Analyzing Airline Codesharing Alliances: Impact on On-Time Performance

This report investigates the effects of airline codesharing agreements on on-time performance (OTP), focusing on the alliance between Delta Air Lines, Northwest Airlines, and Continental Airlines. The study examines how these codeshare agreements impact OTP, considering factors such as pre-alliance competition in a market. The research reveals that codesharing improves alliance partners' on-time performance. Moreover, the size of the codeshare effect on OTP depends on the level of competition before the alliance, with the effect being more significant in markets where the partners competed before the alliance. The report also provides background on the history of airline alliances, particularly in the U.S. domestic market, and discusses the potential benefits and drawbacks of codesharing, including its impact on fares and service quality. The paper contributes to the existing literature by evaluating the net impact of a codeshare alliance on partner carriers' OTP and aims to shed light on whether the emergence of airline alliances has made problems related to airline OTP better or worse.

Airline code-sharing and its effects on on-time performance
Jules O.Yimga*
Embry-Riddle Aeronautical University,Department of Business,3700 Willow Creek Road,Prescott,AZ 86301,United States
a r t i c l e i n f o
Article history:
Received 22 May 2016
Received in revised form
30 September 2016
Accepted 2 October 2016
Keywords:
Airline on-time performance
Airline codeshare alliances
JEL classification codes
L13
L93
a b s t r a c t
Although airline on-time performance has always received much attention,we are unaware of any
empirical research that measures the on-time performance effects of domestic airline alliances.In this
study, we empirically investigate the on-time performance effects of the largest domestic alli-
ancedbetween Delta Air Lines, Northwest Airlines and Continental Airlines. We find evidence that code-
sharing improves alliance partners' on-time performance and that the size of the codeshare effect on on-
time performance depends on pre-alliance competition in a market,with the effect being larger in
markets where the partners competed in prior to the alliance.
© 2016 Elsevier Ltd.All rights reserved.
1. Introduction
The public outcry and media coverage that ensued in the 1980s
over increasing air traffic delays attracted congressional attention
on airline on-time performance (OTP). Since 1988, the U.S.
Departmentof Transportation's (DOT)Bureau of Transportation
Statistics (BTS) tracks the on-time performance of domestic flights
operated by large air carriers. It is now mandatory for airlines with
at least one percent of all domestic traffic to disclose flight-by-flight
information on delays (Mayer and Sinai, 2003).
Interestingly, even with this flight-by-flight data disclosure, the
DOT's Office of Aviation Enforcement and Proceedings showed that
the most prevailing consumer air travel complaint in the year 2000,
stems from flight problems namely cancellations,delays and
missed connections.1 In fact, 1 out of 4 flights was either delayed,
canceled or diverted (Rupp et al.,2006).According to Mayer and
Sinai (2003), in 2000, flights that arrived at their destination
within 15 minutes of their scheduled arrivaltime and without
being canceled or diverted, accounted for less than 70 percent.
Given the concerns over OTP and the recent trend ofairline
alliance formation as a dominant feature of the airline industry,a
new and interesting question is:how do airline alliances affect
partners' OTP? Answering this question would shed light on
whether the recent emergence of airline alliances has made prob-
lems related to airline OTP better or worse.
Airline alliances vary from limited cooperation,such as recip-
rocal frequent flyer programs, to more enhanced agreements, such
as code sharing.2 A codeshare agreement(CSA) is a reciprocal
agreementbetween two or more airlines, through which one
airline can sell seats on its codeshare partners' flights using its own
reservation code.3
Airline alliance formation has a long history in the international
air travel market but this practice is a relatively new phenomenon
among U.S.domestic carriers.Since the mid-1990s,major airlines
that serve the U.S.domestic markethave increasingly found it
appealing to form alliances.In 1995,Northwest and Hawaiian Air-
lines announced their intention to create an alliance (Ito and Lee,
2005). In the first half of 1998,the six largest U.S.carriers fol-
lowed suit with their own codeshare alliance proposals,with
Continental Airlines and Northwest Airlines making alliance
announcement in January 1998; Delta Airlines/United Airlines and
American Airlines/US Airways followed in April 1998 (Bamberger
et al., 2001).This practice proliferated with the implementation
of subsequentalliance partnerships such as Alaska/Hawaiian in
October 2001,American West/Hawaiian in October 2002,United/
* Corresponding author.
E-mail address: [e-mail].
1 US Department of Transportation Office ofAviation Enforcementand Pro-
ceedings (2001).USDTOAEP.Feb.2001 p.34.
2 US General Accounting Office (1999).
3 The International Air Transport Association (IATA) uses two-character codes to
identify all airlines; for example the code DL is assigned to Delta Airlines.
Contents lists available at ScienceDirect
Journal of Air Transport Management
j o u r n a lhomepage: w w w . e l s e v i e r . c o m / l o c a t e / j a i r t r a m a n
http://dx.doi.org/10.1016/j.jairtraman.2016.10.001
0969-6997/© 2016 Elsevier Ltd.All rights reserved.
Journal of Air Transport Management 58 (2017) 76e90

US Airways in January 2003 and Delta/Northwest/Continentalin
June 2003, among others.
For illustration purposes,a CSA between Alaska Airlines (AS)
and Hawaiian Airlines (HA) for instance, allows Alaska Airlines
(referred to as the “ticketing carrier” or “marketing carrier”)to
market and sell seats on thousands of flights operated by Hawaiian
Airlines (referred to as the “operating carrier”) and vice-versa.In
this example,Alaska Airlines may place its code (AS) on this Ha-
waiian's flight and selltickets for seats on this flight as if Alaska
Airlines operates the flight. So, this same flight will be listed twice
in computer reservation systems, once under Alaska Airlines' code
(AS) and again under Hawaiian Airlines'code (HA).Therefore,un-
der a CSA,partner airlines are able to expand their flight offerings
without adding planes.
We make the following argument.Alliance partners typically
coordinate in an attempt to achieve seamless integration of their
route networks, which potentially results in more travel-
convenient routing across partner carriers'networks.The interde-
pendence across partner carriers'networks caused by the alliance
may in turn influence each partner's OTP.It is not clear a priori
whether the alliance will improve or worsen a given partner's OTP.
On the one hand, a carrier may have a greater incentive to provide
better OTP when it joins a codeshare alliance because its OTP not
only affects the timeliness of connections within its own network,
but also affects the timeliness of connections between its network
with its partner carriers'networks.On the other hand,a carrier's
OTP could worsen after joining a codeshare alliance since an extra
source of a carrier's delay can be due to its partner carriers'delay.
While not attempting to study the incentives to form an alliance,
the primary objective of this paper is to evaluate the net impact of a
codeshare alliance on partner carriers' OTP.
An extensive literature exists on the pricing effects of strategic
alliances (Brueckner and Whalen,2000; Brueckner,2001,2003;
Bamberger etal., 2001; Ito and Lee, 2005; Gayle, 2008, 2013;
Gayle and Brown,2014).The above-mentioned studies argue on
one hand that airline cooperation due to an alliance exerts a
downward pressure on airfares in interline markets due to product
complementarity and the mitigation ofdouble marginalization.
Brueckner and Whalen (2000)and Brueckner (2003)note that
code-sharing allows airlines to eliminate a double markup on
itineraries with multiple operators, resulting in lower fares. Ito and
Lee (2005) also show code-sharing to be associated with lower
fares. On the other hand, an alliance can also reduce competition in
markets where the partners' route networks overlap (typically their
inter-hub markets), and this in turn, puts pressure on airfares to rise
in these markets.Zou et al.(2011) argue that it is possible that an
alliance causes fares to increase even in markets where the part-
ners'route segments are complementary rather than overlapping.
They make the argument that the improvement in the quality of
interline connections that comes with an alliance,would subse-
quently increase demand.
Other studies have examined the effects of frequent flyer alli-
ances.For instance, Bilotkach (2005) examines airline alliance for-
mation using transatlantic markets to determine ifcode-sharing
with and without antitrust immunity decreases fares for interline
trips equally.The findings suggest that code-sharing and alliance
formation both have fare-decreasing effects,however the code-
sharing effect turns out to be more than twice the magnitude of
the alliance effect.
Concerns overpoor on-time performance may therefore be
exacerbated or improved by airline alliances.Few authors have
explored and analyzed the relationship between airline alliances
and service quality,both theoretically and empirically.The empir-
ical literature has been largely inconclusive,with some studies
suggesting that airline alliances increase product quality,others
suggesting that airline alliances decrease product quality, and some
studies found no relationship between airline alliances and product
quality (Hassin and Shy, 2004; Gayle and Thomas, 2015; Gayle and
Yimga, 2014; Goh and Uncles, 2003; Tiernan et al., 2008; Tsantoulis
and Palmer,2008).At the center of these diverging empirical re-
sults, reside two main issues: (1) the difficulty in defining quality in
a way that is mathematically tractable and (2) the sensitivity of
results to assumptions of a particular theoretical model (Prince and
Simon,2009; Park, 1997).
With respect to the first issue, some measures of service quality
have been explored.Goh and Uncles (2003) empirically study the
perceptions that business travelers have of the benefits of global
alliances. To measure quality, they use a cross-sectional self-
completion survey that was administered to a sample of Australian
business travelers.Tsantoulis and Palmer (2008) examine service
quality effects of a co-brand alliance where service quality is
proxied by a quality index they constructed based on some tech-
nical and functionalaspects ofquality.Gayle and Yimga (2014)
empirically investigated the routing quality effects ofthe Delta/
Northwest/Continental codeshare alliance, while Gayle and
Thomas (2015) investigated the routing quality effect of global al-
liances, antitrust immunity,and domestic mergers.4
Another service quality measure is an airline's on-time perfor-
mance.Almost no research has been conducted to examine the
impact of a codeshare alliance on the on-time performance of its
partner members. An exception is the work by Tiernan et al. (2008).
They investigate the service quality ofE.U. and U.S.members of
main airline alliances.Three specific measures ofairline service
quality were used in their study: on-time flight arrival percentage,
percentage offlights not canceled and percentage ofpassengers
filing baggage reports (bags lostdamaged,delayed or pilfered).
Their examination of the international airline alliances indicates no
significant differences in the quality of service indicators.
Apart from Tiernan et al.(2008) who looked at the linkage be-
tween on-time performance and international airline alliance, most
studies on on-time performance have focused on its relationship
with competition,multimarket contact,prices and entry or threat
of entry (Mayer and Sinai, 2003; Mazzeo,2003; Rupp et al.,2006;
Prince and Simon,2009, 2014; Forbes,2008; Prince and Simon,
2014).
To examine whether and how codeshare partners' product
quality provision change in response to a codeshare agreement, we
focus on the Delta Air Lines (DL),Northwest Airlines (NW) and
Continental Airlines (CO) Codeshare Alliance.We choose this
codeshare alliance for the following reasons: (i) it involves three
major carriers in the U.S. domestic airline industry; (ii) the alliance
was the largest ever approved in the history of the U.S. commercial
aviation; and (iii) the alliance turned out to be the most contentious
alliance in the U.S.domestic airline industry.
The contribution of this paper is to assess how Delta Air Lines
(DL), Northwest Airlines (NW) and Continental Airlines'(CO) on-
time performance change in response to their codeshare agree-
ment of August 23,2003.We find that the codeshare agreement
(CSA) improved OTP for the alliance firms,and that this improve-
ment occurs in both markets where the codeshare partners had
competed prior to the CSA and markets where they did not.
However, markets in which the codeshare partners competed prior
to the CSA,witnessed larger OTP effects.
The rest of the paper is organized as follows.The next section
provides an overview of the Delta,Northwest,Continentalcode-
share alliance.Section 3 describes the data used for analysis.
4 Routing Quality is defined as the ratio of nonstop flight distance to the product's
itinerary flight distance used to get passengers from the origin to destination.
J.O.Yimga / Journal of Air Transport Management 58 (2017) 76e90 77

Section 4 discusses the research methodology and estimation
technique used to analyze the OTP effects of the alliance. Results are
presented and discussed in Section 5, while concluding remarks are
gathered in Section 6.
2. Delta/Northwest/Continental codeshare alliance
2.1. Background
In August 2002,Delta Air Lines,Northwest Airlines and Conti-
nental Airlines submitted the largest domestic codeshare agree-
ment proposalin the United States.This agreement grants some
privileges to the partner airlines like reciprocal frequent-flyer
programs and reciprocalaccess to airportlounges.The partner
airlines'managers claimed that the CSA willgenerate benefits to
consumers such as increased flightfrequencies,broader travel
options,improved frequent flyer programs and better route con-
nections.They also claimed thatcost savings from the alliance
members will be passed on to consumers in terms of lower airfares.
Despite,initial assurances by the partner airlines,policy makers
have expressed a greatdeal of skepticism when appraising the
Delta/Northwest/Continentalalliance proposal, which policy
makers believed did not adhere to certain antitrust laws and reg-
ulations because of its potential to yield anti-competitive effects:
“The Department has determined that the agreements,if
implemented as presented by the three airlines,could result in a
significant adverse impact on airline competition,unless the air-
lines formally accept and abide by certain conditions that are
intended to limit the likelihood of competitive harm. If the airlines
choose to implement the agreements without accepting those
conditions,the department will direct its Aviation Enforcement
office to institute a formal enforcement proceeding regarding the
matter”5
.
Likewise, the General Accounting Office (GAO)stressed the
adverse effects of alliances on competition:
“[Proposed alliances] willreduce competition on hundreds of
domestic routes ifthe alliance partners do not compete with
each other or compete less vigorously than they did when they
were unaffiliated … It will be critical to determine if an alliance
retains or reduces incentives for alliance partners to compete on
price”6
The three-way alliance between Delta,Northwest and Conti-
nental was the largest domestic alliance at the time, accounting for
almost 30% of domestic origin-destination passengers (Ito and Lee,
2005). Given this recent trend towards increased alliance formation
and the fact that at the time,carriers comprising the three largest
U.S. alliances (Continental/Northwest/Delta,United/US Airways
and American/Alaska) accounted for approximately two thirds of
all domestic origin and destination passenger traffic, we posit that
there could be legitimate policy apprehensions regarding the
impact of these cooperative agreements on on-time performance
delivery.
Using data on OTP and factors that are likely to influence OTP,
this paper uses a reduced-form regression analysis to investigate
whether codeshare partners'OTP is impacted by a CSA.While ar-
guments can be made to support both views, there is currently no
empirical evidence that supports either.
2.2. Mechanisms at play
Code-sharing allows airlines to sell seats on one another's flights
as if these seats were their own. A codeshare agreement provides a
way for airlines to expand their network of flights without adding
planes to serving costly small regional markets with low-volume
commuters.
Over the years,as airlines continue to struggle financially after
the Deregulation of1978,they are increasingly relying on code
sharing to sell more tickets and increase their profitability.
The interdependence across partner carriers'networks caused
by the alliance may in turn influence each partner's OTP.Some of
the mechanism(s) through which a CSA helps improves OTP are as
follows.7 Once partner airlines codeshare on a route: i) they remove
some flights with poor OTP, ii) they remove flights taking off at peak
hours (which may be more likely to be late); hub operations in
particular may struggle with OTP especially during peak times and
fixing this problem requires that airlines explore new ways to share
the airspace with other carriers, iii) they remove previously
competing flights taking off at almost the same time,iv) they use
better planes less likely to have mechanical issues.
3. Definitions and dataset construction
3.1.Definitions
A market is defined as a directional air travel between an origin
and destination city during a specific time period.By directional,
we mean that an air travel trip from Miami to Las Vegas is a distinct
market from an air travel trip from Las Vegas to Miami.This con-
trols for the number of passengers traveling between the origin and
destination.8 There are two types of carriers in the datadticketing
carrier and operating carrier. The ticketing carrier is the airline that
issues the flight reservation or ticket to consumers.The operating
carrier is the airline that operates the aircraft,in other words,the
airline that actually transports the passengers. An air travel product
is defined as a unique combination of ticketing carrier,operating
carrier(s) and itinerary. Gayle (2007) and Ito and Lee (2005) define
three types of air travelproducts: pure online; traditionalcode-
share; and virtual codeshare.For a pure online product,the same
airline is the ticketing and operating carrier on all segments of the
trip. For example,a two-segment ticket with both segments mar-
keted by Delta Air Lines and both segments of the itinerary are also
operated by Delta Air Lines. An air travel product is said to be code-
shared if the operating and ticketing carriers for thatitinerary
differ.
A traditional codeshare product has a single ticketing carrier, but
multiple operating carriers, one of which is the ticketing carrier. For
example,a connecting itinerary operated by Delta Air Lines (DL)
and Northwest Airlines (NW) but marketed solely by Delta Air Lines
(DL) is a traditional codeshare product.A virtual codeshare air
travel product has the same operating carrier for all segments of the
itinerary,but the ticketing carrier is different from the operating
carrier.For example,a connecting itinerary operated entirely by
Northwest Airlines (NW) but marketed solely by Delta Air Lines
(DL) is a virtual codeshare product.
For proper identification of the different types of productdpure
online and virtualcodeshared we do not recode regionalfeeder
5 Departmentof Transportation.Office of the Secretary Termination Review
Under 49 U.S.C.41,720 ofthe Delta/Northwest/ContinentalAgreements.Federal
Register.Vol 68,No.15.Thursday,January 23,2003.Notices.
6 US General Accounting Office. Aviation Competition: Proposed Domestic Airline
Alliances Raise Serious Issues. 1998.
7 We thank an anonymous referee for pointing out some of these mechanisms.
8 See Berry et al.(2006), 2006) and Gayle (2007). However,unlike these studies,
some flights could be segments of itineraries with intermediate stop(s).
J.O.Yimga / Journal of Air Transport Management 58 (2017) 76e9078

carriers to have their major carriers' code. This makes sure that we
do not attribute regional feeders' OTP to the main ticketing carriers.
For instance,if a product that involves Delta Air Lines (DL) and
Comair Delta Connection (OH), where one of them is the ticketing
carrier and the other the operating carrier,instead of recoding
Comair Delta Connection as Delta Air Lines (DL), we simply drop the
product. If not dropped, this product would mistakenly be
considered a codeshare product because the ticketing and oper-
ating carriers are different.
Given that our analysis focuses on OTP,we only consider pure
online and virtual codeshare flights.
3.2. Dataset construction
We use data gathered and published by U.S.Departmentof
Transportation (DOT) Bureau of Transportation Statistics (BTS). The
BTS requires all U.S. domestic carriers with revenues from domestic
passenger flights of at least one percent of total industry revenues
to report flight on-time performance data.
The data cover scheduled-service flights between points within
the United States.The data frequency is monthly.A record in this
survey represents a flight.Each record or flight contains informa-
tion on the operating carrier,the origin and destination airports,
miles flown, flight times, and departure/arrival delay information.9
Moreover,because on-time performance is only measured for in-
dividual flights,we restricted our analysis to nonstop service.We
collect monthly data for every non-stop domestic flight for the
third and fourth quarters of2002 and 2004 for 19 U.S.carriers.
Table 1 reports a list of carriers in the data sample. All variables are
constructed from the originaldata set of 6,274,848 flights in the
sample.We omitted all canceled and diverted flights.
Given that the Delta/Northwest/Continental codeshare alliance
was formed in August of 2003, the third and fourth quarters of 2002
representthe pre-alliance period whereas the third and fourth
quarters of 2004 represent the post-alliance period (using data
from the same quarters for both years willcontrol for potential
seasonal effects in OTP).We choose this particular time period to
balance the “before” and “after” periods around the codeshare
event and avoid data right after the September11th terrorist
attacks.
To enable a more manageable-sized data set,we place some
restrictions on the raw data.We follow the same procedures used
by Aguirregabiria and Ho (2012) for the selection of markets.We
focus on air travel amongst the 63 largest U.S.cities.Table 2 pre-
sents a list of the cities and corresponding population sizes.
Incomplete data reporting in addition to missing/incorrect on-time
performance data slightly reduces the sample.
We use the geometric mean of the populations at the origin and
destination to help measure the impact of potentialmarket size.
Unlike Aguirregabiria and Ho (2012),we do not group cities that
Table 1
Airlines in sample.
Code Airline
AA American Airlines
AS Alaska Airlines
B6 JetBlue Airways
CO Continental Air Lines
DH Independence Air
DL Delta Air Lines
EV Atlantic Southwest
FL AirTran Airways
HA Hawaiian Airlines
HP America West Airlines
MQ American Eagle
NW Northwest Airlines
CO Comair
OO SkyWest
RU ExpressJet
TZ ATA Airlines
UA United Air Lines
US US Airways
WN Southwest Airlines
Table 2
Cities,airports and population.
City,state Airports City population
2002 2004
New Yorka LGA,JFK, EWR 8,606,988 8,682,908
Los,Angeles,CA LAX,BUR 3,786,010 3,796,018
Chicago,IL ORD,MDW 2,886,634 2,848,996
Dallas,TXb DAL,DFW 2,362,046 2,439,703
Houston,TX HOU, IAH, EFD 2,002,144 2,058,645
Phoenix,AZc PHX 1,951,642 2,032,803
Philadelphia,PA PHL 1,486,712 1,514,658
San Antonio,TX SAT 1,192,591 1,239,011
San Diego,CA SAN 1,251,808 1,274,878
San Jose,CA SJC 896,076 901,283
Denver-Aurora,CO DEN 841,722 848,227
Detroit,MI DTW 922,727 924,016
San Francisco,CA SFO 761,983 773,284
Jacksonville,FL JAX 758,513 778,078
Indianapolis,IN IND 783,028 787,198
Austin,TX AUS 671,486 696,384
Columbus,OH CMH 723,246 735,971
Charlotte,NC CLT 577,191 614,446
Memphis,TN MEM 674,478 681,573
Minneapolis-St.Paul,MN MSP 660,771 653,872
Boston,MA BOS 585,366 607,367
Baltimore,MD BWI 636,141 641,004
Raleigh-Durham,NC RDU 503,524 534,599
El Paso,TX ELP 574,337 582,952
Seattle,WA SEA 570,166 570,961
Nashville,TN BNA 544,375 570,068
Milwaukee,WI MKE 589,975 601,081
Washington,DC DCA,IAD 564,643 579,796
Las Vegas,NV LAS 506,695 534,168
Louisville,KY SDF 553,049 558,389
Portland,OR PDX 537,752 533,120
Oklahoma City,OK OKC 518,516 526,939
Tucson,AZ TUS 501,332 517,246
Atlanta,GA ATL 419,476 468,839
Albuquerque,NM ABQ 464,178 486,319
Kansas City,MO MCI 443,390 458,618
Sacramento,CA SMF 433,801 446,295
Long Beach,CA LGB 470,398 470,620
Omaha,NE OMA 399,081 426,549
Miami, FL MIA 371,953 378,946
Cleveland,OH CLE 468,126 455,798
Oakland,CA OAK 401,348 394,433
Colorado Springs, CO COS 369,945 388,097
Tula,OK TUL 390,991 382,709
Wichita, KS ICT 354,306 353,292
St.Louis, MO STL 347,252 350,705
New Orleans,LA MSY 472,540 461,915
Tampa,FL TPA 315,151 320,713
Santa Ana,CA SNA 341,411 339,319
Cincinnati,OH CVG 322,278 331,717
Pittsburg,PA PIT 327,652 320,394
Lexington,KY LEX 262,706 274,581
Buffalo,NY BUF 287,469 281,757
Norfolk,VA ORF 238,343 241,979
Ontario,CA ONT 164,734 168,068
a New York-Newark-Jersey.
b Dallas-Arlington-Fort Worth-Plano, TX.
c Phoenix-Temple-Mesa,AZ.9 Some flights could be segments of itineraries with intermediate stop(s).
J.O.Yimga / Journal of Air Transport Management 58 (2017) 76e90 79

belong to the same metropolitan areas and share the same airport
for two reasons: (1) airport grouping will lessen the heterogeneity
in OTP data and (2) observations in dataset may not be products but
are individualflights,most of which are segments on itineraries
with intermediate stop(s).
3.3. On-time performance (OTP) measures
We directly use measures of on-time performance from the U.S.
DOT BTS0 dataset.According to the U.S.DOT,flights that do not
arrive at (depart from) the gate within 15 minutes of scheduled
arrival (departure) time are late arrivals (departures).This repre-
sents performance measured against airlines' published schedules.
The three main measures are arrival minutes late, the percentage of
flights arriving at least 15 minutes late and the percentage of flights
arriving at least 30 minutes late.
Following Prince and Simon (2009), our analysis is conducted at
the carrier-route-month-year level,so we use the average arrival
delay over all of a carrier's flights,on a particular route,during a
month and year.10 Given that in a specific month,an airline can
operate a specific origin-destination multiple times, with different
OTP values,we construct our OTP measures by averaging the OTP
values for a given origin-destination for a given carrier in a given
month and year. We then collapse the data by carrier-origin-
destination-month-year combinations.Explanatory variables are
averaged and collapsed using the same approach. Our final working
data set has 31,748 usable observations, where an observation is at
the level of carrier-origin-destination-month-year combination.
We construct the first OTP measure based on the arrival delay of
a flight, in other words, the difference between scheduled and
actual arrival time.
The second OTP measure is constructed from the arrival delay
indicator in the dataset for flights arriving at the gate at least
15 minutes late. We use this arrival delay indicator to compute the
proportion of a carrier's flights on a route in a month that arrived at
least 15 minutes late.
The third OTP measure uses the 30 minutes and more arrival
delay indicator.Similar to the second measure,we use this arrival
delay indicator to compute the proportion of a carrier's flights on a
route in a month that arrived at least 30 minutes late.We use
analogous measures for departure OTP.The same 15- and 30-mi-
nutes rules apply to departure delay.Table 3 summarizes OTP
measures.Overall,arrival delays are longer than departure delays
for all measures,supporting the findings from the Bureau of
Transportation Statistics (2011) that indicate that on-time arrival
performance has the greatest impact on passengers.Also, arrival
measures tend to vary more than departure measures.
Figs. 1 and 2 display the frequency of observations in 15-minutes
intervals around their scheduled arrival(departure)time. It is
surprising to note that a sizeable portion of the flights in our sample
were “early”d55.4 percent of flights arrived at their gate prior to
the scheduled arrivaltime while 47.6 percent of flights departed
from their gate prior to the scheduled departure time.This is
indicative of a certain amount of slack that may be built into the
airlines' schedules. Prince and Simon (2009) suggests that this may
be done strategically by airlines. On the other hand, 17.6 percent of
flights in the dataset arrived 15 minutes or more late while 14.7
percent departed from their gate 15 minutes or more late.
Tables 4 and 5 summarize OTP by month.For all measures,the
percentage of flights arriving late peaks in winter. We only consider
flights arriving at their destination and do not include cancellations
or diversions even though cancellations tend to rise during the
winter months in the face of severe weather (Bureau of
Transportation Statistics,2011).
Throughout our study, early arrivals are treated as a delay of zero
minutes rather than as a negative delay. Counting early arrivals and
departures as zero delays assumes that passengers derive disutility
from late arrivals/departures butno utility from early arrivals/
departures.
Tables 6 and 7 summarize OTP by carrier.Hawaiian Airlines
performs better than allcarriers on arrivaldelay minutes,while
Independence Air has the worst arrival delay minutes.Northwest
Airlines has the shortest departure delay minutes while SkyWest
has the longest departure delay minutes.
4. Empirical method and estimation
To examine whether partner firms' on-time performance is
impacted by their participation in a codeshare alliance, we estimate
reduced-form regression equations of the various measures of OTP
described above.Possible codeshare alliance effects on OTP are
identified using a difference-in-differences strategy.This strategy
enables us to compare pre-post alliance periods' changes in OTP of
flights operated by the alliance firms, relative to changes in OTP of
flights operated by non-alliance firms over the same pre-post alli-
ance periods.We specify our empiricalmodel of product quality
effects due to code-sharing. On-time performance (OTP) is used to
proxy product quality.
We specify a linear regression model in which an OTP measure is
a function of: (1) timing of implementation of the codeshare alli-
ance; (2) carrier and airport characteristics; and (3) market structure
characteristics. Furthermore, we also examine whether the effect of
code-sharing on OTP depends on the existence ofpre-alliance
competition between alliance firms. Variables are defined in Table 8.
The baseline reduced-form specification ofthe arrival OTP of
airline a in market m in time period t is as follows:
OTPamt ¼ a þ bX amt þ gZamt þ dW amt þ l a þ ht þ originm
þ destm þ εamt (1)
where Xamt representflight characteristics,Zamt include market
Table 3
On-time performance summary statistics.
Obs Mean Std.Dev. Min Max
Arrival
Arrival Delay (in minutes) 31,748 10.71 8.94 0 440
Fraction of flights arriving at least 15 minutes late (percent) 31,748 12.13 9.47 0 100
Fraction of flights arriving at least 30 minutes late (percent) 31,748 6.59 6.10 0 100
Departure
Departure Delay (in minutes) 31,748 9.20 8.90 0 415
Fraction of flights departing at least 15 minutes late (percent) 31,748 10.22 8.94 0 100
Fraction of flights departing at least 30 minutes late (percent) 31,748 5.85 5.75 0 100
Note: Early arrivals are counted as zero delays.
10 Early arrivals are counted as zero delays.
J.O.Yimga / Journal of Air Transport Management 58 (2017) 76e9080

characteristics,W amt is a vector of dummy variables representing
the codeshare effects.l a’s are airline specific fixed effectsht’s are
time specific fixed effects,origin and destination airport specific
fixed effects are denoted by originm and destm, εamt is the unob-
served part of OTP.The reduced-form OTP regression is estimated
using Ordinary Least Squares (OLS). We provide further description
of the explanatory variables in the following section.
5. Empirical results
5.1.Estimates from reduced-form arrival OTP equation
In this section,we present empirical analyses of the impact of
code-sharing on OTP.We start with on-time arrivalperformance
since it has greatest impact on passengers (Bureau of
Fig. 1. Histogram of arrival “minutes late” in the dataset.
Fig. 2. Histogram of departure “minutes late” in the dataset.
Table 4
Mean values of arrival delay measures by month.
year Month Mean arrival Delay (minutes) Percentage of flights delayed
more than 15 minutes
Percentage of flights delayed
more than 30 minutes
2002 July 10.9 13.9 7.8
August 9 11.9 6.3
September 6.1 8 4.2
October 7.5 11 5.2
November 7.4 10.6 4.9
December 11.8 14.8 8.1
2004 July 15.2 14.1 8.4
August 13.3 13.9 7.6
September 7.9 7.9 4.2
October 9.4 11 5.4
November 11.6 12.2 6.5
December 15.8 16.8 9.9
Note: Early arrivals are counted as zero delays.
J.O.Yimga / Journal of Air Transport Management 58 (2017) 76e90 81

Table 5
Mean values of departure delay measures by month.
Year Month Mean departure Delay (minutes) Percentage of flights delayed
more than 15 minutes
Percentage of flights delayed
more than 30 minutes
2002 July 9.6 12.3 6.9
August 7.9 10.3 5.7
September 5.1 6.5 3.7
October 5.9 8.2 4.3
November 5.9 7.7 4.1
December 10.2 12.8 7.2
2004 July 13.2 12.3 7.5
August 11.3 11.1 6.6
September 6.9 7 4
October 7.7 8.7 4.7
November 9.9 10.1 5.8
December 14.2 14.8 9
Note: Early arrivals are counted as zero delays.
Table 6
Airlines' mean arrival delay.
Code Airline Arrival Delay (minutes) Proportion of flights arriving at
least 15 minutes Late (%)
Proportion of flights arriving at
least 30 minutes Late (%)
HA Hawaiian Airlines 0.0 0.00 0.00
WN Southwest Airlines 8.9 13.92 7.41
UA United Air Lines 9.4 8.83 5.00
NW Northwest Airlines 9.5 14.88 7.15
US US Airways 9.8 13.62 7.45
B6 JetBlue Airways 10.0 12.17 6.61
AA American Airlines 10.7 10.46 6.27
DL Delta Air Lines 10.7 13.19 6.60
CO Continental Air Lines 11.1 12.85 6.76
TZ ATA Airlines 11.1 12.84 7.08
AS Alaska Airlines 11.2 12.33 6.65
HP America West Airlines 11.2 9.64 4.63
RU ExpressJet 12.9 13.28 7.70
OH Comair 13.3 12.86 7.68
EV Atlantic Southwest 13.4 5.05 3.15
FL AirTran Airways 14.0 5.73 3.99
MQ American Eagle 14.4 13.38 8.09
OO SkyWest 16.2 7.73 5.23
DH Independence Air 17.5 13.01 8.59
Note: Early arrivals are counted as zero delays.
Table 7
Airlines' mean departure delay.
Code Airline Departure Delay (minutes) Proportion of flights departing at
least 15 minutes Late (%)
Proportion of flights departing at
least 15 minutes Late (%)
NW Northwest Airlines 7.0 9.75 5.55
CO Continental Air Lines 7.6 7.69 4.58
TZ ATA Airlines 7.8 9.01 5.09
UA United Air Lines 7.8 7.03 4.35
DL Delta Air Lines 8.3 9.25 5.13
US US Airways 8.5 11.37 6.42
AA American Airlines 8.8 8.39 5.4
HA Hawaiian Airlines 8.8 12.53 12.5
RU ExpressJet 8.9 8.47 5.63
B6 JetBlue Airways 9.3 11.36 5.33
HP America West Airlines 9.5 6.97 3.89
WN Southwest Airlines 9.7 16.57 8.33
AS Alaska Airlines 10.8 11.81 6.69
MQ American Eagle 11.9 11.23 7.06
FL AirTran Airways 12.2 5.61 3.62
OH Comair 12.3 12.18 7.5
EV Atlantic Southwest 13.2 5.4 3.38
DH Independence Air 16.6 12.63 8.3
OO SkyWest 17.0 7.66 5.13
Note: Early arrivals are counted as zero delays.
J.O.Yimga / Journal of Air Transport Management 58 (2017) 76e9082

Transportation Statistics,2011).
5.1.1.Determinants of arrival OTP
Airport congestion has an influentialrole in determining on-
time performance.11 One way to controlfor airport congestion is
controlling for hubbing. Effective hubbing implies that flights from
different origin airports known as “spokes” of a network arrive at
the “hub” airport roughly at the same time. The aircraft at the hub
waits for these spoke flights and facilitates the transfer of passen-
gers and baggage. Subsequently, flights depart from the hub airport
in quick sequence back out along the spokes.
Essentially,passengers departing from any non-hub origin to
other destinations in the network generally proceed first to the hub.
Table 9 shows that 14 out of the 19 carriers possess at least one hub
and 11 have at least 3 hubs. These airlines co-ordinate arrivals and
departures at their hubs in order to minimize delays for passengers
continuing through the hub to final destinations on spokes other
than the one on which they originated. We include a control for hub
airlines (INTOHUB).This measure is carrier-specific and captures
the effect of effective hubbing on arrival OTP. INTOHUB is a dummy
variable that equals unity if destination airport is a hub for that
carrier. As expected,regression results in Table 10 revealthat
INTOHUB is a predictor of arrival OTP.The coefficient estimate on
INTOHUB is negative and statistically significant suggesting shorter
arrival delays for carriers flying into their hubs. Carriers flying into
their hubs have a greater incentive to make sure that passengers ge
to their intermediate stop on time for their connecting flights since
the cost of a missing flight may be quite substantial from rebooking
passengers onto new connections to handling missed connection
luggage.The disutility experienced by the passenger in terms of
inconvenience and frustration may result in loss of future business.
Mazzeo (2003) finds interestingly that flights out ofthe hub
have a longer than scheduled flight time on average,whereas
flights into the hub do not. He partly attributes these differences to
the logistical difficulties associated with turning around large
banks of flights at busy hub airports. We were able to obtain similar
results from an estimation not shown in this paper.
Since carriers often have hubs ofdifferent sizes,a particular
airport might be a major hub for the airline while another airport
might be a medium-size hub. However, the INTOHUB variable does
not capture the heterogeneity in hub sizes for a given carrier since it
is a dummy variable. Thus, to capture this heterogeneity in carrier's
hub sizes, we also include a continuous variable DPRESCOST which
counts the number of different cities that an airline serves using
nonstop flight from the destination city of the market.Including
DPRESCOST controls for (dis)economies of scope and hubbing ef-
fects associated with offering multiple routes from the same
destination airport. The coefficient estimate on DPRESCOST is
negative and statistically significant as expected suggesting that a
carrier's arrival delay decreases with the number of cities it con-
nects to from a given destination airport.
Columns 3 through 6 of Table 10 re-estimate the model using
Table 8
Variable definitions and summary statistics.
Variable Definition Mean Std.dev.
Codeshare Event
Tdnc
t Time period dummy variable,equals unity for post-alliance period. 0.566 0.496
Flight, airport and market characteristics
DPRESCOSTNumber of different cities that an airline flies to from the destination city of the market using nonstop flight 30.171 30.840
OPRESCOSTNumber of different cities that an airline offers flights from going into the origin city of the market 30.095 30.792
INTOHUB Dummy Variable ¼ 1 if destination is a hub for that carrier (list of hub/airline combination in Table 9) 0.387 0.487
OUTOFHUB Dummy Variable ¼ 1 if origin is a hub for that carrier (list of hub/airline combination in Table 9) 0.387 0.487
DISTANCE Nonstop flying distance (in miles) between the origin and destination. 937.118 635.545
RELSPEED Carrier mean speed across its flights in a market as a ratio of market average speeda 1.000 0.016
MKTdnc Market-specific dummy variable,equals unity for O&Db markets in which any two of three allied carriers competed prior to alliance.0.022 0.146
MKTSIZE (logged) Geometric mean of the populations at both endpoint airports 13.483 0.536
DNCamt Zero-one dummy variable that takes the value one when the carrier is one of the three alliance carriers,DL,NW, or CO 0.264 0.441
Market Structure
MONOMKT Dummy Variable ¼ 1 if only 1 airline serves the city-pair market Non-stop 0.430 0.495
NUMCOMP Number of competitors in a market 1.817 0.877
a Speed is measured as distance divided by flight air time.
b O&D ¼ origin and destination.
Table 9
Airline carriers and their hubs.
Code Carrier Hub airports
AA American Airlines Dallas,O'Hare,Miami, New York,Los Angeles
AS Alaska Airlines Seattle,Portland,Los Angeles, San Francisco
B6 JetBlue Airways New York
CO Continental Air Lines Houston, Cleveland,Newark
DL Delta Air Lines Atlanta,Cincinnati,New York,Boston,Los Angeles,Minneapolis, Detroit,Seattle
EV Atlantic Southwest Dallas,O'Hare,Atlanta,Detroit,Cleveland,Houston,Denver,Kansas City,Newark, Dulles
HP America West Airlines Los Angeles, Phoenix
MQ American Eagle Dallas,O'Hare,Miami, New York
NW Northwest Airlines Minneapolis. Detroit,Memphis
OO SkyWest O'Hare,Seattle,Portland,Los Angeles, San Francisco,Detroit,Minneapolis,Denver,Houston, San Francisco, Phoenix
TZ ATA Airlines O'Hare,Indianapolis
UA United Air Lines Houston, O'Hare,San Francisco, Houston,Denver,Los Angeles, Newark
US US Airways Cleveland,Philadelphia,Phoenix,Washington
WN Southwest Airlines Atlanta,Washington, Chicago,Dallas,Los Angeles,Las Vegas,Houston, Phoenix,Oakland
11 Flores-Fillol(2010) and Rupp and Sayanak (2008),among others,investigate
this relationship.
J.O.Yimga / Journal of Air Transport Management 58 (2017) 76e90 83

different measures ofOTPdthe percentage offlights arriving at
least 15 minutes late and the percentage of flights arriving at least
30 minutes late. These other measures look at delay over a certain
threshold.
The coefficient estimate on DPRESCOST is positive and statisti-
cally significant when the dependent variable is the percentage of
flights arriving at least 15 minutes late. The coefficient estimate on
DPRESCOST in columns 3 and 4 of Table 10 indicates that for flights
over a certain delay threshold (15 minutes and more),an airline's
OTP worsens with increases in the number of distinct cities that an
airline has nonstop flights to, going out of the destination airport. In
other words,for flights into destination airport that are at least
15 minutes late, carriers' arrival OTP worsens, the larger the scale of
operations at the destination airport. This result is potentially
driven by logistical difficulties associated with turning around large
banks of flights. The same reasoning applies when we use the
percentage of flights arriving at least 30 minutes late as the
dependent variable,however the estimates on DPRESCOST in col-
umns 3 and 4 of Table 10 are not statistically significant.
In addition, we also control for market size (MKTSIZE), measured
as the (logged) geometric mean of the populations at both market
endpoints.The net impact of market size on OTP may either be
negative or positive.On one hand, larger marketsizes may be
associated with higher demand for air travel and thus more airport
congestion resulting in more delays.On the other hand,in larger
markets airlines have more incentive to be on time because more
people will be affected if they are not,resulting in future loss of
business.Thus, in the latter case,arrival OTP may improve with
increasing market size. Therefore, we argue that MKTSIZE captures
the net effect of these conflicting forces. The coefficient estimate on
MKTSIZE is positive and statistically significant,suggesting that
larger markets are associated with worse arrival on-time
performance.Thus, the airport congestion effect dominates.
On-time performance is also influenced by flight distance and
the relative speed of the flight.The variable DISTANCE represents
the flight's distance in miles. The parameter estimate on DISTANCE
is negative and statistically significant, suggesting that carriers have
some ability to “make time up in the air”on longer flights.This
ability to “make time up in the air” improves arrival OTP.
We also include a measure for the carrier's relative speed
(RELSPEED) defined as the average speed of a carrier's flights in the
market divided by the average speed of allflights in the market.
RELSPEED captures how fast a carrier is,relative to the typical car-
rier's velocity in a market.The parameter estimate on RELSPEED is
negative and statistically significant, suggesting that airline carriers
with above-average flying speed tend to have better arrival OTP.
Building on extant research exploring the relationship between
service quality and competitive conditions,we investigate how
route competition affects carriers'arrival OTP.12 We control for
route-level competition by including a measure of market structure
(MONOMKT), which is a monopoly dummy variable that equals one
if there is only one carrier serving a given market.The coefficient
estimate on MONOMKT is positive and statistically significant. This
result is consistent with our expectations,suggesting that arrival
delays are greater on less competitive routes.This result is also
consistent with findings by Mazzeo (2003) and Rupp et al.(2006)
who posit that airlines provide worse on-time performance on
less competitive routes.To go a step further,consider how the
degree of market competitiveness,as measured by the number of
competitors (NUMCOMP) in a given market, affects the arrival OTP.
NUMCOMP represents a more heterogeneous measure ofmarket
Table 10
Arrival on-time performance estimation results.
Variables Arrival Delay in
minutes
Arrival Delay in
minutes
% of flights arriving at least
15 minutes late
% of flights arriving at least
15 minutes late
% of flights arriving at
least 30 minutes late
% of flights arriving at
least 30 minutes late
(1) (2) (3) (4) (5) (6)
INTOHUB 1.3499*** 1.3478*** 1.1886*** 1.1787*** 0.7450*** 0.7502***
(0.2065) (0.2066) (0.1585) (0.1586) (0.1079) (0.1079)
DPRESCOST 0.0208*** 0.0208*** 0.0083*** 0.0079*** 0.0018 0.002
(0.0037) (0.0037) (0.0029) (0.0029) (0.0019) (0.0019)
RELSPEED 37.0368*** 37.3472*** 4.3097 4.0099 3.8671** 3.7672**
(3.4189) (3.4205) (2.6244) (2.6253) (1.7857) (1.7866)
DISTANCE 0.0022*** 0.0022*** 0.0001 0.0001 0.0006*** 0.0006***
(0.0001) (0.0001) (0.0001) (0.0001) (0.0001) (0.0001)
MONOMKT 0.6765*** 0.6999*** 6.1969*** 6.2091*** 3.2302*** 3.2455***
(0.1934) (0.1938) (0.1484) (0.1488) (0.1010) (0.1012)
NUMCOMP 0.0609 0.0373 3.0513*** 3.0253*** 1.7877*** 1.7826***
(0.1188) (0.1191) (0.0912) (0.0914) (0.0620) (0.0622)
MKTSIZE 20.6967*** 21.0915*** 17.8731*** 18.0976*** 12.2403*** 12.4839***
(5.4601) (5.4639) (4.1912) (4.1936) (2.8517) (2.8538)
DNCamt 1.4607*** 1.3976*** 0.6141*** 0.5186** 0.4826*** 0.4771***
(0.2733) (0.2755) (0.2098) (0.2115) (0.1427) (0.1439)
Tdnc
t 4.2784*** 4.2674*** 2.9316*** 2.9217*** 2.0867*** 2.0827***
(0.1695) (0.1696) (0.1301) (0.1301) (0.0885) (0.0886)
Tdnc
t DNC amt 1.4460*** 1.3132*** 0.4744** 0.3517* 0.6666*** 0.6198***
(0.2507) (0.2555) (0.1925) (0.1961) (0.1310) (0.1335)
MKTdnc 0.5691 0.0625 0.5453**
(0.5150) (0.3953) (0.2690)
Tdnc
t DNC amt MKT dnc 2.1015*** 2.0021*** 0.6964*
(0.7669) (0.5886) (0.4005)
CONSTANT 226.7955*** 231.6940*** 224.3223*** 227.0041*** 155.5177*** 158.6174***
(71.4461) (71.4932) (54.8431) (54.8718) (37.3152) (37.3417)
No. of Obs. 31,748 31,748 31,748 31,748 31,748 31,748
R2 0.23 0.23 0.41 0.41 0.35 0.35
The equations are estimated using ordinary least squares.Fixed effects are included in each specification but were not reported for brevity.
Note: Standard errors are in parentheses.***p < 0.01; **p < 0.05; *p < 0.10.
12 Studies by Mazzeo (2003) and Rupp et al.(2006) examine this relationship.
J.O.Yimga / Journal of Air Transport Management 58 (2017) 76e9084

structure compared to the MONOMKT dummy variable.As ex-
pected,arrival OTP improves with increasing number of competi-
tors. Though the coefficient estimate on NUMCOMP has the
expected sign, it is not statistically significant in columns 1 and 2 of
Table 10.
5.1.2.Codeshare effects on arrival on-time performance
The remaining rows ofTable 10 contain the key variables of
interest in evaluating the codeshare effects of a codeshare agree-
ment on arrivalOTP.We also focus on changes in arrivalOTP in
certain types of marketsdmarkets where any two of the three
alliance firms had competed prior to the alliance.
To examine persistent differences in OTP of flights operated by
the alliance partners,we include a dummy variable DNCamt which
equals unity for flights operated by any of the alliance carriers. The
coefficient estimate on DNCamt is positive and statistically signifi-
cant in columns 1e4 in Table 10,indicating that throughout the
sample period the mean arrival delay of flights operated by Delta,
Northwest and Continental is greater than the mean arrival delay of
flights operated by other carriers in the sample.
We also define a dummy variable Tdnc
t to help identify the OTP
effects of the codeshare alliance.Tdnc
t is a time period dummy
variable,which equals unity in the post-alliance period.The coef-
ficient estimate on Tdnc
t measures,on average,how arrival delay
changes over the pre-post codeshare alliance period for flights that
are not associated with Delta,Northwest or ContinentalAirlines.
The positive coefficient estimate on Tdnc
t indicates that the mean
arrival delay of flights operated by airlines other than Delta,
Northwest and Continental airlines increased (OTP worsens) from
pre-to post-alliance periods.
Finally, we include the interaction between the DNCamt and Tdnc
t
variables. The coefficient estimate on this new variable
Tdnc
t DNC amt represents the difference-in-differences estimate
that identifies whether arrival delay of flights operated by any of
the alliance carriers changed differently relative to arrival delay of
flights operated by other airlines over the pre- and post-alliance
periods.It captures changes in arrivaldelay in DL/NW/CO flights
(relative to non-DL/NW/CO flights) due to the alliance.The coeffi-
cient estimate is negative and statistically significant,suggesting
that the alliance caused the mean arrivaldelay for DL/NW/CO
flights to fall compared to the mean arrival delay for non-DL/NW/
CO flights over the pree and postealliance periods.In a nutshell,
the codeshare alliance is associated with improved arrival OTP for
the alliance firms relative to other carriers.
This result is supported by Table A1 in the Appendix.Table A1
reports mean arrival(departure) delay minutes before and after
the CSA for alliance partners versus other carriers.Table A1 in-
dicates that even though OTP worsens overallover the pre-post
alliance periods for allcarriers on average,the increase in delay
minutes is smaller for the alliance partners' flights. We test for the
difference in mean arrival (departure)delay minutes between
alliance partners and other carriers. All tests of difference in means
are statistically significant at 1% level.
5.1.3.Codeshare effects on arrival OTP based on existence of pre-
alliance competition between alliance firms
To examine whether changes in partner carriers' arrival OTP are
explained by the existence ofpre-alliance competition between
alliance firms,we construct and include a market-specific dummy
variable, MKTdnc that equals to one for origin-destination markets in
which any two of the three alliance partners competed prior to the
alliance.Thus,we are able to examine whether the codeshare ef-
fects on OTP differ in markets where the alliance partners
competed prior to the alliance.Columns 2,4 and 6 in Table 10
reproduce the baseline arrival OTP regressions with the inclusion
of the market dummy variable and some interactions with this
dummy variable.
The effects of the DL/NW/CO codeshare alliance on OTP in
markets where the alliance firms competed before the alliance is
determined by summing the coefficients on the interaction vari-
ables Tdnc
t DNC amt and Tdnc
t DNC amt MKT dnc in Specification 2
and doing the same for Specifications 4 and 6 (columns 2, 4 and 6 in
Table 10).Summing the coefficients yields a negative estimate,
indicating an improvement in arrival OTP of flights operated by the
alliance firms in the markets where they competed with each other
prior to their alliance.
The coefficient estimate on the interaction variable
Tdnc
t DNC amt in columns 2,4 and 6 of Table 10,has a different
interpretation. In fact, the coefficient estimate captures changes in
arrival delay in DL/NW/CO flights due to the codeshare alliance in
markets where they did not compete prior to the alliance.The
coefficient estimate on Tdnc
t DNC amt in columns 2, 4 and 6 of
Table 10,is negative and statistically significant at conventional
levels of significance,suggesting that the codeshare alliance also
improved arrival on-time performance in markets where the alli-
ance firms did not compete prior to alliance. Thus, evidence shows
that the alliance caused the alliance firms to improve arrival OTP
regardless of whether they competed or not in markets prior to the
alliance,but the partners'arrival OTP improvements are relatively
larger in markets that the partners competed in prior to the
alliance.
5.2. Estimates from reduced-form departure OTP equation
To further isolate the source of delays, we investigate the effect
of code-sharing on departure delay.Similar to arrival OTP, we
consider three different measures of departure OTP. In Table 11, we
report results for the three measures of departure delay in the data.
We also controlfor flight and market structure characteristics as
well as airline, month,year and airport-specific fixed effects.
5.2.1.Determinants of departure OTP
We now analyze factors that influence airlines'departure OTP,
with the ultimate goal of understanding how this OTP measure is
influenced by a codeshare alliance.To control for airport conges-
tion, we include the OUTOFHUB dummy variable that equals one if
the origin airport is a hub for that carrier.13 Similarly to INTOHUB in
Table 10, OUTOFHUB captures the hubbing effect on departure OTP.
For all three measures ofdeparture OTP,the hubbing effect is
positive and statistically significant, indicating that airlines produce
poor departure OTP on flights originating from their hubs.Flights
originating from an airline's hub are often spoke flights that are
heading to passengers'final destination (spoke airport).At spoke
airports, there are no interdependencies between airlines' aircrafts
since few arrive or depart and passengers do not connect,hub
carriers may have less incentive to improve OTP (Mayer and Sinai,
2003).
Given its binary nature,OUTOFHUB fails to capture heteroge-
neity in airline's hub sizes. To solve this problem, we include a more
reasonable measure of hubbing effects (OPRESCOST) in the depar-
ture delay regressions.OPRESCOST counts the number of different
cities that an airline offers flights from, going into the origin city of
the market using a nonstop flight. The coefficientestimate on
OPRESCOST is positive and statistically significant for all measures
of departure delay as expected.In particular,a carrier's departure
delay increases with the number of different cities it offers direct
flights from, that go into the origin city of a given market.
13 See list of hub/airline combination in Table 9.
J.O.Yimga / Journal of Air Transport Management 58 (2017) 76e90 85

The negative coefficient estimate on DISTANCE in the departure
OTP regression in Table 11 suggests that longer flights tend to have
shorter departure delays. On longer flights, carriers have an
incentive to depart on time to minimize the likelihood of late arrival
(or late departure for a subsequent connecting flight). Even though
carriers departing late can “make time up” during a flight, there is a
downside to that. “Making time up” means accelerating which ends
up burning substantially more fuel and adding thousands of dollars
to the overall flight expense.Thus, carriers have an incentive to
reduce departure delay so as to avoid additional costs in “making
time up.” Some studies show that pilots do try to make up time in
the air,but only for delays that fall into a particular sweet spot.
Recall that market size is measured as the (logged) geometric
mean of the populations at both market endpoints. The coefficient
estimate on MKTSIZE in the departure delay regression is positive
and statistically significant suggesting that larger markets deteri-
orate departure OTP.Thus,the airport congestion argument pre-
vails just like in the arrival delay results.
The market structure variables show similar results to arrival
delay regressions.Once again,the coefficient estimate on the mo-
nopoly dummy variable MONOMKT is positive and statistically
significant,while the coefficient estimate on the number of com-
petitors in a given market NUMCOMP,is negative and statistically
significant for two of the departure delay measures.These results
are consistent with the premise that less competitive markets tend
to have poor departure OTP because of less competitive pressure.
Borenstein and Netz (1999) show that before airlines choose their
departure time,they take into consideration the number of other
non-stop competitors on a route.
5.2.2.Codeshare effects on departure OTP
The remaining rows of Table 11 display key variables of interest
that examine the codeshare effects of the Delta/Northwest/Conti-
nental codeshare alliance on the partner carriers'departure OTP.
Changes in departure OTP are investigated in markets where any
two of the three alliance firms had competed prior to the alliance.
Similarly to the arrivalOTP regressions,we include a dummy
variable DNCamt which equals unity for flights operated by any of
the alliance carriers to examine persistent differences in departure
OTP of flights offered by the alliance partners.The coefficient es-
timate on DNCamt is negative and statistically significantacross
estimations,indicating that the mean departure delay offlights
operated by Delta,Northwest and Continental airlines is less than
the mean departure delay of flights operated by other carriers in
the sample.
The time period dummy Tdnc
t has a positive coefficient estimate
suggesting thatthe mean departure OTP offlights operated by
airlines other than Delta, Northwest and Continental airlines
increased (OTP worsens) from pre-to post-alliance periods.
The coefficient estimate on the interaction variable
Tdnc
t DNC amt represents the difference-in-differences estimate
that identifies whether departure OTP of flights operated by any of
the alliance carriers changed differently relative to departure OTP
of flights operated by other airlines over the pre- and post-alliance
periods.The coefficient estimate is negative and statistically sig-
nificant across estimations, suggesting that the alliance caused the
mean departure OTP for DL/NW/CO flights to increase relative to
the mean departure OTP for non-DL/NW/CO flights over the pre-
and post-alliance periods.In a nutshell, the codeshare alliance
improved departure OTP for the alliance firms relative to other
carriers.
5.2.3.Codeshare effects on departure OTP based on existence of
pre-alliance competition between alliance firms
Columns 2, 4 and 6 in Table 11 reproduce the baseline departure
OTP regressions with the inclusion of the MKTdnc dummy variable.
The effects of the Delta/Northwest/Continental codeshare alliance
on departure OTP in markets where the alliance firms competed
Table 11
Departure on-time performance estimation results.
Variables Departure Delay
in minutes
Departure Delay
in minutes
% of flights departing at
least 15 minutes late
% of flights departing at
least 15 minutes late
% of flights departing at
least 30 minutes late
% of flights departing at
least 30 minutes late
(1) (2) (3) (4) (5) (6)
OUTOFHUB 0.8131*** 0.8090*** 1.2659*** 1.2491*** 0.8167*** 0.8035***
(0.1803) (0.1803) (0.1441) (0.1442) (0.1011) (0.1011)
OPRESCOST 0.0238*** 0.0240*** 0.0287*** 0.0294*** 0.0127*** 0.0132***
(0.0032) (0.0033) (0.0026) (0.0026) (0.0018) (0.0018)
DISTANCE 2.98e-5 2.83e-5 0.0006*** 0.0006*** 0.0005*** 0.0005***
(0.0001) (0.0001) (0.0001) (0.0001) (0.0001) (0.0001)
MONOMKT 0.4188** 0.4313** 4.9897*** 5.0294*** 2.8408*** 2.8667***
(0.1694) (0.1698) (0.1355) (0.1358) (0.0950) (0.0952)
NUMCOMP 0.0438 0.0455 2.4309*** 2.4319*** 1.4493*** 1.4535***
(0.1041) (0.1043) (0.0832) (0.0834) (0.0583) (0.0585)
MKTSIZE 18.5741*** 18.7712*** 25.8526*** 26.4670*** 13.7093*** 14.1043***
(4.7830) (4.7869) (3.8244) (3.8262) (2.6817) (2.6829)
DNCamt 0.6325*** 0.6173** 1.0298*** 0.9519*** 1.0913*** 1.0239***
(0.2394) (0.2413) (0.1914) (0.1929) (0.1342) (0.1353)
Tdnc
t 2.9635*** 2.9608*** 2.6752*** 2.6689*** 1.9002*** 1.8973***
(0.1485) (0.1485) (0.1187) (0.1187) (0.0832) (0.0832)
Tdnc
t DNC amt 1.3356*** 1.3065*** 0.7249*** 0.6581*** 0.6411*** 0.6118***
(0.2197) (0.2239) (0.1756) (0.1789) (0.1232) (0.1255)
MKTdnc 0.486 1.6518*** 1.1398***
(0.4509) (0.3604) (0.2527)
Tdnc
t DNC amt MKT dnc 0.4129 0.8777 0.3321
(0.6715) (0.5367) (0.3764)
CONSTANT 237.3860*** 239.9722*** 325.2973*** 333.3471*** 171.8417*** 177.0110***
(62.4722) (62.5240) (49.9509) (49.9755) (35.0266) (35.0426)
No. of Obs. 31,748 31,748 31,748 31,748 31,748 31,748
R2 0.24 0.24 0.45 0.45 0.35 0.35
The equations are estimated using ordinary least squares.Fixed effects are included in each specification but were not reported for brevity.
Note: Standard errors are in parentheses.***p < 0.01; **p < 0.05; *p < 0.10.
J.O.Yimga / Journal of Air Transport Management 58 (2017) 76e9086

before the alliance formation is determined by summing the co-
efficients on the interaction variables Tdnc
t DNC amt and
Tdnc
t DNC amt MKT dnc in Specification 2 and doing the same for
Specifications 4 and 6 (columns 2, 4 and 6 in Table 11). Even though
the coefficient estimate on Tdnc
t DNC amt MKT dnc has the same
sign as in the arrival delay regression, it is not statistically
significant.
The coefficient estimate on the interaction variable
Tdnc
t DNC amt in columns 2,4 and 6 of Table 11,has a different
interpretation. In fact, the coefficient estimate captures changes in
departure OTP in DL/NW/CO flights due to the codeshare alliance in
markets where they did not compete prior to the alliance.The
coefficient estimate is negative and statistically significant at con-
ventional levels of significance,suggesting thatthe codeshare
alliance improved departure OTP in markets where the alliance
firms did not compete prior to alliance.
Thus, evidence shows that the alliance caused the alliance firms
to improve departure OTP regardless of whether they competed or
not in different markets prior to the alliance.
5.3. Endogeneity concerns and robustness checks
Every empirical study has some endogeneity issues that should
at least be discussed and possibly addressed. The fact that a carrier
may abandon an airport if one of his alliance members assumes the
role of operating carrier, raises potential endogeneity issues of the
variable DPRESCOST.Furthermore,NUMCOMP may be endogenous
especially in markets where the alliance partners were all present
prior to the alliance formation.14
It is worth mentioning that possible endogeneity concerns of
NUMCOMP are mitigated by the fact that a carrier's presence in a
given market requires large investment and the menu of products it
offers are not routinely and easily changed during a short period of
time, which mitigates the influence of on-time performance shocks
Table 12
Arrival on-time performance 2SLS estimation results.
Variables Arrival Delay
in minutes
Arrival Delay
in minutes
% of flights arriving at
least 15 minutes late
% of flights arriving at
least 15 minutes late
% of flights arriving at
least 30 minutes late
% of flights arriving at
least 30 minutes late
(1) (2) (3) (4) (5) (6)
INTOHUB 8.5670*** 8.9483*** 9.5429** 9.9469** 10.6056*** 10.9533***
(2.9277) (3.0613) (4.7112) (4.8637) (3.7333) (3.8721)
DPRESCOST 0.3011*** 0.3122*** 0.2437* 0.2553* 0.2801*** 0.2903***
(0.0826) (0.0864) (0.1331) (0.1375) (0.1055) (0.1095)
MKTSIZE 22.1574*** 21.7516*** 17.7223*** 18.5435*** 11.5098*** 12.5064***
(5.9421) (5.9705) (4.6592) (4.6695) (3.6920) (3.7175)
DISTANCE 0.0024*** 0.0024*** 7.41E-06 3.25E-06 0.0004*** 0.0004***
(0.0001) (0.0001) (0.0001) (0.0001) (0.0001) (0.0001)
RELSPEED 55.3358*** 56.2324*** 19.7368** 20.1206** 22.0697*** 22.5005***
(6.5426) (6.7328) (9.1572) (9.3830) (7.2563) (7.4700)
MONOMKT 0.7706*** 0.7612*** 5.8027*** 5.8564*** 2.9261*** 2.9830***
(0.2190) (0.2184) (0.1818) (0.1766) (0.1440) (0.1406)
NUMCOMP 0.7471*** 0.7205*** 2.7601*** 2.7177*** 1.2972*** 1.2896***
(0.2275) (0.2303) (0.3178) (0.3184) (0.2518) (0.2535)
DNCamt 0.7715** 0.533 1.1661*** 1.2321*** 0.1836 0.3643
(0.3585) (0.3945) (0.3986) (0.4695) (0.3159) (0.3737)
Tdnc
t 4.3251*** 4.3191*** 2.9191*** 2.9038*** 2.0582*** 2.0491***
(0.1845) (0.1857) (0.1448) (0.1464) (0.1147) (0.1166)
Tdnc
t DNC amt 1.0565*** 0.9567*** 0.7899*** 0.6507** 1.0455*** 0.9700***
(0.2951) (0.2983) (0.2807) (0.2752) (0.2224) (0.2191)
MKTdnc 1.8046** 2.0632* 2.8836***
(0.9000) (1.2021) (0.9570)
Tdnc
t DNC amt MKT dnc 1.6609** 2.2595*** 1.0538*
(0.8468) (0.6839) (0.5445)
CONSTANT 220.5515*** 214.1751*** 243.1204*** 254.5643*** 170.7772*** 184.4638***
(77.6543) (78.3253) (60.6062) (61.7263) (48.0255) (49.1421)
Endogeneity Test: F(2,31,591)¼ F(2,31,589)¼ F(2,31,591)¼ F(2,31,589)¼ F(2,31,591)¼ F(2,31,589)¼
H0: NUMCOMP and
DPRESCOST are
exogenous.
8.503*** 8.092*** 65.115*** 62.133*** 62.701*** 61.411***
Wu-Hausman (p ¼ 0.0002) (p ¼ 0.0003) (p ¼ 0.0000) (p ¼ 0.0000) (p ¼ 0.0000) (p ¼ 0.0000)
No. of Obs. 31,748 31,748 31,748 31,748 31,748 31,748
The equations are estimated using ordinary least squares.Fixed effects are included in each specification but were not reported for brevity.
Note: Standard errors are in parentheses.***p < 0.01; **p < 0.05; *p < 0.10.
Fig. 3. Marginal effect of number of competitors in a market on arrivalOTP with a
pointwise 95% confidence interval,from a linear regression model.
14 We thank an anonymous referee for pointing this out.
J.O.Yimga / Journal of Air Transport Management 58 (2017) 76e90 87