Harzer wirtschaftswissenschaftliche Schriften
Georg Westermann Editor
Data Envelopment Analysis in the Service Sector
Westermann
Data Envelopment Analysis in the Service Sector
GABLER EDITION WISSENSCHAFT
Harzer wi rtschaftswi ssenschaftl iche Schriften Herausgegeben vom Fachbereich Wirtschaftswissenschaften der FH Harz
In den ,Herzer wirtschaftswissenschaftlichen Schriften" werden Beitrage zu aktuellen okonomischen Fragestellungen veroffentlicht. Die FH Harz in Wernigerode, an der ehemaligen Nahtstelle zwischen Ost und West gelegen, leistet mit dieser Reihe des Fachbereichs Wirtschaftswissenschaften einen Beitrag zur Erfullung der Bruckenfunktion zwischen Theorie und Praxis, zwischen Wirtschaft, Technik und Kultur.
Georg Westermann (Ed.)
Data Envelopment Analysis in the Service Sector
Springer Fachmedien Wiesbaden GmbH
Die Deutsche Bibliothek - CIP-Einheitsaufnahme
Westermann, Georg: Data envelopment analysis in the service sector/ Georg Westermann (ed.). - Wiesbaden: Dt. Univ.-Verl.; Wiesbaden: Gabler, 1999 (Gabler Edition Wissenschaft : Herzer wirtschaftswissenschaftliche Schriften)
Aile Rechte vorbeholten
© Springer Fachmedien Wiesbaden 1999 Ursprünglich erschienen bei Deutscher Universitäts-Verlag 1999
Lektorot: Ute Wrosmonn
Dos Werk einschliel31ich oiler seiner Teile ist urheberrechtlich geschutzt. Jede Verwertung aul3erhalb der engen Grenzen des Urheberrechtsgesetzes ist ohne Zustimmung des Verlag~~ unzuli:issig und strafbar. Dos gilt insbesondere fur Vervielfi:iltigungen, Ubersetzungen, Mikroverfilmungen und die Einspeicherung und Verarbeitung in elektronischen Systemen.
hHp:/ /www.gabler.de hHp:/ /www.duv.de
H&hste inhaldiche und technische Oualiti:it unserer Produkte ist unser Ziel. Bei der Produktion und Verbreitung unserer Bucher wollen wir die Umwelt schonen. Dieses Buch ist deshalb auf si:iurefreiem und chlorfrei gebleichtem PaP.ier gedruckt. Die Einschweii3Folie besteht aus Polyi:ithylen und demit aus organischen Grundstolten, die wader bei der Herstellung noch bei der Verbrennung Schadstoffe freisetzen .
Die Wiedergabe von Gebrauchsnamen, Handelsnamen, Warenbezeichnungen usw. in diesem Werk berechtigt ouch ohne besondere Kennzeichnung nicht zu der Annahme, dal3 solche Nomen im Sinne der Worenzeichen- und Markenschutz-Gesetzgebung als frei zu betrachten wi:iren und daher von jedermann benutzt warden durften.
ISBN 978-3-8244-7012-9 ISBN 978-3-663-08343-6 (eBook) DOI 10.1007/978-3-663-08343-6
Data Envelopment Analysis in the Public and Private Service Sector v
Preface to "Efficiency in the Private and Public Service Sector"
The Aims of this Book
The October 1998 Symposium m Wemigerode, Germany, was the successive European Conference on Data Envelopment Analysis (DEA) to the 1997 meeting in Marseilles. The label, "European", has to be seen with respect to two features:
First, and very obviously, the conference site, The Harz University of Applied Studies and Research, is located in the middle of the European continent. Secondly, an explicitly formulated goal of the symposium was the formation of a European Network of researchers who are concerned with the development and application of the DEA methodology. Looking at the list of participants, this aim has been fulfilled. Over and above that, contacts with colleagues from North and South America as well as Australia and Asia were invigorated and enlarged.
Although DEA has often and effectively been applied to measure efficiency in North American companies and public institutions, DEA is largely unknown amongst European practitioners and academics. Especially the public and private service sector shows some specificity that classical measurement and benchmarking instruments normally fail to serve. Missing prices for (non-marketed) public goods or distinct firmspecific solutions to the same problem - and ,thus, different production techniques -are only two of the frequently arising problems.
Efficiency in the Private and Public Service Sector - Recent Developments in the Application and Methodology of Data Envelopment Analysis was chosen as the title of the conference volume in order to emphasize the adjustment and development of the DEA methodology to the requirements of the service sector and to highlight the widespread application of the instrument - especially to problems the orthodox methods fail to solve or to emerging fields within the service sector.
A striking observation with respect to the DEA literature is that empirical applications in most cases are lagging far behind the theoretical methodological developments. There is often a clear cut distinction between contributions from mathematicians and/or programmers and papers by DEA users. This distinction might not be confined only to the DEA methodology. But the fact should nevertheless be noticed and the theory-to-application-lag should not be allowed to become too large.
VI Data Envelopment Analysis in the Public and Private Service Sector
Context of the Contributions
The number of presentations during the Wemigerode Symposium clearly exceeds the scope of this volume. We decided to include papers that fit into the service sector context either from the methodological or from the empirical point of view. Thus, papers not published in this book are not necessarily of inferior quality.
Lawrence M. Seiford summarizes the development of the DEA methodology over the past two decades. In addition, he provides an agenda for the future in this field of research. The need for a "stochastic" DEA is emphasized and a comprehensive literature survey is given. Thus, this first article may serve as an introduction.
The contributions of Rajiv D. Banker et al. and of Rob Ball and Elizabeth Roberts are splendid examples for successfully applying recent methodological developments in order to solve actual empirical problems in the public accounting industry or in the health sector. Other papers concentrate more or less on theoretical or methodological issues and can be summarized as follows:
Matthias Staat argues in his paper that including productivity relevant but individually uncontrollable variables " ... may lead to comparisons of qualitatively different DMUs." He contrasts the effects of different model specifications for empirical analyses.
Laurens Cherye and Tom van Puyenbroeck point to the fact that for empirical investigations the use of the radial efficiency measure in combination with the
existence of zero data and/or slacks might " ... result in wrong management conclusions." They develop a modified additive model to overcome these difficulties.
Holger Scheel investigates whether the BCC model shows the property of continuity. Continuity ensures that small data errors cause only small perturbations in the efficiency measure. Again the case of zero data is shown to be of crucial importance.
Wenbin Liu and John Sharp explain the empirical problems arising from a mixture of positive and negative responding outputs to changes in inputs. They derive new DEA models from a goal programming approach, which might be applied in cases showing the above mentioned properties.
Data Envelopment Analysis in the Public and Private Service Sector VII
Dieter Gstach compares the performance of the stochastic DEA+ model to the Ray
Frontier Approach by Lothgren (1997). This contribution concentrates on simulations
in order to demonstrate how both models behave when bounded or unbounded noise is
assumed.
Rolf Fare et al. demonstrate how DEA models might be used for computing shadow
prices for firm outputs. This seems especially useful in cases when " ... a market price is
unavailable, administered or distorted." The public service sector is an example for
these non-marketed goods.
Five mainly application oriented contributions build the final part of the book:
Georg Westermann and Gerhard Johnson apply DEA efficiency scores from
different input-output combinations to construct strategic management portfolios for
social service units. The transformation of inputs into capacity, production and final
effectiveness is analyzed.
Katrin Allen summarizes the state-of-the-art in applying DEA to studies in the
ecological context. This paper corresponds very well to the methodological aspects by
Liu and Sharp within this volume.
M. C. Sampaio de Sousa and Francisco S. Ramos measure the efficiency of public
spending in Brazilian Municipalities. The huge database applied, the input-output
combinations and the results are of interest for all researchers engaged in accounting
for public administration performance.
Finn R. Forsund and Kjell 0. Kalhagen are concerned with the evaluation of
Norwegian colleges and trace the respective productivity developments for three years.
This contribution convinces by its input-output model and the database on the
department level..
P.-Y. Badillo compares the DEA scores for technical efficiency to the financial
performance of telecommunication operators in Europe and the USA over the period
from 1986 to 1997. The effects of monopoly and deregulation on both indicators are
discussed.
VIII Data Envelopment Analysis in the Public and Private Service Sector
Acknowledgements
This book would not have been completed without the help of many people to whom I am most grateful.
The ministry of education of Sachsen-Anhalt funded the travel expenses of the invited speakers of the conference. The Hochschule Harz made its facilities available to the symposium and provided financial support for the printing of this volume.
The help of Caren Labohm, Ricarda Otto and Stefan Schneider in organizing and running the conference in Wemigerode is also acknowledged with many thanks.
As editor I am indebted to Stefan Schneider who was responsible for collecting the papers from the authors and completing the manuscript.
While I am giving away thanks for advice and help, I am still responsible for remaining mistakes.
Georg Westermann
Data Envelopment Analysis in the Public and Private Service Sector
Contents
Lawrence. M. Seiford:
Data Envelopment Analysis: Twenty Years Out
Matthias Staat:
Treating non-discretionary variables one way or the other:
implications for efficiency scores and their interpretation
Laurens Cherchye, Tom van Puyenbroeck:
Non-Radial Efficiency as Semi-Radial Efficiency
Holger Scheel:
Continuity of the BCC Efficiency Measure
Wenbin Liu, John Sharp:
DEA Models via Goal Programming
Dieter Gstach:
Bounded vs. Unbounded Noise in Efficiency Estimation:
Performance of Alternative Estimators
Rolf Fare, Shawna Grosskopf, Pontus Roos:
Price Indexes for Nonmarketed Goods
Rob Ball, Elizabeth Roberts, Richard Thomas:
Lessons Learned for DEA Practice from Health Care
Applications in the UK
Rajiv D. Banker, Hsihui Chang, Reba Cunningham, Ram Natarajan:
Recent Advances in Data Envelopment Analysis:
An Illustrative Application to the U.S. Public Accounting
Industry
IX
1
23
51
65
79
103
121
133
159
X Data Envelopment Analysis in the Public and Private Service Sector
Georg Westermann, Gerhard Johnson:
Combining DEA and "Transformation Stages":
Management Strategies for the Disability Service
Units of the St. Georg Association
Katrin Allen:
DEA in the ecological context- An overview
Maria da Concei-;;ao Sampaio de Sousa, Francisco S. Ramos:
Measuring Public Spending Efficiency in Brazilian
Municipalities: A Nonparametric Approach
Finn R. F0rsund, Kjell Ove Kalhagen:
Efficiency and Productivity of Norwegian Colleges
Patrick-Y. Badillo:
Efficiency and financial performances in telecommunications
183
203
237
269
309
Data Envelopment Analysis: Twenty Years Out
Lawrence M. Seiford'
Abstract
This paper briefly traces the evolution of DEA from the initial publication by Charnes, Cooper and Rhodes (1978) to the current state-of-the-art (SOA). The state of development of DEA is characterized at four points in time to provide a perspective in both directions-past and future. An evolution map is provided which illustrates DEA growth during the twenty year period, the timing of the major events, and the interconnections and influences between topics. An extensive DEA bibliography is provided.
(Portions of this paper are based on Seiford, Lawrence M. ,Data Envelopment Analysis: The Evolution of the State-of-the-Art (1978--1995)." Journal of Productivity Analysis 7, no. 213 (1996):pp. 99-138.)
1 Mechanical and Industrial Engineering, University of Massachusetts, Amherst, MA 01003 USA
2 Data Envelopment Analysis: Twenty Years Out
Structure
Introduction
2 In the beginning
3 Evolution ofDEA
References
Data Envelopment Analysis: Twenty Years Out 3
1 Introduction
This paper briefly traces the evolution of DEA from the initial publication by Chames, Cooper and Rhodes (1978) to the current state-of-the-art (SOA). In characterizing the development of DEA over the past 20 years, I describe the then-current SOA at four points in time. These "snapshots" of the development provide a perspective in both directions, i.e., in terms of what was known as well as what remained unknown. The particular milestones (1980, 1985, 1990 & 1995) were chosen to allow a broad-brush description of incremental growth, enhancement and improvement of the methodology. For accuracy and ease of verification, the development and timeline is with respect to published articles (instead of working papers or oral tradition).
In the discussion that follows the reader is referred to the Evolution Map (Figure 1) as a pictorial guide.
2 In the beginning
Before DEA could flourish, the necessary foundations to support its growth and development had to be in place. Antecedents include works of Afriat (1972), Aigner and Chu (1968), Shephard (1970), Debreu (1951), and Farrell (1957)2, the conceptual definitions of Koopmans (1951) and Pareto (1927}, and the Linear Fractional transformation ofChames and Cooper (1962).
With these pieces in place, DEA actually started with Rhodes' dissertation topic-how to evaluate Program Follow-Through in U.S. education. The first published article describing the methodology and labeling the approach as Data Envelopment Analysis was Chames, Cooper, and Rhodes (1978). (Some researchers still recall the TIMS XXIV International Meeting in Hawaii in June 1979 where Chames and Cooper gave the first presentation on DEA.)
For several years I heard rumors of early DEA-type work from the mid-60s. Through the assistance of Knox Lovell, I finally obtained copies of four papers, Bressler ( 1966), Boles ( 1966), Seitz ( 1966), and Sitorus ( 1966), which appeared in the Proceedings of the 39th Annual Meeting of the Western Farm Economics Association. These four papers were presented in a single session at the meeting. The purpose of the session was to recall Farrell's approach to the attention of economists and statisticians. (It was
2 Contrary to numerous statements in the literature, Farrell did not employ LP in his 1957 paper. In fact, it was A. J. Hoffman, one of the discussants for the paper, who pointed out to Farrell (after the fact) that the problem he had described could be formulated and solved as an LP. In a later paper, Farrell and Fieldhouse (1962) provide the LP formulation for the single output case which Hoffman suggested.
4 Data Envelopment Analysis: Twenty Years Out
well received in 1957 but there had been virtually no further application of the method.) The four papers: (i) summarize Farrell's ideas; (ii) provide LP formulations and efficient computational procedures for a variety of problems in technical efficiency including the multiple output case; and give illustrative applications to (iii) steamelectric generating plants and (iv) aggregate census data, respectively. Although it appears that DEA was anticipated over a decade earlier, the technique again lay dormant until the ratio interpretation and introduction of the methodology to the ORIMS community by Charnes, Cooper, and Rhodes (1978). As we shall see, whether due to the timing, the packaging, or the proponents, this time the methodology flourished as evidenced by the rapid growth and widespread diffusion across disciplines.
3 Evolution of DEA
In our attempt to characterize the development of DEA over the past 20years we will examine the state-of-the-art (SOA) at several milestones in the evolution of the methodology For each we will contrast what was then known, what remained unknown, the state of DEA computation, range of application areas, and what was considered to be then-current SOA.
3.1 State-of-the-art circa 1980
In 1980, the SOA ofDEA was much simpler than it is today. Model choice was limited to the single constant-returns-to-scale model of Charnes, Cooper, and Rhodes (1978) which measured only technical efficiency. The few published applications were publicsector, not-for-profit and primarily in education (Bessent and Bessent, 1980; Banker, 1980; Charnes and Cooper, 1980; Charnes, Cooper and Rhodes, 1980; Schinnar, 1980). The focus was on relative efficiency of organizational units; other uses for the methodology were not yet discerned although a game-theoretic interpretation had been proposed by Banker (1980).
DEA computation in 1980 was extremely primitive. SOA DEA codes were based on a
naive implementation of epsilon as 10-6. (As shown in Ali and Seiford (1993), this can produce unreliable results.) The most significant break-through in advancing SOA applications at this time was the Program Follow-Through/ Non-Follow-Through evaluation that formed the basis of Ed Rhodes' dissertation and was published in Management Science in 1981.
Data Envelopment Analysis: Twenty Years Out 5
3.2 State-of-the-art circa 1985
By 1985, DEA theory was considerably more advanced. Model selection had expanded to encompass a wide range of models. The constant-returns-to-scale model of Charnes, Cooper, and Rhodes (1978) was joined by the variable-returns-to-scale model of Banker, Charnes, and Cooper (1984) for measuring scale efficiency, Multiplicative models for piecewise log-linear frontiers (Charnes et al., 1982, 1983), and the nonoriented Additive model (Charnes et al., 1985). A firm link to production theory was established through the theoretical characterizations of the inherent structure and capabilities of Pareto-Koopmans (empirical) frontier production functions given in Charnes et al. (1985).
The primary focus remained on relative efficiency but application areas now included hospitals (Bedard, 1985; Nunamaker, 1983; Sherman, 1981, 1984), post offices (Deprins et al., 1984), electric utilities (F\"are et al., 1983, 1985; Thomas, 1985), banking (Gold, 1982; Joseph et al., 1983; Sherman and Gold, 1985), mass transit (Kusbiantoro, 1985), courts (Lewin et al., 1982), agriculture (F\"are et a!., 1985), maintenance (Bowlin, 1984), mining (Byrnes et a!., 1984), pharmacies (Capettini, 1985), and USAF fighter wings (Charnes et al., 1985). Applications in education were now numerous due to the efforts of the Bessants' Educational Productivity Council at UT Austin (Bessent eta!., 1981, 1983, 1984, 1985; Blair, 1983; Garrett, 1985; Katims, 1985; Reaves, 1983; Splitek, 1981; Stone, 1984; Thorogood, 1983).
The perspective on DEA was widening; for example, issues of ownership versus efficiency were being examined (Byrnes, 1985; F\"are eta!., 1985) and DEA was also making inroads into marketing (Charnes et a!., 1985; Eechambadi, 1985). Links between DEA and basic production theory were established in Byrnes eta!. (1984) and F\"are et a!. (1985). The first of several comparisons of DEA with regression (Bowlin et a!., 1985) had appeared and the controversy or misunderstanding over the NonArchimedean (epsilon) had arisen (Boyd and F\"are, 1984; Charnes and Cooper, 1984)'. Researchers were beginning to look at stochastic issues (Sengupta, 1982).
However, DEA advances in this period were, for the most part, limited to models and theoretical enhancements. DEA computation had not yet progressed beyond the early stages. Examples of then-current SOA include most productive scale size (MPSS) (Banker, 1984), the Additive model and associated Pareto-Koopmans foundation for DEA (Charnes eta!., 1985), and window analysis (Charnes eta!., 1985).
3 The role of the NonArchimedean epsilon in detecting non-proportional inefficiencies (slacks) is much better understood today. See Ali and Seiford (1993) and Ali, Lerme, and Seiford (1995).
6 Data Envelopment Analysis: Twenty Years Out
3.3 State-of-the-art circa 1990
By 1990, DEA was becoming fully developed. Significant advances had been made on all fronts: models, extensions, computation, and practice. The UNC conference (1988) and the Austin DEA conference ( 1989) had contributed greatly to this development. Dialog and collaboration between researchers from Economics and Operations Research/Management Science had been initiated. The Journal of Econometrics Special Issue resulting from the UNC conference appeared with lead articles by Bauer (1990) and Seiford and Thrall (1990) providing dual perspectives.
Theoretical refinements and advances were numerous. Studies comparing the various DEA models (Ahn et al., 1988; Charnes et al., 1990; Epstein et al., 1989; F\"are et al., 1988; Seiford et al., 1990) provided a framework for understanding implicit assumptions and requirements. A Malmquist index (F\"are et al., 1989) had been developed to examine components of productivity growth, and technical, scale, and allocative efficiency (Banker and Maindiretta, 1988; Morey et al., 1990; RetzlaffRoberts, 1990) had been compared and contrasted. Non-convex models had been introduced by Petersen (1990) which enlarged the perspective on basic assumptions of DEA. A number of significant model extensions had been developed including: the capability to handle nondiscretionary variables and/or categorical variables (Banker and Morey, 1986); the ability to incorporate judgement (restricting multipliers (Dyson and Thanassoulis, 1988; Wong and Beasley, 1990), the Cone Ratio model (Charnes et al., 1989, 1990), and Assurance Regions (Thompson et al., 1986, 1990)); and model ordinal relationships (Golany, 1988). Connections were being established with the field of decision analysis via DEA-inspired consensus ranking approaches (Cook et al., 1990) and game theoretic interpretations (Banker et al., 1989, Charnes et al., 1989, 1990; Clarke, 1988). Finally, sensitivity and stability studies (Charnes et al., 1989, 1990) and translation invariance (Ali and Seiford, 1990) round out the theoretical contributions.
Computational issues had surfaced and been addressed (Ali, 1990). The nonArchimedean models had been correctly implemented in a DEA code (e.g., IDEAS, 1989) as a two-stage preemptive procedure. Other computational requirements, specific to DEA, had been recognized and production-quality DEA codes were available that employed specialized pricing rules, anticycling techniques to address degeneracy, and eschewed sparse matrix techniques.
Published applications from this period addressed more complex issues involving property tax valuation (Adolphson et al., 1987, 1989), software development (Banker et al., 1987, 1989), institutions of higher learning (Ahn, 1987; Ahn et al., 1987, 1988, 1989), university departments (Beasley, 1990; Tomkins, 1988), energy use (Baxter et al., 1986), DRG reimbursement (Borden, 1986, 1988), site selection (Bowen, 1990; Desai et al., 1990), spatial efficiency (Desai and Storbeck, 1990), farming (Byrnes et
Data Envelopment Analysis: Twenty Years Out 7
a!., 1987), unions (Byrnes et a!., 1988), sports (Camm, 1988), electric cooperatives (Chames et a!., 1989), Chinese cities (Chames et a!., 1989), individual physicians (Chilingerian, 1989, 1990), highway maintenance (Cook eta!., 1988, 1990), regulatory environments (F\"are et a!., 1986, 1989), organizational slack (Golden, 1989), airlines (Johnston, 1990), logistics systems (Kleinsorge eta!., 1989), parks (Rhodes, 1986), pubs (Sant, 1989), construction (Shash, 1988), telecommunications (Majumdar, 1990), and US Army recruiting (Thomas, 1990).
Examples of SOA include relaxed non-convex assumptions (Petersen, 1990), various approaches to placing restriction on the possible range of multipliers for incorporating judgement or managerial preference (see earlier references), a more balanced perspective on DEA ( Epstein et a!., 1989; Stolp, 1990), and connections with Econometrics (Varian, 1990).
3.4 State-of-the-art circa 1995
In the recent years (1990-1995) there have been significant theoretical advances but it has been DEA practice that has evolved the most extensively. DEA is now recognized as a versatile and effective tool for data analysis and is often used as an exploratory technique (E-DEA) for "visualizing" the data. Applications are frequently large-scale requiring significant computational power, and a flexible user interface for data management and model management has become as important as a robust and accurate optimizer. Parallel processing environments and new solution approach can dramatically reduce solution times as reported in Barr and Durcholz. (1997).
The applications continue to become more sophisticated and recent studies have focused on revenue transfers (local aid) (Ali et a!., 1993), TQM (Bailey, 1993), benchmarking and identification of best-practice (Collier and Storbeck, 1993; Chilingerian, 1995; Golany and Thore, 1997), forecasting bank failures (Barr et a!., 1993, 1994; Siems, 1991 ), strategy (Day et a!., 1994, 1995; Ali and Lerme, 1997), pollution (Haynes et a!., 1994), improved performance indexes in sports (Anderson, 1997), and X-efficiency (Bohnet and Beck, 1990; Frantz, 1992; Leibenstein and Maital, 1992). At the same time DEA has moved from being an esoteric research methodology to a more mainstream analytical tool as evidenced by (i) its inclusion in an introductory MBA OR textbook (Anderson, Sweeney, and Williams, 1991), and (ii) its being featured in Fortune magazine (Norton, 1994).
Significant theoretical contributions from this period would include the free disposal hull (FDH) (non-convex) model of Tulkens (1993) as well as work on the statistical foundations of DEA (Banker, 1993; Simar, 1992) and the chance-constrained framework introduced in Land eta!. (1993). Examples of current SOA would be the graphical backend for visualization of DEA results developed by Paradi et a!. at the
8 Data Envelopment Analysis: Twenty Years Out
University of Toronto, large-scale benchmarking studies for the Federal Reserve Bank and USAF medical treatment facilities by Barr, Seiford and others, cost containment in HMOs (Chilingerian and Sherman, 1997), and cone analysis to reveal preferred practice regions (Ward eta!., 1997).
3.5 Current view of State-of-the-art in DEA
The previous sections have described the development and migration of DEA. We tum now to the question: what is the current state-of-the-art? SOA DEA models would include CCR, BCC, Additive, and FDH. The interrelationships between these models are best explained within the framework given in Ali, Lerme, & Seiford (1995) and the interested reader is referred there.
Any of the convex models can be combined with various theoretical extensions (e.g., nondiscretionary variables, categorical variables, ordinal relationships, etc.) and multiplier restrictions can be introduced to incorporate judgement. For cross sectional/time series/panel data one could employ window analysis or a Malmquist index to examine changes across time periods.
However, some modeling issues are still not easily handled. Negative inputs or outputs cause difficulty; ranking units can be problematic. With large scale application becoming more frequent, computational issues become even more important. Stochastic issues are being addressed and statistical tools are being developed but much work remains to be done.
3.6 Future issues in DEA
It's usually the case that new methodologies and generalizations give rise to new possibilities and new questions and DEA is certainly no exception. DEA has become an important and widespread analytical tool. With increasing use in real-world largescale complex applications, the need for validation of these studies becomes more critical.
Software has become increasingly important for the large-scale and complex DEA studies now being conducted. This dependence on software raises additional issues. How can one insure the availability of robust accurate DEA software? How should one validate DEA codes as producing accurate results? If validation should be on a suite of test problems, what types and how wide a range of conditions are sufficient? This validation issue remains important whether one is using specialized DEA software or a standard LP package, e.g., SAS, GAMS, LINDO, etc., to perform the analysis.
Data Envelopment Analysis: Twenty Years Out 9
Along with the issue of accurate results, there is a need to develop more effective means for presentation of the results. What visualization/presentation formats should be implemented for viewing solution results? These graphical back-ends would be used both for E-DEA (exploratory DEA) and for management presentations. The visualization capability becomes critical for large-scale problems particularly if one is searching for patterns or explanations of inefficiency across units and subunits.
The final topic, Stochastic DEA, should come as no surprise. It appears on almost everyone's list of Future Research Areas for DEA and presents a formidable challenge. The essential problem is noise (e.g., measurement error) in the underlying data. Promising approaches and partial solutions are given in Banker (1993), Simar (1992), Korostelev eta!. (forthcoming), Land et a!. (1993), and Olesen and Petersen (1995). However, the development of Stochastic DEA which can incorporate measurement error and other sources of noise that inevitably contaminate the data used in an analysis is far from complete. Lovell (1994) eloquently states the importance of this problem. "Until a stochastic DEA is developed, statisticians and econometricians will remain skeptical of the managerial and policy implications drawn from DEA."
In my opinion, it is this last topic, Stochastic DEA, which is the most critical and the most difficult future issue in DEA. However, I'm confident that when we reach the next progress milestone, the year 2000, it will be listed among the SOA accomplishments of the prior five years.
References
Adolphson, Donald L., Gary C. Cornia, and Lawrence C. Walters (1989): Railroad Property
Valuation Using Data Envelopment Analysis, Interfaces 19, no. 3, p. 18-26.
Adolphson, Donald L., Gary C. Cornia, and Lawrence C. Walters (1987): The Relative Efficiency of
Railroads and Obsolescence, Proceedings of the Seventeenth Annual Program on the Appraisal
of Utilities and Railroad Property for Ad Valorem Taxation, 97-130. Wichita, KS: Wichita
State University.
Afriat, S. N. (1972): Efficiency Estimation of Production Functions, International Economic review
13, no. 3, p. 568-598.
Ahn, Tae Sik. (1987): Efficiency and Related Issues in Higher Education: A Data Envelopment
Analysis Approach, Ph. D. dissertation, Graduate School of Business, University of Texas.
Ahn, Tae Sik, V. Arnold, A. Charnes, and W. W. Cooper (1989): DEA and Ratio Efficiency Analyses
for Public Institutions of Higher Learning in Texas, in: Research in Governmental and
Nonprofit Accounting, 165-185. Editors James L. Chan, and James M. Patton. Greenwich, CT:
JAI Press.
10 Data Envelopment Analysis: Twenty Years Out
Ahn, Tae Sik, A. Charnes, and W. W. Cooper. (1988): Efficiency Characterizations in Different DEA Models, Socio-Economic Planning Sciences 22, no. 6, p. 253-257.
Ahn, Tae Sik, and Lawrence M. Seiford (1993): Sensitivity ofDEA to Models and Variable Sets in a Hypothesis Test Setting: The Efficiency of University Operations, in: Creative and Innovative Approaches to the Science of Management, 191-208. editor Yuji Ijiri. New York: Quorum Books.
Aigner, D. J., and S. -F Chu {1968): On Estimating the Industry Production Function, American Economic Review 58, no. 4, p. 826-839.
Ali, Agha Iqbal {1990): Data envelopment analysis: computational issues, Computers, Environment and Urban Systems 14, no. 2, p. 157-165.
Ali, Agha Iqbal, and Catherine S. Lerme {1997): Comparative Advantage and Disadvantage in DEA, Annals of Operations Research (73 ).
Ali, Agha Iqbal, Catherine S. Lerme, and Robert A. Nakosteen {1993): Assessment of Intergovernmental Revenue Transfers, Socio-Economic Planning Sciences 27, no. 2, p. 109-118.
Ali, Agha Iqbal, Catherine S. Lerme, and Lawrence M. Seiford (1995): Components of Efficiency Evaluation in Data Envelopment Analysis, European Journal of Operational Research 80, no. 3, p. 462-473.
Ali, Agha Iqbal, and Lawrence M. Seiford (November 1993): Computational Accuracy and Infinitesimals in Data Envelopment Analysis, INFOR 31, no. 4, p. 290-297.
Ali, Agha Iqbal, and Lawrence M. Seiford (November 1990): Translation Invariance in Data Envelopment Analysis, Operations Research Letters 9, no. 6, p. 403-405.
Anderson, David R., Dennis J. Sweeney, and Thomas A. Williams (1991): Linear Programming Applications: Data Envelopment Analysis, in: An Introduction to Management Science: Quantitative Approaches to Decision Making, p. 147-152. Sixth ed. St Paul, MN: West Publishing Company.
Anderson, Timothy R., and Gunter P. Sharp {1997): A New Measure of Baseball Batters Using DEA, Annals of Operations Research (73).
Bailey, Marshall Hamilton, III. (1993): Public Administration Efficiency Through Total Quality Management, Ph. D. dissertation, George Mason University.
Banker, Rajiv D. {1980): A Game Theoretic Approach to Measuring Efficiency, European Journal of Operational Research 5, p. 262-266.
Banker, Rajiv D. (July 1984): Estimating Most Productive Scale Size Using Data Envelopment Analysis, European Journal of Operational Research 17, no. I, p. 35-44.
Banker, Rajiv D. (October 1993): Maximum Likelihood, Consistency and Data Envelopment Analysis: A Statistical Foundation, Management Science 39, no. 10, p. 1265-1273.
Data Envelopment Analysis: Twenty Years Out 11
Banker, Rajiv D. (1980): Studies in Cost Allocation and Efficiency Evaluation, D. B. A. dissertation, Graduate School of Business Administration, Harvard University.
Banker, Rajiv D., A. Chames, Richard L. Clarke, and W. W. Cooper (1989): Erratum: Constrained game formulations and interpretations for data envelopment analysis, European Journal of Operational Research 42.
Banker, Rajiv D., A. Chames, and W. W. Cooper (1984): Some Models for Estimating Technical and Scale Inefficiencies in Data Envelopment Analysis, Management Science 30, no. 9, p. 1078-1092.
Banker, Rajiv D., A. Chames, W. W. Cooper, and Richard L. Clarke (June 1989): Constrained Game Formulations and Interpretations for Data Envelopment Analysis, European Journal of Operational Research 40, no. 3, p. 299-308.
Banker, Rajiv D., A. Chames, W. W. Cooper, and Ajay Maindiratta (1988): A Comparison of DEA and Translog Estimates of Production Frontiers Using Simulated Observations From a Known Technology, in: Applications of Modem Production Theory: Efficiency and Productivity, editors Ali Dogramaci, and RolfHre. Boston: Kluwer Academic Publishers.
Banker, Rajiv D., A. Chames, W. W. Cooper, and A. P. Schinnar (1981): A Bi-Extremal Principle for Frontier Estimation and Efficiency Evaluations, Management Science 27, no. 12, p. 1370-1382.
Banker, Rajiv D., R. F. Conrad, and R. P. Strauss (January 1986): A Comparative Application of Data Envelopment Analysis and Translog Methods: An Illustrative Study of Hospital Production, Management Science 32, no. I, p. 30-44.
Banker, Rajiv D., and Holly Hanson Johnston (1994): Evaluating the Impacts of Operating Strategies on Efficiency in the U.S. Airline Industry, in: Data Envelopment Analysis: Theory, Methodology, and Applications, editors A. Chames, W. W. Cooper, Arie Y. Lewin, and Lawrence M. Seiford. Boston: Kluwer Academic Publishers.
Banker, Rajiv D., and Chris F. Kemerer (October 1989): Scale Economies in New Software Development, IEEE Transactions on Software Engineering 15, no. 10, p. 1199-1205.
Banker, Rajiv D., and Ajay Maindiratta (1988): Nonparametic Analysis of Technical and Alloctive Efficiencies in Production, Econometrica 56, no. 6, p. 1315-1332.
Banker, Rajiv D., and Ajay Maindiratta (1986): Piecewise Loglinear Estimation of Efficient Production Surfaces, Management Science 32, no. I, p. 126-135.
Banker, Rajiv D., and Richard C. Morey (July 1986): Efficiency Analysis for Exogenously Fixed Inputs and Outputs, Operations Research 34, no. 4, p. 513-521.
Banker, Rajiv D., and Richard C. Morey (December 1986): The Use of Categorical Variables in Data Envelopment Analysis, Management Science 32, no. 12, p. 1613-1627.
Banker, Rajiv D., and Robert M. Thrall (9 October 1992): Estimation of Returns to Scale Using Data Envelopment Analysis, European Journal of Operational Research 62, no. I.
12 Data Envelopment Analysis: Twenty Years Out
Banker, Rajiv D., Srikant M. Datar, and Chris F. Kemerer (December 1987): Factors Affecting Software Maintenance Productivity: An Exploratory Study, in: Proceedings of the 8th. International Conference on Information Systems, p. 160-175. Pittsburgh.
Barr, Richard S., and Matthew L. Durcholz (1997): Parallel and Hierarchical Decomposition Approaches for Solving Large-Scale Data Envelopment Analysis Models, Annals of Operations Research (73).
Barr, Richard S., Lawrence M. Seiford, and Thomas F. Siems (1993): An Envelopment-Analysis Approach to Measuring the Managerial Efficiency of Banks, Annals of Operations Research 45,p.l-19.
Barr, Richard S., Lawrence M. Seiford, and Thomas F. Siems (1994): Forcasting Bank Failure: A Non-Parametric Frontier Estimation Approach, Recherches Economiques de Louvain 60, no. 4, p. 417-429.
Bauer, Paul W. (1990): Recent developments in the econometric estimation of frontiers, Journal of Econometrics 46, no. y,, p. 39-56.
Baxter, Lester W., Stephen L. Feldman, Arie P. Schinnar, and Robert M. Wirtshafter (April 1986): An Efficiency Analysis of Household Energy Use, Energy Economics, p. 62-73.
Beasley, J. E. (1990): Comparing University Departments, Omega 18, no. 2.
Bedard, Jean Catherine (1985): Use of Data Envelopment Analysis in Accounting Applications: Evaluation and Illustration by Prospective Hospital Reimbursement, Ph. D. dissertation, Graduate School of Business, University of Wisconsin.
Bessent, Authella M., and E. Wailand Bessent (December 1981): A Fractional Programming Model for Determining the Efficiency of Decision Making Units, ERIC Clearinghouse on Educational Management, University of Oregon, Eugene, Oregon.
Bessent, Authella M., and E. Wailand Bessent. ,Determining the Comparative Efficiency of Schools Through Data Envelopment Analysis." Educational Administration Quarterly 16, no. 2 (March 1980): 57-75.
Bessent, Authella M., E. Wailand Bessent, A. Charnes, W. W. Cooper, and N. Thorogood (March 1983): Evaluation of Educational Program Proposals by Means of Data Envelopment Analysis, Educational Administration Quarterly 19, no. 2, p. 82-107.
Bessent, Authella M., E. Wailand Bessent, Joyce Elam, and D. Long (1984): Educational Productivity Council Employs Management Science Methods to Improve Educational Quality, Interfaces 14, no. 6, p. 1-8.
Blair, Larry Delwood (1983): A Comparative Analysis of the Financial Practices of School Districts Selected by Data Envelopment Analysis Efficiency Indices, Ph. D. dissertation, College of Education, University of Texas.
Bohnet, A., and M. Beck (1990): The Impact of the Income Tax on Work Effort and X-inefficiency in Enterprises, in: Studies in Economic Rationality: X-Efficiency Examined and Extolled. Essays written in the tradition of and to honor Harvey Leibenstein, 227-251. editors K. Weiermair, and M. Perlman. Ann Arbor: 1990.
Data Envelopment Analysis: Twenty Years Out 13
Boles, James N. (1966): Efficiency Squared - Efficient Computation of Efficiency Indexes, Proceedings of the Thirty Ninth Annual Meeting of the Western Farm Economics Association, p. 137-142.
Borden, James Patrick (1986): An Assessment of the Impact of Diagnosis Related Group (DRG)based Reimbursement on the Technical Efficiency of New Jersey Hospitals, Ph. D. dissertation, Drexel University.
Borden, James Patrick (June 1988): An Assessment of the Impact of Diagnosis-Related Group (DRG)-Based Reimbursement on the Technical Efficiency of New Jersey Hospitals Using Data Envelopment Analysis, Journal of Accounting and Public Policy 7, no. 2, p. 77-96.
Bowen, William M. (1990): The Nuclear Waste Site Selection Decision--a Comparison of Two Decision-Aiding Models, Ph. D. dissertation, Indiana University.
Bowlin, William Frank (1984): A Data Envelopment Analysis Approach to Performance Evaluation in Not-for-profit Entities with an Illustrative Application to the U.S. Air Force, Ph. D. dissertation, Graduate School of Business, University of Texas.
Bowlin, William F., A. Charnes, W. W. Cooper, and H. David Sherman (1985): Data Envelopment Analysis and Regression Approaches to Efficiency Estimation and Evaluation, Annals of Operations Research 2, no. I, p. 113-138.
Boyd, G., and Rolf Fare (1984): Measuring the Efficiency of Decision Making Units: A Comment, European Journal of Operational Research 15, p. 331-332.
Bressler, R. G. (1966): The Measurement of Productivity Efficiency, Proceedings of the Thirty Ninth Annual Meeting of the Western Farm Economics Association, p. 129-136.
Byrnes, P. (1985): Ownership and Efficiency in the Water Supply Industry: An Application of the Nonparametric Programming Approach to Efficiency Measurement, Ph. D. dissertation, Southern Illinois University.
Byrnes, P., Rolf Fare, and S. Grosskopf ( 1984): Measuring Productive Efficiency: An Application to Illinois Strip Mines, Management Science 30, no. 6, p. 671-681.
Byrnes, P., Rolf Fare, S. Grosskopf, and C. A. Knox Lovell (1988): The Effect of Unions on Productivity: U.S. Surface Mining of Coal, Management Science 34, no. 9, p. 1037-1053.
Byrnes, P., Rolf Fare, S. Grosskopf, and S. Kraft (1987): Technical Efficiency and Size: The Case of 1980 Illinois Grain Farms, European Review of Agricultural Economics 14, no. 4, p. 367-381.
Camm, J.D., and T. J. Grogan (November 1988): An Application of Frontier Analysis: Handicapping Running Races, Interfaces 18, no. 6, p. 52-60.
Capettini, Robert, David A. Dittman, and Richard C. Morey (June 1985): Reimbursement Rate Setting for Medicaid Prescription Drugs Based on Relative Efficiencies, Journal of Accounting and Public Policy 4, no. 2, p. 83-110.
Charnes, A., Charles T. Clark, W. W. Cooper, and Boaz Golany (1985): A Developmental Study of Data Envelopment Analysis in Measuring the Efficiency of Maintenance Units in the U.S. Air
14 Data Envelopment Analysis: Twenty Years Out
Forces, in: Annals of Operation Research, 95-112. editors Russell G. Thompson, and Robert M.
Thrall.
Chames, A., Richard L. Clarke, and W. W. Cooper (1989): An Approach to Testing for Organizational Slack with R. Banker's Game Theoretic Formulation of DEA, in: Research in Governmental and Nonprofit Accounting, 211-230. editors James L. Chan, and James M.
Patton. Greenwich, CT: JAI Press.
Chames, A., and W. W. Cooper (1980): Auditing and Accounting for Program Efficiency and Management Efficiency in Not-For-Profit Entities, Accounting, Organizations and Society 5,
no. I, p. 87-107.
Chames, A., and W. W. Cooper (1962): Programming With Linear Fractional Functionals, Naval
Research Logistics Quarterly 9, p. 181-186.
Chames, A., and W. W. Cooper (1984): The Non-Archimedean CCR Ratio for Efficiency Analysis: A Rejoinder to Boyd and Fiire, European Journal of Operational Research 15, no. 3, p. 333-334.
Chames, A., W. W. Cooper, D. Divine, T. W. Ruefli, and D. Thomas (1989): Comparisons of DEA and Existing Ratio and Regression Systems for Effecting Efficiency Evaluations of Regulated
Electric Cooperatives in Texas, in: Research in Governmental and Nonprofit Accounting, p. 187-210. editors James L. Chan, and James M. Patton. Greenwich, CT: JAI Press.
Chames, A., W. W. Cooper, Boaz Golany, Lawrence M. Seiford, and J. Stutz (October 1985):
Foundations of Data Envelopment Analysis for Pareto-Koopmans Efficient Empirical Production Functions, Journal of Econometrics 30, no. 1/2, p. 91-107.
Chames, A., W. W. Cooper, David B. Leamer, and Fred Y. Phillips (March 1985): Management Science and Marketing Management, Journal of Marketing 49, no. 3, p. 93-105.
Chames, A., W. W. Cooper, Arie Y. Lewin, Richard C. Morey, and John J. Rousseau (1985): Sensitivity and Stability Analysis in DEA, Annals of Operations Research 2, p. 139-156.
Chames, A., W. W. Cooper, Arie Y. Lewin, and Lawrence M. Seiford (1994): Data Envelopment
Analysis: Theory, Methodology, and Applications, Boston: Kluwer Academic Publishers.
Chames, A., W. W. Cooper, and S. Li (1989): Using DEA to Evaluate Relative Efficiencies in the Economic Performance of Chinese Cities, Socio-Economic Planning Sciences 23, no. 6, p. 325-344.
Chames, A., W. W. Cooper, and Edwardo L. Rhodes (1980): An Efficiency Opening for Managerial Accounting in Not-For-Profit Entities, in: Management Accounting 1980: Proceedings of the
University of Illinois Management Accounting Symposium, editor H. P. Holzer, 21-4 7. University of Illinois, Urbana, Illinois.
Chames, A., W. W. Cooper, and Edwardo L. Rhodes (June 1981): Evaluating Program and Managerial Efficiency: An Application of Data Envelopment Analysis to Program Follow Through, Management Science 27, no. 6, p. 668-697.
Chames, A., W. W. Cooper, and Edwardo L. Rhodes (1978): Measuring the Efficiency of Decision
Making Units, European Journal of Operational Research 2, no. 6, p. 429-444.
Data Envelopment Analysis: Twenty Years Out 15
Charnes, A., W. W. Cooper, Lawrence M. Seiford, and J. Stutz (1982): A Multiplicative Model for Efficiency Analysis, Socio-Economic Planning Sciences 16, no. 5, p. 223-224.
Charnes, A., W. W. Cooper, Lawrence M. Seiford, and J. Stutz (1983): Invariant Multiplicative Efficiency and Piecewise Cobb-Douglas Envelopments, Operations Research Letters 2, no. 3, p. 101-103.
Charnes, A., W. W. Cooper, Quan Ling Wei, and Z. M. Huang (1989): Cone Ratio Data Envelopment Analysis and Multi-objective Programming, International Journal of Systems Science 20, no. 7, p. 1099-1118.
Charnes, A., W. W. Cooper, Quan Ling Wei, and Z. M. Huang (1990): Fundamental Theorems of Nondominated Solutions Assocaited with Cones in Normed Linear Spaces, Journal of Mathematical Analysis and Applications 150, no. I, p. 54-78.
Charnes, A., Stephen Haag, Patrick V. Jaska, and John Semple (1992): Sensitivity of efficiency classifications in the additive model of data envelopment analysis, International Journal of Systems Science 23, no. 5, p. 789-798.
Charnes, A., Z. M. Huang, John J. Rousseau, and Quan Ling Wei (1990): Cone Extremal Solutions of Multi-Payoff Games with Cross-Constrained Strategy Sets, Optimization 21, no. I, p. 51-69.
Chames, A., Z. M. Huang, John Semple, T. Song, and D. Thomas (1990): Origins and Research in Data Envelopment Analysis, The Arabian Journal for Science and Engineering 15, no. 4B, p. 617-625.
Chames, A., and L. Neralic (1989): Sensitivity Analysis in Data Envelopment Analysis-Part I,
Glasnik Matemativcki. Serija III 24 (44), no. I, p. 211-226.
Charnes, A., and L. Neralic ( 1989): Sensitivity Analysis in Data Envelopment Analysis-Part 2,
Glasnik Mathemativcki. Serija III 24 (44), no. 2/3, p. 449-463.
Charnes, A., and L. Neralic (1990): Sensitivity Analysis of the Additive Model in Data Envelopment
Analysis, European Journal of Operational Research 48, no. 3, p. 332-341.
Chilingerian, Jon A. (1995): Evaluating Physician Efficiency in Hospitals: A Multivariate Analysis of Best Practice, European Journal of Operational Research 80, no. 3, p. 548-574.
Chilingerian, Jon A. (1989): Investigating Non-Medical Factors Associated with the Technical Efficiency of Physicians in the Provision of Hospital Services: A Pilot Study, Annual Best Paper Proceedings of the Academy of Management, p. 85-89.
Chilingerian, Jon A., and David H. Sherman (1997): DEA and Physician Report Cards: Using Assurance Regions to Benchmark Clinical Best Practices in an HMO, Annals of Operations Research (73).
Chilingerian, Jon A., and H. David Sherman (1990): Managing Physician Efficiency and Effectiveness in Providing Hospital Services, Health Serv Manage Res 3, no. I, p. 3-15.
16 Data Envelopment Analysis: Twenty Years Out
Clarke, Richard Lee (1988): Effects of Repeated Applications of Data Envelopment Analysis on Efficiency of Air Force Vehicle Maintenance Units in the Tactical Air Command and a Test for the Presence of Organizational Slack Using Rajiv Banker's Game Theory Formulations, Ph. D. dissertation, Graduate School of Business, University of Texas.
Collier, David, and James Storbeck (1993): Monitoring of Continuous Improvement Performance Using Data Envelopment Analysis, Proceedings of Decision Sciences Institute, p. 1925-1927.
Cook, Wade D., and Moshe Kress (1990): A Data Envelopment Model for Aggregating Preference Rankings, Management Science 36, no. II, p. 1302-1310.
Cook, Wade D., and Moshe Kress (1990): A m-th Generation Model for Weak Ranking of Players in a Tournament, Journal of the Operational Research Society 41, no. 12, p. 1111-1119.
Cook, Wade D., Moshe Kress, and Lawrence M. Seiford (1993): On the Use of Ordinal Data in Data Envelopment Analysis, Journal of the Operational Research Society 44, no. 2, p. 133-140.
Cook, Wade D., Yaakov Roll, and Alex Kazakov (1990): A DEA Model for Measuring the Relative Efficiency of Highway Maintenance Patrols, INFOR 28, no. 2, p. 113-124.
Cook, Wade D., Yaakov Roll, and Alex Kazakov (October 1988): Measurement of the Relative Efficiency of Highway Maintenance Patrols in Ontario, Ministry of Transportation, Toronto, Ontario, Canada.
Day, D. L., A. Y. Lewin, and H. Li (1995): Strategic Leaders or Strategic Groups: A Longitudinal Data Envelopment Analysis of the U. S. Brewing Industry, European Journal of Operational Research 80, no. 3, p. 619-638.
Day, D. L., Arie Y. Lewin, R. J. Salazar, and Hongyu Li (1994): Strategic Leaders in the U.S. Brewing Industry: A Longitudinal Analysis of Outliers, in: Data Envelopment Analysis: Theory, Methodology, and Applications, editors A. Charnes, W. W. Cooper, Arie Y. Lewin, and Lawrence M. Seiford. Boston: Kluwer Academic Publishers.
Debreu, G. (1951): The Coefficient of Resource Utilization, Econometrica 19, no. 3, p. 273-292.
Deprins, D., L. Simar, and H. Tulkens (1984): Measuring Labor-Efficiency in Post Offices, in: The Performance of Public Enterprises: Concepts and Measurement, p. 243-267. editors M. Marchand, P. Pestieau, and H. Tulkens. North-Holland: Elsevier Science Publishers B.V ..
Desai, Anand, and James E. Storbeck (1990): A Data Envelopment Analysis for Spatial Efficiency, Computers, Environment and Urban Systems 14, no. 2, p. 145-156.
Desai, Anand, James E. Storbeck, Kingsley E. Haynes, Homee F. E. Shroff, and Yan Xiao (1990): Extending Multiple Objective Programming for Siting Decision Sensitivity, Modeling and Simulation 22, p. 153-158.
Dyson, R. G., and E. Thanassoulis (1988): Reducing Weight Flexibility in Data Envelopment Analysis, Journal of the Operational Research Society 39, no. 6, p. 563-576.
Eechambadi, Narasimhan Varadarajan (1985): Efficiency Analysis of Market Response and the Marketing Mix: Extending Data Envelopment Analysis to a Competitive Environment, Ph. D. dissertation, Graduate School of Business, University of Texas.
Data Envelopment Analysis: Twenty Years Out 17
Epstein, Michael K., and John C. Henderson (1989): Data Envelopment Analysis for Managerial Control and Diagnosis, Decision Sciences 20, no. I, p. 90-119.
Fare, Rolf, R. Grabowski, and S. Grosskopf (1985): Technical Efficiency of Philippine Agriculture, Applied Economics 17, p. 205-214.
Fare, Rolf, S. Grosskopf, B. Lindgren, and P. Roos (1994): Productivity Developments in Swedish Hospitals: A Malmquist Output Index Approach, in: Data Envelopment Analysis: Theory, Methodology, and Applications, editors A. Charnes, W. W. Cooper, Arie Y. Lewin, and Lawrence M. Seiford. Boston: Kluwer Academic Publishers.
Fare, Rolf, S. Grosskopf, and James Logan (1983): The Relative Efficiency of Illinois Electric Utilities, Resources and Energy 5, no. 4, p. 349-367.
Fare, Rolf, S. Grosskopf, and James Logan (1985): The Relative Performance of Publically-Owned and Privately-Owned Electric Utilities, Journal of Public Economics 26, p. 89-106.
Fare, Rolf, S. Grosskopf, and C. A. K. Lovell (1994): Production Frontiers, London: Cambridge University Press.
Fare, Rolf, S. Grosskopf, and C. A. Knox Lovell (1985): The Measurement of Efficiency of Production, Boston: Kluwer-NijhoffPublishing, Kluwer Academic Publishers.
Fare, Rolf, S. Grosskopf, and D. Njinkeu (1988): On Piecewise Reference Technologies, Management Science 34, no. 12, p. 1507-1511.
Farrell, M. J. (1957): The Measurement of Productive Efficiency, Journal of the Royal Statistical Society, Series A 120, no. Part 3, p. 253-290.
Farrell, M. J., and M. Fieldhouse (1962): Estimating Efficient Production Frontiers Under Increasing Returns to Scale, Journal of the Royal Statistical Society, Series A Part II, p. 252-267.
Frantz, Roger (May 1992): X-Efficiency and Allocative Efficiency: What Have We Learned? American Economic Review 82, no. 2, p. 434-438.
Fried, H., C. A. Knox Lovell, and S. Schmidt (editors), (1993): The Measurement of Productive Efficiency: Techniques and Applications, London: Oxford University Press.
Garrett, Allan Warren (1985): Constrained Facet Analysis and Related Linear Programming Models: Tools for the Evaluation of the Efficiency, Productivity and Effectiveness of School Classrooms, Ph. D. dissertation, College of Education, University of Texas.
Golany, Boaz (1988): A Note on Including Ordinal Relations Among Multipliers in Data Envelopment Analysis, Management Science 34, no. 8, p. 1029-1033.
Golany, Boaz, and Sten Thore (1997): Restricted Best Practice Selection in DEA: An Overview with a Case Study Evaluating the Socio-Economic Performance of Nations, Annals of Operations Research (73).
Gold, Franklin Harold (1982): Data Envelopment Analysis: An Application to a Savings and Loan Association, M.S. thesis, Alfred P. Sloan School of Management, M.I.T..
18 Data Envelopment Analysis: Twenty Years Out
Golden, Peggy A. (1989): Measuring Organizational Slack and an Application of the Slack Construct to the Prediction of Merger And Acquisition, D. B. A. dissertation, University of Kentucky.
Haynes, Kingsley E., Samuel Ratick, and James Cummings-Saxton (1994): Toward a Pollution Abatement Monitoring Policy: Measurements, Model Mechanics, and Data Requirements, The Environmental Professional 16, p. 292-303.
Johnston, Holly Hanson (1990): Empirical Studies in Management Accounting: Three Essays on the U.S. Airline Industry, 1981-1985, Ph. D. dissertation, Carnegie-Mellon University.
Joseph, D. A., R. P. Cerveny, and Edwardo L. Rhodes (September 1983): Application Architecture and Technologic Efficiency : A Comparative Study of Computerized Bank Transaction Systems, Proceedings of the AIDS National Meeting.
Katims, Michael Allen (1985): Using Efficiency Analysis to Evaluate Program Effects of Educational Intervention, Ph. D. dissertation, College of Education, University of Texas.
Kleinsorge, Ilene K., Phillip B. Schary, and Ray D. Tanner (1989): Evaluating Logistics Decisions, International Journal of Physical Distribution and Materials Management 19, no. 12.
Koopmans, T. C. (1951): Analysis of Production as an Efficient Combination of Activities, in: Activity Analysis of Production and Allocation, editor T. C. Koopmans. New York: Wiley.
Korostelev, A. P., L. Simar, and A. B. Tsybakov (1995): On Estimation of Monotone and Convex Boundaries, Pub. Inst. Stat. Univ. Paris 34, no. I, p. 3-18.
Kusbiantoro (1985): A Study of Urban Mass Transit Performance: Concept, Measurement, and Explanation, Ph. D. dissertation, University of Pennsylvania.
Land, Kenneth C., C. A. Knox Lovell, and Sten Thore (1993): Chance-constrained Data Envelopment Analysis, Managerial and Decision Economics 14, no. 6, p. 541-554.
Leibenstein, Harvey, and Shlomo Maital (May 1992): Empirical Estimation and Partitioning of XInefficiency: A Data-Envelopment Approach, American Economic Review 82, no. 2, p. 428-434.
Lewin, Arie Y., Richard C. Morey, and T. J. Cook (1982): Evaluating the Administrative Efficiency of Courts, Omega 10, no. 4, p. 401-411.
Lovell, C. A. K. (1994): Linear Programming Approaches to the Measurement and Analysis of Productive Efficiency, TOP 2, no. 2, p. 175-248.
Lovell, C. A. Knox, Lawrence C. Walters, and L. L. Wood (1994): Stratified Models of Education Production Using Modified DEA and Regression Analysis, in: Data Envelopment Analysis: Theory, Methodology, and Applications, editors A. Charnes, W. W. Cooper, Arie Y. Lewin, and Lawrence M. Seiford. Boston: Kluwer Academic Publishers.
Majumdar, Sumit Kumar (1990): The Impact of a Competitive Environment on Corporate Performance in U.S. Telecommunications, Ph. D. dissertation, University of Minnesota.
Morey, Richard C., D. J. Fine, and S. W. Loree (1990): Comparing the Allocative Efficiencies of Hospitals, Omega 18, no. I, p. 71-83.
Data Envelopment Analysis: Twenty Years Out 19
Norman, Michael, and Barry Stoker (1991): Data Envelopment Analysis: The Assessment of performance. Chichester, England: John Wiley.
Norton, Rob (October 1994): Economics for Managers: Which Offices or Stores Really Perform Best? A New Tool Tells, Fortune 38.
Nunamaker, Thomas R. (June 1983): Measuring Routine Nursing Service Efficiency: A Comparison of Cost Per Patient Day and Data Envelopment Analysis Models, Health Services Research 18, no. 2, part I, p. 183-205.
Olesen, 0., and N. C. Petersen (March 1995): Chance-Constrained Efficiency Evaluation, Management Science 41, no. 3, p. 442-457.
Pareto, V. (1927): Manuel d'economie politique, deuxieme edition. Paris: Marcerl Giard.
Petersen, Niels Christian ( 1990): Data Envelopment Analysis on a Relaxed Set of Assumptions, Management Science 36, no. 3, p. 305-314.
Reaves, Linda Jean (1983): Using Data Envelopment Analysis to Operationalize the Concept of Equal Education Opportunity, Ph. D. dissertation, College of Education, University of Texas.
Retzlaff-Roberts, Donna Lynn (1990): Incorporating Uncertainty Into Allocative Data Envelopment Analysis, Ph. D. dissertation, University of Cincinnati.
Rhodes, Edwardo L. (1986): An Exploratory Analysis of Variations in Performance Among U.S. National Parks, in: Measuring Efficiency: An Assessment of Data Envelopment Analysis, p. 47-71, Editor Richard H. Silkman. New Directions for Program Evaluation, No. 32, San Francisco, Jossey Bass, Inc.: American Evaluation Association,.
Rhodes, Edwardo Lao (1978): Data Envelopment Analysis and Approaches for Measuring the Efficiency of Decision-making Units with an Application to Program Follow-Through in U.S. Education, Ph. D. dissertation, School of Urban and Public Affairs, Carnegie-Mellon University.
Sant, R. (1989): Measuring the Efficiency of Pubs for Allied Breweries Ltd Using DEA, Unpublished M. Sc. thesis, University of Warwick.
Schinnar, A. P. (1980): Frameworks for Social Accounting and Monitoring of Invariance, Efficiency and Heterogeneity, in: Models for Alternative Development Strategies, Hague, The Netherlands: Institute of Social Studies.
Seiford, Lawrence M. ( 1990): Models, Extensions, and Applications of Data Envelopment Analysis: A Selected Reference Set, Computers, Environment and Urban Systems 14, no. 2.
Seiford, Lawrence M., and Robert M. Thrall (October 1990): Recent Developments in DEA: The Mathematical Programming Approach to Frontier Analysis, Journal of Econometrics 46, no. 1-2, p. 7-38.
Seitz, Wesley D. (1966): Efficiency Measures for Steam-Electric Generating Plants, Proceedings of the Thirty Ninth Annual Meeting of the Western Farm Economics Association, p. 143-151.
20 Data Envelopment Analysis: Twenty Years Out
Sengupta, Jati K. (1982): Efficiency Measurement in Stochastic Input-Output Systems, International Journal of Systems Science 13, p. 273-287.
Sengupta, Jati K. (1989): Efficiency Analysis by Production Frontiers: The Nonparametric Approach, Kluwer Academic Publishers Group, Dordrecht.
Shash, Ali H. (1988): A Probabilistic Model for U.S. Nuclear Power Construction Times, Ph. D. dissertation, Department of Civil Engineering, University of Texas.
Shephard, R. W. (1970): Theory of Cost and Production Functions. Princeton, NJ: Princeton University Press.
Sherman, H. David, and Franklin Gold (June 1985): Bank Branch Operating Efficiency: Evaluation with Data Envelopment Analysis, Journal of Banking and Finance 9, no. 2, p. 297-315.
Sherman, H. David (October 1984): Hospital Efficiency Measurement and Evaluation: Empirical Test of a New Technique, Med Care 22, no. 10, p. 922-38.
Sherman, H. David (1981): Measurement of Hospital Technical Efficiency: A Comparative Evaluation of Data Envelopment Analysis and Other Efficiency Measurement Techniques for Measuring and Locating Inefficiency in Health Care Organizations, D. B. A. dissertation, Graduate School of Business Administration, Harvard University.
Siems, Thomas F. (1991): An Envelopment Analysis Approach to Measuring Management Quality and Predicting Bank Failure, Ph. D. dissertation, Southern Methodist University.
Silkman, Richard H. (Editor), ( 1986): Measuring Efficiency: An Assessment of Data Envelopment Analysis, New Directions for Program Evaluation, No. 32, San Francisco, Jossey Bass, Inc.: American Evaluation Association.
Simar, Leopold (1992): Estimating Efficiencies from Frontier Models with Panel Data: A Comparison of Parametric, Non-Parametric and Semi-Parametric Methods with Bootstrapping, The Journal of Productivity Analysis 3, no. 1/2, p. 171-203.
Sitorus, Bistok L. ( 1966): Productive Efficiency and Redundant Factors of Production in Traditional Agriculture of Underdeveloped Countries: A Note on Measurement, Proceedings of the Thirty Ninth Annual Meeting of the Western Farm Economics Association, p. 153-158.
Splitek, David Franklin (I 981 ): A Study of the Production Efficiency of Texas Public Elementary Schools, Ph. D. dissertation, College of Education, University of Texas.
Stolp, Chandler (1990): Strengths and Weaknesses of Data Envelopment Analysis. An Urban and Regional Perspective, Computers, Environment and Urban Systems 14, no. 2, p. 103-116.
Stone, Martha Jean (1984): A Comparative Analysis of the Personnel Practices of School Districts Selected by Data Envelopment Analysis Efficiency Indices, Ph. D. dissertation, College of Education, University of Texas.
Thomas, David Alan (1990): Data Envelopment Analysis Methods in the Management of Personnel Recruitment Under Competition in the Context of U.S. Army Recruiting, Ph. D. dissertation, Graduate School of Business, University of Texas.
Data Envelopment Analysis: Twenty Years Out 21
Thomas, Dennis Lee (1985): Auditing the Efficiency of Regulated Companies Through the Use of Data Envelopment Analysis: An Application to Electric Cooperatives, Ph. D. dissertation, Graduate School of Business, University of Texas.
Thompson, Russell G., P. S. Dharmapala, and Robert M. Thrall (1994): Sensitivity Analysis of Efficiency Measures With Applications to Kansas Farming and Illinois Coal Mining, in: Data Envelopment Analysis: Theory, Methodology, and Applications, editors A. Charnes, W. W. Cooper, Arie Y. Lewin, and Lawrence M. Seiford. Boston: Kluwer Academic Publishers.
Thompson, Russell G., Larry N. Langemeier, Chih-Tah Lee, Euntaik Lee, and Robert M. Thrall (1990): The Role of Multiplier Bounds in Efficiency Analysis with Application to Kansas Farming, Journal of Econometrics 46, no. 1,2, p. 93-108.
Thompson, Russell G., F. D. , Jr. Singleton, Robert M. Thrall, and Barton A. Smith (November 1986): Comparative Site Evaluation for Locating a High-Energy Physics Lab in Texas, Interfaces 16, no. 6, p. 35-49.
Thorogood, Nellie Jean Carr (1983): The Application and Utilization of Data Envelopment Analysis for Decision Support in the Administration of Instructional Programs for an Urban Community College, Ph. D. dissertation, College of Education, University of Texas.
Tomkins, Cyril, and R. H. Green (June 1988): An Experiment in the Use of Data Envelopment Analysis for Evaluating the Efficiency of UK University Departments of Accounting, Financial Accountability and Management 4, no. 2, p. 147-164.
Tulkens, Henry (1993): On FDH Efficiency Analysis: Some Methodological Issues and Application to Retail Banking, Courts, and Urban Transit, The Journal of Productivity Analysis 4, no. 1/2, p. 183-210.
Varian, Hal R. (1990): Goodness-of-Fit in Optimizing Models, Journal of Econometrics 46, no. 1/2, p. 125-140.
Ward, Peter, James E. Storbeck, Stephen L. Mangum, and Patricia E. Byrnes (1997): An Analysis of Staffing Efficiency in U.S. Manufacturing: 1983 and 1989, Annals of Operations Research (73).
Wong, Y. -H B., and J. E. Beasley (September 1990): Restricting Weight Flexibility in Data Envelopment Analysis, Journal of the Operational Research Society 41, no. 9, p. 829-835.
f f
; t ~ }
C'CI ~ Q) .
> I ~ · t:: ~ Q) ll ~ ~j .. Jl. < I W 1 c h
Treating non-discretionary variables one way or the other:
implications for efficiency scores and their interpretation
Matthias Staat'
Abstract
This paper explains the main DEA-techniques to model continuous and categorical non-dis
cretionary variables as well as a related two-stage approach. The implications of using either
alternative are demonstrated in practice using the pharmacy data from the original study by
Banker and Morey (1986b) on categorical non-discretionary variables.
It is argued that the model appropriate for continuous non-discretionary variables rests on
rather restrictive assumptions about the production technology. The model for categorical
non-discretionary variables does not result in higher efficiency scores, i. e. a more robust
assessment of the inefficiency of production units, as Banker and Morey claim. In addition, its
efficiency scores can not be compared across observations with different values for the cate
gorical variable as each category is evaluated by a differently sized data set. The bias
resulting from this practice is discussed in Zhang and Bartels (1998).
The two-stage approach to modelling non-discretionary variables does not rest on more
restrictive theoretical assumptions than the model for categorical non-discretionary
variables. It does, however, use the full data set for the evaluation of each unit and is
therefore recommended for the empirical analysis when non-discretionary variables are a
relevant factor.
1 Mannheim University, Lehrstuhl fur VWL, ins b. Mikrookonomie, D-68131 Mannheim, Germany
24 Treating non-discretionary variables one way or the other ...
Structure
Data Envelopment Analysis, non-discretionary variables, efficiency ranking
2 Introduction
3 Model formulation
4 Comparison of results
5 Conclusion
References
Treating non-discretionary variables one way or the other ...
1 Data Envelopment Analysis, non-discretionary variables,
efficiency ranking.
25
Despite potentially important practical implications for the efficiency evaluation of
decision making units (DMUs) non-discretionary variables -variables that that have a
productivity relevant influence on the production process but can not be controlled by
the individual DMU- have remained a field of primarily theoretical interest. Ignoring
non-discretionary variables in practical applications either by leaving them out of the
data or by treating them as controllable variables may lead to comparisons of
qualitatively different DMUs. The efficiency scores obtained will then have little
meaning.
The reason why techniques handling non-discretionary variables are rarely applied
may be the confusion over how to specify models for non-discretionary variables and
doubts about what is implied by the various specifications. For instance, Banker and
Morey (1986a) introduced a model for continuous non-discretionary variables which
leads to lower efficiency scores compared to the standard model. Ruggiero (1996)
shows that their approach is not fully consistent with production theory and that the
efficiency scores generated by it may in fact be too low. Banker and Morey (1986b)
demonstrate how to model the effects of categorical non-discretionary variables. They
suggest that this model restricts the set of peers more stringently than the approach for
the continuous case and therefore should tend to result in higher efficiency scores. An
alternative two-stage approach DEA-regression procedure can be used to analyse the
effects of non discretionary variables.
This study tries to clarify these points showing the implications of several
specifications on the original Banker and Morey (1986b) data. Exceptions to the
assertion that the categorical approach leads to higher efficiency scores than the
continuous approach are demonstrated. Also, if categories indicating a small value of a
non-discretionary input contain only a few observations there may in fact be too few
peers to generate meaningful efficiency scores. In addition, each of the categories is
evaluated by a different number of potential peers. This may distort the results in
general and allows only a very limited interpretation of the efficiency scores. Several
alternative ways of treating non-discretionary variables are contrasted using the
Banker and Morey ( 1986b) data.
26 Treating non-discretionary variables one way or the other ...
2 Introduction
Non-discretionary variables are productivity relevant inputs or outputs which are not
(fully) under the control of the decision making unit (DMU) in charge of the
production process. The term fixed or contextual variables is used as well. In addition,
there may be factors which are neither inputs nor outputs to the production process but
nevertheless influence performance. Lovell (1994) labels them environmental factors.'
If a non-discretionary variable is the cause of a DMU being rated inefficient the DMU
would be unable to improve its efficiency to the maximum value of 1 because it can
not or only partially influence the level of that variable. Since the level of such a fixed
variable can not be influenced by the DMU its efficiency score should not depend on
it.
Using DEA, a DMU is evaluated via comparison with a production frontier made up
of actual observations. Therefore it is critical that all observations being part of the
frontier are comparable to DMU0. Regression methods can -theoretically- fit more
than one hypothetical production frontier on a sample of heterogeneous DMUs,
allowing e. g. for different slope parameters for different subsamples. With DEA the
same effect has to be achieved although it is only possible to identify one production
frontier per sample.
There are two ways out of this problem: One is to come up with additional criteria for
the comparability of DMUs in a given sample. These result in restrictions for the
reference technologies. The other way is to alter the sample according to ones beliefs
about which DMUs are comparable with DMU0. Only the first strategy is associated
with non-discretionary variables by Lovell (1994, section 8.2); the second with what
he calls environmental factors (section 8.3).
Non-discretionary variables are a potentially important factor in almost any DEA
analysis. However, the topic is hardly ever touched by practitioners. The usual set of
parameters for the evaluation of, for example, hospital productivity -a standard
application of DEA models- includes type of ownership, size of the hospital and case
mix. These factors are productivity relevant in different ways. For instance, it is often
found that non-profit hospitals are run less efficiently than for-profit hospitals. The
type of ownership can usually not be influenced by the hospital management but on
2 Other authors use the non-discretionary and environmental as synonyms.
Treating non-discretionary variables one way or the other ... 27
the other hand it has no direct influence on the functions performed in the hospital. No
special treatment for this variable is necessary. 3
The variable case mix has a different character: if some hospitals treat a mix of cases
which is more demanding in terms of therapy than others this should be reflected by
the fact that none of the hospitals with standard case mixes should be part of the set of
peers of the disadvantaged hospitals. Put differently, if an analysis disregarding case
mixes reveals that hospitals with difficult case mixes are less efficient this can hardly
be interpreted as managerial slack.
Finally, the size of a hospital can not be changed in the short run by hospital managers.
Assuming full capacity use, one could consider larger hospitals to have an advantage
over smaller ones since they can usually afford higher investment in equipment that
allows for more efficient treatment.• This should be a reason to consider non
discretionary techniques.
The classical examples for the application of techniques treating non-discretionary
variables from the Banker and Morey (1986a, 1986b) papers are a restaurant chain
where advertising budgets are set by the central marketing division rather than
individual restaurant managers and pharmacies located in communities of different
sizes. The size of the community and the size of its marketing budget cannot be
controlled by the DMUs' managers; therefore these variables are labelled non-discre
tionary. Not all restaurants of the chain can be compared to each other as, for instance,
the restaurants with a temporary promotion may have an easier time selling certain
products than restaurants which sell them for the regular price. Pharmacies in smaller
communities may have a more limited sales potential than pharmacies in larger
communities. Therefore, comparisons between pharmacies from communities of
different sizes may not make sense.
It is, however, desirable to include as many as possible DMUs in the analysis. For
instance, all the available information on restaurant performance should be used to
evaluate the performance of a long established restaurant. Its performance can be
3 This is only true if public hospitals do not have functions in the health sector different from private hospitals.
4 Of course, one could also imagine situations in which smaller hospitals have an advantage over larger ones. Lovell ( 1994, p. 213) describes procedures which can be applied the direction of the effect is not known ex ante.
28 Treating non-discretionary variables one way or the other ...
compared to all restaurants which have been established around the time it went into
business or later.
It is important to recognise non-discretionary variables as such and to treat them
accordingly because the credibility of DEA results depends critically on whether the
set of potential peers comprises only DMUs which can be compared to the DMU being
evaluated. Comparing DMUs operating in different environments without controlling
for environmental factors will lead to differences in the environment being wrongly in
terpreted as differences in efficiency.
The following section will present the basic5 approaches to handle different types of
non-discretionary variables. These basic principles apply to all direct extensions and
refinements as well as to a number of related approaches. Some examples illustrate
how the models discussed work in practice and what is implied by that. A pragmatic
aside asserts in a non-technical way which model may be best used in a given
situation. The implications of handling the problem in different ways are demonstrated
on the original Banker and Morey ( 1986b) data in the next part of the paper. A final
section will take stock of the results.
3 Model formulation
3.1 Standard Model
Banker and Morey ( 1986a) begin their discussion of possible alternatives to the
standard model by noting that one element of the input vector xij, i = 1,2, may not be
under the control of DMU0 to be evaluated by comparison with the DMUs
j=l, ... ,N.•
The standard model is displayed here in its input-oriented, variable-returns-to-scale
(VRS) form as formula (1). The value of the efficiency parameter /%, 0< 80 $1,
describes what fraction of the actual inputs of DMU0 would suffice to produce its
5 Refinements and variants of these basic models will only be mentioned in passing as the purpose of this paper is to structure the problem of treating non-discretionary variables instead of the enumeration all of its aspects.
6 To make the exposition as transparent as possible the two-input-one-output example from Banker and Morey (1986a) is used. The generalisation is a matter of notation.
Treating non-discretionary variables one way or the other ... 29
output were it to employ an efficient technology. There may be additional positive
slacks for some of the inputs and outputs, s1- ,s•, respectively. The efficient reference
technology is formed by a A.-weighted average of input-output combinations of other
efficient DMUs.
min z0 = B0 - es•- es; s,;.,,+ ,,-
N
s. t. LY;A.1 -s• = Y, j
N
:~::XuA. 1 - s; = B0x10
I N
LAJ =I j
o •. A. j,s+,s,- ~ 0
3.2 Continuous non-discretionary variables
(1)
Banker and Morey (1986a) now point out that if any of the inputs in the above formu
lation can not be controlled by the DMUs then DMU0 has no possibility of reducing
that input to a fraction ~ of its original level. They therefore suggest the following
reformulation of model (1), where the superscript ''f' in B& indicates that the case of
continuous non-discretionary or fixed variables is treated. 7
The input vector again consists of two inputs, one discretionary (d), the other one fixed
(f). Only the discretionary input directly enters the determination of the maximum
input reduction possible for inefficient units. The parameter ~ has vanished from the
constraint for Xfi so has the slack of that variable from the objective function.
The role now played by the non-discretionary input is reduced to ensuring that the
reference technology has on average no advantage over DMU0 with respect to the
variable x1. This is how it assures comparability of DMU0 and its reference technology;
it corresponds to the first strategy of solving the problem that was mentioned in the
introduction. It is tantamount to relaxing the original set of constraints as the
A.-weighted average of the fixed inputs now only has to meet the condition "5. x 10 "
7 The superscript ''!' will be used to indicate a continuous fixed variable whereas the superscript "c" will be reserved for the categorical variables discussed in section 3.3.
30 Treating non-discretionary variables one way or the other ...
instead of the stronger condition "$ B0x10". Banker and Morey (1986a, p. 515) point to
the fact that softening the original condition "enriches the comparison set". As a con
sequence, Banker and Morey's proposition 1 states that 80 :2: B[. 8
N
s. t. LYy11 -s+ =Yo J
N
:~.::X.q-1 1 - sd- = B~ X do j
N
LXDAJ -sf- =xfo
(2)
The fact that the reference technology on average has to have the same value for the
non-discretionary indicator as DMU0 implies that, for example, a restaurant with one
week of promotions can be sensibly evaluated by a reference technology made up of
two restaurants each with A. = 0.5 where one has no, the other two weeks of
promotions in a given period. This point will be taken up again in section 4.
Lovell (1994, section 8.3) points out that one could also ensure comparability of the
DMUs by excluding all DMUs for which x0 > x10 holds from the reference technology
of DMU0 as higher levels of the non-discretionary input indicate an advantageous
position. The character of the restriction changes: Conditions on the reference technol
ogy are replaced by eligibility conditions on individual DMUs for inclusion in the
sample (and therefore in the reference technology as well). This leads to
8 An extension of this specification to the treatment of simultaneous continuous non-discretionary inputs and outputs is mentioned in section 2.4.
Treating non-discretionary variables one way or the other ...
min Z0 = 9~ -ES+ -esd-e,.>..,.r* .sd-
N
s. t. L>?·1 -s• = Y, J
N
~>djA j - Sd- = 8~Xdo 1
31
(3)
and is identical with the model Ruggiero (1996) proposes. Also, this corresponds to
the second strategy mentioned in the introduction: restricting the sample to DMUs
comparable with DMU0.
Ruggiero ( 1996) concludes that e,f may be somewhat too low due to the inclusion of
DMUs in the reference technology which do not belong there because they operate in
an advantageous environment compared to DMU0 and therefore e~ > e~ (see
proposition 1, p. 559). Whether ~ ("/" indicating the Lovell/Ruggiero specification)
will be lower or higher than 80 can not be determined a priori as the original
restriction is replaced by a different type of restriction.
As pointed out by Lovell (1994, section 8.3), one consequence of this type of
restriction is that unlike in the Banker and Morey (1986a) method discussed so far, the
sample size changes from evaluation of one DMU to the next. For DMUs with small
values of x1 the number of potential peers may be only a fraction of the number
available for the evaluation of DMUs with large values for the fixed parameter. This
point will also be discussed more in-depth in section 4.
Table 1 gives some examples for a DMU0 with both the discretionary and the fixed
input and the output equal to 1 being evaluated by different peer units. The first data
column in Table 1 contains values for DMU0. The following columns list data for
efficient peers. To keep the examples simple, a reference technology is made up of just
as many efficient peers as necessary to demonstrate a certain effect. The first row of
data contains the values for the output, followed by the discretionary and the fixed
32 Treating non-discretionary variables one way or the other ...
input in row 2 and 3, respectively. Row 4 contains the weight for the efficient peer and
rows 5 to 7 contain the efficiency scores calculated by the different models.
Table 1: Efficiency scores for various methods w. r. t non-discretionary variables
DMU0 Peers (a) (b) (c) y 1 1 118 4.5 1
Xd 1 0.5 118 2 0.5
Xf 0.5 118 2 1
A. 1 0.8 0.2
Bo 0.5 .5 1 ()f 0.5 .5 0.5
0
B' 0 0.5 1 (?) 0.5
Example (a) demonstrates how all different models lead to the same efficiency score.
Suppose 00 is .5 since there existed a DMU with the same output as DMU0 using only
half of all inputs to produce that output. Assuming there are no other dominant
(combinations of) peers e~ would also be .5 as xd of the peer unit is half of xod and the
fixed parameter naturally meets the softer condition of the non-discretionary approach
(see section 3.2). In this case, €fo would also be .5 since the only member of the
reference technology has no advantage over DMU0 with respect to Xp
The case that is relevant for the Levell/Ruggiero specification is that of a "large" peer
DMU with 4.5 times the output and twice the input ofDMU0 and another "small" peer
with values that are just 118 of DMU0 's. This situation is considered in example (b).
According to Levell/Ruggiero DMU0 should not be evaluated by a reference
technology containing the large DMU. Therefore, DMU0 is rated as efficient as long as
no other efficient peer (combination) dominates it. The other two approaches would
permit the large DMU in the reference technology and thus give each an efficiency
score of0.5.
Finally, there are cases when B~ = B~ '* 00 • This happens when the efficiency of DMU0
critically depends on x1 when evaluated by means of the standard model as in example
(c). Here, only x1 is used in the same way by the technologies of DMU0 and the
technology of the efficient peer but the latter is more efficient in using xd. Due to
x,d = x0d, 00 =I holds. Softening the constraint for x1 the performance with respect to
Treating non-discretionary variables one way or the other ... 33
xd now becomes critical and B~ drops from 1 to 0.5. Since there is no advantage of the
peer over DMU0 with respect to the fixed parameter B'r, is also .5; i.e. lower than the
original value ~ = 1. There can be no general a priori ranking for the values of 90 and
B~.
3.3 Categorical non-discretionary variables
Banker and Morey (1986b) consider the case when non-discretionary variables are not
continuous. Only the category into which the value for the non-discretionary variable
falls is known." The example they use to illustrate their model are pharmacies located
in communities of different sizes. Pharmacies in smaller communities naturally have a
lower sales potential and should therefore not be evaluated through peers located in
communities. Banker and Morey ( 1986b) classify the communities into eleven
categories which were also used by the U.S. Census of Population and Housing,l980. 10
This classification is supposed to reflect differences in market size defined by
population figures.
Banker and Morey ( 1986b) generate dummy variables which are set to 1 if a com
munity is of equal or smaller size than indicated by the dummy. If a community
belongs to, say, population size category 3, the dummies for population sizes 1 to 3 are
set to one and the dummies for category 4 and above are set to 0. Banker and Morey
(1986b) then treat the dummies like the non-discretionary variables in specification (2)
above. It is obvious that the reference technology for certain pharmacies can only
consist of pharmacies located in communities of the same or of smaller size. This is so
because x01, x = 1 , . . . , 11, assumes the value 0 for all classes indicating larger
communities (markets) than the one of DMU0 and any pharmacy from a larger
community will violate the restrictions for the reference technology with respect to x1
The same could be accomplished by using a variable with values I to 11 and applying
the Lovell/Ruggiero specification (3) above. In fact, Ruggiero notes that his model for
continuous non-discretionary variables can be interpreted as a variant of Banker and
Morey's (1986b) approach for categorical non-discretionary variables. However, it
would be more in the spirit of the method developed by Ruggiero ( 1996) to base the
decision about whether to accept a DMU in the set of peers on actual instead of
9 The same study contains a model for discretionary categorical variables.
10 The upper limits of the II classes are 199,499,999, 1499, 1999,9999, 19999,24999,49999,99999,249999.
34 Treating non-discretionary variables one way or the other ...
categorised population sizes. Again, this point will be discussed in more detail in
section 4.
Banker and Morey (1986b, p. 1614) motivate their approach by the following
example:
However, suppose we are attempting to estimate the resources (such as
labour and capital) that a branch of a bank needs to obtain a given level of
deposits, given a population base of say 100.000, with a specific income,
age, and other demographic characteristics. Then in DEA the branch in
question might well be compared to a composite branch built from a branch
with a population of 80,000 and another with a population of 120,000, both weighted equally. While this may seem like a very reasonable
approximation, it is clear that the branches employed for this comparison
would be less controversial if we were to insure that the peer group
consisted only of branches with a population of 100,000 or less.
What is desired for the above situations is a method for insuring that the composite reference members be constructed from members which are in
the same or possibly from those in a category which is deemed to be operating in an even more difficult or unfavourable situation.
The first paragraph clearly calls for the approach suggested by Ruggiero ( 1996). In the
second paragraph, the conditions for being a member in the reference group are
somewhat softened because conditioning on "the same ... situation" can lead to
comparisons of rather different DMUs. For instance both the community of I 00.000
and the community of 120.000 in Banker and Morey's example belong to the same
market size category (11) which ranges from 100.000 to 249.999 (see footnote 10).
Thus, the categorical approach in this case makes exactly the kind of comparisons
possible it was designed to prevent.
Maybe the above comparison is allowed for good reason and the classification puts
communities that constitute markets of the same size into the same categories. But
maybe this categorisation was - as categorisations often are - designed such that there
are sufficient members in each cell. This consideration - legitimate as it may be - is
Treatmg non-discretionary variables one way or the other ... 35
certainly not related to market size and the categories above could be irrelevant or
misleading in the given context."
Also, using this classification of DMUs, reference groups now become possible that
would not be possible applying any other approach. For instance, a DMU from a community of 100000 could now have a reference technology consisting exclusively
of peers from communities with a population between 200000 and 220000.
Of course, one could "solve" the problem by choosing only peers which operate under
more difficult circumstances for comparison. This would leave some inefficiencies
undetected that can only by demonstrated by comparison with DMUs which operate
under like circumstances. Efficiency scores generated in this way could be considered
upper bounds to the "true" scores.
4 Comparison of results
4.1 General aspects
At this point it seems fitting to collect some of the basic results and discuss some implications of using one of the two (three) methods designed to handle non
discretionary variables.
A slight modification of the standard model is necessary to handle continuous non-dis
cretionary variables according to Banker and Morey (1986a, see formula (2)). This
leads to lower efficiency scores since it results in "enriching the comparison set". It
was also mentioned that this implies that the reference technology and DMU0 be on
average the same with respect to the non-discretionary variable. The model is
11 Often, no ready made categorisation will be available and one has to make up a classification ad
hoc. Consider international comparisons which are listed among the most important novel applications
in Seiford ( 1996). Suppose country size matters: There is no obvious answer to the question into how
many categories a sample of countries should be divided by size. Even if there is an answer, say three
categories, what is a small, a medium size and a large country w. r. t. the matter of interest may not
always be simple to decide. This reintroduces a certain arbitrariness into the analysis.
36 Treating non-discretionary variables one way or the other ...
applicable for a technology like the one discussed in Ray (1988, p. 171, equation (11))
for which multiplicative separability in the discretionary and the fixed factors holds. 12
The modifications to the standard model called for by the models described in Banker
and Morey ( 1986b) as well as Ruggiero ( 1996) are more significant than the ones just
described (see formula (3) above). However, these approaches are easy to implement.
The model can be applied by selecting all admissible peer-DMUs for each DMU and
then using the standard model without any further modifications.
In order to apply the Lovell/Ruggiero model one would have to be convinced that the
non-discretionary parameter affects productivity in a direct way. For instance, in
Banker and Morey's (1986b) study (see also section 4.2.1 of this paper) population size
was used as a surrogate for sales potential. They mentioned, however, that sales
potential depends also on unknown demographic characteristics of the population such
as age and income distribution, education etc. (see the quote on p. 34). In a case like
this, it would not make much sense to apply this model using actual population size
since all that is known is that communities of the same size have about the same sales
potential. 13
What is called for in this case is the approach suggested by Banker and Morey (1986b)
with broad categories of population sizes. This model, too, is not without peculiar
aspects. Most notably, it would allow for reference technologies that consist of peers
which all have a higher sales potential than DMU0• 14 Of course, any empirical method
would suffer from imprecise information. But unlike regression which transforms
imprecise information - interpreting it as a measurement error - into imprecise results,
DEA will translate it into biased results about the efficiency ofDMUs.
The following table collects some characteristics of the models discussed.
12 Ray {1988) points out that his formulation amounts to modelling the non-discretionary factor like a parameter for Hicks-neutral technical progress.
13 The model would be more appropriate for a technical parameter of a production schedule like investment where higher investment means better facilities and in tum better capabilities of production.
14 In the pharmacy data set, actual population sizes are known such that an additional restriction preventing that could be implemented. In cases when only the categories are known, however, this will not be possible.
Treating non-discretionary variables one way or the other ... 37
Table 2: Models, characteristics, implications
Model Relative Apply if I problems Ease of use
(formula) efficiency
BM 1986a f)~ ()f DMU0 and reference technology on average the same Code modification
(2) w.r.t. indicator. of standard model necessary
BM 1986b Claim: Only categorical indicator available. Standard model with
(3) ()'"' ~ ()f Continuous indicator with loose connection to varying data sets
but see productivity relevant factor.
section May lead to reference technologies with peers that are
4.2. all advantaged w.r.t. Xf
Ruggiero ()/ ~ ()f Continuous indicator. Standard model with
(3) Close correspondence between indicator and en-varying data sets
vironment!technology.
4.2 Banker and Morey (1986b) data revisited: Part I
4.2.1 Data
The above models are now applied to the original data set used in Banker and Morey
(1986b, Appendix A, p. 1624 ff. ). The purpose of their study was to demonstrate the
difference between modelling a non-discretionary variable as a continuous indicator
vs. modelling it as a discrete indicator, i.e. specification (2) vs. a variant of
specification (3) above. Their results will be reproduced below. 15 In addition, the
results for the standard specification (1) will be reported. They will serve as a baseline
for specification (2). According to Banker and Morey (1986a), specification (2) will
yield lower efficiency scores than specification ( 1) whereas specification (3) will yield
a higher score than specification (2).
The data consist of 69 pharmacies located in communities of different sizes ranging
from only 500 to more than 200.000 inhabitants. The population count is the so-called
non-discretionary parameter as it can not be controlled by the pharmacy owner once
the pharmacy has been established. It is modelled as an input since Banker and Morey
15 The results are slightly sensitive to the scaling of the data. As Banker and Morey do not indicate how they scaled the data their results are not to the third digit past the dot the same as the ones reported here. However, this does not produce spurious effects for the comparisons.
38 Treating non-discretionary variables one way or the other ...
(1986b) assume that the larger the community, the easier it is to achieve an efficient
input/output combination.
The other three (discretionary) inputs are labour as well as other operating costs and
the average value of the inventory.
The two outputs are number of prescriptions and their value.
4.2.2 Results
Table 3 lists the results for selected DMUs. The data columns contain the efficiency
score and in order to economise on space only the population classes of the peers of
the inefficient DMUs are listed. The different rows refer to the different methods of
handling non-discretionary variables just discussed.
Pharmacies #15 and #52 were chosen by Banker and Morey in their study to dem
onstrate the effect of treating population size as a categorical non-discretionary
variable vs. treating it as a continuous non-discretionary indicator. Pharmacy #15 is
located in a community which has a population of 2500 and therefore belongs into
population category 6.
When treating population size as a continuous variable one of #15 peers belongs to
category 5, i.e. it serves an even smaller population whereas the second peer belongs
to category nine with a population of over 30000. When population is treated as a
categorical indicator the set of peers does contain again two observations, the one from
category 5 also contained in the set of efficient DMUs when population is treated as a
continuous indicator and another peer from population category 6. The efficiency
score of #15 rises slightly from .56 to .57 using the categorical model. This is in line
with the authors' expectations.
The peer-DMU from population size category 6 that was contained in the second of
the above reference technologies is located in a community with a larger population
(over 4000) than the community of pharmacy #15. Therefore, the Lovell/Ruggiero
approach which - when applied rigorously - only allows peers that have the same or
smaller population again produces a new set of peers from categories 4 and 5 and a
still higher efficiency score of .58.
Treating non-discretionary variables one way or the other ... 39
Table 3: Banker and Morey (1986b) results compared
Pharmacy (Category) #15 (6) #52 (8) #26 (6)
Model specifications
Peer categories 4, 5, 6 4,6,9,9
non-discretionary, continuous Efficiency score 0.56 0.65
Peer category 5,9 4, 6, 9, 9, II
non-discretionary, categorical (II} Efficiency score 0.57 0.86 0.95
Peer category 5,6 5,5,6 6,6,6,6
non-discretionary, Ruggiero Efficiency score 0.58 0.86
Peer category 4,5 5, 5, 6
non-discretionary, categorical (2) Efficiency score 0.57 0.65 0.77
Peer category 5,6 3, 9, II 4,6, 6, 7
Standard model (baseline) Efficiency score 0.62 0.66 I
Peer category 3, 5, 6 4, 6, 9, 9, II
The highest efficiency score is generated by the baseline specification treating
population size as a continuous discretionary variable. This is in line with proposition
1 in Banker and Morey (1986a). While the restrictions for the non-discretionary cases
allow all reference technologies which have a (weighted) population - measured in
actual numbers or in categories - of less than the community of pharmacy #15 the
baseline specification only allows reference technologies with population less than or
equal to 80*(population of#15). The latter turns out to be much more restrictive in this
case (see section 3.2).
While one can clearly identify the effect of using different specifications the actual
efficiency scores are very similar when looking at the results for #15. Banker and
Morey (1986b) also give an example for a major change of the efficiency score due to
a change in specification. Pharmacy #52 serves a population of 23.166 and therefore
belongs to category 8. The reference set for the continuous non-discretionary approach
consists of five pharmacies, two from smaller and three from larger communities than
its own. Except for the pharmacy in the community of category 6 none of the original
set of peers remains in the reference technology when the categorical approach is used.
Thus, the efficiency score rises from 0.65 to 0.86. Disregarding the non-discretionary
character of the variable population size altogether leads to virtually the same results
as treating it as a continuous non-discretionary variable (see baseline).
40 Treating non-discretionary variables one way or the other ...
For these two pharmacies the propositions in the papers cited hold. The continuous
non-discretionary approach leads to lower efficiency scores than the discretionary
approach, the Lovell/Ruggiero specification leads to higher scores than the continuous
approach as does the categorical non-discretionary specification. Note, however, that
for the latter to happen in all cases, the categorical approach would have to result in a
more stringent restriction than the continuous one. As discussed in section 3.3 of the
paper, this is not the case. With the categorical approach it would theoretically be
possible for #15 (or #52) to have a set of peers consisting of pharmacies from
communities larger than its own (but not larger than 99999). This would not be
possible under the continuos non-discretionary approach.
A look at the results for #26 confirms that this does indeed happen when using the
categorical specification. This pharmacy is located in a community of 2217. This
community is slightly smaller than the one #15 is located in but belongs to the same
category, 6. The continuous non-discretionary specification rates #26 as efficient. The
categorical case gives an efficiency score of only 0.95 for #26. All other specifications
again lead to an efficiency score of I. The peers forming the reference technology in
the categorical specification belong to the same category as #26 does but all are lo
cated in larger communities. The actual populations range from 2718 to 5607. There
fore it is not necessarily true that, as Banker and Morey (1986b, p. 1619) claim: "the
number of DMUs identified as technically inefficient will not increase under the
categorical treatment, ... ".
This can be demonstrated by changing the number of categories used. One could
naively assume that it only matters whether a pharmacy is located in a small (less than
25000 inhabitants) or a large community (25000 and over). The results for this
specification are also reported in Table 3. Banker and Morey (1986b) report 41
inefficient DMUs for the continuous and 36 inefficient for categorical (11 categories)
case. However, reducing the number of categories to 2 results in 42 inefficient DMUs
which shows that this does not hold. For #26 the efficiency score drops to 0.77 using
the variant with two categories.
Finally, the fact that for the latter two methods the sample size changes from
optimisation to optimisation may raise concern about whether there are sufficient
Treating non-discretionary variables one way or the other ... 41
observations in each of the categories'• and whether the efficiency rankings can be
compared across categories. Since efficiency measures vary with the ratio of
parameters" to observations only the B.; for the DMUs of one and the same category
can be compared. In a model of type (3) with non-discretionary inputs there will be
more potential peers for DMUs with large values of that input and less for DMUs with
small values.
Before some further empirical illustrations of the models are presented some
extensions of the DEA model to incorporate non-discretionary variables are briefly
mentioned.
4.3 Further Extensions
4.3.1 Simultaneous non-discretionary inputs and outputs
The techniques described up to now were developed to handle either non-discretionary
inputs or non-discretionary outputs. Golany and Roll (1993) develop a model that
handles non-discretionary inputs and outputs simultaneously. They note that the same
linearised model can be derived from different ratio forms. By choosing a convenient
ratio form (see Golany and Roll, 1993, p. 423f) they are able to extend the basic model
to simultaneous non-discretionary inputs and outputs.
Formula (4) describes a problem with two inputs and two outputs, one discretionary
and one fixed each. This model also differs from the above specifications in that
constant returns to scale are assumed. For a variable-returns-to-scale specification
L 1 A 1 =I must hold and in an input-oriented model there would be no difference in the
treatment of discretionary and fixed outputs. This results in:
16 This concern is expressed in Lovell (1994, section 8.3). It is obvious that the DMU with the smallest value for any non-discretionary input can not have a reference technology matching that value in a VRS specification.
17 See Zhang and Bartels ( 1998) for Monte Carlo evidence on this point .
42 Treating non-discretionary variables one way or the other ...
N
s. t. LYJ1A.1 -sd • = ~0 j
N N
LYpA; -s/ = LA;lfo
N
LXdJAJ- sd- =~X do j
N N
LXpA j- s,- = LAJXfO
(4)
Golany and Roll (1993) also contains an extension of the basic model to partially
controllable variables. The method can be applied by using model (2) above after the
variables have been transformed accordingly.
4.3.2 DEA-regression two-stage procedures
Ray (1988, 1991)" treats non-discretionary variables combining DEA and regression.
He performs standard DEA on the discretionary subset of the variables in the first
stage and regresses the efficiency scores derived, h, on the non-discretionary factors in
the second stage.
In Ray (1991), the second stage regression is used to predict maximum efficiency, h*, given the set of non-discretionary variables. This adjusted maximum efficiency may be
well below 1. The difference h *-h is then interpreted as managerial inefficiency. The
maximum efficiency measure given x1 is derived by adding the highest positive value
of the error term to the intercept in the regression. The residuals derived using this
adjusted intercept will all be non-positive. Thus, the predicted value for the efficiency
measure will at least be equal to the observed value derived in the first stage DEA and
18 Chames et al. (1981) mention some earlier studies using two-stage procedures.
Treating non-discretionary variables one way or the other ... 43
consequently the inefficiency measure is lower than the first stage results would
suggest. 19
While Ray (1988) justifies his empirical procedure on theoretical grounds Fried et al.
(1993) propose a similar two-stage procedure for reasons of empirical practicability.
They analyse a sample of American credit unions. The two stages of their approach
comprise a free-disposal hull (FDH) model on the first and a logit regression on the
second stage. The logit model uses a binary indicator on whether a DMU was found to
be inefficient or efficient in the first stage analysis as the dependent variable. The first
stage relies solely on discretionary indicators while the second stage tests whether the
first stage findings can be explained by the non-discretionary indicators disregarded in
the first stage analysis. Fried et al. (1993) also apply a SURE system estimator to the
slacks of each variable used in the first round as the dependent variable and regress
them on the set of non-discretionary indicators. While the logit (single equation)
approach results in some plausible parameters explaining the variation of efficiency
through differences in the environment, the SURE approach yields no clear cut results.
As pointed out by Lovell (1994, section 8.3) both methods have one advantage over
the Lovell/Ruggiero (1996) and Banker and Morey (1986b) type of models: there is no
reductions in sample size for particular DMUs. Both models are also reminiscent of
partial regression (see Greene, 1997, Ch. 6) in that the variation not explained in a first
round analysis is analysed in a second stage.20
To further analyse what factors drive the results originally obtained by Banker and
Morey (1986b) a two stage procedure will be applied in the next section.
19 It should be noted that the proper estimation procedure in case of a censored dependent variable -the efficiency score is censored at 1- would be a !obit specification. Therefore, Ray's estimates give biased results (see Greene, 1997, Ch. 20).
20 Using partial regression, however, the effects of the non-discretionary factors would be "netted out" during the first stage of the procedure and the "pure" efficiency effects would be analysed during the second stage. This may be an alternative to the two approaches described.
44 Treating non-discretionary variables one way or the other ...
4.4 Banker and Morey (1986b) revisited: Part II
4.4.1 The effects of changing sample size
The fact that for the model relating to categorical variables the sample size changes
from optimisation to optimisation may for one raise concern about whether there are
sufficient observations in each of the categories" and also whether the efficiency
rankings can be compared across categories. Since Zhang and Bartels ( 1998)
demonstrate that efficiency measures vary with the ratio of parameters to observations
-8 rises c. p. if more parameters are added to the model while the number of
observations remains constant just as the R2 of a standard regression model rises under
the same circumstances- only the ~ for the DMUs of one and the same category can
be compared.
To demonstrate the effect of changing the ratio of parameters and sample size the
following experiment is conducted. The sample is first sorted by population size, then
it is split into two halves in two different ways. First, two samples are created, one
consisting of all odd numbered pharmacies, the other of all even numbered ones. Next,
two samples, one containing pharmacies 1 to 37, the other pharmacies 38 - 69 are
created." The reason for this exercise is to find out how structural efficiency is
influenced by changing the number of DMUs in this sample by deleting observations
in a random versus deleting them in a systematic way.
Deleting randomly for instance the odd or the even numbered pharmacies, corresponds
to the results obtained by Zhang and Bartels (1998) about the relationship between
sample size, number of parameter and efficiency scores. Deleting systematically, i. e.
deleting the small or the large pharmacies from the sample, corresponds to restricting
the sample to comparable DMUs. The Banker and Morey (1986b) as well as the
Lovell!Ruggiero method do therefore suffer from the same type of bias for the
efficiency parameters that can be detected by deleting observations systematically.
This bias may in fact be smaller than the one create by randomly deleting observations
as one deletes systematically DMUs which are less likely to be an efficient peer for the
21 This concern is expressed in Lovell (1994, section 8.3). It is obvious that the DMU with the smallest value for any non-discretionary input can not have a reference technology matching that value in a VRS specification.
22 The small pharmacy sample contains 37 pharmacies as there is a jump in population size between pharmacies 3 7 and 3 8.
Treating non-discretionary variables one way or the other ... 45
DMUs left in the sample. From the point of view of the pharmacies in the small
sample it may not matter all that much whether the largest five pharmacies are taken
out of the sample or not.
Table 4: Sample size effects
model specifications Average score using
Standard variables no variable for size
all observations .88 .82
Even .90 .88
Odd .92 .84
Small .92 .90
Large .91 .84
Table 4 above shows the results for the two experiments. As was to be expected in
view of the results obtained by Zhang and Bartels (1998) the efficiency scores derived
with the split samples suggest lower structural inefficiency than those for the
resprective full samples. However, there does not seem to be a difference for this
particular data set between deleting systematically of randomly as the efficiency scores
of the respective subsamples are similar. Hence, the sample size effect seems to carry
through.
4.4.2 The determinants of efficiency
Table 5 presents the original Banker and Morey (1986b) results by population size
category. There were no observations in the first two categories of the classification
they used such that nine categories remain. As discussed in section 3.3 the pharmacies
in smaller communities can only be evaluated by other pharmacies in about as small or
even smaller communities.
Therefore, the 3 pharmacies in category 3 will only be compared with each other. The
7 observations in category 4 -see the column "nr. of obs."- can be compared with each
other and with the 3 smaller peers in category 3. This results in 10 potential peers
altogether- see the column entitled "nr. of peers". There are 14 potential peers for the
pharmacies in category 5.
46 Treating non-discretionary variables one way or the other ...
Table 5: Banker and Morey (1986b) data revisited
POPCAT nr. ofobs. nr. of peers Theta_8
3 <1000 3
4 <1500 7 10 .98 .99
5 <2000 4 14
6 <10000 23 37 .88
7 <20000 2 39 .94 .88
8 <25000 40 .86
9 <50000 10 50 .79
10 <100000 9 59 .83 .81
II <250000 10 69 .81
Total .88
As a consequence, the efficiency scores are in effect derived from three different
DEAs, one based on 3, another on 10 and the third on 14 observations. Not
surprisingly, given the fact that six variables are used in the analysis, only one of the
fourteen smallest peers is inefficient and the average efficiency of this group is .99.
For the middle three categories evaluated by 37 to 40 peers the average or structural
inefficiency is .88 whereas it is .81 for the largest categories which is based on 50 to
69 peers. The overall average being again .88. Of course, smaller pharmacies might be
more efficient than larger ones but in this case one can not draw this type of
conclusion as long as the effects of sample size may still play a role.
To demonstrate the properties of their new model Banker and Morey ( 1986b) compare
their results to findings generated by a model for continuous non-discretionary
variables. To generate these results they replace the population categories by a variable
on actual population size and employ model (1) above. They point out the fact that the
categorical approach generate less (36) inefficient DMUs than the model bases on the
continuous variable and note that this is in line with their theoretical assertion that the
categorical approach will result in lower inefficiency but that the inefficiency detected
will have a higher credibility than the inefficiency derived by the continuous type
model. Interestingly three out of the five newly efficient DMUs are among the first 14
DMUs where efficiency scores are derived on the basis of extremely small sample
sizes.
Treating non-discretionary variables one way or the other ... 47
To gain some insight into whether the results are driven by sample size or population
size, a DEA on the full sample without using the variable for population size is carried
out. Next, the efficiency scores given by that specification will be regressed on an
indicator for the population category." The results of the second stage tobit-regression
are shown in Table 6. It is clear that by not using an indicator for population size
during the DEA-stage, all the efficiency relevant information relating to size should
still be contained in the efficiency scores. However, the tobit model does not detect a
significant relationship between size and efficiency. Probably even more surprising, if
the effect found was a significant one, its direction would be just the opposite of what
Banker and Morey ( 1986b) assumed: Here, pharmacies in larger communities seem to
be in a disadvantaged position whereas Banker and Morey ( 1986b) considered a larger
community to be advantaged with respect to sales potential.
Table 6: "Determinants" of efficiency: Tobit model I
Dep. V ar. theta _p
Coefficient
!-value
Population categories
-0.02
-1.55
Constant
1.02
9.96
s. e.
0.25
The next table shows the results of a similar analysis. This time the efficiency scores
presented in Table 5 are regressed on the number of peers also listed in that table. This
gives the -in view of the Zhang and Bartels (1998) results not at all surprising- finding
that sample size matters when taking account of population size in the way Banker and
Morey (1986b) do.
23 This corresponds to the strategy applied by different researchers to model the effect of nondiscretionary variables: a two-stage procedure consisting of a DEA model using only discretionary variables on the first stage and some regression procedure applied to the scores of the first round analysis and the non-discretionary factors on the second (see section 4.3.2).
48 Treating non-discretionary variables one way or the other ...
Table 7: "Determinants" of efficiency: Tobit model II
Dep. Var. Theta_8 m. of peers _cons _se
coef. -0.01 1.24 0.21
t-value -3.95 14.56
This suggests that all results based on model specifications that apply restrictions on
individuals peers such that each DMU0 will have a different number of potential peers
will generate biased results.
5 Conclusion
The paper discussed several aspects related to the treatment of non-discretionary
variables. To assure compatibility of the empirical models with the theory of
production functions one has to assume multiplicative separability between
discretionary and non-discretionary factors. One way of doing so is interpreting non
discretionary factors as Hicks-neutral technical progress (see Ray, 1988). A more
general way of modelling the influence of non-discretionary factors seems desirable.
On the empirical side, some models possess the undesirable property that the
efficiency scores are derived using subsamples of different sizes. In view of the results
obtained by Zhang and Bartels (1988), these results have to be considered as biased.
Especially if it seems inappropriate to evaluate DMUs by reference technologies
which are only on average comparable to DMU= but which contain individual peers
that are advantaged with respect to a non-discretionary variable, the only alternative
are two-stage approaches. This is a field where future research efforts should be
directed.
References
Banker, R. D. and R. C. Morey (1986a): Efficiency analysis for exogenously fixed inputs and outputs,
Operations Research 34 (4): 513-21.
Treating non-discretionary variables one way or the other ... 49
Banker. R. D. and R. C. Morey (1986b): The use of categorical variables in data envelopment analysis,
Management Science 32: 1613-27.
Charnes, A., W. Cooper, A. Y. Lewin and L. M. Seiford (1997): Data Envelopment Analysis: Theory,
Methodology and Applications", 3 ed.
Fried, H. 0., C. A. K. Lovell and P. Vanden Eeckhaut (1993): Evaluating the performance of US
credit unions, Journal of Banking and Finance 17: 251-265.
Golany, B. and J. Roll (1993): Some Extensions of Techniques to Handle Non-Discretionary Factors
in Data Envelopment Analysis, Journal of Productivity Analysis 4: 419-432.
Golany, B. and J. Roll (1997): Restricted best practice selection in DEA: An overview with a case
study evaluating the socio-economic performance of nations, Annals of operations research 73:
117-140.
Greene, W. H. (1997): Econometric Analysis, New York.
Lovell, K. (1995): Linear programming approaches to the measurement of productive efficiency, Top
2: 174-248.
Ray, S. (1988): Data Envelopment Analysis, Nondiscretionary Inputs and Efficiency: An Alternative
Interpretation, Socio-Economic Planning Sciences 22: 167- 176.
Ray, S. (1991): Resource-Use Efficiency in Public Schools: A Study of Connecticut Data,
Management Science 37: 1620-28.
Ruggiero, J. (1996): On the measurement of technical efficiency in the public sector, European Journal
of Operations Research 90: 553- 565.
Seiford, L. M. (1996): Data Envelopment Analysis: The Evolution of the State of the Art (1978 -
1995), Journal of Productivity Analysis 7: 89-138.
Zhang, Y. and R. Bartels (1998): The Effect of Sample Size on Mean Efficiency in DEA with an
Application to Electricity Distribution in Australia, Sweden and New Zealand, Journal of
Productivity Analysis 9: 187- 204.
Non-Radial Efficiency as Semi-Radial Efficiency
Laurens Cherchye and Tom Van Puyenbroeck'
Abstract
Existing non-radial efficiency measures have focused on solving the slack problem associated
with the Debreu-Farrell notion of efficiency, sometimes at the expense of the underlying
economic intuition as regards their eventual projection. In this paper we take another
perspective. Building on the idea that any reference point can be obtained using a semi-radial
method, we start by imposing some (desirable) properties on the reference. Specifically, we
require the projection to be as close as possible to the radial one while still belonging to the
efficient subset of the production technology. In a second step efficiency scores are then
computed by reference to the obtained projection. These scores can be decomposed in a
Debreu-Farrell component and a factor that captures the divergence between reference and
evaluated input-output mixes. This second component can also be interpreted as measuring
the extent to which the radial projection deviates from the closest Pareto-Koopmans efficient
feasible point. This new way of measuring non-radial efficiency allows to maintain the
attractive interpretation of radial measures while avoiding the slack problem.
1 Centre for Economic Studies, Catholic University of Leuven, Naamsestraat 69, B-3000 Leuven, Belgium
52 Non-Radial Efficiency as Semi-Radial Efficiency
Structure
Introduction
2 Decomposing non-radial efficiency
3 Towards some desirable properties of the projection direction
4 Conclusion
References
Non-Radial Efficiency as Semi-Radial Efficiency 53
1 Introduction
The purpose of the two benchmark DEA models of Chames, Cooper and Rhodes
[CCR (1978, 1979)] and Banker, Chames and Cooper [BCC (1984)] was to put into
practice the notion of efficiency as expressed in the earlier work of Debreu ( 1951) and
Farrell ( 1957). As the inverse of Shephard's distance function the associated "Debreu
Farrell" measure is closely linked to the microeconomic theory of production.
Moreover, this measure has a straightforward cost interpretation as it can be written as
the ratio of reference to actual costs (input orientation) or actual to reference revenues
(output interpretation), independently of the price vector that is used [see Russell
(1985)].
These convenient properties follow from the radial projection of inefficient decision
making units (DMUs) on the reference frontier. It is well known that the
equiproportionate nature of comparisons has a drawback however since it implies that
Debreu-Farrell efficiency does not necessarily coincide with the more general Pareto
Koopmans efficiency concept [as introduced by Koopmans (1951)]. This is the so
called slack problem: "efficient" projections can sometimes increase their technical
efficiency by a further non-radial change of some input or output dimensions. This
problem is likely to increase when the number of dimensions becomes big relative to
the number of observations. It is even more pronounced when the proportionality
(CCR) or convexity (BCC) assumption is dropped, so that only the deterministic and
free disposal postulates are maintained like in the increasingly popular Free Disposal
Hull (FDH) model [see Tulkens (1993)].
In some instances the slack problem can influence the efficiency scores considerably,
which in tum could induce wrong management conclusions. In such cases it seems
more appropriate to call for measures that ensure inefficient DMUs are compared to
Pareto-Koopmans undominated references by projecting in a non-radial way. As their
name suggests, these non-radial measures essentially differ from their radial
counterparts in that the input (output) mix of the input (output) efficient projection may
deviate from the mix of the inefficient observation. However, this feature is as such not
addressed in well-known non-radial variants such as the Russell measure, introduced
by Hire and Lovell ( 1978), and the additive measure of Chames et a!. (1985) and many
54 Non-Radial Efficiency as Semi-Radial Efficiency
of their successors. An explicit analysis of such mix deviations relative to the radial
projection may be rewarding due to the latter's attractive interpretation.
In this paper we present a measure which aims at preserving the nice economic
intuition of the radial efficiency estimates while still computing efficiency with respect
to a Pareto-Koopmans efficient projection. In doing so, we exploit the characteristic
that any non-radial measure can be re-interpreted as a semi-radial efficiency measure.
The associated efficiency scores can thus be decomposed in an appealing way.
Specifically, they are the product of a classical Debreu-Farrell measure and a factor
that captures differences in input-output mixes of an evaluated DMU and its reference.
These latter projection points are determined starting from some axiomatic properties.
Essentially, we require that the reference should be "as close as possible" to the radial
projection in terms of mix deviation. In a second step then, we calculate efficiency
scores using the procedure outlined in an earlier paper [Cherchye and Van
Puyenbroeck (1998)].
The plan of the paper is as follows. In section 2 we shortly repeat the procedure to
compute the efficiency scores. In section 3 we propose a set of desirable properties for
an efficient projection, and show how such a projection can be obtained in a DEA
framework. Concluding remarks follow in section 4.
2 Decomposing non-radial efficiency
In Cherchye and Van Puyenbroeck (1998) we introduced an (input and output
oriented) decomposable measure which can be computed as soon as one disposes of a
reference point for the DMU under study. For simplicity, we will mainly concentrate
on its input orientation, but the intuition straightforwardly extends to the output
orientation. Throughout the paper we consider a general situation where N
observations use an m-dimensional semi-positive input vector x = (xh ... , xm) to
produce an s-dimensional semi-positive output vector y = (y1, ... , y,).
Assume for the moment that an input vector x is projected on an "arbitrary" reference
xR. We propose to compute the input efficiency score for x as follows:
Non-Radial Efficiency as Semi-Radial Efficiency 55
(1)
In Cherchye and Van Puyenbroeck (1998) it was shown how the input measure (1) can
be decomposed in an intuitive way by exploiting its semi-radial nature, as will now be
recaptured briefly. In figure 1 we present an input vector x. (=(7,4)) with xb (=(2,3)) as
the reference vectors. The vectors differ in two dimensions. First, there is the obvious
fact that xb Pareto-Koopmans dominates x •. Second, the input mixes of x. and xb differ.
Indeed, the proportion x/x1 for x. is 4/7 whereas it amounts to 3/2 for xb. The same can
also be expressed by means of the vectors v. = (1, 4/7) and vb = (1, 3/2), where the first
input is taken as the numeraire.
Applying (1) gives an efficiency score of 44.72% for x •. This score can now be
decomposed in two factors. To see this we first construct the lines perpendicular on the
radials through x. and xb. These lines take the form fi = v/ X; (i=a,b). Both
perpendiculars are depicted in figure 1, together with a line through x. which is parallel
to the perpendicular through xb. This allows us to identify xc, which is the closest
projection of x. on the radial through xb. In fact xc corrects for the deviation between v.
and vb. In order to get from x. to xb one first needs to adjust the input-mix (x. ~ xc), so
as to apply an equiproportionate reduction of all inputs (xc ~ xb) afterwards. The
inefficiency of x. thus results from (i) a deviation from the reference input mix and (ii)
a cost level which is too high, even after adjustment for the mix deviation.
Figure 1: decomposing non-radial efficiency
8
6
2
· ,~26/2)~x, + (312)x,
',
''-...,X
',,, :8 .. · '',(13/2)~x, +(3/2)x,
0 J_~~--·_· ---+------+----'"'",-'-'>..--------;
0 4 Input x1
6 8
56 Non-Radial Efficiency as Semi-Radial Efficiency
The cosine of the angle between xb (or xc) and X a constitutes a measure for (i). The ratio
of the cost level of xb over that of xc can be used to estimate (ii). Summarising, measure
(1) is equivalent to:
cost level x b . = x cosO, Wtth 8 the angle between X a and Xb (or Xc)
cost level x c
T T = vb xb x Xa xb
v/ X a llxa II x llxb II = 0.5 X 0.8944
= 44.72%
The decomposition of measure (1) thus yields two factors. The first one captures
radial inefficiency due to a higher cost level after adjustment for differences in input
proportions (Debreu-Farrell component (DF)). The second factor estimates the
inefficiency due to the deviation between the actual and the reference input mix (mix
correspondence component (MC)). Its value will equal one if both the evaluated and
reference input vectors lie on the same radial, and will be smaller than one otherwise.
In the general case an input vector x will be projected on a reference xR. Denoting the
vectors which define the perpendiculars on the radials through x and xR as respectively
v and vR, we get:
· . llxRII v/ xR xr xR mput effictency score= Txif = v R r x x llxll x llx R II (2)
An analogous formulation holds for the slack based output efficiency measure:2
T T . UR Y Y YR
output effictency score= u R r y R x IIYII x IIY R II (3)
2 One cannot take ~~~;111 as a measure for output efficiency as this ratio can be decomposed in the
following way: ~~;1( ::::, x[li( 11J:t~.IIJl Deviations from the reference output proportions would
thus be rewarded, which is clearly undesirable.
Non-Radial Efficiency as Semi-Radial Efficiency 57
where y and YR denote the actual and reference output vector, respectively. The vectors
u and uR define the perpendiculars to the radials. The interpretation of the output
measure (3) follows immediately.
Two additional comments are in order here. First, note that both the input and output
measures allow for zero entries in the data set. Second, the above decomposition was
conducted for an "arbitrary" reference input- or outputvector. By this we mean that the
reference could have been obtained in various ways, e.g. as an exogenously given
"most prefered target", after applying an additive DEA-model, following a Fare-Lovell
procedure, etc. It can also be computed for radial projections, which would however
be a rather trivial exercise given that the MC component in such a setting is evidently
equal to 1. In all cases, it is fundamentally the semi-radial character of the projection
that allows to resurrect (mix-corrected) Debreu-Farrell efficiency measures and their
associated economic interpretation.
3 Towards some desirable properties of the projection direction
Using this last basic concept of semi-radiality, one can actually proceed in the opposite
way. In particular, by imposing restrictions on the nature of the mix correspondences
one may narrow down the set of allowable non-radial projections. In this section we
provide a simple application of such an approach that eventually leads to a novel non
radial measure. It must be noted at the outset that this way of proceeding contrasts
with the usual approach in the literature, where the focus has mainly been on axiomatic
properties of non-radial measures instead of the reference points themselves. This of
course resulted in nicely behaving measures, but the economic intuition associated
with the efficient projections is not always obvious. Here we take the other perspective
and start from some desirable characteristics of the reference projection.
One such a characteristic is obviously that it should belong to the Pareto-Koopmans
(input or output) efficient subset of the reference technology (C.l ). A second property
we propose is related to the natural theoretical and empirical benchmark status of
Debreu-Farrell efficiency measures. Specifically, it is considered desirable that the
direction of projecting is the closest possible to the radial one (with the difference
measured by the cosine of the angle between actual and reference vectors), so as to
58 Non-Radial Efficiency as Semi-Radial Efficiency
maintain the attractive interpretation of radial measures as much as possible (C.2).
Intuititively, we thus seek for a reference point that is as near as possible to the
Debreu-Farrel reference while in any case being freed of the slack problem.
For ease of exposition we again concentrate on the input orientation, but extensions to
the output oriented projection are straightforward. Consider an input-output
combination (X0 J10 ). Let L(y.) denote the input correspondence that assigns the subset
of all input vectors x which allow to produce the output vector y/ Further define Eff
L(y.) as the efficient subset of L(y.): 4
The following conditions should then be met by a reference vector xR for X 0 :
C.l XR E Ef!L(y.)
In fact, C.l becomes redundant in view of C.2. We choose to state both properties
separately however to illustrate the difference between the projection introduced here
and other non-radial projections in the literature which, as far as the reference point is
concerned, exclusively focus on C.l. The reference projection satisfying C.2 can be
identified in two steps: one first locates the efficient subset of L(y) and then looks for
its cosine minimising element in a second step. Once the reference is identified, one
can of course compute an efficiency score using (2).
In a DEA context specific assumptions are made about L(y.). For the constant returns
to scale (CRS) model with free disposability of both inputs and outputs (FD) it is
defined in the following way (fork= l, ... ,s; I= l, ... ,m and}= l, ... ,N):
3 See e.g. Fiire, Grosskopf and Lovell (1985) for a discussion of L(yj in a general setting. 4 We use the symbol<=(>=) to indicate "smaller than or equal to", whiles(<::) means<=(>=) and*·
Non-Radial Efficiency as Semi-Radial Efficiency 59
When the assumption of constant returns to scale is relaxed to variable returns to scale
(VRS), we get:
In some instances the convexity assumption is hardly defensible. Then it seems
advisable to use an FDH reference technology, which differs from the previous model
in that the intensity vector e can only contain zeros or ones [see Tulkens (1993)]:
To obtain a projection under EJJ L(yJFDH is relatively easy. One only needs to identify
those observations that are not Pareto-Koopmans dominated by another observed
input-output combination (applying e.g. the additive FDH model [see Bardhan et a!.
(1996)]). The reference that satisfies C.2 is then the one corresponding to the smallest
angle with x •. 5
The algorithm to be applied for L(yJFD-cRs and L(yJFD-YRS is analogous. The main
difference is that efficient subsets do no longer consist of individual points but instead
of efficient facets constituting the corresponding convex polyhedrons. Algorithms and
software codes to identify these facets have been developed in the domain of multiple
objective linear programming (one can e.g. use the ADBASE code [Steuer (1995)] or
one of its variants [e.g. Armand and Malivert (1991) or Strijbosch eta!. (1991)]).6 Each
facet F can be represented as the set of all convex combinations of P efficient extreme
points {(xEI> Y£1), ... ,(xEP• YEP)} defining it. The associated set IF of input values can
thus be reconstructed as follows (for I =1, ... , m and}= 1, ... , P):
' Note that for the FDH reference technology the efficient subset is normally discontinuous. In certain cases this may lead to a reference observation identified under C.2 that does not Pareto-Koopmans dominate (x.,. y.). If one wishes that the cosine minimising reference is not only Pareto-Koopmans efficient but also guaranteed to be Pareto-Koopmans dominant, this is accomplished by adding to C.2 the following (FDH-specific) condition: xR <= x •.
'See Steuer (1986, 1994) for an overview of algorithms and software codes.
60 Non-Radial Efficiency as Semi-Radial Efficiency
For each F one has to identify that element of IF which maximises the cosine with X0 •
This boils down to the following mathematical programming problem:
(4) subject to: N z>j =1 j=l
aj >= 0
It can easily be checked that the objective of (4) is concave with respect to the
elements of a, implying that a maximum will always be obtained. Problem (4) can thus
be solved by means of simple Lagrangian techniques.
Summarising, for L(y,)FD-cRs and L(yJFD-vRs the projection satisfying C.2 can be found
following a three-step procedure: (i) identify the facets that together constitute the
efficient subset, (ii) solve problem (4) for each of these facets, (iii) look for the
minimum of the corresponding objective function values. Of course, these three steps
need only be executed when slacks are associated with radial projections. Moreover,
when locating the efficient facets it suffices to consider only those observations that
are found Pareto-Koopmans efficient by the constant or variable returns to scale
additive DEA models ofChames et al. (1985).7
To illustrate the new measure, consider the example presented in figure 2. For
simplicity it is assumed that all six DMUs produce the same output level. Table 1
shows the scores (and their MC and DF components) for the inefficient observations d,
7 Sometimes it is computationally more efficient to proceed by identifying all efficient facets of the production possibility set following e.g. the algorithm of Yu et al. (1996). This procedure is straightforwardly analogous to the above. Of course, for an inefficient observation it then suffices to consider only those efficient facets that are constructed from at least one vertex that ParetoKoopmans dominates it.
Non-Radial Efficiency as Semi-Radial Efficiency 61
e andf We first consider the FD-CRS and FD-VRS technologies in which cases the
reference frontier is represented by the full line. Obviously there is a slack problem
associated with the radial projection for observation f Applying the above procedure
yields c as the appropriate reference under C.2. As f is not "naturally enveloped" the
MC score is below one.8 Note that the DF score off coincides with that of e. In fact, e
is the closest projection off on the radial through c. When an FDH technology is
assumed, the slack problem is usually more pronounced. This is also illustrated here
(see the dotted line). Now d also has to adjust its input mix if it is to become a member
of the Pareto-Koopmans efficient subset. Correspondingly, its MC score is below 1.
The needed input mix adjustment, as captured by the cosine of the angle between a and
d, is even greater than for f
Figure 2: A hypothetical example
....
121 10 t
8
;; 6 (1.7
! 4 ~
Table 1: Efficiency results
DMU CCRIBCC score MC
D 0.586 1.000
E 0.625 1.000
F 0.623 0.997
-· d (3,/0)
2 4
DF 0.586
0.625
0.625
-c-(6,3)
6 8
Input I
FDH
score
0.677
0.625
0.623
_ e e (9.6. 4.8) -• f(8,6) ----
10 12
MC DF 0.989 0.685
1.000 0.625
0.997 0.625
8 The term "not naturally enveloped" is taken from Bessent et al. (1988) which use it to indicate inefficient units that have a mix of inputs or outputs different from any point belonging to the efficient subset.
62 Non-Radial Efficiency as Semi-Radial Efficiency
4 Conclusion
While most non-radial proposals allow to solve the slack problem associated with
radial efficiency measurement, they do not possess an immediate interpretation like the
prevalent Debreu-Farrell measures. The literature has mainly concentrated on the
axiomatic properties of the non-radial measure itself rather than on any economic
intuition associated with the eventual projection.
In this paper we built on the semi-radial re-interpretation of non-radial measures to
change the perspective: an axiomatisation of the (non-radial) direction of projection is
proposed which ensures it to be as close as possible to the radial one while always
selecting a Pareto-Koopmans undominated reference. Consequently, the ensuing
projection will equal the equiproportionate one as long as the latter belongs to the
efficient subset. In the other case it will minimise adjustments in the input-output
structure. Once the "best" reference is obtained, the decomposable measure we
introduced in an earlier paper can be used to obtain nicely interpretable efficiency
scores. Specifically, it enables to split up total inefficiency in a Debreu-Farrell and a
mix correspondence component.
The obtained efficiency results capture useful additional information, especially when
radial efficiency measurement is associated with slacks. The mix correspondence score
then gives a measure for the least extent to which the input (output) proportions need
adjustment in order for the evaluated observation to become "naturally enveloped".
One could also consider it as a composite relative measure for the slacks associated
with radial projection. The Debreu-Farrell component, on the other hand, estimates the
residual inefficiency after correcting for the input (output) mix deviations.
The analysis presented above can be extended in some ways. For example, other
desirable properties for reference points or mix-deviation factors can be imposed to
provide a characterisation of other (existing as well as novel) efficiency measures. In
any case, it seems worthwile to recognize what we have subsumed here under the
general header of semi-radial efficiency measurement. As DEA problems are mostly
formulated in terms of multiple inputs and outputs, the possible divergence between
reference and actual input or output proportions at least potentially becomes a relevant
dimension of the problem setting. Inefficiency does then no longer only depend on a
ratio between minimum (maximum) to actual input (output), but also on the proportion
Non-Radial Efficiency as Semi-Radial Efficiency 63
between the several inputs (outputs). In this sense, the particular projection procedure
and associated measure introduced here may constitute a valuable alternative when
slacks indeed appear systematically if efficiency is estimated in a radial way.
References
Armand, P. and C. Malivert (1991): Determination of the Efficient Set in Multiple Objective Linear
programming, Journal of Optimization Theory and Applications, 70 (3), p. 467-490.
Banker, R.D., Charnes, A. and W.W. Cooper {1984): Some Models for Estimating Technical and
Scale Inefficiencies in Data Envelopment Analysis, Management Science 30 (9), p. I 078-1092.
Bessent, A., Bessent, W., Clark, T. and J. Elam (1988): Efficiency Frontier Estimation by Constrained
Facet Analysis, Operations Research 36, p. 785-796.
Bardhan, 1., Bowlin, W.F., Cooper, W.W. and T. Sueyoshi {1996): Models and Measures for
Efficiency Dominance in DEA: Part I: Additive Models and MED Measures, Journal of
Operations Research Society of Japan 39, p. 322-332.
Charnes, A., Cooper, W.W. and E. Rhodes {1978): Measuring the Efficiency of Decision Making
Units, European Journal of Operational Research 2, p. 429-444.
Charnes, A., Cooper, W.W. and E. Rhodes {1979): Measuring the Efficiency of Decision Making
Units, European Journal of Operational Research 3, p. 239.
Charnes. A., Cooper, W.W., Golany, B., Seiford, L. and J. Stutz (1985): Foundations of Data
Envelopment Analysis for Pareto-Koopmans Efficient Empirical Production Functions, Journal
of Econometrics 30, p. 91-107.
Cherchye, L. and T. van Puyenbroeck {1998): Learning from Input-Output Mixes in DEA: A
Proportional Measure for Slack-Based Efficient Projections, Managerial and Decision
Economics, forthcoming.
Debreu, G. {1951): The Coefficient of Resource Utilisation, Econometrica 19(3), p. 273-292.
Fare, R., Grosskopf, S. and C.A.K. Lovell {I 985): The Measurement of Efficiency of Production,
Kluwer-NijhoffPublishing, Boston.
Fare, R. and C.A.K. Lovell (1978): Measuring the Technical Efficiency of Production, Journal of
Economic Theory 19 (I), p.J50-162.
Farrell, M.J. (1957): The Measurement of Productive Efficiency, Journal of the Royal Statistical
Society Series A 120 (3), p 253-281.
64 Non-Radial Efficiency as Semi-Radial Efficiency
Koopmans, T. (1951): Analysis of Production as an Efficient Combination of Activities, in:
Koopmans, T. (ed.), Activity Analysis of Production and Allocation: Proceedings of a
Conference, Yale University Press, New Haven, p. 33-97.
Russell, R. (1985): Measures ofTechnical Efficiency, Journal of Economic Theory 35(1), p. 109-126.
Steuer, R.E. (1986): Multiple Criteria Optimization: Theory, Computation and Application, John
Wiley, New York.
Steuer, R.E. (1994): Random Problem Generation and the Computation of Efficient Extreme Points,
in: Multiple Objective Linear Programming, Computational Optimization and Applications 3,
p. 333-347.
Steuer, R.E. (1995): Manual for the ADBASE Multiple Objective Linear Programming Package,
Faculty of Management Science, University of Georgia, Athens.
Strijbosch, L.W.G., van Doorne, A.G.M. and W.J. Selen (1991): A Simplified MOLP Algorithm: The
MOLP-S Procedure, Computers and Operations research 18 (8), 709-716.
Tulkens, H. (1993): On FDH Efficiency Analysis: Some Methodological Issues and Applications to
Retail Banking, Courts and Urban Transit, Journal of Productivity Analysis 4, p. 183-210.
Yu, G., Wei, Q., Brockett, P. and L. Zhou (1996): Construction of All DEA Efficient Surfaces of the
Production Possibility Set under the Generalised Data Envelopment Analysis Model, European
Journal of Operational Research 95, p. 491-510.
Continuity of the BCC Efficiency Measure
Holger Scheel 1
Abstract
Continuity is a desirable property of an efficiency measure. It ensures that small data errors
cause only small errors in the efficiency measure. In this paper continuity properties of the
BCC efficiency measure are studied. Examples are given where this measure 'jumps" under
arbitrary small data perturbations. However, it is shown that under weak assumptions it
depends continuously on the input output data. Implications to the stability of efficiency
classifications are discussed, and the results are illustrated by an empirical example.
1 Operations Research und Wirtschaftsinforrnatik, Universitlit Dortmund, D-44221 Dortmund, Germany
66 Continuity of the BCC Efficiency Measure
Structure
Introduction
2 Discontinuities of the BCC measure
3 Stability of the BCC measure
4 An illustrative example
Appendix
References
Continuity of the BCC Efficiency Measure 67
1 Introduction
Data Envelopment Analysis (DEA) has become a standard approach to efficiency
measurement. It is a family of methods for the evaluation of relative efficiency of
"decision making units" (DMUs). We refer the interested reader to Seiford (1996) for a
recent survey. The classical DEA models developed by Chames et a!. (1978) and
Banker eta!. (1984) classify each DMU to be evaluated as "efficient" or "inefficient".
Moreover, for each inefficient DMU a measure is computed which indicates the
proportional input reduction which is necessary, ceteris paribus, to change its
classification from inefficient to (radially) efficient.
Let each DMU k in the set of n DMUs be characterized by its input output data
collected in the row vector (X",r*), where we suppose all entries to be nonnegative and
at least one input and one output to be positive. Let (X,Y) denote the matrix of input
output data, where each row represents one DMU and each column represents one
input or output. We assume that each column contains at least one positive element.
Following Banker eta!. (1984) we assume that the production possibility set exhibits
variable returns to scale and is thus given by
T(X,Y)= {(x,y) I ..17X :s; x, ..17Y ;e: y, ..1.Te =1, A ;e: 0},
where e = ( 1 , ... , 1 )r in appropriate dimension. Recall that DMU k or, synonymously,
the input output vector (X,Y') is called (radially) input-efficient in T if it is impossible
to reduce inputs proportionally without reducing any output, i. e. if there is no B < 1
such that (BX",Y') E T. The BCC efficiency measure hk(X,Y) ofDMU k is defined as
the optimal value of the linear program
minimize e
s.t. AT X :::; ex' ATY "= Y' ATe 1
(1)
A "= 0,
cf. Banker et a!. (1984). Notice that by defining the BCC measure hk in this way we
focus on radial efficiency measurement, i. e. we ignore "mix inefficiency" which is
often included in DEA measures by a non-Archimedian component.
68 Continuity of the BCC Efficiency Measure
In economic theory an efficiency measure is supposed to satisfy certain conditions.
Weak versions of the axioms introduced2 by Fare and Lovell ( 1978) can be formulated
as follows:
(I) Indication of efficiency:
hiX. Y) = 1 if and only ifDMU k is (radially) input-efficient.
(H) Homogeneity:
hix",Y) = ~ hk(X,Y) for every e with ( eX',Y") E T(X,Y). [x" denotes the matrix X
where row X' is replaced by eX'.]
(M) Monotony:
If X'::;; x and Y" ~ y then hiX,Y) ;?: hk( X, Y ). [(X, Y) denotes the matrix (X,Y)
where row (X,Y") is replaced by (x, y ).]
Notice that the homogeneity property (H) provides a key for interpreting efficiency
measures. In fact, ifDMU k is assigned an efficiency measure of hiX,Y) = e< 1, then
a proportional reduction of inputs to ex suffices to become efficient, i.e. hix",Y) = 1.
Hence an efficiency measure satisfying (H) may be viewed as a stability measure
which indicates the feasible input change preserving the efficiency status. It is easily
verified that the BCC measure satisfies (I), (H) and (M).
Russell ( 1990) introduced "continuity" into the above axiom system.
(C) Continuity:
hiX,Y) depends continuously on the data matrix (X,Y).
This property, defined formally below, is desirable in particular from a practitioner's
perspective: It ensures that small perturbations of data cause only small changes in the
efficiency measure. Russell (1990) formulated conditions for general production
possibility sets that ensure continuity of radial efficiency measures. Recently Kuntz
2 Fiire and Lovell ( 1978) proposed four axioms. However, Russell ( 1985) showed that their property "E.2" is implied by the other three conditions.
Continuity of the BCC Efficiency Measure 69
and Scholtes (1996) and Scheel and Scholtes (1998) studied continuity of radial
measures assuming a constant returns to scale technology. Following the route taken in
the latter paper we study in the sequel the continuity property (C) of the BCC measure
which is based on the variable returns to scale technology T.
Motivating our results we present in the next section two examples where the BCC
measure violates (C). Section 3 contains the main continuity results. Moreover, some
implications to the stability of efficiency classifications are discussed. In the final
section the results are summarized by means of an illustrative empirical example.
2 Discontinuities of the BCC measure
Whenever in applications an input or output is measured on a continuous scale then
inaccuracies in the data matrix (X,Y) cannot be avoided since a measurement tolerance
must be chosen a priori. It is possible to choose an arbitrary small measurement
tolerance, but it is impossible to choose a zero tolerance. Since higher precision usually
induces higher measurement cost, it is desirable that improved data quality yields to a
better estimation of the "true" efficiency measure, i.e. if a sequence of data matrices
(X,,Y1) converges by increasing measurement precision to the "true" data (X,Y), then the
corresponding sequence of BCC measures hk(X,,Y,) should converge to the "true"
measure hiX,Y). This continuity property, however, may not be satisfied by the BCC
measure.
Notice first that zero data may cause discontinuities of hk as can be illustrated by the
following example adapted from Scheel and Scholtes ( 1998). Consider a data matrix
for two inputs and one output given by
The BCC measure of DMU 1 is h1(X,,Y1) = 1 for every t e (0, 1] and h1(X,,Y,) = 0.75
for t = 0. Assuming that t = 0 is the "true" value then even choosing an arbitrary high
measurement precision cannot ensure that the true efficiency measure can be
computed. Moreover, DMU 1 may be classified falsely as efficient.
70 Continuity of the BCC Efficiency Measure
This example may suggest that - like in the constant returns to scale case studied in
Scheel and Scholtes (1998) - discontinuities of the efficiency measure are always
caused by zeroes in the data matrix. However, the BCC measure h* can jump although
the data matrix is strictly positive.
y
40
2 3 30
20
10
10 20 40 50 X
Figure 1: Discontinuity with positive data matrix
Consider the production possibility set in Figure 1, where the underlying data matrix is
strictly positive. The efficiency measure of DMU 3 is h3 = 0.4 but jumps to unity after
an arbitrary small increase of output Y 3•
Notice that continuity is not only desirable in the presence of measurement errors but
also in view of the reliability of measures computed by linear programming software.
3 Stability of the BCC measure
Continuity means that small data perturbations cause only small changes in efficiency
measures. In fact we shall require that these changes are bounded by the size of
perturbations. Following Scheel and Scholtes (1998), h* is called continuous at (X,Y)
with respect to certain admissible perturbations if there exists a positive scalar y such
Continuity of the BCC Efficiency Measure 71
that lhiX.Y) - hk (X',Y11 ~ r II(X,Y)-(X',Y)II for every data matrix (X',Y1 in a
neighborhood of(X,Y) which is an admissible perturbation.
As indicated by the example above, zeroes in the data matrix may play an important
role for the continuity of efficiency measures. It can be shown that the BCC measure
may jump if an inefficient DMU has a zero input. In fact, if hk(X, Y) < 1 and X contains
a zero entry X/ then by adding for every DMU I* k an arbitrary small positive number
to X,' we obtain A-1 = 0 for I * k and A.k = 1 for every feasible A. in the linear program (I),
whence hk jumps to 1.
Remark 1 The BCC measure of an inefficient DMU with a zero input is not continuous
if perturbations in the corresponding input column of the data matrix are defined as
admissible.
Hence, if a DMU with zero inputs appears inefficient, one has to look carefully at the
data quality before using this classification. However, in most applications zeroes in
the input data occur only if these inputs are not used by this DMU at all. This means
that the zero values do not suffer from a positive measurement tolerance, i.e. one can
expect that a zero value is exactly zero. For a useful notion of stability it is thus not
necessary to consider arbitrary data perturbations but only perturbations of the
nonvanishing data. Following Scheel and Scholtes (1998), we call an efficiency
measure stable if it is continuous with respect to perturbations of non vanishing entries
of the data matrix.
In order to derive conditions which ensure stability of the BCC measure recall that a
DMU is called radially output-efficient if it is impossible to increase all outputs
proportionally, i.e. the output oriented BCC measure which is defined as the optimal
value of the linear program
maximize¢
s.t. ATX ~ x* ATY ~ ¢Y* Are I
(2)
A ~ 0
is unity. Notice that in Figure 1 DMU 3 is radially output-efficient but not radially
input-efficient. The following theorem shows that indeed this is a necessary and
72 Continuity of the BCC Efficiency Measure
sufficient condition for instability of the BCC measure. The proof can be found in the
appendix.
Theorem 1 hk is stable at (X,Y) if and only if output-efficiency of DMU k implies input
efficiency.
A violation of the assumption of the theorem can be detected by computing both the
input oriented and output oriented BCC measure. Notice that if the assumption of
Theorem 1 is violated then there exists a solution for program (2) where all input
slacks are positive. However, since the condition "output-efficiency implies input
efficiency" is satisfied for almost every data matrix3, one can expect in practical cases
the BCC measure to be stable.
Efficiency measures which satisfy the indication property (I) measure inefficiency
rather than efficiency since they assign unity to every efficient DMU. Replacing (I) by
the requirement that an efficiency measure greater than or equal to unity indicates
efficiency, Andersen and Petersen (1993) introduced an extension to DEA measures
which allows the ranking of efficient DMUs. We will later use this extension to obtain
stability results for the classification of DMUs as efficient.
The extended BCC measure, also called "superefficiency measure", can be defined4
by
Hk(X. Y) =sup{ e I (eX', Y*) is input-efficient in T(X,Y)},
where X denotes the input matrix X with row X replaced by eX'. The measure Hk
yields for every data matrix (X,Y) an efficiency measure in the interval (O,oo], where
infinity means that efficiency is preserved under arbitrary proportional increase of
inputs.
' This means that the condition is satisfied for all matrices in an open and dense subset of data matrices. It can be shown that {(X,Y) I H,(X,Y) * I, G,(X,Y) * I} is an appropriate subset with H, given by (3) and the corresponding output oriented extended BCC measure G, given by (2) where the k-th row of (X,Y) is removed.
4 In contrast to Andersen and Petersen (1993) we include this definition explicitly, since some authors argued that their measure may not exist in some cases where the corresponding linear program (3) is infeasible, cf. e.g. Wilson (1995).
Contmuity of the BCC Efficiency Measure 73
Note that for inefficient DMUs H* coincides with the BCC measure hk> whereas for
efficient DMUs H* '2 hk = 1. Indeed, if the efficiency measure is Hk >1 then DMU k
turns inefficient if it increases its inputs more than to H* X'. Clearly Hk coincides with
the optimal value of the linear program
minimize B
s.t. x-• A s BX* y-• A '2 y* Are
A '2 0,
(3)
where (X- *, Y- k) is obtained by deleting row k of (X, Y) and the optimal value is set to
infinity if the feasible set is empty.
A sufficient condition for stability of the extended BCC measure has to be stronger
than for the standard BCC measure. This can be seen from Figure 1 where DMU 2 is
output-efficient and input-efficient, hence the assumption of Theorem 1 holds.
However, when decreasing output y3 of DMU3 the efficiency measure H2 jumps from
2.5 to infinity. Note that such jumps can only occur if Hk '2 1, since for inefficient
DMUs the values coincide with the BCC measure which is continuous provided the
assumption of Theorem 1 holds.
We shall study now the "direction" of possible jumps of BCC measures. In the
examples given before, discontinuities never appear as downward steps. The following
theorem shows that such steps are indeed impossible.
Theorem 2 Hkcanjump only upwards.5
The theorem means that efficient DMUs either remain efficient under small data
perturbations or tum inefficient with a continuous decrease of their efficiency measure.
Moreover it follows that if the assumption of Theorem 1 does not hold then jumps of
the standard BCC measure hk are always upwards. These results have useful
implications for the stability of the DMUs' efficiency status which we shall discuss
now.
' A formal statement of the theorem can be found in the Appendix.
74 Continuity of the BCC Efficiency Measure
Given an input output data matrix, the set
E(X,Y) := {k I DMU k is input-efficient in T(X,Y)}
is called efficiency classification. Since the simple efficiency status of a DMU attracts
much interest in practical efficiency valuations it is desirable that this classification is
stable, i.e. if (X',Y1 is a data matrix in a neighborhood of (X,Y) where only
non vanishing entries are perturbed then the equation
E(X,Y) = E(X',Y')
holds. A stable efficiency classification ensures that efficient DMUs remain efficient
and inefficient DMUs remain inefficient under small perturbations of nonvanishing
data.
In the spirit of DEA as a "fair" methodology for efficiency measurement, cf. Epstein
and Henderson (1989), a DMU should be classified as inefficient only if a positive
minimal distance to the efficient frontier can be quantified. This means that every
inefficient DMU should remain inefficient under small data perturbations, i.e. E(X,Y)
;;;;< E(X', n for every perturbed matrix (X', Y1. In view of the theorems stated above this
inclusion holds if and only if none of the inefficient DMUs is output-efficient.
To ensure that the efficiency classification E is stable, the reverse implication E(X,Y) ~
E(X',Y1 must hold for every (X', Y1 near (X,Y) as well. This inclusion6 means that
efficient DMUs remain efficient under small data perturbations. It can easily be shown
that a DMU k remains efficient under data perturbations if and only if HiX. Y) > I,
where the "if' follows from Theorem 2. To see the "only if', assume HiX,Y) = I.
Then any proportional increase of the input data X' will decrease () in the first
constraint of program (3), and Hk which coincides with the optimal value of (3) will
decrease as well. Hence, DMU k turns inefficient.
Remark 2 The BCC efficiency classification E(X,Y) is stable if and only if HiX,Y) ;t. 1
for each DMU k and there is no output-efficient DMU which is input-inefficient.
' If this inclusion does not hold for the whole set E, there may be a subset cr(X, Y) of E which satisfies a(X,Y) c E(X',Y1 for every (X',Y1 near (X,Y), i.e. cr denotes the set of stable efficient DMUs. This classification is closely connected to the classification introduced by Charnes eta!. (1986). Following their notation, we have cr = E and E = E u E' u F.
Continuity of the BCC Efficiency Measure 75
4 An illustrative example
In order to illustrate the results of the previous section let us consider an empirical
application where the efficiency of 63 agencies of an insurance company with respect
to launching a new contract was evaluated. Table 1 contains the inputs and outputs
used for the evaluation.
Inputs No. of clients Type A No. of clients Type B
No. of clients Type C Potential new premiums (DM)
Outputs No. of new contracts Sum of new premiums (DM)
Table 1: Inputs and outputs of insurance agencies
The aim of the agencies is to sell as many contracts with as many premiums as
possible which is indicated by the two outputs.
The clients served by the agencies and possibly buy the new contract are classified into
several types which reflect the current insurance coverage, e.g. the total policy value
etc. The potential new premiums depend on the clients' current coverage, hence it is
included as an additional input. The data set contains some zeroes in the input "No. of
clients Type C". This means that there are no such clients, i.e. there is no measurement
tolerance for this input. As expected, each output-efficient DMU is input-efficient,
hence the BCC measure is stable. Statistics of the results7 can be found in Table 2.
No. ofDMUs MinH,
MaxH,
Mean H,
Std. deviation H,
Table 2: Results
Efficient II
1.029
Inefficient 52
0.292
0.976
0.651 0.209
7 The original data are available from the author on request.
76 Continuity of the BCC Efficiency Measure
The results show that all efficient DMUs achieve optimal values greater than unity,
whence each computed efficiency measure is stable. Moreover, we can conclude from
Remark 2 that the efficiency classification is stable as well.
To sum up the paper it can be stated that the BCC measure and the corresponding
efficiency classification are stable in most practical situations. The case where
discontinuities may occur can be detected by computing both the input oriented and
output oriented BCC measure. Additional information about the stability of the
efficiency classification can be obtained by the extended BCC measure.
Appendix:
Proofs of the theorems
Theorem 1 hk is stable at (X, Y) if and only if output-efficiency of DMU k implies input
efficiency.
Proof If there is an input i with X/= 0 then it follows from the first constraint that A-1 =
0 for every DMU l with X/> 0. We conclude that the optimal value of (1) is left
unaltered under small perturbations of nonvanishing data if we delete all these rows l
and input column i from the data matrix. Moreover, we can delete every output column
j with Y/ = 0 since the corresponding constraints in program ( 1) are redundant. Hence
we can assume without loss of generality that (.x*,Y") is strictly positive.
If DMU k is output-efficient but not input-efficient then we have max { ¢ I (X,¢ Y") E
T(X,Y)} = 1 and (X,Y") E T(X-k,Y-k) whence max{¢ I (X,¢ Y") E T(X-k,Y-k)} =I. For
every arbitrary small e> 0 it follows (X,(1+e)Y") it' T(x-k,y-k), thus for every feasible
solution of the corresponding program (1) with e > 0 we have Ak = I, i.e. the BCC
efficiency measure jumps to unity. This proves the "only if'.
To prove the "if'-part we consider two cases. First, suppose that there exists a
nonnegative vector A with Are = 1 and Y -k.A_ > Y". Note first that the solution set of
Contmuity of the BCC Efficiency Measure 77
program ( 1) is nonempty and bounded since (} is unique and A. is bounded by the
convexity constraint. In view of Robinson ( 1977), it remains to show Othat the
constraints of (1) are regular, i.e. there is a A.> 0 with A_Te = 1 such that all constraints
are strictly satisfied. By assumption there exists a nonnegative vector A. with A.re = 1
and y-k A.> Y". This inequality remains valid if we replace A. by (A.+ee)/(1 +ne) withe>
0 sufficiently small, therefore we assume A. > 0. Choosing (}sufficiently large the first
inequality is strictly satisfied as well.
Now we tum to the second case, i.e. for every nonnegative vector A. with A.re = 1 there
is an output} such that Yf"?.lj-kA.. It follows that max{¢ I (X,¢ Y") e T(X-k,Y-")} = 1,
i.e. DMU k is output-efficient and thus input-efficient by assumption. Now let (X,,Y,)
be a sequence of data matrices tending to (X. Y) and (A., f),) be a corresponding sequence
of solutions. Since the sequence is bounded we may consider convergent
subsequences. Assume that (A-,,8,) converges to some vector (A.,{}). Recall that the
solution set of ( 1) is nonempty and bounded whence we conclude applying Lemma 2
of Robinson (1977) that (A.,(}) is feasible for (1). Since DMU k is input-efficient the
optimal value of ( 1) is unity and thus (} = 1. This is true for every subsequence whence
for the complete sequence, i.e. hk is continuous.
Theorem 2 "Hk can jump only upwards". I.e., Hk is lower semicontinuous:
liminf HiX,,Y,) "?. Hk(X,Y). (X,.Y,)-+(X,Y)
Proof Starting from the dual program of ( 1) which is
maximize Y"q + m
s.t. X'p
Yq +w-Xp :s; 0
p,q ~ 0
the assertion follows analogously to Lemma 1 of Scheel and Scholtes ( 1998), setting m
=0.
78 Continuity of the BCC Efficiency Measure
References
Andersen, P. and N.C. Petersen (1993): A procedure for ranking efficient units in Data Envelopment
Analysis, Management Science 39, 1261-1264.
Banker, R. D., A. Charnes, and W. W. Cooper (1984): Some models for estimating technical and
scale inefficiencies in Data Envelopment Analysis, Management Science 30, 1078-1092.
Charnes, A., W. W. Cooper, and E. Rhodes (1978): Measuring the efficiency of decision making
units, European Journal of Operational Research 2, 429-444.
Charnes, A., W. W. Cooper, and R. M. Thrall (1986): Classifying and characterizing efficiencies and
inefficiencies in Data Envelopment Analysis, Operations Research Letters 5, 105-110.
Epstein, M. K. and J. C. Henderson (1989): Data Envelopment Analysis for managerial control and
diagnosis, Decision Sciences 20,90-119.
Fare, R. and C. A. K. Lovell (1978): Measuring the technical efficiency of production, Journal of
Economic Theory 19, 150-162.
Kuntz, L. and S. Scholtes (1996): Sensitivity of efficient technologies in Data Envelopment Analysis,
Technical report, University of Cambridge, Cambridge CB2 lPZ, England.
Robinson, S. M. (1977): A characterization of stability in linear programming, Operations Research
25, 435-447.
Russell, R. R. (1985): Measures of technical efficiency, Journal of Economic Theory 35, 109-126.
Russell, R. R. (1990): Continuity of measures of technical efficiency, Journal of Economic Theory
51,255-267.
Scheel, H. and S. Scholtes (1998): Stability ofDEA efficiency scores, Judge Institute of Management
Studies Working Paper Series 36/1998, University of Cambridge, Cambridge CB2 lPZ,
England.
Seiford, L. M. ( 1996): Data Envelopment Analysis: The evolution of the state of the art (1978-1995),
Journal of Productivity Analysis 7, 99-137.
Wilson, P. W. (1995): Detecting influential observations in Data Envelopment Analysis, The Journal
of Productivity Analysis 6, 27-45.
DEA Models via Goal Programming
Wenbin Liu and John Sharp'
Abstract
In this paper, we investigate the relationship between Data Envelopment Analysis (DEA) and
Multiple Criteria Decision Making Theory. We re-examine DEA models from a goal
programming perspective. It has been shown in this work that many known DEA models and
new ones, can be derived via this approach.
As an illustrative application of the approach, the effectiveness of some antidepressant
pharmacotherapies is examined using one of the DEA models derived in this work.
1 Canterbury Business School, University of Kent, Canterbury, CT2, 7PE, UK
80 DEA Models via Goal Programming
Structure
Introduction
2 Goal Programming And Input-Output Systems
3 Merit Functions - Measures of Performance
4 DEA Models with General Goals
5 An Application
References
DEA Models via Goal Programming 81
1 Introduction
Data Envelopment Analysis (DEA) has become a standard non-parametric approach to
productivity analysis, especially to relative efficiency analysis of Decision Making
Units (DMUs). Since the introduction of the first DEA model CCR in 1978, it has been
widely used in efficiency analysis of many business and industry evaluation
procedures. Excellent literature survey can be found in, for instance [7] and [ 15].
Many DEA models exist that are designed to cope with various situations, such as the
CCR model [4], the BCC model [3], Additive model [5], and Cone Ratio model [6]
among the most well-known DEA models. Most of the fundamental DEA models are
derived from economic efficiency theory including Debreu-Farrell efficiency, Pareto
Koopmans efficiency, and more general technical efficiency axiomatic approaches (see
[8], and [13]). These models have then been modified to handled more complicated
applications in various ways.
In this paper we intend to demonstrate that many DEA models, known or new, can be
derived directly from goal programming. We have no intention of exhausting every
possibility, rather to show the essential ideas of the approach, through a small
illustrative application. There seems to be no similar work in the literature though
some relations between the DEA and multiple criteria decision making have been
noticed (see, for instance [ 17]).
2 Goal Programming And Input-Output Systems
In this section we very briefly introduce some basic concepts in multiple criteria
decision making theory and goal programming. We then examine input-output
systems from a point of view of goal programming, since they are the essential
elements of a DEA model.
Goal programming is a different way of seeking "good" solutions in multiple criteria
decision making processes. In this approach, instead of optimising multiple objective
functions, we set up a group of goals to be achieved. It may be impossible to achieve
all these goals simultaneously. Goal programming (GP) provides a mathematical tool
to investigate whether these goals can be simultaneously achieved, if not, to find some
82 DEA Models via Goal Programming
compromise solutions. The details of conventional goal programming theory may be
found in, e.g., [10] and [II]. Here we are mainly concerned on whether these goals
have been set properly or can be further improved.
In the followings we very briefly recall some basic ideas in (GP). We first examine
goal setting.
To this end, it is very useful to recall an important concept in multiple criteria decision
making theory: preference. It can be viewed as an order relation, and is closely related
with goal setting.
Definition 2.1 Let Y be a set and let y,,y2 E Y. A preference or an order> on Y is a
subset of YxY denoted by {>}such that y, > y 2 iff (y,,y2 ) E {>} Similarly one can
define a preference or an order ::? , < and .5:
Normally we require that the preference is transitive, etc, see [10]- [11] for the details.
The most frequently used order m DEA IS Pareto preference. Let
X=(x,,x.),Y=(y.,,y.) e R•. Then m Pareto preference
X> Yiff X; ;::y; {i = 1,2,,n)and X"" Y. X <Yiff- X> - Y. In this paper, we assume the
Pareto preference is selected if not explicitly stated otherwise. However we have to
emphasise that other orders such as K-cone order and lexicographic order are also very
useful in DEA model building. The former can, for instance, lead to the well-known
Cone Ratio model ([6]) and the latter may let us build DEA models which are able to
express the preferences of the evaluators.
It is reasonable to assume that one has a preference selected before setting up goals.
Setting up goals for a particular application is by no means trivial, and has been
extensively studied in the literature, see, for instance, [10] and [11]. Here we only
examine some cases relevant to our later discussions. Let A = {..t,, ... , ..tJ' be the n
dimensional decision making variable and S be a constraint set of the variable. Assume
that we wish to maximise the quantity
where x; (i = l, .... ,n) are known constants. Instead of solving (a linear programming
problem)
DEA Models via Goal Programming
n
max"' X·A.. .<eS f I I
we may set the goal to find ..1. = (i!,, ... , ..tJ such that
i,x;A; ~g 0 , AES, I
83
where g 0 is referred to as goal level. This type of goal will be referred to as positive
response type (PRT), as a higher level of the goal is desirable. Similarly one can set up
the following goal:
±x;A; ~g0 , A ES. I
This type of goal will be referred to as negative response type (NRT).
Similarly we can set multiple goals. In this work we only examine the following two
basic types of goals (and their mixtures):
n n
(L x:A; ~ g? , ... , l:x;"A; ~ g~), A E S (PRT) I I
or
<±xi A;~ g? , ... ,i,x;" A;~ g~), A E S. (NRT) I I
It is clear that one has to select a preference before setting any goals. Again we
emphasise that there are other types of goals very useful in DEA theory.
2.1 Input-Output Systems
In the following we shall apply the basic ideas introduced above to examine input
output systems which are fundamental to DEA theory. In the discussion to follow, it is
assumed that there are n decision-making units to be evaluated. Each DMU consumes
varying amounts of m-different inputs to produce s-different outputs. Let the m-
84 DEA Models via Goal Programming
dimensional vector X 1 = (x11 ,x21 , ... ,x.,)' denote the inputs of DMUi, and the s
dimensional vector Y1 = (y11 ,y21 ,. .. ,y,)' denote the outputs of the DMUi. Let
X=(X,X2 , ... ,XJ be the input matrix and Y=(J';,Y2 ... ,Y,) be the output matrix. In
DEA, each of the DMUs is viewed as an input-output system with a goal pre-set by
evaluators. In a sense, DEA evaluates the efficiency of the DMU by finding whether
not this goal level (or its performance) can be further improved. This point will be
further expanded in Section 4. In the following we try to examine the DMUs
according to their desired input-output response.
Positive Response Systems (PRS)
In many applications, a DMU is expected to yield a higher level of outputs when its
input level is increased, if the unit is being operated efficiently. The efficiency of this
DMU is judged by how much more output can be produced by the unit. The higher the
extra output the unit yields, the more efficient it is considered to be. Such a DMU will
be referred to as a Positive Response System (PRS).
To be more precise we have to specify a preference. If the Pareto preference ts
selected, then a frequently used goal type associated with a (PRS) is
Unless explicitly stated otherwise, we in this paper shall always use this goal type with
a (PRS). Of course one may select other preferences according to the needs of an
individual application, and set up different types of goals for (PRS).
Negative Response Systems (NRS)
In some situations, an efficient DMU may be expected to yield a lower level of its
outputs when its input level is increased. An example of such outputs is the pollution
DEA Models via Goal Programming 85
level of a factory, taking investment as the input. Such a DMU is referred to as a
Negative Response System (NRS).
A frequently used (Pareto) goal type associated with a (NRS) is
Unless explicitly stated otherwise, we in this paper shall always use this goal type with
a (NRS). Again one may select other preferences according to the needs of an
individual application.
It is clear that the type of an input-output-system does not depend on the sign of the
input or output data (i.e., whether they are positive or negative).
Mixed Response Systems (MRS)
In practical problems it is most likely that the desired input-output relations of the
DMUs show a mixture of both the above responses. For instance, it may be that the
outputs of the DMUs can be divided into two groups: outputl and output2, and the
desired input-outputl response is positive and the desired input-output2 response is
negative. Such a DMU is referred as to a Mixed Response System (MRS). For a power
station, it is more practical to consider total amount of electricity produced, total
profits, and overall pollution level as the outputs, instead of considering only the
pollution level. If one takes its staff numbers and investment as the inputs, then this
system is a mixed response system.
It is sometimes possible to transfer a NRS to, say, a PRS, by redefining the outputs
(e.g., by setting new-output =-old-output or new-output= 1/old-output, etc). However
such a transformation may completely change the nature of the original input-output
system (e.g., CRS, etc), so that the classic DEA models may not be suitable for the
new input-output systems - as no existing DEA models are fully translation invariant.
Not only the efficiency scores (see [I]) but also the classification as efficient or
inefficient (see [2]), if a non-convex technology is used, may be changed by such
86 DEA Models via Goal Programming
translations. Hence one should use the original data whenever possible (see [12]).
Furthermore, it may be very difficult or impossible to transfer between other types of
goals and preferences.
Currently most of the existing DEA models only deal with positive response systems.
It is possible to modify them to suit other types of applications, see, e.g., [2] and [14].
It is the purpose of this work to address this issue in a more systematic way. We
approach this problem by re-examining DEA models from a goal programming
perspective. To this end we have to introduce more concepts from multiple criteria
decision making theory.
3 Merit Functions - Measures of Performance
It is natural and important to ask whether it is possible to express our preference over
the outcomes in terms of numbers so that the larger the number the better the
performance, and more importantly to quantify any extra achievement or performance
that is beyond the initial goals.
Suppose that we select a preference> on a set Y, A merit function m(.)is a function
from Y toR+ such that m(y1)>m(yJ if y 1>y2; that is, it is a (strictly) monotone
function on Y. We here only examine some merit functions closely related to DEA
models. It is therefore plausible to require them to satisfy some of the economic
efficiency axioms. We shall however not discuss this complex issue here (see, e.g. [9]).
Let us first look at the single objective case to illustrate the ideas. Let A= (Aw·· A,)'
be the n-dimensional vector of decision variables and S be a constraint set of these
variables.
We wish to find A= (A" ... ,A. )' to achieve the following (PRT) goal:
Ix;A; ~ g 0 ,A. E S, I
where g0 is referred to as the goal level. The goal may not be achievable. We shall
only examine the case where the goal can be achieved. When the goal is indeed
achievable, one wishes to know how much extra performance beyond the initial goal is
DEA Models via Goal Programming 87
achievable and how to quantify this extra performance by using, for example, a
suitable merit function. One may wish to know what is the best goal level achievable,
as the current goal level may be set too loosely. It is clear that these questions are
closely related to DEA models.
In what follows we only consider those A, e S which are able to achieve the initial goal
level. For the particular goal set above, let s• = L~ x1A.1 - g 0 ~ 0 and let
where w > 0. Then m is clearly a merit function and the extra performance of a strategy
A. can be measured by m(A.) for this PRT goal. Therefore for the single objective PRT
goal, the highest achievable goal level can be found by solving
subject to
"'" 1 - + - 0 + > 0 1 s L..t X;A; S - g ,S _ , A E .
If the optimal solution is zero, then the current goal level is the best and no extra
performance can be possibly achieved with the current constraint. Otherwise, it is
possible to improve on the current goal. The highest achievable goal level is, g0 + s+,
and s+ is the extra performance beyond the initial goal when the highest goal is
achieved. This merit function is referred to as additive type, since
When the goal level g0 is positive, one can define a radial type merit function by
88 DEA Models via Goal Programming
m(2) = w(},
where L~X;A; - s+ = ~0 ,(} ~ 0. or define an "almost" radial type merit function by
where we[O,l] is a small positive weight and L~X;A; -s+ =~0 ,s+ ~0,(}~0. We
define the almost radial merit function because the pure radial one is not strictly
monotone in the multiple objective case in the Pareto preference, but the almost one is.
We can easily define such a merit function for the case g 0 < 0 in a similar way, though
we shall not get into the details here.
Possible extra performance can then be found by solving
max (}
subject to
or
subject to
where E is a very small positive number. There are many more useful merit functions
though we shall not introduce them as here we only intend to illustrate the ideas. One
can similarly introduce merit functions for the NRT goals.
DEA Models via Goal Programming 89
3.1 Multiple Objective Cases
The above discussions can be easily generated to multiple goals. Let us examine the
following two basic types:
n n
(Lx;A.; ~g~, ... ,L:xf"A.; ~g~), A.eS, (PRS) I I
or
n n
(Lx;A.; ~g~·····Lx;mA.; ~g~, A.eS. (NRS) I I
Possible extra performance beyond the goal level for a PRS goal may be found by
solving (using the additive merit function)
subject to
"'" 1 , _ + _ 0 "'" m ' _ + _ 0 + > 0 · - 1 L..IX1 A 1 S1 -gi•···•L.,IX;A; Sm-gm,Sj_ ,)- , ... ,m,
or if g,0 > 0 (using the almost radial merit function)
max ()+e(st + ... +s~),
subject to
for a NRS goal, by solving (using the additive merit function)
90 DEA Models via Goal Programming
subject to
"'n I 1 - _ 0 "'n m 1 - _ 0 - > 0 · _ 1 L.,1X;A; +s1 - g 1 , ••• ,L.,1x; A; +sm- gm,sJ _ ,J- , ... ,m,
or if g~ > 0 and X; ~ 0 (using the almost radial merit function)
max 1-B+e(s) + ... s;;;),
subject to
where Eisa very small positive number. We mention again it is very easy to introduce
almost radial type of merit functions for the case g~ < 0 (i =I ,2, ... m). For instance, for
a NRT goal with g~ < 0, one can find the extra performance by using the following
problem (using the almost merit function):
maxB+&(s) + ... +s;;;),
subject to
There are many more possible merit functions can be introduced to suit a particular
application. For instance, one can easily introduce more non-radial types of merit
functions using the Fare-Lovell's (or Russell's) measure and Zieschang's measure
(see, [8] and [18]). We now examine the mixture of the above two types.
DEA Models via Goal Programming 91
Mixed Goal Types
In general let
and
Then a possible mixture goal setting reads:
It follows from the merit functions introduced above that using the additive merit
function, the possible extra performance can be found by solving
max( S PR WPR + S NR WNR)
subject to
where s PR and s NR are nonnegative vectors and ( wPR, wNR) are weight vectors. One
can also use the almost radial merit function if g ~R, g ~R > 0 and finds the possible
extra performance by solving
subject to
92 DEA Models via Goal Programming
where s PR and s NR are nonnegative vectors and 1.1 is the [I norm.
In the next section we shall derive some possible DEA models using the above ideas.
4 DEA Models with General Goals
We assume that there are n decision-making units to be evaluated. Each DMU uses m
different inputs to produce s-different outputs. Let the m-dimensional vector
X 1 = (x11 ,x,1, ... ,xmi)' denote the inputs of DMU1, and the s-dimensional vector
Y1 =(y11 ,y,1 , ... ,y,)'denote the outputs ofDMU1. Let X=(X,,X,, ... ,X.) be the input
matrix and Y = (Y,, Y, ... , r.) be the output matrix. It is reasonable to assume that the
goal type of this input-output system has been pre-set. If, for instance, one wishes to
minimise the input and maximise the output, then it is a positive response system. We
extend this input-output system to the following virtual input-output system:
n n
X(A.) =_LA.; X;, Y(A.) =_LA.; Y,, I I
where A.=(A,, ... ,A..)are the mixture vectors in a technology set Sto be specified in
applications. The set S should at least contain the unit vectors e" ... ,e •. These input
output units are referred to as virtual DMUs, or simply virtual units. Now assume that
we wish to evaluate efficiency of DMUi' One of the essential assumptions in DEA is
that if the initial goal level (X1,Y)cannot be further improved, or no extra
performance can be achieved, then DMU1 is efficient. Otherwise it is inefficient.
Assume that the input-output system is a PRS, for instance. Then the initial goal will
be
DEA Models via Goal Programming 93
If this goal level is found to be the best in the sense that it cannot be improved further
in a merit function, then this DMU is considered to be efficient. Of course, this
procedure will depend on the preferences and merit functions used, and thus leads to
different DEA models, depending on how these are chosen.
As an illustration of this general approach, we shall derive a few DEA models from the
above procedure. We do not emphasise however the importance of the actual DEA
models to be derived, but rather the systematic way to derive them.
4.1 Additive Types
In this subsection we use additive merit functions. We make no assumption as to
whether the input or output date are positive or negative. We first examine the PRS
case. In this case the initial goal can be stated as to find 'A such that
Of course we know that this goal is indeed achievable by letting -1-1 = 1,-1-1 = 0 (i i' )).
Then our next step is to find whether the system is able to produce superior
performance, or improve the goal level.
Indeed if we can find a virtual unit which can produce superior performance, then this
DMU is considered to be inefficient by DEA. Otherwise it is considered to be efficient.
From the discussion about finding extra performance in Section 3, for a PRS we have
the following linear programming problem to find whether the j-th unit is efficient or
not: (ADD-PRSI)
max WJP Sfp + WPR SpR
subject to
94 DEA Models via Goal Programming
where YA and XA. are the virtual inputs and outputs with A being in a selected
technology set S and w'P, wPR are nonnegative weight vectors. The values
'L:S;P and 'L:S~R identify the amounts of extra performance that the evaluated system
should be able to produce if ran efficiently. If they are zero, the j-th unit is classified to
be efficient by this DEA model. It follows that if a unit is efficient for a particular set
of non-zero weights (i.e. none of the weights is zero), then the unit is efficient for any
set of non-zero weights. Therefore we only consider non-zero weight sets here. If we
treat all the inputs and outputs equally then all the weights should be equal.
Equivalently we can take all weights to be one. Then the above is the well known
Additive Model of DEA if one assumes a convex technology set. If one makes the
weight wRP(w'R)very small, then one has the input (output) oriented Additive DEA
model.
Let us now examine the NRS case. Following the same argument as above, it is clear
that one can solve the following linear programming problem to find whether the j-th
unit is efficient or not.
max WIP S JP + WNR S NR
subject to
where again w1P and W'R are positive weight vectors. Then we have derived a DEA
model for the NRS case. It should be noted that here we make no assumption about
sign of the input or output data.
It is interesting to note that if one redefines the outputs as -Y that is, using negative
outputs if the original outputs are positive, then this DEA model become identical
with the PRS model obtained above. Therefore one can apply the Additive Model to
NRS provided one uses "negative" outputs.
Again one can have input or output oriented models as well. It is more complicated to
deal with the MRS case. Here we only examine the case where the output of j-th unit
can be decomposed into
DEA Models via Goal Programming 95
with {PR},{NR} being fixed index sets independent of j, such that the input-output
system Xi ~ Yf" is a PRS, and Xi ~ rtR is a NRS. In such a case, let
(yPR J (yPR J
yj = ~NR 'y = y NR
Then the extra performance of the virtual units can be found by solving the following
linear programming problem: (ADD-MRS I)
max WIP SIP + WPR S PR + WNR S NR
subject to
where w 1P, wPR, wNR are again nonnegative weight vectors. The j-th unit is efficient if
and only if the maximum is zero. One again can have input or output oriented model.
4.2 Almost Radial Types or Combinations
In this section we use the almost radial type of merit functions or the combinations of
the additive and the radial. There are many possible combinations even for the case
with only two different kinds of merit functions. We shall not examine all of these
possibilities.
96 DEA Models via Goal Programming
Again we first examine the PRS case. From the discussions in the last section,
assuming that we do not discriminate between input and output efficiency, we can find
possible extra performance by solving: (RAD-PRSI)
subject to
where 1.1 is the /1 norm, w1 (=) w2 are two positive weights and E is a small positive
number. In this model we have assumed that all the inputs and outputs are positive. If
one makes w1, w2 different, then one has oriented models. We can also use the
combinations of the additive and almost radial types of merit functions. For instance,
we can use an almost radial merit function for the outputs and an additive one for the
inputs. This is particularly useful for the oriented models. We only examine one
example here- an output oriented DEA model:
subject to
This is of course the output oriented CCR model. If we take S = R: and it is the output
oriented BCC model if we assume a convex technology set. We here only assume that
the outputs are positive. One can similarly derive the input oriented CCR and BCC
model.
Let us now examine the NRS case. Similarly we have the following two DEA models,
using the almost radial merit function for both inputs and outputs: (RAD-NRS 1)
DEA Models via Goal Programming 97
max w, (1 - (}in) + Wz (1 -(}out) + e(ls IP I, + Is NR I,) subject to
or using the additive merit function for the outputs and the almost radial for the inputs:
(RAD-ADD-NRS 1)
subject to
where 1.1 is the /1 norm, w1 (=) w2 are two positive weights and E is a small positive
number. In the first model we have assumed that all the inputs and outputs are positive.
We only assume that the inputs are positive in the second. It is not difficult to write
down DEA models using the almost radial merit functions for the case where the
inputs or the outputs are negative (see e.g. [14]).
Finally let us examine the MRS case. Again we only examine the case where the
output of j-th unit can be decomposed into
with {PR}, {NR} being fixed index sets independent of j, such that the input-output
system X 1 ~ Yf" is a PRS, and X 1 ~ r;R is a NRS. For the sake of simplicity, we
only introduce one input oriented DEA model for this case where we use the almost
98 DEA Models via Goal Programming
radial merit for the inputs and an additive one for the outputs. Then the extra
performance of the virtual units can be found by solving the following linear
programming problem: (RAD-ADD-MRSl)
subject to
Here we assume that the inputs are positive. This model will be used in the next
section.
5 An Application
As an illustrative application of the new DEA model (RAD-ADD-MRS1), we now
examine the effectiveness of antidepressant pharmacotherapies using amitriptyline,
imipramine, sertraline, and paroxetine respectively, from a cost saving perspective.
This has been studied via a decision tree method in [16], where the clinical statistics
have been summarised.
Using the DEA model (RAD-ADD-MRS1), we offer an alternative for evaluation of
the therapies. In this model, we use the (selected) costs of the therapies as the inputs.
The selected outputs are the probabilities of dropout, efficacy, and relapse, see [ 16] for
the details. These three outputs are used to measure the degree of success of the
therapies in [16]. The data can be found in Tables 3-4 in [16] and are summarised in
the following table.
DEA Models via Goal Programming 99
Table 1: Summary of inputs (in Pounds) and outputs (in probability) of the DEA
model
Drug Costs Efficacy Dropout Relapse
Amitriptyline 47.16 0.55 0.27 0.22
Imipramine 59.16 0.65 0.27 0.22
Sertraline 138.72 0.70 0.19 0.25
Paroxetine 120.72 0.708 0.19 0.25
For an effective therapy, one would wish to minimise the probabilities of dropout and
relapse and to maximise that of efficacy. Although all the data are positive, this input
output system is actually a MRS. If we choose the efficacy as the first output, and the
dropout and relapse as the second and third output respectively, the sub input-output
system: input ---+output! is a PRS and the sub system: input ---+ (output2, output3) is a
NRS.
We have
and
X 1 = (47.16),X2 = (59,76),X3 = (139.72),X4 = (120.72).
Y1= (0.55, 0.27, 0.22)', Y2= (0.65, 0.27, 0.22)', Y3= (0.7, 0.19, 0.25)', Y4
= (0.7 0.19, 0.25)1
We now apply the model (RAD-ADD-MRSI). It is clear that we have
and
yPR = (0.55,0.65,0.7,0.708).
100 DEA Models via Goal Programming
It is then possible to solve the following problem to find whether the j-th drug 1s
effective or not:
subject to
where we have taken S = {A: A; ~ 0 (i = 1,2,3,4); I~ A; = 1}; that is, the convex
technology set, or S ={A: A; ~ 0 (i = 1,2,3,4)}. The same efficiency results have
been found with these two technology sets.
The above programming problems are solved by a linear programming solver. It
appears that the therapies using amitriptyline, or inipramine, or paroxetine are efficient
and the therapy using sertraline has an 87% efficiency rate. This finding seems to
agree with the conclusion from a decision tree analysis in [16]: There seems no clear
cost argument demonstrated for switching between the therapies.
References
Ali, A. and L.M. Seiford ( 1990): Translation in variance in data envelopment analysis, Operations
Research Letters 10,403-405.
Allen, K. (1998): DEA in the ecological context - An overview, working paper in European
Symposium on Data Envelopment Analysis, Fachhochschule Harz, Wemigerode.
Banker, R.D. (1984): Chames, A. and W.W. Cooper, Some models for estimating technical and scales
inefficiencies in data envelopment analysis, Management Science 30, I 078-1092.
Chames, A., Cooper, W.W. and E. Rhodes (1978): Measuring the efficiency of decision making units,
European Journal of Operational Research 2, 429-222.
DEA Models via Goal Programming 101
Charnes, A., Cooper, W.W., Golany, B., Seiford, L. and J. Stutz (1985): Foundations of data
envelopment analysis for Pareto-Koopmans efficient empirical production functions, Journal of
Econometrics 30, 91-107.
Charnes, A., Cooper, W.W., Wei, Q.L. and Z.M. Huhng (1989): Cone ratio data envelopment analysis
and multi-objective programming, Int. J. Systems. Sci. 20, I 099-1118.
Charnes, A., Cooper, W.W., Lewin, A.Y. and L.M. Seiford (1994): Data envelopment analysis,
Kluwer Academic Publishers, Dordrecht.
Fare, R. and C.A. Lovell (1978): Measuring the technical efficiency of production, Journal of
Economic Theory 19, 150-162.
Ferrier, G.D., Kerstens, K and P.V. Vorden Eeckaut (1994): Radial and nonradial technical efficiency
measures on a data reference technology, Recherches Economiques de Louvain 60, 449-479.
Liu, P.L. (1985): Multiple Criteria Decision Making, Plenum, New York.
Nemhauser, G.L., Khan, A.R. and M.J. Todd (1989): Handbooks in O.R. and Management Sciences,
Vol. I, Chapter 10, North Holland.
Pastor, J.T. (1996): Translation invariance in data envelopment analysis, Annals of Operations
Research, Vol. 66,93-102.
Russell, R. (1988): On the axiomatic approach to the measurement of technical efficiency, in:
Eichhorn, W. (ed), Measurement in Economics, Heidelberg, Physica-Verlag, 207-217.
Scheel, H. (1998): Negative data and undesirable outputs in DEA, working paper in EURO Summer
Institute.
Seiford, L.M. ( 1996): Data envelopment analysis: evolution of the state-of-the-art (1978-1995),
Journal of Productivity Analysis 7, 99-137.
Stewart, A. (1994): Antidepressant pharmacotherapy: cost comparison of SSR!s and TCAs, British
Journal of Medical Economics 4, 67-79.
Stewart, T. (1996): Relationships between data envelopment analysis and multi-criteria decision
analysis, Journal of the Operational Research Society 47, 654-665.
Zieschang, K. (1984): An extended Farrell efficiency measure, Journal of Economic Theory 33, 387-
396.
Bounded vs. Unbounded Noise in Efficiency Estimation:
Performance of Alternative Estimators
Dieter Gstach1
Abstract
Whether we assume that noise, which might disturb the productio~ process, is unbounded or
bounded opens methodologically different possibilities of estimating a production frontier and
thus efficiency. The present paper provides simulation evidence about statistical properties of
technical efficiency estimators for multi-input, multi-output technologies under these two
possible assumptions: The Ray Frontier Approach (RFA) from Lothgren (1997) designed for
unbounded noise and DEA+ proposed in Gstach (1998), developed for bounded noise. RFA,
unlike earlier approaches in the realm of stochastic frontier analysis, is capable of efficiency
estimation in the case of multiple outputs as well and lends itself for comparison with DEA+.
Several settings with varying sample sizes, types of distributions and noise to signal ratios are
investigated.
1 Vienna University of Economics, Dept. VW 6, Augasse 2-6, 1090 Vienna, Austria
104 Bounded vs. Unbounded Noise in Efficiency Estimation ...
Structure
Introduction
2 Production Model and Data Generating Process
3 Simulation Settings
4 Results
5 Conclusions
References
Bounded vs. Unbounded Noise in Efficiency Estimation ... 105
1 Introduction
In this paper evidence from Monte-Carlo simulations is provided concerning some
statistical properties of two different approaches to technical efficiency estimation for
multiple-output production under noisy conditions. These approaches are the Ray
Frontier Approach (RFA) from Lothgren (1997) and DEA+ proposed in Gstach
( 1998). An alternative to these techniques is the distance function approach as
proposed in Grosskopf and Hayes (1993), Hetemiiki (1994) and Coelli and Perelman
(1996), but comparison here is confined to DEA+ and RFA.
Earlier work by Gstach (1997) yielded encouraging evidence about the small sample
properties of DEA+ estimators for the single output case and motivated extension of
research to the multiple output setting. Corresponding properties of RF A, which is a
multiple-output extension of the Aigner, Lovell and Schmidt (1977) and Meeusen and
Van den Broeck (1977) stochastic frontier analysis (SF A) approach, have to my
knowledge not yet been investigated. Because of this close link between SFA and
RF A, results from Monte Carlo studies about SFA performance (starting with Olson,
Schmidt and Waldman (1980)) are also relevant in the RFA context.
Both estimation techniques share two common properties: First they are based on the
assumption of proportional impact of noise and efficiency on all outputs. Secondly
they require parametrization of the error structure to distinguish between the noise and
the inefficiency component, unlike for example the panel data approach presented in
Kneip and Simar (1996).
The distinguishing features between RFA and DEA+ concern their respective domain
of noise and the way of moulding production. In case of RF A it is assumed that noise
is unbounded like in all stochastic frontier models. Furthermore RFA requires
parametrization of the production relationship. DEA+ on the other hand assumes noise
to be bounded, while making no assumption about production parameters, both points
reflecting the descent from DEA.
In some sense thus the two approaches compared here appear as competitors: This
would be the case if RFA were based on a flexible functional form (for example
translog) and if a normal distribution were accepted as proxy for a symmetrical,
unimodal and thin tailed and possibly bounded distribution. If, on the other hand, one
of these conditions is violated the two approaches should rather be considered as
complements, each with its own shortcomings and strengths.
106 Bounded vs. Unbounded Noise in Efficiency Estimation ...
The remainder of this paper is organized as follows: In Section 2 the production
structure and the data generating process will be defined. Section 3 describes the
various settings and steps of the Monte-Carlo-Simulations in detail. Exposition and
discussion of the results is contained in Section 4 and Section 5 concludes.
2 Production Model and Data Generating Process
Estimation of efficiency is based on observed outputs, denoted y, which must be
distinguished from often unobserved quasi deterministic outputs, denoted y . But only
the latter should be used to appreciate the performance of some decision making unit
(DMU). Based on this distinction output (production possibility) sets of a multiple
output technology can be correctly specified as
(I) Y(x) = {.Y e 9{~1 xcan produce y} with S > 1.
It is assumed that these sets satisfy strict convexity and monotonicity. The efficient
frontier in Farrells sense corresponding to input mix x is then defined as set
and the corresponding Farrell output-efficiency e of a point (x,y) as
(3) e(x,y) =max{.~. I A.ji e yF (x))}
Determining this efficiency thus amounts to find an output vector in Y(x) with output
proportions equal to those of y , which has maximum norm. But fixed output
proportions can also be guaranteed by keeping the angles 0, used in a polar-coordinate
Bounded vs. Unbounded Noise in Efficiency Estimation ... 107
representation of ji = I.YI m(B) fixed. The maximum norm one is looking for may thus be
defined as a function
(4) h(x,B) = max{r I r m(B) e Y(x)
of inputs and polar-coordinate angles of the output vector. This representation is used
in Lothgrens RF A approach, which gives rise to an alternative definition of the Farrell
efficient frontier, analogous to (2):
(5) yF (x) = {y e Y(x) I h(x,B) = I.YI}
and corresponding efficiency measure
( 6) c(x, ji) = I .Jill h(x, B)
With either (3) or (6) the efficient output level corresponding to production (x,ji) may
be defined as
(7) yF (x) = ji I c(x,ji)
Finding the efficient frontier is thus the crucial step in either approach to efficiency
measurement. In that respect DEA+ and RFA follow completely different routes. The
data generating process (DGP) motivating the use of DEA+ is assumed to consist of
density functions for noise v and inefficiency u and a function relating frontier outputs
to observed outputs as follows
108 Bounded vs. Unbounded Noise in Efficiency Estimation ...
The crucial assumptions of this DGP are thus a bounded noise term and proportional
errors. Because of this boundedness of the noise term equation reference to:
production has a natural deterministic equivalent:
A BCC type of DEA applied to {x" y1 };.1 will then yield estimates .V (x,) and
furthermore i$, of pseudo-frontier and pseudo error terms. The essential step ofDEA+
is then to estimate the parameters determining g,(.) (most importantly vmax and gu(.)
from the convoluted density of the { i$, t terms (with known domain [o, oo) !) via
maximization of the corresponding likelihood function. So the DEA+ estimate of the
true frontier is finally defined as
To motivate the use of RFA on the other hand, the DGP described above must be
modified with regard to noise. More precisely (8) must be replaced by
Bounded vs. Unbounded Noise in Efficiency Estimation ... 109
But with errors applying to all outputs equi-proportionally as defined in (10), frontier
estimation might be conducted via regression of the norms of observed output vectors
on values of x and () . This can be seen by combining the definitions of
m(h),h(x,B~/(x) andy from above, leading to the formulation IY11 = h(x1,B;)e'·-··.
Using a flexible translog function in the explaining parameters x1 and B1 as proxy for
the true functional form h(.) and ln~y1 1)-values as outputs, MLE is applied to this
regression equation to estimate the functional and the distributional parameters. This
leads to frontier point estimates
This procedure thus is an extension to the techniques investigated in Aigner, Lovell
and Schmidt (1977) and Meeusen and Van den Broeck (1977) as mentioned in the
introduction.
3 Simulation Settings
As common basis for the simulations conducted a technology with two inputs, two
outputs and decreasing returns to scale was chosen. Furthermore an input/output
separable form of this technology was assumed for ease of generating data (see for
example Kumbhakar (1996), where such functions are analyzed in an SFA context).
The output side exhibits constant elasticity of transformation while the input side is
modeled as a constant elasticity of substitution function. The specific relationship
between inputs, inefficiency, noise and outputs for DMU i then reads as follows:
( 15) ~0.5y 2 + 0.5y 2 = x0·2 x05 e'•-•• 1,1 1,2 1,1 1,2
All simulations are carried out on the basis of a fixed design of the input side,
characterized by xu values, where i = l... .. n and j = 1,2. This should be kept in mind,
when interpreting the results, as it reduces variability of the results compared to a
110 Bounded vs. Unbounded Noise in Efficiency Estimation ...
random design setting. The advantage of the fixed design is better comparability of the
results between the RFA and the DEA+ approach.
The xij values are fixed as a grid of order statistics. To construct the latter a normally
distributed random variable r with E(r) = 10 and STD(r) = 3 is defined first. The
sample size n was chosen to be the square of some integer. This then gives rise to
order statistics rk for k = I ... Fn with corresponding expected values denoted r,. Now
all possible variations of elements r, of order two with replacement yield an [n x 2]matrix, which I use as input-matrix. The first vector of this input-matrix is accordingly
given through k., x1,2 ) = (~, ~ ), while the last vector is (x,,., x • .J = ~~. ,r,J Thus I
have
With these fixed inputs and some random values for noise term v, and inefficiency
term u, the norm of the output of DMU is fully specified as
~ y~, + y~2 = J2 x~i2 x~i e'·-·· . This norm is then split by introducing a random
technology parameter a, to yield output values defined by
- ~ 0.2 o.s ,,-u; -~1 a 0.2 05 ,,-u, y 1 - a. x., x. 2 e , y. 2 - - . x. 1 x. 2 e . I, I /, I, I, I I, I,
These technology parameters are drawn independently from a [0,1]-truncated-normal
with
(17) E(a,) = 0.5, STD (a,)= 0.2
All stochastic models used are parametrized via mean inefficiency f.J. = E(u) and
noise-to-signal ratio 1J = VAR(v)/[VAR(v)+ VAR(u)]. The latter will simply be referred to
as noise in the sequel. In case of the bounded noise term underlying the DEA+ model
Bounded vs. Unbounded Noise in Efficiency Estimation ... 111
an extra parameter is needed to fully specify the two distributions involved (see
below).
For the inefficiency term u two different distributions are considered:
u- iid exponential with parameter B, u e [O,+oo ). The value of B is derived from given
f.J, via f.J. = 1/B, so VAR(u) = 1/82 = f.J:.
u- iid halfnormal with parameter a •. ue[O,+oo). The value of a, is derived from the
relationship f.J, = a • ..J2fi, while VAR(u) = (1 - 2/ 7r )a: = (1r /2- 1) f.J: .
The distribution of the bounded noise term v used in evaluating DEA+ performance is
v-iid Symmetrical Beta with Parameters a and vma,,ve[-vma,,+vmaJ. The first
coefficient is set to a = 2. This together with values for f.J, and 1J and the definition
VAR(v) = v!,)(2a + 1) then determines vmax = J1J(2a + 1)VAR(u)/(1-1J).
The distribution of the unbounded noise term $v$ used in evaluating RFA performance
IS
v- iid Normal with Parameter a,. The specific value of a, is derived from the
specifications of f.J, and 1J via a,= J VAR(u)/(1-1]).
I performed simulations for a fixed mean inefficiency of f.J, =0.2, sample sizes of
n=JOO, 225, 400, noise-to-signal-ratios of IJ= 0.1, 0.2, 0.4 and exponential and
halfnormal inefficiency distributions. Simulations for each RFA setting were
replicated 300 times and 200 times for each DEA+ setting. The extremely CPU time
consuming bootstrapping step used to bias-correct the naive DEA estimates ruled out
greater number of replications so far.
The performance of both approaches was measured with the following three statistics:
112 Bounded vs. Unbounded Noise in Efficiency Estimation ...
( 18) Bias of Mean inefficiency estimator: A, - f..l,
(19) Bias ofNoise estimator: r,, -1],
(20) Mean absolute Deviation (MAD) of frontier estimates: Elln(V (xJ)-ln(/ (x, ))I
For applied work the performance of the mean inefficiency estimator possibly is of
most interest. The other two statistics were added primarily for diagnostic purposes.
The results for these statistics should be compared in magnitude to the fixed mean 0.2
of inefficiency and the standard deviations of 0.2 resp 0.15 (for exponential resp.
halfnormal inefficiency, all on log-scale), as this sheds light on the identification task
of the estimators. Interpreting the results keep in mind, that due do the parametrization
chosen, overall variability V AR(w) changes across inefficiency specifications: The
composed error w = v-u in the halfnormal specification has only about 80% of the
variability compared to the exponential specification.
4 Results
Let's start with the case of exponential inefficiency and look at the results for Mean
Inefficiency Bias of the two approaches under consideration as given in Table 1.
Except for the most unfavorable n=100, noise=40% case, the permanently negative
bias ofRFA has magnitude less than 5% (-.01) in terms of the target figure of0.2 with
standard deviations across MC-trials of at most 15% (0.03) again in terms of target
value. I carried out simulations for settings with different population inefficiencies (0.1
and 0.4) as well with similar results, so I just documented the intermediate case of 0.2
true mean inefficiency.
In the case of bounded noise and consequent application of DEA+ matters deteriorate
slightly. Dropping like before the worst case ofn=100 with 40% noise the maximum
bias is roughly 10% with maximum standard deviations ranging up to 25% in the 40%
noise scenarios. For less noisy settings (up to 20%) DEA+ has maximum standard
deviations of around 15% like RF A. Convergence is rather poor, due to the boundary
estimation step involved in DEA+ , an observation in line with the findings m
Korostelev, Simar and Tsybakov (1995) or Park, Simar and Weiner (1997).
Bounded vs. Unbounded Noise in Efficiency Estimation ... 113
Looking more closely at these Monte-Carlo distributions and running Kolmogorov
tests with H 0 :jJ ""normally distributed I found that this hypothesis in only 3 of 18
cases considered in Table 1 could be rejected at significance levels of at least 10%.
Therefore the tabulated standard deviations may be used to construct approximate
confidence bands to test for difference in mean inefficiencies of two samples.
Noise estimation ofRFA for the smallest samples is negatively biased in all except one
(n=100, 77 = 0.1) case, thus generally attributing too much weight to the noise
component in total variability and too little weight to inefficiency (see Table 2). In a
much more pronounced fashion amounting to complete failure the same holds true for
DEA+ , which is mainly due to a highly overestimated shape-parameter of the Beta
distribution. Thus through the DEA+ lens noise appears as extremely peaked
distribution with long, thin tails as opposed to the true inverse U-shape as given by the
true shape-parameter value f3 = 2.
Standard deviations of the RF A noise estimators on the other hand are very large for
n=lOO, 77 < 0.2 compared to DEA+. Thus using mean squared errors as goodness-of-fit
criterion most of the advantage ofRFA with moderate noise and small samples is lost.
Only when noise may appropriately be modeled as normal and samples are at least of
size n=400, RFA may be used with more confidence. For illustration: The range
between the 10% and the 90% quantiles of the distribution of the RF A noise estimator
~, when actually 77 = 0.4, shrinks from [0.16,0.64] for n=1 00 to [0.30,0.50] for n=400.
In a paper by Coelli ( 1994) about the finite sample properties of stochastic production
functions a simpler model with a constant term but no regressors and normal noise,
halfnormal inefficiency distribution is investigated along the same lines. Coelli's
model is parametrized with r = 0'; /(0'; + 0';) instead of the noise parameter 77 used
here. Because of MLE's invariance property the 77 -estimates may be transformed via
r (1 -77 )/(!- 27] ") to yield parameter estimates comparable with Coelli's results. Table
7 contains the corresponding figures, all of which are expressed in percentage terms of
the true parameter values for r .
The standard deviations of the estimators are indeed of similar size, although Coelli's
results are more accurate owing to more replications (1000 vs. 300). The fact, that
standard deviations of the r -estimator are smaller for RF A than in Coelli's setting
when n=400, despite a noticeable loss in degrees of freedom ( 12 parameters here vs. 3
in Coelli) has to do with different 0' 2 = 0'; + 0'; values employed. The 10% noise case
114 Bounded vs. Unbounded Noise in Efficiency Estimation ...
here corresponds to o- 2 =0.0882, 20% to o- 2= 0.1199 and 40% to o- 2 = 0.2150. Coelli,
in the light of the invariance results of Olson et a!. (1980) keeps o- 2 constant at 0.25
and uses mean inefficiency values of around 0.4, depending on the y value chosen. So
my 40% case comes closest to his choice of a and this is nicely reflected in the
tabulated standard deviations.
Comparing Table 5 with Table 7 you will also note, that percentage bias of y for the
RF A model in terms of target values differs from the corresponding figures for TJ
roughly by factors between 0.2 (TJ=0.4, n=100) and 30 (TJ=O.l, n=400). This finding
of course is easily explained by different definitions of parameters, but it tells, that
information about statistical properties of y are of little if any use in trying to assess
the relative contributions of noise and inefficiency to total signal variability in applied
work.
Turning now to some overall measure of fit of the estimated frontier points, the mean
absolute deviations of a samples estimated frontier points from their true values is
informative (see Table 3). As the deviations are measured in log-scale, they may be
compared directly with the bias figures for the inefficiency estimator. Not surprisingly,
these mean absolute deviations are typically two to three times larger than the bias of
the inefficiency estimators, as the former focuses on point estimates. Again the
assumption of unbounded noise and thus application of RF A leads to better estimators
in terms of mean squared errors as compared to the bounded noise case.
The case of halfnormal inefficiency appears quite similar in terms of statistical
properties as a glance at the figures in Tables 4 - 6 reveals. Thus the general
observations made above for exponential inefficiency still hold. As far as mean
inefficiency is concerned, DEA+ and RFA again exhibit moderate differences
between. Mean squared errors (MSE) of the DEA+ estimator are smaller than the MSE
of RF A in three instances (n= 100, TJ 5, 0.2 and n=225, TJ = 0.1) and vice versa for the
other settings. Permanent underestimation of inefficiency via RF A is in line with
Coelli's findings.
Bias estimates for noise given in Table 5 on the other hand show, that Coelli's findings
(with equal error structure) of permanent underestimation ofthe y-parameter, which is
inversely related to the TJ -parameter used here, does not generalize to this setting with
multiple outputs. Increasing sample size in my case is accompanied by switching bias
from positive to negative, although at decreasing magnitude. This means that the noise
Bounded vs. Unbounded Noise in Efficiency Estimation ... 115
contribution to total variability is underestimated on average for larger sample sizes,
while it is constantly overestimated in Coelli's setting (see also Table 7 for the
corresponding r representation).
5 Conclusions
This paper investigated the finite sample properties of two basic models for technical
efficiency estimation with multiple outputs. These models are Lothgren's ray frontier
approach (RFA) and Gstach's adaptation of DEA to noisy settings (DEA+}, both
requiring a parametrization of the error structure to identify inefficiency contained in
noisy performance signals. But the approaches differ basically in their respective idea
of noise: RF A assumes unbounded (log-normal) noise, as has become standard in
stochastic frontier estimation. DEA+ on the other hand is developed under the
assumption of bounded (log-symmetrical-beta) noise, to exploit the non-parametric
structure of DEA.
For benchmarking I used estimators for mean sample inefficiency, for noise-to-signal
ratio and for mean absolute deviations of frontier point estimates from their true values
(MAD). The performance of these estimators was analyzed for a known technology
with two inputs and two outputs and known error structure in a Monte-Carlo
experiment with fixed mean inefficiency of 20%. Settings with different noise to
signal ratios (10%, 20% and 40%}, different sample sizes (100, 225, 400) and different
inefficiency distributions (exponential, halfnormal) were considered.
Mean inefficiency estimation with DEA+ for the smallest sample size (n=lOO) and
moderate noise ( ~ 0.2) has slight advantages in terms of mean squared errors
compared to RF A. But better convergence of RF A makes it a superior choice for
larger samples and more noise, given that the unbounded noise assumption is valid.
With n=400 the RFA mean inefficiency estimator has maximum bias (across noise
specifications and inefficiency distributions) of ~ 1% and maximum standard
deviation ~ 14% in terms of the single fixed target value of 0.2. Bias of DEA+ is
predominantly positive, while RF A on average underestimates inefficiency in all
settings.
Noise estimation with a sample size of n=!OO is a hopeless task in a statistical sense
for practical purposes with bias of up to 98% (DEA+) and standard deviations of up to
116 Bounded vs. Unbounded Noise in Efficiency Estimation ...
150% (RFA) in terms of (varying) target values. For sample size n=400 at least for
RFA matters improve somewhat, while DEA+ simply fails as far as noise estimation is
concerned. But even RF A noise estimators with n=400 have bias of up to 11% and
standard deviations of up to 36%.
Overall accuracy of DEA+ frontier estimates assessed via mean absolute deviations
from the frontier (log-scaled) is about as good as the corresponding RF A estimates for
n=100 and noise :5 20%. But RFA converges much faster leading to roughly twice the
accuracy of DEA+ for n=400. For noise=20% and n=100 RFA on average filters
around 80% of the error and DEA+ around 70%.
Further research will have to investigate the sensitivity of the two models to
inappropriate noise specifications, i.e. employing RF A when noise in fact is bounded
resp DEA+ when noise is unbounded. Analyzing the first possibility would indicate
the quality of the Normal approximation to other kinds of noise, while analyzing the
second would shed some light on the issue of DEA+ as sort of a multidimensional
order statistics estimator.
Appendix
DEA+ RFA
n-100 N-225 n=400 n-100 n-225 n-400
Noise 10%, Bias -.007 .010 .011 -.009 -.000 -.000
STD .022 .019 .015 .029 .018 .014
20%, Bias .019 .025 .013 -.003 -.000 -.000
STD .030 .029 .021 .035 .021 .017
40%, Bias .032 .021 .005 -.005 -.003 -.001
STD .064 .049 .025 .048 .031 .020
Table 1: Monte-Carlo Distribution of Mean Inefficiency Estimator for Log
Exponential Inefficiency with Mean 0.2.
Bounded vs. Unbounded Noise in Efficiency Estimation ... 117
DEA+ RFA
N-100 N -225 n-400 n-100 N -225 n-400
Noise 10%, Bias -.094 -.085 -.074 .020 -.015 -.009
STD .012 .017 .019 .114 .039 .031
20%, Bias -.178 -.144 -.097 -.018 -.014 -.007
SID .047 .061 .050 .131 .068 .056
40%, Bias -.226 -.136 -.073 -.027 -.006 -.004
STD .228 .153 .097 .186 .120 .080
Table 2: Monte-Carlo Distribution of Noise Estimator for Log-Exponential
Inefficiency with Mean 0.2.
DEA+ RFA
N-100 n-225 n=400 n-100 n-225 n=400
Noise 10%, MAD .053 .045 .038 .043 .023 .016 STD
.008 .006 .007 .015 .006 .005
20%, .069 .059 .047 .047 .029 .021
MAD STD .011 .014 .Oll .016 .008 .006
40%, .105 .076 .054 .062 .041 .029
MAD SID .027 .024 .009 .022 .012 .009
Table 3: Monte-Carlo Distribution of Frontier Estimators assessed via Mean
Absolute Deviations from true frontier points for Log-Exponential Inefficiency
with Mean 0.2.
DEA+ RFA
N-100 N-225 n-400 n-100 n-225 N-400
Noise 10%, Bias -.020 -.001 .005 -.019 -.002 -.000
STD .014 .Oll .009 .027 .014 .Oll
20%, Bias -.000 .018 .021 -.011 -.001 -.001
STD .016 .012 .014 .031 .019 .013
40%, Bias .027 .040 .035 -.016 -.003 -.000
STD .056 .036 .031 .059 .032 .022
Table 4: Monte-Carlo Distribution of Mean Inefficiency Estimator for Log
Halfnormal Inefficiency with Mean 0.2.
118 Bounded vs. Unbounded Noise in Efficiency Estimation ...
DEA+ RFA
N-100 n -225 n-400 n -100 n-225 n-400
Noise 10%, Bias -.098 -.096 -.090 .068 -.013 -.011
STD .003 .006 .012 .156 .061 .036
20%, Bias -.195 -.189 -.166 .014 -.016 -.010
STD .012 .022 .036 .161 .095 .059
40%, Bias -.311 -.301 -.246 -.005 -.014 -.019
STD .249 .164 .144 .245 .150 .105
Table 5: Monte-Carlo Distribution of Noise Estimator for Log-Halfnormal
Inefficiency with Mean 0.2.
DEA+ RFA
N-100 n-225 n-400 n-100 n-225 n-400
Noise 10%, MAD .054 .038 .034 .042 .022 .016 STD
.008 .005 .005 .017 .007 .005
20%, .057 .052 .047 .043 .026 .019
MAD STD .008 .007 .009 .016 .008 .006
40%, .091 .079 .067 .054 .037 .027
MAD sm .036 .019 .018 .024 .017 .011
Table 6: Monte-Carlo Distribution of Frontier Estimators assessed via Mean
Absolute Deviations from true frontier points for Log-Halfnormal Inefficiency
with Mean 0.2.
RFA Coelli (1994)
n-100 n-400 N-100 n-400
Noise -4.1 0.4 -0.7 -0.2
10%, BIAS 9.1 1.5 4.8 1.8
STD
20%, BIAS -2.0 0.4 -1.9 -0.6
STD 11.6 3.1 10.2 3.3
40%, BIAS -5.4 0.7 -8.9 -2.6
STD 29.1 9.5 28.8 10.2
Table 7: Monte-Carlo Distribution of Gamma Estimator y =a~ /(u~ + u;) for
Log-Halfnormal inefficiency with Mean 0.2. All figures in percentage terms of
true gamma values (=0.9612 for 10% noise, 0.9167 for 20% and 0.8050 for 40%).
Bounded vs. Unbounded Noise in Efficiency Estimation ... 119
References
Aigner, D.J., Lovell, C.A.K and P. Sclunidt (1977): Formulation and estimation of frontier production
function models, Journal of Econometrics, 6: 21-37.
Coelli, T. J. (1994): A monte carlo analysis of the stochastic frontier production function, Working
paper, Econometrics Department, University of New England.
Coelli, T. J. and S. Perelman (1996): Efficiency measurement, multiple-output technologies and
distance functions: With application to european railways, CREPP Discussion Paper 96/05,
University of Liege, Belgium.
Grosskopf, Shawna and K. J. Hayes (1993): Local public sector bureaucrats and their input choices,
Journal of Urban Economics, 33: 151-166.
Gstach, D. (1997): Using {DEA+ to estimate multiple output technologies: First Monte Carlo
evidence, Journal of the International Atlantic Economic Society, 3(4), Research Note.
Gstach, D. (1998): Another approach to data envelopment analysis in noisy environments: {DEA+,
Journal of Productivity Analysis, 9: 161-176.
Hetemaki, L. (1994): The impact of pollution control on firm production technology: A stochastic
distance function approach, in: Brlinnlund, R., Kristriim, B., Lofgren, K. G., and L. Mattson,
Environmental Economics, Swedish University of Agricultural Sciences, Department of Forest
Economics, Report No. I 06, Umea.
Kneip, A. and L. Simar (1996): A general framework for frontier estimation with panel data, Journal
of Productivity Analysis, 7(2/3): 187-212.
Korostelev, A., Simar, L. and A. Tsybakov (1995): Efficient estimation of monotone boundaries, The
Annals of Statistics, 23: 476-489.
Kumbhakar, S. C. (1996): Efficiency measurement with multiple outputs and multiple inputs, Journal
of Productivity Analysis, 7(2/3): 225-255,.
Liithgren, M. (1997): A multiple output stochastic ray frontier production model, Working Paper
Series in Economics and Finance 158, Stockholm School of Economics.
Meeusen, W. and J. van den Broeck (1977): Efficiency estimation from Cobb-Douglas production
functions with composed error, International Economic Review, 18: 435-444.
Olson, J.A., Sclunidt, P. and D.M. Waldman (1980): A Monte Carlo study of the stochastic frontier
production function, Journal of Econometrics, 13: 67-82.
Park, B.U., Simar, L. and C. Weiner (1997): FDH efficiency scores from a stochastic point of view,
Discussion Paper 9715, Universite Catholique de Louvain, Institut de Statistique.
Price Indexes for Nonmarketed Goods
Rolf Fiire', Shawna Grosskopf and Pontus Roos3
Abstract
The purpose of this paper is to outline theoretical and empirical guidelines for specifying and
estimating a price index for goods or services for which there is no market price available (or where those prices are administered or distorted or distorted). The case we are thinking of
include many public services such as publicly provided education and health services.
Obviously, one of the difficulties in constructing price indexes for nonmarketed goods is
finding a proxy for output prices to use in the price index. This we address by turning to shadow prices. We outline several theoretical alternatives, as well as providing some
practical approaches to estimation of shadow prices ofnonmarketed goods.
We also provide linear programming, or DEA, method of estimating the price index for
individual observations.
1 Department of Economics and Department of Agricultural and Resource Economics, Oregon State University, Corvallis, OR 97331-3612, USA
2 Department of Economics, Oregon State University, Corvallis, OR 97331-3612, USA
3 Institute for Health Economics (IHE), Box 2127,220 02 Lund, Sweden
122 Price Indexes for Nonmarketed Goods
Structure
Introduction
2 Output Price Indexes
3 Shadow Prices
4 Summary
References
Price Indexes for Nonmarketed Goods 123
1 Introduction
The purpose of this paper is to outline theoretical and empirical guidelines for
specifying and estimating a price index for goods or services for which a market price
is unavailable, administered, or distorted. This includes for instance, such services as
publicly provided education and health care. One may also think of nonmarketed
goods such as enviromental quality, though we do not address directly some of the
more complicated issues which arise in that context.4
The principal difficulty in constructing price indexes for nonmarketed goods is finding
a suitable proxy for output prices. This we address by turning to shadow prices. In this
paper we outline several theoretical approaches as well as providing some practical
empirical approaches to estimation of shadow prices of nonmarketed goods.
2 Output Price Indexes
We define our output price indexes in terms of production. In particular, we define the
indexes as ratios of revenue functions, which is in the spirit of Koniis (1924) 'true'
index of the cost of living. We also provide a linear programming or DEA5 method of
computing the index for individual observations.
More specifically, suppose we have k = I, ... ,K units of observations such as
government offices, schools or stores that supply a vector of services over which we
wish to compute price indexes. We show how these observations may be used in
calculating the revenue functions that enter into the price indexes. Since we are
especially interested in sectors in which outputs are not marketed, this also involves
identification of shadow prices for those outputs. Finally, using output shares, we
illustrate how the K individual price indexes can be aggregated into one index.
We begin with specification of the technology, which at time t, consists of the set of all
input vectors x' = (x; , .... x~ )e ~:and output vectors y' = (y; , .... y~ )e ~7 such that x' can produce y'. We denote this set by T', and we assume that it satisfies standard
4 For a discussion of some of these, see Fare and Grosskopf(l998).
5 The term DEA (Data Envelopment Analysis) was coined by Chames, Cooper and Rhodes (1978).
124 Price Indexes for Nonmarketed Goods
regularity conditions, such as those in, e.g., Fare and Primont (1995). Given output
prices6 p 1 = (p; , .... p'.. )e 9\~ we can define the revenue functions as
(1) R1 (x 1 ,p1 )= max{p 1y 1 : ~~ ,y1 )e T 1 } y'
This function is the maximal revenue that can be achieved at prices p 1 with the given
input vector x 1 • If prices change, say to p 1• 1 , another maximal revenue can be
calculated relative to technology T 1 with the same input vector x 1 , which we write as
R1 (x 1 , p 1• 1 ). The ratio of the two revenue functions is a Fisher and Shell (1992) output
price index.
Here, period t technology T 1 , and the input vector X1 are fixed while prices change
from p 1 to p 1• 1 • This output price index is the analog to the Kontis cost of living
index; both are defined as ratios of support functions.
The revenue function is homogeneous of degree + 1 in output prices, thus by Euler's
theorem we may write
(3) Rl(xl PI)= ~8R 1 (x 1 ,p 1 L 1
' L...J a I Pm' msl 'Pm
and
(4) M 8RI( I 1+1)
Rl(xl pi•')=" x ,px 1+'. , L..... a 1+1 Pm
m•l 'Pm
By Shephard's lemma (Shephard , 1970) the partial derivatives are the optimal
supplies of outputs, i.e.,
6 We address the issue of how to obtain these prices in Section 3 of the paper.
Price Indexes for Nonmarketed Goods
(5)
and
(6)
aR 1k.P 1L I (I 1\ ---'---'-'---'-Ym X ,p ;.m-l, ... ,M, ap~
aR I( I 1+1) X 'p - I ( I I+ I\ - 1 M --'--:-1+7-1--'- Ym X ,p ;.m- , ... , .
apm
125
The outputs y~ (x 1 , p 1 ) can be interpreted as the observed outputs at period t, but the
mixed period outputs y~ (x 1 , p,.1 ) in (6) are not observed. However, if the two vectors
of outputs are approximately the same or if we take y~(x',p')=y~(x',p,.1 ), then the
price index (2) is a Laspeyres output price index (Laspeyres, 1871)
(7) ""M t t+l
P'( I t t+l)- L.Jm•lYmPm X 'p 'p - "M I I
L..m•IYmPm
The price index (2) can be computed using linear programming or DEA methods. For
this purpose we assume that there are k = l, ... ,K observations of inputs
x'·k = (x~ 1 , •••• x~N) and output vectors y1·' = {y~ 1 , •••• y~). These data may be used to form
a reference technology, i.e.,
T' ={k.y'):
(8)
K
y~ ~2:z;y~,m=l, ... ,M, k•i
K
x: ~ 2:z~x~, n = l, ... ,N, k•l
z; ~ O,k = l, .... ,K}
The technology in (8) satisfies constant returns to scale, i.e., .lT1 = T', .l> 0 with strong
disposability of both inputs and outputs, i.e., for X1 ~ x' ,(x1 ,y') E T 1 then (x 1 ,y' )e T'
and .Y' ~ y' ,(x', y 1 ) e T' then (x', y') e T', respectively.
126 Price Indexes for Nonmarketed Goods
Now suppose that output prices are known for each k = 1, ... , K and t, • l,k ( I I ) 1.e.,p = PkJ•···oPw is a known vector of prices.7 Then for each k', the revenue
functions that form the price index in (2) may be estimated as the solutions to the
linear programming problems below.
K
s.t. y~ ;<;; :~:z;y~, m = l, ... M, k•l
K
x; ~ L:z;x~, n = I, ... N, k•l
z: ~ 0, k = l, .. .K.
R 1 (k' ,x*'.l ,p*'.l+l )= M
m.a~ LP;~~y~ Y ,z m=l
K
s.t. y~ :<> Z:z;y~, m = l, ... M, k·l
K
x; ~ Z:z;x~, n = I, ... N, k•l
z; ~ 0, k = l, ... K.
The results from (9) and (10) are used to obtain the output price index
P1(k,xk·1 ,pk·1 ,p*·,.') for each observation k=l, ... ,K. These individual price indexes
may be aggregated into a total index using revenue shares, where the share for k' is
defined by
(11) R 1(k' xk'.l p*-.1 )
s(k') = ' ' z::., R 1 (k, x*·1 , p*·1 )'
7 We shall relax this assumption in Section 3.
Pnce Indexes for Nonmarketed Goods 127
The overall index is then given by
K
(12) p 1 = :Ls(k')P'(k' ,xk',t ,pk',t ,pk',t+l ). k'•l
After some manipulation this index can be shown to take the simple form of
( 13) "K R' (k xk,t k,t+l)
P' = L.k., • ,p "K R' (k k,t k,t) ~k-=1 ,X ,p
Thus the overall output price index is the ratio of sums of revenue functions, one for
the period t+ 1 prices and one for the period t prices.
We note also that parallel to our discussion of the Laspeyres form of the output price
index, one may also specify a Paasche index, and therefore a Fisher type index.
3 Shadow Prices
In this section we summarize work done on identifying shadow prices for nonmarketed
outputs. This section borrows heavily from Fare and Grosskopf ( 1998).
Fare and Grosskopf (1990) developed a method for computing prices for firm inputs
based on Shephard's (1953) input distance function. Here, following Fare, Grosskopf,
Lovell and Yaisawamg (1993) and Fare and Primont (1995) we show how their
method can be used to compute shadow prices for firm outputs.
Denote inputs by x' = (x: , .... x~ )E 9l~ and outputs y' = (y: , .... y;, )E 9l~ .8 The output
distance function is defined by
(14) D.(x,y)=inf{B:(x,y/B)ET}
Where T is the technology given by
8 We suppress the time index for simplicity in this section.
128 Price Indexes for Nonmarketed Goods
(15) T = {(x,y): x can produce y}.
Fare and Primont (1995, p.22) show that T can be recovered from the output distance
function, i.e.,
(16) T = {(x,y):D.(x,y)s!},
if and only if outputs are weakly disposable in the sense of
(17) (x,y)e T and 0:58:51 => (x,~)e T.
Let p e 9i~ ,p * 0 be a vector of output prices and consider the revenue maximization
problem
R(x, p) =max py s.t. D. (x, y) :51. y
The Lagrangian problem associated with (18) is
(19) r(y,B) = py + B(l- D0 (x,y )~
and the first order conditions are
It can be shown, see e.g., Fare and Primont (1995, p. 54) that the optimal Lagrangian
equals maximum revenue, i.e.,
(21) B = R(x,p}
Combining expressions (19) and (20) proves that the output gradient of the output
distance function equals revenue- deflated output shadow prices, i.e.,
Equivalently, output shadow prices equal the output gradient of the output distance
function times maximal revenue, i.e.,
Price Indexes for Nonmarketed Goods 129
p = R(x,p)'\1 YD.(x,y}
For two different outputs m and m' it follows that their relative prices equals the
corresponding ratio of distance function derivatives,
(24) Pm· _ 8D.(x,y)/8ym' Pm - 8D.(x,y)/8ym
Thus, one may compute shadow prices of outputs using only data on input and output
quantities. These shadow prices are normalized by maximal revenue. 'Absolute'
output shadow prices may be computed in relative terms as above, or if at least one
observed output price is known and one is willing to assume that it equals its shadow
price, absolute prices may be retrieved following Fare, Grosskopf, Lovell and
Yaisawarng (1993). These shadow prices are derived based solely on production
technology (in conjunction with price information in the aforementioned case).
Since the production technology is often characterized in dual terms, typically the cost
function, one may ask whether shadow prices also be computed based on the cost
function. Let we 91~ denote input prices, then the cost function is defined as
(25) C(y, w)= m}n{wx: (x,y)e T}.
As is well- known from duality theory, the technology may be recovered from the cost
function.
From the profit maximization problem
max{py- C(y, w}} y
it follows that
p = '\1 yc{y, w),
i.e., output shadow prices are equal to their associated marginal costs. In contrast to the
distance function approach to shadow pricing, input prices rather than input quantities,
are the required data.
In a similar vein, one may also derive output shadow prices based on a cost indirect
output distance function
13 0 Price Indexes for Nonmarketed Goods
(28) ID.(w/c,y)=inf{B:0'eT,wx:>c},
where c represents the budget or target costs. Based on the duality between the cost
indirect output distance function and the cost indirect revenue function (see Fare and
Grosskopf, 1994),
(29) IR(wjc,p)= sup{py: ID.(w/c,y):> 1},
and specifying the Lagrangian and following the logic used with the output distance
function above we have
(30) p = IR(w/ c,p )'V YID. (w/c,y)
Thus, duality theory provides several alternative models which can be employed to
derive shadow prices of outputs. Implementation of these generally requires
specification of a (differentiable) functional form of technology, estimation and
derivation of the shadow prices based on the derivative properties derived above. This
is a approach taken in Fare, Grosskopf, Lovell and Yaisawarng (1993 ), among others.
Shephard (1970) provides an alternative approach that is closely related to the
approach proposed for the computing of our price indexes, namely a linear
programming approach. The general idea is to solve for maximal revenue, but instead
of solving for optimal output vectors given output prices, he solves for optimal output
prices given outputs and inputs. Specifically, for each observation k'= I, .. .K, solve
( k' ) k' H y ,w*. =maxpy p,w
M
(31) s.t. LPmYtm- L w,xkn :> 0, k = I, ... ,K,
w, ~ wk',n' n = l, ... N,
p ~ 0, w~ 0.
The solution vector gives the revenue maximizing shadow price of outputs given
inputs and outputs, p(x,y).
Price Indexes for Nonmarketed Goods 131
4 Summary
In summary, we have specified a producer based output price index as ratios of
revenue functions, which we propose for use in computing price indexes for
nonmarketed goods such as public services. We suggest computing these as solutions
to simple linear programming problems. The difficulty, of course, is identifying output
prices when the outputs or services are not marketed. We outline several ways in
which shadow prices of these services may be derived using duality theory. We also
discuss a model proposed by Shephard ( 1970) which solves for output shadow prices
in a linear programming framework. All of the approaches outlined here are applicable
to the multiple service/output case which characterizes, for example, most public
service providers.
References
Aigner, D. and Chu, S.F. (1968): On estimating the industry production function, American Economic
Review 58, 226-239.
Althin, R. (1995): Three essays on efficiency measurement, PhD dissertation, Lund, Sweden.
Charnes, A., Cooper, W.W., and Rhodes, E. (1978): Measuring the efficiency of decision making
units, European Journal of Operational Research, 2:6, November, 429-444.
Clement, J., Grosskopf, S. and Valdmanis, V. (1996): A comparison of shadow prices and
reimbursement rates of hospital services, Annals of Operations Research, 67, 163-182.
Fare, R. and Grosskopf, S. (1994): Cost and Revenue Constrained Production, Bilkent University
Lecture Series, New York: Springer- Verlag.
Fare, R. and Grosskopf, S. (1990): A distance function approach to price efficiency, Journal of Public
Economics, 43, 123-126.
Fare, R. and Grosskopf, S.( 1998): Shadow pricing of good and bad comodoities, American Journal of
Agricultural Economics, 43, forthcoming.
Fare, R. and Grosskopf, S. Lovell, C.A.K., and Yaisawarng, S. (1993): Derivation of shadow prices
for undesirable outputs: a distance function approach, The Review of Economics and Statistics,
374-380.
132 Price Indexes for Nonmarketed Goods
Fare, R. and Primont, D. ( 1995): Multi-Output production and duality: Theory and applications,
Boston: Kluwer Academic Publishers.
Fisher, F. and Shell, K. (1972): The pure theory of the national output deflator, in: F.M. Fisher and K.
Shell, The Economic Theory of Price Indices, New York: Academic Press, 49-113.
Grosskopf, S. and Hayes, K. ( 1993): Local public sector bureaucrats and their input choices, Journal of
Urban Economics, 33, 151-166.
Grosskopf, S., Hayes, K., Taylor, L., and Weber, W. (1997): Budget- Constrained Frontier Measures
of Fiscal Equality and Efficiency in Schooling, Review of Economics and Statistics, 79: I, 116-
124.
Konils, A.A. (1924): The problem of the True Index of the Cost of Living, Translated Econometrica,
7 (1939), 10-29.
Laspeyres, E. (1871): Die Berechnung einer mittleren Warenpreissteigerung, Jahrbilcher fUr
Nationalokonomie und Statistik, 16,296-314.
Paasche, H. (1874): Ober die Preisentwicklung der letzen Jahre nach den Hamburger
Borsennotirungen, Jarbilcher fur Nationalokonomie und Statistik, 23, 168-178.
Shephard, R. W. (1953): Cost and Production Functions, Princeton University Press.
Shephard, R. W. ( 1970): Theory of Cost and Production Functions, Princeton University Press.
Lessons Learned for DEA Practice from Health Care
Applications in the UK
Rob Ball (Head of Department), Elizabeth Roberts (Doctoral Research Student) and
Richard Thomas 1
Abstract
The literature on DEA shows limited practical application to public sector services in the UK.
We have applied DEA to compare the efficiency and performance of a number of hospitals in
Scotland. This has given us insight on a number of key issues related to the application of
DEA, which have been further extended through discussions with a variety of managers in the
field. This paper discusses these issues, including the introduction of weight limitations,
issues of comparability of DMUs, the robustness of the models and the sensitivity of the
results, illustrating them using real data from Scotland.
1 Department of Management and Organization, University of Stirling, Stirling, FK9 4LA, UK
Robert Ball: [email protected]
Elizabeth Roberts:
Richard Thomas:
134 Lessons Learned for DEA Practice from Health Care Applications in the UK
Structure
Introduction
2 DEA Modelling in a Health Service Context
3 Development of the Sample ofDMUs
4 Choosing the DEA Model
5 Identification of Inputs and Outputs
6 Results from the Basic Model
7 Introduction of Weight Restrictions
8 Evaluation of the Approach
Appendix 1
Appendix 2
References
Lessons Learned for DEA Practice from Health Care Applications in the UK 135
1 Introduction
Technical and theoretical developments to the DEA methodology have been widely
discussed over recent years, following its instigation by Charnes, Cooper and Rhodes
(1978). However, it has been observed that, despite the fact that DEA appears to be
ideally suited to the measurement of efficiency in health care organizations, there are
few examples to be found showing DEA being used as a practical tool for efficiency
assessment by managers. There are only a handful of papers relating its application to
the provision of health services in the UK, although there are a greater number using
data from the United States. In many cases, the DEA literature relating to health care
organizations appears to use the health data to illustrate facets of the DEA technique,
rather than using the DEA technique to investigate health care data, such as the
situation commented on by McMillan (1987).
Coincidentally, the search for adequate and appropriate methods for the measurement
of efficiency and performance has become a 'central preoccupation for public sector
organizations' (Flynn, 1986). The potential reasons for this development are
numerous:
1. There are now severe restrictions on the level of public finance
available and constraints have been placed on all areas of spending by
government;
2. Public sector services have been perceived to be inherently inefficient
and in need of new measures to target these inefficiencies (Mooney
and Ludbrook, 1984);
3. The role of management within the NHS has changed, giving more
emphasis to the aim of improving the efficiency with which resources
are used (Levitt and Joyce, 1987);
4. Government agencies, such as the Audit Commission, have adopted
the pursuit of efficiency as one of their primary responsibilities.
Consultation with health care managers, and the presentation of a discussion paper at
the Young OR Conference in the UK, has enabled the authors to gain an insight into
the suitability of the DEA technique for addressing the performance assessment
136 Lessons Learned for DEA Practice from Health Care Applications in the UK
requirements of the Health Services in the UK. This paper explores these issues using
data from the NHS in Scotland.
2 DEA Modelling in a Health Service Context
If DEA is to be used to successfully examine the efficiency of a group of hospitals, or
other health care organizations, there are several key stages to the modelling process.
The following diagram, adapted from Golany and Roll ( 1989), has been proposed as a
guide as to what these different stages should be:
1 Define Population ofDMUs J ~
I Set Goals for Analysis I ~
.I Select DMUs to be Compared ~~;::.:::-------1 ·1 ~ I
~ ! ' ; i
List Relevant Factors I '
l j i
~ I I I Set Factors' Examine Factors I
I ... J
Measurement Scales (Judgement)
I
l l y Define Production Examine Factors II I Relationships (Correlations) I
+ • I
Formalise Initial
I I Examine Factors III
1-Model (Trial Runs) --- --· ------------·
I l Formalise Final Model I
l Computer Environment I Present Initial Results I
I ~ ~ + I Analyse by Factor I I General Conclusions I
and Special Analyses I Analyse by I
Individual DMU
Figure 1: An Application of Data Envelopment Analysis to Health Care Data
Lessons Learned for DEA Practice from Health Care Applications in the UK 137
Whilst the above diagram is useful as an indicator of the complexity of the DEA
modelling process, a more simplified structure will be used here, with the key stages as
follows:
1. Definition of the Sample ofDMUs;
2. Formulation of DEA Model Type and Orientation;
3. Selection of Variables to be used as Input and Output Factors;
4. Results of the Preliminary Analysis;
5. Revisions to the Model - Sample Size, Variables Included, Addition
of Weight Restrictions;
6. Results from the Revised Model;
7. Evaluation of Results- Sensitivity and Robustness.
3 Development of the Sample of DMUs
As was referred to above, the investigation presented is based on the analysis of
data from the NHS in Scotland. However, this is an extremely complex
organization, as can be seen in the diagram below from the Scottish Office
(1998), depicting its organizational structure:
138 Lessons Learned for DEA Practice from Health Care Applications in the UK
Health Boards X 15
Figure 2: The NHS in Scotland (1998)
- Policy
- Performance Management
- Local Strategy
- Monitoring of Implementation
There are many levels of decision-making in the Scottish NHS and analyses using
DEA could be carried out at a variety of levels, such as, investigating the relative
efficiency of the health boards, of which there are fifteen. However, given the previous
experience of the authors and the availability of the relevant data, the decision was
taken to focus on the analysis of individual hospitals. This is the main point at which
services are delivered. This also reflects the current attention given to this by the UK
Government and the Department of Health.
Lessons Learned for DEA Practice from Health Care Applications in the UK 139
The majority of hospital-based activity in Scotland is managed through acute hospital
trusts, as referred to in figure 2. (Some trusts concentrate on specific areas of health care
such as community, mental health and dental services, or provide only one type of
service such as outpatient facilities, and are not included here.) The majority of the acute
trusts provide a cross section of services at the hospital sites under their control, with the
size and scope of the trusts varying quite considerably.
According to the viewpoints of the health care managers interviewed, there are many
different types of trust, providing services at a number of hospital sites. Therefore, to use
all the trusts in a single DEA sample would not be appropriate or useful.
Therefore, the level of service provision to be used for comparison will be the hospital
level, one level down from the trusts in the organizational structure. However, it is still
necessary to reduce the sample still further, as there are over 200 hospitals in Scotland,
which also vary markedly in size and the services they offer. The approaches in the
literature show how the sample of all hospitals can be reduced according to some
specified characteristics. For example, Byrnes and Valdmanis (1995) focused on
community hospitals, Sherman (1984) used a sample of teaching hospitals and Grosskopf
and Valdmanis (1987) used large, urban hospitals.
These types of study, however, developed their sample from very large groupings of
hospitals. With Scottish data, refining the sample according to very strict characteristics,
such as location, size and type, would result in a large number of very small samples.
There are only a handful of hospitals in each area, which are often of different types, such
as community, acute, teaching, children's, maternity and geriatric. The hospitals in
Scotland can actually be divided into some 49 categories, called functional
classifications, and defined by the ISD (1996). It is necessary to make some
generalisations in order to obtain an appropriate sample with an adequate number of
DMUs.
Therefore, following the approach of Parkin and Hollingsworth (1996) and Ozcan et a!.
( 1992}, the sample chosen for the evaluation consisted of those hospitals that are
classified as 'acute', of which there are 74 in Scotland. In terms of sample size, the figure
of 74 DMUs is consistent with the sample of California hospitals used by Grosskopf and
Valdmanis (1987). These 'acute' hospitals cover 15 functional classifications and the
hospitals in these classes can be divided into four broad categories:
140 Lessons Learned for DEA Practice from Health Care Applications in the UK
1. General Hospitals - may have a teaching element and a maternity
department, will cover the majority of surgical functions but not
necessarily all of them and may vary in size quite considerably;
2. GP Practitioner Cottage Hospitals - will have a limited surgical
function, may have a maternity department but no teaching element;
3. Mixed Specialist Hospital - may have a maternity department, will
cover the basic surgical functions but not highly specialised treatment;
4. Specialist Children's Hospitals - may have teaching element, will
cover complete range of paediatric services.
The initial sample of 74 DMUs, the acute hospitals in Scotland, on which the primary
stages of the analysis were carried out (and referred to in Roberts and Ball, 1997) was
found to have several limitations. These could seriously affect the usefulness of the
results obtained and their acceptability in a health care environment. The widely held
belief within many hospitals, which was stressed by two of the health care managers
interviewed, was that each hospital is 'unique' in its operation, facilities and
environment. This is to such an extent that they believe it cannot be sensibly compared
with any other hospital, unless they can be convinced that the other hospital operates
under the same or very similar constraints.
Analysis of the actual data relating to this sample also identified a further problem in
terms of the large variations in the level of services offered by each hospital. Many of
the hospitals did not over the complete range of services included as output factors.
Therefore, the output data in particular was found to contain many zero values, which
could potentially have a distorting effect on the calculation of efficiency scores using
DEA.
In particular, several DMUs were identified with unusual and even unique output
patterns, which reflected the concerns of the health service managers. To counteract
this, the sample was reduced to 47 DMUs, which were, therefore, a more acceptable
set of DMUs in terms of their homogeneity, based on analysis of their functional
classifications. Further revisions to the sample of DMUs, and some alternative
approaches to the measurement of hospital-based activity, have also been considered
Lessons Learned for DEA Practice from Health Care Applications in the UK 141
and will be debated following the identification of appropriate input and output factors
and the development of an initial DEA model.
4 Choosing the DEA Model
In the majority of investigations into health service organizations, the original DEA
model developed by Charnes, Cooper and Rhodes (1978) and denoted by CCR has
been used for the analysis of efficiency. (See Hollingsworth and Parkin (1995),
Sherman (1984) and Ehreth (1994).) Thus, the majority of models have been
developed with the assumption of constant returns to scale (CRS) and the free
allocation of weights. Parkin and Hollingsworth ( 1996) introduced the idea of
examining the efficiency of hospitals, whilst taking account of varying returns to scale
(VRS). However, the analysis here, will employ the CCR model for two important
reasons:
1. As mentioned above, most models used by previous authors to
measure hospital efficiency have used the CCR model, probably
because it is the simplest form of the DEA methodology, being easier
to understand and interpret.
2. The data sample employed here, which consists of 'acute' hospitals,
can be said to be homogeneous, meaning the constant returns to scale
assumption should still be valid (Smith and Mayston, 1987).
In addition to the choice of model, it is also necessary to determine its orientation, be it
input minimisation or output maximisation. In this case, the input minimisation model
will be utilised, since most of the thinking in the health care sector has reflected a wish
to minimise costs without reducing services. Figure 3 represents the DEA model to be
solved in the following analysis, expressed in ratio form.
142 Lessons Learned for DEA Practice from Health Care Applications in the UK
m
MINIMISE: ea = I V d i a , i= 1
m S
SUBJECT TO: I V d i} - I UrOrJ '2:. 0, j = l.. .. a .... n i= I
s
AND: I U rOra = 1 r=l
U, and V; > E FOR ALL rAND i (E being a very small positive number)
s = number of outputs
m = number of inputs
Figure 3: Input Minimisation Orientation of the CCR Model
5 Identification of Inputs and Outputs
The selection of the variables is an extremely important element of the modelling
process - if inappropriate variables are chosen, the results of the analysis will yield no
benefit. In Scottish Acute Hospitals, the cases can be divided into six main categories:
inpatients, consultant outpatients, PAM (Professions Allied to Medicine) outpatients,
day cases, day patients and Accident and Emergency attendances.
The first step in the selection of variables is the identification of all possible factors
and those for which for the data is not available must then be eliminated from the
model. In this analysis, a potential input factor would be the 'total number of trained,
learning and other nurses', as used by Hoiiingsworth and Parkin (1995).
Unfortunately, such data was not available. Nevertheless, there are a large number of
inputs and output factors for which the data was available that could therefore be
employed, as shown in Table 1.
Lessons Learned for DEA Practice from Health Care Applications in the UK
Inputs 'Average No. of Staffed Beds' 'Capital Charge' 'Total Costs' 'Total Direct
Costs', 'Total Allocated Costs' 'Total Inpatient Expenditure' 'Total Outpatient
(8) Expenditure' 'Total Daycase Expenditure'
Outputs 'Total Inpatient Discharges' 'Inpatient Discharges, Surgical' 'Inpatient
(30) Discharges, ITU' 'Inpatient Discharges, Medical' 'Inpatient Discharges,
Obstetrics and Gynaecology' 'Inpatient Discharges, Other' 'Total Inpatient
Days' 'Inpatient Days, Surgical', 'Inpatient Days, Medical' 'Inpatient Days,
Obs and Gynae' 'Inpatient Days, Other' 'Inpatient Days, ITU' 'Consultant
Outpatients Attendances, Other including ITU' 'Consultant Outpatients
Attendances, Surgical' 'Consultant Outpatients Attendances, Medical' 'Total
Consultant Outpatient Attendances' 'Total New Consultant Outpatient
Attendances' 'Total A&E Attendances' 'Total New A&E Attendances' 'Total
Daycase Attendances' 'Daycase Attendances, Medical' 'Daycase Attendances,
Surgical' 'Daycase Attendances, Other' 'Total Daypatient Attendances'
'Daypatient Attendances, Medical' 'Daypatient Attendances, Other including
Mental Health' 'Total PAM Outpatient Attendances' 'Total New PAM
Outpatient Attendances' 'Total Outpatient Attendances, A&E, Consultant and
PAM'
Table 1: Potential Variables
143
A case could be made for the inclusion of all of the above, based on their previous use
in the DEA literature, the fact that they intuitively describe some element of hospital
performance or they are acceptable to health care experts having been used in official
internal investigations. Nevertheless, to employ all 38 factors would be inappropriate
and unnecessary. It is therefore essential to reduce the list to more manageable
proportions and several approaches have been identified to narrow down the list of
potential variables. These are expert advice, previous use in the literature, evaluation
and manipulation of the data using statistical methods and, finally, heuristics. Each of
these is discussed below.
(i) Opinion of Experts: It is our belief that this should be an essential part of the
model building process. A member of the NHS Executive with whom the issues
surrounding DEA application were discussed suggested that 'inpatient discharges'
should always be used as an output in place of 'inpatient days', which is not seen as a
reflector of efficiency. Support was also expressed for disaggregating inpatient activity
factors, as 'inpatients' may be seen as too broad a category of hospital activity. The
144 Lessons Learned for DEA Practice from Health Care Applications in the UK
importance of including some measure of day case activity was also stressed. In terms
of inputs, apart from favouring financial measures, the possible redundancy of the
'average staffed beds' was highlighted as this is covered to a certain degree by 'capital
charge'.
(ii) Previous Use in the Literature: The use of 'inpatient discharges' rather than
'inpatient days', as proposed by the experts, is supported by the literature (See Ehreth,
1994). Different levels of disaggregation of the inpatient factor have been identified,
such as, the three categories (acute, ICU and surgical) used by Grosskopf and
Valdmanis (1987) and the four categories (surgical, medical, obstetrics and others)
used by Parkin and Hollingsworth (1995). Outpatient discharges and A&E attendances
have also frequently been used as output factors in the literature. Turning to the input
factors, evidence from the literature suggests that these should reflect three main
characteristics, to be measured in a number of ways:
1. Hospital Size - number of staffed beds, net plant assets, number of
admissions, fixed assets, bed days available.
2. Staffing Levels - number of full time employees, number of staff in
each category, payroll expenditure, direct salary costs.
3. Supply Expenses - operational costs, total value of supplies, cost of
drug supplies.
(iii) Evaluation of the Data: Disaggregation of the output factor 'inpatient discharges'
was not practical, as many of the DMUs do not have data entries for all categories.
There were several zero or missing values for some of the output categories,
particularly 'day patients' and 'PAM outpatients', as these services were not provided
at all of the hospitals. Statistical analysis concluded that there were several pairs of
highly correlated output factors. The policy of removing one of each pair of highly
correlated factors is not however as generally accepted in DEA as it is in regression (as
discussed by Nunamaker, 1985).
A further statistical technique is the use of regression to establish that the factors to be
included are related technical efficiency, rather than being arbitrary measures of input
and output. Golany and Roll (1989) advocated regression as a means of 'eliminating
Lessons Learned for DEA Practice from Health Care Applications in the UK 145
redundancies and reducing the list of factors', to be used not as 'reliable rules but
merely as indicators for a need to examine some of the factors more closely.' Using a
multiple regression procedure, the key output factors were identified as the measures
of inpatient, day case and outpatient activity.
(iv) Heuristics: The final approach is to run several alternative models and then analyse
the results to determine which groupings of inputs and outputs best describe the health
care situation being investigated. This is an approach frequently found in the literature,
notably Ehreth (1994).
Taking on board the important factors identified above, and trying several DEA models,
the mix of variables chosen for this DEA application were determined to be:
INPUTS: 'total direct costs' (TDC), 'total allocated costs' (T AC) and
'capital charge' (CAP).
OUTPUTS: 'total inpatient discharges' (TID), 'total consultant
outpatient attendances' (COA), 'total accident and emergency
attendances' (AEA) and 'total day case attendances' (DCA).
6 Results from the Basic Model
Following the selection of the variables and the DEA model type, the results from the
DEA model can be obtained; in this case specially written DEA software package
developed at the University of Stirling was utilised.
The model selected, as we have seen from the previous section, had four outputs ('total
inpatient discharges', 'total consultant outpatient attendances', 'total accident and
emergency attendances' and 'total day case attendances') and three inputs ('capital
charge', 'total direct costs' and 'total allocated costs'). The simplest form of the DEA
model, the CCR model with an input minimisation orientation, was used and the
results from this model were investigated without employing weight restrictions.
Of the 47 DMUs in the sample, 20 were rated as 100% efficient, with the mean
efficiency score being 87.98%, with a standard deviation of 15.08. A full breakdown
of the efficiency analysis is given in appendix 1. Further to this, the results from the
model can be analysed in numerous ways, such as, focusing on the least efficient
146 Lessons Learned for DEA Practice from Health Care Applications in the UK
DMUs, in order to attempt to categorise the characteristics of inefficiency. Table 2
summarises the DEA results for the five least efficient units and it can be seen that
each of them is dominated by the contribution of the 'total inpatient discharges' factor.
DMU Efficiency Non-zero Dominant Dominant RatioofTDC Score Output Input Factor Output Factor toTAC
Categories (Virtual) (Virtual)
#40 66.0 COA TDC TID 3:2
#34 53.8 COA TDC TID 3:2
#20 52.0 COA,AEA TDC TID 7:4
#41 51.4 COA,AEA TDC TID 5:4
#25 45.2 COA,AEA, TAC TID 4:1 DCA
Table 2: Analysis of the Inefficient DMUs
Similarly, the efficient DMUs can also be analysed. 12 of the 20 efficient DMUs have
more than 300 staffed beds but the DMU ranked as number 1 in the overall list of
efficiency scores is one of the smaller hospitals, with 102 'average staffed beds'. (The
ranking of efficient DMUs is based on the number of occasions they appear in the
reference groups of the inefficient units). Also, 4 of the DMUs are included in the
reference set for 10 or more of the inefficient units (#30, #12, #13 and #14). These all
have 'total direct costs' as the dominant input factor and either 'total inpatient
discharges' or 'total day case attendances' as the dominant output factor (based on
virtual input and output values). The efficient DMUs can be further analysed by
refinements to the DEA model, such as super-efficiency and the development of cross
efficiency matrices.
Inefficient DMUs can be analysed in tum to determine areas for potential
improvement, which is useful for health care managers. For example, DMU #1 has an
efficiency rating of 95.23% and could achieve a 100% efficiency score by reducing
'total direct costs' from £59, 679k to £56, 833k, that is, a decrease of 3 million pounds.
Alternatively, the results from the model can be analysed in more general terms,
primarily in terms of the allocation of the factor weights. Table 3 summarises the
contributions made by each of the factors, according to the virtual weights calculated
for each of them, in order to identify the most influential of the inputs and outputs.
Values have been calculated by excluding the zero values and hence the means do not
sum to 100 for the inputs and outputs, as might be expected. It shows that the 'total
Lessons Learned for DEA Practice from Health Care Applications in the UK 147
direct costs' factor is found to be the dominant input, contributing most to the
efficiency score for 43 of the 47 hospitals.
This is confirmed by the mean value for the TDC virtual input being 89.7%. For the
outputs, it is 'total inpatient discharges' that appears to be the most dominant factor, as
it contributes most to the efficiency scores for over half of the DMUs.
CAP TDC TAC COA AEA TID DCA
mean 9.88333 89.7093 35.6 46.86 28.8611 59.6436 21.9742
stdev 6.85766 11.4837 37.135 24.8859 32.9552 33.282 23.5668
min 1.1 57.7 2.4 12.7 0.3 0.2 0.7
max 32.5 100 100 99.3 92.1 100 100
count 24 43 17 25 18 39 31
dom 0 43 4 11 5 25 6
KEY: Mean/StDev/Max/Min: calculated by excluding zero values for virtual weights. Count: Number of times input/output factor is used in efficiency calculations. Dom: Count of the number of times that factors contributes most to the efficiency score calculation. Factors codes as defined above.
Table 3: Analysis of the Virtual Weights
Table 4 illustrates the distribution of the numbers of factors involved in the efficiency
calculations, showing that none of the efficiency scores have been calculated using all
seven factors - for every DMU, at least one factor has been given a virtual weighting
of zero.
I No. of Factors 2 3 4 5 6 7
Count 3 8 17 15 4 0
Table 4: Distribution of Factors Used to Calculate Efficiency Scores.
Taking tables 3 and 4 in combination illustrates one of the major issues for debate in
the application of the DEA methodology, that is, the free allocation of weights. The
information they contain can be used in the development of weight restrictions, in
conjunction with the views of the health care managers, to be examined next.
148 Lessons Learned for DEA Practice from Health Care Applications in the UK
7 Introduction of Weight Restrictions
As was seen in table 3, the efficiency scores for each DMU have been calculated based
on contributions from some but not all of the input and output factors. Table 5 to
follow, which presents a selection of both efficient and inefficient DMUs, shows the
widely varying patterns of factor weights associated with the basic DEA methodology.
Virtual Factor Weights(%)
DMU CAP TDC TAC COA AEA TID DCA
#1 7.8 92.2 0 31.3 0 68.7 0
#7 7.7 0 92.3 13.4 85.0 0 1.6
#12 0 100 0 12.7 0 43.1 44.3
#16 0 64.9 35.1 69.5 0 12.0 18.5
#26 0 100 0 33.8 0 37.7 28.5
#42 11.3 88.7 0 20.9 1.9 73.0 4.2
Table 5: Virtual Factor Weights for Selected DMUs
According to Wilkinson ( 1991 ), these widely varying weighting patterns, and the fact
that some of the selected factors are completely excluded from the efficiency
calculations, are likely to produce results of limited value. Thus, the applicability of
weight restrictions has become a contentious issue in DEA debate. A variety of
methods have been proposed to actually impose weight restrictions, ranging from the
introduction of factor inequalities to the development of closed cones, within which
the factor weights may vary to a proscribed degree. Some of the key papers on this
issue are Dyson and Thanassoulis (1987), Thompson et al (1990), Wong and Beasley
(1990), Charnes eta!. (1989), Thanassoulis eta!. (1995), Roll and Golany (1993) and
Roll eta!. (1991). The reasons for the introduction of weight restrictions in a health
service context have been summarised as follows by Ball et a!. ( 1997):
• A DMU that has specialised in a particular area to the neglect of
others currently has more chance of being classified as efficient than the
good all-rounder;
• The lack of discrimination, given a reasonable number of inputs and
outputs, is unsatisfactory, as most DMUs will be 100% efficient.
Lessons Learned for DEA Practice from Health Care Applications in the UK
Eliminating factors is conceptually unsound and a very crude form of
weight limitation - a variable gets a weight of either zero or one;
• In many problems, not all inputs contribute to the production of every
output. This raises the possibility of reaching 100% efficiency on the
basis of a meaningless ratio;
• Allowing some inputs and outputs to be more highly weighted than
others may be appropriate, where specialist knowledge or policy
suggests this to be sensible.
149
In this investigation, weight restrictions have been introduced by attaching constraints
to the virtual inputs and outputs, following the approach presented in Ball and Roberts
( 1998), in which several different scenarios are developed to determine the impact a
range of weight restriction options.
However, the dilemma in applying weight restrictions is in finding the right balance. If
the restrictions are too loose then their adoption has no effect. On the other hand, very
tight restrictions leave no scope for flexibility and the resulting model could be
infeasible.
In this case, five additional scenarios have been developed to incorporate weight
restrictions, based on the information obtained in the initial analysis and the viewpoint
of the health care managers. Table 6 summarises the five scenarios, with the results
from the weighted models shown in summary form in table 7 and in greater detail in
appendix 2. For those factors where the maximum weight specified is 100%, the
weight restriction applied was simply a minimum constraint. It will obviously not be
possible for that factor to achieve a virtual weight of 100%, due to the minimum
constraints applied to the remaining factors.
Weighting on Virtual Weight for Each Factor (Min- Max%)
Scenario CAP TDC TAC COA AEA TID DCA
I 5- 25 50- 100 5-25 20-35 20-35 20-35 20-35
2 10- !00 10- 100 10- 100 5- 100 5- 100 25- 100 25-100
3 10- 50 10-50 10-50 5-50 5- 50 25-50 25-50
4 I I I 5- 50 5- 50 25-50 25-50
5 I I I I I 25-50 25-50
Table 6: Weight Restriction Scenarios
150 Lessons Learned for DEA Practice from Health Care Applications in the UK
In relation to the outputs, the following scenarios are explored: each output should
contribute equally to the efficiency score, each output should contribute something to
the efficiency score and finally, some of the outputs should contribute more than
others to the efficiency score for each DMU. These represent different approaches to
health care provision in hospitals, in terms of the priorities of patient care. The input
weighting strategies reflect the importance attached to the direct costs, although
allocated costs should not be ignored, as they represent the implied cost of
'bureaucracy'.
Model or Weighting Scenario
Basic 1 2 3 4 5
Mean Score 87.98 67.74 70.49 65.07 72.70 74.53
St. Dev 15.08 22.61 21.97 22.48 21.84 22.11
No. of Eff. DMUs 20 4 4 4 6 7
Minimum Score 45.2 9.7 7.7 7.4 8.0 8.0
Table 7: Impact of Weight Restrictions
Just four DMUs are efficient in all scenarios (#2, #3, #13 and #43), suggesting that
their rating as efficient is robust. However, the DMU ranked number 1 with no
restrictions to the weights (#30) does not achieve an efficiency score under any of the
alternate weighting scenarios, dropping to 29 in the ranking under scenario 5. The
most significant impact of introducing weight restrictions is on DMU # 35, which is
efficient under the basic model and scores less than 10% under all the different
scenarios and is ranked 47th. This is due to the fact that it can no longer base its
efficiency score on a dominant contribution (99.7% in the basic model) from just one
output, 'total consultant outpatient attendances'.
Detailed analysis of the results from the scenarios can be used to illustrate aspects of
performance for each of the hospitals, such as, identifying the good all-rounder or
those particularly efficient in one aspect of service provision. Additional scenarios
could also be developed to investigate this further, in order to find the most robustly
efficient DMUs who remain unaffected under all scenarios. In addition, this type of
analysis may raise such as, whether all of the DMUs in the sample are appropriate for
comparison- a point stressed firmly by many of the health care managers interviewed.
Further analysis of the data may also be appropriate at this stage, as strong weight
restrictions will significantly affect some of the DMUs with more unusual data
Lessons Learned for DEA Practice from Health Care Applications in the UK 151
patterns or a strong emphasis in one particular area. For example, table 8 shows how
efficiency scores are related to functional class for the basic model and for the weight
restricted model denoted by scenario 5. The hospitals in class 2 appear much more
resilient to the inclusion of weight restrictions, with just a small reduction in the mean
efficiency score, when compared with the hospitals in classes 5 and 7+.
Functional No. of Mean Score No. of Mean Score No. of Efficient Codes DMUs
(Basic Model) Efficient (Scenario 5) DMUs (Seen. 5)
DMUs (Basic)
1 7 97 4 90 2
2 10 94 6 90 2
5 10 78 2 56 0
7+ 8 89 5 55 0
11 6 77 0 72 0
12 6 95 3 90 3
Overall 47 88 20 75 7
Table 8: Efficiency Scores and Functional Class
The result of such analysis may be to redefine the samples being used for evaluation,
by, for example, assessing efficiency for the hospitals in each functional class in turn,
if appropriate sample sizes can be obtained.
8 Evaluation of the Approach
Interpreting the above analysis, in terms of evaluating the choice of model, sample and
variables and the results obtained, is clearly an important aspect of the DEA
application procedure. From the perspective of the DEA technology, the sample
provided an appropriate number of DMUs, in relation to the number of variables
included, and the results could be interpreted in a variety of ways. However, by taking
the perspective of the health care manager and their views on the methodology, those
aspects of the DEA technique which are most significant to its acceptance in the health
services have been identified:
1. The selection of the variables - ensuring that all those included are
relevant and contribute to the efficiency calculations for each of the
152 Lessons Learned for DEA Practice from Health Care Applications in the UK
DMUs, leading to the introduction of weight restrictions, devised
through consultation rather than applied arbitrarily.
2. The definition of appropriate samples - developing carefully
selected samples of DMUs, chosen for their inherent similarity rather
than according to externally given definitions. These may represent
one aspect of hospital activity, rather than comparing across large
groups of hospitals as a whole, or be much smaller samples.
3. The presentation of the results - moving away from generalising
about average efficiency scores and focusing on the performance of
each individual DMU, with particular reference to target setting for
improving efficiency and the importance of identifying peer groups.
Appendix 1:
DEA Results for the Basic Model
Efficiency Reference Reference
DMU Score Ranking Set Count Group
I 95.2 23 0 2 II 14
2 100 7(1) 6 2
3 100 10(1) 3 3
4 85 32 0 2 3 14 16
5 95.7 21 0 2 3 16
6 100 16(1) 2 6
7 100 17(1) I 7
8 77.5 38 0 2 6 13 16
9 79 37 0 2141643
10 100 20(1) I 10
II 100 6(1) 7 II
12 100 2(1) 13 12
13 100 3(1) 12 13
14 100 4(1) 10 14
Lessons Learned for DEA Practice from Health Care Applications in the UK 153
15 91.1 26 0 II 14 30
16 100 5(1) 7 16
17 93.8 24 0 13163547
18 83.9 33 0 12 30
19 77.2 39 0 12 30
20 52 45 0 12 30
21 86.2 30 0 14 30 33
22 71.5 41 0 12 13 27 30
23 100 18(1) 23
24 71 42 0 27 31 35 45
25 45.2 47 0 13
26 90.1 28 0 35 45 47
27 100 9(1) 4 27
28 73.8 40 0 II 13 30
29 83.3 34 0 II 14 30
30 100 I (I) 18 30
31 100 14(1) 2 31
32 100 19(1) 32
33 100 15(1) 2 33
34 53.8 44 0 12 27 30
35 100 8(1) 6 35
36 92 25 0 13
37 85.6 31 0 12 13 30 35
38 80.1 36 0 13 14 16 30 35
39 86.9 29 0 12 30
40 66 43 0 12 13 14 30
41 51.4 46 0 12 30
42 90.3 27 0 1112143043
43 100 II (I) 3 43
44 95.2 22 0 II 12 13 30
45 100 12(1) 3 45
46 82.2 35 0 12 13 30
47 100 13(1) 3 47
154 Lessons Learned for DEA Practice from Health Care Applications in the UK
Appendix 2:
Efficiency Results Using Alternative Weighting Scenarios
Basic Scenario 1 Scenario 2 Scenario 3 Scenario 4 ScenarioS
DMU Score Rank Score Rank Score Rank Score Rank Score Rank Score Rank
1 95.2 23 84.1 12 82.9 15 79.9 13 84.7 16 84.9 17
2 100 7(1) 100 3(1) 100 2(1) 100 2(1) 100 5(1) 100 5(1)
3 100 10(1) 100 4(1) 100 4(1) 100 4(1) 100 6(1) 100 7(1)
4 85 32 72.2 23 70.8 27 70.7 20 70.8 29 71.2 30
s 95.7 21 21.2 46 53 38 51.4 37 53.9 38 86.3 16
6 100 16(1) 62.3 30 95.7 9 81 12 89.4 II 96.8 II
7 100 17(1) 85.7 10 85.5 12 85.4 9 86.1 14 87.4 15
8 77.5 38 72.6 22 75.4 23 72.6 19 76.6 23 77.3 26
9 79 37 74.9 21 75.7 22 73.7 17 76.4 24 77.1 27
10 100 20(1) 76.7 17 86.2 II 81.3 II 86 15 88.2 14
11 100 6(1) 96.5 6 97.1 7 96.6 6 97.6 8 98.1 10
12 100 2(1) 91.7 8 97.6 6 92.6 7 100 4(1) 100 3(1)
13 100 3(1) 100 I (I) 100 1(1) 100 I (I) 100 1(1) 100 1(1)
14 100 4(1) 97.4 5 97.9 5 97.9 5 97.9 7 98.6 8
IS 91.1 26 76.2 18 72.7 24 69 22 76.3 25 76.4 28
16 100 5(1) 88.9 9 93.4 10 91.5 8 95.4 9 98.2 9
17 93.8 24 85 II 84 13 76.6 15 88.5 13 89.7 13
18 83.9 33 50.6 38 48 39 37.5 40 49.7 39 49.9 39
19 77.2 39 57.6 33 64.2 33 49.9 38 69.4 30 69.4 33
20 52 45 37.6 42 41 41 34.3 42 45.4 41 46.1 41
21 86.2 30 46.1 39 40.7 42 35.5 41 42.2 43 42.8 43
22 71.5 41 67.5 28 64.8 31 59.3 29 67.5 33 67.9 34
23 100 18(1) 31.3 44 26 46 25.5 46 26.1 46 26.2 46
24 71 42 68.2 26 64.5 32 56.2 32 71 28 71 31
25 45.2 47 23.7 45 33.1 45 28.7 45 35.4 45 36.5 45
26 90.1 28 71.6 25 78.4 18 65.5 25 89.5 10 90.1 12
27 100 9(1) 67 29 60.8 35 56 33 61.2 36 61.9 37
28 73.8 40 57.3 34 56.6 37 52.6 36 57.5 37 57.7 38
29 83.3 34 67.6 27 67.6 29 67.3 23 69.2 31 70.5 32
30 100 1(1) 80 15 70.9 26 61.9 27 73.7 27 73.7 29
31 100 14(1) 60.8 32 61.4 34 53.2 35 63.9 35 63.9 36
Lessons Learned for DEA Practice from Health Care Applications in the UK 155
32 100 19(1) 82.4 13 70.5 28 60.4 28 78.5 21 78.5 23
33 100 15(1) 53.7 37 46.4 40 44.2 39 46.9 40 47.2 40
34 53.8 44 37.8 41 35.1 44 30.3 44 36.9 44 37.3 44
35 100 8(1) 9.7 47 7.7 47 7.4 47 8 47 8 47
36 92 25 75.5 20 81.2 16 79 14 82.2 19 83.3 19
37 85.6 31 75.5 19 77.1 20 66.2 24 83.1 17 84.6 18
38 80.1 36 77.7 16 76.4 21 73.4 18 78.1 22 78.2 24
39 86.9 29 54.3 36 71.7 25 58.4 30 75.2 26 78.7 22
40 66 43 62.3 31 60.6 36 54.8 34 64.5 34 64.5 35
41 51.4 46 34 43 39.4 43 32.1 43 43.5 42 45.8 42
42 90.3 27 80.5 14 78.6 17 74.1 16 82.3 18 82.4 20
43 100 II (I) 100 2(1) 100 3(1) 100 3(1) 100 3(1) 100 6(1)
44 95.2 22 72 24 77.5 19 64.1 26 78.7 20 78.9 21
45 100 12(1) 96 7 96 8 82.8 10 100 2(1) 100 4(1)
46 82.2 35 43.2 40 65 30 57.1 31 68.7 32 77.9 25
47 100 13(1) 57.1 35 83.8 14 70.2 21 88.9 12 100 2(1)
References
MacMillan, W. D. (1987): The Measurement of Efficiency in Multiunit Public Services, Environment
and Planning, Vol. 19, pp. 1511-1524.
Flynn, N. (1986): Performance Measures in Public Sector Services, Policy and Politics, Vol. 14, No.3,
pp. 389-404.
Mooney, G. H. and Ludbrook, A. (1984): The NHS: Efficiency need not be a Dirty Word, British
Medical Journal, Vol. 288, No. 6433, pp. 1817-1818.
Levitt, M.S. and Joyce, M.A. S. (1987): The Growth and Efficiency of Public Spending, Cambridge
University Press.
Golany, B. and Roll, Y. (1989): An Application Procedure for DEA, Omega, Vol. 17, No. 3, pp. 237-
250.
Byrnes, P. and Valdmanis, V. (1995): Analyzing Technical and Allocative Efficiency of Hospitals, in:
Charnes, A., Cooper, W. W., Lewin, A. Y. and Seiford, L. M (eds.): Data Envelopment
Analysis: Theory, Methodology and Applications, Kluwer.
Sherman, H. D. (1984): Hospital Efficiency Measurement and Evaluation, Medical Care, Vol. 22, pp.
927-938.
!56 Lessons Learned for DEA Practice from Health Care Applications in the UK
Grosskopf, S. and Valdmanis, V. (1987): Measuring Hospital Performance: A Non-parametric
Approach, Journal of Health Economics, Vol. 6, No.2, pp. 89-107.
Information and Statistics Division (1996): Scottish Health Service Costs 1995-96, lSD, NHS in
Scotland.
Parkin, D. and Hollingsworth, B. (1996): Measuring Production Efficiency of Acute Hospitals in
Scotland, (1991-1994): Validity Issues in: Data Envelopment Analysis, Working Paper,
Department ofEpiderniology and Public Health, University of Newcastle.
Ozcan, Y. A., Luke, R. and Haksever, C. (1992): Ownership and Organizational Performance: a
Comparison of Technical Efficiency Across Hospital Types, Medical Care, Vol. 30, pp. 781-
794.
Roberts, E. and Ball, R. (1997): Efficiency and Performance Assessment in the Health Service- Using
Ideas of Policy to Develop Practical DEA Models, Paper Presented at the First DEA
Symposium in France, Marseille, June 26th - 28th, 1997.
Charnes, A., Cooper, W. W. and Rhodes, E. (1978): Measuring the Efficiency of Decision- Making
Units, European Journal of Operational Research, Vol. 2, No.6, pp. 429-444.
Hollingsworth, B. and Parkin, D. (1995): The Efficiency of Scottish Acute Hospitals: An Application
of Data Envelopment Analysis, IMA Journal of Mathematics Applied in Medicine and Biology,
Vol. 12, pp. 161-173.
Ehreth, J. L. (1994): The Development and Evaluation of Hospital Performance Measures for Policy
Analysis, Medical Care, Vol. 32, No.6, pp. 568-587.
Smith, P. and Mayston, D. (1987): Measuring Efficiency in the Public Sector, Omega, Vol. 15, No.3,
pp. 181-189.
Nunamaker, T. (1985): Using Data Envelopment Analysis to Measure the Efficiency of Non-profit
Organizations: a Critical Evaluation, Managerial and Business Economics, Vol. 6, No. I, pp.
50-58.
Dyson, R. G. and Thanassoulis, E. (1988): Reducing Weight Flexibility in DEA, Journal of the
Operational Research Society, Vol. 39, No.6, pp. 563-576.
Thompson, R. G., Langemeier, L. N., Lee, C. T., Lee, E. and Thrall, R. M. (1990): The Role of
Multiplier Bounds in Efficiency Analysis with Application to Kansas Farming, Journal of
Econometrica, Vol. 46, pp. 93-108.
Wong, Y-H. B. and Beasley, J. (1990): Restricting Weight Flexibility in DEA, Journal of the
Operational Research Society, Vol. 41, No.9, pp. 829-835.
Lessons Learned for DEA Practice from Health Care Applications in the UK !57
Charnes, A., Cooper, W. W., Wei, Q. L. and Huang, Z. M. (1989): Cone-ratio Data Envelopment
Analysis and Multi-objective Programming, International Journal of Systems Science, Vol. 20,
pp. 1099-1118.
Thanassoulis, E., Boussofiane, A. and Dyson, R. G. (1995): Exploring Output Quality Targets in the
Provision of Perinatal Care in England Using DEA, European Journal of Operational Research,
Volume 60, pp. 588-608.
Roll, Y. and Golany, B. (1993): Alternate Methods for Treating Factor Weights in DEA, Omega, Vol.
21, pp. 99-109.
Roll, Y., Cook, W., and Golany, B. (1991): Controlling Factor Weights in DEA, liE Transactions, No.
23, pp. 2-9.
Ball, R., Monaghan, C., Thomas, R. E. and Wagner, R. (1997): Data Envelopment Analysis: A
Practical Tool for Policy Makers?, University of Stirling, Departmental Working Paper.
Ball, R. and Roberts, E. (1998): The Relevance ofDEA in the Public Sector- A Discussion Paper for
Young OR 10, Young OR 10 Keynote Papers, Operational Research Society.
Recent Advances in Data Envelopment Analysis:
An Illustrative Application to the U.S. Public Accounting
Industry
Rajiv D. Banker, Hsihui Chang, Reba Cunningham and Ram Natarajan1
Abstract
In this paper we present some recent methodological innovations in Data Envelopment
Analysis and empirical results from the application of these innovations to the U.S. public
accounting industry. This paper draws on three different working papers: Banker, Chang
and Cunningham (1999), Banker, Chang and Natarajan (1999) and Banker and Natarajan
(1999). We describe how a consistent estimator of aggregate technical and a/locative
inefficiency can be obtained using DEA models and how it can be used to derive firm-specific
estimates of a/locative inefficiency. We also provide a statistical foundation for the various
two-stage methods used in the prior DEA literature to estimate the impact of contextual
variables on productivity. Finally, we document the presence of significant technical and
a/locative inefficiencies in the U.S. public accounting industry and explain the variation in
productivity across firms through a set of contextual variables.
1 School of Management, The University of Texas at Dallas, Richardson, TX 75083-0688, U.S.A.
160 Recent Advances in Data Envelopment Analysis ...
Structure
Introduction
2 Valuating Allocative Inefficiency Using DEA Models
3 Evaluating Contextual Variables Affecting Productivity Using DEA
4 Evaluating the Productivity of Public Accounting Firms
5 Conclusion
Appendix
References
Recent Advances in Data Envelopment Analysis ... 161
1 Introduction
In less than 20 years smce its inception, Data Envelopment Analysis (DEA) has
become an important and widespread analytical tool for evaluating factors affecting
efficiency. Seiford (1996) surveys the evolution of DEA from the publication of the
Chames, Cooper and Rhodes ( 1978) study to the current state of art. He identifies
statistical tests for model specification and stochastic DEA as two important areas for
future research in DEA. While the original DEA models specify the production set
relating inputs to outputs only in terms of properties such as convexity and
monotonicity and do not impose any explicit parametric structure on the production set
or the distribution of efficiency of individual observations, statistical properties can be
derived for the DEA estimator and a variety of statistical tests can be devised if
additional structure is specified (Banker 1993, 1996).
A model specification issue in DEA that has received inadequate attention is the
evaluation of allocative inefficiency in multiple output - multiple input production
models. A recent paper (Banker, Chang and Natarajan 1999, BCN hereafter) shows
that the DEA technical inefficiency measure using a single aggregate output variable,
constructed from multiple outputs weighted by their prices, reflects the aggregate
technical and allocative inefficiency. BCN employ this result to construct statistical
tests of the null hypothesis of no allocative inefficiency analogous to those of the null
hypothesis of constant returns to scale described in Banker (1996). BCN apply the
above methodology to an analysis of the productivity of firms in the public accounting
industry in the U.S. and document the presence of significant technical, scale and
allocative inefficiencies, but do not find any significant changes in these efficiency
measures over time.
Many studies have used a two-stage procedure to evaluate a set of contextual variables
believed to explain the variation in DEA inefficiency scores (Grosskopf 1996). A
recent study (Banker and Natarajan 1999, BN hereafter) investigates assumptions
about the underlying stochastic processes that generate the data about the inputs,
outputs and contextual factors to theoretically justify procedures such as OLS
(Ordinary Least Squares), COLS (Corrected Ordinary Least Squares), TOBIT
(Tobin's, 1958, Censored Regression Model) and MLE (Maximum Likelihood
Estimation) used in the second stage analysis ofDEA inefficiency scores. In addition,
BN use the DEA+ framework developed by Gsatch (1998) and conditions for the
consistency of ML estimators identified by Greene ( 1980) to propose a methodology
162 Recent Advances in Data Envelopment Analysis ...
for a statistically consistent two-stage estimation of the impact of contextual variables
on inefficiency. Banker, Chang and Cunningham (1999, BCC hereafter) identify and
analyze contextual variables affecting the productivity of the U.S. public accounting
industry.
In this paper, we collate selected text and results from BCN, BN and BCC to describe
recent advances in DEA methodology and their application to the U.S. public
accounting industry. The remainder of the paper has the following structure. Section
2 describes the methodological advances in the BCN paper. Section 3 summarizes
salient methodological aspects of the BN paper. Section 4 describes the empirical
findings in BCN and BCC based on the application of these advances in DEA to the
U.S. public accounting industry. Concluding remarks are provided in section 5.
2 Evaluating Allocative Inefficiency Using DEA Models
Schmidt and Lovell ( 1979) describe two ways in which an observation about a
production process can exhibit inefficiency. It can be technically inefficient in the
sense that it fails to produce the maximum level of outputs from a given level of
inputs, or it can be allocatively inefficient in the sense that the marginal revenue
product of an input is not equal to the marginal cost of that input. Using a stochastic
production frontier estimation (SFE) approach, Schmidt and Lovell ( 1979) and
Kumbhakar (1987) extend the analysis of Aigner, Lovell and Schmidt (1977) and
Meussen and van den Broeck (1977) to describe the calculation of aggregate
inefficiency and its technical and allocative components. Kumbhakar ( 1996) discusses
the modeling of technical and allocative inefficiencies in both cost-minimizing and
profit-maximizing frameworks with special emphasis on multiple inputs and multiple
outputs. Kumbhakar's paper also uses a stochastic frontier framework.
Unlike the SFE-based models, which provide both estimation methods and statistical
tests for allocative efficiency, prior research dealing with allocative inefficiency in
Data Envelopment Analysis has focused only on its measurement. Banker and
Maindiratta ( 1988) describe the calculation of aggregate, technical and allocative
inefficiency through DEA-based linear programs for situations when the set of
observed output, input and price data is not consistent with profit-maximization (or
cost-minimization) for at least one firm in the sample. They define allocative
inefficiency as aggregate inefficiency divided by technical inefficiency.
Recent Advances in Data Envelopment Analysis ... 163
Implementing Banker and Maindiratta's (1988) programs involves complete
knowledge of both input and output prices as well as input and output quantities.
While the researcher may have input and output quantity data, he may not have
information about individual output or input prices, except for the knowledge that the
firms operate in the same competitive market place. Very often, the available data
consists only of aggregate revenues and costs, and quantities of multiple inputs and
multiple outputs. In such a situation, BCN show how an aggregate technical and
allocative inefficiency measure, equivalent to the one described by Banker and
Maindiratta (1988), can be calculated. They then measure allocative inefficiency as
aggregate inefficiency divided by technical inefficiency as in Banker and Maindiratta
(1988). More importantly, the BCN study describes statistical tests for determining
the presence of allocative inefficiency in the observed sample ofDMUs.
Let Yi = (Yii•· ··Yri····YR) ~ 0 and Xi = (xii····Xij··· .x1j) ~ 0, j=l, .... N be the observed
output and input vectors generated from an underlying production possibility set T=
{(X,Y)J outputs Y can be produced from inputs X} for a sample ofN firms in the same
industry.2 Each output r is sold by all firms in the same competitive market at a price
Pr- Let P=(p~. ..... , PR) be the vector of output prices. The revenue from output r for
firm j is then n~ = PrY~· Denote the aggregate revenues as ni = L p,y,1 . The technical
inefficiency e; ~ 1 of an observation (Xj,Yj) E T, measured radially by the reciprocal of
Shephard's (1970) distance function, is given by BJ =t{Xi,Yi) = sup{OJ(Xi, BYj)ET}.
Assume that the production set T is monotonically increasing and convex and the 1+8
probability density function f(8) is such that f(8)=0 if 0< 1 and fJ(O)dB>O for 8>0. I
Then, a consistent estimator of BJ is obtained as BJ by solving the following Banker,
Chames and Cooper (1984) and Banker (1993) model:
BJ= Max 0
s.t. LA., n,, ~ e n,j Vr = I, ... R
' L A.,x,, :o; xu Vi= 1, .. .1
'
2 At least one output y~ and one input x;i are assumed to be strictly positive.
(1.0)
(1.1)
(1.2)
164 Recent Advances in Data Envelopment Analysis ...
(1.3)
(1.4)
Aggregate technical and allocative inefficiency is measured in Banker and Maindiratta
( 1988) as follows:
(2.0)
Vr = i, ... R (2.1)
Vi= 1, .. .1 (2.2)
(2.3)
B,A.k :?:0 (2.4)
BCN consider the linear program in (1) after replacing the R constraints in (1.1) by a
single constraint, I A.k nk :?: Bn 1 for the aggregate revenue, and denote the resulting k
DEA technical inefficiency measure as BJ. They show that BJ is a consistent
estimator of the technical inefficiency for a derived production set with a single
aggregate output. More importantly, they also prove that BJ = BJ. The allocative
inefficiency, e; , is then calculated as BJ I e: . The equivalence of BJ and BJ is useful
for establishing the statistical consistency of the aggregate technical and allocative
inefficiency estimator.
Based on Banker ( 1993) and analogous to the tests of constant returns to scale described
in Banker (1996), BCN develop a variety of procedures to test the null hypothesis of no
allocative inefficiency against the alternative of the presence of such inefficiency based
on assumed structure for the distribution of t(9), where t(.) is an appropriate
transformation function. The following illustrate the test procedures for the case of
t(.)=ln(.):
Recent Advances in Data Envelopment Analysis ... 165
(i) If In( BJ) is distributed as exponential over [0, oo ), then under the null hypothesis of
N
z)n<B)) no allocative inefficiency, the test statistic is calculated as ~1~:;-'-- and is evaluated
z)n<BJ) j.:l
relative to the critical value of the F-distribution with (2N, 2N) degrees of freedom.
(ii) If In( BJ) is distributed as half-normal over the range [0, oo ), then under the null
N
L(1n(B)))' hypothesis of no allocative inefficiency, the test statistic is calculated as ";1;:;:.'--
L(ln(BJ))' j=l
and is evaluated relative to the critical value of the F-distribution with (N, N) degrees of
freedom.
(iii) If no such assumptions are maintained about the probability distribution of
inefficiency, then a nonparametric Smimov's test statistic given by max {FA(ln( e:)) -F8(ln( e;)) I j = 1, ... N}, where FA(.) and F8 (.) refer to the empirical distributions of
In( B) ) and In( BJ ), is used.
However, as Banker (1993, 1996) points out, while ln(Bj) and ln(BJ} are
asymptotically independent of each other they need not be independently distributed
for finite samples. For finite samples, they need not also follow the true distribution of
In( e;) and In( Bf ). An important caveat, therefore, is that the above tests are designed
for large samples, their small sample performance need to be evaluated with Monte
Carlo experimentation.
3 Evaluating Contextual Variables Affecting Productivity Using DEA
Many studies have assessed the impact of contextual factors on DEA efficiency scores
(Grosskopf 1996). In these studies, the relative efficiency of each organization is
evaluated in the first stage based on data about their input consumption and output
production. The efficiency score is then regressed on the potential contextual factors in
the second stage to identify the factors whose impact on productivity is statistically
166 Recent Advances in Data Envelopment Analysis ...
significant. Alternative second stage methods have included the use of logarithmic
transform of the relative efficiency score as dependent variable in an OLS regression,
and the use of a TOBIT procedure to reflect the fact that the logarithm of the DEA
efficiency score is bounded above by zero.
A question that has not been addressed in earlier studies is whether such a two-stage
approach is statistically valid to assess the significance of individual contextual
variables. BN address this issue by specifying the basic data generating process to
derive the appropriate statistical estimation models. They also make a contribution to
the stochastic DEA literature by presenting estimators of the inefficiency of individual
DMUs conditional on the estimated value of the composite error comprising the one
sided inefficiency component and a two-sided random noise component.
Consider observations on j = l, .... N decision making units (DMUs) containing a
vector of outputs Yi = (Yti• ... YRi), a vector of inputs Xi = (xtj, ... X 1j) and a vector of
contextual variables Zi = (ztj, ... Zsj) that may influence the overall productivity in
transforming the inputs into the outputs. Thus, for instance, in Farrell's (1957) original
setting for efficiency evaluation, the outputs Y may represent a farm's production
measured in tons of grain, the inputs X may be labor, capital and materials and the
contextual variables Z that influence productivity may be factors such as ownership of
the farm and management methods.'
The basic model in BN is described for the case of a single output, y, to maintain
direct contact with parametric stochastic frontier models. The extension to the multiple
outputs case is straightforward.4 The data generating process (DGP) described in BN
specifies the "true" production function g(X) and the generation of random variables
representing the inputs X and productivity p :<> 1. The production function g(.) is
monotone increasing and concave, and relates the input vector X to a single output y as
specified by the equation
y = g(X) * p (3)
3 Contextual variables should not be confused with non-discretionary or exogeneously fixed inputs. 4 Our extension to the multiple output case involves an additional vector of random variables specifying the proportion of each output. The data generating process then determines the output vector Yi as in the single output case on the ray defined by the vector of random variables specifying the output mix.
Recent Advances in Data Envelopment Analysis ... 167
The random variable representing productivity p is itself generated by the process
In p =- h(Z)- u + v (4)
where Z is the vector of contextual variables hypothesized to affect productivity, h(.)
;::: 0 is a monotone increasing and convex function, u represents technical inefficiency
and has a one-sided distribution, and v represents random noise and has a two-sided
distribution. The following structure is imposed on the probability density functions
generating these variables:
fz,( z,) = 0 for all z, < 0 (Sa)
fu(u) = 0 for all u < 0 (5b)
fv(v) = 0 for all v > yM (5c)
(5d)
Substituting equation (4) into (3), we obtain
In y = In g(X) - h(Z) + E (6)
where E = v - u is a composed error term whose p.d.f. is given by
v" fE(E) = Jr,(v)fu(v-E)dv (7)
If, in addition, a parametric structure is imposed for the two functions g(.) and h(.)
such that In g(X; a) is linear in its parameter a, and h(Z; ~) is linear in its parameter ~.
then
In y =In g(X; a)- h(Z; ~) + E (8)
It is evident from an inspection of (8) that it is of the form specified in parametric
stochastic frontier estimation (PSFE). Therefore, ordinary least squares estimation of
(8) will yield unbiased estimators of all a and ~ except the intercept term (Schmidt
1976). The impact of individual contextual variables z, can be assessed by evaluating
the significance of the corresponding ~' if, for instance, h(Z; ~) = ~T Z. The OLS
estimator of the intercept term in such a parametric production function, however, is
biased. This is addressed in PSFE by employing corrected ordinary least squares
(COLS), where the estimator for the intercept is obtained by solving the moment
168 Recent Advances in Data Envelopment Analysis ...
equations (Olson, Schmidt and Waldman 1980). Alternatively, maximum likelihood
estimators (MLE) can be obtained for all the parameters in (8) as in Aigner, Lovell and
Schmidt ( 1977) or Meeusen and van den Broeck ( 1977) by specifying a parametric
form for the p.d.f. ofu and v consistent with (5). These estimators, however, may not
be consistent and asymptotically normally distributed because of the upper truncation
for the p.d.f. ofv, unless Greene's (1980) conditions are satisfied.
To develop a DEA based estimation procedure that is consistent with the DGP
described in (3)- (5), BN adapt the DEA +method introduced by Gstach (1998). This
involves defining:
In g(X) = In g(X) + yM and (9a)
u = u + (VM-v) +h(Z) ~0 (9b)
Since g(X) is derived from g(X) by multiplication with a constant factor, g(X) is also
monotone increasing and concave. Substituting (9a) and (9b) into (6) yields
In y =In g(x) - u (10)
Therefore, the logarithm of the DEA inefficiency estimator, lnB, obtained by
performing DEA on the inputs-output observation (Xi, Yi) , j = 1, ... N, IS a
consistent estimator of u . As observed earlier, these concepts extend directly to the
multi-output case. See, also, Banker, Janakiraman and Natarajan (1999) for
statistically consistent estimation of general monotone and concave or convex
functional relationships.
To estimate the impact of individual contextual variables z,, BN rely on the following
relation motivated by (9b):
In e = h(Z) +E' (11)
where e = u + (VM - v) ~ 0. Since h(.) is a monotonically increasing and convex
function, a second stage DEA estimation on the pseudo "input-outputs" observations
(In 01 , Z 1 )yields the inefficiency estimator fi/j and a consistent estimator of £1 can be
(rji j - 1 )in Bj calculated as . To evaluate the statistical significance of individual z,, a
rjlj
Recent Advances in Data Envelopment Analysis ... 169
third stage DEA estimation is first performed on the pseudo observations (1nB1, Zj')
where z-s is the original Z vector without the Zs variable. Let the resulting inefficiency
estimator be lj/1-'. Under the null hypothesis that the marginal impact of Zs (i.e.
8h(Z)/8z5) is zero, the asymptotic distributions of lji and V' are identical (Banker
1996). If the asymptotic distribution of c1 is assumed to be exponential or half
normal, the null hypothesis of no impact of Zs is tested by comparing the ratios
N
I (v; j - 1 )In ej I VI j against critical values obtained
;=I
;=I
from F-
distributions with (2N,2N) or (N,N) degrees of freedom, respectively.
However, in general, such a distributional assumption for c1 may not be justified.
Therefore, BN propose an alternative approach of regressing lji-' on lji:
(12)
Since, lj/-' and lji are asymptotically identically distributed if the null hypothesis is
true, an asymptotic test of the null hypothesis is obtained by evaluating the hypothesis
~I= J.
To relate this analysis back to the initial objective of rationalizing the two stage
approach used in several prior empirical studies that regress the logarithm of the
inefficiency on the contextual variables, some additional structure is imposed by
assuming h(Z) = ~ T Z, so that the evaluation of the impact of a variable z5 is reduced to
the test of significance of the hypothesis ~s = 0. It is evident from (11) that estimators
of~ may be obtained from the relation:
- T M In B = ~ Z + V - (v- u) (13)
An inspection of the relation in ( 13) indicates that it is analogous to the estimation of
stochastic production frontiers, although the dependent and independent variables are
not inputs or outputs. Therefore, standard results for parametric stochastic frontier
170 Recent Advances in Data Envelopment Analysis ...
estimation apply in this case. Thus, OLS estimation of (13) yields unbiased estimators
/3 (Schmidt 1976), but the OLS estimator Po consistently estimates yM- E(v-u).
It may appear that TOBIT analysis could be justified as MLE if the distribution of E
were truncated Normal. In general, however, MLE yields consistent estimators if
Greene's ( 1980) conditions are satisfied. But, if the distribution of E is truncated
Normal then Greene's conditions are not satisfied, and therefore consistency property
cannot be established for the TOBIT estimator using this approach.
Greene ( 1980) identifies the following two sufficient conditions to ensure that the ML
estimator is consistent and its asymptotic distribution is Normal:
v" /,(c)= fJ,(v)f.(v-c)dv=O atE=VM
The first condition is always satisfied at E = VM. The second condition implies that
fv(VM) = 0 [e.g. Beta] or fu(O) = 0 [e.g. Gamma or Lognormal].
Suppose u is distributed Gamma (2,A.) and v as N(O,cr/) truncated above at vM. Let
( c a J ( VM a J ~ • . c,= -+-2. , c,= -+-2. , and r (.)and F (.)be the standard normal density and a, A. a, A.
distribution funtions, respectively. BN derive the p.d.f. of E = v- u as
.aJ + &
2..tl A.
f(E)= a,e( MJ[{f(c,}-j'(c,}}+c,{F'(c,}-F'(c,}}] A.' F' V
a,
(14)
Since (13) can also be expressed as E = ~Tz + yM -In B, the log-likelihood function
can be formed as 2)nf(,BrZ+VM-lnB) using (14) and this function can then be
maximized with respect to ~. A., crv and yM to consistently estimate the unknown
parameters.
BN also derive the conditional distribution of the inefficiency u given E and show that
it is truncated between 0 and yM- E. The conditional mean E(uiE) = E(t) + Var(t)/E(t),
Recent Advances in Data Envelopment Analysis ... 171
where t is distributed as N(- (e +~),a; J truncated between 0 and yM- E, is a
consistent estimator of u given E. Also, the conditional mode M(ule) is a MLE as in
Jondrow et al. (1982). The mode is given by the following:
M I Ife ~ V -~ then M(ule) = yM- E (15)
+ cr; A.
(16)
4 Evaluating the Productivity of Public Accounting Firms
In this section, we discuss the main empirical results from BCN and BCC. These
studies employ the analytical models described in sections 2 and 3 to estimate the
relationship between input, output and contextual variables for the U.S. public
accounting industry. The public accounting industry has undergone many changes in
recent years. The economic slowdown of the early 1990s, advances in information
technology and considerable growth in the demand for consulting services are believed
to have Jed to an increase in competition among the firms and significantly impacted
the economics of the profession. While there is intense debate among academics,
practitioners and regulators about the effect of competition on public accounting firms,
very little empirical evidence exists on whether they efficiently utilize their resources,
whether they allocate these resources to the different types of services to generate
maximum revenue and what firm-specific factors affect their productivity.
The data used in the two studies were obtained from Accounting Today's annual
surveys of the top 100 accounting firms for the three years 1995, 1996 and 1997.
Accounting Today constructed the database from firms' responses and its own research
and estimates. This annual survey of the profession's largest practices has now
become one of the most often cited benchmarks in the business. All data reported in
the annual surveys are for domestic U.S. operations and exclude foreign holdings.
Non-CPA firms such as American Express Inc., Padgett Business Services and H&R
172 Recent Advances in Data Envelopment Analysis ...
Block are eliminated from the database. There are a total of 93, 92, and 93
observations for the years 1995, 1996 and 1997, respectively.
BCN focus on the production correspondence between service revenues generated and
human resources employed by public accounting firms. There are three service
outputs: Accounting and Auditing (A&A), Tax Services (TAX), and Management
Advisory Services (MAS), each measured in millions of dollars of revenues. A&A
includes compilations, special reports, and reviews in addition to engagements
involving the attest function. TAX includes tax research, planning and preparation
work. MAS is defined as consulting, systems development, integrating and reselling
computer equipment and software, and any other management assistance. The three
human resource input variables considered are the number of partners (PARTNERS),
the number of other professionals (PROFESSIONALS) and the number of other
employees (OTHERS). The designation PARTNERS includes all owners and
shareholders. PROFESSIONALS includes those trained to perform the accounting
and other services offered by the firm. Generally, these include staff accountants,
senior accountants, and managers. OTHERS are clerical and support personnel;
usually involved in administration, printing of reports, record keeping and the like.
Personnel costs constitute a significant fraction of total costs for the public accounting
firms. A recent national survey indicates that employee costs and partner
compensation account for 74.5% of the revenues, while capital costs are less than 7%,
for accounting practices with revenues in excess of one million dollars (Texas Society
of Certified Public Accountants 1997).
Recall the programs described in section 2 for the calculation of the technical
inefficiency BJ and the aggregate technical and allocative inefficiency e: . BCN use
the revenues from the three services as output variables and the number of partners,
professionals and other employees as input variables in the linear program described in
( 1) to calculate BJ . In contrast, the calculation of e: involves the use of the aggregate
revenue from all three services as the single output variable. The estimation of the
inefficiency measures is carried out on a year by year basis. The allocative
inefficiency, e;, is then calculated as e: I BJ.
The cross-sectional distribution of the technical, aggregate and allocative inefficiencies
is shown in table 1. The mean (median) values of technical inefficiency BJ are 1.2213
(1.1496), 1.1825 (1.1043) and 1.1941 (1.1520) for 1995, 1996 and 1997, respectively.
Recent Advances in Data Envelopment Analysis ... 173
The corresponding values for allocative inefficiency, e;, are 1.1671 (1.1459), 1.2266
(1.1907) and 1.2072 (1.1787), for 1995, 1996 and 1997, respectively. Both
distributions are skewed to the left and more than 25% of the sample firms are found
to be technically efficient. While the distributions are quite informative, additional
assumptions on the inefficiency variables are necessary to make inferences about the
level of inefficiency in the industry.
Panel A of table 2 provides approximate 95% confidence intervals for the logarithm of
the technical and allocative inefficiency measures, InBJ and !no;, respectively. Four
different types of confidence intervals are provided. The first three are for the mean of
the log-transformed inefficiency measure whereas the fourth one is calculated for the
median. CI -1 is calculated as mean(In B)± 1.96 sy};ln B) where SD(ln B) is the sample
standard deviation of lnB. Cl-2 is calculated as mean(InB)± 1.96 m:Ji(inB) assuming
lnB is distributed as exponential over [0, oo ). CI-3 is calculated as mean(InB)
±1.96mean(inB)~"- 2 assuming a half-normal distribution for lnB. CI-4 is 7rN
calculated by first sorting the observations in ascending order and then choosing the
observations given by 0.5(N+ 1) ± 0.98JN as the bounds for the interval (Lehmann
1975, p.184 ). For all three years of data, none of these intervals is found to include
zero implying that the public accounting industry operated under significant technical
and allocative inefficiency.
The null hypothesis of no allocative inefficiency is tested against the alternative of
existence of inefficiency by using F-statistics similar to those described in Banker
( 1996). Panel B of table 2 provides two tests to evaluate the null hypothesis. Test I is
based on the assumption of identical exponential distributions for In BJ and In e: under the null hypothesis while test II assumes identical half-normal disributions. All
the F-statistics presented in panel B reject the null hypothesis of no allocative
inefficiency at the 5% level for all three years of data suggesting that the public
accounting industry exhibited significant allocative inefficiency.
To test for efficiency improvements over time, BCN first pool the observations from
all three years after converting all revenue data to constant 1995 dollars and then re-
estimate e:' the aggregate inefficiency, and e;' the technical inefficiency, using a
174 Recent Advances in Data Envelopment Analysis ...
single program for all three years of data. Statistical tests of differences in In( B/ ) (or
In( Bf} as the case may be) are performed for each pair of years. For example, BCN
evaluate whether the parameter describing the distribution of firm inefficiency in 1995
equals that which describes the distribution in 1996, against the alternative that the two
parameters are unequal. The test statistics are similar to those described in Banker
(1993) to evaluate the null hypothesis of no difference in the inefficiency distributions
of two sub-samples. The tests are based on assumed exponential or half-normal
distributions for the inefficiency variable. The tests do not reject the null hypothesis of
no change in efficiency for all three years and for both distributional assumptions at
the 5% significance level. That is, there is no evidence of efficiency improvement
over the three-year period. The results are robust for the restricted sample of 79 firms
for which data are available for all three years.
BCC examine whether the variation in firm productivity can be explained by
contextual variables such as the percentage of revenues from management advisory
service business (MAS%), percentage of revenues from tax service business (TAX%),
service diversity (HERFINDEX), the number of offices (OFFICES) and whether the
firm is a Big Six accounting firm (BIG6). Herfindahl index for a CPA firm's services
is calculated as [(A&A%/100)2 + (TAX%/100)2 + (MAS%1100)2]. The higher the
HERFINDEX, the lower is the service diversity.
To evaluate the impact of different contextual variables identified by BCC, the
following version of the linear model specified in (13) is used:
(17)
lnB = ~o +~~TAX%+ ~2MAS% + ~3 HERFINDEX + ~4 /nOFFICES + ~s BIG6 + E
This model is estimated using two different methods. The first method uses OLS
estimation and the second allows the error term E to consist of two components where
E = u - v + yM where u is distributed as Gamma (2,/...) and v as N(O,cr/) truncated
above at yM and uses MLE. As described in BN and explained in section 3, while the
OLS yields consistent estimators of the coefficients of the contextual variables, the
MLE method provides consistent estimates of all the parameters for the data
generating process specified in (3) to ( 5).
Recent Advances in Data Envelopment Analysis ... 175
Tables 3 and 4 describe the results from these two estimation methods. The following
salient points emerge from the analysis of the results. Both methods produce very
similar results. This is because the MLE method estimates low values for the
parameters of the one-sided disturbance variable, u, and it estimates values of the
upper truncation point, VM, for the truncated Normal random variable, v, at 2.65, 3.07
and 3.91 standard deviations away from zero for the three years of data. It appears that
except for the identified systematic variation in inefficiency across different public
accounting firms attributable to the contextual variables, the variation in their
performance is due only to the two-sided random noise term.
Of the five contextual variables, the percentage of revenues derived from taxes
(TAX%) and the logarithm of the number of offices (lnOFFICES) have significantly
negative impact on productivity, while the Big Six status variable (BIG6) has a
significantly positive impact, under both models at the 5% significance level. MAS%
is not significantly associated with productivity. Given that the correlation between
TAX% and MAS% is of the order of -0.6 for all the three years of data, it is possible
that MAS% is not able to explain any additional variation in productivity over and
above that explained by TAX%. Accounting firms with more offices do not appear to
be more productive. In fact, there is strong evidence to the contrary. Productivity is
not found to be affected by service diversity, measured by the HERFINDEX variable,
during all the three years.
5 Conclusion
The main objective of this paper is to present some recent methodological innovations
in Data Envelopment Analysis and empirical results from the application of these
innovations to the U.S. public accounting industry. This paper draws on three
different working papers: Banker, Chang and Cunningham (1999), Banker, Chang and
Natarajan (1999) and Banker and Natarajan (1999). It documents the presence of
significant technical and allocative inefficiencies in the public accounting industry and
explains the variation in productivity across the firms through a set of contextual
variables.
We describe in this paper how a consistent estimator of aggregate technical and
allocative inefficiency can be obtained and how it can be used to derive firm-specific
estimates of allocative inefficiency. We also provide formal statistical tests to evaluate
176 Recent Advances in Data Envelopment Analysis ...
the null hypothesis of no allocative inefficency against the alternative of the presence
of such allocative inefficiency. While such tests have been available for parametric
stochastic frontier estimation methods, the BCN paper is perhaps the first in the DEA
literature to propose appropriate statistical tests for allocative inefficiency. The
theoretical advance in BN also contributes to the evolving new field of stochastic
DEA. We introduce contextual variables in Gsatch's (1998) DEA+ framework to
provide a statistical foundation for the two-stage methods used in the prior DEA
literature to analyze the impact of contextual variables on productivity. We also show
that a simple OLS regression of the logarithm of the DEA inefficiency score on the
contextual variables can provide consistent estimators of the productivity impact of
these variables.
The empirical results in BCC and BCN have important implications for the public
accounting industry. Our analysis of the data for the top I 00 U.S. public accounting
firms for the period 1995 to 1997 indicates that, contrary to common beliefs in the
prior literature, the public accounting industry operated under significant technical,
scale and allocative inefficiencies. This finding implies that public accounting firms
have not fully reaped scale economies from mergers and productivity gains from
investments in information technology and that they can generate significant cost
savings by better utilizing their human resources.
Appendix
Table 1: Cross-Sectional Distribution of the Various Inefficiency Estimators
Panel A: Technical inefficiency, Bf
YEAR Mean Std. Dev. 25% Median 75%
1995 (N-93) 1.2213 0.2449 1.0000 1.1496 1.3632
1996 (N-92) 1.1825 0.2175 1.0000 1.1043 1.3088
1997 (N-93) 1.1941 0.2189 1.0000 1.1520 1.2816
Recent Advances in Data Envelopment Analysis ... 177
Panel B: Aggregate technical and allocative inefficiency, B/ YEAR Mean Std. Dev. 25% Median 75%
1995 (N-93) 1.4239 0.2913 1.1894 1.3888 1.6140
1996 (N-92) 1.4439 0.3001 1.2047 1.4656 1.6248
1997 (N-93) 1.4367 0.2995 1.2038 1.4074 1.6831
Panel C: Allocative inefficiency, B; = B/ I B;.
YEAR Mean Std. Dev. 25% Median 75%
1995 (N-93) 1.1671 0.1457 1.0751 1.1459 1.2188
1996 (N-92) 1.2266 0.1939 1.0764 1.1907 1.3480
1997 (N-93) 1.2072 0.1838 1.0817 1.1787 1.3104
Table 2: Tests of the Existence of Various Inefficiencies in the Public Accounting
Industry
Panel A: 95% confidence intervals for lnB; and lnB;. CI-1 is calculated as
mean(lnB )± 1.96SD(1nB)/ JN where SD(lnB) is the sample standard deviation of lnB,
CI-2 as mean (In B)± 1.96 mean(in B)/ JN, CI-3 as mean(lnB) ± 1.96mean(ln B)~ n- 2 nN
and CI-4 is calculated by first sorting the observations in ascending order and then
choosing the observations given by 0.5(N+ 1) ± 0.98Jl.i as the bounds for the interval.
YEAR Interval Type 'B lnB1
'v lnB1
1995 (N-93) CI-1 (0.144,0.220) (0.125,0.171)
CI-2 (0.145,0.219) (0.118,0.178)
CI-3 (0.154,0.21 0) (0.125,0.171)
CI-4 (0.037,0.080) (0.048,0.067)
1996 (N-92) CI-1 (0.118,0.187) (0.163,0.223)
CI-2 (0.121,0.184) (0.154,0.233)
CI-3 (0.129,0.176) (0.163,0.223)
CI-4 (0.009,0.080) (0.058,0.088)
1997 (N-93) CI-1 (0.128,0.197) (0.151 ,0.206)
CI-2 (0.130,0.196) (0.142,0.215)
CI-3 (0.138,0.188) (0.151 ,0.206)
CI-4 (0.028,0.087) (0.057 ,0.089)
178 Recent Advances in Data Envelopment Analysis ...
Panel B: F-statistics for testing the null hypothesis of no allocative inefficiency against
the presence of such inefficiency. Test I computes the F-statistic as
N IN ~ 1n(BJ} ~ ln(Bf) and this is evaluated relative to the critical value of the F-
distribution with (2N, 2N) degrees of freedom. Test II computes the F-statistic as
N IN ~ (ln(B: )) 2 f.; (ln(B: )) 2 and this is evaluated relative to the critical value of the F-
distribution with (N, N) degrees of freedom.
YEAR Test I Test II
1995 (N-93) 1.830 (0.00 I) 2.246 (0.00 I)
1996 (N-92) 2.265 (0.001) 3.142 (0.001)
1997 (N-93) 2.098 (0.00 1) 2.898 (0.00 1)
Table 3
Impact of Contextual Variables Affecting Productivity in Public Accounting
Firms -- Ordinary Least Squares Estimation
The model estimated is: In e: = Po + PI TAX% + Pz MAS% + P3 HERFINDEX + p4
/nOFFICES + Ps BIG6 + E, where In e: is the logarithm of the aggregate technical
and allocative efficiency estimator for firm j, TAX% is revenues from tax services
expressed as a percentage of total revenues, MAS% is revenues from management
advisory services expressed as a percentage of total revenues, HERFINDEX is
[(A&A%/100)2 + (TAX%/100)2 + (MAS%/100)2], lnOFFICES is the natural logarithm
of the number of offices and BIG6=1 if the firm is one of the Big Six firms, and 0
otherwise. The values given below the coefficient estimates are p-values for two-sided
tests. * indicates significance at the 5% level.
Variables Coefficient 1995 1996 1997
Estimates (N=93) (N=92) (N=93)
Intercept [3. 0.3622 0.1872 0.1924
(0.003) (0.095) (0.079)
TAX% [3. 0.0027 0.0056' 0.0058'
(0.175) (0.009) (0.005)
Recent Advances in Data Envelopment Analysis ... 179
MAS% 13> -0.0018 0.0003 -0.0012
(0.144) (0.818) (0.353)
HERFINDEX 13> -0.3473 -0.1960 -0.0720
(0.098) (0.285) (0.692)
LnOFFICES ~ 0.0725' 0.0655' 0.0406'
(0.001) (0.001) (0.022)
BIG6 13. -0.5001' -0.5151' -0.3621'
(0.001) (0.001) (0.001)
F-value 11.837 11.48 9.922
Adj. R2 0.371 0.366 0.327
Table 4
Impact of Contextual Variables Affecting Productivity in Public Accounting
Firms -- Maximum Likelihood Estimation
The model estimated is: lnBf = ~0 + ~ 1 TAX%+ ~2 MAS% + ~3 HERFINDEX + ~4
/nOFFICES + ~5 BIG6 + E, where In e: is the logarithm of the aggregate technical
and allocative inefficiency estimator for firm j, TAX% is revenues from tax services
expressed as a percentage of total revenues, MAS% is revenues from management
advisory services expressed as a percentage of total revenues, HERFINDEX is
[(A&A%/100)2 + (TAX%/100)2 + (MAS%/100h lnOFFICES is the natural logarithm
of the number of offices and BIG6=1 if the firm is one of the Big Six firms, and 0
otherwise. The error term is defined as E = u-v+VM where u is distributed as Gamma
(2,A.) and vas N(O,o}) truncated above at VM. The values given below the coefficient
estimates are p-values for two-sided tests. * indicates significance at the 5% level.
Variables Coefficient 1995 1996 1997
Estimates (N=93) (N=92) (N=93)
Intercept 13. -0.2163' -0.3296' -0.4467
(0.044) (0.001) (0.143)
TAX% 0.0041' 0.0056' 0.0063'
(0.031) (0.001) (0.008)
MAS% -0.0012 0.0006 -0.0013
(0.247) (0.560) (0.381)
180 Recent Advances in Data Envelopment Analysis ...
HERFINDEX (3. -0.1107 -0.1848 -0.0539
(0.425) (0.196) (0.741)
LnOFFICES ll 0.0695' 0.0617' 0.0415'
(0.001) (0.001) (0.054)
BIG6 (3. -0.4686' -0.4535' -0.3631'
(0.001) (0.001) (0.001)
a, 0.1633' 0.1682' 0.1614'
(0.001) (0.001) (0.001)
I. o.oo5o' 0.005!' 0.0050'
(0.001) (0.001) (0.001)
yM 0.4322' 0.5164' 0.6303
(0.001) (0.001) (0.118)
References
Aigner, D. J., Lovell, C. A. K. and Sclunidt, P. (1977): Formulation of Estimation of Stochastic
Frontier Production Function Models, Journal of Econometrics, 6, 21-37.
Banker, R. D. (1993): Maximum Likelihood, Consistency and Data Envelopment Analysis: A
Statistical Foundation, Management Science, October, 1265-1273.
Banker, R. D. (1996): Hypothesis Tests Using Data Envelopment Analysis, Journal of Productivity
Analysis, 7, 139-159.
Banker, R. D., Chang, H. and Cunningham, R. (1999): The Public Accounting Industry Production
Function, Working Paper, The University of Texas at Dallas.
Banker, R. D., Chang, H. and Natarajan, R. (1999): Efficiency of Public Accounting Firms, Working
Paper, The University of Texas at Dallas.
Banker, R. D., Chames, A. and Cooper, W. W. (1984): Models for the Estimation of Technical and
Scale Inefficiencies in Data Envelopment Analysis, Management Science, 30, 1078-1092.
Banker, R. D., Janakiraman, S. and Natarajan, R. (1999): Estimation of Monotone and Concave or
Convex Functions, Working Paper, The University of Texas at Dallas.
Recent Advances in Data Envelopment Analysis ... 181
Banker, R. D. and Maindiratta, A. (1988): Nonparametric Analysis of Technical and Allocative
Efficiencies in Production Econometrica, November, 1315-1332.
Banker, R. D. and Natarajan, R. (I 999): Evaluating Contextual Variables Affecting Productivity Using
Data Envelopment Analysis, Working Paper, The University of Texas at Dallas.
Charnes, A., Cooper, W. W., and Rhodes, E. (1978): Measuring the Efficiency of Decision Making
Units, European Journal of Operational Research, 429-444.
Craswell A., Francis J. R., and Taylor, L. (1995): Auditor Brand Name Reputations and Industry
Specializations, Journal of Accounting and Economics, 20, 297-322.
Farrell, M. J. (1957): The Measurement of Productive Efficiency, Journal of the Royal Statistical
Society (A,general) 120, pt. 3, 253-290.
Greene, W. H. (1980): Maximum Likelihood Estimation of Econometric Frontier Production
Functions, Journal of Econometrics, 13,27-56.
Grosskopf, S. (1996): Statistical Inference and Nonparametric Efficiency: A Selective Survey, Journal
of Productivity Analysis, 7, 161-176.
Gstach, D. (1998): Another Approach to Data Envelopment Analysis in Noisy Environments: DEA+,
Journal of Productivity Analysis, 9,161-176.
Jondrow, J., Lovell, C. A. K., Materov, I. S. and Schmidt, P. (1982): On The Estimation of Technical
Inefficiency in the Stochastic Frontier Production Function Model, Journal of Econometrics, 19,
233-238.
Kumbhakar, S. ( 1987): The Specification of Technical and Allocative Inefficiency in Stochastic
Production and Profit Frontiers, Journal of Econometrics, 34, 335-348.
Kumbhakar, S. ( 1996): Efficiency Measurement with Multiple Outputs and Multiple Inputs, Journal of
Productivity Analysis, 7, 225-255.
Lehmann, E. L. (1975): Nonparametrics - Statistical Methods Based on Ranks, Holden-Day and
Mcgraw-Hill.
Meeusen, W. and van den Broeck, J. (1977): Efficiency Estimation from Cobb-Douglas Production
Functions with Composed Error, International Economic Review, June, 435-444.
Olson, J. A., Schmidt, P. and Waldman, D. A. (1980): A Monte Carlo Study of Estimators of
Stochastic Frontier Production Functions, Journal of Econometrics, 13,67-82.
Schmidt, P. (1976): On the Statistical Estimation of Parametric Frontier Production Functions, Review
of Economics and Statistics, May, 238-239.
182 Recent Advances in Data Envelopment Analysis ...
Schmidt, P. and Lovell, C. A. K. (1979): Estimating Technical and Allocative Inefficiency Relative to
Stochastic Production and Cost Frontiers, Journal of Econometrics, 9, 343-366.
Seiford, L. M. (1996): Data Envelopment Analysis: The Evolution of the State of the Art, Journal of
Productivity Analysis, 7, 99-13 7.
Shephard, R. W. (1970): Theory of Cost and Production Functions, Princeton, N. J., Princeton
University Press.
Texas Society of Certified Public Accountants (1997): Management of an Accounting Practice
Survey, Dallas.
Tobin, J. (1958): Estimation of Relationships for Limited Dependent Variables, Econometrica,
January, 24-36.
Combining DEA and "Transformation-Stages":
Management Strategies for the Disability Service Units of
the St. Georg Association
Georg Westermann and Gerhard Johnson'
Abstract
In this paper we analyze the efficiency of social service units in a way very similar to the
design of hospital efficiency studies. The houses of the St. Georg Association care for
mentally disabled persons. The explicitly formulated goal of the association is to help the
patients to reach a higher quality of daily-life. Our approach shows the possibility to include
qualitative measures into health sector productivity analysis. We design our analysis in such
a way as to provide management information for controlling the service units. This is
accomplished with the help of portfolio techniques and norm strategies.
1 Gerhard Johnson is Professor for Human Resource Management at the Hochschule Harz, University of Applied Studies and Research.
Georg Westermann is Professor for Public Sector Management at the Hochschule Harz, University of Applied Studies and Research.
184 Combining DEA and "Transformation-Stages" ...
Structure
Introduction
2 Questions to be answered
3 St. Georg Association
4 Design of the Investigation
5 Results of the Investigation
6 Conclusion
References
Combining DEA and "Transformation-Stages" ... 185
1 Introduction
The health sector in almost all industrialized countries in the world is consuming a
steadily growing share of those nations' GDP.' It is therefore not surprising that more
and more scientific efforts are being made to develop suitable management tools.
Governmental health authorities responsible for granting budgets to the different
institutions within the sector are often especially interested in discovering inefficient
units.
Developing adequate instruments for measuring the efficiency of different institutions
within the health sector is still an ongoing task. GrosskopfNaldmanis (1987)
convincingly state that "... empirical analysis of productive performance ... is
complicated by the nature of the 'productive' process behind health care delivery.
Clearly, the conceptual output- improved health status - is difficult to measure as an
output."
During the last 15 years there has been a fast growing literature suggesting Data
Envelopment Analysis DEA as an appropriate methodology to measure the
performance of health providing institutions. 3 There seem to be two main advantages
of DEA that convince more and more researchers to apply this linear programming
method:
(I) DEA is able to process with multiple inputs and outputs that are expressed in technical
terms rather than in (often unavailable) cost terms.
(2) DEA allows for different hospitals to employ different production techniques.
The following table (1) provides a roughly structured picture of the DEA efficiency
studies accomplished in the health sector'. It becomes evident that hospitals and
nursing homes are of special interest. This paper is to be seen within this strand of
2 For an empirical illustration see Ferlie E. et al. (1996).
3 A discussion and an overview can be found in Banker/Das/Datar (1989), Brooks (1995) or Breyer/Zweifel (1996).
4 This is, of course, not a complete overview.
186 Combining DEA and "Transformation-Stages" ...
literature when it examines the performance of social service units m a way very
similar to the hospital productivity studies.5
Table (1) DEA Efficiency Studies in the Health Sector
\ "I h· II ' ' '·" I ) \II lnpuh ()utpuh I llh.l l' lll..~ ( Ul l lqll
Banker (1984) Nursing hours Patients under 14 years Labor/Capital
Hospitals General service hours Patients between 14 and 65 efficiency
Ancillary service hours Patients older than 65 Quantity oriented
Beds
Meyer/Wohlmann- Cost per case Patient judgement Cost efficiency stener ( 1985) Investment per patient % cases without Quantity/Quality Hypothetical hospitals Patients under 60 years
complications oriented
No emergency patients Differentiated inputs
Banker/Conrad/Straus Nursing services Patients under 14 years Labor/Capital s (1986)
Other services Patients between 14 and 65 efficiency
Hospitals Capital Patients older than 65 Quantity oriented
Grollkopf/Valdmanis Physicians Acute care Labor/Capital ( 1987)
Non-Physicians Intensive care efficiency
Hospitals Admission Surgeries Quantity oriented
Plant assets Ambulatory&emergency care
Borden ( 1988) Beds Cases in 9 DRG categories "Total" efficiency
Hospitals Nonpayroll expenses Quantity oriented
Staff
Nurses
Nyman/Bricker Nursing hours SNF patients Labor efficiency (1989) Social service worker ICF patients Quantity oriented Nursing homes hours Limited care patients Differentiated outputs
Therapist hours Personal care patients Other worker hours
Residential care patients
Banker/Das/Datar 6 total departmental 8 inpatient services Cost efficiency
5 This seems to be appropriate because according to Weisbrod (1991 ) the main characteristics of health care are the strong effects on the quality of life and the complexity of the production processes. Thts also holds for social service units.
Combining DEA and "Transformation-Stages" ... 187
(1989) cost categories 2 outpatient services Quantity oriented
Hospitals
Thanassoulis ( 1993) Total cost Teaching units Cost Efficiency
Hypothetical hospitals Regular patients Quantity oriented
Severe patients Differentiated outputs
Donni (1994) Nurses Infants Labor efficiency
Day-Care Centers Cleaning/cooking staff Toddlers Quantity oriented
Qualified staff
ByrnesNaldmanis Registered Nurses Medical-surgical acute Labor/Capital (1994)
Practical Nurses discharges efficiency
Hospitals Management Staff
Medical-surgical intensive Quantity oriented care
Technical Staff Maternity discharges
Aides & orderlies
Beds
Chilingerian (1994) Average length of stay Low-severity cases Cost efficiency
Hospital Physicians Cost ancillary services discharged in a healthier Quality oriented state
High-severity cases discharged in a healthier state
Fare/Grosskopf/ (Estimated) real labor Inpatient discharges Labor/material Lindgren/Roos (1994)
(Estimated) real other Long-term patient bed efficiency
Hospitals input (exc. capital) days Quantity oriented
Ambulatory doctor visits
Norman/Stoker Headquarters Non-psychiatric in-patients Cost efficiency (1991) administration cost discharges&deaths Quantity oriented District Health 7 hospital services cost Psychiatric in-patients bed-Authorities 2 community health days
services cost Attendance day-patients
Day cases
Attendance out-patients
Roos (1997) Total cost Patients Cost Efficiency
Hypothetical eye Change in daily life Strongly quality surgery departments activities oriented
The layout of the paper is as follows. In section 2 we will pose and explain the main
questions within our investigation, while section 3 describes St. Georg Association as
the research object. In section 4 the implementation and design of the public sector
"transformation stages concept" within DEA is of special interest, while section 5
188 Combining DEA and "Transformation-Stages" ...
introduces the results of our calculations. Section 6 concludes with some general
remarks.
2 Questions to be answered
The questions we asked when we started our research project were m principle
twofold. On the one hand we wanted to know whether it was possible to trace
efficiency even in a field as difficult as the care for disabled people. The last column
of table (1) above soundly demonstrates that most of the recent studies neglect the
quality of the output produced in the health sector. 6 Thus, our study aims at explicitly
defining and measuring quantitative and qualitative efficiency concepts for the social
service units. On the other hand we were seeking to explore the possibilities of
obtaining results that could be used to provide a management tool for controlling the
system of different service units within the St. Georg Association. Here we wanted to
be as close as possible to the methodologies known by business managers.
3 St. Georg Association
St. Georg is a non-profit organization that cares for mentally disabled persons of
different ages with a variety handicaps. Most of the disabled people live together in
groups that are structured similarly to families. The groups live together in units that
are called ,houses". And there is always a caregiving person to whom the group
members relate most closely.
In addition to the life in the ,families" the disabled persons go to work - when possible
- or attend treatment by psychologists, psychiatrists or other therapists. The ultimate
goal is to enable the group members to lead a life as independently as possible.
The different ,houses" are grouped together within three regions and are controlled by
regional managers who have to report to the central management. In 1997 the ,total
sales" of these three units amounted to 124 Mio. German Marks. Within the next few
6 Moreover, the few studies considering qualitative efficiency concepts are based on the analysis of hypothetical hospitals.
Combining DEA and "Transformation-Stages" ... 189
years the organization will be restructured, which should lead to a group structure with
the regional organizations as fully owned subsidiaries.
The aim of the central management for starting this investigation was to develop
instruments to control the system of different houses with respect to the efficient
production of the services provided. In other words, management was looking for a
system for benchmarking the houses. The DEA methodology seems to be an adequate
instrument within that context.
4 Design of the Investigation
Measuring the efficiency in the public sector is in itself a difficult task. Measuring the
efficiency of organizations that provide social services is even more difficult because
of the measurement problems -especially with respect to the output-side. Most of the
outputs lack reliable market prices.
In order to grasp the structure of the production processes in this sector we rely on a
theoretical approach going back to Bradford, Malt and Oates (1969). They divide the
process of the production of public social services into several, so-called, ,stages of
transformation". The word transformation is chosen with respect to the transformation
of inputs into outputs. The distinction between the stages is especially based on the
consideration of different outputs.
For our analysis we concentrate on the following three different ,transformation
stages" given in table (2).
Table (2): Transformation Stages
TS 1: Readiness To Produce Services
TS II: Production of Services
TS III: Effect of Services on Customers
Stage I describes the ability of the different units to efficiently transform the respective
factors of production into the desired capacity. In our context, Stage I is the readiness
to care for a certain and fixed number of disabled patients. Stage II describes the
efficiency of transforming inputs into ,sold" output. Below we will define what ,sold"
190 Combining DEA and "Transformation-Stages" ...
output means within this context. Stage III is designed to compare the final outcome of
care on the disabled persons with respect to the required inputs.
Following this approach means to be more precise in the definition of what is the goal
of production in the health sector. All studies in the above table (1) that use "quantity
oriented" efficiency concepts aim at measuring productivity in stages I or II.
GrosskopfNaldmanis (1987) are well aware of this constraint when they label their
approach as a model of " ... the production of the intermediate good- health services."'
Heads of health administrations, politicians and consultants often forget about the
existence of stage III when they aim mainly at input saving activities.
As an analytical instrument to measure the efficiency of the service units, we apply the
Data Envelopment Analysis. We chose DEA especially for two well-known reasons:
(a) DEA offers the possibility to include simultaneously variables that are measured in
different units, such as hours of care, number of beds, staff costs etc. There is no need
to transform or weigh those variables in order to accomplish an evaluation. This is
quite advantageous especially with respect to the output of social service units.
(b) DEA identifies structural differences between the evaluated units as well as efficiency
differentials. The effect is that non-efficient units are compared only against efficient
units or combinations of efficient units that apply the most similar production
technique.
For our computations we apply the following well known CCR8 (or constant returns to
scale) model ofDEA:
mine1-&er s7 -&er si s.t.:
S/
+ s, ;.,,,s7.si~O
0
7 Grosskopfi'Valdmanis (1987), p. 90.
8 It is based on the seminal work of Charnes/Cooper/Rhodes ( 1978).
(I)
Combming DEA and "Transformation-Stages" ... 191
Y1 and X1 are the r- and s-vectors of outputs and inputs respectively of firm 1; Y and X
are the matrices of outputs and inputs of all firms within the sample. The parameter 81
to be minimized accounts for efficiency, the n-vector A.1 provides information about
reference sets, s+ and s· are the excess inputs and output slacks respectively, vector eT
contains only elements 1, and E is the positive so-called Non-Archimedean constane.
We calculate input oriented models (TS I and TS II) as well as an output-oriented
model (TS III). In order to clearly rank the houses we additionally include the possible
input and output slacks into the efficiency measure. Adding the slacks to the necessary
proportional reduction is accomplished in the way suggested by Ali/Lerme ( 1990) or
Chang/Guh ( 1991).
Before the presentation of some of the results we have to define the input/output
models that we apply to describe the production processes on the three transformation
stages. We decided to calculate separate models for each transformation stage because
they shed light on totally different managerial problems.10 The following table (3)
provides the details.
Table 3: 1/0 models with respect to transformation stages
;\I odd Inputs Outputs
TS I (capacity) Sta!T(FTE) Maximum capacity (beds•days)
Other costs (DM)
TS II (production) Staff(FTE) Days charged to customers
Ot.hcr costs (DM)
TS Ill (effect) Staff(FTE) Hours that ,parent persons" care for group
Other costs (DM) members
Days charged to customers
9 See Chames/Cooper (1984).
10 Additionally, the output side seems to provide only constrained possibilities for the management of the DMUs to substitute between outcomes.
192 Combining DEA and "Transformation-Stages" ...
We are aware of the fact that especially the variable on the output side of our model
TS III is a very rough proxy for the effect of care on the disabled persons.'' The
validity of that variable is based on the assumption that the more hours of care by a
parent person a disabled person experiences the stronger is the positive effect on the
personal development.
The management of the St. Georg Association is about to introduce a system of
assigning an outside tutelary to each of the disabled patients. Management plans to
direct a questionnaire to those neutral guardians in order to collect information about
the development of the patients. After receiving those questionnaires, the TS III model
can be re-calculated with more precise data.
5 Results of the Investigation
In this section we are going to present the results of our investigation. We proceed by
providing the efficiency scores of the DEA runs for TS I and TS II first. Then those
results are combined into a strategic management portfolio. This portfolio serves as a
basis to develop management norm strategies. The findings of TS III are then added
and interpreted with respect to the outcomes of TS I and TS II.
Table (4) shows how efficiently the different houses produce their respective
maximum capacity. This represents the efficiency concept connected with stage I. In
column 1 we plotted the position of a certain house within this ranking. Column 2
contains the identification number of the houses, and column 3 the input oriented DEA
efficiency score. Columns 4 and 5 are concerned with the number of staff members
and the amount of other costs that would have to be saved to become efficient. In the
last column (Peer House) we display the houses that should serve as yardsticks for the
inefficient ones. Those were identified by using the weights of the efficient units
within the "A vectors of the inefficient houses.
We do not want to extend the interpretations of those numbers any further, especially
because this first run is very conventional and the possibilities to save inputs are
evident.
11 For more details and especially the variable "care-output-unit" see Johnson et al. (1999).
Combming DEA and "Transformation-Stages" ... 193
With respect to the heterogeneous production techniques, the two efficient houses
represent production that is both more capital intensive (DMU #11) and more staff (or
labor) intensive (DMU #6). This differentiation was confirmed by the intuitive
judgements of the members ofthe central management.
Table 4: Results TS I (capacity)
I'OSI I I< J:\ 1101 Sl·. 1· 1·1·1< W\<' Sl \H 0111.1( I'•. I( s \\I'<; ('OS IS ll<ll ......
S\\ I\(,
l 6 1,00000 0,0 0,0 6
I II 1,00000 0,0 0,0 II
3 10 0,98288 0,2 4703,0 II
4 31 0,94887 0,6 15 177,7 II
5 28 0,94455 1,1 15091,3 6
6 17 0,93614 0,9 19912,1 II
7 4 0,93340 2,7 35878,2 6
8 27 0,80799 4,0 61580,7 6
9 21 0,78799 8,2 154260,6 II
10 5 0,75016 4,7 117123,5 II
11 20 0,74672 4,6 118707,8 II
12 22 0,74234 10,8 289018,0 II
13 I 0,73662 11,9 262986,0 II
14 9 0,73023 6,8 146458,7 11
15 25 0,73019 4,9 62984,7 6
16 23 0,72676 10,0 203501,8 II
17 12 0,71258 9,9 167522,2 11
18 2 0,70986 9,8 249881,3 11
19 24 0,69967 7,2 96810,2 6
20 19 0,67702 11,6 200210,1 II
2 1 26 0,67297 2,9 184511,7 II
22 3 0,67141 4,4 98097,6 II
23 14 0,66830 5,3 204715,0 II
24 8 0,66827 6,2 100646,2 6
194 Combining DEA and "Transformation-Stages" ...
25 16 0,66425 7,1 152950,6 II
26 29 0,66077 8,9 128310,0 6
27 18 0,65257 14,2 277628,9 II
28 13 0,65161 8,7 135980,8 6
29 30 0,62609 9,0 146981,8 6
30 15 0,59932 14,4 393121 ,6 II
31 7 0,59261 12,0 246703,6 II
In the next table (5) we present the results of the TS II run. Here we want to analyze
how efficiently the houses adapted the consumption of inputs to the amount of care
that could really be sold to customers. For the managers of St. Georg Association it
contains information about re-scaling possibilities for the inefficient DMUs. The
structure of this table is exactly the same as in table (4) above.
The most important result of this table is that there seem to be differences in the ability
of the managers of the houses to adapt to the demand for care. Some houses produce
the readiness to care quite efficiently (house #6) but fail in adjusting to the demanded
, bed-days" by customers. Others seem to offer their capacity with too high amounts of
inputs but afterwards produce the sold ,care days" efficiently.
It should become clear that identifying those different patterns of strength and
weaknesses leads to developing management strategies for the respective houses. We
have applied the portfolio technique to assess the positions of the different houses with
respect to their performances.
Table 5: Results TS II (production)
I'OSIIIO' 1101 ..... : Fl FIC IF '< \ S I \I+ 0 IIIH( 1'. FR S\\ I'C; ( "OSIS ll<ll SF
S\\I'C ; I 3 1,00000 0,0 0,0 3
I 11 1,00000 0,0 0,0 II
I 28 1,00000 0,0 0,0 28
4 4 0,99178 0,3 4428,2 28
5 17 0,96659 0,5 10417,5 3
6 10 0,90337 1,1 26545,2 3
7 31 0,89758 1,2 30402,9 3
Combining DEA and "Transformation-Stages" ... 195
8 9 0,85700 3,6 77635,0 3
9 27 0,85122 3,1 47716,1 28
10 25 0,82618 3,1 40576,7 28
II 6 0,81664 6,2 74978,5 28
12 21 0,79868 7,8 146482,4 3
13 22 0,79663 8,5 228120,7 II
14 I 0,78070 9,9 218971,9 3
15 20 0,76722 4,2 109099,8 3
16 5 0,76304 4,5 111085,4 3
17 24 0,75913 5,8 77643,5 28
18 23 0,75419 9,0 183072,7 3
19 12 0,74754 8,7 147145,8 28
20 30 0,72423 6,6 108403,5 28
21 19 0,71078 10,4 179282,9 28
22 2 0,70677 9,9 252542,6 3
23 29 0,70338 7,8 112193,3 28
24 16 0,69599 6,4 138491,5 3
25 13 0,69104 7,7 120590,8 28
26 18 0,68704 12,8 250084,1 3
27 8 0,66919 6,2 100367,1 28
28 14 0,66504 5,4 185319,6 II
29 26 0,63890 3,2 132714,7 II
30 15 0,63662 13,1 356525,2 II
31 7 0,58951 12,1 248580,8 3
The next figure (1) is a first attempt to plot the strategic positions of the DMUs.
Drawing lines with respect to the average performance in the two calculations under
consideration, one can assign each house to one of four strategic fields. 12 One is now
able to discuss with the management of the St. Georg Association the development of
norm strategies for these fields. Norm strategies are going to serve as an instrument for
the strategic controlling of the different houses. Moves of DMUs from one field into
another field indicate changes in their respective efficiency performances and deserve
modified strategies.
12 For an application of this approach to bank branches see Westermann (1999).
196 Combining DEA and "Transformation-Stages" ...
Strategies for the fields:
• Houses assigned to field I supply their capacity with above average efficiency. But the
production of the sold output shows weaknesses. Here it seems to be necessary either to
intensify the efforts to gain new customers or to downscale the house.
• Houses in field II should try a clear strategy of contraction in inputs because they are too
input consuming in both respects: readiness to care and production of care.
• Houses situated in field III should be able to expand their capacities without a
proportionate input increase.
• The units in the fourth field are above average within both respects. They should expand
their businesses if that is possible. They also should serve as benchmarks for inefficient
houses.
Thus, the analysis accomplished is still quite orthodox with respect to the efficiency
concept pursued. In a next step we try to integrate the results (see table 6) from the TS
III DEA run. This computation is output oriented and aims at displaying the
possibilities of the houses to dedicate further time to the disabled persons. The best
DMUs efficiently transform the three inputs Staff (FTE), Other Costs (DM) and Days
Charged to Customers into the single output Hours , Parent Persons" Care for Group
Members. 13
Figure (1): Strategy Portfolio
Strategy Portfolio
Ill >. 1,1 IV Ill u _., Ql c 0,9 c G)
:s u Ill ca !f 0,7 • 2! G)
0,5 0,5 0,7 0,9 1,1
production efficiency
13 For more details and especially the variable "care-output-unit" see Johnson et al. (1999).
Combining DEA and "Transformation-Stages" ... 197
Table 6: Results TS III (outcome)
I'OSIIII 1"\ 1101 sr I· HU I I-"'<\ I'< HI" \SI-" 1'1· 1"1{ .. 1'\IH' I " 1101 Sl 1'1· HSO' II
I 6 1,00000 0,0 6
I 7 1,00000 0,0 7
I 9 1,00000 0,0 9
4 29 1,00000 0,0 29
5 4 1,07474 2876,0 29
6 27 1,08416 1887,6 29
7 28 1,13794 2528,6 29
8 25 1,16942 2582,5 6
9 18 1,26890 10957,8 7
10 19 1,28238 9585,0 29
11 12 1,32613 10566,7 29
12 5 1,35236 7829,4 9
13 17 1,43528 7050,2 9
14 10 1,45590 6359,5 9
15 30 1,48588 9382,5 29
16 21 1,49442 17828,5 9
17 8 1,51976 7143,3 29
18 1 1,54622 24066,4 9
19 16 1,54752 10123,9 7
20 2 1,55580 18348,5 9
21 15 1,66954 21109,4 9
22 23 1,67482 20568,7 9
23 13 1,70357 11569,7 29
24 31 1,72456 9212,7 9
25 20 1,76058 12763,7 9
26 22 1,93421 39759,3 9
27 24 1,96320 12027,8 29
28 3 2,12601 11784,5 9
29 11 2,13517 11447,8 9
30 26 3,02627 17708,1 9
31 14 3,37857 28205,3 9
198 Combining DEA and "Transformation-Stages" ...
What can social service managers learn from this third strand of computations? It
worth to notice is for example that house #6 that was efficient in producing capacity to
care is also efficient in caring for the patients. House #7 in contrast performs very
badly in capacity production but is efficient in dedicating time to the group members.
For reasons of a comprehensive insight in the efficiency structures of St. Georg
Association it seems to be appropriate to construct a strategic portfolio similar to
figure ( 1 ). In order to avoid a three-dimensional plot in table (7) we apply a three-digit
approach to indicate high/low efficiency in the three DEA runs TS I, TS II and TS III.
Column 5 in table (7) additionally provides the allocation of houses across the
respective fields.
Table 7: Strategic Fields in the Three Dimensional Case
·1 Sl rttkil•nc~ 'I Sll Eltkil·nc~ I Sill l·: lficil·m·~ l· idcl 1-'idcll'upulatinn
High Low Low I -High High Low 2 2
High High High 3 7
High Low High 4 -Low Low Low 5 2
Low High Low 6 2
Low High High 7 2
Low Low High 8 II
It is obvious that a remarkable percentage (27%) of houses is situated in field 3 and
thus can be labeled as being above average in their efficiency with respect to all three
efficiency concepts. Another global feature seems to be the trade-off between TSIITSII
and TSIII. Exactly 50% (field 2 and 8) of the DMUs trade capacity and production
efficiency against outcome efficiency.
With respect to the strategies that should be pursued for the houses in the fields 1-8 we
will concentrate only on one illustrating example. The distinguishing feature of the
most populated field 8 is the low efficiency in the transformation of staff and other
cost into capacity and sold beds. In those houses the ,parent persons" can avoid
management and auxiliary tasks because other staff are responsible for that.
Combming DEA and "Transformation-Stages" ... 199
Management should explore the possibilities of especially reducing auxiliary staff
without decreasing the extraordinary high level of care. 14
6 Conclusion
In this paper we analyze the efficiency of social service units in a way very similar to
the design of hospital efficiency studies. The houses of the St. Georg Association care
for mentally disabled person. The explicitly formulated goal of the association is to
help the patients to reach a higher quality of daily-life. Thus, it is not reasonable to
tackle the efficiency problem with bare quantitative concepts.
Our approach shows the possibility of including qualitative measures into health sector
productivity analysis. Moreover, we design our analysis such that it will provide
management with information for controlling the service units. This is accomplished
with the help of portfolio techniques and norm strategies.
Our variable for the effects of care on the disabled persons is a very rough proxy. The
planned questionnaire containing a neutral judgement on the condition and
development of the disabled persons is going to provide a more adequate database. A
second weakness of our present analysis is the absence of data with respect to the
structure of handicaps and the differences between the houses. This problem is going
to be solved by assigning the group members to different handicap or case groups. The
number of persons within those groups can then be used as output variables in further
DEA calculations.
We are aware of the fact that the present analysis is not at all perfect but we hope to be
on a way that it nevertheless can help the management of the St. Georg Association to
care more efficiently for their customers. But caring more efficiently must not neglect
all those non-economic variables inherent in human relations.
14 In addition, it would be worth analyzing the situation in those houses with a more decomposed staff variable.
200 Combining DEA and "Transformation-Stages" ...
References
Ali A.l. and C.S. Lerme (1990): Data Envelopment Analysis Models: A Framework, School of
Management, University of Massachusetts at Amherst, Working Paper.
Banker, Rajiv D. (1984): Estimating Most Productive Scale Size Using Data Envelopment Analysis,
European Journal of Operational Research, Vol. 17, pp. 35-44.
Banker, Rajiv D., Conrad, Robert F. and Strauss, Robert P. (1986): A Comparative Application of
Data Envelopment Analysis and Translog Methods: An Illustrative Study of Hospital
Production, in: Management Science, Vol. 32, No. I, pp. 30-44.
Banker R.D., D. Somnath and S.M. Datar (1989): Analysis of Cost Variances for Management Control
in Hospitals, Research in Governmental and Nonprofit Accounting, vol.5, 1989,pp.269-291.
Borden, J. P. (1988): An Assessment of the Impact of Diagnosis-Related Group (DRG)-Based
Reimbursement on the Technical Efficiency of New Jersey Hospitals Using Data Envelopment
Analysis, Journal of Accounting and Public Policy, Vol.7, No.I, pp.77-96.
Bradford, D.F., R.A. Malt and W.E. Oates (1969): The Rising Cost of Social Public Services: Some
Evidence and Reflections, National Tax Journal, 22 (1969), pp. 185-213.
Breyer, F. and Zweifel P. (1996): Health Economics, Oxford University Press, Oxford.
Brooks, R. ( 1995): Health Status Measurement: A Perspective on Change, MacMillan Press.
Byrnes P. and V. Valdmanis (1994): Analyzing Technical and Allocative Efficiency of Hospitals, in:
Charnes A., W.W. Cooper, A.Y. Lewin and L.M. Seiford (1994): Data Envelopment Analysis:
Theory, Methodology and Application, Kluwer Academic Publishers, Boston, pp. 129-44.
Chang, K.-P. and Guh, Y.-Y. (1991): Linear Production Functions and the Data Envelopment
Analysis, European Journal of Operational Research, Vol. 52, pp. 215-223.
Charnes, A. and Cooper, W.W. (1984): The non-Archimedean CCR Ratio for Efficiency Analysis: A
Rejoinder to Boyd and Flire, European Journal of Operational Research, Vol. 15, pp.333-334.
Charnes, A., Cooper, W.W. and Rhodes, E. (1978): Measuring the Efficiency of Decision-Making
Units, European Journal of Operational Research, Vol. 2, pp. 429-444.
Chilingerian, Jon A. (1994): Exploring Why Some Physicians' Hospital Practices Are More Efficient:
Taking DEA Inside the Hospital, in: Charnes A., W.W. Cooper, A.Y. Lewin and L.M. Seiford
(1994): Data Envelopment Analysis: Theory, Methodology and Application, Kluwer Academic
Publishers, Boston, pp. 1167-193.
Donni, Olivier (1994): Efficiency of Day-Care Centers in Belgium, Paper presented at the European
Economic Association, Ninth Annual Congress, Mastricht.
Combining DEA and "Transformation-Stages" ... 201
Fare R., S. GroBkopf, B. Lindgren and P. Roos (1994): Productivity Developments in Swedish
Hospitals: A Malmquist Output Index Approach, in: Charnes A., W.W. Cooper, A.Y. Lewin
and L.M. Seiford (1994): Data Envelopment Analysis: Theory, Methodology and Application,
Kluwer Academic Publishers, Boston, pp.253-72.
Ferlie E., A. Pettigrew, L. Ashburner and Louise Fitzgerald (1996): The New Public Management in
Action, Oxford University Press, Oxford, p. 2.
Grosskopf, S. and Valdmanis, V. (1987): Measuring Hospital Performance: A Non-Parametric
Approach, Journal of Health Economics, Vol. 6, pp. 89-107.
Johnson G., H. Johnson and R. Klaus (1999), Personalcontrolling im sozialen Bereich -
Qualitatssicherung bei der Betreuung behinderter Menschen, in: PERSONAL, 51. Jg. 1999,
Volume 4.
Meyer, Manfred and Wohlmannstetter, Viktor (1985): Effizienzmessung in Krankenh!iusern, in: ZfB,
Vol. 55, No. 3, pp. 262-280.
Norman, Michael and Stoker, Barry (1991): Data Envelopment Analysis - The Assessment of
Performance, Wiley, Chichester.
Nyman, J. and Bricker, D. (1989): Profit Incentives and Technical Efficiency in the Production of
Nursing Home Care, Review of Economics and Statistics, pp. 586-594.
Roos, Pontus ( 1997): Measurement of Output and Productivity of Hospital Services: A Discussion of
the Malmquist Index Approach, The Swedish Institute for Health Economics Working Paper,
1997:7.
Thanassoulis, E. (1993): A Comparison of Regression Analysis and Data Envelopment Analysis as
Alternative Methods for Performance Assessments, Journal of the Operational Research Soc.,
Vol. 44, No. II, pp. 1129-1144.
Westermann G. (1999): De Ia comptabilite d'efficience a le developpement de Ia strategie -
Application de Ia methode DEA aux succursales des banques, forthcoming in: Badillo P. -Y.
and J.C. Paradi (eds.).
DEA in the ecological context- An overview•
Katrin Allen'
Abstract
Data envelopment analysis has a high application potential for environmental manage
ment/ecological controlling, especially for eco-benchmarking projects. The aim of this paper
is to provide some basics for DEA applications within this scope. DEA assumes that inputs
and outputs are 'goods ', but from an ecological perspective also 'bads ' have to be consid
ered. The potential and state-of-the-art of DEA in the ecological context are outlined. Model
ling alternatives, axiomatic modifications, and application driven extensions are described,
followed by a concept for DEA applications in environmental performance measurement as
well as a framework for future research.
1 Financial support by the Deutsche Forschungsgemeinschaft (DFG) is gratefully acknowledged.
2 Technical University of Aachen, Templergraben 64, D-52064 Aachen, Germany
204 DEA in the ecological context- An overview
Structure
Introduction
2 Environmental Management and Ecological Controlling
3 DEA in the ecological context
4 Ecological extension ofDEA
5 A concept for applications ofDEA in EPM
6 Summary and outlook
References
DEA in the ecological context- An overview 205
1 Introduction
Data envelopment analysis (DEA) is a collection of models measuring the relative effi
ciency of decision making units (DMUs) with linear programming techniques.' DEA
shows a high potential for applications within the field of environmental manage
ment/ecological controlling (EM/EC). Private and non-profit organisations look for
practical, flexible, and objective tools supporting EM/EC, due to the ecological and
economic benefits to be gained by environmental protection activities. The aim of this
paper is to provide some basics for applications ofDEA in EM/EC. The main problem
is that DEA assumes inputs and outputs are 'goods'. This assumption no longer holds,
when - from an ecological perspective - consideration of outputs to be minimised (e.g.
S02-emissions) and inputs to be maximised (e.g. recycling of wastepaper) is required.
The paper is organised as follows: Section 2 starts with an overview of EM/EC, its
problems and instruments. Section 3 points out the potential and limitations ofDEA in
the ecological context and provides a literature review. Selected aspects of an ecologi
cal extension of DEA are examined in section 4, grouped into modelling alternatives,
axiomatic modifications, and application driven extensions. A concept for applications
in environmental performance measurement is suggested in section 5 considering the
potential and limitations outlined so far. The final section provides a summary and a
framework for future research.
2 Environmental Management and Ecological Controlling
"Organizations of all kinds are increasingly concerned to achieve and demonstrate
sound environmental performance by controlling the impact of their activities, prod
ucts or services on the environment [ ... ]".' Environmental protection has become a
'trendy' topic not just because companies want to develop and achieve the idea of
3 For an introduction into DEA see e.g. Seiford {1996) and Chames et al. {1994), chapters I to 3.
4 Introduction to the ISO 14000 series of standards and guidelines in the field of environment developed from the International Organization for Standardization. For details see http://www.iso.ch.
206 DEA in the ecological context- An overview
'sustainable development''. It has become an important competitive factor as there are
increasingly more economic benefits to be gained:
• General growth from the stakeholders about environmental matters
and a sustainable development.
• Implementation of increasingly stringent environmental legislation
and economic tools, e.g. taxes, tradable permits.
• Identifying and taking advantage of ecological factors of perform
ance, e.g. improvement of the corporate image, cost savings.
• Identifying and taking countermeasures against potential ecological
risks, e.g. fines, bad publicity.
Environmental management (EM) is the part of the overall management that has to
pursue environmental objectives in the planning, implementation, and control of inter
actions in all parts of a company. For this the EM has to implement a corresponding
organisational structure and incentives for the ecological motivation of the employees.•
The increasing importance of EM is underlined by the efforts of the European Union.
The objective of EMAS' is to promote continuous improvements in the environmental
performance of industrial activities by committing sites to evaluate and improve their
environmental performance and provide relevant information to the public. Addition
ally, the norm ISO 14001' "[ ... ] specifies the requirements for an environmental man
agement system, to enable an organization to formulate a policy and objectives taking
into account legislative requirements and information about significant environmental
impacts."
5 "Sustainable development is development that meets the needs of the present without compromising the ability of future generations to meet their own needs." World Commission on Environment and Development (1987), p. 43.
6 For an introduction into EM see e.g. Dyckhoff(l998b).
7 Council regulation EEC No 1836/93 of 29 June 1993 allowing voluntary participation by companies in the industrial sector in a Community ~;co-management and audit .scheme. The full text is provided on http://www.emas.lu.
8 ISO 14001 'Environmental management systems- Specification with guidance for use', see fn. 4.
DEA in the ecological context- An overview 207
A specific ecological controlling can support the extensive information and co-ordi
nation needs of environmental management. Ecological controlling (EC) can be de
fined as a subsystem of the overall management and the overall controlling system. Its
task is to co-ordinate the development and coupling of the ecologically-oriented plan
ning, control, and information search as subsystems of the EM. By doing this, it sup
ports the ability for co-ordination, reactions, and adaptability of the strategic and op
erational EM and therefore guarantees an improvement in the efficiency and effective
ness of the ecologically relevant decisions and activities of the overall management.•
To support the EM, the EC has to develop various powerful instruments10• Strategic
instruments are e.g. ecological early warning systems, strength-weakness-profiles, risk
management, the estimation of potential impacts and acceptability of new technologies
or products, ecological value chain analysis, and ecological portfolio analysis. Typical
operational instruments are ecological checklists and life cycle assessment. Addition
ally several concepts for environmental accounting and ecological budgeting can be
mentioned. Eco-benchmarking is usually based on environmental performance indica
tors, which also represent an original EC instrument. Both can show a strategic or op
erational orientation. With respect to the focus of this paper three instruments are de
tailed below:
Life cycle assessment (LCA)" is a process to analyse and assess the environmental
impacts of a product, process or activity over its whole life cycle. A LCA study usually
contains four steps:
I. Goal definition and scoping
2. Inventory analysis
All material and energy flows that cross the system boundary are
identified and quantified. The result is the so-called material and en
ergy balance (inventory table).
9 Ecological modification of Horvath (1994), p. 144, suggested by Rudiger (1998), p. 283.
1° For extensive overviews see e.g. Gunther (1994) and Rudiger (1998). The classification strategic/operational is not always clear.
11 For an introduction into LCA see e.g. Miettinen!HiimiiHiinen ( 1997).
208 DEA in the ecological context- An overview
3. Impact assessment
The material and energy balance is translated into potential environ
mental impacts. Usually weights are used to aggregate some repre
sentative inputs and outputs into one/several indicators.
4. Improvement assessment
Options for reducing the environmental impacts of the regarded pro
duct, process or activity are identified and evaluated.
Environmental performance indicators (EPis) 12 can be defined as numbers, which
capture a quantifiable aspect of ecological relevance in a concentrated manner. EPis
appear as absolute or relative numbers based on monetary, non-monetary or partially
monetary size (e.g. cost for waste disposal, amount of waste in tons, cost for waste dis
posal per ton of waste). EPis are integrated in several instruments of EC, but they also
represent an original instrument, supporting information, planning, control, and
checking purposes on a strategic as well as on an operational level. They can be used
for either target-performance comparisons, comparisons over time, or benchmarking
projects. ISO 14031 differs between environmental condition indicators and environ
mental performance indicators, the latter including management performance indica
tors and operational performance indicators. 13 Environmental operational performance
indicators, which are based on the material and energy balance, are most commonly
used. Isolated EPis can not reflect complex systems and interrelations. Therefore it is
recommended to choose some meaningful, logically-related EPis to develop an envi
ronmental performance indicator system (EPIS). 1' EPis are meaningless without com
parative values, which can be determined in benchmarking projects.
Eco-benchmarking can be defined as the continuous process of ecological improve
ments of products, processes, methods, and functions by comparison with the 'best of
the class' and realisation of the 'best ecological practice' in their own company. 1' Eco
benchmarking projects can focus on comparisons over time, comparisons within a
company, and comparisons among companies operating in similar or different
12 For details see Seidel eta!. (1998).
13 ISO 14031 'Guidelines on environmental performance evaluation', see fn. 4.
14 See e.g. the EPIS in Loew/Kottmann (1996), p. 12.
1' See Goldmann!Schellens (1995).
DEA in the ecological context- An overview 209
branches of industry. In analogy to the ISO 14031 indicators, objects of an ecological
benchmarking project can be:
1. Environmental impacts
Benchmarking of adverse of beneficial changes to the environment
resulting from the organisations' activities, products, or services.
2. Environmental practice
Benchmarking of normative/strategic/operative elements of the EM
system.
3. Environmental performance
Benchmarking of measurable results of the EM system, related to the
organisation's control of its environmental aspects.
There are numerous problems with which EM/EC are generally associated, especially
in regard to the instruments LCA, EPis, and eco-benchmarking: The database needed
for the instruments is extensive, novel, and complex. Especially as there are multiple
inputs and outputs which are measured on different scales. For a full description of a
product, sometimes thousands of inputs and outputs are required. The functional rela
tionships between the inputs and outputs are often unknown. Usually there are no mar
ket prices or objective preferences for the consumption of natural resources and the
emission of harmful chemicals, to allow for the aggregation of the inputs and outputs.
A lot of approaches in the literature build on weight coefficients, equivalence numbers,
or monetary conversion. Although, the assessment of such values turns out to be very
difficult due to subjective and political influences. Further problems stem from the da
tabase. The standardisation of the inventory analysis step in LCA and of EPis is an
important presupposition for comparisons, especially for eco-benchmarking. Though
so far, standardisation is at a low level and countermeasures like ISO 14000 are still in
their infancy. However, even if companies have reliable and well-developed environ
mental data, it might not be available to the public.
For these reasons there is an increasing need for practical, flexible, and objective tools
to support the EM/EC. In particular, one looks for methods which allow for considera
tion of multiple inputs and outputs without requiring difficult value judgements for
ecologically-oriented comparisons of plants, processes, products, management systems
or functions.
210 DEA in the ecological context- An overview
3 DEA in the ecological context
3.1 Problem framework
DEA and activity analysis approaches in general assume that inputs and outputs are
'goods', i.e. objects with a positive value, where a positive value does not necessarily
mean a monetary advantage. In the environmental context, we also have to consider
'bads', i.e. objects with a certain negative value. Furthermore, 'neutrals' as objects
without any kind of value with respect to the decision problem at hand, may also arise.
Combining these three classes with the criteria 'position as input or output in the pro
duction process' results in six categories of different ecological desirability, which are
shown injigure 3.1.
neutral bad
Input reduct
output product
D D desirable indifferent undesirable
Figure 3.1 Categorisation of ecologically relevant objects according to the
'standard case'
The 'standard case' specification represents a weak preference order for a certain cate
gorisation of ecologically relevant objects. It will be illustrated in the following by
means of the example of a waste-burning power plant as shown in figure 3.2. 16
16 Regarding the 'standard case' and the waste burning power plant example see Dyckhoff (1994) p. 65 and Dyckhoff (1998a) p. 97, respectively.
DEA in the ecological context- An overview 211
/ / 470 kWh
970kWh
1000 kg .. 6000m'
waste 7001 8001 .. heating
60ka power
330kg 6000m ' ., plant 890 kWh
/
Figure 3.2 Categorisation for a waste heating power plant
Factors and products as goods on the input and output side respectively, represent the
classical DEA case. For a waste-burning power plant, the use of water should be kept
to a minimum. The products, electric power and heat, which are the aims of the trans
formation process, should be maximised. A bad on the output side is undesirable, and
therefore its quantity should be minimised, in this example pollutants such as waste
gas, waste water, scrap, and cinders. The expression 'reduct' is used for the symmetric
case of bads on the input side. Waste to be burned at the power plant represents such
an undesirable object whose input should be maximised. If one is completely indiffer
ent towards air on the inputs side, and residual heat on the output side, they represent a
by-factor and by-product respectively.
The 'standard case' specification of ecologically relevant objects is chosen here as the
problem framework for an ecological extension of DEA. Though it should be noted
that the categorisation can be subjective and arbitrary. Furthermore, the categorisation
can change due to time, place, information, quantity or quality. Imagine for example,
that the output residual heat could be sold to a nearby swimming centre or that there is
a penalty for emitting more than a given limit.
3.2 Potential and limitations
Miettinen/Hamalainen (1997) "[ ... ]see DEA as a promising tool for LCA." Within the
scope of air pollution monitoring, Cooper et a!. (1996) point out that comparitive
evaluations of performance like DEA "[ ... ]could lead to improvements from learning
212 DEA in the ecological context- An overview
how the very best performers deal with satisfying constraints." The idea to apply DEA
within the scope ofEM/EC builds on two trains of thought:
Ignoring the ecological perspective for the moment, DEA has in general proved its
'controlling' potential. Following Wemer/Brokemper (1996) we are able to differen
tiate applications for information, planning, and control systems. Within the area of
intelligence activity, DEA is predominantly used for weak point analyses and for the
preparation and analysis steps in benchmarking projects." DEA can also support the
planning of future activities, e.g. the assessment of minimal cost input combinations,
the development of investment strategies or the choice of location." The efficiency
score represents a comprised feedback, which can be used for control purposes, espe
cially for control over time in DEA window analysis." Furthermore, DEA may support
the generation of control models.'"
From a different perspective, there are some arguments underlining the 'technical'
potential of DEA in the ecological context: Activity analysis based approaches like
DEA show special qualities for EM/EC. Especially as their structure is comparable to
that ofthe inventory table in LCA, as an important database ofEM/EC." DEA explic
itly allows for the consideration of multiple inputs and outputs measured on different
scales. DEA does not require information on functional relationships or a priori
weights for the aggregation of different impacts of production processes on the envi
ronment. The general questions addressed by DEA are important for ecological com
parisons. For example, which inputs and outputs cause the inefficiency of a DMU, and
to what extent? Finally, DEA is flexible regarding modelling needs and a lot of exten
sions to the basic DEA models are relevant for applications to the context at hand.
Apart from this confident judgement, conceptual limitations and pitfalls ofDEA have
to be considered, e.g. its sensitivity to outliers or model choice, and unreasonable re
sults due to the flexibility of the weights. 22 The outlined potential is further endangered
17 Schefczyk (1993) and Cook et al. (1992).
18 Ray/Kim (1995) and Athanassopoulosffhanassoulis (1995).
19 Day et al. (1995).
20 Epstein/Henderson (1989).
21 See for detailed arguments Fl!re et al. (1996), p. 162, and Souren!Rildiger (1998), p. 306-308.
22 See e.g. Stolp ( 1990).
DEA in the ecological context- An overview 213
by the practical problems occurring within the field of EM/EC mentioned in section 2.
Applications in the ecological context will have to cope with a high number of inputs
and outputs in combination with a small number of comparative objects at a low level
of standardisation respectively with incomplete data. This relationship may weaken the
differentiation power of DEA. Though first, solutions for modelling 'bad' objects in
contrast to the classical DEA focus on 'goods' have to be found. The following litera
ture review reveals different approaches as to how to address this problem.
3.3 Literature review
The DEA literature already includes a number of articles dealing with either applica
tions or theoretical investigations within the ecological context. To the best of my
knowledge and using the 'published' data available (in English), table 3.1 gives an
overview. The articles are sorted by 'year of publication' as well as 'connection in
contents'. The columns 'type' and 'essentials' do not summarise the complete articles,
but the facts relevant for the ecological extension of DEA. Some of the listed articles
do not explicitly mention DEA, though there are many more related papers (e.g. fo
cusing on the use of distance function and multi-criteria approaches in the ecological
context), which could not be included in this review."
Only five articles exclusively focus on the theoretical background of undesirable out
puts in DEA. A fact that is all the more surprising when considering that undesirable
outputs do not only appear in the ecological context, but also, for example, within
health care (the complications of medical operations) or within business (tax pay
ments). Furthermore, the articles show very different and sometimes arbitrary ap
proaches regarding the treatment of undesirable outputs. In section 4, selected aspects
mentioned in table 3.1 are explained and discussed in detail.
23 See e.g. Fare et al. (1993) or Cooper et al. (1996).
I Aut
hor(
s)
I Typ
e I Es
sent
ials
I A
ppli
cati
on
I Rem
ark
s I
Fir
e e
t al.
(19
89)
T+
a I
stro
ng/w
eak
disp
osab
ilit
y o
f und
esir
able
out
puts
U
.S. p
aper
mil
ls
o se
vera
l non
·rad
ial h
yper
boli
c ef
fici
ency
mea
sure
s tr
eati
ng i
nput
s, d
esir
able
out
puts
and
un
desi
rabl
e ou
tput
s as
ymm
etri
call
y o
conv
ersi
on in
to L
P ap
prox
imat
ions
B
alle
t al.
(19
94)
T+
a o
4 m
odel
s w
ith
line
ar o
r hy
perb
olic
eff
icie
ncy
mea
sure
on
prod
ucts
or
prod
ucts
and
U
.S. a
gric
ultu
re
poll
utan
ts,
assu
min
g w
eak
disp
osab
ilit
y o
f und
esir
able
out
puts
o
calc
ulat
ed s
hado
w v
alue
s us
ed to
adj
ust a
con
vent
iona
l tot
al f
acto
r pr
oduc
tivi
ty g
row
th
mea
sure
for
the
gene
rati
on o
f und
esir
able
out
put
Kao
/Y an
g (1
991)
T
+a
desi
rabl
e ou
tput
'for
est s
tock
ing
in m
'lha
' cho
sen
as a
mea
sure
for
soi
l con
vers
atio
n fo
rest
dis
tric
ts i
n T
aiw
an
repl
acin
g th
e un
desi
rabl
e ou
tput
'soi
l ero
sion
'
Hay
nes
et a
l. (
1993
) t
max
imis
atio
n o
f rec
ipro
cal o
fCC
R-m
odel
'nor
mal
isat
ion'
wit
h ch
emic
al r
esid
uals
take
n as
(th
e on
ly)
inpu
ts
Hay
nes
et a
l. (
1994
) T
+a
max
imis
atio
n o
f rec
ipro
cal
ofC
CR
-mod
el 'n
orm
alis
atio
n' w
ith
chem
ical
res
idua
ls ta
ken
poll
utio
n pr
even
tion
in
as (
the
only
) in
puts
N
.J. c
hem
ical
pla
nts
Gol
any
et a
l. (
1994
) a
II S
02-
emis
sion
s tr
eate
d as
a c
ateg
oric
al v
aria
ble
on th
ree
leve
ls s
uch
that
a D
MU
can
po
wer
pla
nts
in Is
rael
on
ly b
e co
mpa
red
to D
MU
's in
its
or a
bet
ter c
ateg
ory
o in
corp
orat
ion
of s
tand
ards
Lov
ell e
t al.
(19
95)
a II
Car
bon
and
nitr
ogen
em
issi
ons
are
conv
erte
d by
taki
ng th
eir
reci
proc
al
mac
roec
onom
ic p
erfo
r-m
ance
OE
CD
cou
ntri
es
Fir
eJG
ross
ko
pf (
1995
) t
II st
rong
/wea
k di
spos
abil
ity
of u
ndes
irab
le o
utpu
ts
o in
corp
orat
ion
of j
oint
ness
bet
wee
n de
sira
ble
and
unde
sira
ble
outp
uts
I o
disc
ussi
on o
f Hay
nes
et a
l. (1
993)
I
o no
n-ra
dial
mod
el a
ccou
ntin
g fo
r de
sira
ble
outp
uts
to b
e in
crea
sed
and
at th
e sa
me
tim
e
I fo
r un
desi
rabl
e ou
tput
s an
d in
puts
to b
e de
crea
sed
at d
iffe
rent
rat
es
Tyt
eca
(199
6)
t o
CC
R-m
odel
s 'n
et p
rodu
ctio
n',
'pol
luta
nt=
fact
or',
and
'nor
mal
isat
ion'
as
EP
is
lite
ratu
re r
evie
w
/ o
disc
ussi
on o
fFD
H a
nd 'i
deal
' fro
ntie
r in
com
pari
son
to th
e D
EA
mod
els
on
EP
M
Fir
e e
t al
. (19
96)
T+
a II
wea
k di
spos
abil
ity
for
unde
sira
ble
outp
uts
U.S
. fo
ssil
fue
l-fi
red
com
pare
d to
o
deri
ve a
n E
PI
from
dec
ompo
sing
ove
rall
eff
icie
ncy
into
an
inpu
t pro
duct
ive
effi
cien
cy
elec
tric
uti
liti
es
IF-i
ndex
an
d an
env
iron
men
tal
inde
x T
ytec
a (1
997)
T
+a
CC
R-m
odel
s 'n
et p
rodu
ctio
n',
'pol
luta
nt=
fact
or',
and
'nor
mal
isat
ion'
as
EP
is
U.S
. fo
ssil
fuel
-fir
ed
com
pare
d to
el
ectr
ic u
tili
ties
IF
-ind
ex
Tyt
eca
(199
8)
T+
a 7
sust
aina
ble
deve
lopm
ent i
ndic
ator
s st
ress
ing
econ
omic
, soc
ial,
and/
or e
nvir
onm
enta
l U
.S.
foss
il fu
el-f
ired
co
mpa
red
to
aspe
cts
elec
tric
uti
liti
es
IF-i
ndic
ator
s
Cal
lens
ffyt
eca
(199
8)
t C
CR
-mod
el f
or a
n un
ique
agg
rega
te s
usta
inab
le d
evel
opm
ent
indi
cato
r th
at s
houl
d be
de
vide
d in
to p
artia
l in
dica
tors
str
essi
ng e
cono
mic
, so
cial
, an
d/or
env
iron
men
tal a
spec
ts
Gal
lez/
Tyt
eca
(199
8)
T+a
C
CR
-mod
els
'net
pro
duct
ion'
and
'pol
luta
nt=
fact
or' a
s E
P!s
Van
den
Eec
kaut
et
a!.
(199
7)
T+
a 14
EP!
s de
rive
d by
com
bini
ng I
. E
colo
gica
l vs
. ra
dial
/non
-rad
ial
econ
omic
-eco
logi
cal
impr
ovem
ents
, 2.
DE
A v
s. FD
H,
3. s
tron
g vs
. w
eak
disp
osab
ility
of u
ndes
irab
le o
utpu
ts,
4. C
RS
vs.
VR
S
Cou
rcel
le e
t a!
. (1
998)
T
+a
• st
art
from
uni
que
aggr
egat
e su
stai
nabl
e de
velo
pmen
t in
dica
tor
• FD
H m
odel
'pol
luta
nt=
fact
or' w
ith u
ndes
irab
le o
utpu
t res
idue
rat
e is
com
pare
d to
C
CR
-mod
el w
ithou
t und
esir
able
out
put
Pio
t-L
epet
it e
ta!.
(19
97)
T+
a •
focu
s on
red
ucin
g fe
rtili
sers
/pes
ticid
es f
or c
ausi
ng e
xter
nal e
nvir
onm
enta
l eff
ects
•
use
subv
ecto
r and
non
-rad
ial t
echn
ical
eff
icie
ncy
mea
sure
s du
e to
qua
si f
ixed
fac
tors
Pio
t-L
epet
it!V
erm
ersc
h (1
998)
T
+a
• B
CC
-mod
el a
nd s
lack
adj
uste
d m
easu
re 'p
ollu
tant
=fa
ctor
' with
und
esir
able
out
put
'org
anic
nit
roge
n'
• st
rong
/wea
k di
spos
abil
ity
of u
ndes
irab
le o
utpu
t •
deri
vati
on o
f a s
hado
w p
rice
of t
he u
ndes
irab
le o
utpu
t
Ath
anas
sopo
ulos
et
al. (
1998
) T
+a
• ch
oose
'pol
luta
nt=
fact
or' m
odel
ling
for
the
und
esir
able
out
puts
con
trol
labl
e em
issi
ons
and
num
ber
of a
ccid
ents
as
a m
easu
re f
or s
afet
y •
DE
A u
sed
to g
ener
ate
targ
et s
etti
ng s
cena
rios
e.g
. fo
r po
llut
ion
emis
sion
s
All
en (
1998
b)
t •
EM
IEC
app
lica
tion
pot
enti
al o
fDE
A a
nd li
tera
ture
rev
iew
•
com
pari
son
of a
n ec
olog
ical
ly e
xten
ded
Add
VR
S m
odel
and
mod
elli
ng a
lter
nati
ves
• st
ruct
ure
for
ecol
ogic
ally
rel
evan
t ext
ensi
ons
and
axio
mat
ic m
odif
icat
ions
ofD
EA
•
a co
ncep
t for
EPM
app
licat
ions
Dyc
khof
f/ A
llen
(19
98)
t de
riva
tion
of a
gen
eral
ised
Add
VR
S m
odel
whi
ch is
eco
logi
call
y sp
ecif
ied
acco
rdin
g to
th
e 's
tand
ard
case
'
I Schee
l (19
98)
I t
I effi
cien
cy m
easu
re c
onsi
ders
tha
t any
cha
nge
of o
utpu
ts i
nvol
ves
both
des
irab
le a
nd
unde
sira
ble
outp
uts
(ass
umpt
ions
: st
rong
dis
posa
bilit
y, r
ay r
egul
arit
y)
Tab
le 3
.1
Ove
rvie
w o
fDE
A li
tera
ture
in t
he
JF: J
aggi
/Fre
edm
an (
1992
) t:
theo
ry
ecol
ogic
al c
onte
xt
a:
appl
icat
ion
U.S
. fo
ssil
fuel
-fir
ed
com
pare
d to
el
ectr
ic u
tiliti
es (
verb
al)
JF-i
ndex
U
.S.
foss
il fu
el-f
ired
el
ectr
ic u
tiliti
es
EU
mun
icip
al s
olid
w
aste
col
lect
ion
and
sort
ing
prog
ram
mes
Fren
ch c
erea
l far
ms
Fre
nch
pig
farm
s
UK
ele
ctri
city
gen
erat
ing
plan
ts
Eur
opea
n ec
onom
ies
'net
pro
du
ctio
n'
'pol
luta
nt=
fact
or'
} se
e se
ctio
n 4.
1 'n
orm
alis
atio
n'
216 DEA in the ecological context- An overview
4 Ecological extension of DEA
This section discusses selected aspects of an ecological extension of DEA. Section 4.1
shows possible extensions of basic DEA models whereas section 4.2 also considers
axiomatic modifications. Section 4.3 outlines the relevance of some well-known DEA
developments for applications in the ecological context. These reflections do not claim
completeness or profundity but represent a brainstorming to highlight the potential and
flexibility ofDEA for applications in EM/EC.
4.1 Modelling alternatives
At first sight, the integration of 'bads' in DEA does not seem to be a problem. Without
going into mathematical details, the main characteristics, limitations, and possible im
provements for the modelling alternatives are discussed in this section."
4.1.1 Data transformation
An intuitive approach to integrate 'bads' into DEA is shown for example in Lovell et
a!. (1995). Carbon and nitrogen emissions are converted by taking their reciprocal and
regarded as usual outputs afterwards. Conversely, a reduct may be treated as a normal
product after this kind of data transformation. The advantages of this approach are that
the selected DEA model can be applied without any modification and the ordinal
ranking of DMUs according to the inverted quantity stays the same. Though the scale
and intervals of the original data get lost and the reciprocal of zero values is not feasi
ble. Furthermore, it can be shown that the efficiency classifications and rankings when
choosing the reciprocal can differ from those of the alternatives 'translation' and 'pol
lutant=factor'. 25
Similar arguments appear when changing the data by translation, i.e. by subtracting
the pollutant (reduct) quantity from a positive constant and regarding it as an usual
product (factor). The constant is usually the maximal quantity of all DMUs regarding
24 For more details and an empirical illustration see Allen (1998b).
25 An example was given by Prof. Robert Dyson in the presentation on "Pitfalls and protocols in DEA'' at the European Summer Institute XVI "DEA- 20 years on", University of Warwick, England, 16--26 August 1998.
DEA in the ecological context- An overview 217
that pollutant (reduct) plus 1, in order to avoid zero values. As previously pointed out,
the DEA models are suitable without any modification. Although, on the other hand,
translation results in moving the zero to a different position and can yield results devi
ating from those of the alternatives.'•
Even if reciprocal or translation are applied, their use is dependent on the chosen effi
ciency measure. For example, if carbon emissions are modelled by taking the recipro
cal and an input-oriented model is chosen, ecological inefficiencies remain disre
garded. Such problems are important arguments in the following discussion of model
transformations.
4.1.2 Model transformation
Oriented models
Table 4.1 shows the structure of the ratio forms for the oriented models that have been
suggested for DEA applications in the environmental context." None of these models
accounts for 'reducts'. It is therefore assumed that the inputs exclusively consist of the
factors (F), the outputs comprise the products (P) and the pollutants (A). The transfor
mation of non-linear ratio forms into DEA LPs is demonstrated in Chames et a!.
(1978).
Model Input orientation output orientation
'pollutant=factor' p F+A max-- (I) min--
F+A p
'normalisation' A min A'=-p
'net production' P-F A (and 'profitability') max A minP-F
'joint output' P-A F max-F- min--
P-A
Table 4.1 Oriented models with pollutants
26 Please note the difference to the 'translation invariance' problem addressed e.g. in Pastor (1996), where negative variables are focused on without changing the minimax direction.
27 Some of these models are described in detail in the unpublished work of Prof. Mikulas Luptacik. I am grateful for his support.
218 DEA in the ecological context- An overview
As an intuitive approach to integrate pollutants into DEA, one can treat them as usual
factors. From an activity analysis perspective, the 'pollutant=factor' approach has the
advantage of a 'natural' min/max direction for each object category. Model ( 1) shows a
BCC model for input orientation according to the ratio shown in table 4.1.
(1)
min 8,A,s
s. t.
6-& L sk kEFvAvP
7t
L I.P xf + S; = ex? (i E F) p=l
7t
LI.Pyj +sj = eyJ (j E A) p=l
7t
LI.Pyj -s1 = yJ (j E P) p=l
I.P ~0 (p= 1, ... ,7t) ;s; ,s1 ~0 (i eF,j e Pu A) ;6 free
An inefficient DMU becomes efficient when it radially reduces its factors and pollut
ants according to e. Ecologically relevant impacts caused within the factors (consump
tion of non-renewable materials/'resource nature') and pollutants are accounted for,
but the interpretation of their proportional decrease may be difficult. The model is ap
plied e.g. in Piot-Lepetit/Vermersch (1998).
The corresponding output-oriented model aims at increasing the products. From an
ecological perspective it is not relevant. The model can be reduced to the minimisation
of a weighted sum of normalised pollutants as suggested by Tyteca ( 1996) and simi
larly by Haynes eta!. (1994). Factors and products are no longer explicitly considered.
The pollutant quantities are divided by a quantity measuring the DMU's activity level,
usually a production quantity. This simplified ecological efficiency measure is suitable
when technical efficiency is almost given and low ecological impacts are caused within
the factors.
DEA m the ecological context- An overview 219
In the input-oriented model 'net production' the pollutants are regarded as peculiar
outputs which should be minimised. An inefficient DMU becomes efficient when it
radially reduces its pollutants according to e. Possible improvements in the products
and factors which form a kind of 'net production' are neglected. Therefore, the model
is suitable for DMUs operating near to technical efficiency but on different pollution
levels. Though the model shows limitations, when ecologically relevant impacts are
caused within the factors. Changing to the output-oriented model results in a neglect of
inefficiencies within the pollutants. LP formulations of the 'net production' model are
presented as a CCR model e.g. in Tyteca (1996). An interpretation of 'net production'
(a weighted sum of factors subtracted from a weighted sum of products) might be con
troversial unless prices for factors and products are known. In this case, the efficiency
score can be regarded as a kind of ecological profit.
The main idea of the model 'joint output' is to consider the fact that undesirable out
puts m outputs of the production process, which is preferable from a material/energy
balance perspective. The input-oriented model considers inefficiencies of factors what
could be reasonable for production processes characterised by an intensive consume of
the 'resource nature'. The pollutants are indirectly considered by subtraction from the
product quantities. A change to the output-oriented model causes difficulty with inter
pretation. In this case, the objective function value yields the necessary radial increase
in a weighted sum of pollutants subtracted from a weighted sum of products, whereas
factor inefficiencies are neglected. The idea of a 'joint output' is underlying the models
in e.g. Fare eta!. ( 1989).
The neglect of ecological inefficiencies and the interpretation problems of the oriented
models would even deteriorate if reducts had to be integrated. For an improvement,
efficiency measures are needed, where individual, i.e. asymmetric, non-proportional
minimax directions for~ category respectively object can be defined. As a first step
target setting could be used. Thanassoulis/Dyson (1992) present models where a) one
input or output is given pre-emptive priority to improve, b) a general preference struc
ture can be specified that attaches different degrees of importance to (non
proportional) input and/or output changes, or c) a DMU can specify target levels it
would ideally wish to adopt. Inputs and outputs which are not included in the target set
are considered in the objective function only by means of a non-archimedean value.
220 DEA in the ecological context- An overview
An additive model
Being ecologically-motivated and taking a production theory, particularly an activity
analysis perspective, in Dyckhoff/ Allen ( 1998) a weighted additive model with vari
able returns to scale is derived. The integration of a multi-dimensional function allows
for the extension of this generalised AddVRS model for more complex preference
structures. Model (2) shows the resulting LP formulation, when specifying the model
according to the 'standard case'. Reducts are neglected, in order to simplify a compari
son with the oriented models below. By also neglecting the pollutants and having the
special case of all weights gk fixed on a value of one, this model (2) would then be
come an AddVRS model as presented by Charnes et a!. ( 1985).
max Po L gksk keFuAuP
7t
s. t. LIJ'xP + s I I = x? (i e F) p=l
7t
L"P p o YJ +s1 =y1 (j e A) (2) p=l
7t
Lt.Pyj -s1 = yJ (j e P) p=l
7t
LAP =1 p=l
t.P ~ o (p = 1, ... ,7t); s1 ,s1 ~ 0 (i eF,j ePu A)
The 'ECO-AddVRS' model (2) considers any kind of technical and ecological
(in)efficiencies. The underlying L1-measure aims at a maximisation of product, pollut
ant, as well as factor slacks. The ecological extension of the additive model is theoreti
cally founded and includes some modelling alternatives as special cases. However, the
main disadvantage is that the efficiency score is dependent on the input/output scales
and its interpretation is therefore not meaningful. Though the influence of scales can be
DEA in the ecological context- An overview 221
compensated by the suitable choice of weights gk, or by using efficiency measures
holding for units in variance."
The alternative approaches based on data and model transformations indicate that the
modelling of pollutants causes additional complexity and is obviously not a trivial task.
Apart from the aforementioned indications for improvements, further theoretical in
vestigation is advisable: What are the impacts of the alternative approaches? Which
relationships exist between the alternative approaches? And under which presupposi
tions should a certain approach be selected?
4.2 Axiomatic modifications
4.2.1 Relaxing the convexity assumption (FDH)
One fundamental assumption of DEA is that convex combinations of the 1t DMUs can
be technically realised." This assumption may be relaxed by using the Free Disposal
Hull (FDH) technology as described in Deprins eta!. (1984) and Tulkens (1993). In
this case, the DMUs are compared between themselves and classified 'efficient' as
soon as there do not exist dominating real DMUs. In contrast to classical DEA, ineffi
cient DMUs can not be opposed to virtual, i.e. non-existing convex combinations of
observed DMUs. In technical terms, DEA models with variable returns to scale (VRS)
like (1) and (2) become FDH models by appending constraints incorporating integer
variables as follows:
(3) f...PE{0,1} (p=l, ... ,7t)
Figure 4.1 contrasts the resulting staircase-like frontier (with more efficient DMUs) to
a DEAvRs frontier.
28 E.g. Lovell/Pastor (1995).
29 Postulate I in Banker et al. (1984).
222 DEA in the ecological context- An overview
product
.........
•
factor or •-------L...-..J pollutant
DEAvRs
FDH
'ideal' frontier
Figure 4.1 DEAvRS• FDH, and 'ideal' frontier (adapted from Tyteca 1996)
Tulkens ( 1993) describes the advantages of FD H over D EA as I) reference can not be
made to virtual DMUs, 2) FDH is less sensitive to outliers than DEA, 3) FDH is suit
able for situations in which reference has to be made to 'good practice', 4) efficiency
measures and necessary improvements of inefficient DMUs can be more explicitly
quantified, 5) the final choice of a benchmark for an inefficient DMUs stays with the
decision maker, and 6) computation is easier. Though Tyteca (1996) finds an interest
ing argument for preferring classical DEA over FDH in the ecological context "namely
that [ ... ] we do not worry about the fact that we compare existing points to an abstract,
'artificial' frontier since we would instead refer to a frontier that is eventually even
located outside the convex envelope of the existing points." He continues that the defi
nition of such an 'ideal' frontier- e.g. the dotted line in figure 4.1 - could be based on
either the best available technology ("technological definition") or a quasi zero-waste
state ("thermodynamic definition"). Neither a best practice nor a best available tech
nology concept, but just a dynamically adapted zero-waste state frontier fully corre
sponds to the idea of sustainability.30
4.2.2 Strong versus weak disposability
Table 3.1 showed several articles modifying the classical DEA assumption of strong
disposability for the pollutants, e.g. Fiire et al. (1996). Using the present paper's spe-
30 Tyteca (1996), pp. 295-298, and Callensfryteca (1998), last section.
DEA tn the ecological context- An overview 223
cific terms in their definitions, a factor is strongly disposable, if the same level of out
puts can be produced at no cost with higher quantities of factors. Products are assumed
strongly disposable, if it is possible to reduce their quantities at no cost using the same
inputs. For pollutants, the assumptions of weak disposability is introduced. Pollutants
hold for weak disposability, if their production can be reduced only at the expense of a
joint reduction in some other products, or a joint increase in the use of some factors,
where the joint movements are considered as proportional. Figure 4.2 opposes the dif
ferent technologies resulting from 1) regarding the pollutant as a strongly disposable
factor for the DEAvRS case or 2) assuming weak disposability for the pollutant."
product product (strong disp.) (strong disp.)
•
pollutant •-------'-----' (strong disp.)
1)
Figure 4.2 Strong versus weak disposability
• "--------'----1~ pollutant
(weak disp.)
2)
Vanden Eeckaut et a!. ( 1997) give the following estimation which approach should be
chosen: "The answer to this question relies on the nature of the problem analyzed.
When there is a strict relation between pollution and the inputs (due to regulation or
technological characteristics) then the first model has to be preferred [weak dispos
ability] .... However if it is possible to substitute pollution with inputs then pollution
has to be considered as an input [strong disposability]. An example of this substitution
31 Regarding the disposability assumptions and for further illustrations see e.g. Fare et al. (1996), Fare et al. (1989), Vanden Eeckaut et al. (1997), and Scheel (1998).
224 DEA in the ecological context- An overview
is the increase of labor which improves the quality of the control inside the plant and
by consequence the emission of pollution."
4.2.3 Further axiomatic modifications
The most important characteristic of production processes including ecologically rele
vant objects are strong relationships between factors and pollutants as well as between
products and pollutants. The influence of such relationships on the DEA results could
possibly be detected by correlation analyses before selecting inputs and outputs for the
DEA model. Also, axiomatic modifications could cope with such relationships.
Fiire/Grosskopf (1995) implement the notion of nulljointness- i.e. to produce 'good'
outputs one will ~ to produce some 'bad' outputs - into models with weakly dis
posable pollutants. Further interesting results can be expected when relaxing the DEA
assumption of fixed input/output relationships. In this case, a DMU can achieve effi
ciency not just by reducing undesirable and increasing desirable objects, but also by
'changing the mix' of its inputs and outputs.
4.3 Application driven extensions
In this section some well-known DEA developments are presented, which show a high
potential for DEA applications in EM/EC."
4.3.1 Categorical data
The convexity assumption of DEA causes problems in cases where input or output
variables appear as categorical variables. In such cases the reference units should be
constructed from DMUs which are in the same category or possibly from those in an
even more unfavourable category. Models for an incorporation of categorical variables
were presented by Banker/Morey (1986b) and Kamakura (1988).
In the ecological context, categorical variables could be useful e.g. to compare compa
nies on different levels of environmental certification, to describe pressure levels of
stakeholders or the level of competition surrounding a company, or to consider differ
ent technologies of e.g. waste burning plants. Golany et a!. (1994) treat S02-emissions
32 For the LP formulations, an empirical illustration, and further ideas see Allen (1998b).
DEA in the ecological context- An overview 225
as a categorical output. It should be noted that some of the variables mentioned here
may be exogenously fixed (see below).
4.3.2 Ordinal data
DEA assumes that all inputs and outputs can be measured on a cardinal scale. Though
it could happen that one or several factors may only be measurable on an ordinal scale
or that the DMUs can only be ranked relatively to such ordinal data. The incorporation
of ordinal data in DEA is addressed in Cook et al. (1993) and (1996).
Ordinal data may appear in EM/EC for example when waste coming out of a process
can be used in a downstream recycling process only if it reaches a pre-defined quality
level. For a pollutant coming out of a production process it may be useful to model its
toxicity class. This ordinal information can also express a rough estimation of the cor
responding disposal costs. Regarding data availability and standardisation problems in
the LCA context, experts could suitably substitute imprecise or non-existing quantita
tive or qualitative information on emissions by ordinal approximations.
4.3.3 Exogenously fixed inputs or outputs
In practical applications managers must often deal with situations where some inputs
or outputs are beyond their discretionary control. Banker/Morey (1986a) and Go
lany/Roll (1993) developed DEA models in order to estimate the extent to which the
controllable inputs/outputs can be improved while keeping the exogenously fixed in
puts/outputs constant.
In ecologically-oriented applications of DEA exogenously fixed inputs or outputs may
appear when describing the surroundings of a company, for example regarding the
strictness and costs of local environmental regulations, the availability of alternative
energy sources, the existence of purification plants or facilities for disposal of toxic
wastes, the population number in the neighbourhood, their and other stakeholders' at
titudes, as well as characteristics of soil, topology and weather.
4.3.4 Incorporation of weights restrictions and value judgements
The incorporation of weight restrictions and value judgements covers a considerable
part of the DEA literature, an overview is given in R. Allen et al. (1997). The flexibili
ty of the weights in DEA sometimes causes problems, for example when the DEA re-
226 DEA in the ecological context- An overview
suits crucially differ from prior views of the decision maker, when the discrimination
power of DEA is weak, or when the efficiency classification of a DMU results from
just one extreme input or output. To minimise these problems, even partial information
about the preferences of each kind should be incorporated into the analysis. In section
2, it was pointed out that the assessment of weights on ecologically relevant objects is
complex and difficult. Therefore, it is likely that in the ecological context 'simple'
more than 'sophisticated' approaches are used for an incorporation of weights restric
tions and value judgements. For example, decision makers might be able to express at
least ordinal preferences, assurance regions (i.e. relative relationships between
weights), or some absolute weight restrictions on pollutants, when information or an
agreement on their toxicity is not available.
4.3.5 Incorporating standards
Apart from the empirically observed DMUs one may sometimes want to integrate ac
cepted standards or benchmarks of performance into DEA because"[ ... ] it may happen
that DMUs forming the 'efficient frontier' are themselves lagging in efficiency behind
some known 'standard'." The integration of standard DMUs results in a 'standard en
velope' that can be composed of observed as well as standard DMUs and is usually
located outside the former 'self envelope'. The scores of inefficient DMUs will be
lower and previously efficient DMUs might now be evaluated as inefficient. It is rec
ommended to set several multi-dimensional standards instead of one single standard
for each input and output dimension."
From an ecological perspective, standards can be used for example to model environ
mental regulations. In this case, the standard DMUs consist of the input/output vector
of the observed DMUs, where the pollutant under regulation is replaced by the regula
tion's limiting value. Furthermore, in analogy to the discussion of an 'ideal' frontier in
section 4.2, one can define standards as best available or zero-waste technologies. In
both cases the incorporation of standards changes the DEA perspective from past ob
servations (static tool) towards projected values (prospective tool). A distance measure
representing the relative position of such standard envelopes and the best practice
frontier can provide useful information on the ecological position of the observed
DMUs. In their application on power plants in Israel Golany eta!. (1994) incorporated
33 For details see Golany/Roll ( 1994).
DEA m the ecological context- An overview 227
one standard DMU for each observed site. Management knowledge, government
regulations and technical data were taken as sources for the definition of the standard
DMUs.
4.3.6 Ecological efficiency over time
Information on developments over time are important for the EM/EC and could be
provided in the DEA framework for example by using window analysis or Malmquist
indices. Though it should be noted that regarding the impacts of pollutants on nature
absolute quantitative reductions are preferable compared to relative improvements,
where the production level and therefore the level of negative impacts of the pollutants
may be increased. This fact could be considered by selecting a model just focused on
pollutant minimisation or by including an ordinal variable which is high (low) for
DMUs that reduced (increased) the absolute pollutant quantities over time.
Although it is sometimes recommended to use window analysis to overcome the dis
advantage of a low number of DMUs due to data availability problems, this argument
should be regarded critically from an ecological perspective. A DMU could estimate
its ecological position as positive although being compared to old-fashioned technolo
gies or despite not accounting actively for environmental activities in the past.
5 A concept for applications of DEA in EPM
In section 2, environmental performance measurement (EPM), especially at the opera
tional level and by means of indicators and benchmarking, was described as an impor
tant task of EM/EC. The assessment of weights in this context is an extremely difficult
task. DEA does not need such a priori weights, but the state-of-the-art approaches de
scribed in section 4.1 still contain some technical or interpretation problems. Unless
more theoretical insight, improved models, and axiomatic modifications are available,
the potential of DEA for applications in the ecological context can be exploited by re
ducing the complexity caused by the integration of ecologically relevant objects. Fig
ure 5.1 illustrates a concept for applications of DEA in EPM.34
34 See Allen (1998a).
228 DEA in the ecological context- An overview
In step 1 the DMUs' inventory tables are collected to serve as the database. In the case
of numerous inputs and outputs, a selection of the most relevant inputs/outputs as well
as an aggregation of inputs/outputs should be undertaken as far as necessary and possi
ble. In step 2 the inventory tables are transformed into a categorised inventory table
according to the 'standard case'. The main idea of the concept (steps 3 to 5) is to de
velop a system of indicators, and then to apply DEA to each indicator. In terms of the
underlying ratio form, each indicator is constructed as a weighted sum of objects from
one desirable category and a weighted sum of objects from one undesirable category.
This structure is typical for EPis in practice. Either the indicators just consist of for
example one pollutant over one product quantity, or weights have to be assessed to
aggregate for example several factor over several product quantities. As an advantage
over common EPis, in this concept no weights are required to achieve an aggregated
efficiency score for multiple objects. Though if value judgement on inputs/outputs ex
ist they can and should be integrated in step 4. As an advantage over the unique indi
cators coming out of the models in section 4.1, these indicators allow for concrete in
terpretations. Especially, the choice of a core indicator set allows for consideration of
ecological and technical efficiency measures.
The concept can not solve general problems of EPM (e.g. data availability and stan
dardisation, see section 2) and is not unproblematic within itself. Neither the selection
and categorisation of inputs/outputs nor the development of the indicator set are trivial
tasks. Though experts are more likely to agree on these aspects than on weights for
inputs/outputs. The concept represents a simplified possibility to apply DEA for EPM
if one wants to avoid intensive theoretical investigations or if first insights are quickly
required to find out where detailed analyses should be focused. For applications of
DEA in EPM, for example for an ecological product benchmarking, a certain degree of
simplification and aggregation will always be necessary.
The line of thought underlying this concept is confirmed by proceedings and argu
ments in some of the first DEA applications in the EM/EC context, e.g. Callens/Tyteca
(1998): "[ ... ] it is important not to base decisions on one unique, aggregate
sustainability indicator; instead, it is suggested to develop two or three partial indica
tors that stress different aspects of the problem.""
" Theoretical arguments are given in Dyckhoff/Rildiger ( 1997).
DEA m the ecological context- An overview
3 Inventorv table OMU1 2
Inputs Outputs t!!!'l!.•~l ___ _J~rj!liw;ti_ __
lJ!I~t---- Jio.li.d.l!'UJL _ ~~r ____ wut;_w_ile_L-
r.~e..-v --- ~~:~~.--
'core' env. lndicaiOr system material: raw m:ue:nal~ r _____ ~~CIQ\!!Oiill~ __
solid \Y3Ste: wastes tor rccjfhmz ______ ~~e_9.u~tt.!!_~ __
waste water: water consumgqon _ _ _ _ _ l!!_odUC:t_gl!!!l.!!!!,C!.. - -
waste air: <» f"SP\qm!?Jons ______ e!.odJ!C!,9'!!,0~1~ __
energy: 'Pi~it0:~~m:n
r~ft;re!!.C'< . ..U!!!'L _____ _ '!_l~a_!!c~ o.f:. iJ2p'!,l,si~u!P~ __
~r~a,!!o.!! '![ ~j.:_c~~~~~i!!_e~ funher non·DEA analyses
Step l Inventory analysis
tep 2
categorisation of inputs/outputs
('standard case')
tep 3 choice of 'core' EPis (desirable over undesirable object)
tep 4 DEA on each 'core' indicator
Step 5
interpretation/conclusions
Figure 5.1 A concept for applications ofDEA in EPM
6 Summary and outlook
229
Starting from an overview on environmental management/environmental controlling
and their main problems, the potential of DEA applications in this scope was pointed
230 DEA in the ecological context- An overview
out. A literature review underlined this potential, but also revealed different and some
times arbitrary approaches regarding the integration of ecologically relevant objects in
DEA. Each of the 'state-of-the-art' approaches shows conceptual limitations or inter
pretation problems. It is therefore vital to go back into DEA's axioms and foundations
in the theory of production and particularly activity analysis background for a system
atic and consistent derivation of ecologically extended models. For example, in
Dyckhoff/ Allen ( 1998) a generalised AddVRS model is derived and ecologically
specified according to the 'standard case'. As several examples proved in section 4, it
is unproblematic to extend such theoretically-founded models according to the real
EM/EC application context.
To my mind, new flexible efficiency measures, axiomatic modifications as well as the
exploitation of synergies with multi-criteria and distance function approaches show an
extraordinary potential for DEA applications in the ecological context. Figure 6.1 in
cludes these keywords in a framework on three levels of different theoretical and prac
tical contents.
DEA for EM/EC
ecological extension of basic DEA models integration of e.g. weights, standards, targets, specific data types handle incomplete and unstandardised data, low number of OM Us specific pitfalls, aggregation levels, and needs ofEM/ EC
Puu·lt .llwn in tht.nr~
analysis of alternative approaches flexible efficiency measures axiomatic modifications, e.g. disposability, changing the mix absolute/relative development over time
parametric approaches stochastic approaches
,.......----~ --MCDM J ratio analysis
distance function approaches
applicatioo
Figure 6.1 Framework for future research
DEA in the ecological context- An overview 231
References
Allen, K. (1998a): Moglichkeiten und Grenzen der Data Envelopment Analysis im Rahmen des Oko
Controlling; in: Dyckhoff, H./Ahn, H. (Hrsg.): Produktentstehung, Controlling und Um
weltschutz, Heidelberg, 327-352.
Allen, K. ( 1998b ): Basics of DEA for applications in environmental management/ecological control
ling: Framework, foundations, ecological extensions, and prospects; lnstitut flir
Wirtschaftswissenschaften, RWTH Aachen, working paper 98105.
Allen, R. I Athanassopoulos, A. D. I Dyson, R. G. I Thanassoulis, E. (1997): Weights restrictions and
value judgements in data envelopment analysis - Evolution, development and future directions;
in: Annals of Operational Research 73, 13-34.
Athanassopoulos, A. D. I Thanassoulis, E. ( 1995): Separating market efficiency from profitability and
its implications for planning; in: Journal of the Operational Research Society 46, 20-34.
Athanassopoulos, A. D. I Lambroukos, N. I Seiford, L. (1998): Data envelopment scenario analysis
for setting targets to electricity generating plants, submitted working paper.
Ball, V. E. I Lovell, C. A. K. I Nehring, R. F. I Somwaru, A. (1994): Incorporating undesirable out
puts into models of production: An application to US agriculture; in: Cahiers d'economie et so
ciologie rurales 31, 60-74.
Banker, R. D. I Charnes, A. I Cooper, W. W. (1984): Some models for estimating technical and scale
inefficiencies in data envelopment analysis; in: Management Science 30, I 078-1092.
Banker, R. D. I Morey, R. C. (1986a): Efficiency analysis for exogenously fixed inputs and outputs;
in: Operations Research 34, 513-521.
Banker, R. D. I Morey, R. C. (1986b): The use of categorical variables in data envelopment analysis;
in: Management Science 32, 1613-1627.
Callens, I. I Tyteca, D. (1998): Towards indicators of sustainable development for firms - a produc
tive efficiency perspective; in: Ecological Economics.
Charnes, A. /Cooper, W. W. I Golany, B. I Seiford, L.M. I Stutz, J. (1985): Foundations of Data En
velopment Analysis for Pareto-Koopmans efficient empirical production functions; in: Journal
of Econometrics 30, 91-107.
Charnes, A. I Cooper, W. W. I Lewin, A. Y. I Seiford, L. M. (1994): DEA- Theory, methodology and
applications, Boston/Dordrecht/London.
232 DEA in the ecological context- An overview
Charnes, A. I Cooper, W. W. I Rhodes, E. (1978): Measuring the efficiency of decision making
units; in: European Journal of Operational Research 2, 429-444.
Cook, W. D. I Johnston, D. A. I McCutcheon, D. (1992): Implementation of robotics: Identifying
efficient implementors; in: Omega 20, 227-239.
Cook, W. D. I Kress, M. I Seiford, L. M. (1993): On the use of ordinal data in data envelopment
analysis; in: Journal of the Operational Research Society 44, 133-140.
Cook, W. D. I Kress, M. I Seiford, L. M. (1996): Data envelopment analysis in the presence of both
quantitative and qualitative factors; in: Journal of the Operational Research Society 47,
945-953.
Cooper, W. W. I Hemphill, H. I Huang, Z. I Li, S. I Lelas, V. I Sullivan, D. W. (1996): Survey of
mathematical programming models in air pollution management; in: European Journal of Op
erational Research 96, 1-35.
Courcelle, C. I Kestemont, M.P. I Tyteca, D. I Installe, M. (1998): Assessing the economic and envi
ronmental performance of municipal solid waste collection and sorting programmes; in: Waste
Management & Research 16(3), 253-263.
Day, D. L. I Lewin, A. Y. I Li, H. (1995): Strategic leaders of strategic groups: A longitudinal Data
Envelopment Analysis of the U.S. brewing industry; in: European Journal of Operational Re
search 80, 619-638.
Deprins, D. I Simar, L. I Tulkens, H. (1984): Measuring labor-efficiency in post offices; in: Mar
chand, M. I Pestiau, P. I Tulkens, H. (Eds.): The performance of public enterprises: Concepts
and measurement, North Holland, 243-267.
Dyckhoff, H. (1994): Betriebliche Produktion: Theoretische Grundlagen einer umweltorientierten
Produktionswirtschaft, 2"• ed., Berlin.
Dyckhoff, H. (1998a): GrundzUge der Produktionswirtschaft, 2"• ed., Berlin.
Dyckhoff, H. (1998b): Umweltrnanagement; in: Berndt, R. I Altobelli, C. F. I Schuster, P. (Eds.):
Springers Handbuch der Betriebswirtschaftslehre 2, Berlin, 389-431.
Dyckhoff, H. I Allen, K. (1998): An ecologically motivated generalisation of data envelopment analy
sis (DEA); submitted working paper.
Dyckhoff, H. I RUdiger, C. (1997): Ein M6glichkeitstheorem fur Oko-Indizes; in: Zeitschrift fur
Angewandte Umweltforschung (ZAU) 10, 58-65.
Epstein, M. K. I Henderson, J. C. (1989): Data Envelopment Analysis for managerial control and
diagnosis; in: Decision Sciences 20,90-119.
DEA in the ecological context- An overview 233
Flire, R. I Grosskopf, S. (1995): Environmental decision models with joint outputs; submitted working
paper.
Fare, R. I Grosskopf, S. I Lovell, C. A. K I Pasurka, C. (1989): Multilateral productivity comparisons
when some outputs are undesirable: A non-parametric approach; in: The Review of Economics
and Statistics, 90-98.
Fare, R. I Grosskopf, S. I Lovell, C. A. K. I Yaisawamg, S. (1993): Derivation of shadow prices for
undesirable outputs: A distance function approach; in: The Review of Economics and Statistics
75(2), 374-380.
Fare, R. I Grosskopf, S. I Tyteca, D. (1996): An activity analysis model of the environmental per
formance of firms - application to fossil-fuel-fired electric utilities; in: Ecological Economics
18,161-175.
Gallez, C. I Tyteca, D. (1998): Explaining the environmental performance of firms with indicators; in:
Us6, J. L. I Brebbia, C. A. I Power, H. (Eds.): Ecosystems and sustainable development- Ad
vances in ecological sciences, Vol. I, 317-330.
Golany, B. /Roll, Y. (1993): Some extensions of techniques to handle non-discretionary factors in
data envelopment analysis; in: The Journal of Productivity Analysis 4 ( 4), 419-432.
Golany, B. I Roll, Y. (1994): Incorporating standards via DEA; in: Chames, A. I Cooper, W. W.
Lewin, A. Y. I Seiford, L. M. (Eds.): DEA - Theory, methodology and applications, Bos
ton/Dordrecht/London, 313-328.
Golany, B. I Roll, Y. I Rybak, D. (1994): Measuring efficiency of power plants in Israel by data en
velopment analysis; in: IEEE Transactions on Engineering Management 41(3), 291-301.
Goldmann, B. I Schellens, J. (1995): Betriebliche Umweltkennzahlen und llkologisches Benchmark
ing; Schriftenreihe Wirtschaft und Umwelt, Bd. 6, Kllln.
Giinther, E. (1994): 6kologieorientiertes Controlling- Konzeption eines Systems zur l:lkologieorien
tierten Steuerung und empirische Validierung, Miinchen.
Haynes, K. E. I Ratick, S. I Bowen, W. M. I Cummings-Saxton, J. (1993): Environmental decision
models: U.S. experience and a new approach to pollution management; in: Environmental In
temationa119, 261-275.
Haynes, K. E. I Ratick, S. I Cummings-Saxton, J. (1994): Toward a pollution abatement monitoring
policy: Measurements, model mechanics, and data requirements; in: The Environmental Pro
fessional 16, 292-303.
Horvath, P. (1994): Controlling, 5" ed., Miinchen.
234 DEA in the ecological context- An overview
Jaggi, B. I Freedman, M. (1992): An examination of the impact of pollution performance on eco
nomic and market performance: Pulp and paper firms; in: Journal of Business Finance and Ac
counting 19,697-713.
Kamakura, W. A. (1988): A note on "the use of categorical variables in data envelopment analysis";
in: Management Science 34, 1273-1276.
Kao, C. I Yang, Y. C. (1991): Measuring the efficiency of forest management; in: Forest Science
37(5), 1239-1252.
Loew, T. I Kottmann, H. (1996): Kennzahlen im Umweltmanagement; in: Okologisches Wirtschaften
2, 10-12.
Lovell, C. A. K. I Pastor, J. T. (1995): Units invariant and translation invariant DEA models; in: Op
erations Research Letters 18, 14 7-151.
Lovell, C. A. K. I Pastor, J. T. I Turner, J. A. (1995): Measuring macroeconomic performance in the
OECD: A comparison of European and non-European countries; in: European Journal of Op
erational Research 87, 507-518.
Miettinen, P. I Hlimalainen, R. P. (1997): How to benefit from decision analysis in environmental life
cycle assessment (LCA); in: European Journal of Operational Research I 02, 279-294.
Pastor, J. T. (1996): Translation invariance in data envelopment analysis: A generalization; in: Annals
of Operations Research 66, 93-102.
Piot-Lepetit, I. I Vermersch, D. I Weaver, R. D. (1997): Agriculture's environmental externalities:
DEA evidence for French agriculture; in: Applied Economics 29, 331-338.
Piot-Lepetit, I. I Vermersch, D. (1998): Pricing organic nitrogen under weak disposability assump
tion: An application to the French pig sector; in: Journal of Agricultural Economics 49(1 ),
85-99.
Ray, S. C. I Kim, H. J. (1995): Cost efficiency in the US steel industry: A nonparametric analysis
using Data Envelopment Analysis; in: European Journal of Operational Research 80, 654-671.
Rudiger, C. (1998): Controlling und Umweltschutz: Grundziige eines koordinationsorientierten Oko
Controlling; in: Dyckhoff, H. I Ahn, H. (Hrsg.): Produktentstehung, Controlling und Um
weltschutz, Heidelberg, 271-298.
Scheel, H. (1998): Undesirable outputs in efficiency valuations; submitted working paper.
Schefczyk, M. (1993): Operational performance of airlines: An extension of traditional measurement
paradigms; in: Strategic Management Journal 14, 301-307.
Seidel, E. I Clausen, J. I Seifert, E. K. (1998): Umweltkennzahlen, Miinchen.
DEA in the ecological context- An overview 235
Sci ford, L. M. (1996): Data Envelopment Analysis: the evolution of the state-of-the-art (1978-1995);
in: The Journal of Productivity Analysis 7, 99-137.
Souren, R. I Rudiger, C. (1998): Produktionstheoretische Grundlagen der Stoff- und Energiebilan
zierung: Eine Analyse aus Sicht des 6ko-Controlling; in: Dyckhoff, H. I Ahn, H. (Hrsg.): Pro
duktentstehung, Controlling und Umweltschutz, Heidelberg, 299-326.
Stolp, C. (1990): Strengths and weaknesses of data envelopment analysis. An urban and regional per
spective; in: Computers, Environment and Urban Systems 14(2), 103-1 f6.
Thanassoulis, E. I Dyson, R. G. (1992): Estimating preferred target input-output levels using data
envelopment analysis; in: European Journal of Operational Research 56, 80-97.
Tulkens, H. (1993): On FDH efficiency analysis: Some methodological issues and applications to
retail banking, courts, and urban transit; in: The Journal of Productivity Analysis 4, 183-210.
Tyteca, D. ( 1996): On the measurement of the environmental performance of firms - A literature
review and a productive efficiency perspective; in: Journal of Environmental Management 46,
281-308.
Tyteca, D. (1997): Linear programming models for the measurement of environmental performance of
firms- Concepts and empirical results; in: Journal of Productivity Analysis 8, 183-197.
Tyteca, D. (1998): On sustainability indicators at the firm level: Pollution and resource efficiency as a
necessary condition towards sustainability; in: Journal oflndustrial Ecology 2(4).
Vanden Eeckaut, P. I Tyteca, D. I Courcelle, C. (1997): A simple business environmental perform
ance indicator based on productive efficiency, submitted working paper.
Werner, T. I Brokemper, A. (1996): Leistungsmessung mit System- Data Envelopment Analysis als
Instrument des Controlling; in: Controlling, Heft 3, 164-170.
World Commission on Environment and Development (1987): Our common future, Oxford.
Measuring Public Spending Efficiency in Brazilian
Municipalities: A Nonparametric Approach
Maria da Conceiyao Sampaio de Sousa and Francisco S. Ramos 1
Abstract
The importance of public expenditure in such social services as education and public health
represent a decisive contribution to a nation 's progress. A major problem is how to allocate
government spending in such a way as to provide public services efficiently. In a political
federation as the Brazil, the critical issue becomes the choice of the degree of decentralization
in the provision of public services. Decentralization is advantageous in the production and
distribution of public services due to: i) the proximity of users facilitates the ranking of the
priorities; ii) it is easier to control the use of resources; iii) the requirements of managerial
capacity are lower.
In this paper we assess the performance of Brazilian municipalities regarding the utilization
of public revenue. The paper inquires whether, for a given availability of services, local
governments minimize the expenditure needed to finance those services. To answer these
questions, a cost-efficiency frontier will be determined by using various techniques of
efficiency analysis: two DEA variants- DEA-F and DEA-V- and the FDH approach.
Our results suggest that the Brazilian recent municipal decentralization policy does not lead
to an efficient use of public resources. The outcome of this policy was a proliferation of small
municipalities that, due to their size, do not benefit from the "economies of scale inherent to
the production of certain public services. They tend to operate with higher average costs thus
bringing about a considerable waste of resources, which can be inferred by estimating the
excessive public spending that characterizes those cities.
1 Universidade de Brasilia and Universidade Federal de Pernambuco. We are indebted to CNPq (Conselho Nacional de Desenvolvimento Cientifico e Tecnol6gico) and IPEA (Instituto de Pesquisas Econ6micas e Sociais) for support during the elaboration of this paper. We would also like to thank Maria da Conceiyao Silva for providing the data set
238 Measuring Public Spending Efficiency in Brazilian Municipalities ...
Structure
Introduction
2 Methodology
3 Data
4 Efficiency Results
5 Efficiency and Returns to Scale: Does Decentralization Benefits Brazil?
6 Concluding Remarks
References
Measuring Public Spending Efficiency in Brazilian Municipalities ... 239
1 Introduction
The importance of public expenditure policies for economic development is by now
widely recognized.' In particular, investments in such social services as education and
public health represent a decisive contribution to a nation's progress. Regarding
income distribution, recent research shows that policies of public spending have played
an important role in creating and preserving a more equitable form of economic
growth. 3 Indeed, when the objective is reducing poverty and income disparities, those
policies seem to be more efficient than redistributive tax policies.' The latter point is
particularly relevant for Brazil, a country characterized by such extreme inequalities,
and correspondingly enormous social debt, that seriously jeopardize the long run
growth prospects of the economy. An adequate public provision of certain key goods
and services may contribute to reduce significantly those disparities.
In those circumstances, a major problem is how to allocate government spending in
such a way as to provide public services efficiently and attenuate income disparities. In
a political federation as is the case of Brazil, the critical issue becomes the choice of
the degree of decentralization in the provision of public services. There is now a broad
consensus that such decentralization enjoys a presumption of being substantially
advantageous in the production and distribution of public services. Various reasons are
given for local governments having a superior advantage in providing such services.
First, the proximity of users facilitates the identification of the priorities and thus
allows an easily ordering of the services to be supplied. Second, it is easier to control
the use of resources if they do not need to run through long bureaucratic channels
before reaching their destination. Therefore, decentralizing resources reduces the
opportunities for corruption and inefficiencies so commonly associated with the public
sector. Third, the requirements of managerial capacity are lower, which is essential for
the good performance of projects in small and medium sized municipalities.
2 See, for instance, Van de Walle and Nead (1995), Harberger (1977), Goode (1984).
3 See Selowsky (1979), Meerrnan (1979), Anand and Ravaillon (1993), Lipton and Ravaillon (1995).
4 Several studies have shown that the tax system is inefficient to correct income disparities generated by successive interventionist policies. See Andie (1977), Bird and de Wullf(1978).
240 Measuring Public Spending Efficiency in Brazilian Municipalities ...
In short, the need to make local governments responsible for the provision of a
significant share of public goods is, by now, largely recognized. However, the mere
decentralization of public spending does not guarantee, per se, a satisfactory provision
of the corresponding services. It is also necessary to ensure that public funds are used
in the best possible way by the municipalities. Yet, this condition is not easy to fulfill.
For instance, in Brazil, it is by no means clear that the additional resources granted by
the 1988 Constitution to the municipalities are being used to maximize the
community's welfare as measured by the array of public services made available to the
population5. So, it is necessary to establish efficiency criteria that could be used to
evaluate how well or poorly are local governments spending public resources.
The objective of this paper is to assess the performance of Brazilian municipalities
regarding the utilization of public revenue. The paper inquires whether local
governments allocate resources so as to maximize the supply of public goods;
alternatively, whether, for a given availability of services, they minimize the
expenditure needed to finance those services. To answer these and other related
questions, this paper analyzes the relationship between aggregate municipal spending
and the corresponding provision of public services. In particular, a cost-efficiency
frontier will be determined by using various techniques of efficiency analysis.
This paper is organized as follows. Section 2 presents and discusses the methodology
used to compute the efficiency levels of Brazilian municipalities. Section 3 presents
the database and discusses the physical indicators chosen as proxies for the supply of
public services. Section 4 provides nonparametric efficiency measures consisting of
FDH and DEA calculations. Section 5 covers the relationship between efficiency and
returns to scale. Finally, Section 6 draws some lessons and conclusions from the
experience of the Brazilian municipalities.
2 Methodology
Public expenditure patterns may be analyzed from a number of perspectives. Among
them we find studies which emphasize the role played by demand on the provision of
5In 1991, municipal revenues represented 5,8% of GDP compared with only 2,9% for the period 1980-87, before the implementation of the 1988 Constitution (Afonso (1994)).
Measuring Public Spending Efficiency in Brazilian Municipalities ... 241
public goods and those who dwell in aspects of the technical efficiency associated with
the production of those goods. The first approach is based upon the pioneer work of
Bergstrom and Goodman (1973), who see the effective supply ofloca1 public services
as a response of local governments to the demand for such services. Therefore, they
associate municipal expenditures with typical demand variables such as income,
relative prices, and other socioeconomic variables. This type of analysis is particularly
appropriate to explain the levels and distribution of the various local services. An
alternative approach highlights the efficiency aspects of local public spending as stated
in the work of Vanden Eeckaut, Jamar and Tulkens (1991). Here, the main concern is
to identify the minimum cost at which a given quantity of goods and services could be
provided. Therefore, it becomes a classical problem of cost minimization. Under this
approach, the emphasis is placed on the determination of a cost-efficiency frontier
associated with the production of public goods that would permit to estimate the
efficiency levels of the various municipalities. That approach was used in this study.
2.1 Measuring efficiency levels: parametric and nonparametric approaches
Efficiency levels may be measured by using different methodologies. The distinction
among those methods lies on the type of technique applied to depict the frontier. For
instance, parametric and nonparametric approaches could be used to determine the
boundaries of the production set. In the first case - the parametric approach - the
frontier is supposed to be described as a function characterized by constant parameters.
Firstly used by Aigner and Chu (1968), this method specifies a production function and
then determines its parameters so that the estimated function envelops the data from
above and minimizes the distance between the observations and the graph of the
function.' The obvious shortcoming of this method comes from the fact that the
computed efficiency levels depend, crucially, on the particular specification attached a
priori to the production function. The very definition of efficiency is intrinsically
linked to the functional form chosen.
' For a detailed discussion concerning the estimation of parametric production frontiers see Forsund, Lovell and Schmidt (1980) and Bauer (1990).
242 Measuring Public Spending Efficiency in Brazilian Municipalities ...
Nonparametric approaches have a more flexible character as they impose no functional
forms a priori on the underlying technology. They require only that the production set
fulfill such properties as free disposal, convexity, and piece-wise linearity of the
technology. Under this approach, the central problem is to determine, according to pre
established hypotheses, which observations could be considered elements of the
frontier. Those elements are found by solving, for each observation, a system of linear
equations. In the vast literature relative to this topic, the method of determining the
efficient points is known as Data Envelopment Analysis (DEA).7
Farrell (1957) was the precursor in the use of this type of methodology. In his seminal
paper, this author assumed proportionality between inputs and outputs (constant
returns to scale). Along the same lines, Charnes, Cooper and Rhodes (1978)
established a DEA method to evaluate cost-efficiency frontiers. The production set
generated by this approach - known as the CCR model, after their authors - is a convex
hull of the data set. For each observation, the input-efficiency is measured as the
horizontal distance between the observed point and the appropriate facet of the
polyhedron. In this procedure, the efficiency of an observed decision making unit with
respect to best observed practice in the sample of n observations is computed by
solving a nonlinear problem'.
This first DEA version - referred to as DEA-C from now on - implies strong
restrictions concerning the form of the productive set. In particular, it imposes constant
returns to scale and convexity. To tackle such problems, this approach was
subsequently extended to incorporate non-constant returns of scale. Two variants of
this method were proposed: one characterized by non-increasing returns of scale
(DEA-CD) and another allowing for the existence of variable returns (DEA-V).
Non-increasing returns to scale, as found in Fare, Grosskopf and Lovell (1985, 1994),
is modeled by adding an additional constraint implying that the production set is
convex, includes the origin and satisfies the condition of free disposal: it allows for
constant and decreasing returns to scale. If the restriction is modified to exclude the
origin, we have the DEA-V measurement of efficiency: it allows for variable returns to
7 See Chames, Cooper and Rhodes (1978, 1981 ), Fare, Grosskpof and Lovell (1990), Seiford and Thrall (1990).
'See Chames, Cooper and Rhodes (1978).
Measuring Public Spending Efficiency in Brazilian Municipalities ... 243
scale, with increasing returns for lower levels of output and decreasing returns for
higher levels of output. This model is known in the literature as the BCC model, after
Banker, Charnes and Cooper (1984)9 •
Although representing a clear advance relative to parametric techniques, these
approaches still rely on hypotheses very restrictive on the structure of the production
set. Deprins, Simar and Tulkens (1984) have proposed weaker assumptions. They
postulate that the frontier of the production set is simply the boundary of the free
disposal hull (FDH) of the data set. Strong disposability of inputs and outputs is
maintained as well as variable returns to scale but no convexity hypothesis is required.
In this method - henceforth referred to as FDH - the frontier is obtained by comparing
inputs and outputs so as to establish the dominant points. An observation is declared
inefficient if it is dominated by at least another observation, domination, here,
meaning the ability to produce more output with less input. Consequently, if an
observation is not dominated by any other it is declared FDH efficient. In terms of the
DEA linear programming approach, this implies adding to problem the n+ 1
restrictions L~., A.j ~ 1 and~ e {0,1 }, j =1, ... , n. 10
Figure 2.1 illustrate well this methodology for the one-input one-output case and allow
its comparison with the DEA techniques (DEA-C, DEA-CD and DEA-V). It shows the
FDH frontier as well as the DEA-C, DEA-CD, and DEA-V frontiers. The implications
of the convexity assumption are clearly perceived. The proportionality hypothesis in
Farrell's frontier presupposes constant returns to scale. The segment OF gives the
frontier of the polyhedral production set. For the DEA-C method, characterized by
constant and decreasing returns of scale, the frontier takes the piecewise linear form
represented by line OABC. Finally, excluding the origin, the segment VABC
corresponding to the variant DEA-V, which presents variable returns to scale, gives the
frontier of the production set.
We can evaluate quantitatively the efficiency level associated with the different
productive activities by using the concept of distance. Considering observation M on
figure 2.1, it is easy to compute the efficiency measures implied by the different
9 See Fare, Grosskopf and Lovell (1985, 1994)
10 For more technical details see Deprins, Simar and Tulkens (1984) and Tulkens (1990).
244 Measuring Public Spending Efficiency in Brazilian Municipalities ...
methodologies as the distance between this point and the various calculated frontiers.
Hence, MF, MY and MH correspond, respectively to the efficiency measures
computed by using Farrell (DEA-C), DEA-Y and FDH methods.
Figure 2.1: Alternative forms oftbe production frontier
output
_...er------e
MF->DEA-C
MY-> DEA-CD and DEA-Y
MH->FDH
Fonte: Tulkens (1990)
2.2 Computational Aspects
To assess the efficiency levels of the Brazilian municipalities we consider one input,
the aggregate cost of the various public services supplied by the community, and
several outputs here represented by the indicators of the different public services.
Those variables are described in detail in Section 3.
Measuring Public Spending Efficiency in Brazilian Municipalities ... 245
(2.1)
subject to
j = l, ... n
where ck and cj represent aggregated total expenditures of municipality k and j,
respectively; Ykr and y1, are the quantities of the rth output indicator of these
municipalities. The variables 'YJ represent the weights attached to the municipalities
against which the municipality k is being evaluated. The value of the objective function
gives the efficiency degree in terms of inputs. DEA variants - DEA-C and DEA-V -
calculations were obtained by adding to program (2.7) the appropriate constraints. To
compute the several variants of the linear programs embedded in DEA methods, we
used the MINOS solver incorporated in the GAMS (General Algebraic Modeling
Systems) language."
As for the FDH, in spite of the fact that this methodology was expressed in terms of an
integer programming approach, its implementation is very easy and does not require
solving any linear program. Exhaustive vector comparisons allow the computation of
all relevant variables. FDH efficiency measurements were then obtained by using a
special program written in the GAMS programming language. This program evaluates
the efficiency degree of the municipalities as follows.
An observation is said to be inefficient if one or more municipalities dominate it.
Recall that domination in this context means that (a) there are other municipalities that
spend less than the dominated and (b) all the indicators of public services of those
11 See Brooke and Meerhaus (1988).
246 Measuring Public Spending Efficiency in Brazilian Municipalities ...
others municipalities are equal or greater than those associated with the dominated
municipality. Conversely, a municipality is said to be efficient if it is undominated.
Using these definitions, the efficiency degree and the excess spending for each
municipality can be computed." Thus, if a municipality is inefficient and dominated by
more than one other municipality, the most dominant is the one with the lowest
expenditure. The degree of cost-efficiency for an inefficient municipality is computed
as the ratio between the expenses of its most-dominating municipality and its own
expenses. By construction, this ratio is smaller than one and greater than zero. The
excess spending, for each cost-inefficient municipality, is calculated multiplying the
complement to unity of its cost-efficiency degree (1-degree of efficiency) by the its
total expenses. These excess spending estimates society's loss "due to the inadequate
use of municipal resources.
3 Data
The implementation of the methodologies outlined above requires information about
aggregate total costs (inputs) and the quantity of public services available to the
population for the municipalities (outputs). In view of the large number of
municipalities and the interregional income disparities typical of Brazil, as well as to
obtain meaningful comparisons, the municipalities were aggregated by state of the
federation. In some cases, when the number of observations (municipalities) was small
thus restricting the possibilities of comparison, data were aggregated by region".
Initially, information for the 4984 Brazilians municipalities was collected. After a
detailed scrutiny of the data, the sample was purged by deleting the data for those
municipalities for which some key information was invalid or missing. This way, 620
communes were dropped since 493 of them had a recorded population of zero and 127
others had no data on current expenditure. 14 Subsequently, cities having more than
12 See Deprins, Simar and Tulkens (1984) and Vanden Ecckaut, Tulkens and Jamar ( 1991 ).
13 See subsection 3.2, for details.
14 The reason why this happens lies on the flourishing creation of new municipalities in Brazil. Indeed, some municipalities, although legally created, have not yet being dismembered from the mother-commune. As a result, they do not report output indicators. In addition, null values for
Measuring Public Spending Efficiency in Brazilian Municipalities ... 247
100,000 inhabitants were dropped to increase data homogeneity. Those large
municipalities will be analyzed late in a separate paper to determine their relevant
characteristics. The final data set was composed of 3756 municipalities aggregated into
15 states and 1 region (the Northern Region)".
3.1 Input and output indicators for the Brazilian municipalities
Aggregate total costs were computed as the value of municipal current spending
(input). 16 As for output measurements, due to the impossibility of quantifying directly
the supply of public services, they were approximated by a set of selected indicators,
which are observable factors taken as proxies for the services supplied. After a careful
choice, six output indicators were retained. A list of inputs and outputs are provided in
Table 3.1.
Table 3.1 - Input and Output Indicators and the Corresponding Municipal
Services
Indicators Code Source Municipal Services of which indicators serve as proxies
Input Indicator
I. Current Spending I.DSP92_C STN I. Aggregate total costs
Output Indicators
!.Total resident population I. POP91_T IBGE' I. Administrative services
2. Domiciles with access to safe 2. DOM91_A IBGE 2. Public health services water
3. Domiciles served by garbage 3. DOM9l_L IBGE 3. Public health services collection
current expenses may be also explained by the fact that the data were not yet been reported by the STN (National Treasure Secretary ).
1' For more details on the data set see Sampaio de Sousa, and Silva (1997).
16 We have dismissed investment expenditures as they are more erratic and would certainly jeopardize the comparison among municipalities. Such expenses will be considered when the study will be repeated for several years.
248 Measuring Public Spending Efficiency in Brazilian Municipalities ...
4. Illiterate population 4.ANALF91 IBGE 4. Educational services
5. Enrollment in primary and 5. EDUC MEC' 5. Educational services secondary municipal schools
Sources: 1: STN - Secretaria do Tesouro Nacional (National Treasure Secretariat); 2:
IBGE - Instituto Brasileiro de Estatistica, Censo Demognifico de 1991; 3: MEC -
Ministerio da Educa91io e Desportos.
Data for current spending were obtained from the Secretaria do Tesouro Nacional
(STN). The data for output indicators Ns. 1 to 4, on Table 3.1, were obtained from the
1991 Population Census. Data on enrollment in primary and secondary municipal
schools (indicator 5) refer to 1991 and were provided by the Coordenadoria Geral do
SEEC/INEP of the Ministerio da Educa91io e dos Desportos (MEC). Below, those
indicators will be briefly discussed.
Current spending (DSP92 _C) corresponds to the total amount of municipal current
expenditure for 1992 as defined by Law 4320/64. It is expressed in cruzeiros, at prices
of 1992." Total population (POP91_T) refers to the population residing between
August 31 and November 1, 1991. The variable Domiciles with safe water supply
(DOM91_A) includes the number of those served by Rede Geral (potable water
network). Domiciles served by garbage collection (DOM91_L) includes those where
the garbage is collected directly or indirectly by public or private services. In the first
case, the garbage is collected at the domicile whereas in the second case the garbage is
dropped in garbage dumps and subsequently removed. Data on illiteracy (ANAL91)
include people who are five years old or older and are unable to read and write a
simple letter in their current idiom. Finally, student enrollment (EDUC) is the sum of
students enrolled in the preprimary, primary, and secondary municipal schools.
17 This variable represented, in 1992, 75% of the total municipal expenditures.
Measuring Public Spending Efficiency in Brazilian Municipalities ... 249
4 Efficiency Results
The methodologies presented in Section 2 were used to measure technical efficiency
for the Brazilian municipalities. Due to obvious space constraints, we will discuss, step
by step, only results for the states of Bahia and Minas Gerais. The analysis can be
applied mutatis mutandis to the other states".
4.1 Measuring efficiency levels for the municipalities of Bahia and Minas
Gerais
In the analysis of the state of Minas Gerais, the sample consisted of only the
municipalities having less than 100,000 inhabitants. Hence, from the 723 communes
included in the database, we considered only 701 municipalities. In the case of the state
of Bahia, we exclude municipalities with population greater than 40,000 inhabitants.
That is, from a total of 411 municipalities, only 402 were included in the sample.
Tables 4.1 to 4.6 summarize the results obtained. Tables 4.1 and 4.2 categorize 701
municipalities of Minas Gerais and 402 municipalities of Bahia into efficient and
inefficient ones, according to three distinct methodologies: Farrell's method (DEA-C),
DEA-V, and FDH. Results for the variant DEA-C, which allows for constant and
decreasing returns, are not reported as they are highly correlated with those produced
by the DEA-V methodology.
Table 4.1 - Rating of the Municipalities of Minas Gerais by Methodology
Population Number of Efficient Municipalities Inefficient Municipalities
Classes Observa-
tions
DEA-C DEA-V FDH DEA-C DEA-V FDH
# T% # 1% # T% # 1% # 1% # 1%
18 Detailed results for all the Brazilian municipalities are shown in Sampaio de Sousa, Ramos and Silva ( 1997).
250 Measuring Public Spending Efficiency in Brazilian Municipalities ...
0-4999 198 I 0,51 2 1,01 100 50,5 196 99,0 197 99,5 98 49,5
5000-9999 191 0 0,00 3 1,57 110 57,6 188 98,4 191 100 81 42,4
I 0000·14999 99 I 1,01 2 2,02 62 62,6 97 98,0 98 99,0 37 37,4
15000-19999 74 I 1,35 I 1,35 58 78,4 73 98,6 73 98,6 16 21,6
20000-24999 35 2 5,71 2 5,71 26 74,3 33 94,3 33 94,3 9 25,7
25000-29999 22 I 4,55 2 9,09 15 68,2 20 90,9 21 95,4 7 31,8
30000-34999 18 I 5,56 2 11,1 14 77,8 16 88,9 17 94,4 4 22,2
>35000 64 3 4,69 18 28,1 53 82,8 46 71,9 61 95,3 II 17,2
TOTAL 701 10 1,43 32 4,56 438 62,5 669 95,4 691 98,6 263 37,5
Table 4.2- Rating of the Municipalities of Bahia by Methodology
Population Number of Efficient Municipalities Inefficient Municipalities
Classes Observa-
lions
DEA-C DEA-V FDH DEA-C DEA-V FDH
# % # % # % # % # % # %
0-4999 6 I 16,7 2 33,3 6 100 5 83,3 4 66,7 0 0,0
5000-9999 67 0 0,0 0 0,0 41 61,2 67 100 67 100 26 38,8
I 0000-14999 113 0 0,0 0 0,0 65 57,5 113 100 113 100 48 42,3
15000-19999 72 I 1,39 I 1,39 45 62,5 71 98,6 71 98,6 27 37,5
20000-24999 58 0 0,0 I 1,72 36 62,1 58 100 57 98,3 22 37,9
25000·29999 22 0 0,0 0 0,0 16 72,7 22 100 22 100 6 27,3
30000·34999 19 I 5.26 I 5,26 II 57,9 18 94,7 18 94,7 8 42,1
>35000 45 3 6.67 10 22,2 40 88,9 42 93,3 35 77,8 5 11,1
TOTAL 402 66 1,49 15 3,73 260 64,7 396 98,5 387 96,3 142 35,3
It is apparent from Tables 4.1 and 4.2 that the FDH results strongly diverge from those
obtained by using the DEA-C and DEA-V methods. Whereas the FDH methodology
shows a large number of efficient municipalities, on the other extreme Farrell's method
finds only 1.4% and 1.5% efficient communes, in the states of Minas Gerais and Bahia
respectively. This proportion slightly increases when the DEA-V method is used.
Several reasons could be invoked to explain such divergence.
Measuring Public Spending Efficiency in Brazilian Municipalities ... 251
Firstly, the frontier of cost efficiency computed by using FDH is based on a stricter
concept of domination than the frontiers calculated by the DEA methods. A
municipality is FDH dominated if and only if all his output indicators are inferior to
the ones of an efficient municipality - effectively observed- with which it is compared,
and its current spending is equal or superior to the expenses of this dominant
municipality. On the other hand, a municipality is DEA-dominated by a fictitious
observation defined as being a linear combination (convex if we use DEA-V) of a
group of efficient municipalities. Therefore, FDH frontiers depend on the possibilities
of factual comparison whereas DEA frontiers "fabricate" their own possibilities of
comparison. When, by lack of information, comparison is impossible, the FDH method
declares the observation efficient by default leading to an increase in the number of
municipalities ranked as efficient.
Furthermore, the convexity hypothesis incorporated in DEA frontiers puts an
unnecessary restriction on the underlying technology for producing public services. In
particular, Farrell's method (DEA-C), that adjusts a cost-efficiency frontier
characterized by constant returns, ignores local non-convexities, hence underestimates
systematically the efficiency degree of the municipalities. So, it may not constitute an
adequate adjustment of the production frontier. Considering that, the data's suggestion
that in most states there are scale economies for small municipalities, is a shortcoming
that may significantly reduce the credibility efficiency measurements based on DEA
C.1' This point will be further discussed (see Section 5 below).
Finally, recall from Section 2 that the method FDH envelops the data tightly while the
frontiers engendered by FDH, DEA-V, and DEA-C are "nested" in one another, with
the FDH frontier staying more close to the data and the DEA-C frontier being the
farthest away." Such a close envelopment results in having more municipalities being
ranked as efficient.
19 Vanden Eeckaut, Tulkens, and Jamar (1991) already pointed out the poor adjustment of Farrell's frontier for the Belgian municipalities they have examined.
20 See section 2.
252 Measuring Public Spending Efficiency in Brazilian Municipalities ...
4.2 Characteristics of the FDH methodology: efficiency by default and
outliers
More detailed FDH results are presented in tables 4.3 and 4.4. At this point, two
aspects of the FDH methodology deserve special attention, viz. efficiency by default
and outliers.
Efficiency by default: Recall that in the absence of a sufficient number of similar
municipalities ("pairs") with which one given municipality can be compared, this
municipality, instead of creating a relationship of the type dominant/dominated, is
declared efficient by default. This ranking of efficiency does not result from any
effective superiority but is due to the lack of information that would allow pertinent
comparisons. In addition, by construction, the FDH concept of efficiency by default
applies both to the municipality that presents the lowest level of spending and to those
with the highest values for at least one output indicator. Notice that the occurrence of
efficiency by default is higher for large municipalities. This is explained by the fact
that the number of large communes is relatively limited whereas there are plenty of
small municipalities. As the FDH method rests on the possibility of comparing
observed municipalities, some large municipalities are declared efficient merely due to
the impossibility of finding similar communes with which they could be compared.
This does not mean that those municipalities are inefficient; it only means that if there
were more municipalities in the relevant range, results could be different.
This extreme form of the sparsity bias that characterizes the FDH technique certainly
leads to a significant overestimation of the number of efficient municipalities and thus
constitutes a serious shortcoming of the FDH approach. Such a limitation is
particularly constraining when there are only a few observations and/or the data are
highly heterogeneous as is the case of the Brazilian municipalities. To partly correct
this problem, we suggested to use the results obtained with the use of other
methodologies. Thus, municipalities declared FDH efficient by default, and also found
to be efficient when the method DEA have been applied, were considered as
effectively efficient.
On the other hand, the FDH methodology is particularly suited to detect the most
obvious cases of inefficiency as this technique is very assertive regarding the
measurement of inefficiency. To each municipality declared FDH-inefficient, it is
Measuring Public Spending Efficiency in Brazilian Municipalities ... 253
possible to find at least one municipality in the sample that presents a superior
performance relative to the first (dominated) municipality.
Table 4.3- FDH Characterization of Efficient Municipalities: Minas Gerais
Classes of Population Number Efficient Efficient and Municipalities ofMuni- Munici- Dominating Efficient by Default cipalities palities Municipalities
Abs % (3/2) Abs % (5/2)
(I) (2) (3) (4) (5) (6)
0-4999 198 100 60 60,0 40 40,0
5000-9999 191 110 76 69,1 34 30,9
10000-14999 99 62 31 50,0 31 50,0
15000-19999 74 58 32 55,2 26 44,8
20000-24999 35 26 14 53,8 12 46,2
25000-29999 22 IS 9 60,0 6 40,0
30000-34999 18 14 5 35,7 9 64,3
35000 + 64 53 15 28,3 38 71,7
TOTAL 701 438 242 55,2 196 44,8
Outliers: By definition, nonparametric frontiers are defined by the extreme values of
the dimensional space of inputs and outputs. Thus, the appearance of outliers, atypical
observations that differ significantly from the rest of the data, may considerably
influence efficiency computations. It is thus necessary to verify whether the divergence
does not result from evaluation errors. However, once one is convinced of the
reliability of the data set, this kind of information may provide valuable information.
Table 4.4- FDH Characterization of Efficient Municipalities: Bahia
Classes of Population Number Efficient Efficient and Municipalities ofMuni- Munici- Dominating Efficient by Default cipalities palities Municipalities
Abs % (3/2) Abs % (5/2)
(1) (2) (3) (4) (5) (6)
0-4999 6 6 0 0,0 6 100,0
254 Measuring Public Spending Efficiency in Brazilian Municipalities ...
5000-9999 67 41 19 46,3 22 53,7
10000-14999 113 65 39 60,0 26 40,0
15000-19999 72 45 27 60,0 18 40,0
20000-24999 58 36 20 55,6 16 44,4
25000-29999 22 16 7 43,8 9 56,2
30000-34999 19 II I 9,1 10 90,9
35000 + 45 40 7 17,5 33 82,5
Total 402 260 120 46,2 140 53,8
For instance, one can identify among the efficient municipalities some that not only
dominate several others but also appear, systematically, as the most dominant
municipality (the one with the lowest expenditure ) and, hence, have a decisive
influence on the FDH measurement of the efficiency levels. These municipalities,
somehow, compared with their pairs, possess the "best technology" thus defining the
technological frontier. Removing one of those municipalities reduces the requirements
to belong to the frontier so those municipalities previously declared inefficient may
tum out to be ranked as efficient.
This point is illustrate in Table 4.5, where some selected municipalities of the state of
Minas Gerais are outliers in the sense discussed above, e.g. they dominate several other
cities thus having a unambiguous influence upon efficiency computations. Consider
first the case of Capiro Branco, which dominates 46 other cities. This municipality,
with a lower spending, has all its output indicators exceeding those of the 46 cities. In
addition, Capiro Branco was also ranked as most dominatini in 18 cases.
Besides Capiro Branco, the municipalities of Alto Jequitiba and Brasopolis also
contribute to build up the boundaries of the technological frontier. Indeed, for 17 out of
the 20 municipalities dominated by Alto Jequitiba, this commune appears as the IllQS!
dominating, thus determining the efficiency rank of the concerned municipalities. As
for Brasopolis, this proportion is still higher as it functions as the most dominating in
10 out of 12 cases. Indisputably, those cities contribute to define the best-practice
fmntkr for the production of public services. Removing them from the sample will
substantially increase the number of efficient municipalities. This is due to the fact that
the concept of efficiency is a relative one and depends not solely on the performance of
the municipality analyzed but also on the achievements of the other cities with which
they are compared. Notice that Itapeva and Jacui, in spite of the fact that each one of
Measuring Public Spending Efficiency in Brazilian Municipalities ... 255
them dominates 9 municipalities, do not influence the assessment of the efficiency
levels as there is no case where they function as most-dominating municipality.
Table 4.5 - Selected Dominating Municipalities: Minas Gerais
Dominating Municipalities Population # of Dominated # of cases where
( 1000 inhabitants) Municipalities the municipalities is ID.QS1: dominatinll
Capim Branco 6334 46 18
Concei9i!o do Rio do Ouro 7695 31 15
Alto Jequitiba 7435 20 17
Lumimirias 5193 23 12
Natercia 4361 20 5
Prados 7371 20 I
Resende 9706 20 8
Florestal 5053 17 I
Guimarania 5739 17 2
Lagoa Dourada 10118 14 2
Ribeirao Verde 3614 14 9
Borda da Mala 15410 13 6
Santa Rita de Caldas 9258 13 I
Brasopolis 13711 12 10
Bandeira do Sui 4100 II 3
Perdigao 4546 II 4
Crucilandia 4579 10 7
Muzambinho 17887 10 6
Sao Tiago 9642 10 I
ltaguara 10671 9 I
ltapeva 5529 9 0
Jacui 6616 9 0
It is worth to point out that many of those outliers are small cities with less than I 0,000
inhabitants. This may be explained by the fact that the possibilities of comparison are
higher for those municipalities but also suggest that small cities are typically
inefficient. This point will be further discussed in Section 5.
256 Measuring Public Spending Efficiency in Brazilian Municipalities ...
Another type of outlier applies to inefficient municipalities. Among those
municipalities, the outliers are the ones, which show extremely low efficiency levels
and/or are dominated by several other municipalities. They represent the most obvious
cases of inefficiency. Table 4.6lists some of those municipalities for the state of Minas
Gerais.21 A typical example is the city of Dores de Guanha. This municipality is
dominated by 55 other communes. Furthermore, to finance the public services supplied
to its population, this municipality, if efficient, needed to spend only 52,8% of its
present expenses. This amount corresponds to the expenses of the municipality of
Cipotiinea - its most dominating municipality - that presents all output indicators
superior to those of Dores de Guanha. From an administrative point of view, this
information has a great relevance. Indeed, this result indicates that Dores de Guanha
could make better use of its resources as several other municipalities - and Cipotiinea
in particular - do. This analysis applies also to the other municipalities listed in Table
4.6.
Table 4.6 -Method FDH: Selected Dominated Municipalities: Minas Gerais
Municipaliti Current Population Efficiency Excess # of dominant Most-es spending degree spending municipalities dominating
municipality
Cr$ 10• mil
Habitantes
Dores de 3159 5403 0,5277 1492 55 Cipotll.nea Guanha
Bel oriente 2413 16718 0,2118 17666 41 Montesia
Joanesia 2886 6941 0,6268 1077 30 Alto Jequitiba
Monjolos 2375 2941 0,5562 1054 28 Acaiaca
Sao 4130 8322 0,5562 1833 25 Resende Gons:alodo Rio Preto
Cajuri 2639 3721 0,6559 908 24 Fortalez
Alpercata 3485 6752 0,5191 1676 22 Alto Jequitiba
Nacip 2259 4003 0,761 540 20 CapelaNova Raydan
Joaquim 2651 4441 0,6722 869 19 CruciHindia Felicio
2' The complete list of the dominated cities is presented on appendix I
Measuring Public Spending Efficiency in Brazilian Municipalities ... 257
Marlieria 2410 3540 0,7154 686 IS Carranca
Iturama 33856 45699 0,3472 22100 16 Vi yo sa
Sao Miguel 2667 6315 0,6783 858 14 Alto Jequitiba
Nova Uniao 2520 4865 0,7726 573 13 Lumimirias
Santa Rita 2273 4064 0,784 491 12 Crucilfmdia do Ibitipoca
Serranos 2562 2036 0,6077 1005 II Rio Doce
Delfinop6lis 3836 6698 0,585 1592 10 Conceiyao do Ouro
Juramenta 3141 6389 0,5759 1332 9 Alto Jequitiba
Conceicao 3839 3839 0,4993 1922 8 Perdigao do Para
Felix 6478 11926 0,4731 3413 6 Brasopolis
Sao 5863 4941 0,3321 3916 5 Luminarias Francisco de Sales
Santa 26510 25931 0,407 15720 3 Saloure Barbara
Pratapolis 3402 9395 0,3894 8183 2 ltanhandu
5 Efficiency and Returns to Scale: Does Decentralization Benefits
Brazil?
The efficiency results for Brazilian municipalities pose a question that deserves to be
carefully examined: the relationship between the size of the municipality and its
efficiency. Indeed, the different methodologies utilized in this research seem to
indicate that smaller cities tend to be less efficient than larger ones. Both under the
FDH method and the DEA variants, the quality of the frontier adjustment improves
significantly as the size of the municipality increases. Figure 5.1 shows that the
proportion of inefficient municipalities is reduced as we consider municipality classes
with larger populations.
258 Measuring Public Spending Efficiency in Brazilian Municipalities ...
Figure 5.1 - Inefficient municipalities grouped according population in alternative
methodologies
~
I ~ ; -0
"1-
BAHIA
100
90 80 70
60 50 40
30 20 10
0-4999 5000-9999 10000- 15000- 20000- 25000- 30000- 34500+ 14999 19999 24999 29999 34999
Cla11 of populaUon
MINAS GERAIS
0-4999 5000-9999 IOOIJO.. 15000- 200IJO.. 25000- 30000· >35000 14999 19999 24999 29999 34999
Class of pepul>lion
I _ OEA.C 1-DEA-V
-FDH c____
Such a result is not uniquely explained by the lack of possibilities of comparison as
stated in the "efficiency by default problem" since it pertains also to the DEA methods,
particularly the DEA-C version, for which the phenomenon of"efficiency by default",
while existent, is much less important. Its justification lies rather on the existence of
local increasing returns to scale prevalent among small municipalities. For those
communes, a given proportional increase in all output indicators could be achieved
with a proportionally inferior augmentation of current expenditures. This implies that it
would be possible to increase the size of the typical Brazilian municipality and yet
Measuring Public Spending Efficiency in Brazilian Municipalities ... 259
provide the required public services to these expanded communities without incurring
in an equivalent increase in public expenditure.
Local non-convexities arise because cities too small are unable to exploit the
economies of scale that characterizes the production of certain public services, hence
do not use the available resources optimally. In the case of educational services, there
is ample evidence that operating costs decrease with enrollment due to existence of
high fixed costs. 22 Consequently, larger schools tend to be more cost-efficient because
the fixed costs are diluted among a higher number of students. This fact, clearly,
discriminates against small municipalities as their schools have, very probably, only a
few students on average and thus tend to present excessively high average costs. Were
those cities larger, they would be able to enroll a greater number of students and
reduce the cost per student without significant loss of educational quality.
A similar explanation applies to other local public services. For instance, the existence
of important fixed costs involved in the production of administrative services (e.g. the
creation and maintenance of a physical and human administrative infrastructure)
explains why the per capita cost of those services is probably higher for smaller
municipalities. A recent paper by Maia Gomes and MacDowell (1997) corroborate this
point. Examining the Brazilian case, they show that municipal per capita expenses with
personnel are substantially higher in small cities. For municipalities with 50,000 or less
inhabitants, those authors show that the average cost curve for personnel expenses is
clearly decreasing. Hence, the dismembering of municipalities creates unnecessary
administrative costs and pushes those communes to work on the decreasing portion of
their average cost curves23 .
Local increasing returns to scale are also responsible for the precarious adjustment of
Farrell's frontier. Indeed, only five municipalities and 3 among those with less than
30,000 thousands inhabitants, in Minas Gerais and Bahia respectively, are at the cost
efficiency frontier (Tables 4.1 and 4.2); the fitting does not significantly improve with
increases in the size of the municipality . This phenomen can be better observed by
22 See Tan e Mingat (1992)) and Sampaio de Sousa (1996).
23 For a good account of the dismembering of Brazilian municipalities, see Maia Gomes and MacDowell ( 1997).
260 Measuring Public Spending Efficiency in Brazilian Municipalities ...
carefully examining Farrell's multipliers.24 Within the DEA-C approach, the presence
of local non-convexities may be computed by the value of the sum of the weights as
stated in the linear program described in Section 2, LYi, evaluated at the optimal
solution. Thus. LYi < 1 implies that locally (for the municipality considered) returns to
scale are increasing; Iyi > 1 points out to decreasing returns; and when LYi = 1, returns
are constant25. Figures 5.2 and 5.3 illustrate this point. They show the logarithm of LYJ
for the municipalities of Minas Gerais and Bahia as a function of the population. We
observe that, in both states, for the majority of small municipalities, the dependent
variable (logarithm of LYi) is negative, indicating the existence of economies of scale.
Those results are maintained for the other states of the Federation as well. A brieflook
at the Farrell's multipliers for the other states indicates the existence of increasing
returns in municipalities with population under 50,000. Thus Farrell's approach, by
imposing proportionality between inputs and outputs, captures as inefficiency what
actually is local increasing returns.
Notice that there is no contradiction between these results and the prevalent decreasing
returns found in the aggregate analysis conducted in Section 3. Indeed, the smooth
frontier computed from aggregate data fails to acknowledge the considerable local
nonconvexities that characterize a large number of Brazilian municipalities.
Fig. 5.2- Minas Gerais - Returns to scale
1.5 .-----------------------,
.::.. 0.5
~ 0 r .0.5
j -1:: -2
••• •
• 80000 100000
• ·2.5 ...._ _____________________ __.
24 The weights Y; defined on section 2.
" See Banker, Chames and Cooper (1984).
Population
Measuring Public Spending Efficiency in Brazilian Municipalities .. .
Fig. 5.3- Bahia -Returns to scale
2,----------------------------------,
·;; e e .. .. e ~
1,5
... 0 .s
•
• • • • • ••
• •• • • 63000
Populalion
' •
•
83000
261
Summing up, the proliferation of small communes, resulting from the intense
dismemberment of municipalities since 1985, reduces the average efficiency levels of
the Brazilian cities and results in a considerable waste of resources. Tables 5.2 and 5.3
present estimates for those losses. The results show that the inefficiency losses are
substantial for municipalities under 10,000 inhabitants. Depending on the methodology
used, the waste varies from 11% to 50% of the available resources.
Table 5.2 - Excess Spending Under Different Methodologies: Minas Gerais
Classes of Number of Spending Excess of Spending Popu-lation Municipa-
lities
(1000 CR$) FDH DEA-V DEA-C
Abs % Abs % Abs % Abs %
0-4999 198 28,24 440996 47041 10,67 183241 41,55 223457 50,67
5000-9999 191 27,24 526056 58220 11,07 208046 39,55 258637 49,17
I 0000-14999 99 14,12 404209 32765 8,11 167093 41,34 184998 45,77
I 5000- 19999 74 10,56 452040 43289 9,58 206083 45,59 214979 47,56
20000-24999 35 5,00 246718 12976 5,26 103406 41,91 107158 43,43
25000-29999 22 3,14 226392 32381 14,30 104028 45,95 105984 46,81
30000-34999 18 2,57 159682 3249 2,04 52371 32,80 54180 33,93
>35000 64 9,12 1385493 135371 9,77 430306 31,06 552989 39,91
TOTAL 701 100,00 3841586 365296 9,5 1 1454576 37,86 1702385 44,31
262 Measuring Public Spending Efficiency in Brazilian Municipalities ...
Table 5.3 - Excess of Spending Associated with Different Methodologies: Bahia
Classes of #Munici- Spending Excess of Spending Popu-lation palities
{1000 CR$) FDH DEA-V DEA-C
Abs % Abs % Abs % Abs %
0-4999 6 1,49 12786 0 0,00 7245 56,66 7945 62,14
5000-9999 67 16,67 186591 64398 34,51 101085 54,17 88744 47,56
I 0000-14999 113 28,11 377071 129631 34,38 174535 46,29 154960 41,10
15000-19999 72 17,91 307973 107675 34,96 132526 43,03 123297 40,04
20000-24999 58 14,43 395502 110195 27,86 163176 41,26 160091 40,48
25000-29999 22 5,47 131466 34541 26,27 70537 53,65 67380 51,25
30000-34999 19 4,73 151509 51288 33,85 75118 49,58 70149 46,30
>35000 45 11,19 570848 43797 7,67 245270 42,97 295397 51,75
TOTAL 402 100,00 2133746 541527 25,38 969495 45,44 967967 45,36
These results indicate that the smaller municipalities are seriously handicapped
regarding the efficient provision of public goods and services. Although further
research is needed on this matter, the prevailing concentration of the Brazilian cities
into the population bracket of under 10,000 inhabitants represents a significant extra
cost for the country as a whole. Very probably, the typical size of those cities is far
below the optimal size required to minimize the cost of the production of public
services'•. Hence, a more rational utilization of public funds should consider not
dismembering but regrouping municipalities. Unfortunately, opportunistic political
considerations may well prevent any serious initiative in this direction. It would take
changing the present array of political and economic incentives to stop and possibly to
revert the dismemberment process.
6 Concluding Remarks
In this paper we have attempted to appraise, quantitatively, the efficiency levels of the
Brazilian municipalities. The objective was to evaluate the performance of the local
26 The paper by Maia Gomes and MacDowell also questions the economic and fiscal viability of the smaller communes.
Measuring Public Spending Efficiency in Brazilian Municipalities ... 263
governments regarding the utilization of public resources. For that purpose, the paper
analyzed the relationship between aggregate municipal current spending and various
indicators of the production of local public services by constructing nonparametric
cost-efficiency frontiers. Different techniques of efficiency analysis were used to
determine this frontier: two DEA variants - DEA-C and DEA-V - and the FDH
approach.
Results obtained by the different methodologies were compared and its main
advantages and shortcomings, discussed. Emphasis special was given to the FDH
procedure. Compared to the DEA techniques, this method is less restrictive as it is
based on weaker hypothesis. Furthermore, instead of calculating an abstract frontier by
referring to a fictitious combination of municipalities, as DEA methods do, this
procedure build up its cost-efficiency frontier by contrasting actually observed
municipalities. That gives to the efficiency scores, mainly those applying to inefficient
municipalities, a credibility that the DEA methods lack. Yet, a main concern with FDH
lies on the fact that, by lack of comparability, this methodology tends to declare a great
number of municipalities efficient by default thus providing limited discriminatory
power. This problem is particularly important when the pattern of observations is
relatively heterogeneous as is the case with most of the Brazilian municipalities. When
this heterogeneity was combined with a relatively small number of municipalities, the
FDH method nearly collapsed by declaring most of the communes efficient by default.
In such cases, the DEA variants proved to be more effective as they handle better this
kind of problem. In short, each method has its advantages and disadvantages. The
appropriateness of their use depends on the particular question being examined and
should not be determined a priori.
As for economies of scale, our results suggest that the Brazilian recent municipal
decentralization policy does not lead to an efficient use of public resources. The
outcome of this policy was a proliferation of small municipalities. Due to their size,
these communes do not benefit from the economies of scale inherent to the production
of certain publics services. They tend to operate with higher average costs thus
bringing about a considerable waste of resources, which can be inferred by estimating
the excessive public spending that characterizes those cities. Therefore, to prevent
further losses in the overall efficiency of local public spending, this excessive
dismembering of communes should be avoided.
264 Measuring Public Spending Efficiency in Brazilian Municipalities ...
It is important to stress the exploratory nature of this study. Efficiency scores should be
used carefully as more detailed analysis is required to determine if the measured scores
reflect genuine technical inefficiencies or if they are explained by the action of others
factors. For instance, in some cases, dominated municipalities may well be intrinsically
different from the dominating ones, and what is regarded as inefficiency could
correspond simply to the effects of such municipality-specific characteristics. In
particular, no attempt was made to include variables reflecting the quality of public
services. Therefore, the low efficiency scores found for some municipalities could well
result from the higher quality of the services provided. Furthermore, indications of
inefficiency arise not only out of administrative incompetence or the lack of
appropriate incentives. They may also be due to the fragility of the data set. Indeed,
due to the deterministic nature of nonparametric models, the computed efficiency
levels crucially depend on the quality of the information used. Missing variables,
measurement errors, and other statistical discrepancies may significantly reduce the
credibility of the estimated scores. Hence, a high priority should be conferred to
improving the quality of the data set by using any suitable information available.
Finally, a close examination of the data set seems to indicate that grouping
municipalities by state of the federation does not apprehend all the complexity of the
economic and social relationship that characterize the Brazilian municipalities. This
problem is particularly relevant when estimating cost-efficiency frontiers that are based
on peer comparisons. Significant differences among municipalities of the same state
substantially increase the degree of heterogeneity of the information, restrict the range
of comparability and, thus, distort the relative basis on which those models are
established. Hence, it is essential to redefine the aggregation base by using criteria
other than the ones implied by the traditional geopolitical division.
References
Aigner, D. J. and S. F. Chu (1968): On Estimating the Industry Production Function, American
Economic Review 58: 826-839.
Anand, Sudhir and Ravaillon, M. (1993): Human Development in Poor Countries: On the Role of
Private Incomes and Public Services, Journal of Economic Perspectives 7: 133-50.
Measuring Public Spending Efficiency in Brazilian Municipalities ... 265
Bauer, P. W. (1990): A Survey of Recent Econometric Development in Frontier Estimation, Journal
of Econometrics 46:39-56.
Bergstrom, T. C. and Goodman, R. P. (1973): Private Demand for Public Goods, American Economic
Review 63: 280-296.
Bird, R. and De Wulff, L. H. (1978): Taxation and Income Distribution in Latin America: A Critical
Review of Empirical Studies, IMF Staff Papers : 639-682.
Broke, A, Kendrick, D. and Meeraus, A (1988): Gams: A User Guide, The Scientific Press, USA.
Charnes, A., Cooper, W. W. and Rhodes, E. (1978): Measuring Efficiency of Decision Making Units,
European Journal of Operational Research I: 429-44.
Charnes, A., Cooper, W. W. and Rhodes, E. (1981): Evaluating Program and Managerial Efficiency:
An Application of Data Envelopment Analysis to Program Follow Through, Management
Science 27: 668-97.
CIDE - Funda9ao Centro de Informayoes e Dados do Rio de Janeiro (1996): Anmirio Estatistico do
Estado do Rio de Janeiro 93-94, Rio de Janeiro.
Debreu, G. (1951): The Coefficient of Resource Utilization, Econometrica 19: 273-92. New York:
Wiley and Sons.
Debreu, G. (1959): Theory of Value, New York: Wiley and Sons.
Deprins, D., Simar, L. and Tulkens, H. (1984): Measuring Labor Efficiency in Post Offices, m:
Marchand, M., Pestiau, P. and Tulkens, H. eds.: The Performance of Public Enterprises :
Concepts and Measurements, Amsterdam: North Holland.
Flire, R. Grosskpof, S. and Lovell, C. K. (1985): The Measurement of Efficiency of Production,
Boston-Dordrech: Kluwer-Nijhoff Publishing,.
Fare, R. Grosskpof, S. and Lovell, C. K.(l994): Production Frontiers, Cambridge: Cambridge
University Press.
Forsund, F., Lovell, C. A. K. and Schmidt, P. (1980): A Survey of Frontier Production Functions and
of their Relationship to Efficiency Measurement, Journal of Econometrics 13: 5-25.
Goode, R. (1984): Government Finance in Developing Countries, Washington: The Brookings
Institution.
Harberger, A. (1977): Fiscal Policy and Income Distribution, in Franck, C. R. and Webb, R.C. eds.:
Income Distribution and Growth in Less Developed Countries, Washington: The Brookings
Institution.
266 Measuring Public Spending Efficiency in Brazilian Municipalities ...
IBGE, (1994): Censo Demogn\fico de 1991, Numero I, Brasil, Rio de Janeiro.
IBGE/DPE/DECNA (1996): Regionaliza9ilo das Transa9oes do Setor Publico: Resultados das
Administra9oes Publicas e da Atividade Empresarial do Govemo, 1992, Rio de Janeiro.
Koopmans, T. C. (1957): Three Essays on the State of Economic Science, New York: McGraw-Hill.
Lovell, C. A. K. (1993): Productions Frontiers and Productive Efficiency, in: Fried, H. 0, Lovell, C.
A. K., and Schmidt, S. S. (1993): The Measurement of Productive Efficiency, Oxford
University Press.
Lipton, M. and Ravaillon, M. (1995): Poverty and Policy, in: Behrman, J. R. and Srinivasan, T.N.,
eds.: Handbook of Development Economics, vol. 3, Amsterdam: North Holland.
MacFadden, D. (1978): Cost Revenue and Profit Functions, in Fuss and MacFadden, eds.: Production
Economics: A Dual Approach to Theory and Applications Amsterdam: North Holland.
Maia Gomes, G. e MacDowell, C. (1997): Os Elos Fn\geis da Descentraliza9ilo: Observa9oes sobre as
Finan9as dos Municipios Brasileiros, 1995, Anais do XXV Encontro Nacional de Economia,
Recife, PE, pp. 645-660.
MEC/SEDIAE/INEP (1996): Estatisticas da Educa91io no Brasil, Brasilia, DF.
Meerman, J. (1979): Public Expenditure in Malaysia: Who Benefits and Why, New York: Oxford
University Press.
MF/STN (1996): Finan9as do Brasil: Receita e Despesa dos Municipios, Anode 1995, Volume XLI,
Brasilia, DF.
Perelman, S. (1986): Frontieres d'Efficacite et Performance Technique des Chemins de Fer, Annals of
Public and Cooperative Economics 4: 449-59.
Pestieau, P. and Tulkens, H. (1990): Assessing the Performance of Public Sector Activities: Some
Recent Evidence From the Productive Efficiency Viewpoint.
Sampaio de Sousa, M. C. (1997): Efficiency and Equity Aspects of Social Spending in Selected
Countries of Latin America and East Asia: A Comparative Approach, Anais do XXV Encontro
Nacional de Economia, Recife, PE, pp. 1328-1347.
Sampaio de Sousa, M. C and Silva, M. C. ( 1997): lndicadores de Servicos Publicos para o Brasil: .
Uma Analise em Nivel de Municipios, IPEA, Brasilia.
Schmidt, P.: On the Statistical Estimation of Parametric Frontier Production Functions, The Review
of Economics and Statistics 58: 238-289.
Measuring Public Spending Efficiency in Brazilian Municipalities ... 267
Seiford, L. M. and Thrall, R. M. (1990): Recent Developments in DEA: The Math Programming
Approach to Frontier Analysis, Journal of Econometrics 46:7-38
Selowsky, M. (1979): Who Benefits from Government Expenditures: A Case Study of Colombia,
New York: Oxford University Press.
Shepard, R. W. (1970): Theory of Cost and Production Functions, Princeton: Princeton University
Press.
Simar, L. (1992): Estimating Efficiencies from Frontiers Models with Panel Data: A Comparison of
Parametric, Non-Parametric and Semi-Parametric Methods with Bootstrapping, The Journal of
Productivity Analysis 3: 171-203.
Tulkens, H. (1990): The Measurement of Productive Efficiency by FDH Frontiers, Document de
Travail, CORE, Universite Catholique de Louvain, Louvain-la-Neuve.
Vanden Eeckaut, Tulkens, H. and Jamar, M.A. (1991): A Study of Cost-Efficiency and Returns of
Scale for 235 Municipalities in Belgium, Core Discussion Paper n• 9158, CORE, Universite
Catholique de Louvain.
Van de Walle, D. and Nead, K. (1995): Public Spending and the Poor: Theory and Evidence, The
World Bank.
Efficiency and Productivity of Norwegian Colleges
Finn R. F0rsund1 and Kjell Ove Kalhagen2
Abstract:
Regional colleges in Norway were reorganised in 1994 with the purpose of promoting
efficiency and productivity. This is the first effort of checking what actually has happened
afterwards with efficiency and productivity. DEA and Malmquist index approaches are used.
Data for three years, 1994, 1995 and 1996 at a department level for about 100 units where
collected by questionnaire and direct contacts. The three outputs where final exams
distributed on two types; short- and long studies, and research publications. inputs where
number of academic and non-academic staff in full time equivalents, current expenses other
than salaries, and building size in square metres. Typical cross section efficiency results show
a large share of efficient departments, with a disproportionate number of efficient
departments giving theoretical general education, and a large variation within the group of
inefficient units. The difference between professional and arts and science departments may
be explained by the nature of the teaching production function, but calculations for a sub
sample of professional departments (e.g. nurses, engineers, teachers) show almost the same
variation within this group. The productivity change each year was mainly positive, with most
departments experiencing a positive productivity effect from frontier shift, but a greater
variation from positive to negative as regards the contribution from catching up.
1 Department of Economics University of Oslo and The Frisch Centre
2 The Frisch Centre
270 Efficiency and Productivity of Norwegian Colleges
Structure
Background
2 Measures of outputs and inputs
3 The method
4 Data
5 Efficiency results
6 The productivity development
7 Further research
References
Efficiency and Productivity of Norwegian Colleges 271
1 Background
Pressure on public sector expenditures has generated interest in performance indicators
the last decades. Higher education in Norway is almost exclusively state run. The
sector consists of colleges and universities. Recent interest in overhauling the
performance of the public sector of Norway resulted in the creation of a Parliamentary
Commission looking into cost efficiency. Performance of the college sector was paid
special attention, because with effect from October 1994, 98 colleges were merged
into 26 new ones. One purpose of the reform was to obtain a more efficient use of the
resources according to educational- and research policy objectives. The task of the
Commission in 1997 was to find out if this potential has been realised.
The new state run colleges consist of totally 109 departments, varying from 1 to 8 with
an average of 4,5 departments. The colleges offers a lot of studies; professional studies
(health and social studies, teacher training, engineering, media, and degrees of
Bachelor of Commerce and graduate engineer), university subjects (minor and major
subjects), or arts and science in general. The colleges are fully financed by the
Ministry of Education, Research and Church Affairs.
In contrast to universities, colleges are relatively more teaching intensive. Another
difference is that the colleges, although required to carry out research, do not have a
national responsibility for performing basic research.
As a part of the work of the Commission the Frisch Centre has undertaken to
investigate the efficiency and productivity of colleges for the relevant time period.
The department level turned out to be the most disaggregated level suitable for data
collection. In our analysis we will regard each department in the colleges as
comparable production units producing education and research. A more ideal level
would have been each study organised under departments.
The initial plan was to collect data for a suitable number of years before the reform
and up to the latest available year, 1996. But it turned out to be impossible to get data
for the pre-reform period for enough departments, leaving us with data for the years
1994, 1995 and 1996. With such a limited number of years our intention with the
productivity part of the study is more to explore the possible methods and result
presentations rather than offer conclusive insights. In defence of the exercise it may be
underlined that this is the first time such an exercise is performed with the applied
272 Efficiency and Productivity of Norwegian Colleges
methodology, and it may serve as a catalyst for improving the data production in the
sector, or as Rhodes and Southwick (1993, p.146) expressed it: " .. our intention in this
exploratory exercise is to identify areas for more thorough investigation and to bring
some light, however dim, on a question of relative performance that has received little
previous exposure".
When studying inefficiency there are two methodological problems that should be
separated: i) establishing a frame of reference for efficient operations, ii) defining the
efficiency measures. As to the former we will use the non-parametric approach of
DEA, as introduced by Chames et al. (1978) based on an idea of Farrell (1957),
assuming a piecewise linear frontier production structure, and as to the latter we will
use the Farrell (1957) efficiency measures. The motivation for imposing a minimal
structure on the production possibilities is that the technology for college production is
rather unknown, and typically multi-output. Furthermore, there are no prices on
outputs; they are not traded in markets.
Among previous studies using DEA for analysing efficiency in higher education
related to our study we would like to mention Tomkins and Green (1988), Ahn et al.
(1989), Beasley (1990), Rhodes and Southwick (1993), Johnes and Johnes (1993) and
(1995), Doyle and Arthurs (1995), and Sarafoglou and Haynes (1996). Typically, all
studies have used proxies for the ideal output variables (Flemming, 1991 ), such as
number of students at different levels, exam points, number of research publications of
various categories, and research grants. Inputs used have been number of employees
of different categories, especially faculty- and administrative staff, wage bill, building
and equipment investments, expenditure general and maintenance, equipment, support
functions, and research grants. Only Ahn et al. (1989) use data for several years, but
do not calculate productivity changes, but focus on changes in efficiency scores by use
of "windows analysis". We will explicitly calculate productivity changes. The studies
all show a significant dispersion of efficiency scores, and deal with sensitivity analyses
in different ways to illustrate the impact of choice of model specifications. We may
note that quality issues seldom have been dealt with, probably due to Jack of data, but
Rhodes and Southwick (1993) do a two-stage analysis with quality-related variables in
the second stage of correlating efficiency scores with explanatory variables.
Efficiency and Productivity of Norwegian Colleges 273
Conceptual issues in defining outputs and inputs are dealt with in Section 2. The DEA
method and Malmquist index are presented in Section 3, and the data structure is
shown in Section 4. The efficiency distributions are given in Section 5, and
productivity results and a more detailed analysis of productivity determinants
performed in Section 6. Some remarks on further research are offered in Section 7.
2 Measures of outputs and inputs
When studying productivity and efficiency the key to success is, first of all, to base the
study on theoretically satisfactorily definitions of outputs and inputs, and then to
operationalise these definitions without compromising too much. A fruitful approach
to understand what the institutions in question are producing, is to inspect the
objectives of the activities. In general terms a college produces educational services,
research, and dissemination of knowledge in society at large. Ideal measures of outputs
may be measures of the human capital added for students taking degrees as to
education, addition to scientific knowledge as to research (person-specific knowledge
and general knowledge, according to Beasley, 1990), and increase in enlightenment of
society at large as to interactions college - society (and contribution to "national
culture" according to Higgins, 1989). Operational measures of the first category may
be number and type of exams. Research may be measures by number of research
publications of different types; from prestigious international journals to national
language local working papers (see e.g. Johnes and Johnes (1993) for a classification).
Interacting activities may be measures by newspaper articles, media appearances,
participation of scientific staff in public commissions, and consulting for public and
private sector. Ideal and most commonly used measures are presented in Table I.
The classification of inputs can in general be cast in the KLEM format, i.e. Capital3,
Labour, Energy and Materials. Ideal measures of inputs are hours of labour of different
types, such as scientific faculty, administration and support staff, building space,
various categories of equipment, and current inputs such as energy, cleaning,
maintenance, postage, telephones, stationary. It is usually possible to operationalise
Labour straightforwardly by hours worked by different categories. Areas of buildings
3 K is used instead of C due to tradition.
274 Efficiency and Productivity of Norwegian Colleges
Table 1: Ideal output measures and operationalisations
Variables
Education
Research
Interaction society
Ideal measures
Addition to human capital
Addition to scientific knowledge
Increase in general knowledge, impact on decision-making
Operationalisations
Stock of students, Flow of exams, degrees
Research publications External research funds PhD's
Newspaper articles, media exposure, participation in public commissions, consultancies
may be supplemented with year of building to indicate functionality. Equipment
should include PCs, but these are difficult to operationalise because ideally we are
interested in the potential productivity of the PC, and actual purchase or replacement
value do not correspond well to the role of the equipment in research. May be capacity
in Bytes and speed in Herz could serve. Usually one has to use purchase figures, and
we have to cope with the distortions created.
The quality dimensions are of especial importance for college outputs. Number and
types of exams do not tell us the full story of the addition to human capital. One way
of capturing the quality dimension of exams would be to have a measure of the
success of the candidates after graduation. In a society where wages are strongly
influenced by productivity a measure of lifetime income would serve as a quality
measure of education. But such information is very difficult to come by, and the
egalitarian structure of Norwegian wages makes the quality signals very weak. A
more limited measure would be the time it takes for students to get jobs after
graduation, assuming that people from the most prestigious colleges get jobs first
(see e.g. Johnes et a!. ,1987). But such measures, which are possible to get from
special labour market surveys, depend heavily on the state of the relevant labour
market. With a low rates of unemployment, as in Norway in the relevant years, many
candidates experience such low waiting times that a correlation breaks down, e.g.
Efficiency and Productivity of Norwegian Colleges
Table 2: Quality dimensions
Variables
Education
Research
Interactions society
Student material
Staff material
Quality measures
Time before getting first job Income level Reputation of college
Citations Peer recognition
Impact on decisions
Qualifications at entry Number of hours studying
Degrees Seniority Participation in networks International conferences
275
because a need for a holiday before entering the labour market may be more
influential than the expected quality of the education.
Quality of research could be captured by influence measures by citation indices (but
see e.g. Flemming (1991) and Higgins (1989) on problems using these). The extent
(national/international) and type of networks of faculty could represent quality, and
also international conference participation. Where relevant the diffusion of research
into practical adaptations in business could be a measure of quality.
It is very difficult to measure the quality of the interactions with society. Impacts
through citations of media exposure could be one way.
The role of students should be paid particular attention. Students are the "carriers" of
education outputs, but are also inputs. The personal qualities of the students determine
how much human capital is actually absorbed during the education. The number of
hours used by students studying will obviously also influence the build-up of human
capital.
Quality of staff may also be of importance. Measures used have been years of
experience, seniority, etc. (see e.g. Johnes and Johnes (1993) p. 343).
276 Efficiency and Productivity of Norwegian Colleges
The use of proxies for the ideal variables, as portrayed in Table 1, makes necessary
explicit measures of quality. Some suggestions are provided in Table 2.
3 The method
3.1 The DEA Approach
The technology set, S, can in general be written:
S = { (y,x) ly can be produced by x} (1)
where y is the vector of M outputs and x a vector of R inputs. It is assumed that the set
is convex and exhibiting free disposability of outputs and inputs. Farrell (1957)
technical efficiency measures can be defined with respect to this set, and they are
identical to distance functions (introduced to economics in Shephard, 1953) or their
inverse. The input-oriented technical efficiency measure, E1j for unitj is:
(2)
i.e. we seek the maximal uniform proportional contraction of all observed inputs
allowed by the feasible technology set.
Introducing a set ofN observations the set, S, is estimated as a piecewise linear set by:
S={(y,x) I LAnYnm ~Ym (mEM),x, ~ LA.nxnr (rER),A.. ~O(nEN)} (3) neN neN
where An is the weight for observation n when defining the reference point on the
frontier, and N, M, R are also used as symbols for the index sets. It is assumed that the
Efficiency and Productivity of Norwegian Colleges 277
envelopment of the data is done as "tight" as possible, i.e. minimum extrapolation and
inclusion of all observations are assumed. Further, constant returns to scale (CRS) is
specified. A special form of variable returns to scale (VRS) is obtained by restricting
the sum of the weights to be I:
(4)
A piecewise linear production set with (4) included was first formulated in Afriat
( 1972) as the relevant set for efficiency analysis.
The estimator for the input-saving efficiency measure for observationj is then:
£ 1•1 = (5)
Min{e I LAnYnm ~ Yjm (Vm EM), Bx1, ~ LAnxn,(Vr E R), LAn =I, An~ 0 (Vn EN}} .<,0 neN neN neN
This problem is a linear programming problem with M+R (CRS) (+1 if VRS)
constraints, and can be solved in a standard wa/. Following Charnes et al. (1978) this
is called the DEA model. The VRS case was reintroduced by Banker et a1.(1984),
without reference to Afriat ( 1972).
The Farrell technical efficiency measures are radial, and measure the relative distance
to the frontier from an observation. There are two natural directions: keeping output
fixed and input-orient the measure, and keeping input fixed and output-orient the
measure. The efficiency measures can be interpreted as total factor productivity
measures in the standard meaning of an index of outputs on an index of inputs. The
input-oriented (or input-saving) measure is the ratio of the productivity of the
observation and the corresponding reference point on the frontier, keeping outputs
constant, the output-oriented (or output-increasing) measure is the ratio of the
productivity of the observation and the corresponding reference point on the frontier,
keeping inputs constant. Since the numerators (denominators) of the productivity
4 We are using an in-house program of the Frisch Centre.
278 Efficiency and Productivity of Norwegian Colleges
indices in the input-oriented (output-oriented) case are identical, we do not have to
worry about how the output (input) index is constructed. The efficiency score is based
on proportional change of all magnitudes. Assuming that the input (output) index is
homogenous of degree 1 in the inputs (outputs) the unknown input (output) index for
the observation cancels out, and we are left with the efficiency score (see F0rsund
(1997) for further explorations).
For a VRS frontier technology the basic efficiency measures are extended to cover
scale (see, F0rsund and Hjalmarsson 1974, 1979). A sort of a scale measure, termed
gross scale measure in F0rsund and Hjalmarsson (1979), but here renamed more
appropriately technical productivity measure, is defined as the ratio of the productivity
of the observation and the productivity at the corresponding ,(i.e. keeping observed
output ratios and input ratios) technically optimal scale point on the frontier. We know
(see Frisch (1965) or e.g. F0rsund, 1996a) that the latter productivity is maximal. The
pure scale measures defined in F0rsund and Hjalmarsson (1979), here simplified to
scale measures, may also be interpreted as productivity measures by forming ratios of
productivities with the input- and output corrected reference points respectively on the
frontier and optimal scale point. To realise that also in these cases we do not have to
know the productivity indices is a little more involved, and require the introduction of
the enclosure of the VRS production function by the smallest cone, i.e. a CRS
technology. We will return to this explanation after the graphical presentation of the
DEA frontier and the efficiency measures provided in Figure 1.
Two inefficient units, P1 and P2 are shown in Figure 1, and the concepts used in DEA
analysis are introduced. The efficiency measures for observations P1 are:
Input- saving efficiency: E1 = Xp /x 1
Output- increasing efficiency: E2 = Y1 /yG,
Scale efficiency, input orientation: E4 = E3 I E1 = (y 1 /xp) I (Ya lxa) ,
Scale efficiency, output orientation E5 = E3 I E2 = (yG/x1) I (y8 /x8 ).
Efficiency and Productivity of Norwegian Colleges 279
Output CRS-frontier
Referencing units, peers ( E1)
y
YG
\ VRS-fronti"
YB
YI
Yz
Figure 1:
0
Self evaluator
!~ Frontier reference point (E1 )
•A!
L__ Output-slack (E 1) . r~~- .. , Input
DEA frontier, concepts and efficiency measures
The way these measures are defined they are all between zero and one. The
productivity- and scale measures can be expressed as ratios of productivity of the
observation, P1, and its two corresponding frontier points, F and G respectively, and
the maximal productivity at the frontier at B. These measures can also be expressed as
ratios of the slopes of the rays from the origin through these points and the slope of
the ray to the point of maximal productivity, B. Returning to the productivity
interpretation above for the E3, E4 and E5 measures in general, note that the
productivity measure is identical to the input- and output-oriented efficiency measures
with the CRS support technology as the frontier reference technology, as stated above
for Figure I. But this is a general result because with more dimensions we require that
observed output ratios and input ratios are kept fixed. Therefore, the last two relations
are also general. These can than be used to give E4 and E5 productivity interpretations.
The two main technologies, CRS and VRS are shown in the figure. We note the
special feature ofVRS in the DEA case: the technology does not include the origin. A
280 Efficiency and Productivity of Norwegian Colleges
non-increasing returns to scale technology (NIRS) could also be specified, in Figure I
with OBCD as graph.
The terminology we will use is indicated in Figure 1. The efficient units when
calculating the efficiency score for an inefficient unit are termed referencing units, or
peers i.e. the efficient units with positive 8-weights in (5), and the point on the frontier
is the reference point. Calculating, in the VRS case, E1 for unit Ph units A and B are
referencing units (peers) and F is the reference point. Unit D is efficient, but is a self
evaluator calculating both input- and output- oriented measures.
We know slacks are an integral part of a LP problem. In Figure 1 we have an output
slack when calculating E1 for unit P2• With more dimensions we can also have input
(output)- slacks when calculating input (output)-oriented efficiency, and we have a
choice of presenting the radial efficiency measures, or non-radial ones including slacks
(see e.g. Torgersen et al. (1996) for an overview).
Finally, the LP programme also calculates the duals and gives us all the shadow prices,
which can be utilised to calculate marginal transformation rates and productivities.
The Farrell technical efficiency measure in the CRS case (E1 = E2 ) is the most used,
but also the extended Farrell measures have been used in the literature under various
names. However, the comprehensive scheme offered above, predating this literature,
based on F0rsund and Hjalmarsson (1974) and (1979), seems to have gone mainly
unobserved5. Since the student enrolments of colleges are determined by the
Government it is most relevant to calculate input-saving efficiency measures here.
3.2 The Malmquist productivity index
The productivity index is based on binary comparisons for a production unit between
two time points (or between two different units at the same point in time). The time
5 For instance, Banker et al. (1984) call E3 for "technical and scale efficiency", and E4 for "(input) scale efficiency", while Fare and Lovell (1978), Fare et al. (1985), Fl!re et al. (1994a) do not recognise E3 as a scale measure, but as a technical efficiency measure for CRS technology, probably due to E3 = E1 (CRS) = E2(CRS), and call E4 input scale efficiency measure and E5 output scale efficiency measure.
Efficiency and Productivity of Norwegian Colleges 281
periods to be compared, are denoted 1 and 2 for short. Only quantities are involved,
and at least one technology has to be known. As a convention we will compare a unit
observed in period 2 with the same unit observed in period 1, i.e. expressions
involving period 2 observations will be in the numerator and expressions involving
period 1 observations will be in the denominator.
Introducing cross-section data sets for several years the technology set, S, has to be
dated, e.g. S\ t 0 T, where T is the set of years. Caves et al. (1982) introduced
productivity indices for discrete observations based on Malmquist (1953). The basic
idea is to utilise Farrell efficiency measures, or distance functions, for the two
observations against a common reference frontier. An efficiency measure can itself be
interpreted as a ratio of the observed productivity and the productivity at the
corresponding point on the reference technology. The Malmquist productivity index,
~1 1·2 , for comparison between two time periods 1 and 2 for a unit j with frontier
technology from period 1 as reference, based on input-oriented efficiency measures, is:
(6)
1,2ET,jEN
The index notation system is that observation years are shown as superscripts, and
technology year as subscript. We have picked out two years called 1 and 2 as
observation years, and used one of them, 1, as technology reference. In general, any
year in the set t can be used as technology reference. The numerator shows the
proportional adjustment, by the scalar 22, of the observed input vector of the period 2
observation required to be on the frontier function of the reference period 1 with
observed outputs. The denominator shows correspondingly the adjustment by i of
the observed input vector of period 1 for the observation to be on the same period 1
frontier function. Note that the measure with different time reference for year of
observation and reference technology now may be greater than one, if the observation
is not feasible within the technology in question. In fact, the measure itself may be
infeasible to calculate. If M1j 1•2 > (<) 1, then the observation in period 2 is more (less)
productive than the observation in period 1.
An output-oriented Malmquist index can be defined in a similar way. Under the CRS
assumption it would be equal in value to the input-oriented index, and the efficiency
282 Efficiency and Productivity of Norwegian Colleges
measures will always be feasible in principle. If we want to interpret the index as a
total factor productivity index we must base the efficiency measures on comparing
the observations with the corresponding optimal scale points, i.e. we must use the
measure we have termed E3; the technical productivity measure. In practical
applications this is as if we use a CRS reference technology enveloping the actual VRS
technology (enveloping the VRS technology with the smallest cone) (see F0rsund
(1997) for further discussion).
In the presence of inefficient observations change in productivity is the combined
effect of change in efficiency and shift in the frontier production function6. Fare et al.
(1994b) 7 showed how the CCD index in the case of inefficient observations could be
decomposed when there are two time periods and one of them is used as reference
technology. The Malmquist productivity index, ~1 u, can be multiplicatively
decomposed into two parts showing the catching up, MCj/. 2 u, and the technology
shift, MFJ,f
£2 £2 £2 M 1,2 - ~- __.!l:_' __}_}_- MC1,2 . MF2
}1 - £1 - £1 £2 - }1,2 }1,2 }1 }1 }2
1,2 E T (7)
The catching-up effect, MC11,2 '· 2 , expresses the relative movement of the observed
unit to the frontier, a higher (lower) "contemporary" efficiency score for the second
period implying increased (decreased) efficiency. The frontier technology change is
expressed by the ratio of the efficiency scores for the same second period observation
relative to the two technologies. The numerator expresses the scaling of period 2
inputs in order to be on period I technology, while the denominator expresses the
scaling of the same input vector in order to be on period 2 technology, in both cases
subject to period 2 observed outputs. This then serves as a measure of technology shift,
and is greater than one if period 2 technology is more efficient relative to period I
technology for the input-output mix of the period 2 observation.
6 See e.g. Nishimizu and Page (1982) for such a decomposition in the parametric frontier case.
7 Originally circulated as a working paper in 1989.
Efficiency and Productivity of Norwegian Colleges 283
If another year than the two observation years is chosen as basis for reference
technology, the expression for frontier shift is slightly complicated by imposing
chaining (see Berget a!., 1992):
£2 £2 MF I,2 = )I ;2
}I 2 I I , i,1,2 E T , j E N ' E E ji jl
(8)
The chained frontier technology change is a relative change between period i
technology and period 2 technology in the numerator, and period i technology and
period 1 technology in the denominator.
4 Data
We shall concern ourselves here with data for the 1994, 1995 and 1996 academic years
(the latest year for which data are available at the time of writing). In addition to
public data (NSD, 1997), the data used in the present paper was collected by the
Foundation for Research in Economics and Business Administration (SNF) (now the
Frisch Centre), at college department level. We sent out questionnaires to the 26
regional colleges comprising 109 departments and received data from 23 of them
comprising 99 departments. Unfortunately the project had a very tight time schedule,
so the quality of the data may be negatively influenced by this. In order to secure
quality we followed up the questionnaire by telephone contact and gave all colleges
the opportunity to see our first version of the data for themselves and communicate
any corrections. Although there was some problems with interpretations of our
variable definitions and the tight time schedule, in our opinion the data are of
sufficient quality to express reliable structural features and trends in the regional
college sector.
4.1 Output measures
As proxy-variables for research output (R&D) we asked for information according to
the following typology:
284 Efficiency and Productivity of Norwegian Colleges
(i) Papers in professional journals
(ii) Papers in academic journals
(iii) Authored books
( iv) Contributions to edited books.
Due to lack of information, some departments could not split up their research
production into these four categories. We have therefore summed (no weighting) these
four categories into one category called R&D.
As output-measure for person-specific increased knowledge we have used total
number of exam credits8 (product of candidates and exam credits). We have split this
measure into two categories due to typical difference in resource usage between short
and long education:
(i) Short education: Studies that is stipulated from 6 months up to 2 years,
plus one year extension course.
(ii) Long education: Studies that are stipulated for 3 years or more.
4.2 Input measures
Four input-measures are used in the analysis:
(i) Faculty staff: Number of faculty staff man-labour year
(ii) Administrative staff: Number of administrative staff man-labour year
(iii) Net operating expenses: operating expenses minus wage costs
(iv) Building capital: Number of square meters.
Two measures of staff man-labour year are used in the analysis. One for the staff with
a solely research and teaching functions and one for staff with only administrative
functions.
8 By stipulation full time students will obtain 20 exam credits per year.
Efficiency and Productivity of Norwegian Colleges 285
In the analysis we use net operating expenses (operating expenses subtracted by wage
costs) for other inputs, such as cleaning, heating, stationary, telephones and postage,
maintenance9.
We have included size of building in the analysis. It is not obvious that this is an
interesting input in our context. Of course, some minimum space is needed for the
educational process and research, but above that it is difficult to argue that more space
promotes the production of our outputs. Effects of space like it is more expensive to
clean rooms in large buildings, and higher costs associated with central heating the
more space, would be captured by operating expenses. Table 3 summarises the
variables used in the analysis.
Table 3: Variables and definitions in DEA model
(v) Variables
Output measures:
Shortedu
Longedu
R&D
Input measures:
F acuity staff
Admmistrative staff
Net operating expenses
Size of building:
Definitions
Studies that are stipulated from 6 months up to 2 years,
Plus one year extension courses
Studies that are stipulated for 3 years or more
Research publications
Number of faculty staff man-labour year
Number of administrative staff man-labour year
Operating expenses minus wage costs
Number of square meter building
9 Smce Buildings are represented by area rent should have been taken out of the expenditure figures. It could also be argued that maintenance should be taken out, since it could be used as a proxy for buildings, see Ahn et al. (1989).
286 Efficiency and Productivity of Norwegian Colleges
4.3 Structure of data
The structure of data can graphically be illustrated by joining variables in pairs as
shown in Figure 2. Each histogram represents one department, and the width of the
histogram represents the relative size of the department measured by the number of
full time student equivalents. By ranking the departments in increasing order by the
ratios we obtain information about the total variation in the distribution, the shape of
the distribution, and the localisation of large and small units. The extent of the outlier
problem will be revealed, and data to be double-checked are pinpointed. With totally
seven variables, there are a lot of possible combinations to be shown. We have focused
on six. We have done the calculations for all three years, but will only show the
structure for 1996.
The three first distributions, Panels a, b and c, shows the ratio between the three
products short education, long education, R&D and the input faculty staff
. ..__ ...................... 0 U U U ~ ~ U U ~ ~ I
Panel a: Long education/faculty staff
Panel c: R&D/faculty staff
Panel b: Short education/faculty staff
11000
eooo 1000
i eooo ,,.., 'g <OOO
f:rm ;roo
·ooo ~"'""" ... :Qillli):ID(mmUIEWlJJJ
Panel d: Expendituresllong education
Efficiency and Productivity of Norwegian Colleges
IIIXXXl
IIIXXXl
'"""" i,lOOX> I,"""" jiiXXXl ~SliD . """"
Panel e: Expenditure/shan
education
287
0 ~ ~ ~ ~ ~ ~ ~ ~ ~ I fWettv....O.II~rilt-..df.,._.
Panel f: Administrative stafl7faculry staff
Figure 2:
equivalents.
Salter diagrams 1996. Relative size measured by full time student
Exam credits for the product long education per faculty staff varies gradually from 16
to 314. A little tail of departments representing about 9% of the population of students,
has no long education at all. There is a tendency that medium-sized departments
dominate the most "productive" part of the distribution, but with exceptions.
Exam credits for the product short education per faculty staff (Panel b) varies from 4
to 538 with a median of 35. Also for short education we have departments
representing about 9% of the population of students, with zero output. The distribution
has a different shape with a large share of departments having modest productivity.
Middle-sized departments dominate the most productive part of the distribution, which
has a more marked "best practice" tail than for long education. Small and large
departments dominate the part of the distribution with lowest productivity.
The distribution for R&D per faculty staff (Panel c) is somewhat skewed like the one
for Short education. There are 12 departments that representing about 7% of the total
number of students, with no R&D production at all. On the other side of the
distribution a group of departments that represents 7% of the students has extreme
high R&D production. These units are smaller than the average measured by relative
288 Efficiency and Productivity of Norwegian Colleges
student population. One of the two most extreme departments is very small and has
over two R&D contributions per faculty staff. Generally we observe that small
departments have R&D productions characterised by larger variance than larger
departments.
In Panel d we have the ratio between operating expenses and exam credits for the
product long education. The distribution shows large variation, from 99 to 8410 NOK
per exam credits. The median is about 522. One department is extreme within a tail
representing about 5% of the students.
In Panel e we have the ratio between operating expenses and exam credits for the
product short education. The distribution shows large variation, from 63 to 159 275.
The median is 1067.We recognise the same extreme department having almost 20
times as high ratio than the median. The distribution is visually dominated by this
observation. Double-checking revealed that the department had had extremely low
number of exams of both types that year.
In Panel f we look at the ratio between the inputs administrative staff and faculty staff
We find a smooth distribution with no extreme outliers, but the most extreme
department has a somewhat higher ratio than the next one. We would expect to see a
mix of economies of scale and professional departments needing more technical
laboratory or equipment staff classified as administrative (not teaching). There is a
relatively even mix of small and large departments in the distribution, but the lowest
ratios are dominated by small departments, indicating diseconomies of scale, while
around the median value medium-sized units dominate. Some large departments have
relatively high ratios. These are professional departments and the technical staff effect
could dominate. But it should be remembered that the distributions are all partial and
that the simultaneous approach below is needed for a proper look into issues like
economies of scale.
Effictency and Productivity of Norwegian Colleges 289
5 Efficiency results
5.1 Efficiency distributions
The technical efficiency of a college reflects the potential for increasing the college
output without increasing the use of resources (output efficiency) or the potential for
reducing the use of resources without reducing the school output (input efficiency).
The analysis makes use of the input efficiency definition. This is due to the fact that
student capacities are regarded as exogenous in the short run. We allow for variable
returns to scale, which means we believe size of college is of importance calculating
the efficiency scores.
The technical measure and the scale measure for 1996 are presented in Figure 3,
Panels a and b. Along the horizontal axis we have all the 99 departments. Each
histogram represents a department and the width is the ratio between student mass at a
department related to the total student population in the college sector. Efficiency is
measured along the vertical axis. The departments are ranked according to increasing
efficiency score.
The distribution for the input saving technical efficiency measure shows that 47
departments of99 are technically efficient (score equal to 1), and these best practice
0 Ot 0.2 Ol 0.4 O.S 06 07 0.8 Ot 1 Relative department size
Panel a: Input-saving efficiency
1996
•.. o,e 0,1
0,0
:a o.• o.• 0.3
0.2
0,1
. ~~~~~~~~~
o 0.1 0.2 o.3 o • o.s o.& 0.1 o._a o.e ' Relative department size
Panel b: Technical productivity
measure 1996
290 Efficiency and Productivity of Norwegian Colleges
Eflkiency score
"/ ,..:r-. ,.,, ....
Rela1ive dcpartmenl size
Panel c: Efficiency distributions. Input-saving measure, 1994, 1995, 1996
Figure 3: Efficiency measure distributions. Relative size measured by
full time student equivalents 1996.
(BP) departments have a share of students at about 55% . Worst practice departments
(WP) have a share of students at about I 0% when WP is defined efficiency scores
lower than 60% (or 0,6). From the figure we see that the BP units mainly consists of
small and big departments, while WP mainly consists of medium-sized departments.
Panel b shows the distribution for the technical productivity measure. Of 99
departments 31 are scale efficient, and the optimal scale departments have a student
mass at about 33%. The scale efficient units consist mainly of small and medium-sized
departments. WP productivity departments have a student mass at about 15%
efficiency, when WP is defined as efficiency lower than 60% (or 0.6). WP mainly
consists of small and medium-sized departments, but the extreme worst tail consist of
small ones.
Panel c shows the shift of the distribution for input-saving efficiency over the years
1994, 1995 and 1996. The tops of the histogram distributions like in Panel a are
exhibited as step curves. We see that the shape and location of the distributions for
1994 and 1996 are quite similar (but note that movements of individual departments
cannot be seen), and that the distribution for 1995 shows somewhat higher
Effictency and Productivity of Norwegian Colleges 291
inefficiencies that year. The share of students at efficient departments is remarkably
stable10.
5.2 The Peer index
Panel c of Figure 3 shows us that the share of students at efficient units is relatively
high at about a level of 50% for all years for input-saving efficiency. These units are
the peers that inefficient units may study in order to improve their performance. The
efficient units cannot be further ranked as to efficiency score. This has been pointed
out as a problem in the literature, and ways of ranking them have been introduced (see
Andersen and Petersen, 1993). We will here prefer to show an alternative ranking
introduced in Torgersen eta!. (1996). For each efficient unit we have in Figure 4
calculated the share of total potential input saving as to faculty staff that is represented
by the inefficient units that have the efficient unit in question as a peer. We know that
in general there may be several peers for an inefficient unit (in Figure 1 units A and B
are peers for unit P1). The potential input saving is therefore weighted by the weight
of the peer in the calculation of the frontier reference point (the 8n in Eq. (5)). The
peer index is input (or output) specific. We are only showing the index for faculty staff
for the input-saving measure for the three years, identifying the ten most important
peers.
Panel a
10 Note that this approach is different from "window analysis" (Ahn et al. , 1989), where different cross section sets are created by dropping and adding years.
292 Efficiency and Productivity of Norwegian Colleges
Panel b
....
Panel c ( 1996)
Figure 4: Peer index for faculty staff (input saving efficiency). Ten most important peers
5.3 Stability
A very important opportunity provided by times-series cross section data is to check
on the stability of best practice units. If the turnover is very high then the yearly
efficiency results are driven by time-specific conditions and it is difficult to learn from
the exercise as to policy implications. If the set of best practice departments is fairly
stable, then one has a much more reliable basis as to required policy actions in order
to improve efficiency.
Efficiency and Productivity of Norwegian Colleges 293
The VRS model yields a fairly high proportion of best practice departments for all
years, 52% for 1994, 45% for 1995, and 43% for 1996. Such relations caution us to
look for self-evaluators. There are 11 in 1994, 9 in 1995 and 10 in 1996, or a little in
excess of 1/5 of the best practice departments each year. Of the best practice units in
1994, about 2/3 are also best practice ones in 1995, and of the remaining efficient
ones a little less than 2/3 remains efficient also in 1996. Of the efficient units i 1995
above 2/3 remain efficient in 1996. The set of units remaining efficient in all years
represents somewhat above Yz of the best practice units each year, or varying from 27
to 24% of the total number of departments. In this set no unit is a self-evaluator in all
years, and only two are for two years, while the percentage of self-evaluators varies
from 1/5 to Jess than 1/10 for each year.
Another way of looking at stability is to inspect the group of most influential best
practice departments. The Peer index for each year in Figure 4 shows us the most
influential peers. Choosing the faculty-oriented index, we have that of the 10 most
influential peers each year, 6, 7, and 4 of the units in the years 1994, 1995, and 1996
respectively remaining efficient all the years belong to the 10 most influential. Of
these, two units, no. I and 67, remain in the top-ten set all years, while five units are in
the top-ten set two of the years. Although not based on any formal test, we conclude
that there is enough stability in our results to claim that the study has revealed some
structural features worth while pursuing for policy purposes.
6 The productivity development
6.1 The Malmquist productivity index
The strength of our approach to calculate productivity growth is that we get the
development for each unit. As a background for a discussion of distributions of
productivity change it may be useful to inspect the average changes of the variables,
set out in Table 4.
294 Efficiency and Productivity of Norwegian Colleges
Table 4: Percentage change in variables
Variable (95-94)/94% (96-95)/95% (96-94)/94%
Short edu 25.5 0.1 25.6
Long edu 3.4 15.9 23.0
R%D 16.0 10.8 28.6
Faculty 0.4 1.9 2.2
Adm. Staff -0.5 2.9 0.0
Expenditure -19.4 -26.2 -36.1
m2 -1.9 0.3 -1.6
Regarding the three outputs we see a strong average growth in short education in the
first period and a moderate increase in long education, while short education is at a
standstill in the second period while long education has strong growth. A strong
substitution is indicated. Research and development has a high growth in both periods.
As to the inputs all except expenditures (net of wages) are more or less at a standstill.
The expenditures decrease quite strongly. This average development points to
productivity increase on the average driven by output growth and expenditure
decrease. The individual variability was demonstrated in Section 4. The variability in
the outputs short education and research and development, and in the input
expenditures, is much stronger than in the other variables.
Figure 5 shows productivity distributions for pairs of years (1994-95, 1995-96 and
1994-96) in Panels a-c. The frontier for the starting year 1994 is used as reference
technology. Since we are assuming VRS-technology, the Malmquist-index is based on
the technical productivity measures. The productivity index is calibrated such that
productivity estimates lower than I are indicating decreased productivity, and larger
than one increased productivity. If a unit obtains 1.10 this shall be interpreted as a I 0%
productivity growth. The width of the histogram is still proportional to the relative size
measured by the number of full time student equivalents.
Efficiency and Productivity of Norwegian Colleges
'" ! ~ ... E
~
3.5
2,5
1.5
o.5
3.5
if 3
~ 2.5
i 2
~ 1.5
~ 1
0,5
0.1 0.2 o.3 o.• o.5 o.e o.1 o e o.9
RtlaUvt department size
Panel a: Productivity growth 1994-95
0,1 0,2 0 ,3 0,4 0,5 0.6 o. 7 0.6 0.9
Rel•tlvt department aln
Panel b: Productivity growth 1995-96
0.1 0.2 0.3 0,4 0.5 0,6 0, 7 0.6 0.9
Relallvt department t in
Panel c: Productivity growth 1994-96
Figure 5: The Malmquist productivity index
295
296 Efficiency and Productivity of Norwegian Colleges
Panel a shows the productivity growth in 1994-95. Departments with decreased
productivity represent about 35% of the student mass (in 1996), and departments with
increased productivity growth represent 40%. These variations are large taking into
account the short period, and as expected from the average changes set out in Table 4
and individual variability illustrated in Figure 2. There is a group of departments with
almost no productivity growth covering about 25% of the students. We have a mix of
medium-sized and small departments here. Small and medium sized departments also
dominates the WP group with decreased productivity, and then some large
departments, while medium-sized departments dominates the top group with
productivity growth.
Panel b shows productivity growth distribution for the period 1995-96. Departments
with positive productivity growth represent about 55% of the students (in 1996). In
contrast to Panel a, there is no longer a group of departments with constant
productivity. Large and small departments dominates the group with productivity
decline, while small and medium-sized departments dominates the group with
productivity growth, the latter again in the maximum growth group.
In Panel c we show the productivity growth for the whole period 1994-96. Since we
are applying an index that is chained, productivity growth is simply the multiplication
of the two corresponding numbers for a unit in Panels a and b. Therefore it is not
surprising that we observe different trends regarding which type of departments having
productivity growth. The share of departments with positive productivity growth
increases further, with over 70% of the students at departments with productivity
growth. For 1994-96 we observe no clear pattern indicating whether there are small or
big departments dominating the group with productivity growth, but the positive
productivity growth part of the distribution starts with large and medium-sized units
dominating, then a part with small units, and lastly some large and small units at the
top end. Note that the numbers are rather large for such a short period. The three large
units in the top group in Panel c have an growth in productivity of about 150 % , while
the small best practice outlier has a growth of almost 300%. But the significant
changes in average values revealed in Table 4, and the large individual variation
illustrated in Section 4 support the reliability of the results.
Efficiency and Productivity of Norwegian Colleges 297
6.2 Decomposition of the Malmquist-index
In Figure 6 we have decomposed the productivity growth from 1995 until 1996 into a
part called "frontier shift" (Panel a) and a part called "catching up" (Panel b), in
accordance with Equations (7) and (8). From Panel a we can see that most of the
departments have gained from a positive shift in the frontier transformation function.
About 67% of the departments (relative size measured by full time student equivalents
in 1996) have benefited from the frontier function shift. The large units have the most
modest impact, while the top group consists of small departments. As to decline
through frontier shift all the groups are represented, with medium-sized departments
dominating the group with most modest impacts, and then large departments. There is
a little tail (about 4% of the students) with a marked contribution in decrease in
productivity from frontier shift.
The "catching up" effect (Panel b) shows large variations, especially at the upper end
of positive productivity growth contribution. The departments that are catching up the
best practice departments represents about 45 % of the students. Large and medium
sized departments, and some small ones, dominate the group with productivity growth,
with the latter group clearly dominating the top part. A share at about 20% shows now
change. This group consists of the departments that are on the frontier both in 1995
and 1996. The units with reduced productivity growth represent about 35% of the
students. Some large departments belong to this group with productivity decline
contributed by catching-up, except from the worst practice group where small
departments dominate (worst practice defined as productivity growth lower than 0.6,
i.e. productivity decline of 40%). Summing up, it seems that positive frontier shift is
most important for small units, negative shifts most important for large units, while
positive catching-up is most important for medium-sized units, and negative catching
up for both small and large units.
6.3 Characterisation of productivity change
It is interesting to examine to what extent changes in the variables from 1995 to 1996
effects the estimated Malmquist index. The classic hypothesis of Verdoom (see
F 0rsund, 1996b) is that there must be growth in output in order to realise productivity
growth. In the spirit of Verdoom we want to investigate the average relationship
between productivity and changes in all the variables. We have made a regression
298 Efficiency and Productivity of Norwegian Colleges
analysis were the regressors represent the percentage change in the variables from
1995 to 1996. The dependent variable is the estimated Malmquist index from in the
relevant period.
1,8
1,6
;;; 1.4 "1 "' 1, 2 ;
1 = -- ------- ---- -=---...... - .......-n1'1'"f'Fl!!lT-JIII':nlllriiU ~Ul :;:
O.B w ~ ~ 0,6
£ 0,4
0,2
0 0 0,1 0,2 0.3 0.4 0,5 0,6 o. 7 0.6 0.9
Panel a: Malmquist decomposition, frontier shift
2.5
g. 1,5
!!' £ 1 ··-···------------- - -
""' nil O~LU~JJ~W-~~~
0 0,1 0,2 0.3 0.4 0,5 0,6 0.7 0,8 0,9
Pan el b : Malmquist decomposition, catching up
Figure 6: Decomposition of the Malmquist productivity index for the period 1995- 1996. Relative size measured by full time student equivalents.
Efficiency and Productivity of Norwegian Colleges 299
Table 5: Drivers for the Malmquist productivity index 1995-96. % change in the DEA variables as explanatory variables
Variable: Estimate St.dev. t-value p-value
Shortedu 0.002 0.002 0.96 0.34
Longedu 0.012 0.001 8.54 0.00
R&D 0.349 0.363 0.96 0.34
Faculty staff -1.362 0.813 -1.67 0.10
Adm. staff 0.099 1.614 0.06 0.95
Net oper. exp. 0.000 0.000 -1.06 0.29
Buildmg (m2) 0.031 0.052 0.60 0.55
Total number of observations: 89, R-squared: 0,449, F-value: II ,247
The results from the estimation process are presented in Table 5. In general one would
expect positive signs for output growth, and negative for input growth. But we observe
that there are only two variables having a significant effect on the estimated
Malmquist index choosing a 10% rejection level. These are long education and
faculty staff As expected there is a positive correlation between growth in long
education and the Malmquist index and a negative correlation for faculty staff It is
surprising that changes in operating expenditures are not significant, but this illustrates
the great variance of this variable. The picture above is also relevant for the period
1994 - 1995, and also looking at the decomposition of the Malmquist index into
"catching up". For "frontier shift" it is interesting to note that there are no significant
correlations. We therefore conclude that especially changes in the long education
product and also the faculty staff input are the main drivers behind average
productivity growth.
6.4 Anatomy of productivity change
The development over time for each department that lies behind the average relations
analysed in Table 5, can be illustrated graphically following the classification in Table
6. In Quadrant I we have departments that obtain both positive productivity growth
and positive output growth. These departments have an efficient expansion because
output is growing faster than inputs. In Quadrant II we have departments that
300 Efficiency and Productivity of Norwegian Colleges
Table 6: Characterization of change
II I
Positive adjustment capability Efficient expansion
Positive productivity growth Positive productivity growth Negative output growth Positive output growth
M > 1, output growth in % < 0 M > 1, output growth in % > 0 "Lean and Hungry" "Top Dog"
Negative adjustment capability Inefficient expansion
Negative productivity growth Negative productivity growth Negative output growth Positive output growth
M < 1, output growth in % < 0 M < 1, output growth in % > 0 "Fat Cat" "Puppy Dog"
III IV
combines positive productivity growth with negative output growth. This is only
possible if inputs are reduced more than outputs. These departments have positive
adjustment capability. In Quadrant III we have departments that obtains a decrease in
both productivity growth and output growth. These departments also have less
reductions in inputs than in outputs, i.e. negative adjustment capability, because the
reductions in inputs are not sufficient to obtain positive productivity growth. In
quadrant IV we have departments that combines negative productivity growth with
positive output growth. These have inefficient expansion because inputs are increasing
more than outputs.
In Figure 7, Panel a we have shown the distribution on the four quadrants when
productivity is linked with growth in R&D. Each square represents one department and
the size of the square is proportional to the number of full time student equivalents in
1996. We can see that the departments are distributed on all quadrants. We observe
units with both negative, zero and high R&D growth, the range is from 100% decline
to 400% increase. (Units going from zero to a positive number have been excluded,
Efficiency and Productivity of Norwegian Colleges
2_4
o24 III
100 0
M(95-96)
2.4
II
0.24
0 IV
400 R&D% Panel a: Research
D E C0 0
0 0 0
0
IV
250 Longedu%
Panel b: Long education
301
302 Efficiency and Productivity of Norwegian Colleges
2 ,4
0
IV 0 .2•
-100 0
Panel c: Short education
Figure 7: Productivity- and output growth
and units going from a positive number to zero have been given the figure 1 00). Some
units have remarkable high productivity growth and reduction in R&D growth. This
may indicate a substitution effect towards more teaching, meaning an increase in the
number of grade points. But we should have in mind the possibility of lag effects
between faculty input and R&D. One cannot expect a stable relationship year by year.
Panel b shows the distribution when we focus on the long education product. We
observe a longitudinal pattern; growth in grade points is the main driver behind
productivity growth. This is accordance to the average structure revealed in Table 5.
There are relative few units in quadrants II and IV. The majority of departments
experience an increase in long education, but there are also a number of departments
with negative adjustment capability.
In Panel c we are comparing the productivity growth with growth in the short
education product. We no longer find the longitudinal pattern as in Figure 6, in
accordance with the insignificant coefficient in Table 5. Departments are spread over
all quadrants. The average growth in short education is about zero, and it is noticeable
that many departments show positive adjustment capability. There may be a
substitution effect here: the departments with positive adjustment capability have
Efficiency and Productivity of Norwegian Colleges 303
managed to increase long education sufficiently to achieve positive productivity
growth.
7 Conclusions and further research
In view of the variables we have had to use in the study and the ideal variables set out
in Section 2, it is obvious that the study is far from perfect. However, in order to
generate sufficient interest in engaging in the hard work at the institutional level of
collecting new types of data we believe the study has been worth while. The proxies
used for the three outputs where final exams distributed on two types; short- and long
studies, and research publications. The four inputs where number of academic and
non-academic staff in full time equivalents, current expenses other than salaries, and
building size in square metres.
Typical cross section efficiency results show a large share of efficient departments,
with a disproportionate number of efficient departments giving theoretical general
education, and a large variation within the group of inefficient units. The difference
between professional and arts and science departments may be explained by the nature
of the teaching production function, but calculations for a sub-sample of professional
departments (e.g. nurses, engineers, teachers) show almost the same variation within
this group. The productivity change was mainly positive, with most departments
experiencing a positive productivity effect from frontier shift, but a greater variation
from positive to negative as regards the contribution from catching up. Positive
frontier shift is most important for small units, negative shifts most important for large
units, while positive catching-up is most important for medium-sized units, and
negative catching-up for both small and large units.
Although some doubt has been voiced as to the legitimacy of the present study
representing "true" efficiency, at least the structural differences between departments
as to efficiency and productivity warrant further research.
There are several ways of improving upon the analysis:
304 Efficiency and Productivity of Nmwegian Colleges
7.1 Stage two- analysis
In order to address the question of why units differ in efficiency a second set of
explanatory variables may be introduced (see e.g. Rhodes and Southwick, 1993). The
stage two analysis tries to capture other variables that may affect the efficiency scores.
In order for the procedure to be statistically sound, the new set of explanatory
variables must be uncorrelated with the variables used in the first stage. It is usual to
focus on non-discretionary variables outside the control of the units. We have tested
the quality of staff by position, using as dependent variable the technical input-saving
efficiency score obtained by DEA. It had a (weakly) significant effect on efficiency
scores for two of the years. Number of individual studies offered by a department was
not significant any year, but here we have a covariation problem with inputs used in
the first stage. Other variables could be the location of the college (urban - rural, co
location with other institutions of higher learning), concentration or not of campus
(spread out on different locations or in one location).
7.2 Separating professional and arts and science departments
It may be legitimate to question the assumption of the same technology for all types of
departments. We have experimented with a subgroup of departments giving only
professional education, since the lion's share of efficient departments are arts and
science, and the underlying technology characterising professional education, like
small student groups, need for laboratories, practice outside the college, etc. may well
indicate different technologies. It turned out that the difference in efficiency scores and
the shape of the distribution was very much alike the one for the total sample for 1996.
Further investigations as to teaching technology is warranted. Are small teaching
groups necessary, or just tradition, etc.
7.3 Quality variables
There is an obvious need for variables capturing quality aspects, as discussed in
Section 2. There is also room for improvement of the variables used. The research
output can be more elaborately designed by weighting, and research for departments
Efficiency and Productivity of Norwegian Colleges 305
like Music and Media must be introduced. Only written reports have been used in this
study.
References
Afriat, S. (1972): Efficiency estimation of production functions, International Economic Review
13(3), 568-598.
Ahn, T., V. Arnold, A. Chames and W. W. Cooper (1989): DEA and ratio efficiency analyses for
public institutions of higher learning in Texas, Research in Governmental and Nonprofit
Accounting, 5, 165-185.
Andersen, P. and N.C. Petersen (1993): A procedure for ranking efficient units in Data Envelopment
Analysis, Management Science, 39, 1261-1264.
Banker, R. D., Charnes, A. and W. W. Cooper (1984): Some models for estimating technical and scale
inefficiencies, Management Science 39, 1261-1264.
Beasley, J. E. (1990): Comparing university departments, OMEGA Int. Journal of Management Sci.
18(2), 171-183.
BergS. A., F. R. flll'sund and E. Jansen (1992): Malmquist indices of productivity growth during the
deregulation of Norwegian banking, 1980- 1989, Scandinavian Journal of Economics, 94,211-
228.
Busch, T., L. Fallan and A. Pettersen (1997): Diciplinary differences in job satisfaction, self-efficacy,
goal commitment, and organizational commitment among faculty employees in Norwegian
colleges: an empirical assessment of indicators of performance, Rapport i T0H-serien 1997:2,
Avdeling for 0konomisk-Administrativ utdanning, H0gskolen i Sill'-Tmndelag.
Caves, D. W., L. R. Christensen and W. E. Diewert (1982): The economic theory of index numbers
and the measurement of input, output and productivity, Econometrica, 50, 1393-1414.
Charnes, A., W. W. Cooper and E. Rhodes (1978): Measuring the efficiency of decision-making units,
European Journal of Operational Research 2, 429-444.
306 Efficiency and Productivity of Norwegian Colleges
Doyle, J. R. and A. J. Arthurs (1995): Judging the quality of research in business schools: the UK case
study, Omega 23, 257-270.
Erlandsen, E., F. R. F01'sund og K. 0. Kalhagen (1998): Effektivitet og produktivitet i de statlige
heyskoler [Efficiency and productivity in the public colleges], SNF- rapport 14/98, Oslo.
Farrell, M. (1957): The measurement of productive efficiency, Journal of the Royal Statistical Society,
Series A (General), 120 (III), 253-281 (290)
Flemming, J. (1991): The use of assessments of British university teaching, and especially research,
for the allocation of resources- A personal view, European Economic Review 35,612-618.
Frisch, R. (1965): Theory of production, Dordrecht: D. Reidel.
Fare, R. and C. A. K. Lovell (1978): Measuring the technial efficiency of production, Journal of
Economic Theory 19, 150-162.
Fare, R., S. Grosskopf and C. A. K. Lovell (1985): The measurement of efficiency of production,
Boston: Kluwer- Nijhoff.
Fare, R., S. Grosskopf and C. A. K. Lovell (1994a): Production frontiers, Cambridge: Cambridge
University Press.
Fare, R., S. Grosskopf, B. Lindgren and P. Roos (1994b): Productivity developments in Swedish
hospitals: a Malmquist output index approach, in Chames, A., W. W. Cooper, A.Y. Lewin and
L. M. Seiford (eds.), Data Envelopment Analysis: theory, methodology, and applications,
Boston!Dordrecht/London: Kluwer Academic Publishers, 253-272.
Fersund, F. R. (1996a): On the calculation of the scale elasticity in DEA models, Journal of
Productivity Analysis, 7(2/3), 283-302, 1996.
Fersund, F. R. (1996b): Productivity of Norwegian establishments: a Malmquist index approach, in
D. G. Mayes (ed.) : Sources of productivity growth, Cambridge: Cambridge University Press,
315-330.
Fersund, F. R. (1997): The Malmquist productivity index, TFP and scale, Memorandum no. 233, Dept.
of Economics, School of Economics and Commercial Law, Gllteborg University.
Efficiency and Productivity of Norwegian Colleges 307
F0rsund, F. R. and L. Hjalmarsson (1974): On the measurement of productive efficiency, Swedish
Journal of Economics 76 (2), 141-154.
F0rsund, F. R. and L. Hjalmarsson (1979): Generalised Farrell measures of efficiency: an application
to milk processing in Swedish dairy plants, Economic Journal 89,294-315.
Higgins, J. C. (1989): Performance measurement in universities, European Journal of Operational
Research 38, 358-368.
Johnes, G. ( 1990): Measures of research output: university departments of economics in the UK,
1983-88, The Economic Journal, 100, 556-560.
Johnes, G. (1997): Costs and industrial structure in contemporary British higher education, The
Economic Journal, 107, 727-737.
Johnes, J. and G. Johnes (1993): Measuring the research performance of UK economics departments:
an application of Data Envelopment Analysis, Oxford Economic Papers, 45, 332-347.
Johnes, J. and G. Johnes (1995): Research funding and performance in U.K. university departments of
economics: a frontier analysis, Economics of Education Review, 14, 3, 301-314.
Johnes, J., J. Taylor and G. Ferguson (1993): The employability of new graduates: a study of
differences between UK universities, Applied Economics 19,695-710.
Johnes, J., J. Taylor and B. Francis (1993): The research performance of UK universities: a statistical
Aanalysis of the results of the 1989 Research Selectivity Exercise, J.R. Statistical Society, 156,
part 2, 271-286.
Kalhagen, K. 0. (1998): En analyse av teknisk effektivitet og produktivitet i h0gskolesektoren baser!
pa Data Envelopment Analysis [An analysis of technical efficiency and productivity based on
Data Envelopment Analysis], SNF-arbeidnotat nr. 38/98, Oslo.
MalmqUist, S. (1953): Index numbers and indifference surfaces, Trabajos de Estadistica, 4, 209-242.
Nishimtzu, M. and J. M. Page (1982): Total factor productivity growth, technological progress and
technical efficiency change: dimensions of productivity change in Yugoslavia 1965-78,
Economic Journal 92, 920-936.
308 Efficiency and Productivity of Norwegian Colleges
Norsk Samfunnsvitenskaplige Datatjeneste (1997): Statistikk om h0gre utdanning 1997. 0konomi,
Studenter, Ansatte. NSD-publikasjon av desember 1997.
Rhodes, E. L. and L. Southwick (1993): Variations in public and private university efficiency,
Applications of Management Science, Public Policy Applications of Management Science 7,
145-170.
Sarafoglou, N and Haynes, K. E. (1996): University production in Sweden: a demonstration and
explanatory analysis for economics and business programs, The Annals of Regional Science,
30, 285-304.
Shephard, R. W. (1953): Cost and production functions, Princeton: Princeton University Press.
Sinuany-Stem, Z., A. Mehrez and A. Barboy (1994): Academic departments efficiency via DEA,
Computers Ops Res. 21, 543-556.
Tomkins, C. and Green, R. (1988): An experiment in the use of Data Envelopement Analysis for
evaluating the efficiency of UK university departments of accounting, Financial Accountability
& Management 4, 147-164.
Torgersen, A.M., F. R. Fmsund and S. A. C. Kittelsen (1996): Slack-adjusted efficiency measures and
ranking of efficient units, Journal of Productivity Analysis, 7, 379-398.
Efficiency and financial performances in
telecommunications
P.-Y. Badillo'
Abstract
Since the break-up of AT & Tin 1984, the telecommunications sector has been in a process of
quick and very important changes. In this paper the evolution of the performances in the
telecommunications sector are analysed by combining two approaches : the DEA (Data
Envelopment Analysis) method and the use of financial indicators. Various DEA analysis,
among which some « window analysis », are applied and throw light and shade on different
facets of the performances of the operators at the international level. Then a brief financial
analysis gives some indications about the financial positioning and dynamics of the
operators. The first main conclusion concerns the use of the network : the European
operators, especially the French operator, offer a limited access to the network with high
prices. The evolving situations of the operators are the second issue : in the United States, in
a competitive context the differences between operators have increased and some
restructurations have occurred; in Europe, the less deregulated operators, Deutsche Telecom
and France Telecom, are behind British Telecom and Telecom Italia.
' Professor at the University of Aix-Marseille; Director of the EJCM (Ecole de Journalisme et de Communication de Marseille), Universite de Ia Mediterranee, CEFI (Centre d'Economie et de Finances Internationales), CID (Recherche sur le Calcul economique, !'Innovation et Ia Decision), France
310 Efficiency and financial performances in telecommunications
Structure
Introduction
2 Monopoly, Deregulation and Efficiency
3 Performances of the Telcos: the DEA Method
4 The Financial Analysis
5 Conclusions: Key Factors for Performances
Annex
References
Efficiency and financial performances m telecommunications 311
I Introduction
For the last fifteen years the telecommunications sector has been in a process of quick
and very important changes. In this paper the evolution of the performances in the
telecommunications sector will be analysed combining two approaches : the DEA
(Data Envelopment Analysis) method and the use of financial indicators. The study
relies on a very complete and original database including financial and technical data
since 1986 on telecommunications operators. The database includes the main operators
in the world (United States of America and European operators, about 12 operators)
and, when the statistics are available, for some indicators less important operators.
So we shall combine two approaches on a quite long period (1986-1997) in order to
examine one of the most strategic sectors of the contemporary society. The first results
are presented hereafter and the research is developing. By observing the efficiency
scores and indicators of financial performances for different firms and countries, we
shall be able to compare the different trajectories and conclude on their relative
efficiency, trying to establish a link with the type of (de)regulation. With such
methodological grounds and with a very rich database we shall try to answer two very
important questions :
• is there a same movement from the point of view of efficiency in production activities and from the point of view of a financial analysis ?
• is there a parallel in the evolution of the different firms or on the contrary a
very contrasted scheme ?
In our paper we shall begin with a brief presentation of the stakes and difficulties
behind an analysis of the performances of the telecommunications operators : are
monopolies inefficient and what are the effects of deregulation, especially from the
point of view of efficiency '? (2). We shall quickly expose the main questions before
defining our methodology based on the DEA method and financial analysis. Then
different possible measures of performances in telecommunications will be proposed
and we shall compare the performances of different telecommunications operators
using the DEA method (3). In a fourth part, a brief financial analysis will be made,
which will give some indications about the financial positioning and dynamics of the
operators ( 4). Our results will emphasize the evolution of the performances of the
operators especially for the recent period.
312 Efficiency and financial performances in telecommunications
2 Monopoly, Deregulation and Efficiency
Natural monopoly is a special case from the point of view of economic theory. The
market economy which is omnipresent today has admitted for a long time the existence
of natural monopolies. Until the sixties in highly capital intensive industries -like the
electric power industry, telecommunications, airlines- the regulated monopoly
appeared as the guarantee of economic efficiency -thanks to increasing returns-. Yet as
far back as 1962 two American economists attached their names to what is now called
the Averch-Johnson effect. They showed that the regulated monopoly AT&T suffered
a phenomenon of overinvestment. Being not in a competitive situation AT&T would
have invested beyond the optimum. So the community should have « overpaid » the
telecommunications services. The Averch-Johnson effect was the first line of attack
against the regulated monopoly. On the occasion of the AT&T break up and since then
many new arguments in favour of deregulation have appeared. It is nevertheless
possible to find a common denominator to all these analyses : the regulated monopoly
would be inefficient -overinvestment, bureaucratic management, plethoric manpower
and the inefficiency would appear obvious especially thanks to productivity indicators
or performance analyses. Since about forty years one of the main themes of economic
research on regulation has been the effects of regulatory policies, as noted by R.G.
Noll (1989, p. 1254). Nevertheless studies concerned by the effects of regulation on
productivity have been few. In their article on the effects of economic regulation
published in 1989, P.L. Joskow and N.L. Rose were distressed that« so little effort has
been devoted to measuring the effects of regulation on innovation and productivity
growth ». Since then, regarding telecommunications, the impact of regulation on
productivity or efficiency has been examined in the case of the United States, but less
has been done on other countries and on comparisons between countries.
Thus today productivity and performance analyses are a very important task for firms
as well as for public authorities. Productivity and performances mean efficiency and
competitive advantage, but are to a certain extent an unexplored domain, especially in
the field of an international comparison. Moreover and paradoxically productivity can
be analysed from many different point of views and can give rise to highly shaded or
even opposite measures. The telecommunications sector does not escape this paradox.
Yet on the basis of productivity ratios some economists and decision makers
sometimes recommend massive dismissals measures in order to « improve >>
Effic1ency and financial performances m telecommumcations 313
productivity. Besides, in a more deregulated and competitive context, management and
strategic decisions are often based on some indicators of financial performances.
2.1 The effects of regulation on performances
Previous studies which have covered a vast range of effects of regulation - on prices,
costs of production, investment, service quality, productivity, profits - do not show
undisputable results (for a review of literature see Badillo [ 1999]). In the case of
telecommunications, the break-up of AT&T, the changes in US regulatory rules since
then, and the changes in other countries such as the United Kingdom and France, have
questioned the effects of regulation on productivity and more generally on
performances ; the empirical findings on this issue are various and may differ a lot
from a methodological point of view and on the results. The long debates about natural
monopoly have proved how difficult it is to evaluate, even theoretically, the
advantages and losses associated with different types of market structures and
regulatory constraints (see the abundant theoretical literature; for example A.E. Kahn
(1971], W.J. Baumol, J.C. Panzar and R.D. Willig [1982], R.R. Braeutigam [1989],
J .J. Laffont and J. Tirole [ 1993], J .J. Lafont [ 1994]). The difficulties in estimating a
production function or a cost function have long been the origins of different empirical
findings. In the recent years similar difficulties appear in the discussion about the
comparison between price-cap regulation and rate-of-return regulation (see for
example R.R. Braeutigam and J.C. Panzar [1993]).
In order to evaluate empirically the effects of regulation on productivity and
performances, methodological choices are crucial. P.L. Joskow and N.L. Rose (1989)
point out four basic empirical methodologies for measuring the effects of regulations :
comparing regulated a~d unregulated firms or markets, using variation in the intensity
of regulation, using controlled environment experiments, and structural
estimation/simulation models of regulated firms and markets. The approach by using
controlled environment experiments is certainly promising but is very specific and our
studies have not relied upon it. Structural models of behaviour or performance,
combined with simulation techniques, may be also useful ; they need careful
implementation : for example, the results depend upon the accurate identification and
estimation of demand and cost functions. The first and second basic methodologies
comparing regulated and unregulated firms and using variation in the intensity of
314 Efficiency and fmancial performances in telecommumcations
regulation may be based on cross-sectional or on time-series analyses ; the main
question is the specification of the regulatory regimes. In the first approach the
dependent variable of interest, which may be prices, costs or the rate of technical
change, is defined as a function of exogenous economic characteristics which
influence performance independently of regulation and a variable - generally a dummy
variable - indicating the influence of regulation. In the second approach, variations in
the regulatory constraints over time and space must be estimated. In these two
approaches it is also important to take care of non-regulatory differences between
firms or markets and of the possible interactions between regulatory structures,
economic characteristics of firms or markets and performances. In fact, it is generally
difficult to determine causal relationships between different variables, and more
particularly for telecommunications industry it may be difficult to distinguish the
effects of regulation from the effects of competition. D. Kridel, D. Sappington and
D. Weisman (1996) call this difficulty the "Competition Effect Pitfall" in the
implementation and interpretation of empirical findings in telecommunications ; they
show that the results of econometric models which use proxies for competitive
pressure may be biased and specifically the true impact of a change in regulatory
regime may be over-estimated.
In this paper, our aim is to measure different facets of the performances of
telecommunications operators at the international level. This approach may give some
indications about the link between regulation and performances. In any case, this does
not mean a causality between regulation and performances ; many pitfalls effects could
explain performances of the firms independently of the regime of regulation.
2.2 Productivity, efficiency, performance and financial results
Productivity is an ambiguous concept and we need to define it as precisely as possible.
We can find an early definition in J.B. Say in the "Traite d'Economie Politique"
(1803): "To obtain more product with the same amount of work is the main purpose of
the industry". A common approach is to consider productivity, in the case of a
production unit, as the ratio of its output to his input : according to Z. Griliches ( 1987,
p. I 0 I 0), "productivity is a ratio of some measure of output to some index of input use.
The meaning and quality of such a measure depends on the definition and quality of its
ingredients and on the particular formula and the associated weights used to aggregate
EflicJency and financial performances m telecommunications 315
the various components into one output or input index". So, as far as we have one input
and one output productivity is very easy to evaluate; but generally a unit of production
uses several inputs and produces several outputs. In this case it is necessary to
aggregate inputs and outputs, and the problem arises. Moreover, as C.A. Knox Lovell
(1993, p. 3) points out, "Productivity varies due to differences in production
technology, differences in the efficiency of the production process, and differences in
the environment in which production occurs".
Since we are interested not only by productivity, but by performances, a large
definition of performances might include many aspects : "Economic performance is the
term used to measure how well industries accomplish their economic tasks in society's
interest" (W.K. Viscusi, J.M. Vernon and J.E. Harrington Jr [1995], p. 73). From a
theoretical point of view this refers essentially to efficiency, which may be analyzed in
a static or dynamic way (technical progress), with an approach centered on different
facets of efficiency such as allocative efficiency or/and X-efficiency (H. Leibenstein
[ 1966 ]). From an empirical point of view the emphasis will be put in this paper on
some measures of firms' performances in order to have an approach of productive
efficiency and also of financial results of the firms.
From these general definitions, we can precise our limited goals in the present paper :
in the second section we shall analyse the productive efficiency of the different firms
or countries (the level will change with the availability of the data; we shall called
Telcos the telecommunications companies ; the operators or the countries observed are
described in annex). It is usually recognized that productive efficiency has two
components : technical efficiency which "refers to the ability to avoid waste by
producing as much output as input usage allows, or by using as little input as output
production allows", and allocl!tive efficiency which "refers to the ability to combine
inputs and outputs in optimal proportions in light of prevailing prices" (C.A. Knox
Lovell, op. cit., pp. 9-1 0). Our approach will be based on efficiency through the DEA
method while a specific and complementary financial analysis will be proposed in the
third section. In this paper we do not present technical aspects ofDEA method (see for
example A. Chames, W. W. Cooper, A. Y. Lewin and L. M. Seiford, chapter 1 in P.-Y.
Badillo et J. C. Paradi (Eds.) [1999]) and we only briefly schematize our financial
analysis (see section 3).
316 Efficiency and financial performances in telecommunications
3 Performances of the Telcos: the DEA Method
We shall first define the outputs and inputs that can be used for DEA analysis (§2.1),
before presenting our results. The performances of the Telcos will be studied from
different points of view : after measuring their ability to obtain a high turnover (§2.2),
especially on the recent period (§2.3 ), we shall perform another and more original
series of DEA analysis which will permit to estimate the performances of the operators
to develop the use of their network (§2.4), especially the use of new
telecommunications services (§2.5).
3.1 Definition of outputs and inputs for the DEA method
OUTPUTS
As far as OUTPUT (Telecommunications Services) is concerned, it is possible to take
TOTAL TURNOVER or evaluate it in PHYSICAL TERMS (e.g. number of
subscribers, minutes of toll use, optional services sold, etc.).
At first glance, Total Turnover does not seem to pose any particular problem from a
data gathering point of view. But, when the data are examined closely, it is remarkable
how different the methods of collecting the statistics is for each telecommunications
company (Telco). When comparing data obtained from the U.S. with data from Europe
for example, the Europeans publish data that include the total economic activities of
the Telco - as turnover also comprises revenues apart from its monopoly activities -
while the FCC (Federal Communications Commission) in the U.S. collects data strictly
from the Telcos' telecommunications activities.
An evaluation of the "physical" output is similarly troublesome for two reasons : first,
the enterprises do produce multiple forms of outputs, second, all the operators do not
furnish the same statistics on the same set of outputs. Hence, one can envisage at least
two indicators of physical output : the total number of calls or the total number of
minutes of communications on the network.
Eftlc1ency and financial performances in telecommumcations 317
The total number of calls completed on the network
The statistics relative to this measure have been utilized in other studies for calculating
productivity. We consider that productivity ratios constructed this way (using the
number of calls as the numerator) must be treated with the utmost prudence for the
following reasons :
a. First of all, the number of calls may well be evidence of data representing volume which is very sensitive to the evolution of the tariff structure. This means that as the price is lowered in the framework of "regulation", as it was in the United Kingdom, the number of calls increases. In this instance we cannot report the results as productivity improvement due to either a change in capital structure or labour deployment, but merely productivity changes resulting from output volume changes (calls completed).
b. The other major effect on the number of calls is the size of the network. In effect, a telephone network offers to its subscribers the potential for calls which is an exponential function with respect to network size (a function of the type: f{n(n-1)12}).
Thus, for a network four times larger, the number of potential calls is much greater. And effectively there is more domestic communications traffic in the U.S. than in Europe (for example, a U.S. company will have many subsidiaries and branch offices all over the United States and moreover the intra-country tariffs in Europe are much higher than inter-state tariffs are in the U.S.).
c. Furthermore, the number of calls also depends on the range of services offered on the network. The American Telcos have offered a whole series of services, introduced well ahead of the Europeans, that have generated a large number of calls (but not necessarily for longer duration).
The total number of minutes used on the network
This data is very interesting when characterizing output in physical terms. Quite
obviously, it also poses problems, particularly when it comes to obtaining comparable
figures from different countries.
318 Efficiency and financial performances in telecommunications
INPUTS
In accordance with the input selection, it is possible to construct partial or total factor
productivity. The accuracy of the indicators LABOUR PRODUCTIVITY and
CAPITAL PRODUCTIVITY is affected as both labour and capital are evaluated
differently by each Telco.
The factor - Labour
When evaluating the LABOUR FACTOR we shall use two methods : the total number
of employees which represents a "physical" evaluation of labour and gives the
productivity figures on a per employee basis. The wage and salary values, which
provide an economic evaluation of productivity and permit us to calculate, on average,
the productivity per $1 spent on wages and salaries.
The factor- Capital
Dealing with the CAPITAL FACTOR we shall use, once agam, two types of
calculations : the number of lines installed which represents the "physical" capital and
permits us to evaluate the Telco's infrastructure; and the tangible assets represented by
the estimate of net plant and equipment, to provide an "economic" measure of the
stock and equipment used in production (taking into account depreciation of the
equipment and its functionality as relating to their age and technical obsolescence).
The possibility of choosing different inputs and outputs means that multiple DEA
analyses can be conducted, and depending on the indicators chosen, different results
can be obtained.
We shall now give and discuss the main results of our analyses based on the DEA
method. First we shall present a standard analysis which includes allocative efficiency
because it contents data evaluated with the weights of prices. In a second part we shall
propose a more robust analysis because we shall use only physical inputs and outputs.
The synthesis will enlighten the main results.
Efficiency and financial performances in telecommunications 319
3.2 A first approach : the performances of the firms to obtain a high
turnover
Our approach is input oriented with the variable returns to scale hypothesis. We
observe one output : the turnover for the year 1994 (In the following paragraph 2.3 a
complementary analysis will be proposed for the recent period 1992-1997). This
output is obtained through the two main inputs : capital and labour. According to the
data used for evaluating capital and labour, different DEA analysis can be performed.
The table hereafter indicates the three DEA analysis that aim to measure the
performances of the operators to obtain a high turnover in 1994.
Summary of the DEA analysis performed in the paragraph 3.2
Inputs Output
DEA 1 Wages and salaries* and net plant and Turnover* equipment*
DEA2 Wages and salaries* and number of Turnover* main lines
DEA3 Number of employees and number of Turnover* main lines
* in constant $
If we evaluate the inputs in money, the evaluation of labour is done by wages and
salaries, and capital corresponds to the net plant and equipment. The results are given
in figure I as DEA 1 :
European Telcos are in very good position : the efficient firms are ALL (Deutsche Telekom), OK, IRL, ITA, UK (British Telecom);
very near these efficient firms we find : FR (France Telecom) and some BOCs (SW and BA). ESP and AUT are behind.
These first results seem to indicate that a link between regulatory regime and
performance is not obvious : some BOCs as well as some European operators obtain a
good level of performance.
320 Efficiency and financial performances in telecommunications
However we can ask ourselves about the input in capital : capital is a very specific and
difficult indicator to evaluate; and the framework of an international comparison
reinforces this measurement difficulty (with the problems linked to different account
systems, depreciation and so on).
So we have used a physical indicator of capital : the number of main lines which
constitutes the basis of the capital of an operator in telecommunications. Thus a new
DEA analysis is conducted (see DEA 2 in figure I) with still the same output
(turnover) and the following inputs : labour evaluated by wages and salaries and
capital through the number of main lines. The results are very near the first analysis,
and this physical evaluation of capital (number of lines), even if it is not a perfect one,
suffers less critical comments than an evaluation in money.
Output : Turnover If now we substitute N (number of employees) to wages and salaries
we obtain new results which show how the inputs labour and capital evaluated in
physical terms are used by the different operators to obtain their turnover : in this case
(DEA 3 in figure 1), the BOCs appear in a relative better situation and the following
operators obtain worse performances than in DEA I : BEL, DK, FR and UK.
Figure 1: Efficiency scores: DEA I to 4
Inputs: Capital and labour Output: Turnover
1.2 .---------------------------------------------------------,
0.8 -
0.6 -
0.4
0.2
Effic1ency and financial performances in telecommunications 321
3.3 The evolution for the recent period 1992-1997: a "window analysis"
In order to study the evolution of the Telcos performances, we first made a DEA
analysis for the year 1989 which permits the comparison with DEA 3. We estimate the
turnovers of the different operators for 1989, after converting current data into
statistics expressed in fixed value dollars (in $1994 and in price 1994). Every impact of
dynamics due to inflation or exchange rates has been eliminated. This DEA analysis
(DEA 4 in figure 1) does not show a very important evolution between 1989 and 1994.
The BOCs were already in a position of leader in 1989 and the other European
operators have improved their scores of efficiency during the period 1989-1994. This
is a very preliminary result about the dynamic analysis, which is coherent with the fact
that the period 1989-1994 corresponds to a first stage of deregulation.
It is more interesting to observe the recent evolution. We made a senes of DEA
analysis similar to the DEA 1 to 3 for the period 1992-1997. The number of operators
included in our analysis has been limited by the availability of data. So we have
compared the main European operators (France Telecom, Deutsche Telekom, Telecom
ltalia, British Telecom) and some American Telcos for which data have been available,
that is to say Ameritech, Bell Atlantic (before and after the merger with Nynex), US
West. Given the small number of operators, and in order to have enough data for an
analysis, we have performed a "window analysis" using a three-year window, on the
period l 992-1997. As a whole, the European operators keep a good score for DEA I,
while the American operators are still the leaders for DEA 3. In the recent period the
hierarchy is the following one for DEA 3 (from the weakest to the highest score) :
France Telecom, Deutsche Telekom, British Telecom, Telecom Italia, then the
American Telcos.
Our conclusion arises now : if the European operators obtain good performances,
when we evaluate their capacity to have an high turnover with inputs evaluated in
money, their situation is worse when labour and capital are not weighed by
prices. We can note that Switzerland (not shown in the above figure) has a good
performance in any case with high salaries : probably it is due to a very high turnover
explained by the specific position of this operator (a monopoly with a lot of
international communications).
With all these DEA analysis we don't show a discriminated evolution between the
main operators, which would be linked more specifically to different regulatory
322 Efficiency and financtal performances in telecommunications
schemes. These first results can be compared to those ofT. Sueshoyi [ 1994] who uses
OCDE data but only for the year 1987 ; Sueyoshi applies a stochastic frontier
production analysis and takes into account three inputs (the number of telephone main
lines, the number of employees and the total amount of capital investment) and one
output (the total amount of telecommunications service revenues). He shows that four
countries, i. e. Iceland, Norway, Switzerland and the United States (taken as a whole,
the firms are not studied), were efficient according to the method employed, while
Germany had a bad score (0,70) as well as Italy (0,71), and the United Kingdom and
France reached intermediate scores (around 0,85). The results of Sueyoshi based on the
same type of inputs and output are comparable to those we obtain with DEA 4,
especially concerning the position of the American operators and of Switzerland (the
case of Germany is specific since the German operator was included in Deutsche
Bundespost in 1987, which induces difficulties for some statistics). By studying the
more recent period (1992-1997) we bring to light the favorable dynamics of an
operator, Telecom Italia, while among the European operators the position of France
Telecom seems to become a little less good, Deutsche Telecom has regressed a little,
and some movements have caracterized the situation of British Telecom because of the
regulatory evolution (price-cap regulation) which has affected directly its turnover.
However there has been no great gap, nor very significant evolution.
We propose now a complementary study based on "physical" data which enlightens
technical efficiency.
3.4 Performances of the operators : their capacity to use the network
We shall now use only physical indicators and try to appreciate how the different
operators have developed the accessibility of their network. Two inputs will be
considered : labour measured by the number of employees and the number of main
lines as an estimation of the capital of each operator.
We have chosen two different outputs : the first often quoted is the number of calls.
The second is a statistics very difficult to obtain : the number of minutes of use of the
network; this statistics is the best approach to appreciate the real use of the network.
Effictency and financial performances in telecommunications 323
First, we shall provide a large analysis for 19 operators in 1993, for the output number
of calls. Then we will try to grasp the dynamics of performances : so a "window
analysis" will be performed for a limited number of operators with the number of
minutes of communications as the output. Finally, the same type of analysis will be
conducted but with the number of calls as the output.
Summary of the DEA analysis performed in the paragraph 3.4
Inputs Output
DEA 5 Number of employees and number of number of calls main lines
DEA 6 Number of employees and number of number of minutes of communications et 7 main lines
DEA 8 Number of employees and number of number of calls et 9 main lines
DEA 5 (see figure 2) corresponds to a large analysis for 19 operators in 1993 with the
number of calls as the output. The BOCs appear in the first position. Only Finland,
Denmark and Switzerland obtain a good score. Other European operators are very far
behind: with a score of only 25% Germany, France and the United Kingdom have to
do a lot to improve the use of their network.
Figure 2: Efficiency scores : DEA 5
Number of calls (output) and number of employees
and of lines (inputs)
1.2
0.8
0,6
0 ,4 - --• I
0.2
0
• I • I I I I
ALL OK ESP AM BA BS NY PA SW US BOC FIN FR ITA NTT NDL POL UK SUI
324 Efficiency and financial performances in telecommunications
One may wonder if there has been an evolution during the recent years. In order to
give an answer we propose a "window analysis" for the periods 1988-1990 (DEA 6)
and 1995-1997 (DEA 7). The figure 3 gives the scores corresponding to DEA 6 and
DEA 7 with only one point by firm at the begining and at the end of the period. The
preceding result is confirmed : the American telcos are far behind the European
operators concerning the efficiency in giving access to the network. The score of the
European firms is about 40% at both the beginning and the end of the period ( 1988-
1990 or 1995-1997). The DEA 8 and 9 analysis give similar results and therefore are
not presented here.
It seems that the different regimes of regulation have some impact on productivity or
on performances of the different firms. We can verify that the first place belongs to the
BOCs and that the United Kingdom, which was the last in 1989, is before France
Telecom and Deutsche Telekom in 1995-1997. British Telecom, submitted to an
evolution of the regulation, has very much improved its relative position ; as a matter
of fact, this improvement has been correlated to an important slowdown in the number
of employees (246000 employees in 1989, 124 700 in 1998).
We shall be able to confirm (or infirm) these first conclusions by a new analysis which
takes into account the development of telecommunications industry from a general
point of view.
Figure 3: Efficiency scores
DEA 6 and 7 (1988-1990 and 1995-1997). The use of the network (number of minutes of communications as the output).
Efficiency and financial performances in telecommunications 325
3.5 The development of the telecommunications services
We have to precise our analysis because it is a new approach of the problem. In the
previous paragraphs we estimated performances on a classical basis with outputs like
turnover or number of calls. Now we consider that for the last decades, even in the
USA or in the United Kingdom where deregulation or privatization have taken place,
national operators have been in a situation of quasi monopoly in their area (it is the
case for the BOCs, and in the United Kingdom BT had not still a strong competition
from Mercury until the recent years). In such a situation it is interesting to appreciate
how some new services have been developing in different countries. The outputs for
the following analyses will be the numbers of Fax, of pagers and of mobiles which
indicate the level of development of telecommunications in each country, compared to
the number of lines and the revenue from telecommunications services. In other words
we shall analyze in which proportion an operator with a large infrastructure in lines
and an important revenue contributes to the development of the new services of
telecommunications.
Summary of the DEA analysis performed in the paragraph 3.5
Inputs Outputs
DEA 10 and DEA Turnover and number of main Fax, Pagers and Mobiles II lines
Figure 4: Efficiency scores DEA 10 and DEA 11
The contribution to the development of new services (Fax, Pagers
and Mobiles, 1990-1996).
ALL AUT BEL ESP FIN FR ITA NDL PORT UK UE USA JAP
326 Efficiency and financial performances in telecommunications
The results for 13 countries for 1990 and 1996 (DEA 10 and DEA 11 in figure 4) are
the following: the three leaders in 1990 as well as in 1996 are : USA, Japan and
Finland. We can observe that the three main European operators are very far from the
level of these leaders. Nevertheless France and Italy have progressed very quickly
during the period.
4 The Financial Analysis
4.1 A brief presentation of the indicators
Our goal is now to observe the performances of the Telcos from a financial point of
view. We need not a detailed financial analysis for each operator ; we have chosen
some indicators in order both to compare the firms at an international level and to have
an idea of the evolution of the main financial criteria. From the database which has
been constituted different ratios can be calculated; the three ratios which are reported
hereafter are very significant of the financial situations of the Telcos.
I. The first ratio is a classical measure of the profitability of a firm from the point of view of equity owners. The return on equity measures the ability of the firm to generate profits available to common shareholders. This ratio is the following one:
Rr= Net Income I shareholders' equity
2. Profitability can be analysed from another point of view, which insists on some economical aspects. Among the potential indicators, one has appeared very significant for comparing the operators from the point of view of their ability to generate earnings thanks to their operating activity. We shall report here a ratio with the sum of the operating income and depreciation and amortization in the numerator and the revenues from sales in the denominator. So the numerator is an estimation of earnings before interest, taxes, depreciation and amortization, which is noted EBITDA. This statistics is often considered as the most important indicator of their operating performances by US firms.
Efficiency and financial performances in telecommunications 327
The EBITDA margin or ratio is the following one :
EBITDA I Turnover
3. It is also useful to compare the performances of the tel cos from the point of view of their financial independence, which means to have an idea of the influence of debt. The third ratio reported here takes into account the capacity of the firm to generate Cash Flow in order to cover its debt. More precisely, the numerator corresponds to the indicator called "Capacite d' Autofinancement" in the French conception which differs a little from the American "Cash Flows from Operating Activities" ; of course, the estimation of this statistics, noted CAF, is not easy on an international level because of the differences in the presentation of financial statements from country to country. In the denominator we use only the longterm debt noted D, which are the loans and other borrowings with amounts falling due after more than one year. It is clear that the ratio is decreasing when the long-term debt of a firm tends to increase while the company has difficulty to increase or even maintain its cash-flow.
So the third indicator will be called Cash Flow to long-term debt ratio and noted
CAF I D.
4.3 Main results of the financial analysis
It is necessary to insist again on the great precautions which have to be taken in order
to make international comparisons because of the difficulties in collecting comparable
data. Moreover, the financial time series are not always homogeneous; in many cases
we can observe some statistical breaks : either some important changes in accounting
principles were introduced in a country or for a firm and the series became not
comparable from one year to another or there happened a special event, such as for
Deutsche Telecom the reunification of East and West Deutschland and the
corresponding changes in the financial statements.
We can give an example of the difficulties to analyse financial performances in
dynamics. If we are studying France Telecom, we must remember that there are at
least four important discontinuities in the time series :
328 Efficiency and financial performances in telecommunications
the introduction of the value added tax (TV A) at the end of 1987 leads to very large changes in data for 1988, especially the falls in the turnover and the net income, which have repercussions on ratios such as Rr
there is also a discontinuity in the capital time series because of a change in the accounts in 1991. This explains the decline of the ratio Rr between 1990 and 1991.
In 1994 for the first time the accounts were published for the group France Telecom. Moreover, there was a change in the fiscal system applied to France Telecom. For 1995 and 1996 the only collected data concern the group and no direct comparison is possible between the data for the company France Telecom (head office) until 1994 and the group from 1994.
Similarly the regulation of France Telecom changed at the end of 1996 and this affects almost all the data.
However, we have elaborated an original database on over ten years which permits to
measure the financial performances. In order to have a clear interpretation we present
and analyse herafter the three chosen indicators for some operators which constitute a
good representative sample of the main financial situations in the telecommunications
sector at an international level.
The study of the three financial ratios (see figure 5 hereafter which gives results for
1988 and 19972) leads to three main results :
I. The differences between operators are important. In particular, we can observe that the less deregulated operators (deregulation appears only at the end of the period of observation), France Telecom, Telecom Italia and Deutsche Telecom, obtain less good results in terms of return to equity and, at a weaker degree, in terms of the CAF I D ratio (except for the last year concerning CAF I D), but better results with the EBITDA ratio which reflects operating performances. We can propose the following explanation : deregulation needs good financial performances for equity holders and a capacity of the firm to resist to competition pressures, while a regulated monopoly is not very much concerned by return to equity, nor by the capacity to reimburse the debt but can obtain a high turnover and consequently high operating ratios without price constraint. Price regulation has been in effect only recently for France Telecom, Deutsche Telekom and Telecom Italia.
! For Nynex and Pacific Teles1s the data are for 1988 and 1996 smce the two compames merged respectively w1th Bell Atlantic and SBC m 1997.
Efficiency and financial performances in telecommunications 329
However, differences inside the group of the Europeans and inside the group of the Americans have increased and are to-day important. So, while we have observed, thanks to DEA analysis, a great contrast between the European and the American operators concerning the use of their network, the financial indicators of performances show a weaker opposition, especially at the end of the period.
2. In fact, if we observe the dynamics of each operator, some significant changes appear in financial performances. One of the most obvious changes is the progression of the financial performances of BT and Telecom Italia among the Europen operators. BT has to-day a quite high level of return to equity, a good EBITDA ratio and above all a high cash flow compared to its long-term debt. The Rr and CAF I D ratios ofTelecom Italia have also increased.
In a very slight deregulated context, France Telecom seemed until recently to have good tinancial performances if we refered to the EBITDA ratio, thanks to a price policy which, nevertheless, limited the uses of telecommunications services; but, on the one hand, the CAF I D ratio has been at very lower levels than those of BT or Telecom Italia, and on the other hand, in dynamics the EBITDA ratio has been declining. The situation of Deutsche Telecom is the worst of all the European operators concerning the EBITDA ratio. So from a financial point of view DT and France Telecom have inferior performances by comparison to BT and Telecom Italia.
Figure 5: Three financial ratios 1988 and 1997: Rf, Ebitda ratio, CAJt'ffi
120%
100%
80%
80%
• 0%
20'>\
0% Nl. BA BS NY PA SoN US BT OT F'T ITA
•Rree . RIV7 [] El>itdl88
C E-97
. CAf/088
. CAf/097
__ j
3. If deregulation is an incentive for better performances, it has not the same effects on every firm. In a competitive context some firms can be in trouble. In the United States the increased competition and the evolution of the regulation rules
330 Efficiency and financial performances in telecommunications
have been favorable to some BOCs, but others have had difficulty in adapting themselves. According to a normal competitive process mergers have taken place; in 1997 Bell Atlantic entered into a merger agreement with Nynex and SBC Communications Inc., formerly called Southwestern Bell, merged with Pacific Telesis. In 1998 SBC entered into another merger with Ameritech and a merger between Bell Atlantic and GTM was announced. If we consider the two mergers in 1997, the BOCs with very good financial performances were in a position of leaders in these mergers with the BOCs which had worse financial performances (for example, Nynex had a very bad CAF I D ratio in the three years preceding the merger with Bell Atlantic).
The analysis of the financial performances shows quite well the evolution of the telecommunications sector at the international level : in the United States, in a competitive context the differences between operators have increased and some restructurations have occurred; in Europe, Deutsche Telecom and France Telecom are behind British Telecom and Telecom Italia.
5 Conclusions: Key Factors for Performances
We can enlighten the progress and the interest of our study at two main levels :
- methodology,
- results concerning the behaviour of the Telcos, their performances, and the
impact of the regulatory regime on the development of telecommunications uses
(in a large acceptation).
1. DEA appears to be a very interesting method to evaluate performances, from the
productive efficiency point of view, of the observed firms by taking into account
many inputs and many outputs. Moreover, we propose an analysis on a very
recent period for some indicators.
A financial analysis gives some complementary indications about the relative
forces and difficulties of the Tel cos.
At this stage of our research we want to precise that the DEA analyses and the
financial ratios have to be interpreted with precaution : of course the results are
dependent of the panel of units we observe and of the data. A very difficult task
is to obtain and organize a reliable database for a long period. This is why we ask
Efficiency and financial performances in telecommunications 331
the reader to consider some results as preliminary (the DEA scores may be
influenced, for example, by too narrow a panel, and so on).
2. Keeping this remark in mind, we can give some conclusions about the
performances of the telecommunications operators.
The performing Telco will be the one which :
- encourages the growth of the number of calls,
- favours an increased use of its network from the point of view of the number
of minutes of communications,
- reduces its tariffs,
- contributes to the development of new services.
This is an original approach of the performances by comparison with many works
which relied on a global indicator based on productivity or cost, as well as with
the study of Sueyoshi [ 1994] which does not take into account the use of the
network and the development of services. Not only we have shown that the use of
the network is very low in Europe comparatively to the USA and Japan if we
consider the number of calls as well as the number of minutes, but we have also
given an evaluation of the contribution of the Telcos to the development of
telecommunications uses in a very large acceptation. With this last consideration
it appears that the USA and Japan are far ahead the other Telcos; probably this is,
to a certain extent and with a lot of precaution, due to the deregulation regime of
these countries. The United Kingdom is a special case among Europeans, since
this country was very far behind the other European Telcos in the 80's and today
BT is at the same level as the other European Telcos.
To emphasize the differences in the uses of telecommunications we have
constructed hereafter a last figure : the horizontal axis indicates the use of the
network and the vertical axis the cost of the access to the network (with adapted
scales in the figure below). Clearly we have two groups of situations : a group
with a very important use and a very low cost of the access (it is the group of the
BOCs), and the group 2 with less uses and a cost which is very high and variable
(from 20 to 80 and even 120 -for Switzerland-). Note that we obtain the same
332 Efficiency and financial performances in telecommunications
type of figure with the number of calls or the number of minutes of
communications on the horizontal axis. The figure above corresponds to the year
1993 for which many data have been available, but for the more recent years,
1996-1997, similar results can be noticed: France telecom has a very high
turnover by line and a very weak number of minutes by line (high prices and low
access) ; the other European Telcos have approximately the same number of
minutes by line, but for Deutsche Telekom and British Telecom the turnover by
line is 3 to 8 times lower (low access but lower prices, comparatively to France
Telecom); the American operators have both a low turnover by line and a large
number of minutes by line (low prices and best access).
Figure 6: The use of the network (horizontal axis : number of calls
per line) and an indicator of price (vertical axis: turnover per line) 140
I 120 •
I ·-
- f-. ·-
100
80
• •j Group I . ~
60
• 40 •---I I • Groupe • • ~lA ' • • 20
0 I 0 2 3 4
Undoubtedly deregulation will lead the group 2 in the direction of the group 1. Even if
the link between deregulation and performance is not obvious to establish, we think
that by many ways we have throw light on it.
Some complementary results are provided by the financial analysis. In a deregulated
context such as in the United States, the performing firms both in productive and
financial terms absorb less performing operators. The situation of European operators
is contrasted : DT is in a course of a restructuring processs with relatively bad
performances, while FT has relatively good productive performances but a weak
development of services and not very good financial performances. BT and Telecom
Efficiency and financial performances in telecommunications 333
Italia are in a better situation. BT was very far from the point of view of physical
performances in the mid-1980s but has improved these performances, and has obtained
good financial performances with a net increase of the Cash Flow to long-term debt
ratio.
Two issues are emerging :
firstly in a deregulated framework telecommunications operators clearly have
good productive performances, especially in the use of the network, in the
diffusion of new services; but the weakest firms increase their debts and finally
have to enter into mergers; so, in the USA where deregulation is already far in
advance, the recent process of concentration raises the question of a new
deregulation ;
secondly in a country such as France where the deregulation is only at an early
stage, France Telecom has long obtained relatively good productive
performances with high prices but has not impulsed strongly the use of the
network and new services and has not favored a quick development of the
information society (the more obvious example is at the present time the case of
Internet) ; moreover the situation of France Telecom is evolving with financial
performances inferior to those of British Telecom or Telecom Italia.
There is some correlation (but we are not able to show a causality) between market
organization productive performances (DEA method), and financial performances : the
efficient BOCs and BT have a high return on equity, with a low CAF/D ratio, and a
good diffusion of services (only for the BOCs) while operators in an early stage of
deregulation (France Telecom or Deutsche Telekom) have less good performances.
Finally, we have to keep in mind that the evolution in the telecommunications sector is
continuous. For example, taking into account the burst of the number of mobiles in
France during the last twelve months would probably lead to new analysis.
Nevertheless it remains that the European operators, especially France Telecom, offer
consumers a limited access to the network with high prices. In the context of the
emerging global information society (see the Bangemann report), in which the
diffusion of information gives rise to knowledge and competitivity, this situation has to
change. It is probably one of the most important challenges that the European
regulation organizations have to take up.
334 Efficiency and financial performances in telecommunications
Annex: database
The database includes the telecommunications operators of the following countries (the abbreviations used in the paper are indicated in brackets)
GERMANY (ALL) or Deutsche Telekom AUSTRIA (AUT) or OPT
BELGIUM (BEL) or BelgacomDENMARK (DK) or Tele Danemark
SPAIN (ESP) or Telefonica FINLAND (FIN)
FRANCE (FR) or France telecom NETHERLANDS (NDL) or PTI Nederland
IRELAND (IRL) or Telecom Eireann PORTUGAL (PORT)
ITALY (ITA) or Telecom ItaliaPOLAND (POL)
JAPAN (JAP) or NTI UNITED KINGDOM (UK) or British Telecom
SWITZERLAND (SUI) or PTI SUISSE
For the DEA analyses I to 9 the data are those of the main operator of each country (for example, British Telecom for UK); for the DEA analyses 10 and II the data are those of the countries.
For the United States data on each of the seven Bell Operating Companies as well as on the total of the BOCs were collected :
AMERITECH (AM) PACIFIC TELESIS (PA)
BELL ATLANTIC (BA) SOUTHWESTERN BELL (SW)
BELL SOUTH (BS) US WEST(US)
NYNEX(NY)
The group of the seven BOCs has been symbolized by BOC.
References
Badillo, P.-Y. (1994): Les productivites des operateurs de telecommunications, Communication at the
XLth International Conference of the Applied Econometrics Association, Osaka, Japan, 24 and
25 March.
Badillo, P.-Y. (1997): Has deregulation an impact on productivity and performances? -Some
preliminary results for telecommunications-, Communication at the EURO CPR '97, 23-25
March 1997, Venice, published in Communications et Strategies, 1997.
Badillo, P.-Y. (1999): L'efficience des operateurs de telecommunications : une comparaison
intemationale, in: Badillo, P.-Y. and Paradi J.-C. (Eds.) (1999): La methode DEA, analyse des
performances, Paris, Hermes.
Efficiency and financial performances in telecommunications 335
Badillo, P.-Y. and Paradi J.-C. (Eds.) (1999): La methode DEA, analyse des performances, Paris,
Hermes.
Baumol, W. J., Panzar, J. C. and Willig, R. D. (1982): Contestable Markets and The Theory of
Industry Structure, New york, Harcourt Brace Jovanovich.
Braeutigam, R. R. (1989): Optimal Policies for Natural monopolies, in: Schmalensee, R. and Willig,
R. (Eds), Handbook of Industrial Organization, Amsterdam, North-Holland, Volume II, pp.
1289-1346.
Braeutigam, R. R. and Panzar, J. C. (1993): Effects of the Change from Rate-of-Return to Price-Cap
Regulation, American Economic Review, mai 1993,83 (2), pp. 191-198.
Charnes, A., Cooper, W. W., Lewin, A. Y. and Seiford, L. M. (Eds.) (1994): Data Envelopment
Analysis : Theory, Methodology and Applications, Boston, Kluwer Academic Publishers.
Gnltches, Z. (1987): Productivity : measurement problem, in: Eatwell, J., Milgate, M., Newman, P.
(Eds.), The New Palgrave Dictionary of Economics, London, Macmillan Press, pp. 1010-1013.
Joskow, P. L. and Rose, N. L. (1989): The Effects of Economic Regulation, in: Schmalensee, R. and
Willtg, R. D. (Eds.) (1989): Handbook of Industrial Organization, Volume II, Elsevier Science
Publishers, pp.1449-1506.
Kahn, A. E. (1971): The Economics of Regulation, Principles and Institutions, Second Printing 1989,
Cambridge, Mass, MIT Press.
Kndel, D., Sappington, D. and Weisman, D. (1996): The Effects of Incentive Regulation in the
Telecommunications Industry: A Survey, Journal of Regulatory Economics, 9 (3), pp. 269-306.
Lafont, J. J. and Tirole, J. (1993): The Theory of Incentives in Procurement and Regulation,
Cambndge, Mass, MIT Press.
Lafont, J. J. ( 1994): The New Economics of Regulation Ten Years After, Econometrica, 62 (3), May,
pp. 507-537.
Leibenstein, H. (1966): Allocative Efficiency vs. X-Inefficiency, American Economic Review, 56 (6),
June, pp.392-415.
Lovell, K. C. A. (1993): Production Frontiers and Productive Efficiency, in: Lovell et al. (Eds.)
(1993): The measurement of Productive Efficiency, New York Oxford, Oxford University
Press, pp.3-67.
Noll, R. G. (1989): Economic Perspectives on the Politics of Regulation, in: Schmalensee, R. and
Willig, R. D. (Eds.) (1989): Handbook of Industrial Organization, Volume II, Elsevier Science
Publishers, pp.l253-1287.
336 Efficiency and financial performances in telecommunications
Schaffnit, C., Rosen, D. and Paradi, J. C. (1995): Best Practice Analysis of Bank Branches : An
Application ofDEA in a large Canadian Bank, Working Paper.
Seiford, L. M. (1996): Data Envelopment Analysis : The Evolution of the State of the Art (1978-
1995), Journal of Productivity Analysis, Vol. 7 n° 2/3, July 1996, pp. 99-137.
Sueyoshi, T. (1994): Stochastic frontier production analysis: Measuring performance of public
telecommunications in 24 OECD countries, European Journal of Operational Research, 74,
1994, pp. 466-478.
Viscusi, W. K., Vernon, J. M. and Harrington, J. E. Jr (1995): Economics of Regulation and Antitrust,
Cambridge, Mass., The MIT Press, Second Edition.