Hub-Spoke Network Choice Under Competition ... - Semantic Scholar

TRANSPORTATION SCIENCE

informs

Vol. 39, No. 1, February 2005, pp. 58–72 issn 0041-1655 eissn 1526-5447 05 3901 0058

®

doi 10.1287/trsc.1030.0081 © 2005 INFORMS

Hub-Spoke Network Choice Under Competition with an Application to Western Europe Nicole Adler

School of Business Administration, Hebrew University of Jerusalem, Mount Scopus, Jerusalem 91905, Israel, [email protected]

T

he aim of this paper is to present a model structure that analyzes the hub-spoke network design issue within a competitive framework. Under deregulation, airlines have developed hub-and-spoke networks, enabling them to increase frequency by aggregating demand and to prevent entry into the marketplace by reducing airfares. While liberalization in the United States and Europe was undertaken to increase competition, the results in this direction are unclear. This research evaluates airline profits based on a microeconomic theory of airline behavior under deregulation and the effect on hub-and-spoke networks. Through a two-stage, Nash best-response game, we search for equilibria in the air transportation industry. The game is applied to Western Europe, where profitable hubs and monopolistic equilibria are clearly identifiable, and duopolistic equilibria are potentially viable, given sufficient demand. Key words: air transportation; hub-and-spoke network design; competition; location; game theory History: Received: March 2000; revision received: October 2002; accepted: May 2003.

Introduction

Dobson and Lederer (1993), Nero (1996), Hendricks et al. (1999), and Marianov et al. (1999). Hansen (1990) developed an n-player, noncooperative game in which the airline’s sole strategy set is frequency of service. The airlines were split into two categories: hub airlines and point-to-point (or fully connected) airlines. The set of simplifying assumptions included fixed airfares, adequate capacity, inelastic demand to price and service level, and consideration of nonstop and one-stop services only. Using regression analysis, Hansen could not prove the existence of an equilibrium; his application to the U.S. air transportation industry showed “quasiequilibrium.” Hong and Harker (1992) developed a two-stage, game-theoretic representation of an air traffic network market mechanism for slot allocation. Assuming an oligopolistic air transport market model, their solution derived the flight patterns, ticket prices, routes, and carrier choice for passengers, as well as landing fees paid to the airports. Using a quasivariational inequality technique to solve a Cournot-Nash model, they proved the existence of a unique solution. Given a total origin-destination (O-D) trip demand matrix, scheduled flight times, aircraft size and capacity, average operating costs, and fixed maintenance costs for each flight and a utility function for the passengers on each O-D pair, they solved for a three-node example. Dobson and Lederer (1993) developed a mathematical program to study the competitive choice of flight

In this paper we attempt to evaluate the most appropriate hub-and-spoke (HS) network for an airline to develop in a competitive environment. A mathematical program has been developed for this purpose, which is then applied within a game-theoretic framework. An application discusses potential equilibria in the West European air transportation industry today. Numerous studies have attempted to analyze the use of HS networks in a deregulated market. See, for example, Bailey et al. (1985), Bittlingmayer (1990), Borenstein (1989, 1992), Brueckner and Spiller (1994), Button and Pitfield (1991), Hendricks et al. (1992), Johnson (1985), Keeler (1991), Morrison and Winston (1986), Reynolds-Feighan and Hewings (1990), and Spiller (1989). Call and Keeler (1985) argued that in the long-run airfares will fall in a competitive market, despite the marketing and network advantages enjoyed by the large, established carriers. They argued that, slowly, “oligopoly models may give way to models of competition (or possibly monopolistic competition) in explaining air fares and service.” Until the early 1990s, airline competition and airline network strategies were generally treated as separate subjects in the literature. Ghobrial and Kanafani (1985) did seek to identify equilibrium in an airline network, but they restricted the case to single-hub networks. Several papers have since been written in the field of airline competition using HS networks, including Hansen (1990), Hong and Harker (1992), 58

Adler: Hub-Spoke Network Choice

Transportation Science 39(1), pp. 58–72, © 2005 INFORMS

schedules and route prices by airlines operating in a pure HS (i.e., single hub) system. Utilizing a subgame perfect Nash equilibrium for a two-stage game, they found equilibria in a five-node network example. The profit-maximizing heuristic developed searches for the best flights, routes, and prices in a three-level hierarchical process, but they cannot be assured of a globally optimal solution. Assumptions in their model included: A single aircraft size and one class of customers; no traffic originates or is destined for the hub airport; airline variable costs are dependent on flying time alone; and zero variable passenger costs. An additional assumption requires that duopolists serve the identical set of spoke cities using the same hub, however an airline using a hub generally requires at least 50% of the slots available. Marianov et al. (1999) discussed the relocation of hubs in a competitive environment given changes in the O-D demand matrix over time. Demand, in terms of flow, is captured through a minimum cost breakdown to avoid the use of prices. A tabu search heuristic is developed to solve the maximal flow optimization model. Hendricks et al. (1999) wrote a theoretical paper discussing equilibria in airline networks. They studied duopolies using a two-stage game in which two carriers simultaneously chose their networks and then competed for travelers. They argued that when carriers compete aggressively (e.g., Bertrand-like behavior), monopoly is an equilibrium outcome and no duopoly equilibrium exists if both carriers choose HS networks. They further attested that duopoly equilibria exist either in non-HS networks or when HSnetwork carriers do not price aggressively and there are a sufficient number of nodes. They assumed that interlining (when a passenger travels via two carriers) is not feasible, that a carrier’s operating profits are additively separable across city-pair markets, and that there are no capacity constraints. The majority of papers cited above utilize discrete choice models to evaluate demand for routes, based on an aggregation of individual travelers’ preferences. Ben-Akiva and Lerman (1985) defined discrete choice analysis as “the modeling of choice from a set of mutually exclusive and collectively exhaustible alternatives Briefly, a decision-maker is modeled as selecting the alternative with the highest utility among those available at the time a choice is made.” Hansen (1990) used a market share model based on the logit function, including price, log of frequency, and the travelers’ preference for a direct service as variables. Hansen argued that the logarithmic form of service frequency is preferable because “one would expect diminishing returns with respect to the gain in service attractiveness from adding additional flights.”

59 Hong and Harker (1992) utilized a linear combination of flying time, connection delay, price, frequency, aircraft size, and the number of legs flown in their disutility function based on Morrison and Winston’s (1986) study. Dobson and Lederer (1993) assumed that a traveler’s demand for each route was based on their cost function, which could be split into three elements that are subsequently aggregated linearly: the cost of departing at a time that differs from the customer’s most preferred departure time, the costs associated with the route duration, and the airfare. Alamdari and Black (1992) discussed the use of the logit model in evaluating the influence of liberalization on passenger demand. They argued that “simple ‘all or nothing’ models that assume the cheapest airline is chosen by all the passengers are not suitable for determining airlines’ market share. Passenger demand is influenced by a combination of fare and the many attributes that make up the quality of service provided From an empirical modeling viewpoint, not all the factors that affect travelers’ choice are known, and of those that are known, only some are measurable.” In addition, the share model can only predict the airline’s share as a function of the chosen variables; it does not predict changes in total demand as a result of the levels specified. They refered to nine studies that evaluated diverse variables to explain airline demand within the logit framework. Six of the studies found fares to be an influence on the share of passengers, perhaps because of the absence of a significant variation in the airfares, and all studies found relative frequency of flights to be an important influence. This work presents an operations research–based modeling framework for analyzing specific, potential HS networks given the existing competition. The research includes four new elements beyond that of the existing literature. While many papers have discussed the practice of “hubbing,” none have developed an n-node representation to specify which precise network would be most profitable to an airline. In addition, this is the first reasonably sized problem to be analyzed in the literature, from this point of view. Second, while the U.S. aviation market has been researched quite substantially, little or no work has been undertaken in Europe, where 45 member states collate data, if data are accessible at all, using different practices and accounting terminology. The data gathered for the purposes of examining Western Europe were collected over a period of several years across numerous countries. Third, the transportation literature has so far assessed equivalents of the subgame, either under oligopolistic or duopolistic market conditions. However, no paper has viewed the wider game or attempted to evaluate specific, profitable HS networks in a competitive marketplace. While the ambitious target of identifying all networks


60


most likely to survive under deregulation in Western Europe was not achieved due to the size of the problem, the framework presented here can be used to identify the most appropriate response to competitive carriers and to suggest suitable adaptations necessary to ensure stability and profitability. The mix of discrete choice, optimization, and game-theoretic models can also be used to evaluate potential strategic partners connecting via alliances or mergers. Finally, the managerial, modeling approach can be used by potential new entrants to assess the advantages of entering new markets or setting up a new network. A glossary of terminology (notations) can be found in §1, and assumptions of the models developed in this research can be found in §2. A multinomial logit (MNL) model is described in §3 and a nonlinear mathematical program, which evaluates the most profitable airline network, is developed in §4. A twostage, Nash-type, best-response game is derived in §5, and an application to the Western European aviation industry is developed in §6, with a listing of all airports in the sample set in Appendix A and a presentation of the collated demand dataset in Appendix B. A summary can be found in §7.

1.

Notation

A set of airlines a index that represents a specific airline CA = A ∪ 0 set of airline alternatives available to a passenger and the option not to fly N set of all existing nodes in the network configuration Arc a set of all existing legs in airline a’s specific network configuration Hub a set of all hub nodes (a node that is connected to more than one other node) for airline a ∈ A i j ∈ N node indices s ∈ b nb s represents the type of traveler, either business b or nonbusiness (nb) k ∈ Arc a leg index parameter in cost function reflecting scale economies parameter in cost function reflecting network economies s parameter in utility function reflecting frequency elasticity of demand of type s traveler s parameter in utility function representing willingness of type s traveler to pay for a direct flight s parameter in utility function translating frequency into a price comparable value for type s traveler parameter in cost function translating the CobbDouglas function into monetary terms fka number of flights per week on leg k for airline a

dijs O-D demand from node i to node j per type s traveler (passengers per week) pijsa airfare from node i to node j for passenger type s traveling with airline a MSijsa market share of airline a over O-D trip i j

for passenger type s RFija reduction factor on path i j for airline a c k a commencement node on leg k for airline a e k a end node of leg k for airline a h k a plane size index for leg k for airline a Ci maximum aircraft movement (ACM) capacity of node i (ACM/week) n total number of nodes in network m total number of paths = 2(n − 1) given HS configuration (one or two hubs) passenger-related charge from departing PCnt i node i for nontransfer passenger PCti passenger-related charge from departing node i for transfer passenger TPCia total passenger charges paid by airline a at departing node i LCi landing related charge per plane ton at node i MTOW maximum take-off weight (MTOW) of airplane in tons PSSEAT h k a plane size in terms of number of seats on leg k for airline a plane size in terms of MTOW on leg k for PSMTOW h k a

airline a

2.

Assumptions

It is assumed that an HS network reduces total costs to the airline due to economies of scope and scale and that passengers will be prepared to travel over two or three legs if necessary. The U.S. air transportation market today would appear to show these assumptions to be reasonable. The model developed requires airlines to choose an HS system with one or two potential hubs. In reality, an airline may choose a specific HS network and add a few direct connections where demand justifies such a decision. Inclusion of these arcs would not add to the complexity of the solution procedures described. It is further assumed that an airline is entirely free to choose its most preferable network and purchase the rights to land and take off as required. Presently, slot allocations at European Union (EU) airports are governed by historical grandfather rights, whereby a “home carrier” representative serves as a slot coordinator in (IATA’s) annual negotiations. New EU regulations may marginally affect the distribution of slots, in particular at highly congested airports. With the slow disappearance of the flag-carrier phenomenon, our assumption may be considered reasonable.


61


Passenger utility is defined as a function of frequency, willingness to pay for a direct flight, and airfare. The utility of a traveler flying via a hub is based on the minimum frequency along one of the legs of the route, because this will cause the most important restriction on choice. The utility function also ensures that passengers forced to travel via one or two hubs will pay less than those traveling directly. While this clearly does not include all possible factors affecting a passenger’s decision to travel, it does ensure tractability of the problem outlined. We assume that the maximum, total, O-D passenger demand is given, and the carrier’s strategy therefore affects the number of passengers carried and not the total volume of demand. Sensitivity analysis can be used to understand the effects of an increase or decrease in maximal demand. For example, demand could be increased across the board equally as shown in §6; alternatively, demand could be adapted at a particular airport to understand its effect on hub choice. Another alternative would be to add an additional model to the framework developed here, such as time series analysis or a gravity model to adapt demand according to the decision variables available. We assume that the MNL model’s utility function describes a traveler’s preferences sufficiently well to determine the market share of an airline, when compared to its competitors. Alamdari and Black (1992) argued that there is strong evidence to show the broad validity of the logit share model in analyzing market share in the air transportation market. Two fares exist as decision variables—one for business travelers and the other for nonbusiness travelers—from node i to node j, per airline. This is assumed for tractability of the problem space and results in two prices per O-D per airline, which can be increased if a more complex analysis of airfares is required. The fare from origin i to destination j is the same as that of j to i per airline. This also reduces the number of airfare decision variables to 1 n n − 1 for computational tractability and can be 2 easily removed if necessary. Airlines have developed complicated pricing strategies based on yield management technology; this technology has been ignored because we are looking at the networks from a strategic rather than a tactical point of view. This model could be enlarged or connected to a separate pricing model for more specific revenue results. The revenue results presented here will therefore be conservative, since only two passenger types are considered. No interlining is permitted; each traveler type chooses an airline or not to fly based on price, frequency, and value of time. In the real world, airlines attempt to dif-

ferentiate their product from the competition in order to avoid interlining through pricing, minimum connecting times, and marketing techniques such as frequent flyer programs. In addition, the likelihood of lost baggage and missed connections increases with interlining, thus encouraging travelers to avoid this practice. We assume that the operating costs of the carrier are a function of frequency and include parameters for scale and network economies. This assumption could be considered reasonable if the majority of flights are a similar stage length (within Western Europe, given an HS network, this may be considered tolerable). The fixed costs of an airline, such as the purchasing of aircraft and accrued capital costs, have been ignored. This study is attempting to choose between potential networks; if the fixed costs are constant, irrespective of the network chosen, then this information is irrelevant. In addition, plane sizes are limited so that the same size aircraft is used in both directions on a single path, which is a clearly reasonable assumption given that fleet size variation is naturally restricted. Finally, the game-theoretic form used to evaluate the effects of explicit competition requires the following assumption: Airlines choose their strategic network and decision variables in two stages. First, all airlines choose their networks and whether or not to offer services concurrently, then they choose frequencies, plane sizes, and airfares simultaneously, based on the knowledge of all the airlines’ choices in the first stage.

3.

Market Share Model

A multinomial logit (MNL) model has been developed to compute the market share of the airlines, given a maximal level of demand. Ben-Akiva and Lerman (1985) defined the MNL market share model as MSijsa = eVijsa thus

0 ≤ MSijsa ≤ 1 a∈A

eVijsc

∀a ∈ A

c∈CA

∀a ∈ A

MSijsa ≤ 1

If we assume that the utility achieved by traveler s from alternative c is Uijsc = Vijsc + !ijsc and that the disturbances !ijsc are independently and identically Gumbel distributed with a location parameter and a scale parameter " > 0, then Vijsc is the


62


deterministic utility attained by a type s traveler in O-D market i j from alternative c. The utility of a single passenger represents the passenger’s preferences on frequency, nonstop service, and price; it is drawn from Adler and Berechman (2001), where the parameters of the model were computed.

Vijsa =

 s wijsa mink∈Rija fka s     traveling from i to j via airline a    0   

traveling via a different mode or not traveling

(1)

4.

where Rija = k k ∈ Arc leg k belongs to route from node i to node j for airline a 1 + s direct travel wijsa = 1 indirect travel. The utility function can then be introduced into the logit model such that the market share of airline a for the O-D market i j per traveler type can be computed as follows: MSijsa

captures the efficiency, hence costs, of the carrier directly and the length of flying time indirectly. The additional value-of-time variable can be interpreted as the extra dollar amount that the consumer is willing to pay for direct, nonstop service. It can thus be used as a supplementary separation between the carriers, above and beyond the price and frequency, since it establishes whether an airline has a competitive advantage over other carriers with respect to its network combination and, specifically, its choice of hubs. Only passengers traveling directly will pay the additional expense, which is an immediate result of the airlines’ choice of hubs.

s

= exp s wijsa min fka − pijsa k∈Rija

s

−1

· 1+ exp s wijsa min fka

− pijsa

a ∈A

k∈Rija

The addition of one in the denominator of the market share function provides the traveler with the additional choice of not traveling, or at least not flying, consequently attaining zero utility. According to Hong and Harker (1992), the inclusion of the “no purchase” option captures the price elasticity of total demand and prevents airlines from overcharging. Numerous variables have been considered in the utility functions of various papers, in an attempt to capture all the decision rules of the passengers (see Alamdari and Black 1992). We have chosen three variables: airfare, service frequency, and value of time. The first two variables were considered relatively important in most studies. We would suggest that minimum service frequency is a variable that can capture more than simply the surface understanding of an HS network. It is a strategic variable enabling an airline to differentiate between its product and that of the competition, and also acts as an entry deterrent in the market. The level of frequency also

Mathematical Program

A mathematical program has been developed based on a microeconomic model of the behavior of an airline acting under deregulation (see Adler 2001). The decision variables of the economic model include airfare per O-D trip, per traveler type, per airline pijsa , aircraft size per leg per airline PSSEAT h k a , and frequency per directed leg per airline fka . The nonlinear mathematical program is solved per airline per network chosen. An airline’s profit function is maximized subject to three sets of constraints. The profit function is based on an airline’s revenue and cost functions. The revenue function computes receipts based on airfares, maximal demand, and the airline’s market share, which in turn is based on the travelers’ utility function. It is assumed that the airline’s operating cost function can be defined as a constant elasticity of substitution, Cobb-Douglas function as follows: C f ≡ fk > 0 (2) k ∈ Arc

This class of function is general enough to capture the cost of operating different types of networks with a varying number of routes and layout patterns. Starr and Stinchcombe (1992) in their economic analysis of HS and related systems argued that these systems were optimal under a variety of costand-demand configurations, typically demonstrating large, pervasive economies of scale. The cost function in Equation (2) is monotonically increasing in frequency and exhibits increasing returns to scale if < 1. Thus, as frequency increases, the additional cost per flight decreases. It is generally worthwhile increasing frequency in an HS system as opposed to a fully connected network, because it enables the airline to deter entry into the marketplace by carriers attempting to provide a direct service between i and j, where this route does not yet exist. Additional costs result from payments made by the airline to the airports in the form of landing and passenger charges. The landing charges (LC) are paid to


63


the arrival airport and are based on the maximum take-off weight of the airplane type. The passenger charges (PC) are paid to the departure airport for each passenger carried. Two PC have been included: the full tariff paid at the first departure airport and a transfer charge paid at subsequent hubs, when the passenger is carried on two or more legs to reach her chosen destination. This pricing system is in line with the present rules of most international airports. Because there are many different types of charges, the LC and PC have been modified to include other relevant charges such as handling, night, and noise charges, and will be referred to as landing-related and passenger-related charges. A secondary result of the model developed includes an analysis of the effects of the HS system chosen on the hub airports themselves. Both fixed and variable airport costs are evaluated in the model. The variable costs are based on aircraft movement (ACM)–related costs and the number of passengers using the terminal. In addition to airport costs and revenues, airport capacity and quality may affect the airlines’ choice of hubs. Airport capacity can be split into several categories, namely runway, apron (or parking), air traffic control, and terminal capacities. These are reflected in the constraints of the mathematical program. Airport quality is beyond the scope of this paper and is discussed in detail in Adler and Golany (2001). The mathematical program’s objective function for airline a is stated in Equation (3). Max'a =

min 1 min k∈Rija

i j s i =j

y z

fka PSSEAT h k a

y =z k∈Ryza s dyzs MSyzs a

· dijs MSijsa pijsa − PCint +

−

k∈Rija

fkaa

−

k∈Rija

k∈Rija c k a =i

PCtc k a

a1 +a2 PSSEAT h k a LCe k a fka

(3)

Airline revenue has been further tempered by the number of seats physically available. If there is excess demand that cannot be carried at the chosen levels of frequency and plane size, all passengers traveling along the arc are reduced equally, irrespective of their ultimate destination or passenger type (see §2 on assumptions of the model). Clearly, the airlines will try to manage demand in a more effective manner, but this is currently beyond the scope of the model. The double minimum in the objective function is used to prevent an airline from charging passengers without having available airline seats, caused by capacity constraints at airports and a maximum aircraft size. Furthermore, it may not be preferable to increase tariffs because this may cause groups of passengers to

choose an alternative. The constraints of the mathematical program include Minimum number of seats ≤ PSSEAT h k a ≤ Maximum number of seats

∀ k ∈ Arc a a ∈ A

a k∈Arc a

fka ≥ 0 pijsa ≥ 0 pijsa = pjisa

fka ≤ Ci

∀i ∈ N

(4) (5)

∀ k ∈ Arc a a ∈ A

(6)

∀ i j ∈ N s ∈ b nb a ∈ A

(7)

∀ s ∀ i j ∈ N i = j ∀ a ∈ A

(8)

Equation (4) specifies that each airline chooses its own fleet size and structure independently, per undirected leg. Equation (5) requires that each airport’s maximum runway capacity constraint be considered, based on the summed aircraft movements of all airlines using the airport facilities. Equations (6) and (7) ensure that frequency and prices are nonnegative, respectively. Furthermore, because the prices have been set such that the airfare from airport i to airport j is the same as the airfare in the opposite direction, Equation (8) is also necessary in the analysis. The mathematical program can be solved using several nonlinear optimization techniques, such as subgradient approaches (Bazaraa et al. 1993), smoothing approximations (Zang 1980), or conjugate gradient projection (Goldfarb 1969). The mathematical program solves the objective function per airline, given all other airlines’ hypothesized strategies, using a relevant algorithm for nonlinear objective functions and linear constraints. However, it should be noted that due to the nonconvexity of the objective function, a global optimum is not necessarily found under any of these techniques. Thus each problem must be solved from several initial points in an attempt to search for the globally optimal solution. For a detailed analysis, see Adler and Berechman (2001).

5.

Game Theory Model of Explicit Competition

The competitive model can now be defined as a multiairline, noncooperative, two-stage game. A game is defined by a set of players; each player uses a set of alternative actions or strategies to maximize his or her own payoff function, whose value depends on each of the players’ simultaneous actions (Von Neuman and Morgenstern 1953). The game described here consists of a set A of airlines whose strategy sets include the HS network, frequency of service, aircraft size, and airfares. In the first stage, airlines simultaneously


64 choose an HS network and whether or not to participate. Once the HS connections are set, they cannot be changed within the game. In other words, an airline can choose not to connect a spoke, through low frequency, but the spoke cannot be attached to a different hub once the first stage has been completed. In the second stage, after the network choice is revealed, each airline attempts to maximize its profit function (Equation (3)) through its choice of service frequency, aircraft sizes, and airfares, given the decisions of all other airlines. The airlines alone have been chosen as players in the game for several reasons, one being the lack of an airport strategy set. The airports are, except in the case of London, government owned. The individual European governments set their own airport pricing policy. In addition, the airport’s LC and PC are generally too low to affect the airlines’ network choice decision problem because they generally account for approximately 7%–12% of the airlines’ total costs (see Adler and Berechman 2001). One of the most significant elements—the size, hence the capacity—of the airports is again not a decision of the airport itself, but is based on political and environmental considerations of the relevant government agencies. Furthermore, the time lag between a decision to expand an airport’s capacity and the construction are considerable and beyond the scope of this model. The other potential set of players that are not accounted for explicitly in the game is the passengers themselves. However, the passengers’ utility functions are the basis for assessing the airfares and market share of the airline in the airlines’ payoff functions, thus the travelers are accounted for indirectly. Given explicit competition, subgame perfect Nash equilibria of the noncooperative game are sought with the following algorithm. An initial strategy profile is specified for the airlines choosing to participate. Then a specific sequence of airlines is chosen and the mathematical program is solved for the first airline in the sequence. Once an optimal solution is found, given all other airlines’ decision variables, the mathematical program is then solved for the second airline, and so on. Once a solution has been computed for all airlines, the program begins again with the first airline in the sequence. The iterative process continues until one of the following occurs. (1) A second-stage, subgame, Nash solution is found and an entire iteration has been computed such that no airline has changed its best response to any other airline. This solution could be of several forms: (1.1) all airlines have positive frequencies and serve the market in some capacity; (1.2) two or more airlines take over the market, forcing the remaining airlines to lose market share and profits, consequently going bankrupt through


continual losses, or merging, as appropriate. This could result in a duopoly, oligopoly, or competitive market; (1.3) one airline takes over the entire market, resulting in a monopoly. (2) A quasi-equilibrium is achieved, which can be defined as two distinct possibilities: (2.1) the program cycles around two or more possible solutions, without converging; (2.2) the majority of decision variables achieve convergence, but a few remain divergent. (3) No equilibrium exists and no convergence is attained. A Nash equilibrium can be defined as a strategy profile in which each player’s choice solution is as good a response to the other airlines’ choices as any other strategy available to that player (Kreps 1990, p. 404). Sufficient conditions for a Nash equilibrium to exist include: (i) the strategy set of each player is bounded, convex, and closed; (ii) the payoff function for each player is concave with respect to the player’s strategy set assuming fixed competitor strategies; and (iii) all payoff functions are continuous over the strategy sets of all players. While the strategy set of each player is bounded, convex, and closed and the payoff functions are continuous, the profit function is not concave. Therefore, we cannot prove that a subgame equilibrium exists. In addition, if an equilibrium is found, we cannot be sure that it is the only subgame equilibrium solution that exists. Thus, the entire process is run from several different initial strategy profiles. Once the equilibria of all the subgames have been evaluated, the existence of a subgame perfect equilibrium of the overall game can be analyzed. It is assumed that any airline making losses will choose not to play, thus any equilibrium of the subgame in which one or more airlines make losses will not be considered a subgame perfect equilibrium of the overall game. Subgame perfect equilibria exist if those airlines choosing to play cannot find an alternative that will provide greater profits and it is not worthwhile for airlines that choose not to participate to enter the market.

6.

Application to the Western European Aviation Industry

We examine the results of the model with respect to Western Europe and subsequently compare them to results published in the literature. First, we discuss the results of a base run in which three airlines choose to compete or not, each with different, two-hub network combinations. The results analyze competitive, duopolistic, and monopolistic equilibria. Subsequently, policy runs are developed for three-airline,


65


doubled O-D demand matrices over double-hub scenarios, and for a four-airline, single-hub scenario. The mathematical program was applied to a threeairline, Western European market, whereby each airline chooses a two-hub, balanced network. In this illustrative example, half the European spokes were attached to each of the two hubs according to minimum distance. The first airline consisted of a London Heathrow (LHR)–Zurich (ZRH) hub combination, the second used an Amsterdam (AMS)–Madrid (MAD) hub combination, and the third airline used a Frankfurt (FRA)–Barcelona (BCN) hub combination. For convenience, we refer to the first airline as BA, the second airline as KLM, and the third airline as Lufthansa. The smaller hubs were chosen as secondary, intra-European hubs, and the larger hubs, (LHR, AMS, and FRA), acted as international gateways with connections to the Far East and the United States. See Appendix A for a list of airports in the sample and Appendix B for the maximal O-D demand data. The mathematical program was run from many different initial strategy profiles and airline sequences. In several cases, while one solution was preferable for one airline, another starting point resulted in a more profitable solution for a different airline. Consequently, we report all the solutions computed, unless a solution is strictly dominated for all airlines. One of the solutions of the subgame, 1a in Table 1, presents the result attained after choosing an airline sequence beginning with BA, whereby initial aircraft sizes were fixed at the upper bound, i.e., 401 passenger seats, demand was split evenly between the three airlines, and load factors were set at 100%, thus frequency could be computed automatically. The initial airfare was fixed for business and nonbusiness, intraEuropean, or international, according to four constant values. Solution 1b presents the results whereby, with the same sequence of airlines, the initial strategy profile was set with a fixed aircraft size of 250 passenger Table 1

seats. Furthermore, game theory suggests that different solutions may be attained according to the order of the cycle, thus beginning the optimization procedure with the second airline instead of the first may achieve different results. Solutions 2a and 2b represent the results attained when KLM was evaluated first, and Solutions 3a and 3b are achieved when Lufthansa was evaluated first. To present the summary results in a game-theoretic format for the three airlines, two matrices are required. In each cell of Table 1, representing different subgames, the payoffs are in millions of U.S. dollars per week in the following order: BA, KLM, and Lufthansa, respectively. Results within a subgame that were strictly dominated for all airlines are not reported. In this game there is a single, pure-strategy, subgame perfect equilibrium. Given all solutions of the subgames, the BA monopolistic solution is the overall subgame perfect equilibrium, because it is worthwhile for BA to choose to play ($97,000,000 versus 0 payoff), but it is not worthwhile for KLM or Lufthansa (0 versus −$10,000,000) to change their strategies of not playing. For BA, the play strategy dominates the not-play strategy, because the airline attains positive profits independent of the other airlines’ strategy choice, given all the subgame combinations. Having removed BA’s not-play strategy, all remaining cells report losses for the other airlines except when they choose not-play, hence the monopolistic solution. In the unique, subgame perfect equilibrium, the LHR-ZRH network combination is strong enough to push the other two airlines out of the market. In this solution, although LHR and ZRH are each connected to nine European airports initially, ZRH only has positive frequency in connection with FCO, FRA, LIN, and LHR. The plane sizes are all around 250 seats except for the JFK-LHR and LHR-ZRH arcs, where the optimal aircraft size is 401 seats. The average airfare within Europe is $800 for business travelers and $600 for nonbusiness travelers. The prices are high in comparison to other solutions, suggesting

Potential Outcomes of the Airline Competition Game Lufthansa

BA

Play

Not

KLM

KLM

Play

Not

Play

Competition: (1a) 33 −9 −11 (1b) 52 −32 −9 (2a) 17 −10 −8 (3a) 41 −8 −19

Duopoly: (1a) 50 0 −10 (3a) 51 0 −12

Not

Duopoly: (2a) 0 33 −7

Monopoly: (3a) 0, 0, 32

BA

Play

Not

Play

Duopoly: (1a) 74 −10 0

Monopoly: (1b) 97, 0, 0

Not

Monopoly: (2a) 0, 57, 0

No Participation 0, 0, 0


66


that BA is taking advantage of its monopolistic status. Finally, it may be interesting to note that, although no subgame perfect equilibrium was found under oligopolistic market conditions, only one airline attained profits in all the enumerations. The duopoly solutions also show only one airline reaching profitable status. BA is always the strongest, and in the KLM-Lufthansa duopoly, Lufthansa is the loss leader. These results validate the theoretical conclusions reached in Hendricks et al. (1999). They argued that in a model of two carriers, each choosing HS networks and competing aggressively, Bertrand style, the only equilibrium solution outcome is that of a monopoly. Dobson and Lederer (1993), in a duopolistic setting, where each airline chooses the same HS network, found multiple equilibria in their subgame scheduling and pricing problem. In the overall game, they found one equilibrium solution in which both airlines compete because “There is too much demand for just one firm.” When they reduce the aircraft size (in their model there is only one exogenous aircraft size), ceteris paribus, they found two subgame perfect equilibria: one monopolistic and the other duopolistic. Given the knowledge that the basic O-D demand matrices represent approximately 45% of the actual demand carried in the base year of 1992, when taking into account both scheduled and charter flights to all airports in Western Europe, we doubled the demand matrices and attained the subgame equilibria of Table 2. The same three HS-network combinations compete in the two-stage, noncooperative game. Demand would now appear to be sufficient to sustain two airlines with two-hub networks in certain circumstances, a similar result to that of Dobson and Lederer (1993). The two subgame perfect equilibria of this game consist of two duopolies: BA verTable 2

sus KLM, where the competitive subgame equilibrium outcome is (1a), (1b), or (2a); and BA versus Lufthansa, where the competitive subgame equilibrium outcome is (3b). None of the equilibria of the three-airline subgame is sustainable because one airline loses in each case. None of the monopoly situations is sustainable because it is always worthwhile for one of the nonparticipating airlines to join the game. The duopoly game between Lufthansa and KLM is not sustainable because it is BA’s best response strategy to always play. In the BA-KLM duopolistic equilibria, both airlines connect just over half the spokes to their hubs, as shown in Figure 1. BA ignores almost all the ZRH connections and the LHR-AMS path. KLM ignores five connections to AMS and the MAD-ZRH path. Given that each airline uses a different two-HS scenario, the potential direct competition exists only on the paths from the hubs of one airline to the hubs of the other. In this case, KLM connects LHR-AMS and BA connects LHR-MAD and there are no direct connections between ZRH and AMS or MAD. Thus, the two airlines avoid direct competition in all four potential cases. The average plane size for BA is 370 passenger seats, and for KLM 375 passenger seats. BA achieves average load factors of around 90% and KLM achieves approximately 80%. The airfares set by each airline are similar, with business travelers within Europe being charged on average $900 and nonbusiness travelers around $450. The international routes cost the business traveler on average $1,200 and the nonbusiness traveler approximately $800. Table 3 presents the input parameters; Table 4 presents the financial reports of the two airlines; Table 5 presents the frequencies (because demand is fairly symmetric, only one set of frequencies is shown), and seat capacities; and Table 6 presents the summary of airfares charged. It should be noted that in this illustration only two airlines serve the entire Western Euro-

Potential Outcomes in the Doubled-Demand Airline Competition Game Lufthansa

BA

Play

Not

KLM

KLM

Play

Not

Play

Competition: (1a) 150 36 −38 (1b) 159 3 −61 (2a) 89 11 −36 (3b) 116 −27 4

Duopoly: (1a) 176, 0, 92

Not

Duopoly: (2a) 0 58 64 (2b) 0 86 −11 (3a) 0 33 66

Monopoly: (3a) 0 0 238

BA

Play

Not

Play

Duopoly: (2a) 163, 129, 0


Not


No Participation 0 0 0


67


BA KLM

JFK NRT

Figure 1

Duopoly Subgame Perfect Equilibrium Between BA and KLM

pean demand, hence the relatively high frequencies for example. Furthermore, the aircraft seat size in the illustration was limited to between 150 and 401 seats. Table 3

Input Parameters

Parameter input data b nb b nb

Values 12 07 073 03 08 017

In the BA-Lufthansa duopolistic equilibrium, each airline connects just over half the spokes to its respective hubs. BA ignores the LHR-FBU and LHRMAN paths, and ZRH is almost unconnected to the relevant spokes. Lufthansa fully connects BCN, but FRA is connected only to MUN, BCN, and the two international airports. Lufthansa serves the BCN-LHR and BCN-ZRH paths alone and no direct link exists between the LHR-FRA or ZRH-FRA routes, thus once again the airlines have avoided direct competition. BA chooses an average plane size of 390 passenger seats and attains 90% load factors. Lufthansa’s optimal average plane size is 370 passenger seats resulting


68


Table 4

Financial Reports for the Two Airlines

Financial reports in dollars per week Total revenue Total landing charges Total passengers charges Station expenses Total charges from all airports Total operating costs Total costs Gross potential profit

Table 6

BA

BA

KLM

439638368 10822159 6299121

398763680 10880468 4631278

39384596 56505872 (20% of airline’s costs) 220424480 276930368 162708000

35682136 51193884 (19% of airline’s costs) 218794512 269988384 128775296

in load factors of approximately 70%. Once again, the airfares set by each airline are fairly similar to each other and to the previous equilibria. General conclusions that can be drawn from the two equilibria include the following. (a) Under duopolistic conditions, each airline chooses not to compete directly on any one path and reduces the size of its total network to not compete aggressively. (b) The major distinction between the two competitors lies in their choice of network. Most travelers do not have a choice of two airlines in traveling from their origin to their destination. For example, in the BA-KLM duopoly situation, if no interlining is permitted, only two intra-European traveler choices exist, between LHR-MAD and MAD-FCO. Passengers traveling LHR-MAD can fly directly with BA or indirectly

Table 5

Summary of Frequencies and Seat Capacity for the Two Airlines Per Week BA

Frequencies per week LHR-AMS: LHR-ARN: LHR-BRU: LHR-CDG: LHR-CPH: LHR-FBU: LHR-MAD: LHR-MAN: LHR-JFK: LHR-NRT: ZRH-ATH: ZRH-BCN: ZRH-FCO: ZRH-FRA: ZRH-GVA: ZRH-LIN: ZRH-MUC: ZRH-VIE: ZRH-LHR:

00 1400 1348 1400 1400 1400 1268 24 1400 1336 00 00 1345 00 00 1220 00 00 1400

KLM Seat capacity — 355 346 401 353 353 352 385 401 304 — — 314 — — 318 — — 401

Frequencies per week MAD-BCN: MAD-CDG: MAD-FCO: MAD-GVA: MAD-LIN: MAD-MUC: MAD-VIE: MAD-ZRH: AMS-ARN: AMS-ATH: AMS-BRU: AMS-CPH: AMS-FBU: AMS-FRA: AMS-LHR: AMS-MAN: AMS-JFK: AMS-NRT: AMS-MAD:

Seat capacity 1400 1377 1400 928 1400 1006 647 00 00 703 00 52 00 1400 1395 00 1400 1303 1400

Summary Description of Airfares in U.S. Dollars

363 323 401 383 401 378 384 — — 391 — 382 — 353 376 — 401 310 401

EU

Average Standard deviation

KLM

International

EU

International

B

NB

B

NB

B

NB

B

NB

917 292

492 166

1251 534

837 189

860 410

483 203

1201 327

836 159

via AMS with KLM. Passengers flying MAD-FCO can travel directly with KLM or indirectly via LHR with BA. The only other real competition lies on the international routes, which are connected to LHR and AMS. However, the “Northern” airports (ARN, CPH, and FBU), MAN, and BRU are only connected to LHR, and most of the “Eastern” airports are connected via KLM alone, including ATH, GVA, MUC, and VIE. Thus, competition has been severely limited in this solution outcome. (c) The prices offered by each airline are fairly similar, once again enabling the airlines not to compete aggressively. (d) The load factors achieved are very high in comparison to real-world statistics today, thus each airline is relatively efficient. In addition, we have analyzed a four-airline competitive subgame that represents a simplification of reality. Each airline has a pure HS network based at one of the four major Western European hubs: AMS, CDG, FRA, and LHR. The results are presented in Table 7. In all cases, except Case 2a, BA alone achieves a profitable status. Despite the doubled demand, the addition of a fourth airline, and the use of a single hub, not more than one airline manages to profit from the marketplace. According to the frequency decision variables, BA chooses not to serve six intraEuropean routes, thus a merger with one or more of the other airlines would enable BA to attract more passengers. In Case 2a, three airlines achieve profits, suggesting that a three-airline competitive market could exist. Only KLM fails to reach a profitable status, with almost zero frequency on all routes. BA serves only JFK and Manchester airports, becoming Table 7

Potential Outcomes in a Four-Airline Subgame Potential airline profit in millions of dollars per week

Airline

Hub

(1a)

(2a)

(3b)

(4a)

KLM Air France Lufthansa BA

AMS CDG FRA LHR

−11 −86 −9 105

−9 55 35 46

−104 −33 −131 170

−80 −11 −10 242


69


a specialist serving mainly the business community, charging high fares for transatlantic flights, and not competing with either of the two remaining airlines. Lufthansa serves most of the cities, while Air France connects half of the spokes to the hub. In other words, demand is sufficient to sustain one specialist airline, one small airline (one hub connected to nine airports), and one large airline that connects 18 of the 20 airports in the sample. Clearly, however, Case 2a is not an equilibrium solution because KLM attains losses. It should be recalled that we have analyzed the solutions of the subgame in this instance and not the subgame perfect equilibria of the entire game.

7.

Summary

In the model framework presented in this paper, an airline profit-maximizing objective function contains a discrete choice model, in which each airline’s market share is computed based on a function of frequency, airfare, value of time, and the decision variables of the other airlines. Each airline’s decision variables include hub choice and subsequently frequency, aircraft size, and airfare. To place the airlines’ decisions in a competitive context, a game-theoretic approach has been developed in the form of a two-stage, noncooperative game. In the first stage, the airlines simultaneously choose their network and whether or not to participate. In the second stage, each airline competes for market share, given the other airlines’ decisions. The iterative process of backward induction evaluates each airline’s best-response strategy. An equilibrium does not necessarily exist, because the payoff function is not concave. The subgame, best-response program consists of a nonlinear objective function and linear constraints. Locally optimal solutions can be found, but global equilibria cannot be guaranteed due to the nonconvexity of the objective function. It is not possible to analyze all potential, alternative competitors within the scope of this research, thus we developed a three-airline game with specific two-hub networks to evaluate whether one or more equilibria exist. In the base run, a single, monopolistic subgame perfect equilibrium existed in which the only airline to achieve profitability and stability was based at LHR-ZRH. In the doubled-demand sensitivity analysis, two-duopolistic subgame perfect equilibria exist. In other words, with sufficient demand, two airlines can organize their networks to attain profitability; it is not worthwhile for either to change its strategy or for a third airline to enter the market. These results validate the theoretical conclusions reached in Hendricks et al. (1999) as well as those of Dobson and Lederer (1993), in which both monopolistic and duopolistic equilibria are discovered in a six-node example. In the duopoly equilibria outcomes, the LHR-ZRH

hub–based network are also always partners, indicating that demand plays an important role in the solution outcome, especially since LHR is geographically placed on the outer edge of Western Europe. Clearly, London is the largest city in Western Europe, with approximately 7.4 million people, with Berlin a distant second with approximately 3.3 million. Even accounting for the transfer phenomenon resulting from the hub network (for example, 60% of AMS passengers today are transfer passengers), this demand disparity affects the results and ensures robustness of the solution set. All the results presented here are based on specific HS networks. Thus, future research attempting to choose a new HS network needs to consider all possible enumerations to pick the network with the greatest possibility of survival and profitability. An alternative approach would be to model the existing aviation situation and aid one of the airlines in adapting their network, for example choosing a secondary hub, to maximize their profits, given their competitors’ HS network combinations. Finally, given the international alliances occurring today, it may be of interest to extend this model to account for several hubs on different continents. Acknowledgments

The author would like to thank the anonymous reviewers for excellent comments and the Recanati Foundation for partial funding.

Appendix A.

Airport Sample Set

The airports that are considered in the mathematical models and results include the following. European AMS ARN ATH BCN BRU CDG CPH FBU FCO FRA GVA LHR LIN MAD MAN MUC VIE ZRH

Airports Schiphol, Amsterdam Stockholm Athens Barcelona Brussels Charles de Gaulle, Paris Copenhagen Oslo Fumicino, Rome Frankfurt Geneva London Heathrow Milan Madrid Manchester Munich Vienna Zurich

Non-European Airports JFK John F. Kennedy, New York NRT Narita, Tokyo

0 502 617 501 480 1550 615 395 903 717 520 4144 1050 729 556 645 555 1044 3538 1367

503 0 0 72 402 784 2613 1968 20 391 339 1658 381 138 13 116 278 649 594 255

ARN

587 0 0 115 619 401 275 9 1940 750 364 1829 773 301 36 356 494 812 867 310

ATH

Business Demand Data

522 109 119 0 514 2 271 42 776 517 417 1364 1065 8454 97 290 268 550 572 127

BCN 484 407 645 502 0 1007 585 298 929 812 865 3102 1192 1004 323 507 484 887 2206 475

BRU 1501 745 495 6 615 0 776 513 2830 1289 1976 8331 2894 50 575 1230 1039 1781 4099 2468

CDG 616 2542 311 283 581 974 0 2369 336 523 276 1573 578 506 262 330 404 644 872 972

CPH 392 1904 17 22 304 523 2394 0 66 222 45 1593 172 64 0 13 0 336 499 49

FBU 910 22 2056 782 916 2712 321 61 0 844 691 3252 10655 1689 140 589 515 1043 3465 727

FCO 721 424 797 527 803 1327 541 224 859 0 505 2487 1193 769 315 2900 1217 1053 4205 1961

FRA

Passenger Origin-Destination Demand Matrices for 1992

Source. Adler and Berechman (1996).

AMS ARN ATH BCN BRU CDG CPH FBU FCO FRA GVA LHR LIN MAD MAN MUC VIE ZRH JFK NRT

AMS

Table B.1

Appendix B.

508 358 406 468 849 1885 315 44 676 494 0 2710 206 518 164 207 197 992 287 127

GVA 4198 1585 1976 1360 3068 8805 1590 1550 3037 2405 2721 0 3169 2316 2188 2452 1819 3877 11527 4904

LHR 1025 309 806 982 1171 2900 571 152 10157 1197 215 3171 0 1390 361 431 483 1155 391 218

LIN 730 166 297 8477 986 35 480 53 1689 779 582 2399 1457 0 182 444 430 756 1981 184

MAD 579 26 46 53 323 592 252 0 144 319 191 2188 371 137 0 217 79 462 1011 137

MAN

636 71 354 276 461 1171 367 10 621 2928 200 2386 400 412 194 0 666 993 1032 165

MUC

531 278 544 216 439 998 381 0 496 1229 188 1887 524 403 66 641 0 1649 787 248

VIE

1104 668 863 490 873 1664 621 313 1091 1043 813 3901 1078 768 424 1002 1657 0 2434 660

ZRH

3816 605 934 619 1668 4212 1265 489 3879 4659 540 12659 706 2083 1144 1402 1046 2336 0 7212

JFK

1368 217 324 197 724 2757 1048 59 969 2140 195 5637 386 405 261 459 368 1036 7133 0

NRT

70 Transportation Science 39(1), pp. 58–72, © 2005 INFORMS


0 215 264 215 206 664 264 169 387 308 223 1776 450 312 238 277 238 448 1516 586

215 0 0 31 172 336 1120 843 8 168 145 711 163 59 6 50 119 278 255 109

ARN

252 0 0 49 265 172 118 4 831 322 156 784 331 129 15 153 212 348 371 133

ATH

224 47 51 0 220 1 116 18 333 222 179 584 456 3623 42 124 115 236 245 55

BCN

Nonbusiness Demand Data

Source. Adler and Berechman (1996).

AMS ARN ATH BCN BRU CDG CPH FBU FCO FRA GVA LHR LIN MAD MAN MUC VIE ZRH JFK NRT

AMS

Table B.2

208 174 276 215 0 432 251 128 398 348 371 1329 511 430 138 218 208 380 946 204

BRU 643 319 212 2 264 0 333 220 1213 552 847 3570 1240 21 247 527 445 764 1757 1058

CDG 264 1090 133 121 249 417 0 1015 144 224 118 674 248 217 112 142 173 276 374 417

CPH 168 816 7 9 131 224 1026 0 28 95 19 683 74 27 0 6 0 144 214 21

FBU 390 9 881 335 393 1162 138 26 0 362 296 1394 4566 724 60 252 221 447 1485 312

FCO 309 182 342 226 344 569 232 96 368 0 216 1066 511 330 135 1243 522 451 1802 840

FRA 218 153 174 201 364 808 135 19 290 212 0 1161 88 222 70 89 85 425 123 54

GVA 1799 679 847 583 1315 3774 681 664 1302 1031 1166 0 1358 992 938 1051 780 1662 4940 2102

LHR 439 132 345 421 502 1243 245 65 4353 513 92 1359 0 596 155 185 207 495 167 93

LIN 313 71 127 3633 423 15 206 23 724 334 249 1028 624 0 78 190 184 324 849 79

MAD 248 11 20 23 138 254 108 0 62 137 82 938 159 59 0 93 34 198 433 59

MAN 272 30 152 118 198 502 157 4 266 1255 86 1023 172 177 83 0 285 425 442 71

MUC

228 119 233 92 188 428 163 0 213 527 81 809 224 173 28 275 0 707 337 106

VIE

473 286 370 210 374 713 266 134 467 447 349 1672 462 329 182 430 710 0 1043 283

ZRH

1635 259 401 265 715 1805 542 210 1662 1997 231 5425 303 893 490 601 448 1001 0 3091

JFK

586 93 139 84 310 1182 449 25 415 917 84 2416 166 174 112 197 158 444 3057 0

NRT



71

72 References Adler, N. 2001. Competition in a deregulated air transportation market. Eur. J. Oper. Res. 129(2) 337–345. Adler, N., J. Berechman. 1996. Optimal hub policy affecting aircraft movement and hub choice by airlines. Final Report 2 Vol. The Netherlands Ministry of Transportation, Directorate General of Aviation. Adler, N., J. Berechman. 2001. Evaluating optimal multi-hub networks in a deregulated aviation market with an application to Western Europe. Transportation Res. A 35(5) 337–345. Adler, N., B. Golany. 2001. Evaluation of deregulated airline networks using data envelopment analysis combined with principal component analysis with an application to Western Europe. Eur. J. Oper. Res. 132(2) 260–273. Alamdari, F. E., I. G. Black. 1992. Passengers’ choice of airline under competition: The use of the logit model. Transport Rev. 12(2) 153–170. Bailey, E., D. Graham, D. Kaplan. 1985. Deregulating the Airlines. MIT Press, Cambridge, MA. Bazaraa, M. S., H. D. Sherali, C. M. Shetty. 1993. Nonlinear Programming: Theories and Algorithms. Wiley and Sons, New York. Ben-Akiva, M., S. R. Lerman. 1985. Discrete Choice Analysis: Theory and Application to Travel Demand. MIT Press, Cambridge, MA. Bittlingmayer, G. 1990. Efficiency and entry in a simple airline network. Internat. J. Indust. Econom. 8 245–257. Borenstein, S. 1989. Hubs and high fares: Dominance and market power in the U.S. airline industry. RAND J. Econom. 20(3) 344–365. Borenstein, S. 1992. The evolution of U.S. airline competition. J. Econom. Perspectives 6 45–74. Brueckner, J. K., P. T. Spiller. 1994. Economies of traffic density in the deregulated airline industry. J. Law Econom. 37 379–415. Button, K., D. Pitfield. 1991. Transport Deregulation. Macmillan, London, U.K. Call, G. D., T. E. Keeler. 1985. Airline deregulation, fares and market behavior: Some empirical evidence. A. F. Daughety, ed. Analytical Studies in Transport Economics. Cambridge University Press, Cambridge, U.K., 221–247. Dobson, G., P. J. Lederer. 1993. Airline scheduling and routing in a hub-and-spoke system. Transportation Sci. 27(3) 281–297. Ghobrial, A., A. Kanafani. 1985. Airline hubbing: Some implications for airline economics. Transportation Res. 18A(4) 15–27.



Goldfarb, D. 1969. Extension of Davidon’s variable metric method to maximization under linear inequality and equality constraints. SIAM J. Appl. Math. 17 739–762. Hansen, M. 1990. Airline competition in a hub-dominated environment: An application of noncooperative game theory. Transportation Res. 24B(1) 27–43. Hendricks, K., M. Piccione, G. Tan. 1992. The economics of hubs— The case of monopoly. Rev. Econom. Stud. 62 83–99. Hendricks, K., M. Piccione, G. Tan. 1999. Equilibria in networks. Econometrica 67(6) 1407–1434. Hong, S., P. T. Harker. 1992. Air traffic network equilibrium: Toward frequency, price and slot priority analysis. Transportation Res. 26B(4) 307–323. Johnson, R. L. 1985. Networking and market entry in the airline industry. J. Transport Econom. Policy 19(3) 299–304. Keeler, T. E. 1991. Airline deregulation and market performance: The economic basis for regulatory reform and lessons from the U.S. experience. D. Banister, K. Button, eds. Transport in a Free Market Economy. Macmillan, London, U.K., 121–169. Kreps, D. M. 1990. A Course in Microeconomic Theory. Harvester Wheatsheaf, New York. Marianov, V., D. Serra, C. ReVelle. 1999. Location of hubs in a competitive environment. Eur. J. Oper. Res. 114 363–371. Morrison, S., C. Winston. 1986. The Economic Effects of Airline Deregulation. The Brookings Institution, Washington, D.C. Nero, G. 1996. A structural model of intra European Union duopoly airline competition. J. Transport Econom. Policy 30 137–155. Reynolds-Feighan, J., J. D. Hewings. 1990. Airline network structure and regional economic development—U.S. case studies. K. Peschel, ed. Infrastructure and the Space Economy. SpringerVerlag, New York. Spiller, P. 1989. A note on pricing of hub-and-spoke network. Econom. Lett. 30 165–169. Starr, R., M. Stinchcombe. 1992. Efficient transportation routing and natural monopoly in the airline industry: An economic analysis of hub-spoke and related systems. Working paper 92-25, Department of Economics, University of California, San Diego, CA. Von Neuman, J., O. Morgenstern. 1953. Theory of Games and Economic Behavior. Princeton University Press, Princeton, NJ. Zang, I. 1980. A smoothing-out technique for min-max optimization. Math. Programming 19 61–77.