Flying Under the Radar: Multimarket Contact and Tacit Collusion in the U.S. Airline Industry
A SENIOR THESIS
Submitted to the Faculty in the Department of Economics at Miami University in Partial Fulfillment of the Requirements of Departmental Honors
By
Henry Jameson Shaneyfelt* Miami University Spring 2021
ABSTRACT This paper serves to identify the effects of multimarket contact on tacit collusion through the measures of average price and price dispersion. Implementing methods used in Evans and Kessides (1994) and Ciliberto and Williams (2014), I employ gate-use data to instrument for the average multimarket contact variable and address any potential bias from endogeneity. Additionally, squared gate-use instruments, which are novel to the existing literature, are included in the analysis. When using the novel instruments, a significant positive relationship is found between average multimarket contact and both prices and price dispersion.
Advisor: Dr. Charles Moul
*I would like to sincerely thank Dr. Moul for all the help he has provided me throughout this process. I would also like to thank the Miami University Economics Department for giving me the opportunity to complete this research project.
1 Introduction The majority of classical microeconomic theory is based in the analysis of market behavior that does not extend beyond the market’s scope. External markets and their effects on firm behavior are typically not studied at length. However, the modern-day economy is seeing an increasing number of firms competing with rival firms in numerous differing markets at a given time. An intriguing attribute of this newly structured economy is the ability for firms to interact with rivals across markets. The competitive pricing of a firm in one market may now result in a rival firm responding in a different, mutually served market. This occurrence, known as multimarket contact, has been a topic of interest for many economists who posit that it substantially increases the probability of collusive behavior. While this may seem difficult to follow at first blush, the intuition behind this theory makes it clear why this outcome is a plausible possibility. The general notion driving this theory is that multimarket contact creates the potential for larger-scale retaliations to firms who may try deviating from collusive outcomes. For the sake of example, imagine there are two firms, A and B, who simultaneously compete in markets 1, 2, and 3. If firm A were to attempt to deviate from a collusive outcome in market 1 by cutting prices, firm B could retaliate by not just dropping prices in market 1 but by also dropping prices in markets 2 and 3. Lest this outcome be realized, firm A will find it optimal to price less aggressively (i.e., collude) in all simultaneously served markets. Such is the essence of tacit collusion, where non-competitive outcomes are not explicitly coordinated but mutually understood and maintained. As the multimarket contact between two firms increases, the potential retaliation does the same. Due to this, firms that experience higher levels of multimarket contact may be less likely to behave competitively. As this probability of tacit collusion rises, theory indicates that monopolistic behavior is more likely to be observed. The specific characteristics of this behavior that I consider are higher prices and more price discrimination. These serve as strong indicators for tacit collusion, as they would be impossible to occur in a competitive setting. Within a market containing competitive firms, the ability for a firm to price discriminate is ultimately eroded by their rivals’ pricing. Having established these indicators, the challenge then becomes finding an industry that exhibits these characteristics. The domestic airline industry is rich with multimarket contact, as there are few routes in which only one carrier serves. In this paper, I measure the effects of increasing multimarket contact on average price and price discrimination. To identify price discrimination, I measure the magnitude of price dispersion in a given market. Initial results replicating Evans and Kessides (1994) yield
2 results that fail to follow theory and intuition. However, I address the potential endogeneity of the multimarket contact variable by instrumenting for it with gate-use data. Ciliberto and Williams (2014) introduced this application of gates, and I add to it by accounting for non-linearity in the relationship between multimarket contact and gate utilization. When employing this technique, I find that increasing multimarket contact increases both average price and price dispersion. The results from this paper ultimately identify the characteristics of tacit collusion occurring at higher levels of multimarket contact, consistent with the prevailing theory.
2 Literature Review 2.1 Theory Bernheim and Whinston (1990) was the first paper to address the potential for increased multimarket contact causing non-competitive outcomes. They provide a mathematical framework to produce an irrelevance result. This irrelevance result proved that multimarket contact would not have an effect on the optimal strategy of the firm given there are identical markets, identical firms, and constant returns to scale in production technology. Building upon this, the authors then prove that removing any of these three criteria results in multimarket contact having nontrivial effects on equilibrium behavior. The authors prove in this paper that the external threat of retaliation across all other simultaneously served markets was enough to alter a firm’s competitive strategy in a single market. Given that in reality, industries typically do not meet these three mentioned criteria simultaneously, this theory has clear applicative power. The task since this paper was published has been finding significant empirical evidence to test this theory.
2.2 Empirical One of the first attempts at such a test was Evans and Kessides (1994). This was the first paper to assert that the airline industry is a strong candidate for measuring the effects of multimarket contact on average prices. The main result of this paper is that multimarket contact does indeed have a positive relationship with average prices. The most notable of this paper’s contributions are the construction of the average multimarket contact variable, which depends on the amount of overlap among airlines’ routes, and the use of city-pair controls in OLS analysis. While Evans and Kessides (1990) found evidence that increased multimarket contact has a positive relationship with average price, their methodology did not address any potential endogeneity
3 issues stemming from the multimarket contact variable itself. The first paper to do this was Ciliberto and Williams (2014). In this paper, the authors claim that the potential endogeneity of multimarket contact may come from the unobservable heterogeneity that effects the decision to either enter or exit a market. To address this, they instrumented for multimarket contact with gate ownership variables, asserting that the long-term leasing of airport gates makes them plausibly exogenous. This implementation led to substantially larger estimated impacts of contact on analysis, and it laid the foundation for the later structural work that related conduct to contact. Along with average price, there is an existing parallel literature looking at price dispersion in the domestic airline industry. Borenstein and Rose (1994) looked at the effects of increased competition on price dispersion by comparing measures of market concentration to a pricing Gini coefficient, a well-known measure of statistical dispersion. The main findings of this paper are that price dispersion cannot be explained by cost variation alone, and increased competition is associated with higher levels of price dispersion. The authors contest that a significant portion of the price dispersion was coming from price discrimination arising in a context of monopolistic competition. Not all results in the literature concerning price dispersion were the same, however. In response to Borenstein and Rose (1994), Gerardi and Shapiro (2009) replicate their study and find contrasting results. Namely, this paper finds that price dispersion actually decreases as competition increases, which is more aligned with standard oligopoly theory. The authors claim that the cause of the different results was primarily that the cross-sectioned data of Borenstein and Rose (1994) omitted distance from their controls. This omission is claimed to have positively biased their results substantially.
2.3 Contemporary Work Kim, Kim, and Tan (2019) attempted a current application of Evans and Kessides (1994) by looking at the effects of multimarket contact on prices and also price dispersion. The authors specifically looked at the effects of Southwest Airlines on collusive outcomes. The paper found that multimarket contact is positively associated with prices but fails to find the same effect with price dispersion. The authors also noted that the presence of Southwest Airlines on a route results in multimarket contact not having a significant association with price dispersion. In another application of Evans and Kessides (1994), Chiang and Liou (2018) tried to measure the effects of multimarket contact on price dispersion in the airline industry. This paper also implemented the Gini coefficient to measure this dispersion. The main finding from this work
4 was that differing market sizes result in differing effects of multimarket contact on dispersion. However, the paper ultimately fails to identify a positive relationship between multimarket contact and price dispersion.
3 Data 3.1 Data Sources A great majority of the data collected and utilized comes from the Bureau of Transportation Statistics’ DB1B. The DB1B is a 10% random sample of flight itineraries from reporting carriers. The sample used for this analysis is quarterly data spanning from the beginning of 2014 to the end of 2019. Observations of prices, market distances, and the types of itineraries for the 24 quarters were taken directly from the DB1B. Additionally, several of the variables used in this analysis were constructed using this data. Unlike pricing and market data, airport gates data are not easily accessible. In fact, all data related to airport gates are privately owned by several large companies. One of the largest of these companies is Cirium, owner of the website FlightStats which contains flight records going as far back as 1995. Included in these data are the routes, carriers, and gates that were utilized in the flights of any date. To extract gate data, I developed a sophisticated web scraper. With the web scraper, a cross-section of 33 airports was collected from a date in the second quarter of 2016, which falls in the middle of the sample being analyzed. To further justify the utilization of the cross-sectional data with the panel pricing and market data, the gates data were compared to Airport Competition Plans (ACPs). The Wendell H. Ford Aviation Investment and Reform Act for the 21st Century, also known as AIR21, mandated that medium and large hubs, as defined by the Federal Aviation Administration (FAA), are required to submit competition plans. These plans include information on leasing, airport expansion, and gate allocation. The plans for the 33 airports in the sample were found online and cross-referenced with the collected data to ensure accuracy. Berry (1992) also suggests that airports sign long-term leases to aid in capital investment while maintaining low interest on debt issues, an assertion also employed by Ciliberto, Murry, and Tamer (2020). This institutional detail allows for cross-sectional data to be used with confidence that gate-use variation over time is likely small.
5 3.2 Observations Observations in this sample are indexed by firm, city-pair, and quarter. A market is defined as a unidirectional route between two airports in a given time period. An example of this would be a flight from Columbus (CMH) to Atlanta (ATL) in the first quarter of 2014. Observations from the data are dropped if there are fewer than 100 passengers flying with a specific firm in a specific market. Since the DB1B is a 10% sample, the threshold becomes 10 sampled passengers. Observations are also dropped if the price is either lower than $25 or higher than $2,500, as these are likely incorrectly inputted data. These are the same thresholds used in Ciliberto and Williams (2014). Additionally, there are 24 distinct periods in the data that are indexed by t ϵ {1, …, T}, with T = 24. I sample 7 of the largest airlines in the data that are indexed by j ϵ {1, …, J}, with J = 7. These airlines are American Airlines (AA), Alaskan Airlines (AS), Delta (DL), Frontier (F9), Spirit (NK), United Airlines (UA), and Southwest Airlines (WN). The omissions of smaller airlines from this sample, such as Allegiant Airlines, have the potential to hinder the estimation of the effects of multimarket contact on collusive measures. While the magnitude of this effect is unknown, it is worth noting before analyzing the results. With the seven airlines in the sample, there are 1,056 distinct markets in the data that are indexed by m ϵ {1, …, M}, with M = 1,056. The summary statistics for all variables can be seen in Table 1.
3.3 Dependent Variables The dependent variables in this analysis are the Gini coefficient and the natural log of the average fare. As stated previously, the Gini coefficient is a measure of statistical dispersion which is typically used to measure income inequality. However, it lends itself well to measuring price dispersion. It ranges from 0, which indicates all consumers pay the same fare, to 1, which indicates one consumer paid a positive fare and all others flew for free. In the analysis, the Gini coefficient is scaled up by 100. The natural log of average price will then be used to measure the percent change in prices.
3.4 Average Multimarket Contact Variable For the average multimarket contact variable, we use the construction introduced in Evans and Kessides (1994). To calculate this, a contact matrix must be constructed first. The contact matrix Ct = (cjmt) for every market and period pair, where
6
c = δ δ for j, k ϵ {1, … , J } The δs are binary variables indicating whether or not firm j or firm k is serving that market during that time period. The resulting matrix contains elements with the total amount of simultaneously served markets between two firms. An example of the contact matrix can be seen in Table 2. In this matrix, we see that Delta (DL) is a rival on 97.6% of routes flown by American Airlines (AA) during that quarter. The expectation from this is that prices and price dispersion will be higher in their mutually served markets. The opposite is expected between Alaskan Airlines (AS) and Spirit Airlines (NK), where Spirit is only a rival in 30.3% of Alaskan’s routes in that quarter. The average multimarket contact variable is constructed in the following equation: 1 ������� = � δ δ [� (� − 1)]/2
In this equation, � is the number of firms serving market m at time t. This variable is then scaled down just as it is in Evans and Kessides (1994). Ciliberto and Williams (2014) acknowledges the possibility of endogeneity when it comes to the multimarket contact variable. This issue would necessitate the use of instrumental variables to remove any biases from the estimation. Ciliberto, Murry, and Tamer (2020) observe that market- specific multimarket contact may be endogenous due to some of the unobservable heterogeneity correlated with the decision to enter and exit a market.
3.5 Control Variables All of the controls used in this analysis come from Evans and Kessides (1994) and Ciliberto and Williams (2014). The controls are the natural log of distance, the natural log of distance squared, route market share, airport market share, percent of tickets that are direct, percentage of tickets that are roundtrip, and hub status. Including distance captures the consumer’s decision to travel, as it may change with differing distances. The intuition justifying the inclusion of the natural log of distance squared is that, as the distance of travel increase, flying becomes a much more attractive option relative to driving. That being the case, there is likely going to be some non-linearity associated with the distance variable. The route market share variable is calculated as the percent of tickets sold by a firm in a specific market-period pair. The airport market share variable is the average of the route market
7 share variables at both endpoints in a market. Ciliberto, Murry, and Tamer (2020) also discuss the endogeneity of variables of these nature with the same intuition that was used for claiming the endogeneity of multimarket contact. To account for possible pricing variation across itinerary types, the percentage of tickets which are direct and the percent of tickets which are roundtrip are utilized as controls. Also, the hub variable is the percent of flights in which either the destination or origin of the flight was a hub for the carrier. This control accounts for any cost advantages a firm may have at a hub airport.
3.6 Instrumental Variables The instruments used for multimarket contact include those used in Ciliberto and Williams (2014). These include the percentage of gates used by an airline at both the origin and destination, the percentage of gates used by an airline just at the origin, the percentage of gates used by an airline just at the destination, the percentage of low-cost carrier used gates at the origin and destination, and the percentage of gates utilized by Southwest at the origin and destination. The intuition behind these is again that the number of gates utilized by a carrier is resistant to shocks due to the long-term nature of airline-airport leasing agreements. There are three additional instrumental variables used in this analysis that are new to the literature. These variables are the percentage of gates used at both origin and destination squared, the percentage of gates used at just the origin squared, and the percentage of gates used at just the destination squared. While Ciliberto and Williams (2014) is correct in the usage of gates as an instrument for multimarket contact, the following reasoning suggests that they did not examine the full relationship between the two variables. The intuition behind these inclusions is that as the number of gates utilized increases, the more likely the airport is going to be a regional airport in which one airline will serve primarily. Following this logic, the highest level of multimarket contact should occur when all carriers are equally competing in the same market. Additionally, the notion of non-linearity is justified by the fact that a firm with a monopoly in a market has an average multimarket contact of 0. An application of Rolle’s theorem proves that there must be at least one turning point in the effect of gates on average multimarket contact. While it is not clear whether or not the relationship is purely quadratic, the use of the squared instrument is essential in capturing non-linearity. Additional justification comes from Dieterle and Snell (2016), which looked at the effects of the inclusion of squared instrumental variables in previously linear instrumental variable
8 analyses. The authors find that several of the papers they studied have statistically different results when the squared instrument is included.
4 Results 4.1 Models The base models used are reduced form and follow the approach of Ciliberto and Williams (2014). The dependent variables are regressed onto average multimarket contact, the vector of control variables, and the fixed effect variables. The following equations describe the OLS models:
ln(Average Fare) = β ∗ Contact + γ ∗ Controls + ψ + µ + η + ϵ
Gini Coefficient = β ∗ Contact + γ ∗ Controls + ψ + µ + η + ϵ
In these models, β is the coefficient of the average multimarket contact variables, and γ is the vector of coefficients of the vector of controls. The fixed effects for firm, city-pair, and quarter are represented by ψ , µ , and η , respectively. Finally, the idiosyncratic error term is represented by
ϵ . Note that half of the specifications in Table 3a and Table 3b include city-pair fixed effects. Since these variables are included, the distance variables must be dropped to prevent perfect multicollinearity. For the other half, the opposite is the case where city-pair fixed effects are included, and distance variables are dropped. In addition to the OLS, there are first and second stage regressions for the instrumental variable analysis. Several first stage regressions were run yielding similar results, but the most noteworthy two specifications are presented. As mentioned earlier, there is potential for the route market share and airport market share variables to be endogenous. As a result of this, these variables are excluded from all of the specifications. Also, the lack of variation in the gates data results in multicollinearity with the city-pair fixed effects. The city-pair fixed effects are consequently dropped from the instrumental variable analysis to deal with this. In the second stage specifications, average multimarket contact is instrumented with gates, gates squared, and Southwest gates. The justification for this specification appears in the first stage results section.
9 4.2 OLS Regressions To understand the interpretation of these results, the nature of the unscaled average multimarket contact variable must be understood. To increase this variable by a unit, enough additional markets must become simultaneously served to move the average up by 1. Since this variable is an average, the number of additional markets needed for this one-unit increase is equal to the number of firms competing in the market (e.g., a market with 2 competing carriers only needs 2 additional simultaneously served markets). Having acknowledged this, interpreting these results becomes much less complicated. The results of the six specifications in Table 3a tell conflicting stories. While average multimarket contact has a positive relationship with the natural log of average fares when city-pair fixed effects are not included, there is a negative relationship between the two when they are. Another glaring feature of these results is that all of the estimates are economically negligible. For the average price, a unit increase in multimarket contact results in no more than a 1% change. Compared to Evans and Kessides (1994) and Ciliberto and Williams (2014), this is consistent with the results of initial OLS regressions in each paper. The results of Table 3b which had regressions with the Gini coefficient as the dependent variable are similar to those of the previous table. The signs for the coefficients of average multimarket contact flip again when controlling for city-pair fixed effects. Having acknowledged that, the results do not appear to be any less trivial than the previous results. Looking at the R- squared statistics across tables, the models with the natural log of average fare as the dependent variable again seem to have substantially more of the variation explained than the model with the Gini coefficient as the dependent variable. Also, the R-squared statistics reveal that the specifications with distance variables explained more of the variation than those with city-pair fixed effects. Overall, these initial regressions would point to average multimarket contact not having any real impact on prices or price dispersion.
4.3 IV First Stage Regressions The most noteworthy result of the first stage regressions in Table 4 is the effect that the squared instrumental variable has on the gates variable. The correction of the negative bias on the gate variables is clearly substantial. Considering the slopes implied by the second specification, the maximum amount of multimarket contact occurs when about 17% of gates are utilized by a carrier. An equal distribution of gates between the sampled airlines has them each utilizing 14% of the gates.
10 These similar figures follow the intuition that the highest multimarket contact would occur when all carriers were competing in a market. These observations justify the claim that non-linearity in the gates variable exists. Having not been implemented in this literature before, this is an important finding. The coefficient on the gates variables, when accounting for non-linearity, has a significant positive association with the average multimarket contact variable. On the other hand, the Southwest gates variable has a negative association with average multimarket contact. The intuitive response to this result is that the carrier in this variable has, on average, a lower number of gates and presence in the domestic airline industry than the rest of the sample. The increase in the percentage of gates Southwest has means that there are lower percentages of gates held by carriers with larger presences such as American, Delta, and United. It is also worth noting that multimarket contact is decreasing with the distance of the city-pair over all relevant distances
4.4 IV Second Stage Regressions The second stage regressions in Table 5 yield some relevant results. In contrast to the OLS results, average multimarket contact has a significantly positive effect on both average price and price dispersion. Given that this is not a result that has been able to be achieved in contemporary research, the results have significant meaning. Interpreting the results, a one unit increase in average multimarket is associated with an increase in prices of 56%, ceteris paribus. The percent change in prices appears to be quite large, but given the volatility of airline prices, this result is not inconceivable. This result matches the results of Ciliberto and Williams (2014), which found substantial increases in average price with increases in multimarket contact. Looking at the effects of distance, average fares appear to increase with distance for all flights. Along with this result, a one unit increase in average multimarket is associated with an increase an increase in price dispersion of 2.2 percentage points, ceteris paribus. This result deviates from the contemporary literature, as this result reveals a statistically significant and positive relationship between average multimarket contact and price dispersion. Looking at the effects of distance again, price dispersion increases with distance for all short flights (Distance<1,500 miles) and decreases with all long flights (Distance>1,500 miles). This result indicates that price dispersion may not be as relevant in coast-to-coast flights. Along with these results, there were two tests used to validate the results of the instrumental variable regressions. The first of these two is the Stock-Yogo test, which tests to see if the
11 instruments are strong predictors of the variable that is being instrumented. The threshold for this test is an F-stat of 10, so none of the instruments used were weak predictors. The other test implemented was the Wu-Hausman test, in which the endogeneity of the average multimarket contact variable is measured. If the estimates that came from the base model were similar enough to those of the second stage regressions, then average multimarket contact may not be endogenous at all. The results show that average multimarket contact was indeed endogenous and warranted the employment of instrumental variables.
5 Conclusion Firms respond to incentives. Multimarket contact unquestionably creates an incentive for firms to refrain from competitive behavior. The goal since the advent of Bernheim and Whinston (1990) has been to prove that the theory occurs in the real world. Paper by paper, the empirical work in this literature has been using different methods in attempts to align with theory. This paper not only is consistent with theory, but it also implements a novel method that significantly augments the accuracy of the empirical side of this literature. There are many positive implications that can be derived from these results. The most apparent of these is that the empirical evidence matches theory. Falling in line with the majority of the theoretical and empirical literature, the results indicate that there is a significantly positive relationship between multimarket contact and both price and price dispersion. It is also abundantly clear that the effects of multimarket contact are meaningful in this context. The estimates of this paper indicate that increased multimarket contact has a conspicuous impact on the city-pair markets in which it is present. The domestic airline industry is one that has been plagued with market inefficiency stemming from non-competitive pricing. These results are consistent with the notion that this inefficiency is still present. Another implication that is certainly worth noting is that of the squared instrumental variables. The implementation of a squared instrumental variable in this literature is entirely unprecedented. The intuition justifying its inclusion and the potency it showcases in this analysis would lead one to the conclusion that this an important contribution to the literature. With improved gate data, the potential of this new instrumental variable can only grow. There are several limitations in this paper to be acknowledged as well. While the sample generated from the seven carriers and 1,056 unique city-pair markets is substantial, the addition of other large airlines and city-pairs can only improve the robustness of the results. The inclusion of
12 other large low-cost carriers such as JetBlue and Allegiant would make for a compelling extension. The inclusion of these variables could further account for lower levels of multimarket contact, giving a more accurate picture of the effect of multimarket contact across all levels. Given the accessibility of the data, the effort required to make these addenda would be trivial. While the limitation in data from the DB1B is noteworthy, the most glaring data deficit lies with the gates. Subject to financial and temporal constraints, cross-sectional data was the most that could have been collected. The improvements that can come from panel gate data are by no means negligible. Being able to account for changes in gate utilization over the observed time periods would add an entirely new dimension to the analysis. Consequently, the changes in gate use would undoubtedly correct inaccuracies that the instruments in this paper may have. Though the corrections may not be crucial, increasing the estimates’ proximity to the true values is the ultimate goal.
13 Appendix
Table 1: Summary Statistics (N=93,791)
Statistic Mean St.Dev. Min. Max.
Gini Coefficient (x100) 23.876 6.172 5.014 64.090 Average Fare ($) 252.25 101.56 30.39 1,128 Contact 0.218 0.274 0 1.5 Distance 1,398 930 12 5,093 Route Mkt Share 0.259 0.280 0 1 Airport Mkt Share 0.219 0.145 0.005 0.935 Direct 0.311 0.406 0 1 Roundtrip 0.689 0.191 0 1 Hub 0.385 0.487 0 1 Gates (%) 0.060 0.051 0.002 1 Gates Origin (%) 0.067 0.084 0.001 1 Gates Dest. (%) 0.067 0.083 0.001 1 WNGates (%) 0.031 0.045 0 0.614 LCCGates (%) 0.067 0.061 0.007 0.811
Table 2: Number of Common Markets in 2016-Q2 AA AS DL F9 NK UA WN AA 862 249 841 309 209 773 609 AS 249 251 249 104 76 225 179 DL 841 249 973 309 210 782 711 F9 309 104 309 309 114 302 247 NK 209 76 210 114 210 181 156 UA 773 225 782 302 181 806 566 WN 609 179 711 247 156 566 734 Note: The off-diagonal numbers represent the number of markets served simultaneously by the carrier in the row and the carrier in the column. The numbers on the diagonal are the total number of markets served by a carrier. Airlines listed: American Airlines (AA), Alaskan Airlines (AS), Delta Airlines (DL), Frontier Airlines (F9), Spirit Airlines (NK), United Airlines (UA), and Southwest Airlines (WN).
14 Table 3a: Prices & Multimarket Contact (OLS, N=93,791) Dependent variable: ln(Average Fare) (1) (2) (3) (4) (5) (6) Contact 0.002** 0.005*** 0.004*** -0.010*** -0.009*** -0.009*** (0.001) (0.001) (0.001) (0.001) (0.002) (0.001) ln(Distance) -1.008*** -1.090*** (0.030) (0.030) ln(Distance)2 0.094*** 0.101*** (0.002) (0.002) Route Mkt Sh. 0.038*** -0.044*** 0.037** -0.063*** (0.006) (0.006) (0.015) (0.015) Airport Mkt Sh. 0.686*** 0.617*** 0.685*** 0.636** (0.013) (0.013) (0.031) (0.032) Direct -0.129*** -0.037*** -0.103*** 0.019** (0.003) (0.003) (0.008) (0.007) Roundtrip -0.089*** -0.089*** -0.148*** -0.129*** (0.006) (0.006) (0.010) (0.011) Hub -0.020*** 0.015*** -0.056*** 0.014*** 0.063*** -0.012*** (0.002) (0.002) (0.002) (0.005) (0.005) (0.005) City-Pair FE No No No Yes Yes Yes R2 0.735 0.709 0.724 0.648 0.618 0.635 Note: All regressions account for airline and year-quarter fixed effects. Heteroskedastic robust standard errors are in parentheses. *p<0.10, **p<0.05, ***p<0.01
15 Table 3b: Price Dispersion & Multimarket Contact (OLS, N=93,971) Dependent variable: Gini Coefficient (1) (2) (3) (4) (5) (6) Contact -0.097*** -0.080*** -0.194*** 0.061** 0.067** 0.066*** (0.018) (0.018) (0.018) (0.027) (0.028) (0.029) ln(Distance) 19.157*** 18.775*** 19.505*** (0.652) (0.634) (0.692) ln(Distance)2 -1.346*** -1.316*** -1.428*** (0.046) (0.045) (0.049) Route Mkt Sh. -1.297*** 0.362*** -0.198 1.089*** (0.142) (0.140) (0.309) (0.290) Airport Mkt Sh. 6.560*** 8.227*** 9.683*** 11.159*** (0.297) (0.313) (0.636) (0.689) Direct 5.345*** 5.717*** 3.266*** 4.685*** (0.070) (0.065) (0.159) (0.134) Roundtrip -5.436*** -5.242*** -3.317*** -3.131*** (0.165) (0.164) (0.232) (0.240) Hub 1.042*** 1.277*** 2.295*** 0.180* 0.812*** 0.581*** (0.044) (0.043) (0.042) (0.103) (0.108) (0.099) City-Pair FE No No No Yes Yes Yes R2 0.257 0.252 0.199 0.208 0.192 0.190 Note: All regressions account for airline and year-quarter fixed effects. Heteroskedastic robust standard errors are in parentheses. *p<0.10, **p<0.05, ***p<0.01
16 Table 4: Multimarket Contact & Gates (OLS, N=93,971) Dependent variable: Contact (1) (2) Gates 0.327*** 1.788*** (0.109) (0.198) Gates2 -5.312*** (0.629) WN Gates -1.680*** -1.792*** (0.090) (0.091) ln(Distance) -1.266*** -1.182*** (0.150) (0.149) ln(Distance)2 0.072*** 0.065*** (0.011) (0.011) Direct -0.318*** -0.342*** (0.015) (0.015) Roundtrip 0.339*** 0.330*** (0.027) (0.027) Hub -0.089*** -0.094***
(0.010) (0.010) R2 0.809 0.809 Note: All regressions account for airline and year-quarter fixed effects. Heteroskedastic robust standard errors are in parentheses. *p<0.10, **p<0.05, ***p<0.01
17 Table 5: Effects of Multimarket Contact on Collusion Indicators (2SLS, N=93,971) Dependent variables: Gini Coefficient ln(Average Fare) (1) (2) Contact 2.181*** 0.560*** (0.279) (0.028) ln(Distance) 22.224*** -0.377*** (0.944) (0.093) ln(Distance)2 -1.522*** 0.059*** (0.065) (0.006) Direct 6.332*** 0.114*** (0.103) (0.011) Roundtrip -6.035*** -0.231*** (0.202) (0.019) Hub 1.440*** 0.055*** (0.052) (0.006) Stock-Yogo 138*** 138*** Wu-Hausman 73*** 1803*** Note: Instruments used are Gates, Gates2, and WN Gates. All regressions account for airline and year-quarter fixed effects. Heteroskedastic robust standard errors are in parentheses. *p<0.10, **p<0.05, ***p<0.01.
18 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] Borenstein, S. and Rose, N. “Competition and Price Dispersion in the U.S. Airline Industry.” Journal of Political Economy, Vol102(4) (1994), pp. 653-683.
[4] Chiang, P. and Liou, T. “Does Multimarket Contact Affect Price Dispersion? Evidence from the airline industry.” Working paper, 2018.
[5] Ciliberto, Murry, and Tamer. “Market Structure and Competition in Airline Markets.” SSRN. February 14, 2020.
[6] Ciliberto, F. and Williams, J. “Does Multimarket Contact Facilitate Tacit Collusion? Inference on Conduct Parameters from the Airline Industry.” RAND Journal of Economics, Vol45(4) (2014), pp. 764-791.
[7] Dieterle, Snell. “A Simple Diagnostic to Investigate Instrument Validity and Heterogeneous Effects When Using a Single Instrument” Labour Economics, Volume 42, 2016, Pages 76-86.
[8] 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.
[9] 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.
[10] Kim, Kim, and Tan. “Tacit Collusion and Price Dispersion in the Presence of Southwest Airlines.” SSRN, 2020.
[11] Stock, J.H. and Yogo, M. “Testing for Weak Instruments in Linear IV Regression.” Identifi- cation and Inference for Econometric Models, 2005.
19