ABSTRACT

MULTIMARKET CONTACT ON TACIT COLLUSION: EVIDENCE FROM THE AIRLINE INDUSTRY

by Christopher Granquist

This paper examines the effect of multimarket contact on tacit collusion using empirical evidence from the airline industry. We replicate a model of multimarket contact applied to price and apply it to price dispersion, and address potential endogeneity in our measure of multimarket contact through fixed effects. We find three early results: (i) Increased concentration in the U.S. airline industry has reduced the effect of multimarket contact on prices, but the effect remains relatively strong and positive; (ii) The effect of multimarket contact on price dispersion may have either a positive or negative effect based on specification; (iii) First-stage regression results suggest that the percentage of gates owned by a carrier or their low-cost competition at the destination airport on a route is a strong instrumental variable for future extension. MULTIMARKET CONTACT ON TACIT COLLUSION: EVIDENCE FROM THE AIRLINE INDUSTRY

Thesis

Submitted to the

Faculty of Miami University

in partial fulfillment of

the requirements for the degree of

Master of Arts in Economics

by

Christopher Granquist

Miami University

Oxford, Ohio

2020

Advisor: Dr. Charles Moul Reader: Reader:

c 2020 Christopher Granquist

ii This thesis titled

MULTIMARKET CONTACT ON TACIT COLLUSION: EVIDENCE FROM THE AIRLINE INDUSTRY

by

Christopher Granquist

has been approved for publication by

Farmer School of Business

and

Department of Economics

iii Contents

1 Introduction 1 1.1 Introduction ...... 1

2 Literature Review 2 2.1 Multimarket Contact ...... 2 2.2 Price Dispersion ...... 3 2.3 Contemporary Work ...... 3 2.4 Contribution ...... 4

3 Data 4 3.1 Data Sources ...... 4 3.2 Observations ...... 5 3.3 Dependent Variables ...... 6 3.4 Control Variables ...... 7 3.5 Multimarket Contact ...... 8 3.6 Instrumental Variables ...... 10

4 Results 11 4.1 Replication ...... 11 4.2 Price Dispersion Results ...... 12

5 First-Stage Regression Analysis 14 5.1 Instrumental Variables ...... 14 5.2 Results ...... 16

iv List of Figures

1 Summary Statistics Table ...... 19 2 Contact Between Airline Pairs, Q1 2014 ...... 20 3 Comparison of Average Contact Summary Statistics ...... 20 4 Log Price Regression Results Table ...... 21 5 Comparison of Average Contact Results ...... 22 6 Gini Coefficient Regression Results Table ...... 23 7 Within-Obs Std. Deviation Regression Results Table ...... 24 8 First-Stage Regression Results Table ...... 25

List of Tables

v Dedication

vi Acknowledgments

vii 1 Introduction

1.1 Introduction

Unilateral market power is a well-established theoretical and empirical phenomenon. In the absence of unilateral power, however, it is possible for multiple firms to collude as a cartel to increase market power. The purest mechanism of collusion would be legally-binding collusive agreements, but the real world does not favor legal collusion in markets. Instead, firms may engage in tacit collusion by enacting a set of credible strategies that allow firms an avenue of punishment to those who break a tacitly collusive equilibrium. Mechanisms that allow tacit collusion amount to threats such as price wars. The determinants of tacit collusion is a relatively recent chapter in industrial organization literature, especially concerning firms operating in a single market. However, real world firms operate across many markets, and the resulting multimarket contact that arises allows for new avenues of punishment to enable tacit collusion. Generally speaking, market power creates market inefficiencies. The specific effect of multimarket contact creating market power through tacit collusion may create these inef- ficiencies through multiple routes. The literature finds empirical evidence for the effect of multimarket contact in allowing firms the power to raise prices in a market. Instead of statically increasing prices, tacit collusion may also enable firms to engage in price discrimi- nation. If price discrimination does not increase the quantity of goods sold in a market, then it creates market inefficiencies primarily in the terms of decreased welfare in the market, but also induces a shift of existing welfare from consumers to producers. If price discrimination does exist in a market, it would come in the form of price dispersion, which would indicate different prices being charged to different consumers in the market. The airline industry is a favored industry of industrial organization due to publicly avail- able data and well-defined markets. Evidence for the effect of multimarket contact and its role in facilitating collusion begins with Evans and Kessides (1994), who use airline data to find evidence of a positive effect of multimarket contact on prices. At the same time, Borenstein and Rose (1994) use the airline industry to investigate the effects of competi- tion on price dispersion. We believe it is a natural extension to use airline industry data to examine the effect of multimarket contact on price dispersion as a mechanism of tacit collusion, since strong empirical evidence exists connecting the effect of multimarket contact as a specific mechanism of competition on prices as an effect of tacit collusion, and empirical evidence exists connecting the effect of competition within a market on price dispersion as a mechanism of tacit collusion. We replicate models from the literature, and then apply the models to price dispersion as an independent variable.

1 2 Literature Review

2.1 Multimarket Contact

The microeconomic theory of multimarket contacted was formalized by Bernheim and Whin- ston (1990). Bernheim and Whinston produce an irrelevance result for multimarket contact; competition across markets would not affect firm strategy under assumptions of identical markets, identical firms, and constant returns to scale. In the absence of these assumptions, competition across markets may have an effect on firm strategies. Evans and Kessides (1994) find that the airline industry is an ideal industry to test for multimarket contact. Since city- pair markets are well-defined in the airline industry, there are no ad-hoc assumptions made by the researcher. The hub system for airlines encourages airport dominance, which differ- entiates production costs across markets. Firms also sort into national, regional, or low-cost carriers, with varying access to capital and production costs across markets. There is also significant discretized returns to scale as the cost to fly a plane faces only marginally increases fuel costs between flying a single passenger or flying at full capacity. Evans and Kessides (1994) then test the hypothesis that prices in the airline industry are affected by multimarket competition using panel data from 1984 to 1988. EK use an OLS regression on log prices on a suite of controls and a constructed variable to measure multimarket contact, controlling for fixed effects. The multimarket contacted variable con- structed in EK (1994), labeled AverageContact, is a monumental construction in future works addressing multimarket contact in the airline industry. The results conclude that empirical evidence upholds the BW theory of multimarket contact and tacit collusion, such that markets whose firms compete in a number of other markets find significantly higher prices overall as compared to markets whose firms face a low amount of multimarket con- tact. Evans and Kessides also find that market fixed-effects have a significant negative effect on the magnitude of multimarket contact. CW (2014) reexamine multimarket contact in the airline industry by revisiting EK (1994). Using updated data from 2004 to 2007 combined with survey data concerning gate leases, CW (2014) explore structural parameters of tacit collusion. The results find that firms with little multimarket contact do not collude and marginal changes in multimarket contact are only significant for firms at low levels of contact, as well as reinforcing the previous results that assuming Bertrand-Nash competition for firms will lead to biased results. Relevant to our paper, they replicate EK (1994) reduced-form analysis and Ciliberto and Williams suggest that the AverageContact variable as constructed by EK is itself endogenous at the beginning of their research. Ciliberto and Williams also explore a wider variety of control specifications as well as using number of gates leased by a firm and their competition at airports in the market as an instrumental variable for the AverageContact regressor.

2 2.2 Price Dispersion

Concurrently to Evans and Kessides in 1994, Borenstein and Rose (1994) examine the airline industry. Instead of analyzing mutltimarket contact, however, Borenstein and Rose investi- gate price dispersion in the airline industry. BR (1994) uses the Gini coefficient (defined in Section 3.4) as a measure of price dispersion. The results find that price dispersion cannot be entirely explained by cost variation, and instead that price discrimination is driving a significant effect of price variation. BR further explain possible sources of price dispersion, emphasizing that price discrimination may increase in the face of increasing competition. Borenstein and Rose also use an OLS regression of Gini on a suite of controls as well as several variables to measure competition, and then instrument the competition variable. The result of BR is a finding of a positive association of competition within a market to price dispersion, of which a significant portion of the price dispersion is attributed to price discrimination. GS (2009) revisit Borenstein and Rose, using panel data and replicated cross-sectional data to find contrasting results to BR (1994). Similar to BR (1994), Gerardi and Shapiro use an OLS regression of Gini on controls and various measures of competition, and then instrument the competition variable. Geradi and Shapiro’s results suggest that the actual effect of competition on price dispersion is negative, and that the difference between BR (1994) and GS (2009) may be reconciled by omitted variable bias in the cross-sectional work of Borenstein and Rose, specifically the effect of omitting distance, may cause a significant positive bias to the effect of competition on price dispersion.

2.3 Contemporary Work

Several working papers have been recently released addressing price dispersion using Evans and Kessides multimarket contact model. Chiang and Liou (2018) explicitly examine the effect of contact across markets on price dispersion in the airline industry using the EK (1994) average contact variable as a regressor to the Gini coefficient as used in BR (1994). Chiang and Liou find results varying based on market size, and find that larger markets are more susceptible to increases in price dispersion from multiple sources as well as finding that multimarket contact and price dispersion behave in opposite directions. Kim, Kim, and Tan (2019) produce a similar effort, explicitly using EK (1994) AverageContact model to explain the Gini coefficient as a proxy for price dispersion. Kim, Kim, and Tan demonstrate specific interest to the effect of , and find that when South- west Airlines is present on a route, multimarket contact fails to have a significant effect on price dispersion. However, in the absence of Southwest, price dispersion decreases in a market as firms in the market exhibit more competition across markets.

3 2.4 Contribution

We contribute to the literature by examining the effect of multimarket contact on price dispersion. Currently, the literature is weak in this field, and only recently has empirical research began. Most current literature finds insignificant results, and we contribute stronger evidence. Furthermore, the literature uses Evans and Kessides (1994) multimarket contact measure, without addressing the endogeneity concerns as raised by Ciliberto and Williams (2014). We therefore find stronger causality of our results through use of these instrumen- tal variables. As a secondary contribution to the literature, we also provide more context for Ciliberto and Williams (2014) reduced-form results by providing tests for instrumental variables to formally suggest that the instrumental variables used are good instruments.

3 Data

3.1 Data Sources

We draw the majority of our data from the Bureau of Transportation Statistics’ DB1B database. The DB1B is a 10% random sample of flight itineraries from reporting carriers. A flight itinerary is a schedule of airline transit for a consumer, including data on origin, destination, layover airports, carriers, passengers, market fare, and other parameters of in- terest. Reporting carriers may be either operating carriers, which provide the actual service of the flight, or ticketing carriers, which perform the itinerary transaction to the consumer. Operating carriers may differ from ticketing carriers as contracted regional airlines that do not ticket consumers directly but instead fly under aliases for multiple national carriers. Skywest Airlines, for example, has contracts with , , Delta Airlines, and Alaskan Airlines. We note that the DB1B itinerary information comes from reporting carriers, which arbitrarily may be either operating carriers or ticketing carriers. We also construct original measures for absolute distance between airports, status by airport, and airline gate leasing information by airport. Hubs are used by airlines to concentrate operations, and the hub status of an airport is designated by carriers. Gate leasing data are originally constructed. When available, we observe the total number of gates at an airport, number of common-use gates at an airport, and number of gates leased to each airline at an airport. We use information from airport competition plans submitted to the Federal Aviation Administration (FAA) combined with information obtained directly from airports. The Wendell H. Ford Aviation Investment and Reform Act for the 21st Century, colloquially known as AIR21, mandated that medium and large hubs, as defined by the FAA, are required to submit competition plans to the FAA. These competition plans include information on gate leasing procedures, airport expansion, access to common-use

4 gates, and other leasing information. An update letter to the FAA must be sent by each competition-plan-covered airport annually, and new plans may be triggered on condition. These federal regulations help ensure accuracy of our gate leasing data as drawn from airport information. We note that competition plans are not readily available from the FAA, and must be collected individually. Furthermore, competition plans are erratically located and available, making it realistically impossible to scrape for competition plan data, let alone find consistent annual entries. Due to information constraints, we only observe one instance of data for each airport and airline across all time periods. We reference the Government Accounting Office (GAO, 1990) report, as used by CW (2014). This report states that 41% of gate leases are for over twenty years of duration, 25% for eleven to twenty years of duration, and 22% for three to eleven years of duration. Berry (1990) suggests that airports sign long-term leases to aid in capital investment while maintaining low-interest on debt issues, and that airlines sign long-term leases to integrate the airport into a network. We also note that it is difficult for airlines to adjust airport leases as termination is bilateral and subleasing is expensive due to imposed limits by airports (Ciliberto and Williams 2010). As an example of the inability of unilateral termination, Ciliberto and Williams describe an incident where, prior to 2010, Dallas Love airport declined an attempt by American Airlines to terminate a gate lease early, and the latter was forced to maintain payments until lease expiration in 2011. The limited access to modifications to airport capital combined with the limit on sublease fees discourage airlines from subleasing out to competitors. We finally note that gate capital is resistant to demand shocks in a single market, as single markets rarely generate a significant portion of the revenue from the airport as compared to the value of other markets in the airline’s network at that airport.

3.2 Observations

We use the DB1B to generate observations indexed by time, firm, and market. We define a market as a unidirectional route between two airports. Hence, Chicago O’Hare (ORD) to Atlanta (ATL) would be considered one route, and ATL to ORD would be considered a distinct route. After filtering data, some markets are dropped while the converse market is not dropped, which drives some portion of variation in our data. We observe data quarterly from 2014 to 2018, generating twenty periods. We index quarter-yearly periods by t ∈ {1,...,T }, with T = 20. We finally observe twenty-two airports. We are limited to these airports as a restriction of major or large hubs with recently updated competition plans as so that gate leasing information observed may be reliable. Our definition of a market as a unidirectional route between two airports therefore generates 484 possible markets. However, twenty-two of

5 these markets are same-city pairs, such as ORD to ORD, and we exclude uninteresting data. Generally, we index market by m ∈ {1,...,M}, with M = 484. Therefore, an observation

is at the airline-city-pair-quarter level, indexed jmt. After filtering by observations with nonzero prices, distances, and within-observation standard deviations, we generate 49877 observations. We list our summary statistics for all generated variables in Table 1. We observe thirteen significant carriers across our observations. We use carrier, airline, and firm as interchangeable notation in this paper. We drop airlines included the in DB1B that do not report at least 1000 passengers in at least one period. Our observed airlines are Alaskan Airlines (AS), American Airlines (AA), Delta Airlines (DL), (F9), (YV), Shuttle America (S5), SkyWest Airlines (OO), Southwest Airlines (WN), (NK), (SY), Trans States Airlines (AX), United Airlines (UA), and US Airways (US). We note that through 2014, US Airways was in the process of being merged with America Airlines, and therefore we do not have reliable gate leasing information on US Airways, nor do they appear in the majority of our observations. We identify the national carriers as AA, DL, and UA; the low-cost carriers as WN, F9, NK, and AX; and the regional carriers as AS, OO, SY, S5, and YV. We index firm by j ∈ {1,...,J}, with J = 13.

3.3 Dependent Variables

We create three dependent variables of interest. We first create prices as a replication effort for EK (1994) and CW (2014). We then create two measures of price dispersion: the Gini Coefficient, as used by BR (1994), and within-observation standard deviation. The DB1B observes market fare by itinerary. As per CW (2014), we discard extreme- priced observations below $25 and above $2500. We also halve the listed market fare for round-trip itineraries. We then calculate the simple average of market fares indexed by firm- market-period. We then deflate fares by the Consumer Price Index to 2014$ standard, and define this as our measure of prices as a replication of EK (1994) and CW (2014). P rice has a mean of 171.56, a standard deviation of 88.49, and a median of 152.42 in our data. Our primary topic of interest in this paper is examining the effect of multimarket com- petition on price dispersion. Hence, we similarly generate a Gini Coefficient at the firm- market-period level using market fares as observed in the DB1B. The Gini Coefficient is mathematically defined as twice the area between the 45-degree line of absolute equality and the Lorenz Curve when sorting by prices as a function of passengers. The Lorenz Curve, in this context, is the cumulative amount of fares paid by the lowest x% of passengers. The Gini Coefficient is bounded between zero and one, with zero being perfect inequality in a market and one being perfect equality. In context, a zero Gini would represent all fares in a market being paid by a single passenger, and converse a one would represent every passenger

6 paying the same fares. Therefore, as described in BR (1994), we use the Gini Coefficient as a measure of price dispersion. This measure is equivalent to twice the expected absolute difference between two randomly drawn ticket prices (Geradi, 2009). Gini has a mean of 0.206, a standard deviation of 0.093, and a median of 0.225. As another measure of price dispersion, we construct within-observation standard devi- ations. Given that the DB1B is a 10% sample of flight itineraries, we may observe multiple itineraries for each of our carrier-city-pair-quarter level observations. Hence, we may expect to see variation in the market fare of each itinerary, and hence can construct the standard deviation of all market fares indexed by firm, market, and period. We define this standard de- viation of market fares across itineraries in a working observations as the within-observation standard deviation. It operates simply as an alternative measure of price dispersion of one observations similar to the Gini coefficient, and we expect both Gini coefficient and within- observation standard deviation to behave in a similar way. We denote within-observation standard deviation as sd, and it has a mean of 114.06, a standard deviation of 77.37, and a median of 107.14.

3.4 Control Variables

We construct a number of control variables from EK (1994) and CW (2014). These controls include distance, round-trip percentage, direct-flight percentage, hub, route market share, airport market share, HHI, and network size. Distance is the absolute linear distance between two airports. Distance is an important component of consumer travel decisions. Distance may have a nonlinear relationship, how- ever, as air travel becomes an increasingly attractive mode of transport relative to alternative forms of travel as distances increase. Therefore, we follow the literature and also include a measure of distance squared Distance2, to account for this non-linearity. Distance has a mean of 1659.2, a standard deviation of 1215.1, and a median of 1198.7. RoundT rip is the percentage of passenger itineraries which are round-trip. A round trip is an itinerary in which the flight and return flight are purchased together. We include the fraction of flights in one firm-market-period observation that are round trip. Direct is the percentage of itineraries which are direct flights. A direct flight is an itinerary that flies straight between origin and destination airport without any layovers. We include the fraction of flights in one firm-market-period observation that are direct. Hubs are airports that are central to a network for an airline. Hub status is designated by airlines; we construct our measure of hubs by airline designation. Since a market is a unidirectional city pair with both an origin and destination airport, we construct three measures of hub. All measures are at the firm-market-time level. HubOrigin is a dummy variable that has a value of one if and only if the origin airport is a hub for the specified

7 airline, and similarly HubDest has a value of one if and only if the destination airport is a hub. HubEither is a dummy variable with a value of one if either the origin or destination airport is a hub. We note that HubEither is the control used by CW (2014). HubEither has a mean of 0.263, a standard deviation of 0.440, and a median of 0. Note that a median of 0 is reasonable in context, as any given airport is not likely to be a hub for any given carrier. RouteMktShare is a simple measure of route market share observed at the firm-market- time level. We define it as specified in EK (1994) and CW (2014) as the percentage of quantity of passengers served in one market in one period by an airline. RouteMktShare is a simple measure of competition across routes, and is similar to the measures of competition used by Borenstein and Rose (1994). Similarly, HHI is the Herfindahl-Hirschman Index (HHI) of one market in one period. The HHI of a market measures the level of concentration in a market. It is mathematically formulated as the sum of the square of market shares of all firms in the market. We calculate shares as 100 times the fraction of quantity of passengers served by an airline relative to the total amount of passengers served in the market. Hence, HHI is bounded between 1 and 10,000, where 1 is perfect competition and 10,000 is monopoly in the market. RouteMktShare has a mean of 0.207, a standard deviation of 0.267, and a median of 0.077. We also construct a measure of market share at an airport. Similar to the hub variables, we generate three controls of airport market share: share at the origin airport (AptShareOrigin), share at the destination airport )AptShareDest), and average share between both endpoint airports, (AptShareAvg). The airport market share for an airport is calculated as the frac- tion of passengers served by an airline in a period relative to the total number of passengers, and hence is observed at the firm-market-period level. EK (1994) use AptShareAvg as their measure of airport market share. In our data, AptShareAvg has a mean of 0.152, a standard deviation of 0.038, and a median of 0.119. Network is a measure of the firm’s network size for the origin airport. This is measured as the fraction of routes an airline serves at an airport relative to the total possible routes at the airport. Network has a mean of 0.614, a standard deviation of 0.311, and a median of 0.614.

3.5 Multimarket Contact

Evans and Kessides (1994) create a measure of average route contact as an attempt to capture the amount of cross-market contact of all firms operating in a specific market. We

label this measure AverageContact. For any given market m in period t, there are fmt firms

operating in the market. Let Djmt be a dummy variable that has a value of one if firm j is

operating in market m at time t. We construct a matrix At = (ajkt) for each market-time

8 pair, with M X ajkt = DjmtDkmt for j, k ∈ {1,...,J} m=1 Thus each element in matrix A is the number of markets concurrently served by airlines j and k in market m and period t. The diagonal of this matrix is the total number of routes served by airline j in market m at time t, since j = k. From this, we construct our AverageContact variable.

J J 1 X X AverageContactmt = ajkmtDjmtDkmt [fmt(fmt − 1)]/2 j=1 k=j+1

This measure represents the average amount of contact in market m across all markets between any two firms operating in the market m at time t. We note the properties of this measure as described in EK (1994). A monopolistic route will have an average contact value of zero. AverageContact is also positively correlated with firm size, since large firms will operate across more markets. It also only measures the exposure of airlines in the market to other airlines by route presence, not magnitude of the exposure. Table 2 provides an illustration of the amount of contact between firms in first quarter 2014. To provide an example for intuition, suppose the ORD to ATL route is served by only American Airlines, United Airlines, and Southwest Airlines. Suppose American and United compete on 300 routes, American and Southwest 200 routes, and United and Southwest 100 routes. Then our measure of multimarket contact would be calculated as: 1 AverageContact = (300 + 200 + 100) = 200 [3(3 − 1)]/2

We note that AverageContact is a difficult measure to interpret at a glance. It roughly interprets as the average amount of contact between any pair of firms in a market and period. Hence, before scaling, a one unit increase in AverageContact may be effectively interpreted as the average number of routes that any two firms share in a market increasing by one. Table 3 provides a comparison on summary statistics of AverageContact as compare to EK (1994) and CW (2014), who both use the same measure of multimarket contact as in this paper. We find in our data a mean of AverageContact of 0.19, a median of 0.18, and a standard deviation of 0.07 when scaled to the level of EK (1994). Ciliberto and Williams (2014) discuss the endogeneity behind prices, market shares, and average contact. Since we consider only a reduced-form analysis by regressing our depen- dent variables directly onto average contact, our endogeneity concern is of average multimar- ket contact. Ciliberto, Murry, and Tamer (2018) observe that market-specific multimarket

9 contact may be endogenous due to unobservables correlated with pricing, entry, and exit decisions. Since for any given time the average contact across all firms is fixed, variation across markets originates from the set of operating firms in a market. Variation within a market across time is driven by a change in set of operating firms and a change in multi- market contact. This variation comes from market structure and it may be correlated with the unobservables mentioned prior (Ciliberto 2014). Griliches and Mairesse (1995) suggest fixed effects perform poorly when endogeneity may is driven at a market and time level, CW suggest that instrumental variables should be used to address the possible endogeneity present in AverageContact.

3.6 Instrumental Variables

As per Ciliberto and Williams (2014), we intend to use the access to gates at an airport as an instrumental variable. We construct OwnGatesOrig as the percentage of gates leased by a carrier at the origin airport, OwnGatesDest as the percentage of gates leased by a carrier at the destination airport, and OwnGatesAvg as the average percent of gates leased by a carrier at the origin and destination airports. We also construct LCCGatesOrig as the percentage of gates leased by low-cost carriers excluding the specified carrier at the origin airport, LCCGatesDest as percentage of gates leased by low-cost carriers excluding the specified carrier at the destination airport, and LCCGatesAvg as the simple average of percentage of gates owned by low-cost carriers excluding the specified carrier at the origin and destination airports. As mentioned in our data sourcing, we find that gate leases are resistant to shocks, subleases, and termination, and therefore are unlikely to change over the course of our sample despite only being observed at one period. Summary statistics for instrumental variables are described in Table 1. Ciliberto and Williams use only the average percentages of gate leases at both endpoints. While the qualitative concept backing percentage of gate leases is valid, they fail to express the power of the instrumental variables. Without first-stage regressions or power tests for the IVs, it is difficult to justify why a specific suite is chosen. We explore the strength vari- ous combinations of instrumental variables and the resulting effect on multimarket contact to validate the formal power of the instrumental variables, as well as to select the most appropriate instruments to use.

10 4 Results

4.1 Replication

In order to determine the effect of multimarket contact on price dispersion. For this paper, we are only interested in reduced-form analysis to find a direct impact of multimarket contact on price dispersion. Hence, we cite the Evans and Kessides (1994) model for reduced-form price equation.

ln(P ricejmt) = AverageContactjmt ∗ β + Controls ∗ ~γ + aj + bm + ct + ujmt with β as our coefficient of interest for our key regressor AverageContact, ~γ as the vec- tor of coefficients for our controls, fixed-effects denoted aj, bm, ct for firm, market, and pe- riod fixed-effects respectively, and idiosyncratic error ujmt. We expect the coefficient of AverageContact in this model to be positive, as we take the empirical evidence found in the literature as our null (EK 1994, CW 2014). Although we are not interested in the effect of multimarket contact on price in the scope of this paper, we are interested in an empirically validated model. As to not blindly apply this model to our data without precaution, we replicate the reduced-form fixed-effects models of EK (1994) and CW (2014). Finding similar results within tolerance of changes over time to the airline industry would allow us to validate our independence assumptions of our AverageContact measure. Hence, we create five specifications of the control vector to test against five relevant specifications from the literature, all within the same reduced- form model. All results are robust to heteroskedasticity via White-corrected standard errors. We note that EK (1994) makes no robustness checks, and CW (2014) correct with White standard errors on fixed-effect regressions. We illustrate these results in Table 4, with each column a specification of control variables as follows. Specification 1 is a replication of the OLS regression from EK (1994). We also control for firm and period fixed-effects. Specification 2 is a replication of the main fixed-effects regression from EK (1994). We also control for firm, period, and market fixed-effects. We find a significant coefficient of AverageContact of 0.173 with a standard error of 0.037. Specification 3 is a replication of the fixed-effects replication regression from CW (2014). We find a significant coefficient of AverageContact of 0.163 with a standard error of 0.037. Specification 4 is a replication of the OLS regression from CW (2014). We find a significant coefficient of AverageContact of 0.159 with a standard error of 0.037. Specification 5 is a replication of the OLS regression from CW (2014). We find a significant coefficient of AverageContact of 1.155 with a standard error of 0.033. We acknowledge that this fixed-effects model alone may be insufficient for causality, even

11 as we take our set of controls alongside fixed-effects as a plausible host of variables that may control for the independence of our AverageContact measure. Since Ciliberto and Williams (2014) address possible endogeneity in the construction of AverageContact arising from variation in market structure that may change over market and time, we cannot take our results as absolutely causal. The prior results provide entirely replication work to validate the model as suggested by the literature. Our early key results on price dispersion will suffer the same issue, but we find an indication of the direction of the effect of multimarket contact that matches our predicted hypothesis and the related literature. We use the replication results to directly compare the coefficients on AverageContact to the literature. Table 5 provides the comparison of the coefficient of AverageContact from our results to the literature we replicated. The replicated literature is Evans and Kessides (1994) for specifications 1 and 2, and Ciliberto and Williams for specifications 3, 4, and 5. We note that specification 5 is not directly comparable to the literature as CW (2014) use instrumental variables, whereas we do not here. We overall find a similar effect of multimarket contact as modeled by AverageContact on prices to the literature. We make several observations on these results. First, it appears that, when controlling for market fixed-effects, the effect of multimarket contact has decreased over time. We believe that this is due to increasing concentration in the airline industry. We note EK (1994) use 33 carriers, CW (2014) use 17 carriers, and we identify only 13 carriers. Furthermore, US Airways merges with American Airlines in late 2014, while Shuttle America is acquired by Mesa Airlines at the end of our sample. Hence, we believe that increasing concentration in the airline industry may result in a decreasing effect of multimarket contact not due to a decrease in collusion but instead due to decreasing variation in multimarket contact. As firms conglomerate, there are fewer routes that are not being served by any two airlines, and hence the contact between any two airlines may converge in variation.

4.2 Price Dispersion Results

Aside from the endogeneity of AverageContact, we now believe this model may be effectively applied to our data to consider the impact on price dispersion. Hence, we create two new models that only differ from our replication model in dependent variable. We first regress the Gini Coefficient onto our prior regressors

Ginijmt = AverageContactjmt ∗ β + Controls ∗ ~γ + aj + bm + ct + ujmt

We also use the log of within-observation standard deviation, denoted sd, as a dependent variable. We note that the Gini Coefficient is bounded between zero and one, where zero is perfect equality in fares and one is perfect inequality in fares. Comparatively, log standard

12 deviation may be any positive real number but with appropriate scaling. Zero still represents perfect equality in fares paid, but there is no cap on perfect inequality for standard deviation. Hence, we expect standard deviation and the Gini to work in similar directions. However, standard deviation is more easily intepretable, and hence we are interested in its effects as well. We also believe that linear regression more appropriate fits an unbounded variable, and hence we craft our model of standard deviation

ln(sdjmt) = AverageContactjmt ∗ β + Controls ∗ ~γ + aj + bm + ct + ujmt

For the ln(sd) model, we filter any standard deviation equal to zero. We are more interested in the effects of price dispersion on markets that may have price discrimination and cost variation, and hence discard perfectly equally priced routes as uninteresting. This allows us to use a log transformation, and accounts for the missing observations for this model. We generate the same set of specifications as in our replication work. We provide our results on Gini in Table 6 and ln(sd) in Table 7. All results our robust to heteroskedasticity via White-corrected standard errors. We present our results as follows Specification 1 has a significant coefficient of AverageContact on Gini of -0.035 with a standard error of 0.007. We also find a significant coefficient of AverageContact on ln(sd) of 1.000 with a standard error of 0.051. Specification 2 finds a significant coefficient of AverageContact on Gini of 0.057 with a standard error of 0.011. We also find an insignificant coefficient of AverageContact on ln(sd) of 0.128 with a standard error of 0.072. Specification 3 results in a significant coefficient of AverageContact on Gini of 0.042 with a standard error of 0.010. We also find an insignificant coefficient of AverageContact on ln(sd) of 0.112 with a standard error of 0.072. Specification 4 finds an insignificant coefficient of AverageContact on Gini of 0.025 with a standard error of 0.011. We also find an insignificant coefficient of AverageContact on ln(sd) of 0.043 with a standard error of 0.073. Specification 5 has a significant coefficient of AverageContact on Gini of -0.039 with a standard error of 0.006. We also find a significant coefficient of AverageContact on ln(sd) of 1.094 with a standard error of 0.049. We make several observations on our results. In every instance, both Gini and ln(sd) have the same direction and comparative magnitude, so we find both of our measures of price dispersion are appropriate at a surface level. We also not a large decrease in significance, especially when controlling for market-level fixed-effects. AverageContact demonstrated the same scaled-down magnitude for market fixed-effects, but it appears that the effect of AverageContact is weaker or insignificant overall. More interestingly, it appears that the direction of the coefficient of AverageContact is not stable. We explain this as omitted variable bias as conceived by Geradi and Shapiro (2009). Borenstein and Rose (1994) use cross-sectional data and find a positive estimate for

13 the effect of competition on price dispersion. GS (2009) reevaluate those findings with panel- level data and new instruments to find a negative effect instead. They attribute this difference to be omitted variable bias in cross sectional data attributed to not controlling for distance, a significant source of variation across markets. We note that our negative results control for distance and imitate GS (2009), while our positive results do not control for distance and imitate BR (1994). As we believe the mechanism for within-market competition would reflect the same mechanisms for multimarket competition, we believe our results are comparable to the literature. However, it is difficult to dissect the actual effect of the omitted variable bias, as it would make sense if distance effects were absorbed by the market-level fixed effects. Therefore, while we do not find any strong causal findings in these early results, we are aware of existing problems and begin to make correcting efforts.

5 First-Stage Regression Analysis

5.1 Instrumental Variables

Due to the endogeneity concerns present in AverageContact, we seek to address this endo- geneity with instrumental variables. We generate a similar instrumental database to Ciliberto and Williams (2014). Ciliberto and Williams create a suite of four instrumental variables: a carrier’s percentage of own gates, percentage of competition’s gates, percentage of South- west’s gates, and low-cost-carrier’s gates. All instrumental variables are calculated as the simple average of percentages at the origin and the destination airports. However, CW (2014) only qualitatively explains the instrumental variable. We believe that a simple average of gate percents at origin and destination airport may distort the effects of either the origin gates or destination gates. Furthermore, if both origin and destination are significant, a simple average may mute effect, and would limit variation in the data. Therefore, we run first-stage regressions of AverageContact onto various specifications of instrumental variables to test the strength of our suggested instrumental variables. Our instrumental variables are a carrier’s percentage of own gates at the origin airport, destination airport, or simple average of both, and low-cost-carrier percentage of gates leased excluding own-gates at the origin airport, destination airport, or simple average of both. Our low- cost carriers include Southwest Airlines, Frontier Airlines, Spirit Airlines, and Sun Country Airlines. We also note a slight difference in summary statistics of our origin and destination in- strumental variables. By defining markets as unidirectional routes and filtering out markets by nonzero prices, we drop some observations of markets while not dropping their converse. Hence, we exploit some variation in the data, while marginal, by a difference in origin and

14 destination gate numbers. We test the strength of instrumental variables on AverageContact by the following re- gression.

AverageContactjmt = β ∗ Instrumentsjmt + γ ∗ Controlsjmt + aj + bt + umt

Where Controls is a vector of Distance, Distance2, Direct, RoundT rip, and Network. Controls are chosen as to replicate the set of controls used by CW (2014) in their instru-

mented regression. We include fixed effects for firm, aj, and for time, bt. While AverageContact is calculated at the market-time level, we include an observations for each carrier in the market, as carrier’s differ in their amount of gate ownership at each airport. Similarly, while we only observe one period of observation for each instrumental variable, we do not believe that there is a significant change in the set of number of gates owned in an airport over time. Hence, we assume that the instrumental variable indexed by market and firm is identical for each period. We provide our results to first stage regressions in Table 8. Standard errors are robust to heteroskedasticity via White-correction. We create eight specifications as combinations of instrumental variables. Specification 1 regresses AverageContact on OwnGatesAvg, controls, and fixed effects. We find an insignificant effect of OwnGatesAvg with a coefficient of 0.029 and a standard error of 0.023. We find significant effects for Direct, Round, Distance, and Distance2. We note the especially high significance of Direct with a coefficient of -0.174 and a standard error of 0.008. We believe that this significant effect on AverageContact may be due to the contact encounters with other airlines at layover airports. Specification 2 regresses AverageContact on OwnGatesOrig, controls, and fixed effects. We find a highly insignificant effect of OwnGatesOrig with a coefficient of -0.002 and a standard error of 0.016. Specification 3 regresses AverageContact on OwnGatesDest, controls, and fixed effects. We find a somewhat significant positive effect of OwnGatesDest, with a coefficient of 0.028 and a standard error of 0.016. We note that destination gates appear to have a much greater effect than origin gates. Specification 4 regresses AverageContact on OwnGatesAvg, LCCGatesAvg, controls, and fixed effects. We find an insignificant coefficient of OwnGatesAvg of 0.034 with a stan- dard error of 0.023. We find that OwnGatesAvg has slightly stronger effect and significance than in the specification 1, leading to an initial postulation that own gates and low-cost carrier gates work better as a pair. We also note that the standard error is slightly smaller than in specification 1, which suggests that low-cost carrier gates are much more strongly correlated with AverageContact, and very weakly correlated with a carrier’s own gates. We

15 also find a significant effect of LCCGatesAvg of 0.063 with a standard error of 0.024. Low- cost carrier gates may have a stronger effect on multimarket contact as an increase in the number of low-cost carrier gates at an airport implies that there will be more competition at that airport by way of more low-cost carriers operating. Specification 5 regresses AverageContact on OwnGatesOrig, LCCGatesOrig, controls, and fixed effects. We find, once again, an insignificant effect of OwnGatesOrigin of -0.004 with a standard error of 0.016. We also find an insignificant effect of LCCGatesOrig with a coefficient of -0.023 and a standard error of 0.018. We do note that low-cost carrier gates, despite having no real statistical significance, still bear much greater significance than a carrier’s own gates. Furthermore, despite statistical insignificance, results for both own gates and low-cost-carrier gates at the origin are still negative. This suggest that multimarket contact may be negatively impacted by gates at the origin, or that there is a negative-biased omitted variable accounting for the difference. Specification 6 regresses AverageContact on OwnGatesDest, LCCGatesDest, controls, and fixed effects. Unlike in specification 3, we find an insignificant effect of OwnGatesDest, with both coefficient and standard error of 0.036. However, we find a very significant effect of LCCGatesDest of 0.081 with standard error of 0.017, and therefore find once again that both low-cost carriers and destination gates bear much more significant effect on multimarket contact as compared to own gates and origin gates. Specification 7 regresses AverageContact on OwnGatesOrig, OwnGatesDest, controls, and fixed effects. With a coefficient of 0.001 and a standard error of 0.016, we find almost no effect of OwnGatesOrig. We also find little significance of OwnGatesDest, with a coefficient of 0.028 and a standard error of 0.016. Specification 8 regresses AverageContact on OwnGatesOrig, OwnGatesDest, LCCGatesOrig, LCCGatesDest, controls, and fixed effects. This regression provides a consensus of the re- sults we found in the previous regressions. OwnGatesOrig has insignificant, minimal effect. LCCGatesOrig also bears an insignificant effect of coefficient. OwnGatesDest carries a coef- ficient of 0.036 and a standard error of 0.016 for reasonable significance, and LCCGatesDest has high significance with a coefficient of 0.079 and a standard error of 0.017. All regressions are robust to heteroskedasticity via White-corrected standard errors. We acknowledge the possibility of clustering of standard errors, but do not correct for them at this time.

5.2 Results

Overall, we believe that using a simple average of gate lease percent at origin and destination significantly impacts the effect of the instrumental variables. It is statistically more powerful to use only destination gate lease percent as an instrument by the first stage regression.

16 However, for future extension, we wish to expand on our instrumental variable tests by ap- plying the instruments, and ultimately determining the best set of instrumental variables. We believe that using specification 8, the full suite of instrumental variables we have con- structed, may provide the most powerful instrument despite the individual weakness of two instruments. Furthermore, we find some interesting results. First, low-cost carriers are generally a much stronger instrument than a carrier’s own gates. As suggested earlier, the fact the more potential competition exists at an airport if low-cost carriers own gates does give an intuition at this result. More strangely, destination gates have a far greater statistical significance than origin gates. Furthermore, origin gates appear to have a negative though insignificant impact on multimarket contact. For future extension, we wish to analyze a greater set of instrumental variables. In particular, we imagine that national carriers American, United, and Delta may have a strong impact as an instrument. We also see a significant interest in the impact of Southwest in the literature. It may be that low-cost carrier gates are driven in significance by Southwest, or it may be that Southwest gates and low-cost carriers excluding Southwest both have an impact. We also imagine that the amount of common gates at an airport may provide a valuable instrument for future research.

17 References

[1] Bernheim, B.D. and Whinston, M.D. ”Multimarket Contact and Collusive Behaviour.” RAND Journal of Economics, Vol21(1)(1990), pp. 1-26. [2] Berry, S. ”Estimation of a Model of Entry in the Airline Industry.” Econometrica, Vol60(4)(1992), pp. 889-917. [3] Bilotkach, V. ”Multimarket Contact and Intensity of Competition: Evidence from an Airline Merger.” Review of Industrial Organization Vol 38(1)(2011), pp. 95-115. [4] Borenstein, S. ”Hubs and High Fares: Dominance and Market Power in the U.S. Airline Industry.” RAND Journal of Economics, Vol20(3)(1989), pp. 344-365. [5] Borenstein, S. and Rose, N. ”Competition and Price Dispersion in the U.S. Airline In- dustry.” Journal of Political Economy, Vol102(4)(1994), pp. 653-683. [6] Chiang, P. and Liou, T. ”Does multimarket contact affect price dispersion? Evidence from the airline industry.” Working paper, 2018. [7] Ciliberto, F., Murry, C., and Tamer, E. ”Inference on Market Power in Markets with Multiple Equilibria.” Working paper, 2012. [8] Ciliberto, F. and Williams, J. ”Limited Access to Airport Facilities and Market Power in the Airline Industry.” Journal of Law and Economics, Vol53(3)(2010), pp. 467-495. [9] Ciliberto, F. and Williams, J. ”Does multimarket contact facilitate tacit collusion? Infer- ence on conduct parameters from the airline industry.” RAND Journal of Economics, Vol45(4)(2014), pp. 764-791. [10] Evans, W. and Kessides, I. ”Living by the ’Golden Rule’: Multimarket Contact in the U.S. Airline Industry.” Quarterly Journal of Economics, Vol109(2)(1994, pp. 341-366. [11] Gerardi, K.S. and Shapiro, A.H. ”Does Competition Reduce Price Dispersion? New Evidence from the Airline Industry.” Journal of Political Economy, Vol117(1)(2009), pp. 1-37. [12] Griliches, Z. and Mairesse, J. ”Production Functions: The Search for Identification.” Working paper, 1995. [13] Jans, I. and Rosenbaum, D. ”Multimarket Contact and Pricing: Evidence from the U.S. Cement Industry.” International Journal of Industrial Organization, Vol15(3)(1997), pp. 391-412. [14] Kim, Kim, and Tan. ”Tacit Collusion and Price Dispersion in the Presence of Southwest Airlines.” Working paper, 2019.

18 Figure 1: Summary Statistics Table

19 Figure 2: Contact Between Airline Pairs, Q1 2014

Figure 3: Comparison of Average Contact Summary Statistics

20 Figure 4: Log Price Regression Results Table

21 Figure 5: Comparison of Average Contact Results

22 Figure 6: Gini Coefficient Regression Results Table

23 Figure 7: Within-Obs Std. Deviation Regression Results Table

24 Figure 8: First-Stage Regression Results Table

25