THREE ESSAYS IN INDUSTRIAL ORGANIZATION: ALLIANCES, MERGERS, AND PRICING IN COMMERCIAL AVIATION DAVID R. BROWN. B.A., Hastings College, 2005

Transcription

1 THREE ESSAYS IN INDUSTRIAL ORGANIZATION: ALLIANCES MERGERS AND PRICING IN COMMERCIAL AVIATION by DAVID R. BROWN B.A. Hastings College 2 AN ABSTRACT OF A DISSERTATION submitted in artial fulfillment of the reuirements for the degree DOCTOR OF PHILOSOPHY Deartment of Economics College of Arts and Sciences KANSAS STATE UNIVERSITY Manhattan Kansas 2

2 Abstract My research focuses rimarily on industrial organization and alied microeconomics. Secifically I have extensively studied the airline industry. My first essay considers the effect of the Delta/Continental/Northwest codeshare alliance. Codeshare agreements can benefit airlines due to network exansion and benefit consumers by eliminating a double marku on flight itineraries with multile oerating carriers. However olicymakers have exressed concern that an alliance between airlines may facilitate rice and service collusion in markets where codeshare artners services overla. I develo a structural econometric model that is able to searately identify suly and demand factors as sources of rice-uantity changes caused by the creation of the alliance. The estimates from the model show both collusive and demand increasing effects associated with the codeshare alliance. However the demand increasing effect is larger than the collusive effect. My second essay considers the effects of the recent Delta/Northwest merger. This merger is of articular interest because the two airlines are codeshare artners. Using re-merger data a counterfactual simulation is erformed in which Delta and Northwest are assumed to merge. The results indicate that codeshare roducts owned by the merging firms exerience higher redicted rice increases relative to ure online roducts. In addition the mean redicted rice increases are relatively small across most markets. I also examine re-merger redictions with -merger data and analysis and find that the re-merger redictions roughly accord with de-merger simulated effects using -merger data. My third essay takes an extended look at airline mergers. When the Delta/Northwest merger was aroved by the Deartment of Justice consumer grous and olicymakers were concerned that the merger and oor economic outlook would act as a catalyst for more mergers. This aer examines this ossible scenario using simulations to model the effects of other codeshare artners merging in addition to Delta and Northwest. Results indicate that the redicted rice increases for all mergers exhibit relatively small averages but large variances across markets. Further the largest redicted rice increases affect a small ercent of roducts and an even smaller ercent of assengers who choose roducts owned by a merging firm.

3 THREE ESSAYS IN INDUSTRIAL ORGANIZATION: ALLIANCES MERGERS AND PRICING IN COMMERCIAL AVIATION by DAVID R. BROWN B.A. Hastings College 2 A DISSERTATION submitted in artial fulfillment of the reuirements for the degree DOCTOR OF PHILOSOPHY Deartment of Economics College of Arts and Sciences KANSAS STATE UNIVERSITY Manhattan Kansas 2 Aroved by: Maor Professor Phili G. Gayle

4 Coyright DAVID R. BROWN 2

5 Abstract My research focuses rimarily on industrial organization and alied microeconomics. Secifically I have extensively studied the airline industry. My first essay considers the effect of the Delta/Continental/Northwest codeshare alliance. Codeshare agreements can benefit airlines due to network exansion and benefit consumers by eliminating a double marku on flight itineraries with multile oerating carriers. However olicymakers have exressed concern that an alliance between airlines may facilitate rice and service collusion in markets where codeshare artners services overla. I develo a structural econometric model that is able to searately identify suly and demand factors as sources of rice-uantity changes caused by the creation of the alliance. The estimates from the model show both collusive and demand increasing effects associated with the codeshare alliance. However the demand increasing effect is larger than the collusive effect. My second essay considers the effects of the recent Delta/Northwest merger. This merger is of articular interest because the two airlines are codeshare artners. Using re-merger data a counterfactual simulation is erformed in which Delta and Northwest are assumed to merge. The results indicate that codeshare roducts owned by the merging firms exerience higher redicted rice increases relative to ure online roducts. In addition the mean redicted rice increases are relatively small across most markets. I also examine re-merger redictions with -merger data and analysis and find that the re-merger redictions roughly accord with de-merger simulated effects using -merger data. My third essay takes an extended look at airline mergers. When the Delta/Northwest merger was aroved by the Deartment of Justice consumer grous and olicymakers were concerned that the merger and oor economic outlook would act as a catalyst for more mergers. This aer examines this ossible scenario using simulations to model the effects of other codeshare artners merging in addition to Delta and Northwest. Results indicate that the redicted rice increases for all mergers exhibit relatively small averages but large variances across markets. Further the largest redicted rice increases affect a small ercent of roducts and an even smaller ercent of assengers who choose roducts owned by a merging firm.

6 Table of Contents List of Figures...ix List of Tables... x Acknowledgements... xi CHAPTER - Airline Strategic Alliances in Overlaing Markets: Should Policymakers be Concerned?.... INTRODUCTION DEFINITIONS MODEL... 7 Demand... 7 Suly... Cometitive Interactions... 3 Secializing the Suly Euation..... Estimation... 8 Instruments Data Results... 2 Results from the System GMM Estimation Counterfactual Simulations Conclusions... 3 CHAPTER 2 - Analyzing the Effect of a Merger between Airline Codeshare Partners Introduction The Model... 2 Definitions... 2 Demand... 3 Product Pricing and Marginal Cost Estimation... Instruments Methodology used for Merger Analysis... 8 vi

7 2.. Data... 9 Collasing the Data... Creation of Other Variables Results... 2 Variable signs and significance... 3 Demand Elasticities... Merger Analysis... Post-Merger Data Analysis Conclusions... CHAPTER 3 - Mergers Triggering Mergers: A Simulation Analysis of Merging Airline Codeshare Partners Introduction and Background The Model... 7 Definitions... 7 Demand... 8 Suly Estimation... 7 Instruments Data Results... 7 Demand Parameters... 7 Merger Analysis Conclusions Weaknesses of the model Policy Imlications & Future Research... 8 References... 8 Aendix A - Three Airline Single Market Examle Strategic Interaction in the Pre-Alliance Period Strategic Interaction in the Post-Alliance Period Strategic interaction in market when v : Strategic interaction in market when v :... 9 vii

8 Aendix B - Additional Chater 2 Tables Aendix C - Additional Chater 3 Tables... 9 viii

9 List of Figures Figure. Route Network Diagram... Figure.2 Modified Route Network Diagram... Figure 2. Distribution of Predicted Price Increases... 8 Figure 2.2 Distribution of Predicted Price Increases among Codeshare Tyes... ix

10 List of Tables Table. List of Airlines in the Dataset Table.2 Summary statistics broken down by time eriod and market tye where market tye is based on whether or not DL CO and NW virtual codeshare together in the market... 2 Table.3 Summary statistics for roducts offered by DL CO or NW. These statistics are broken down by time eriod and market tye where market tye is based on whether or not DL CO and NW virtual codeshare together in the market... 2 Table. Demand Parameter Estimates from Single Euation Estimation... 2 Table. Demand and Suly Parameter Estimates... 3 Table 2. Airlines reresented in the dataset... Table 2.2 Dual Presence of Delta and Northwest in Markets... 2 Table 2.3 Summary Statistics... 2 Table 2. Demand Regression Results... Table 2. Summary Statistics for Merger Analysis... 7 Table 2. Decomosition of Predicted Price Increases... 9 Table 2.7 Determinants of Predicted Price Increases and Pre-merger Prices when Delta/Northwest Codeshare Products are Decomosed into Grous... 2 Table 2.8 Percent Predicted Price Increases at Various Market Percentiles... 3 Table 2.9 Summary Statistics for Post Merger Data... Table 3. Airlines reresented in the dataset Table 3.2 Summary Statistics... 7 Table 3.3 Demand Regression Results... 7 Table 3. Summary Statistics for Merger Variables Table 3. Decomosition of Predicted Price Increases Among Merger Products... 8 Table 3. Distribution of Predicted Price Increases... 8 Table B. Instrument Validity Test Table B.2 Post-Merger Data Demand Results Table C. Instrument Validity Test... 9 Table C.2 List of Hubs for each Airline... 9 x

11 Acknowledgements I would like to thank my maor advisor Phili Gayle for assisting me in finding a starting oint and for roviding excellent guidance throughout the dissertation rocess. In addition he heled me stay motivated to work hard and got me involved in resenting my research at various conferences. I greatly thank my wife Jaime for her direct hel by assisting with data collection and roofreading. I also thank her for her indirect hel through all the love and suort she gave to me throughout the busiest and most difficult moments of research. I thank my arents Mark and JoAnn and my brother Kerry for instilling a great work ethic in me allowing me to attend graduate school in the first lace and ultimately finish my degree. I thank Casey Abington for being a great friend officemate and classmate. Finally I thank the rest of my suervisory committee (Dennis Weisman Tian Xia and Dong Li) for taking time to offer helful suggestions during my roosal and final defense. I also would like to acknowledge and thank Lawrence White Robert Porter Volodymyr Bilotkach Karunakaran Sudhir the articiants at the 29 International Industrial Organization Conference the articiants at the 29 Missouri Economics Conference the articiants at the 29 Southern Economics Association Conference and two anonymous referees for helful insights and suggestions. xi

12 CHAPTER - Airline Strategic Alliances in Overlaing Markets: Should Policymakers be Concerned?. INTRODUCTION Policy analysts have exressed sketicism when reviewing airlines' alications to form a codeshare alliance in the event that such an alliance involves otential artners that have significant overla in their route networks. The heart of the concern is that these otential artners are direct cometitors in the segments of their networks that overla and an alliance between them which often reuires broad discussions between artners to make their interline service seamless could facilitate collusion (exlicit or tacit) on rices and/or service levels in the artners' overlaing markets. Before ultimately aroving the Delta/Continental/Northwest alliance which was formed in June 23 the U.S. Deartment of Transortation (DOT) exressed these concerns. 2 The DOT's review of this roosed alliance oints out that the three airlines' service overla in 32 markets accounting for aroximately 8 million annual assengers which is in contrast to the next largest alliance between United Airlines and US Airways with overlaing service in only 3 markets accounting for. million annual assengers. So unlike much of the earlier literature that focuses on international airline alliances [Brueckner (23) Brueckner and Whalen (2) Bilotkach (27)] we focus on a U.S. domestic alliance [Bamberger Carlton and Neumann (2) Armantier and Richard (2) Armantier and Richard (28) Ito and Lee (27)]. The central uestion that this aer sets out to answer is: Has the Delta/Continental/Northwest alliance reduced cometition between these carriers in their overlaing markets? Using a reduced-form econometric model similar to that in Bamberger Carlton and Neumann (2) Gayle (28) has shed some light on this uestion. In articular Gayle (28) found that the alliance is associated with a marginal rice increase which by itself oints to ossible collusive effects. But a marginal rice increase is also consistent with increased demand and there is good reason to believe that an alliance has a demand-increasing Interline means that at some oint in the tri when assengers change lanes they also change airlines. 2 See "Termination of review under 9U.S.C. 72 of Delta/Northwest/Continental Agreements" ublished by Office of the Secretary Deartment of Transortation January 23.

13 effect associated with it. For examle assengers that are members of an airline's freuent-flyer rogram may cumulatively earn and redeem freuent-flyer miles across any artner in the alliance. The new oortunities for assengers to earn and redeem miles will likely increase demand for the alliance artners' roducts. In fact Gayle (28) also found that the alliance is associated with increased traffic (measured by number of assengers) a finding which casts doubt on the effectiveness of collusive behavior if at all resent. It is imortant to note however that Gayle's (28) conecture about the ossible absence of collusive effects is based on the estimated change in euilibrium rice and uantity over the re- and -alliance eriods. While the aroach is an informative and suggestive first ste it is not definitive in regards to ruling out collusive effects unless we are able to disentangle the demand and suly sources of the euilibrium rice-uantity changes. In other words we cannot be certain that the rice and uantity increases are solely driven by demand-increasing factors of the codeshare alliance. If collusive effects are also resent then the uantity increase is smaller and rice increase larger than they would have been were it not for collusion. So the main contribution of our resent aer is to derive and estimate a structural econometric model that allows us to disentangle the demand and suly sources of euilibrium rice-uantity changes that are associated with a codeshare alliance. Second the ability to do welfare analysis in a straightforward way is another advantage of the structural econometric model in this aer comared to the reduced-form econometric model in Gayle (28). Third since our econometric model can be alied to study the market effects of codeshare alliances other than the Delta/Continental/Northwest alliance this aer contributes to the methodology of analyzing the cometitive effects of codeshare alliances. Our key findings are as follows: First the econometric estimates for the air travel demand euation suggest that the Delta/Continental/Northwest alliance has a demand-increasing effect associated with it. Second the econometric estimates for the air travel suly euation suggest that the alliance is associated with a softening of cometition between the three airlines in their overlaing markets relative to cometition between other cometing airlines in these markets. To the best of our knowledge this is the first aer to exlicitly isolate and find evidence of a collusive effect associated with a domestic codeshare alliance. We must oint out that we are not the first to resent a structural econometric model to examine codeshare alliances. Armantier and Richard (28) also use a structural econometric 2

14 model to examine a codeshare alliance. However a fundamental difference between our model and the model in Armantier and Richard (28) is that we model both demand and suly asects of codesharing while Armantier and Richard (28) only model the demand side. This crucial methodological difference affords us the advantage of being able to searately identify demand and suly effects of codesharing which further allows us to more meticulously examine short-run market effects within a market euilibrium framework. For examle we use our model to erform counterfactual exeriments by removing the demand-increasing and collusive effects of the alliance to see how the artners' euilibrium rices number of assengers and consumer surlus would be affected. These exeriments reveal that the demand-increasing and collusive effects of codesharing reinforce each other in their influence of the artners' rices with collusive effects accounting for a larger ortion of the relatively small rice increases in the case of the Delta/Continental/Northwest codeshare alliance. On the other hand the demand-increasing and collusive effects of codesharing oose each other in their influence on the artners' number of assengers with the demand-increasing effect dominating and leading to increases in the artners' number of assengers. The exeriments also reveal that the codeshare alliance seems to have imroved consumer surlus owing to the domination of demand-increasing effects over collusive effects which should be reassuring to olicymakers who aroved the alliance. Armantier and Richard (28) examine the effects on consumer surlus resulting from the earlier Continental/Northwest codeshare alliance imlemented in 999. They find that the Continental/Northwest alliance increased the welfare of assengers on connecting flights but decreased the welfare of assengers on nonsto flights which resulted in no significant imact on overall consumer surlus. However there is a fundamental difference in how we simulate and measure welfare effects versus how these effects are measured in Armantier and Richard (28). Essentially Armantier and Richard (28) use their demand model to estimate and comare consumer surlus in the re- and -alliance eriods in markets where the two airlines codeshare versus markets in which the two airlines do not codeshare. Thus any temoral change in consumer surlus that is different across codeshare versus non-codeshare markets is attributed to the alliance. In our aroach we first econometrically identify both demand and suly effects of codesharing which then gives us the advantage of erforming well-controlled counterfactual welfare exeriments. For examle in one counterfactual exeriment we hold 3

15 market conditions in the -alliance eriod constant and simulate new market euilibria in the event that only the demand-increasing effect of codesharing is artificially removed. As such for a given market at a given oint in time by comaring actual market euilibrium to simulated market euilibrium we can assess how consumer surlus would change if only the demandincreasing effect of codesharing is removed. The rest of the aer is organized as follows: In the next section we make some key definitions which build the foundation for imortant issues we subseuently model analyze and discuss. In Section.3 we resent the structural econometric model while the estimation strategy is discussed in Section.. We discuss the characteristics of our data in Section. and results are shown and discussed in Section.. Concluding remarks are offered in Section DEFINITIONS A market is defined as directional round-tri air travel between an origin and a destination city during a articular eriod. The assumtion that markets are directional imlies that a round-tri air travel from Atlanta to Detroit is a distinct market comared to round-tri air travel from Detroit to Atlanta. Furthermore this directional assumtion allows for the ossibility that origin city characteristics may influence market demand [see Gayle (27a 27b 27c) Berry Carnall and Siller (2)]. A flight itinerary is defined as a secific seuence of airort stos while traveling from the origin to destination city. An air travel roduct is defined as a uniue combination of airline(s) and flight itinerary. Following Ito and Lee (27) a ure online roduct means that the same airline markets and oerates all segments of a round tri. For examle three searate ure online roducts are: () a non-sto round tri from Atlanta to Detroit marketed and oerated by Delta Air Lines; (2) a round tri from Atlanta to Detroit with one sto in Minneaolis marketed and oerated by Delta Air Lines; and (3) a non-sto round tri from Atlanta to Detroit marketed and oerated by Northwest Air Lines. Note that all three roducts are in the same market. A codeshare agreement effectively allows one carrier (called the "ticketing carrier" or "marketing carrier") to sell seats on its artners' lanes as if these seats are owned by the carrier selling the seats. The carrier whose lane that actually transorts the assenger is referred to as the "oerating carrier". For examle Northwest may sell tickets for a subset of seats on the Delta oerated flight between Atlanta and Detroit as if the lane were owned by Northwest. So a

16 assenger that uses a codeshare itinerary may have urchased the round tri ticket from Northwest but actually flies on a Delta lane. The tye of codeshare roduct we focus on in this research is referred to as "virtual" codeshare. 3 A assenger using a virtual codeshare itinerary remains on a single oerating carrier's lane(s) for the entire round tri but the ticket for the tri was marketed and sold by a artner ticketing carrier. We focus on virtual codeshare roducts because Gayle (28) found that this was the only tye of codeshare roduct that is strongly associated with rice increases. Figure. gives an examle in which two airlines' route networks overla and they may virtual codeshare together in the origin-destination market. The figure shows that Northwest and Delta both oerate non-sto flights in the Atlanta to Detroit market. If they virtual codeshare together in this market then a subset of the assengers on the Delta lane would have bought their tickets from Northwest while a subset of the assengers on the Northwest lane would have bought their tickets from Delta. Figure. Route Network Diagram Figure.2 shows an alternate situation in which the airlines' route networks may overla. In Figure.2 Northwest oerates a non-sto flight in the Atlanta to Detroit market while Delta oerates a one-sto itinerary in the Atlanta to Detroit market but unlike Figure. Delta does 3 See Gayle (28) and Ito and Lee (27) for discussions of the main tyes of codeshare roducts in the U.S. domestic market.

17 not oerate a non-sto flight in this market. Northwest and Delta's networks are still considered to be overlaing in Figure.2 even though Delta oerates a one-sto itinerary while Northwest oerates a non-sto itinerary. Both carriers may virtual codeshare together in Figure.2. Figure.2 Modified Route Network Diagram In Figure.2 it might seem counter-intuitive that a assenger would choose a one-sto itinerary even though a non-sto flight between the origin and destination is available. However assengers often choose less convenient routes (flight itineraries that reuire intermediate stos) to get from their origin to destination when such alternate routing is cometitively riced. In other words within reasonable bounds some assengers are willing to trade-off travel itinerary convenience for a lower rice. Figure.2 can also be used to illustrate a situation in which virtual codesharing is likely to have a demand-increasing effect associated with it. In the event that Northwest and Delta do not have a codeshare alliance Northwest can only offer its Atlanta-based customers (some of whom may be members of Northwest's freuent-flyer rogram) a non-sto flight to Detroit. However an alliance with Delta allows Northwest to offer its Atlanta-based customers both a non-sto flight on its own lane and a one-sto virtual codeshare itinerary oerated solely by Delta. While assengers in Atlanta already had the otion rior to an alliance to urchase either a ure online one-sto itinerary from Delta or a ure online non-sto flight from Northwest

18 Northwest's freuent-flyers could not accumulate freuent-flyer miles on the Delta oerated flights. Thus the alliance created a new oortunity for Northwest freuent-flyers to accumulate miles on a Delta oerated one-sto itinerary. Similarly Delta freuent-flyers that would like to travel on the non-sto Northwest flight also have a new oortunity to accumulate freuent-flyer miles on the Northwest oerated flight. The new oortunity for assengers to accumulate freuent-flyer miles across artner carriers is one reason we exect a demand-increasing effect to be associated with a codeshare alliance. Our econometric model is designed to isolate and test for this otential demand-increasing effect. Figure.2 is also illustrates the main concern that the DOT exressed in its review of the roosed alliance between Delta Continental and Northwest. Since Delta and Northwest were cometitors in the market shown in Figure.2 the DOT was concerned that forming an alliance could reduce how fiercely the two airlines comete with each other. The econometric model we resent below is designed to isolate and test how cometitive interactions between the three airlines differed in the -alliance eriod comared to the re-alliance eriod..3. MODEL We roceed by first describing the demand side of the model. The suly side is then laid out which is where we model cometitive interactions between airlines. Demand In the sirit of Peters (2) Berry and Jia (29) Berry Carnall and Siller (2) and Armantier and Richard (28) air travel demand is modeled using a discrete choice framework. Secifically we use a nested logit model. τ faces a choice between Potential assenger i in market l during time eriod J τ + alternatives. There are J τ + alternatives because we allow l assengers the otion ( the outside good) not to choose either one of the l differentiated air travel roducts considered in the emirical model. Since the model focuses on ure online J l τ We concede that a nested logit model is not as flexible and therefore less desirable comared to a random coefficients logit model. However it is well-known that the random coefficients model is more comutationally demanding to estimate relative to the nested logit model. This estimation burden becomes articularly severe in our case since we have a large data set and lan to ointly estimate demand and suly arameters. As such to make estimation feasible we had to choose the nested logit model over the random coefficients logit model. In addition it is subseuently shown that the suly euation is highly nonlinear in its arameters which also makes estimation comutationally burdensome. 7

19 and virtual codeshare roducts the outside good includes interline air travel roducts and other means of getting from the origin to destination city besides air travel. Products in a market are assumed to be organized into G + exhaustive mutually exclusive grous g...g in which the outside good is assumed to be the only member of grou. A grou here refers to the set of roducts offered by an airline within a market. A assenger solves the following otimization roblem: Max Ui l l i l g ( ) i l {... J l } τ δ τ + σζ τ + σ ε τ τ (.) where U iτ l is the level of utility assenger i will obtain if roduct is chosen while δ is the τ l mean level of utility across assengers that consume roduct. δ is a function of the τ l characteristics of roduct which we subseuently describe. ζ is a random comonent of utility that is common to all roducts in grou g whereas the random term ε is secific to roduct and is assumed to have an extreme value distribution. The arameter σ lies between and and measures the correlation of the consumers utility across roducts belonging to the same grou. Since roducts are groued by airlines σ can also be thought of as measuring the correlation of the consumers' utility across roducts offered by a given airline. As σ aroaches the correlation of references among roducts offered by the same airline within a market increases. Conversely as σ decreases the correlation of references for roducts offered by the same airline within a market decreases. The rationale for the roduct grouing structure above is to cature the ossibility that assengers view an airline's roducts as closer substitutes for each other comared to the substitutability of these roducts across airlines [Gayle (27b)]. One reason why this could be the case is that a assenger may be heavily invested (accumulated miles flown) in a given airline's freuent-flyer rogram and therefore on the margin would refer to choose among alternate flights offered by this airline in order to build u accumulated miles towards the reuired threshold necessary for a discounted tri. Second some consumers may ust have a strong brand loyalty to a given airline based on ast exerience. In any event since σ is a arameter we estimate the data will reveal whether or not a sufficient number of assengers are brand-loyal to render σ >. i τ l g i τ l 8

20 The mean level of utility obtained across the oulation of consumers that consume roduct is given by: δ τ l x τ l β α τ l + a r + λ T + λ T DCN + λ T DCN Codeshare _ mkt + ξ 3 2 τ l (.2) where x τ l is a vector of observed roduct characteristics (the number of intermediate stos used by an itinerary an alternate measure of itinerary convenience whether or not the origin is a hub for the carrier offering the roduct for sale and whether or not the roduct is ure online or virtual codeshare) β is a vector of consumer taste arameters (marginal utilities) associated with the roduct characteristics in x is the rice of roduct α is a measure of the τ l τ l marginal utility of rice a r are airline fixed effects where subscrit r indexes ticketing carriers (ticketing carrier dummies) T is a zero-one time dummy eual to if the itinerary occurred in the -alliance eriod DCN is a zero-one dummy eual to if roduct is being offered for sale by either Delta Continental or Northwest Codeshare_mkt is a zero-one dummy eual to if a virtual codeshare roduct between Delta Continental or Northwest was offered in the origin-destination market and ξ τ l catures unobserved (by the econometricians but observed by assengers) roduct characteristics. λ λ 2 and λ 3 are taste arameters to be estimated. λ catures the change in mean utility over the re- and -alliance eriods for roducts offered by airlines other than Delta Continental or Northwest. λ + λ2 catures the change in mean utility over the re- and alliance eriods for roducts offered by Delta Continental or Northwest in markets where any two of the three airlines do not virtual codeshare together. Therefore when we focus on markets in which Delta Continental and Northwest do not virtual codeshare together λ 2 measures how the change in mean utility differs for roducts offered by Delta Continental or Northwest comared to roducts offered by other airlines. λ + + catures the change in mean utility over the re- and -alliance eriods λ2 λ3 for roducts offered by Delta Continental or Northwest in markets where any two of the three airlines virtual codeshare together. Therefore λ 3 measures how the change in mean utility for roducts offered by Delta Continental or Northwest differs in markets where any two of the three airlines virtual codeshare together comared to markets in which they do not virtual 9

21 codeshare together. In other words λ 3 > imlies that virtual codesharing has a demandincreasing effect associated with it which is one of the main hyotheses we want to test. The discussion above reveals that a key comonent of our demand secification that allows us to identify demand effects associated with the Delta/Continental/Northwest codeshare alliance ( λ 3 ) is that euation (2) effectively comares consumers' choice behavior before and after imlementation of the alliance in markets where the three airlines virtual codeshare together ("treatment" markets) versus markets in which they do not virtual codeshare together ("control" markets). A reasonable criticism to raise at this oint is that Codeshare_mkt in euation (2) is not strictly exogenous since airlines choose which markets to codeshare in and which markets not to codeshare in. However one argument in suort of using Codeshare_mkt as is in euation (2) is that airlines' decisions about whether or not to codeshare in a market is likely redetermined at the time when a tyical consumer decides on their utility maximizing roduct. In other words it may be reasonable to assume that Codeshare_mkt is a redetermined variable in euation (2) and therefore estimates of λ 3 will not be biased. We now comlete the derivation of our demand euation. In what follows we dro the market and time subscrits (l and τ ) only to avoid a clutter of notation. Let there be G g roducts in grou g. If roduct is in grou g the well-known formula for roduct s redicted grou share or euivalently the conditional robability of choosing roduct given that grou g is chosen is given by s / g δ ( e D σ ) g where share is D g e G g δ ( σ ). The robability of choosing grou g or euivalently grou g s redicted s g D σ g G σ D + g D σ g (.3)

22 σ Recall that the outside good is the only good in grou therefore D e we normalize the mean utility of the outside good to zero i.e. δ. This imlies that and euation (.3) can be written as s σ g g G + D g D σ g δ. As usual D σ The unconditional robability of choosing roduct or euivalently the redicted market share of roduct is given by s δ (σ ) σ e Dg s / g sg G Dg σ + Dg g or s D σ g e + δ (σ ) G g D σ g Euation (.) is the well-known nested logit formula. Finally the demand for roduct is given by d M s ( d x ξ θ ) (.) where M is a measure of market size which we assume to be the size of the oulation in the origin city ( ) s is the redicted roduct share function described in euation (.) x and are vectors of observed non-rice roduct characteristics and rice resectively ξ is a vector of unobserved (by the researchers) roduct characteristics and θ d ( β α λ σ ) is the vector of demand arameters to be estimated. Suly What is commonly known about how a codeshare agreement works is that the ticketing carrier markets and sets the final rice for the round-tri ticket and comensates the oerating carrier for oerating services rovided. Details on comensation mechanisms actually used by artner airlines are not usually made known to the ublic and may even vary across artnershis.

23 Therefore we face the challenge of coming u with a modeling aroach that catures our basic understanding of what is commonly known about a codeshare agreement without imosing too much structure on a contracting rocess about which we have few facts. We concede that the following is ossibly a simlistic aroximation of the actual contracting used by artners to comensate each other for services needed to rovide a codeshare roduct. One way to roceed as ointed out in Gayle (27c) is to think of a codeshare agreement as a rivately negotiated ricing contract between artners ( w Γ) where w is a erassenger rice the ticketing carrier ays over to an oerating carrier for transorting the assenger while Γ reresents a otential lum sum transfer between artners that determines how the oint surlus is distributed. As we develo the suly side of the model further it will become clear that only the level of w affects euilibrium final roduct rices. Since for the uroses of this aer we are not concerned how the surlus is distributed between artners through the lum sum transfer Γ we do not attemt to derive an euilibrium value of Γ. We assume that the final rice of a codeshare roduct is determined within a seuential rice-setting game. In the first stage of the seuential rocess the oerating carrier sets the rice for transorting a assenger w and rivately makes it known to its artner ticketing carrier. In the second stage conditional on the agreed uon rice w for services sulied by the oerating carrier the ticketing carrier sets the final round-tri rice for the codeshare roduct. The final subgame in this seuential rice-setting game is layed between ticketing carriers. Let r... R index cometing ticketing carriers in a market and let f...f index the corresonding oerating carriers. Further let F r be a subset of the J roducts that are offered for sale by ticketing carrier r in the origin-destination market. Carrier r solves the following rofit maximization roblem for each F r Fr : f Max ( w ) (.) where () in euilibrium is the uantity of roduct offered for sale on the market d d () is market demand for roduct reresents a J vector of final rices and as defined See Chen and Gayle (27) for a similar theoretical modeling aroach of an airline codeshare agreement. For most of the subseuent euations we intentionally omit a market subscrit for variables and euations only to avoid notational clutter. Notwithstanding our omission of market subscrits the reader should continue to interret euations in a market-secific way. 2

24 reviously f w is the rice the ticketing carrier ays to oerating carrier f for its transortation services. So f w is the effective marginal cost that ticketing carrier r incurs by offering codeshare roduct for sale. 7 Since in the first stage of the seuential rice-setting game f f oerating carriers each otimally choose we know that the euilibrium level of deends w on the marginal cost of the oerating carrier that offers transortation services for codeshare w roduct. Therefore we can secify f w as a function of factors that shift the marginal cost of the oerating carrier. Secifically we osit that w f ex( W γ + a + η ) f where W is a vector of variables that shift marginal cost of the oerating carrier (itinerary distance itinerary distance suared time fixed effect) and γ is the associated vector of arameters a f catures the oerating carrier-secific ortion of marginal cost and η is a mean-zero random error term that catures unobserved determinants of marginal cost. In the event that roduct is a ure online roduct instead of a virtual codeshare roduct f r r we relace in rofit function (.) with where is the marginal cost that carrier r incurs w c by offering roduct. Note that in such a case carrier r is the sole ticketing and oerating carrier of roduct. The marginal cost function is therefore f w if is virtual codeshare ex( W γ + a f + η ) (.) r c if is ure online Cometitive Interactions In the sirit of Sudhir (2) we measure cooerative or aggressive behavior by the degree to which euilibrium rices may deviate from Bertrand-Nash rices. We now augment the rofit function (.) above to allow for such analysis. Let A r be a subset of the J roducts that are offered for sale by the alliance artners of c carrier r while B r is a subset of the J roducts that are offered for sale by carriers that are not artners of carrier r. Note that sets F A and B do not have any elements in common but r r r 7 We imlicitly assume here that the ticketing carrier of a virtual codeshare roduct only incurs fixed exenses in marketing the roduct to otential assengers. 3

25 together they contain all the roducts in the market. We can now secify the augmented rofit function for carrier r: Π r Fr f f f ( w ) + ( w ) + ( w ) (.7) a Ar where a is the weight an airline uts on its artners' rofits while a is the weight an airline uts on non-artners' rofits. Note that in the augmented rofit function (.7) yields the rofit function in (.) which would generate Bertrand Nash rices. Therefore > imlies cooerative behavior relative to Bertrand Nash while < imlies more aggressive cometitive behavior relative to Bertrand Nash. a Br A ure strategy Nash euilibrium in final rices reuires that of any roduct offered by carrier r must satisfy the first-order condition: Π r d f d k ( ) f d k ( ) f d k ( ) ( ) + ( wk ) + a ( wk ) + a ( wk ) k F r k A r The first-order conditions are a set of J euations one for each roduct. A few additional definitions allow for a more convenient reresentation of the first-order conditions using matrix notation. First let own Ω be a structure of the J roducts. Let Ω own Second let ( k ) Ωa be a k B J J matrix which describes the ticketing carriers' ownershi own Ω ( k ) denote an element in otherwise own Ω where if distinct roducts k and are offered by the same carrier J J matrix which contains zeros and a which is the arameter describing the degree of cometition between artner carriers. Let Ω ( k ) denote an element in where Ω a Ω a Third let a ( k ) Ωa be a if distinct roducts k and are offered by artner carriers otherwise J J matrix which contains zeros and a which is the arameter describing the degree of cometition between carriers that do not belong to the same alliance. Let Ωa ( k ) denote an element in Ω a where a r

26 Ω a -a ( k ) Fourth let Δ be a resect to final rices where element if distinct roducts k and are offered by carriers that do not belong to the same alliance otherwise J J matrix of first-order derivatives of roduct market shares with d Δ ( k ). In vector notation the system of J first- order conditions for the ticketing carriers can now conveniently be exressed as own w + [( Ω + Ω + Ω ).* Δ] d( ) a a k (.8) where d( ) and w are J vectors of roduct demands final rices and ticketing carriers' effective marginal costs resectively while.* means element-by-element multilication of two matrices. Euation (.8) says the set of rices that satisfy a ure strategy Nash euilibrium is a own function of effective marginal costs w and roduct markus [( Ω + Ω + Ω ).* Δ] d( ). a a Secializing the Suly Euation We now secialize euation (.8) to facilitate analyzing the cometitive effects of virtual codesharing between Delta Continental and Northwest in markets where at least two of them are direct cometitors both before and subseuent to forming the alliance. Aendix A rovides a simle three-airline single market examle of the suly model laid out below. As we discuss further in the data section our samle is restricted to markets in which at least two of the three airlines (Delta Continental and Northwest) offer their own cometing ure online roducts. Furthermore the time san of the samle is deliberately chosen to include realliance and -alliance eriods. Let re be the weight Delta Continental and Northwest ut on each others' rofits in the re-alliance eriod while re is the weight cometing airlines ut on each others' rofits assuming that at least one airline in the air of cometitors is neither Delta Continental or Northwest. For examle Delta's rofit function in a market during the re-alliance eriod is: Π Delta FDelta f re f f ( w ) + ( w ) + ( w ) (.9) ADelta re BDelta where F Delta is the set of roducts offered for sale by Delta airlines in the secific origindestination market during the re-alliance eriod A Delta is the set of roducts offered either by

27 Northwest or Continental airlines in the said market during the re-alliance eriod and B Delta is the set of roducts offered by other cometing airlines in the said market during the re-alliance eriod. Based on our matrix notation convention described above this market would have the reown recom following matrices associated with it during the re-alliance eriod: Ω Ω and recom Ω reown where Ω is a J J matrix of zeros and ones that describes the ticketing carriers' ownershi structure of the J roducts in the re-alliance eriod Ω recom is a matrix of Ω re recom re zeros and while is a matrix of zeros and. The system of roduct marku euations during the re-alliance eriod for a market is: reown recom recom re re [( Ω + Ω + Ω ).* Δ ] ( ) re m d During the -alliance eriod Delta Continental and Northwest only virtual codeshare together in a subset of the markets in our samle. So in our samle some markets lay the role of a treatment grou ( markets in which the three airlines directly comete and virtual codeshare together) while the other markets lay the role of a control grou ( markets in which the three airlines directly comete but do not virtual codeshare together). In markets where Delta Continental and Northwest do codeshare together in the alliance eriod let be the weight that each of the three airlines ut on each others rofit while is the weight cometing airlines ut on each others rofits assuming that at least one airline in the air of cometitors is neither Delta Continental or Northwest. In the case of markets in which Delta Continental and Northwest do not codeshare together in the alliance eriod let be the weight that each of the three airlines ut on each others rofit while is the weight cometing airlines ut on each others rofits assuming that at least one airline in the air of cometitors is neither Delta Continental or Northwest. For a given market in the -alliance eriod the marku euation is: where V is a m + own com com { V.* [( Ω + Ω + Ω ).* Δ ] } own com com ( Υ V ).* [( Ω + Ω + Ω ).* Δ ] d( ) { } d( ) J J indicator matrix whose elements are all ones if for this market but if for the market then all elements in V are zeros. Υ is a J J matrix of ones. To comlete

28 com our descrition of m note that is a J matrix that contains zeros and com Ω J Ω J Ω J com is a J matrix that contains zeros and is a J matrix that com contains zeros and and is a J matrix that contains zeros and. Ω J re See Aendix A for an illustration of what the matrices in m and m look like in the case of a three-airline single market examle. The suly euation used to evaluate the cometitive effects of virtual codesharing between Delta Continental and Northwest is given by: w + m (.) τ l τ l τ l where τ indexes time eriod l indexes market and m τ l is the roduct marku term for market l during eriod τ. From above we know that m τ l is determined by the exression for either re m or m. Our task is to use euation (.) along with our demand euation to obtain econometric re re estimates of and. We can then do airwise comarisons of these arameters. For examle if then the degree of cometitiveness between Delta Continental and Northwest differs in markets where they directly comete and virtual codeshare together comared to markets in which they directly comete but do not virtual codeshare together. Second if then the degree of cometitiveness between Delta Continental and Northwest differs in the -alliance eriod re relative to the re-alliance eriod. Most imortant the airwise comarisons of re and can hel us determine whether airwise comarisons of and are identifying the cometitive effects of virtual codesharing between Delta Continental and Northwest or ust reflecting market wide trends in cometitive behavior. Imlicit in our discussion above is that a key comonent of our strategy to identify (anti-) cometitive effects associated with the Delta/Continental/Northwest codeshare alliance is to comare Delta Continental Northwest and other airlines' ricing behavior before and after imlementation of the alliance in markets where the three carriers virtual codeshare together ("treatment" markets) versus markets in which they comete but do not virtual codeshare together ("control" markets). Again we acknowledge that the airlines also choose which 7 re

29 markets to codeshare in but we resume that such codeshare choices would have been decided rior to making otimal ricing decisions for the menu of roducts the airlines already lanned to offer assengers. Our analysis is therefore focused on short-run ricing behavior conditional on the airlines' redetermined menu of roduct offerings. In summary the vector of suly arameters are θ ( γ ) s... Estimation The arameters to be estimated are θ d ( β α λ σ ) for demand and θ s ( γ ) for suly. Following Berry (99) the estimation strategy for demand arameters involves choosing arameter values such that observed roduct shares S are eual to redicted roduct shares s that is S s ( δ σ ) (.) We comute observed roduct shares based on oulation in the origin city and S M where M is the size of the is the actual number of travel tickets sold for a articular itinerary-airline(s) combination called roduct. As Berry (99) shows in the case of the nested logit model we can analytically solve for the mean levels of utility ( δ ) that satisfy euation (.). The analytical solution is: δ ln( S ) ln( S) σ ln( S / g ) (.2) where S is the observed share of the outside otion and is the observed within grou share of roduct. 8 Using euation (.2) to substitute for δ in euation (.2) and rearranging terms S / g yields: ln( S ) ln( S ) x β α + σ ln( S / g ) + a + λ T DCN Codeshare _ mkt + ξ 3 r + λ T + λ T DCN 2 (.3) 8 The observed share of the outside otion is comuted by S S g where S g S. The observed S within grou share of roduct is comuted by S g S /. g 8 G g G g

30 where ξ is the structural demand error term. Provided we have valid instruments for and euation (.3) is straightforward to estimate using a linear instrumental variables techniue such as two-stage least suares (2SLS). However we are also interested in estimating suly arameters and there is some efficiency benefit in estimating demand and suly arameters ointly. Since both demand and suly arameters enter the suly euation in a nonlinear way we need a nonlinear system estimation techniue to estimate the arameters ointly. We use system Generalized Methods of Moments (GMM) for oint estimation of the arameters. S / g Using the marginal cost function in euation (.) to substitute for w in euation (.) yields the following econometric secification of the suly euation: or euivalently ex( W γ + a + η ) + m [ m ( θ d ) ] W γ + a f + η f ( θ ) ln (.) where η is the structural suly error term and m ( θ d ) is the roduct marku function that deends on demand and cometition arameters θ d and. The GMM estimates of the arameters are obtained by solving the following roblem Min ψ ZΦ θ θ d s Z ψ ξ where ψ is a vector of the demand and suly structural error terms ψ Φ is a η ositive definite weight matrix while Z is a block diagonal matrix of instruments for the demand and suly euations that are assumed orthogonal to the error vector ψ. 9 d Instruments First focusing on demand euation (.3) we recognize that a roduct's rice and its within grou share ( and resectively) are likely to be correlated with the ortion of the S / g 9 In articular Zd Z where Zd is an N Ld matrix of demand instruments and Z s is an N Ls matrix Z s of suly instruments. N is the samle size L is the number of demand instruments while L is the number of suly instruments. d 9 s