ra

AD-A283 214 iNTATION PAGE 1 I

Appoved
0M8 No. 0704-0188
%

tor reviewing in$tructoni setrcti ng enssing data

""3 3333
atedo
min
.1 f.a 1 hOUr ocr resporte. Incluing the time
Send comment, etcgarding this burden estrnmate or any other awaet Of this
I ,*esi g the .o~llection of inform ation
soares.

s burden to WVashington HeadQuarters Services. Directorate for information Operaton$ arid Reports. It 1 jeffetsfo%
......
. .- . at "u ime Office of Management and Budget. PaPerwork RItduction Proje.t (0704-0 181). Washington, DC 20503.

1. AGENCYUSE ONLY(Leave blank) J2. REPORT DATE ' 3. REPORT TYPE AND DATES COVERED

D-mA "8WTm M AMA L4/ .si S. FUNDING NUMBERS

YAi0p&LS P#,1bP&Arf$ APPIJM i0"~S

6. AUTHOR(S)

A~A~JL .b)jr'!ho lz ___

7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION

REPORT NUMBER

9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS-(ES) 1 10. SPONSORING/MONITORING

DEPARTMENT OF THE, AIR FORCE AGENCY REPORT NUMBER
AFtT/CI
2950 P STREET
WRIGHT-PATTERSON AFB OH 45433-7765
M J1 I
11. SUPPLEMENTARY NOTES

QG 1 2 199449
12a. DISTRIBUTION/AVAILABILITY STATEMENT lAW 190-1 12b. DISTRI CODE
Release
Approved for Public
Distribution Unlimited
MICHEAL M. BRICKER, SMSgt, USAF
Cheif Administration

13. ABSTRACT (MaxrMU.nM 200 words)

14. SUBJECT TERI'MS 11. NUMBER OF PAGES

16. PRCECODE E

17. SECURITY CLASSIFICATION 1Ie. SECURITY CLASSIFICATION 19. SECURITY CLASSIFICATION 20. LIMITATION OF ABSTRACT
OF REPORT OF THIS PAGE OF ABSTRACT

V77 ."'. ' . .-.. ,

NSN 7 540-O1-2B W-5500
 GENERAL INSTRUCTIONS FOR COMPLETING SF 298
The Report Documentation Page (RDP) is used in announcing and cataloging reports. It is important
that this information be consistent with the rest of the report, particularly the cover and title page.
Instructions for filling in each block of the form follow. It is important to stay within the lines to meet
optical scanningrequirements.

Block 1. Agency Use Only (Leave blank). Block 12a. Distribution/Availability Statement.
Denotes public availability or limitations. Cite any
Block 2. Report Date. Full publication date av.ailability to the public. Enter additional
including day, month, and year, if available (e.g. 1 limitations or special markings in all capitals (e.g.
Jan 88). Must cite at least the year. NOFORN, REL, ITAR).

Block 3. Type of Report and Dates Covered. DOD See DoDD 5230.24, "Distribution
State whether report is interim, final, etc. If S e e o n Technical
applicable, enter Jun
inclusive 7-30Jun
report dates (e.g. 10
8).Documents."- Statmentso c
Jun 87 - 30 Jun 88). DOE See authorities.
Block 4. Title and Subtitle. A title is taken from NASA - See Handbook NHB 2200.2.
the part of the report that provides the most NTIS - Leave blank.
meaningful and complete information. When a
report is prepared in more than one volume, Block 12b. Distribution Code.
repeat the primary title, add volume number, and
include subtitle for the specific volume. On
classified documents enter the title classification DOE - Eave blank.
in parentheses. DOE Enter DOE distribution categories
from the Standard Distribution for
Block 5. Funding Numbers. To include contract Unclassified Scientific and Technical
and grant numbers; may include program Reports.
element number(s), project nLrmber(s), task NASA - Leave blank.
number(s), and work unit number(s). Use the NTIS - Leave blank.
following labels:
C - Contract PR - Project Block 13. Abstract. Include a brief (Maximum
G - Grant TA - Task 200 words) factual summary of the most
PE - Program WU - Work Unit significant information contained in the report.
Element Accession No.

Block 6. Author(s). Name(s) of person(s) Block 14. Subject Terms. Keywords or phrases
responsible for writing the report, performing identifying major subjects in the report.
the research, or credited with the content of the
report. If editor or compiler, this should follow
the name(s). Block 15. Number of Pages. Enter the total
number of pages.
Block 7. Performino Oroanization Name(s) and
Address(es). Self-explanatory
Block 16. Price Code. Enter appropriate price
Block 8. Performing Orcanization Report code (NTIS only).
Number. Enter the unique alphanrumeric report
number(s) assigned by the organization Blocks 17.- 19. Security Classifications. Self-
performing the report. expiaratory. Enter U.S. Security Classification in

Block 9. Sporsorino/rlorinior' q A.oerncv Name(s) accordance with U.S. Security Regulations (i.e.,
and Address(es). Self-exparatory UNCLASSIFIED). If form contains classified
Bnformatuon, stamp classification on the top and
Block 10. SponsorinqlM.'on;Loriia Agency iDbottom of the page.
Report Number. (If known)

Block 11. Suoplementarv Notes. Enter Block 20. Limitation of Abstract. This block must
information not included elsewhere such as: be completed to assign a limitation to the
Prepared in cooperation with...; Trans. of...; To be abstract. Enter either UL (unlimited) or SAR (same
published in.... When a report is revised, include as report). An entry in this block is necessary if
a statement whether the new report supersedes the abstract is to be limited. If blank, the abstract
or supplements the older report. is assumed to be unlimited.
 LARGE-SCALE DATA ENVELOPMENT ANALYSIS

MODELS AND RELATED APPLICATIONS

A Dissertation Presented to the Graduate Faculty of the

School of Engineering and Applied Science

Southern Methodist University

Accesion For
in
NTIS CRn&I
F Partial Fulfillment of the Requirements

DUGl 1AI [] for the degree of

U;: iiir,o _J

.. Doctor of Philosophy
By witha
Dist: ib:!tion ! wih

A,.aiiEb-iity .oc-ccs Major in Operations Research

Ava!9 ar,; d/or
DistI Spctal by

Matthew L. Durchholz

(B.S., United States Air Force Academy, 1981)

(M.A., Southern Methodist University, 1987)
(M.S., Southern Methodist University, 1991)

May 21, 1994

\\Ar e94-25408

48 08
LARGE-SCALE DATA ENVELOPMENT ANALYSIS
MODELS AND RELATED APPLICATIONS

Approved by:

DILRIHRD S.BARR

DR. RICHARD V. HEL kS( /

DN

DRI FFERY L. KENNINCTOX

DR. MARION G. SOBOL

Durchholz, Matthew L. B.S., United States Air Force Academy, 1981
M.A., Southern Methodist University, 1987
M.S., Southern Methodist University, 1991

Larfe-Scale Data Envelopment Analysis

Models and Related AD2lications
Advisor: Associate Professor Richard S. Barr

Doctor of Philosophy degree conferred May 21, 1994

Dissertation completed October 18, 1993

This dissertation presents several advances in data envelopment analysis (DEA),

a method for assessing the efficiency of decision units through the identification of empir-
ical best-practice frontiers. First, a new hierarchical decomposition approach for solving
large-scale problems is described with results of computational testing in both serial and
parallel environments, that dramatically reduces the solution time for realistic DEA ap-

plications. Second, a new set of models for stratifying and ranking decision units provides
important newer insights into the relationships among the units than what was possible
with traditional frontier analysis. Because of the intensive computational requirements
of these models, their practicality builds on the effectiveness of the hierarchical process.

Finally, a new means of assessing the robustness of a decision-unit's efficiency is given

which spans all current models and assists managers in their evaluation of process and
organizational improvement options. It is expected that these advances will permit

practitioners and researchers to be more expansive and ambitious in their use of this
important class of models, and will hopefully encourage new and even more exciting
applications of DEA.

iii
 TABLE OF CONTENTS

LIST OF ILLUSTRATIONS .......... ...................... ... vi

LIST OF TABLES ........... .......................... ... viii

ACKNOWLEDEGEMENTSS ............. ...................... x

Chapter
I. MEASURING EFFICIENCY WITH
DATA ENVELOPMENT ANALYSIS ............ ................ 1

1.1. Overview ................. ........................ 1

1.2. Notation and Conventions ............ ................. 5

1.3. Characteristics of DEA Efficiency Measures ....... .......... 6

1.4. A History of Frontier Estimation Models ....... ........... 8

1.5. Parallel Computing Fundamentals ..... .............. ... 18

1.6. Objectives ........... ........................ .. 26

II. NEW ADVANCES IN COMPUTATIONAL DEA .... ........... ... 28

2.1. Computational DEA ............................... 29

2.2. PIONEER Code ......... ..................... .... 36

2.3. Parallel PIONEER Code ........ .................. ... 42

2.4. Hierarchical Decomposition ...... ................. ... 47

2.5. Conclusions ............ ....................... ... 71

III. RANK ORDERING DMUS USING TIERED DEA .... .......... .. 72

3.1. Need for Ranking Procedure ........... ................ 75

3.2. Tiered DEA ............ ....................... ... 78

3.3: Ranking Within Tiers ........ ................... ... 88

3.4. Example of Tiered DEA ........ .................. ... 90

3.5. Computational Considerations ...... ................ ... 93

3.6. Additional Benefits of Ranking ........... ............... 94

iv
 3.7. Conclusions . . . . . . . . . . . . . . . . . . . . . . . 95

WV. DEA SENSITIVITY ANALYSIS ........ .................. ... 97.

4.1. Motivation ........... ........................ ... 97

4.2. BCC and CCR DEA Sensitivity Analysis ............... ... 100

4.3. CCR and BCC Sensitivity Procedures ......... ............ 105

4.4. Additive Model Sensitivity Analysis ..... ............. .. 108

4.5. Expanded Sensitivity Analysis ...... ................ .. 114

4.6. Summary ............ ........................ .. 117

V. SUMMARY AND CONCLUSIONS .......... ................. 119

APPENDIX

A. MULTINORMAL DATA GENERATOR ..... ............... ... 121

B. DEA-GEN ............. ........................... ... 123

BIBLIOGRAPHY .............. .......................... .. 126

V
 LIST OF ILLUSTRATIONS

Figure Page
1.1. Total product curve ................ ...................... 7
1.2. Decision-making unit ............... ...................... 9
1.3. Technically efficient DMUs ....... .................... ... 15

1.4. Maximum-attainable speedup as a function of a and p ........... .. 22

2.1. Solution time for subsets of BANK-8 ...... ............... .... 41
2.2. Cumulative solution time for subsets of BANK-8 ...... ........... 42
2.3. Solution time versus problem size ...... ................. ... 48
2.4. Hierarchical decomposition schematic .......... ............... 50
2.5. Total solution time vs. blocksize for BANK-8 .... ............ ... 52
3.1. Normalized data isoquant plot ............ .................. 73
3.2. Example tiered results - two inputs, one output .... ........... ... 79
3.3. Example tiered results - one input, two outputs .... ........... ... 80
3.4. Example 1 tiered results - one input, one output ...... ........... 84
3.5. Example 2 tiered results - one input, one output ...... ........... 86
3.6. Example CCR tiered results - one input, one output ............ ... 87
3.7. Extremal DEA example ......... ..................... ... 89
3.8. TEDEA examples ............ ....................... ... 92
4.1. Normalized data isoquant plot ............ .................. 113
4.2. Total region of stability .......... ..................... .. 115
4.3. Expanded sensitivity analysis ....... ................... .. 116

vi
4.4. Inefficient DMU expanded sensitivity analysis .... ............ ... 117
A.1. Correlations of multinormal data variables ..... ............. ... 122

vii
 LIST OF TABLES

Table Page
1.1. Envelopment models . .. .. .. . . .. . . . . . . . . . . . 19
1.2. Multiplier models ....... ... ....................... .... 19
1.3. Necessary Amdahl fraction a levels ....... ................ ... 23
2.1. PIONEER solution times on test set A ..... ............... ... 40
2.2. Test problems parallel run times (min.) .... ... .............. ... 45
2.3. Test problems parallel speedups ...... .................. ... 46
2.4. Test problems parallel efficiencies ...... ................. ... 46
2.5. Hierarchical test problem results ......... ................. ... 53
2.6. Bank hierarchical level test problem results ..... ............. ... 55
2.7. Multinormal hierarchical level test problem results .... .......... ... 56
2.8. DEA-GENa hierarchical level test problem results .... .......... ... 57
2.9. DEA-GENb hierarchical level test problem results ... .......... ... 58
2.10. DEA-GENc hierarchical level test problem results ....... .......... 59
2.11. Bank data results - linear programs required ..... ............ ... 61
2.12. Multi-Normal data results - linear programs required ............ ... 62
2.13. DEA-GENa data results - linear programs required ..... .......... 63
2.14. DEA-GENb data results - linear programs required ..... .......... 64
2.15. DEA-GENc data results - linear programs required ............ .... 65
2.16. Non-hierarchical and hierarchical comparisons 8,000 DMU problems . 66
3.1. large data set tiered DEA results ...... ................. ... 81

viii
3.2. Tiered DEA example data ........... .................... 90
3.3. Results of the tiered rank ordering procedure ..... ............ ... 91

4.1. Sensitivity analysis example .......... ................... 111

4.2. Comparative sensitivity results .......... .................. 11
4.3. Normed data example ............. ...................... 112
4.4. Normed data sensitivity results ........ .................. .. 113
4.5. Expanded sensitivity example results ....... ............... ... 116
A.1. Variance-Covariance matrix for multinormal data .... .......... .. 121
B.1. Parameter values for DEA-GEN code .......... ............... 125

ix
 ACKNOWLEDGMENTS

The completion of this effort marks the attainment of a goal that I set while I was

an undergraduate student at the United States Air Force Academy. The process that
allowed me to complete this work came about because of the influence and guidance of
many people to whom I am greatly indebted.

My sincere thanks goes to the U.S. Air Force which sponsored all my graduate
education. I am particularly grateful to Colonel Daniel Litwhiler and Colonel Raymond
Franck for making the opportunity to pursue the Ph.D. possible.

I owe a great debt to my dissertation committee, especially my advisor and friend,

Dr. Richard Barr. The committee's guidance and recommendations made the seemingly
impossible task of completing a Ph.D. program in three years a reality. I also acknowl-

edge Dr. Lawrence Seiford who offered support and provided valuable recommendations
which kept my work on schedule.
Without the encouragement of my family and friends, I would not have had the

endurance or drive to complete this work. I appreciate the members of my community

group from Fellowship Bible Church who welcomed me as a family member. Their

prayers and emotional support gave me the strength to see this effort through.
Finally, I'm thankful to my love, Leila McLendon (soon to be Durchholz). God has
blessed me with such a wonderful friend during the most intense periods of my research.
I look forward to having her as my wife.

x
 CHAPTER I
MEASURING EFFICIENCY WITH
DATA ENVELOPMENT ANALYSIS

1.1. Overview
In both competitive and regulated markets, measuring the productive efficiency

of firms and industries is important to economic theorists, policy makers, and production
managers. For example, the measurement of productive efficiency before and after
policy changes are enacted on a firm's operations will reveal the effects of those changes.
These measures of efficiency can also assist managers in identifying those portions of the

production process that, if made more efficient, would increase output, without absorbing
more input resources. For private industry, this can allow the competitive firm to increase
market share, effectively translate raw resources to consumable outputs, and pass the
savings on to the customer. For regulated industries, inefficiencies can be revealed and
policy effectiveness measured.

1.1.1. Firm Productivity: Early

Efficiency Measures
One early approach to efficiency measurement had been to provide a mathemati-

cal representation of a firm's economic activity, based on a set of behavioral assumptions.

The actual behavior of a firm is compared to the model to determine the effectiveness of

the firm at achieving productive efficiency.

Unfortunately, establishing a mathematical representation for even a simple pro-

duction process can be a difficult task. Since the measured results of the model de-

pend directly on the behavioral assumptions, the validity of the assumptions must be

1
scrutinized carefully. Moreover, in practice, policy makers are seldom able to agree on

which factors to include in the model formulation and as different models are employed,.

results can vary greatly. Consequently the results of the mathematical model may not

reflect true economic activity as much as the degree to which they reflect the assumptions

made in designing the model.

Another limitation of this modeling approach is that there axe few common be-

havioral assumptions that exist across a variety of industries. Therefore, a new complex

mathematical model must be developed for each industrial application. In addition, once

a model is established for an industry, the model may be applicable only to a specific time

period. This lack of a unifying methodology, based on a limited number of assumptions,

for all industries and time frames, proved to be unsatisfactory. Therefore, this approach

to measuring efficiency has not achieved wide acceptance.

Because of the above limitations, other approaches were developed which did not

involve building mathematical models, but centered directly on measuring the efficiency

of a firm by observing the quantity of inputs consumed and the quantity of outputs

produced. Those firms producing more output per unit of input, compared to other firms,

would naturally be considered more efficient. Such engineering-style ratios of output to

input could also be used to rank the firms by order of efficiency.

For the case of a single input and a single output this ratio analysis is straightfor-

ward. But for the more common case of a firm with multiple inputs and multiple outputs,

the measurement task becomes more complex. One approach has been to form a variety

of simple ratios and rank the firms using a prioritization of the ratios. Determining

the relevant ratios and assigning the appropriate priority to each proved to be tenuous.

Attempts to aggregate the outputs into a single weighted measure and form a ratio with

aggregated inputs has also been offered. However, no consensus emerged regarding the

appropriate weighted index value to apply to each output and input.

2
 1.1.2. Modem Efficiency Measurement

In 1957, M. J. Farrell [51] proposed a solution to the problem of identifying a

single measure of productive efficiency for processes which transform multiple inputs into

multiple outputs. Based on extensions of ratio analysis, he was able to avoid the index
number problems that characterized previous approaches. Because of the computational
resources available at the time, Farrell restricted his attention to applications involving a

single output but with multiple inputs.

As computer technology advanced, Farrell's methodology for determining produc-

tive efficiency drew increased attention. In 1978, Charnes, Cooper, and Rhodes [37] used
fractional programming techniques to demonstrate the equivalence of Farrell's non-linear
measure to a well-defined linear-programming formulation. Whereas Farrell's metric had
been limited to the case of a single output because of the complexity of solving the

more generalized approach, the Charnes, Cooper, and Rhodes (CCR) method offered an
elegant reformulation and solution procedure for the multiple-output multiple-input case.

Theorists, practitioners, policy makers, and managers now had a means of measuring the
relative efficiency for all firms of interest. This technique is called data envelopment

analysis (DEA). Since the original CCR paper, over 400 articles have been published re-
garding usage and extensions of DEA [80]. This growing acceptance among theoreticians

and practitioners reflects an emerging interest in obtaining a single efficiency measure to

evaluate a firm's productivity.
As DEA has gained in popularity, the need to examine the computational char-
acteristics of the DEA formulations has become evident. Some general-purpose linear
programming codes, when applied to DEA, have proven to be unreliable, computationally
expensive (in terms of time to solution), and limited in the size of problems that can be
solved; as a result, they limit the full usefulness of the advantages that the DEA approach

3
offers. Surprisingly, with the explosive growth of DEA over the last decade, little has

been written to specifically address the issues concerning the computational aspects of-

DEA models.

1.1.3. Objectives of the Research

This dissertation contributes to the development of the DEA methodology by

examining the computational issues of the foundational models. One purpose of this work

is to identify the special structural characteristics associated with the DEA formulations
which, when exploited, can result in significant computational improvements in solution

times.
As newer computer technology is used, these improvements become more relevant.

Experience indicates that advances in mathematical programming have been closely

linked to advances in computer technology. Consequently, the breadth of applications
for data envelopment analysis will grow wider as the computational advantages brought
by newer technologies are identified and employed. Also, since DEA spans the fields

of Operations Research and Economics, improved computational tools will magnify the
synergism existing in the interdisciplinary nature of DEA, yielding new insights and
expanded applications.
Specifically, this dissertation introduces a parallel programming approach to solve
large-scale DEA problems. Because of their nature, DEA models are ideal candidates
for exploiting the relatively new parallel computing machinery. During the process of
adapting DEA to this new technology, new insights into the DEA problems were realized.
This dissertation will identify improvements that can be made to all DEA codes (both
serial and parallel), answer problematic concerns that have been revealed in DEA, and
present new approaches for DEA analysis that are possible with the new computationally

efficient codes.

4
 In addition, a new DEA model is presented for grouping and rank ordering the

firm's operations and accounting for possible outliers in the data. Also, a new approach-

to DEA sensitivity analysis helps paint a picture of the robustness of the operations of a
particular evaluation unit. Both of these advances build upon the gains realized in the

computational research.

1.2. Notation and Conventions

Because the DEA measure of efficiency is derived from the solution of a math-
ematical programming model, a description of the notation and conventions which will
be used throughout this dissertation is necessary. As much as possible, these notations

will conform to the standard emerging in the DEA literature. Matrices are denote
bold upper case Latin letters. The jth column of the matrix X is denoted by Xj anG

element in the ith row and jth column is described as xij. Small Greek and Latin bold
letters represent vectors; the ith element of vector A is given by AX.The Greek letter e is
an arbitrarily small positive non-Archimedean scalar value, i.e., the value is not explicitly
defined. The vector of all zeroes is denoted by 0. The vector of all ones is denoted as
1. Surplus and slack vectors of the linear programming problems will be denoted by a.

Superscripts will denote the constraints to which the surplus or slack variables are to be
associated. Scalars are denoted by small italicized Latin and Greek letters. Subscripts on
scalars are used to distinguish two or more variables that share common characteristics
in the problem but assume different values. Any variable superscripted by `*"(e.g., z*),
represents the value the variable takes on as part of the optimal solution for the stated
problem. All vectors and matrices are treated in a manner such that the dimensions

are conformbble for multiplication. A vector is treated as a row when it appears on

the left of a matrix and as a column when it appears on the right; thus, alleviating the
need to distinguish between row and column vectors. The inner product of two vectors
I&and Yo is denoted by sYo. Point sets are denoted by capital Latin letters. Both

5
subscripts and superscripts may uniquely identify a set described within a given context.

Elements of those sets are denoted by small Latin letters with appropriate superscripts

and subscripts. The cardinality of set E will be denoted by IEI.

1.3. Characteristics of DEA Efficiency Measures

It is important to understand what is meant by efficiency with reference to DEA,

since economists have a multitude of definitions and measures. In simplest terms, a

measure of how well a firm succeeds in producing as much of its outputs as possible for

a given amount of its inputs is called the efficiency of operations for that firm. But what

is the basis of comparison? One approach is to establish an absolute standard, usually

based on a mathematical model, by which all firms will be evaluated. The shortcomings

of this approach have already been discussed. An alternate approach would be to

measure a firm's performance relative to other firms which are producing like outputs

using the same type of inputs. Of course, the quantity of outputs produced as well as

the amount of inputs consumed in the production process will vary from firm to firm.

It is precisely this variation that will indicate a firm's relative efficiency compared to its

contemporaries.

Data envelopment analysis identifies: those firms producing the most output for

a given amount of input and those firms consuming the least input to produce a given

level of output. Figure 1.1 depi,;ts the performance of a group of firms (each shown as a

point in the scatter diagram) that produce one type of output and consume one type of

input. The efficient firms are termed technically efficient and form an upper boundary by

which the remaining inefficient firms are compared. 1 This empirically derived boundary,

1 Farrell [51] gives a complete description of technical efficiency which measures

a firm's success in producing maximum output for a given amount of inputs. This differs
from other measures of efficiency, such as price efficiency, which measures the firm's
success of choosing an optimal set of inputs to produce the output. The efficient firms
identified by the Farrell measure are a subset of the technically efficient firms.

6
 J ....- Jinefficient

X=Input

Figure 1.1. Total product curve

together with the connecting line segments, is called the efficient frontier or efficient
production surface.
Because DEA identifies technical efficiency in general terms of quantities of out-

puts and inputs, the methodology can be applied to competitive firms, regulated firms,
and nonprofit organizations. This broad applicability contributes to the growing use
of DEA. However, the approach also has shortcomings. First of all, the technically

efficient subset of firms under analysis may not be very effectie in translating inputs
into outputs. Most would agree that a firm engaged in a poor production process, even

though performing better than contemporaries, should still not be considered "effective."
Farrell [51], as well as Charnes and Cooper [31], recognized this shortcoming and state
dearly that the resulting DEA measure concerns relative efficiency while the question of

effectiveness, or absolute efficiency, remains with the policy makers and analysts.
However, the attractiveness of DEA is that it compares performance with the

best actually achieved rather than some unattainable ideal. If one is concerned with
the question of effectiveness, DEA could first be used to identify the technically efficient
firms. The analyst could then limit the scope to these firms to address the question of

effectiveness. Consequently, DEA continues to be of value even when the effectiveness

of firms is drawn into question. Likewise, if a parametric mathematical representation

7
is desired, the analyst may limit the scope only to the best performing firms from

which to build the model. Consequently, DEA can be of value in legitimizing a given

parametric approach. For the case where prices are not available for either the outputs or

the inputs (e.g., improvement in test scores for elementary school students participating

in Headstart programs [391), DEA provides an ideal approach to measuring productive

efficiency. However, for cases where prices, costs, measures of quality, etc., are available,

DEA can still be useful in developing alternate measures of efficiency which use this

additional data.

Because the economic term firm offers the connotation of for-profit organizations,

a more general definition is used in DEA literature. Since each organization must decide

how best to transform available inputs into desired outputs, each organization will be

referred to as a decision-making unit (DMU).

1.4. A History of Frontier Estimation Models

1.4.1. The Technical Efficiency Measure

If the set N = {1, 2,... , n} is the index set of DMUs being compared, we may

define the efficiency of each DMUj,j E N, as 2.=- A DMUj then is technically

efficient if Ej = z = mac{E}, thus having the highest output-to-input ratio. For the
kEN
case of multiple inputs and/or multiple outputs, Ej can be defined in a manner similar

to the one-input one-output case.

The production process for DMUs with two inputs and one output can be viewed

as in Figure 1.2. The definition of efficiency can be extended to the case of two inputs
and one output
utpt ne bby re-defining
and redefnin E, tj = V1 J
VlXlj+V2X2j where vq represents a weighting factor

for input q. The efficient DMUs are determined as before and result in the greatest ratio

of output to weighted inputs.

Naturally, the above metrics can be generalized to incorporate an arbitrary num-

ber of inputs and outputs. If R = {1,... , s} is an ordered index set of s outputs and

8
 Input x,
Production Output Yl
Input x2 Process

Figure 1.2. Decision-making unit

Q .1,..., m} is an ordered index set of q inputs, then the efficiency for DMUj,j E N,
can be defined as:

- UrYlrj

E- rER
Z VqXqj
qEQ

where u, and vq are weights placed on output r E R and input q E Q, respectively.

A critical concern of this approach is determining the appropriate weights. The-
orists and policy makers could assign specific values to the weights, but their choices
would be open to criticism and the validity of the resulting efficiency measure could be

suspect.
An alternate approach, presented by Farrell, avoids this problem. If the efficient
DMUs were known, the weights could be defined so that Ej would equal 1 for all efficient

DMUs and would be less than 1 for all inefficient DMUs. Consequently, the problem of
finding the relative technical efficiency measure for a given decision making unit, DMUo,
can therefore be described as:

max z E (1.1)
s.t. Ej ,1 jEN (1.2)

ur>O, rER (1.3)

vq2!O, qEQ (1.4)

9
 To find all of the efficient decision-making units in N, the above fractional pro-

gramming problem must be Slved n times, once for each DMU of the set. The computa-

tional requirements to solve the fractional programming problem- at least in this form

- make solving anything other than simple problems impractical.

1.4.2. Data Envelopment Analysis Approach

1.4.2.1. The CCR Multiplier Model

The work of Charnes, Cooper, and Rhodes [37,38] transformed Farrell's nonlinear
fractional programming problem for determining Ej into a linear programming problem.
The computational advantages that the linear programming formulation offered made it
possible to solve for the efficiency of DMUs with many inputs and outputs. The new
formulation to find Ej has become known as the CCR Multiplier Model 2 and is written

as:

(CC-R--o) max z =pYo (1.5)

s.t. JY - VX < 0 (1.6)

vX 0 =1 (1.7)

)A_ 61 (1.8)

v > 61 (1.9)

In this model, let there be n DMUs with X = [X 1 ,..., Xn] and Y = [Y 1,...,yn].

Vectors Xj, Yj denote the observed data where, Xj = (xlj,. .. , xmj) is a column vector

of observed inputs and Yj = (ylj,. . . , Ysj) is a column vector of observed outputs for each

DMU. In this model, 1A is the vector of multiplier weights applied to the outputs and v
the multiplier weights applied to the inputs. To insure that the multipliers remain strictly

This formulation is termed input-oriented; a similar output-oriented version

2
will be described in a later discussion.

10
positive, a non-Archimedean scalar, e, is used in constraints (1.8) and (1.9). Each DMUj

is assumed to have the same inputs and outputs but in varying amounts. Originally, all
observed inputs and outputs were assumed to be strictly positive. Later, it was shown

[421 that the CCR multiplier model holds for inputs and outputs that are strictly non-
negative.
An attractive feature of this linear programming formulation, consistent with

Farrell's development, is that no prior knowledge of the weighted values is necessary

to identify the efficiency of a DMU. Here, the maximization problem finds the most
favorable weights possible for DMUo while satisfying all constraints. As a result, the
weights, A*and v*, are found to make DMUo look as efficient as possible while insuring
that the ratio * r
*= , v*xij for all other DMUs never exceeds one. The

normalization constraint aXo = 1 (1.7) linearizes Farrel's objective function and, with
(1.6), insures that the optimal objective function value, z*, never exceeds 1. Indeed

optimal solutions to the problem will find 0 < z* < 1 with efficient DMUs indicated by
=*-1.

Note that the non-Archimedean constant, e, represents an arbitrarily small,

strictly positive value that insures positivity of the optimal multiplier values. Without

the non-Archimedean term, a particular DMUo could have an optimal objective value
z* = 1 with one or more of the multipliers equal to zero, indicating DMUo is relatively

efficient if one or more of the positive input or output variables are ignored. In this case,

the DMU will be on the efficient frontier but may not be Pareto-Koopmans efficient [42].
For example, let the inputs of DMUo consist of Xo > 0 and the outputs result in Yo > 0.

Let DMUo be identical to DMUo in its production of outputs Yo = Yo but with Xo :_ Xo

where the equality holds for all x4o, i = 1,..., m, except for x4o < xko where 1 < k < m.
Assume at optimality, DMUo is determined to be efficient with A* > 0 and v* > 0
with only vZ- = 0. In this case, DMU* would also be efficient. It is clear, however, that

11
DMUo is inefficient relative to DMUo because both DMUs produce the identical levels

of output but DMUo consumes less of one input. DMUo cannot be Pareto-Koopmans

efficient if, while maintaining the same levels of all outputs, the production possibility
set allows for a decrease of an input while holding all other input levels constant. A
similar example can be constructed when one of the multipliers of an output variable is
equal to zero at optimality. A DMUo with z* = 1 but with at least one multiplier, g,
or vi*', equal to zero in all alternate solutions is termed "weakly efficient" by Charnes,
Cooper, and Thrall [42].

1.4.2.2. CCR Envelopment Model

The linear programming CCR multiplier model has an associated dual formula-
tion which has become known as the (input-oriented) CCR envelopment model. This
model is written as:

(CCRO) min w = B - e(Is t + is)) (1.10)

s.t. YA - s = YO (1.11)

OX0 - x - 8X= o (1.12)

A, 8, 0 > 0 (1.13)

Ofree (1.14)

where 80 and 8i are the slacks associated with the output and input constraints, respec-
tively. This is the most widely discussed DEA formulation in current literature. Even

though the CCR multiplier and envelopment models yield equivalent optimal objective
values, the envelopment model is more efficient, computationally, because of the smaller
number of constraints. 3

3With the envelopment model, the number of constraints equals the number of
outputs and inputs observed for each DMU. The number of variables equals the number

12
 The non-Archimedean term, e, of the CCR models poses a problem. In theory,

this term is denoted by an arbitrarily small positive value; in practice though, it must be

explicitly defined. Ali and Seiford [9] show that the choice of e can dramatically effect
the solution of problems and produce erroneous efficiency scores. This concern can be
overcome as described in the next section.

1.4.2.3. Ali's IDEAS Formulation

Ali [31 presented a variation of the CCR envelopment model which eliminated the
need to explicitly define e. This new approach for the CCR/, formulation, called the
integrated data envelopment analysis system (IDEAS), can be written as the following
pair of linear programs:

(IDEAS Phase I) min 0 (1.15)

s.t. YA - 80 = Yo (1.16)

OX0 - Xx - si = 0 (1.17)

A, si, 'S0> 0 (1.18)

Ofree (1.19)

(IDEAS Phase II) max 8 + si (1.20)

s.t. YX - 80 = Yo (1.21)

OX0- XA - si = 0 (1.22)

0=0* (1.23)

A, 8, '0 _>0 (1.24)

o free (1.25)

of DMUs. In data envelopment analysis, the number of DMUs being observed, n, is

typically much larger than the number of output and input variables (s + m). If this
were not the case, all DMUs would likely be efficient.
13
The 0* in Phase II is the optimal solution found in Phase I. It has been shown that a

DMU is efficient if and only if 0* = 1, 80* = 0, si* = 0 for the problem above [3,9,31,52].

Otherwise, a weakly efficient DMU would have 0* = 1 and at least one positive slack

[33], or a DMU with 0* < 1 would be inefficient with 0* measuring the extent of the

inefficiency.

The non-Archimedean constant has been removed from the problem formulation

but at a possible computational cost of solving an additional linear program for each

DMU during the two-stage approach. 4 Later, this paper will show how the computational

costs of this approach can be minimized to achieve the benefit of eliminating the non-

Archimedean term.

1.4.3. DEA Model Variations

The above efficiency measures all share a common shortcoming: they only ac-

count for constant returns to scale. As a result, only a subset of the DMUs on the

efficient production surface may be revealed. This can best be seen by means of an

illustration. Figure 1.3 depicts the case of a collection of DMUs, each with a single

output and single input.

The efficient production surface is outlined by DMUs a, b, c, d, and e. Only

DMUb, and DMUc exhibit constant returns to scale. In economics terms, these are the

DMUs that result in the greatest average product and represent the most productive

scale size in terms of the level of output produced. 5 The points which exhibit constant

returns to scale on the production surface can be found by drawing a line through the

origin which is tangent to the production surface. The other points on the production

4 Just as the so-called "Big M" value is not explicitly defined in the "Phase I-
II' approach to solving linear programming problems, IDEAS uses a two-stage process
to circumvent the need to explicitly define E.
5 Banker specifically identifies relationships between the most productive scale
size and DEA in his 1984 paper [101.

14
 j o//

X=Input

Figure 1.3. Technically efficient DMUs

surface, ara technically efficient, but are not scale efficient at the most-productive scale

size. Technical efficiency implies that for a given level of output, the efficient DMUs

consume the least amount of input.

If constant scale efficiency implies a most-productive scale size, how can DMUb

and DMUc both exhibit constant returns when they produce different levels of outputs?

Even though DMUb, and DMUc produce different levels of output, the amount of the

difference is exactly proportional to the difference in the amount of inputs. Consequently,

the two DMUs are called scale equivalent. DMUd and DMUe represent decreasing returns
to scale compared to the most productive scale-efficient DMUs because they consume
more input, but produce less than a proportional increase in output. Moving from

DMU. to the most productive scale efficient DMUs represents increasing returns to sca'
because the percentage increase in the amount of output produced exceeds the percentage

increase in the amount of inputs consumed.

All of the above discussion extends to higher dimensions. By limiting the scope to
constant returns to scale, only a portion of the efficient production surface is identified.

Consequently, extensions to the CCR data envelopment analysis measure were developed
to identify all technically efficient DMUs.

15
 Farrell and Fieldhouse 152] noticed this limitation to the Farrell measure in 1962.

They presented a "grouping method" to account for increasing and decreasing returns to

scale. In this methodology, DMUs with similar input and output amounts were grouped

together and efficiency scores determined with respect to the other DMUs within the

particular group. This method never gained wide acceptance, and in the case of multiple

outputs and multiple inputs, the process became too complex and unreliable.

1.4.3.1. BCC Variable-Returns-to-Scale Model

In 1984, Banker, Charnes, and Cooper [11] successfully modified the original CCR

envelopment formulation to allow for "variable returns to scale. The BCC input-oriented

envelopment model can be written as:

(BCC',) min 0 - E(lsi + 180) (1.26)

s.t. YA - 8 = Yo (1.27)

OX0 - XA - s8 = 0 (1.28)

1A = 1 (1.29)

A, 8 , SO > 0 (1.30)

0 free (1.31)

In their paper, Banker, Charnes, and Cooper, show how this new formulation

reveals the entire efficient production surface and all technically efficient DMUs. 6 In a

sense, the additional constraint automatically groups DMUs with similar output levels

and determines the appropriate efficiency score. For inefficient DMUf in Figure 1.3, the

efficiency score will be less than 1. The overall efficient score found by the CCR model

6 One assumption of this formulation is that regions of increasing returns to

scale are followed by regions of constant or decreasing returns to scale. This is beneficial
because it satisfies the normal economic assumption of convexity of the production
function.

16
will be calculated as the ratio of the lengths of the line segments W t . In a like manner

the technical efficiency score, found by the BCC model, will be L-. The scale efficient
LM can be found by dividing the CCR result by the BCC result. Obviously,
measure, 1W,
if a DMU is both scale and technically efficient, the efficiency measure will be 1 in both
cases. Notice that for some DMUs, such as DMUg, the scale and technical inefficiency

scores will be identical. In other cases, such as DMUd, the DMU will be technically
efficient but not scale efficient. Therefore, to determine the technical, scale, and overall

efficiency, two linear programs, the CCR and the BCC models, must be solved for each

DMU.

1.4.4. Compendium of Data Envelopment Models

The formulations of efficiency measure discussed so fax are all input-oriented,

that is, the efficient DMUs are discovered by observing an output level and identifying
the DMU that consumes the least input. Inefficient DMUs are projected onto the

efficient surface by proportionally reducing all inputs while holding output constant. The
projection distance yields the efficiency score.
However, a different perspective can be accomplished by viewing efficiency from

an output-oriented perspective. In this case, the efficient DMUs are discovered by

observing a given level of input and identifying the DMU that produces the most output.
The inefficient DMUs are projected onto the efficient surface by proportionally increasing
all outputs while holding inputs constant. The projection distance is the output-oriented
inefficiency score. It is important to note that the input- and output-oriented schemes of
the BCC model identify the same efficient surface, which is composed of all technically
efficient DMUs. However, the scores for the inefficient DMUs may vary between the two
formulations. For the CCR model, the input- and output-oriented schemes will identify
the same most-productive-scale-size efficient surface and the scores for the inefficient

17
DMUs of the input-oriented model will be the reciprocals of the scores for the output-

oriented model. Seiford and Thrall [81] present a generalization of the BCC and CCR

models with input- and output-oriented schemes and define models with non-decreasing
returns to scale (NDRS) and non-increasing returns to scale (NIRS). Table 1.1 anid Table

1.2 describe the envelopment and multiplier model formulations.

The analysis that follows in this paper will be based on the input- and output-
oriented models presented here. To identify all types of efficiency for a given DMU,

eight linear programs must be solved. This involves identifying the scale and technical
efficiency measures for both the input- and output-oriented models using the two-phase
approach of the IDEAS model. Other variations of these models have been developed
which further increase the computational complexity of determining the DEA efficiency

measures for the DMUs of interest.

Although applications containing tens of thousands of DMUs are now beginning
to arise in practice, the computational issues associated with such large-scale models

have not been addressed in the literature. This research explores parallel processing as a
means of approaching these challenging problems. The next section presents an overview
of the salient concepts associated with parallel programming as a background for this

discussion.

1.5. Parallel Computing Fundamentals

A relatively new and expanding field that offers great promise for solving difficult
large-scale problems is application-level parallel processing. Computational gains can be
achieved in a multitasking setting where the power of multiple processing elements act
concurrently to find the solution for a single problem. It is important to distinguish the

multitasking of a single application from multiprogramming wherein an operating system

allows a computer to execute multiple unrelated programs.

18
Table 1. 1. - Envelopment models

(Input-oriented)
CGRV mai {6j - 6(1.8' +18 0 )IYA - ' jOjXj - XA - 8'=0}
BCCj, mun {O, - C(l8, + is') IYA - 8'= Yj, OX - XA - 8 = Ol A =1}

NDRSj: min 1jO c(18i+ laO)IY,\- o =Yj jXj- XA-8' =O0,1A > 11
e(lsi + 18)IYA -s 0 OjXj-XA\-s t =0,1A\<1}
NIRSj nun {B - =Y
X,81,80!0,ej

(Output-oriented)
CCRj' min {~j + E(l8ss+la)IXA\+ 8i = j, 0y, - 'A + 80 =01
BCCj' min {10j + C(lsi + 180)IXA + 8i = Xj jj- 'A + 80 =0, 1A = 0}
NDRSj': nid {q~j + e(l.9 + 180)IXA\ + 8 = xj j- Yx+ so = 0,1ix 0}
NIRSj* min {kj, +(ls + 1s 0 )IXA\+ '= Xj,oj1',-YA\+s 0 =0,1AO5}
A,Si,sOO,.0,
(Non-oriented)
ADD 3 min {-1s'-- 180YA - 0 =Yj, -Xx - 8i= -xi,,ix1}

Table 1.2. - Multiplier models

(Input-oriented)
UC--Rj' maxfZ = gh~MY - VX <0;I'Xj = 1; JA , V >ell
BC max{z = tYj + u*AY - vX +u*1 <_0; vX = 1;IA El, v >e, u* free}
#A,FJU*
7Ti.maX{z=IAYj+u*IPAY-VX+u*1 0;vX,= 1;1A Ely,>e1,u* !01

(Output-oriented)
CC minfz = vXI&Y - vX<0,MYj = 1,I >el, >el}

NDS min {z=vXj+v4jjY-vX-v*1 0,gAYj = 1, JA 6, V >1,v *:501

7N7TA3S 1' i~nn {z=L'X+v~lpY-vX-v*10,JAYj =1,p ElyI> d,v* 0}
(Non-oriented)
ADDj max{jA3-vXj+wIjAY-vX+1w>0,-jA s-1,-v >-1, w free}

19
 The basic premise in multitasking in a parallel environment is that, for any

given problem, a large number of tasks must be completed to successfully arrive at a

correct solution. If these tasks can be scheduled for simultaneous execution on multiple

processors, opportunities exist for achieving dramatic reductions in solution times. This

fundamental motivation for parallel programming suggests great performance gains can

be made using new hardware technology if the work can be decomposed effectively.

Substantial performance gains can be difficult to achieve. While modest gains

are possible for many applications in a parallel setting, outstanding results occur only

for a much smaller set of problems. To achieve outstanding results, the majority of

steps of the algorithm must be done in parallel, and the algorithm must closely match

the architect, ,re of the underlying machine. As a result, advances in algorithm perfor-

mance are tied to how well they can exploit the computing machinery. As advances

in technoiogy offer new challenges in developing improved algorithms to match new

hardware, insights into computational aspects of the problem can be gained as the new

algorithms are developed. These problem insights can be useful both to the parallel and

serial environments. Therefore, even though the challenge to achieve outstanding gains

through parallel programming may seem formidable, the challenge also has the potential

to provide new insights into the solution process that may otherwise go unnoticed.

1.5.1. Measuring Parallel Performance

Parallel processing consists of coordinating the power of multiple processors to

accomplish a given task concerning a given set of data [20]. To measure the impact of

parallelization, a metric is needed to serve as a basis of comparison between serial and

parallel performance.

Of particular interest is a measure of the ability of parallel processing to improve

the real time to solve a particular problem. Such improvement permits existing prob-

lems to be solved faster, and may render solvable problems whose dimensions exceed

20
traditional computational capacities. Regardless of the motivation, a measure is needed

to reflect the computational improvement offered by the parallel algorithm over a serial
algorithm to accomplish the same task.
Speedup, of which there are several definitions [201, is the most common measure

of the effects of parallelization on a problem. In this research report relative speedup

is used. Relative speedup, S(p) is the ratio of a problem's solution time using the serial

version of the algorithm to the time using the parallel version of the same algorithm with
p processors on the same machine [62]. In the development of any parallel algorithm, a
design objective is to use all resources efficiently to achieve linear speedup where S(p) = p.

The possibility of achieving superlinear speedup, S(p) > p, is open for debate [201.
Many factors determine the speedup potential for any parallel algorithm. In 1967,
G.M. Amdahl formulated a law describing the maximum possible speedup on a parallel

computer with p processors. Amdahl's law can be written as:

11
s(p) < a + (I1- a,)/p

where S(p) represents speedup using p processors, and a is the fraction of code that must

be run in serial (i.e., is not parallelizable). From the law we can see

1
lim S(p)< -.
p-o0 a

Therefore, according to Amdahl's law, linear speedup may only be possible when a small
portion of the programming steps is restricted to the serial case.
Figure 1.4 depicts the maximum-attainable speedup, according to Amdahl's law,

for various ranges of p and a. This illustrates the inherent difficulty of achieving linear

21
 -0 20

S
10 p
e
e
5 d
U
p
V 0
0 0.1 0.2 0.3 0.4 0.5
Serial Fraction, a

Figure 1.4. Maximum-attainable speedup as a function of a and p

speedup. It becomes obvious, that the achievement of linear speedup as the number of

processors increases demands almost complete parallelization of the algorithm's steps.

Efficiency, defined as E(p) = Ep is another measure commonly used to deter-

mine the performance of the parallel algorithm, reflecting speedup normalized by the

number of processors. An efficiency of 1 reflects linear speedup; a value less than 1 is the

fraction of linear speedup attained.

Table 1.3 reveals the different values of a that are necessary to achieve a given

level of efficiency for various number of processors. This table underscores the difficulty of

achieving linear speedup, which rises proportionately with the number of processors used.

With more processors, the fraction of tasks that require serial implementation must be

very small. However, as Amdahl noticed, the fraction, a often depends on the problem

size. If n measures problem size, then a(n) --+ 0 as n --+ oo since, for many problems, the

amount of work to be done in serial is at a fixed level regardless of problem size. This

fixed amount of work becomes negligible for large problem sizes. As a result, for a fixed

22
 Table 1.3.- Necessary Amdahl fraction a levels

Number of Efficiency
processors .8 .9 .95 .99
5 .0625 .0278 .0132 .0025
10 .0278 .0123 .0058 .0011
15 .0179 .0079 .0038 .0007
20 .0132 .0058 .0028 .0005

p, S(p) -* p as n --. oo. Consequently, linear speedup may be possible for large, highly

decomposable problems.

1.5.2. Parallel Computer Architectures

When approaching a problem for parallel implementation, the designer must in-
sure the algorithm takes into account the advantages of the specific hardware on which
the algorithm will be executed. Unlike single-processor machines, where serial algorithms
will execute in a more or less common fashion across many different types of computers,

the parallel implementation must be tailored to the specific hardware. Parallel architec-
tures take on many different configurations, each offering its own unique set of advantages
and disadvantages. Consequently, an algorithm designed for a particular parallel machine
may be completely inappropriate for another type of parallel computer. In 1972 Flynn
[56) offered a classification scheme for all types of parallel computers. The four classifi-

cations are dependent on the type of instruction stream and data stream the computer
is capable of controlling. The instruction stream is the sequence of programming steps
carried out on the computer to arrive at a solution to the problem. The objects on which
the instructions act are called the data stream. The single-instruction, single-data (SISD)
computers are the traditional uniprocessor computers which range from a personal com-
puter to the Cray-1 supercomputer. The single-instruction, multiple-data computers

(SIMD) are represented by multi-processor computers where each processor executes

23
the same instruction sequences on different sets of data. The best-known machine in

this category is the Connection Machine (models CM-1 and CM-2) that contains up to

64,000 processors. Currently, there are no known general-purpose multiple-instruction,

single-data (MISD) computers. The most common parallel architecture is the multiple-

instruction, multiple-data (MIMD) computer which can simultaneously operate multiple

independently executing processors on different data sets.

Both the SIMD and MIMD architectures require communication between the

processors to insure integrity of the solution procedure. The communication can accur

either through shared-memory access via a central switch (bus) or by message-passing

through an interconnected network in a distributed system. The shared-memory, or

tightly coupled, design offers the advantage of minimal communication overhead and

simplified programming, but the bus bandwidth limits the number of processors. With

tightly coupled shared memory, each processor takes the same amount of time to access

a particular memory location. Although distributed processing systems pose few limits

on the number of processors, they have higher communication overhead and require more

elaborate programming techniques.

All computational testing for this research was conducted on a tightly coupled

MIMD Sequent Symmetry S81 (Rev B) at Southern Methodist University. It is config-

ured with 20 Intel 80386 processors and Weitek numeric coprocessors and 32 megabytes

of sharable memory. A Unix operating system is used to schedule the parallel tasks.

1.5.3. Programming MIMD Multiprocessors

Parallel processing consists of concurrent manipulation of data among several

processors. The goal of parallel algorithm design is to minimize a while maintaining

a balance of computational load across all processors. The overall goal of the use of

parallelism is to increase the speed in solving the problem.

24
 Several approaches exist in parallel programming to distribute the work load

among the processors. The two primary methods of performing parallel operations are

functional partitioning and data partitioning [751. The two may be used separately or

in combination with any algorithm on a MIMD machine. With functional partitioning,

the various processors perform different functions on the same data set to complete the

problem. An analogy of this process can be made with the construction of a house.

Several tasks are needed to complete the structure. The foundation must be poured,

framing done, plumbing and electrical work completed, dry wall hung, bricks laid, and

the finished work painted. While some tasks depend on the completion of others, many

can be done concurrently. In the this case, the electricians play the role of one processor,

the plumbers another, etc. each operating differently on different aspects of the data set,

the house.

Data partitioning or domain decomposition is fundamentally different from func-

tional partitioning. In data partitioning, each processor is assigned a similar task to

be conducted on uifferent parts of the data. Again, analogies can be drawn with the

construction of the house. As with multiple processors, multiple bricklayers can work

together to either complete a single wall or multiple walls faster than a single worker.

Again, for successful completion, communication must occur between the different pro-

cessors to insure a correct outcome.

Determining whether to use functional partitioning or data partitioning is critical

to the successful implementation of the parallel algorithm. An incorrect choice could

limit useful speedup because of the added time for synchronization and communication

between processors. Additionally, load imbalances can occur where some processors wait

for others to complete their assigned work. In the DEA study reported here, data

partitioning was used because of special characteristics of the DEA linear-programming

problems; exceptional speedup was accomplished across a large number of processors.

25
 Once a functional-partitioning or data-partitioning approach is chosen, a decision

must be made on how to schedule the tasks on the available processors. Two methods

of scheduling the jobs are pre-scheduling and self-scheduling [75]. With pre-scheduling,

the distribution of work across the processors is set prior to compilation and requires

little overhead. This type of scheduling is ideal when the tasks require the same amount

of time. If the tasks require differing amounts of time, the workload could be severely

imbalanced across the processors. A second method to schedule the tasks is known as

self-scheduling or dynamic scheduling. This method allows each processor to schedule its

own tasks at run time by selecting work from a task queue; once the task is completed,

the processor returns to the queue to acquire additional work. The overhead of coordi-

nating the tasks and maintaining the work queue is more time consuming than with the

pre-scheduled approach, but self-scheduling maintains a more balanced workload across

all processors. For the DEA code described herein, the self-scheduling approach was used

since tasks required varying amounts of time, and workload balance was a high priority

in the algorithm design.

1.6. Objectives

This dissertation presents several advances in data envelopment analysis. First,

a new computational approach for solving large scale problems is described, with testing

presented for both serial and parallel environments. Second, a new DEA modeling and

evaluation scheme is presented that gives deeper insights into the relationships among

the decision-making units than traditional analysis. Finally, a new means of assessing

the robustness of a decision-unit's efficiency is given which spans all current models. All

of the new models described can build on the computational advances to make practical

their implementation in industrial and governmental applications.

Chapter II describes a state-of-the-art parallel approach to solving DEA prob-

lems. By using linear programming theory, a computationally efficient approach is

26
introduced and the solution to previously unsolvable problems will be analyzed. The

parallel approach motivates a new hierarchical decomposition methodology for solving

DEA by first isolating the efficient DMUs and then determining the DEA scores for the

inefficient DMUs. The new approach requires fewer computing resources and improves

solution time by one to two orders of magnitude for large-scale DEA applications. The

hierarchical approach also reduces the need to duplicate effort when solving variations

of the DEA models on the same set of data. This opens the door to a new range of

applications for the family of DEA models.

These computational improvements make more practical a new approach to rank-

ing DMUs as discussed in Chapter III. This approach may be useful to managers by

providing a different "picture" of the efficient surface and sublayers. This will allow

managers to observe more options for improving production techniques by striving not

only for improved efficiency scores, but by improving its rank order among competitors.

The approach will also demonstrate how outliers, which may distort the DEA efficiency

measure, can be accounted for and corrected.

In Chapter IV an improved sensitivity analysis approach is presented which helps

develop a picture for the robustness of the operations of a particular DMU. The moti-

vation for the chapter resulted from the need to know how much the use of inputs and

production of outputs can change before a DMU changes its efficiency status.

Chapter V closes with observations and conclusions drawn from this research

effort. Future research directions are discussed along with a summary of the advances

made in this report.

27
 CHAPTER II
NEW ADVANCES IN COMPUTATIONAL DEA

Data envelopment analysis (DEA) has become an established nonparametric

methodology for assessing the relative efficiency of comparable production and other
decision-making units. As such, it has evolved into an approach to evaluating certain

types of data sets: namely, those that can be viewed as emanating from a (economic)
production function. The range of applicability turns out to be a wide one and, while

the approach is new-relative to other multivariate analysis techniques-its usage and

variety of applications has grown rapidly since its introduction in 1978.

Surprisingly, little has been published on the computational aspects of DEA

[3,4,6,7,9,46,76). Since DEA typically involves the solution of a large number of linear

programs, many practitioners and researchers assume that the repeated use of standard
linear programming codes is sufficient for an analysis. Unfortunately this is not the

case. Specialized codes are needed to correctly handle the preemptive prioritized multiple
objectives (reflecting the models' non-Archimedian infinitesimal) and to coordinate and
expedite the solution of the large number of related linear programs.
The state of the art for DEA codes has also been limited in the size of data sets

that can be evaluated in a reasonable time: typically in the hundreds of data points.
(Codes that handle 1,000 observations have been reported, but run times can be on the

order of days [5].) This st .idy was motivated by large applications we have encountered:
franchise analysis (e.g., over 8,000 McDonald's restaurants, 6,500 Century 21 real-estate
offices, and approximately 5,000 H&R Block tax-preparation service centers), the Federal

28
Reserve Bank's efficiency study of 8,742 U.S. banks, a V.A. study of over 20,000 hospi-

tals, and U.S. Postal Service evaluations of over 30,000 branches. These problem sizes

are clearly beyond the limits of current DEA codes.

This chapter describes a new code for solving large-scale DEA problems in a

reasonable amount of time, demonstrate its performance on a real-world application, and

report on its ability to exploit parallel processing to further accelerate solution time. We

also introduce a new decomposition algorithm that streamlines the solution of problem

sets and provide in-depth computational testing of the code on small- and large-scale

problems, including the largest DEA problems reported to date.

2.1. Computational DEA

Key to the development of software for putting DEA into practice are the math-

ematical underpinnings and associated means of exploiting domain-specific structure for

computational gain. We now briefly summarize relevant DEA concepts, describe efficient

implementation techniques, and present the results of computational testing of a new

DEA research code.

2.1.1. DEA Fundamentals

As described previously, data envelopment analysis is a family of models for

assessing and analyzing the relative transformational efficiency of similar decision-making

units. A decision-making unit (DMU) is any organization, process, or other entity that

uses a set of resources (inputs) to produce a set of products or services (outputs). All

DEA models have the same data requirements: for each DMU j, a nonegative observed

value for each of the s outputs (Y,) and m inputs (Xj), with at least one positive input

and output.

A DMU's efficiency is defined in the tradition of engineering and financial anal-

ysis: the ratio of the weighted total output produced to the total input consumed, or

29
IAYj/vXj. For each DMU in the analysis there is associated a separate linear program

which determines such an efficiency ratio 0, where 0 < 0 <- 1 is typically implied.

The various DEA models separate the set of n DMUs (D) into efficient (E*) and
inefficient (1*) sets. A DMU is efficient if 0 = 1, a* = 0, and 80 = 0. Mathematically, the

members of E* are extreme points, extreme rays, or lie on a convex surface which, when

taken with their associated facets, form a piecewise-linear empirical production surface,
or efficient frontier. Managerially, these units exhibit best practice of transformational

efficiency, relative to their inefficient peers.

Otherwise, a DMU is deemed inefficient (i.e., when 0 < 1, S' 4 0, or s' $ 0),

and the observation lies within, not on, the efficient frontier. In this case, the unit
uses more input for the same level of output (or produces lower output for the same
level of input) as another unit, or some combination of units. For each DMU d E P*,

the corresponding linear program's solution suggests a set of inputs and outputs that
would make d efficient. This virtual DMU is obtained by a projection (not necessarily

orthogonal) onto the efficient frontier, and is expressed as a linear combination of d's

reference set. The reference set contains those efficient DMUs with which d is being
directly compared. The virtual DMU's inputs and outputs are determined as k = YA*

and Xi - XA*.

Although there are many formulations of DEA models (see Table 1.1 and Table

1.2), the useful results in each case are similar in nature and are to be determined in
practice by a DEA code. A complete data envelopment analysis produces, for each

DMU: (1) its efficiency score, 0; (2) a set of weights, JA and V, for the outputs and
inputs, respectively; and, for inefficient units, (3) values for the slack variables sO and

si, (4) a set of reference DMUs, and (5) values for A, from which a virtual DMU can be
computed. In the next section, we describe some of the techniques that have been used
to expedite the calculation of these items.

30
 2.1.2. Survey of Efficiency Techniques

In this discussion, the following notation will be used. We define: D = { 1,... n,

the index set of DMUs; B (BT-) is the set of basic (nonbasic) variables associated with

an optimal basis for an envelopment-form DEA model applied to DMU j E D; 0, 80*,

and 3.* are the values of 0, a, and si, respectively, in B;; Aj = {i E DIAi E Bj} ; A- is

the reduced cost of weight A, in a given optimal solution; E* = {j E DljG = 1, sP* = 0,

and s7, = 0}; and I* = D - E*.

2.1.2.1. Choice of Model Form

For each DEA model (CCR, BCC, Additive), either the envelopment or the mul-

tiplier forms may be used to compute solutions. Since typically (s + m) < n, there will

be fewer constraints and a smaller basis inverse to be maintained with the envelopment

form, used by virtually all DEA codes. This results in (n + s + m) variables and (s + m)

or (s + m + 1) constraints.

Since the number and difficulty of pivots in a linear program grows more rapidly

in the number of constraints than in the number of variables, this choice of the envelop-

ment model becomes an increasingly favorable one as the size of the data set expands.

This advantage can be further exploited as the analysis reveals the memberships of E*

and I*.

2.1.2.2. Early Identification of Efficient DMUs

It has been observed in [6,331 that if i E D and i E B1, j E D, then DMU, is

efficient. This is shown formally, as follows.

Theorem 1. For any j E D, A.7 C E* .

Proof. (CCR version) Let the optimal solution to the primal (envelopment) problem

have a set of basic variables B*. As shown in 141], B* must contain at least one Aj > 0,

j E D. If ,j > 0, then, by complementary slackness, the corresponding constraint of the

31
dual problem must be strictly satisfied and 1Y, - vXj = 0. Hence ,sYj/vXj = 1, and a

multiplier set has been found for which the dual LP associated with DMU, would result

in z = 1, the maximum attainable. Hence, DMUj is efficient. I

This theorem holds for all standard DEA models and has many consequences

for computational efficiency in data envelopment analysis. The early identification of

members of E C E* permits: (1) bypassing unsolved subproblems associated with

members of E, if values of the weights and slacks are not needed; (2) near-optimal

advanced starting solutions for known efficient DMUs; and (3) pricing strategies that

give priority to members of E. The use of observation 1 can reduce the number of linear

programming subproblems to be solved in the analysis, while the observations 2 and 3

can expedite the solution of subproblems that must be optimized.

In software implementations, the status of each DMUj, and its associated A,, can

be easily maintained as: member of E C E*, member of I C I*, or unknown (member

of U = D - E - 1). When the subproblem associated with DMUj is solved, j can be

added to E or I. In addition, from Theorem 1, B; may reveal new members of E-from

the reference set, if DMUj is inefficient, or from degenerate basic variables, if DMU, is

efficient.

Because of the computational value of early status identification, the following

corollary is also useful for streamlining the solution of DEA problems.

Corollary 1. If, at optimality, Ai = 0, i E D, then i E E*.

Proof. Let z* be the optimal solution to the DEA linear programming problem for

DMUj. Let z* = z* - ZkENA&Ak where N is the current set of indices of nonbasic

variables. Let i E N with Ai = 0. By letting Ai enter the basis, z* does not change in

value. By Theorem 1, DMUi must be an efficient DMU. I

32
 This corollary supports an extended search for members of E associated with

the nonbasic variables. Our experience is that there are often a substantial number of

alternate optimal solutions and nonbasic variables with zero reduced costs.

2.1.2.3. Restricted Basis Entry

Another consequence of Theorem 1 is that, since any Ai associated with an

optimal basis corresponds to an efficient DMUj, we need only consider those variables

associated with D-I when pricing. This was publically reported by the authors in [18,191
but discovered independently by Ali and published in (6]. This computationally powerful
finding permits continual restriction of the pricing operation; as inefficient DMUs are
identified, their corresponding decision variables are dropped from subsequent problems.
This narrowing of focus to the D - I set has a ratchet-like effect on the size
of the linear programs being solved. As the membership of I increases, the effective
sizes of unsolved subproblems (in terms of number of variables) become progressively

smaller as well. Hence the "later" subproblems in a series tend to solve more quickly
than the "earlier" ones. As will be shown, this has a dramatic impact on the solution
effort required, especially on large-scale problems.

2.1.2.4. Candidate List

Candidate lists, and other multi-pricing schemes, are standard computational

procedures for expediting the solution of linear programs [64,67,71,72,74]. These basically
involve pricing a set of nonbasic variables to form a short list of promising candidates for

entry into the solution, and repeatedly selecting incoming variables from the list (re-
pricing members between pivots) before replenishing with fresh candidates.

33
 This mechanism may be specialized to the structure of data envelopment analyses

as well. First, the restricted basis entry procedure eliminates from consideration variables

associated with I. Further, priority for list inclusion can be given to members of E over

U, since any basis will contain only slacks and members of E* (61.

2.1.2.5. Degeneracy and Anti-Cvcline Logic

Note that if DMUj is efficient, the optimal basic solution for the envelopment

model can be comprised of 0 = Aj = 1, with all other variables equal to zero. Ac-

companying such highly degenerate solutions is a greater likelihood of degenerate pivots,

as when the minimum ratio is zero. When such degeneracy exists, so does the possi-

bility of cycling, or a repetitioD of bases for the same extreme point solution leading

to non-convergence. Although practitioners often dismiss this possibility as remote, our

experience indicates that cycling in DEA codes is not only prevalent but likely, in the

absence of anti-cycling procedures.

Various methods have been proposed for the avoidance of cycling within the

simplex method for linear programming. The leading approaches-lexicographic ordering

130], perturbation [301, and Bland's rule [26J--can be computationally demanding and

hence are usually invoked only under certain conditions. Some researchers support

the use of lexicographic ordering in order to avoid a temporary stalling of the solution

process, noting a substantial decrease in the number of pivots when employed 146]. Ali

(61 proposes a specialized ratio test that gives preference to A variables over slacks when

breaking ties for zero minimum ratios, and yields decreases in the number of pivots and

solution times on reported test problems.

Our experience indicates that, for many problems, simple scaling of the problem

data-by variables and by constraints, normalizing by averages 125]-is most effective

34

41
in lowering the incidence of stalling and cycling 150,64,83). In those cases where a lack
of progress is detected, the invocation of the lexicographic ordering procedure quickly

remedies the situation.

2.1.2.6. Non-Archimedean Infinitesimal. e

The non-Archimedean infinitesimal, e, in the objective functions of envelopment-

form models was represented in early DEA codes using relatively small positive values

(e.g., 10-6). Unfortunately, the effect for CCR and BCC models was to make the
results of any analysis dependent on the value chosen, leading to incorrect results in
many cases [9].

More recently, algorithmic changes have been proposed to avoid such data depen-
dencies and ensure correctness of envelopment analyses. Specifically, the CCR and BCC
objective functions are recast in a preemptive priority form: minimize P10 + P2(-18i
ls). Proposed implementation schemes include two-stage optimization [3], and a multi-
objective pricing form of the simplex method [46]. This approach appears to have re-
solved the infinitesimal issue in the DEA community.

2.1.2.7. Preprocessing of Observations

An analysis of the DMUs' observations can lead to identification of members of

I and E prior to the application of optimization. Also, in some instances, the likelihood
of encountering efficient or inefficient units early in the solution process can be increased
when DMUs are processed in a particular order.
As shown in [6], DMUi is inefficient if it is dominated by any DMUj, that is if
Xj _5X 2 and Yj >_ Y,. In the BCC and additive models, DMUj is efficient if one of
its outputs (inputs) has a unique maximum (minimum) value, or it has the largest ratio

Oj = Iy/1xj.

35
 Processing DMUs in Oj order is reported to identify members of E* earlier than

random order. This may be helpful, especially when the number of efficient DMUs is

small, but it is unclear whether it is generally preferable to solve subproblems associ-

ated with E* or I* in the early part of the process, since solutions to inefficient DMU

subproblems tend to reveal more members of E* than the solutions to efficient DMUs,

and subproblems for members of E* tend to be more degenerate and are more prone to

cycling.

2.1.2.8. Advanced Starting Bases and

Reoptimization

In many algorithms that involve the solution of a series of related linear programs,

or subproblems, the use of reoptimization can lead to substantial time savings. This is

accomplished by using B* as an initial basic feasible solution for the LP associated with

DMUj, as is possible when subproblerns differ by one or more parameters (objective

function or right-hand-side values) but otherwise have the same mathematical structure.

DEA subproblems can employ reoptimization, but its effectiveness is dependent

on the mathematical "closeness" of the two solution points, B* and B. Ali 171 suggests

the use of this technique only for j E I*.

2.2. PIONEER Code

We have developed and tested PIONEER, a new DEA code which embodies most

of the above efficiency techniques. The objective was to build a flexible research code for

testing and evaluating alternative solution approaches to large-scale problems.

2.2.1. Implementation Overview

PIONEER's optimization kernel is based on the XMP library, a collection of

portable, reliable, and widely used Fortran subroutines for the solution of linear pro-

grams using the primal and dual simplex methods 1701. The LP basis is maintained

36
in LU-factorized form with pivoting routines from the Harwell library. The availability

of source code permitted customization of algorithmic steps for DEA problems, and

the clean, documented, modularized design of the routines provided a flexibility that
supported and simplified experimentation.

Anti-cycling logic was added to the XMP ratio test. While the aforementioned

data scaling procedure substantially reduced the incidence of stalling, lexicographic or-
dering is invoked when lack of progress is detected.
The PIONEER code implements all varieties of DEA models described in 181]:

CCR, BCC, and additive with both input and output orientations and two variable-
returns-to-scale models allowing for either increasing or decreasing returns to scale. All
of the efficiency techniques in Section 1.2 are employed, except for preprocessing and
reoptimization. (Our testing showed that reoptimization was of uneven value--sometimes
reducing, sometimes increasing, run times-hence was not used in the final testing.) The

candidate list is not prioritized, and the two-stage optimization approach to the non-
Archimedean infinitesimal issue is employed.
Auxiliary data structures maintain the E, I, or U status of each DMU, as indicated

by the following outline of the code's logic:

Procedure PIONEER
1. Initialize U = D, E =0,I=0.
2. While U $ 0:
a. Select i E U.

b. Solve the subproblem for DMUi.

c. If DMUi is efficient, E = E U {i), else I = I U {i}.

d. Update: E= EUAU{jl-j =0}, and U = D-E-I.

The strength of step 2d depends on Theorem 1, Corollary 1, and the distribution of

observations in R("+m), relative to the efficient frontier.

37
 It should be noted that while the XMP data structures are designed for sparse

problems, DEA subproblems are almost totally dense. This results in an unnecessary
processing overhead in the current implementation that can be eliminated with recoding.
Hence the execution times reported in this paper should be considered worst case, or

at least conservative, measures of performance, relative to a fully optimized and tuned

implementation.

2.2.2. Baseline Computational Testing

To evaluate the efficiency of the PIONEER research code, and to provide a

set of baseline measurements from which to compare algorithmic and implementational

enhancements, a series of test runs were made on medium- and large-scale problems.
Testing with and without individual efficiency techniques lent insight into their impact
on solution speed.

2.2.2.1. Test Data

Although the code was validated for solution accuracy on small problems from

the open literature, of primary interest was its effectiveness on large-scale problem sets.

For testing purposes, real life data, randomly generated data, and data based on the
economic Cobb-Douglas functional form were employed.
First, the Federal Reserve Bank of Dallas provided a data set of 8,742 banks

from its Southwest district, with variables consisting of the six inputs and three outputs
described in [20,211. This challenging problem from industry was clearly beyond the

state of the art of DEA codes, and was a prime motivator for this research. Testing was
performed on smaller subsets by selecting the first n observations from the unordered file.
Randomly generated observations from a multinormal population provided a sec-

ond source of test data. Using the DRNMVN routine from the IMSL library and the
procedures given in Appendix A, large data sets could be easily created. The variables

38
were partitioned into 3 inputs and 4 outputs so as to observe positive correlations be-

tween the "input" and "output" sets.

Finally, a random problem generator, DEA-GEN, was written to create obser-

vations based on the classic Cobb-Douglas form of production surfaces. As detailed in

Appendix B, the program gives the user a measure of control over the proportion of

efficient points, and creates data sets that more closely approximate realistic economic
processes than the multinormal generator.

2.2.2.2. Test Environment

The PIONEER code was tested on Southern Methodist University's Sequent

Symmetry S81B with 32MB of internal storage and processing units consisting of 16-MHz
Intel 80386s with Weitek coprocessors. The software is written entirely in Fortran and

executed under Dynix 3.0.12, a BSD-Unix-based operating system. While the processors
are rated at 4 million operations per second, in terms of current technology they are
equivalent to relatively slow personal computers.

2.2.2.3. Experimental Results

The PIONEER code was applied to problems from each of the three sources. The

solution times reported below are "wall-clock" times, or elapsed real execution times,

exclusive of problem input and output on the Sequent Symmetry. Unsolved subproblems
associated with members of E (as determined by Step 2d) were bypassed.
Table 2.1 describes test set A, which was made up of banking, multinormal, and

DEA-GEN problems. Three problems from Federal Reserve banking data were examined:
the first 1,000, the first 2,000, and the first 8,000 DMUs of the 8,742-bank data set. Two
multinormal data sets, for n = 1,000 and n = 2,000 were generated using the variance-

covariance matrix and procedures given in Appendix B. Six problems, with n = 1,000
and n = 2,000, were created using DEA-GEN and the parameters given in Appendix C.

39
 Table 2.1.-- PIONEER solution times on test set A

Source s m DMUs No RBE Time RBE Time Ratio

FR Bank 3 6 1000 33.29 min 17.04 min 0.526
FR Bank 3 6 2000 179.91 min 95.59 min 0.532
FR Bank 3 6 8000 44.21 hour 19.80 hour 0.448
Multinormal 4 6 1000 31.96 min 18.72 min 0.586
Multinormal 4 6 2000 106.90 min 61.65 min 0.577
DEA-GENa 4 6 1000 57.51 min 33.13 win 0.576
DEA-GENa 5 3 2000 130.02 min 75.94 win 0.584
DEA-GENb 7 4 1000 61.49 min 37.96 win 0.617
DEA-GENb 7 2 2000 151.27 win 92.42 win 0.611
DEA-GENc 6 6 1000 72.83 win 43.29 win 0.594
DEA-GENc 6 5 2000 213.61 win 126.50 win 0.592
Average 0.567

Solution times for these problems with the PIONEER code and the CCR model are given
in Table 2.1.

The code was run using the CCR0 model with and without the use of restricted

basis entry (RBE) and early identification of efficient units (EIE) to examine its impact
on solution time, as shown in Table 2.1. In all cases, the RBE procedure had a strong
impact, cutting solution times roughly in half. The 17.04-minute time for the 1,000-
DMU problem indicated that the PIONEER code is reasonably efficient for medium-

sized problems. But, even with the help of RBE, the 19.8-hour solution time for the

8,000-DMU problem is excessive for practical usage.

A closer examination of the 8000-DMU bank problem (BANK-8) solution process
gives insight into the sources of the speed improvements. Figure 2.1 gives the time to

solve each set of 1,000 subproblems in the 8,000 total, both with and without RBE logic.
Note that when RBE is not used, the time to solve each set of 1,000 subproblems is

roughly the same (around 4.1 hours). But when RBE is employed, the last group of
1,000 subproblems is solved almost four times faster than the first group. By restricting

40
the known inefficient DMUs from entering the basis, the later subproblems are smaller,
and easier to solve. In addition, the early identification of efficient DMUs results in fewer

subproblems to be solved. Typically, fewer than 20% of all DMUs are efficient, so the
faster solution time can be attributed mainly to restricted basis entry.

4.4 o RBE/EIE 0 RBE/EIE

3.9,T

3.31,

2.81 1 1 I 0

S2.3

- 1.8,

1.31

.
0.8

1000 2000 3000 4000 5000 6000 7000 8000

Number of DMU's

Figure 2.1. Solution time for subsets of BANK-8

Figure 2.2 shows the cumulative effect of RBE/EIE on solution time for the 8,000-

DMU bank problem. While the performance improvement is less pronounced in the
earlier subproblem groups, the 2:1 ratio becomes evident in the last few groups, and

trends indicate an even greater disparity might result for larger problem sets.
Although PIONEER's solution times are encouraging and indicate that the code
may be comparable to others described in the literature, other performance improve-

ments are possible. The next section describes the use of parallel processing to further
decrease solution times for these and other problems.

41
 45-

40 NoRBE/EIE
-0 RBE/EJE ... oO~-"'
35-
30--
25--

5-
III !I I I
1000 2000 3000 4000 5000 6000 7000 8000
Cumulative Number of DMUs

Figure 2.2. Cumulative solution time for subsets of BANK-8

2.3. Parallel PIONEER Code

Because of the nature of the DEA formulation, an advance in computer technol-

ogy, parallel processing, can be utilized to achieve dramatic reductions in solution times.

With parallel processing, the solution for a given problem is found by simultaneously

coordinating the efforts of multiple processors to accomplish a given task on a common

body of data. If the algorithm associated with the problem solution process consists of

multiple independent tasks that can be properly assigned to separate processors, dra-

matic reduction in solution times may be possible.

Of the numerous varieties of parallel machine architectures, the most prevalent

commercial design is multiple-instruction, multiple-data (MIMD) 119,56]. Parallel com-

puter systems are composed of multiple processors which have the ability to act inde-

pendently on problem data sets. To arrive at a correct solution, coordination of all

processors and their tasks is necessary. Two basic techniques are used to accomplish

42
this coordination activity in the parallel environment. The processors can communicate

through shared memory. If the time to access a particular memory location is identical

for all of the processors, the system is called tightly coupled, or else, the system is termed

loosely coupled. If the parallel machine contains no shared memory capabilities, coordi-

nation of processor activities must be accomplished through message passing between all

of the processors.

As with traditional single-processor (serial) machines, solution efficiencies are

directly tied to how well the algorithmic steps match the architecture of the underlying

machine. Because a DEA problem involves the optimization of many separate linear-

programming subproblems, the use of MIMD-style parallelism to speed solution appears,

on the surface, to be a "natural" one. In fact, the mapping of the DEA solution process

to a tightly coupled MIMD architecture turns out to be an ideal example of the use of

parallel processing.

2.3.1. Parallel Code Design

The application of parallel processing to DEA problems was first reported in

Phillips, Parsons, Donoho [76], where four transputers were run from a Macintosh IIcx

on a 54-DMU problem. Times were reduced by a factor of three in this loosely coupled

MIMD implementation. The next section describes a very different computing environ-

ment and how the PIONEER code was modified to use this form of parallelism.

2.3.1.1. Target Machine Environment

As with software designed for vector processors, parallel codes must be structured

to match the architecture of the target machine. Our test machine was the same Sequent

Symmetry S81B that was employed for serial testing, but which can be programmed for

parallel processing. The system has a tightly coupled MIMD design, with 20 16-MHz

80386 processors, Weitek coprocessors, and 32MB of sharable memory.

43
 The Sequent's operating system permits the programmer to create multiple, in-

dependent processes which it schedules on the available processors. The processes can

have both private and shared data-spaces. The shared spaces can be read and written by

all designated processes, and can be used for interprocess communication and to avoid

duplication of common data.

2.3.1.2. Parallelization of the DEA Alkorithm

This type of parallel machine is designed to be most effective with work that

can be decomposed into large, independent tasks. The DEA solution process can be

organized in this manner through the use of data partitioning, where multiple identical

processes apply the same program steps to different data. By considering each DMU's LP

subproblem to be a separate task, such large-granularity work decomposition is possible.

In the parallel PIONEER code, a self-scheduling approach is used, where pro-

cesses select and execute tasks from a shared work list, on a first-come-first-served basis.

Although incurring a minor amount of coordination overhead, such self-scheduling per-

mits a balanced distribution of the workload across processes, an important characteristic

when individual task times vary.

Each process solves its LP subproblems in its private memory; shared memory

stores the original problem data and the status of each DMU. Since a DMU's status-in

terms of membership in E, I, or U-may change, restricted basis entry becomes dynamic

and time-based. At one moment, a given Aj variable may be part of the E U U pricing set

and, an instant later, be found to be inefficient and ineligible for basis entry. The shared

status array automatically communicates this information to all processes. This is an

instance of a race condition, or timing-dependent code, which can result in stochastic

solution statistics when the order of events differ from run to run due to minute timing

differences.

44
 2.3.2. Testing Parallel PIONEER

To test the parallel implementation of the PIONEER code, a variety of problems

were chosen for analysis. Table 2.5 describes the characteristics of the problems chosen.

The original bank data consisted of 8,742 DMUs. For each of the bank problems, the

DMUs composing the data set were randomly chosen from the original data set. The

multi-normal and DEA-GEN data sets are described in Appendices A and B. These

problems were generated to simulate real life data of large-scale DEA problems. The

CCR1 model was used for all runs.

The "wall clock" time, measured in minutes, to solve each DEA problem, exclu-

sive of input output times, are given in Table 2.2. For the problems consisting of 8,000

and 4,000 DMUs, the number of processors used was limited by the available memory.

Tables 2.3 and 2.4 reveal the speedup and efficiency of the parallel approach. In all

cases, relative speedup was nearly linear. Isolated examples indicate that superlinear

speedup may be possible because of the use of RBE/EIE in the PIONEER code. Since

the PIONEER code is asynchronous, i.e., each LP can be solved independently of all

other LPs, the parallelization is highly effective at improving solution times.

Table 2.2.- Test problems parallel run times (min.)

# Procs Bank Multi-Normal DEA-GENa DEA-GENb DEA-GENc

1000 4000 8000 4000 2000 2000 2000
1 17.04 283.22 1189.81 242.96 75.94 92.42 126.50
2 8.70 151.23 589.22 122.35 38.20 46.52 63.63
3 5.68 101.42 389.46 81.81 25.58 31.06 42.48
4 4.27 73.26 61.39 19.19 23.39 31.91
5 3.40 61.27 49.14 15.35 18.71 25.57
6 2.86 12.82 15.60 21.32
7 2.45 11.02 13.38 18.28
8 2.15 9.65 11.74 16.02
9 1.96 8.57 10.45 14.26
10 1.72 7.72 9.40 12.80
15 1.17 5.19 6.33 8.61

45
Table 2.3.- Test problems parallel speedups

# Procs Bank Multi-Normal DEA-GENa DEA-GENb DEA-GENc

1000 4000 8000 4000 2000 2000 2000
1 1.00 1.00 1.00 1.00 1.00 1.00 1.00
2 1.96 1.87 2.02 1.99 1.99 1.99 1.99
3 3.00 2.79 3.06 2.97 2.97 2.98 2.98
4 3.99 3.87 3.96 3.96 3.95 3.96
5 5.01 4.62 4.94 4.95 4.94 4.95
6 5.96 5.92 5.92 5.93
7 6.96 6.89 6.91 6.92
8 7.93 7.87 7.87 7.90
9 8.69 8.86 8.84 8.87
10 9.91 9.84 9.83 9.88
15 14.56 1 14.63 14.60 14.69

Table 2.4.- Test problems parallel efficiencies

# Procs Bank Multi-Normal DEA-GENa DEA-GENb DEA-GENc

1000 4000 8000 4000 2000 2000 2000
1 1.000 1.000 1.000 1.000 1.000 1.000 1.000
2 0.979 0.936 0.979 0.993 0.994 0.993 0.994
3 1.000 0.931 1.000 0.990 0.990 0.992 0.993
4 0.998 0.966 0.989 0.989 0.988 0.991
5 1.002 0.924 0.989 0.989 0.988 0.989
6 0.993 0.987 0.987 0.989
7 0.994 0.984 0.987 0.989
8 0.991 0.984 0.984 0.987
9 0.966 0.985 0.983 0.986
10 0.991 0.984 0.983 0.988
15 0.971 0.975 0.973 0.979

46
 2.3.3. Limits of Parallelism

The computational testing indicated that the parallel PIONEER code was a

highly scalable MIMD application that fully utilized the multiprocessing capability of

the given computer system. We feel that the software would also exhibit much of the

same efficiency if implemented in a loosely coupled MIMD environment. In the latter

setting, changes in a DMU's status would have to be broadcast to all processors, thus

incurring additional overhead and a latency in communicating the information. While

unnecessary work (i.e., avoidable in a shared-memory environment) might result, the use

of a larger number of processors could more than offset this disadvantage.

In our testing, memory size limited the dimensions of problems that could use the

full parallelism of this system. We believe that additional internal storage would permit

the same excellent speedups on the larger problems as was observed on the smaller ones.

(This notion was verified by preliminary testing on a larger system.)

Even with these encouraging results, we felt that further improvements were

needed and possible. In fact, a close examination of the parallel solution statistics led

to a new procedure which further reduced all times-both serial and parallel-by up to

an order of magnitude.

2.4. Hierarchical Decomposition

Experimentation with sets of problems that were identical except for the number

of DMUs yielded solution times such as those given in Figure 2.3. Not only did the

memory requirements of larger problems limit the amount of usable parallelism, but

run times grew exponentially in n. Hence if a larger problem could be decomposed

into a series of smaller ones, lower memory requirements, greater parallelism, and faster

individual solution times might offset any additional work that might be required.

47
 1000 2WOO 3M0 4000 5MW 6O00 7600 8OOO
Number of DMUs

Figure 2.3. Solution time versus problem size

2.4.1. DEA Decomposition Background

Consider a partitioning of the set of DMUs into a series of k mutually exclusive

and collectively exhaustive subsets Di C D, where D = UCEK Di, ReK Di = 0, and

K = {1,..., k}. Define E(D1 ) and I(D 1 ) to be the index sets of DMUs in Di that are

efficient and inefficient, respectively, relative to Di (i.e., Di = E(D,) U I(D 1 )), based on

the meaning of efficiency in the DEA model of interest.

Theorem 2. If Di g D, then I(D1 ) C I* and E* C UiEK E(D,).

Proof (C--CR version). Define z;(Di) to be the optimal value of z for DMU j relative to

the set Di. We know that z(D) < z;(Di),Vj E Di in this maximization problem, since

there are more model constraints associated with D than Di.

Assertion I(Di) C I*:

DMU j is inefficient (efficient) relative to set S if z,*(S) < 1 (= 1). Assume j E Di has

zj*(Ds) < 1, and hence is inefficient. Since zj(D) _<z;(D1 ), j will be inefficient relative to

D.

48
 Assertion E* C- Ui'K E(D,):

Since E* C D = UiEKDi, then E* C U, (E(D,) U I(Di)). From above, E* n uI(D,) = 0,

therefore the assertion must be true. I

Hence if a DMU is inefficient in a subproblem, it is inefficient for the full prob-

lem; if it is efficient for a subproblem, it may or may not be efficient overall. These

relationships provide the foundation for the following decomposition procedure for DEA
models.

2.4.2. The Hierarchical Decomposition Procedure

The following approach to DEA problems solves a series of subproblems which are
arranged in a hierarchy. The subproblems are themselves DEA problems created from
subsets of the DMUs in D. They can be solved independently, but information about set
I membership generated from the subproblems can accelerate the solution of others.

First we define a procedure for creating and solving the subproblems. It partitions
the set of DMUs whose efficiency is unknown (U) into a series of subsets, or blocks, of

size b. Recall that U = D - E - I and IUI is the cardinality of U.

Procedure SolveBlocks(b, e, I):

1. Partition U into k = rFUi/b] mutually exclusive, approximately equal-sized blocks

of DMUs.
2. For each block Bi, i E K, K = {1,..., k}:
(a) Apply a DEA envelopment model to compute Ej = E(Bi), using early
identification of efficiency and restricted basis entry.

(b) Set I =IUI(B,).

This is used in the following hierarchical decomposition procedure (HDEA). A

graphical representation of the process is given in Figure 2.4.

49
 Blockj Block 2 Block 3 -------------------------- Blockk

EiE

H2

Figure 2.4. Hierarchical decomposition schematic

Procedure HDEA(b, a, 0):

Level 1: (Initial pass)

1. 1 =O, U = D,t +- 1.

2. SolveBlocks(b, f,I).

Level 2: (Identify E* and 1*)

while ( U $ 0 ) do:

1. f -'-+l.

2. u +- IUI.

3. SolveBlocks(b, fj).

4. if IUI/u > a, b - IUI, else b -- /3b.

Level 3: (Resolution of I*)

Re-solve the DEA model (with basis entry restricted to E*) for members of I* to

compute correct solution values.

50
 2.4.3. Implementation Considerations

HDEA is a divide-and-conquer algorithm similar in nature to merge sorting and

sorting networks [49,661: the original problem i, broken into several smaller subproblems,

which are solved recursively and the results cu,.bined to solve the original problem.

The blocksize parameter, b, defines sets of linear programs with the same number of

constraints as the original problem, but many fewer variables. This results in lower

memory requirements and faster solutions for the subproblems, although more linear

programs may be involved. Because of this increased speed and Theorem 2, these

easier problems can eliminate inefficient-DMU variables earlier than with non-hierarchical

approaches.

The HDEA procedure focuses initially on isolation of E* (in levels 1 and 2),

so as to expedite solution of the subproblems associated with P* (in level 3). The

method should be particularly effective when IE*I < JI*, as is typically the case.

The decomposition into subproblems with minimal data communication requirements is

highly attractive from a parallel processing standpoint.

Memory requirements are a function of max{b, IE*I}. If primary storage is at a

premium, its use can be minimized by paging in DMU data from external storage for

each subproblem separately. In parallel implementations, participating processes might

place in shared memory a copy of the X and Y data, plus a DMU-status array; each

process would also need its own private storage for solver-specific information such as a

simplex basis inverse and candidate list.

The choice of blocksize also affects overall solution time. Since b influences the

tradeoff between the size and the number of subproblems solved, computational testing

for an appropriate setting is required. Figure 2.5 shows the total solution time in minutes

for various values of b on an example problem.

51
 820
800
780
"760
740
.720
-'700
680
660
640
620

50 75 100 150 200 250 300

BlockSize

Figure 2.5. Total solution time vs. blocksize for BANK-8

2.4.4. Computational Testing

For testing the HDEA procedure, real life data, randomly generated data, and

data based on the economic Cobb-Douglas functional form were employed. For each
test problem, an array containing the status of each DMU was maintained in shared
memory. A private copy of the appropriate data was given to each processor in parallel
so that each DMU could be solved independently. Since the status of each DMU could be
updated without conflict from other processors, no locks were needed during the status

update and as a result, the parallel implementation is purely asynchronous.

All times reported axe wall clock times, in minutes, to solve the problems ex-
clusive of any input/output. The single processor times represent the best achieved
serial times across the differing-sized DEA problems. The parallel case is a direct im-
plementation of the serial formulation, requiring only a few modifications for the parallel

environment to accomplish the self-scheduling parallel implementation. It is important

to note, none of the runs were made in a dedicated environment, and hence, times are

subject to system load. However, precautions were taken to conduct the runs during non-
peak hours so as to minimize the confounding of outside factors on the solution times.

52
 m~000 Me0M0 C4 ) t (M0000e
t"
Uat:a Coqq q O

0O 4 C4 -4 C4
-4 LO~' C4 CD (
m t- ~o6
o6 o6 t- o6 eq
w v o6o6 o6 o o6m 6
-D 4
oo% 4M-4 C4C400 44 4C

(12 eqtU R 0 -0 r- 0-4t0Co -400

- .4 4 C4 C4

0o t- s 00Co;o 0 m= Co000m
0t-0
0nC4-V - 4O)l C -t-C0 D
0 - 0~C0-4t- 4

-4 1 M-40 -4'" 00 t-t- -MC

412 cq L6; o6 6o6 L6 -i vi - L ;oo vi 000

Cq C4
CD
00c q 0 q O0 4C
Cc -44C cnMC
-44-
q m C l eq - -0 o LaCDrD 0
fDtf = mC4 Ceaq o t-4V-4CD 0) 004

LO I"C o00C 00 4 0-4 t- 4Vt

O4 V~ A -4 U00ho m C4 U.) C4 0 M ODt-e t- 0-4m t

C) .4
-4C 0t 4 4( M( 0-e U'3~ C.4 4m ( UM qm
U* "oL 00100m 100m 0 - 000 C

o.4 M~ to
-4 -4 q A -4 4 -Ot -4O

m CD

v-C 0000L oC4-uj 00m000t 0mt

C) " ( 0 04 t 4M 04C4
Q (

m -o m0 A 0t -C
F-O _ _Oo_ 00_ Vt_ 4C -LI ot _Om _

cl C CC4 4 OC4oom00It o53C

 For the test problems shown, a blocksize of 250 was used for the 8,000 and 4,000

DMU cases, 125 for the 2,000 DMU cases, and 100 for the 1,000 DMU cases. The effects

of different blocksizes must be investigated, hence, these results should not be viewed as

the best possible. During level 2 of the HDEA procedure, /3 = 1.5. As a result, each

subsequent block at level 2 grows by 50% until all DMUs have been classified as either

efficient or inefficient.

Table 2.5 shows that for all test runs the speedup was nearly linear. Larger more

difficult problems resulted in better speedup than smaller problems. With the smaller

problems, the overall solution time in parallel is so small that the overhead of maintaining

the self-scheduling tasks, as well as possible workload imbalance between processors,

becomes apparent. Consequently, although the parallel results for small problems are

quite good, the parallel implementation will be more efficient for large, difficult problems.

Table 2.5 also gives the total number of pivots and pricing operations executed to

find the DEA scores for each data set. These numbers vary with the number of processors

because of the method used to update the status of each DMU. As expected, problems

with a greater number of DMUs, and those with a larger density of efficient DMUs,

require more pivots and pricing operations and, consequently, take longer to solve.

Tables 2.6 through 2.10 contain the results of each test problem for each level

of the HDEA procedure. The speedup remains relatively consistent across all problems

across all levels and is nearly linear.

The effects of using early identification of efficient DMUs to reduce the number of

linear programming problems that must be solved are shown in Tables 2.11 through 2.15.

With the HDEA procedure, the total number of linear programs that must be solved

is slightly more than twice the original number of DMUs contained in each data set.

However, since the HDEA linear programs are smaller than other DEA methods, overall

performance is improved. EIE improves performance further by reducing the number of

54
4)
t- 0 4 m U'
w m m

-4 4 o -4 t- M
40 (

tz~ Go 0 00 -0 0

-4j ". -e4 LO m m

10 oo mto c

c-4 oi -4a6c4 t

t) m 00 0~ uo " 00 -0 m "-4v

-
o'jt-U

Go ' -414

LO C r co

-R- - - - -

__ CD 4 CD_ 0_00t _mr_QC

!) to' AW W - "a L
m.-4 0

55- L
 -4hf C M C4

i -! 'IT'r

Noo
2 o0 00M0
t- 6Za 6-6d 6d ;o

-4;M 0 4 t t- M "'M 4

0 V 10V.4

0~ 0 1"0
C>~a O t
o QV
L q T q - c4o '-T -4

-4 C4 V CD

-l 10 p4 O c'00--UM

4 -4 -4

v-M -4 4

C4 11
cqc
ho I c'
_ 1 _ -4 -4

4)4
z0
56mUDt
0- 4M0
 0

o o V -4 U.

U ~ ' ~ 0 - 00-

0060000
ci 6 00 C

r-L~ 0~ CD -4 w0 M

0lo qC 0V M XMt

0-t- w mw6o

LO a) -4t-

0 ~~ ~oa)t ~ ~0t-m4~0 mt 040

0C)4 4 - M b C1 b-00 M M

v-CD -400 ao- ( UD -aCOC 0r-

mV
+m -- 4 1-4 C40 4m t0 ~00

t-q 4Q " - t~~0

-

1- C - CC00t- m v mt
S 04 -r- 4wc-4 0 4C
0 Om4- .4 o

o m4=m
~1-4 ~-4. "0
O0 =
1- 0ot-
00 m 00
m -4
.0=$ ' 4C O 0 C o V M I t-

w -4 w 4 "C14- 0 00 C4 k
_ _- -4 o t r-

*0000 co l r 4 0 1 -c

% _ - 4 ;5. 'o q

0 IUOD8 U-3L, t-00a

LO
00 0 - D57) 0 4 C
 0 t- U13 00 M~ m 1- 00 LO 0 0 CD

UDU

>1 4C~~. C4 Mt

U ~ t ~ cdm co~
L-c

t 0 0'l C) aj 0 0t0000a mG

1-4 W M -4 10 -4 O -4 t-

-40C l0C 00 a

S-4- 00 t 4-0

04 COC 4COt wI
o4 (Mc 0 m t D
rI C-4 -4 -L -4

-4W00( C4 t- 04 4L M W

0)
C4ot- qw0
oooh~-to k~ CCO
-4 1"_ t_ 0 t-4 -4

to 4w o 00

P!4 C'
Lo- -0q 0 o 'og M
CD Lo P, uw)C
.0to

t- __D t__ o- cmCO :

CoC4- t -gCD8;jt

w -o L 58C
I"
 4) 000000l-000 000U b

t- to eq t.4 0O eqt-
m 00

-~C C -4C~-
000

>1~~ ~ 4 ~ t-0 -4q O(

0 -:o ) 7!C!C !o

t-WmWoMMM ==

'.4 eq=
-l ..-
.0 0t O)0 C.. C4 t- M

o -W cz m e =

t-~
teqC

oq
00,q C Ci
h q
Lo r-4 co m "-4L L
-4 -4 -4 C h C C C)

LO "4 U13 -4

-4 - -
L q f -4

m ~ e _ _ _4IV_

59-
80 0 O14
linear programs at level 1 and level 2 by 10-25%. Early identification has no effect at

level 3 since only DEA scores for inefficient DMUs are found at that level.

Table 2.16 shows the dramatic effect of the hierarchical procedure on improving

solution time for the 8,000 DMU cases over the best DEA approach reported to date.

Only three processors are used in the parallel implementation because of memory limi-

tations of the non-hierarchical procedures. The HDEA procedure results in a 6- to 12-

fold increase in speed over the non-hierarchical procedure when both cases are run with

the same number of processors. Because the HDEA procedure requires less memory,

permitting use of more processors, the cases with 15 processors are 75 to 125 times faster

than the serial non-hierarchical procedure.

Although the test problems were limited to 8,000 DMUs because of the availabil-

ity of real life data, the HDEA procedure can accommodate much larger problems. A

Cobb-Douglas test case (Appendix B) consisting of 25,000 DMUs with 3 inputs and 3

outputs was solved using 15 processors in 19.26 minutes. Because of memory limitations,

it was not possible to solve this problem with the non-hierarchical procedure.

As seen with the test problems, the HDEA procedure allows for the solution of

large-scale DEA problems much faster than previously reported DEA methods. When

coupled with a parallel environment, the HDEA procedure yields solutions to problems

in a matter of minutes which previously, would have taken days. Additionally, the

HDEA procedure allows for the solution of very large-scale DEA problems that remain

unsolvable (in a practical sense) with any other reported approaches.

2.4.5. Further Advantages of the HDEA Approach

Accompanying the HDEA method's advances in sheer solution speed are a variety

of additional computational advantages in other settings. The hierarchical structure can

be exploited when solving multiple DEA models and further streamlining opportunities

await exploration.

60
o 4 4C o V4' -U) mqr0 qc

4-4

Q Q1 f -V4V4( - -4 0 0 C

C.) M DU' MV4l

~OC14-4(MM0
U8l t4 Ov- C400 C4 C40 MM
4-D C C CD eOOO-4t-

Q0C40 O,- -
00 Ot'- CD

-r-

-4 r .4 00 4
__ I"_
CD M_ m _ _ t___ ___

ID Cqts 61
 o 4 -4C 4 4V4 -- 4 V MM

00 Nmm = D -40 w4

C-4 m m00 - O tO to

t- WCq
C V-4
MM Cq O t-= -4

Cq M
- M ce) 0 t--4 -4

4
04 ao'
cm-. ( "c4-4 ot4 - - t 4C
-4 OO,-4c)0
MO4O -- -C

C*4 1 moo 0-4 -4

-4

e-4 0 4 0 00 4 t

~ ~~
H __ _ - ___ 4_

DCC4
CD00" M M 3 W DL62O-
 00 t- 00M

tR 10 000 DC -

o9

-4 v 0t-V4
4
0"C
to t- m
8-~ t-~ ' D0lc
v-4 c
OI- -t-C t

C.) -4- 4m 0

.M m o00 t- -4-

63
 o CDt OU. m 4( - (
(D - C4 = AA (

0 C4 t cz C4t-cv

to LO00
m l) a
00t- e

- ~O~r- ~ CCCO v

LO o 1-4 LO LOQ( D 0ot

C.) CD o -V O 0 -)C

(D -.4 14 C14e ,4"

- u C
m00 00C4 t~-00

C 0DO
oo o c.o c, 00 Vto c, co ~t-m
""-4 9-4 Lo V, m C4 t-00

T T,0>>

-0C 00

64
 t,-- eq C4 q A % t-

P.4 -4m j -t- -4

r-
V cc wt- C-
C14 to -4 - C

Otht

( .4-4 C4e C4- 4MI 4C~4 t

oo4C t

-40 MC

65
 QV 4 -4

-4
-Q~ 00 4

aC' -! t": Cl
M~

-4

-4

4 V (D00 M00t
0 4 -. 4 -4 -4

00 0 0 Q00oom t-
-400 ' 4qt-
Lo
-4-4 -4 04 C4

m~ C1 4t

oo tcD

U3-4 U5QIC -4

Q C -4 (D CD0

0) C)C)o

" C~-4
W 00 t-

-4 -C*4 -4

z -

-14

66
2.4.5.1. Solving Multinle Models

There are numerous instances in which multiple models must be solved for each

DMU under consideration. The HDEA approach can be easily modified and extended to

efficiently determine all of the required values.

" To solve for technical, scale, and overall efficiency (inefficiency), both the BCC
and CCR models must be solved [85]. With HDEA, We can use the BCC or

ADD model through level 2 to find E* . On only the E* set we can apply the

CCR model formulation to find the E*cr (9 E;) set of DMUs. During level 3,

while determining the inefficient scores, we can solve the values for I*cr. Then,

as an advanced starting basis, we can allow all E* DMUs to enter the basis and

solve for V~. (Alternately, we could solve for the I c values first, then relax

the BCC 1A 1 constraint, tighten the basis entry rules, and solve for the I,,

scores). Once the BCC and CCR values are found, we can rapidly solve for the

three efficiency values desired. Note that this prevents a duplication of effort to

find the inefficient values.

"* Byrnes, Fire, and Grosskopf [28] argue that a "no increasing returns to scale"
(NIRS) model must be solved to determine if the DMU (or the virtual DMU if

inefficient) is in a region of increasing or decreasing returns to scale. Knowing

this may help indicate the direction the DMU should go to become efficient,

either expand operations through output augmentation, or downsize through

input contraction. We know that E*, C_E*irs C E-,. Consequently, the NIRS

strategy could be added to the strategy above to find these scores.

"* Besides solving for BCC, CCR, NIRS models, we can also apply assurance region
restrictions [84,85] once the E* for each model is found at level 2. The various

assurance region approaches can be solved at level 2 then extended to level 3. In

this way, a series of models can be solved rapidly without duplication of effort.

67
 We know E*r _ E* for any of the above models. Note: this same approach can

be used for the cone-ratio model of Charnes, et al [34,35].

"* For the CCR and BCC models, there is an associated orientation, either input

or output. The inefficiency scores will vary depending on the orientation, but

the E;, and E,*r sets will remain the same regardless of the orientation. Con-

sequently, if both orientations are needed, then only the level 3 values must be

resolved to obtain both.

"* Some versions of the model may provide faster solutions than others depending on

the number of output and input variables that are used. For example, if the num-

ber of outputs exceeds the number of inputs, the output-oriented model solves

faster. Also, the CCRO model has an initial feasible solution, thus bypassing

Phase I of the simplex method. Since the output model may solve faster, this can

be used through level 2, then the input model applied in level 3 calculations if the

input values are needed. HDEA provides the flexibility to incorporate the fastest

model to identify E* through level 2, then at level 3 can use the model of choice

to determine the inefficiency scores.

2.4.5.2. Implementation Issues

"* As noted above, any method can be used to find the set of DMUs belonging to E*

and I* such as sorting, preprocessing, domination theory, etc. These can possibly

enhance the HDEA procedure to expedite the level 2 and level 3 efficiency score

values.

"* Computational concerns have arisen over the two-phase [4] or 146] approaches,
which can be lessened via HDEA. We know that the efficiency scores do not vary

with the sum-of-slack solutions. So in HDEA, there is no need to solve for the

sum of slacks during level 1 or most of level 2. Once potential E* DMUs are

68
 identified, the sum of slacks must be solved only for these DMUs to determine
if they are weakly efficient. If there are any weakly efficient DMUs, they are
removed from E*. During level 3, there is no need to solve for the sum of slacks for
the inefficient DMUs since a proper projection to the efficient surface is obtained
since no weakly efficient DMUs can enter the basis. Solving for the sum of slacks

for the inefficient DMUs may simply identify possible alternate optimal solutions
that offer no new information when weakly efficient DMUs are not in the basis.

* As the number of input/output variables grows, the likelihood of cycling also

increases because of severe degeneracy involved with the solution of some DMUs

(especially efficient DMUs). Anti-cycling procedures invoke a high computational

cost. Also, the possibility of cycling can increase with a larger number of DMUs
(variables) in the model. HDEA can help reduce the cost of cycling by main-

taining smaller problems. Additionally, at level 1 and most of level 2 there is

no need to invoke the anti-cycling rules. Any DMUs that exhibit cycling can be
deferred. Only when all potential members of E* are identified at level 2, must
the anticycling rules be invoked. By passing over the DMUs that cycle, the basic
variables of other DMUs' solutions may later reveal the problematic DMUs as

efficient. In this manner, the DMUs will not need to be solved and the cost of

applying anti-cycling rules is reduced.

2.4.5.3. Exploiting the Levels

* As we progress through level 1 to level 2 to level 3, information at each level can

be used to enhance the solutions at the next level. DMUs that frequently show
up in efficient reference sets can be given priority during the pricing procedure
to enter the basis. By choosing the most attractive variables to enter the basis,
the number of pivots to solve the LPs may be reduced. This may also reduce

69
 the potential for cycling. This also indicates the possibility of using different

heuristics for the entering and leaving variables for the different levels of the

HDEA process.

* One advantage of the HDEA process over domination and other preprocessing

techniques is that it can be used to find DEA scores for categorical variables

and window analysis enroute to determining overall DEA scores. By blocking

appropriately, the efficient DMUs for each category or window can be recorded

along the way. In this way, solving overall scores as well as those within particular

categories or across various categories can be accomplished without duplication of

effort.

2.4.5.4. Parallel Implementation

"* The HDEA procedure can expand the parallel approach across many platforms.
For example, a SIMD distributed network can be used to solve each block of

DMUs. Since the blocks are smaller, memory requirements are smaller. The

advantage of updated restricted basis entry and early identification of DMU

status within blocks will be lost, but the basic process will still lead to a solution.

The basis entry and identification schemes will hold between levels.

"* For the parallel case, Amdahl's law assumes that the same information available

for the single processor as for multiple processors. But with RBE and early

identification, the interaction between processors ran enhance the solution process

resulting in high levels of efficiency when multiple processors are used. This is not

an HDEA-exclusive advantage, but it does hold for the HDEA process.

70
 2.5. Conclusions

We have described a new hierarchical decomposition procedure for solving DEA

problems that advances the state of the art for computational data envelopment analysis.

As demonstrated with medium- and large-scale test sets, this approach can have dramatic

benefits over traditional methods-in both single-processor and parallel settings-and

permits enormous problems to be optimized in a modest amount of time. The ability to

routinely solve problems with thousands of DMUs permits researchers and practitioners

to be more ambitious in their application of this important class of models and, we hope,

will encourage new and even more exciting applications of DEA.

71
 CHAPTER III

RANK ORDERING DMUS USING TIERED DEA

Data envelopment analysis was originally developed for efficiency measurement.

The primary focus to date has been on the development of a set of DEA models to

identify the efficient and inefficient DMUs under different sets of assumptions designed

to measure various types of inefficiencies. In spite of the tremendous growth and success

of the DEA methodology, there has been little research into methodologies for ordering

the DMUs according to the efficiency of their production and management practices.

The inability to rank DMUs by the comparative degree of efficiency or inefficiency limits

the potential for DEA to fully characterize successful and unsuccessful management

practices.

Although inefficient DMUs receive a DEA score which reflects the degree of

inefficiency, a direct comparison of DMUs is problematic unless both DMUs have the

same efficient reference set (ERS). To illustrate, Figure 3.1 depicts a unit isoquant plot

of a set of firms with one output and two inputs. Since each point represents a firm

and all firms produce the same level of output, the more efficient units are closer to the

origin. The efficient frontier is formed by DMUs A, B, C and D. DMUs E, F, and G are

all inefficient. The efficiency of each point is determined by a ratio whose denominator

is the length of the line segment from the point to the origin and the numerator is the

length of the line segment from the origin to the efficient boundary. The line connecting

an inefficient DMU's point and the origin will intersect one of the line segments forming

the efficient frontier. The end points of this line segment, composed of efficient DMU

points, form the efficient reference set for the inefficient DMU.

72
 According to Charnes and Cooper [32], to compare DMUs with different efficient

reference sets would require assumptions of the weighting (or pricing) scheme used by

DEA. But it is precisely this lack of restrictions on the weighting scheme that makes

the DEA methodology so attractive. Consequently, in general such assumptions are

undesirable. In our example, DMUF can be compared to DMUG because they share the

same ERS consisting of DMUc and DMUD. In this case DMUG with an efficiency score

of 0.800 is more efficient than DMUF with an efficiency score of 0.774. However, neither

of these DMUs should be compared with DMUE, with an efficiency score of 0.733, which

has a different ERS composed of DMUs B and C. Hence, for inefficient DMUs, a new

approach is necessary to further discriminate and allow comparisons across all inefficient

DMUs.

X2
4
A E.
3
B' O
2 CC

1 2 3 4 5 X,

Figure 3.1. Normalized data isoquant plot

For efficient DMUs, the relative importance of each DMU is difficult to discern.

Because each such unit has a DEA score of 1, there exists no variation in scores to

determine a relative value. Charnes and Cooper [32] suggested a tool which they called

the envelopment map to characterize the magnitude of the importance of each efficient

DMU. This method consists of counting the number of times each efficient DMU ap-

peared as a member of an ERS. Those DMUs occurring more often would be considered

73
more "consistently efficient." However, there are at least two problems with this measure.

First, to correctly count all occurrences of an efficient DMU in an ERS, all alternate

optimal solutions of the DEA i.odels would need to be identified. This can be compu-

tationally expensive and difficult to track. Secondly, this counting offers only a limited

amount of useful information. An efficient DMU that occurs often in an ERS merely

indicates that the DMU helps define part of the efficient surface which overshadows a

high concentration of inefficient DMUs. Firms utilizing new production techniques may

be extremely efficient, yet operate far from the "crowd" of other DMUs. As a result,

these efficient firms do not occur often in the efficient reference sets of the inefficient

units. Consequently, these maverick DMUs may not be deemed as important as they

should be.

To discriminate between and identify the successful and unsuccessful production

practices of the DMUs, a new procedure is necessary that provides a more detailed

classification of DMUs than the ordinary DEA efficiency scores offer. This procedure

should result in a rank ordering of DMUs which serves as a proxy measure for managerial

efficiency. In this way, managers and analysts will have a useful tool to identify manage-

ment practices that both accentuate and detract from productive efficiency by observing

the practices of the higher and lower ranked DMUs. Additionally, experts may want to

state, a priori to the DEA analysis, what they believe to be the most efficient firms in

the industry. The rank ordering procedure can then be used to determine how the DEA

results compare with the experts' opinions. As in standard DEA analysis, the ordering

should allow either constant or variable returns-to-scale envelopments. The remainder of

this chapter presents just such a rank ordering procedure which is easy to implement and

meets all of the above criteria.

74
 3.1. Need for a Ranking Procedure

Until recently, most modem writing on production theory assumed that all pro-

ducers were efficient. The premise was that in a competitive market place the inefficient
producers would realize the direct costs and indirect opportunity costs of continued pro-

duction and would leave the market to pursue more profitable adventures. However,
economic analysts have come to accept that inefficient production occurs in the market
place and its causes vary. Inefficient production can occur because information on the
most productive methods is neither perfect nor free. As a result, some firms may be

slower to respond to changing market conditions than others. Along with imperfect
information, market uncertainty influences the production process. The organization's

(or the manager's) position towards risk will dictate the rapidity with which the firm
will respond to change in the shadow of this uncertainty. Additionally, because perfect

competition is rarely (if ever) seen, regulations, and other exogenous constraints may
induce inefficiencies in the production process. Because of the social costs associated

with inefficient conversion of input resources to output goods, there has been a growing
interest in identifying and quantifying the inefficient processes. Data envelopment anal-

ysis has proven useful in measuring various types of the production inefficiencies which
may be attributed to inefficient managerial practices.

The original DEA models focused on identifying technical inefficiency or scale

inefficiency [111. FMre, Grosskopf, and Lovell [53], relaxed the usual DEA assumption

of strong disposability of inputs and outputs to further identify inefficiencies due to

congestion of resources. 1 In addition, Fare, Grosskopf, and Lovell, following the lead
of Farrell 151], introduced prices of inputs and outputs to identify allocative inefficiencies.

I Congestion of inputs implies that as at least one input increases, at least one
output is reduced. That is, there is not a positive correlation between all outputs and
inputs.

75
A firm demonstrates allocative inefficiency when it departs from its predefined goal such

as maximizing profits or minimizing costs.

Even though technical, scale, and allocative inefficiencies can be measured, little

has been written to formulate a means of ranking the DMUs based on the types of inef-

ficiencies they demonstrate. As a result, even though the inefficiencies can be identified,

they have not fully been related to various aspects of producer behavior. With a ranking

system, the DMUs exhibiting the best production processes, in terms of efficiency, could

be compared to those characterized by the worst production techniques. The manage-

ment practices of the best producing DMUs could then be compared to the worst DMUs

in order to identify the underlying managerial inefficiencies. Once identified, the less

efficient firms could adopt the practices of the best firms to improve the productivity of

their operations.

The purpose of this study is to present a new approach to rank order or stratify

the DMUs to more clearly relate the efficiency (or inefficiency) of a given DMU to all

others in the set. The intent is not to suggest this approach as the only valid rank

ordering scheme. Indeed, any set of items can be ranked by any subjective means.

However, the purpose is to present a methodology, with theoretical underpinnings, which

can result in a meaningful ordering system. It is hoped that this methodology may

stimulate research into other possible ranking procedures so that the richness of the DEA

methodology can be more fully utilized.

3.1.1. Basis of the Ranking Procedure

Data envelopment analysis defines efficiency based on empirical observations of

each evaluation unit's behavior. Each DMU consumes multiple inputs in order to pro-

duce one or more outputs. The implicit assumption in DEA is that the DMUs transform

the inputs into outputs by means of a well-behaved production technology. A feasible

input-output vector for a given DMU is technically efficient (Pareto-Koopmans efficient)

76
if it is technologically impossible to increase any output and/or to reduce any input

without simultaneously reducing at least one other output and/or increasing at least one

other input. The empirical models of DEA use observed data to identify the DMUs which

form the Pareto-Koopmans efficient surface, and hence, are best at using the current

technology to convert the input resources to output goods. The level of technological

achievement revealed by the efficient surface will be highly dependent upon the choice

of DMUs used hi the study since the methodology only measures relative (not absolute)

efficiency.

The ranking method presented in this study separates DMUs into groups based

on the level of technological achievement which they demonstrate. It will be shown that

this aggregation reveals additional aspects of inefficiencies not available with traditional

DEA measures. For example, a DMU that appears to be very inefficient by the standard

DEA measures, may rank well, compared to other firms, when viewed in terms of the

DMU's ability to employ the most recent technological advances. Once the DMUs are

separated into these achievemem levels, a procedure will be presented to rank the DMUs

within each level. The ranking within a level will be determined by the contribution the

DMU makes to defining the shape of the efficient surface of that level.

An attractive feature of the proposed ranking procedure is that it can be used

across the many formulations of the DEA models. In this paper, the ranking procedure

is introduced and applied to both the input- and output-oriented BCC (variable-returns-

to-scale) models. The input- and output-oriented models achieve identical stratification

of DMUs into tiers of common technological achievement levels. However, the ranking

within each tier will differ according to the orientation used. Next, the ranking procedure

is applied to the CCR (constant returns to scale) model. In this case, the input- and

output-oriented schemes result in identical rankings. The CCR model adds interesting

interpretations of the most productive scale size to achieve each technological level.

77
 3.2. Tiered DEA

At the heart of data envelopment analysis is the separation of evaluation units

into (relatively) efficient and inefficient sets. The models' objective function values, 0 or
z, have been used as metrics for the degree of inefficiency for comparative and predictive
purposes [21,22]. Since these values may be incompatible, from an economic point-of-

view, we present a different approach to comparing DMUs that has significant appeal,

from both intuitive and economic-theoretic standpoints.

3.2.1. Tiering Algorithm

The following tiered DEA (TDEA) procedure stratifies decision units into tiers,

or layers, of comparable productive efficiency, as measured by any standard DEA model.

Procedure TDEA

1. Initialize: t +- 1, D[1] +-- D.

2. While Dft] 6 0 do:

a. Apply a DEA model to the DMUs in set D[t] to identify Ett.
blit] = DIt] - E[t].
b.I*-D E
c. t+-t+l.
d. DitI =I*.
where t is a tier index and Et*and I*tI are the sets of efficient and inefficient DMUs on

tier t, respectively, relative to set D[tl.

The TDEA procedure begins with a traditional data envelopment analysis, then
progressively strips away productioa surfaces, revealing a series of frontiers of decreasing
productivity. Specifically, at tier 1 all of the DMUs in the data set are analyzed using

a standard DEA model, thus separating them into efficient and inefficient sets. The
efficient units are then assigned to the current tier and the inefficient ones become the
data set of interest for the next higher tier; this process is applied recursively until all

78
DMUs are assigned to a tier. In this way, the tier levels represent successive layers

of relatively efficient production surfaces where the DMUs at any given tier are less

productively efficient than those of "outer" (lower-numbered) tiers and more efficient

than DMUs at "inner" (higher-numbered) tiers.

An example application of the TDEA procedure can be seen in Figure 3.2 where

10 DMUs, each with 1 output and 2 inputs, are plotted. All DMUs have the same output

level so only the inputs are shown. DMUs A, B, and C are DEA efficient since the line

segments joining these DMUs envelop the other DMUs from below. If firms A, B, and C

were removed from the data set, a new efficient frontier would be formed by DMUs D, E,

F, and G. The TDEA procedure reveals that the data contains three production surface

layers.

X24~ D H

K
JB

I I II I I,I I I I
1 2 3 4 5 X1

Figure 3.2. Example tiered results - two inputs, one output

Figure 3.3 shows 10 DMUs each with 1 input and 2 outputs. In this case, all

DMUs use the same level of input so only the outputs are shown. Again, DMUs A,

B, and C are efficient since they form a boundary that envelops the other DMUs from

above. TDEA reveals three production surface layers.

79
 D B

1 K
O-r C

1 2 3 4 5 Y1

Figure 3.3. Example tiered results - one input, two outputs

3.2.2. Example TDEA Applications

The TDEA procedure was applied to three 8,000-DMU data sets, one from in-

dustry and the others randomly generated. The "Banking" data represents a selection

of banks from the Federal Reserve Bank's Southwest district, with 6 input and 3 output

values, as described in 121J. A set of "Cobb-Douglas" data with 5 input and 4 output val-

ues per observation was created with DEA-GEN using constant returns to scale and the

parameter set a = (.20, .20, .20, .20, .20). (See Appendix B for the generation procedure.)

Also, a "Multi-Normal" data set, with 4 inputs and 3 outputs and the variance-covariance

matrix given in Appendix A, was generated using the DRNMVN routine from the IMSL

library.

For this test bed, the TDEA procedure used the CCR' model to assign DMUs

to layers. Table 3.1 reports, for the first 20 tiers, the number of DMUs assigned to

each layer ("IE[t I") and their maximum, minimum, and mean tier 1 efficiency scores

(0). Note that, in each case, that the mean and maximum efficiencies drop with each

successive interior layer, as might be expected. However, the TDEA provides insight

that DEA cannot provide. As noted by the minimum values, some DMUs possess a

low tier 1 efficiency score, yet fall on an outer tier, indicating an efficient use of current

80
 El"a koC 0 -- CD CD M~ C14 M~LM~ 0 M 0m "W M 0 0 O to C4

o i 00 00 000 0- 0- 00 0- 0 0 0I 0 0OLOL
0Z 03

CD Ll.) 000400ccI 4(=00e0 O c mt-v Y3-00

0) C M oo 00 00 00 00 t- t-t e 00
-C c t A o LO~

X M vo cz 00i m~ 00 M- Lo M~ 0m 00U-3 vO CO
GW 0- m~ 00 CD CV) = C4 =4 w~ 4m
0 4 m c 000=

5-4 -- 4 1- 4 C -1m m m e

t-C ) Oe t-4-I OCo 000wC'

_p C? CK pq: - lC!C C ?

C) ( (D CD D (D(z) ) Q D Q C) C) ( C C

00 C1 qrU'
- ~
Dt
O k (m 0k0 czUl 0CO mm 00r 0e 0CI C
to q-4 40c D(m( )QQC)0C 000l
)C
c; i000c;00 c0C6c; s 0c 606 66 660

-Dw=e 4mv -4ulC -L q 0t o L I

00C OV MC)C4C 4- - - DC
-0ci c; ;0c0c;C6 C 0 06 00;C6 ; 0c

t--vc m oC4e Mt qCl-

<-Vr oM4qLOt 4Mr4t t O t >C

Mv 4 0O=WMWWt-

I - - ;_ _

81
technology. Conversely, the maximum values indicate that while other DMUs may seem

fairly efficient, from a 0-standpoint, they fall on an inner tier because of less productive

uses of current technology. It is precisely this behavior that should prove useful in

identifying a DMU's true level of efficiency.

Of interest also are the differences between these three problems. The frequency

counts reveal that the banking data has only 18 tiers, while the Cobb-Douglas data still

has 3,286 of its 8,000 points still un-stratified after 20 tiers. The multi-normal problem

has more populous tiers than the Cobb-Douglas, but less than most of the bank's. The

banking and multi-normal data sets have tier-i-efficient DMUs on interior layers, but

Cobb-Douglas does not. The banking and Cobb-Douglas data have much larger mean-

efficiency drops between tiers 1 and 2, relative to the multi-normal. Each problem seems

to have a very different structure from the others, as uncovered by the TDEA process.

3.2.3. Tiered DEA and the BCC Model

The examples of the figures above illustrate the result of applying the tiering

procedure to the BCCj and BCC0 models. These models, used to identify the Pareto-

Koopmans efficient production surface, with no restrictions on returns to scale, were first

proposed by Banker, Charnes, and Cooper [11]. The BCC,, model can be written as:

(BCCj) m+ 0 - +ls180) (3.1)

s.t. YA - 80 = Yj (3.2)

0Xj - XA - =8
=0 (3.3)

1A = 1 (3.4)

A, a' > 0 (3.5)

0 free (3.6)

82
 In Figure 3.2 and Figure 3.3, the outer layers designated by DMUA, DMUB,

and DMUc represent the efficient set which demonstrate best practice in the production

process. However, this empirical production surface is dependent on the data observed.

Suppose the original set had not included B. The models would have identified a

different, yet legitimate, production surface revealing the current technology consisting of

DMUs A, E, F, and C. If data on B later became available and was added to the set of

observations, it would markedly alter the shape of the efficient production surface. B now

reveals a new production process with a corresponding level (, technological achievement

that was previously unseen. It is precisely this realization that motivates the tiered

ranking procedure.

In DEA, all of the efficient DMUs share a common characteristic: they all demon-

strate an ability to make the best use of the current technology to conduct their produc-

tion process (i.e., they demonstrate best practice behavior). Once these DMUs have

been identified, they can be temporarily removed, and the remaining DMUs form a new,

valid DEA data set. When the BCC" model is applied to this new data set a new

efficient production surface is revealed. Had the data for DMUs of the outer tier not

been available originally, this new production surface would legitimately characterize the

best technological achievement level observed. Consequently, the DMUs comprising this

new efficient surface share a common level of success of utilizing the currently revealed

technology. Repeating this process groups the DMUs according to common achievement

levels. DMUs on outer tiers reveal a technological advance not realized by DMUs on

inner tiers.

Of primary significance is that the tiering procedure provides greater discrim-

inatory power in determining managerial efficiency than previous DEA measures. To

appreciate the importance of the new measure, an understanding of what causes a DMU

to be inefficient is helpful. By stratifying across different tiers, the new measure provides

83
a more complete description of the DMU's managerial efficiency relative to its contem-

poraries. One reason a DMU may be characterized as relatively inefficient in DEA is

that it is dominated by a few highly efficient DMUs. A high concentration of other

DMUs, with similar, but superior, management practices can cause the inferior DMUs

to attain a low tier assignment. Figure 3.4 depicts a set of DMUs with one output and

one input. The tiering procedure produces three layers. In this case, J is enveloped

by a concentration of DMUs using nearly the same level of inputs to produce a similar

level of outputs. However, J consistently under-achieves compared to the other DMUs,

consequently, it would receive a relatively low ranking. Even though the original DEA

scores for DMUs F, G, and J are similar, TDEA further discriminates J as less productive

given the available technology than DMUs F and G.

Y
4

3 D Tierl
N--- ----- * Tier2

2A0 _ Tier3
B1F

A E
I i I I I II I i i
1 2 3 4 5 X

Figure 3.4. Example 1 tiered results - one input, one output

A second reason an inefficient DMU may result in a low efficiency score is that it

operates in a production process "away from the crowd." These maverick DMUs may be

leaders in the introduction of new production technologies or management methods into

the market place. The transition to these new procedures is penalized under traditional

DEA analysis because it appears that the DMUs introducing the new inputs consume

84
more than the peers. The dominant DMU of this shift may be efficient, but the other

DMUs can appear to be very inefficient. This situation can be seen with H. Notice,
in this case H has a technical efficiency score seen as the ratio NH-. J has a higher

technical efficiency score seen as the ratio = but is ranked lower than H in terms of
tier assignment. H may represent a risk-taker that shows a short-term reduction in its

DEA efficiency score, to take advantage of new technology, by introducing new inputs.
However, even with a short term loss of efficiency, the DMU could see a rise in its rank

ordering by moving to a higher tier level. Once the new technology is fully integrated into
the production process, the DMU may witness substantial increases in its DEA efficiency
score. Traditional DEA analysis would penalize the management decision, in the short
run, to introduce the new technology into the production process. Consequently, the

decision to incur possible short term losses to achieve long term gains would appear
unfavorable in a traditional DEA analysis. However, the stratified ranking of DMUs
reveals the success of this management decision.
The above scenario can occur often in a free competitive market. Under dynamic

conditions, the firms in the competitive market must adapt to maintain market share. A
DMU that adopts a management style and production process similar to other DMUs,

but consistently under-performs, may result in a seemingly high relative efficiency score,
but with a low rank ordering. The inability of this competitive firm to find its niche and
distinguish itself from the "competition" may result in its failure in the market place.

Current DEA methodologies are inadequate to reveal such conditions; TDEA offers an
attractive means to discriminate between these managerial behaviors.
DMUs with low efficiency scores but with relatively high rank order are not
restricted to DMUs at "fringe" production levels. Data outliers can strongly affect the

shape of the efficient surface. Figure 3.5 shows the effect that outlier B has on the
production surface. The outlier significantly distorts the surface making F seem relatively

85
inefficient. Had B not been in the data set, F would be efficient. In spite of F's low
efficiency score, TDEA ranks F relatively high. Consequently, the performance of F may

not be as poor as indicated by the DEA score.

Y
4

3 B D Tier I
2 M
- - F G H Tier2
3
2M " - "Tier
I I I

416

, ,: , ,I I I I
5
1 2 3 4 X

Figure 3.5. Example 2 tiered results - one input, one output

A key advantage of stratifying DMUs into tiers is that it allows the DEA method-
ology to more closely describe true managerial efficiency that may be masked by tradi-

tional DEA analysis. As a result, managers of inefficient DMUs have increased flexibility
in improving production operations. The mant,&ger's long-term goal may be to achieve
efficiency. The DMU can strive to accomplish this by improving the short run rank
ordering without a myopic focus on its DEA efficiency score.

3.2.4. Tiered DEA and the CCR Model

The ranking procedure can also be applied to the CCR models. For these models,
which assume constant returns to scale, the efficient DMUs not only are operating

most efficiently (with the greatest level of technological achievement) but they are also
operating at the most productive scale size (MPSS). For the single input and single
output model, the MPSS is determined by the DMUs yielding the highest ratio of
quantity output to quantity input. In economic terms, this equates to the DMUs yielding

86
the highest average product. Banker 1101 demonstrated that the CCR model revealed the

efficient DMUs operating at the MPSS in multi-dimensional processes.

The MPSS units which form the frontier for the CCR models are dependent on

the observed set of data. In a traditional DEA analysis, the CCR scores for the inefficient
DMUs reflect both technical and scale inefficiencies. To separate the technical from the

scale inefficiency, the BCC model must also be run. The tiered procedure, when applied
to the CCR model, presents a different picture of scale inefficiency. As each outer tier

is removed from the data set, the set of DMUs which composes the most productive

scale size changes. As a result, the values of the scale inefficiencies for the other DMUs
also change. Therefore traditional DEA measures may overstate the amount of scale

inefficiency that a DMU demonstrates.

Y
4 Tier 1

3 D u- o'Tier2
C ,'HO

2 ,C K

0
E
'A

1 2 3 4 5 X

Figure 3.6. Example OCR tiered results - one input, one output

Figure 3.6 shows two tiers of scale-efficient DMUs for the same set of data used in

Figure 3.4, where DMUs A, B, C, and D were shown to be technically efficient. However,
Fig 3.6 indicates B and C are scale efficient while A and D are both scale inefficient.

Because D falls far from the most productive scale boundary, the CCOV DEA measure

87
would indicate it is more scale inefficient than A. Yet, with the tiering procedure, both

become scale efficient at a common level. Consequently, the two DMUs may not be as

different, in terms of scale, as first suggested by the DEA scores.

3.3. Ranking Within Tiers

The TDEA approach extends traditional DEA to allow for a stratification of

DMUs reflecting different productive efficiency layers. However, a ranking system within

each layer is still needed. This section describes a new procedure that can provide a
useful ranking.
Each tier defines a set of DMUs that forms a production surface (layer). To

determine the relative rank ordering among the DMUs at each tier, one could measure
the contribution of the DMU to the shape of its layer. DMUs that significantly distort

the production surface layer on which they are assigned play a more prominent role in
determining the shape of the production surface. Consequently, if these distortions can
be measured, the DMUs could be ranked by the degree of distortion they contribute to
the production layer.

The following model describes an extremal DEA (EDEA) approach for the MCr
model. However, the formulation also applies to the output-oriented models and to
either constant- or variable-returns-to-scale formulations. With this methodology, the
DEA formulation is modified slightly to measure how far a DMU is from the resulting
production surface when the DMU itself is removed from the data set of interest. The

EDEA procedure can be described as follows. Let X be an (m x n) matrix and Y be

an (s x n) matrix of the observed input and output values, respectively, of all the DMUs
of interest. Select a DMU to be observed, in this case DMUo. Let X{IP and Y101 be X

and Y, respectively, with the observations of DMUo removed. Then the extremal DEA
measure 0 is computed by the following model.

88
(EDEA Model) minO (3.7)

s.t. y[oJ, -. So = Yo (3.8)

OXo - XIO]x - 8i = 0 (3.9)

A, Si, 8 0 > o (3.10)

0 free (3.11)

In contrast to a traditional DEA analysis, 0 can now be greater than one. For
0> 1.0 the DMU will be efficient and 0 measures the allowable proportional increase
in inputs for the DMU to remain efficient. If 0 = 1, the corresponding DMU may be
weakly efficient from a DEA standpoint. However, for the inefficient DMUs, the EDEA
scores will be identical to traditional DEA scores, (e.g., 0 = 0), since the inefficient DMU
itself will never be a member of an optimal basis (this can be shown in the same manner
as Theorem 1 of Chapter 11). Consequently, EDEA has an advantage over traditional
DEA models in that it provides greater meaning to the scores for the efficient DMUs by
allowing additional variability in the efficient score values.

X2
4
D
D H
3
B* *J
2 . , K
1 K

1 2 3 4 5 X,

Figure 3.7. Extremal DEA example

89
 Figure 3.7 illustrates the EDEA procedure for B when applied to the data in
Table 3.2. Here, B is projected onto the "new" efficient surface using EDEA. The

resultant objective value, 0" = measures how "far" B is from the efficient surface.

Viewing the problem from a different perspective, 0" reflects the degree to which B would

contribute to the shape of the new efficiency surface if it was added to the data set.
DMUs that cause significant and important distortions of the efficiency surface
will result in a high EDEA objective value. Those DMUs that have little influence on

the shape of the production surface will have objective values close to 1. Consequently,
the DMUs can be ranked by order of influence at each tier level based on the EDEA

scores.

An overall ranking of all DMUs can be achieved by: (1) using TDEA to to stratify
all DMUs into tiers, (2) applying EDEA to each tier level, (3) ranking each tier's DMUs

by 0, (4) then finding each unit's overall rank by ranking those DMUs on outer tiers as
more important than those on inner tiers. The following section presents an illustrative
example of this procedure and provides some numerical interpretations.

3.4. Example of Tiered DEA

The illustration shown in Figure 3.2 is based on the data in Table 3.2. As illus-

trated in Figure 3.2, the TDEA approach yields three production surfaces. Those DMUs
on tier 1 are the same efficient units found in a standard envelopment analysis. The
DMUs of succeeding inner tiers compose less efficient production surface layers. Notice

though, the inefficient DMUs can now be further distinguished by the production surface
on which they fall. DMUs on outer tier levels represent more successful production
processes. Consequently, the DMUs on outer tiers can be ranked more efficient than

DMUs on inner tiers. As will be demonstrated, this does not necessarily coincide with

90
 Table 3.2. - Tiered DEA example data

DMU
A B C D E F G H J K
y 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
x, 0.50 1.00 5.00 1.00 1.50 2.00 4.50 2.00 3.00 4.50
x2 4.00 1.50 1.00 3.50 2.50 2.00 1.50 3.50 2.50 1.80

Table 3.3. - Results of the tiered rank ordering procedure

DMU
A B C D E F G H J K
DEA 1.00 1.00 1.00 .765 .650 .722 .788 .482 .565 .688
TDEA 1 1 1 2 2 2 2 3 3 3
TEDEA1 2.00 1.60 1.50 .765 .650 .722 .788 .482 .565 .688
TEDEA2 - - - 1.50 1.05 1.13 1.33 .733 .774 .889
TEDEA3- -. -- -- -- - - - 1.50 1.13 1.39
RANK1 1 2 3 . . . . . . .
RANK2 - - - 1 4 3 2 - - -
RANK3 -- ---- -- -- -- -1 3 2
RANK 1 2 3 4 7 6 5 8 10 9
DEARANK 1 I1 1 5 8 6 4 10 9 7

the traditional DEA score. The results of the TDEA coupled with the EDEA are shown
in Table 3.3 and labeled as TEDEA1 through TEDEA3.

The DEA row lists the CCRI 0 scores for each unit in the data set. The TDEA
row indicates the tier level to which each DMU is assigned as a result of the TDEA
procedure. TEDEAt rows give the EDEA i values for all DMUs on tier t or higher.

DMUs in tier t are ranked within tier in the RANKt rows. The overall rank for the entire
set of DMUs is given in the RANK row. This can be compared to the ranking the DMUs
would have been given, listed in DEARANK, had they been ordered by the CCkI 0
value.
Some important observations can be made concerning these results. Notice, the
DEA and the TEDEA1 results are identical for the inefficient DMUs. Likewise, the

91
efficient DMUs are appropriately identified with values greater or equal to 1. Notice that

the higher inefficiency scores do not necessarily indicate on what tier level a DMU may

fall. For example, K has a higher 0 than E but falls on a lower tier level. By observing

the results at TEDEA3 one notices that the ranking of H, J, and K does not correspond

with the ranking the DMUs would receive if the DEA efficiency scores were used. In

fact, H has the lowest DEA efficiency score, but the highest rank of tier 3. Figure 3.8

illustrates why this is so. DMU J does not significantly distort the efficiency surface that

exists when it is not present. This is not true for H which significantly distorts the shape

of the production surface; DMU H is more influential than DMU J and thus is ranked

higher.

x2 x2 I
4 4

3H J3-

2 2*
K K

T V
5
1 2 3 4 Xl 1 2 3 4 5 X,
Example A Example B

Figure 3.8. TEDEA examples

A major advantage of TEDEA, besides ranking each DMU by its influence, is

that it can help paint a numerical picture of the environment in which a DMU operates.

As each tier level and associated production surface is removed, the new DEA scores for

the remaining DMUs can be calculated. In this way, the migration of the DEA scores

for a particular DMU can be traced through a series of tiers. A rapid rise in the scores

may indicate that the DMU is in a region with a relative low density of other DMUs

but is dominated by a few highly efficient ones. A slow rise in the score may indicate

92
that the DMU is surrounded by a larger density of other DMUs which may have similar
management styles or environments but are operating more efficiently. This information

can assist the analyst and managers in determining appropriate courses of actions to
improve either the rank ordering or the efficiency score.

The rank ordering procedure may also prove useful in window analysis. If DMU

behavior is tracked over time, the changes in rank ordering should reflect the relative
effectiveness of on-going managerial decision-making. These managerial changes may
remain hidden from traditional DEA analysis unless the changing practices result in a

change in efficient or inefficient status of the DMU. Consequently, the rank ordering
methodology may provide prompt managerial feedback as to how a DMU compares with

the competition as a result of implemented changes.

3.5. Computational Considerations

It is important to note, unlike CCR or BCC models, with the EDEA method the

basic elements in the optimal solution need not represent efficient DMUs. Figure 3.7

depicts such a situation. When B is removed from the data set, the resulting effi-
cient surface includes E and F, both of which are inefficient according to a traditional
DEA. Therefore, the advantage of computational efficient techniques for DEA, such as
restricted basis entry and early identification of efficient DMUs, cannot be maintained

for the EDEA models. However, other computationally efficient procedures are possible
for the EDEA model. Computational efficiency can be maintained by combining TDEA
and EDEA into a single formulation, TEDEA*, based on the following observation.
When a DMU is removed from tier t's data set, the resulting production surface
will consist of DMUs belonging to either tier t or t + 1.

In this way, the stratification of DMUs proves to be a valuable computational

tool for the EDEA method. Since the number of DMUs at tier t and t + I is typically
small compared to the entire set, the linear programming problems remain small when

93
determining the EDEA scores. As a result, TDEA and EDEA can be combined to form
a computationally efficient rank ordering method. Let Xt. and Yt. be the matrices of
input and output vectors of DMUs belonging to tier t or t + 1. Choose a DMU from tier
t for analysis, let this be DMUo. The TEDEA* model can be written as:

(TEDEA*) min 0 (3.12)

s.t. Yt0,_- 80 = YO (3.13)

x0o- X.]A- si = 0 (3.14)

A, 82, 80 > 0 (3.15)

0free (3.16)

Because the TEDEA* problem consists of LP's much smaller than EDEA, computational

efficiency is maintained.

3.6. Additional Benefits of Ranking

As mentioned previously, the empirical production surface of the DEA model is

greatly influenced by possible outliers in the data. In 1971, Timmer [861 assumed a
Cobb-Douglas form to develop a probabilistic frontier production function for observed
data. Using the Cobb-Douglas form, Timmer translated the problem into a linear
programming model with a striking similarity to the DEA model. Since he assumed

a single output, multi-input case, he was able to compare the frontier analysis with the
traditional econometric model. He showed that by eliminating the top 2% of observations
which appeared to be outliers, the linear programming frontier model yielded results that
could be supported by the econometric models.
In DEA, no functional form is assumed. Consequently, potential outliers have

bewn difficult to identify. In this case, the efficient DMUs may not represent normal
production activities and therefore may not serve as a legitimate basis of comparison.

94
Consequently, DEA efficiency scores of outliers could distort the true picture of manage-

rial efficiency of the inefficient DMUs. As a result, identification of possible outlier or

extremely dominant DMUs is essential. With TEDEA*, the outliers will be located on

outer tiers. If possible outliers axe revealed, the problematic DMUs can be removed from

the analysis. By following this procedure, a better estimation of managerial efficiency

can be achieved.

3.7. Conclusions

This chapter has outlined a rank ordering of DMUs based on the influence they

serve in forming empirical production layers for the data set. The intent is not to present

the only valid system or rank ordering, but to stimulate thought as to how DEA methods

may be modified to make the results more meaningful to practitioners. One concern with

the rank ordering approach is the large computational requirement necessary to achieve

the results. On some layers the density of efficient DMUs will be high and on other

layers the density may be low. The code must be computationally efficient across this

wide variety of conditions. In addition, the code must be flexible to allow for constant

or variable returns to scale in either the input or output models. Flexibility to switch

between models at different layers may also be desirable. The TEDEA* approach meets

the requirements to effectively and efficiently perform such tasks.

The rank order procedure presented in this discussion classifies DMUs by their

ability to use current technology to efficiently conduct the production process. This

classification can serve as a proxy measure of management efficiency that provides more

detailed information than offered by standard DEA models. By identifying the best and

worst performing DMUs, analysts can pinpoint which DMUs should serve as a basis of

comparison from which to determine which management practices enhance and which

deter from productive efficiency. In addition, the rank ordering methodology presented

95
here reveals the need to further charac~terize the computational issues of DEA and

provide efficient codes to accomplish these new tasks.

96
 CHAPTER lV
DEA SENSITIVITY ANALYSIS

4.1. Motivation
In 1978, Charnes, Cooper, and Rhodes (CCR) [37] developed a technique, which

they named data envelopment analysis (DEA), to determine the relative efficiency of

a production unit (commonly known as a decision making unit, or DMU) compared

to other units using the same kinds of inputs to produce the same kinds of outputs.
There are three popular linear programming formulations which indicate whether a DMU
is relatively efficient or inefficient. The CCR model [37] assesses whether the DMU

is efficient at the most-productive scale-size (MPSS). This model is perhaps the most
widely used in DEA applications. The Banker, Chs nes, and Cooper (BCC) model t111

and the additive model [33] determine whether the DMU is Pareto-Koopmans efficient.
In economic contexts, the BCC and additive models find the entire efficient empirical
production surface while the CCR model finds a proper subset of that surface which is
composed of efficient units at the MPSS. For a single-output, single-input DMU set, the
MPSS is the size that yields the greatest average product (i.e, the greatest ratio of output

to input).
Within the context of one of these models, a DMU's status is either efficient or

not efficient. To determine this status, a separate linear program must be solved, either
directly of implicitly. Since the DEA methodology is empirically based, values of the
inputs and outputs must be measured or estimated at a given point in time for each

DMU. A concern then is the accuracy of the measures or estimates, the comparability of

the decision units (particularly on a temporal basis), and any uncertainty associated with

97
the observed values. Any of these factors can affect the validity of an efficiency analysis

and influence the status of all of the decision units under consideration.
Sensitivity analysis is an important element of any DEA study because it assesses

the robustness of each DMU's status when variations in the inputs and outputs are

possible. This is more than a simple extension of traditional LP sensitivity theory

because of the nature of the DEA models.
Traditional sensitivity analysis in linear programming concentrates on finding a

new optimal solution value to a problem, as updated estimates of some of the sample

data become available, without the expense of resolving the problem from scratch. This
may take the form of determining ranges within which data may be varied without
requiring a change in the set of vectors composing the optimal basis. For the DEA model

formulation, this traditional analysis is inappropriate because a change in the values of

the inputs and outputs of a particular DMU simultaneously affects structural and right-
hand-side coefficients for all LPs in the analysis. Additionally, the assumptions of non-
degeneracy of the optimal basis and uniqueness of the optimal solutions are violated for
the DEA problems. Also, in DEA, the primary concern is the effect that perturbations

may have on the status of a DMU, not simply on the solution value of the underlying
linear programming problem. Consequently, new approaches for sensitivity analysis must
be developed when considering DEA models.
To date, a number of studies address sensitivity analysis in DEA [33,36,43,44],

most with respect to the additive model. Charnes, Cooper, Lewin, Morey, and Rousseau
[36] present a limited study of perturbations to a single output or single input using the
CCR model. This study was extended by Charnes, Haag, Jaska, and Semple (CHJS) [43]
to identify regions of stability using the additive model, in which allowable perturbations
cause the DMU to retain its status. Because the additive model finds the Pareto-
Koopmans efficient frontier without respect to an input or output orientation, the CHJS

98
example is restricted to simultaneous perturbations of both the inputs and outputs of

each DMU in the model.

This chapter presents a new methodology that allows for sensitivity analysis

in the case of proportional changes to either the outputs alone or the inputs alone

for the BCC and CCR models. The approach will identify a range of values such
that all allowable perturbations for the DMU within the range, ceteris paribus, will

preserve the DMU's current status. Since the analysis can be conducted in conjunction
with determining the DEA efficiency score, minimal additional computational costs are

required. Also presented is an update to the CHJS approach that overcomes inherent
problems with their additive model approaches.
The remainder of this chapter is divided into four sections. Section 4.2 devel-
ops the framework for conducting sensitivity analysis on DEA problems. Section 4.3
introduces the methodology to conduct sensitivity analysis on each DMU for the BCC

and CCR models. Although the sensitivity analysis is based on proportional changes
in either outputs or inputs, nonproportional changes will also be investigated. Section

4.4 compares the methodology presented in Section 4.3 to the CHJS sensitivity analysis
approach. A problematic concern of the CHJS method is presented and a demonstration
of how the new methodology compliments the CHJS to meet that concern. Section 4.5
introduces an expanded sensitivity analysis approach based on the models introduced

in Section 4.3. Although the expanded sensitivity analysis may be computationally in-

tensive, it can reveal information to guide the management in determining what future
actions should be taken to maintain or establish efficiency. Conclusions are given in

Section 4.6.

99
 4.2. BCC and CCR DEA Sensitivity Analysis

A new method for sensitivity analysis in DEA can be constructed from simple

modifications to the CCR and BCC DEA formulations. In this approach, the sensitivity
of a DMU's status is measured in terms of either proportional changes in outputs or
inputs. The concerns in DEA sensitivity analysis differ dramatically between DMUs

which are efficient and those that are inefficient. For efficient DMUs, an increase of any

output cannot result in a lower efficiency rating; likewise, a decrease in any input cannot
worsen the DEA efficiency score. Therefore, with efficient DMUs, primary attention
is given to determining what reductions in outputs are allowable, or what increases in
inputs are permissible in order to insure the DMU retains its efficient status. For efficient
DMUs one is concerned with identifying how bad the production process can get before
the DMU loses its efficient status.
This motivation fundamentally differs for the inefficient DMUs, for which the

primary concern is the identification of changes that will improve the efficiency rating.
Generally, one would like to identify those areas that require the smallest improvements
which would cause a given DMU to become relatively efficient. For inefficient DMUs
this would consist of either increasing outputs or decreasing inputs. We now present an
elegant formulation for analyzing both efficient and inefficient DMUs in the CCR and
BCC models.

4.2.1. CCR and BCC Sensitivity: A Background

For the sensitivity analysis, let DMUk be the DMU of interest. In this model,

we define D = {1,...,n}, the index set of DMUs, with X = 1X1 ,...,Xn] and Y =
[Y,..., Yn]. Vectors (Xj, 1') denote the observed data where Xj = (Xlj, ... , Xij, ... , Xj)

is a column vector of observed input values with xij _0 (at least one positive) and
S--- (ylj,..., Yrj,...,. yj) is a column vector of observed outputs with Yrj 0 (at
least one positive). Let X[kJ be the X matrix with the Xk column removed. Similarly

100
define Y[k] to be Y with the Yk column removed. A new approach called extremal DEA

(EDEA) will form the basis for the DEA sensitivity analysis. The approach allows for
both an input and output oriented analysis and can be used within the CCR and BCC
modeling contexts. The CCR-based EDEA models can be formulated as follows, with
ECCRi and ECCR0 having input and output orientations, respectively.

(ECCR4) min w= (4.1)

s.t. Y[kJA - B = Yk (4.2)

6Xk -x[klA,- 8 = 0 (4.3)

A, 8 ,0SO 0 (4.4)

6 free 4.5)

(ECCRO) min w = 4 (4.6)

s.t. X'k]A + 8' = Xk (4.7)

4Yk _ y[kIA +so = 0 (4.8,

A,s, 80 > 0 (4.9)

4 free (4.10)

The BCC versions of the above EDEA models - EBCC&and EBCC', respectively -
can be obtained by appending the constraint: 1A = 1.
In the discussion that follows, the term DEA models refers to the original CCR,

BCC, and additive models. In standard DEA form, the index set of frontier-efficient
DMUs of the BCC and CCR models which are Pareto-Koopmans-efficient is designated
as E* = {j E D16* = 1 (or 0* = 1), a** = 0, and a'* = 0}. The set F is the index set

of frontier-efficient but not Pareto-Koopmans efficient DMUs, where F = {j E DJO* = 1

or (* = 1) and 80* # 0 or a8* # 0. For a DMU to be Pareto-Koopmans efficient, it

101
must not be possible to decrease any input value without also decreasing at least one

output or increasing at least one other input value. In addition, to be Pareto-Koopmans-

efficient, it must not be possible to increase any output value of DMUk without also
increasing at least one input value or decreasing at least one other output. All DMUs

on the frontier, but which are not Pareto-Koopmans efficient, i.e., the members of F,
are labeled weakly efficient. Inefficient DMUs have 0* < 1 for the input-oriented model
and 0* > 1 for the output-oriented model. With EDEA, d* and 4* values for inefficient

DMUs are identical to their corresponding variables, 0* and 0*, in the DEA models.
However, frontier efficient DMUs in EDEA are characterized by 0 > I or 4* < 1. To
understand this, a discussion of the DEA formulations is in order.
The description of the DEA formulations gives insight into the meaning of 0*
(*). The DEA models identify the set of efficient DMUs, '*, as those forming an
empirical production surface which envelops all DMUs in D, the set under investigation,
111,31,33,39,41]. For the CCR, BCC, and additive models, all DMUs in D are evaluated
against the DMUs which are members of E'*. For the ECCR and EBCC models, the
problem reference set is denoted as D - k, where DMUk is the unit under evaluation.
If k U F, then DMUk is inefficient and 0*
=8* or = *. However, if k E"E* U F,

and since k 9 D, then 0* > 1 or 4* < 1. This results because DMUk will compare to the
remaining frontier formed by the efficient DMUs of D. In this case 0* can be interpreted
in terms of the Farrell measure [511; DMUk may increase its inputs by a factor of 0* and

remain frontier efficient. Correspondingly, for the output-oriented model, DMUk may

decrease its outputs by a factor of 4* and remain frontier efficient.

Charnes, Cooper, and Thrall place the scale-efficient DMUs for the CCR model
into three categories which they label E, E', and F [411. For the input-oriented DEA
model, DMUk E E U El U F if and only if 8* = 1 (i.e., DMUk is frontier efficient).
Mathematically, the members of E are extreme rays (CCR model) or extreme points

102
(BCC model) which, when taken with their associated facets, form a piecewise-empirical
production surface, or Pareto-Koopmans frontier 1331. For the ECCR and EBCC models,
only members of E will have 0" > 1. Members of E' are non-extreme rays (or non-
extreme points) but depict DMUs which fall on the frontier and are Pareto-Koopmans

efficient. For the ECCR and EBCC models, only members of E' will have 0 = 1, a' = 0,
and 80 = 0 with an efficient reference set consisting of DMUs which axe members of
E-*. DMUs which axe elements of F represent DMUs on the frontier but which are not

Pareto-Koopmans efficient. For the ECCR and EBCC models only members of F will
have 0* = 1 and s' 54 0 or 80 54 0 with an efficient reference set consisting of DMUs

which axe members of E*. It is important to note that if DMUk is a member of F with
an efficient reference set containing other members of F then, for the ECCR and EBCC
models, the optimal solution for DMUk may result in 0" 1, si = 0, and 80 = 0. In this

case, members of E and F can be distinguished if the efficient reference set is retained in
the solution of ECCR and EBCC, for once the solution for all DMUs in D are found, all

members of F, E, and E* can be identified. Therefore, the two-phase approach suggested

by Ali [31 to identify members of F is not necessary.

Some important observations can be made concerning the above classification

scheme. With input-oriented DEA analysis, the transformation

Xk* +- O*Xk - 8 *, k +'8*

Yk +- YE (4.11)

projects DMUk onto an associated Pareto-Koopmans frontier-efficient facet. Correspond-

ingly, the transformation

X +.--Xk-*, Y4-- O*Yk + S* (4.12)

103
projects DMUk onto an associated Pareto-Koopmans frontier-efficient facet for the output-

oriented models 1331. In either case, the projected DMU* consisting of (Yk*, XZ) is called

the virtual representation of DMUk 137].

The virtual representation DMUk at (Ykr, XZ) determines an efficient projection

that is a member of E U E', i.e., Pareto-Koopmans efficient. However, other transforma-

tions are possible that provide important insights. The transformation

Xk +-- i*Xk, Y* +-- Yk (4.13)

will produce an efficient projection for the ECCR' and EBCCi input-oriented models,

which will be a member of E U E'U F (i.e., be frontier efficient). Likewise, for the output-

oriented models ECCRO and EBCCO, the transformation

XkZ*+- Xk, Y* *- *Yk (4.14)

will also produce an efficient projection, DMU*, which is a member of E U E' U F.

Consequently, the primary purpose of the slacks in DEA models is to insure that the

virtual transformation DMU* is Pareto-Koopmans efficient and not simply a DMU on

the efficient frontier.

However, the restriction to limit projections to Pareto-Koopmans-efficient points

is not necessary when determining the allowable perturbations that will cause a DMU to

change status, for any weakly efficient DMU that has all inputs or all outputs perturbed

by a small amount will become either Pareto-Koopmans efficient or completely inefficient.

This has important implications for the EDEA models and corresponding sensitivity

analysis. For the ECCR' and EBCC' input models, 0" determines the maximum

proportional changes in all inputs of DMUk which allow it to maintain its current status.

104
For ECCR' and EBCC, the output oriented models, * determines the maximum

proportional changes in all outputs of DMUk which allow it to maintain its current

status. Consequently, these facts can be beneficial in determining the allowable variations
in inputs or outputs which will not change the status of DMUk.

4.3. CCR and BCC Sensitivity Procedures

The following procedures are used to determine the maximum allowable changes

to inputs and outputs before a DMU changes its status. The procedures differ slightly for
the CCR and BCC input- and output-oriented models. The procedure must be applied
to each DMU in the set in which sensitivity analysis is desired.
Set C = Xj and 17 = Yj of the DMU j under consideration.
Procedure SensCCR(j,,C, v):

1. Set Xj - andYj4 -v.

2. Solve ECCR,'.
3. Let X*- O*X.

4. AXj = X - Xj.

Procedure SensCCRP(j, C, j):

1. Set Xj -- and Yj +- 71.

2. Solve ECCR~j.
3. Let +"
4.

Procedure SensBCC'(j, C,qi):

1. Set Xj -- C and Yj ---l.

2. Solve EBCC'.
3. If EBCC" is infeasible then:
a. DMUj is efficient.

105
 b. AX, = oo.

Else:
a. Let X*' dX.

b. AX,= X*-X.

Procedure SensBCC*(j, C, q):

1. Set Xji -- C and Yj 4- l.

2. Solve EBCCjO.

3. If EBCCj' is infeasible then:

a. DMUj is efficient.

b. AYj = Yj.

Else:
a. Let 1* *- *l.

b. AY = Yj* - Y,.

In the EDEA input models, if 0> _ 1, then DMUk is efficient, and the maximum

increase of inputs that allows the DMU to remain efficient is AXk = O*Xk - Xk. If

0* < 1, then DMUk is inefficient and AXk determines the necessary decrease of inputs

to cause the DMU to become efficient. Note that AXk identifies ranges such that if all

perturbations remain less than the AXk values, the DMU will retain its current status.

Therefore, the effects of non-proportional variations of the inputs on the DMU's status

can be evaluated. It is also important to note that only when all perturbations exceed

the AXk ranges will the DMU be guaranteed to change its status. If even one input

perturbation is smaller than the allowable range, a change in status is not assured for the

DMU. This will be discussed further in Section 4.4.

A similar interpretation can be made concerning the EDEA output-oriented mod-

els. In these models, allowable proportional changes in outputs are determined by

106
AYk = t*Yk - Yk. If * < 1, DMUk is efficient and the maximum allowable decrease of

outputs is given by AYk. For * > 1, DMUk is inefficient and AYk identifies the required

increases of the outputs of DMUk which will cause the DMU status to change.
In the BCC model, the EDEA linear programming formulations may be infeasible

for the DMU of interest, DMUk, in either the input or the output model. For the input-
oriented model, this would mean that DMUk is producing one or more outputs in an

amount greater than any other DMU. Consequently, regardless of how much the inputs
of DMUk are increased, DMUk will remain Pareto-Koopmans efficient at its own scale
level. For the output-oriented model, an infeasible solution indicates there is at least one

input for which DMUk consumes lcss than any other DMU. In this case, outputs can be
reduced to zero and the DMU will remain efficient at its own scale level. Because of their

special characteristics, these DMUs will be termed critically efficient because eliminating
them from the data set critically alters the feasible region. Pre-processing of the data can
reveal these DMUs prior to performing any sensitivity analysis. It is important to note,

that the problem of infeasibility is unique to the BCC model and not the CCR model

because of the feasible regions: the feasible region of the BCC model is a closed convex
set, whereas the feasible region of the CCR model is a convex cone.
The advantage of EDEA over previous models for conducting sensitivity analysis

is that it separates the effects of perturbations of inputs from those of the perturbations
of outputs. In general, the results are only meaningful for measuring effects of pertur-

bations in inputs alone, or outputs alone, but not both simultaneously. Additionally,
the method requires few computational resources and can be used simultaneously with
determining the desired DEA efficiency scores.

107
 4.4. Additive Model Sensitivity Analysis

Charnes, Haag, Jaska and Semple [43] (CHJS) presented the following procedure

for performing sensitivity analysis in DEA based on the additive model. Given the
empirical points (Xj,Yj), j 1,... ,n, with input vector Xj > 0 with at least one
xij positive and output vector Yj >_0 with at least one yrj positive, Charnes, et. al. [43],
define the empirical production set as

n n
PE= y ,, 'nI(y,ax) =Z i(YiXi); Z ,i = 1, A >_o
i=1 i=1

where \j are real scalar coefficienis.

The efficiency of a DMU is determined by comparing its vector of outputs and
inputs with the empirical production set. The DMU of interest is efficient if its vector is
Pareto-Koopmans efficient with respect to PE.
A unique feature of the additive model is that it does not require an input or

output orientation. The additive model is given as:

(ADDk) max z = 180 + 1.I (4.15)

s.t. YX - S' = Yk (4.16)

XA + 8i = Xk (4.17)

1A = 1 (4.18)

X, ssi > 0. (4.19)

DMUk is efficient if and only if z* = 0 at optimality. As in all DEA formulations, with

this method, a linear program must be solved for each DMU tested.

The purpose of the CHJS paper is to present a new procedure to conduct sen-
sitivity analysis for a specific class of DEA models which allows for a simultaneous

108
proportional change in outputs and inputs. For each DMU, CHJS calculate a "region

of stability," that is, a symmetric cell such that all perturbations within the cell maintain
the DMU's status. The shape of the cell is affected by the norm used. The authors
present an analysis based on the L. norm as well as the L1 norm. Only the Lo norm

will be discussed since the same conclusions can be reached for the L1 norm.

To understand the CHJS procedure, a few definitions are necessary. Given

DMUk, the radius of stability is the largest number, rp, such that arbitrary perturbations

to DMUk with p-norm strictly less than rp preserve the DMU's current state. The L..
norm is defined as tIzIll = max Izil.
A characteristic of the CHJS approach is that the analysis must be carried out

differently for efficient DMUs and inefficient DMUs. In the CHJS approach, a procedure
is needed to first determine the efficiency status of each DMU. Once this status is
determined, sensitivity analysis can be conducted on each DMU with an appropriate
procedure, depending if the DMU is efficient or inefficient. Note, with the EDEA
formulation, computational efficiency is attained because one process determines the

efficiency status and the necessary information to conduct sensitivity analysis for each
DMU. The CHJS sensititivity analysis can be described as follows. For efficient DMUk,

the authors describe the L.,e model as:

(Looe-norm Model) min 0 (4.20)

s.t. yfk1]\- 8_ + 10 = y, (4.21)

X[k]A + Si - 10 = Xk (4.22)

1A = 1 (4.23)

x, so, 8i, 0 > 0. (4.24)

If DMUk is inefficient, the Loi model must be used. It is written as:

109
(Li-norm Model) max 0 (4.25)

s.t. y[k]A - _ - 10 Yk (4.26)

X[k]A + si + 10 =X (4.27)

1= 1 (4.28)

Ak, 8,8', 0 > 0. (4.29)

The L,,e model determines the maximum permissible decrease in any output with

a simuLIaneous increase in any input before the efficient DMU becomes inefficient. On the

other hand, the Loo model finds the required decrease of inputs along with an increase

in outputs that the inefficient DMU must witness before it changes status and becomes

efficient.

The authors present proofs demonstrating that 0* for both models is the radius

of stability r.. Having determined 0*, the authors claim that allowable perturbations to

an efficient DMU of the form

lXAJ xk + lr

where r E [0, 0*), preserve efficiency for the DMU. No corresponding analysis is presented

for the inefficient DMUs, but the allowable perturbations for the inefficient DMUs would

take the form

where 7- E [0, 0*). Only when r > 0* will a change in status be assured for the inefficient

DMU.

110
 The authors claim that the projected points found when r = 0* always form an

unstable point, and therefore prove the region of stability is valid. A point (Y,, XA) is

unstable if and only if for all c > 0, the open ball centered at (Y1,Xj) with radius E con-

tains efficient and inefficient points with respect to the associated empirical production

possibility set.

4.4.1. Sensitivity Analysis Example

The results of the CHJS approach will now be compared with the ECCRi,

ECCRO, EBCC1 and EBCC0 models' results using the data presented in the CHJS

paper. !n this example, there axe three DMUs, each with one output and two inputs, as

shown in Table 4.1. The results of an analysis of efficient DMUA are given in Table 4.2.

Table 4.1. -- Sensitivity analysis example

DMUA DMUB DMUC

Y!j 1.00 0.25 0.25
Xlj 1.00 0.25 1.00
X2J 1.00 1.00 0.25

Table 4.2.-- Comparative sensitivity results

EDEAi
ECCRI EBCCi ECCRO EBCCO CHJS
Ay1A I - - - 0.643 -0.750 -0,750
AXIA 1.800 infeas. - - 0.750
AX2A 1.800 infeas. - 0.750

To make proper comparisons between the EDEA models and the CHJS approach,

the EBCC models should be observed. The infeasibility of the EBCCt model indicates

that DMUA is efficient at its own scale size, therefore, the inputs can be increased

indefinitely and the DMU will remain efficient. The EBCC0 model indicates how far

111
the output can be reduced for DMUA to remain efficient. In this case, as long as

YIA > 0.25, DMUA will remain efficient. With the information from both the input-
and output models, one could conclude that simultaneous changes can occur in both the
inputs and outputs and DMUA will remain efficient. As long as YlA > 0.25, inputs

can increase indefinitely and DMUA'S status will remain unchanged. Notice, the CHJS

approach yields similar, although more limited, results. The CHJS results indicate that
the inputs and outputs can simultaneously be perturbed; as long as YlA > 0.25 and

XIA < 1.75, and X2A < 1.75, DMUA will be efficient. Compared to the EDEA models, it

is evident that the CHJS results place an artificial limit on the allowable perturbations

of the inputs.

An additional benefit of the EDEA sensitivity analysis approach is that the

1 and ECCR models reveal information on the perturbations which allow DMUA
ECCR

to remain scale efficient. Information on any aspects of scale efficiency are unavailable

with the CHJS approach.

To compare the CHJS approach to the ECCRi and ECCR0 models, a simple

scaling of the data, so that all DMUs are at the same scale size, is useful. Since there is

only one output, this scaling can be accomplished by dividing all data for each DMU by

its output value. The normed data are presented in Table 4.3.

Since all DMUs produce a single output at the same level, a unit isoquant plot of

the data is possible as depicted in Figure 4.1. In this example, DMU A is efficient, B is

weakly efficient, and C is inefficient.

Table 4.3. -- Normed data example

DMUA DMUB DMUc

Ylj 1.00 1.00 1.00
Xlj 1.00 1.00 4.00
X2J 1.00 4.00 2.00

112
 X2
4 B 4,

3 -- - - - - -- -

2S

I - lAe
*o I

I I I: I1 I I
1 2 3 4 5 x,

Figure 4.1. Normalized data isoquant plot

Table 4.4.-- Normed data sensitivity results

EDEA
ECCRI EBCC' ECCRO EBCC0 CHJS
ATYIA - - -0.643 infeas. -1.800
AZ1lA 1.800 1.800 - - 1.800
AX2A 1.800 1.800 - 1.800

Results for the EDEA and CHJS sensitivity models are given in Table 4.4. The

infeasibility of the EBCC' model indicates that reductions in output are possible; as

long as output is greater than zero, the DMU will remain Pareto-Koopmans efficient.

Note however, if YlA < (1 - 0.643) = 0.357, DMUA will no longer remain scale efficient.

The ECCR' input-oriented models indicate that as long as YA is held fixed, inputs can

be increased to 2.8 and DMU A will remain both scale and Pareto-Koopmans efficient.

In fact, when both the ECCR' and ECCRO models are paired, one would conclude

that simultaneous perturbations of both outputs and inputs are possible. As long as

inputs remain less than 2.8, output can be reduced to 0.357 and DMUA will remain scale

113
efficient. DMUA will remain Pareto-Koopmans efficient even if output is reduced to 0, as
long as inputs remains less than 2.8.

The results of CHJS models reveal a potential problem with their approach.

The results indicate as long as the inputs remain less than 2.8, DMU A will remain

Pareto-Koopmans efficient, even with simultaneous reductions in outputs. Further, A

will remain efficient, according to CHJS, even if output is reduced to i1A = -0.800. This

clearly is not true. The causes of the problem can be identified. The authors claim the

projection

1i.[vi, - 0* [ '0.81
X* X/o+ 0*= 2.8
L J [ X2o + 0* 2.8

is an unstable point. The projection must be an unstable point for the region of stability

to be legitimate. Generally, it is assumed that at least one output must be positive

in efficient production analysis, such as DEA. Therefore, the projected point cannot

be unstable. In fact, the projected point cannot be a valid member of the modified

empirical production possibility set PE. Consequently, although the CHJS approach

provides important insights into sensitivity analysis for DEA, there are special cases that

are problematic with that approach.

4.5. Expanded Sensitivity Analysis

An attractive feature of the EDEA model is its simple implementation using

existing DEA methodologies. However, a concern of this approach is that the resulting

region of stability may be a small subset of the true allowable perturbations for the DMU

to maintain its status. Figure 4.2 illustrates this. The shaded area above points A, E,

El and below points A, B, C, D represents the allowable perturbations for which DMU

114
 X,26 A

4
Y3

1 2 3 4 5 6 7

Figure 4.2. Total region of stability

E remains efficient. The EDEA method reveals a region of stability represented by the

shaded box.

A larger region of stability is revealed by repeated application of the EDEA

methodology, as described below. Procedures ESA-I and ESA-O characterize a series of

stable regions for changes in the input and output vectors, respectively, for a given DMU

j to the degree of refinement desired. Recall that R - {1,. .. ,s} and Q - {l,...,m}

are the respective index sets for output and input measures. Initially, let - X7 and

Procedure ESA-I(j, C,Y)):

1. Apply procedure SensCCRW(j, C, 17) or SensBCCi(j, C, i7) to compute AX).

2. If the region of interest is not fully characterized, then:

For all q E Q: solve ESA-I(j, C,v + eqAxqj)

Procedure ESA-O(j, C,ii):

1. Apply procedure SensCCR(j, C, i) or SensBCCO(j, C,q) to compute Al'.

2. If the region of interest is not fully characterized, then:

For all r E R: solve ESA-O(j, C,, q + erAyrj)

where ek is a column vector of zeros with a 1 in the kth row.

115
 X2

,,".. ,-" El

1 2 3 4 5 6 7 XI

Figure 4.3. Expanded sensitivity analysis

Figure 4.3 reveals how one iteration of this procedure expands the region of

stability for an efficient DMU, E. Because this DMU contains only one output, only the

ESA-1 portion of the algorithm is relevant. Notice, the ESA algorithm can be applied

iteratively to yield larger regions of stability if desired. However, the number of linear

programs required to trace larger regions grows rapidly. Table 4.5 shows the input and

output values for the DMUs depicted in Figure 4.3.

Table 4.5. -- Expanded sensitivity example results

DMUA DMUB DMUC DMUD DMUE DMUE. DMUE* DMUE.

Ylj 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
Xij 1.00 2.50 4.00 7.00 1.50 3.25 5.84 2.00
x2j 6.00 3.50 3.00 2.50 1.50 3.25 2.69 4.33

The results of the expanded analysis now yield greater information than the

EDEA sensitivity analysis. With straight EDEA sensitivity analysis, results indicate

that for DMU E, both inputs can be increased from 1.5 to 3.25 before the DMU becomes

inefficient. However, the expanded analysis yields more detailed information. DMU E*

indicates that as long as X2E _52.69, xlE can be increased to 5.84 and E will remain

116
efficient. Likewise, DMU q indicates X2E can be increased to 4.33 as long as XIE < 2.00.

As can be seen in this example, the expanded sensitivity analysis reveals a greater region

of stability in which non-proportional increases in inputs can be witnessed while the

DMU maintains its status. This same analysis can apply to an inefficient DMUs as

depicted by DMU I in Figure 4.4.

x2

52

1 2 3 4 5 6 7x

Figure 4.4. Inefficient DMU expanded sensitivity analysis

Because both input and output models must be solved for each DMU, the number

of linear programs required to trace the expanded region grows rapidly. For each DMU,

(s + m) additional LPs must be solved for each additional increase in the region of

stability that is measured. Note however, the CHJS L1 norm technique for inefficient

DMUs also requires (s + m) LPs to be solved.

4.6. Summary

The motivation for the sensitivity analysis for the efficient DMUs differs substan-

tially from that for the inefficient DMUs. The managers of the efficient DMUs desire

to know how much loss in productive efficiency is allowable before the DMU becomes

relatively inefficient compared to its peers. The inputs that permit only small increases

117
before efficiency is lost play a different role in the production process than the inputs
which can suffer large perturbations and not result in a change of state. This information

can give the manager a sense of the relative worth, or price value, of each input. In this
way, managers can plan for future markets by shifting to production technologies that
provide greater regions of stability. From the output-oriented model, the manager is able

to gain a more complete picture of the role of each output. By knowing which outputs
are most critical to maintain relative efficiency, the manager is more able to evaluate the

opportunity costs associated with producing each output.

For the inefficient DMUs, the manager wants to identify the smallest region of
stability. This will reveal how best to allocate resources to move to efficiency. Instead of
down-sizing across all inputs with proportional reductions, the manager may be able to
focus on reducing those resources that will more rapidly bring the DMU to the efficient
frontier. Likewise, by focusing on output expansion, the manager may be able to identify

a market niche with a few of outputs in which the DMU can achieve relative efficiency.
The sensitivity analysis presented in this chapter presents the manager with
useful information about the role each output and input variable plays in determining
the efficiency of the production process. In this manner, the manager can concentrate
attention on those factors which can more quickly result in increased productive efficiency
when considering policy decisions.

118
 CHAPTER V
SUMMARY AND CONCLUSIONS

This dissertation contributes to the development of solution techniques for solving

large-scale data envelopment analysis problems which involve the solution of thousands
of linear programs. Although the focus is on large-scale problems, the new algorithms
presented can be used to effectively solve a wide range of DEA problems.
The analysis presented in Chapter II describes a new problem decomposition
procedure which dramatically expedites the solution of the computationally intensive
DEA problems and fully exploits parallel processing environments. Computational ex-

periments indicate that by combining linear programming theory with known character-
istics of the DEA formulations, solution times can be cut in half compared to general
linear programming approaches. When utilizing a parallel processing environment, the
efficiency remains above 0.97 indicating that the solution time will improve by the same
magnitude as the number of processors used. A new decomposition algorithm is pre-

sented which causes a 6- to 10-fold improvement in the solution time for the serial case.
When the decomposition procedure is coupled with the parallel processing environment,
solution times can be increased by up to two orders of magnitude.
Besides effectively reducing the time to solve large-scale DEA problems, the de-

composition approach offers many additional advantages over prior DEA solution meth-
ods. It allows for the solution of a large number of different DEA formulations in rapid

succession. The decomposition approach opens a new means of conducting window anal-
ysis as well as studying categorical variables. Additionally, the decomposition approach

helps avoid problems, such as cycling, which frequently occur in larger DEA problems.

119
 Building on these improvements to the DEA solution process, Chapter III outlines

a new rank ordering procedure to classify DMUs by their ability to apply current technol-

ogy to the production process. This new ranking procedure acts as a proxy to measure

managerial effectiveness in a way not possible by traditional DEA methods. Best and

worst performing DMUs are easily classified with the new ordering procedure. In this

way, the analyst can compare the management practices of the DMUs from both clas-

sifications to assist in determining which policies improve and detract from productive

efficiency.

In Chapter IV, we presented a computationally intensive means of conducting

sensitivity analysis on the DEA results. In the chapter, the new sensitivity analysis

procedure is compared with previous approaches. The new procedure is shown to avoid

problems that may arise with the previously reported methods. The key advantage

of the sensitivity analysis is that it helps clarify which factors can cause significant

improvements to the production process.

The study of large-scale DEA problems reveals many new issues that have gone

unnoticed with the study of smaller problems. The insights gained by investigating these

new problems can continue to result in new analytical tools to broaden the usefulness of

DEA. The techniques presented in this dissertation open the door to new research which

will further expand the range of applications for data envelopment analysis.

120
 APPENDIX A
MULTINORMAL DATA GENERATOR

Since few large-scale DEA problems are currently available, randomly generated
data sets were needed to simulate possible real life data. One procedure to generate such
data is to draw the samples from a multinormal distribution. To this end, the DRNMVN
routine from the IMSL library was used to generate data for 10,000 DMUs. The variance
covariance matrix of Table A.1 was used to generate the data for the input and output
variables for each DMU. The code to generate this data is described by Hickman[62].

Table A. 1.- Variance-Covariance matrix for multinormal data

X1 X2 X3 X4 X5 X6 X7 X8
X1 303.4
X2 2.5 2.5
X3 12493.9 6093.9 112750625.0
X4 4062.6 484.2 -2890405.7 31033992.2
X5 265.2 45.3 -1346076.8 -455158.7 7629765.5
X6 19597.2 -195.0 3436081.9 -270405.4 -674642.7 86492323.0
X7 36418.8 6428.4 111950224.4 27415022.3 5153887.3 88983356.8 233502490.7
x8 3647.0 1001.0 10815783.1 112293.8 -64824.7 - 319883.0 10543369.3 7812141.0

From the randomly generated data, the variables representing inputs and outputs
needed to be carefully chosen. To coincide with sound economic theory, all outputs
should be positively correlated with all the inputs. In Fig. A.1, all negatively correlated
variables are connected with a line.
By eliminating variable z4, variables xl, X3, and X7 could represent outputs

since they would be positively correlated with all the other remaining variables that

121
 Figure A.1. Correlations of multinormal data variables

would represent inputs. In this way, each DMU would be comprised of 3 outputs

and 4 inputs. Notice also that inputs positively correlated would be compliments in

the production process and those negatively correlated would be substitutes. Since the

input variables cover a wide range of compliment and substitute relationship cases, the

randomly generated data further simulates possible real life data.

122
 APPENDIX B
DEA-GEN

While random data generators can provide large problems, a more systematic

approach is need to represent real life scenarios. To this end, economic theory was used
to generate large-scale DEA problem sets. In production economics, the most widely used
functional form is known as the Cobb-Douglas production function [241. This function is
written as:

m
y,= a. x4, xij >0,=l .n
2=l

Here, yj is the single aggregate output produced by DMUj, xij is the value of input i
used by DMUj in the production process, ai is the elasticity factor for input i, and ao
is a constant scale factor. If E'n1 a2 = 1 then only constant returns to scale exist in

the production process. For '= ai < 1 decreasing returns to scale are present, while

ET, ai > 1 indicates increasing returns to scale. In the DEA studies, increasing returns
to scale are not used because the function results in only a few DEA efficient points.
Although this does not pose a problem, it does not realistically represent true life data.
For the single output model, this production function has many desirable prop-

erties. If the inputs are randomly generated, the function generates output values that

will always lie on the production possibility frontier, i.e., they will be DEA efficient for

IxI a2 < 1. This frontier is central to the theory of economic growth and measures
the rate of technological progress. The Cobb-Douglas frontier represents the best use

123
of technology as well as the best management practices to achievc efficient production.

This coincides with the practical use of DEA. Unlike DEA, the quest in economic studies

is to attempt to estimate the values of ai by fitting the function to the observed data.

By choosing the ai values a priori and then randomly generating the input values, the

output values can be determined so that they coincide with widely accepted economic

theory. To insure that not all generated data sets fall on the efficient frontier surface,

the a. scale factor can be randomly generated. The efficient DMUs will consist of those

generated where a, takes on it maximum value. Control over the number of DMUs that

are efficient in the data set can be maintained by limiting the number of DMUs generated

with max a,.

In DEA analysis, the single-output, multiple-input scenario is not of primary

interest. DEA was developed to analyze the case of multiple-output, multiple-input

studies. Faire and Grosskopf [54] point out that the existence of a joint production

function has not been established. That is, the multiple-output, multiple-input model

does not produce values strictly on the efficient production frontier. To do so, would

require strict assumptions that would not provide the desired realism for the DEA study.

However, a joint model without the strict assumptions would simulate economically

sound production processes. Consequently, a joint model was developed for the DEA-

GEN problem generator.

For each of the problem sets, the input values were generated from a uniform

distribution. The ai values were chosen to simulate constant returns to scale processes as

well as a variety of cases representing different levels of decreasing returns to scale. The

a. value was randomly generated, but with a modest control of the number of efficient

DMUs that could be present in the problem data.

Once the single aggregate output level was calculated, the individual output levels

were determined by assigning each individual output as a percentage of the aggregate.

124
The percentages for each individual output were drawn from normal distributions with

predetermined means and standard deviations. The means of the normal distributions

were chosen so that the percentages sum to one. Table B.1 lists the ai values as well

as the means and standard deviations that were used to generate twelve different Cobb-

Douglas data sets.

Table B. .-- Parameter values for DEA-GEN code

#DMUs a Values Mean Values Standard Dev.

DEA-GENa 1,000 .3,.2,.2,.3 .12,.13,.13,.14,.11 .03,.04,.02,.01,.01
DEA-GENa 2,000 .3,.2,.2,.2,.1 .20,.30,.15 .03,.04,.02
DEA-GENa 4,000 .3,.3,.2,.2 .12,.12,.15,.14 .03,.04,.02,.O1
DEA-GENa 8,000 .4,.2,.1,.1,.1 .5 .1
DEA-GENb 1,000 .03,.07,.12,.03,.12,.1,.02 .2,.3,.15 ')1,.05,.01
DEA-GENb 2,000 .1,.1,.1,.1,.1,.1,.1 .52,.13 .03,.04
DEA-GENb 4,000 .4,.3,.1 .4,.3,.15 .13,.08,.02
DEA-GENb 8,000 .3,.17,.12,.03,.12,.2 .2,.3,.15,.14 .01,.05,.01,.01
DEA-GENc 1,000 .18,.2,.14,.1,.1,.1 .2,.12,.14,.11,.11 .03,.04,.02,.01..01
DEA-GENc 2,000 .13,.18,.2,.14,.1,.1 .2,18,.15,19 .03,.04,.02,.01
DEA-GENc 4,000 .13,.2,.12,.13 .2,.3,.15,.14,.11 .03,.04,.02,.01,.01
DEA-GENc 8,000 .08,.12,.21,.07,.11 .3,.3,.3,.1 .04,.04,.1,.03
DEA-GENd 25,000 .3,.3,.3 .3,.3 .01,.01

Because of the economic foundations of the DEA-GEN code, the data generated

resembles a class of problems that should more closely simulate realistic economic data

than what would be possible from the data sampled from a multinormal distribution.

125
 BIBLIOGRAPHY

[1] D.L. Adolphson, G.C. Cornia and L.C. Walters, A unified framework for classifying
DEA models, Operational Research '90 (1990)647-657.

12] D.L. Adolphson, G.C. Corria and L.C. Walters, Railroad property valuations using
data envelopment analysis, Interfaces 19(1989)18-26.
[3] A.I. Ali, IDEAS: Integrated data envelopment analysis system, Technical Report
(Department of General Business and Finance, University of Massachusetts
Amherst, MA, 1989).
[4] A.I. Ali, Data envelopment analysis: computational issues, Comput. Environ. and
Urban Systems 14(1990)157-165.
[5] A.I. Ali, Facilitating DEA methodology, ORSA/TIMS Joint National Meeting, San
Francisco, CA, 1992.

161 A.I. Ali, Streamlined computation for data envelopment analysis, European Journal
of Operational Research 64(1993)61-67.
[7] A.I. Ali, Computational aspects of data envelopment analysis, to appear in
A. Chaxnes, W.W. Cooper, A. Lewin and L.M. Seiford, DEA: Theory, Methodology,
and Applications (Kluwer Academic Publishers, Boston, MA, 1994).

[8] Al1. Ali, W.D. Cook and L.M. Seiford, Strict vs. weak ordinal relations for
multipliers in data envelopment analysis, Management Science 37(1991)733-738.

[9] A.I. Ali and L.M. Seiford, Computational accuracy and infinitesimals in data
envelopment analysis, Technical Report (Department of General Business and
Finance, University of Massachusetts, Amherst, MA, 1991).

[101 R.D. Banker, Estimating most productive scale size using data envelopment
analysis, European Journal of Operations Research 17(1984)35-44.

[11] R.D. Banker, A. Charnes and W.W. Cooper, Some models for estimating
technical and scale inefficiencies in data envelopment analysis, Management Science
30(1984)1078-1094.

[12] R.D. Banker, A. Charnes, W.W. Cooper and A.P. Schinner, A bi-extremal
principle for frontier estimation and efficiency evaluations, Management Science
27(1981)1370-1389.

126
1131 R.D. Banker, R.F. Conrad and R.P. Strauss, A comparative application of data
envelopment analysis and translog methods: an illustrative study of hospital
production, Management Science 32(1986)30-44.

[14] R.D. Banker, S.M. Datar and C.F. Kemerer, A model to evaluate variables
impacting the productivity of software maintenance projects, Management Science
37(1991)1-18.

[151 R.D. Banker and A. Maindiratta, Piecewise loglinear estimation of efficient

production surfaces, Management Science 32(1986)126-135.

116] R.D. Banker and R.C. Morey, Efficiency analysis for exogenously fixed inputs and
outputs, Operations Research (1985)513-521.

[17] R.D. Banker and R.C. Morey, The use of categorical variables in data envelopment
analysis, Management Science 32(1986)1613-1627.
[181 R.S. Barr and M.L. Durchholz, Parallel software for large-scale data envelopment
analysis, ORSA/TIMS Joint National Meeting, San Francisco, CA, 1992.

[19] R.S. Barr and M.L. Durchholz, Parallel and hierarchical decomposition approaches
for solving large-scale DEA models, ORSA/TIMS Joint National Meeting, Chicago,
IL, 1993.
[20] R.S. Barr and B.L. Hickman, Reporting computational experiments with parallel
algorithms: issues, measures, and experts' opinions, ORSA Journal on Computing
5(1993)2-18.
[21] R.S. Barr, L.M. Seiford and T. Siems, An envelopment-analysis approach to
measuring the managerial quality of banks, Annals of Operations Research
42(1993)1-19.
[22] R.S. Barr and T. Siems, Predicting bank failure using DEA to quantify management
quality, Technical Report (Department of Computer Science and Engineering,
Southern Methodist University, Dallas, TX, 1993).
[231 P.W. Bauer, A.N. Berger and D.B. Humphrey, Inefficiency and productivity growth
in banking: a comparison of stochastic econometric and thick frontier methods,
Financial Studies Section, Federal Reserve Board Washington DC (1991).

[24] M.S. Bazaraa, J.J. Jarvis and H.D. Sherali, Linear Programming and Network Flows
(John Wiley & Sons, NY, 1990).

[25] R.E. Bixby, Implementing the simplex method: the initial basis, ORSA Journal on
Computing 4(1992)267-284.
[26] R.G. Bland, New finite pivoting rules for the simplex method, Mathematics of
Operations Research 2(1977)103-107.
[27] W.F. Bowlin, A. Charnes, W.W. Cooper and H.D. Sherman, Data envelopment
analysis and regression approaches to efficiency estimation and evaluation, Annals
of Operations Research 2(1985)113-138.

127
[281 P. Byrnes, R. Fare and S. Groaskopf, Measuring production efficiency: an
application to Illinois strip mines, Management Science 30(1984)671-680.

1291 P. Byrnes, S. Grosskopf and K. Hayes, Efficiency and ownership: further evidence,
The Review of Economics and Statistics 68(1986)337-341.
[30] A. Charnes and W.W. Cooper, Management models and industrial applications of
linear programming (John Wiley & Sons, NY, 1961).

[31] A. Charnes and W.W. Cooper, Preface to topics in data envelopment analysis,
Annals of Operations Research 2(1985)59-94.

[32] A. Charnes and W.W. Cooper, Data envelopment analysis, Operational

Research '90 (1990)641-646.
[33] A. Charnes, W.W. Cooper, B. Golany, L. Seiford and J. Stutz, Foundations of data
envelopment analysis for Pareto-Koopmans efficient empirical production functions,
Journal of Econometrics 30(1985)91-107.
[34] A. Charnes, W.W. Cooper and Z.M. Huang, Polyhedral cone-ratio DEA models
with an illustrative application to large commercial banks, Journal of Econometrics
46(1990)73-91.
[35] A. Charnes, W.W. Cooper, Z.M. Huang and D.B. Sun, Relations between half-space
and finitely generated cones in polyhedral cone-ratio DEA models, International
Journal of Systems Science 22(1991)2057-2077.
[36] A. Charnes, W.W. Cooper, A.Y. Lewin, R.C. Morey and J. Rousseau, Sensitivity
and stability analysis in DEA, Annals of Operations Research 2(1985)139-156.
[37] A. Charnes, W.W. Cooper and E. Rhodes, Measuring the efficiency of decision
making units, European Journal of Operational Research 2(1978)429-444.
[38] A. Charnes, W.W. Cooper and E. Rhodes, Short communication: measuring the
efficiency of decision making units, European Journal of Operational Research
(1979)339.
[39] A. Charnes, W.W. Cooper and E. Rhodes, Evaluating program and managerial
efficiency: an application of data envelopment analysis to program follow through,
Management Science 27(1981)668-697.

[40] A. Charnes, W.W. Cooper, L. Seiford and J. Stutz, A multiplicative model for
efficiency analysis, Journal of Socio-Economic Planning Science 16(1981)223-235.
[41] A. Charnes, W.W. Cooper and R.M. Thrall, Classifying and characterizing
efficiencies and inefficiencies in data envelopment analysis, Operations Research
Letters 5(1986)105-110.
[42] A. Charnes, W.W. Cooper and R.M. Thrall, A structure for classifying and
characterizing efficiency and inefficiency in data envelopment analysis, The Journal
of Productivity Analysis 2(1991)197-237.

128
[43] A. Charnes, S. Haag, P. Jaska and J. Semple, Sensitivity of efficiency classifications
in the additive model of data envelopment analysis, International Journal of
Systems Science 23(1992)789-798.

[44] A. Charnes and L. Neralid, Sensitivity analysis of the additive model in data
envelopment analysis, European Journal of Operational Research 48(1990)332-341.

[45] A. Charnes and L. Nerali6, Sensitivity analysis of the simultaneous proportionate

change of inputs and outputs in data envelopment analysis, Research Report CCS
667 (Center for Cybernetic Studies, University of Texas at Austin, 1991).

[46] A. Charnes, J. Rousseau and J. Semple, An effective non-archimedean anti-

degeneracy/cycling linear programming method especially for data envelopment
analysis and like models, Annals of Operations Research 32(to appear 1994).

[47] C.W. Cobb and P.H. Douglas, A Theory of Production, American Economic Review
(1928)139-165.

[48] W.D. Cook and M. Kress, A data envelopment model for aggregation preference
rankings, Management Science 36(1990)1302-1310.

[49] T.H. Cormen, C.E. Leiverson and R.L. Rivest, Introduction to Algorithms (MIT
Press, Cambridge, MA, 1990).

[50] CPLEX Optimization Inc., Using the CPLEX Linear Optimizer (CPLEX
Optimization Inc., Incline Village, NV, 1990).

[51] M.J. Farrell, The measurement of productive efficiency, Journal of Royal Statistical
Society Series A 120(1957)253-290.

[52] M.J. Farrell and M. Fieldhouse, Estimating efficient production functions under
increasing returns to scale, Journal of Royal Statistical Society Series A (1962)252-
267.

[53] R. Fdre, S. Grosskopf, C.A.K. Lovell, The measurement of efficiency of production

Kluwer-Nijhoff Publishing, MA, (1985)1-216.

[541 R. FMre and W. Hunsaker, Notions of efficiency and their reference sets,
Management Science 32(1986)237-243.

[55] G.D. Ferrier and C.A.K. Lovell, Measuring cost efficiency in banking, econometric
and linear programming evidence, Journal of Econometrics 46(1990)229-245.

[561 M.J. Flynn, Very high-speed computing systems, Proceedings of the IEEE
54(1966)1901-1909.

[57] F.R. Fcrsund, C.A.K. Lovell and P. Schmidt, A survey of frontier production
functions and of their relationship to efficiency measurement, Journal of
Econometrics 13(1980)5-25.

[58] T. Gal, H. Kruse and P. Zonig, Survey of solved and open problems in the
degeneracy phenomenon, Mathematical Programming 42(1988)125-133.

129
[591 S.I. Gass, Comments on the possibility of cycling with the simplex method,
Operations Research 27(1979)848-852.

[601 P. Hall, W. Hirdle and L. Simar, Iterated bootstrap with applications to frontier
models, CORE Discussion Paper No. 9121 (Center for Operations Research and
Econometrics, Belgium, 1991).
[61] B. Hattersley and J. Wilson, A dual approach to primal degeneracy, Mathematical
Programming 42(1988)135-145.

162] B.L. Hickman, Parallel algorithms for pure network problems and related
applications, Ph.D. thesis (Southern Methodist Universitv. Dallas, TX, 1991).
[63] IBM Corporation, Mathematical Programming System Extended/370 Program
Reference Manual (IBM Corporation, Kingston, NY, 1979).
[64] IBM Corporation, Optimization Subroutine Library Reference (IBM Corporation,
Kingston, NY, 1979).
[65] W.A. Kamakura, A note on the use of categorical variables in data envelopment
analysis, Management Science 34(1988)1273-1276.
166] D.E. Knuth, The Art of Computer Programming Volume 3 / Sorting and Searching
(Addison-Wesley, Reading, MA, 1973).

[67] L.S. Lasdon, Optimization Theory for Large Systems (Macmillan, NY, 1970).
[68] A.M. Law and W.D. Kelton, Simulation Modeling and Analysis, 2nd ed. (McGraw-
Hill, NY, 1990).
[69] A. Maindiratta, Largest size-efficient scale and size efficiencies of decision-making
units in data envelopment analysis, Journal of Econometrics 46(1990)57-72.
[701 R. Marsten, The design of the XMP linear programming library, ACM Transactions
on Mathematical Software 7(1981)481-297.
[71] J.M. Mulvey, Pivot strategies for primal-simplex network codes, Journal of the
ACM 25(1978)206-270.
[72] B.A. Murtagh, Advanced Linear Programming (McGraw-Hill, NY, 1981).

173] E.D. Nering and A.W. Tucker, Linear Programs and Related Problems (Harcourt
Brace Jovanovich, Boston, MA, 1993).

[74] Orchard-Hays, Advanced Linear Programming Computing Techniques (McGraw-

Hill, NY, 1968).

[751 A. Osterhaug, Guide to Parallel Programming On Sequent Computer Systems

(Sequent Computer Systems, Inc., OR, 1986).

130
[761 F. Phillips, R.G. Parsons and A. Donoho, Parallel microcomputing for data
envelopment analysis, Computing, Environment, and Urban Systems 14(1990)167-
170.

[77] D.L. Retzlaff-Roberts and R.C. Morey, A goal-programming method of stochastic

allocative data envelopment analysis, Technical Report (Memphis State University,
TN, 1992).
[78] J.J. Rousseau and J.H. Semple, Categorical outputs in data envelopment analysis,
Management Science (forthcoming).

179] J.J. Rousseau and J.H. Semple, Non-archimedean infinitesimals and categorical
inputs in data envelopment analysis, International Journal of Systems Science
(forthcoming).
180] L.M. Seiford, A DEA bibliography (1978-1992), to appear in Charnes, A., W.
W. Cooper, A. Y. Lewin, and L. M. Seiford, DEA: Theory, Methodology, and
Applications (Kluwer Academic Publishers, Boston, MA, 1994).
[81] L.M. Seiford and R.M. Thrall, Recent developments in DEA: the mathematical
programming approach to frontier analysis, Journal of Econometrics 46(1990)7-38.

[82] L. Simar, Estimating efficiencies from frontier models with panel data: a
comparison of parametric, non-parametric and semi-parametric methods with
bootstrapping, CORE Discussion Paper No. 9126 (Center for Operations Research
and Econometrics, Belgium, 1991).
[83] Sperry Univac Corporation, Functional Mathematical Programming System
Programmer Reference (Sperry Univac Corporation, St. Paul, MN, 1974).
[84] R.G. Thompson, L.N. Langemeier, C. Lee, E. Lee and R. M. Thrall, The role of
multiplier bounds in efficiency analysis with application to Kansas farming, Journal
of Ecoaometrics 46(1990)93-108.
[85] R.G. Thompson, F.D. Singleton, Jr., R.M. Thrall and B.A. Smith, Comparative site
evaluations for locating a high-energy physics lab in Texas, Interfaces 16(1986)35-
49.
[861 C.P. Timmer, Using a probabilistic frontier production function to measure
technical efficiency, Journal of Political Economy 79(1971)776-794.

131