Coordinated Capacity Reductions and Public Communication in the Airline Industry
CCoordinated Capacity Reductions and PublicCommunication in the Airline Industry ∗ Gaurab Aryal † , Federico Ciliberto ‡ , and Benjamin T. Leyden § February 8, 2021
Abstract
We investigate the allegation that legacy U.S. airlines communicated via earningscalls to coordinate with other legacy airlines in offering fewer seats on competitiveroutes. To this end, we first use text analytics to build a novel dataset on commu-nication among airlines about their capacity choices. Estimates from our preferredspecification show that the number of offered seats is 2% lower when all legacy airlinesin a market discuss the concept of “capacity discipline.” We verify that this reductionmaterializes only when legacy airlines communicate concurrently, and that it cannotbe explained by other possibilities, including that airlines are simply announcing toinvestors their unilateral plans to reduce capacity, and then following through on thoseannouncements.(JEL: D22, L13, L41, L93) ∗ This paper was previously circulated under the title “Public Communication and Collusion in the AirlineIndustry.” We thank Yu Awaya, David Byrne, Karim Chalak, Marco Cosconati, Kenneth G. Elzinga, LeslieMarx, Robert Porter, Mar Reguant, D. Daniel Sokol, and the seminar/conference participants at the DOJ,University of Florida, UVa, the 16th IIOC, 2018 BFI Media and Communication Conference, 2018 DC IODay, NBER IO SI 2018, EARIE 2018, 2018 FTC Microeconomics Conference, 2018 PSU-Cornell Conference,8th EIEF-UNIBO-IGIER Bocconi IO Workshop, and 2020 Next Generation of Antitrust Scholars Conferencefor their constructive feedback. We thank the Buckner W. Clay Dean of A&S and the VP for Research atUVa for financial support. Also, Aryal and Ciliberto acknowledge the Bankard Fund for Political Economyat the University of Virginia for support. We also thank Divya Menon for outstanding research assistance. † Department of Economics, University of Virginia, [email protected] . ‡ Department of Economics, University of Virginia; CEPR, London; DIW, Berlin, [email protected] . § Dyson School of Applied Economics and Management, Cornell University; CESifo [email protected] . a r X i v : . [ ec on . GN ] F e b Introduction
There are two legal paradigms in most OECD countries meant to promote market efficiency,but that are potentially at odds with one another. On the one hand, antitrust laws forbidfirms from communicating their strategic choices with each other to deter collusion. Onthe other hand, financial regulations promote open and transparent communication betweenpublicly traded firms and their investors. While these latter regulations are intended tolevel the playing field among investors, policymakers have raised concerns that they mayalso facilitate anticompetitive behaviors. For example, the OECD Competition Committeenotes that there are pro-competitive benefits from increased transparency, but “informationexchanges can ... offer firms points of coordination or focal points,” while also “allow[ing]firms to monitor adherence to the collusive arrangement” [OECD, 2011]. Thus, firms can betransparent about their future strategies in their public communications to investors—e.g.,by announcing their intention to rein capacity—which can foster coordination among firmsin offering fewer seats. In this paper, we contribute to this overarching research and policy issue by investigatingwhether the data are consistent with the hypothesis that top managers of legacy U.S. airlinesused their quarterly earnings calls to communicate with other legacy airlines to coordinatein reducing the number of seats offered. Specifically, we investigate whether legacy airlinesused keywords associated with the notion of “capacity discipline” in their earnings calls tocommunicate to their counterparts their willingness to reduce offered seats in markets wherethey compete head-to-head. The airline industry is an appropriate industry to investigate the relationship between Similar situations, where one set of laws is at odds with another, generating unanticipated consequences,often as antitrust violations, occur in many industries. For example, in the U.S. pharmaceutical industry,the tension between the FDA laws and patent law led to the Drug Price Competition and Patent TermRestoration Act (colloquially known as the Hatch-Waxman Act). This Act aims to reduce entry barriers forgeneric drugs, but it incentivized incumbent firms to “Pay-for-Delay” of generic drugs and stifle competition.For more, see Feldman and Frondorf [2017]. Other cases include Byrne and de Roos [2019] who documentthat gasoline retailers in Australia used a price transparency program called
Fuelwatch to initiate andsustain collusion. Furthermore, Bourveau, She and ˇZaldokas [2020] document that with the increase incartel enforcement, firms in the U.S. start sharing more detailed information in their financial disclosureabout their customers, contracts, and products, which may allow tacit coordination in product markets. An earnings call is a teleconference in which a publicly-traded company discusses its performance andfuture expectations with financial analysts and news reporters. Legacy carriers are Alaska Airlines (AS),American Airlines (AA), Continental Airlines (CO), Delta Airlines (DL), Northwest Airlines (NW), UnitedAirlines (UA) and US Airways (US), and the low-cost carriers (LCC) are AirTran Airways (FL), JetBlue(B6), Southwest (WN) and Spirit Airlines (NK). This idea that “capacity discipline” is used by airlines to signal their alleged intention to restrict supplyhas been applied in class-action lawsuits filed against a few airlines. Sharkey [2012] and Glusac [2017] providecoverage of this concept in the popular press. See Rosenfield, Carlton and Gertner [1997] and Kaplow [2013]for antitrust issues related to communication among competing firms. private and noisy monitoring, both of which make coordination difficultwithout communication. Demand is stochastic, not least because of exogenous local events,such as the weather, unforeseen events at the airport, and spillovers from other airports.Monitoring is private and noisy because, first, airlines do not instantaneously observe oth-ers’ actions; second, they use connecting passengers to manage their load factors; and third,they observe only each other’s list prices, not transaction prices.Recently, Awaya and Krishna [2016], Awaya and Krishna [2019] and Spector [2020] haveshown that firms may be able to use cheap talk–unverifiable and non-binding communication–to sustain collusion in environments with private and noisy monitoring, where collusion isotherwise unsustainable. In our context, airlines have access to public communication tech-nology, their quarterly earnings calls, through which they can simultaneously communicatewith other airlines. We build an original and novel dataset on the content of public communication fromearnings calls to measure communication and assess its relationship with capacity. TheSecurities and Exchange Commission (SEC) requires all publicly traded companies in theU.S. to file a quarterly report, which is accompanied by an earnings call—a public conferencecall where top executives discuss the report’s content with analysts and financial journalists.We collected transcripts of all such calls for 11 airlines from 2002:Q4 to 2016:Q4. We thenclassified each earnings call as either pertinent or not pertinent, depending on whether theexecutives on the call declared their intention of engaging in capacity-discipline or not. There is a precedence of accusation against the airlines for using communication technologies to coordi-nate. For example, in 1992, the U.S. DOJ alleged that airlines used the Airline Tariff Publishing Company’selectronic fare system to communicate and collude, see, for example, Borenstein [2004] and Miller [2010]. There is a vast literature on market conduct and the behavior of cartels; see Harrington [2006]; Mailathand Samuelson [2006]; Harrington and Skrzypacz [2011], and Marshall and Marx [2014]. Among others,Porter [1983]; Green and Porter [1984] study collusion under imperfect monitoring where all firms observethe same (noisy signal) price. In their setting, access to communication technology does not have any effectbecause the profits from public perfect equilibrium with or without communication are the same. Someexamples where communication helped collusion are Genesove and Mullin [2001], Wang [2008, 2009], Clarkand Houde [2014], and Byrne and de Roos [2019], among others. Airlines may have other avenues for coordination, e.g., via industry conferences and trade organizationevents [Awaya and Krishna, 2020] and common-ownership [Azar, Schmalz and Tecu, 2018]. However, quar-terly earnings call are ideal for our purpose because they occur at regular intervals, every publicly listedairline uses them, and we observe the conversation. Our decision to consider only communication throughearnings calls can be viewed as conservative because we cannot account for any amount of relevant com-munication outside this medium and underestimate the negative relationship between communication andcapacity. And lastly, we focus only on simultaneous messaging among (legacy) airlines and do not distinguishintra-quarter timing of airlines because determining if there is a “leader” among the airlines by following,say, Byrne and de Roos [2019], would require higher-frequency (e.g., daily) data on communication. Other papers that use “ text as data ” [Gentzkow, Kelly and Taddy, 2019], include Leyden [2019], whouses text descriptions of smartphone and tablet apps to define relevant markets, Gentzkow and Shapiro[2014], who use phrases from the
Congressional Record to measure the slant of news media, and Hoberg and
3e estimate the relationship between communication and carriers’ monthly, market-levelcapacity decisions using the Bureau of Transportation Statistics’s T-100 Domestic Segmentdataset, which contains domestic non-stop segment data reported by both U.S. and foreignair carriers. To that end, we regress the log of seats offered by an airline in a market in amonth on an indicator of whether all legacy carriers operating in that market discuss capacitydiscipline. Given that airlines’ capacity decisions depend on a wide variety of market-specificand overall economic conditions, we include covariates to control for such variation acrossmarkets and carriers over time.We find that when all legacy carriers operating in an airport-pair market communicateabout capacity discipline, the average number of seats offered in that market is 2.02% lower.To put this in perspective, we note that the average change in capacity among legacy carriersis 3.67%. So a 2.02% reduction in capacity associated with the phrase “capacity discipline”accounts for more than 50% of this average change, which is economically significant.Capacity reductions could benefit consumers if they reduce congestion at the airportswithout affecting ticket fares. However, we (i) do not find evidence to support the hypoth-esis that carriers reduce airport congestion, but (ii) find that communication is positivelyassociated with fares. So, even though we do not estimate the social value of communication[Myatt and Wallace, 2015], our estimates suggest that the carriers’ capacity reductions areeconomically significant and they most likely harm consumers.Nonetheless, we face two primary identification challenges in investigating the accusationthat legacy U.S. carriers are using their earnings calls to coordinate capacity reduction. First,there may be a more straightforward, alternative explanation for our findings. In particular,it might be that airline executives are communicating to investors their intention to reducecapacity, not because they want to coordinate, but because reducing capacity is the bestresponse to negative demand forecasts. In other words, our results may be evidence thatearnings calls are serving their ostensible purpose.We address this concern in three ways. First, we find that legacy carriers unilaterallydiscussing capacity discipline is not associated with them reducing capacities. Second, wefind that the capacity is not lower in monopoly markets when legacy carriers discuss capacitydiscipline. Finally, we find that legacy carriers do not decrease their capacity when all butone of the legacy carriers serving a market have discussed capacity discipline. Supposediscussions of capacity discipline were meant to inform investors about the carrier’s futureactions. In that case, we should see a reduction in all three of these cases.Second, an airline could be using earnings calls to truthfully share its payoff relevantprivate information with other airlines, which, when others do the same, could induce cor-
Philips [2016], who use the text descriptions of businesses included in financial filings to define markets. not have an incentive to share their payoffrelevant private information about demand with others unless they intend to coordinate onan action, e.g., capacity choice. Second, if this hypothesis is correct, then it implies that thelikelihood of us observing a reduction in capacity by an airline would increase with the num-ber of legacy airlines communicating, irrespective of the said airline’s private information. Ifairlines were only sharing their information, then an airline should be responsive to others’announcements. We show that, contrary to this information-sharing hypothesis, even whenall of a legacy carrier’s legacy competitors in a market communicate, if the carrier itself doesnot communicate, then it does not reduce its capacity. However, this result is consistentwith airlines using earnings calls to coordinate on their capacities.
In this section we introduce the legal cases that motivate our analysis, explain how we useNatural Language Processing (NLP) techniques to quantify communication among airlines,present our data on the airline industry, and show that airlines have flexible capacity at themarket level.
On July 1, 2015, the
Washington Post reported that the U.S. Department of Justice (DOJ)was investigating possible collusion to limit available seats and maintain higher fares in U.S.domestic airline markets by American, Delta, Southwest, and United (Continental) [Harwell,Halsey III and Moore, 2015]. It was also reported that the major carriers had received CivilInvestigative Demands (CID) from the DOJ requesting copies, dating back to January 2010,of all communications the airlines had had with each other, Wall Street analysts, and majorshareholders concerning their plans for seat capacity and any statements to restrict it. TheCID requests were subsequently confirmed by the airlines in their quarterly reports. Concurrently, several consumers filed lawsuits accusing American, Delta, Southwest, andUnited of fixing prices, which were later consolidated in a multi-district litigation. The case In Appendix G we consider whether our results vary before and after the January, 2010 threshold, andthe July, 2015 reporting of the DOJ investigation.
5s currently being tried in the U.S. District Court for the District of Columbia. Anothercase, filed on August 24, 2015, in the U.S. District Court of Minnesota against American,Delta, Southwest Airlines, and United/Continental, alleges that the companies conspired tofix, raise, and maintain the price of domestic air travel services in violation of Section 1 ofthe Sherman Antitrust Act. The lawsuits allege that the airline carriers collusively impose “capacity discipline” inthe form of limiting flights and seats despite increased demand and lower costs , and that thefour airlines implement and police the agreement through public signaling of future capacitydecisions. In particular, one of the consumers’ lawsuits reported several statements made bythe top managers of American, Delta, Southwest, United, and other airlines. The statementswere made during quarterly earnings calls and various conferences. These lawsuits provide the foundation to build a vocabulary from the earnings calls thatcan capture legacy airlines’ (alleged) intention to restrict their offered capacity. To thatend, we have to consider both the semantics (airlines’ intention to rein in capacity) and thesyntax (which keywords are used) of the earnings call reports. Next, we explain the stepswe take to measure communication.
All publicly traded companies in the U.S. are required to file a quarterly report with the SEC.These reports are typically accompanied by an earnings call, which is a publicly availableconference call between the firm’s top management and the analysts and reporters coveringthe firm. Earnings calls begin with statements from some or all of the corporate participants,followed by a question-and-answer session with the analysts on the call. Transcripts of calls This is the “Domestic Airline Travel Antitrust Litigation” case, numbered 1:15-mc-01404 in the USDistrict Court, DC. Case 0:15-cv-03358-PJS-TNL, filed 8/24/2015 in the US District Court, District of Minnesota. In Novem-ber 2015, this case was transferred to the District Court in DC. At the time of this writing, American Airlinesand Southwest have settled the class action lawsuits. The consumers’ lawsuits also stress the role of financial analysts who participate at the quarterly earningscall. See Azar, Schmalz and Tecu [2018] for a recent work on the role of institutional investors on marketconduct. We instructed our research assistant (RA) to find all instances where institutional investors werethe first to bring up capacity discipline. The RA found only three such instances. Therefore, we decided notto consider the role of institutional investors in our analysis. For example, during the US Airways 2012:Q1 earnings call, the CFO of US Airways Derrick Kerr said“.. mainline passenger revenue were $ - Q - Q QuarterAAASCODLNWUAUSB6FLNKWN C a rr i e r CollectedPrivately heldPre-mergerPost-mergerBankruptMissing
Notes. This figure shows the availability or non-availability of transcripts for 11 airlines. The x-axis denotesthe time year and quarter, and the y-axis denote the name of the airline. Each color/shade denotes thestatus of the transcript. are readily available, and we assume that carriers observe their competitors’ calls.We collected earnings call transcripts for 11 airlines, for all quarters from 2002:Q4 to2016:Q4 from LexisNexis (an online database service) and Seeking Alpha (an investmentnews website). Figure 1 indicates the availability of transcripts in our sample for each of the11 airlines. As the figure shows, transcripts are available for most of the quarters except under(i) Bankruptcy—five carriers entered bankruptcy at least once during the sample period; (ii)Mergers and acquisitions—airlines did not hold earnings calls in the interim period betweenthe announcement of a merger and the full operation of the merger; (iii) Private airlines—Spirit Airlines, which was privately held until May 2011, neither submitted reports norconducted earnings calls prior to its initial public offering; and (iv) Other reasons—in a fewinstances the transcripts were unavailable for an unknown reason. In all cases where a callis unavailable, we assume the carrier cannot communicate to its competitors. The key step of our empirical analysis is to codify the informational content in thesequarterly earnings calls into a dataset that can be used to see how capacity choices changeover time in response to communication among legacy carriers. Before delving into theconceptual challenges, we note two preliminary steps. Every statement made by the operatorof the call and the analysts is removed from the transcripts, as are common English “stopwords” such as “and” and “the.” Then we tokenize (convert a body of text into a set of wordsor phrases) and lemmatize (reduce words to their dictionary form) the text from the earnings Of course, the airlines may have other means to communicate, that we do not observe (e.g., see Foot-note 6). To the extent to which airlines use other, unobserved, means of communications when earnings callsare unavailable our estimate will be biased toward zero (or positive). { discipline , airline , executive , discuss , capacity , discipline } . This process allows us to abstract from the inflectional and derivationallyrelated forms of words to better focus on the substance/meaning of the transcripts.To do so, we use a combination of NLP techniques and manual review to identify alist of words or phrases that are potentially indicative of managers communicating theirintention to cooperate with others in restricting their capacity. Although in most casesmanagers specifically use the term “capacity discipline,” managers sometimes use other wordcombinations when discussing capacity discipline. This identification is a time-consumingprocess, and it is the focus of the remainder of this section. To codify the use of the phrase “capacity discipline” and other combinations of wordsthat carry an analogous meaning, we begin by coding “capacity discipline” with a categoricalvariable
Carrier-Capacity-Discipline j,t , which takes the value 1 if that phrase appearsin the earnings call transcript of carrier j in the year-quarter preceding the month t and 0otherwise.In many instances airline executives do not use the exact phrase “capacity discipline,”but the content of their statements is closely related to the notion of capacity discipline, asillustrated in the following text (emphasis added):“We intend to at least maintain our competitive position. And so, what’s neededhere, given fuel prices, is a proportionate reduction in capacity across all carriersin any given market . And as we said in the prepared remarks, we’re going toinitiate some reductions and we’re going to see what happens competitively. Andif we find ourselves going backwards then we will be very capable of reversingthose actions. So, this is a real fluid situation but clearly what has to happenacross the industry is more reductions from where we are given where fuel isrunning.” – Alaska Airlines, 2008:Q2.Our view is that this instance and other similar ones should be interpreted as conceptuallyanalogous to uses of the phrase “capacity discipline.” Yet in other cases it is arguable whetherthe content is conceptually analogous to the one of “capacity discipline,” even though thewording would suggest so. For example, consider the following cases:“We are taking a disciplined approach to matching our plan capacity levels withanticipated levels of demand” – American Airlines, 2017:Q3 In Appendix A, we also use NLP to identify words that can be used to evaluate conditional exogeneityin our setting. On one hand, the “anticipated levels of demand” dependon the competitors’ decisions, and thus one could interpret this statement as a signal tocompetitors to maintain capacity discipline. On the other hand, an airline should not putmore capacity than what is demanded because that implies higher costs and lower profits.We take a conservative approach and code all these instances as ones where the categor-ical variable
Carrier-Capacity-Discipline j,t is equal to 1. This approach is conservativebecause it assumes that the airlines are coordinating their strategic choices more often thantheir words would imply, and would work against finding a negative relation. In other words,we design our coding to err on the side of finding false negatives (failing to reject the nullhypothesis that communication is not correlated with a decrease in capacity), rather thanerring on the side of finding false positives. We take this approach because our analysisincludes variables that control for year, market, and year-quarter-carrier specific effects thatcontrol for many sources of unobserved heterogeneity that might explain a reduction of ca-pacity driven by a softening of demand. Therefore, our coding approach makes us less likelyto find evidence of coordination even when airlines are coordinating.In practice, to identify all the instances where the notion of capacity discipline was presentbut the phrase “capacity discipline” was not used, we used NLP to process all transcripts andflag those transcripts where the word “capacity” was used in conjunction with either the word“demand” or “GDP.” This filter identified 248 transcripts, which we read manually to classifyas either pertinent or not pertinent for capacity discipline. If the transcript was identified byall three of us as pertinent, then we set the variable
Carrier-Capacity-Discipline j,t = 1,and zero otherwise. Out of the 248 transcripts, 105 contained statements that we deemedpertinent. Table 1 presents the summary statistics of
Carrier-Capacity-Discipline j,t . We have320 earnings calls transcripts for the legacy carriers, and 40.9% include content associatedwith the notion of capacity discipline. We have fewer transcripts for LCCs, JetBlue and Airlines can change the capacity across markets in multiple ways. They can remove an aircraft from adomestic market and keep it in a hangar, or they can move it to serve an international route, or they canreassign that plane to another domestic market. The airlines can also change the “gauge” of an aircraft,i.e., increase or decrease the number of seats or change the ratio of business to coach seats. Additionally, inmarkets where carriers outsource some flights and/or routes to regional carriers, moving capacity should beeven easier. All of these options are discussed in conference calls. Besides the coding approach described above, we had a research assistant independently code all tran-scripts, and coded all transcripts only using the automated, NLP approach. We discuss these approaches,and the results of estimating our primary model with these datasets, in Appendix E.
Carrier Type
Legacy 0.409 0.492 320LCC 0.124 0.331 89Jet Blue 0.109 0.315 55Southwest 0.071 0.260 56
Total
Notes. Fraction of earnings calls where
Carrier-Capacity-Discipline is equal to one.
Southwest, and content associated with capacity discipline is much less frequent. Overall,we have 520 transcripts and
Carrier-Capacity-Discipline j,t = 1 in 29 .
2% of them. Table1 suggests that the LCCs, including Southwest (WN), are much less likely to talk publiclyabout capacity discipline. In view of this data feature, in our empirical exercise, we focusonly on communication by legacy carriers.
We use three datasets for the airline industry: the Bureau of Transportation Statistics’s(BTS) T-100 Domestic Segment for U.S. carriers, the BTS’s Airline On-Time Performancedatabase, and a selected sample from the OAG Market Intelligence-Schedules dataset. Weconsider the months between 2003:Q1 and 2016:Q3 (inclusive). The BTS’s T-100 DomesticSegment for U.S. carriers contains domestic non-stop segment (i.e., route) data reported byU.S. carriers, including the operating carrier, origin, destination, available capacity, and loadfactor. The BTS’s Airline On-Time Performance database reports flight times.In many instances, regional carriers, such as SkyWest or PSA, also operate on behalf ofthe ticketing carriers. The regional carriers might be subsidiaries fully owned by the nationalairlines, e.g., Piedmont, which is owned by American (and prior to that by U.S. Airways),or they might operate independently but contract with one or more national carrier(s), e.g.,SkyWest. To allocate capacity to the appropriate ticketing carriers, we merge our data withthe data from the OAG Market Intelligence, which contains information about the operatingand the ticketing carrier for each segment at the quarterly level. Using this merged dataset,we allocate the available capacity in each route in the U.S. to the ticketing carriers, which arethe carriers of interest. We consider only routes between airports located in the proximityof a Metropolitan Statistical Area in the U.S. We use the U.S. DOC’s 2012 data to identify Metropolitan Statistical Areas in the U.S. We also perform .4 Alignment of Earnings Call and Airline Data Our analysis in this paper is focused on understanding how communication in the airlines’quarterly earnings calls relates to their subsequent capacity decisions. An airline’s earningscall for a particular quarter takes place following the conclusion of that quarter, so we wishto associate the call for a given quarter with the monthly capacity data for the next quarter.That is, we wish to associate communication with the airline behavior that occurs followingthe call. Typically, these calls take place in the middle of the first month of the followingquarter. For example, calls for the first quarter (January to March) take place in the middleof April. This presents a challenge for merging the communication data with the monthlyairline data. Our approach for doing so is to associate the content of an earnings call withthe three full months following the call. For example, we use the content from a call for Q1 ,occurring in mid-April, to define
Carrier-Capacity-Discipline j,t for the months of May,June, and July. We say that legacy airlines are communicating with each other when all of the legacy airlinesserving a market with at least two legacy carriers discuss capacity discipline. Letting J Legacy m,t be the set of legacy carriers in market m at time t , we define a new variable, Capacity-Discipline m,t = (cid:8) Carrier-Capacity-Discipline j,t = 1 ∀ j ∈ J Legacy m,t (cid:9) , | J Legacy m,t | ≥ , | J Legacy m,t | < Capacity-Discipline m,t indicates whether all of the legacy carriers in m discussedcapacity discipline, conditional on two or more legacy carriers serving that market for thatmonth. In cases where fewer than two legacy carriers serve a market,
Capacity-Discipline m,t is set equal to 0. While
Carrier-Capacity-Discipline j,t varies by carrier and year-month, the empirical analysis where markets are defined by the origin and destination cities, rather than airports inAppendix D. An alternative approach would be to associate the Q1 call taking place in mid-April with the capacitydata for
April , May, and June. In Appendix Section F we present our primary results under this alternativeapproach. The results are similar to what we find under our preferred approach. In Awaya and Krishna [2016, 2019] firms communicate simultaneously, and it is crucial for the con-struction of their equilibrium. For example, Awaya and Krishna write, “The basic idea is that players canmonitor each other not only by what they ‘see’—the signals—but also by what they ‘hear’—the messagesthat are exchanged” [Awaya and Krishna, 2019, page 515]. In equilibrium, firms cross-check the messagesagainst each other, and under the asymmetric-correlation information structure, concurrent communicationensures that the signal is the most informative. - Q - Q QuarterAAASCODLNWUAUSB6FLNKWN C a rr i e r Collected (Talk)Collected (No Talk)Privately heldPre-mergerPost-mergerBankruptMissing
Notes. This figure shows the availability of transcripts and the prevalence of “Capacity Discipline” for11 airlines. The x-axis denotes years and quarters, and the y-axis denotes the name of the airline. Eachcolor/shade denotes the status of the transcript. Collected (Talk) means the transcript is available and theairline discussed capacity discipline, and Collected (No Talk) means the transcript is available but the airlinedid not discuss capacity discipline.
Table 2: Summary Statistics
Seats Cap. Discipline Talk Eligible Monopoly Market Missing ReportMean SD Median Mean SD Mean SD Mean SD Mean SD N
Carrier Type
Legacy 11,757.894 12,264.478 7,364.000 0.089 0.285 0.311 0.463 0.546 0.498 0.262 0.440 562,469LCC 11,255.056 10,467.260 8,220.000 0.032 0.177 0.106 0.307 0.471 0.499 0.098 0.297 279,522
Market Participants
Mixed Market 13,349.373 12,749.700 8,990.000 0.058 0.235 0.197 0.398 0.321 0.467 0.147 0.354 410,888Legacy Market 9,915.007 10,330.230 6,282.000 0.082 0.274 0.287 0.452 0.713 0.452 0.265 0.441 431,103
Total
Notes. Table of summary statistic for all key variables. Observations are at the carrier-market-month levelfor airport-pair markets. our treatment
Capacity-Discipline m,t varies by market and year-month. This is an impor-tant distinction for the empirical analysis, where the observations are at the market-carrier-year-month level.Figure 2 shows the occurrence of
Carrier-Capacity-Discipline j,t in our data. Eachrow corresponds to one airline and shows the periods for which the carrier discussed capacitydiscipline. There is variation in communication across both airlines and time, which isnecessary for the identification. Even though the reports do not vary within a quarter,the composition of airlines operating in markets—market structure—varies both within aquarter and across quarters, causing the dummy variable
Capacity-Discipline m,t to varyby month. 12able 2 provides a summary of this airline data. Legacy carriers offer, on average, 11,757.9seats in a month, while LCCs offer 11,255.1. Consistent with our focus on the communi-cation among only the legacy carriers, we find that legacy carriers are far more likely to bein a market where
Capacity-Discipline is equal to 1. We define the categorical variable
Talk-Eligible m,t ∈ { , } to be equal to 1 if thereare at least two legacy carriers in market m in period t and 0 otherwise. This variablecontrols for the possibility that markets where legacy carriers could engage in coordinatingcommunication may be fundamentally different from markets where such communicationsare not possible. Not including this control variable would confound the correlation betweentalking and seats. Table 2 shows that, on average, 24% of the observations in our samplehave the potential for coordinating communications. In a similar vein, markets served by asingle carrier could differ from non-monopoly markets. We account for this possibility byintroducing a categorical variable MonopolyMarket m,t , which is equal to 1 if in t , market m is served by only one firm and equal to 0 otherwise. We also see that, on average, 52%of observations are monopoly markets, and that legacy carriers are more likely to servemonopoly markets than LCCs.As discussed above, we take special note of markets where we were unable to collectan earnings call transcript. To account for that, we introduce a categorical variable
MissingReport m,t ∈ { , } is equal to 1 if at least one of the legacy carriers serving market m in period t did not hold an earnings call for the quarter prior to month t . Table 2 showsthat legacy carriers are more likely to operate in a market that is missing a report—a resultof the bankruptcy by many of the legacies. Following Jones [1996], in our regression, we use MissingReport and its interactions with other covariates to account for missing reports.
In this section, we specify and estimate a model to investigate whether the data are consistentwith the allegation that U.S. legacy carriers used their quarterly earnings calls to coordinatecapacity reductions. We begin with the premise that airlines have access to communication We use the seats variable in the T-100 dataset, which corresponds to the scheduled seats transported ina month between two airports. If we use seats weighted by the share of performed departures over scheduleddepartures, the main empirical findings do not change. Despite the lawsuit, we do not include Southwest (WN) when assessing communication because it isknown to have a different cost structure and business model than the legacy carriers, and, more importantly,the notion of capacity discipline appears only four times in the entire Southwest’s transcripts; see row WNin Fig. 2. See Section 2.2 for a discussion of when and why we were unable to collect a transcript. Transcripts aremissing, mostly for legacy carriers, largely due to their increased prevalence of bankruptcies and mergers.
We examine the relationship between communication among legacy airlines and the seatsthey offer between 2003:Q1 and 2016:Q3 (inclusive). We use panel data model to estimatethese relationships by estimating the following model using the within-group estimator:ln( seats j,m,t ) = β × Capacity-Discipline m,t + β × Talk-Eligible m,t + β × Monopoly m,t + β × MissingReport m,t + β × Talk-Eligible m,t × MissingReport m,t + β × Monopoly m,t × MissingReport m,t + µ j,m + µ j,yr,q + γ origin,t + γ destination,t + ε j,m,t , (1)where the dependent variable is the log of total seats made available by airline j in (airport-pair) market m in month t . Our variable of interest is Capacity-Discipline m,t , which isthe dummy variable introduced in Section 2.2 is equal to 1 if there are at least two legacycarriers in market m and month t , and they all communicated about capacity discipline intheir previous quarter’s earnings calls, and 0 otherwise.The idea behind capacity discipline is that airlines restricted seats even when there wasadequate demand, which can vary across both markets and time. To control for theseunseen factors, we include carrier-market fixed effects, µ j,m , and carrier-year-quarter fixedeffects, µ j,yr,q . These fixed effects allow airlines to provide different levels of capacity acrossdifferent markets and time. During our sample period, we observe several mergers (seeFig. 2). Since it is possible that a carrier’s relationship to a specific market could changein a meaningful way after a merger, we redefine the carrier as the merged entity in orderto allow greater flexibility in these fixed effects. For example, the fixed effect for AmericanAirlines serving the ITH-PHL market is allowed to differ before and after American mergeswith US Airways. Lastly, to control for time-dependent changes in demand we use origin-and destination-airport specific time trends, γ origin,t and γ destination,t . Implicitly, we are assuming that our panel data model satisfies the strict-exogeneity assumption.We performed a diagnostic test proposed by [Wooldridge, 2001, page 285] by including the lead market market structure DL reports communicating
Cap-Dis Report Monopoly Talk-Eligible parameters { DL } no n/a 0 1 1 0 β + β { DL } yes n/a 0 0 1 0 β { DL, UA } yes { DL, UA } β + β { DL, UA, US } no { US } or { UA } or { US, UA } β + β { DL, UA, US } yes { US, UA } β { DL, UA, US } yes { DL, UA, US } β + β { DL, UA, US, F9 } yes { DL, UA, US } β + β { DL, F9 } yes n/a 0 0 0 0 - Notes. An example to show identification from the perspective of Delta, i.e., when j = DL , and here UAand US are legacy carriers while F9 is an LCC. Next, we explain the identification strategy for Eq. (1). To highlight the key sources ofvariation in the data, we fix an airline—say, Delta (i.e., j = DL )—and consider differentpotential market structures and communication scenarios in Table 3. In markets m = 1 , Capacity-Discipline ,t = Capacity-Discipline ,t = 0. Then we can use variation in whether a report is available(for m = 2) or not (for m = 1) to identify β and β , as shown in the last column. Market m = 3 is served by both DL and UA and both discuss “capacity discipline” in the previousquarter, so Capacity-Discipline ,t = 1, which identifies β + β . The same identificationargument applies to identifying β + β in markets m = 6 , m = 7). In contrast, formarket m = 4, even when both US and UA discuss capacity discipline, we identify β + β because DL did not have a transcript.Lastly, we identify the fixed effects using the deviation from the mean. Therefore, the keysource of identification is the variation in Capacity-Discipline across markets and overtime (see Figure 2), which in turn depends on the variation in market structure and commu-nication. We also assume that conditional on all control variables,
Capacity-Discipline is uncorrelated with the error, and this conditional exogeneity of treatment is sufficient toidentify the relationship between
Capacity-Discipline and log-seats [Rosenbaum, 1984].We present the estimation of the semi-elasticity from Eq. (1) in column (1) of Table 4. Using our model, we find that when all of the legacy carriers in a talk-eligible market com-municate with each other about capacity discipline, there is a subsequent reduction in thenumber of seats offered by an average of 2.02%. This estimate is a weighted average of the
Capacity-Discipline m,t +1 as an additional regressor. This regressor’s estimated coefficient was +0.007and statistically significant at the 5 percent level, which suggests that the assumption of strict exogeneity isreasonable in our context. Throughout the paper, for a binary regressor, we present its estimated semi-elasticity. If the estimatedcoefficient of a dummy variable in a semilogarithmic regression is ˆ β , then the effect of the dummy variableon the outcome variable is 100 × (exp( ˆ β ) − (1) (2) (3) (4) (5) (6)Log Seats Log Seats Log Seats Log Seats Log Seats Log SeatsCapacity Discipline -0.0202 -0.0150(0.0060) (0.0049)Capacity Discipline 2 -0.0193 -0.0160(0.0064) (0.0052)Capacity Discipline 3 -0.0285 -0.0109(0.0112) (0.0103)Capacity Discipline 4 -0.0332 -0.0018(0.0541) (0.0342)Legacy Market x Capacity Discipline -0.0169 -0.0197(0.0074) (0.0063)Mixed Market x Capacity Discipline (Legacy) -0.0195 0.0010(0.0115) (0.0085)Mixed Market x Capacity Discipline (LCC) -0.0341 -0.0221(0.0155) (0.0107)Carrier-Market FE’s (cid:88) (cid:88) (cid:88) Carrier-Market-Structure FE’s (cid:88) (cid:88) (cid:88)
R-squared 0.088 0.083 0.088 0.083 0.088 0.084N 841,991 841,991 841,991 841,991 841,991 841,991
Notes. We report semi-elasticities (see Footnote 24), with standard errors clustered at the bi-directionalmarket level in parentheses. Other control variables included in all regressions, but whose coefficients arenot reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicatorwith
Talk-Eligible and
Monopoly . In columns 2 and 3, these coefficients are allowed to vary based on thenumber of legacy carriers in the market (either 0 or 1, 2, 3, 4, or 5 legacy carriers). In columns 5 and 6,these coefficients are allowed to vary across legacy and mixed markets, and within mixed markets for legacycarriers and LCCs. Additionally, all regressions include origin- and destination-airport annual time trends,and carrier-year-quarter fixed effects. Columns 1, 3, and 5 include carrier-market fixed effects, and columns2, 4, and 6 include carrier-market-structure fixed effects. parameter estimates across markets, time, and types of carriers and should be interpreted asa level decrease in capacities. The standard errors are clustered at the bi-directional marketlevel. To determine the estimate’s economic significance, we can compare it to the averagechange in capacities across talk-eligible markets when there is no communication. In thesecomparison markets, the average percentage change in capacities is 3.67%. So, wheneverlegacy airlines communicate, their capacities drop by more than 50% of the average changein capacities, a significant reduction.While we attempt to capture some of the differences in market structures that permitcommunication (via the
Talk-Eligible variable), this may not adequately capture the man-ner in which competitive behavior may respond to market structure, either in terms of thenumber or type of carriers, or the specific set of carriers serving a market. To address thisconcern, we re-estimate our primary specification Eq. (1), but control for specific market Following de Chaisemartin and D’Haultfœuille [2020], we estimated the weights for each group, and only0.07% of those weights were negative, suggesting that negative weights do not drive our estimate. structure fixed effects.To best understand the carrier-market-structure fixed effects, consider an example of theIthaca (ITH) to Philadelphia (PHL) market. Suppose we observe this market for four periods,and during this time the market structures are { AA, DL } , { AA } , { AA, UA } , and { AA, DL } .The carrier-market fixed effects for a given carrier would be constant across all periods inwhich they compete in the ITH-PHL market, but carrier-market-structure fixed effects allowAmerican (AA) to behave differently when in a duopoly with Delta (DL) compared to when itis competing in a duopoly with United (UA). In Table 4-column (2) we present the estimationresults from this alternative specification. We find that communication is associated with a1.50% reduction in offered capacity. Next, we consider whether the relationship between communication and capacity varieswith the number of communicating airlines. Let
Capacity-Discipline- k m,t ∈ { , } be 1 ifmarket m in period t is talk eligible, is served by exactly k legacy carriers, and all k of themuse capacity discipline. Then we estimate Eq. (1) after replacing Capacity-Discipline m,t with three (additively separable) indicators { Capacity-Discipline- k m,t : k = 2 , , } . Theestimation results using the carrier-market fixed effects and the carrier-market-structure fixedeffects are in columns (3) and (4) of Table 4, respectively. In column (3) of Table 4, we find that with the carrier-market fixed effects, the associationbetween communication and capacity reductions are increasing in the number of legacycarriers serving the market. In particular, we find that communication is associated with areduction in capacity of 1 . .
85% and 3 .
32% in markets with two, three and four legacycarriers, respectively. Although, because there are few markets with four legacy carriers, thiscoefficient is imprecisely estimated. With the carrier-market-structure fixed effects, however,we find that communication is associated with a reduction in capacity by 1 .
60% in marketswith two legacy carriers. For the markets with three or four legacy carriers, the coefficientsare imprecisely estimated with no effect.Lastly, we explore how the estimate change between markets with only legacy carriersand mixed markets with both legacy and LCCs. We present summary statistics for thesetwo types of markets in Table 2. We present the results from this exercise using the carrier-market fixed effects and the carrier-market-structure fixed effects in columns (5) and (6)of Table 4, respectively. With the carrier-market fixed effects, we find that communicationabout capacity discipline is associated with a 1.69% decrease in the number of seats offered. Under the carrier-market-structure fixed effects,
Talk-Eligible and
Monopoly are redundant and aretherefore excluded from the regressions. While we do observe some markets with five legacy carriers,
Capacity-Discipline m,t is always zero inthese markets, and so we do not include an additional variable for this case.
17n mixed-markets, we find that communication is associated with a 1.95% decrease in legacyseats and a 3.41% decrease in LCCs seats.In summary, we find that capacity is lower when all legacy carriers serving a talk-eligiblemarket discuss capacity discipline, a finding which is consistent with the allegation that U.S.legacy carriers used their quarterly earnings calls to coordinate capacity reductions. Onaverage, we find that capacity is between 1.50% and 2.02% lower when this communicationoccurs, though we find this varies with the number of legacy carriers in a market, and thepresence of LCCs. In the analysis that follows, we use the specification outlined in Eq. (1), which employscarrier-market fixed-effects, as our primary specification because it takes advantage of bothimportant sources of variation in
Capacity-Discipline , namely, that
Capacity-Discipline can turn on or off as a result of a legacy carrier beginning or ending communication, or whena legacy carrier enters or leaves a market. For completeness, we provide correspondingestimates in Appendix I for everything that follows using carrier-market-structure fixed ef-fects.
We now turn our attention to other, consumer-welfare relevant outcomes. While thesereductions in capacities (relative to the demand) probably reduce welfare, such reductionscould also reduce congestion at airports—because airlines might also coordinate the timing offlights—and therefore benefit consumers. Additionally, these capacity reductions might notultimately affect prices, thus limiting the impact on consumer welfare. Although estimatingthe welfare effect of communication is beyond the scope of this paper, we can determine (i)if conditional on reducing capacity, airlines change their departure times and reduce airportcongestion; and (ii) if communication is associated with higher average fares. As we shownext, we find no evidence to support the hypothesis that congestion has decreased, andwe find preliminary evidence that fares may have risen, both of which show that capacitydiscipline likely hurt consumers. In Appendix C we explore how the relationship between capacity and communication varies with thesize of a market and the amount of business travel in a market. An additional concern with including carrier-market-structure fixed effects is that market structure maycorrelate with the unobservable in Eq. (1). As noted in Section 4.3, and with more details in Appendix B,using a control function approach we find that our primary results are robust to this concern. For instance, Armantier and Richard [2003] consider the effect of information exchanges between UAand AA out of O’Hare airport and find that while airlines benefit, it only moderately hurts consumers. Theyconclude, “Hence, a marketing alliance between AA and UA, with the sole objective of exchanging costinformation, would be advantageous to airlines without significantly hurting consumers.” .2.1 Airport Congestion First, we examine if, conditional on reducing capacity, legacy airlines change their departuretimes and reduce congestion at the airport. To measure congestion in an airport, we usethe following measure proposed by Borenstein and Netz [1999].On a route with n daily departures departing d , . . . , d n minutes after the midnight, theaverage time difference between two flights is given by Average-Time-Difference := 2 n − n (cid:88) i =1 n (cid:88) j>i (cid:113) min {| d i − d j | , − | d i − d j |} . To make this measure comparable across markets with different n , we normalize it by themaximum time difference if the flights were equally spaced throughout a day, such that valuesclose to 1 corresponds to the least crowded flights. Although we use the normalized measure,for notational ease, we continue to refer it as Average-Time-Difference . To calculate
Average-Time-Difference we use the Bureau of Transportation Statistics’s Airline On-Time Performance database, which records flight times.We estimate a fixed effects model, where the dependent variable is
Average-Time-Difference and the regressors are the same as in Eq. (1), plus two additional variables: the totallog-seats offered in the market and an interaction term between the total log-seats and
Capacity-Discipline . Congestion is at the market-level, so we replace the carrier-marketfixed effect with market fixed effects. Our primary variable of interest is the interaction termbecause it estimates the relationship between log-seats and the changes in the average timedifference with communication. If, conditional on reducing offered seats, airlines were in-creasing the average time between their flights and reducing congestion, then this interactionterm’s coefficient would be positive.We present the estimation results in column (1) of Table 5, under the heading “Conges-tion.” As we can see, the coefficient for the interaction term is − . Next, we consider estimating the relationship between communication and prices. If, when-ever airlines communicate, they lower their offered capacities, then, unless capacities never In additional, but unreported analyses, we find suggestive evidence that both overall market capacity andthe number of flights are lower when all legacy airlines communicate about capacity discipline in talk-eligiblemarkets. These results are available from the authors upon request. (1) (2) (3)Congestion Price PriceCapacity Discipline x Log Market Seats -0.0043(0.0026)Capacity Discipline 0.0501 0.0061(0.0290) (0.0033)Capacity Discipline (Legacy) 0.0049(0.0035)Capacity Discipline (LCC) 0.0187(0.0058)Log Market Seats 0.0743(0.0031)Market-Structure FE’s 0.077 0.133 0.138Carrier-Market FE’s 463,951 649,204 649,204
Notes. Congestion refers to the average difference between two flights’ departure times within an airport,and P refers to the log of average fares. We report semi-elasticities (see Footnote 24), with standard errorsclustered at the bi-directional market level in parentheses. Other control variables included in all regressions,but whose coefficients are not reported are are
Talk-Eligible , Monopoly , MissingReport , interactions ofthe
MissingReport indicator with
Talk-Eligible and
Monopoly , origin- and destination- airport annualtime trends, and carrier-market fixed effects. Columns 1 and 2 include year-quarter fixed effects, and columns3 and 4 include carrier-year-quarter fixed effects. bind, it is reasonable to expect that prices would rise due to communication.Even though it might seem straightforward to estimate this relationship, for example, byestimating Eq. (1) after replacing the log of offered seats as the dependent variable with thelog of the prices as the dependent variable, this empirical strategy is infeasible. Airlines selltickets for origin to final -destination pairs, but the offered capacities and communication areat the direct-segment level. Thus, to understand the relationship between communicationand prices, we must first construct a new dataset of prices and communication.Connecting tickets involve flights that go through different nonstop segments, possiblywith different market structures in each segment. Thus, while the prices are at the origin-destination level, capacity plans and our communication measure are at the nonstop segmentslevel. So, we have to aggregate capacity and communication from the segment level to theorigin-destination level. For example, consider flights traveling from A to C via a connect-ing airport, B. In particular, assume that in segment A–B, two airlines are talking, but insegment B–C, there are three airlines but the third airline is not talking. Our aggregationmust account for how to define
Capacity-Discipline in these and similar situations. Fur-thermore, airlines may use multiple routes for the same market (i.e., use multiple airportsto connect a given origin and destination), adding additional complexity to our problem.20ext, we define how we aggregate communication in segments A–B and B–C to determinecommunication in the origin-destination pair A to C. First, we follow Borenstein [1989] andconstruct a dataset of prices, but instead of aggregating at the market level (e.g., marketA to C), we aggregate them at the market-route level. For example, consider a ticketingcarrier, say UA, serving A to C via two routes, AB-BC and AB-BD-DC. In this case, wetreat these two routes separately, even though they are have the same origin and destination.At the end of this aggregation, we have average prices and the total number of passengerstransported by each airline for each market-route. We then use the number of passengerstransported to determine weighted average prices and
Capacity-Discipline , weighted bythe number of passengers in those combinations defined at the carrier-market level.In particular, to determine
Capacity-Discipline m,t at the route level, we calculate
Capacity-Discipline for every nonstop segment. Then we merge the price data with thesenew communication data and restrict the sample in the price data to those markets we ob-serve in our primary analysis. Note that the number of carriers serving a market in our pricedataset weakly exceeds the number of carriers serving that market in our primary analysisbecause they include carriers that serve the origin and destination pair via a connection.We can then aggregate the dummy variable
Capacity-Discipline that we definedpreviously from the segment level to the origin-destination level. In particular, if thevariable
Capacity-Discipline = 1 in all nonstop segments of a route, then we define
Capacity-Discipline = 1 for that route. For the market missing report variable, we takethe opposite approach: if it is 1 for at least one segment, then it is 1 for the route. Finally, weconstruct a
Capacity-Discipline variable for each market by taking the passenger weightedaverage of
Capacity-Discipline for each route through which a carrier serves that market.To better understand this approach, consider the following stylized example. Suppose acarrier serves a market-quarter { m, t } via three different routes, and Capacity-Discipline variable is 1 , , and 1 for these three routes. Furthermore, if the carrier sends 25% of its pas-sengers along route 1, 25% along route 2, and 50% along route 3, then Capacity-Discipline variable for the carrier in { m, t } is equal to 1 × .
25 + 0 × .
25 + 1 × . .
75. We use thesame approach to calculate the
Talk-Eligible , Monopoly , and
Missing-Report variables.Using these variables in a panel data model like Eq. (1) we estimate the relationshipbetween
Capacity-Discipline and the log of (average) route-level prices. The results arein columns (3) and (4) in Table 5. The estimates suggest that the average price increasedby 0.59%, and that this increment is mostly due to LCCs, whose average prices increased by1.80%. For instance, as ITH-CHO is not served nonstop by any airline, it does not appear in our primaryanalysis. We drop this market from this analysis, even though there are connecting flights between them.
21n summary, we fail to find evidence to support the view that the capacity reductionseither improved or did not affect consumer welfare, as congestion did not decrease, and wefind preliminary evidence that prices may have risen. In Section 3, we found that whenever all of the legacy carriers in a market discuss capacitydiscipline, capacity is on average 2% lower in the next quarter, a finding which is consistentwith the accusation that legacy carriers used their earnings calls to coordinate with othercarriers to reduce capacity. In this section, we perform a series of robustness exercises toaddress other possible explanations for this finding.
We have shown that we observe lower capacity when all legacy carriers in a market discusscapacity discipline. Of course, it could be that airlines are not coordinating but are simplyannouncing their unilateral intentions to reduce capacity in response to demand forecastsor for other reasons specific to themselves. That is, the airlines may be using the quarterlyearnings call for its ostensible purpose: to inform investors about the state of their businesses.If this is the case, then the number of seats offered by an airline would also fall whenthe airline is communicating, but its competitors are not. That is not what we find. We donot find evidence that a carrier reduces capacity when it discusses capacity discipline, butits legacy competitors do not. Additionally, carriers do not reduce capacity in monopolymarkets, where we would also expect to find capacity reductions following communication.Finally, we find no evidence of capacity reductions when all but one of the legacy carriersserving a market discuss capacity discipline.To investigate whether airlines decrease capacity when they are the only one discussing Our analysis treats capacity choices as strategic substitutes. It is reasonable to consider the possibilitythat if consumers care about departures, and if this preference is strong enough, that may soften competitionto the effect that the capacity choices become strategic complements and not strategic substitutes. However,we do not believe this to be the case because airlines’ departures and capacity choices are interlinked. Thus,even after setting aside airlines’ communication decisions, we would have to consider three choices (departuretimes, capacity choices, and airfares) together in our model. There are several ways to model departures.One of them is the Salop/Vickrey circular city model, as in Gupta et al. [2004], which is also consistentwith Borenstein and Netz [1999]. We can then embed this model within a Kreps and Scheinkman [1983]framework, which results in a game played by airlines in three stages. First, they choose departure times,then the capacities and the prices. However, conditional on the circle locations (i.e., the departure times),capacities are still strategic substitutes. seats j,m,t ) = β × Only-j-Talks j,m,t + β × Talk-Eligible m,t + β × Monopoly m,t + β × MissingReport m,t + β (cid:62) × MissingReport-Interactions m,t + µ j,m + µ j,yr,q + γ origin,t + γ destination,t + ε j,m,t , (2)where our variable of interest is Only-j-Talks j,m,t defined as
Only-j-talks j,m,t = (cid:8) Carrier-Capacity-Discipline j,t = 1 ∧ Carrier-Capacity-Discipline k,t = 0 | J Legacy m,t | ≥ ∀ k (cid:54) = j ∈ J Legacy m,t (cid:111) | J Legacy m,t | < . That is,
Only-j-Talks j,m,t indicates whether carrier j is the only legacy carrier in market m that discussed capacity discipline, conditional on there being at least two legacy carriers.The parameter β shows the extent to which a legacy carrier that discusses capacity disciplinewhen none of its market-level competitors discussed capacity discipline changes capacity. Ifdiscussion of capacity discipline is meant to inform investors about future strategic behavior, β should be negative and, likely, close to -2.02%. We present the estimation results fromEq. (2) in column (1) of Table 6. As we can see from the estimates in the first row of column(1), there is no evidence of a decline in the capacity associated with the unilateral discussionof capacity discipline. We find the opposite: the number of offered seats is 1.72% higherwhen airlines communicate unilaterally.A second approach to addressing the concern mentioned above is to look at capacity de-cisions in monopoly markets. If carriers discuss capacity discipline to inform investors abouttheir plans to reduce capacity, presumably independent of what other airlines are doing, weshould expect to see reductions in monopoly markets following those discussions. To esti-mate the role of “monopoly capacity discipline” we estimate our primary model Eq. (1), butusing the treatment Monopoly-Capacity-Discipline m,t , which is equal to 1 when a carrierin a monopoly market discussed capacity discipline and 0 otherwise. We estimate this modelusing both our full sample and a sample that consists of only monopoly markets, and presentthe results in columns (2) and (3) of Table 6, respectively. In the full sample we find theopposite—capacities are higher after a monopoly airline discusses capacity discipline—butfor the monopoly markets sample we find no evidence of an effect.23able 6: Financial Transparency and Information Sharing (1) (2) (3) (4) (5)Log Seats Log Seats Log Seats Log Seats Log SeatsOnly j Talks 0.0172(0.0066)Monopoly Capacity Discipline 0.0260 0.0027(0.0074) (0.0050)Capacity Discipline N − j ” -0.0010(0.0074)R-squared 0.088 0.088 0.061 0.087 0.087N 841,991 841,991 438,980 841,991 841,991 Notes. Notes. We report semi-elasticities (see Footnote 24), with standard errors clustered at the bi-directional market level in parentheses. Other control variables included in all regressions, but whose coeffi-cients are not reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicator with
Talk-Eligible and
Monopoly , origin- and destination-airport annual time trends, carrier-year-quarter fixed effects, and carrier-market fixed effects. Column 3 omits the
Talk-Eligible and
Monopoly variables.
Finally, we consider whether carriers reduce capacity in cases where all but one of thelegacy carriers serving the market discuss capacity discipline. To do so, we estimate Eq. (1)with the treatment variable
Capacity-Discipline-N-1 m,t defined as
Capacity-Discipline-N-1 m,t = (cid:80) j ∈ J Legacy m,t (cid:8) Carrier-Capacity-Discipline j,t (cid:9) = | J Legacy m,t | − , | J Legacy m,t | ≥ , | J Legacy m,t | < , (3)which is equal to 1 when all but one of the legacy carriers in a Talk-Eligible marketdiscuss capacity discipline, and 0 otherwise. We present this estimation results in column(4) of Table 6. We find no evidence of a relationship between communication and capacitywhen all but one of the legacy carriers serving a market discuss capacity discipline. In lightof these exercises—looking at markets where one carrier speaks but its competitors do not,looking at capacity decisions in monopoly markets, and looking at markets where all butone legacy carrier speak—we conclude that discussion of capacity discipline is not simply abona fide announcement of future, unilateral intentions.24 .2 Information Sharing
So far, we have shown that when all legacy carriers in a market discuss capacity discipline,capacity is subsequently lower, and, if any one of the legacy carriers is not discussing capacitydiscipline while the others are, their number of offered seats does not change (Table 6, column(4)). While these two results are consistent with coordination, they could also be consistentwith the idea that (for some historical reason) airlines use correlated strategies. That is,when they announce their intention to engage in capacity reduction during the earnings call,they share their private information about the aggregate airline demand.In fact, our previous finding that the level of capacity reduction is increasing in the num-ber of legacy carriers serving the market (Table 4, columns (3) and (4)) provides suggestivesupport for such an alternative hypothesis: when more airlines are communicating, the preci-sion of the aggregate signal gets better, which in turn induces stronger correlation in capacitychoices. Thus, this alternative “information sharing” model interprets the communicationas being payoff relevant, unlike in Awaya and Krishna [2016] wherein capacity discipline ischeap talk, but it does not require firms to coordinate on any action.To better understand this alternative theory, consider the following. Suppose that withprobability θ ∈ (0 ,
1) there is a negative demand shock. Each airline receives a private signal θ i of the actual θ and publicly announces its θ i during its earnings call, and airlines thenbase their decisions on all the announced θ ’s. So, airlines reduce capacity when all signalsare unfavorable compared to when only one firm received a negative signal because of thecorrelation in their strategies induced by information sharing. This alternative model assumes that airlines always have an incentive to share theirinformation about aggregate demand. Clarke [1983], Gal-Or [1985], and Li [1985], however,show that firms do not have an incentive to share their private information about marketdemand with others unless, as Clarke [1983] shows, they can use that information to collude. To verify the validity of the alternative model, we test its implication that absent its signalabout low demand airline j would still reduce capacity in the presence of a strong, aggregatesignal from others. To that end, we estimate the effect of “everyone except airline j talking”on j ’s capacity choice next quarter. Let Capacity-Discipline-(not − j ) m,t ∈ { , } be adummy variable equal to 1 if the market m in period t is talk eligible and if every legacy carrierserving m except airline j discusses capacity discipline, and 0 otherwise. Then we estimateEq. (1) after replacing Capacity-Discipline m,t with
Capacity-Discipline-(not − j ) m,t and present the results in column (5) of Table 6. We find that even when everyone else This alternative model makes a stronger assumption—airlines cannot misrepresent their information.Under our cheap-talk interpretation, however, it is moot whether or not a message is truthful. For more on the role of information-sharing on collusion see [Vives, 2008; Sugaya and Wolitzky, 2018]. j is communicating, it does not affect j ’s capacity. Although this “no-effect” resultis inconsistent with the information-sharing model, it is consistent with the allegation thatlegacy carriers communicate to coordinate capacity reductions. In addition to the work described above, we conduct three additional robustness exercises.For brevity, we present these results in Appendices A, B, and D. First, in Appendix A,we provide evidence that is consistent with the view that our analysis does not violate theconditional exogeneity assumption regarding how we define our communication variable.To this end, we follow White and Chalak [2010]. Heuristically, suppose we observe thatairlines use some other words as frequently as “capacity discipline” in their earnings callsand suppose these words are contextually similar to “capacity discipline.” For our modelto be consistent with conditional exogeneity, controlling for communication about “capacitydiscipline,” capacities should not depend on the use of any of these new words.To implement this diagnostic test, we first identify words in the corpus of earnings calltranscripts that are contextually similar to capacity discipline and are equally likely to occurwhen carriers discuss it. For that, we employ the word2vec model, a neural network modelcommonly used in computational linguistics; see Mikolov et al. [2013]. The word2vec modelidentifies a set of six words that satisfy these criteria. For each of these six words, wedefine a dummy variable equal to one if all legacy airlines in a talk-eligible market use thesaid word and include this new measure of communication as an additional control in ourmain specification (Eq. (1)). We fail to find evidence of a negative association betweenthe six words and capacity choices. Additionally, in all six cases, the coefficient on ourprimary variable of interest
Capacity-Discipline does not meaningfully change. Theseresults provide additional assurance that our communication variable is consistent with theconditional exogeneity assumption.Second, in Appendix B, we consider the possibility that market structure can be en-dogenous because a factor that affects capacity decisions can also affect airlines’ decisionsto serve a market. If a market structure is endogenous, then
Capacity-Discipline will beendogenous as well. To address this, we use a control function approach, where the excludedvariables are functions of the geographical distances between a market’s endpoints and eachcarrier’s closest hub, which we define as an airport with “sufficiently” many connections.The identification assumption is that an airport’s distance to the airline’s nearest hub is aproxy for entry cost and is therefore correlated with the market structure, but is less likely tobe directly correlated with capacity decisions [Ciliberto and Tamer, 2009]. In other words,26his approach leverages a timing assumption, namely, that unobservables that affect anairline’s network are not contemporaneously correlated with the unobservables that affectthe carrier’s capacity decisions. Under this approach, we find that legacy carriers reducetheir seats by 2.02% on average when they communicate, similar to our preliminary result.Additionally, the results in Appendix B help to validate our specifications that use carrier-market-structure fixed effects, as the carrier-market-structure fixed effects would violate thestrict conditional exogeneity assumption if the market structure is endogenous, which resultsin biased estimates.Finally, throughout this paper, we have defined markets as origin and destination airportpairs, an approach commonly used in the literature. A second approach would be to definemarkets as a directional pair of cities, as discussed in detail in Brueckner, Lee and Singer[2014]. In Appendix D, we define markets using the city-pair approach and re-estimate ourprimary specification. Under this approach to defining markets, we fail to find evidence ofa relationship between communication and capacity choices. However, this appears to bedue to the two three-airport cities in our sample: Washington D.C. and New York City, andexcluding them produce results consistent with our primary findings.
In this paper, we investigate whether legacy airlines use public communication to sustaincooperation in offering fewer seats in a market. We maintain that airlines communicated,with each other, whenever all legacy carriers serving a market talked about capacity disciplinein their earnings calls. Using natural language processing methods, we converted quarterlyearnings call transcripts into numeric data to measure communication among legacy carriers.Our estimate is consistent with the allegation that legacy carriers who communicate about“capacity discipline” offer 2% fewer seats, on average, across markets and time.Even though we do not estimate the social value of communication, our estimates sug-gest that the carriers’ capacity reductions are economically significant and most likely harmconsumers because (i) we fail to find evidence that airport congestion has reduced; and (ii)simultaneous communication is positively associated with average fares. While we find thatthese estimates are consistent with anticompetitive behavior, we are aware that communi-cation is not exogenous, and so we have to exercise caution in interpreting these estimationresults as proof of collusion.We address various threats to the identification of our primary model. First, while ourestimated reduction in capacity after carriers discuss capacity discipline is consistent withairlines coordinating, we do not find it consistent with an alternative hypothesis that earnings27alls are serving their intended purpose of making markets more transparent. We also verifythat the way we have defined communication in our model is consistent with conditionalexogeneity, and finally, we use a control function approach to confirm that our estimates arenot affected by endogenous market structure. Thus, we cannot rule out the possibility thatpublic communication allows legacy airlines to coordinate.Our finding is relevant for the current policy debate about the social value of informa-tion and the correct response to increasing information about firms in social media andincreasing market concentration across industries. We have provided evidence that in theairline industry, the SEC’s transparency regulations are at odds with antitrust laws—a factthat policymakers should be cognizant of. While the value of public quarterly earnings callsremains debatable, economists and policymakers view the public disclosure of informationthrough these calls as beneficial for investors. At the same time, the competitive effects ofthis increased transparency are theoretically ambiguous and under-studied. We contributeto this literature and hope that this paper will spur further empirical research on this topic.While, in some cases, communication helps in equilibrium selection, its broader implica-tions for welfare are unknown. For instance, to determine if a public communication channelis anticompetitive, one must understand how the coordination mechanism depends on thenature of communication. While we find results consistent with the alleged claim that thecommunication channel enables anticompetitive behavior in the airline industry, there arestill many compelling research questions about how these results came to be and the ex-tent to which these results generalize to other industries and methods of communicationthat remain unanswered. Answers to these questions will help design laws related to publiccommunication and antitrust policy.In our context of airlines, these questions require the estimation of a flexible oligopolymodel, where firms can choose capacity and prices, whether to collude or compete and wherestrategic behavior can be influenced by public communication. As we mentioned earlier, oneapproach could consist of developing and estimating a model that incorporates both pricesand capacity decisions in the airline industry, in the vein of Kreps and Scheinkman [1983], butwith differentiated products, and extend it to allow collusion [Brock and Scheinkman, 1985;Benoit and Krishna, 1987; Davidson and Deneckere, 1990] with communication. An evenmore ambitious step would be to allow consumers to care about departures `a la Salop/Vickreycircular city model of Gupta et al. [2004]. While these models have been studied in isolation,their interactions pose challenges that have not yet been explored and we leave that forfuture research. 28 eferences
Armantier, Olivier, and Oliver Richard.
RAND Journal of Economics , 34(3): 461–477.
Awaya, Yu, and Vijay Krishna.
AmericanEconomic Review , 106(2): 285–315.
Awaya, Yu, and Vijay Krishna.
Theoretical Economics , 14(2): 513–553.
Awaya, Yu, and Vijay Krishna.
RAND Journalof Economics , 51(2): 421–446.
Azar, Jos´e, Martin C. Schmalz, and Isabel Tecu.
Journal of Finance , 73(4): 1513–1565.
Benoit, Jean-Pierre, and Vijay Krishna.
Review of Economic Studies , 54(1): 23–35.
Borenstein, Severin.
RAND Journal of Economics , 20(3): 344–365.
Borenstein, Severin.
The Antitrust Revolution: Economics, Competition andPolicy . . 4th ed., , ed. John Kwoka Jr. and Lawrence White, Chapter 9, 233–51. OxfordUniversity Press.
Borenstein, Severin, and Janet Netz.
International Journal ofIndustrial Organization , 17: 611–640.
Bourveau, Thomas, Guoman She, and Alminas ˇZaldokas.
Journal of Accounting Research , 58(2): 295–332.
Brock, William A., and Jos´e A. Scheinkman.
Review of Economic Studies , 52(3): 371–382.
Brueckner, Jan K., Darin Lee, and Ethan Singer.
Review of IndustrialOrganization , 44(1): 1–25. 29 yrne, David, and Nicolas de Roos.
American Economic Review , 109(2): 591–619.
Ciliberto, Federico, and Elie Tamer.
Econometrica , 77(6): 1791–1828.
Clarke, Richard N.
TheBell Journal of Economics , 14(2): 383–394.
Clark, Robert, and Jean-Fran¸cois Houde.
The Journal of IndustrialEconomics , 62(2): 191–228.
Davidson, Carl, and Raymond Deneckere.
International Economic Review , 31(3): 521–541. de Chaisemartin, Cl´ement, and Xavier D’Haultfœuille.
American Economic Review ,110(9): 2964–2996.
Feldman, Robin, and Evan Frondorf.
Drug Wars: How Big Pharma Raises Pricesand Keeps Generics off the Market.
New York City:Cambridge University Press.
Gal-Or, Esther.
Econometrica , 53(2): 329–343.
Genesove, D., and W. P. Mullin.
American Economic Review , 91(3): 379–398.
Gentzkow, Matthew, and Jessie Shapiro.
Econometrica , 62(1): 35–71.
Gentzkow, Matthew, Bryan T. Kelly, and Matt Taddy.
Journalof Economic Literature , 57(3): 535–574.
Glusac, Elaine.
The New YorkTimes . Green, Edward J., and Robert H. Porter.
Econometrica , 52(1): 87–100.
Gupta, Barnali, Fu-Chuan Lai, Debashis Pal, Jyotirmoy Sarkar, and Chia-MingYu.
International Journal of Industrial Orga-nization , 22(6): 759–782. 30 alvorsen, Robert, and Raymond Palmquist.
American Economic Review , 70(3): 474–475.
Harrington, Joseph E.
How Do Cartles Operate? Foundations and Trends in Mi-croeconomics , Now Publishers Inc.
Harrington, Joseph E., and Andrzej Skrzypacz.
American Economic Re-view , 101(6): 1–25.
Harwell, Drew, Ashley Halsey III, and Thad Moore.
The Washington Post . Hoberg, Gerard, and G. Philips.
Journal of Political Economy , 124(5): 1423–1465.
Jones, Michael P.
Journal of the American Statistical Association ,91(433): 222–230.
Kaplow, Louis.
Competition Policy and Price Fixing.
Princeton University Press.
Kreps, David M., and Jos´e A. Scheinkman.
The Bell Journal of Economics ,14(2): 326–337.
Leyden, Benjamin T.
Working Paper . Li, Lode.
RAND Journal of Eco-nomics , 16(4): 521–536.
Mailath, George J., and Larry Samuelson.
Repeated Games and Reputations:Long-Run Relationships.
Oxford University Press.
Marshall, Robert C., and Leslie M. Marx.
The Economics of Collusion: Cartelsand Bidding Rings.
MIT Press.
Mikolov, T., K. Chen, G. Corrado, and J. Dean.
ArXiv e-prints . Miller, Amalia R.
TheJournal of Law and Economics , 53(3): 569–586.31 yatt, David P., and Chris Wallace.
Journal of Economic Theory , 158(Part B): 466–506.
OECD. . Porter, Robert H.
The Bell Journal of Economics , 14(2): 301–314.
Rosenfield, Andrew M., Dennis W. Carlton, and Robert H. Gertner.
George Mason Law Review , 5(3): 423–440.
Sharkey, Joe.
The New York Times . Spector, David.
Mimeo . Sugaya, Takuo, and Alexander Wolitzky.
Jour-nal of Political Economy , 126(6): 2569–2607.
Vives, Xavier.
The New Palgrave Dictionaryof Economics: Volume 1 – 8 , , ed. Steven N. Durlauf and Lawrence E. Blume, 3053–3055.London:Palgrave Macmillan UK.
Wang, Zhongmin.
Review of Industrial Organization , 32(1): 35–52.
Wang, Zhongmin.
Journal of Political Economy ,117(6): 987–1030.
White, Halbert, and Karim Chalak.
Economic Letters , 109(2): 88–90.
Wooldridge, Jeffrey.
Econnometric Analysis of Cross Section and Panel Data.
MITPress. 32 ppendix A Conditional Exogeneity
Although we employ a rich set of fixed effects and other covariates as control variables, it isstill desirable to explore the possibility that our finding is driven by a missing communication-related variable positively correlated with capacity discipline and negatively correlated withoffered seats. To this end, we propose to run a diagnostic test `a la White and Chalak [2010].We can explain this approach using an example. Suppose we define an additional com-munication variable equal to 1 whenever all legacy airlines use the word “stable” , and zerootherwise. Furthermore, suppose that the occurrence of “stable” is positively correlated with,and occurs as frequently as, the discussion of “capacity discipline.”
Then, under this diagno-sis, we verify that adding this new dummy variable that captures the discussion of “stable” asan additional regressor in Eq. (1) neither affects the estimated relationship between capacitydiscipline and offered seats nor is it negatively correlated with offered seats.Although intuitive, to implement this diagnostic test, we have first to find all relevanttokens (e.g., “stable” ). Given the large amount of text data we have, it is a nontrivial taskto find such tokens objectively. To do so, we use methods from computational linguisticsto search our entire text and identify tokens or keywords that (i) are “close” in terms ofcontext to the discussion of capacity discipline, and (ii) occur approximately as frequentlyas “capacity discipline.” Then, for each token, we define a dummy variable Z m,t that is equalto 1 only if all legacy carriers in talk-eligible market m use it in period t and include it asan additional regressor in Eq. (1). Then, we test if the estimated coefficient for each Z m,t isstatistically negative or not, and verify whether the coefficient of capacity discipline changeswith the introduction of Z m,t .To construct such a set of tokens, we identify three tokens that are essential to theconcept of capacity discipline: “capacity discipline,” “demand,” and “gdp.” Then, we usethe word2vec model from computational linguistics [Mikolov et al., 2013] to determine othertokens that are close to these three tokens, using a distance metric that we define shortlybelow. word2vec allows us to be objective in determining the tokens.Broadly, the word2vec model is a neural network that maps each unique token we observein the earnings call transcripts to an N -dimensional vector space (in our analysis, N = 300)in such a way as to preserve the contextual relationships between the tokens. The vectorrepresentation of each token is such that contextually similar tokens are located “close” toeach other, and tokens that are dissimilar are located “far” from each other. This senseof “closeness” reflects the likelihood that the given tokens appear near each other in the The word2vec model was developed at Google in 2013 [Mikolov et al., 2013] to analyze text data. Foran intuitive and accessible explanation, see Goldberg and Levy [2014]. We use the gensim implementationof the word2vec model [ ˇReh˚uˇrek and Sojka, 2010]. capacity disciplineholiday =13590180 270
88 5
Notes. A schematic illustration of a hypothetical word2vec model. Tokens are mapped to a vector space,such that the cosine of the angle between two tokens represents the level of “similarity” between those tokens.In the case above, “holiday” is seen to be very dissimilar to “capacity discipline.” earnings call transcripts. Thus, if “discipline” and “stable” are close, then the discussion ofone term in an earnings call is likely given a discussion of the other. We directly train the word2vec model using our transcript data, so the derived relationships between words arespecific to the context of airlines’ earnings calls, as opposed to a more general context. Forexample, if airline executives use the word “discipline” in a contextually different mannerthan used in more general conversation or writing, our model will account for that.To measure the similarity of two tokens in the word2vec vector space, we use a commonlyused metric called the cosine similarity metric.
This metric is equal to the cosine of the anglebetween the vector representation of the two tokens Singhal [2001], such that for any twonormalized vectors associated with two tokens, k , and (cid:96) , this measure of similarity is d cos ( (cid:96), k ) = k T (cid:96) || k || · || (cid:96) || , where || · || is the L norm. When two vectors are the same, cosine similarity is 1, and whenthey are independent (i.e., perpendicular to each other), it is 0. To understand our use of cosine similarity, consider Fig. A.1, which displays a hypotheti-cal example of training the word2vec model in a 2-dimensional space. The word2vec modelmaps all of the tokens in our vocabulary to this space. For example, the token “capacitydiscipline” is represented by the vector (5 , Note that the cosine metric is a measure of orientation and not magnitude. This metric is appropriatein our cases, as we are interested in comparing the contextual meaning of the words, not in comparing thefrequency of the words. (1) (2) (3) (4) (5) (6)slow weakness domestically internationally stable paceZ Token -0.0025 0.0016 0.0187 0.0030 0.0027 0.0052(0.0063) (0.0061) (0.0068) (0.0054) (0.0099) (0.0076)Capacity Discipline -0.0200 -0.0202 -0.0187 -0.0202 -0.0204 -0.0207(0.0059) (0.0060) (0.0059) (0.0059) (0.0060) (0.0059)N 841,991 841,991 841,991 841,991 841,991 841,991
Notes. Estimation results from including new tokens an additional regressors in Eq. (1). The table showsthe coefficient estimates for each token, and for
Capacity-Discipline . We report semi-elasticities (seeFootnote 24), with standard errors clustered at the bi-directional market level in parentheses. Other controlvariables included in all regressions, but whose coefficients are not reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicator with
Talk-Eligible and
Monopoly , origin-and destination-airport annual time trends, carrier-year-quarter fixed effects, and carrier-market fixed effects. vector ( − , θ = 135 ◦ , so d cos ( holiday , capacity discipline ) = − . k ∈ { capacity discipline, demand, gdp } , we define the set: L k ( d, d ) = (cid:8) (cid:96) ∈ L : d ≤ d cos ( (cid:96), k ) ≤ d (cid:9) , where L is the set of all tokens. To satisfy the second criterion, we restrict the token to besuch that at least 50% of the time it appears in the same report as these three keywords.In Table A.1, we present all the tokens that satisfy the above two criteria. For each token,we define Z m,t as we did for Capacity-Discipline m,t and use it as an additional regressorin Eq. (1). The estimated coefficients for the tokens are in the first row, with the estimatedcoefficient for
Capacity-Discipline m,t in the second row. As we can see, five out of six atokens have no relationship with log seats, and even then the coefficient of “domestically” ispositive which shows that, if anything, our results understate the true relationship betweenthe discussion of capacity discipline and capacity. What is also reassuring is that for allthe tokens, the estimates for
Capacity-Discipline are stable, with estimates close to ourprimary result of − . Appendix B Control Function Approach
In this section, we present results from using a control function approach to estimate ourmodel. 35ur treatment,
Capacity-Discipline m,t , is the product of
Talk-Eligible m,t and whetherall of the legacy carriers in m discussed capacity discipline in their most recent earnings calls.By construction, Talk-Eligible m,t is a function of the market structure (the set of airlineswho serve market m in month t ). An airline’s decision to serve m , among other factors,will depend on the cost of serving it, which is unobserved and might not be captured by thefixed effects. So it is possible that Talk-Eligible m,t is endogenous, which in turn means
Capacity-Discipline m,t would be endogenous too. And because
Talk-Eligible m,t , andhence
Capacity-Discipline m,t , are negatively correlated with the cost of serving m in t ,our estimator in Eq. (1) might exaggerate the negative effect of communication on capacity.Finding an IV for our regression is not a simple task because decisions across marketsare interconnected in a network industry. For example, Hendricks, Piccione and Tan [1999]consider a one-shot two-stage model where two carriers incur fixed costs and simultaneouslychoose their networks and compete (Bertrand or Cournot) for passengers. Still, we believethat the endogeneity of market structure could bias our results, and so we propose an instru-mental variable that exploits a plausible timing assumption. Leveraging a timing assumptionin this way is common in empirical studies of market competition; see, for example, Olleyand Pakes [1996] and Eizenberg [2014].In particular, we propose to use a measure of the distance between a market’s endpointsand the carriers’ closest hubs, henceforth, “hub-distances,” as an instrumental variable formarket structure. The distance of a market’s endpoints to a carrier’s closest hub is a proxyfor the fixed cost that a carrier has to face to serve that market [Ciliberto and Tamer, 2009].This is the direct effect of the distance on an airline’s decision to serve a market. Distancesto the hubs also indirectly affect the market structure through competition: An airline’sprobability of serving a market should increase with its competitors’ distances.Conditional on including the distance between the origin and destination airport, whichis captured by the carrier-market fixed effects in our model, hub-distance should not affectconsumer demand and the carriers’ variable costs. We are aware that, for a given networkstructure of the industry, the distance to hubs might correlate with capacity decisions, butwe believe that if the relationship exists, it is weaker than the one between the distancesfrom the hubs and the entry decision. Indeed, the variable hub-distance is not included inthe standard structural models of demand and supply for the airline industry, see Berry andJia [2010]. Moreover, the fact that we measure the impact of communication on market-levelcapacity choices and not on the aggregate capacities further suggests that hub-distance isuncorrelated with the capacity choice.Finally, as mentioned above, our instruments rely on a plausible timing assumption,namely that the unobservables that affect the development of an airline network are not36 ontemporaneously correlated with the unobservables that affect prices and capacity deci-sions. This assumption relies on a crucial institutional feature of the airline industry, wherebynetwork service is fundamentally dependent on the ability of airlines to enplane and deplanetravelers at airports, and they can only do this if they have access to gates. As discussedin Ciliberto and Williams [2014], “a substantial majority of gates are leased on an exclusiveor preferential basis, and for many years.” In addition, Ciliberto and Williams note that“it is difficult to adjust access to airport facilities in response to unexpected changes in de-mand and costs.” This institutional feature of the airline industry is particularly true for anairline’s hubs. Therefore, the development of an airline network, with its determination ofits hub-and-spoke structure, is considerably slower than an airline’s ability to enter and exitmarkets, and to change capacities and prices. Despite this, we are aware that there mightbe persistent components of the unobservables, but we maintain that those are captured bythe market-carrier and the carrier-year-quarter fixed effects.To measure the role of airline networks as determinants of market structure we proceedas follows. First, for each airline, we compute the air-distance of an airport to the airline’s“hubs” (which are defined based on connectedness of the time-varying network of marketsserved by an airline, defined shortly below). Data on the distances between airports arefrom the data set
Aviation Support Tables: Master Coordinate , available from the NationalTransportation Library. Then for each carrier j , market m , and month t in our sample, wecalculate that carrier’s hub-distance D j,m,t as the sum of the distance from the origin airportto the carrier’s nearest hub, and the distance from the destination airport to the carrier’snearest hub. We use these hub-distances as instrumental variables for Talk-Eligible and,in turn, for
Capacity-Discipline . In Fig. B.1 we display the histograms for the within carrier-market variances of these dis-tances, measured in thousands of miles. Fig. B.1a displays the entire sample while Fig. B.1brestricts the sample to only those with positive variance in distances. Both these figures andtable show that there is substantial variation in distances. We also present the summarystatistics of these distances by carriers in Table B.1.Using the calculated hub-distances, we employ a control function approach to estimatethe effect of communication on capacity [Imbens and Wooldridge, 2007]. In the first-stage, The concept of connectedness is from the theoretical literature on networks. See Section B.1 for additionaldetails on the calculation of the set of hubs for each airline. We thank Mar Reguant for suggesting this approach, that when an endogenous variable is a interactionthen we can use one of the two variables as an instrument for the product, and if that variable is alsoendogenous then an instrument for that variable will still be a valid instrument for the product. It is similarto the approach used in Fabra and Reguant [2014]. Our approach also controls for an (unlikely) event thatlegacy carriers discussing capacity correlates with the unobserved cost of serving a market, as long as thatevent is uncorrelated with the instrumental variable. (a) All Values (b) Positive Values
Notes. Observations constructed by calculating the standard deviation of the hub-distance for each carrier-market. Hub-distance is measured in thousands of miles. Panel (a) includes all carrier-market observations,and panel (b) only includes carrier-market observations where the value is non-zero.
Table B.1: Summary Statistic of Hub-distances by Carriers
Mean Sd Median NAirline
AA 1.274 0.629 1.192 613,673AS 3.550 1.113 3.801 613,673CO 1.315 0.770 1.150 613,673DL 1.059 0.506 0.987 613,673LCC 0.956 0.608 0.837 613,673NW 1.258 0.711 1.054 613,673UA 1.089 0.543 1.038 613,673US 1.215 0.745 1.068 613,673
Total
Notes. Each row displays the mean, standard deviation, median and number of observations of air-distancesto closest hubs for a carrier. Distances are measured in thousands of miles. LCC is the average of distancesfor all LCCs. we estimate
Talk-Eligible m,t = (cid:88) j ∈ J CF σ j D j,m,t + α × Monopoly m,t + α × MissingReport m,t + α × Monopoly m,t × MissingReport m,t + µ m + µ yr,q + γ origin,t + γ destination,t + r m,t , (B.1)where J CF is the set of legacy carriers, and an aggregated LCC carrier. We aggregate thelow-cost carriers by setting D LCC,m,t to the shortest hub-distance for all of the LCCs formarket m in month t . We present the results of estimating Eq. (B.1) in Table B.2. Having38able B.2: First-stage Regression for Control Function Approach: Communication andAvailable Seats (1)Talk EligibleAA Distance -0.0195(0.0014)CO Distance 0.0088(0.0009)DL Distance -0.0281(0.0016)LCC Distance -0.0196(0.0007)NW Distance -0.0051(0.0012)UA Distance -0.0011(0.0013)US Distance -0.0158(0.0006)F-statistic (instruments) 265.9674N 613,673 Notes. Observations are at the market-month level. Bootstrapped standard errors clustered at the bi-directional market level are reported in parentheses. Other control variables included in the regression,but whose coefficients are not reported are
Monopoly , MissingReport , the interaction interaction of the
MissingReport and
Monopoly variables, origin- and destination-airport annual time trends, year-quarterfixed effects, and market fixed effects. estimated Eq. (B.1) we recover the residuals ˆ r m,t . Then, in the second step, we re-estimatethe parameters in Eq. (1) with ˆ r m,t as an additional covariate. We present the second-stageresults in column (1) of Table B.3, and replicate our primary results in column (2) to facilitatecomparison. We can see that when legacy carriers communicate, they reduce their capacityby 2 . B.1 Determination of Airline Hubs
In this section, we explain how we determine hubs for each airline, and provide evidence ofvariations in our instruments. To identify hubs over time, we follow Ciliberto, Cook andWilliams [2019]. They show that using the shortest path between two airports to determinethe betweenness centrality measure identifies the hub airports well.To illustrate this measure of centrality, consider Figure B.2, which displays a networkof airports served by an airline. Betweenness centrality for CHO measures the number of39able B.3: Control Function Approach: Communication and Available Seats (1) (2)Log Seats Log SeatsCapacity-Discipline -0.0201 -0.0202(0.0060) (0.0060)Talk Eligible 0.1478 -0.0415(0.2331) (0.0150)Residual -0.1904(0.2349)N 841,353 841,991
Notes. We report semi-elasticities (see Footnote 24), with bootstrapped standard errors clustered at thebi-directional market level in parentheses. Column (1) reports the results of the control function approach,and column (2) replicates our primary estimates of Eq. (1) to facilitate comparison. Other control vari-ables included in all regressions, but whose coefficients are not reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicator with
Talk-Eligible and
Monopoly , origin-and destination-airport annual time trends, carrier-year-quarter fixed effects, and carrier-market fixed effects.
Figure B.2: Network for an Airline
DFW CLTJFKORDLAX CHOPHXSFO
Notes. A schematic representation of airports-network served by an airline. times CHO is the shortest connection between two other airports. In this example, CHO isnever in the shortest path between any two airports, so the betweenness centrality for CHOis zero. Similarly, the betweenness centrality for PHX is also zero. However, DFW will havehigher betweenness centrality because it is in a stop of multiple airports, like PHX and SFC.Similarly, the betweenness centrality for CLT and LAX will be high.Formally, the betweenness measure for an airport k , for airline j is B jk := (cid:88) (cid:96) (cid:54) = (cid:96) (cid:48) ,k (cid:54)∈{ (cid:96),(cid:96) (cid:48) } N j − N j − P jk ( (cid:96), (cid:96) (cid:48) ) P j ( (cid:96), (cid:96) (cid:48) ) , N j is the number of airports served by airline j , P jk ( (cid:96), (cid:96) (cid:48) ) is the number of shortestpaths between airports (cid:96) and (cid:96) (cid:48) with a stop at k , and P j ( (cid:96), (cid:96) (cid:48) ) is the total number of shortestpaths between (cid:96) and (cid:96) (cid:48) . If there is only one shortest path between (cid:96) and (cid:96) (cid:48) , then the ratio is1, and if there are multiple paths, then this measure gives equal weight to each path. Themeasure is rescaled by dividing through by the number of pairs of nodes not including k ,so that B jk ∈ [0 , j andevery period t we choose the airports with the betweenness centrality that is at least 0 . j ’s “hubs.” By this definition, the hubs in Figure B.2 are { DF W, CLT, LAX } . Appendix C Market Heterogeneity
In this section, we consider the role of market sizes and the composition of passengers indetermining the relationship between communication and capacity choices.
C.1 The Role of Market Size
First, we explore how airlines’ reductions in capacity differ by market size. Carriers’ abilityto coordinate on capacity can vary by market, depending on the ability of legacy airlines tomonitor each other and their markets’ contestability. We follow Berry, Carnall and Spiller[2006] and define market size as the geometric mean of the Core-based statistical area popu-lation of the end-point cities. The annual population data are from the U.S. Census Bureau.We define markets with a population larger than the 75 th percentile of the market populationdistribution as large, markets with a population in the range of (25 th , th ] percentiles of thepopulation as medium, and those at or below the 25 th percentile as small markets. Table C.1 shows that the average number of seats a carrier offers, the likelihood ofthe treatment
Capacity-Discipline = 1, and the likelihood of
Talk Eligible = 1 areincreasing with the size of a market. Perhaps unsurprisingly, the likelihood that a market isa monopoly market decreases with the size of the market.To assess if the intensity of coordinated capacity reduction varies with market size, we re-estimate Eq. (1), interacting
Capacity-Discipline m,t with indicators of whether a marketis small, medium, or large. We present these estimation results in column (1) of Table C.2. When classifying markets as small, medium, or large, we use the average market population over oursample period so that a market’s size classification does not change across time. The 25 th percentile cutoffis 1.27 million people, and the 75 th percentile cutoff is 3.25 million people. Although not reported, we also allow the impacts of
Talk-Eligible , Monopoly , and
MissingReport tovary with market size.
Seats Cap. Discipline Talk Eligible Monopoly Market Missing ReportMean SD Median Mean SD Mean SD Mean SD Mean SD N
Market Size
Small 5,301.915 5,469.491 3,844.000 0.005 0.069 0.027 0.163 0.842 0.365 0.192 0.394 115,293Medium 9,890.301 9,373.000 7,145.000 0.046 0.209 0.156 0.362 0.588 0.492 0.191 0.393 420,714Large 16,298.972 14,270.025 11,880.000 0.129 0.336 0.445 0.497 0.309 0.462 0.236 0.425 305,984
Business Travel
Low Business 11,249.436 11,404.137 7,535.000 0.067 0.250 0.216 0.411 0.445 0.497 0.202 0.402 175,445Medium Business 12,005.773 12,151.691 7,949.000 0.089 0.285 0.296 0.456 0.461 0.499 0.232 0.422 295,324High Business 11,605.024 11,495.510 7,883.000 0.061 0.239 0.216 0.412 0.599 0.490 0.218 0.413 149,993
Total
Notes: Observations are at the carrier-market-month level.
Table C.2: Communication and Available Seats: The Role of Market Size and BusinessTravel (1) (2)Log Seats Log SeatsCapacity Discipline x Small Population -0.0146(0.0276)Capacity Discipline x Medium Population -0.0134(0.0108)Capacity Discipline x Large Population -0.0251(0.0069)Capacity Discipline x Low Business -0.0087(0.0098)Capacity Discipline x Medium Business -0.0266(0.0078)Capacity Discipline x High Business -0.0172(0.0130)R-squared 0.088 0.086N 841,991 620,762
Notes. We report semi-elasticities (see Footnote 24), with standard errors clustered at the bi-directionalmarket level in parentheses. Other control variables included in all regressions, but whose coefficients arenot reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicatorwith
Talk-Eligible and
Monopoly . These coefficients are allowed to vary based on the market size orbusiness travel classifiers. Additionally, all regressions include origin- and destination-airport annual timetrends, carrier-year-quarter fixed effects, and carrier-market fixed effects.
We find that communication among legacy carriers is associated with, on average, a 1 . .
40% reduction in seats supplied in smaller and medium markets, respectively, butthat the coefficients are imprecisely estimated and, as a result, we cannot reject the nullhypothesis that in these two types of markets, communication and the number of seatsoffered are uncorrelated. However, for the large markets, we find that communication amonglegacy carriers is associated with a 2 .
42% reduction in seats supplied.42 .2 The Role of Business Travelers
Next, we investigate whether the composition of the market demand in business and leisuretravelers is associated with the degree to which carriers respond to communication. Businesstravelers tend to have a higher willingness to pay for a ticket and have less elastic demandfor air travel than leisure travelers. So, all else equal, markets with a relatively high numberof business travelers should have higher mark-ups and be more lucrative for coordination.We follow Borenstein [2010] and Ciliberto and Williams [2014] and use a business indexconstructed using the 1995 American Travel Survey (ATS). The ATS was conducted by theBureau of Transportation Statistics (BTS) to obtain information about the long-distancetravel of people living in the U.S., and it collected quarterly information related to thecharacteristics of persons, households, and trips of 100 miles or more for approximately80,000 American households. We use the survey to compute an index that measures thefraction of passengers traveling for business out of an airport.We define a market’s business travel index as the computed travel index for its originairport. In classifying markets based on their business travel level, we follow the sameapproach as in our market size classifications. Low business markets are those with an indexvalue at or below the 25 th percentile, medium business markets have an index value in the(25 th , th ] percentiles, and high business markets are those with an index above the 75 th percentile. The average number of seats offered in a market is relatively constant across ourbusiness travel classifications, but coordinated communication is more common in low andmedium business markets than in high business markets. Having constructed our businessclassifications, we estimate a model interacting Capacity-Discipline m,t with indicators forthe three levels of business travel.We present the results from this regression column (2) of Table C.2. The last threerows present the estimated relationship between communication and capacity choices inlow-business, medium-business, and high-business markets, respectively. As we can see,communication is associated with a 0.09%, 2.66%, and 1.72% decrease in the number ofseats offered, respectively, although the estimates for low- and high-business markets areimprecisely estimated.
Appendix D An Alternative Approach to Defining Mar-kets: City Pairs
In the main paper, we have followed Borenstein [1989]; Kim and Singal [1993]; Borenstein andRose [1994]; Gerardi and Shapiro [2009]; Ciliberto and Tamer [2009]; Berry and Jia [2010];43iliberto and Williams [2010]; and Ciliberto and Williams [2014], and defined a market bythe origin and destination airport pairs. An alternative argument maintains that marketsare to be defined by the origin and destination cities , rather than airports. This alternativemarket definition has been followed by, among others, Berry [1990, 1992]; Brueckner andSpiller [1994]; Evans and Kessides [1994]; and Bamberger, Carlton and Neumann [2004].Table D.3: Summary Statistics for City-Pair Markets
Seats Cap. Discipline Talk Eligible Monopoly Market Missing ReportMean SD Median Mean SD Mean SD Mean SD Mean SD N
Carrier Type
Legacy 12,751.265 15,306.754 7,440.000 0.108 0.311 0.395 0.489 0.445 0.497 0.279 0.448 518,609LCC 11,694.473 11,826.904 8,220.000 0.076 0.265 0.270 0.444 0.295 0.456 0.165 0.372 269,019
Market Participants
Mixed Market 14,919.775 16,107.449 9,381.000 0.115 0.319 0.415 0.493 0.166 0.372 0.224 0.417 478,473Legacy Market 8,475.512 9,413.635 5,428.000 0.070 0.256 0.254 0.435 0.746 0.435 0.265 0.442 309,155
Total
Notes. Table of summary statistic for all key variables. Observations are at the carrier-market-month levelfor city-pair markets.
The city-pair market aggregates possibly more than one airport-pair market. For il-lustration, consider two flights flying out of Piedmont Triad International Airport (GSO),located in Greensboro, NC, with one flying to O’Hare International Airport (ORD) and theother flying to Midway International Airport (MDW), both located in Chicago, IL. Underthe airport-pair market definition, these flights operate in separate markets—the first is inthe GSO-ORD market, and the second is in the GSO-MDW market. Under the city-pairsmarket definition, these flights operate in the same Greensboro to Chicago market. InTable D.3 we present the city-pair analogue of Table 2.How to define airline markets is of interest in antitrust policies. While the airport-pairapproach is often used in academic research on the airline industry, antitrust practitionersuse the city-pair approach. Using the city-pair approach leads to larger markets, which, forantitrust purposes, provides a more robust basis for government intervention if there is anyevidence of anticompetitive effects.However, with the city-pair definition, we should expect the effect of communication onthe capacity to change. As an example, consider Table D.4, which lists four flights fromGreensboro, NC to Chicago, IL. Under the airport-pair definition of markets, this tablepresents two markets: GSO-ORD and GSO-MDW. The first, GSO-ORD, is served by twolegacy carriers (AA and DL) and is, therefore, a “talk eligible” market. Since both carrierstalked about capacity discipline,
Capacity-Discipline is equal to 1. However, the second In our empirical analysis, to identify the airports under the city-pair definition, we follow Brueckner,Lee and Singer [2014].
City Airport
Talk-Eligible Capacity-Discipline
Origin Destination Origin Destination Carrier Communication Airport-pair City-pair Airport-pair City-pairGreensboro, NC Chicago, IL GSO ORD AA (legacy) 1 1 1 1 0DL (legacy) 1GSO MDW UA (legacy) 0 0 0B6 (lcc) N/A
Notes. Table shows an example that highlight changes in our definition of communication when we movefrom airport-pair definition to city-pair definition of a market.
Table D.5: Communication and Available Seats for City-Pair Markets (1) (2)Log Seats Log SeatsCapacity Discipline 0.0015 -0.0185(0.0056) (0.0086)Exclude NYC & DC No YesR-squared 0.090 0.097N 787,628 628,022
Notes. We report semi-elasticities (see Footnote 24), with standard errors clustered at the bi-directionalmarket level in parentheses. Other control variables included in all regressions, but whose coefficients arenot reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicatorwith
Talk-Eligible and
Monopoly , origin- and destination-airport annual time trends, carrier-year-quarterfixed effects, and carrier-market fixed effects. All markets that include New York City, NY, or WashingtonD.C. are excluded in column 2. market, GSO-MDW, is served by one legacy, which is not discussing capacity discipline, andone low-cost carrier. Since the market is not talk-eligible,
Capacity-Discipline equals 0.As can be seen, under the airport-pair approach to defining markets, we have one marketwhere coordinated communication is taking place and one where it is not.Now consider the city-pair approach to defining markets. Under this approach, thetable shows a single market, Greensboro to Chicago, served by four carriers. Three legacycarriers serve the market, so this city-pair market is talk-eligible. However, one of the legacycarriers did not discuss capacity discipline (UA), so
Capacity-Discipline is equal to zero.This example shows how the frequency of
Capacity-Discipline m,t = 1 can differ betweenairport and city markets. Moreover, depending on the relative passenger volume throughGSO-ORD and GSO-MDW, we can get a different result. If a city has three airports, thenthe association between communication and capacity will become even more ambiguous andcannot be predicted by looking at what is happening in those three airports individually.Only two cities, Washington, D.C., and New York City, are served by three airports. Thus,the effects of communication on capacity may vary with market definitions.We use the same specification as Eq. (1), except for the markets’ city-pair definition. The45rimary results are in Table D.5, column (1). The interpretations of all variables are thesame as before, and the coefficient of interest is the first row, which shows that under thisalternative approach to defining markets, communication does not correlate with the offeredseats.To further shed light on why communication seems to be uncorrelated with capacity, webegin by observing that only two cities have three airports. Our result may be driven bywhat is happening in those two cities. So we re-estimate the model, but without Washington,D.C., (which includes BWI, DCA, and IAD) and New York City (EWR, JFK, and LGA),and present these results in column (2) of Table D.5. As we can see, in city-pairs served byat most two airports, the capacity discipline parameter estimates are equal to -1.85%, whichis similar to the -2.02% we found for the airport-pair markets. Thus, these two cities withthree airports (Washington, D.C., and New York City) appear to be driving the differencesbetween our primary, airport-pair market results and these city-pair market results. Tounderstand the reason behind these differences, we need to understand the role of airportsin the coordination mechanism, which is beyond our paper’s scope.
Appendix E Independent Verification
In Section 2.2 we detail the process we employ to code whether carriers discuss capacitydiscipline in each transcript. In this appendix, we consider two approaches to ensure thatour results are not affected by the way we coded.In the first approach, we hired an undergraduate student majoring in economics from theUniversity of Virginia. We provided the student with our definition of “capacity discipline,”and then had the student read every transcript and independently decide whether an earn-ings call discussed capacity discipline. Similar to our approach in Section 2.2, the studentclassified cases where the exact words “capacity discipline” were used and where the exactwords do not appear, but the concept of “capacity discipline” was discussed. A detaileddescription of the RA’s coding and the associated table is available upon request.In the second approach, we used natural language processing tools to automatically codeeach transcript based on whether a variation of the phrase “capacity discipline” was used.In this approach, we relied entirely on the automatic processing of the transcripts, ratherthan augmenting that work with human inspection of transcripts.Table E.1 shows the results of estimating our primary model under these two approaches.The first column shows the results of estimating this model using the RA’s transcript codingdata, and the second column shows the results of using the machine-coded transcripts. Toaid in comparison, we reproduce our main results, from the first column of Table 4, in the46able E.1: Estimates from Independently Classified Data (1) (2) (3)Log Seats Log Seats Log SeatsCapacity Discipline -0.0215 -0.0197 -0.0202(0.0077) (0.0071) (0.0060)R-squared 0.088 0.088 0.088N 841,991 841,991 841,991
Notes. In column 1, we present the results of estimating Eq. (1) using a communication variable that wasindependently coded by an RA, and in column 2 we use a communication variable that was automaticallycoded using natural language processing tools. Column 3 reports our primary estimates, Table 4, column 1to aid comparisons across these three approaches. We report semi-elasticities (see Footnote 24), withstandard errors clustered at the bi-directional market level in parentheses. Other control variables includedin all regressions, but whose coefficients are not reported are
Talk-Eligible , Monopoly , MissingReport ,interactions of the
MissingReport indicator with
Talk-Eligible and
Monopoly , origin- anddestination-airport annual time trends, carrier-year-quarter fixed effects, and carrier-market fixed effects. third column of Table E.1. We find similar estimates to what we present in Table 4 underboth the RA and Automatic coding approaches.
Appendix F Alternative Alignment of Earnings Calland Airline Data
As discussed in Section 2.4, an airline’s earnings call about a specific quarter takes placefollowing the conclusion of that quarter, and throughout the paper, we associate the contentof an earnings call with the three full months following the call. For example, we use thecontent from a Q1 call from mid-April to define
Carrier-Capacity-Discipline j,t for May,June, and July. Alternatively, we could associate the Q1 call, taking place in mid-April, withthe capacity data for
April , May, and June. In Appendix F.2 we reproduce all the results inTable 4 using this alternative definition. As we can see, the estimates are similar to those inTable 4, suggesting that our findings are robust with respect to this definition.
Appendix G Communication and Capacity Responsesand the DOJ Investigation
We investigate whether the airlines appear to have behaved differently before and after twokey moments in the legal cases regarding capacity discipline. First, we investigate whetherthe estimates vary before and after January 2010, reportedly the earliest date in the records47able F.2: Communication and Available Seats: Alternative Timing (1) (2) (3)Log Seats Log Seats Log SeatsCapacity-Discipline -0.0210(0.0058)Capacity Discipline 2 -0.0202(0.0063)Capacity Discipline 3 -0.0278(0.0106)Capacity Discipline 4 -0.0468(0.0378)Legacy Market x Capacity-Discipline -0.0194(0.0075)Mixed Market x Capacity Discipline (Legacy) -0.0182(0.0117)Mixed Market x Capacity Discipline (LCC) -0.0327(0.0149)R-squared 0.088 0.088 0.089N 841,991 841,991 841,991
Notes. This table replicates the primary estimates (columns 1, 3, and 5) from Table 4, except we nowassociate, e.g., the Q1 call taking place in mid-April with the airline capacity data for April, May, andJune. We report semi-elasticities (see Footnote 24), with standard errors clustered at the bi-directionalmarket level in parentheses. Other control variables included in all regressions, but whose coefficients arenot reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicatorwith
Talk-Eligible and
Monopoly . In column 2, these coefficients are allowed to vary based on the numberof legacy carriers in the market (either 0 or 1, 2, 3, 4, or 5 legacy carriers). In column 3, these coefficients areallowed to vary across legacy and mixed markets, and within mixed markets for legacy carriers and LCCs.Additionally, all regressions include origin- and destination-airport annual time trends, carrier-year-quarterfixed effects, and carrier-market fixed effects. requests the DOJ sent to the airlines (c.f. Section 2.1). To this end, we estimate Eq. (1),allowing the role of
Capacity-Discipline before January 2010 to be different from the oneafter January 2010. The estimates are presented in column (1) of Table G.1. We find thatthe difference in the estimates before and after January, 2010 is not statistically significant.Next, we consider whether the estimated parameter estimates vary before and after the
Washington Post article reporting the DOJ investigation was published in July 2015. TheDOJ investigation began at approximately the same time. As before, we estimate Eq. (1),allowing the relationship before the
Washington Post article to be different from the oneafter the article was published. We present these results in column (2) of Table G.1. We findthat the relationship we estimated in the paper between capacity choices and communication48able G.1: Communication and Available Seats: Before & After Key Moments in DOJInvestigation (1) (2)Log Seats Log SeatsPre-2010 Capacity Discipline -0.0206(0.0094)Post-2010 Capacity Discipline -0.0193(0.0068)Pre-WaPo Capacity Discipline -0.0164(0.0058)Post-WaPo Capacity Discipline 0.0597(0.0231)R-squared 0.088 0.088N 841,991 841,991
Notes. We report semi-elasticities (see Footnote 24), with standard errors clustered at the bi-directionalmarket level in parentheses. Other control variables included in all regressions, but whose coefficients are notreported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicator with
Talk-Eligible and
Monopoly . These coefficients are allowed to vary before and after 2010 or the publicationof the article in The Washington Post. Additionally, all regressions include origin- and destination-airportannual time trends, carrier-year-quarter fixed effects, and carrier-market fixed effects. does not appear to persist beyond the July 2015 announcement of the DOJ investigation.That said, it is important to note that Washington Post announcement occurs close to theend of our sample (which ends at the close of the third quarter of 2016), and so we observesubstantially fewer carrier-market-months in the post period of this particular analysis.
Appendix H Examples of Contents in Earnings Calls
In this section, we discuss the contents of the earnings calls pertaining to capacity discipline,which can shed light on what the airline executives generally say when they discuss capacitydiscipline. Airlines typically mention capacity discipline as part of a broader discussion oftheir capacity plans or broader strategic goals, but what is said depends on several factorsthat are carrier-specific, such as their networks of airports served, exposure to the fluctuationsin fuel costs, expectations about future demand, contracts with regional carriers and theirlabor contracts. The following quotes provide some context for how the topic of capacitydiscipline appear in discussions of an airline’s strategic goals.“. . . and while we still have a long way to go, we believe we are moving down the49ight track by continuing our capacity discipline while we strengthen our balancesheet and reinvest in key products, services, and in our fleet.”– American, 2007Q2“To get there, we’re focused on these key points: growing diversified revenues,treating our people well in a culture of positive employee relations, continuing ourcapacity discipline, keep our costs under control, running an airline customersworldwide prefer, deleveraging the business and limiting capital spending throughinvestments with high IRRs.” – Delta, 2011 Q3In this example, we see the airline specifically relate capacity discipline to cancellationdecisions.“And in addition, and kind of in line with our capacity discipline strategy, we’retaking a lot more aggressive approach on kind of day/week cancellations, par-ticularly in the sub UA network this Thanksgiving versus the past. ” – United,2011 Q3At times, airlines specifically note that they are comparing capacity growth to GDPgrowth.“As you can see, we remain committed to keeping our capacity discipline in checkand our capacity growth within GDP rates.” – Delta, 2010 Q3From their conversations, we can also deduce that airlines understand that there arebenefits from capacity discipline but that these benefits accrue only if their competitors alsoexercise capacity discipline. For instance, consider the following quotes from Alaska andUnited:“So we mentioned that Delta is trending upward in our markets. But we areactually seeing really good capacity discipline from other carriers on the WestCoast, in particular from United, from Virgin, and from Southwest making prettymaterial reductions in our network.” – Alaska, 2014 Q1“So again, I think our capacity discipline, as well as the industry discipline, whatwe’ve seen, I think, we’ve done a good job of not—the traffic that we’re missingis the low yield price-sensitive traffic and we’re doing a good job of not dilutingthe higher-end traffic. And I think the capacity discipline has allowed us to dothat.” – United, 2011 Q3 50n other cases, we see airlines discuss capacity decisions in the context of their competi-tors’ behavior, though without specifically raising the phrase “capacity discipline.”“We have taken steps to further trim our domestic capacity for 2003. But I thinkAmerican on its own making small incremental reductions in capacity don’t re-ally help solve the overall industry imbalance between capacity and demandand just put us at a further competitive disadvantage. . . We also continue toplan for reduced capacity on a year-over-year basis. For the third quarter weexpect mainline capacity to be down more than 5% from last year’s third quar-ter. . . Additionally, with a better alignment of capacity and demand this year theindustry may well benefit from a reasonably stable pricing environment.” – AA,2002 Q4As we can see in Fig. 2, some airlines discuss capacity discipline less frequently thanothers. For instance, AS discusses capacity discipline less frequently than AA, but their“messages” are similar whenever they discuss. The differences in what they say appear to bea function of the differences in the markets they serve—AS’s business is mostly concentratedin the Northwest region, as exemplified below (slightly edited for clarity).“[W]hat you are referring to are reductions announced by Delta and Northwest.We have almost no overlap with them so their capacity reductions really don’thelp us. But you might hypothesize that a capacity reduction in other markets inthe country might cause competitors’ capacity to move to fill that void and thatmight moderate what we would have seen otherwise in competitive incursions inour geography. But I would say the impact that we expect from those capacityreductions on AS is very small. And we are not a player in the transcontinentalexcept from Seattle and there hasn’t been any big reduction capacity in thetranscontinental from the Seattle market.” – AS, 2005 Q3During the period we study, several airlines file for bankruptcies, and we find that theircompetitors appear to keep track of their capacity plans. This concern is captured nicely inthe following:“We pulled down a fair bit of capacity this summer. ...[O]ne of the questions forthe whole industry is at significantly higher ticket prices, what does the demandpicture look like and then how much excess capacity is there? It’s exacerbateda little bit by the movement in competitor’s capacity, ...while domestic capacityis down about 5% in 2006, that’s not what we’re seeing within our geography.Within our geography we’re seeing competitive capacity [up about] 3%. But51able I.1: Financial Transparency and Information Sharing (1) (2) (3) (4) (5)Log Seats Log Seats Log Seats Log Seats Log SeatsOnly j Talks 0.0136(0.0054)Monopoly Capacity Discipline 0.0294 0.0027(0.0064) (0.0050)Capacity Discipline N − j ” -0.0026(0.0060)R-squared 0.083 0.084 0.061 0.083 0.083N 841,991 841,991 438,980 841,991 841,991 Notes. We report semi-elasticities (see Footnote 24), with standard errors clustered at the bi-directionalmarket level in parentheses. Other control variables included in all regressions, but whose coefficients arenot reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicatorwith
Talk-Eligible and
Monopoly , origin- and destination-airport annual time trends, carrier-year-quarterfixed effects, and carrier-market-structure fixed effects. Column 3 omits the
Talk-Eligible and
Monopoly variables. you know, we’re hopeful, and we have got to see what happens to the rest ofthe capacity and how carriers [act with] bankruptcy for this year. We’re sort ofwatching what’s happening, with Independence Air going away, and with Deltaand Northwest bankruptcy, their shrinking capacity in the Heartland and onthe East Coast, and we’re not big players in either of those markets... They’removing some capacity in the West Coast markets that they pulled during thebankruptcy, so we are a little bit concerned about that.” –AS, 2005 Q4In discussing capacity discipline, airlines also discuss various ways in which they mightget rid of their “excess” capacity. These methods include a mix of reducing capacity buyingplan, re-writing contracts with their regional partners, expediting retirement of aircraft,delaying future aircrafts deliveries, re-allocating capacities to international markets, wherethe mix and thus the savings vary across airlines.
Appendix I Additional Results
In this section, we present the results from estimating models in Tables 6, C.2, A.1, D.5, E.1and F.2, with carrier market-structure fixed effects.52able I.2: Estimates for Conditional Exogeneity (Carrier-Market-Structure Fixed Effects) (1) (2) (3) (4) (5) (6)slow weakness domestically internationally stable paceZ Token -0.0026 0.0022 0.0131 -0.0008 -0.0041 0.0057(0.0056) (0.0060) (0.0060) (0.0047) (0.0087) (0.0065)Capacity Discipline -0.0148 -0.0150 -0.0144 -0.0150 -0.0148 -0.0154(0.0049) (0.0049) (0.0049) (0.0049) (0.0049) (0.0050)N 841,991 841,991 841,991 841,991 841,991 841,991
Notes. Estimation results from including new tokens an additional regressors in Eq. (1). The table showsthe coefficient estimates for each token, and for
Capacity-Discipline . We report semi-elasticities (seeFootnote 24), with standard errors clustered at the bi-directional market level in parentheses. Other controlvariables included in all regressions, but whose coefficients are not reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicator with
Talk-Eligible and
Monopoly , origin-and destination-airport annual time trends, carrier-year-quarter fixed effects, and carrier-market-structurefixed effects.
Table I.3: Communication and Available Seats: The Role of Market Size and Business Travel (1) (2)Log Seats Log SeatsCapacity Discipline x Small Population -0.0060(0.0252)Capacity Discipline x Medium Population -0.0083(0.0091)Capacity Discipline x Large Population -0.0189(0.0054)Capacity Discipline x Low Business -0.0131(0.0080)Capacity Discipline x Medium Business -0.0122(0.0066)Capacity Discipline x High Business -0.0233(0.0090)R-squared 0.084 0.082N 841,991 620,762
Notes. We report semi-elasticities (see Footnote 24), with standard errors clustered at the bi-directionalmarket level in parentheses. Other control variables included in all regressions, but whose coefficients arenot reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicatorwith
Talk-Eligible and
Monopoly . These coefficients are allowed to vary based on the market size orbusiness travel classifiers. Additionally, all regressions include origin- and destination-airport annual timetrends, carrier-year-quarter fixed effects, and carrier-market-structure fixed effects. (1) (2)Log Seats Log SeatsCapacity Discipline -0.0020 -0.0088(0.0040) (0.0050)Exclude NYC & DC No YesR-squared 0.086 0.092N 787,628 628,022
Notes. We report semi-elasticities (see Footnote 24), with standard errors clustered at the bi-directionalmarket level in parentheses. Other control variables included in all regressions, but whose coefficients arenot reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicatorwith
Talk-Eligible and
Monopoly , origin- and destination-airport annual time trends, carrier-year-quarterfixed effects, and carrier-market-structure fixed effects. All markets that include New York City, NY, orWashington D.C. are excluded in column 2.
Table I.5: Estimates from Independently Classified Data (Carrier-Market-Structure FixedEffects) (1) (2) (3)Log Seats Log Seats Log SeatsCapacity Discipline -0.0169 -0.0181 -0.0150(0.0061) (0.0057) (0.0049)R-squared 0.083 0.083 0.083N 841,991 841,991 841,991
Notes. In column 1, we present the results of estimating Eq. (1) using a communication variable that wasindependently coded by an RA, and in column 2 we use a communication variable that was automaticallycoded using natural language processing tools. Column 3 reports our primary estimates, Table 4, column 1to aid comparisons across these three approaches. We report semi-elasticities (see Footnote 24), withstandard errors clustered at the bi-directional market level in parentheses. Other control variables includedin all regressions, but whose coefficients are not reported are
Talk-Eligible , Monopoly , MissingReport ,interactions of the
MissingReport indicator with
Talk-Eligible and
Monopoly , origin- anddestination-airport annual time trends, carrier-year-quarter fixed effects, and carrier-market-structure fixedeffects. (1) (2) (3)Log Seats Log Seats Log SeatsCapacity-Discipline -0.0166(0.0049)Capacity Discipline 2 -0.0171(0.0052)Capacity Discipline 3 -0.0152(0.0101)Capacity Discipline 4 -0.0286(0.0221)Legacy Market x Capacity-Discipline -0.0223(0.0065)Mixed Market x Capacity Discipline (Legacy) 0.0003(0.0086)Mixed Market x Capacity Discipline (LCC) -0.0213(0.0095)R-squared 0.084 0.084 0.084N 841,991 841,991 841,991
Notes. This table replicates the primary estimates (columns 1, 3, and 5) from Table 4, except we nowassociate, e.g., the Q1 call taking place in mid-April with the airline capacity data for April, May, andJune. We report semi-elasticities (see Footnote 24), with standard errors clustered at the bi-directionalmarket level in parentheses. Other control variables included in all regressions, but whose coefficients arenot reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicatorwith
Talk-Eligible and
Monopoly . In column 2, these coefficients are allowed to vary based on thenumber of legacy carriers in the market (either 0 or 1, 2, 3, 4, or 5 legacy carriers). In column 3, thesecoefficients are allowed to vary across legacy and mixed markets, and within mixed markets for legacycarriers and LCCs. Additionally, all regressions include origin- and destination-airport annual time trends,carrier-year-quarter fixed effects, and carrier-market-structure fixed effects. (1) (2)Log Seats Log SeatsPre-2010 Capacity Discipline -0.0229(0.0081)Post-2010 Capacity Discipline -0.0110(0.0061)Pre-WaPo Capacity Discipline -0.0157(0.0050)Post-WaPo Capacity Discipline 0.0776(0.0220)R-squared 0.084 0.084N 841,991 841,991
Notes. We report semi-elasticities (see Footnote 24), with standard errors clustered at the bi-directionalmarket level in parentheses. Other control variables included in all regressions, but whose coefficients arenot reported are
Talk-Eligible , Monopoly , MissingReport , interactions of the
MissingReport indicatorwith
Talk-Eligible and
Monopoly . These coefficients are allowed to vary before and after 2010 or thepublication of the article in The Washington Post. Additionally, all regressions include origin- anddestination-airport annual time trends, carrier-year-quarter fixed effects, and carrier-market-structure fixedeffects. ppendix References Awaya, Yu, and Vijay Krishna.
AmericanEconomic Review , 106(2): 285–315.
Awaya, Yu, and Vijay Krishna.
Theoretical Economics , 14(2): 513–553.
Bamberger, Gustavo E., Dennis W. Carlton, and Lynette R. Neumann.
The Journal of Law and Economics , 47(1): 195–222.
Berry, Steven, and Panle Jia.
AEJ: Microeconomics , 2: 1–43.
Berry, Steven, Michael Carnall, and Pablo T. Spiller.
Advances in Airline Eco-nomics: Competition Policy and Antitrust . Vol. 1, , ed. Darin Lee, 183–214. Amsterdam:Elsevier.
Berry, Steven T.
American EconomicReview , 80(2): 394–399.
Berry, Steven T.
Econo-metrica , 60(4): 889–917.
Borenstein, Severin.
RAND Journal of Economics , 20(3): 344–365.
Borenstein, Severin.
NBER Working Paper . Borenstein, Severin, and Nancy Rose.
Journal of Political Economy , 102(4): 653–683.
Brueckner, Jan K., and Pablo T. Spiller.
Journal of Law and Economics , 37(2): 379–415.
Brueckner, Jan K., Darin Lee, and Ethan Singer.
Review of IndustrialOrganization , 44(1): 1–25. 57 iliberto, Federico, and Elie Tamer.
Econometrica , 77(6): 1791–1828.
Ciliberto, Federico, and Jonathan W. Williams.
Journal of Law and Economics ,53(3): 467–495.
Ciliberto, Federico, and Jonathan W. Williams.
RAND Journal of Economics , 45(4): 764–791.
Ciliberto, Federico, Emily E. Cook, and Jonathan W. Williams.
Review of IndustrialOrganization , 54(3): 3–36.
Eizenberg, Alon.
Review of Economic Studies , 81(3): 1003–1045.
Evans, W.N., and I. N. Kessides.
Quarterly Journal of Economics , 109(2): 341–366.
Fabra, Natalia, and Mar Reguant.
American Economic Review , 104(9): 2872–2899.
Gerardi, Krisopher S., and Adam Hale Shapiro.
Journal of Political Economy ,117(1): 1–37.
Goldberg, Y., and O. Levy.
ArXiv e-prints . Hendricks, Ken, Michele Piccione, and Guofu Tan.
Econometrica , 67(6): 1407–1434.
Imbens, Guido W., and M. Wooldridge.
Kim, E. Han, and Vijay Singal.
American Economic Review , 83(3): 546–569.
Mikolov, T., K. Chen, G. Corrado, and J. Dean.
ArXiv e-prints .58 lley, G. Steven, and Ariel Pakes.
Econometrica , 64(6): 1263–1297. ˇReh˚uˇrek, Radim, and Petr Sojka.
Rosenbaum, P.
Journal of American StatisticalAssociation , 79(385): 41–48.
Singhal, Amit.
Bulletin of theIEEE Computer Society Technical Committee on Data Engineering , 24(4): 35–42.
White, Halbert, and Karim Chalak.