Tian Luo, Vikrant Vaze 1

Transcription

1 Tian Luo, Vikrant Vaze 1 Impacts of Airline Mergers on Passenger Welfare Tian Luo Thayer School of Engineering at Dartmouth College 14 Engineering Drive, Hanover, NH Tel: ; tian.luo.th@dartmouth.edu Vikrant Vaze, Corresponding Author Thayer School of Engineering at Dartmouth College 14 Engineering Drive, Hanover, NH Tel: ; Fax: ; vikrant.s.vaze@dartmouth.edu Word count: 7221 words text + 1 tables x 250 words = 7471 words Submission Date: July 31, 2016

2 Tian Luo, Vikrant Vaze 2 ABSTRACT Over the last decade, US domestic airline industry has undergone a series of consolidations. We provide a comprehensive assessment of the overall effects of each of the five major recent mergers on passenger welfare as evaluated through consumer surplus changes. We develop discrete choice models with fare, nonstop and one-stop service frequency, travel time, and other carrier and route attributes as parameters. The consumer surplus, as a function of these parameters, is calculated for each market as the measure of passengers welfare. By using the markets not affected by the mergers as a control group, we are able to separate out the welfare effects of mergers from those of other extrinsic factors such as oil price changes, changes in economic conditions, etc. Several new insights are obtained. We find that mergers of legacy network carriers with a significant proportion of overlapping markets are generally accompanied by flight reallocation and network reorganization, which in turn, contribute to an increase in passenger welfare. However, overall passenger welfare for very small communities declined after the mergers. Also, overall passenger welfare in markets with many competitors declined, consistent with the classic economic theory of consolidation-induced welfare losses. We also find that the welfare gain from mergers of legacy network carriers with significant proportion of overlapping markets progressively decreased as the number of existing major domestic carriers decreased, and that after the most recent mergers, any further potential mergers of legacy network carriers are likely to result in welfare losses. Key words: Airline Merger, Consumer Welfare, One-stop Service Frequency, Multinomial Logit, Consumer Surplus

3 Tian Luo, Vikrant Vaze 3 1. INTRODUCTION Since its deregulation in 1978, U.S. airline industry has experienced two major waves of airline consolidations. During the first wave of mergers and bankruptcies, starting right after deregulation and lasting until the early 1990s, the number of major domestic carriers dropped from 23 to 8 (1). The second wave of mergers started with the US Airways America West Airlines merger in 2005 and ended with the most recent American Airlines US Airways merger, during which the number of mega-carriers (defined as carriers that carry at least 5% of all U.S. domestic passengers) decreased from seven to four. The overall effects of airline consolidations on air travelers are of considerable interest to researchers and policy makers alike. There is a large amount of literature studying the impacts of mergers on passengers. However, most previous studies (2,3,4,5,6) have almost exclusively focused on the fare changes caused by the mergers, and most have focused on the impacts of only one or two of these mergers in any single study. Comparative analysis of multiple mergers has not been performed before. Moreover, those previous studies which did focus on service frequency changes due to the mergers have been limited to assessing the merger impact on only the nonstop service frequency. In actuality, passengers also consider flights/routes with one or more stops when making travel decisions. For example, approximately 30% of the domestic U.S. passengers traveling within the 48 contiguous states in the year 2015 chose a one-stop itinerary (7). In this study, for the first time, we analyze the impacts on overall passenger welfare due to each of the five mergers of major carriers, starting with the US Airways America West Airlines merger in 2005 and ending with the American Airlines US Airways merger in We develop a discrete choice model with fare, service frequency (both nonstop and one-stop), travel time, and other carrier and route attributes as parameters. The consumer surplus, also incorporating fare, service frequency, travel time and other attributes, is calculated for each market as a measure of passengers welfare. In order to evaluate the welfare impacts of these mergers, we compare the difference between consumer surplus before and after the merger in markets affected by the merger with the difference between consumer surplus before and after the merger in markets not affected by the merger. In other words, we use the markets not affected by the merger as a control group, and therefore can separate out the effects of mergers from those of changes in other extrinsic factors such as oil price changes, changes in economic conditions, etc. Our study makes three main contributions. 1) We provide a holistic assessment of the overall passenger impact of the mergers by capturing not only fare changes but also changes in service quality as measured by both nonstop and one-stop service frequency, travel times, and other attributes of the routes and carriers. We demonstrate that in addition to the effects of fare changes, the effects of frequency changes also play a very important role in determining the welfare consequences. The welfare consequences in our study are calculated by incorporating these multidimensional attributes. 2) Ours is the first study to define, calculate and use the changes in one-stop service frequency as a part of passenger welfare changes. 3) Previous studies focused on only one or two mergers at a time. However, we analyze the impact of all five major mergers in the second wave of mergers and study general impacts that the mergers bring to the passengers, by observing the similarities and differences in the effects of these five mergers. The rest of this paper is organized as follows. Section 2 describes the data sources used in our analyses and provide details of some key data pre-processing steps. Section 3 discusses the passenger choice model used in this study and summarizes the passenger choice model estimation results. Section 4 describes the calculation of consumer surplus as well as the difference-in-

4 Tian Luo, Vikrant Vaze 4 differences estimator used to measure the welfare change caused by the merger. Section 5 presents a series of key findings of this research and Section 6 provides the conclusions and discussion. 2. DATA SOURCES AND PRE-PROCESSING In this study we examined all five mergers of major carriers that happened between 2005 and 2015, including US Airways (US)-America West Airlines (HP), Delta Airlines (DL)-Northwest Airlines (NW), United Airlines (UA)-Continental Airlines (CO), Southwest Airlines (WN)-AirTran Airways (FL) and American Airlines (AA)-US Airways (US). The pre-merger and post-merger periods for each merger are selected as follow: US-HP: 2005 Q1-Q4 and 2008 Q1-Q4, DL-NW: 2008 Q1-Q4 and 2010 Q1-Q4, UA-CO: 2010 Q1-Q4 and 2012 Q1-Q4, WN-FL: 2012 Q1-Q2 and 2015 Q1-Q2, AA-US: 2013 Q3-Q4 and 2015 Q3-Q4. The selection of these periods is based on a number of considerations. First, the post-merger period should begin after the merging of the FAA operating certificate as reflected in the Airline On-Time Performance (AOTP) database (8) and also in the Airline Origin Destination Survey (DB1B) database (7) on the Bureau of Transportation Statistics (BTS) website. Second, the pre-merger period should be no later than the year when the DOJ approved the merger. Third, the overlap of the pre- and post-merger periods across different mergers should be as little as possible. Fourth, the gap between pre- and post-merger periods should not be too long. Note that the case of the WN FL merger is somewhat different than the rest. Although they received an FAA single operating carrier certificate as early as in March 2012, the merging carriers operated separately (ticketing systems had not merged and joint itineraries were not sold) until the first quarter of This difference is reflected in our choice of pre- and post-merger periods. Because of the overlap between the AA-US and WN-FL merger timelines, we also performed a joint analysis of these two mergers. Through this joint analysis, we obtained results that were very similar to those obtained from a separate analysis of each merger. Therefore, for simplicity of presentation, we only present the results of the separate analyses in this paper. 2.1 Data Sources Two databases from the Bureau of Transportation Statistics (BTS) website were used to acquire all the data required for our analysis. The Airline Origin and Destination Survey (DB1B) database provides a 10% sample of the U.S. domestic passenger tickets for each quarter of a year (with flight date and time information removed) (9). The DB1B Market table in this database contains the origin and destination of each directional market. This table provides information on number of passengers, fares and ticketing carrier of each sampled ticket, but does not include scheduling information beyond the year and quarter of a flight. By aggregating this data we get the total quarterly number of passengers and average fares for each combination of carrier and route (where a route is defined as a combination of origin airport, connection airport (if any), and destination airport). More importantly, this dataset provides information on connection airport codes for all one-stop passengers. The second database we depend on is the Airline On-Time Performance (AOTP) database, which provides schedule and operational information about individual U.S. domestic flights by major carriers (8). Reporting to this dataset is mandatory for carriers that have at least 1% of the total U.S. domestic scheduled-service passenger revenues. AOTP is particularly useful to us because, combined with the DB1B datasets, it allows us to calculate the nonstop and one-stop service frequency and average total travel time. The process used for calculation of these and other attributes is described next in Section 2.2.

5 Tian Luo, Vikrant Vaze Data Pre-processing Before describing the data pre-processing steps, we first define two terms that we will use repeatedly in this section and throughout the rest of the paper. A carrier-route is defined as a combination of carrier, origin airport, connection airport (only for one-stop routes), and destination airport, representing the flight path a passenger could select to travel from his/her origin to destination on a specific carrier. An itinerary is defined as a sequence of connecting flight(s) representing a one-way trip. According to our definitions, each valid itinerary is associated with exactly one carrier-route but each carrier-route is typically associated with several valid itineraries across that quarter. Note that our analysis ignores all one-stop itineraries that use two different carriers that are not the mainline-regional partners of each other. However, the total passenger share of such itineraries is very small, accounting for approximately 0.34% of all passengers. Three major data pre-processing steps are performed sequentially. The first step is to match the mainline carriers to their regional partners that help the mainline carriers in taking passengers from smaller airports to their major hubs, and back. The flights of the regional carriers are appended to the flight list of their respective mainline partners for all subsequent analyses. The second step is to generate the (nonstop and one-stop) itineraries for each carrier-route for each quarter. The third step is to calculate the values of the attributes for each carrier-route in each market Mainline-Regional Carrier Match From the aforementioned AOTP database, we generated lists of flights belonging to each carrier for each quarter. Since the regional partners help in taking passengers to and from hub airports of the mainline carriers, we need to include the regional carriers flights to avoid underestimation of service frequency especially for those small communities which depend heavily on the regional carrier service. In order to match the mainline carriers with their regional partners, we performed the following two steps: 1. We used the DB1B market dataset to calculate the number of passengers each regional (operating) carrier delivered for the mainline (ticketing) carrier, and we ranked the regional carriers by the number of passengers that they delivered in each quarter. 2. The regional carriers that delivered at least the minimum of 10,000 (10% sampled) passengers per quarter or 5% of the mainline carrier s number of passengers in that quarter were labeled as partner regional carriers. If the partner regional carrier reported to the AOTP database that quarter, then their flights are added to the flight list of their mainline partner. Note that the thresholds of 10,000 sampled passengers and 5% of the mainline carrier s passengers are both arbitrary. However, upon conducting several additional experiments while varying these two thresholds we found that our main results remained unaltered. Since in our data, on an average, the proportion of passengers on code-share flights accounted for less than 1% of the total passengers flying with that ticketing carrier, we found that this does not have any significant effect on our results. Therefore, we did not treat code-share agreement flights any differently than other flights in our datasets Itinerary Generation In this study we only consider the nonstop and one-stop carrier-routes, and those with more than one stop are excluded from both the pre- and post-merger data for all mergers. This simplification makes our analysis considerably easier to perform and understand, and is not expected to affect the main results of our study significantly, since itineraries with more than one stop account for

6 Tian Luo, Vikrant Vaze 6 less than 3% of all passengers in the DB1B database for the time period considered in this analysis. Generation of nonstop itineraries is straightforward, since each flight in AOTP is associated with a single nonstop itinerary. A one-stop itinerary is generated for each pair of flights in the AOTP satisfying the following rules: 1. The carrier-route associated with the flight pair should exist in the DB1B Market table, and that carrier-route should have, on an average, at least 2 passengers per day, that is, 18 (10% sampled) passengers each quarter. 2. The planned connection time (that is, the difference between the scheduled departure time of the second flight and the scheduled arrival time of the first flight) should be no less than 30 minutes and no longer than 5 hours. 3. The date of the two flights (defined based on the planned departure times of the two legs of the one-stop itinerary) should be the same (we do not consider the overnight connections) and both flights should belong to the flight list of the same mainline carrier associated with the carrierroute (thus flights from regional partner carriers are also included). Our main results were found to be highly insensitive to small variations in these data preprocessing assumptions Carrier-Route Attributes Calculation Following the generation of itineraries, we calculated the average fare, total travel time and service frequency at the carrier-route level. Average fare was calculated by aggregating the tickets prices in the DB1B Market table. The average value of the total travel time for each carrier-route was obtained by averaging the total travel time of all generated itineraries belonging to that carrierroute. The total travel time of a nonstop itinerary is the scheduled elapsed time (that is, the difference between the schedule arrival time and the scheduled departure time) of that flight. Total travel time of a one-stop itinerary is defined as the sum of the scheduled elapsed time for both flights in that itinerary plus the planned connection time (that is, the difference between the scheduled departure time of the second flight and the scheduled arrival time of the first flight). For a nonstop carrier-route, service frequency is the number of nonstop flights (itineraries) associated with that carrier-route in each quarter. For a one-stop carrier-route, service frequency is defined as the number of distinct first-leg flights from the origin airport to the connection airport such that there exists at least one second-leg flight (from the connection airport to the destination airport) operated by the same carrier or its regional partners, and the planned connection time is between 30 minutes and 5 hours. 3. PASSENGER CHOICE MODEL In our model a market is defined as an ordered pair of airports in a specific quarter, such that, on an average, at least 30 passengers travel from origin to destination of this market per day. In other words, there should be at least 270 (10% sampled) DB1B-listed passengers traveling in that market in that quarter. Market is directional in our model. For example, Boston to Los Angeles and Los Angeles to Boston in the same quarter are treated as two different markets. Products are carrierroutes that link the origin and destination of that market, and a valid carrier-route in a market should have, on an average, at least 2 passengers per day, and should have, on an average, a service frequency of at least 1 per week. In our model, we only consider carrier-routes with at most one connection, which means that the carrier-routes are either nonstop or one-stop. The choice set of a passenger in a market is the set of all products (valid carrier-routes) in that market.

7 Tian Luo, Vikrant Vaze 7 The indirect or random utility derived by a passenger from a product (carrier-route) j in market m can be expressed as U j m = V j m + ϵ j m where V j m represents the deterministic part of the utility, and the error term ϵ j m is the random or disturbance component of the utility, which represents the unobserved preferences of passengers for product j in market m. We assume that the ϵ j m terms are independently and identically distributed (i.i.d.) and follow the type-i extreme value (Gumbel) distribution. Each passenger in a market selects the product corresponding to the largest random utility across all available products in that market, which leads to the well-known multinomial logit formula for the probability that product j is selected by a passenger in market m given by Pr m (j) = Pr (U m j U m j, j J m, j j) = exp(v m j ) exp (V m j ) where J m denotes the choice set, i.e., the set of all products in market m. The deterministic utility of product j is given by V m j = β 0 P m j + β 1 ln(f m j ) + β 2 ln(f m m j ) I non stop,j m + β a I a,j m + β 3 I non stop,j j J m + β 4 T j m + β h a A V j m depends on the following attributes. P j m : average fare of carrier-route j in market m, calculated as the weighted average (weighted by the number of passengers) of all tickets prices corresponding to that carrier-route and that quarter. F j m : quarterly total service frequency of product j in market m. This variable denotes the total service frequency of the carrier-routes aggregated quarterly. T j m : average value of the total travel time of product j in market m, calculated by taking the average of the scheduled travel time of all itineraries associated with carrier-route j in market m. I m non stop,j: binary dummy variable that identifies the non-stop carrier-routes. This variable is set to 1 for non-stop carrier-routes and is 0 otherwise. H : set of three hub-related identifiers, namely the hub_connection, hub_origin and hub_destination. In other words, H = {hub_connection, hub_origin, hub_destination}. I m h,j : binary dummy variables corresponding to subscript h, namely I m hub connection,j, I m hub origin,j, and I m hub dest,j, which respectively identify whether the connection, origin or destination airport is the main hub of the carrier. It is set to 1 in case of a hub airport and is 0 otherwise. For all the nonstop carrier-routes I m hub connection,j = 0. A: set of carriers. The merged carrier is considered to be different from either of the merging carriers. For example, UA-CO is used to designate the post-merger carrier in case of the merger between UA and CO. I m a,j: binary dummy variable corresponding to carrier a. For example, for the carrier-route DL- BOS-LAX, the dummy for Delta Airlines I m DL,j equals to 1 and all other I m a,j dummies are set to 0. We estimated a separate model for each merger. Table 1 summarizes coefficient estimates in the multinomial logit model. All estimates were found to be statistically significant at 0.5% or lower level, and their signs and relative magnitudes were found to be consistent with our intuition (10). h H m I h,j

8 Tian Luo, Vikrant Vaze 8 4. CONSUMER SURPLUS MODEL After estimating the discrete choice model coefficients, we calculated the expected values of the pre-merger and post-merger consumer surplus in each market as the measure of passenger welfare. This approach enables us to aggregate the merger s impacts on fare, service frequency, travel time and other attributes into a single metric, thus making it possible for us to analyze the overall passenger effects of each merger. The expected value of the consumer surplus in a market can be written as (11,12): E(CS m ) = E [ 1 α m max (U m j, j J m )] = E [ 1 j β 0 max (U m j, j J m )] j where CS m is the consumer surplus of a passenger in market m, α m is the marginal utility of income in market m (equal to the absolute value of the coefficient of P m m j, β 0, in our model), U j is the random utility of product j in market m, and J m denotes the choice set in market m. The expectation is taken over all possible values of ϵ m m j. Although the random utility U j is not observable, it is possible to calculate the expected consumer surplus using the observable utility V m j. Hanemann (13) and Small and Rosen (12) demonstrated that, if all ϵ m j are independently and identically distributed (i.i.d.) and follow type-i extreme value distributions, and utility is linear in income, then the expected consumer surplus can be expressed as E(CS m ) = 1 β 0 ln ( exp(v j m )) + C j J m where C is an unknown constant which represents the fact that the absolute value of utility cannot be measured. This is the well-known logsum formula for expected consumer surplus under the multinomial logit model assumptions. The change in consumer surplus for a specific market is then calculated as the difference between the E(CS m ) value after the merger and the E(CS m ) value before the merger: ΔE(CS m ) = 1 β 0 [ln ( exp (V post,m j )) ln ( exp (V pre,m j ))] j J post,m j J pre,m where the superscripts pre and post refer to before and after the merger respectively. The constant C is canceled out (14) since it appears in expressions both before and after the merger. Note that this expression accounts for not only the change in product attributes, but also for any changes in choice sets themselves. Thus, the effects of entry and exit of carriers in a market, new carrierroutes added to a market and old ones removed from a market are captured in our model. To analyze the impacts of mergers, Difference-In-Differences (DID) estimators are generated to remove the temporal trends as well as effects of events other than the mergers. The Basic DID framework can be expressed as the following simple components-of-variance process (15): y m,t = ω m + d t + γ D m,t + υ m,t (1) y m,t represents the outcome variables of interest (namely, ECS, fare, frequency) in market m and during period t. t is a binary variable which equals 0 in the pre-merger period and 1 in the postmerger period. D m,t is a binary dummy variable which equals 1 only when market m is in the treatment group and t = 1, and 0 otherwise. ω m and d t are market-specific and time-specific components respectively. υ m,t is a market-and-time-specific component with mean equal to zero at each time period. The coefficient γ measures the impact of the merger (treatment).

9 Tian Luo, Vikrant Vaze 9 The only observable variables are y m,t and D m,t. Equation (1) can be rewritten (16) for t = 0,1 as: y m,t = μ + τ D m,1 + δ t + γ D m,t + ζ m,t (2) where ζ m,t = ω m E(ω m D m,1 ) + υ m,t δ = d 1 d 0 μ = E(ω m D m,1 = 0) + d 0 τ = E(ω m D m,1 = 1) E(ω m D m,1 = 0) All of the four parameters in Equation (2), namely μ, τ, δ and γ, can be estimated by Ordinary Least Square (OLS). Specifically, the estimate of parameter γ can be written (16): γ = (E(y m,1 D m,1 = 1) E(y m,0 D m,1 = 1)) (E(y m,1 D m,1 = 0) E(y m,0 D m,1 = 0)), and hence the term difference-in-differences. This is the mathematical expression for differencing out the effects of temporal trends and events that are not results of the merger. We estimated the treatment impact γ with an OLS estimator, with each observation (market m) associated with a weight equal to the average (over the pre-merger and post-merger periods) number of passengers in that market. The results presented in the next section are based on the estimation of γ parameter. The percentage changes in frequency are calculated by dividing by the corresponding pre-merger frequency values. Calculation of the percentage change in frequency is consistent with the utility expression in our passenger choice model which is linear in the logarithm of frequency. 5. RESULTS In this section, we describe our major findings summarizing the effects of the mergers on the passenger welfare. The average per-trip value of consumer surplus change for each market due to each merger is calculated as per the methodology described in Section 5. To get the overall weighted average and weighted sums across the markets, we first calculated the average (over the pre- and post-merger periods) annual number of passengers in each market affected by the mergers by simply multiplying the DB1B-reported 10% sampled number of passengers by 10. Then we used these numbers of passengers as the weights to calculate the weighted averages and weighted sums of consumer surplus changes. Following each merger the total consumer surplus change is as follows: US-HP: +$0.10 Billion [+$0.50], DL-NW: + $9.76 Billion [+$42.80], UA-CO: +$1.99Billion [+$10.40], WN-FL: -$0.12 Billion [-$1.30] (Q1-Q2 only) and AA-US: +$0.53 Billion [+$4.10] (Q3-Q4 only). The numbers listed in square brackets are those on a per-trip average basis. The five most significant findings of this research are summarized below as Key Findings 1 through 5. Key Finding 1 The passenger welfare increases after the mergers of the legacy network carriers (DL-NW, UA-CO and AA-US), and remains almost unchanged when at least one of the merging carriers is a low-cost carrier (US-HP, and WN-FL), owing primarily to the percentage change in service frequency. We categorized US, DL, NW, UA, CO and AA as legacy network carriers; and HP, WN and FL as low-cost carriers. We defined the merger markets as the markets with operations of at least one of the merging airlines. We find that following the DL-NW, UA-CO and AA-US mergers, the average per-trip consumer surplus increased by $42.8, $10.4 and $4.1 respectively. In contrast, consumer surplus remained almost unchanged after the US-HP and WN-FL mergers, with merely

10 Tian Luo, Vikrant Vaze 10 a $0.5 increase and $1.3 drop respectively. While seasonality does play a role in affecting these exact values, the positive effects of the DL-NW, UA-CO and AA-US mergers and the mixed effects of the US-HP and WN-FL mergers were found to be consistent across all quarters of the data used in this analysis. The reasons for such changes become clear when we look at the percentage change in service frequency. After the UA-CO merger, the total service frequency (nonstop and one-stop combined) per market increased by 4.6% on average. Same service frequency change (4.6%) was found after the DL-NW merger, while after the AA-US merger, service frequency increased by 2.5%. In contrast, US-HP merger resulted in only a 1.2% service frequency increase and the WN-FL merger was followed by a 2.5% decrease in service frequency. Change in airfare also played an important role in determining the passenger welfare after these mergers. Although the DL-NW and UA-CO mergers experienced the same service frequency increase of 4.6% after merger, consumer surplus increased by $42.8 after the DL-NW merger, which is approximately 4 times as large as that after the UA-CO merger. This can be explained by the difference in fare changes after these two mergers, with an average $5.5 increase after the UA- CO merger compared to an average $12.4 drop after the DL-NW merger. The welfare consequences following the WN-FL merger of low-cost carriers are not surprising given their point-to-point mode of operations, which made it relatively difficult for them to increase service frequency through reorganization of their joint network. The almost unchanged welfare following the US-HP merger likely stems from their highly segregated pre-merger networks, with US largely operating in the eastern part of the United States, while HP operating mostly in the western part. We calculated the percentage of overlapping markets (defined as the number of markets in the treatment group with pre-merger operations of both merging airlines divided by the total number of markets in the treatment group) for each merger. US-HP had only 8.6% overlapping markets, but DL-NW, UA-CO, WN-FL and AA-US had 35.2%, 42.2%, 22.7% and 52.7%, respectively. This low overlap in their networks rendered the mechanism of service frequency increases through network reorganizations (described in more details in the next Key Finding) inapplicable for the US-HP merger, thus causing no welfare gains. We observe that the same argument holds, to less extent, for the WN-FL merger as well. Key Finding 2 Hub-and-spoke carriers reorganized their joint networks, after a merger, by reinforcing the hub airports of the primary carriers, with more nonstop flights diverted to these hubs. Aside from an increase in service frequency, this network reorganization also resulted in significant welfare gain for passengers using these hubs as origins or destinations. We defined hub airports as the former designated hubs of each merging carrier before the merger. The primary carriers (as defined by those with larger number of flights) in the US-HP, DL-NW, UA-CO, WN-FL and AA-US mergers are HP, DL, UA, WN and AA, respectively. We find that hubs of primary carriers were considerably strengthened after the mergers. Specifically, average quarterly nonstop service frequency in markets with primary carrier s hubs as origin and/or destination increased by 40, 90, 51 and 66 after the US-HP, DL-NW, UA-CO and AA-US mergers respectively. The corresponding numbers for secondary carrier s hub markets were -74, 77, 46, and -15 respectively. Previous research has shown that this increased concentration of traffic to/from hubs of primary carriers potentially improves the efficiency of the joint network (17). We found that it also led to an increase in service frequency. Another related finding is that, primary carriers hubs benefited more than the secondary carriers hubs. In the US-HP merger, passengers using HP hubs as origin and/or destination had a consumer surplus increase by $5.3, whereas for US hubs it declined by $3.4. Similarly, UA-CO merger resulted in a $33.1 increase in consumer surplus for passengers with origin and/or

11 Tian Luo, Vikrant Vaze 11 destination at UA hubs but a $25.8 decrease for CO hubs. Even though both DL and NW hubs experienced welfare increases, for passengers to/from the primary carrier s (DL) hubs gained $52.7 whereas those to/from the secondary carrier s (NW) hubs gained only $36.4. Similarly, for the AA-US merger, the corresponding numbers were +$24 for AA hubs but only +$4.3 for US hubs. Since WN does not have designated hubs, we did not include the WN-FL merger in this particular analysis. Key finding 3 When comparing the mergers of hub-and-spoke carriers with significant proportion of pre-merger overlapping markets, the consumer welfare gains decreased as the number of existing major domestic carriers decreased. There is a monotonic relationship between the welfare increases following the DL-NW, UA-CO and AA-US mergers and the number of major domestic carriers just before the time of the merger. From the merger of DL-NW, to that of UA-CO, to the most recent merger of AA-US, the welfare gains declined from $42.8 to $10.4 to $4.1. At the same time, the number of major domestic carriers (defined as mega carriers in Section 1) before each merger, decreased from 7 to 6 to 5. This positive relationship between the number of major domestic carriers and the welfare gain due to the merger implies that passengers gained more from mergers of hub-and-spoke airlines when there were more major domestic carriers. We note that this finding should be treated with some caution because it relies only on 3 data points. However, it is consistent with previous theoretical work (1), which suggests that as the number of carriers decreases, the welfare gains decline, and finally reach an equilibrium where any further mergers would result in zero or even negative welfare gains. Key finding 4 For very small communities, the DL-NW and UA-CO mergers resulted in welfare losses to passengers. Before defining very small communities, we first categorized the airports within the 48 contiguous (i.e., excluding Alaska and Hawaii) U.S. states into 4 classes: large, medium, small and very small, as per the definitions of large hub, medium hub, small hub, and non-hub airports respectively as defined by the Federal Aviation Administration (FAA) (18). As per the FAA, a large hub is defined as an airport accounting for at least 1% of national annual passenger boardings. We simply call them large airports (to avoid confusion over the term hub ). Airports contributing between 0.25% and 1% of the national annual passenger boardings are categorized as medium hubs ( medium airports in our terminology), while those accounting for between 0.05% and 0.25% are denoted as small hubs ( small airports in our terminology). Finally, airports accounting for less than 0.05% of the national annual passenger boardings are defined as non-hubs ( very small airports in our terminology). Using the above criteria, 14 of the total 141 airports in our study are categorized as very small airports, and the number of small, medium and large airports is 56, 40 and 31, respectively. For our analysis, markets with origin and/or destination at a very small airport are denoted as markets serving very small communities. We observed that the merging airlines in only the DL-NW, UA-CO and AA-US mergers operated in any markets serving very small communities and the number of markets serving very small communities in our AA-US merger data is too small (18 total markets including 6 in the control group) to draw any reliable conclusions. In comparison, for the UA-CO merger, there were 94 total markets serving very small communities including 70 in the control group, and for the DL- NW merger there were 140 total markets serving very small communities including 38 in the control group. Therefore, our analyses of merger impacts on very small communities were limited to the DL-NW and UA-CO mergers. In the markets serving the very small communities, the DL- NW and UA-CO mergers lead to per-trip welfare losses of $69.2 and $27.0, respectively.

12 Tian Luo, Vikrant Vaze 12 Interestingly, very different welfare consequences were observed in markets serving slightly larger communities. We define as markets serving small communities those markets which have origin and/or destination at a small airport. All three mergers between two legacy network carriers, namely DL-NW, UA-CO and AA-US, were followed by a per-trip welfare gain of $70, $18 and $39.3, respectively, in the markets serving small communities in contrast to per-trip welfare losses of $35.2 and $29.7 after the US-HP and WN-FL mergers, respectively. Welfare change in markets serving large communities (defined as those markets where the origin and/or destination is a large airport) is similar to the overall impacts after each merger. Per-trip welfare in markets serving large communities increased by $43.6, $9.9 and $3.6 respectively after the DL-NW, UA-CO and AA- US mergers, and remained almost unchanged following US-HP (increased by $0.8) and WN-FL (decreased by $1.7) mergers. These results for the small and very small communities are mostly driven by the fare changes in those respective markets. The welfare losses in markets serving very-small communities were accompanied by fare increases following both the DL-NW and UA-CO mergers. In markets serving small communities, for four out of the five mergers, the welfare consequences were mostly driven by fares changes, with fare increases causing welfare losses and fare decreases causing welfare gains. The only exception to this was the welfare gain despite the fare increase following the AA-US merger. This welfare gain was driven by a significant service frequency increase (22.1%), which was at least twice as big as that for any other of the five mergers. Key finding 5 Passengers from the markets with low or moderate market concentration experienced welfare losses after all five of the mergers. We used the Herfindahl Hirschman Index (HHI) as the measure of market concentration for each market, following the conventional definition of HHI as the square of sums of the market shares of all carriers in that market (19). We define market share as the number of passengers of the carrier in the market divided by the total number of passengers in that market. The markets were then characterized as high concentration (with post-merger HHI of at least 0.18) or moderate/low concentration (with post-merger HHI smaller than 0.18), which includes markets with both low and moderate concentrations as per the definition by the Department of Justice (19). The threshold of 0.18 is chosen for consistency with the definition of Department of Justice and to have slightly more balanced number of markets in the two categories created due to the threshold. After the mergers, the welfare consequences for high concentration markets were very similar to the corresponding trend in overall welfare changes, with increases of $2, $44.2, $12.5, $0.1 and $4.5 for US-HP, DL-NW, UA-CO, WN-FL and AA-US mergers, respectively. However, for the markets with low and moderate concentrations, the five mergers resulted in welfare losses of $44, $61.4, $74.1, $35.7 and $9.1, respectively. From the results, we see that for highly concentrated markets the changes in welfare, fares and service frequency are very similar to the general changes for all markets as analyzed previously in this section. All three of them, namely, passenger welfare, average fares and average service frequencies, almost always go up following a merger. This similarity is not a coincidence, since approximately 95% of all passengers travel in markets with high concentration. This implies that the impacts of, as well as the possible mechanisms behind the mergers discussed in this section so far are essentially for the highly concentrated markets, and passengers on these markets account for the majority of the total traveling population. Also noticeable is that for the low and moderate concentration markets, the consequences for welfare, fares as well as service frequency are very different. Following the mergers, passengers in these markets experienced fare increases (in all cases), service frequency decreases

13 Tian Luo, Vikrant Vaze 13 (in all but one cases), and consequently welfare losses in all cases. The low and moderate concentration markets are very dense markets being served by a large number of competitors. Their characteristics, as compared to those of the high concentration markets, are much closer to the classic definition of perfectly competitive markets. It appears that in these moderate and low concentration markets with a lot of competitors, the reduction in competition due to the mergers does have significant negative effects on prices and service quality as expected by classic economic theories. Moreover, many carriers in these major markets generally offer non-stop services, the frequency of which may not be significantly increased through network reorganizations, unlike one-stop frequency increases. Combining these two factors, these markets experience a significant welfare loss following all five mergers. 6. CONCLUSIONS In this study, we examined the welfare consequences of mergers of major airlines by incorporating effects of fare, service frequency, travel time as well as other relevant attributes of the carriers and routes. Previous research on impacts of mergers focused primarily on the fare effects while ignoring other attributes such as service frequency. We hypothesized that service frequency (for both nonstop and one-stop carrier-routes) plays a very important role in determining the welfare of passengers. Consistent with our hypothesis, we find that service frequency, along with fare, travel time and other attributes, determines the change in consumer welfare following a merger. In general, the merging carriers that operate in a hub-and-spoke mode, are likely to reorganize their joint network following the merger. The network reorganization results in the strengthening of the hub airports of the primary carriers, with more flights channeled through them. For merging airlines with significant percentage (more than 30%) of overlapping markets, the network reorganization (diversion of more flights to the primary merging carrier hubs) leads to a significant increase in service frequency, especially the one-stop service frequency, for the merged carrier. As a result, the effects of increasing service frequency outweigh those brought about by increased fares, leading to net welfare gains for passengers. Compared to the theoretical work of Brueckner and Spiller (20,21), which concluded that consumers may benefit from the efficiency gains and cost reductions achieved by the merging carriers, we stress that the post-merger reorganization of hub-and-spoke networks was found to especially benefit the passengers through increased service frequency. As for the mergers between legacy network carriers, aside from the attributes of the merging airlines such as fare and service frequency, the amount of welfare gain also depends on the existing number of major domestic carriers. From the DL-NW to UA-CO to AA-US merger, the welfare gains shrank along with the decrease in the number of major domestic U.S. carriers. The welfare increase due to the most recent AA-US merger was merely $4.1, only about 10% of that of the DL-NW merger. This seems to indicate that the number of major domestic U.S. carriers after the most recent merger is close to the equilibrium number and further consolidation may not lead to welfare increases or may even result in welfare losses. The mergers of US-HP and WN-FL, where at least one merging carrier was a low-cost carrier, had a negligible effect on passenger welfare. One possible reason is that the networks of US and HP before the merger were segregated with a very small overlap, with HP network mostly in the west and US network mostly in the east, sharing only 8.6% overlapping markets before merger. This segregation of networks likely renders the mechanism of channeling more traffic through main hubs inapplicable for US-HP merger (and to a lesser extent for the WN-FL merger with 22.7%

14 Tian Luo, Vikrant Vaze 14 overlapping markets); while for the other three mergers, it is this mechanism that significantly increased the service frequency, therefore contributing to an increase is passenger welfare. Finally, we note that most aviation markets are highly concentrated in that they are dominated by a few carriers. Mergers in such markets typically yield service frequency increases that more than compensate for any fare increases and result in net welfare gains. However, in the few markets that are fiercely competed for by multiple major carriers using non-stop service, a situation sharing some characteristics with the perfect competition setting, mergers led to fare increases and service frequency reductions, resulting in welfare losses.

15 Tian Luo, Vikrant Vaze 15 REFERENCES 1. Bailey, E., Liu, D., Airline consolidation and consumer welfare. Eastern Economic Journal, Vol. 21, No. 4, 1995, pp Borenstein, S., Airline mergers, airport dominance, and market power. American Economic Review, Vol. 80, No. 2, 1990, pp Kim, E., Singal, V., Mergers and market power: evidence from the airline industry. American Economic Review, Vol. 83, No.3, 1993, pp Morrison, S., Airline mergers: A longer view. Journal of Transport Economics and Policy, Vol. 30, 1996, pp Veldhius, J., Impacts of the Air France-KLM merger for airlines, airports and air transport users. Journal of Air Transport Management, Vol. 11, 2005, pp Peters, C., Evaluating the performance of merger simulations: evidence from the U.S. airline industry. The Journal of Law and Economics, Vol. 49, 2006, pp BTS, RITA BTS Transtats., Accessed June 28, BTS, RITA BTS Transtats., Accessed June 28, BTS, RITA BTS Transtats., Accessed June 28, Luo, T., Impacts of airline mergers on passenger welfare. M.S. Thesis, Thayer School of Engineering, Dartmouth College, McFadden, D., Modelling the choice of residential location. In Karlqvist, A., Lundqvist, L., Snickars, F. and Weibull, J. (eds) Spatial Interaction Theory and Residential Location. North-Holland, Amsterdam. 1978, 12. Small, K., Rosen, H., Applied welfare economics with discrete choice models. Econometrica, Vol. 49, No. 1, 1981, pp Hanemann, W., A methodological and empirical study of the recreation benefits from water quality improvement. Ph.D. Dissertation, Department of Economics, Harvard University, 1978, 14. Train, K. Discrete Choice Methods with Simulation, Cambridge, New York, Ashenfelter, O., Card, D., Using the longitudinal structure of earnings to estimate the effect of training programs. The Review of Economics and Statistics, Vol. 67, 1985, pp Abadie, A., Semiparametric difference-in-differences estimators. Review of Economic Studies Rev Econ Studies, Vol. 72, 2005, pp Bilotkach, V., Fageda, X., Flores-Fillol, R., Airline consolidation and the distribution of traffic between primary and secondary hubs. Regional Science and Urban Economics, Vol. 43, No. 6, 2013, pp FAA, Airport Categories., Accessed June 26, DOJ, 1.5 Concentration And Market Shares., Accessed July 02, 2016.

16 Tian Luo, Vikrant Vaze Brueckner, J., Spiller, P., Competition and mergers in airline networks. International Journal of Industrial Organization, Vol. 9, No. 3, 1991, pp Brueckner, J., Spiller, P., Economies of traffic density in the deregulated airline industry. Journal of Law and Economics, Vol. 37, No. 2, 1994, pp LIST OF TABLES TABLE 1 Multinomial Logit Estimation Results Coefficients\Mergers US-HP DL-NW UA-CO WN-FL AA-US FARE FARE_Q FARE_Q FARE_Q LOG_FREQ LOG_FREQ_NONSTOP NONSTOP TRAVEL_TIME HUB_CONNECTION HUB_ORIGIN HUB_DESTINATION DL NW WN AA US HP UA CO FL B AS VX US+HP DL+NW UA+CO WN+FL AA+US