Omega 41 (2013) 442–458

Contents lists available at SciVerse ScienceDirect

Omega

0305-04

doi:10.1

n Corr

E-m

v.lieber

journal homepage: www.elsevier.com/locate/omega

Benchmarking airports from a managerial perspective

Nicole Adler a,b,n, Vanessa Liebert c, Ekaterina Yazhemsky a

a Hebrew University of Jerusalem, Israelb Northwestern University, United Statesc Jacobs University Bremen, Germany

a r t i c l e i n f o

Article history:

Received 1 February 2011

Accepted 20 February 2012

Processed by B. Levstudies either treat the airport production technology as a black box, or separate the terminal and

airside activities, assessing them individually. In this article we analyze airports as a single unit due to

Available online 28 February 2012

Keywords:

Air transport

DEA

Benchmarking

Business policy

83/$ - see front matter & 2012 Elsevier Ltd. A

016/j.omega.2012.02.004

esponding author.

ail addresses: msnic@huji.ac.il (N. Adler),

t@jacobs-university.de (V. Liebert), katy1@ms

a b s t r a c t

Benchmarking airports is currently popular both in the academic literature and in practice but has

proved rather problematic due to the heterogeneity inherent in any reasonably sized dataset. Most

the direct complementarities, thus avoiding the artificial separation of inputs. Using data envelopment

analysis (DEA), we open the black box in which a network describes the production process, thus

demonstrating the sequential effects that separate final from intermediate outputs, including those

under partial managerial control and those that are known to be non-discretionary. To further improve

the benchmarking process, we identify appropriate peers for a case study of 43 European airports over

10 years, through a restricted reference mechanism according to pre-defined characteristics. Compared

to basic DEA models, the results of the proposed structure provide more meaningful benchmarks with

comparable peer units and target values that are potentially achievable in the medium term. By

identifying each unit’s individual reference set, unique outliers influence the performance measure-

ment less severely than occurs under basic DEA. In addition, the formulations produce an implementa-

tion path that moves the airport towards the Pareto frontier gradually, taking into account the

regulatory and business environment in which the unit is located.

& 2012 Elsevier Ltd. All rights reserved.

1. Introduction

According to the Princeton dictionary, an airport is defined as‘‘an airfield equipped with control tower and hangars as well as

accommodations for passengers and cargo’’. Airports can be definedas an important, basic infrastructure to a society in which aviationis one of the drivers of a modern economy. An alternativeapproach defines an airport as a private production system inwhich society maximizes social welfare by encouraging airportmanagement to maximize profits, whilst simultaneously consid-ering consumer surplus via some form of airport regulation ifdeemed necessary. Consequently, it is unclear whether airportsshould be considered as a not-for-profit, public good, as is thegeneral approach in the United States, or as a private enterprisemaximizing shareholder value. Since it would appear to be truethat large regions of the world are gradually adopting theprivatized form [1] and that independent authorities managingpublic airports in the United States do not behave differently totheir private counterparts with respect to productivity [2], in this

ll rights reserved.

cc.huji.ac.il (E. Yazhemsky).

research we develop an airport benchmarking methodology froman airport manager’s perspective in which we assume that theairport intends to maximize revenues and minimize costs.

Liebert and Niemeier [3] review airport benchmarking studiesapplied to a diverse range of activities using various methodolo-gies. The methods most frequently applied include price indextotal factor productivity [4–6], parametric stochastic frontieranalysis [7,8] and non-parametric data envelopment analysis(DEA). DEA has been used to compare the performance of airportswithin national boundaries, including the U.S. [9,10], U.K. [11],Spain [12,13], Australia [14], Taiwan [15,16] and Portugal [17] aswell as airports around the world [18,19]. It is rather difficult todraw general inferences since many of these articles arrive atdirectly opposing conclusions. For example, Murillo-Melchor [13]show that Spanish airports in their dataset suffer from decreasingreturns to scale whereas Martın et al. [20] concluded increasingreturns to scale for the same set of airports. Abbott and Wu [14]found most Australian airports enjoy increasing returns to scale,Pels et al. [7] argue that European airports operate under constantreturns to scale in air traffic movements and increasing returns toscale on the terminal side and Lin and Hong [19] argue that mostairports are not operating at an optimal scale. Graham and Holvad[21] and Abbott and Wu [14] argue that Australian airports aremore efficient than their European counterparts, Lin and Hong

N. Adler et al. / Omega 41 (2013) 442–458 443

[19] argue that the U.S. and European airports are more efficientthan their Asian and Australian counterparts and Pels et al. [7]conclude that widespread European airport inefficiency is notspecific to a country or region. Consequently, Morrison [22] hascalled for a balanced approach and a dialog between airportmanagers and researchers.

The majority of previous studies have treated airport technol-ogy as a single production process, avoiding the complexityinherent to airport systems. Gillen and Lall [9] and Pels et al. [7]were the first to argue that the airport could be analyzed as twoseparate decision-making processes, one serving airside activitiesand the other serving landside production. Yu [16] is the first topresent an operational framework of airport services as a multi-stage process opening the black box via network DEA [23],utilizing a slack-based measure of airport network performance[24]. Yu [16] analyzed airport operational activities decomposedinto production, airside and landside services, whilst taking intoaccount environmental factors (population). In the productionstage, inputs include labor and quasi-fixed runway, terminal andapron areas. The intermediate products are defined as runwayand terminal capacities. The airside outputs include aircraftmovements and the landside outputs cover cargo and passengertransportation. We also argue that a single black box approachwould be insufficient to capture the rich picture underlying themulti-stage airport technology as demonstrated in Fig. 1. Sincethe liberalization of the aviation industry in Europe in the lateeighties, airports have focused on both aeronautical and com-mercial landside activities. The network DEA approach recognizesthe fact that generalized and fixed costs connected to the two setsof activities can only be split in an artificial manner and that whileaeronautical revenues draw from passengers, cargo and air trafficmovements, the non-aeronautical revenue is more closely tied topassenger throughput. Although airports may have limited con-trol over traffic volume, non-aeronautical revenues drawn fromcommercial activities, such as airport cities, are indeed within thepurview of airport management. As argued in Oum et al. [25], the

Labor (full-time equivalent employees): Managerial Engineering Maintenance CleaningSecurityCapital: Terminal capacity Runway capacity Apron capacity (terminal/remote) Security capacity Baggage handling capacity Airport area Gates Public parking spots Materials and supplies: Outsourcing costs Snow removal equipment Fire truck & stations Hangers Maintenance costs

Business & leisInternational-traInternational noDomestic-transfDomestic non-t

Air Transport MJetsRegional Jets/TGeneral AviatioCargo

Non-traveling c

Fig. 1. Airport netw

omission of outputs such as commercial services is likely to biasefficiency results as it underestimates the productivity of airportswhose managers focus on generating additional revenue sources.Many airports attempt to increase revenues from non-aeronau-tical sources which are not directly related to aviation activities inorder to cross-subsidize aviation charges in turn attracting moreairlines and passengers to their airport [26]. Revenue sourcediversification that exploits demand complementarities acrossaeronautical and non-aeronautical services appears to improveairport productive efficiency [2]. We would argue that it is morereasonable to analyze airports as a single unit because of thedirect complementarities, thus avoiding the need to artificiallyseparate inputs between the terminal and airside. Consequently,in this research we develop a DEA modeling approach in order tomeasure the relative cost or revenue performance of airports withrespect to aeronautical and commercial activities, whereby activ-ities are connected via passengers as the common intermediateproduct.

Another issue that arises in the airport benchmarking litera-ture is the problem of comparability. A base assumption withinthe DEA context that has been questioned in the literature is thehomogeneity of the decision-making unit under analysis and theappropriateness of this assumption with respect to airports [22].The aim of the formulations presented here are to directly broachthe question of airport benchmarking in light of the reasonablelevel of heterogeneity to be found in any multiple airport study,which is necessary to generate sufficient data points for purposesof analysis. In order to ensure comparability, we apply a restrictedreference approach [27] which forms individual reference setsbased on similar mixes of inputs or outputs and intermediateproducts. Certain inputs may be beyond managerial control in theshort to medium term yet affect airport performance [18]. Ingeneral, capital is frequently treated as a non-discretionaryvariable over which airport management has little to no control[28]. In this research, capital has been defined in terms ofdeclared runway and terminal capacity which are agreed upon

ure passengers: nsfer n-transfer er

ransfer

ovements:

urboprops n

ustomers

Undesirable outputs: Non-weather related delays Aircraft noise Air pollution

Concession revenues: Duty free and retail Catering Car parking RentalBanking Entertainment Passenger services Airport cities

Aeronautical revenues:Aircraft landing fees Passenger charges & fees Aircraft parking fees Ground handling fees Cargo fees Centralized infrastructure fees Environmental & Noise surcharges Security charges

ork technology.

1 Constant returns to scale means that the producers are able to linearly scale

the inputs and outputs without increasing or decreasing efficiency.

N. Adler et al. / Omega 41 (2013) 442–458444

within a multiple stakeholder setting and result in a number thataccounts for the airport system configuration. For example, someairports consist of a reasonably large number of runways how-ever, for reasons of weather and/or geographical layout, only asmaller portion may be in use at any given time. Declaredcapacity takes this into account as compared to simply countingthe number of runways. On the terminal side, check-in counters,security queues, passport control and gates together produce athroughput level per hour that is otherwise assumed to be linearwithin a standard DEA framework. We argue that the capacity ofan airport, as a proxy for capital, may be adjusted to a certainextent in the medium term, hence the model restrictions permitterminal and runway declared capacity to change up to a pre-determined level. Pure capital investment is not an appropriatemeasure even within a specific country because the accountingprocedures differ, rendering the information incompatible.Finally, principal component analysis (PCA) combined with DEA[29,30] is applied in the input-oriented model in order to reducethe curse of dimensionality and the resulting bias, reducing theset of peer airports from 49% to 39% in the most conservative datareduction when at least 95% of information is retained in themodel in the current case study described in Section 4.

The aim of this research is to develop a comprehensivemethodology tailored to benchmarking from a managerial per-spective. A comparison with basic DEA results demonstrates thatthe additional reference restrictions lead to more reasonable peercomparisons, permitting an analysis of strategies which couldpotentially be adopted over the short and medium term forplanning purposes. The modeling approach in this research allowsairport managers to include their industry knowledge in the formof comparability with regard to airport size, operating conditionsand restricted variability of capacity, which is encapsulated in therestricted reference approach. Given that our dataset does notinclude all airports in Western Europe due to a lack of publiclyavailable data, the numerical analysis presented is a case studyhighlighting the potential of this modeling framework for airportbenchmarking. To further this aim, we scrutinize three specificairports, namely Vienna, Hanover and Lyon, in order to examinethe usefulness of the framework developed. For example, theresults of the under-utilized airport in Hanover indicate that inthe medium-term the airport could reduce runway capacity by24% without changing terminal capacity. Through the slackanalysis, it is also clear that the airport is capable of handlingadditional domestic passenger and cargo throughput, despite thereduction in capacity. Basic DEA results suggest closing one oftheir three existing runways and reducing terminal capacity by35% or attempting to increase cargo throughput by a magnitude offive. The formulations developed are suitable for assessing appro-priate strategies with respect to aeronautical and commercialactivities not only separately but also in combination. Accordingto the combined DEA restricted reference revenue maximizationapproach, Lyon airport has achieved a sustainable level of aero-nautical revenues and ought to search for appropriate commercialrevenue opportunities as opposed to the results from the basicDEA, which suggest a further increase in aeronautical revenues of40%. The methodology provides an airport manager with the toolsfor both exploratory data analysis and inefficiency estimation,removing the need for additional tests of homogeneity. Further-more, utilizing a declared capacity measure both as a terminal andairside proxy of physical capital appears to be new in airportbenchmarking studies. Compared to the standard quantity mea-sures such as the number of runways or gates, this proxy providesan improved managerial measure of the airport infrastructure as asystem and allows us to consider bottlenecks at an airport.

This article is organized as follows. Section 2 presents details ofthe individual modeling formulations that are combined in Section 3

in order to produce airport benchmarks based on a PCA–DEArestricted reference approach. Section 4 provides a description ofthe public data available for analysis and Section 5 compares theresults of the combined formulations to those of basic DEA modelsand benchmarks a select subset of airports in order to demonstratethe utility of the approach developed in this research. Finally, Section6 concludes and presents recommendations for further research.

2. Methodology

DEA is a non-parametric method of frontier estimation thatmeasures the relative efficiency of decision-making units utilizingmultiple inputs and outputs. DEA accounts for multiple objectivessimultaneously without attaching ex-ante weights to each indicatorand compares each decision-making unit (DMU) to the efficient setof observations, with similar input and output ratios, and assumesneither a specific functional form for the production function, northe inefficiency distribution. DEA was first published in [31] underthe assumption of constant returns-to-scale1 and was extended byBanker et al. [32] to include variable returns-to-scale. This non-parametric approach solves a linear programming formulation perDMU and the weights assigned to each linear aggregation are theresults of the corresponding linear program. The weights are chosenin order to show the specific DMU in as positive a light as possible,under the restriction that no other DMU is more than 100% efficient.Consequently, a Pareto frontier is attained, marked by specificDMUs on the boundary envelope of input–output variable space.Formulation (1a) presents an input-oriented DEA model andformulation (1b) presents an output-oriented DEA model, both ofwhich assume variable returns-to-scale.

Minl,y

y

s:t:XN

n ¼ 1

Xni l

nryXai 8i¼ 1,. . .,m

XN

n ¼ 1

Ynj l

nZYa

j 8j¼ 1,. . .,s

XN

n ¼ 1

ln¼ 1

ln,yZ0 ð1aÞ

Maxl,y

y

s:t:XN

n ¼ 1

Xni l

nrXai 8i¼ 1,. . .,m

XN

n ¼ 1

Ynj l

nZyYa

j 8j¼ 1,. . .,s

XN

n ¼ 1

ln¼ 1

ln,yZ0 ð1bÞ

where superscript a is the index of DMUa the unit under investiga-tion for the complete set of N DMUs; Xi

a represents the input i

values of DMUa; Yja the output j values of DMUa; ln the intensity

variables; m the number of inputs and s the number of outputs. yrepresents the relative efficiency score and in the input-orientedmodel a value of 1 indicates efficiency and a value smaller than1 indicates the level by which the relevant inputs ought to bedecreased in order for DMUa to be deemed relatively efficient. Inthe output-oriented model, y is maximized and a value of

y /x

DMU

DMU2

DMU1

DMU

DMU

DMU

y /x

DMU

DMU

DMU

Fig. 2. Benchmark clustering.

N. Adler et al. / Omega 41 (2013) 442–458 445

1 indicates relative efficiency and a value greater than 1 indicatesthe level by which the relevant outputs ought to be increased inorder for DMUa to be deemed relatively efficient. Removing the lastconstraint changes the assumption to constant returns-to-scale andeither orientation may be considered reasonable when benchmark-ing airports depending on the long run or short run nature of theanalysis [9,7,33].

In Section 2.1 we discuss the restricted reference mechanismthat ensures comparable benchmarks are chosen from a datasetgiven exogenous parameter values. In Section 2.2 we present thenetwork DEA model designed to disaggregate the process ofdecision-making within a unit. Subsequently, in Section 2.3 wediscuss the combination of principal component analysis and dataenvelopment analysis, which reduces over-estimation bias and inSection 2.4 we present a multi-dimensional scaling approach thatproduces a graphical representation of the data. Finally, in Section2.5, we discuss a non-parametric statistical procedure that mea-sures efficiency variation across different groups within thedataset in order to estimate the potential impact of environmen-tal variables on the relative Pareto efficient frontier.

2.1. Restricted best practice frontier

Basic DEA benchmarking may lead to inappropriate targets forimprovement in a dataset in which there are substantial differ-ences in size among the DMUs under analysis. Sarkis and Talluri[34] propose second-stage clustering to identify benchmarks forpoor performers, after applying DEA to determine the relativeefficiencies of airports. This study applies a dynamic clusteringapproach first proposed by Golany and Thore [27] that restrictsthe selection of best practice DMUs according to predefinedboundaries within the DEA framework in a single stage process.The boundaries of the cluster are defined in relative terms, limit-ing the efficient reference set2 to those DMUs whose input–outputvalues are within the distance defined by the proportions.

In Fig. 2 we demonstrate the impact of the cluster restrictionsfor a simplified model with two outputs and a single input. DMUa

is compared to the Pareto frontier (darker line defined by DMUs3 to 6) in a standard DEA formulation, with DMUs 4 and 5 actingas benchmarks for DMUa. In our proposed approach, each ineffi-cient airport may refer to a set of benchmarks that do not liedirectly on the Pareto frontier, rather within the dotted radius. IfDMUa lies far enough away from the Pareto frontier as shown inFig. 2, all potential benchmarks may lie in the interior of theenvelope, resulting in DMUs 1 and 2 acting as benchmarks forDMUa. The assumptions of this approach lead to the conclusion

2 A reference set, or peer group, is defined by a subset of units ‘‘closest’’ to the

unit under evaluation i.e. with similar mixes of inputs and outputs [27].

that DMUa, the hypothetical observation lying on the interiorfrontier, represents a relevant target which is more accessible thanDMUa in the short to medium term. It should be noted that asDMUa is not compared to the overall Pareto frontier, no inferencecan be made with regard to the economic efficiency of this unit asdefined under basic DEA. Instead, the motivation of this model isto find appropriate benchmarks and short to medium term targetsin order to improve airport performance. The restricted clustersrepresent an improvement over the Sarkis and Talluri [34] two-stageprocedure since additional information, such as the importance ofeach target DMU, can be drawn from the one-step procedure.

2.2. Network DEA

Network DEA models were first introduced by Fare [23] andFare and Grosskopf [35,36] and subsequently extended by Lewisand Sexton [37], Emrouznejad and Thanassoulis [38], Golany et al.[39], Chen [40], Kao [41] and Tone and Tsutsui [24]. Opening theblack box permits an analysis of the optimal production structureof DMUs and their priorities, to determine both efficient sub-systems and overall efficiency in order to allocate resourcesefficiently and determine appropriate targets. Castelli et al. [42]provide a classification of DEA models accounting for the internalstructure of DMUs, depending on the assumptions of the model-ing approach and then present mathematical formulations, exten-sions and applications. Fig. 3 demonstrates a network structure inwhich the outputs of some decision making sub-units (DMSU)become inputs for other sub-units, connected serially or inparallel. This framework has been widely applied in manufactur-ing production systems and supply chains. In the transportationliterature, network DEA has been applied recently by Yu and Lin[43] in order to simultaneously estimate passenger and freighttechnical and service efficiency for 20 selected European railwaysand by Yu [16] in order to analyze the operational framework of15 domestic airports in Taiwan.

This research develops a network model that defines a multi-product airport in which capital, labor, materials and outsourcingof services produce traffic volume, in the form of aircraft move-ments, passenger and cargo. This throughput then generatesrevenues from aeronautical charges, paid by airlines per departingpassenger and per landing aircraft, and from commercial term-inal-side services serving passengers. The overall profits of thissystem are driven by services provided by outside parties includ-ing airlines and third party contractors as well as the airportprocesses themselves. Airport management retains reasonablecontrol over labor, materials and levels of outsourcing but limitedcontrol over capital investments. In addition, management con-trols the variety and the pricing policies offered on the non-aeronautical side and (partially) controls aeronautical tariffs,dependent on the regulatory regime of the relevant country.Network DEA lends itself to a more accurate description of thisprocess than standard performance analyses. In this research,

Fig. 3. Network decision making unit and sub-units [42].

N. Adler et al. / Omega 41 (2013) 442–458446

network DEA describes the production process, demonstratingthe sequential effects separating final and intermediate outputsincluding those under partial managerial control and those thatare known to be non-discretionary. We have not defined multi-component parameters for shared inputs, as suggested in Beasley[44] and Cook et al. [45], because the joint resources (costs andairport capacities) could not be separately assigned to intermedi-ate services (passengers, cargo and air traffic movements) andcannot be endogenized due to a causal, non-linear relationshipbetween the overlapping intermediate outputs. Consequently, ourbenchmarking framework does not determine an aggregate effi-ciency score per airport due to airport management controlrestrictions. Instead we concentrate on cost minimization givenintermediate and final outputs and separately on revenue max-imization given the inputs and intermediate output levelsrequired from the system.

2.3. Principal component analysis integrated with DEA

The results of the DEA model may not sufficiently distinguishbetween efficient and inefficient DMUs due to an overestimationbias caused by the curse of dimensionality [30]. PCA–DEA is oneof the methodologies that have been developed to reduce thenumber of inefficient DMUs incorrectly classified as efficient[29,46]. The original variables are replaced with a smaller groupof principal components (PCs), which explain the variance struc-ture of a matrix of data through linear combinations of variables.The principal components are uncorrelated linear combinationsranked by their variances in descending order and those thatexplain little of the variance of the original data may be removedthus reducing the dimensions in the DEA linear program. In orderto use principal components instead of the original data, the DEAmodel needs to be transformed to take into account the linearaggregation. Nataraja and Johnson [47] argue that, unlike othervariable selection techniques, PCA–DEA incorporates some levelof information from each of the original variables and performswell with highly correlated variables (greater 0.8) and small datasets (less than 300 observations). A rule-of-thumb computed in[30] suggests dropping PCs one-by-one until a reasonable level ofdiscrimination is achieved, but not less than 76–80% of theinformation should be retained in the model in order to minimizethe overestimation bias.3 Clearly, if we use less than full informa-tion, we will lose some of the explanatory powers of the data butwe will improve the discriminatory power of the model. It shouldbe noted that as a result of the free sign in the principalcomponent analysis and the transformed constraints in thePCA–DEA model, the targets and benchmarks obtained couldreflect a change in the current mix of input–output levels of theinefficient DMUs, in a similar fashion to that of weight con-strained DEA. In this research, PCA–DEA is applied in order toensure that the set of benchmarks is robust. Target estimationsfor inefficient DMUs are based on the original data drawn fromthe benchmarks that have been determined consistently rela-tively efficient per cluster according to the results of the PCA–DEAformulation.

2.4. Visualizing multiple dimensions

Co-Plot, a variant of multi-dimensional scaling, aids both inexploring the raw data and in visualizing the results of DEA[48,49]. Co-Plot positions each decision-making unit in a two-

3 The rule-of-thumb defines the percentage of retained information required

to balance the trade-off between the two incorrect definitions of (in)efficiency,

namely efficient decision-making units defined as inefficient (under-estimation)

and inefficient DMUs defined as efficient (over-estimation).

dimensional space in which the location of each observation isdetermined by all variables simultaneously according to a corre-lation analysis. The graphical display technique plots observationsas points and variables as arrows, relative to the same arbitrarycenter-of-gravity. Observations are mapped such that similarDMUs are closely located on the plot, signifying that they belongto a group possessing comparable characteristics and behavior.A general rule-of-thumb states that the picture is statisticallysignificant if the coefficient of alienation is less than 0.15 and theaverage of correlations is at least 0.75.4 We apply Co-Plot to theset of variable ratios (each output divided by each input), in orderto align the technique with DEA, such that Co-Plot graphicallydisplays the DEA results in two dimensions. In general, theefficient DMUs appear in the outer circle of the plot signifyingtheir relative achievements and we exogenously determine thecolor of the DMUs in order to clarify the results of the DEA.

2.5. Measuring variation across groups

In order to determine whether there are distinct differencesbetween groups of airports, we apply the program evaluationprocedure outlined in [50,51]. Four steps are required to imple-ment the procedure. In the first step, the complete set of DMUs(n¼1,y,N) are split into k sub-groups and the model is runseparately over each of the k groups. In the second step, for eachof the k individual groups, the inefficient DMUs are moved totheir hypothetical efficient level by projecting them onto theefficient frontier of their relevant group. In the third step, a pooledDEA is run with all n DMUs based on their adjusted variables. Inthe fourth step, a Kruskal–Wallis test is applied to determine ifthe k groups possess the same distribution of efficiency valueswithin the pooled set. If the null hypothesis is correct, we expectto see most of the DMUs from all groups rated as efficient in stepthree. Note that in order to avoid inaccuracy in the nonparametricrank test, the number of observations in each of the k subgroupsshould be of similar size. If this is not the case, the size of thesmallest subgroup is calculated and simple random samplingwithout replacement is applied to form subgroups of equallysmall sized samples.

3. Model formulations

In this section we describe four formulations with restrictedreference sets that are then applied in Section 5. Given the publicdata available for the case study, Fig. 4 presents the airportnetwork technology that we analyze based on a subset ofavailable variables described in Fig. 1. X represent inputs, Y

outputs and I intermediate products. The number in bracketsrepresents a node index in the network.

Model (2) assumes that the airport management is interestedin maximizing revenues, drawing from aeronautical and non-aeronautical activities given airport throughput on the terminaland airsides, which are in turn limited by the physical infra-structure and associated costs available to support the system.Drawing on discussions with airport managers and Pels et al. [7],we assume variable returns to scale with respect to revenues.Larger airports that provide the infrastructure for airport cities,including shopping malls, hotels and other commercial activities,may have relatively greater opportunities to collect revenues than

4 The coefficient of alienation is a single measure of goodness-of-fit for the

configuration of n observations obtained from a smallest space analysis [52]. The

higher the correlation, the better the common direction and order of the

projections of the n points along the arrow. The length of the arrow is proportional

to the correlation.

Inputs (1):Staff costs XOther operating costs XRunway capacity XTerminal capacity X

Intermediate goods (2):International passengers IDomestic passengers I

Intermediate goods (3):Tons of cargo IATM I

Output (4):Non-aeronautical revenues Y

Output (5):Aeronautical revenues Y

Fig. 4. Two-stage airport network technology.

N. Adler et al. / Omega 41 (2013) 442–458 447

their smaller counterparts. The two-stage formulation for theradial, output-oriented, variable returns to scale, linear programapplied in this research is presented in model (2).

Stage I: Restricted References

Na12 � n9alr

Xnj

Xaj

rau 8j¼ 1,. . .,4, blrInk

Iak

rbu 8k¼ 1,2

( )

Na13 � n9alr

Xnj

Xaj

rau 8j¼ 1,. . .,4, blrInm

Iam

rbu 8m¼ 3,4

( )

Stage II: Linear Program

Maxl,y

y1þy2

s:t:X

nANa12

Inkl

n24r Ia

k 8k¼ 1,2

XnANa

12

Yn1l

n24Zy1Ya

1

XnANa

12\Na13

Inj l

n235r Ia

j 8j¼ 1,. . .,4

XnANa

12\Na13

Yn2l

n235Zy2Ya

2

XnANa

12

ln24 ¼ 1

XnANa

12\Na13

ln235 ¼ 1

y1,y2,ln24,ln

235Z0 ð2Þ

where superscript a is the index of DMUa, the unit underinvestigation; Xa represents the input values of DMUa; Ya and Ia

are the output and intermediate values of DMUa respectively andthe subscript of intensities for DMUa, ln

ij, denotes the linkconnecting nodes i and j in the network presented in Fig. 4.y1 and y2 represent the relative performance for the commercialand airside activities respectively, where a value of 1 indicates aposition on the frontier and a value greater than 1 indicates byhow much the relevant revenues ought to be increased in orderfor DMUa to move towards the frontier. Stage I restricts theairport comparators to those that possess similar input andintermediate output levels and in the case study described inSection 4, they are bounded between 10% and 300% of DMUasinputs and between 20% and 200% of DMUas intermediate out-puts. The parameter values are chosen to ensure that a reasonableset of comparators exist in the cluster. It should be noted that

formulation (2) is always feasible provided that the lower bounds(al and bl) are less than one and the upper bounds (au and bu)greater than one, thus ensuring that the airport acts as a self-comparator at the very least because aANa. Sensitivity analyses ofour current dataset suggest that smaller bounds result in exces-sive limitations and pure self-comparisons over time whereas awider set leads to unreasonable benchmarks whereby LondonHeathrow (LHR) and Tallinn (TLL), representing the largest andsmallest airports in the dataset, are considered directly compar-able. We assume that bodies such as the International CivilAviation Organization (ICAO) or Airports Council International(ACI) would be able to utilize the methodology and apply smallerbounds to their more substantial datasets. The first stage cluster-ing approach may also include additional restrictions that arise asa result of the varying environments in which the DMU may belocated.

ln24 connects the number of passengers to non-aeronautical

revenue derived and ln235 connects passengers, cargo and air

traffic movements to aeronautical revenue. Since no trade-offbetween aeronautical and non-aeronautical activities is intro-duced in the model, benchmarks on each side of the airportactivity are determined independently. An equivalent one-stageformulation of model (2), which is in the form of a mixed integerlinear program, is presented in Appendix A.

Alternatively, whilst it may be assumed that a private, unre-gulated airport pursues profit maximization, airports that aresubject to economic regulation may behave as social welfaremaximizers. Hence, maximizing aeronautical revenues may notbe the target of airport management due to regulatory constraintssuch as price capping. Furthermore, even profit maximizers mayconsider lower aeronautical charges as an opportunity to expandnon-aeronautical activities and generate additional commercialrevenues by attracting airlines through lower airport charges.Consequently, the DEA formulation for the radial, output-oriented, variable returns-to-scale, linear program, combiningboth aeronautical and concession activities, is presented in (3).The goal is to maximize non-aeronautical revenue (Y1) giveninternational and domestic passengers, cargo and air trafficmovements. Physical infrastructure (declared terminal and run-way capacity), costs (labor, outsourcing and materials) and inter-mediate outputs define the reference set for each DMU as inmodel (2). Aeronautical revenue (Y2) is included in the analysis asa non-discretionary variable [28]. According to this model, bench-marks consist of airports achieving higher non-aeronautical

N. Adler et al. / Omega 41 (2013) 442–458448

revenues, despite being of similar size and serving similar levelsof demand. In the following we will refer to (2) as the indepen-dent model in which the clusters were independently defined andthe performance of the aeronautical and non-aeronautical sideestimated separately. Formulation (3) presents the combinedmodel since a common set of benchmarks are considered butonly non-aeronautical revenues are maximized.

Stage I: Restricted References

Na� n9alr

Xnj

Xaj

rau 8j¼ 1,. . .,4, blrInk

Iak

rbu 8k¼ 1,. . .,4

( )

Stage II: Linear Program

Maxl,y

y1

s:t:X

nANa

Yn1l

nZy1Ya

1XnANa

Yn2l

nZYa

2XnANa

Inj l

nr Iaj 8j¼ 1,. . .,4

XnANa

ln¼ 1

y1,lnZ0 ð3Þ

The formulation for the radial, input-oriented, variablereturns-to-scale, linear program proposed in this research ispresented in formulation (4). Stage I clusters airports accordingto revenue and traffic mix and the parameter values al, au, bl, bu,wl and wu are set such that aANa in order to guarantee thefeasibility of DEA model in Stage II and balance the trade-offs.

y represents the relative efficiency score, whereby a score of1 means that the airport lies on the (interior) frontier and lessthan 1 indicates the level of the controllable or discretionaryinput retraction required in order to reach the frontier defined bythe relevant benchmarks. As developed in Banker and Morey [28],we require that the composite reference group utilize no more ofthe non-discretionary resource than the unit under evaluation(DMUa). As a result, the first constraint in model (1) is replaced bytwo constraints in model (4), in which k is an index of terminaland runway capacities. We assume that the capacity of an airport,applied as a proxy for capital, is partially controllable and may beadjusted to a certain extent in the medium term. Consequently,the model restrictions permit terminal and runway declaredcapacity to decrease up to a pre-determined level according tothe third constraint of model (4). Since d impacts the choice ofbenchmarks, it will indirectly limit the value of the objectivefunction, y, via the first set of constraints.

The cost minimization linear program assumes variablereturns to scale, since a doubling of output may not necessarilyrequire a doubling of staff, materials and outsourcing costs. Byremoving the last constraint in (4) in which the sum of lambdasequals one, the formulation may be easily adapted to a constantreturns-to-scale approach.

Stage I: Restricted References

Na� n9alr

Ynj

Yaj

rau 8j¼ 1,2, blrInk

Iak

rbu 8k¼ 1,2,4, wlrIn3

Ia3

rwu

( )

Stage II: Linear Program

Minl,y,S

y

s:t:X

nANa

Xnj l

nryXaj 8j¼ 1,2

XnANa

Xnkl

nrXak 8k¼ 3,4

XnANa

Xnkl

nZdXa

k 8k¼ 3,4

XnANa

Inql

nZ Ia

q 8q¼ 1,. . .,4

XnANa

ln¼ 1

y,lnZ0 ð4Þ

Finally, we may need to replace the original variables withprincipal components, in order to aggregate the data should therebe problems with the level of discrimination in the results. ThePCA–DEA formulation of model (4) is presented in the followingequation:

Minl,y,S

y

s:t:X

nANa

PCnj l

nþX2

i ¼ 1

lijSi1 ¼ yPCa

j 8j¼ 1,2

XnANa

PCnkl

nþX2

i ¼ 1

likSi2 ¼ PCa

k 8k¼ 1,2

XnANa

PCnkl

n�X2

i ¼ 1

likSi3 ¼ dPCa

k 8k¼ 1,2

XnANa

PCnql

n�X4

i ¼ 1

liqSi4 ¼ PCa

q 8q¼ 1,. . .,4

XnANa

ln¼ 1

y,ln,S11,S2

1,S12,S2

2,S13,S2

3,S14,S2

4,S34,S4

4Z0 ð5Þ

where S represent slack variables and l are normalized eigenvec-tors based on costs, capacities and intermediate outputs. Theprincipal component PCj combines staff costs and other operatingcosts. The principal component PCk combines terminal and run-way declared capacities. PCq combines domestic and internationalpassengers, ATM and cargo. The PCA–DEA formulation (5) isexactly equivalent to the original linear program (4) if and onlyif the PCs explain 100% of the variance in the original input andoutput matrices, i.e. j¼k¼1,2 and q¼1,y,4 in model (5).Decreasing the dimensions of the model such that jo2, ko2and/or qo4, given the current network described in Fig. 4, leadsto a reduction in the set of peer DMUs. For example, if we retainprincipal components explaining at least 95% of the originalinformation, the comparator set drops from 49% to 39% in ourdataset. If we retain at least 85% of the original information, thecomparator set drops to 26%, which improves the discriminatorypower of the modeling approach substantially.

4. Dataset

The dataset consists of 43 European airports located in 13different countries. In order to increase the likelihood of compar-ability, we focus the case study on European airports. In Europe,passenger terminals are normally operated and maintained by theairport operator whereas in the U.S. airports frequently contractout such activities to airlines or the airline builds and owns aterminal, making cross-comparisons somewhat problematic. Dis-aggregated data for major European airports such as Paris Charlesde Gaulle, Barcelona or Madrid, all of which belong to airportgroups or authorities, are not available because financial data isonly reported across the entire group. Consequently, in order tomaximize our dataset, we have pooled the data to an unbalancedset of 294 observations covering the time period from 1998 to2007 and include all European airports for which financial,capacity and traffic data are publicly available. Appendix B liststhe set of airports under study, the specific timeframe for whichthe data was available and whether ground handling processesare undertaken in-house or outsourced. We identify an increase in

N. Adler et al. / Omega 41 (2013) 442–458 449

costs beyond the changes in traffic levels over time, indicating atechnological retraction possibly caused by the security situation,however due to data limitations, an intertemporal analysis iscurrently beyond the scope of this research. All airports offerdomestic and international routes, however airports located insmaller countries such as the Netherlands generally serve veryfew domestic destinations. The passenger volume varies consid-erably from less than a million passengers at Tallinn and DurhamTees Valley airports up to more than 60 million at LondonHeathrow, the largest European airport in terms of passengerthroughput and number three in the world [53].

Ten variables were collected in total for purposes of analysisbased on publicly available data. The variables are categorizedinto three groups: four inputs (X), four intermediate products (I)and two outputs (Y). Table 1 presents summary information andspecifies the data sources. The variable inputs consist of staff costsand all other non-labor related operating costs, which includesupplies, materials and outsourcing. Although a smaller airportthan London Heathrow in terms of air traffic movements, Frank-furt’s staff costs are highest due to the level of vertical integrationwhereby the airport operates most of the services in-house orthrough wholly-owned subsidiaries. As an example, the airportmanages the ground handling operations which represent one ofthe most labor intensive activities at an airport, a processtraditionally organized by airlines or independent third partyproviders at Heathrow. Consequently, Heathrow’s other operatingcosts are highest reflecting the high levels of outsourcingundertaken.

In the literature to date, physical data such as the number ofrunways, gates, check-in counters and overall terminal size iscollected as a proxy for capital. However, such data is oftenproblematic because the number of runways does not includeinformation on their configuration or the impact of weather onthe number of runways open within a given timeframe. Further-more, the terminal area in square meters is somewhat subjectivesince some airports report gross terminal area including sections ofan airport that are not open to the public. If the dataset covers morethan one country, the monetary measurement of physical capitalcreates difficulties due to different national accounting standardsand depreciation methods or periods across and within countries.For example, the airports of the British Airports Authority (BAA)depreciate their runways over 100 years whereas the airports

Table 1The list of variables in the airport performance analysis.

Variable Description Name Averag

Staff costs Wages and salaries, other staff

costs (PPP US$)

X1 81

Other operating costs Costs of materials, outsourcing

and other (PPP US$)

X2 103

Declared runway capacity Number of total movements

per hour

X3

Terminal capacity Number of total passenger

throughput per hour

X4

International passengers Annual passenger volume I1 1

Domestic passengers Annual passenger volume I2 2

Cargo Metric tons (trucking excluded) I3

Air transport movements Number of total commercial

movements

I4

Non-aeronautical revenues Revenues from concessions of

own retail and restaurants,

rents, utilities and other (PPP

US$)

Y1 117

Aeronautical revenues Landing, passenger and aircraft

parking charges; revenues from

ground handling, cargo

revenues and other (PPP US$)

Y2 175

operated by the Aeroports de Paris depreciate over a period of 10to 20 years [54]. In this research, terminal capacity is defined interms of passengers handled within an hour, thus combining thecapacities of all terminal facilities including check-in counters,security controls, baggage delivery and retail area into one commoncapacity figure. The airside is defined by the declared runwaycapacity, specified as the number of departing and arrival move-ments permitted per hour. The declared per hour capacities that wehave applied include the impact of night curfews and the differentoperating hours of the airport. Yearly instead of standardized hourlyfigures could be applied, but this will not impact the results becauseof the units invariant properties of the models. Airport stakeholdersnegotiate the runway capacity measure every six months and it isprimarily used to avoid congestion at schedule facilitated airports aswell as aid in the allocation of slots at coordinated airports [55]. Theadvantage of using declared capacity is that the parameters accountfor bottlenecks across the terminal and runway systems, providingtwo individual capacity measures. Amsterdam possesses the highestdeclared terminal and runway capacities in our sample with 26,000passengers and 110 movements per hour. Due to their geographicallocation near the coast, they require a special runway configurationto operate as a hub airport. The smallest airport with respect torunway capacity is Florence in the Tuscany region, with a maximumhourly rate of twelve movements. Due to its short, single runwaysystem (1688 m), the airport can handle aircraft up to the size of aBoeing 737 or an Airbus A319 [56].

The annual traffic volume is represented by the number ofpassengers, commercial air traffic movements and tons of cargo(trucking is excluded). The passengers are divided according todomestic and international destinations only because we could notcollect sufficiently disaggregated data to separate the passengersbetween intercontinental, Schengen and domestic flights or accountfor transfer passengers, which would be preferable since thesegroups probably generate different revenue streams. Aviation rev-enues are generated from (often regulated) landing and passengercharges, ground handling undertaken in-house when relevant andcargo activities. Non-aviation revenues include revenues from retailactivities and restaurants, concessions and income from rents andutilities. The largest non-aeronautical revenues were generated atHeathrow, whereas Frankfurt earned the highest aviation revenues.Commercial revenues equal 67% of total airport revenues onaverage in the dataset, clearly supporting the argument that

e Maximum Minimum Source

,704,359 1,080,756,267 5,962,213 Annual reports

,364,471 725,987,196 5,010,381 Annual reports

49 110 12 IATA [58], airport and

coordinator websites

6768 26,000 450 IATA [58], airport websites

0,300,571 61,517,733 355,579 IATA [58], airport websites

,433,287 9,932,208 48 IATA [58], airport websites

214,076 2,190,461 37 IATA [58], airport websites

152,133 492,569 16,000 IATA [58], airport websites

,906,043 1,107,046,057 4,629,813 Annual reports

,507,645 1,739,331,693 7,199,668 Annual reports

N. Adler et al. / Omega 41 (2013) 442–458450

non-aeronautical activities should not be ignored in a productivityanalysis of airports from a managerial perspective.

All financial data is deflated to the year 2000 by dividing thenominal value by the price index and adjusted by the purchasingpower parity according to the United States dollar in order toensure comparability across countries (the International Mone-tary Fund’s World Economic Database 2006). In addition, the datahas been normalized by the standard deviation to limit theinfluence of outliers in the dataset [57].

5. Results

The results suggest that airports that outsource activities aregenerally more efficient from a cost perspective and that manyairports ought to concentrate on non-aeronautical revenue manage-ment in order to achieve greater throughput and higher revenueefficiency. In the following section we identify the impact of verticalintegration and subsequently include the information in therestricted reference sets. In Section 5.2 we compare and contrastthe results of a basic DEA model (formulation (1)) with those of thenew formulations described in Section 3. Section 5.3 discusses thebenchmarking results for a subset of airports, specifically Viennawhich acts as a benchmark in the current case study, Hanover as anexample of an input-oriented inefficient case (formulations (4) and(5)) and finally Lyon, in order to compare and contrast theindependent output model (formulation (2)), which assesses theperformance estimates of both revenue generating activities sepa-rately, and the combined output model (formulation (3)) in whichonly commercial revenues are maximized.

5.1. Efficiency variation across groups

When estimating the relative efficiencies, it would appear thatairports offering in-house ground handling services operate on a

Fig. 5. Kruskal–Wallis ANOVA for outsourcing. (a) Output-orientation, (b) Input-orienta

the third step of the program evaluation procedure (a score of one implies relative effi

different production frontier to airports that outsource thisactivity. This is not immediately obvious because airports provid-ing ground handling services in-house have higher labor costswhereas airports that outsource the activity have higher ‘other’costs. Both business models achieve higher revenues than airportsthat permit third parties to provide the service since in the lattercase no costs appear on the books and only minor concessionalfees may be collected because the contracts themselves do notappear on the airport’s accounting books. In order to analyzewhether airports choosing to provide such services lie on adifferent production frontier than those which do not, the non-parametric program evaluation procedure was applied using abasic DEA model without reference restrictions, in which wecombine both revenue sources into a single output variable (non-aeronautical and ground handling revenues combined). The out-put oriented DEA model (1b) includes nodes {2345} with Ii

representing inputs (i¼1,y,4) and Yj outputs (j¼1,2) accordingto Fig. 4. In the input oriented DEA model (1a) we include nodes{123}, with Xi representing inputs (i¼1,y,4) and Ij outputs(j¼1,y,4). In our sample, 21 airports offer ground handling in-house and 22 outsourced or never offered this service whichtranslates into 156 DMUs in the ground handling group and 138DMUs otherwise (see Appendix B for details). Graph (a) in Fig. 5presents the DEA efficiency scores on the vertical axis for the twogroups from a revenue maximization perspective and (b) showsthe DEA efficiency scores from the cost minimization perspectiveacross the two groups. The results are clear and significant thatairport operators providing ground handling appear to be revenuemaximizers but were highly inefficient in cost minimization relativeto airports from the non-ground handling group. Airports withground handling activities perform on average 10% better in max-imizing their outputs as their aeronautical revenues per passengerare higher whereas in the input-oriented model, airports that do notprovide ground handling achieve on average 10% higher efficienciessince no costs are associated with this service.

tion. (The vertical axis of the boxplot represents the efficiency scores computed in

ciency)).

N. Adler et al. / Omega 41 (2013) 442–458 451

The efficiency variation across groups also proved significantin the basic DEA model in which air traffic movements, passen-gers, cargo and commercial income (including ground-handlingrevenues) were selected as output in order to adjust for the effectof outsourcing (Ij, j¼1,y,4 and Y1). Staff and other operating costsand runway and terminal capacities were defined as inputs (Xi,i¼1,y,4). Having now considered both costs and revenues in theefficiency estimation, the radial variable returns-to-scale, input-oriented model still indicated significant efficiency differencesacross both groups based on the results of a Kruskal–Wallis test.After liberalization in 1996, airports handling more than twomillion passengers per annum and providing ground-handling in-house were required to permit access to competing firms. Munichand Frankfurt airports have consistently claimed substantiallosses in this segment [59,60].5 However, the strong labor unionsin Germany have prevented airport management from eithercutting wages or outsourcing this service to third-party providerswithout guarantees that workers would continue under the sameconditions. Thus, at least in the short-term, the degree of out-sourcing can be regarded as a political factor that is beyondmanagerial control and ground handling is included in the net-work DEA formulation as an environmental variable which willfurther limit the potential benchmark set via clustering. Conse-quently, we analyze airports that offer ground handling servicesseparately from airports outsourcing this activity to airlines orindependent providers, in turn limiting comparisons to airportsoperating under similar managerial strategies.

5.2. Comparison between basic DEA and the proposed formulations

In contrast to formulations (2)–(5), basic DEA neither restrictspotential benchmarks nor does it permit a limited deviation inone or more variables. In order to compare the modelingapproaches in a fair and consistent manner, the dataset was splitaccording to in-house or outsourced ground handling provisionand DEA was applied individually to each category. For the outputorientation, the technology of Fig. 4 reduces to nodes {235} (inwhich Ii represents inputs (i¼1,y,4) and Y2 output) and {24} (inwhich Ii represents inputs (i¼1,2) and Y1 output) with respect toaeronautical and commercial activities respectively, whilst theinput orientation case collapses to nodes {123} (in which Xi

represent inputs (i¼1,y,4) and Ij outputs (j¼1,y,4)). The basicDEA results generate consistently efficient airports that belongeither to the set of smallest airports e.g. Bremen, Florence andLjubljana, which provide ground handling in-house and Malta,Durham Tees Valley and Leeds/Bradford which outsource, or thelargest airports in the sample such as Frankfurt. In none of theoutput-oriented formulations do airports achieve 100% efficiencyover the entire review period, although Salzburg, Ljubljana andMalta appear consistently close to the frontier. A notable excep-tion is Cologne–Bonn, which remains cost efficient over time.Cologne–Bonn is the European hub for the parcel service providerUPS, which rents office space and warehouses from the airport,suggesting a behavior different to others in the sample [61].

Under basic DEA, all airports are compared against a singlePareto frontier and Salzburg represents an important benchmarkfor Vienna, Dusseldorf, Frankfurt, Hamburg and Munich. How-ever, it is doubtful that the management of a primary or

5 Most German airports are fully or at least majority publicly owned and if

ground handling is operated in-house by the parent company, the airport pays

salaries based on public tariffs, which are on average 20% higher compared to

those of private ground handling companies. Some German airports, such as the

minor-private airport Hamburg, outsourced the ground handling segment to a

100% subsidiary in order to set flexible tariffs; however this was not deemed

acceptable by the public shareholders of Munich airport for example.

secondary hub airport would adopt the strategies of an airportthat handles less than 2 million passengers per year with very lowcargo throughput too. Durham Tees Valley, a small airport in EastEngland with less than 700,000 passengers per year, was definedas a benchmark for Lyon, Geneva, Oslo and the hub airport inZurich, which does not occur in the formulations we present inSection 3 due to the restricted reference approach.

Airports in very small clusters are unique in character and inthe extreme case tend to form their own reference set. In thecurrent dataset, these mostly included the smaller and lesscongested airports such as Tallinn, Leeds/Bradford and DurhamTees Valley. These airports can be identified as outliers accordingto the Andersen and Petersen [62] super efficiency procedure[63,64]. Such observations frequently influence the basic DEAPareto frontier, for example Durham Tees Valley appears withinthe reference sets of Oslo and Zurich airports. In comparison, theresults of the cost minimization formulation (4) categorizesLondon Stansted and Copenhagen as peer airports for Oslo andZurich hence the modeling approach indicates benchmarks thatare more homogeneous in character. Another unique exampleincludes Dortmund, which acts as a self benchmark in the costminimization approach from 2003 to 2007, namely after itscapacity expansion and severe reduction in cargo operations.Dortmund is the only airport that exhibits operational losses overthe entire timeframe. The airport is partly owned by the localelectricity distributor and losses are covered by their majorshareholder [65]. Dortmund is located in the Ruhr area with apopulation of more than five million, representing the largestagglomeration in Germany. Airport competition includes Dussel-dorf, Cologne–Bonn and Paderborn which are located in theircatchment area (defined as 90 km around the airport) and inter-modal competition includes high speed rail and the motorway,especially on domestic routes and traffic originating in Benelux.Hence despite their high capital investment, it may be necessaryfor the airport to further decrease their aviation charges in orderto attract airlines and new destinations thereby generating addi-tional commercial revenues.

The network DEA formulations provide the user with anexploratory data analysis that does not exist in the basic DEAresults. The results of formulations (2) and (4) demonstrate thatthe average restricted reference set (cluster size) per inefficientairport was reasonably small because the capacity of airportsvaries considerably across the sample and some airports sufferfrom low utilization rates whereas others are highly congested.The operating costs at highly congested airports were largemostly due to employee costs hence airports with similar capa-cities did not necessarily belong to the same cluster. In general,large clusters indicate that various airports in the sample possesssimilar characteristics which in our dataset included Dusseldorf,Hamburg, Stuttgart, Venice and London Gatwick in the output-oriented model (2) and Manchester and London Stansted in theinput-oriented model (4).

5.3. Benchmarking airports

In this sub-section, we describe the type of results and analysisthat are attainable by applying the formulations described inSection 3. We focus on relatively productive Vienna, relativelyinefficient Hanover and finally Lyon, in order to describethe potential balance between the two revenue streams. Viennais an example of an airport that has gradually improved withrespect to both the cost input and revenue output efficiencyestimates over time. Vienna appears in the reference set ofCologne–Bonn and Dusseldorf in the input-oriented case (models4 and 5). Between 1998 and 2007 Vienna’s costs and revenuesincreased on average by similar proportions (99% and 94%

Fig. 6. Co-plot graphic display of Vienna’s input-oriented strategy. (Coefficient of alienation is 0.06 and average of correlations is 0.87).

N. Adler et al. / Omega 41 (2013) 442–458452

respectively),6 while traffic volume grew by 76% for total passen-gers, 54% for air traffic movements and 37% for cargo. The input-oriented case in Fig. 6 shows that from 1998 to 2003, Vienna liesclose to the arrows that display the ratio of intermediate outputsto costs. In 2004, both staff costs and other operating costssubstantially increased partly due to the introduction of a 100%hold baggage screening policy and the founding of a subsidiary forinfrastructure maintenance [66]. After 2004 greater emphasis hasbeen allocated to the issue of runway utilization, viewed in Fig. 6by the proximity of the later years to the capital asset relatedratios. Vienna airport moves in a positive direction towards animproved utilization of the runways which increased from 48% to66% between 2000 and 2007. Hence, despite substantial costincreases, the airport still managed to improve its relativeperformance as costs per air traffic movement decreased overtime. On the aeronautical output side, the airport managementchanged their regulated tariff structure by increasing passengercharges from an average price of 3.90h in 1999 to 5.90h7 in 2007and decreasing average landing charges by 3%, whilst reducingthem for larger aircraft by up to 20%. In summation, Vienna

6 Staff costs increased by 111% and other operating costs by 88%. Non-

aeronautical revenues increased by 78% and aviation revenues by 109%.7 These values are an average passenger price and were computed by dividing

total passenger revenue charges by passenger throughput as obtained from the

annual reports, hence could deviate from the passenger charges specified in the

charges manual. Both prices are stated in nominal values.

airport increased passenger charges out of the total aviationrevenues collected from 33% in 2001 to 46% in 2007 anddecreased the share of landing fees from 44% to 28% over thesame period [66]. These managerial policies have aided Viennaairport’s revenue productivity.

It should be noted that Vienna’s potential cost reference set isrestricted to 15–25 DMUs, as a result of which Vienna iscategorized as efficient for the entire decade except for 2003.Since the first principal component with respect to costs explains98% of the information, with respect to capacities 59% and withrespect to intermediate outputs 93%, the cost and intermediateoutput variables were replaced with one principal componentrespectively as in formulation (5). As a result, Vienna’s airportappears to be robustly efficient for the years 2006 and 2007among others, which highlights the positive path that the airportmanagement have achieved over time (the light observations inFig. 6 represent the efficient years according to PCA-DEA).

Hanover airport is the ninth largest airport in Germany,handling 5.6 million passengers in 2008 [53]. It is partly ownedby the Fraport AG Company which owns and operates Germany’sgateway hub in Frankfurt/Main. For al¼bl¼wl¼0.1, au¼bu¼3 andwu¼3.5, Hanover’s potential reference set includes clusters of 59–66 DMUs including comparable airports in Bologna, Bremen,Dresden, Florence, Genoa, Hanover, Leipzig, Nuremberg, Stuttgartand Venice. For larger parameter values, the reference clusterincludes 156 DMUs i.e. all airports in the sample that provideground handling services in-house (Appendix B). Consequently,

Fig. 7. Current and target input values for cost minimization models with respect to Hanover.

Fig. 8. Catchment area of Hanover airport (within a 2 h drive).

Source: adapted from ADV [71]. Language: German.

N. Adler et al. / Omega 41 (2013) 442–458 453

we directly restrict the set of comparators in Stage I and throughthe partially-discretionary capacity variables, we restrict theprojection in the medium term. As d ranges from 0.7 to 0.999,the categorization of efficient DMUs does not change but thetarget values and implementation path to the Pareto frontier forHanover does.8

Whilst Hanover’s non-aeronautical revenues from rents andutilities increased over the decade analyzed due to the develop-ment of a large airport city, the relative cost performance yconsistently dropped from 100% in 2000 to 65% in 2007 accordingto model (4) and to 55% according to model (5) when at least 95%of information was retained.9 Hanover’s most important bench-marks (based on l in models (4) and (5)) include Stuttgart (1998)and Hanover (2000) for all periods, which strongly suggests thatthe airport has been moving in the wrong direction over the lastdecade. Over the same time period, passenger volume slightlyincreased by 2%, air traffic movements decreased by 9% and cargodropped by 23% however staff costs and other operating costsincreased by 41% and 95% respectively, as shown in Fig. 7. Fromthe technical perspective, Hanover clearly has the capability toserve much larger throughput due to a declared runway capacityof 60 movements per hour. Capacity utilization at Hanoverremained at a stable 19%, whereas Stuttgart achieves 56% due inpart to an extensive business based catchment area [67]. Compar-ing the results for Hanover with respect to model (4) and basicDEA, it becomes clear that the long term goals for Hanover, ceterisparibus, would be to close a runway.

8 Targets presented in Fig. 7 are calculated for d¼0.7.9 Hanover’s cluster is stable over time, a common set of benchmarks exists

between 1998 and 2007 (Stuttgart, Bremen, Hanover in 2000 and Venice) and a

comparison of efficiency scores over time is possible. According to basic DEA, the

efficiency estimate y decreased from 92% in 2000 to 65% in 2007 and the

benchmarks include Bremen, Dusseldorf, Tallinn, Vienna, Stuttgart and Venice.

It should be noted that Hanover’s management may find itdifficult to increase capacity utilization due to the airport’s highlycompetitive location. The airport faces direct competition fromHamburg and Bremen that are in close proximity, as well asDortmund, Paderborn and Munster–Osnabruck, which primarilyserve charter and low cost carriers and are less than a twohour drive by car, as shown in Fig. 8. Potential competitionincludes regional airports located in Braunschweig–Wolfsburg,

Table 2Output benchmarks for Lyon airport.

Airport Domestic

passengers

(yearly PAX)

International

passengers

(yearly PAX)

Movements Cargo

(tons)

Aeronautical

revenues

(PPP US$)

Non-aeronautical

revenues

(PPP US$)

DMUs under review

LYS1998 2,478,508 2,742,712 106,170 39,749 25,552,145 33,325,320

LYS1999 2,565,033 2,935,515 116,894 39,050 29,032,107 37,403,371

LYS2000 2,715,196 3,311,666 129,373 40,126 35,476,207 40,519,260

LYS2001 2,714,678 3,393,929 132,903 38,902 41,348,698 44,483,257

LYS2002 2,523,982 3,254,242 120,529 35,349 48,933,813 44,982,158

LYS2003 2,571,177 3,368,718 118,489 35,494 58,338,026 47,915,432

LYS2004 2,633,962 3,594,650 123,958 34,874 61,171,391 50,077,362

LYS2005 2,682,123 3,879,242 128,868 38,725 68,845,652 55,678,876

Output benchmarks for Lyon airport in 2005

Benchmark Intensity (l) Domestic

passengers

International

passengers

Movements Cargo Aeronautical

revenues

Non-aeronautical

revenues

Short-term benchmark for non-aeronautical activities (model 3, y1¼1.14)

MLH2002 0.47 792,765 2,264,199 88,000 31,285 34,865,786 45,079,574

NCE2006 0.48 4,332,382 5,615,653 164,617 13,940 100,059,952 83,755,849

MRS2002 0.05 3,718,623 2,059,601 86,677 60,025 37,708,043 40,007,062

Medium-term benchmark for non-aeronautical activities (model 2, y1¼1.15)

NCE2006 0.48 4,332,382 5,615,653 164,617 13,940 100,059,952 83,755,849

MLH2002 0.47 792,765 2,264,199 88,000 31,285 34,865,786 45,079,574

GLA2006 0.05 4,575,124 2,629,876 75,000 5,000 62,015,630 52,572,442

Medium-term benchmark for aeronautical activities (model 2, y2¼1.02)

NCE2006 0.51 4,332,382 5,615,653 164,617 13,940 100,059,952 83,755,849

MLH2004 0.44 651,102 1,893,772 57,915 34,227 30,882,599 38,867,805

GLA2000 0.05 3,568,259 3,453,741 92,000 10,000 69,790,936 38,013,125

Long-term benchmark for both activities (basic DEA, y¼1.40)

LCY2002 0.39 417,551 1,187,449 53,000 1,000 38,509,245 12,353,716

MLH2002 0.31 792,765 2,264,199 88,000 31,285 34,865,786 45,079,574

OSL2007 0.19 9,477,511 9,566,489 223,000 97,000 177,975,321 245,588,135

ATH2007 0.08 5,953,814 10,571,571 205,295 119,000 453,224,152 137,319,560

MLH2005 0.03 741,538 2,570,356 62,251 32,215 34,791,663 47,818,329

N. Adler et al. / Omega 41 (2013) 442–458454

Kassel–Calden and Magdeburg–Cochstedt, none of which cur-rently operate commercially although plans exist to offer com-mercial flights from Kassel–Calden in 2012 [68]. Additionalintermodal competition includes the ICE high speed rail alter-native and a highly connected motorway network. In conclusion,Hanover faces direct, potential and intermodal competition hencethe airport needs to cut costs by as much as 40%, further developnon-airport related activities and attempt to attract cargothroughput. The latter strategy may increase runway utilizationand seems reasonable given the high level of connectivity of thecity and the lack of night flights restrictions due to their location.

Our final analysis presents the benchmarking results for LyonSaint-Exupery, which is the fourth largest airport in France, with apassenger throughput of 7.9 million in 2008 [53]. Similar to themajority of French regional airports, Lyon is fully publicly ownedand operated by the regional Chamber of Commerce. Lyon airportbecame a major regional hub airport for the national carrier AirFrance at the end of the 1990s and today, Easyjet is their secondlargest customer [69]. In the independent revenue maximizingformulation (2), Lyon improved aeronautical performance (y2)from a score of 1.95 to 1.02 between 1999 and 2005,10 benefittingfrom a change in the tariff structure similar to that of Viennaairport. The important peer airports over time include privatizedBAA Glasgow and Basel–Mulhouse, both of which focus on lowcost carrier traffic. Glasgow serves Easyjet and Scottish Loganairin competition with Ryanair at Prestwick, and Basel–Mulhouseserves Easyjet, with a market share of almost 50% in 2007 [70].

10 Lyon’s benchmark clusters are stable over time according to formulation (2).

With respect to non-aeronautical activities, Lyon increased itsperformance (y1) from 1.43 in 2000 to 1.15 in 2005, in part due tothe large increase in car parking revenues, rents and utilitieswhich contributed to a 67% increase in overall commercialrevenues (see Table 2). Benchmarks on the non-aeronautical sideinclude Basel–Mulhouse, Nice and Glasgow.

Given that aeronautical revenue maximization is not necessa-rily an optimal policy, irrespective of ownership form, thecombined formulation (3) defines aeronautical revenue as anon-discretionary output and maximizes commercial revenuealone. The results of this model suggest that Lyon’s shorter termcommercial revenue target in 2005 should be $63 million, anincrease of 14%, given current aeronautical revenues. The med-ium-term target (formulation (2)) suggests an increase of 15% innon-aeronautical revenues to increase performance and thelonger term, standard DEA target requires increases of 40% onboth the commercial and non-aeronautical side (Fig. 9). In thecombined model, Basel–Mulhouse and Glasgow appear as impor-tant benchmarks and Nice acts as a benchmark of increasingintensity over the years. As also shown in Fig. 10, Lyon airport ismoving in the direction of Basel–Mulhouse, Glasgow and Nicehence improving in productivity over time.

In summary, Lyon’s aeronautical revenues reached their short-term targets by 2003 and almost achieved their medium-termbenchmarks by 2005 (Fig. 9), suggesting that their landing andpassenger charges are sufficient and should not be increasedfurther. However, Lyon could still optimize commercial revenuesin order to increase managerial productivity and better managetheir joint revenue streams. The targets obtained from thebasic DEA results would be very challenging in the short- or

Non-aeronautical revenues Aeronautical revenues

15,000,000

25,000,000

35,000,000

45,000,000

55,000,000

65,000,000

75,000,000

85,000,000

95,000,000

105,000,000

basic model

independent model (2)

combined model (3)

current15,000,000

25,000,000

35,000,000

45,000,000

55,000,000

65,000,000

75,000,000

85,000,000

95,000,000

105,000,000

basic model

independent model (2)

combined model (3)

current

Fig. 9. Current and target output values for revenue maximization models with respect to Lyon.

Fig. 10. Co-plot of output maximization results with emphasis on Lyon. (Coefficient of alienation is 0.107 and average of correlations is 0.815. The shorter term targets are

the white observations and the larger, black observations represent the long term targets).

N. Adler et al. / Omega 41 (2013) 442–458 455

medium-term and should therefore be considered only as a long-term target. As shown in the graphs of Fig. 9, several paths to thePareto frontier can be defined for the airport, in which bothaeronautical and commercial revenues could be expanded, eitherequi-proportionally, or with greater emphasis on the non-aero-nautical revenues.

6. Conclusions and directions for future research

DEA has been repeatedly applied to the study of airport produc-tivity however the basic DEA models treat the airport technology asa black box, which reduces the usefulness of the model for purposesof benchmarking. The focus of this paper has been to model the

N. Adler et al. / Omega 41 (2013) 442–458456

airport production process from a managerial perspective in order toprovide a set of models that aid the benchmarking process. Usuallynetwork-DEA is applied to determine both the efficiency of subprocesses and overall efficiency [39] whereas in this research,network DEA acts as a decision aid describing the productionprocess and demonstrating the sequential effects that separate finalfrom intermediate outputs, including those under partial managerialcontrol and those that are known to be non-discretionary. Conse-quently, the approach connects aeronautical and commercial activ-ities via intermediate products.

To ensure that a reasonable set of peer airports are chosen, arestricted reference mechanism that limits the DEA dual variables(benchmark intensities) is applied, ensuring appropriate compar-ability within the heterogeneous dataset. The approach proposedby Golany and Thore [27] restricts the selection of best practiceDMUs according to predefined boundaries within the basic DEAframework. As a result, each DMU optimizes the last stage of thenetwork, whilst taking into account the information from pre-vious stages. In addition, PCA–DEA is applied to reduce thenumber of variables when clusters are too small and variablesare highly positively correlated in order to avoid the curse ofdimensionality. By identifying distinct reference sets, we provideindividual benchmarks for inefficient DMUs that permit identifi-cation of strategy changes over time and account for constraintswith respect to economic regulation and airport infrastructure.The formulation was further adapted to enable partial flexibilitywith respect to an expensive and complicated infrastructuresystem, thus developing an implementation path for the bench-marking process. Finally, the provision of ground handling wasshown to severely affect efficiency estimates, leading to a separa-tion in the comparison of those airports that undertake theprocess in-house compared to those that outsource.

Data proved to be one of the more difficult issues for the casestudy developed in this research. After defining salient variables (asin Fig. 1), we were then forced to reduce the extent of the potentialanalysis due to data availability issues (as in Fig. 4). It would beextremely helpful if government bodies, such as the InternationalCivil Aviation Organization, were to standardize the data collectionand publish data openly as this would enable fair and transparentcomparisons. However, the results have shown that compared withthe basic DEA approach, the formulations developed in this researchprovide more appropriate benchmarks which provide useful infor-mation for airport managers interested in improving performanceover time. In the case of Hanover, we show that in the short run it issufficient to decrease costs and increase utilization, whereas basicDEA benchmarks require the airport to close the equivalent of atleast one runway. In the case of Lyon we demonstrate that in theshort-term the airport earns a sufficient level of aeronauticalrevenues and needs to focus on improvements on the commercialside. In comparison, the results of basic DEA require Lyon to increaseaeronautical revenues by 40% in order to operate efficiently. Itshould be noted that, in contrast to the results from basic DEA,DMUs acting as benchmarks are not necessarily efficient since theymay deviate from the overall Pareto frontier if limited by thereference restrictions. On the other hand, this process provides animplementation path towards the long-term Pareto frontier.

To be in a position to undertake a benchmarking process requiresthe collection and publication of airport related data providedopenly at the federal level, since such information would be ofpublic interest. Furthermore, an airport also produces undesirableoutputs such as delays, pollution and ground access congestion.Besides capacity utilization which has been considered in ourresearch, delay substantially affects airport and airline performanceand should clearly be included in a benchmarking study. Forimproved managerial benchmarking, disaggregated data withregard to non-aeronautical activities would help to identify

successful strategies on the commercial side. Other factors that arebeyond managerial control include the competitive environment,ownership structure and economic regulation. These aspects influ-ence managerial behavior, and accounting for them may furtherimprove comparability and permit the relevant authorities toanalyze the impact of cost or incentive based regulation on manage-rial performance. In order to assess technological changes over asufficient period of time, Malmquist DEA may be applied in futureresearch provided a balanced dataset exists. Furthermore, an airportis typical of an industry with lumpy investments, hence time lagsbetween investment and changes in productivity should be con-sidered, which would require adapting the formulations developedhere in order to account for dynamic changes over time.

Acknowledgments

We would like to thank the German Airport Performance(GAP) and German Aviation Benchmarking (GAB) research pro-jects that are funded by the German Ministry of Education andResearch for providing the data. Further, we thank Prof. GertBrunekreeft, Prof. Peter Forsyth, Prof. David Gillen, Dr. ThomasImmelman, Prof. Hans-Martin Niemeier, Prof. Mike Trethewayand the participants of the German Aviation Research Society(GARS) seminar on Airport Benchmarking in November 2009, theWorld Bank roundtable on Transport Productivity in February2010 and the INFORMS Conference 2010 for fruitful discussions aswell as extremely helpful comments from two referees of thisJournal. Nicole Adler would also like to thank the Recanati Fundfor partial funding of this research.

Appendix A. Mixed integer linear program formulation(model 2)

The following formulation combines the restricted referencesand linear program of Model (2) into a single stage but at theexpense of solving a mixed integer linear program.

Maxl,y

y1þy2

s:t:alXaj l

n12rXn

j ln12rauXa

j ln12 8j¼ 1,. . .,4, n¼ 1,. . .,N

blIakl

n12r In

kln12rbuIa

kln12 8k¼ 1,2, n¼ 1,. . .,N

alXaj l

n13rXn

j ln13rauXa

j ln13 8j¼ 1,. . .,4, n¼ 1,. . .,N

blIaml

n13r In

mln13rbuIa

mln13 8m¼ 3,4, n¼ 1,. . .,NXN

n ¼ 1

Inkl

n24r Ia

k 8k¼ 1,2

XN

n ¼ 1

Yn1l

n24Zy1Ya

1

XN

n ¼ 1

Inj l

n235r Ia

j 8j¼ 1,. . .,4

XN

n ¼ 1

Yn2l

n235Zy2Ya

2

XN

n ¼ 1

ln24 ¼ 1,

XN

n ¼ 1

ln235 ¼ 1

ln24rln

12, ln235rln

12, ln235rln

13

ln12Af0,1g, ln

13Af0,1g binary

y1,y2,ln24,ln

235Z0

where superscript a is the index of DMUa, the unit underinvestigation; Xa represents the input values of DMUa; Ya and Ia

are the output and intermediate values of DMUa respectivelyand the subscript of intensities for DMUa, ln

ij, denotes thelink connecting between nodes i and j in the network

N. Adler et al. / Omega 41 (2013) 442–458 457

presented in Fig. 4. y1 and y2 represent the relative performancefor the commercial and airside activities respectively, where avalue of 1 indicates a position on the interior frontier and a valuegreater than 1 indicates by how much the relevant revenuesought to be increased in order for DMUa to move towards thefrontier. It should be noted that the first four rows of model (2) arenot summed over n, in order to restrict envelopment intensities ln

24

and ln235, thus comparing airports that possess input levels that lie

within a boundary around DMUas intermediate outputs.

Table B1

Code Airport Country Time period Ground handling

ABZ Aberdeen UK 1999 Not provided

AMS Amsterdam Netherlands 1998–2007 Not provided

ATH Athens Greece 2005–2007 Not provided

BHX Birmingham UK 2005 Not provided

BLQ Bologna Italy 2000–2005 Provided

BRE Bremen Germany 1998–2007 Provided

CGN Cologne–Bonn Germany 1998–2007 Provided

CPH Copenhagen Denmark 1998–2007 Not provided

DRS Dresden Germany 1998–2004 Provided

DTM Dortmund Germany 2001–2007 Provided

DUS Dusseldorf Germany 1998–2007 Provided

FLR Florence Italy 2000–2005 Provided

FRA Frankfurt Germany 2002–2007 Provided

GLA Glasgow UK 1999–2006 Not provided

GOA Genoa Italy 2000–2005 Provided

GVA Geneva Switzerland 1998–2007 Not provided

HAJ Hanover Germany 1998–2007 Provided

HAM Hamburg Germany 1998–2007 Provided

LBA Leeds–Bradford UK 1998–2006 Not provided

LCY London City UK 2002 Baggage provided

LEJ Leipzig Germany 1998–2002 Provided

LGW London Gatwick UK 1998–2002 Not provided

LHR London Heathrow UK 1998–2006 Not provided

LJU Ljubljana Slovenia 2004–2007 Provided

LTN London Luton UK 1998–2005 Not provided

LYS Lyon France 1998–2005 Not provided

MAN Manchester UK 1998–2006 Not provided

MLA Malta Malta 2005–2006 Not provided

MLH Basel–Mulhouse France 1998–2007 Not provided

MME Durham Tees Valley UK 2002 Baggage provided

MRS Marseille France 1998–2006 Not provided

MUC Munich Germany 1998–2007 Provided

NCE Nice France 1998–2006 Not provided

NUE Nuremberg Germany 1998–2007 Provided

OSL Oslo Norway 1999–2007 Not provided

RIX Riga Latvia 2004–2006 Provided

STN London Stansted UK 1998–2006 Not provided

STR Stuttgart Germany 1998–2007 Provided

SZG Salzburg Austria 2004–2007 Provided

TLL Tallinn Estonia 2002–2007 Provided

VCE Venice Italy 2000–2005 Provided

VIE Vienna Austria 1998–2007 Provided

ZRH Zurich Switzerland 1998–2007 Not provided

In order to connect the network approach of Fig. 4 and therestricted reference set, ln

12 and ln13 are binary variables and ln

24

and ln235 are non-negative continuous variables. If ln

12 ¼ 1 thenDMUn could be included in the peer group for DMUa on the non-aeronautical side and if ln

12 ¼ ln13 ¼ 1 then DMUn could be

included in the peer group for DMUa on the aeronautical side.ln

12 consequently connects costs to the number of passengersproduced and ln

13 connects costs to cargo and air traffic move-ment production such that DMUs of similar size and coststructure represent potential benchmarks.11 ln

24 connects the

11 The effect of this approach is depicted in Fig. 2 resulting in DMUs 1 and

2 acting as benchmarks for DMUa: l1,212 and l1,2

13 are binary variables equal to

number of passengers to non-aeronautical revenue derived andln

235 connects passengers, cargo and air traffic movements toaeronautical revenue.

Appendix B. Airport Dataset12

See Table B1 below.

References

[1] Zhang A, Zhang Y. Airport charges and capacity expansion: effects ofconcessions and privatization. Journal of Urban Economics 2003;53:54–75.

[2] Oum TH, Adler N, Yu C. Privatization, corporatization, ownership forms andtheir effects on the performance of the world’s major airports. Journal of AirTransport Management 2006;12:109–21.

[3] Liebert V, Niemeier H-M. Benchmarking of airports—a critical assessment. In:Proceedings of World Conference on Transport Research; 2010.

(footnote continued)

1 since they lie within the boundaries of the first four equations in model (2)

whereas l4,512 and l4,5

13 equal 0.12 Source: adapted from SH&E [72] and airport websites.

N. Adler et al. / Omega 41 (2013) 442–458458

[4] Hooper PG, Hensher DA. Measuring total factor productivity of airports—anindex number approach. Transportation Research Part E 1997;33:249–59.

[5] Oum TH, Yu C. Measuring airports’ operating efficiency: a summary of the2003 ATRS global airport benchmarking report. Transportation Research PartE 2004;40:515–32.

[6] Vasigh B, Gorjidooz J. Productivity analysis of public and private airports: acausal investigation. Journal of Air Transportation 2006;11:144–63.

[7] Pels E, Nijkamp P, Rietveld P. Inefficiencies and scale economies of Europeanairport operations. Transportation Research Part E 2003;39:341–61.

[8] Oum TH, Yan J, Yu C. Ownership forms matter for airport efficiency: astochastic frontier investigation of worldwide airports. Journal of UrbanEconomics 2008;64:422–35.

[9] Gillen D, Lall A. Developing measures of airport productivity and perfor-mance: an application of data envelopment analysis. Transportation ResearchPart E 1997;33:261–73.

[10] Sarkis J. An analysis of the operational efficiency of major airports in theUnited States. Journal of Operations Management 2000;18:335–51.

[11] Parker D. The performance of the BAA before and after privatization. Journalof Transport Economics and Policy 1999;33:133–46.

[12] Martın JC, Roman C. The relationship between size and efficiency: a bench-marking analysis of Spanish commercial airports. Journal of Air TransportManagement 2008;2:183–97.

[13] Murillo-Melchor C. An analysis of technical efficiency and productive changein Spanish airports using the Malmquist Index. International Journal ofTransport Economics 1999;26:271–92.

[14] Abbott M, Wu S. Total factor productivity and efficiency of Australianairports. Australian Economic Review 2002;35:244–60.

[15] Yu MM. Measuring physical efficiency of domestic airports in Taiwan withundesirable outputs and environmental factors. Journal of Air TransportManagement 2004;10:295–303.

[16] Yu MM. Assessment of airport performance using the SBM-NDEA model.Omega—International Journal of Management Science 2010;38:440–52.

[17] Barros CP, Sampaio A. Technical and allocative efficiency of airports. Inter-national Journal of Transport Economics 2004;31:355–77.

[18] Adler N, Berechman J. Measuring airport quality from the airlines viewpoint: anapplication of data envelopment analysis. Transportation Policy 2001;8:171–81.

[19] Lin LC, Hong CH. Operational performance evaluation of international majorairports: an application of data envelopment analysis. Journal of Air Trans-port Management 2006;12:342–51.

[20] Martın JC, Roman C, Voltes-Dorta A. A stochastic frontier analysis to estimatethe relative efficiency of Spanish airports. Journal of Productivity Analysis2009;31:163–76.

[21] Graham A, Holvad T. Efficiency measurement for airports. Paper presented at theTrafik Dags PAA Aalborg Universitet 2000 Conference, Aalborg University; 2000.

[22] Morrison WG. Understanding the complexities and challenges of airportperformance benchmarking. Journal of Airport Management 2009;3:145–58.

[23] Fare R. Measuring Farrell efficiency for a firm with intermediate inputs.Academia Economic Papers 1991;19:329–40.

[24] Tone K, Tsutsui M. Network DEA: a slacks-based measure approach. Eur-opean Journal of Operational Research 2009;197:243–52.

[25] Oum TH, Yu C, Fu X. A comparative analysis of productivity performance ofthe world’s major airports: summary report of the ATRS global airportbenchmarking research report––2002. Journal of Air Transport Management2003;9:285–97.

[26] Zhang A, Zhang Y. Airport capacity and congestion pricing with bothaeronautical and commercial operations. Transportation Research Part B2010;44:404–13.

[27] Golany B, Thore S. Restricted best practice selection in DEA: an overview withcase study evaluating the socio-economic performance of nations. Annals ofOperations Research 1997;73:117–40.

[28] Banker RD, Morey RC. Efficiency analysis for exogenously fixed inputs andoutputs. Operations Research 1986;34:513–21.

[29] Adler N, Golany B. Evaluation of deregulated airline networks using dataenvelopment analysis combined with principal component analysis with anapplication to Western Europe. European Journal of Operational Research2001;132:18–31.

[30] Adler N, Yazhemsky E. Improving discrimination in data envelopmentanalysis: PCA–DEA or variable reduction. European Journal of OperationalResearch 2010;202:273–84.

[31] Charnes A, Cooper WW, Rhodes E. Measuring the efficiency of decisionmaking units. European Journal of Operational Research 1978;2:429–44.

[32] Banker RD, Charnes A, Cooper WW. Some models for estimating technicaland scale inefficiencies in data envelopment analysis. Management Science1984;30:1078–92.

[33] Morrison SA. Estimation of long-run prices and investment levels for airportrunways. Research in Transportation Economics 1983;1:103–30.

[34] Sarkis J, Talluri S. Performance based clustering for benchmarking of USairports. Transportation Research part A 2004;38:329–46.

[35] Fare R, Grosskopf S. Productivity and intermediate products: a frontierapproach. Economic Letters 1996;50:65–70.

[36] Fare R, Grosskopf S. Network DEA. Socio Economic Planning Sciences2000;34:35–49.

[37] Lewis HF, Sexton TR. Network DEA: efficiency analysis of organizations withcomplex internal structure. Computers and Operations Research 2004;31:1365–410.

[38] Emrouznejad A, Thanassoulis E. A mathematical model for dynamic effi-ciency using data envelopment analysis. Applied Mathematics and Computa-tion 2005;160:363–78.

[39] Golany B, Hackman ST, Passy U. An efficiency measurement framework for multi-stage production systems. Annals of Operations Research 2006;145:51–68.

[40] Chen CM. A network-DEA model with new efficiency measures to incorpo-rate the dynamic effect in production networks. European Journal of Opera-tional Research 2009;194:687–99.

[41] Kao C. Efficiency decomposition in network data envelopment analysis: arelational model. European Journal of Operational Research 2009;192:949–62.

[42] Castelli L, Pesenti R, Ukovich W. A classification of DEA models when theinternal structure of the decision making units is considered. Annals ofOperations Research 2010;173:207–35.

[43] Yu MM, Lin ET. Efficiency and effectiveness in railway performance using amulti-activity network DEA model. Omega—International Journal of Man-agement Science 2008;36:1005–17.

[44] Beasley J. Determining teaching and research efficiencies. Journal of theOperational Research Society 1995;46:441–52.

[45] Cook WD, Hababou M, Tuenter HJH. Multicomponent efficient measurementand shared inputs in data envelopment analysis: an application to sales andservice performance in bank branches. Journal of Productivity Analysis2000;14:209–24.

[46] Adler N, Golany B. Including principal components weights to improvediscrimination in data envelopment analysis. Journal of the OperationalResearch Society 2002;53:985–91.

[47] Nataraja NR, Johnson AL. Guidelines for using variable selection techniques indata envelopment analysis. European Journal of Operational Research 2011;215:662–9.

[48] Adler N, Raveh A, Yazhemsky E. DEA presented graphically using multi-dimensional scaling. In: Cook W, Zhu J, editors. Modeling data irregularitiesand structural complexities in data envelopment analysis. New York:Springer; 2007. p. 171–87.

[49] Adler N, Raveh A. Presenting DEA graphically. Omega—International Journalof Management Science 2008;36:715–29.

[50] Brockett PL, Golany B. Using rank statistics for determining programmingefficiency differences in data envelopment analysis. Management Science1996;42:466–72.

[51] Sueyoshi T, Aoki S. A use of a nonparametric statistic for DEA frontier shift:the Kruskal and Wallis rank test. Omega—International Journal of Manage-ment Science 2001;29:1–18.

[52] Guttman L. A general nonmetric technique for finding the smallest coordi-nate space for a configuration of points. Psychometrika 1968;33:469–506.

[53] Airports Council International. World airport traffic report 2008. Geneva: ACIWorld; 2009.

[54] Graham A. Airport benchmarking: a review of the current situation. Bench-marking: An International Journal 2005;12:99–111.

[55] International Air Transport Association. World scheduling guidelines, 19thedition, /http://www.iata.org/NR/ContentConnector/CS2000/SiteInterface/sites/whatwedo/scheduling/file/fdc/WSG-12thEd.pdfS; 2010 [accessed 17 March2010].

[56] Aeroporto di Firenze. Technical data, /http://www.aeroporto.firenze.itS;2010.

[57] Lovell CAK, Pastor JT. Units invariant and translation invariant DEA model.Operations Research Letters 1995;18:147–51.

[58] International Air Transport Association. Airport capacity/demand profiles.2003edition Monetral, Geneva: Air Transport Consultancy Services; 2003.

[59] Dietz P. Verluste auf dem Rollfeld. Frankfurter Rundschau; 11 November 2009.[60] Hutter D. 2000 Mitarbeitern droht die Kundigung. Suddeutsche Zeitung; 29

May 2009.[61] Cologne–Bonn Airport. Press office, /http://www.airport-cgn.de/main.

php?lang=2&id=140S; 2010 [accessed 18 March 2010].[62] Andersen P, Petersen NC. A procedure for ranking efficient units in data

envelopment analysis. Management Science 1993;39:1261–4.[63] Banker RD, Gifford JL. A relative efficiency model for the evaluation of public

health nurse productivity. Carnegie: Mellon University Mimeo; 1988.[64] Banker RD, Chang H. The super-efficiency procedure for outlier identification,

not for ranking efficient units. European Journal of Operational Research2006;175(2):1311–20.

[65] Dortmund Airport. Geschaftsbericht 2007. Flughafen Dortmund GmbH; 2007.[66] Vienna International Airport. Geschaftsbericht. Flughafen Wien AG; 2001,

2004, 2007.[67] Gelhausen MC. Modelling the effects of capacity constraints on air travellers’

airport choice. Journal of Air Transport Management 2011;17:116–9.[68] Flughafen Kassel–Calden. Ausbau, /http://www.flughafenkassel.de/t3S;

2010 [accessed 18 March 2010].[69] Lyon Aeroport. Les indicateurs trafic de l’aeroport de Lyon-Saint Exupery.

Aeroport Lyon-Saint Exupery; 2008.[70] Flughafen Basel–Mulhouse. 2007—The year-key figures. Flughafen Basel

Mulhouse; 2007.[71] ADV. Mitglieder/Flughafenkarte, /http://www.adv.aero/mitglieder.htmlS;

2010 [accessed 19 March 2010].[72] SH&E. Appendices: study on the quality and efficiency of ground handling

services at EU airports as a result of the implementation of Council Directive96/67/EC. Report to European Commission. London: SH&E International AirTransport Consultancy; 2002.