Partnering with Competitors – Effects of Alliances on ... · Partnering with Competitors –...

Alliance Center for Global Research and Education

Partnering with Competitors – Effects of Alliances on Airline Entry and Capacity Decisions

_______________

Jun LI Serguei NETESSINE 2011/24/TOM/ACGRE

Partnering with Competitors – Effects of Alliances on

Airline Entry and Capacity Decisions

Jun Li*

Serguei Netessine**

* Doctoral Student at Wharton School, University of Pennsylvania, 3730 Walnut Street

533.3 Jon M. Huntsman Hall, Philadelphia, PA19104, USA. Ph: 215-573-0504 ; Email: l [email protected]

** The Timken Chaired Professor of Global Technology and Innovation, Professor of

Technology and Operations Management, Research Director of the INSEAD-Wharton Alliance at INSEAD Boulevard de Constance 77305 Fontainebleau, France. Ph: 33 (0)1 60 72 92 25 E-mail: [email protected]

A Working Paper is the author’s intellectual property. It is intended as a means to promote research tointerested readers. Its content should not be copied or hosted on any server without written permission from [email protected] Click here to access the INSEAD Working Paper collection

http://www.insead.edu/facultyresearch/research/search_papers.cfm

Partnering with Competitors - Effects of Alliances onAirline Entry and Capacity Decisions

Jun LiThe Wharton School, [email protected]

Serguei NetessineINSEAD, [email protected]

The formation of an airline alliance is one of the most important and difficult decisions that has to be

made by airline management. In particular, domestic airline alliances have always caused controversy due

to the possibility of attenuating the competition, which has been a key concern of policy makers. Before

approving such alliances, government authorities therefore pay a lot of attention to how many overlapping

routes the two airlines attempting to join forces already have. In this paper we go beyond this point of view

and attempt to understand the competitive effects of domestic alliances by analyzing how flight networks

change dynamically after an alliance. We analyze six major US airlines from 1998 to 2006 and estimate the

changes in their entry/exit/stay behavior by adopting a concept analogical to the “difference-in-difference”

estimation approach. We show that, in the post-alliance era, airlines are more likely to enter/stay and build

higher capacity in markets where their alliance partners hold strong market power, compared to markets

where non-partners dominate. We show that this tendency is likely to be induced by higher pricing power in

the markets dominated by the alliance partners rather than by higher demand in those markets. In a typical

duopoly market, the two allied partners are able to charge nearly $9 more per one-way ticket coupon as

compared to non-allied competitors. Our findings have important implications for both policy makers and

airline practitioners. In addition to reviewing the current overlapping routes before an alliance, it is crucial

for policy makers to be aware of how airlines may change their networks dynamically after the alliance,

and how this will affect the competition structure. We also find that the allied airlines may have been over-

optimistic about the demand-increasing effect of alliances on markets shared with partners, since we find a

decrease in load-factors in these markets.

1. Motivation

Airline alliances have come a long way from their early form of limited commuter codeshare agree-

ments to the current form of domestic alliances. Codesharing is a practice whereby the operating

carrier allows one or more other carriers to market its flight and issue tickets as if they were

operating the flight. Commuter codesharing, which started in 1960s, was aimed at serving unprof-

itable short-haul markets. In a similar manner, US and international airlines formed codesharing

partnerships to serve international destinations unreachable by their own network. This spirit was

followed by the first partnership between major domestic airlines Continental and America West.

In 1994, the two airlines “linked their network to serve small peripheral city-pair markets such as

1

2 Li and Netessine: Partnering with Competitors

Tucson, Arizona, Portland and Maine” (Transportation Research Board 1999). All these earlier

codesharing agreements shared the common trait that the partnership introduced new routes which

neither partner had served before. However, this tradition was not followed when Continental and

Northwest started their more comprehensive partnership in 1999. In most codeshared routes, one

or both partners had already offered through-services, so no new competition was introduced. Two

other alliance attempts between United and Delta, American and US Airways, and the subsequent

alliances between United and US Airways, Delta, Northwest and Continental in 2003, caused major

policy concerns regarding their non-competitive nature (General Accounting Office 1998, 1999,

Department of Justice 1999a,b, Department of Transportation 2003). While international alliances

usually involve two vertically connecting operating carriers, domestic alliances are dominated by

codeshared routes operated by a single carrier. As a consequence, though studies on international

alliances showed harmonious effects of increasing passenger volume and decreasing airfares, the

evidence regarding domestic airlines is largely inadequate and discordant.

Not surprisingly, domestic airline alliances have caused a lot of controversy since their inception.

Airlines officials have emphasized the benefits to consumers: broader travel options, increased flight

frequencies, better route connections, improved frequent flyer programs, lower fares due to cost

savings within alliance members. However, consumer advocates have raised many concerns:

“The real purpose of the major global airline alliances is to solidify the oligopoly of their

participants, and to drive smaller non-participants and even large non-aligned airlines out of

business – so that the remaining airlines can raise prices, while travelers are offered fewer

choices.”1

Likewise, policymakers expressed a great deal of scepticism when reviewing alliance proposals

which have to adhere to certain antitrust laws and regulations. For example, three codesharing

proposals were made in 1998, between Northwest and Continental, United and Delta, and Amer-

ican and US Airways, respectively. Only one of them was finally approved by the Department of

Transportation (DOT), yet with conditions imposed: Northwest and Continental had to consent

not to codeshare hub-to-hub routes. Another wave of alliances came in 2002 and 2003, when the

DOT finally approved the two alliances (between United and US Airways, and between Delta,

Continental and Northwest), while expressing serious reservations about the latter 2. At the heart

1 Edward Hasbrouck, consumer advocate. Jan 2006. Airline subsidies, alliances, and code-sharing. Source:http://hasbrouck.org/articles/alliances.html.

2 “The Department has determined that the agreements, if implemented as presented by the three airlines, couldresult in a significant adverse impact on airline competition, unless the airlines formally accept and abide by certainconditions that are intended to limit the likelihood of competitive harm. If the airlines choose to implement theagreements without accepting those conditions, the Department will direct its Aviation Enforcement office to insti-tute a formal enforcement proceeding regarding the matter.” Department of Transportation. Office of the SecretaryTermination of Review Under 49 U.S.C. 41720 of Delta/Northwest/Continental Agreements. Federal Register. Vol.68, No. 15. Thursday, January 23, 2003. Notices.

Li and Netessine: Partnering with Competitors 3

of the concerns of policy makers, including the DOT, the Department of Justice (DOJ) and the

General Accounting Office (GAO), was the adverse effect of alliances on competition:

“[Proposed alliances] will reduce competition on hundreds of domestic routes if the alliance

partners do not compete with each other or compete less vigorously than they did when they

were unaffiliated... It will be critical to determine if an alliance retains or reduces incentives

for alliance partners to compete on price.”3

“Codesharing airlines might compete less aggressively in price or capacity in overlapping mar-

kets, to avoid undermining the agreement on connecting traffic.”4

To summarize the above largely anecdotal evidence, an airline alliance is a double-edged sword

to competition. On one hand, it can be pro-competitive by introducing new services or new com-

petitors; on the other, it can be anti-competitive since it may dampen competition among partners

while raising barriers to prevent outside competitors from entering. Thus, we argue that, to under-

stand how a domestic alliance changes the competitive environment it is not adequate to only look

at fare and traffic changes without looking at the changes in the entry patterns, nor is it ade-

quate to only focus on pre-alliance overlapped markets without considering the future dynamics of

entry/exit which could change the network structure and density in the long run. Network struc-

ture is altered by entry and exit activities, while network density is modified by changes in seat

capacity on each route. In this paper, we evaluate empirically how entry strategies and capacity

decisions have changed since the implementation of codesharing agreements.

The goal of this paper is to investigate how the formation of alliances changes long-term entry

strategy and capacity planning, which in turn change flight network structure, network density

and the degree of competition in local markets. Specifically, we want to know whether airlines are

more likely to enter (or increase capacity) in a city-pair market where their alliance partners are

present and hold strong market power. Alternatively, they might be more likely to reduce capacity

redundancies and avoid overlapping routes by entering markets which are not operated by their

partners but by non-allied competitors. While arguments can be made in support of both views,

there is currently no empirical evidence of what actually happens in practice and why. To uncover

the drivers of these strategic changes, we empirically evaluate the changes induced by alliances in

the average market fare, passenger traffic volume and the load-factor.

Methodologically, we use an analogy of the “difference-in-difference” estimation to identify

changes in the airlines’ entry and capacity planning behaviors before and after three major domes-

tic alliances that occurred in 2003 (United/US Airways, Continental/Delta, and Northwest/Delta).

3 General Accounting Office. Aviation Competition: Proposed Domestic Airline Alliances Raise Serious Issues. 1998.

4 Transportation Research Board report. Entry and Competition in the U.S. Airline Industry: Issues and Opportuni-ties. 1999.


We model the effects of airline entry and changes in capacity on the lagged market power of the

airline itself, the market power of its allied carriers, and the market power of its non-allied car-

riers, while controlling for market characteristics and the network structure. This methodology is

applied to analyze the decisions of six major U.S. domestic airlines, i.e., American, United, Delta,

Continental, Northwest and US Airways, from 1998 to 2006.

Our main results are the following. In the post-alliance era, airlines are more likely to enter/stay

and install higher capacity in the markets where their partners are active than in those where

non-allied carriers are active. This result is robust to multiple alternative model specifications. We

then show that this partner-favoring behavior is likely to be driven by higher average ticket prices

in markets dominated by alliance partners. These effects are highly economically significant. In a

typical duopoly market after an alliance, airlines are able to charge almost $9 more per one-way

ticket in markets shared with partners than in markets with competitors. Overall, our paper raises

important policy questions. Even though a domestic airline alliance may appear acceptable to the

regulators ex-ante, changes to the network structure that are introduced ex-post may run counter

to the public interests. We further find that there is a notable decrease in the load-factor (by about

5.9%) in the markets operated with alliance partners relative to those operated with competitors

which may be due to an over-optimistic increase in capacity in the former markets.

2. Literature Review

Several streams of literature are related to our paper: 1) empirical literature studying operational

and revenue management practices in the airline industry; 2) literature on the effects of airline

alliances and codesharing agreements; 3) literature on airline entry/exit strategies; 4) management

literature on alliance formation.

Increasing efforts have recently been made among the operations management community to

study empirically airlines’ operations and revenue management practices. Cho et al. (2007) show

positive effects of codesharing agreements on revenues and the load-factor. Li et al. (2010) find

that larger-than-predicted frequencies of extremely long delays and cancellations are by far the

most important predictors of a fall in stock price. Ramdas and Williams (2009) show that low-

cost airlines pay over twice as high a penalty (in terms of decrease in utilization) for declines in

on-time performance than full-service airlines. Arikan and Deshpande (2010) study the impact

of flight delays and find that it is systematically “under-emphasized” by airlines as compared to

early arrivals. They also show a significant effect of network structure on the on-time probability.

Arikan et al. (2010) develop a stochastic model based on empirical data in order to measure the

propagation effect of flight delays through the aviation network. Using consumer complaints data

from the airline industry, Lapre and Tsikriktsis (2006) find customer dissatisfaction are U-shaped


function in operating experiences and the organizational learning curves are heterogeneous across

airlines. Granados et al. (2011) find that passengers in online channels are more price elastic than

those in offline channels. In the revenue management realm, discrete choice models have been

suggested recently to improve performance of the revenue management practice (Vulcano et al.

2010, Garrow 2010). Vulcano et al. (2010) propose an algorithm to resolve the unobserved no-

purchases in estimating discrete choice models using transaction data. They demonstrate that this

choice-based model improves revenues by 1-5% on the city-pairs studied in their paper. Newman

et al. (2010) provide a more robust estimation procedure for choice models when using data from

a single firm. Ferguson et al. (2010) apply the choice-based method to hotel revenue management.

Recent modeling papers have attempted to analyze airline alliance agreements in conjunction with

revenue management tools (Wright et al. 2010, Hu et al. 2010, Talluri and van Ryzin 2004).

The economics literature have documented consistent evidence of increasing traffic and declining

fares of 8% to 25% (Oum et al. 1996, Brueckner and Whalen 2000, Brueckner 2003) due to interna-

tional alliances. However, the evidence of the impact of the domestic alliances is more ambiguous.

Some studies (Bamberger et al. 2004, Ito and Lee 2007) find domestic codesharing are associated

with lower fares, while others (Armantier and Richard 2008, Gayle 2008) present somewhat oppo-

site evidences. The findings are also mixed from recent studies using structural models (Gayle 2007,

Shen 2010). Despite the volume of this literature, none of these papers examine changes in entry

or capacity decisions due to an alliance.

The third stream of related work is the economics literature on airline entry. Berry (1992) and

Cilliberto and Tamer (2009) estimate the static entry game, while the latter allows for multiple

equilibria. More recent studies attempt to estimate dynamic entry games. Bajari et al. (2007)

propose a two-stage algorithm in which the equilibrium behavioral probabilities are recovered in

the first stage, and then used in the second stage to obtain the structural parameters. Benkard

et al. (2010) apply this approach to simulate the long-term dynamics of the airline merger. Aguir-

regabiria and Ho (2009) localize the optimization problem to resolve the curse of dimensionality.

Our approach to describing airline entry behavior is close to the first step of Benkard et al. (2010).

However, our focus is on changes of equilibrium behavior before and after alliances, and how they

are associated with the identity of allied vs. non-allied players. Estimating dynamic games will not

shed additional light on the question we aim to answer and is beyond the scope of this paper.

Alliance formation is also a well-established area of interest in the management literature. Adopt-

ing the network perspective, researchers have focused on overlap vs. complementarity in studying

dyadic alliance relations (Gulati 1995, Dyer and Singh 1998, Chung et al. 2000). Evidence from

multiple industries (automobiles, airlines, pharmaceutical, high technologies etc.) suggest that com-

plementarity is generally what makes most alliances successful, by allowing partners to have access


to capabilities which are not otherwise available. Specifically in the airline alliance context, Gimeno

(2004) examines the content and intensity of dyadic relations, and shows that partner selection

is dependent on the extent of alliance co-specialization. While these high-level studies of alliance

formation and partner selection offer insights by recognizing the tension between complementarity

and overlap, questions that go down to the operational level are left unasked– such as the ones

that we ask here – How do airlines adjust their key operational decisions post alliances, i.e., which

flights to operate, how much capacity to implement? Do they pursue complementarity or overlap

with their partners?

3. Hypotheses Development

In this section we develop hypotheses for changes in entry and capacity decisions due to an alliance

by describing the potential drivers of these changes. We reply upon two sets of theories – operational

efficiency and collusive behavior – to draw implications for how airlines might change their entry

strategy and capacity decisions after an alliance has been formed. Operational theories citing

network efficiencies will predict that allied airlines eliminate capacity redundancies and consolidate

overlapping routes. If this is the case, chances are that partners avoid operating in the same

markets since they have the option to codeshare each other’s flights. In this way, each airline

can increase their network coverage and passenger traffic without incurring significant additional

operating costs. Findings in the management literature also support this argument by showing

that firms mainly seek complementarity in alliance formation as a means to combine resources

and opportunities. Conversely, operational theories focusing on cost synergies predict the opposite

– allied airlines are more likely to cooperate in the same markets since cost savings (e.g., due to

aircraft maintenance and joint gate operation) can be expected to be greater in geographically

proximate networks (Reitzes and Moss 2008). Likewise, economic theories of collusive behavior

make contradictory predictions regarding whether airlines are more likely to seek overlapping routes

with their partners’ network or to operate complementary routes. On the one hand, the possibility

arises that partner airlines may collectively decide to divide and allocate markets, which is the

primary concern of the DOJ (Department of Justice 1999a,b). On the other hand, allied airlines

may undertake joint efforts to elevate entry barriers to small airlines and other non-allied carriers.

As a consequence, we might observe more overlaps among partners and less “invasion” of the

territory of out-of-alliance competitors (General Accounting Office 1999, Reitzes and Moss 2008).

The theories and arguments they support are summarized in the following two-by-two table.

Avoid Overlap with partners Seek Overlap with partnersOperational Optimize network by Maximize cost savings onEfficiency eliminating capacity redundancies geographically proximate networksCollusive Divide and allocate markets Elevate entry barriersBehaviors among partners for non-allied competitors


The question of which the above-mentioned rationales ultimately dominates is best analyzed

empirically. Although both arguments may be true, we suspect that it is more likely that airlines

seek overlap with partners after an alliance. Thus we state the hypotheses as follows:

Hypothesis 1 [Entry Hypothesis - Seek Overlap] In the post-alliance vs. pre-alliance era, airlines

are more likely to operate (i.e., enter or stay) in the markets where their alliance partners possess

strong market power, as compared to markets dominated by competitors.

The decision to enter/stay in the market is one manifestation of the airline’s intention to seek

more/less overlap with the partner/competitor. Another manifestation of such an intention can

be an increase or reduction in the capacity on the route, which may or may not be accompanied

by the decisions to stay/exit. For instance, the airline may decide to increase capacity on the

route operated with alliance partners without exiting or entering new routes. Thus, an alternative

way to evaluate post-alliance behavior would be to evaluate changes in capacities. Using the same

reasoning as above, we develop similar hypotheses for capacity decisions.

Hypothesis 2[Capacity Hypothesis - Seek Overlap] In the post-alliance vs. pre-alliance era,

airlines are more likely to build up capacity in the markets where their alliance partners possess

strong market power, as compared to markets dominated by competitors.

4. Model4.1. Entry Model

Carriers, indexed by i, consider an entry strategy into a set of markets (i.e., airport-pair or city-

pair) indexed by m, where m= 1,2,3, ...,M . At time (i.e., year) t, each carrier decides whether or

not to operate a direct flight in the market m. Consider codeshare partnership as the treatment

and note that the treatment is applied at the carrier level. For instance, after United formed a

partnership with US Airways, United’s entry strategy for every market might have been affected

by the fact that US Airways is now the partner. Specifically, we assume that

y∗imt =X ′im(t−1)β+ f(Dit; i,m, t) +λt +αi + εimt, (1)

where y∗imt is the potential profit contribution to carrier i by operating a direct flight on market

m at time t, which not only accounts for immediate profit contribution, but also the long-run

contribution. Moreover, it not only includes the direct revenue contribution from operating the

flight, but also the possible spill-over revenues from other related legs in the carrier’s network.

Xim(t−1) represents the lagged control5 variables for characteristics of the market and of the network.

It includes 1) segment features such as distance, population and per-capita income of both end

5 According to our conversation with airline practitioners, lag of one year is the appropriate time frame for entry/exitdecisions, which is common practice in other related papers as well. Using more lagged years do not add muchexplanatory power but cause collinearity problems.


points, level of competition (including only direct flights), presence of low-cost carriers (LCC),

level of congestion (i.e., the load-factor); 2) network node features (considering cities or airports

as the nodes and connections between them as edges of the network) such as degree of centrality6,

competition level and LCC presence at both cities or airports; 3) network edge features such as

connectivity (number of indirect paths) and level of competition at the city-pair or airport-pair

(See Table 1 for a complete description of variables included.). Dit is a {0,1} variable indicating

that carrier i is in a codeshare partnership at time t if Dit = 1, 0 otherwise. f(Dit; i,m, t) denotes

the effect of an alliance. We allow the effect to vary across carriers, markets and time, and elaborate

on this point later on. λt controls for the time trend that is common to all carriers (e.g., economic

conditions). αi controls for the time-invariant carrier effects. εimt is an idiosyncratic shock which

is observable to decision-makers but not to econometricians. We assume εimt to be i.i.d. across i,

m and t. That is, the shocks are cross-sectionally and serially uncorrelated, an assumption which

we relax subsequently.

We do not actually observe the underlying outcome variable y∗imt. Instead, what we observe is

a {0,1} variable, yimt, which indicates whether carrier i operates a direct flight on market m at

time t. The relationship of the two variables can be formalized as follows (similar to Benkard et al.

(2010)).

yimt = 1{y∗imt ≥ 0|yim(t−1) = 0}, (2)

yimt = 1{y∗imt ≥−γimt|yim(t−1) = 1}. (3)

In this specification, the entry threshold is higher for a potential entrant (Eq. 2) than for the

incumbent (Eq. 3), while the threshold for potential entrants is normalized to zero. Since carriers

with larger market power are usually more capable of surviving lower temporary profits, we allow

the threshold to be dependent on the carrier’s own market power, i.e., γimt = γSim(t−1), where

Sim(t−1) is the market share of carrier i on market m at time t− 1.

To summarize, we observe the following data generating process:

y∗imt = γSim(t−1) +X ′im(t−1)β+ f(Dit; i,m, t) +λt +αi + εimt, (4)

yimt = 1{y∗imt ≥ 0}. (5)

Now we take a closer look at the treatment effect f(Dit; i,m, t). We allow the treatment effect to

depend on the partner’s market share as well as on the competitors’ market share7.

f(Dit; i,m, t) = δDit

6 One way to account for entries due to international connections is to include international gateways as a control.However, this variable is highly correlated with degree of centrality.

7 All major carriers who do not have codeshare partnership with the carrier are included as its competitors. We alsotried an alternative modeling approach in which we define partner’s and competitor’s presence using {0,1} binaryvariable instead of using the market share. The results are consistent.


+ δp1PartnerShareim(t−1) ∗ (1−Dit) + δp2PartnerShareim(t−1) ∗Dit

+ δc1CompetitorShareim(t−1) ∗ (1−Dit) + δc2CompetitorShareim(t−1) ∗Dit, (6)

where δ is the direct treatment effect. δp1 is the effect of the partner’s market share on the entry

probability pre-alliance, and δp2 describes the same effect post-alliance. Similarly, δc1 and δc2 denote

the effects of competitors’ market share on the entry probability before and after the alliance8,

respectively. Ultimately, we are interested in the interaction effects of the alliance and the partner’s

market share and competitors’ market share. The parameters of interest are summarized as follows:

change of partner’s effect δp2− δp1

change of competitor’s effect δc2− δc1

difference-in-difference (δp2− δp1)− (δc2− δc1)

where δp2− δp1 represents the change in the partner’s influence on the carrier’s entry decision after

the codeshare agreements, and δc2−δc1 represents the change in competitors’ influence. Ultimately,

we want to know whether the changes have been different (in direction and magnitude) for partners

and competitors. A significantly positive estimate supports Hypothesis 1, i.e., airlines seek overlaps

with partners after alliances. The identification of the “difference-in-difference” term comes from

the fact that the change in entry probabilities before and after an alliance differ for two groups of

markets: those dominated by the partner vs. those dominated by competitors. The identification

is not directly due to the variation of alliance status among carriers (i.e, American Airlines has not

formed domestic alliance while others have), although this adds further variations for identification.

One additional concern may be selection bias – partners are not assigned randomly but chosen

by airlines. To address this concern, it is important to control for the effect of partner’s market

share even before alliance is formed: the fact that United chooses US Airways as a partner may

reflect complementarities of their networks or some other factors that we, as econometricians, do

not observe, which means United may have appeared to be more friendly or hostile to US Airways

even before their alliance is formed. We provide more discussions around the concerns related to

partner selection in the Results section. The model can be summarized as follows:

y∗imt = γSim(t−1) +X ′im(t−1)β+ δDit + δp1PartnerShareim(t−1) + δc1CompetitorShareim(t−1)

+ (δp2− δp1)PartnerShareim(t−1) ∗Dit + (δc2− δc1)CompetitorShareim(t−1) ∗Dit

+λt +αi + εimt, (7)

yimt = 1{y∗imt ≥ 0}. (8)

8 So far the partner/competitor effect are assumed the same for all carriers, we account for carrier-specific effects inthe robustness test.


4.2. Capacity Model

We further investigate how capacity is adjusted after the alliance, conditional on the airline deciding

to stay in the market. We want to see whether this adjustment in capacity differs for markets

operated together with partners vs. competitors. We use a model that is similar to the entry model

above but with appropriate changes:

Kimt = γSim(t−1) +X ′im(t−1)β+ δDit + δp1PartnerShareim(t−1) + δc1CompetitorShareim(t−1)

+ (δp2− δp1)PartnerShareim(t−1) ∗Dit + (δc2− δc1)CompetitorShareim(t−1) ∗Dit

+λt +αi + εimt, (9)

εimt = µim + ξimt (10)

where Kimt is defined as the log of the carrier’s annual number of seats in the segment. This

model can be estimated under Random Effects and Fixed Effects specifications. Furthermore, an

alternative model specification is the dynamic panel data model with lagged capacity Kim(t−1). We

adopt Arellano-Bond estimator9 (see Greene 2007, Chapter 13) to account for potential correlation

between the error term and predetermined variables Kim(t−1),Xim(t−1).

5. Data

The principal data sources for our study are the Bureau of Transportation Statistics’ T-100 Domes-

tic Segment Data and Airline Origin and Destination Survey (DB1B) which we supplement with

population statistics data from the Bureau of Economic Analysis. The T-100 Domestic Segment

Data provides quarterly information on seat capacity, number of enplaned passengers and the load-

factor. The DB1B data is a 10% quarterly sample of all airline tickets in the United States, which

includes price information. These are standard data sources for the closely related studies.

To accurately measure the impact of the treatment it is best to use a time window neither too

long nor too short before and after the treatment event. The span of our study runs from 1998

through 2006, although since we use lagged control variables, we actually utilize only 8 years of data.

We choose this particular time period to balance the ‘before’ and ‘after’ periods around the major

codeshare events that took place in 200310. We do not use years far ahead because the entry strat-

egy might have changed over a long time-frame due to policy/technology/management/economy

9 Difference GMM is adopted here. An alternative is to use system GMM estimation. However, the assumptions forthe additional moment conditions in system GMM may not be satisfied in this case. Moreover, orthogonal deviationis used instead of first differencing to cope with panels with gaps.

10 We adjust for other major changes in the airline industry during the period of study. 1)9·11 Effect. We correctfor its effect by using year dummies (accounting for industry-wide effect) and 9·11-UA/AA dummies, and replacingmarkets that observe exit in 2002 and re-entry in 2003 as being active in 2002. 2)Acquisition and merger. AmericanAirlines acquired Trans World Airlines in 2001. The routes taken over from Trans World are not counted as entries.Two national airlines, US Airways and America West, merged in 2005. However, America West continued reportingunder its code until 2007. 3)Bankruptcy. All major airlines experienced hard financial situations 2002-2004. Four filed


changes. Moreover, we estimated the decay of the effect using longer horizons, and found that one

to two years after the alliance is the period in which most route adjustments are made11. Although

quarterly data is available in our databases, we use yearly data because this is a more appropri-

ate time-frame in the airline industry to make entry and exit decisions, and yearly aggregation

corrects for the seasonal effects. The carriers of interest are major domestic airlines including AA

(American), UA (United), US (US Airways), CO (Continental), DL (Delta) and NW (Northwest).

Since we use yearly data, a carrier is defined as present in a market in that year if it operates a

direct flight on the market throughout the year. We consider CO and NW in partnership starting

from 1999 (officially approved in November 1998), UA and US in partnership starting from 2003,

and DL/CO, DL/NW in partnership starting from 200412.

We also made the effort to replace regional airlines by their parental major airlines. This modifi-

cation is necessary because during the past decades major airlines gradually gave up direct presence

in many smaller markets and instead contracted with regional partners to operate on these routes.

This does not mean that the major airlines have ceased operations in these markets; they simply

started operating in a more efficient way by utilizing smaller aircrafts in smaller markets. Moreover,

consumers still buy tickets to these destinations under the brand name of the major airlines and

the major airlines control pricing/revenue management systems of the smaller carriers. Without

accounting for these shifts, we would have counted many more exits. Regional airlines accounted for

no less than 20% of all the tickets sold. Details of this correction procedure are found in Appendix

Table A1. Note that the same regional airline may have operated for different major airlines at

different times of its history, e.g., Air Wisconsin started the transition from serving United Airlines

to US Airways in 2003 when United Airlines filed for bankruptcy. Also note that the same regional

airline may contract with two or more major airlines at the same time and even on the same mar-

kets. We use DB1B data to help us correctly identify these markets as well as the percentage of

capacity contracted for each major airline. For example, if both major airlines A and B sell tickets

on the same flight operated by regional airline X, there can be two possibilities. One is that X

only contracts with airline A, but airline B can also sell tickets on flights operated by X through

codesharing agreements with A. The other possibility is that X contracts with both A and B. To

distinguish these two cases, we look at the percentage of tickets sold on A and B. In the former

for bankruptcy protection while continuing their operations. Although we could control for financial situations in themodel, it does not affect the results. The main reason is that these financial shocks are generic to the airline overall,but not specific to markets – there is no particular reason why markets operated by partners or by non-partnersshould be affected more.

11 We also tried to extend the study to longer horizons and the results remained qualitatively unchanged.

12 Officially approved in June 2003. The result is not sensitive to this specification


case, A sells the majority of the tickets. In the latter case, A and B sell comparable portions.

Practically, we use 80% as the dividing point. The result is not sensitive to this specification.

Following standard strategies used, for example, by Benkard et al. (2010) and Aguirregabiria and

Ho (2009), we select the 75 largest U.S. airports, where size is defined by the enplaned passenger

traffic. We then map the 75 airports to Metropolitan Statistical Areas (MSAs). We use Composite

Statistical Area (CSA) or Metropolitan Division when necessary. For example, airports serving the

New York area, JFK, LGA, EWR and ISP are grouped to New York-Newark-Bridgeport CSA.

This grouping accounts for spatial correlation among these airports13 since airports close to each

other usually have correlated demand and supply shocks. This grouping gives us 62 MSAs (see

Appendix Table A2 for details). We supplement the airline data with annual population estimates

and per-capita incomes for these MSAs from the Bureau of Economic Analysis.

We construct the set of markets containing all possible directional combinations14 of these 62

MSAs, which gives us 3,782 markets. Thus, we have a panel data of 22,696 market-carrier dyads

for 8 years. To account for occasional redirection of flights due to unforseen events such as severe

weather, we only count an airline as operating the market in a particular year if it carried more than

36000 passengers (as in Benkard et al. (2010)) in that market-year, which roughly corresponds to

one flight per day15. To focus on the most common types of trips, we place the following restrictions

on the raw data to obtain average market fares, following Ito and Lee (2007): 1) we restrict our

analysis to round-trip, coach class tickets; 2) we limit our analysis to tickets with no more than two

coupons per directional leg; 3) we exclude itineraries with fares per person less than $25 or greater

than $1,500, since they might represent employee tickets or frequent flyer miles tickets or data

errors; 4) we exclude itineraries on which the marketing carrier of either segment was a non-U.S.

carrier. All of these are standard data transformations which are commonly used in the literature

utilizing the same data sources.

Table 1 describes the key variables and Table 2 presents the summary statistics and the correla-

tion information. The left-hand side variables include the operating status, capacity, average fare,

traffic and the load-factor of each carrier on each market-year. The variables of interest include

market shares (for the partner and the competitor), the alliance status, and the interaction terms

between these two sets of variables. The covariates included in this study fall into the following

three categories: market characteristics, network nodal features, and network edge features. As in

13 An alternative approach to deal with the spatial correlation is to control for the distance from each airport to theclosest alternative airports. See Cilliberto and Tamer (2009).

14 Results are similar when using non-directional market definitions.

15 We also tried different cut-off values such as 3600 as used in Berry (1992) and Borenstein and Rose. (1994), whichcorresponds to one regional jet per week. Our results are not sensitive to this cut-off value.


most airline researches, we include market control characteristics such as distance, demographics

at both endpoints, the level of competition and the low-cost carrier’s market share. In addition,

we believe that a key operational measure, i.e., the load-factor, is a critical measure in the airline

entry decision and the price level, since the load-factor reflects the congestion level of the market,

carriers’ operating costs, and their ability to manage demand uncertainty. Inspired by the network

perspective adopted in the alliance formation literature, we add flight network features into this

study16. The second category, i.e., network nodal features, include carriers’ origin and destination

degree, market share, level of competition and presence of low-cost carriers. Airlines’ decisions to

offer direct services on a market also depend on all the connecting possibilities through transferring

at the origin and destination airports. In a hub-and-spoke network, origin and destination degree

of centrality capture these possibilities, and this metric defines to a great extent the position of

a route in the airline’s entire network17. We do not use {0,1} hub or spoke measure since degree

measure already captures the connectivity of each node in a more precise way: for instance, sub-

hubs are emerging in many airlines’ networks but the {0,1} measure would not distinguish them

from hubs. The third category of covariates includes network edge measures, i.e., the number of

one-stop connecting routes between the origin and the destination since airlines’ entry decisions

are also dependent on the alternative existing connecting services. Finally, we include measures of

competition and low-cost carrier presence when accounting for these connecting possibilities. These

network features have not been traditionally included in the related literature, though Benkard

et al. (2010) adopt similar measures in their working paper. We believe that inclusion of these

measures is critical in recognizing that flight operation and capacity decisions are deeply embedded

in the structure of the entire network. We note that our summary statistics are consistent with

those in related studies.

6. Results6.1. Results of the Entry Model

We begin our analysis by providing key summary statistics both for exploratory analysis and to

allow for initial understanding of the competitive landscape in the industry. Table 3 shows the entry

and exit dynamics of the major airlines in the study period. We notice that, over the 8-year period,

on average there has been approximately 10% turnover (entry/exit) each year. However, there was

some turmoil in turnover right after the alliances were formed by most major airlines. United, US

16 Flight networks are different from the relational networks in the management literature. Nodes are represented byairports in the former and by airlines in the latter.

17 We also tried to include other types of network centrality measures, such as closeness centrality and betweennesscentrality. However, these more sophisticated centrality measures did not add much value on top of degree centrality.We believe the reason is that in a hub-spoke network, degree centrality already contains most information.


Air, Delta and Northwest all saw a sizable increase in the number of entries after the alliances

in 2003 and 2004. To get a sense of the extent of overlapping routes between allied partners and

how networks evolved before and after the alliance, we provide another set of summary statistics

in Table 4. Two airlines are considered to overlap on a route if they both operate direct flights

on it. United and US Airways have seen a notable increase in both the absolute number and the

percentage of overlapping routes after the alliance. Although the changes for Delta, Northwest

and Continental are less obvious, we should contrast them with the changes in overlapping routes

between carriers from different alliances. As we show in the third column of Table 4, the number of

overlapping routes among carriers from different alliances decreased after the major alliances were

formed, making the change in overlaps between same-alliance carriers more prominent.

We now move from this preliminary exploratory examination to the statistical analysis. The

three columns in Table 5 represent results from Probit models using an increasing number of

control variables: Model 1 has only demographic and segment-level controls, Model 2 adds some

network features, and Model 3 has a full set of network controls to demonstrate robustness of

our results. Collinearity tests diagnose no multicollinearity problems in all our models18. Models

are estimated with a good model fit, at R2 around 0.8819. We begin our discussion with control

variables whose effects are largely consistent with existing empirical results in the airline industry.

For the demographic variables, we see that distance has a negative effect, and population and per

capita income have positive effects on the entry in Model 1, similar to findings in Benkard et al.

(2010). When network features are controlled in Models 2 and 3, the effect of population is largely

absorbed by the features of the network. Moving on to segment features, we see that a higher

segment congestion level (load-factor) is naturally associated with more entries. Further, entries are

more active when the competition is high (i.e., HHI is low), which can be explained by the higher

underlying profitability of the market. As has been shown in many other recent airline studies,

direct presence of low-cost carriers poses a credible threat and leads to fewer entries. Finally, as

expected, an airline’s own segment load-factor is positively associated with probability of operating,

since a high congestion level indicates high demand relative to capacity. Moving on to the network

features, we observe that they are mostly significant, and that they contribute somewhat to the

explanatory power of the model. The impact of competition and presence of low-cost carriers are

consistent with segment-level controls. Further, a high degree of centrality is associated with a

higher probability of operating. The scale of the coefficient is also similar to Benkard et al. (2010).

18 There is no obvious evidence for multicollinearity. The largest VIFs are 5.75 (own market share) and 5.66 (ownloadfactor), which is within expectation. All the other variables have a VIF smaller than 2.

19 The high R2 is largely due to the strong correlation between previous own market share and the operating decision,which is expected in an industry with relatively slow turnover.


As the long-established literature on the airline industry (Berry 1992, etc.), we also find positive

effects for the airport market share on the decision to operate. However, the effect diminishes as

we include additional controls for network features, which indicates that the airport presence can

actually break down into multiple factors each describing a different aspect of the network features.

We now move on to discuss the variables of main interest, which include the change of the partner

effect, change of the competitor effect and the difference-in-difference term. We focus our discussion

on coefficient estimates for Model 3. We notice first that, after the alliance, partners’ market power

has an increasingly positive effect (change in partner effect= 0.427) on airlines’ decision to operate,

while the competitors’ market power has a diminishing effect (change in competitor effect= - 0.384);

both estimates are highly significant. The difference-in-difference term is positive and also highly

significant (difference-in-difference= 0.811). These results demonstrate that, post alliance, airlines

are more likely to favor a market in which their partners have a strong presence, but disfavor

markets dominated by competitors. We next wish to understand the economic significance of these

effects. Extra caution should be taken to obtain the marginal effects in a non-linear model where

interaction terms are involved. As pointed out in Ai and Norton (2003), “the marginal effects for

the interaction term can be positive, negative or insignificant even though the coefficient of the

probit is estimated to be positive and significant.” We therefore follow the methodology in Ai and

Norton (2003) and Anderson and Newell (2003) to calculate the marginal effects. Figures 1 and 2

illustrate the marginal effects at different market share levels of the operating carrier, partner and

competitor. The X, Y and Z-axes in both figures are the partner’s share, competitor’s share and the

marginal effect on the probability of operating. Figure 1 demonstrates the change of the partner’s

marginal effect (red plane in the middle) and the change of the competitor’s marginal effect (blue

plane in the middle) as well as the 95% confidence intervals of these marginal effects (upper and

lower planes in red and blue, respectively). For the purpose of this illustration, marginal effects

is shown conditional on not operating in the previous year. We observe that there is a clear gap

between the two marginal effects. For most problem parameters the 95% confidence intervals do

not overlap. In other words, the marginal effect of the “difference-in-difference” term is positive

and significant with the exception of the case in which the partner’s share is large and competitors’

share is small (the corner directly in front of the viewer). This observation probably reflects the

diminishing returns of the additional increase in the partner’s share: if the partner’s share is already

very high, the additional 1 percentage point increase in partner’s market share would not have as

much of an effect as otherwise, something we might expect. Quantitatively, for a very conservative

estimate of the marginal difference-in-difference effect, we look at a point with the narrowest gap

between the partner’s (red) plane and the competitor’s (blue) plane, e.g., the left-most point in

Figure 1 (note that difference-in-difference estimates are much larger at other points where the


gaps are wider). The marginal effects can be interpreted as follows. Conditional on not previously

operating in the market, a 1% increase in the partner’s market share after the alliance would induce

an additional 0.005% in the entry probability as compared to the same effect before the alliance.

Conversely, a 1% increase in the competitor’s market share after the alliance would subtract 0.007%

from the entry probability as compared to the same effect before the alliance. The difference-in-

difference of this marginal effect is 0.012 percentage point. Though this estimate may look small at

first glance, it is actually economically significant if one realizes that the baseline entry probability

(inferred from the data) is very low, about 0.005. Thus, a 0.01% increase amounts to a 2% increase

in the entry probability from the baseline level, keeping in mind that this change corresponds

to only a 1% change in the partner’s (as opposed to competitors’) market share. Alternatively,

we could compare two typical types of markets (i.e., average markets from the data) conditional

on the focal carrier not operating: one dominated by a partner (80% market share), the other

by a competitor (80% market share). In the first market, the carrier’s entry probability increases

from 0.0057 to 0.0098 after the alliance, i.e., is almost doubled20. In the second market, the entry

probability drops from 0.0126 to 0.0048, i.e., is more than halved. We further compute the average

effect based on the empirical joint distribution of partner/competitor market share from data – the

alliance is responsible for 4 more entries into partner-dominated markets, while having 18 fewer

entries into competitor-dominated markets annually. The difference-in-difference is 22 entries/year,

which is economically significant (20% of the baseline annual entries).

Furthermore, Figure 2 represents the marginal effects at different levels of operating carriers’

market shares: 0%, 30%, 50% and 80%, respectively. These are the typical levels of market shares

in the following four scenarios: when the airline is a potential entrant, and when the airline is an

incumbent in oligopoly, duopoly, and monopoly settings, respectively. The marginal effect of the

difference-in-difference term on the operating probability ranges from 0.01 to 0.06, as in Figure 2.

Note that this effect is largest in the oligopoly case (30%), where the airline is likely to compete with

one partner and one competitor. In this scenario, the tension between competitors from different

leagues makes it more valuable to gain additional market power. The effect is the smallest in the

monopoly case (80%), in which the operating airline is most likely to operate in such a market

regardless of the market power possessed by the partner, since the operating airline itself already

controls the market to a great extent. The effect for non-existing operating carriers (0%) and the

duopoly case (50%) fall in between, as expected.

The possible endogeneity caused by partner selection merits some discussion. One alternative

argument for the observed entry behavior is that markets operated by partners may be fundamen-

tally different from others, which will lead to the same changes even without alliances. However,

20 The main effect of alliance has been removed to isolate the effect of the interaction term between the alliance andpartner/competitor share.


difference-in-difference approach accounts for the “fundamental difference”, if there is any, that is

present even before an alliance. Second, in the subsequent section we conduct various robustness

tests on whether “the changes would be the same without alliances”, and the results are consistent.

Finally, we conduct placebo analysis using randomly chosen years, and there are no significant

changes of entry strategy in those years other than the alliance years, which suggests that there

appears to be no “fundamentally different” changes on the markets operated by partners due to

reasons other than alliances. Another explanation for what is observed here is “reverse causality” –

is it possible that airlines want to cooperate on certain routes (or to enter certain markets) which

leads to the alliance with a specific partner? This is not very likely according to our conversations

with airline practitioners: although airlines have a general preference on who to partner with based

on market positioning, detailed plans on flight operations and capacities form later. Second, even if

airlines have planned to enter certain markets and then choose to enter an alliance with a partner

based on this decision, it actually reinforces the fact that alliance is what is required to enable these

entries. Without alliance, it would be infeasible or unprofitable for airlines to enter these markets.

That is, the result shows exactly what alliance can achieve, regardless of when (i.e., ex-ante or

ex-post) airlines come to realize the benefits. To summarize, in this section we find strong support

for the Hypothesis 1: post-alliance airline are more likely to operate in the markets in which their

partners possess strong market power.

6.1.1. Robustness Tests To check whether our results are immune to variations in the

assumptions, we conducted the following two sets of major robustness tests. The first set of tests

aims at relaxing assumptions related to the identification of the “difference-in-difference” term.

The second uses the dynamic Probit model to relax assumptions regarding error terms.

Identification of the “Difference-in-Difference” term. Similar to the idea of the classical

difference-in-difference identification strategy, our identification is based on a few implicit assump-

tions: 1) Without the alliance treatment, the effects of the partner’s and competitors’ share on entry

probability would have followed the same trend over time. To check robustness, we allow for differ-

ent trends by adding separate yearly shocks to the partner’s effect (δp1) and the competitor’s effect

(δc1), equivalent to adding interaction terms between yearly dummies and the partner/competitor

market share. 2) Effects of partner’s/competitors’ market share are the same for all airlines. To

check robustness, we allow for carrier-specific “attitudes” towards partners and competitors by

including interaction terms between carrier dummies and partner/competitor market share. 3)

Without the alliance, the change of partner’s/competitors’ effects would have followed the same

trend for every carrier for both treated (allied) and untreated(non-allied) airlines. We relax this

assumption by including a carrier-specific linear trend in the effects of partner’s/competitor’s mar-

ket shares (similar to Besley and Burgess (2004)) which allow carriers to follow different trends in a


limited but revealing manner. The results of all of the aforementioned relaxations are presented in

Table 6. We note that the key results are largely consistent with what we obtained earlier, while the

scale of the difference-in-difference coefficient remains at the same level.21 The additional variables

do not dramatically increase the explanatory power of the model (R2 goes from 0.871 to 0.872).

Robustness Test of the Assumptions on the Error Terms. Recall that in the original

model the error term εimt was assumed to be uncorrelated and homoscedastic across markets, time

and carriers. Of course, even though we include as many relevant covariates as possible, there

may still be some omitted variables, causing the underlying profitability to be correlated within a

carrier-market group over time. To account for such a possibility we decompose the error term into

two parts: an unobserved carrier-market specific term and an idiosyncratic shock εimt = cim + εimt,

where cim can be regarded as the unobserved component of the carrier-market specific profitability

shock. In other words, airlines themselves might have a sense of which market is more profitable

based on their experience in the industry and based on the airline’s strengths and positioning but

this information may not be revealed to econometricians. Traditional random effects model would

require strict exogeneity E(cim|Wimt) = 0, where Wimt represents all the explanatory variables.

This assumption may not be valid since the variables included in Wimt, especially those we are

interested in (i.e., partner’s/competitor’s share), can be correlated with the unobserved carrier-

market profitability shock. The treatment (i.e., the alliance status) is applied at the carrier level

rather than at the market level, which means it cannot be correlated with cim after controlling for

carrier dummies. Nevertheless, the interaction terms which involve partner’s/competitor’s market

share can be correlated with this heterogeneity. A complication here is that, unlike in the linear

models where we can use the fixed effects to remove the unobserved carrier-market specific effect,

we do not have the luxury of doing so in a nonlinear model. To relax this exogeneity assumption,

we adopt the approach proposed by Chamberlain (1980) – Mundlak (1978) for nonlinear panel

data models, i.e., the Correlated Random Effects Probit Models. The essential idea is to explicitly

model the correlation between cim and Wimt in a specific way (see Wooldridge (2010), Chapter 15):

cim ∼Normal(ψ+Wimξ,σ2c ). (11)

where Wim is the average of Wimt over time. It turns out that this estimation can be done in

the traditional random effects framework by adding Wim to the original estimation. By allowing

21 Note that, after including interaction terms of the yearly dummy and the partner’s/competitors’ share, the iden-tification of the change in partner’s effect (δp1 − δp2) comes from the different timing of alliance formation. Theidentification of the change in the competitors’ effect (δc2 − δc1) comes from this different timing and the fact thatAA never joined an alliance. Since the timing is not so different (UA, US in 2003, and DL in 2004), it is not surprisingthat the change in partner’s effect is absorbed by year-specific partner effect. However, the key conclusion stands asthe difference-in-difference term is still significant and at the same scale.


this carrier-market specific effect, we account for the subject-specific heterogeneity (which can be

endogenous), which forces the observations of the same carrier on the same market to be correlated

over time. We further relax this assumption by testing for the serial correlation of the new error

term εimt. The results are displayed in Table 7 and are, once again, consistent with our earlier

findings in terms of the size of the coefficients as well as significance.

6.2. Estimation Results for the Capacity Model

In this section, we continue discussion of our findings for the capacity model. Table 8 shows the

estimation results. The model is estimated under Pooled OLS, Random Effects and Fixed Effects

models when lagged dependent variable is not included, and under dynamic panel data model

(Arellano-Bond estimator) when lagged dependent variable is included. The results are consistent

with our earlier results related to market entry: airlines have given favorable consideration to

routes operated by alliance partners after the alliance as opposed to those operated by competitors.

Thus, airlines build up capacities in the markets where their partner holds strong market power,

in support of Hypothesis 2.

Since both Fixed Effects and Arellano-Bond estimators allow for the correlation of the error

term and the covariates (Random Effects model is rejected under Huasman test), we focus on the

results from the Fixed Effects and Arellano-Bond estimation which also shows the strongest results.

To assess the economic impact of these estimates, we compare two typical duopoly markets: one

operated by the carrier of interest and its partner, the other by the same carrier and one of its

competitors. Each carrier possesses 45% of the market share, and operates 250,000 seats annually.

Note that, under these assumptions, the only difference is the “identity” of the other player (i.e.,

a partner or a competitor). According to our estimates, in the market operated with a partner,

after the alliance, each airline increases seat capacity by 8.69%, which corresponds to 21,700 more

seats annually (418 seats weekly) – roughly 3 additional flights per week. However, if the market

is operated with a competitor, each airline would reduce its annual seat capacity by 6.65%, or

16,625 fewer seats annually (319 fewer seats weekly) – 2 less flight per week. If we look at the

difference-in-difference estimate, the capacity change in the partner’s market over the competitor’s

market is 15%, or 5 flights per week. These capacity changes are clearly economically significant

in addition to being statistically significant.

7. Drivers of the Change in Entry/Capacity Strategies

Both our entry model and capacity model indicate that, after the alliance, carriers give favorable

consideration to markets in which their alliance partner(s) have larger market share. This change

of strategy is likely to be indicative of higher profitability when operating a direct flight in such

markets after the alliance. There are at least three obvious reasons that might contribute to the


higher profitability of these routes: 1) lower entry and operating costs in those markets through

sharing of facilities and personnel and joint activities such as maintenance and procurement; 2)

higher prices, since two codeshare partners may be able to charge passengers more due to better

product quality (e.g., broader travel options), higher demand or through tacit price collusion; 3)

larger demand due to collaborative marketing initiatives and reciprocal frequent flyer programs,

which may induce consumers to switch from non-allied airlines to allied airlines. In this section we

conduct post-hypotheses analysis to better understand drivers of the change in the market entry

strategy. Since in the data available to us we neither observe segment-based nor market-based

costs, it is hard for us to analyze how the cost structure has changed before and after the alliance

without making structural assumptions, which is well beyond the scope of this paper. Instead, we

examine changes in price and demand. We model the impact of an alliance on prices and demand

using a framework which is similar to the one we employed earlier:

zimt =W ′imtβ+ f(Dit; i,m, t) +λt +αi + εimt (12)

The left-hand variable zimt can be the average segment fare or the number of enplaned passengers

or the load-factor of carrier i on market m at time t. Note that the difference from the previous

models is that this equation is contemporaneous rather than lagged.

Since airline fare is only quoted for the entire itinerary which may contain several segments, the

fare proportion that corresponds to each segment is not obvious. We use the distance-weighted

approach to pin down the proportion of the total fare contributed by each segment, and we obtain

the average of those segment fares. This approach is also used by Dana and Orlov (2009). We

are aware that this approach is subject to measurement errors since airline pricing strategy is

not cost-based, but is rather based on the opportunity cost according to the practice of revenue

management. However, the exact breakdown of fares is impossible without detailed information

on the ticket purchase date and associated restrictions. The public data sources we have in our

possession do not provide such data. To assess the sensitivity of our results to this fare breakdown

approach we also performed a robustness analysis on the itinerary-based city-pair average fare and

we obtained similar results (not reported here).

7.1. Estimation Results for Price and Demand

The estimation results for the average segment fare are displayed in Table 9, while the results for

enplaned passengers and load-factors are shown in Table 10. We estimate the model using Pooled

OLS, the Random Effect and the Fixed Effect models. The Hausman Test again rejects the Random

Effects model so we draw the following implications based on the Fixed Effects model. Table 9

indicates that, after the alliance, airlines are able to charge higher fares when their partners are


also present, but charge lower fares when their competitors are present. To put it more concretely,

we again use the typical duopoly market example. After the alliance, airlines are able to charge

$2.3 more on average in the markets operated with the alliance partner, while prices drop by $6.3

in markets with a competitor. Considering the combined difference, an $8.6 premium is charged for

an average one-way coupon fare on markets with partners compared to markets with competitors.

A similar estimation (omitted) on the city-pair average market fare reveals that duopoly markets

dominated by two same-alliance airlines charge an $18 premium for a round-trip itinerary.

Moving to the analysis of traffic changes, we see that the “difference-in-difference” estimation

is positive and significant, but the partner’s and the competitor’s effects deserve some discussion.

After the alliance, the total number of enplaned passengers increases in the markets shared with

partners, but this effect is not statistically significant. Using the same duopoly market example as

before, the point estimate of the increase is 50 passengers per week, even though the capacity is

estimated to increase by 420 seats per week (as we recall from the capacity estimation). Conversely,

in those markets where competitors are present, there is a significant decline in the passenger

number by 260 per week. As a reference point, the capacity is estimated to decrease by 320 seats per

week on these routes. From these estimates it appears that the airlines have been over-optimistic

about the demand in the markets shared with alliance partners and have over-adjusted their

capacity in those markets. We verify that this is the case by analyzing the load-factor as shown

on the right-hand side of Table 10. It appears from this estimation that the traffic increases in the

markets shared with partners as compared to markets shared with competitors, but the load-factor

drops, which is probably indicative of over-capacity on these routes.

To sum up, it appears that higher price might be the reason that drives airlines to enter or

increase capacity in markets where their partners hold strong market power. However, we do not

have sufficient support for the hypothesis that this is driven by higher demand: on the contrary,

it would seem that airlines have been over-adjusting capacities (or build up capacity to deter

new entrants, which is more troublesome from antitrust perspective) without a notable increase

of demand in markets operated with partners. The sizes of all the relevant effects are summarized

below for convenience.

[own share=45%, [own share=45%, Diff-in-Diffpartner share = 45%] competitor share = 45%]

Capacity 3 more flights/week 2 less flight/week 5 more flights/weekLeg Average Fare $2.3 higher $6.3 lower $8.6 preimumEnplaned Passenger 50 more people/week (insignificant) 260 fewer people/week 310 more people/weekLoad-factor down by 3.5 percentage points up by 2.4 percentage points 5.9 percentage points down

8. Concluding Remarks

In this paper, we study changes in airlines’ entry and capacity decisions after joining an alliance.

We further test possible drivers of these changes in the entry behavior. We find that, after an


alliance, airlines are more likely to enter/stay or increase capacity in the markets where their

partners possess strong market power, compared to those in which competitors are present. We

also find that one possible reason for this higher likelihood of “cooperating” with partners is higher

pricing potential rather than higher demand. These findings are highly economically significant.

Our findings have important implications for both policy makers and airline managers. Policy

makers are typically concerned about allowing alliances between airlines that already have many

overlapping markets. We show that this is a somewhat limited and static view of airline networks.

Our study indicates that it is also crucial to look at the longer-term picture. In particular, we

analyze how airlines change their networks dynamically after an alliance is formed. We find that

they tend to increase overlaps with partners post-alliance and the reason might be that the airlines

are able to charge higher prices this way (without much indication of increased demand). Though

this increase in price may reflect better travel products provided after an alliance, there is a

possibility that the increase in price is due to collusive behavior. Further research is needed to shed

light on this issue but it seems certain that policy makers should not only focus on the pre-alliance

network structure. For instance, as a condition of the alliance, regulators might want to require

that airlines do not increase network overlaps with partners on certain routes without sufficient

evidence of higher demand/higher load factors.

Based on our results, an airline manager can better understand the changes that are likely to

follow an alliance among its competitors. For instance, from our analysis it appears that airlines

have been over-confident about the demand increase in the markets operated with partners after

an alliance, and have over-adjusted capacities in these markets (possibly in anticipation of larger

demand). A careful reassessment of the demand in those markets would help airlines better match

supply with demand.

Our findings are somewhat provocative and therefore there is a need for further investigation into

the driving forces of observed changes and into the underlying profit structure of alliance partners.

We are, of course, greatly limited by the data that is available. Future research could use structural

estimation to better understand the reasons for the price increases that we document. Are they the

result of enhanced product quality or of price collusion? How do entry costs and operating costs

change after an alliance? Can we disentangle the size of each effect? More data will be needed to

answer these and other questions.

Acknowledgments

The authors are grateful for the helpful suggestions provided by Saravanan Kesavan, Margaret

Pierson, Fangyun Tan, and Antonio Moreno-Garcia, Mark Ferguson, Laurie Garrow and Nelson

Granados. They also thank the Mack Center for Technological Innovation, the Wharton School,

University of Pennsylvania, for financial support.


References

Aguirregabiria, V., C. Ho. 2009. A Dynamic Oligopoly Game of the US Airline Industry: Estimation and

Policy Experiments. Working paper, University of Toronto, Canada.

Ai, C., C. Norton. 2003. Interaction terms in logit and probit models. Economics Letters 80(1) 123–129.

Anderson, S., R.G. Newell. 2003. Simplified marginal effects in discrete choice models. Economics Letters

81(3) 321–326.

Arikan, M., V. Deshpande. 2010. The impact of airline flight schedules on flight delays. Working paper,

Purdue University, West Lafayette, IN.

Arikan, M., V. Deshpande, M. Sohoni. 2010. Building reliable air-travel infrastructure using empirical data

and stochastic models of airline networks. Working paper, Purdue University, West Lafayette, IN.

Armantier, O., O. Richard. 2008. Domestic airline alliances and consumer welfare. RAND Journal of

Economics 39(3) 875–904.

Bajari, P., L. Benkard, J. Levin. 2007. Estimating dynamic models of imperfect competition. Econometrica

75(5) 1331–1370.

Bamberger, G.E., D. W. Carlton, L.R. Leumann. 2004. An empirical investigation of the competitive effects

of domestic airline alliances. Journal of Law and Economics 47(1) 195–222.

Benkard, C.L., A. Bodoh-Creed, J. Lazarev. 2010. The long run effects of us airline mergers. Working paper,

Yale University, New Haven, CT.

Berry, S. 1992. Estimation of a model of entry in the airline industry. Econometrica 60(4) 889–917.

Besley, T., R. Burgess. 2004. Can labor regulation hinder economic performance? evidence from india.

Quarterly Journal of Economics 119(1) 94–134.

Borenstein, S., N.L. Rose. 1994. Competition and price dispersion in the u.s. airline industry. Journal of

Political Economy 102(4) 653–683.

Brueckner, J.K. 2003. International airfares in the age of alliances: The effect of codesharing and antitrust

immunity. The Review of Economics and Statistics 85(1) 105–118.

Brueckner, J.K., W.T. Whalen. 2000. The price effect of international airline alliances. Journal of Law &

Economics 43(2) 503–545.

Chamberlain, G. 1980. Analysis of covariance with qualitative data. Review of Economics Studies 47(1)

225–238.

Cho, M., M. Fan, Y.P. Zhou. 2007. An empirical study of revenue management practices in the airline

industry. Working Paper, University of Washington, Seattle, WA.

Chung, S., H. Singh, K. Lee. 2000. Complementarity, status similarity and social capital as drivers of alliance

formation. Strategic Management Journal 21(1) 1–22.


Cilliberto, F., E. Tamer. 2009. Market structure and multiple equilibria in airline markets. Econometrica

77(6) 1791–1828.

Dana, J., E. Orlov. 2009. Internet penetration and capacity utilization in the us airline industry. Working

Paper, Northeastern University, Boston, MA.

Department of Justice. 1999a. Competition in the airline industry.

Department of Justice. 1999b. The importance of entry conditions in analyzing airline antitrust issues.

Department of Transportation. 2003. Termination of review under 49 U.S.C. 41720 of

Delta/Northwest/Continental agreements. Office of the Secretary.

Dyer, J.H., H. Singh. 1998. The relational view: Cooperative strategy and sources of interorganizational

competitive advantage. Academy of Management Review 23(4) 660–679.

Ferguson, M.E., L.A. Garrow, M. Meterelliyoz, J.P. Newman. 2010. Multi-resource revenue management

with upgrades: A comparison of emsr-based and choice-based algorithms. Working paper, Georgia

Institute of Technology, Atlanta, GA.

Garrow, L.A. 2010. Discrete Choice Modelling and Air Travel Demand: Theory and Applications. Ashgate

Publishing, Aldershot,United Kingdom.

Gayle, P.G. 2007. Airline code-share alliances and their competitive effects. Journal of Law and Economics

50(4) 781–819.

Gayle, P.G. 2008. An empirical analysis of the competitive effects of the delta/continental/northwest code-

share alliance. Journal of Law and Economics 40(4) 743–766.

General Accounting Office. 1998. Aviation competition- proposed domestic airline alliances raise serious

issues.

General Accounting Office. 1999. Aviation competition - effects on consumers from domestic airline alliances

vary.

Gimeno, J. 2004. Competition within and between networks: The contingent effect of competitive embed-

dedness on alliance formation. The Academy of Management Journal 47(6) 820–842.

Granados, N.F., A. Gupta, R.J. Kauffman. 2011. Online and offline demand and price elasticities: Evidence

from the air travel industry. Information Systems Research Forthcoming.

Greene, W.H. 2007. Econometric Analysis. 6th ed. Prentice Hall, Upper Saddle River, NJ.

Gulati, R. 1995. Social structure and alliance formation patterns: a longitudinal analysis. Administrative

Science Quarterly 40(4) 619–652.

Hu, X., R. Caldentey, G. Vulcano. 2010. Revenue sharing in airline alliances. Working Paper, New York

University, New York City, NY.

Ito, H., D. Lee. 2007. Domestic codesharing, alliances and airfares in the u.s. airline industry. Journal of

Law and Economics 50(2) 355–380.


Lapre, M.A., N. Tsikriktsis. 2006. Organizational learning curves for customer dissatisfaction: Heterogeneity

across airlines. Management Science 52(3) 352–366.

Li, W., M. Lipson, K. Ramdas, J. Williams. 2010. Can stock price movements inform operational improve-

ment efforts? evidence from the airline industry. Working Paper, University of Virginia, Charlottesville,

VA.

Mundlak, Y. 1978. On the pooling of time series and cross section data. Econometica 46(1) 69–85.

Newman, J.P., M.E. Ferguson, L.A. Garrow. 2010. Estimation of choice-based models using sales data from

a single firm. Working paper, Georgia Institute of Technology, Atlanta, GA.

Oum, T.H., J.H. Park, A. Zhang. 1996. The effects of airline codesharing agreements on firm conduct and

international air fares. Journal of Transport Economics and Policy 30(2) 187–202.

Ramdas, K., J. Williams. 2009. An empirical investigation into the tradeoffs that impact on-time performance

in the airline industry. Working Paper, London Business School, London, UK.

Reitzes, J., D. Moss. 2008. Airline alliances and systems competition. Houston Law Review .

Shen, C. 2010. The effects of u.s. domestic airline code sharing on costs, markups and consumer welfare.

Working Paper, Boston University, Boston, MA.

Talluri, K.T., G. J. van Ryzin. 2004. The Theory and Practice of Revenue Management . Kluwer Academic

Publishers, Boston, MA.

Transportation Research Board. 1999. Entry and Competition in the U.S. Airline Industry-Issues and Oppor-

tunities – Special Report 255 . National Academy Press, Washington, D.C.

Vulcano, G., G. van Ryzin, W. Chaar. 2010. Choice-based revenue management: An empirical study of

estimation and optimization. Manufacturing and Service Operations Management 12(3) 371–392.

Wooldridge, J.M. 2010. Econometric Analysis of Cross Section and Panel Data. 2nd ed. The MIT Press,

Cambridge, Massachusetts, London, England.

Wright, C.P., H. Groenevelt, R.A. Shumsky. 2010. Dynamic revenue management in airline alliances. Trans-

portation Science 44(1) 15–37.


Ta

ble

1D

escr

ipti

on

of

Var

iab

les

Cate

gory

Varia

ble

sD

esc

rip

tion

(1)

ow

n{0

,1}

vari

ab

lein

dic

ati

ng

wh

eth

erth

eair

lin

eis

op

erati

ng

inth

em

ark

et(2

)ca

paci

tynu

mb

erof

seats

op

erate

dby

op

erati

ng

air

lin

e(3

)p

ass

enger

nu

mb

erof

pass

enger

sen

pla

ned

on

op

erati

ng

air

lin

e’s

flig

hts

(4)

aver

age

fare

aver

age

on

e-w

ay

dir

ecti

on

al

leg

fare

(ob

tain

edfr

om

itin

erary

fare

,d

ista

nce

-wei

ghte

d)

(5)

allia

nce

{0,1}

vari

ab

lein

dic

ati

ng

wh

eth

erth

eop

erati

ng

carr

ier

has

form

edany

allia

nce

ina

part

icu

lar

yea

r(6

)ow

nsh

are

mark

etsh

are

of

the

op

erati

ng

carr

ier

on

the

dir

ecti

on

al

leg

(in

clu

din

gon

lyd

irec

tfl

ights

)(7

)p

art

ner

share

mark

etsh

are

of

the

all

ian

cep

art

ner

of

the

op

erati

ng

carr

ier

on

the

dir

ecti

on

al

leg

(8)

com

pet

itor

share

mark

etsh

are

of

the

ou

t-of-

all

ian

ceco

mp

etit

ors

of

the

op

erati

ng

carr

ier

on

the

dir

ecti

on

al

leg

(9)

part

ner

share

xallia

nce

inte

ract

ion

term

of

part

ner

share

an

dallia

nce

(10)

com

pet

itor

share

xallia

nce

inte

ract

ion

term

of

com

pet

itor

share

an

dallia

nce

Seg

men

t(1

1)

log(d

ista

nce

)lo

gof

non

-sto

pd

ista

nce

bet

wee

nori

gin

an

dd

esti

nati

on

Fea

ture

s(1

2)

log

sqrt

(pop

1*

pop

2)

log

of

geo

met

ric

aver

age

of

pop

ula

tion

at

the

ori

gin

an

dth

ed

esti

nati

on

MS

As

(13)

log(i

nco

me)

log

of

aver

age

of

per

cap

ita

inco

me

at

the

ori

gin

an

dth

ed

esti

nati

on

MS

As

(14)

load

fact

or

aver

age

con

ges

tion

level

of

the

leg

(tota

l#

of

pass

enger

s/

tota

l#

of

seats

)(1

5)

HH

IH

erfin

dah

l-H

irsc

hm

an

Ind

exof

the

com

pet

itio

nle

vel

on

the

leg

(16)

lcc

mark

etsh

are

of

the

low

-cost

carr

iers

on

the

leg

(17)

ow

nlo

ad

fact

or

con

ges

tion

level

of

the

op

erati

ng

carr

ier

on

the

leg

Nod

e(1

8)

ori

gin

HH

IH

erfi

nd

ah

l-H

irsc

hm

an

Ind

exof

the

com

pet

itio

nle

vel

at

the

ori

gin

Fea

ture

s(1

9)

des

tH

HI

Her

fin

dah

l-H

irsc

hm

an

Ind

exof

the

com

pet

itio

nle

vel

at

the

des

tin

ati

on

(20)

ori

gin

lcc

mark

etsh

are

of

the

low

-cost

carr

iers

at

the

ori

gin

(21)

des

tlc

cm

ark

etsh

are

of

the

low

-cost

carr

iers

at

the

des

tin

ati

on

(22)

ow

nori

gin

deg

ree

aver

age

of

ind

egre

ean

dou

tdeg

ree

of

the

op

erati

ng

carr

ier

at

the

ori

gin

city

(deg

ree

isd

efin

edby

the

nu

mb

erof

dir

ect

flig

hts

)(2

3)

ow

nd

est

deg

ree

aver

age

of

ind

egre

ean

dou

tdeg

ree

of

the

op

erati

ng

carr

ier

at

the

des

tin

ati

on

city

(deg

ree

isd

efin

edby

the

nu

mb

erof

dir

ect

flig

hts

)(2

4)

ow

nori

gin

share

mark

etsh

are

of

the

op

erati

ng

carr

ier

at

the

ori

gin

(25)

ow

nd

est

share

mark

etsh

are

of

the

op

erati

ng

carr

ier

at

the

des

tin

ati

on

Ed

ge

(26)

#of

ind

irec

tp

ath

snu

mb

erof

con

nec

tin

gro

ute

sby

the

op

erati

ng

carr

ier

on

the

city

pair

mark

etF

eatu

res

(27)

city

pair

HH

IH

erfi

nd

ah

l-H

irsc

hm

an

Ind

exof

the

com

pet

itio

nle

vel

at

the

city

pair

mark

et(i

ncl

ud

ing

both

dir

ect

an

dco

nn

ecti

ng

rou

tes)

(28)

city

pair

lcc

mark

etsh

are

of

the

low

-cost

carr

ier

at

city

pair

mark

et(i

ncl

ud

ing

both

dir

ect

an

dco

nn

ecti

ng

rou

tes)


Ta

ble

2S

um

mar

yS

tati

stic

sa

nd

Co

rrel

ati

on

Ta

ble

Mea

nS

td.

Dev

.(1

)(2

)(3

)(4

)(5

)(6

)(7

)(8

)(9

)(1

0)

(11)

(12)

(13)

(14)

(1)

ow

n0.0

78

0.2

68

1.0

0(2

)ca

paci

ty19982

97628

0.6

81.0

0(3

)p

ass

enger

14111

69826

0.6

70.9

91.0

0(4

)aver

age

fare

150.3

076.9

6-0

.07

-0.0

5-0

.05

1.0

0(5

)allia

nce

0.5

63

0.4

96

-0.0

3-0

.05

-0.0

4-0

.07

1.0

0(6

)ow

nsh

are

0.6

32

0.3

20

0.8

80.6

00.6

0-0

.06

0.0

01.0

0(7

)p

art

ner

share

†0.7

01

0.3

04

-0.1

2-0

.09

-0.0

9-0

.12

0.1

6-0

.14

1.0

0(8

)co

mp

etit

or

share

†0.7

35

0.2

84

-0.1

0-0

.07

-0.0

7-0

.03

-0.1

1-0

.18

-0.1

81.0

0(9

)p

art

ner

share

*co

des

hare

†0.3

87

0.4

10

-0.0

9-0

.07

-0.0

7-0

.11

0.2

7-0

.10

0.7

2-0

.14

1.0

0(1

0)

com

pet

itor

share

*co

des

hare

†0.4

20

0.4

16

-0.0

9-0

.07

-0.0

7-0

.04

0.3

7-0

.12

-0.0

90.5

8-0

.05

1.0

0(1

1)

log(d

ista

nce

)6.8

20

0.7

48

-0.1

2-0

.09

-0.0

70.5

50.0

0-0

.11

-0.0

9-0

.07

-0.0

7-0

.02

1.0

0(1

2)

log

sqrt

(pop

1x

pop

2)

14.3

84

0.6

65

0.2

00.2

60.2

50.0

50.0

00.0

70.0

80.2

90.0

40.2

30.1

01.0

0(1

3)

log(i

nco

me)

10.3

94

0.1

28

0.1

00.1

10.1

10.0

10.2

80.0

50.0

80.1

20.1

50.2

00.0

80.4

41.0

0(1

4)

load

fact

or†

0.5

08

0.3

35

0.2

70.1

80.1

9-0

.21

0.0

30.2

40.2

20.3

60.1

70.2

3-0

.16

0.3

30.1

81.0

0(1

5)

HH

I0.6

15

0.3

85

-0.0

3-0

.06

-0.0

6-0

.05

-0.0

10.1

40.1

20.0

30.1

0-0

.03

-0.0

8-0

.19

-0.0

70.1

9(1

6)

lcc

0.2

18

0.3

60

-0.1

1-0

.07

-0.0

6-0

.10

0.0

5-0

.16

-0.1

5-0

.22

-0.1

0-0

.11

0.0

6-0

.01

-0.0

50.0

6(1

7)

ow

nlo

ad

fact

or†

0.7

26

0.1

04

0.9

90.6

80.6

8-0

.07

-0.0

30.8

7-0

.13

-0.1

1-0

.09

-0.0

9-0

.09

0.1

90.1

10.2

8(1

8)

ori

gin

HH

I0.2

92

0.1

68

0.1

30.0

80.0

8-0

.02

0.0

00.1

70.1

30.1

20.0

90.0

4-0

.11

-0.0

4-0

.07

0.1

5(1

9)

des

tH

HI

0.2

92

0.1

68

0.1

10.0

70.0

6-0

.05

-0.0

20.1

60.1

60.1

50.0

90.0

5-0

.13

-0.0

5-0

.07

0.1

6(2

0)

ori

gin

lcc

0.2

89

0.1

91

-0.1

1-0

.09

-0.0

80.0

00.1

0-0

.12

-0.1

1-0

.17

-0.0

4-0

.05

0.1

7-0

.05

0.0

0-0

.07

(21)

des

tlc

c0.2

89

0.1

91

-0.1

1-0

.08

-0.0

70.0

10.1

1-0

.11

-0.1

3-0

.19

-0.0

5-0

.07

0.1

8-0

.04

0.0

0-0

.08

(22)

ow

nori

gin

deg

ree

15.7

48

16.3

12

0.4

20.3

90.3

80.0

2-0

.14

0.3

5-0

.09

0.0

0-0

.09

-0.0

50.0

10.3

00.1

50.1

7(2

3)

ow

nori

gin

deg

ree

15.7

48

16.3

12

0.4

20.3

80.3

80.0

2-0

.13

0.3

5-0

.10

0.0

0-0

.09

-0.0

50.0

00.3

00.1

50.1

5(2

4)

ow

nori

gin

share

0.1

15

0.1

78

0.0

2-0

.01

-0.0

20.0

1-0

.07

0.0

60.0

20.0

30.0

00.0

1-0

.12

-0.0

3-0

.03

-0.0

5(2

5)

ow

nd

est

share

0.1

15

0.1

78

0.0

2-0

.01

-0.0

20.0

0-0

.07

0.0

50.0

30.0

60.0

00.0

2-0

.13

-0.0

3-0

.02

-0.0

3(2

6)

#of

ind

irec

tp

ath

s0.7

78

1.2

63

0.0

20.0

30.0

30.1

0-0

.02

-0.0

10.0

10.1

0-0

.03

0.0

70.2

50.2

20.1

80.1

1(2

7)

city

pair

HH

I0.3

45

0.2

70

0.1

80.1

30.1

2-0

.25

-0.0

10.2

40.2

00.2

30.1

40.1

1-0

.34

0.0

70.0

50.4

3(2

8)

city

pair

lcc

0.1

82

0.2

57

-0.0

4-0

.01

0.0

0-0

.18

0.0

7-0

.12

-0.1

1-0

.15

-0.0

5-0

.05

-0.0

30.0

40.0

00.2

0

Mea

nS

td.

Dev

.(1

5)

(16)

(17)

(18)

(19)

(20)

(21)

(22)

(23)

(24)

(25)

(26)

(27)

(28)

(15)

HH

I0.6

20.3

81.0

0(1

6)

lcc

0.2

20.3

6-0

.07

1.0

0(1

7)

ow

nlo

ad

fact

or†

0.0

50.1

9-0

.03

-0.1

11.0

0(1

8)

ori

gin

HH

I0.2

90.1

70.1

5-0

.19

0.1

21.0

0(1

9)

des

tH

HI

0.2

90.1

70.1

7-0

.21

0.1

0-0

.02

1.0

0(2

0)

ori

gin

lcc

0.2

90.1

9-0

.09

0.4

2-0

.10

-0.3

80.0

11.0

0(2

1)

des

tlc

c0.2

90.1

9-0

.10

0.4

4-0

.09

0.0

1-0

.40

0.0

31.0

0(2

2)

ow

nori

gin

deg

ree

15.7

516.3

1-0

.05

-0.0

30.4

10.0

0-0

.05

-0.1

70.0

61.0

0(2

3)

ow

nori

gin

deg

ree

15.7

516.3

1-0

.06

-0.0

30.4

1-0

.04

-0.0

20.0

5-0

.16

-0.1

11.0

0(2

4)

ow

nori

gin

share

0.1

10.1

80.0

7-0

.16

0.0

10.1

70.0

1-0

.40

-0.0

4-0

.06

0.0

21.0

0(2

5)

ow

nd

est

share

0.1

10.1

80.0

9-0

.16

0.0

10.0

00.2

1-0

.03

-0.4

00.0

2-0

.07

0.0

01.0

0(2

6)

#of

ind

irec

tp

ath

s0.7

81.2

6-0

.03

-0.0

10.0

3-0

.15

-0.1

3-0

.13

-0.1

00.0

80.0

70.2

00.1

91.0

0(2

7)

city

pair

HH

I0.3

40.2

70.3

6-0

.04

0.1

70.2

80.1

8-0

.19

-0.1

30.0

50.0

40.0

80.1

10.0

11.0

0(2

8)

city

pair

lcc

0.1

80.2

6-0

.13

0.7

7-0

.03

-0.1

6-0

.17

0.4

60.4

2-0

.01

0.0

1-0

.22

-0.2

2-0

.08

0.0

81.0

0

†:S

um

mary

stati

stic

ssh

ow

nfo

rth

ese

vari

ab

les

are

con

dit

ion

al

on

op

erati

ng.


Table 3 Airline Route Statistics 1999-2006

UA US AAstay entry exit turnover stay entry exit turnover stay entry exit turnover

1999 282 3 4 0.024 338 18 9 0.078 255 26 2 0.1092000 281 9 4 0.046 338 13 18 0.087 275 46 6 0.1852001 278 5 12 0.059 331 1 20 0.060 297 11 24 0.1092002 259 2 24 0.092 258 2 74 0.229 289 98 19 0.3802003 243 9 14 0.089 228 2 32 0.131 370 17 16 0.0852004 246 25 6 0.123 221 25 9 0.148 334 19 53 0.1862005 257 31 14 0.166 221 29 25 0.220 344 16 9 0.0712006 264 10 24 0.118 241 17 9 0.104 346 7 14 0.058

DL CO NWstay entry exit turnover stay entry exit turnover stay entry exit turnover

1999 388 12 12 0.060 209 11 5 0.075 234 8 2 0.0422000 391 21 9 0.075 219 7 1 0.036 240 5 2 0.0292001 380 5 32 0.090 226 11 0 0.049 244 4 1 0.0202002 363 15 22 0.096 231 3 6 0.038 247 3 1 0.0162003 344 6 32 0.101 229 2 5 0.030 239 3 11 0.0562004 334 30 16 0.131 231 9 0 0.039 239 10 3 0.0542005 316 33 48 0.223 237 6 3 0.038 241 13 8 0.0842006 319 12 30 0.120 243 8 0 0.033 218 3 36 0.154

Note: AA acquired Trans-World Airlines in 2001.

Table 4 U.S. Domestic Airlines Flight Network and Overlap 1998-2006

UA & US DL,NW & CO Airlines from different alliancesoverlapped total overlapped overlapped total overlapped overlapped total overlapped# routes # routes % # routes # routes % # routes # routes %

1998 33 600 5.50% 70 780 8.97% 291 1318 22.08%1999 39 602 6.48% 64 798 8.02% 304 1331 22.84%2000 38 603 6.30% 67 816 8.21% 334 1352 24.70%2001 33 582 5.67% 60 810 7.41% 323 1324 24.40%2002 22 495 4.44% 60 800 7.50% 282 1358 20.77%2003 24 458 5.24% 55 768 7.16% 266 1314 20.24%2004 33 484 6.82% 64 789 8.11% 285 1309 21.77%2005 44 495 8.89% 64 782 8.18% 263 1340 19.63%2006 53 479 11.06% 62 741 8.37% 245 1293 18.95%


Table 5 Probit Model of Entry/Stay/Exit 1998-2006

(1) (2) (3)

Parameters of Interest alliance 0.412*** 0.871*** 0.763***(0.053) (0.079) (0.081)

own share 1.291*** 0.446*** 0.460***(0.072) (0.060) (0.061)

partner share before alliance (δp1) -0.690*** -0.945*** -1.095***(0.090) (0.113) (0.114)

competitor share before alliance (δc1) -0.204*** -0.267*** -0.409***(0.051) (0.060) (0.062)

partner share after alliance (δp2) -0.474*** -0.531*** -0.668***(0.077) (0.100) (0.103)

competitor share after alliance (δc2) -0.797*** -0.645*** -0.793***(0.060) (0.074) (0.076)

change in partner effect (δp2− δp1) 0.216*** 0.414*** 0.427***(0.111) (0.141) (0.141)

change in competitor effect (δc2− δc1) -0.593*** -0.378*** -0.384***(0.070) (0.082) (0.082)

difference-in-difference 0.809*** 0.793*** 0.811***(δp2− δp1)-(δc2− δc1) (0.127) (0.156) (0.156)

Segment Features log(distance) -0.241*** -0.199*** -0.228***(0.015) (0.018) (0.021)

log sqrt(pop1 * pop2) 0.303*** -0.089*** -0.083***(0.023) (0.028) (0.028)

log(income) 0.652*** 0.389*** 0.246(0.156) (0.181) (0.184)

loadfactor 0.779*** 0.970*** 0.985***(0.067) (0.086) (0.089)

HHI -1.056*** -1.154*** -1.330***(0.060) (0.073) (0.078)

lcc -0.839*** -0.273*** -0.314***(0.060) (0.077) (0.117)

own loadfactor 4.601*** 3.902*** 3.898***(0.068) (0.074) (0.075)

Node Features origin HHI 0.757*** 0.822***(0.086) (0.090)

dest HHI 0.770*** 0.912***(0.087) (0.090)

origin lcc -0.410*** -0.211(0.117) (0.122)

dest lcc -0.385*** -0.256(0.116) (0.119)

own origin degree 0.029*** 0.030***(0.001) (0.001)

own dest degree 0.029*** 0.030***(0.001) (0.001)

own origin share 0.240** 0.152*(0.079) (0.082)

own dest share 0.227** 0.092(0.080) (0.083)

Edge Features # indirect paths 0.060***(0.012)

citypair HHI 0.568***(0.074)

citypair lcc -0.224(0.130)

year dummy yes yes yescarrier dummy yes yes yes

# obs 181536 181536 181536Pseudo R-Square 0.8703 0.8917 0.8927

Note: *** p < 0.01, ** p < 0.05, * p < 0.1


Table 6 Robustness Check I - Relaxation of “Difference-in-Difference” Assumptions

Relax 1 Relax 1 & 2 Relax 1, 2 & 3

change in partner effect 0.009 -0.053 -0.096(0.140) (0.156) (0.175)

change in competitor effect -0.731*** -0.643*** -0.811***(0.082) (0.133) (0.225)

difference-in-difference 0.740*** 0.590*** 0.715***(0.157) (0.195) (0.281)

Control for yearly specific coefficient yes yes yesControl for carrier specific base effect no yes yesControl for carrier specific trend (linear) no no yes

# obs 181536 181536 181536Pseudo R-Square 0.8710 0.8715 0.8717

Table 7 Robustness Check II - Results for Subject-Specific Heterogeneity and Serial Correlation

Traditional RE Correlated RE Generalized Estimating Equation Serial CorrelationGEE GEE-AR(1)

change in partner effect 0.233*** 0.343*** 0.328*** 0.275***(0.117) (0.130) (0.114) (0.113)

change in competitor effect -0.614*** -0.564*** -0.450*** -0.509***(0.073) (0.081) (0.074) (0.074)

difference-in-difference 0.848*** 0.907*** 0.778*** 0.784***(0.089) (0.097) (0.087) (0.084)

# obs 181536 181536 181536 181536

Table 8 Regression of Capacity conditional on Stay 1998-2006

Without Lagged Capacity With Lagged CapacityPooled OLS Random Effect Fixed Effect Arellano-Bond Estimator

change in partner effect 0.116* 0.153*** 0.193*** 0.175***(0.062) (0.034) (0.032) (0.054)

change in competitor effect 0.055 -0.160** -0.148*** -0.140***(0.037) (0.021) (0.021) (0.029)

difference-in-difference 0.061 0.313*** 0.341*** 0.315***(0.071) (0.039) (0.038) (0.060)

# observation 13351 13351 13351 13351adj. R-square 0.576 -within 0.212 0.244between 0.454 0.192overall 0.454 0.215


Table 9 Regression of Average Segment Fare 1998-2006

Pooled OLS Random Effect Fixed Effect

change in partner effect 18.885*** 6.673*** 5.080***(3.145) (2.551) (2.587)

change in competitor effect 6.301*** -10.380*** -13.983***(1.827) (1.514) (1.554)

difference-in-difference 12.584*** 17.053*** 19.063***(3.543) (2.887) (2.945)

# obs 14084 14084 14084adj. R-square 0.828within R-sq 0.612 0.621between R-sq 0.818 0.129overall R-sq 0.803 0.163

Table 10 Regression of Traffic and Load-factor 1998-2006

Traffic Load-factor

Pooled OLS Random Effect Fixed Effect Pooled OLS Random Effect Fixed Effectchange in partner effect 0.065 0.016 0.033 -0.003 -0.064*** -0.077***

(0.060) (0.029) (0.028) (0.025) (0.023) (0.024)change in competitor effect -0.009 -0.174*** -0.169*** -0.019 0.048*** 0.053***

(0.035) (0.017) (0.017) (0.014) (0.014) (0.015)difference-in-difference 0.074 0.190*** 0.202*** 0.016 -0.112*** -0.130***

(0.068) (0.032) (0.031) (0.028) (0.026) (0.028)

# obs 14084 14084 14084 14084 14084 14084adj. R-square 0.6088 0.8503within R-sq 0.3902 0.4053 0.7481 0.7499between R-sq 0.4821 0.272 0.8451 0.7627overall R-sq 0.4783 0.2891 0.8467 0.7869

Figure 1 Marginal Effect and 95% Confidence

Interval

Figure 2 Marginal Effect at Different Levels of

Market Share


Table A1 Replacing Regional Carriers by Major Carriers

Regional Regional Ticketing Major Regional Regional Ticketing MajorCarrier Carrier Carrier Carrier Carrier Carrier Carrier CarrierCode Name or LCC Code Name or LCC

16 PSA Airlines Inc. US F8 Freedom Airlines d/b/a HP Expr DL DLCP Compass Airlines NW F8 HP HPCS Continental Micronesia CO NA North American Airlines AA AAEV Atlantic Southeast Airlines DL NA NA NAG7 GoJet Airlines, LLC d/b/a United Express UA OO SkyWest Airlines Inc. OO OOHQ Harmony Airways DL OO CO DLJ7 Valujet Airlines Inc. FL OO DL DLL4 Lynx Aviation d/b/a Frontier Airlines F9 OO NW DLMQ American Eagle Airlines Inc. AA OO UA UAOH Comair Inc. DL OO US UAOW Executive Airlines AA QX Horizon Air AS ASRD Ryan International Airlines FL QX F9 F9RU Expressjet Airlines Inc. CO RW Republic Airlines F9 F9TB USAir Shuttle US RW HP HPU2 UFS Inc. UA RW UA UAXE Expressjet Airlines Inc. CO RW US USXJ Mesaba Airlines NW RW YX YX9E Pinnacle Airlines Inc. CO NW S5 Shuttle America Corp. S5 S59E DL DL S5 CO DL9E NW NW S5 DL DL9E YX YX S5 NW DL9L Colgan Air CO CO S5 UA UA9L UA UA S5 US UA9L US US YV Mesa Airlines Inc. HP HPAX Trans States Airlines AA AA YV LHAX DL DL YV UA UAAX UA UA YV US USAX US US YV YV YVDH Independence Air CO DL ZW Air Wisconsin Airlines Corp UA UADH DH DH ZW US USDH DL DLDH NW DLDH UA UA

Table A2 Airports to MSAs

Airport MSA Airport MSA

ALB Albany-Schenectady-Troy, NY (MSA) MKE Milwaukee-Waukesha-West Allis, WI (MSA)ABQ Albuquerque, NM (MSA) MSP Minneapolis-St. Paul-Bloomington, MN-WI (MSA)ATL Atlanta-Sandy Springs-Marietta, GA (MSA) BNA Nashville-Davidson-Murfreesboro-Franklin, TN (MSA)AUS Austin-Round Rock, TX (MSA) MSY New Orleans-Metairie-Kenner, LA (MSA)BWI Baltimore-Towson, MD (MSA) EWR New York-Newark-Bridgeport, NY-NJ-CT-PA (CSA)BHM Birmingham-Hoover, AL (MSA) ISP New York-Newark-Bridgeport, NY-NJ-CT-PA (CSA)BOI Boise City-Nampa, ID (MSA) JFK New York-Newark-Bridgeport, NY-NJ-CT-PA (CSA)BOS Boston-Cambridge-Quincy, MA-NH (MSA) LGA New York-Newark-Bridgeport, NY-NJ-CT-PA (CSA)BUF Buffalo-Niagara Falls, NY (MSA) OKC Oklahoma City, OK (MSA)RSW Cape Coral-Fort Myers, FL (MSA) OMA Omaha-Council Bluffs, NE-IA (MSA)CLT Charlotte-Gastonia-Concord, NC-SC (MSA) MCO Orlando-Kissimmee, FL (MSA)MDW Chicago-Naperville-Joliet, IL-IN-WI (MSA) PHL Philadelphia-Camden-Wilmington, PA-NJ-DE-MD (MSA)ORD Chicago-Naperville-Joliet, IL-IN-WI (MSA) PHX Phoenix-Mesa-Scottsdale, AZ (MSA)CVG Cincinnati-Middletown, OH-KY-IN (MSA) PIT Pittsburgh, PA (MSA)CLE Cleveland, TN (MSA) PDX Portland-Vancouver-Beaverton, OR-WA (MSA)CMH Columbus, OH (MSA) PVD Providence-New Bedford-Fall River, RI-MA (MSA)DAL Dallas-Fort Worth-Arlington, TX (MSA) RDU Raleigh-Cary, NC (MSA)DFW Dallas-Fort Worth-Arlington, TX (MSA) RNO Reno-Sparks, NV (MSA)DEN Denver-Aurora, CO (MSA) RIC Richmond, VA (MSA)DTW Detroit-Warren-Livonia, MI (MSA) ROC Rochester, NY (MSA)ELP El Paso, TX (MSA) SMF Sacramento-Arden-Arcade-Roseville, CA (MSA)BDL Hartford-West Hartford-East Hartford, CT (MSA) SLC Salt Lake City, UT (MSA)HOU Houston-Sugar Land-Baytown, TX (MSA) SAT San Antonio, TX (MSA)IAH Houston-Sugar Land-Baytown, TX (MSA) SAN San Diego-Carlsbad-San Marcos, CA (MSA)IND Indianapolis-Carmel, IN (MSA) OAK San Jose-San Francisco-Oakland, CA (CSA)JAX Jacksonville, FL (MSA) SFO San Jose-San Francisco-Oakland, CA (CSA)MCI Kansas City, MO-KS (MSA) SJC San Jose-San Francisco-Oakland, CA (CSA)LAS Las Vegas-Paradise, NV (MSA) SEA Seattle-Tacoma-Bellevue, WA (MSA)BUR Los Angeles-Long Beach-Riverside, CA (CSA) GEG Spokane, WA (MSA)LAX Los Angeles-Long Beach-Riverside, CA (CSA) STL St. Louis, MO-IL (MSA)ONT Los Angeles-Long Beach-Riverside, CA (CSA) TPA Tampa-St. Petersburg-Clearwater, FL (MSA)SNA Los Angeles-Long Beach-Riverside, CA (CSA) TUS Tucson, AZ (MSA)LGB Los Angeles-Long Beach-Santa Ana, CA (MSA) TUL Tulsa, OK (MSA)SDF Louisville-Jefferson County, KY-IN (MSA) ORF Virginia Beach-Norfolk-Newport News, VA-NC (MSA)MHT Manchester-Nashua, NH (MSA) DCA Washington-Arlington-Alexandria, DC-VA-MD-WV (MSA)MEM Memphis, TN-MS-AR (MSA) IAD Washington-Arlington-Alexandria, DC-VA-MD-WV (MSA)FLL Miami-Fort Lauderdale-Pompano Beach, FL (MSA) PBI West Palm Beach-Boca Raton-Boynton Beach,MIA Miami-Fort Lauderdale-Pompano Beach, FL (MSA) FL Metropolitan Division

Partnering with Competitors – Effects of Alliances on ... · Partnering with Competitors –...

Documents

Transcript of Partnering with Competitors – Effects of Alliances on ... · Partnering with Competitors –...