OPEN SKIES: ESTIMATING TRAVELERS BENEFITS FROM FREE TRADE IN AIRLINE SERVICES*

Transcription

1 OPEN SKIES: ESTIMATING TRAVELERS BENEFITS FROM FREE TRADE IN AIRLINE SERVICES* Clifford Winston Jia Yan* Abstract. The United States has negotiated bilateral open skies agreeents to deregulate airline copetition on U.S. international routes, but little is known about their effects on travelers welfare and the gains fro the U.S. negotiating agreeents with ore countries. We develop a odel of international airline copetition to estiate the effects of open skies agreeents on fares and flight frequency. We find the agreeents have generated at least $4 billion in annual gains to travelers and that travelers would gain an additional $4 billion if the U.S. negotiated agreeents with other countries that have a significant aount of international passenger traffic. * Winston: Econoic Studies Progra, Brookings Institution, 1775 Massachusetts Avenue, N.W., Washington, D.C (e-ail: CWinston@brookings.edu); Yan: Departent of Econoics, Washington State University, Pullan, Washington (e-ail: Jiay@WSU.edu). We are grateful to John Byerly and Douglas Lavin for their assistance and valuable discussions. We also received very helpful coents fro Jan Brueckner, Kenneth Button, Ashley Langer, Robin Lindsey, Vikra Maheshri, participants in the 2012 International Transport Econoics Association eeting in Berlin, and the referees. Winston gratefully acknowledges financial support fro the Federal Aviation Adinistration.

2 Following Aerica s successful airline deregulation experient in the late 1970s, any countries deregulated their doestic airline arkets. In contrast, deregulation of international airline arkets has occurred ore slowly. At the 1944 Chicago convention, the United States sought to establish ultilateral agreeents whereby arket forces would priarily deterine fares and capacities on international routes. But the effort failed, and ever since, bilateral agreeents have provided the fraework under which fares and service frequency between two countries are deterined. The Carter adinistration prooted the idea of open skies, liberal bilateral agreeents that freed arket forces to be the ost iportant deterinants of fares and capacity. Beginning with a successful agreeent with the Netherlands in 1992 and a recent one with Japan in late 2010, the United States has tended to consuate open skies agreeents with one country at a tie. Other countries have also taken that approach, while ultilateral agreeents aong countries in Africa, South Aerica, and the European Union have allowed participants to serve each others countries, usually without any restrictions on fares. 1 Generally, open skies agreeents are initiated because two countries believe that utual benefits exist fro pricing freedo and having unfettered airline access to each other s gateway airport(s); such agreeents are opposed by countries that seek to protect their flag carrier(s) fro copetition by closely regulating fares, entry, and flight frequency. It is therefore iportant to know whether the open skies agreeents that have been negotiated to date have increased copetition and benefitted air travelers and whether travelers welfare would iprove 1 The United States concluded a ultilateral agreeent in 2001 that superseded bilateral open skies agreeents with several APEC countries, including Singapore and Chile, and in 2007 finalized a coprehensive open skies agreeent with the European Union and its eber states that allowed for open skies between the United States and the United Kingdo and Spain aong other European countries, which previously did not have open skies agreeents with the U.S. 1

3 significantly if ore countries negotiated open skies agreeents. Cristea, Huels, and Roberson (2012) analyzed data fro the U.S. Departent of Transportation that included only U.S. carriers and international routes flown by those carriers and estiated that open skies agreeents have reduced fares, adjusted for changes in flight frequency and new routings, 32 percent copared with fares in arkets that reained regulated. Pierartini and Rousova (2013) found that full adoption of open skies agreeents would increase passenger traffic worldwide 5 percent, but they did not assess the effects on fares. Finally, Micco and Serebrisky (2006) found that open skies agreeents that have been negotiated between 1990 and 2003 and that govern air cargo and passengers have caused a 9 percent drop in the cost of shipping freight by air. A related literature on international airline copetition assesses the effects on travelers of airline alliances where U.S. and foreign carriers have established liited arketing arrangeents, such as a reciprocal frequent flier progra, or an international code-share agreeent that allows an airline to sell seats on a partner s planes as if they were its own. Alliances facilitate interline traffic across the networks of the partners, providing sealess service in city-pair arkets where single-carrier service is not available (Brueckner (2001)). Brueckner, Lee, and Singer (2011) provides recent evidence that alliances reduce fares relative to those offered by two nonaligned carriers by eliinating double arginalization of interline fares. Because alliances could account for soe of the benefits attributable to open skies agreeents, it is iportant to distinguish between the two policies effects on travelers welfare. 2 2 Whalen (2007) attepted to distinguish between the effects on fares of open skies agreeents and code-share alliances that were given antitrust iunity and found that open skies led to soewhat higher fares, which he could not explain. 2

4 In this paper, we draw on a large saple of ajor U.S. and non-u.s. international routes that are served by the world s leading airlines to explore the effects of open skies agreeents on air travelers welfare, accounting for changes in fares and flight frequency. We estiate a odel of airline arket deand, pricing, flight frequency, and arket structure and find that open skies agreeents have generated at least $4 billion in annual gains to travelers on our saple of U.S. international routes, which includes alost a 15 percent reduction in fares and aounts to roughly 20 percent of carriers annual revenues on those routes. Moreover, we find that travelers would reap another $4 billion annually if United States policyakers could overcoe the political obstacles that have prevented the fro negotiating open skies agreeents with other countries that have a significant aount of U.S. international passenger traffic. Given that open skies policies have advanced with little publicized evidence of their benefits to travelers, broad disseination of this (and other) positive evidence ay spur policyakers to eliinate the reaining econoic regulations on foreign airline copetition and to enable the world s airlines to operate efficiently in a fully deregulated environent. I. Overview of the Approach and the Data Set In this analysis, international airline arkets are defined as non-directional airport pairs, such as Washington, D.C. Dulles and London, Heathrow. Our goal is to estiate the effect of open skies agreeents (OSAs) on travelers welfare in those arkets. Traditional analyses of the econoic effects of a regulatory policy specify a duy variable, typically assued to be exogenous, which indicates when the regulatory policy is in effect and captures the policy s effect on a variable related to welfare such as prices (Joskow and Rose (1989)). The analysis 3

5 here is coplicated by several endogenous variables that affect each other and deterine the effects of open skies agreeents on air travelers welfare. We outline the fraework in figure 1. The policy variable, OSAs, eliinates restrictions on entry and fares in a arket and thus affects flight frequency and fares. We classify fares by service segents, such as first class, econoy, and so on. In addition, OSAs can affect arket structure, as easured by the nuber of carriers, which affects flight frequency. Market structure also affects and is affected by fares. Air travel deand is a function of both fares and frequency. We distinguish between top level deand, easured by the nuber of passengers, which affects flight frequency, and botto level deand, which allocates passengers across fare segents and is easured by fare segent expenditure shares. Finally, air travelers deand is used to easure the welfare effects of OSAs based on the copensating variation that is, the change in expenditures that enables travelers to achieve the sae level of utility fro fares and flight frequency before OSAs are ipleented as they do after OSAs are ipleented. Our epirical analysis therefore consists of specifying and estiating a siultaneous equations odel that treats deand, fares, frequency, the regulatory environent, and the nuber of carriers as endogenous and that affect each other as indicated by the figure. 3 To execute the analysis, we purchased data that are provided by the world s leading international airlines to the International Air Transportation Association (IATA). We kept the cost anageable by constructing a saple that consisted of the top 500 non-directional international airport pair routes, including U.S. and non-u.s. routes, based on passengers. 3 Modern epirical industrial organization offers sophisticated structural approaches to derive a odel of arket structure based on airlines strategic behavior (for exaple, Cilberto and Taer (2009)), but we cannot take such an approach here because airlines copete in soe international arkets where entry and fares are tightly regulated. 4

6 It is coon practice in studies of air transportation to construct a saple of routes based on a threshold of the populations of the cities whose airports coprise the routes (e.g., Berry and Jia (2010)) or of the ranking of the routes based on passenger traffic (e.g., Morrison and Winston (2000)) because the largest cities and routes have a disproportionately large share of all airline traffic. Such saples tend to consist of airline travel that would be expected to be generated to a significant extent by a rando saple of airline tickets and should not be seriously affected by selection bias. We copare the findings based on our full saple of routes with the findings based on our priary saple of interest, a subsaple of U.S. international routes that includes the open skies agreeents that were negotiated during the period of study. As shown later, we obtain siilar findings fro the two saples, which is useful validation because both their size and the average characteristics of their routes are different. We obtained onthly suaries of passenger travel during 2005 to According to IATA, the top 500 routes accounted for 26 percent of international airline passengers during The 66 U.S. international routes in the saple accounted for 20 percent of passengers on U.S. international routes. During the period of our saple, the top 500 international routes carried an annual average of 489,660 passengers per route and generated $186 illion in passenger revenues per route and the 66 U.S. international routes carried an annual average of 452,484 passengers per route and generated $306 illion in passenger revenues per route. Because we do not extrapolate the findings to estiate the effects of open skies agreeents on other international routes that are not included in those saples, our conclusions are not subject to selectivity bias. However, as noted, we check the robustness of the paraeter estiates for 4 Although the data set constitutes a viable saple of air travel throughout the world, IATA cannot warrant copleteness or the accuracy of all data eleents. 5

7 U.S. international routes by coparing the with paraeter estiates based on the full saple of 500 international routes. For a given international origin-destination pair, the variables in the data set include average fares plus taxes for five fare classes (discount econoy, full econoy, business, first class, and other), the nuber of passengers by fare class, the nuber of non-stop and connecting flights (hence, we account for non-stop and connecting routes), and the carriers serving the arket with non-stop service and with connecting service. 5 We cobined fares that were siilar into the sae classification and analyzed air travel behavior for three fare classifications: discount econoy and other fares, full econoy, and business and first class fares. The availability of fares and passengers for different fare classifications is a useful feature of the data set because we do not have to restrict travelers preferences to be hoogeneous across those classifications. At the sae tie, we found that soe routes had issing data for particular classes and others periodically had issing data for a onth or so. Hence, our final data set is an unbalanced panel of observations, consisting of coplete data for the three fare classifications for 415 non-directional routes. 6 The treaties that govern aviation policy between two countries fall under the following seven categories: traditional (a non-open skies agreeent that iposes regulatory restrictions on fares, entry, and flight frequency); provisional open skies (functionally open skies, but not yet 5 Fares were provided without taxes. To obtain fares that included taxes, we copiled data provided by IATA on total tax revenue for each arket, each period, and each fare class and added the tax per passenger to the average fares to obtain full (average) fares including taxes. 6 Missing data could arise because the carriers serving a route did not offer service in a particular fare classification or because there were no bookings on a route for a particular fare classification during a given onth. If those observations could be identified, it would be possible to use the in a selectivity odel, but the routes with values of zero for particular fare classifications could not be cobined with routes that had data for all fare classifications to analyze travelers deand because fare substitution patterns would be different for those routes. 6

8 official); open skies (full liberalization of fares, entry, and flight frequency subject to available airport capacity); EU open skies (open skies applying to routes between EU eber countries); US-EU open skies (open skies applying to routes between US and EU eber countries); open skies in force; and transitional (an open skies agreeent has been negotiated but it will be officially in effect at soe future date). 7 We were not able to estiate odels specifying duy variables for each category, so we created three categories by treating traditional and transitional as distinct categories and cobining the various open skies categories. The treaties that the United States has negotiated with other countries are suarized in the U.S. Departent of State s website 8 and the treaties that other countries have negotiated between theselves are copiled by the International Civil Aviation Organization (ICAO). Traditional agreeents govern 63 percent of the routes in our saple, open skies govern 35 percent, and transitional govern 2 percent. Generally, the agreeents reflect the attitude that two countries have toward liberalizing trade with each other. Our final estiations specify a binary duy variable to indicate the presence of an open skies agreeent (OSA), defined as 1 if the regulatory status on a route is open skies or transitional; 0 otherwise. Certain liitations of the data require us to qualify our analysis as likely to understate the benefits of open skies agreeents. First, although our data include the passengers on all the doestic routes that contribute traffic to a given international origin-destination pair (e.g., all the passengers who originate on a U.S route and connect at Washington, D.C., Dulles airport to fly to London, Heathrow airport are included in this D.C.-London international route), we do not 7 Open skies, category 3, ay be pending foralities, such as standard approvals by a non-u.s. country that is involved in the agreeent, while open skies in force, category 6, eans the agreeent is fully bound as a atter of international treaty law. In practice, there is no difference fro the U.S. perspective between open skies and open skies in force

9 easure the benefits to doestic (beyond) traffic generated by OSAs. For exaple, an increase in copetition fro an OSA that reduces fares fro Washington, D.C., Dulles airport to London, Heathrow airport ay also reduce fares on flights fro certain U.S. airports to Dulles airport to attract additional traffic to the United Kingdo. Second, we are not able to estiate a odel to deterine the tiing of an open skies agreeent, but by constructing the OSA duy variable based on the specific date that an open skies agreeent was or about to be in effect, we are likely to understate the benefits of such agreeents because soe liberalization of air travel regulations between two countries ay have occurred before a foral open skies agreeent was negotiated. For exaple, Fisher-Ke and Windle (2012) suarized U.S. aviation negotiations with China during 1999 to 2007, as China gradually agreed to liberalize regulations on the nuber of weekly flights between the two countries, the nuber of carriers that could provide service, and the cities that could be served without negotiating a foral open skies agreeent. Third, we hold the international airline network constant in our analysis, which eans we do not include the benefits fro additional routes between two countries that ay receive service because of an OSA. Finally, as noted, our saple does not include air travelers who ay benefit fro OSAs but who travel on lower density international routes that are not included in our top 500 international routes. This does not ean that our findings are biased due to saple selection. Rather, the benefits fro OSAs on lower density routes could be estiated using a new saple for those routes and the total benefits fro OSAs would then be the su of the benefits fro both saples. We provide soe perspective on the potential additional benefits fro OSAs by coparing findings fro all the international routes in the saple and the U.S. international routes, which carry fewer passengers annually per route. 8

10 Siple suaries indicate that the fare and frequency data are plausible. We show in figures 2a-c that although 2009 yields (average fare per ile) for international routes in all fare classifications are deterined by open skies and regulation, they are consistent with standard suaries of fares in deregulated U.S. arkets (see, for exaple, Morrison and Winston (1995)) by declining with route distance because of the fixed costs of takeoff and landing. As expected, yields for first and business class exceed those for full econoy and discount econoy and the eans of all the yields, which range fro roughly 60 cents per ile to 20 cents per ile, exceed yields on U.S. doestic routes during 2009 of roughly 13 cents per ile. A siple coparison of average fares on international routes with and without open skies agreeents suggests that open skies agreeents have reduced fares for all fare classifications and that the reductions are sizable by 2009 they were roughly 40 percent (table 1). A siilar coparison also shows that routes with open skies agreeents have had ore flights even though they had fewer passengers, with the difference peaking at close to 10 percent during 2007 and 2008 (table 2). Of course, those coparisons do not hold any other influences on fares and flight frequency constant. We do so by specifying a plausible odel of international airline arkets. II. Epirical Specification Our siultaneous equations odel of international airline arkets specifies the deand and supply, including fares and flight frequency, for air transportation, the agreeent (traditional or open skies) negotiated by the two countries that governs arket copetition, and the arket structure, easured by the nuber of carriers on the route. 9

11 A. Deand We easure fares and passenger deand by fare classification, discount econoy, full econoy, and business and first class, where the fare is the weighted (by nuber of passengers) average fare across all airline products in that fare classification and the nuber of passengers is obtained by aggregating the passengers choosing those products. Airline products are defined by carrier (e.g., United Airlines) and airport itinerary (e.g., non-stop between Washington, D.C. Dulles Airport and London Heathrow airport). We use Hausan s (1997) two-level approach, where the top level corresponds to the overall deand for air travel in the arket, and the botto level corresponds to the allocation of total deand aong the three fare classifications (referred to as arket segents), conditional on total expenditures. We odel the botto level using the flexible Alost Ideal Deand Syste (Deaton and Muellbauer (1980)), so deand for a arket segent is given by: (1) s g E g g log P gr gc gy 3 gg g1 gt g log pg g logl s g, g 1,2,3 where sg is the revenue share of segent g in arket (e.g., Washington, D.C., Dulles to London, Heathrow) in onth t ; E is the overall arket expenditure in onth t and P is a price index; pg is the full price (including taxes) of segent g ; and L is the distance between the end-point airports. We include fixed effects duy variables for regions of the world (Europe, North Aerica, and so on), gr, the end-point countries, gc, year, gy, and onth, gt. The regional duy variables indicate routes where both the origin and destination airports are located within a given region so they capture the effects of free trade agreeents (e.g., within the European Union and North Aerica). 10

12 As an illustration of how we specify the regional and country duies, for the Washington, D.C.-Mexico City route we specify one regional duy variable (North Aerica) and two country duy variables (one for the U.S. and another for Mexico). There are a total of 90 countries in our saple. An alternative specification would include country-pair duies, so the U.S. and Mexico would coprise such a duy. There are a total of 242 country-pairs in our saple. As we explain later, our epirical findings are robust with respect to the alternative ways of controlling for country effects. We also specify individual arket effects, g, which we treat as rando and that allow us to identify the coefficients associated with the tie-invariant regressors such as distance and the open-skies duy variables. We discuss the iplications of the rando effects specification s for identification and estiation of the odel later. Finally, The price index is given by the translog functional for: g is an error ter (2) logp 0 g log pg gg log pg log pg g1 g1g 1 Assuing travelers axiize utility, we ipose the following well-known restrictions on the deand paraeters: Adding-up: (3) 3 g1 3 g1 1, g gr 0, 3 g1 3 g1 0, gc g 0, 3 g1 3 g1 gy gg 0 g 1,2,3, 0, and 3 g1 0 gt 3 g1 0, g (4) Hoogeneity: gg 0, g 1,2, 3 3 g1 (5) Syetry: gg gg. 11

13 The adding-up constraints in equation (3) iply that it is appropriate to use only two of the three revenue share equations in estiation to avoid the singularity proble. Because the choice of which two does not affect the estiation results, we drop segent 3, discount econoy. The volue of air travel in a non-directional airport-pair arket captures a portion of the origin and destination countries trade in the aviation service sector; thus, we specify the top level deand as a gravity equation, which is the ost coonly used functional for to odel trade flows: (6) Q 0 1 logp 2 logk 3 logn 4 logi logl log 5 r c y t T Q where Q is the nuber of air travelers in arket at tie t ; P is the price index given in equation (2); K is the nuber of onthly direct and connecting flights; N is the geoetric ean of the populations of the end-point countries at tie t ; I is the geoetric ean of the per-capita-incoes of the end-point countries at tie t ; regional, end-point countries, year, and onth effects; r, c, y, and t are fixed denotes the rando arket effects that are allowed to be correlated with the regressors; and T, where is a rando coponent with zero ean and T denotes the tie trend, captures the rando arket trend. 9 The nuber of flights affects arket deand because ore frequent flights reduce the costs of schedule delay, defined as the difference between travelers preferred departure ties and their actual departure 9 We explored functional specifications that included a linear and a squared ter for distance for the top-level deand equation and for other equations in our odel where distance appeared, but we did not obtain statistically significant estiates of the squared ter. 12

14 ties, which increases the attractiveness of air copared with alternative odes and increases its arket share and the size of the travel arket. 10 B. Supply If airlines operate in an international arket that is not subject to econoic regulations, they decide on the nuber of flights to offer and the fares to charge for those flights to axiize profits. If carriers operate in a regulated arket, they ay or ay not be able to deterine their flight frequencies while fares are set by the regulatory agreeent. Given the constraints iposed on airlines when they operate in a regulated arket, we do not attept to build a structural odel of airline behavior; instead, we siply indicate that our epirical odel resebles a two-stage gae of airlines supply decisions, where airlines first deterine their flight frequency and then set fares given frequency. We then easure the effect of an open skies agreeent on copetition in a arket that ay increase the nuber of flights and reduce fares by reducing carriers costs or their price ark-ups or both. Drawing on the U.S. airline deregulation experience, an open skies agreeent will have an initial and potentially large effect on carriers pricing and operating behavior shortly after it is ipleented and have effects that persist over tie as carriers adjust to the change in the copetitive environent. We therefore capture an open skies agreeent s cuulative effect on fares and flight frequency by specifying duy variables to indicate its effects in the short and long run. Because our saple covers the 2005 to 2009 period, we capture the initial or short-run effects of the 12 open skies agreeents (covering 35 of the 415 routes in the saple) that were 10 The literature is uch less clear on whether the nuber of flights, which does not vary by fare classification, affects the expenditure shares, all else constant, and on the signs of those effects. We explored the atter epirically and found that the nuber of flights was highly correlated with total expenditures and produced very iprecise paraeter estiates and iplausible elasticities. 13

15 signed after 2005 and the long-run effects of the 67 open skies agreeents (covering 144 of the 415 routes) that were signed before Only 3 open skies agreeents (covering 11 routes in the top 500 routes of the initial saple) were signed during 2000 to 2005, preventing us fro capturing the interediate effects of the agreeents because they were collinear with the regional and end-point country duies. We do not directly odel flight frequency because, as indicated by equation (6), it affects arket deand and, subject to the regulatory agreeent, it is adjusted by airlines to respond to changes in deand. Thus, it is very difficult to uniquely identify both deand and frequency, although we later specify instruents for frequency to estiate its effect on deand. Instead, we odel the long-run equilibriu relationship in a arket between deand and flight frequency by drawing on Belobaba, Odoni, and Barnhart s arguent (2009, p. 159) that airlines choose flight frequency to achieve a target load factor as part of their long-run fleet planning process. The load factor, which is defined as the percentage of seats filled by paying passengers, is our easure of capacity utilization that we odel as a function of arket characteristics. We ake the plausible assuption that aircraft size (nuber of seats) in ost international arkets can be taken as given because it is largely deterined by arket characteristics, such as the population at the endpoint cities, distance, and airport size. For a given aircraft size, the nuber of flights is therefore equivalent to the total nuber of seats. Based on airlines long-run fleet planning process, we expect and later verify epirically with Augented Dickey-Fuller tests that passengers and flights are cointegrated that is, soe linear cobination of the is stationary to aintain a long-run equilibriu relationship in capacity utilization. Forally, let 1, 1 denote the noralized cointegrating vector, where cointegration iplies that long-run equilibriu capacity utilization is defined by 14

16 K logq e log 1. In the special case that the cointegrating vector is (1, -1), then e siply the log of flights to deand ratio, which easures the log of the (inverse) load factor. We expect the target (inverse) load factor in a arket to depend on the population and per-capita incoe of the cities that coprise the end-point airports because those variables deterine arket size, and also to depend on the arket structure, regulatory status, the length of haul, and other arket characteristics including the presence of an alliance. We therefore specify: is (7) e log log 8 2 N 3 logi 4 logc 5A K K K L X Γ r c y q OSA 6 l OSA 7 s where population, incoe, and length of haul have been defined previously. We easure arket structure with C, the nuber of carriers, and account for the presence of a ajor airline alliance in arket at tie t with a duy variable A ; l OSA is a duy variable indicating the open skies status of arket in the long run (1 if an open skies agreeent was signed before 2000; 0 otherwise); s OSA is a duy variable indicating the open skies status of arket in the short run (1 if an open skies agreeent was signed after 2005; 0 otherwise); K X is a vector of route-level attributes that affect airlines flight scheduling decisions, including the difference between the historical average onthly rainfall and teperature at the origin and destination airports and the nuber of cities connected to the end-point airports. We also include regional, end-point countries, year, and onthly fixed effects,, and arket effects 15 r c y t and rando, which are allowed to be correlated with the regressors. Those effects K include, for exaple, variations in aircraft size and load factors. Finally, represents the longrun equilibriu error in capacity utilization. Because the error has to be a stationary series, we test whether open skies agreeents have caused long-run equilibriu capacity utilization to undergo a structural change.

17 We specify airlines pricing decisions by first noting that conditional on arket passengers, flight frequency affects air fares through short-run fluctuations in capacity utilization that are captured by e. The reaining direct influences on fares, arket structure, the presence of a ajor airline alliance, and the status of the open skies agreeents affect ark-ups, while carriers operating costs and thus fares are affected by trip distance interacted with the price of crude oil at tie t, O t, and other route-level characteristics, onthly rainfall and teperature. Thus, we specify the fare equations as: f X, including historical average (8) log f logq e logc g 0g 5g 1g l OSA X f Γ f g OSA rg 6g cg s 2g yg 7 g log 3g tg L logl O g 8g f g A 4g, g 1,2,3 t where we also include regional, end-point countries, year, and onthly fixed effects ( rg, cg, and tg ) and rando arket effects ( g ), which are allowed to be correlated with regressors; and f g is an error ter. yg C. Market Structure An open skies agreeent also affects fares by affecting arket structure, naely the nuber of carriers in an international airline arket, because airlines are free to enter the arket to provide service, while they are generally unable to do so in a regulated environent. Our specification includes duy variables indicating whether an open skies agreeent was negotiated in the short run and the long run, fare revenues, which help deterine potential profits, and exogenous arket characteristics. We also include a rando arket tie trend,, where is a rando coponent with zero ean and T is a tie trend, because we expect the T 16

18 evolution of a arket s structure to be tie-persistent and to be different in different arkets an expectation confired by tie series plots of the nuber of carriers in each arket. Our epirical odel of arket structure is therefore: (9) log l s C 0 1OSA π2osa π4 logr 5A M M M logl X Γ T 6 r c y t where R is total fare revenues in arket at tie t and M X is a vector of arket characteristics that are likely to affect post-entry variable profits and the fixed-costs of entry, including the nuber of cities connected to the airports that serve the end-point cities and the nuber of carriers in the end-point countries; the presence of a ajor airline alliance, A, ay affect arket structure by enabling an airline to use its partner s network to serve a arket; r, c, y and t are regional, end-point countries, year, and onthly fixed effects and denotes M rando arket effects correlated with the regressors; and is an error ter. In su, our odeling syste consists of three deand and three fare equations, an equilibriu capacity utilization equation, and a arket structure equation. We account for the coon paraeters that arise in the deand equations because the price index specified in equation (2) appears in the three equations and for the syetry condition in equation (5) that restricts the substitution pattern across segents. The deand equations are also nonlinear in paraeters because the price index is ultiplied by β g in equation (1) and by θ 1 in equation (6). III. Identification and Estiation We use the logic of Difference-in-Differences (DID) ethodology to identify the short run effects of open skies agreeents on arket outcoes because a control group of arkets exists whose regulatory status was unchanged; identification of the long run effects relies on 17

19 cross-sectional variation across arkets. One possible concern with this identification strategy is that OSAs involving the United States tend to be with countries that are ore developed than are other countries, which ay lead to an upward bias in the effects of OSAs. However, we hold constant the difference between the control group of arkets and the treatent group of arkets by including individual end-point country duy variables and observed country characteristics, such as population and incoe, in the specification. In addition, the full saple includes OSAs between countries that are not aong the ost developed English speaking countries. As we report later, the effect of OSAs on travelers fares is actually soewhat larger for the full saple than for the subsaple of US international routes in which the only OSAs involve the United States, which also casts doubt that the estiates of the U.S. OSAs are upward biased. Because open skies agreeents are negotiated at the country-pair level instead of at the route-level, we included duy variables for the end-point countries in the specification of all of the equations to control for oitted group effects at the country level. The estiates obtained fro specifying end-point country duy variables are equivalent to those obtained by specifying country-pair duy variables when we restrict the saple to include only U.S. international routes, which is the basis for our policy siulations. However, when we perfor estiations using the full saple, it is possible that the end-point country duy variables ay not control fully for the effect of free trade agreeents between two countries on fares, arket structure, and capacity utilization and that it would be preferable to specify country-pair duy variables to control for that effect. So, we checked the robustness of our findings by replacing the end-point country duy variables with the country-pair duy variables in our odel and we found that the estiated effects of the OSAs hardly changed. This ay be because the 18

20 regional duy variables that we include, such as for the EU and North Aerica, also capture the effect of free trade agreeents. The rando effects regression equations in our odel can be expressed in general for as: (10) y X B1 WB 2 c T, where y is a dependent variable in arket at tie t that we seek to explain; c represents the individual arket effects that we odel as rando; T is the rando arket trend specified in the top-level deand and arket structure equations; shocks; is the idiosyncratic rando X is a vector of tie-varying regressors that are allowed to be correlated with rando coponents; and W is a vector of exogenous tie-invariant regressors that include route distance, regional, end-point country, year, and onth duies in ost equations. This rando effects specification is siilar to Hausan and Taylor (1981). Challenges to identification arise in our odel because we allow the rando arket effects to be correlated with regressors and because the endogenous variables, passenger deand, fares, flight frequency, arket structure, and regulatory status, are also specified as explanatory variables. When a regressor in c, and ( T X is correlated with the three rando coponents ), we use the deeaned first-order difference of z as its instruent, where z is a variable uncorrelated with and the process of deeaning and first-order differencing reoves its correlation with the rando arket effects and the rando arket trend. When a regressor in X is correlated with c and/or T but not with, deeaning and/or firstorder differencing this variable leads to a valid instruent. 19

21 In soe cases, certain variables can serve as instruents ( z ) because we hold other variables in the specification constant. For exaple, the deeaned incoe and population of the origin and destination countries are valid instruents for arket expenditures in the botto-level deand equation because relative segent prices and expenditures are held constant. Potential siultaneity bias is therefore avoided because fare class choice is not affected by changes in incoe and population. As another exaple, the deeaned total bilateral trade value is a valid instruent for the short-run regulatory status duy variable in the capacity utilization equation (7) and in the fare equation (8) because the nuber of arket passengers is held constant. We suarize the estiable equations, the endogenous variables in those equations, and their instruents in table 3 and discuss our identification strategies in detail in the appendix. We explored first-stage regressions for each equation by regressing the endogenous variables on the instruents and we found that all coefficients were statistically significant, indicating that the instruents are correlated with the endogenous variables. The ain results of those regressions are presented in appendix tables A1 A3. Later we present robustness tests of our estiated odels based on alternative approaches to constructing the instruents. Turning to estiation, let Z denote a vector of instruents including both exogenous regressors and instruents for the endogenous regressors; thus, the regression equations in our odel are identified by the ean independence condition Ec T Z 0. We estiate the paraeters of the odel by Generalized Method of Moents (GMM), which eploys the orthogonal conditions iplied by the ean independence condition as oent functions and accounts for the correlations within an equation that arise fro our rando arket effects specification. Accordingly, identification and estiation of the odel do not rely on any distributional assuptions for the error ters. 20

22 We could further iprove estiation efficiency by accounting for the conteporaneous correlation of the errors across equations but we found that it was not coputationally feasible to siultaneously estiate the large nuber of paraeters that resulted fro specifying eight equations that each included regional, end-point country, year, and onthly duy variables. 11 We therefore estiate the three deand equations jointly to account for their coon paraeters and cross-equation constraints given in equations (3) (5), and then estiate the reaining equations individually. It turns out that the ain paraeters of interest are estiated precisely and that the additional gain fro estiating the eight equations jointly is likely to be sall. We provide a foral presentation of the estiation procedure in the appendix. IV. Estiation Results The United States has negotiated any open skies agreeents with other countries and could negotiate even ore in the future; thus, our ain objective is to estiate travelers benefits fro open skies agreeents on U.S. international routes. We do so by first estiating our odel using a subsaple that contains only U.S. international routes because it is appropriate to use paraeter estiates obtained fro that subsaple to perfor the welfare calculations. As noted, we do not extrapolate our findings to all U.S. international routes and raise the possibility of selectivity bias. However, given the full saple is ore representative of international airline arkets, it is iportant to also estiate our odel using the full saple to check the robustness of the paraeter estiates obtained fro the subsaple. 11 Joint estiation is further coplicated because the arket deand and arket structure equations use lagged variables as instruents and are therefore estiated using subsaples of the full data set. 21

23 GMM paraeter estiates of the deand, capacity utilization, fare, and arket structure equations for both saples are easier to digest if we report the in separate tables. Tables 4 and 5 present the top and botto level (expenditure share) deand equations. The overall price elasticities of air travel deand, conditional on a fixed nuber of flights, are for the full saple of international routes and for the subsaple of U.S. international routes. We copare the unconditional deand elasticities, reported later, with those in the literature. Generally, the other paraeter estiates in the two saples are of siilar agnitude and have the expected sign: a greater nuber of flights increases deand with an elasticity between 0.26 and 0.18; distance has a positive effect on deand, reflecting air s speed advantage over other odes, which reduces travel tie costs for longer distance trips; 12 and deand is stiulated by an increase in the ean population and incoe per capita of the end-point countries. The population and incoe elasticities on U.S. international routes are greater than those on the full saple of routes, in all likelihood because copared with populations in any other countries, a saller share of the U.S. population travels abroad so a change in population will yield a larger deand elasticity and because copared with the United States, other countries have a uch lower per capita incoe so any residents still cannot afford air travel even if their incoe rises. Most of the paraeter estiates in the expenditure share equations shown in table 5 are precisely estiated. Their agnitudes and the differences between the saples are clearer when we use the below to calculate the own and cross-price elasticities of segent deand. The cointegrating vector between arket passengers and the nuber of flights enabled us to construct the short-run fluctuations in equilibriu capacity utilization, which we included in 12 Distance s coefficient is for the average distance in our saple. The effect of distance on deand is likely to weaken for longer distances because alternative odes to air transportation are not viable. 22

24 the airline fare equations. We can estiate the cointegrating vector and other paraeters in the equilibriu capacity utilization equation (7) by regressing log K on Q log and the other explanatory variables. The estiation results presented in Table 6 include specifications in the first and third coluns with passenger deand as the only regressor and specifications in the second and fourth coluns that also include the other regressors. The estiated signs of the regressors are plausible as population, incoe, nuber of carriers, and the nuber of cities connected to the endpoint airports have a positive effect on the nuber of flights, while distance, rainfall, and teperature differences at the endpoint airports have a negative effect. The negative effect of distance indicates that as routes becoe longer, airlines find it ore efficient to operate larger planes with lower frequency than to operate saller planes with greater frequency. In both saples, open skies agreeents increase the nuber of flights in the short run and have a statistically insignificant effect in the long run, while the presence of an alliance has a statistically insignificant effect. We use the paraeter estiates fro the full saple (colun 2) to predict the residuals so that we can test whether passengers and flights are in fact cointegrated (the value of the cointegrating vector was robust to the alternative specifications). We define: (11) eˆ logk logQ, and ipleent the Augented Dickey-Fuller (ADF) test of cointegration by estiating the following fixed-effects odel: where eˆ eˆ (12) j j X ~ p e e ~ eˆ 1 B b, j1 is a vector of regressors including arket characteristics such as population, percapita-incoe, nuber of carriers, and year and onth duies. Under the null hypothesis that X ~ 23

25 the log nuber of flights and the log nuber of arket passengers are not cointegrated, the paraeter 0; 0 under the alternative hypothesis that those two series are conintegrated. We report OLS estiates of based on alternative specifications in the appendix table A4 for the full saple and also for the saple of U.S. international routes (using the paraeter estiates fro colun 4, table 6). All of the coefficient estiates are less than zero and have large t-statistics, providing strong epirical support for the hypothesis that arket passengers and the nuber of flights are cointegrated in both saples. The responsiveness of segent deands, arket deand, and the nuber of flights with respect to a change in segent prices, overall price, and onthly flights can be calculated only nuerically; we describe our approach in the appendix. As shown in table 7, the own-price elasticities for both saples have the correct negative sign and their agnitudes iply that travelers in the full saple who fly first class, business, and full econoy are ore responsive to fare changes in their segents than are travelers who fly discount econoy, in all likelihood because air fares constitute a larger share of the total cost of their trips and because fewer people fly in those segents so a given change in price will produce a larger elasticity. Those factors ay also explain why travelers on U.S. international routes who fly full econoy have the greatest response to fare changes in their segent. The unconditional price-elasticities of arket deand are, as expected, larger than the conditional elasticities obtained previously fro the toplevel deand odel because they account for the change in the nuber of flights. Their agnitudes, for the full saple of international routes and for the subsaple of U.S. international routes, are bounded by the ean price elasticity for business travelers, -0.27, and the ean price elasticity for pleasure travelers, -1.04, that are reported in an extensive survey of air travel deand elasticities by Gillen, Morrison and Stewart (2003). 24

26 Generally, the cross-price elasticities indicate, as expected, that an increase in the fares of one segent increase the deand in the other segents and decrease total deand and onthly flights. An increase in the overall price reduces segent deands and flights. And increases in onthly flights increase segent deands and total deand. The effects of the open skies duies on fares are of particular iportance to our analysis. As shown in table 8, the initial (short-run) effect of an open skies agreeent is to reduce fares approxiately 50 percent or ore in the full saple and approxiately 25 percent or ore in the subsaple of U.S. international routes; its additional long-run effect is to reduce fares approxiately 15 to nearly 30 percent in the full saple and 20 percent or ore in the subsaple of U.S. international routes. 13 Note those estiates hold the nuber of carriers constant when, in fact, open skies agreeents ay enable ore carriers to enter and copete on a route, which according to the paraeter estiates in the table would further decrease fares in each segent, although the effects on full econoy fares in both saples are not statistically significant. We indicated in the introduction that it is iportant to distinguish between the effects on fares of open skies agreeents and an airline alliance between two international carriers. Our specification controls for the effect of an alliance (with antitrust iunity) on fares; thus, our finding that open skies agreeents reduce fares cannot be partly attributed to the presence of alliances. Consistent with previous research, we also find that an airline alliance (with antitrust iunity) on a route generally lower fares. The coefficients, indicating fare reductions of 10 percent to 25 percent, are uch larger and ore precisely estiated for the subsaple of U.S. international routes than for the full saple, which ay reflect the relative effectiveness of 13 We use the ter approxiately because the exact effect of a duy variable in a log linear equation is given by 1-exp(-COEFF), where COEFF is the coefficient of the duy variable. 25

27 alliances that involve a U.S. carrier. We have also noted that we do not account for traffic beyond the origin and destination airports, which ay account for the sall effects of alliances on fares in the full saple. In su, although open skies agreeents reduce travelers fares ore than airline alliances do, the policies are related because the granting of antitrust iunity to a U.S. carrier and its foreign alliance partner is approved by the U.S. Departent of Transportation only if an open-skies agreeent exists between the United States and the foreign partner s country. The reaining paraeter estiates have the expected sign: fares are increased by greater passenger deand, given aircraft and airport capacity constraints, longer distances, higher oil prices, greater rainfall that causes delays and increases operating costs, and larger teperature differences between January and July at the end-point airports that indicate higher operating costs during winter. Fares are reduced by an increase in capacity utilization, although the effect is not precisely estiated in the subsaple of U.S. international routes, possibly because load factors do not vary greatly over tie and across those routes. Finally, the arket structure equation (table 9) indicates that carriers in both saples adjust their networks after the countries at the origin and destination negotiate an open skies agreeent. In the short run, the nuber of carriers on a route falls because inefficient carriers are no longer protected by price and entry regulations and they are driven out by ore efficient carriers. In the long run, other carriers covered by the agreeent have sufficient tie to adjust their networks and take advantage of the opportunities to enter new international arkets thereby increasing the nuber of carriers on a route. Thus, open skies agreeents in the full saple and the subsaple of U.S. international routes have the direct effect of stiulating copetition that reduces fares in the short run and continues to reduce the in the long run and that increases 26