Transportation Research Part A 96 (2017) 64–78

Contents lists available at ScienceDirect

Transportation Research Part A

journal homepage: www.elsevier.com/locate/tra

An airline itinerary choice model that includes the option to

delay the decision
Uzi Freund-Feinstein, Shlomo Bekhor ⇑
Faculty of Civil and Environmental Engineering, Technion – Israel Institute of Technology, Haifa 32000, Israel

a r t i c l e i n f o a b s t r a c t

Article history: Choice situations with variable supply characteristics are found in many applications,
Received 26 June 2015 including airline itinerary selection. This paper discusses the airline itinerary choice prob-
Received in revised form 27 November 2016 lem in dynamic supply settings. The paper develops a specially designed stated preference
Accepted 12 December 2016
(SP) survey, which emulates an air travel website. The survey includes the option to delay
Available online 30 December 2016
the decision to choose an airline itinerary. The rich data set allows the estimation of dis-
crete choice models of airline itinerary choice.
Keywords:
The paper presents selected model estimation and application results for two market
Decision making
Dynamic settings
sectors (tourists and business travelers) and two flight types (medium-haul and long-
Choice modeling haul flights). In addition to expected results for several level-of-service variables such as
Air transportation flight cost, cancellation fees, connection times and punctuality percentages, the model esti-
Marketing mates the expected value of delaying a flight purchase.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction

Choice situations with variable supply characteristics are found in many applications, including airline itinerary selection.
Typically, in airlines and travel agencies’ websites, several itinerary characteristics change frequently, such as ticket prices
and seat availabilities. In these situations, the internet acts as a data source and commerce platform, serving both firms and
consumers. Firms are able to publish the dynamic adaptation of their supply according to demand and production cost, while
consumers can continuously search for the most preferred product which maximizes their utility.
A significant number of air transportation applications using discrete choice modeling (DCM) are concerned with itiner-
ary choices in various settings, mainly in the airline or airport levels. Early DCM studies on air transportation applications
applied multinomial logit (MNL) models to investigate the effect of cost/fare on airline/aircraft choice (Gronau, 1970;
Kanafani and Sadoulet, 1977; Nason, 1980). Following these early examples, a variety of topics related to airline choice were
further investigated, such as the effect of frequent flier program (FFP) membership on choice, using econometric models
(Nako, 1992; Chin, 2002). Airline choice research was done also on domestic and international air travel using MNL models.
These include choice of international airlines in South Korea (Yoo and Ashford, 1996), choice of U.S. domestic airlines
(Proussaloglou and Koppelman, 1999) and choice of domestic airlines in Israel (Bekhor and Freund-Feinstein, 2006).
More advanced DCMs in air transport research were implemented mainly due to their better flexibility. Hess and Polak
(2006) investigated the effect of airline, airport and airport’s access mode, using a nested logit model (NL). Warburg et al.
(2006) focused on the influence of business travelers’ characteristics such as demographics on their itinerary choice using
Mixed MNL (MMNL) and MNL models.

⇑ Corresponding author.
E-mail addresses: uzi.feinstein@gmail.com (U. Freund-Feinstein), sbekhor@technion.ac.il (S. Bekhor).

http://dx.doi.org/10.1016/j.tra.2016.12.004
0965-8564/Ó 2016 Elsevier Ltd. All rights reserved.
 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78 65

Another branch, with relatively limited attention, focused on choice behavior using online websites (Collins et al., 2012)
and dynamic pricing (Lin and Sibdari, 2009), both using MNL models. The effect of departure date proximity to choice date on
consumers’ decision making was recently modeled (Drabas and Wu, 2013). This estimation was conducted on three periods
(90, 30 and 5 days prior to departure) using segment-specific Cross-Nested Logit (CNL) models.
Although almost all studies mentioned above modeled choice in a static fashion, itinerary choice situations are dynamic
in their nature, with pricing, availability and other product characteristics allowed to change over time. Given these product
characteristics, the individuals might delay their decisions, sometimes in anticipation of possible better options or till they
finalize their trip arrangements. This paper develops and estimates DCMs for the flight product choice problem in a dynamic
context using standard model estimation methods. In contrast to previous models developed in the literature for airline itin-
erary choice, the models in this paper take into account the option to delay the purchase during the sales period.
The rest of this paper is organized as follows. Section 2 presents a general model framework for the itinerary choice model
problem. Section 3 describes an empirical investigation of medium-haul and long-haul flights. In this section, a specially
designed stated preference survey (SP) and DCM is described, and estimation results are presented. Finally, Section 4 pro-
vides conclusions regarding the case study results.

2. Itinerary choice model framework

The itinerary choice is being modeled here as a quasi-continuous process affected by the occurrence of airlines’ yield
management and other dynamic processes. A graphic representation of this model framework is depicted in Fig. 1. We start
by defining three fundamental issues namely, the supply and demand sides, and the timeframe allocated to each of them.
Supply is provided by various airlines and online travel agents (OTA’s). We assume that each player’s actions on the sup-
ply side are driven by yield management practices causing price, seat availability and other variables to change dynamically
over a designated timeframe (Talluri and van Ryzin, 2004). In addition to dynamic variables, an array of static variables is
also included in each itinerary (departure and arrival times, in-flight entertainment system, etc.). Consumers, forming the
demand side, are able to monitor price changes online for a given itinerary. We assume that consumers, behaving as utility
maximizers, will try to find the itinerary which best fits them (Ben-Akiva and Lerman, 1985). Defining both players’ types’
actions as a continuous event raises the need to define the timeframe scope in which their actions are conducted. Incorpo-
rating each player’s timeframes into a choice model is not trivial and is unique compared to other air transportation studies.
Both airlines and consumers act during two time scopes, which consist of sales period and choice period, respectively. The
airlines set a period starting with initial presentation of a flight on a website and ends when a flight departs, with most yield
management and consumers’ choice activity taking place during 90 days prior to departure. Therefore, the choice period, in
which consumers choose their itineraries (or decide not to choose at all), is identical to the sales period, or shorter
ðtChoice 6 tSale Þ.

Fig. 1. Airline itinerary model framework.

66 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78

Based on these definitions, our model adapts to itinerary choice the dynamic choice framework formulated by Cirillo and
Xu (2011). In this framework, two elements are specified here in the individual’s decision process. The first is system state
variables (dynamic and static), denoted s; the second is the individual’s decision, denoted d. In relation to system state vari-
ables (which are available in an airline or OTA’s websites), these include dynamic variables such as fare, number of seats for
sale, and on-time performance (in certain websites, e.g., Continental Airlines). Also, theses websites presents also static data
which is the itinerary data that includes, among others, departure/arrival and layover times, aircraft type and IFE system.
Since we deal with discrete problems in a stochastic environment, these variables could change at every pre-determined
time period (e.g., daily, monthly, etc.). Therefore, such dynamic settings allow for utility evaluation at each period. Con-
sumers who do not stop the choice process at period t have to determine their utility again at period t = 1 based on their
utility level at time t (i.e., ut). Given this concept of utility maximization, for each state St, consumers will decide on an opti-
mal decision rule dt at each time period, until they conclude that a choice which maximizes their utility is achieved. Alter-
natively, they can decide to stop and cancel their plans. Such evaluation is allowed to take place until the sales period is over,
or all goods/service capacity is completely sold.
Finally, three constraints are imposed on the choice process model in order to simplify the problem and enhance its real-
ism. The first constraint allows consumers to choose a flight only until the day before the departure date. This constraint is
imposed due to certain airlines’ not allowing online purchases during the day prior to the departure date, and most con-
sumers not needing to do so since they finalize their travel arrangement beforehand. The second constraint refers to the
use of choice sets obtained during search one day or more before a current offering. New choice sets supersedes previous
ones, and consumers are not allowed to go back and choose from earlier choice sets even if they present better utilities, sim-
ilar to reality. The third constraint deals with the travel period and renders departure and return dates fixed, not allowing
them to be changed. With the definitions of this model we now continue and empirically investigate the choice behavior
of tourists and business travelers on both medium-haul and long-haul itineraries.

3. Empirical investigation

In this empirical investigation, we estimate choice behavior in dynamic settings on four popular routes originating from
Tel Aviv Airport (TLV), Israel. These routes are to London and Berlin in Europe, and to New York and Toronto in North Amer-
ica. These four cities were chosen based on seven air transportation related criteria (see Table 1). First criterion, cities served
by local airlines, with at least one major carrier of each country in a city-pair. Second criterion, high service frequency, air-
lines operating the most flights per week between city-pairs. Third criterion, high passenger numbers flown between city-
pairs, based on non-stop flights statistics (Civil Aviation Administration Israel, 2013). Fourth criterion, number of stops, non-
stop and single-stop services only, between city-pairs. Fifth criterion, airline model types, with each city-pair served by full-
service carriers (FSCs) and by low-cost carriers (LCCs) (where available). Sixth criterion, alliance membership, representa-

Table 1
Airline characteristics, European and North-American destinations.

Destination Passengers/flights (year) Air-line Frequency weekly Airporta Route Service model Alliance
Berlin 257,000/1815 LY 13 SXF Non-stop FSC None
AB 16 TXL 1-stop Hybrid Oneworld
AF 21 TXL 1-stop FSC SkyTeam
BA 27 TXL 1-stop FSC Oneworld
4U 7 SXF 1-stop LCC None
LH 147 TXL 1-stop FSC Star Alliance
London 772,000/3698 LY 26 LHR Non-stop FSC None
BA 14 LHR Non-stop FSC Oneworld
U2 7 LTN Non-stop LCC None
AF 39 LHR 1-stop FSC SkyTeam
LH 140 LHR 1-stop FSC Star Alliance
New York 1,194,000/4019 LY 29 EWR/JFK Non-stop FSC None
DL 7 JFK Non-stop FSC SkyTeam
UA 28 EWR Non-stop FSC Star Alliance
AC 21 EWR 1-stop FSC Star Alliance
AF 42 JFK 1-stop FSC SkyTeam
BA 112 EWR/JFK 1-stop FSC Oneworld
LH 28 EWR/JFK 1-stop FSC Star Alliance
Toronto 139,000/784 LY 4 YYZ Non-stop FSC None
AC 14 YYZ Non-stop FSC Star Alliance
AF NA YYZ 1-stop FSC SkyTeam
BA 26 YYZ 1-stop FSC Oneworld
LH 28 YYZ 1-stop FSC Star Alliance
UA 55 YYZ 1-stop FSC Star Alliance

Sources: Civil Aviation Authority Israel (2013)

a
See Appendix A for acronym list.
 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78 67

tives of all three global alliances and airlines not part of an alliance. Finally, seventh criterion, passenger types, cities with
diverse passenger mix mainly business travelers and tourists.

3.1. Itinerary data collection

Investigating consumers’ decision-making in dynamic settings, on two route types and four destinations, requires a broad
data collection over a relatively long period. In order to fulfill such a requirement, itinerary information was gathered
between April and May 2012 (regular months apart from one week holiday in Israel), in order to create a database for
the construction of realistic choice sets. Two main online data sources were used to retrieve airlines’ itinerary data. The first
one was websites of ten airlines specifically selected for this study: Air Berlin (AB), Air Canada (AC), Air France (AF), British
Airways (BA), Delta Airlines (DL), EasyJet (U2), El Al (LY), German Wings (4U), and Lufthansa (LH). These websites contain
available itineraries data for flights operated by the airline that owns the website on variety of route(s). The second source
was an online travel agent (OTA’s) website. For this study the website of travelocity.com was selected, which at the time of
data collection provides itinerary data on all airlines serving each route, except LCC’s. Itineraries with the same flight num-
bers were collected from both data sources, except itineraries of airlines not using Travelocity to distribute their itineraries.
The data collection method followed general guidelines suggested by Pope et al. (2009). However, the use of webbots and
scripts for automated parsing (i.e., web scraping) was found to be unwelcomed by certain websites which actively try to
thwart such activity. Therefore, itinerary data was retrieved manually. All websites were accessed simultaneously on a daily
basis, for a period of 30 days, roughly on the same time, requesting itinerary quotes for the same departure and arrival dates.
When data of itineraries were presented by the website, it was copied to a database with all the relevant details, including
the all flight details and fares. In total, 18,604 itineraries were collected representing all round trip flights serving the chosen
markets.
This very large number of itineraries was not usable as is, and had to be reduced significantly. The reason for such a deci-
sion is that most itineraries offered by to a given destination are very similar, and therefore respondents would find it hard to
evaluate the relative utility of one itinerary over another, yielding problematic model estimation. The goal here was to con-
fine the number of alternatives in each set to a reasonable and representative number. Therefore, no more than ten itiner-
aries per day for each market were chosen, which differ significantly from one to the other in their details (e.g., itineraries
with different departure and arrival times, different layovers, different IFE’s, etc.), leading to reducing the number of itiner-
aries in all routes to 1341 only.
In addition to itineraries, supplemental data was collected in order to incorporate cabin amenities and on-time perfor-
mance (OTP) data. Airline service attributes information was collected from Seatguru.com for each market type
(Seatguru.com, 2012a,b). Airlines punctuality data was acquired from Flightstat.com (2012). Finally, itinerary, amenities
and airlines’ technical data from all of these sources was then combined to create the itineraries database.

3.2. Questionnaire design and distribution

Following the data collecting stage, a detailed set of questionnaires were formulated using Qualtrics software. These
questionnaires included three parts, as described in Fig. 2.
The first part contains instructions and screening questions. These included the number of flights each respondent took in
the past three years, the type of the most recent trip and who paid for that trip. In addition to screening out respondents who
are not familiar enough with air travel and/or usage of travel websites, these questions provided the needed information to
assign the respondent to a specific type of questionnaire. For example, those who indicated that their last flight was mainly
for business purposes were automatically assigned a business trip questionnaire.
The second part was dedicated to the SP questionnaire, which presented itineraries choice sets. This questionnaire was
designed to resemble a generic marketing website in order to increase the realism of the experiments (See Appendix B
for a screen shot of the questionnaire). Four pairs of choice sets were designed according to the destination and marketing
channels, with the same itinerary variables presented in Table 2.
Since choice sets were composed to represent the offerings over a relatively long period of time, and the offerings of cer-
tain city-pairs on two different marketing channels, differences were found accordingly. These differences, which included
price levels and unsold capacity, reflect the result of yield management practices conducted by the airlines, and the different
strategies of each marketing channel. The SP questionnaire was designed to include a random numbers engine, which gov-
erned the choice starting point of each respondent within the sales period. This mechanism provided each respondent with
different questionnaire length (i.e., different number of choice sets) compared to the previous and latter respondents. The
third part was designed to gather demographic information, including personal information such as gender, education
and income.
The key part of the SP experiment is related to the possibility to defer the choice of an itinerary. In this case, the survey
platform sent a message indicating that a day has passed, and in the screen a countdown clock was shown (for a few sec-
onds), to give the respondent a graphical impression of time. After this screen, the respondent was presented again with
a similar set of alternatives, but with possibly different values of the attributes, that were randomly sampled from the real
dataset. For example, seat availability could remain the same, decrease, or even show that there are no more available seats
(meaning that specific itinerary is no longer available).
68 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78

Introduction

Vacation Business

Airline OTA

Berlin London NYC Toronto Berlin London NYC Toronto

Demographics

Fig. 2. Questionnaire structure.

Table 2
Explanatory variables in the questionnaire.

Attribute Levels
Dynamically changing attributes
Ticket fare Medium-haul: 408–1710 USD; Long-haul: 1035–2783 USD
Days till flight date 30–1
Cancellation fee No fee, less than ticket price, same as ticket price
Airline punctuality (OTP) 0–100%
Number of seats for sale 1–10 seats, more than 10 seats
Service and amenities attributes
Departure and Arrival times Early morning, morning, mid-day, afternoon, evening, night
Connection time None, up to two hours, 2–5 h, more than 5 h
Legroom (seat pitch) Low, Medium, High
Inflight entertainment system (IFE) None, Shared TV, Personal TV (PTV)
Socioeconomic variables
Frequent flyer El Al Matmid
Frequent flyer Star Alliance Star Alliance (SA)
Income Above average, little above average, same as average, little below average, below average

Sources: Airlines websites (2012), Flightstat.com (2012), Israel Central Beureu of Statistics (2013), Seatguru.com (2012a, 2012b).

To avoid questionnaire burden, the SP experiment included alternatives with direct flights and one connection only. Note
that in the four destinations selected, there is more than one direct flight option, from a different airline company. The speci-
fic connection airport was not presented in the experiment (only the layover connection time).
Following an extensive technical/content pre-test phase, questionnaires were distributed during a period of six months
among various respondent groups located in Israel and the United States. These included students and faculty members from
all universities in Israel and from MIT in the United States found in electronic mailing lists of the universities, business and
non-business travelers. This sample was chosen according to convenience sampling technique due to the difficulty in recruit-
ing current airline passengers while the availability of the above mentioned participants. The survey platform allows to mon-
itor the main sample characteristics and to send reminders to participants, in an attempt to minimize bias due to
participants’ characteristics.
More than 10,000 questionnaire links were distributed using electronic mailing lists, Facebook, and online forums.
Respondents were asked to fill a single questionnaire, totaling 1825 returned questionnaires. Inspections of questionnaire
answers indicated that in some cases many questions were not answered completely or at all, which reduced that number
to 914 usable questionnaires.
Main sample characteristics for the 914 respondents are as follows: 58% male, 70% are 35 years or younger, and 92% have
at least an undergraduate degree. These characteristics reflect the fact that the internet survey was mainly conducted by stu-
dents and faculty staff. Almost two-thirds of the respondents (64%) provided multiple observations totaling 4141 observa-
tions. Table 3 presents the distribution of respondents by passenger and flight types, and the number of observations for each
flight type.

3.3. Choice model structures

Although simple logit models were widely used in previous studies, they have major limitations (Train, 2009). The main
problem is related to the inability to capture similarity among alternatives. Given that several alternatives are similar, the
 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78 69

Table 3
Distribution of respondents by passenger and flight type.

Flight type Respondents Observations

Tourists Business Total
Medium-haul 298 135 433 (47%) 2198 (53%)
Long-haul 329 152 481 (53%) 1943 (47%)
All 627 287 914 (100%) 4141 (100%)

Cross-Nested Logit (CNL) model is suitable for modeling the choice of airline itinerary. This is mainly because the CNL model
can accommodate an alternative to more than one nest, as opposed to MNL and NL models. In addition, since the question-
naire included more than one choice set per experiment, and a significant number of respondents provided more than a sin-
gle observation, the choice data is of panel type. Therefore, the mixed logit version of the CNL model (MCNL) was chosen
since it is more suited for such experiment.
Two types of utility functions were formulated as part of a single set. One, alternative zero, for which represents the ability
of consumers to defer their choice. Two, utility functions for each itinerary and destination, to be compared with alternative
zero. It is important to note that although alternative zero is included for every choice set in each destination, the number of
utility functions differs from one choice set to another (in each destination) according to the availability of itineraries.
Model structures (see Fig. 3) include 11 airline alternatives on the medium-haul model and 16 alternatives on the long-
haul model. In addition to the itineraries’ alternatives, the alternative zero is included, which represents the option to defer
the choice. The models were estimated separately on route type basis (i.e., medium- or long-haul). Since each respondent
was randomly assigned a specific destination, and was asked to choose from a smaller number of itineraries, the available
choice set for each respondent was set in accordance with the itineraries available for the specific destination. The number of
alternatives in the long-haul model is greater compared with the medium-haul version because some airlines provide ser-
vices to both airports in a multi-airport region (i.e., New-York).
For estimation purposes, it is necessary to specify which alternatives were available to the respondent, in order to be con-
sistent with the SP exercise. These include the choice of an Israeli airline flight (El Al) (one alternative in medium-haul and
one or two alternatives in long-haul), and the choice from a variety of foreign airline’s flights (four or five alternatives on
medium-haul and five or eight alternatives on long-haul).
Tables 4 and 5 present the nesting structure for medium-haul and long-haul flights, respectively. The allocation param-
eters of the CNL model were assumed constant and equal to 0.5. The ‘‘Delay” nest includes a single alternative (postpone
purchase).

Flight

Non
AWSsite TCWsite Alliance Postpone
Alliance

AB LY AF BA 4U LH BA U2 LY AF LH
(TXL) (SXF) (TXL) (TXL) (SXF) (TXL) (LHR) (LTN) (LHR) (LHR) (LHR) Postpone

Flight

El Al AWSsite TCWsite FSC Postpone

DL LY UA AC AF BA LH AC LY AF BA LH UA
(JFK) (EWR) (EWR) (EWR) (JFK) (EWR) (EWR) (YYZ) (YYZ) (YYZ) (YYZ) (YYZ) (YYZ) Postpone

LY BA LH
(JFK) (JFK) (JFK)

Fig. 3. Cross-nested logit model structure, medium-haul and long-haul flights. ⁄ See acronyms list in Appendix A.
70 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78

Table 4
Nesting structure – medium-haul flights.

Alternative Nest
Delay Airline website (AWS) Travel company website (TCW) Alliance Non-alliance/El-Al
0 (Postpone) 1
1 (AB – TXL) 0.5 0.5 1
2 (LY – SXF) 0.5 0.5 1
3 (AF – TXL) 0.5 0.5 1
4 (BA – TXL) 0.5 0.5 1
5 (4U – SXF) 1 1
6 (LH – TXL) 0.5 0.5 1
7 (BA – LHR) 0.5 0.5 1
8 (U2 – LTN) 1 1
9 (LY – LHR) 0.5 0.5 1
10 (AF – LHR) 0.5 0.5 1
11 (LH – LHR) 0.5 0.5 1

Table 5
Nesting structure – long-haul flights.

Alternative Nest
Delay Airline website (AWS) Travel company website (TCW) Alliance Non-alliance/El-Al
0 (Postpone) 1
12 (DL – JFK) 0.5 0.5 1
13 (LY – EWR) 0.5 0.5 1
14 (LY – JFK) 0.5 0.5 1
15 (UA – EWR) 1 1
16 (AC – EWR) 0.5 0.5 1
17 (AF – JFK) 0.5 0.5 1
18 (BA – EWR) 1 1
19 (BA – JFK) 0.5 0.5 1
20 (LH –EWR) 0.5 0.5 1
21 (LH – JFK) 0.5 0.5 1
22 (AC – YYZ) 0.5 0.5 1
23 (LY – YYZ) 0.5 0.5 1
24 (AF – YYZ) 0.5 0.5 1
25 (BA – YYZ) 0.5 0.5 1
26 (LH – YYZ) 0.5 0.5 1
27 (UA – YYZ) 0.5 0.5 1

3.4. Model specification

In a static choice environment with constant choice sets, the utility function is generally specified by the level of service
variables of each alternative. However, as described in Section 3.2 above, the SP experiment included the option to defer the
choice. The simple approach presented in this paper is to include the number of days till the flight date in the utility function
of ‘‘alternative zero”, and level of service variables for the other alternatives. This simple approach assumes that consumers
have the same level of service sensitivity, no matter the number of days remaining for the flight.
To capture this sensitivity, the model specification includes an interaction variable between the ticket fare and number of
days. In this specification, the estimated model parameters will indicate the price sensitivity with respect to longer and
shorter periods relative to 15 days before the flight. This cut-off point was found by successive model estimations performed
for different days, and the best combination for the present dataset was related to 15 days.
Based on the above framework, utility function specifications for medium-haul and long-haul flight models are as
follows:
u0 ¼ b1 Days þ b2 Income

ui ¼ b4 Farei  ðDays P 15Þ þ b5 Farei  ðDays < 15Þ þ b6 No CNLi þ b7 OTPi þ b8 AM i þ b9 Seati þ b10 Lay S T i
þ b11 Lay M T i þ b12 Lay L T i þ b13 Leg i þ b14 IFEi þ b15 FFPLY þ b16 FFPSA
where
Farei – Ticket fare (USD) of flight i.
Days – Days till departure.
No CNLi – No cancellation fees of flight i (dummy variable where 1 indicates if there are no cancellation fees).
OTPi – Airline punctuality (on-time percentage) of flight i.
 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78 71

AMi – Early Morning or Morning arrival time (dummy variable where 1 indicates if the arrival time is between 6:00 and
10:00).
Seati – Number of available seats for flight i (dummy variable where 1 indicates more than 10 available seats).
Lay S T i – Short connection time between flights i at a hub (dummy variable where 1 indicates 1–2 h connection).
Lay M T i – Medium connection time between flights i at a hub (dummy variable where 1 indicates 2–5 h connection).
Lay L T i – Long connection time between flights i at a hub (dummy variable where 1 indicates more than 5 h connection).
Leg i – Legroom size in flight i (dummy variable where 1 indicates large legroom).
IFEi – Inflight entertainment system type i (dummy variable where 1 indicates PTV equipped IFE).
FFPLY – El Al frequent flier member (dummy variable where 1 indicates the respondent being a member).
FFPSA – Star Alliance frequent flier member (dummy variable where 1 indicates the respondent being a member).
Income – Respondent income level (dummy variable where 1 indicates high and very high income levels).

3.5. Estimation results

Estimation of the MCNL model was performed using Biogeme software with 1000 Halton draws (Bierlaire, 2003). As indi-
cated in the model specification, the London and Berlin destinations were pooled to form the medium-haul database, and the
Toronto and New York destinations were pooled to form the long-haul database.
Table 6 shows results for two traveler types (tourists and business travelers) and two route types (medium-haul and long-
haul). The explanatory variables are defined similarly for each model, to allow for comparison and analysis of the results.
The main statistical results indicate that all models converge with significant explanatory power (as indicated by rho-bar
square), and parameters signs are as expected. Two nesting coefficients were estimated: the AWS/TCW nest and the Alliance/
Non-alliance nest. The nesting coefficients are according to theory (between 0 and 1, note that in the estimation procedure
the homogeneity parameter is set to 1).
In the tourists’ models, most estimated coefficients are statistically significantly at 5% level of confidence; while in the
business travelers’ models some coefficients’ are not significant greater than 10%, perhaps due to a lower number of obser-
vations, which corresponds to fewer respondents.
Days till departure is a variable in the zero alternative (i.e., postpone purchase), which represents the tendency of the
respondent to postpone his or her choice to a future time within the choice period. This variable takes into account a variety
of reasons for such action (e.g., price, number of available seats, etc.), As expected, the positive sign of the coefficient indi-
cates that, on average, respondents tend to postpone their purchase if an itinerary is offered to them a long time before their
preferred departure date.

Table 6
Mixed CNL model estimation results, for traveler and route types.

Medium-haul Long-haul
Tourists Business Tourists Business
Coeff. t-test Coeff. t-test Coeff. t-test Coeff. t-test
Ticket fare 1 (days P 15) 0.0052 5.4 0.00427 4.2 0.00307 9.1 0.00276 5.6
Ticket fare 2 (days < 15) 0.00457 5.2 0.00353 3.9 0.0027 10.0 0.00203 5.8
Days till departure 0.0316 1.5 0.0654 1.8 0.046 2.1 0.0111 0.4
No cancellation fee dummy 0.42 2.4 0.403 1.6 0.354 3.8 0.333 2.3
OTP (percent) 0.0151 1.6 0.042 2.5 0.0346 7.1 0.024 4.4
AM arrival time 0.465 1.7 0.412 0.9 0.737 5.0 0.0718 0.4
Seat sale dummy 0.587 2.3 0.732 2.7 0.123 1.3 0.249 2.1
Connection time 1–2 h 1.35 3.6 1.14 1.5 0.462 3.9 0.56 3.4
Connection time 2–5 h 1.6 5.1 1.63 3.0 1.06 6.4 0.925 4.3
Connection time more than 5 h 2.46 3.1 2.57 2.5 1.49 5.5 1.17 2.9
Legroom dummy 0.184 0.7 0.5 1.4 0.637 5.0 0.0248 0.2
IFE dummy 0.489 2.4 0.649 2.3 0.176 1.4 0.0075 0.1
El Al FFP dummy 0.37 1.6 0.402 0.9 0.106 1.1 0.227 1.6
Star Alliance FFP dummy 0.556 1.6 0.49 0.8 0.0914 1.0 0.0209 0.2
Income dummy 0.783 2.8 0.242 0.6 0.902 3.1 0.37 1.0
Nesting coef. choice delay 1 Fixed 1 Fixed 1 Fixed 1 Fixed
Nest coef. airline website (AWS) 0.18 2.3 0.14 2.0 0.22 3.3 0.36 2.6
Nest coef. OTA website (TCW) 1 Fixed 1 Fixed 1 Fixed 1 Fixed
Nest coef. El Al/Non-Alliance 1 Fixed 1 Fixed 1 Fixed 1 Fixed
Nest coef. full service carriers (FSC)/Alliance 1 Fixed 1 Fixed 0.17 3.9 0.16 2.7
Sigma 1.78 10.2 1.66 6.4 2.03 10.5 1.72 7.6
Null likelihood 2823.1 848.9 2848.2 1106.7
Final likelihood 992.1 380.2 1232.8 569.4
Rho-bar square 0.642 0.531 0.56 0.468
Number of observations 1692 506 1395 548
Number of individuals 298 135 329 152
72 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78

Ticket fare is a continuous variable with negative sign. As was found in previous studies, fare has a negative impact on
utility (e.g., Adler et al., 2005; Collins et al., 2013; Hess et al., 2007; Proussaloglou and Koppelman, 1999). Therefore, ticket
fare is expected to receive a negative sign to indicate that rising fare lowers one’s utility. Fare is part of each alternative’s
utility function that is compared to the choice delay alternative; the negative sign means that passengers will tend to delay
their choice when prices are high. Comparing results for both tourists and business travelers reveal that both tourists and
business travelers are more sensitive to price on medium-haul flights, and less sensitive on long-haul flights. A possible
explanation for these results is the combination of factors, including cost of travelling, employees’ seniority, the need to
cut costs and business travel guidelines, as shown in previous research papers (e.g., Gustafson, 2012; Unger, 2013). Business
travelers who must comply with their workplace business travel’ guidelines and endeavor to cut costs might get approval for
their air trip, but are denied when cost is excessive. These restrictions are exacerbated when the employee is not senior,
which applies to most of the business-travel respondents in this research.
The coefficient estimates for periods longer than 15 days (‘‘ticket fare 1”) are higher in comparison to the estimates for
periods shorter than 15 days (‘‘ticket fare 2”). The hypothesis test for equality of coefficients at 10% level cannot be rejected
for the medium-haul models, but for the long-haul models the coefficients are significantly different from each other. There-
fore, for long-haul flights, we found that consumers have different price sensitivities with respect to longer and shorter peri-
ods relative to 15 days before the flight.
No Cancellation fee is a dummy variable with a positive sign, indicating that travelers prefer to choose the itinerary with-
out cancellation fees. This result is in line with the observation of Collins et al. (2013), who found that airlines that charge
fees for itinerary changes provide negative utility.
OTP is a continuous variable with a positive sign. This variable provides information on the proportion (percent) of flights
arriving and departing on-time, and changes over time based on each airline’s performance and exposure to exogenous
events (e.g. an airport strike, bad weather, congestion, etc.).1 As shown in Table 6, good OTP levels provides higher utility levels
compared to lower OTP rates. Such results are rather similar to previous research conducted in the U.S.A (Adler et al., 2005; Hess
et al., 2007; Suzuki, 2007).
AM arrival time is a dummy variable which indicates arrival time between 6:00 and 10:00. Note that the coefficient esti-
mates are significant only for tourists, and also the sign changes between medium-haul flights (negative) and long-haul
flights (positive). Although utility of early arrival may be associated with spare time to spend at the destination, our expla-
nation is related to the time of departure (in the specific case of Israel). Flying to a medium-haul destination and arriving
early requires the traveler to take the early flights (which typically departs between 4:30 and 6:00 AM), and the consequent
need to arrive at the airport at least two hours before the flight. This means that the night before the flight is not spent com-
fortably. However, early arrival on long-haul flights allows requires the traveler to arrive late night and have a full-night rest
on a large aircraft, which is relatively more comfortable. These differences between routes and the equipment used on each
one can explain the different utility perceptions as shown in Table 6.
Seat sale is a dummy variable with a negative sign. Recall that this variable is defined as 1 if more than 10 seats are avail-
able for sale on a given search and 0 otherwise. The negative coefficient sign means that when consumers are looking for an
itinerary, they will not rush to choose if they feel that postponing their decision will not threat achieving their ultimate goal
(which is to purchase a flight).
Connection times are represented by three dummy variables (short, medium and long connection times), setting the base
case as non-stop flight. It is hypothesized that non-stop flights provide the highest utility compared to stop-over ones.
Results indicate that such a hypothesis is true, and as expected, the coefficient estimates show increasing sensitivity to
longer connection times. Although not formulated similarly, the utility level of connections was investigated by several
researchers (e.g., Adler et al., 2005; Collins et al., 2013). In their works, the variable included in the utility functions was
the number of connections. In all of these models, the coefficient estimates were negative.
Legroom and IFE are discrete dummy variables with positive signs. It is expected that consumers will indicate higher util-
ity from provisions of greater legroom dimensions and/or entertainment system of higher quality, similarly to case
researched by Collins et al. (2013). Based on this assumption two dummy variables were formulated. The first dummy is
equal to 1 if legroom is the largest; and the second dummy is equal to 1 if a seat is equipped with a personal screen enter-
tainment system.
However, due to correlations with other variables, only in three market segments significant variables were obtained:
tourist long-haul flights (Legroom) and both tourist and business medium-haul flights (IFE). A possible explanation to this
result is that although travel time is long, these variables are less important in comparison other itinerary characteristics.
In particular, since tourists are more price sensitive, they might be willing to tradeoff comfort for lower fare and minimal
cancellation fees. In addition, the tremendous proliferation of smartphones and tablet computers might act as an adequate
substitute, and may shift the passenger attention to other itinerary characteristics.
El Al FFP and Star Alliance FFP are dummy variables, where 1 indicates a respondent holding this specific airline’s FFP
cards, with positive sign. These variables capture the effect of FFP membership parameter on itinerary choices made by pas-
sengers who are members of the respective FFP club. Therefore, the positive coefficient is expected. As with Legroom and IFE
variables, only few variables were found significant, which may be explained by relatively low importance of these variables

1
Departure and arrival on time are each defined as within 15 min of the scheduled time.
 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78 73

in comparison to other level of service variables. These results are somewhat similar to those found by Hess et al. (2007),
Proussaloglou and Koppelman (1999), and Suzuki (2007), although most results are significant.
Income is a socioeconomic dummy variable, where it is equal to 1 if a respondent has high/very high income level. This
variable was included in the delay choice utility function, which all utility functions are compared to, and therefore a ten-
dency to defer itinerary choice will be estimated with negative signs. It is shown that this variable’s coefficients for both pas-
senger types in both route types indeed receive negative signs, which implies that those with higher than average income
will tend to postpone their itinerary choice. However, it should be noted that this variable is significant only in the case of
tourists. An explanation for the lack of significance is based on the identity of the entity paying for the respondent’s
expenses. In case of tourists, they (or a family member) pay from their own pocket, and therefore have an incentive to post-
pone their choice. Those respondents who have a high income will theoretically still be able to pay for an eventual higher
fare in case postponing their decision resulted in having only high priced itineraries to choose from. Business-travelers
expenses, on the other hand, are paid by their employer, so we can assume that the level of their income is not relevant
to their choices but to their employer’s travel guidelines and to work related needs and occurrences.
Comparison of the estimation results of tourists and business travelers in each route type shows differences between
these passenger groups. Furthermore, comparison between medium-haul and long-haul routes raises similar conclusions.
The following section presents an analysis of monetary values of selected characteristics.

3.6. Monetary values of itinerary characteristics

The estimated models allow the calculation of elasticities that help analyze and understand choice behavior. The value of
time is a well-known parameter that is derived in DCMs by simply calculating the ratio between the coefficients of travel
time and cost parameters. Similarly, monetary values of product characteristics can be calculated with the estimated coef-
ficients. In this paper, we calculate the ratio of the different level of service coefficients with respect to the ticket fare, for
selected level of service variables. Tables 7and 8 respectively summarize the monetary values in USD for medium-haul
and long-haul flights. The variables with significance lower than 90% (2-tailed test) are omitted from the tables.
The confidence intervals for each monetary value are also presented in Tables 7 and 8, respectively for periods longer and
shorter than 15 days. The confidence intervals were estimated according to the asymptotic t-test, as presented in Armstrong
et al. (2001). The following equation is used to calculate the confidence intervals:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi


 2
bt tc ðt t t c  qt2 Þ bt tc ðt t tc  qt2 Þ  ðt 2t  t 2 Þðt2c  t 2 Þ

bc tt ðt c  t Þ
2 2 bc tt ðt 2c  t2 Þ
where tt and tc correspond to the t-statistics for bt and bc respectively; t is the critical value for degree of confidence required
and sample size (in this case, approximately 1.64 for 90% confidence interval); and q is the coefficient of correlation between
both parameter estimates.
As can be seen from the tables above, the confidence intervals are rather large, because of the relatively low values of the
t-statistics. In most monetary values, there is a large overlap between tourists and business travelers, and also between med-
ium and long haul flights. Nevertheless, for choice periods shorter than 15 days (Table 8), the point estimates are higher in
comparison to longer periods (Table 7). This is because the ticket fare coefficients for all market segments are higher for
longer periods. The following paragraphs discuss the results based on the point estimates from Tables 7 and 8.
The equivalent value of one day of choice deferring (based on the point estimates) ranges between 6.1 and 18.5 U.S. dol-
lars. This means that the higher the value, the higher is the consumers’ perceived penalty of delaying their decision. For
medium-haul flights, tourists evaluate such risk as lower in comparison to business travelers, which is expected, given
the differences in fares paid by each passenger group and the tendency to purchase tickets closer to the departure date,
which raises the risk of itineraries being sold out. Note however that for long-haul flights, the delay coefficient is not signif-
icant for long-haul flights. This value is captured in the ticket fare coefficient only.
The comparison of the magnitude of this coefficient for business versus tourists reveals that for medium-haul flights,
tourists tend to postpone their decision for a shorter period compared to business travelers. A possible explanation for this
result is that tourists, given their price sensitivity and lack of knowledge on airlines pricing and inventory allocation behav-
ior, will tend to choose faster.
The monetary value for cancelling a flight is higher in long-haul flights in comparison to medium-haul flights, which is
expected due to higher ticket fare. However, no significant differences were found between tourists and business travelers.
Seat sale value is the consumer’s perceived monetary risk for seat capacity to diminish. In both route types, tourists eval-
uate such risk as lower, compared to business travelers. These results could be explained first by the differences in the rel-
ative choice period during the sales period of each consumer type. Tourists are regarded as early choosers, while business
travelers tend to finalize their trip arrangements much closer to the departure dates, and therefore the risks of not being able
to find an itinerary is different between these groups.
Static itinerary characteristics are assigned monetary values, which differs between the two passengers groups. Flying
non-stop is regarded to be worth a significant amount of money by both passenger types, especially by business travelers
whose choices are driven significantly by comfort and willingness to pay for it. Note the high values obtained for long con-
nection types.
74 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78

Table 7
Selected monetary values and confidence intervals, medium- and long-haul flights, choice period 15 days or longer (in USD).

Medium-haul Long-haul
Tourists Business Tourists Business
Ticket fare 1 1 1 1
Days till departure 6.1 15.3 15 –
[0.1, 12.0] [2.7, 28.9] [4.6, 25.5]
Cancellation fee dummy 80.8 94.4 115.3 120.7
[33.7, 127.6] [5.5, 184.4] [76.3, 151.7] [48.5, 182.2]
OTP (percent) 2.9 9.8 11.3 8.7
[0.2, 4.7] [5.3, 13.5] [9.3, 13.4] [6.1, 11.8]
AM arrival time 89.4 – 240.1 –
[10.5, 182.1] [173.8, 310.6]
Seat sale dummy 112.9 171.4 – 90.2
[42.2, 189.2] [75.9, 305.7] [26.7, 160.6]
Connection time 1–2 h 259.6 – 150.5 202.9
[149.8, 401.7] [100.7, 196.7] [130.0, 273.7]
Connection time 2–5 h 307.7 381.7 345.3 335.1
[206.8, 453.1] [187.7, 673.5] [275.5, 420.6] [230.5, 461.3]
Connection time more than 5 h 473.1 601.9 485.3 423.9
[248.7, 751.8] [242.0, 1091.3] [343.4, 651.8] [204.5, 698.0]
Legroom dummy – – 207.5 –
[144.9, 278.4]
IFE dummy 94 152 – –
[38.3, 155.5] [60.2, 259.9]
El Al FFP dummy 71.2 – – 82.2
[3.7, 144.3] [5.1, 158.2]
Star Alliance FFP dummy – – 61.9 –
[3.5, 225.8]

Table 8
Selected monetary values and confidence intervals, medium- and long-haul flights, choice period shorter than 15 days (in USD).

Medium-haul Long-haul
Tourists Business Tourists Business
Ticket fare 1 1 1 1
Days till departure 6.9 18.5 17 –
[0.1, 14.8] [2.9, 40.3] [5.5, 27.3]
Cancellation fee dummy 91.9 114.2 131.1 164
[37.5, 149.9] [6.5, 230.9] [85.7, 173.7] [63.4, 256.6]
OTP (percent) 3.3 11.9 12.8 11.8
[0.3, 5.7] [6.0, 17.8] [10.2, 15.7] [7.8, 16.9]
AM arrival time 101.8 – 273 –
[12.1, 206.2] [194.4, 357.3]
Seat sale dummy 128.4 207.4 – 122.7
[46.3, 225.3] [88.2, 392.5] [36.7, 215.3]
Connection time 1–2 h 295.4 – 171.1 275.9
[169.5, 463.6] [111.1, 229.2] [170.9, 383.5]
Connection time 2–5 h 350.1 461.8 392.6 455.7
[234.5, 521.8] [224.6, 840.4] [304.4, 489.8] [301.8, 648.8]
Connection time more than 5 h 538.3 728 551.9 576.4
[281.3, 868.2] [291.2, 1353.6] [390.2, 737.9] [279.2, 941.5]
Legroom dummy – – 235.9 –
[161.7, 321.0]
IFE dummy 107 183.9 – –
[44.3, 175.6] [75.7, 308.4]
El Al FFP dummy 81 – – 111.8
[4.2, 169.0] [6.7, 218.8]
Star Alliance FFP dummy 121.7 – – –
[4.0, 259.1]

With respect to connection times, in most cases business travelers give higher value to non-stop flights compared to tour-
ists. These differences may be explained by the fact that travelling on business matters is part of their work, therefore these
travelers value convenience which is manifested (among other) by point-to-point travel which enables the travelers to rest
and work during their trip with little interruptions. In addition, the absence of stopovers eliminates the risk of missing con-
necting flights.
 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78 75

Note that there are some counter-intuitive results with respect business trips: for choice periods longer than 15 days, the
point estimates for medium and long connection times for long-haul flights are higher for tourists. In addition, for long-haul
flights, business travelers value OTP lower in comparison to tourists. This might be explained by the relatively small size and
sample composition of business travelers, which is also reflected in the large confidence intervals.
As indicated in the previous section, other level of service variables such as AM arrival time, legroom and IFE were found
significant in certain market segments, and correspondently the monetary values were calculated for these segments. For
example, legroom monetary value was found significant only for the tourist long-haul flights, while IFE monetary value
was significant for the medium-haul flights. Similarly, airlines’ frequent flier programs were valued for the tourist
medium-haul flights. It should be noted that both significance of the personal variables in the models is lower in comparison
to the level of service variables, therefore additional investigation with respect to socio-economic variables should be
conducted.

3.7. Illustration of model application

As the estimation results of the two MCNL models indicates that both fare and cancellation fees do generally affect the
decision to defer choice, it is interesting to see the effect of these attributes when specific choice periods lengths are desig-
nated. In order to do so, the MCNL model is operationalized using four choice periods, and scales of possible fares. This pro-
cedure was conducted separately on medium-haul and long-haul, business and tourists cases and with/without cancellation
fees. Note that this procedure is similar to the study conducted by Drabas and Wu (2013), but with different reference points.
Their study tested 90, 30 and 5 days prior to departure, compared to 30, 20, 10, and 5 days assumed in the present paper. The
reference points assumed were based on a typical booking curve (e.g., see Carrier, 2008).
In this operationalization, two fare scales were chosen for each route type and passenger types. These are 200–1200 USD
for tourists and 200–1800 USD for business travelers on medium-haul routes, while on long-haul routes fare scales were
500–2000 USD and 500–2850 USD for tourists and business travelers, respectively. These scales were formulated arbitrarily
according to the logic of possible minimum fare on each route and estimation of the maximum each passenger type would
agree to pay on each route type. In accordance with the dummy variable presented in the model, cancellation fees were
either zero (i.e., non-restricted itinerary) or the equivalent to the full amount of the requested fare (i.e., restricted itinerary).
For ease of representation, only 2 days prior to departure are shown (30 days and 5 days prior to departure).
The results of operationalizing the MCNL model for medium-haul flights are shown in Fig. 4, and for long-haul flights are
shown in Fig. 5.
In both medium-haul and long-haul flights, it is visible that when a longer choice period is available to the consumer, the
probability to defer choice is higher, and diminishes as the sales period is getting shorter. This finding is consistent in both

Fig. 4. Effect of fare and cancellation fees on the probability to defer choice – medium-haul routes.
76 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78

Fig. 5. Effect of fare and cancellation fees on the probability to defer choice - long-haul routes.

passenger groups and when fare only changes or when fare and cancellation fees change. A second factor, which affects the
probability to defer choice, is the fare levels and/or the inclusion of cancellation fees. In the case of fare only, it was found
that higher fares result in higher probability to defer choice. In a case when cancellation fees are incorporated into the itin-
erary, it leads to a higher tendency to defer choice, while the opposite was found when cancellation was allowed without a
penalty payment.
Given these findings, it is important to note that although they refer to both passenger groups, there are differences
between tourists and business travelers in terms of the magnitude of these factors affecting the choice deferring behavior.
It was found that the effect of fare and/or cancellation fees on choice behavior is higher on tourists than on business trav-
elers. These differences between passenger groups are in line with the traditional classification of business travelers and
tourists according to monetary sensitivities and the identity of the person/entity paying for the itinerary (i.e., the traveler
himself in case of tourists or the employee’s workplace in case of business travelers). Finally, differences in the probability
to defer choice were found between different route types. As depicted in the graphs it can be seen that the probability to
defer itinerary choice is higher on medium-haul flights than on long-haul flights. However, on both route types and for both
passenger types the level of defer probability is lower as the choice time frame is shorter. An explanation to such differences
is the higher cost of purchasing an itinerary and the higher cancellation fees in the case of long-haul flights compared to
those in medium-haul.

4. Conclusions

This paper presented a model of product choice in dynamic settings with application to airline itinerary choice, and esti-
mation results of a mixed cross-nested logit model. The specially designed questionnaire allowed the option to defer choice.
Both tourists and business travelers’ decision making behavior on medium- and long-haul routes were estimated using an SP
questionnaire which resembled a generic travel website.
Modeling itinerary choice in dynamic settings requires analyzing results in two dimensions. The first dimension regarding
choice is the effect of dynamically changing product characteristics, while the second dimension is the effect of static char-
acteristics. Both of them are two sides of a coin and should be considered as a whole. Analysis results imply that tourists and
business travelers treat both dynamic and static itinerary attributes differently. Dynamic attributes included itinerary fare,
cancellation fees, punctuality (OTP) and number of seats available for sale; in addition, the tendency to postpone the choice
decision was estimated as well.
 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78 77

We found differences between the effect of these dynamic attributes with respect to tourists and business travelers, and
also between medium and long-haul flights. For example, business travelers are more highly affected by seat availability,
cancellation fees and OTP than tourists. In addition, they are willing to postpone their choice for longer periods of time. How-
ever, for long-haul flights, business travelers seem to be less affected by the OTP attribute. It must be noted that these results
are affected by the large number of respondents which are faculty members and graduate students travelling as part of their
work/studies. These respondents were regarded as business passengers (according to the question related to the purpose of
their most recent travel), but their trip related characteristics might be different from regular business persons in terms of
trip guidelines, itinerary flexibilities and budget levels.
We have found further differences with respect to static itinerary attributes, such as service details and frequent flyer pro-
gram participation. The estimation results show that business travelers’ choices are convenience-driven, in contrast with
tourists. In addition, we found significant differences between medium- and long-haul flights with regard to both dynamic
and static variables. Following the estimation of discrete choice models, we have used the coefficients for operationalizing
medium and long-haul models. The results of this procedure reveal the effect of different choice periods as well as fare and
cancellation fees on choice behavior as a whole and the differences between business travelers and tourists.
Our results are in line with choice behavior literature (e.g., Proussaloglou and Koppelman, 1999; Collins et al., 2012;
Drabas and Wu, 2013), where business travelers are less cost-sensitive but relatively more sensitive to convenience, while
tourists are characterized to be the opposite. The inclusion of dynamically changing variables in choice modeling, specifically
the alternative of delaying the purchase, allows a better perception of the level of service variables and provides new
insights. Furthermore, our methodological framework has promising abilities to account for dynamic behavior – a research
area not fully addressed in the literature. From the modeling perspective, the MCNL model structure accounts for a similarity
among alternatives, a feature not captured in simpler model structures.
It should be noted that the models and results obtained in this paper assume that the person travelling is the person that
actually chooses and purchases the flight ticket online. This assumption may be questionable for business travelers, because
someone else may choose and buy the tickets for them. Further research is needed to verify this assumption.
The results of this study were based on a comprehensive stated preference survey. The models estimated in this paper
could be further improved with revealed preference data about flight choice. Another possible line of research is modeling
the threshold values that represent the individual choice set. This paper assumed that every alternative presented to the
respondent was considered. Potential further research will investigate the availability of specific alternatives at the individ-
ual level.

Appendix A. List of acronyms

AB - Air Berlin
AC - Air Canada
AF - Air France
AWS - Airline websites
BA - British Airways
DL - Delta Air Lines
EWR - New York Newark Liberty International Airport
FSC - Full-service carrier
JFK - New York John F. Kennedy International Airport
LCC - Low-cost carrier
LH - Lufthansa
LHR - London Heathrow Airport
LTN - London Luton Airport
LY - El Al
SXF - Berlin Schönefeld Airport
TCW - Travel agents’ website
TXL - Berlin Tegel Airport
UA - United Airlines
U2 - easyJet
YYZ - Toronto Pearson International Airport
4U - German Wings

Appendix B. SP questionnaire interface

B.1. Flight from NYC to Tel Aviv

The following are the flight itineraries available 3 days prior to the flight.
Please consider the flight options presented and decide the extent to which they meet the requirements.
78 U. Freund-Feinstein, S. Bekhor / Transportation Research Part A 96 (2017) 64–78

If you choose to purchase one of the available options, rank your three preferred flight options in order of preference,
where 1 = most preferred itinerary and 3 = least preferred itinerary.
You may choose to postpone your decision. To do this, click ‘‘yes” on the drop-down menu next to the option marked ‘‘I
prefer to postpone my choice of an airline ticket to the next day”.

Flight Departure Arrival Departure Layover Total flight On-Time Leg room In-flight Total price Cancellation fee Availabl
Number Time Time Airport time performance entertainment (USD) e seats
269 00:40 05:30 JFK Direct flight 11:50 76% L (32’’) Shared TV 1,738.00 Non-refundable 4
v
027 10:40 15:55 EWR Direct flight 12:15 26% L (32’’) Personal TV 2,078.99 Free 2
v
001 01:00 05:40 JFK Direct flight 11:40 36% L (32”) Personal TV 1,830.89 Free 10+
v
085 11:25 16:35 EWR Direct flight 12:20 68% M (30’’) Personal TV 1,378.99 Non-refundable 10+
v
085/762 13:10 07:58
(next day)
EWR Toronto, 12:00 25:48 68% L (32’’) Shared TV 2,782.69 Free 10+
v
222/006 08:10 16:10 JFK Paris, 01:45 15:00 95% L (32”) Personal TV 1,695.99 $200 10+
v
162/189 08:05 18:55 EWR London, 04:35 17:50 95% M (30’’) Personal TV 1,748.19 $100 1
v
691/760 05:00 14:00 EWR Frankfurt, 02:50 16:00 82% M (30’’) Personal TV 1,359.89 $240 10+
v
691/400 05:00 12:50 JFK Frankfurt, 01:50 14:50 77% M (30’’) Personal TV 1,189.19 $240 10+
v
I prefer to postpone my choice of an airline ticket to the next day v

References

Adler, T., Falzarano, C., Spitz, G., 2005. Modeling service trade-offs in air itinerary choices. Transport. Res. Rec.: J. Transport. Res. Board 1915, 20–26.
Armstrong, P., Garrido, R., Ortúzar, J.D., 2001. Confidence intervals to bound the value of time. Transport. Res. Part E: Logist. Transport. Rev. 37 (2), 143–161.
Bekhor, S., Freund-Feinstein, U., 2006. Modeling passengers’ preferences on a short-haul domestic airline with rank-ordered data. Transport. Res. Record: J.
Transport. Res. Board 1951, 1–6.
Ben-Akiva, M., Lerman, S.R., 1985. Discrete Choice Analysis: Theory and Application to Travel Demand. MIT Press, Cambridge, Mass.
Bierlaire, M., 2003. BIOGEME: a free package for the estimation of discrete choice models. In: Proceedings of the 3rd Swiss Transportation Research
Conference. Ascona, Switzerland.
Carrier, E., 2008. Modeling the Choice of an Airline Itinerary and Fare Product Using Booking and Seat Availability Data Ph.D. Dissertation. MIT, Cambridge,
Mass..
Chin, A.T., 2002. Impact of frequent flyer programs on the demand for air travel. J. Air Transport. 7 (2), 53–86.
Cirillo, C., Xu, R., 2011. Dynamic discrete choice models for transportation. Transport Rev. 31 (4), 473–494.
Civil Aviation Autority Israel, 2013. Demand, Supply and Competition Level in the Passenger Air Transportation Sector From/to Israel. Civil Aviation
Authority Israel, International Aviation Relations Division. Airport City: CAAI.
Collins, A.T., Rose, J.M., Hess, S., 2012. Interactive stated choice surveys: a study of air travel behavior. Transportation 39, 55–79.
Collins, A.T., Hess, S., Rose, J.M., 2013. Choice modelling with search and sort data from an interactive choice experiment. Transport. Res. Part E: Logist.
Transport. Rev. 56, 36–45.
Drabas, T., Wu, C.-L., 2013. Modelling air carrier choices with a Segment Specific Cross Nested Logit model. J. Air Transport Manage. 32, 8–16.
Flightstat.com. (2012). Flight Status Track. <http://www.flightstats.com/> (Retrieved 15/01–14/02/2012).
Gronau, R., 1970. The Value of Time in Passenger Transportation: The Demand for Air Travel. NBER. UMI, New York.
Gustafson, P., 2012. Managing business travel: developments and dilemmas in corporate travel management. Tourism Manage. 33 (2), 276–284.
Hess, S., Adler, T., Polak, J.W., 2007. Modelling airport and airline choice behaviour with the use of stated preference survey data. Transport. Res. Part E:
Logist. Transport. Rev. 43 (3), 221–233.
Hess, S., Polak, J.W., 2006. Airport, airline and access mode choice in the San Francisco Bay area. Papers Region. Sci. 85 (4), 543–567.
Kanafani, A., Sadoulet, E., 1977. The partitioning of long-haul air traffic – a study in multinomial choice. Transp. Res. 11 (1), 1–8.
Statistical Abstract of Israel, 2013. Households, by size of household, religion and population group - Table 5.1. Stat. Abstr. Israel 64, 256.
Lin, K.Y., Sibdari, S.Y., 2009. Dynamic price competition with discrete customer choices. Eur. J. Oper. Res. 197, 969–980.
Nako, S.M., 1992. Frequent flyer programs and business travellers: an empirical investigation. Logist. Transport. Rev. 28, 395–414.
Nason, S.D., 1980. Analyzing ticket-choice decisions of air travelers. Transp. Res. Rec. 768, 20–24.
Pope, S., Garrow, L., Guin, A., Leonard, J., Bankston, L., Campbell, P., 2009. Conceptual framework for collecting online airline pricing data: challenges,
opportunities, and preliminary results. Transport. Res. Record: J. Transport. Res. Board 2106, 30–37.
Proussaloglou, K., Koppelman, F.S., 1999. The choice of air carrier, flight, and fare class. J. Air Transport Manage. 5, 193–201.
Suzuki, Y., 2007. Modeling and testing the ‘‘two-step’’ decision process of travelers in airport and airline choices. Transp. Res. Part E 43, 1–20.
Talluri, K., van Ryzin, G., 2004. Revenue management under a general discrete choice model of consumer behavior. Manage. Sci. 50 (1), 15–33.
Train, K., 2009. Discrete Choice Methods With Simulation. Cambridge University Press, New York.
Seatguru.com. (2012a). Comparison Charts: Long-Haul Economy Class. <http://www.seatguru.com/> (Retrieved 15/04/2012).
Seatguru.com. (2012b). Comparison Charts: Medium-Haul Economy Class. <http://www.seatguru.com/> (Retrieved 15/04/2012).
Unger, O., 2013. Business Travel Experience. Unpublished Master Thesis. Ben Gurion University, Ben Gurion.
Warburg, V., Bhat, C., Adler, T., 2006. Modeling demographic and unobserved heterogeneity in air passengers’ sensitivity to service attributes in itinerary
choice. Transport. Res. Record: J. Transport. Res. Board 1951, 7–16.
Yoo, K., Ashford, N., 1996. Carrier choices of air passengers in pacific rim: using comparative analysis and complementary interpretation of revealed
preference and stated preference data. Transport. Res. Record: J. Transport. Res. Board 1562, 1–7.