, j 0, j 1, 2,..., n (j k )
(10)
If (1-α) is replaced by θ, model (10) becomes the traditional input-oriented SE-VRS model (2).
Case 2: Input-orientation is given priority with output-orientation retained. This means that the
output deceasing movement is allowed for the evaluated DMUk to reach the frontier formed by
the rest of the DMUs, but such a decrease should be minimized. The (super-) efficiency score is
defined as (1 – α*) omitting the denominator (1 + β*), which means that the efficiency score is
measured by movements of inputs only. Case 2 expresses a modified input-orientation with the
following properties:
1) the standard efficiency score (1 – α*) is equal to θ* in the traditional input-oriented model;
104
2) the SE score is equal to θ* in the traditional input-oriented SE-VRS model when the
traditional input-oriented SE-VRS model is feasible; and
These properties justify that the modified input-oriented model overcomes the problem of
infeasibility while keeping the consistency with the traditional input-oriented model.
1
min
1
s.t. x j ij
(1 ) xik , i 1, 2,..., m
j 1
y j rj
(1 ) yrk , r 1, 2,..., s
j 1
j
1
j 1
j 0, j 1, 2, ..., n
(11)
In model (11), the objective function is minimizing (1-α) and maximizing (1+εβ). In its
solutions, a larger β will result in a smaller α. Increasing β makes essentially no contribution to
the reduction of the objective function, because of the effect of the coefficient ε
(Non-Archimedean infinitesimal), so β will be minimized to zero in the optimal solution. When
this occurs, model (11) is equivalent to the traditional input-oriented VRS model and property
1 is proven.
1
min
1
s.t. x j ij
(1 ) xik , i 1, 2,..., m
j 1
jk
y j rj
(1 ) yrk , r 1, 2,..., s
j 1
j k
105
n
j
1
j 1
jk
0, 0, j 0, j 1, 2, ..., n ( j k )
(12)
In model (12), the objective function is to minimize (1-εα) and to maximize (1+β). In its
solutions, both α and β have non-positive values, and the larger the absolute value of β, the
smaller the absolute value of α. An increase in the absolute value of β will result in an increase
in the objective function, but an increase in the absolute value of α will not result in an increase
in the objective function because of the effect of the coefficient ε (Non-Archimedean
infinitesimal). So, the absolute value of β will be minimized to zero in the optimal solution if
β=0 is a solution. If β=0, model (12) can be formulated as:
min 1
s.t. x j ij
(1 ) xik , i 1, 2,..., m
j 1
jk
y j rj
yrk , r 1, 2,..., s
j 1
jk
j
1
j 1
jk
0, 0, j 0, j 1, 2, ..., n ( j k )
(13)
Because ε is a small positive number, the optimal solution will not change after the replacement
of the objective function with “min (1-α)”. After replacement, the model is equivalent to the
traditional input-oriented SE-VRS model. If β=0 is a solution value in model (12), the
traditional input-oriented SE-VRS model is feasible. Then, property 2 is proved.
When the traditional input-oriented SE-VRS model is infeasible, it means that β cannot be 0 in
model (12). In such cases, model (12) still can yield an optimal solution in which the absolute
value of β is minimized. Property 3 is proved.
Case 3: The generalized model is equivalent to the traditional output-oriented model, and the
(super-) efficiency score 1/(1 + β*) in the generalized model is equal to 1/φ* in the traditional
output-oriented model. This can be proved as case 1.
1) the standard efficiency score, 1/(1 + β*), is equal to 1/φ* in the traditional output-oriented
model;
2) the SE score is equal to 1/φ* in the traditional output-oriented SE-VRS model as well when
it is feasible; and
These properties can be proved by the same procedure that was used in case 2. They mean that
the modified output-oriented model overcomes the problem of infeasibility, and, at the same
time, it keeps the consistency with the traditional output-oriented model.
Although, to deal with infeasibility, in the new approach (case 2 and case 4) both output
decrease and input increase are permitted for the evaluated efficient DMU to reach the frontier
formed by the other DMUs, the objective function requires that the output decrease be
minimized in the modified input-oriented SE-VRS model (case 2), and the input increase be
minimized in the output-oriented SE-VRS model (case 4). To achieve such an idea, the
non-Archimedean infinitesimal is used in the programming. In practice, a small positive
number, for example 10-5, can be used instead of ε, and then the model can be solved by
non-linear programing.
The new models can also be resolved through a two-stage method as follows. To be more
precise, by taking the SE-VRS programming for case 2, i.e. model (12), as an example, at stage
1, the first programming is resolved to get a minimum absolute value of β, namely maximal
non-positive solution β*,:
max
s.t. y j rj
(1 ) yrk , r 1, 2,..., s
j 1
jk
j
1
j 1
jk
0, j 0, j 1, 2, ..., n ( j k )
(14)
And at stage 2, a second programming is solved to get a minimum absolute value of α, namely
maximal non-positive solution α*,
max
107
n
s.t. x j ij
(1 ) xik , i 1, 2,..., m
j 1
jk
y j rj
(1 ) yrk , r 1, 2,..., s
*
j 1
jk
j
1
j 1
jk
0, j 0, j 1, 2, ..., n ( j k )
(15)
The two-stage method has two advantages i.e. 1) it can avoid numerical error caused by
approximation of the non-Archimedean infinitesimal, and 2) the problems in both stages relate
to linear programming, hence the transformation is not needed.
4. Comparison of our approach with the approach proposed by Cook et al. (2009)
Cook et al. (2009) proposed an approach to deal with the problem of infeasibility of the
SE-VRS. Model (16) describes the concept of the approach put forth by Cook et al. to deal with
infeasibility yielded by the input-oriented SE-VRS model.
min
s.t . x j ij
(1 ) xik , i 1, 2,..., m
j 1
jk
y j rj
(1 ) yrk , r 1, 2,..., s
j 1
j k
j
1
j 1
jk
0, j 0, j 1, 2, ..., n ( j k )
(16)
where M is a user-defined large positive number, and the SE score is defined by Cook et al. as:
1+τ*+1/(1-β*). Model (16) is always feasible. The proof of this statement relies on the steps
followed in model (6) for proving its feasibility. Regardless of the feasibility of model (16), two
defects have been identified:
Instance 1: τ*=β*=0
In model (17), τ indicates SE attributed to input increase, and β indicates SE attributed to output
decrease. If both τ and β are zeros, which indicates that neither inputs nor outputs show SE, the
super-efficiency score should be equal to 1. Obviously, the super-efficiency score calculated by
applying the definition: 1+τ*+1/(1-β*) is not equal to 1 in the case of τ*=β*=0, which
contradicts the above statement. Such a case can be easily constructed by adding a DMU to the
sample that has exactly the same input and output values as the evaluated efficient DMUk.
If τ* is a positive number and β* is zero, there must be a feasible optimal solution yielded by the
traditional SE model (according to theorem 1 in Cook et al.’s paper), and the SE score
calculated with the new model should be equal to that calculated with the traditional SE model,
just as stated in Cook et al.’s paper. However, the super-efficiency score in this case calculated
with (17) is 1+τ*+1/(1-β*)=1+τ*+1. It should be noted that 1 + τ* is the SE score in the
traditional SE model (i.e. θ*). This paradox should have been uncovered by the comparison of
the SE scores associated with feasible DMUs calculated using the two models. However, the SE
scores cited in Cook et al.’s paper were calculated with the traditional model rather than with
the new model, so the flaw that was previously pointed out failed to become explicit.
Even if the definition of SE score, i.e. 1+τ*+1/(1-β*), is used solely when the standard SE-VRS
model is infeasible, it still can produce unreasonable results. Namely, let both τ* and β* to be
very small numbers, slightly greater than 0; for example, τ*=0.0104 and β*= 0.0057 (this case
can be found in Table 3 of Cook et al.’s paper). Since τ and β indicate the input “super part” and
the output “super part”, respectively, and since both “super parts” are small enough, it is
obvious that the calculated SE score should be slightly greater than 1. Nevertheless, the SE
score determined by the definition formula put forth by Cook et al. is equal to:
1+0.0104+1/(1-0.0057) = 2.0161, which is a much larger number than was expected.
Furthermore, we elaborate on an extreme example by assuming that τ*=0.00000001 and β*=
0.00000001. In solving the problem, if both τ and β are rounded to zero, the SE score should be
equal to 1, but the calculated score based on the definition formula by Cook et al. is 2.
In fact, the relationship between the two components in the model for the determination of the
SE score should be multiplicative, not additive. In other words, the SE score should be defined
as: (1 + τ*)/(1 - β*). By applying this definition the aforementioned inconsistencies disappear.
Since DMUk is an efficient unit, and τ indicates the input increase needed for DMUk to reach the
n n
frontier, τ must be a non-negative value. If τ<0, there will exist j xij j xij (1 ) xik ,
j 1 j 1
jk
which contradicts the precondition that DMUk is an efficient unit. Therefore, τ > 0 must be
added to model (17) as a constraint. Using the multiplicative definition (1 + τ*)/(1 - β*), if the
constraint τ > 0 is omitted, the SE score of an efficient DMU may be less than 1.
109
Similar defects exist in Cook et al.’s approach for output-oriented SE-VRS models.
The essence of Cook et al.’s approach for dealing with infeasibility was that output decrease is
permitted but must be minimized for the evaluated DMUk to reach the frontier formed by the
rest of the DMUs involved in the input-oriented SE-VRS model.
The model that we developed, described in this paper, mainly uses Cook et al.’s idea for dealing
with infeasibility, and we introduced some modifications in order to overcome the drawbacks
associated with Cook et al.’s approach. Cases 6 and 7 in Table 1 are equivalent to Cook et al.’s
approach for input- and output-oriented SE-VRS models, respectively, provided that the
previously discussed defects are rectified. We recommend that case 2 be used for infeasibility in
the input-oriented SE-VRS model, because the efficiency measure doesn’t incorporate output
decrease, which is in accordance with the meaning of input-orientation; and, for the same
reason, case 4 should be used for infeasibility in the output-oriented SE-VRS model.
5. Illustrative example
We apply the modified input-oriented model and the input-prioritized non-oriented model (case
2 and case 6) to a real dataset as an illustrative example. Case 2 and case 6 have same objective
functions, but different definitions of efficiency score. The objective functions, efficiency
definitions, and WI and WO factors for the two cases are listed in Table 1. In essence, case 2 is
input-oriented, and case 6 is non-oriented.
A real dataset from 20 hospitals in China for the year 2010 is used as an example (Table 2). The
dataset consists of two inputs (number of doctors and number of nurses) and two outputs
(number of outpatient visits and number of inpatient admissions). The number of outpatient
visits of Hospital L is greater than that of any other hospital, and the number of inpatient
admissions of Hospital S is greater than that of the other sample hospitals. According to the
sufficient condition for infeasibility, the input-oriented SE-VRS model will be infeasible for
Hospital L and Hospital S.
Table 3 compares the results from the traditional input-oriented models with those from the
modified input-oriented model (case 2) and the input-prioritized non-oriented model (case 6).
Although Hospital A has neither the largest number of outpatient visits nor the largest number
of inpatient admissions, it still has a problem of infeasibility when the input-oriented SE-VRS
model is applied. This illustrates that the case of one output of a sample DMU being greater
than that of the remaining DMUs is not a necessary condition for infeasibility. The efficiency
component 1 - α* in the new models is equal to θ* in the traditional input-oriented model. Both
the modified input-oriented model (case 2) and the input-prioritized non-oriented model (case 6)
yield SE scores for Hospitals A, L and S, which are infeasible when the traditional
input-oriented SE-VRS model is applied. Additionally, both of the two new models yield SE
values equal to those of the traditional input-oriented SE-VRS model for Hospitals B, C, G, K,
N, and P, which are feasible under the traditional model. Note that in the standard efficiency
model, unlike the modified input-oriented model (case 2) that always yields the same efficiency
scores as those calculated by the traditional input-oriented model, the input-prioritized
non-oriented model (case 6) presents some inconsistencies when compared to the results of the
110
traditional input-oriented model (see efficiency scores of Hospitals H, I, Q and R). Therefore,
the modified input-oriented VRS model (case 2) keeps the complete consistency with the
traditional input-oriented model, so it is suggested that the modified input-oriented VRS model
(case 2) should be applied to address the infeasibility problem in the input-oriented SE-VRS
model.
6. Conclusions
Based on the necessary and sufficient conditions for the infeasibility issue of SE-VRS model
and the idea of Cook et al. (2009) for dealing with the infeasibility problem, in the current paper,
we develop a DEA model with generalized orientation to overcome infeasibility problems in
the SE-VRS models. Among the special cases of the generalized model, two cases, named as
the modified input-oriented model (case 2) and the modified output-oriented model (case 4),
overcome the infeasibility problem in SE-VRS models, while keeping the complete consistency
with the traditional input-oriented model and the traditional output-oriented model, respectively.
This approach can be extended from VRS to NIRS and NDRS.
111
Table 1 Special cases of the standard and the generalized super-efficiency models and their definitions for computing
efficiency scores
Case Model Standard efficiency model Super-efficiency model
wI wO Objective Efficiency wI wO Objective Efficienc
Function Score Function y Score
1 Input-oriented 1 0 1 1 *
1 0 1 1 *
1 1
2 Input-oriented (modified) 1 ε 1 * ε 1 1 *
1 1
1 1 1 1
3 Output-oriented 0 1 0 1
1 1 * 1 1 *
1 1 1 1
4 Output-oriented (modified) ε 1 1 ε
1 1 * 1 1 *
113
References
Andersen, P., & Petersen, N. C. (1993). A procedure for ranking efficient units in data
envelopment analysis. Management Science, 39, 1261-1265.
Aparicio, J., Ruiz, J. L., & Sirvent, I. (2007). Closest targets and minimum distance to
the Pareto-efficient frontier in DEA. Journal of Productivity Analysis, 28,
209-218.
Asmild, M., Paradi, J. C., Aggarwall, V., & Schaffnit, C. (2004). Combining DEA
window analysis with the Malmquist index approach in a study of the Canadian
banking industry. Journal of Productivity Analysis, 21, 67-89.
Banker, R. D., & Chang, H. (2006). The super-efficiency procedure for outlier
identification, not for ranking efficient units. European Journal of Operational
Research, 175, 1311-1320.
Banker, R. D., Charnes, A., & Cooper, W. W. (1984). Some models for estimating
technical and scale inefficiencies in data envelopment analysis. Management
Science, 30, 1078-1092.
Banker, R. D., & Thrall, R. M. (1992). Estimation of Returns to Scale Using Data
Envelopment Analysis. European Journal of Operational Research, 62, 74-84.
Berg, S. A., Forsund, F. R., & Jansen, E. S. (1992). Malmquist Indexes of Productivity
Growth during the Deregulation of Norwegian Banking, 1980-89.
Scandinavian Journal of Economics, 94, S211-S228.
Briec, W. (1999). Holder distance function and measurement of technical efficiency.
Journal of Productivity Analysis, 11, 111-131.
Casu, B., & Thanassoulis, E. (2006). Evaluating cost efficiency in central
administrative services in UK universities. Omega-International Journal of
Management Science, 34, 417-426.
Chambers, R. G., Chung, Y., & Färe, R. (1996). Benefit and Distance Functions.
Journal of Economic Theory, 70, 407-419.
Chambers, R. G., Färe, R., & Grosskopf, S. (1996). Productivity growth in APEC
countries. Pacific Economic Review, 1, 181-190.
Charnes, A., Cooper, W. W., & Rhodes, E. (1978). Measuring the efficiency of decision
making units. European Journal of Operational Research, 2, 429-444.
Charnes, A., Haag, S., Jaska, P., & Semple, J. (1992). Sensitivity of Efficiency
Classifications in the Additive-Model of Data Envelopment Analysis.
International Journal of Systems Science, 23, 789-798.
Charnes, A., Roussea, J. J., & Semple, J. H. (1996). Sensitivity and stability of
efficiency classifications in data envelopment analysis. Journal of Productivity
Analysis, 7, 5-18.
Chen, Y. (2005). Measuring super-efficiency in DEA in the presence of infeasibility.
European Journal of Operational Research, 161, 545-551.
Cheng, G., & Qian, Z. (2012). A Malmquist Model Based on Adjacent Common
Reference. Systems Engineering [Chinese], 30, 105-110.
Cheng, G., Zervopoulos, P., & Qian, Z. (2013). A variant of radial measure capable of
dealing with negative inputs and outputs in data envelopment analysis.
114
European Journal of Operational Research, 225, 100-105.
Chung, Y. (1996). Directional Distance Functions and Undesirable Outputs. Southern
Illinois University at Carbondale
Chung, Y. H., Färe, R., & Grosskopf, S. (1997). Productivity and undesirable outputs: A
directional distance function approach. Journal of Environmental Management,
51, 229-240.
Cook, W. D., Liang, L., Zha, Y., & Zhu, J. (2009). A modified super-efficiency DEA
model for infeasibility. Journal of the Operational Research Society, 60,
276-281.
Cooper, W. W., Seiford, L. M., & Tone, K. (2007). Data envelopment analysis: a
comprehensive text with models, applications, references and DEA-Solver
software (2nd ed.). New York: Springer Science + Business Media.
Deprins, D., Simar, L., & Tulkens, H. (1984). Measuring Labor Efficiency in Post
Offices. In M. G. Marchand, P. Pestieau & H. Tulkens (Eds.), The Performance
of Public Enterprises: Concepts and Measurement (pp. 243-268). Amsterdam,
North Holland: North-Holland.
Doumpos, M., & Zopounidis, C. (2010). A multicriteria decision support system for
bank rating. Decision Support Systems, 50, 55-63.
Färe, R., Grosskopf, S., Lindgren, B., & Roos, P. (1992). Productivity changes in
Swedish pharamacies 1980–1989: A non-parametric Malmquist approach.
Journal of Productivity Analysis, 3, 85-101.
Jahanshahloo, G. R., Vakili, J., & Zarepisheh, M. (2012). A Linear Bilevel
Programming Problem for Obtaining the Closest Targets and Minimum
Distance of a Unit from the Strong Efficient Frontier. Asia-Pacific Journal of
Operational Research, 29.
Kao, C. (2009). Efficiency measurement for parallel production systems. European
Journal of Operational Research, 196, 1107-1112.
Kneip, A., Simar, L., & Wilson, P. W. (2008). Asymptotics and Consistent Bootstraps
for DEA Estimators in Nonparametric Frontier Models. Econometric Theory,
24, 1663-1697.
Lovell, C. A. K., & Rouse, A. P. B. (2003). Equivalent standard DEA models to provide
superefficiency scores. Journal of the Operational Research Society, 54,
101-108.
Luenberger, D. G. (1992). Benefit functions and duality. J Math Econ, 21, 461-481.
Mannino, M., Hong, S. N., & Choi, I. J. (2008). Efficiency evaluation of data
warehouse operations. Decision Support Systems, 44, 883-898.
Oh, D.-h., & Lee, J.-d. (2010). A metafrontier approach for measuring Malmquist
productivity index. Empirical Economics, 38, 47-64.
Pastor, J. T., & Lovell, C. A. K. (2005). A global Malmquist productivity index.
Economics Letters, 88, 266-271.
Podinovski, V. V. (2004). Production trade-offs and weight restrictions in data
envelopment analysis. Journal of the Operational Research Society, 55,
1311-1322.
Portela, M. C. A. S., & Thanassoulis, E. (2007). Comparative efficiency analysis of
115
Portuguese bank branches. European Journal of Operational Research, 177,
1275-1288.
Portela, M. C. A. S., Thanassoulis, E., & Simpson, G. (2004). Negative data in DEA: a
directional distance approach applied to bank branches. Journal of the
Operational Research Society, 55, 1111-1121.
Ray, S. C. (2008). The directional distance function and measurement of
super-efficiency: an application to airlines data. Journal of the Operational
Research Society, 59, 788-797.
Rousseau, J. J., & Semple, J. H. (1995). Two-person ratio efficiency games.
Management Science, 41, 435-441.
Seiford, L. M., & Zhu, J. (1998a). Sensitivity analysis of DEA models for simultaneous
changes in all the data. Journal of the Operational Research Society, 49,
1060-1071.
Seiford, L. M., & Zhu, J. (1998b). Stability regions for maintaining efficiency in data
envelopment analysis. European Journal of Operational Research, 108,
127-139.
Seiford, L. M., & Zhu, J. (1999). Infeasibility of super-efficiency data envelopment
analysis models. Infor, 37, 174-187.
Seiford, L. M., & Zhu, J. (2003). Context-dependent data envelopment analysis -
Measuring attractiveness and progress. Omega-International Journal of
Management Science, 31, 397-408.
Serrano-Cinca, C., Fuertes-Callen, Y., & Mar-Molinero, C. (2005). Measuring DEA
efficiency in Internet companies. Decision Support Systems, 38, 557-573.
Sharp, J. A., Meng, W., & Liu, W. (2007). A modified slacks-based measure model for
data envelopment analysis with 'natural' negative outputs and inputs. Journal of
the Operational Research Society, 58, 1672-1677.
Shestalova, V. (2003). Sequential Malmquist indices of productivity growth: An
application to OECD industrial activities. Journal of Productivity Analysis, 19,
211-226.
Simar, L., & Wilson, P. W. (1998). Sensitivity analysis of efficiency scores: How to
bootstrap in nonparametric frontier models. Management Science, 44, 49-61.
Simar, L., & Wilson, P. W. (1999). Estimating and bootstrapping Malmquist indices.
European Journal of Operational Research, 115, 459-471.
Simar, L., & Wilson, P. W. (2000). A general methodology for bootstrapping in
non-parametric frontier models. Journal of Applied Statistics, 27, 779-802.
Sueyoshi, T., & Aoki, S. (2001). A use of a nonparametric statistic for DEA frontier
shift: the Kruskal and Wallis rank test. Omega-International Journal of
Management Science, 29, 1-18.
Tone, K. (2001). A slacks-based measure of efficiency in data envelopment analysis.
European Journal of Operational Research, 130, 498-509.
Tone, K. (2002). A slacks-based measure of super-efficiency in data envelopment
analysis. European Journal of Operational Research, 143, 32-41.
Tone, K., & Tsutsui, M. (2009). Network DEA: A slacks-based measure approach.
European Journal of Operational Research, 197, 243-252.
116
Tone, K., & Tsutsui, M. (2010a). Dynamic DEA: A slacks-based measure approach.
Omega-International Journal of Management Science, 38, 145-156.
Tone, K., & Tsutsui, M. (2010b). An epsilon-based measure of efficiency in DEA - A
third pole of technical efficiency. European Journal of Operational Research,
207, 1554-1563.
Tulkens, H. (1993). On FDH efficiency analysis: Some methodological issues and
applications to retail banking, courts, and urban transit. Journal of Productivity
Analysis, 4, 183-210.
Tulkens, H., & Vanden Eeckaut, P. (1995). Non-parametric efficiency, progress and
regress measures for panel data: Methodological aspects. European Journal of
Operational Research, 80, 474-499.
Wilson, P. W. (1995). Detecting Influential Observations in Data Envelopment
Analysis. Journal of Productivity Analysis, 6, 27-45.
Xue, M., & Harker, P. T. (2002). Note: Ranking DMUs with infeasible super-efficiency
DEA models. Management Science, 48, 705-710.
Zhu, J. (1996). Robustness of the efficient DMUs in data envelopment analysis.
European Journal of Operational Research, 90, 451-460.
Zhu, J. (2001). Super-efficiency and DEA sensitivity analysis. European Journal of
Operational Research, 129, 443-455.
Zhu, J. (2009). Quantitative Models for Performance Evaluation and Benchmarking:
DEA with Spreadsheets (2nd ed.). Boston: Springer.
117