Mergers and Product Quality: A Silver Lining from De-Hubbing in the U.S. Airline Industry

Transcription

1 Mergers and Product Quality: A Silver Lining from De-Hubbing in the U.S. Airline Industry Nicholas G. Rupp Kerry M. Tan April 2017 Abstract This paper investigates how de-hubbing, which occurs when an airline ceases hub operations, impacts product quality. Examining five cases of de-hubbing following U.S. airline mergers between 1998 and 2015, we analyze how three measures of product quality change following de-hubbing: on-time performance, travel time, and flight cancellations. While state attorney generals have opposed recent airline mergers due to concerns about potential job losses and a reduction in airport service following airport de-hubbing, we find a silver lining from mergers as airlines offer improved product quality at de-hubbed airports due to more reliable flight schedules and shorter travel times. JEL classifications: L15, L93 Keywords: De-Hubbing, Product Quality, Legacy Carriers, Mergers, Airline Competition We would like to thank Gary Fournier, Jonathan Williams, Zhou Zhang, as well as seminar participants at the U.S. Department of Justice and conference participants at the International Industrial Organization Conference and Southern Economic Association Conference. This research was supported by a Summer Research Grant from the Sellinger School of Business and Management at Loyola University Maryland. Department of Economics, East Carolina University, Greenville, NC, 27858; ruppn@ecu.edu. Department of Economics, Loyola University Maryland, Baltimore, MD, 21210; kmtan@loyola.edu. 1

2 1 Introduction Our hub in Cleveland hasn t been profitable for over a decade, and has generated tens of millions of dollars of annual losses in recent years. We simply cannot continue to bear these losses. (former United CEO Jeff Smisek s letter to United s Cleveland employees, 1 February 2014). A series of bankruptcies and mergers in the airline industry since 2004 has trimmed the number of major U.S. airlines from ten down to just four large carriers which dominate the domestic market. These mergers have caused carriers to re-evaluate their flight network structure in search of efficiency gains. A frequent consequence of an airline merger is the decision to reduce the number of hub airports, a phenomenon called de-hubbing. Since 1998, there have been five cases in the U.S. where an airport has been de-hubbed after an airline merger. This paper examines what impact these de-hubbing events have had on product quality. This research also has public policy implications given the testimony by Gerald L. Dillingham, the Director of Civil Aviation for the Government Accounting Office (GAO), to the Subcommittee on Aviation Operations regarding the U.S. Airways - American Airlines merger indicated that the GAO s evaluation of the merger is to determine if the potential benefits for consumers outweigh the potential negative effects (GAO T). Hence, this paper sheds some light on whether airline mergers and subsequent de-hubbing of an airport improves (or worsens) product quality for consumers. Airlines seek to merge for both revenue and cost reasons. From the revenue viewpoint, the merged entity now has a more valuable network of flight offerings and the potential for having greater market power from having one less competitor in overlapping markets. For example, for the recent American and US Airways merger completed in 2015, US Airways estimates that the financial benefits to shareholders following the consummation of the American and US Airways merger provide $1.4 billion in annual benefits with the majority of these benefits coming from additional revenue due to improvements in network connectivity, a more valuable frequent flier network, and optimization in the use of aircraft (GAO T). From a cost standpoint, a merged 2

3 company can also provide potential cost savings. Once again, examining the American and US Airways merger, US Airways executives expect to generate $640 million in annual cost savings by reducing or eliminating duplicative operating costs, including inefficient (or redundant) hubs or routes (GAO T). Due to the frequency in which de-hubbing of airports has occurred following an airline merger, the state Attorney Generals of Arizona, Florida, Michigan, Tennessee, Pennsylvania, Virginia, and District of Columbia filed lawsuits in opposition to the US Airways - American Airlines merger out of fear that their state could potentially suffer substantial job losses at existing hub airports. As a condition for allowing American and US Airways to merge, the AMR Corporation reached a settlement agreement with the Department of Justice and state Attorney Generals to maintain its hubs in Charlotte, New York (Kennedy), Los Angeles, Miami, Chicago (O Hare), Philadelphia, and Phoenix consistent with historical operations for a period of three years (AMR Corporation press release, 12 November 2013). Hence, the possibility still exists that after three years (beginning in 2016:Q4) that American may close one of its existing hub airports. Consequently this research on product quality following a de-hubbing event is of particular interest for travelers at existing American hub airports. There has been considerable research into the so-called hub premium as the network carriers have been shown to charge higher prices to and from hub airports for a variety of reasons, including increased market power (Borenstein, 1989), frequent flier programs (Lederman, 2008), and the mixture of leisure/business passengers (Lee and Luengo-Prado, 2005). More recently, researchers have also investigated the impact of an airport de-hubbing on fares and flight operations. Examining de-hubbing at U.S. airports, Tan and Samuel (2016) find lower fares following the de-hubbing at airports that have a low-cost carrier presence, while higher fares occur after de-hubbing at airports without a low-cost carrier presence. Examining thirty-seven instances of de-hubbing events worldwide, Redondi et al. (2012) find that the typical de-hubbed airport does not recover their original traffic level after five years. 3

4 There has been some recent research exploring the link between airline mergers and product quality. Prince and Simon (2015) find that airline mergers have minimal impact on quality (on-time performance) immediately following the merger, while documenting some evidence of long-run improvements in service quality (between three and five years) after the merger. Our work differs from Prince and Simon (2015) since they examine all flight operations throughout the U.S. following a merger, while this paper focuses solely on the product quality at the de-hubbed airport after a merger. Chen and Gayle (2013) investigate the directness of the itinerary routing as their quality measure following the Delta/Northwest and Continental/United mergers. They find a decrease (increase) in product quality post-merger if the merging firms were competitors (not competitors) in the market. Beyond the airline industry, others have examined the impact on product quality following mergers and acquisitions. The empirical results show considerable variation. For example, using a structural model of convenience store expansion in Japan, Nishida and Yang (2015) report that mergers have a detrimental effect on the underlying unobserved performance dynamics of the merged entity, whereas others have found a neutral effect on quality following a merger. Examining data from Consumer Reports across a variety of brands (e.g., washing machines or vacuum cleaners), Sheen (2014) finds that when two manufacturers of a given product merge, the product quality of their products converge after a two to three year period. There have also been numerous studies which document improvements in product quality and firm performance following a merger. Reviewing 10-K product descriptions following mergers, Hoberg and Phillips (2010) find that the merged entity has improved operational performance as evidenced by an increase in the creation of new products which offer greater product differentiation compared to its rivals. In the Japanese cotton spinning industry, Braguinsky et al. (2015) find that there is a noticeable improvement in the acquired plants productivity and profitability once new management/ownership took control. Maksimovic and Phillips (2001) show that transfer of corporate assets (by mergers, acquisitions, or plant sales) typically improve the allocation of resources and hence, result in an increase in productive efficiency. McGuckin and Nguyen (1995) 4

5 examine more than 28,000 plant ownership changes in the U.S. food manufacturing industry over an eleven year period (in 1970s-80s) and find that a plant ownership change is associated with an improvement in the acquired plants productivity. More recently, Gugler and Siebert (2007) find that mergers in the semi-conductor industry are associated with net efficiency gains. In sum, this paper explores the impact on product quality and mergers by investigating the on-time performance before and after the merged airline de-hubs an airport. We focus on five cases where U.S. airports have been de-hubbed following an airline merger since While the number of airports served typically falls following airport de-hubbing, we find a silver lining since mergers improve product quality as travelers at de-hubbed airports experience more reliable flight schedules and shorter travel times. Therefore, policymakers and state Attorney Generals should consider the efficiency gains from more reliable flight schedules when calculating the costs and benefits from a proposed airline merger. 2 Data The paper utilizes three data bases provided by the Bureau of Transportation Statistics - Airline On-time Performance Data, Airline Origin and Destination Survey (DB1B), and the T-100 Domestic Market data.). 1 The on-time data provide information pertaining to on-time service quality, including domestic flight schedules, origin and destination airports, operating carrier, flight delays, and cancellations. Second, the DB1B data is a 10 percent sample all domestic airline tickets by the reporting airlines. These data provide flight itinerary details including the airline ticket price and passengers transported. Finally, the T-100 data provide information on the number of number of departures along with seating capacity. Since the DB1B data are quarterly observations, we aggregate all three data sets to the quarterly level. Hence, each observation represents a carrier at the route level for each quarter and year. Our data span the period from the 1998:Q1 to 2016:Q4. 1 These data can be downloaded at: 5

6 Our sample includes flights within the contiguous United States for the ten largest US carriers (based on passengers served) during our sample period (in alphabetical order): AirTran Airways, Alaska Airlines, American Airlines, Continental Airlines, Delta Air Lines, JetBlue Airways, Northwest Airlines, Southwest Airlines, United Airlines, and US Airways. In an effort to be consistent with prior related work, we use a similar de-hubbing definition as Tan and Samuel (2016) and Redondi et al. (2012). We compare the performance of the carrier in the before period (four quarters preceding the de-hubbing event) with the post period (four quarters following the de-hubbing event), while excluding the transitional quarter in which de-hubbing occurs. As such, a de-hubbing event occurs in a particular quarter if there is a 30% reduction in the total number of flights compared to the time period prior to de-hubbing. The exclusion of the de-hubbing quarter is necessary since as noted by Jeff Smisek, the former CEO of United Airlines, we have made the difficult decision to substantially reduce our flying from Cleveland. We will make this reduction in stages... in the same 2014 letter to United s Cleveland employees as quoted in the introduction of this paper. In addition to the substantial reduction in the number of flights offered (30%), we also require that the airport be of sufficient size to be considered a hub facility, therefore our second criteria for inclusion as a de-hubbing is that the airport must be among the 50 largest airports in the United States based on the number of enplanements. 2 Table 1: List of Mergers and Related De-Hubbed Airports Merger Airport Ranking Acquirer Acquired Merger Announced Airport De-Hub Date Before After American Reno Air 1999:Q1 SJC 2001:Q American TWA 2001:Q2 STL 2002:Q America West US Airways 2005:Q2 LAS 2008:Q3 7 7 Delta Northwest 2008:Q2 MEM 2012:Q United Continental 2010:Q2 CLE 2015:Q2 47? Table 1 presents the five major mergers in the U.S. airline industry in chronological order of when the merger was publicly announced, as well as the airport that was de-hubbed as a result of the 2 The airport rankings can be found at: allcargo_stats/passenger/. 6

7 merger: Norman Y. Mineta San Jose International Airport (SJC), Lambert-St. Louis International Airport (STL), McCarran International Airport (LAS), Memphis International Airport (MEM), and Cleveland Hopkins International Airport (CLE) 3. Figures 1-5, which are located at the end of the paper, provide a visual representation of the volume of flights at each of the five de-hubbing airport. In each of these graphs we have indicated the number of flight operations by the acquirer airline (red dotted line), the acquired airline (blue dashed line), and both airlines combined (black solid line) for our sample period of 1998:Q1 to 2016:Q4. We have denoted the quarter in which de-hubbing occurs by using a vertical gray dashed line. The shaded regions demarcate the before and after de-hubbing periods used in the regression analysis. As expected, Figures 1-5 shows a substantial drop in the number of flights surrounding the de-hubbing event and in most instances during the four quarters after de-hubbing occurs, the number of flight operations continues to fall. 3 Empirical Analysis Our empirical approach to estimate the effect of the de-hubbing on product quality is to conduct a difference-in-differences (DID) estimation. The advantage of this DID specification is that it enables us to conduct a before and after comparison of how de-hubbing effects product quality in hub markets. We also use a difference-in-difference-in-difference (DDD) estimation to determine how the de-hubbed airline performs in comparison with rival airlines at the same de-hubbed airport. The regression results generally imply that de-hubbed airports following mergers typically experience an increase in product quality due to more reliable flight schedules. While our data span nineteen years, for each of the five de-hubbing cases, we restrict the sample to just eight quarters: only including the four quarters before and after the de-hubbing event. Since Figures 1-5 suggests that de-hubbing does not occur abruptly and immediately, the transitional quarter in which the de-hubbing occurs is also omitted. The result is a data set of 277,792 quarterly 3 The FAA has not released its 2016 airport rankings as of the time of this writing. 7

8 observations involving 8,875 routes. Summary statistics appear in Table 2. Approximately onefifth of flights in the sample arrived late (15+ minutes after the scheduled arrival time). The average travel time, which represents the difference between the actual arrival time and scheduled departure time, is 165 minutes. Flight cancellations are somewhat rare events occurring in just 1.4% of the sample. The average one-way fare is $ These are similar to the summary statistics in related papers. Table 2: Summary Statistics Variable Definition Mean (Std. Dev.) pdelay i jt Proportion of flights with delayed arrivals for carrier i on route j in time period t (0.111) traveltime i jt Average number of minutes (actual arrival time - scheduled departure time) for carrier i to fly route j in time period t (84.873) pcancel i jt Proportion of cancelled flights for carrier i on route j in time period t (0.032) origin f lights jt Number of flights at origin airport of route j in time period t 18, Note: lnorigin f lights = ln(origin f lights) (23,118.22) dest f lights jt Number of flights at destination airport of route j in time period t 18, Note: lndest f lights = ln(dest f lights) (23,109.17) ncom jt Number of carriers operating on route j in time period t (1.202) marketshare i jt Market share for carrier i on route j in time period t Note: marketshare for monopolist = 1.0 (0.308) yield i jt Yield for carrier i on route j in time period t price distance (0.312) Note: yield = price i jt Average one-way fare for carrier i on route j in time period t (71.84) distance j One-way distance (in miles) between the endpoints of route j ( ) Routes Number of routes in the sample 8,875 N Number of observations 277,792 In both the DID estimations (Tables 3, 5, and 6) and DDD estimations (Tables 7, 8, and 9), we employ route-carrier fixed effects and year-quarter fixed effects to determine the relationship between de-hubbing and product quality. We consider three different dependent variables: (1) the proportion of flights with delayed arrivals (pdelay), (2) the average travel time in minutes 8

9 (traveltime), 4 and (3) the proportion of cancelled flights (pcancel). Explanatory variables include airport congestion measures: the natural log of the number of flights at the route s origin airport (lnorigin f lights) and the route s destination airport (lndest f lights). There are also two competitive measures in the estimations: a count of the number of carriers serving the route (ncom) and route-level market share by carrier (marketshare). As a proxy for route profitability, we include route yield for the carrier (yield), which is constructed by dividing the average one-way airfare by flight distance. The airport indicator variable equals one if the de-hubbed airport is either the origin or destination airport, and zero if neither endpoint airport is the de-hubbed airport and at least one of the endpoint airports is ranked as a top 50 airport in the year that de-hubbing occured. 5 A second indicator variable dehub takes the value of one during the post-de-hubbing period, and zero during the pre-de-hubbing period. Hence, the interaction term airport dehub only has the value of one at the de-hubbed airport during the post-de-hubbing period. The following difference-in-differences (DID) specification is used in our analysis: y i jt = α + δ i j + τ t + β 1 X i jt + β 2 airport j + β 3 dehub t + β 4 (airport j dehub t ) + ε i jt, (1) with y i jt being the proportion of delayed arrivals (pdelay i jt ), or average minutes of travel time (traveltime i jt ), or the proportion of flight cancellations (pcancel i jt ) for airline i along route j at time t, δ i j represents the carrier-route fixed effects, τ t represents year-quarter fixed effects, and X i jt indicates the presence of additional explanatory variables previous mentioned. Since the airport variable does not vary independent of the carrier-route fixed effects, it becomes absorbed in the estimation. In a similar fashion, the dehub variable is absorbed with the inclusion of year-quarter fixed effects. Standard errors are clustered at the carrier-route level due to the potential of correlation between the carriers on the route over time. We are not interested in the airport and dehub variables since they represent the treatment variable and time variable in our 4 The traveltime variable is defined as the actual arrival time - scheduled departure time and hence, this measure cannot be manipulated by the carriers. 5 Since Rupp and Holmes (2006) find flight cancellations are more prevalent during adverse conditions at the origin airport, the airport variable indicates whether the de-hubbed airport is the origin airport of the route only in the pcancel estimation. Regardless, removing this constraint produces qualitatively similar results. 9

10 DID specification, respectively. The key variable of interest is the interaction term airport dehub. Should we find a negative and statistically significant coefficient for the interaction term, this would suggest that product quality improves following the de-hubbing event (recall that a negative value represents a reduction in the flight delays, or shorter travel times, or a reduction in flight cancellations at departure). For sake of clarity, we note that, for each of the five de-hubbed airport events, we run separate regressions; hence, airport and dehub dummies represent are unique to the individual de-hubbing event. Table 3: Difference-in-Differences Estimation Results Proportion of Delayed Arrivals (pdelay) SJC STL LAS MEM CLE (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) airport dehub (0.007) (0.008) (0.004) (0.004) (0.004) (0.004) (0.008) (0.008) (0.012) (0.007) lnorigin f lights (0.003) (0.003) (0.005) (0.004) (0.005) lndest f lights (0.003) (0.003) (0.005) (0.004) (0.004) ncom (0.002) (0.001) (0.002) (0.001) (0.001) marketshare (0.009) (0.009) (0.012) (0.010) (0.009) yield (0.008) (0.012) (0.012) (0.012) (0.009) N 27,526 25,623 26,786 24,406 29,137 26,478 28,476 25,745 30,164 26,951 Note: All DID regressions follow the specification from (Equation 1) and include both carrier-route and year-quarter fixed effects which are not reported. Standard errors, in parentheses, are clustered by carrier-route. indicates significance at 10% level. * indicates significance at 5% level. ** indicates significance at 1% level. Using the proportion of delayed arrivals (pdelay) as the dependent variable, Table 3 presents the results for the DID regressions for each of the five de-hubbed airports. We report two estimations for each de-hubbing event. On the left of each column, we include results for the interaction term of interest airport dehub along with both carrier-route fixed effects and year-quarter fixed effects (which are not reported). On the right of each column, we report results for airport dehub and additional controls for flight operations, market competition, and route profitability that have been shown to influence on-time service quality (Rupp et al. (2006)). These estimations also include carrier-route fixed effects and year-quarter fixed effects. The key finding from Table 3 is that the interaction term airport dehub is negative and statis- 10

11 tically significant in seven of ten DID specifications, which suggests that product quality improvements following an airport de-hubbing. Specifically, we find that the standard DID specifications for SJC (Column (1)), STL (Column (3)), and LAS (Column (5)) indicate that de-hubbing contributed to a 4.4%, 2.2%, and 1.4% reduction in the proportion of delayed arrivals, respectively, for all airlines servicing the de-hubbed airport, whereas the estimated coefficient for airport dehub is marginally significant (at the 10% level) for MEM (Column (7)) and statistically insignificant only for CLE (Column (9)). 6 Although some of the competition variables are statistically significant, their inclusion in the regressions did not have a qualitative impact on the estimated coefficient for the airport dehub interaction term. The results in Table 3 for the DID specifications that include competition variables for SJC (Column (2)), STL (Column (4)), LAS (Column (6)), and MEM (Column (8)) suggest that de-hubbing contributed to a 4.4%, 2.5%, 1.2%, and 1.9% reduction in the proportion of delayed arrivals, respectively. Note that the estimate for the airport dehub interaction term associated with MEM is now statistically significant at the 5% level with the inclusion of the competition variables. The estimated coefficient for airport dehub remains negative, yet statistically insignificant for CLE (Column (10)). Regardless, the results generally suggests that one of the positive ramifications for mergers is an increase in product quality through fewer delays and hence, more reliable flight schedules. Table 4: The Number of Spoke Airports Serviced by the De-Hubbed Airline Airport (Airline) SJC (AA) STL (AA) LAS (US) MEM (DL) CLE (UA) Before De-Hubbing After De-Hubbing Percent Change -29.4% -22.7% -50.0% -62.5% -14.3% Note: This table reports the number of spoke airports that the de-hubbed airline serviced. The before de-hubbing time period is the year preceding de-hubbing, whereas the after de-hubbing time period is the year following de-hubbing. 6 We focus our attention on delayed arrivals instead of delayed departures since pilots can make up time while airborne following a delayed departure. Airlines can also pad their scheduled departure and arrival times in order to reduce the likelihood of delays and avoid potential fines from the FAA associated with prolonged delays. In other words, flights can depart behind schedule, yet still arrive at the destination on time. See Rupp (2009) for an in-depth look at flight delays. 11

12 It is also worth noting that a de-hubbed airport also significantly reduces the number of nonstop flight offerings by the de-hubbed airline, as Table 4 shows that the magnitude of this reduction averages 35.8% and ranges considerably from 14.3% (CLE) to 62.5% (MEM). In sum, less congested airports improve schedule reliability and reduce travel time; however, they significantly reduce the number of non-stop offerings by the de-hubbing airline. Table 5: Difference-in-Differences Results Average Minutes of Travel Time (traveltime) SJC STL LAS MEM CLE (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) airport dehub (0.624) (0.638) (0.301) (0.305) (0.360) (0.383) (0.950) (0.976) (1.166) (0.796) lnorigin f lights (0.283) (0.349) (0.608) (0.430) (0.469) lndest f lights (0.239) (0.225) (0.420) (0.346) (0.407) ncom (0.166) (0.116) (0.177) (0.146) (0.139) marketshare (0.892) (0.773) (1.674) (1.042) (1.024) yield (0.725) (1.035) (1.143) (1.173) (0.830) N 27,514 25,261 26,753 24,389 29,123 26,474 28,467 25,742 30,159 26,948 Note: All DID regressions follow the specification from (Equation 1) and include both carrier-route and year-quarter fixed effects which are not reported. Standard errors, in parentheses, are clustered by carrier-route. indicates significance at 10% level. * indicates significance at 5% level. ** indicates significance at 1% level. Table 5 shows the DID regression results following our five cases of de-hubbing with traveltime serving as our quality measure. In a similar fashion as with Table 3, we estimate regressions for each de-hubbed airport with the column on the left including just the fixed effects and airport dehub interaction term, whereas the column on the right includes additional controls for route competition, airport congestion, and route profitability. We are especially interested in the airport dehub coefficient which once again is found to be negative and statistically significant in eight of ten specifications. This result suggests that product quality as measured by travel time falls after an airport has been de-hubbed. Perhaps due to less airport congestion, travelers experience shorter travel times following de-hubbing. The DID results for the basic specifications for SJC (Column (1)), STL (Column (3)), and CLE (Column (9)) suggest that de-hubbing contributed to a 1.6, 1.7, and 3.4 minute reduction in travel time (calculated as actual arrival time - scheduled de- 12

13 parture time), respectively, for all airlines servicing the de-hubbed airport, whereas the estimated coefficient for airport dehub is negative, yet statistically insignificant for LAS (Column (5)) and MEM (Column (7)). Although some of the competition variables are statistically significant, their inclusion in the regressions did not have a qualitative impact on the estimated coefficient for the airport dehub interaction term. The results for the DID specifications that include competition variables for SJC (Column (1)), STL (Column (3)), and CLE (Column (9)) suggest that de-hubbing contributed to a 1.3, 2.1, and 2.1 minute reduction in travel time, respectively. After including the competition variables, the estimated coefficient for airport dehub for MEM (Column (8)) becomes marginally significant (at the 10% level), yet remains statistically insignificant for CLE (Column (10)) after including the competition variables. Nonetheless, the combined regression results reported in both Tables 3 and 5 provide evidence that mergers improve product quality as travelers at de-hubbed airports experience fewer delays and hence, more reliable flight schedules and shorter travel times. Table 6: Difference-in-Differences Results Proportion of Flight Cancellations: (pcancel) SJC STL LAS MEM CLE (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) airport dehub (0.003) (0.003) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) lnorigin f lights (0.001) (0.001) (0.001) (0.001) (0.001) lndest f lights (0.002) (0.000) (0.001) (0.001) (0.002) ncom (0.000) (0.000) (0.000) (0.000) (0.000) marketshare (0.003) (0.002) (0.002) (0.002) (0.001) yield (0.005) (0.015) (0.005) (0.002) (0.002) N 27,526 25,263 26,786 24,406 29,137 26,478 28,476 25,745 30,164 26,951 Note: All DID regressions follow the specification from (Equation 1) and include both carrier-route and year-quarter fixed effects which are not reported. Standard errors, in parentheses, are clustered by carrier-route. indicates significance at 10% level. * indicates significance at 5% level. ** indicates significance at 1% level. Our third and final measure of quality involves flight cancellations. The DID estimations for the proportion of flight cancellations appear in Table 6. Using the same approach, we present 13

14 two regression estimations with just the fixed effects and airport dehub interaction term (on the left of the column) and a second specification that includes controls for competition, airport congestion, and route profitability (on the right of the column). Regardless of the inclusion for the number of airport operations, route competition, and route profitability, we find no significant changes in the proportion of flight cancellations following the de-hubbing event for four of the five airports, which is not too surprising given that cancellations are typically caused by inclement weather conditions or IT system failure. Unlike with delays and travel time, flight cancellations are not likely to be improved with a reduction in airport congestion. The DID approach is used to get a pooled effect of the impact of de-hubbing on product quality; however, a difference-in-difference-in-differences (DDD) approach can tease out whether the dehubbing airline experiences a different effect on product quality compared to its rivals at the same airport being de-hubbed. The general specification for the DDD regressions is as follows: y i jt = α + δ i j + τ t + β 1 X i jt + β 2 airport j + β 3 dehub t + β 4 (airport j dehub t ) + β 5 carrier i + β 6 (carrier i dehub t ) + β 7 (carrier i airport j ) + β 8 (carrier i airport j dehub t ) + ε i jt. (2) Since all of the above variables have been previously defined in the difference-in-difference specification, we refer the reader back to Equation (1) for these definitions. Once again, we include carrier-route fixed effects, which absorb the airport, carrier, and carrier airport variables, while the inclusion of year-quarter fixed effects absorb the dehub variable. Standard errors continue to be clustered at the carrier-route level due to the potential of correlation between the carriers on the route over time. As with the DID approach, our variable of interest is the interaction term airport dehub; however, we are now additionally interested in the difference-in-difference-in-differences estimator carrier airport dehub. Given the inclusion of the carrier variable in our DDD analysis, the interpretation of our original variable of interest differs slightly. Now, a negative and statistically 14

15 significant coefficient for the airport dehub interaction term implies that rival firms product quality increases (due to either a reduction in the proportion of delayed arrivals, a shorter travel time, or a reduction in the proportion of cancelled departures), on average, after the airport has been de-hubbed, whereas this variable represented a pooled effect of the de-hubbing airline and its rival airlines in the DID analysis. Moreover, the carrier airport dehub interaction term is interpreted as the difference between the difference-in-differences estimators for the de-hubbing airline and its rival airlines. As such, a negative and statistically significant coefficient for the carrier airport dehub interaction term implies that product quality for the de-hubbing airline improves relatively more than product quality for rival airlines. We conduct separate regressions for each de-hubbing case so consequently the airport and dehub dummies are specific to each of the five cases of de-hubbing in our sample. Table 7: Difference-in-Difference-in-Differences Results Proportion of Delayed Arrivals (pdelay) SJC STL LAS MEM CLE (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) airport dehub (0.009) (0.010) (0.006) (0.006) (0.004) (0.005) (0.013) (0.013) (0.021) (0.012) carrier dehub (0.003) (0.003) (0.003) (0.003) (0.004) (0.004) (0.003) (0.003) (0.004) (0.004) carrier airport dehub (0.014) (0.015) (0.008) (0.008) (0.009) (0.009) (0.016) (0.016) (0.025) (0.015) lnorigin f lights (0.003) (0.003) (0.005) (0.004) (0.005) lndest f lights (0.003) (0.003) (0.005) (0.004) (0.004) ncom (0.002) (0.001) (0.001) (0.001) (0.001) marketshare (0.009) (0.009) (0.012) (0.010) (0.009) yield (0.008) (0.012) (0.012) (0.012) (0.009) N 27,526 25,263 26,786 24,406 29,137 26,478 28,476 25,745 30,164 26,951 Note: All triple difference regressions follow the specification from (Equation 2) and include both carrier-route and year-quarter fixed effects which are not reported. Standard errors, in parentheses, are clustered by carrier-route. indicates significance at 10% level. * indicates significance at 5% level. ** indicates significance at 1% level. Table 7 shows the DDD estimation results using pdelay as our dependent variable. In this estimation, there are two coefficients of interest: airport dehub interaction term and the carrier airport dehub estimator. First, the airport dehub term shows how rival firms delay rates change at the de-hubbed airport compared to non-de-hubbed airports. The odd number columns 15

16 (1-9) represent specifications for just the fixed effects and interaction terms, while the even number columns (2-10) also include controls for airport congestion, route competition, and route profitability measures. Table 7 shows that the DDD estimations are generally robust to whether we include (or exclude) the additional controls for airport congestion, route competition, and route profitability. Examining the performance of rival carriers at the de-hubbed airport, we find that rival carriers experience significant improvements in the proportion of flight delays in three of the five de-hubbed cases (SJC, STL, and CLE) compared to the other U.S. airports where they operate which were not de-hubbed. Columns (1), (3), and (9) show a reduction in the proportion of flight delays by rival carriers of 3.4%, 1.5%, and 5.6% at SJC, STL, and CLE, respectively. This effect is marginally significant (at the 10% level) at LAS and statistically insignificant at MEM. One possible explanation for these positive spillovers to rival carriers from a de-hubbed airport is that airport congestion is reduced when the hub carrier cuts the number of flights. Similar results are recorded when we include the additional controls for airport congestion, route competition, and route profitability (see the even numbered estimations for the airport dehub interaction term in Table 7). The carrier airport dehub term provides insight into the performance of the de-hubbed airline following the de-hubbing event and these hub airline results are compared to the rival carriers to determine the relative change in performance. As discussed previously, we know that in three of the five de-hubbing events, rival carriers experienced a reduction in delays. We find that the de-hubbed airline at two de-hubbing airports (STL and LAS) registers significantly better on-time performance compared to rival carriers at the de-hubbed airport. Columns (3) and (5) in Table 7 show the magnitude of these changes are 2.8% and 4.3% reduction in the proportion of flight delays at STL and LAS, respectively. These findings combined with the previously discussed results for the rival carriers suggest both rival and hub carriers experience improved service quality. With that said, at de-hubbed airports it is the hub airline that benefits the most in terms of improved service quality. We should also note that at two de-hubbed airports (SJC and MEM) we find no improvement in service quality by the de-hubbed airline. Interestingly, the carrier airport dehub 16

17 is positive and statistically significant for CLE, whereas the airport dehub interaction terms is negative and statistically significant. This suggests that de-hubbing is associated with an increase in the de-hubbed airline s flight delays, whereas rival airlines servicing CLE experienced an improvement in their on-time performance. Similar quantitative and qualitative results also occur for the even column specifications which include additional controls for airport congestion, route competition, and route profitability. We now turn to the triple difference results for travel time. Table 8: Difference-in-Difference-in-Differences Results Average Minutes of Travel Time (traveltime) SJC STL LAS MEM CLE (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) airport dehub (0.646) (0.686) (0.485) (0.483) (0.379) (0.393) (1.588) (1.614) (2,184) (1.368) dehub carrier (0.365) (0.292) (0.270) (0.283) (0.334) (0.305) (0.260) (0.241) (0.491) (0.531) carrier airport dehub (1.302) (1.304) (0.647) (0.652) (0.870) (0.850) (1.903) (1.942) (2.449) (1.624) lnorigin f lights (0.292) (0.348) (0.611) (0.428) (0.465) lndest f lights (0.239) (0.225) (0.421) (0.342) (0.403) ncom (0.163) (0.116) (0.177) (0.144) (0.138) marketshare (0.884) (0.775) (1.685) (1.032) (1.025) yield (0.737) (1.035) (1.142) (1.156) (0.832) N 27,514 25,261 26,753 24,389 29,123 26,474 28,467 25,742 30,159 26,948 Note: All triple difference regressions follow the specification from (Equation 2) and include both carrier-route and year-quarter fixed effects which are not reported. Standard errors, in parentheses, are clustered by carrier-route. indicates significance at 10% level. * indicates significance at 5% level. ** indicates significance at 1% level. Table 8 presents the travel time estimation results for the difference-in-difference-in-differences. The airport dehub interaction term is negative and significant at both STL and CLE, which suggests that rival carriers experience relatively shorter travel times following the de-hubbing event compared to the rival carriers performance at other non-de-hubbed U.S. airports. We follow the same specification as in the previous DDD table, where the odd number columns include fixed effects and interaction terms, while the even number columns include additional controls for airport congestion, route competition, and route profitability. Given that the results are comparable among the two specifications, we focus our attention on the odd number column estimation results. Columns (3) and (9) indicate a reduction in travel time of 1.9 and 7.4 minutes by rival carriers fol- 17

18 lowing de-hubbing at STL and CLE, respectively. These results are consistent with the previous DDD specification for probability of delays since rival carriers at STL and CLE were also significantly less likely to be delayed following de-hubbing. Hence, rivals appear to benefit from a reduction in airport congestion following de-hubbing. The result that rivals benefit from shorter travel time does not hold for every de-hubbing event. The airport dehub interaction term is statistically insignificant for SJC (Column (1)), LAS (Column (5)), and MEM (Column (7)). Turning our attention to the de-hubbed airline performance, the carrier airport dehub estimates in Table 8 reveal that the de-hubbed airline at LAS experienced significantly shorter travel times compared to rival airlines. As in Table 7, the de-hubbed airline at CLE experienced an increase in flight time compared to rival airlines. The remaining three airports (SJC, STL, and MEM) have no significant change in travel times for the de-hubbed airline. Once again, similar results were obtained with the additional controls for airport congestion, route competition, and route profitability controls. Table 9: Difference-in-Difference-in-Differences Results Proportion of Flight Cancellations: (pcancel) SJC STL LAS MEM CLE (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) airport dehub (0.003) (0.003) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) carrier dehub (0.001) (0.001) (0.000) (0.001) (0.001) (0.000) (0.000) (0.000) (0.001) (0.001) carrier airport dehub (0.006) (0.005) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) lnorigin f lights (0.001) (0.001) (0.002) (0.001) (0.001) lndest f lights (0.002) (0.001) (0.001) (0.001) (0.001) ncom (0.000) (0.000) (0.000) (0.000) (0.000) marketshare (0.003) (0.002) (0.002) (0.002) (0.001) yield (0.005) (0.015) (0.005) (0.000) (0.002) N 27,526 25,263 26,786 24,406 29,137 26,478 28,896 25,745 30,164 26,951 Note: Note: All triple difference regressions follow the specification from (Equation 2) and include both carrier-route and year-quarter fixed effects which are not reported. Standard errors, in parentheses, are clustered by carrier-route. indicates significance at 10% level. * indicates significance at 5% level. ** indicates significance at 1% level. Our final DDD specification tracks the proportion of flight cancellations at de-hubbed airports with the results appearing on Table 9. The airport dehub interaction term is negative and sig- 18

19 nificant at STL (-0.3% in Column (3)), which indicates that rival airlines have fewer cancellations at STL compared to other U.S. airports that the rival airline operates. On the other hand, flight cancellations actually increased for rival airlines servicing LAS (0.3% in Column (5)) compared to other comparable U.S. airports that the rival airlines operates. The remaining three de-hubbed airports (SJC, MEM, and CLE) show no significant change in the cancellation rates of rival airlines. Similar results are recorded for odd and even column specifications, which again differ only by the inclusion (or exclusion) of the additional airport and route controls. Examining the performance of the de-hubbed airline the carrier airport dehub estimator shows no statistically significant reduction in cancellations for the de-hubbed airline at any of the five de-hubbed airports. Table 10: List of De-Hubbed Airports Unrelated to Mergers Airport Ranking De-Hub Airline Airport De-Hub Date Before After US Airways BWI 2001:Q Delta MCO 2001:Q Delta DFW 2004:Q2 4 4 US Airways PIT 2004:Q Delta CVG 2005:Q So far we have been focusing on five de-hubbing cases that were a result of a merger in the U.S. airline industry. However, one can argue that our before de-hubbing time period could capture a post-merger effect. In order to tease out the effect of de-hubbing on product quality from the merger effect on product quality. Following the same definition for de-hubbing outlined in Section 2, we identify five de-hubbing cases that are not attributed to a merger and analyze these de-hubbing cases as a robustness check to our main results. Since these airports are unaffected by a merger effect, then we can confidently conclude that de-hubbing improves product quality if we attain qualitatively similar results when analyzing the five de-hubbing cases unrelated to mergers. These five airports (in chronological order of the de-hub date) are Baltimore/Washington International Thurgood Marshall Airport (BWI), Orlando International Airport (MCO), Dallas/Fort Worth International Airport (DFW), Pittsburgh International Airport (PIT), and Cincinnati/Northern Kentucky International Airport (CVG). Table 10 lists these airports, their de-hub date, and airport 19

20 ranking before and after de-hubbing. The cause of these mergers range from bankruptcy (CVG) to weak demand (MCO) to issues with the airport authority (PIT). As with Figures 1-5 for the de-hubbing cases associated with mergers, Figures 6-10, which are also located at the end of the paper, provide a visual representation of the volume of flights at each of the five de-hubbing airports that are unrelated to a merger. In each of these graphs, we have indicated the number of flight operations the de-hubbing airline (black solid line) for our sample period of 1998:Q1 to 2016:Q4. We have denoted the quarter in which de-hubbing occurs by using a vertical gray dashed line. The shaded regions demarcate the before and after de-hubbing periods used in the regression analysis. Similar to Figures 1-5, Figures 6-10 show a substantial drop in the number of flights surrounding the de-hubbing event and in most instances during the four quarters after de-hubbing occurs, the number of flight operations continues to fall. Table 11: Difference-in-Differences Estimation Results Proportion of Delayed Arrivals (pdelay) BWI MCO DFW PIT CVG (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) airport dehub (0.006) (0.006) (0.005) (0.005) (0.003) (0.004) (0.006) (0.006) (0.007) (0.007) lnorigin f lights (0.003) (0.003) (0.003) (0.007) (0.007) lndest f lights (0.003) (0.003) (0.003) (0.007) (0.006) ncom (0.002) (0.002) (0.002) (0.001) (0.001) marketshare (0.011) (0.011) (0.009) (0.013) (0.011) yield (0.012) (0.012) (0.008) (0.009) (0.010) N 27,944 25,650 27,944 25,650 27,395 25,131 27,821 25,083 28,469 25,387 Note: All DID regressions follow the specification from (Equation 1) and include both carrier-route and year-quarter fixed effects which are not reported. Standard errors, in parentheses, are clustered by carrier-route. indicates significance at 10% level. * indicates significance at 5% level. ** indicates significance at 1% level. Table 11 presents the DID regression results using the proportion of delayed flights (pdelay) as the dependent variable for the five de-hubbing cases unrelated to mergers. As with Table 3, the odd numbered columns are associated with the baseline DID regression specification, whereas the even numbered columns include additional controls for flight operations, market competition, and 20

21 route profitability. Four of the five de-hubbing cases are associated with a negative and statistically significant estimate for the airport dehub interaction term. The results in Column (3), Column (5), Column (7), and Column (9) shows that de-hubbing is associated with a 2.0%, 1.4%, 2.3%, and 5.0% reduction in flight delays for MCO, DFW, PIT, and CVG, respectively. However, the results in Column (1) suggest that flight delays increased at BWI following de-hubbing. The inclusion of the competition variables produces qualitatively similar results. Since these results are unaffected by a post merger effect, yet yield similar findings to the results in Table 3, we can mitigate the concerns of our main results being disentangled with a post merger effect. Table 12: Difference-in-Difference-in-Differences Results Proportion of Delayed Arrivals (pdelay) BWI MCO DFW PIT CVG (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) airport dehub (0.007) (0.007) (0.006) (0.006) (0.004) (0.004) (0.013) (0.013) (0.020) (0.021) carrier dehub (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.004) (0.004) (0.004) (0.004) carrier airport dehub (0.013) (0.013) (0.012) (0.012) (0.008) (0.008) (0.015) (0.015) (0.022) (0.022) lnorigin f lights (0.003) (0.003) (0.003) (0.007) (0.007) lndest f lights (0.003) (0.003) (0.003) (0.007) (0.006) ncom (0.002) (0.002) (0.002) (0.001) (0.001) marketshare (0.011) (0.011) (0.009) (0.013) (0.011) yield (0.012) (0.012) (0.008) (0.010) (0.010) N 27,944 25,650 27,944 25,650 27,395 25,131 27,821 25,083 28,469 25,387 Note: All triple difference regressions follow the specification from (Equation 2) and include both carrier-route and year-quarter fixed effects which are not reported. Standard errors, in parentheses, are clustered by carrier-route. indicates significance at 10% level. * indicates significance at 5% level. ** indicates significance at 1% level. We also want to reassure that our difference-in-difference-indifferences specification does not suffer from the same issue of the post-merger effect blending in with our before de-hubbing effect. Table 12 presents the DDD regression results for flight delays at the five de-hubbed airports that are unrelated to mergers. As in Table 7, the results for the rival firms are mixed. Based on the estimates for airport dehub, rival airlines experienced an improvement in on-time performance at DFW (Column (5)) and PIT (Column (7)) relative to other similar airports that they serviced. 21