Technical Report Documentation Page
1. Report No.
SWUTC/03/167520-1
2. Government Accession No.
3. Recipient's Catalog No.
5. Report Date
October 2003
4. Title and Subtitle
A Parameterized Consideration Set Model for Airport Choice: An
Application to the San Francisco Bay Area
6. Performing Organization Code
7. Author(s)
Gözen Basar and Chandra Bhat
8. Performing Organization Report No.
Report 167520-1
10. Work Unit No. (TRAIS)
9. Performing Organization Name and Address
Center for Transportation Research
The University of Texas at Austin
3208 Red River, Suite 200
Austin, Texas 78705-2650
11. Contract or Grant No.
10727
13. Type of Report and Period Covered
12. Sponsoring Agency Name and Address
Southwest Region University Transportation Center
Texas Transportation Institute
Texas A&M University System
College Station, Texas 77843-3135
14. Sponsoring Agency Code
15. Supplementary Notes
Supported by general revenues from the State of Texas.
16. Abstract
Airport choice is an important air travel-related decision in multiple airport regions. This report proposes
the use of a probabilistic choice set multinomial logit (PCMNL) model for airport choice that generalizes
the multinomial logit model used in all earlier airport choice studies. This study discusses the properties of
the PCMNL model, and applies it to examine airport choice of business travelers residing in the San
Francisco Bay Area. Substantive policy implications of the results are discussed. Overall, the results
indicate that it is important to analyze the choice (consideration) set formation of travelers. Failure to
recognize consideration effects of air travelers can lead to biased model parameters, misleading evaluation
of the effects of policy action, and a diminished data fit.
17. Key Words
Air Travel, Metropolitan Area Planning, Discrete-
Choice Models, Hazard Duration Models, Traveler
Behavior
18. Distribution Statement
No restrictions. This document is available to the
public through NTIS:
National Technical Information Service
5285 Port Royal Road
Springfield, Virginia 22161
19. Security Classif.(of this report)
Unclassified
20. Security Classif.(of this page)
Unclassified
21. No. of Pages
54
22. Price
Form DOT F 1700.7 (8-72) Reproduction of completed page authorized
A Parameterized Consideration Set Model for Airport Choice:
An Application to the San Francisco Bay Area
by
Research Report SWUTC/03/167520-1
Southwest Regional University Transportation Center
Center for Transportation Research
The University of Texas at Austin
Austin, Texas 78712
October 2003
Gözen Amber B
asar
and Chandra R. Bhat
iv
DISCLAIMER
The contents of this report reflect the views of the authors, who are responsible for the
facts and the accuracy of the information presented herein. This document is disseminated under
the sponsorship of the Department of Transportation, University Transportation Centers
Program, in the interest of information exchange. Mention of trade names or commercial
products does not constitute endorsement or recommendation for use.
v
ABSTRACT
Airport choice is an important air travel-related decision in multiple airport regions. This
report proposes the use of a probabilistic choice set multinomial logit (PCMNL) model for
airport choice that generalizes the multinomial logit model used in all earlier airport choice
studies. This study discusses the properties of the PCMNL model, and applies it to examine
airport choice of business travelers residing in the San Francisco Bay Area. Substantive policy
implications of the results are discussed. Overall, the results indicate that it is important to
analyze the choice (consideration) set formation of travelers. Failure to recognize consideration
effects of air travelers can lead to biased model parameters, misleading evaluation of the effects
of policy action, and a diminished data fit.
vi
ACKNOWLEDGEMENTS
The authors recognize that support for this research was provided by a grant from the
U.S. Department of Transportation, University Transportation Centers Program to the Southwest
Region University Transportation Center which is funded 50% with general revenue funds from
the State of Texas. The authors would like to thank Ken Vaughn and Chuck Purvis of the
Metropolitan Transportation Commissions (MTC) in Oakland for providing help with data
related issues.
vii
EXECUTIVE SUMMARY
In contrast to the increasing contribution of air travel to urban travel, airport-related travel
is still treated in a rather coarse and simplified manner within the urban travel modeling
framework of most Metropolitan Planning Organizations in the State and the Country. In
particular, airports are identified as “special attractors” and assigned a certain number of trip
attractions, without adequate systematic analysis of the spatial and temporal patterns of the trip
attractions. It is important for transportation agencies to consider a more systematic approach to
analyze and forecast airport-related personal travel, so that improved predictions of traffic
characteristics and traffic levels on urban roadways may be achieved. A systematic analysis of
airport travel is also important for mobile-source emissions forecasting.
An important choice dimension related to airport travel is the origin departure airport
choice in a multiple airport region. A multiple airport region is one in which a passenger living
within has the option of departing and/or arriving from more than one airport. Common
examples that have been used as regions of study in the past include New York City, the San
Francisco Bay Area, Chicago, and the Washington, D.C./Baltimore region. A good
understanding of the factors underlying a passenger’s origin airport choice in multiple airport
regions can enable airport management and airline carriers to attract passengers, upgrade airport
facilities and equipment to meet projected air travel demands, and determine airport staffing
needs. It can also aid Metropolitan Planning Organizations in forecasting travel demand in the
urban region, and in planning transportation networks to/from airports.
The research in this report proposes the use of a probabilistic choice set multinomial logit
(PCMNL) model for airport choice that generalizes the multinomial logit model used in all
earlier airport choice studies. This study discusses the properties of the PCMNL model, and
applies it to examine airport choice of business travelers residing in the San Francisco Bay Area.
Substantive policy implications of the results are discussed. Overall, the results indicate that it is
important to analyze the choice (consideration) set formation of travelers. Failure to recognize
consideration effects of air travelers can lead to biased model parameters, misleading evaluation
of the effects of policy action, and a diminished data fit.
viii
ix
TABLE OF CONTENTS
CHAPTER 1. INTRODUCTION.................................................................................................1
CHAPTER 2. PREVIOUS WORK..............................................................................................5
2.1 Background..........................................................................................................................5
2.2 Airport Choice in Isolation ..................................................................................................5
2.3 Airport Choice Along With Other Dimensions of Air Travel.............................................7
2.4 Choice Set Formation ..........................................................................................................9
CHAPTER 3. MODEL STRUCTURE......................................................................................11
3.1 Background........................................................................................................................11
3.2 Formulation .......................................................................................................................12
3.3 Properties ...........................................................................................................................13
CHAPTER 4. DATA SOURCES ...............................................................................................15
4.1 Primary Data Source..........................................................................................................15
4.2 Secondary Data Sources ....................................................................................................17
CHAPTER 5. EMPIRICAL ANALYSIS ..................................................................................19
5.1 Variable Specification .......................................................................................................19
5.2 Estimation Results .............................................................................................................20
5.2.1 The MNL Model Results ...........................................................................................22
5.2.2 The PCMNL Model Results ......................................................................................22
5.3 Trade-off Between Access Time and Frequency of Service .............................................23
5.4 Substantive Policy Implications ........................................................................................25
5.5 Measures of Data Fit..........................................................................................................27
CHAPTER 6. SUMMARY AND CONCLUSIONS .................................................................31
REFERENCES ............................................................................................................................33
Appendix A. Sample of Choices Involved in Air Travel..........................................................35
Appendix B. Literature Review Table.......................................................................................36
Appendix C. Questions in the MTC Air Passenger Survey.....................................................39
Appendix D. Data Screening Process.........................................................................................41
Appendix E. Top Thirty Domestic Destinations.......................................................................42
Appendix F. Variables Used to Come to a Preferred Specification........................................43
x
LIST OF FIGURE
Figure 1. Study Area......................................................................................................................16
LIST OF TABLES
Table 1. Estimation Sample Shares, Market Shares, and Weights................................................17
Table 2. Estimation Results...........................................................................................................21
Table 3. Time Value of Frequency of Service at Choice Stage ....................................................24
Table 4. Elasticity Effects of Quality of Service Improvements...................................................26
Table 5. Measures of Fit in Estimation and Validation Sample....................................................29
1
CHAPTER 1. INTRODUCTION
Since airline deregulation in 1978, there has been a dramatic increase in the number of
passengers flown per year. Airline deregulation has generated substantial economic benefits for
the vast majority of the traveling public. Because of lower fares and better overall level of
service, demand for air travel has increased. Within the context of intercity travel, air travel is
the fastest growing travel mode in the United States. Notwithstanding the events of September
11, 2001, projections suggest that the number of air travelers in the U.S. will double in this first
decade of the 21
st
century. Further, airports are increasingly serving as freight gateways to
facilitate long-distance commodity movement nationally and internationally. As the number of
air travelers and amount of air freight movements increase, so will the contribution of airport-
related travel to overall urban traffic levels. In addition, increases in person travel and freight
lead to higher staffing needs at airport, thus increasing commuting travel to/from airports.
In contrast to the increasing contribution of air travel to urban travel, airport-related travel
is still treated in a rather coarse and simplified manner within the urban travel modeling
framework of most Metropolitan Planning Organizations in the State and the Country. In
particular, airports are identified as “special attractors” and assigned a certain number of trip
attractions, without adequate systematic analysis of the spatial and temporal patterns of the trip
attractions. It is important for transportation agencies to consider a more systematic approach to
analyze and forecast airport-related personal travel, so that improved predictions of traffic
characteristics and traffic levels on urban roadways may be achieved. A systematic analysis of
airport travel is also important for mobile-source emissions forecasting.
There are several dimensions characterizing air traveler decisions that impact the spatial
and temporal distribution of trips to the airport. For residents of an urban area, some of the first
decisions regarding inter-urban travel may include whether to travel away from the urban area
and to where, the duration of the trip, and the mode for the inter-urban trip (i.e., whether to travel
by air, or some other mode). If air is the mode of choice, the relevant decisions include the
destination airport, the origin airport in a multi-airport urban area, the desired arrival time at the
destination (which impacts the desired flight departure time at the origin), the location and
departure time to the origin airport, and the access mode of transport to the airport. In addition to
2
these choices, other air traveler decisions that would be of relevance to air carriers and airport
management include air carrier choice, fare class of travel, and method of purchase of tickets
1
.
The many dimensions of air travel identified above are clearly inter-related. Ideally, the
analyst would prefer a modeling structure that models all these dimensions jointly. But such a
joint framework is infeasible in practice, and thus the analyst needs to adopt a sequential
structure that may be assumed to reasonably represent the air travel choice process. For one
possible choice hierarchy, please refer to Appendix A. This flowchart represents only one
possible hierarchy of decisions within the context of air travel. The hierarchy of decision
depends on several factors including a passenger’s travel purpose and a passenger’s sensitivity to
variables such as time and cost. For example, if a passenger is extremely price sensitive then he
or she might first jointly choose an airline and travel destination based on special deals at the
time, and then choose the vacation time period depending on when it is cheapest to fly. In
contrast, a passenger traveling on business often has a specific time and day on which he or she
must fly, so they choose to fly the airline that offers the most convenient schedule.
An important choice dimension, which precedes most other air travel decisions in the
choice framework, is the origin departure airport choice in a multiple airport region.
Specifically, a multiple airport region is one in which a passenger living within has the option of
departing and/or arriving from more than one airport. Common examples that have been used as
regions of study in the past include New York City, the San Francisco Bay Area, Chicago, and
the Washington, D.C./Baltimore region. A good understanding of the factors underlying a
passenger’s origin airport choice in multiple airport regions can enable airport management and
airline carriers to attract passengers, upgrade airport facilities and equipment to meet projected
air travel demands, and determine airport staffing needs. It can also aid Metropolitan Planning
Organizations in forecasting travel demand in the urban region, and in planning transportation
networks to/from airports.
Multiple airport regions can be classified into one of two categories. The first of these is
a metropolitan area where there is more than one airport, and where the airports all tend to be
hubs or large-scale operations offering similar services. The second type is that in which
regional airports compete with larger, neighboring airports. The two cases can be analyzed in
similar ways, though it is interesting to note that different factors defining a passenger’s choice
1
Refer to Appendix A for a sample hierarchy of choices involved in air travel.
3
prevail in each scenario. For example, when departing from a regional airport one usually
connects through one of the neighboring airports, depending on the destination. Passengers can
instead opt to travel to these larger airports by personal vehicle, rail, or bus and travel directly
from them. In this case, regional airports lose passengers to the larger airports, but equally
importantly, regional airports lose passengers to various other modes including personal vehicle,
rail, and bus. Passengers might choose to forego the services of their regional airports, and travel
long distances to the larger airports because of factors such as availability of nonstop flights, jet
service, or lower ticket prices.
In the first scenario, where multiple airports compete with one another in large,
metropolitan areas, these same factors (jet service, ticket prices, etc) may not come into play.
For the most part, when dealing with larger airports, the variability of services to destinations is
not as apparent, therefore passengers may choose airports based on specific departure times of
flights, specific airline availability, or because of airport familiarity. The focus in this study is on
the first of the two scenarios, multiple airports in a metropolitan area competing with one
another.
The rest of this study is structured as follows. The next section discusses previous work
in the area of airport choice. Section 3 presents the model structure. Section 4 discusses the data
source and sample formulation procedures. Section 5 describes the empirical results. The final
section highlights the important findings of this study.
4
5
CHAPTER 2. PREVIOUS WORK
2.1 Background
Several earlier studies have examined airport choice in a multiple airport region. Some
of these studies have focused on airport choice in isolation, while others have examined airport
choice along with other dimensions of air travel. These earlier studies have focused on different
urban areas and, sometimes, different population groups (such as business travelers versus
leisure travelers and residents versus non-residents). Following is a detailed review of many of
the previous studies in the area of airport choice.
2
2.2 Airport Choice in Isolation
One of the first airport choice models was developed by Skinner (1976). The area of
study was the Baltimore-Washington, D.C. region, which includes three major airports
(Baltimore, Washington Dulles, and Reagan National). A multinomial logit model was
estimated with variables for flight frequency and ground accessibility of each airport. Skinner
stratified the passengers into two groups: business and non-business. He concluded that
passengers are more sensitive to airport accessibility than to flight frequency.
Harvey (1987) estimated a passenger airport choice model using data from the San
Francisco Bay Area. He used a multinomial logit structure with three airport alternatives (San
Jose International, San Francisco International, and Oakland International). Passengers were
stratified into resident business and resident non-business. Airport access time and flight
frequency were found to be significant determinants of airport choice. Harvey’s conclusions
were that the value of time is lower for non-business travelers while their value of funds is higher
relative to business travelers. Another conclusion was that all travelers prefer direct flights to
commuter and connecting flights. As for future work, Harvey suggested extending the analysis
to include lower-level choices such as access mode in a nested logit framework.
Ashford and Benchemam (1987) estimated a multinomial logit model for airport choice
in Central England. The five airport alternatives were Manchester, Birmingham, East Midlands,
Luton, and London Heathrow. The passengers were segmented into domestic, international
2
For a detailed table of the literature review of previous work within the context of airport choice, please refer to
Appendix B.
6
business, international non-business, and international inclusive tours travelers. The final
variables in the model were travel time to the airport and flight frequency for international
business and international inclusive tours travelers. Flight frequency, travel time to the airport,
and airfare were the final variables in the model for the remaining market segments. Ashford
and Benchemam concluded that business travelers are most sensitive to airport access time, and
that leisure travelers are most sensitive to both airfare and airport access time relative to the other
variables.
Ozoka and Ashford (1987) studied air traveler behavior in Nigeria as a comparison to air
traveler behavior in both the United Kingdom and the United States. Passengers traveling from
two airports, Enugu and Benn, to one common destination were the focus of the study. The
development of a multinomial logit model allowed for the prediction of the effect of a third
airport in the area. Variables considered were weekly flight frequency from each airport to the
destination, economy class airfare, and airport access travel time. Flight frequency and airfare
were found to be insignificant, probably due to the fact that the two airports offered similar flight
schedules and airfares to the same destination. Airport access time was found to be significant,
implying that any improvements to airport access would greatly influence passengers’ choice of
airport. The significance of airport access time indicates that airports in Nigeria, much like the
rest of the world, do compete with one another.
Innes and Doucet (1990) used a binary logit structure for airport choice in the northern
province of New Brunswick, Canada. Their binary structure gave a passenger the choice of
flying from the closest airport to them and the second closest. Their region of study was unique
because it was the northern part of the province of New Brunswick, Canada, where travelers are
faced with the choice of flying out of regional airports, or traveling to another airport where there
are not as many restrictions compared to their local airport. Their region of study is a good
example of the second type of area that can be analyzed (discussed in section 1): where regional
airports compete with neighboring hubs. Original variables were ticket type, who the ticket was
paid by, length of stay at destination, type of aircraft, availability of nonstop service, and
difference in flying time to destination (again comparing the nearest airport to the next closest).
The distance variables were eventually dropped from the model, and Innes and Doucet focused
on level of service. They found that type of aircraft plays a significant role in airport choice, and
7
that air travelers are willing to travel long distances in order to have access to jet service. Also,
Innes and Doucet found that passengers prefer direct flights to connecting flights.
Thompson and Caves (1993) estimated a multinomial logit model to forecast the potential
market share for a new airport in North England that would serve six destinations. Passenger
survey data as well as data for flight services (average fares, flight frequency, aircraft type) from
1983 were used to compare the predicted services of Sheffield (the new airport) with those of
Birmingham, East Midlands, and Manchester airports. The final variables used in the estimation
were access time to the airport, daily flight frequency, and the maximum number of seats
available on an aircraft serving a specific destination; passengers were stratified into business
and leisure groups. With the assumptions made regarding flight frequency and fare offered from
Sheffield airport, Sheffield was projected to take 86% of business travelers (traveling to the six
destinations) living within an hour of the airport. Additionally, Thompson and Caves found that
individuals living closer to an airport value access time significantly more than any other
variable, and are only slightly affected by changes in flight frequency and fare, compared with
people living further away who show higher sensitivity to changes in flight frequency and
airfare.
Windle and Dresner (1995) estimated a multinomial logit model with weekly flight
frequency data and airport access times. They also included a chooser specific variable that
indicated how many times during the past year a passenger had used each of the airports. Their
analysis was in the Baltimore-Washington, D.C. region, including 30 domestic destinations. As
in many previous studies, they found that the level of service variables were significant, and that
the choice specific variable was highly significant. The more an individual used an airport, the
more likely they were to use it in the future.
2.3 Airport Choice Along With Other Dimensions of Air Travel
In addition to airport choice in isolation, many studies throughout the years have focused
on origin airport choice within the context of other air travel choice dimensions. These include,
but are not limited to, destination airport, ground access mode to the airport, and airline choice.
Following is a summary of some of these studies.
Ndoh, Pitfield, and Caves (1990) compared the multinomial logit (MNL) structure with
that of the nested multinomial logit (NMNL) structure to analyze passenger route choice in
8
Central England. They found that the NMNL model where the selection process is route type,
followed by choice of hub airport, and then departure airport, is statistically superior to the MNL
model with the alternatives being different routes. Business travelers were found to value access
time the most, followed by weekly flight frequency on a direct route in their choice of route.
Other significant variables in the model were average journey time, average connection time to
hub airport, and weekly available aircraft seats on each route.
Furuichi and Koppelman (1994) studied air travelers’ departure airport and destination
choice behavior using data from an international air traveler behavior survey administered in
Japan in 1989. Their study included passengers traveling on nonstop flights to international
destinations from four major airports in Japan. A nested logit structure was used for the choice
of departure airport and destination among both business and pleasure passengers. The preferred
model specification for business and pleasure travelers was the same; the variables used for
airport choice were access time and costs to the airport, line-haul time and cost from an
individual’s departure airport to their destination, and the relative flight frequency (flights from
departure airport to chosen destination to the sum of flights at all other possible departure
airports). The variables used in the destination choice model were the log sum variable for
access and line-haul service and the log of the trade value (international trade between Japan and
destination area).
One finding of this study was that both business and pleasure travelers value access travel
cost to the airport more than line-haul travel cost; this finding implies that cost could be valued
differently depending on the type of expenditure. Additional findings were that all travelers
place a high value on flight frequency, access and line-haul time.
Pels, Nijkamp, and Rietveld (2000, 2001) developed a joint airport-airline choice model
for the San Francisco Bay Area using a nested logit structure. Flight frequency and ground
access time were found to be significant. They used number of seats in an aircraft as a proxy for
comfort, and found this variable to be significant as well. Contrary to previous findings, they
found there to be little difference between the estimations for business and leisure travelers.
They also found the choice hierarchy of passengers first choosing departure airport, and then
choosing airline to be more statistically favorable than the opposite case.
Pels, Nijkamp, and Rietveld (2002) estimated an access mode – airport choice model for
the San Francisco Bay Area. A nested logit structure was used, with airport choice at the top
9
level and access mode choice at the lower level being the preferred structure. Resident business
and resident non-business travelers from the Bay Area were examined. Access mode choice was
a function of access cost and access time, while airport choice was a function of airfare and flight
frequency. They found that business travelers have a higher value of time than leisure travelers,
and that access time to the airport is one of the most important factors in airport choice.
A common finding in all these studies (airport choice in isolation and airport choice in
the context of other air travel dimensions) is that access time to the airport and frequency of
service from the airport to the desired destination are the dominant factors affecting airport
choice. Several of these studies also suggest that a simple measure of access time to the airport;
i.e., auto access time; performs as well as more complex formulations that consider multiple
modes and both access time and access cost. In addition, many earlier studies find that airfare is
not a significant factor in airport choice for business travelers, though a few studies find airfare
to affect airport choice for non-business travelers.
2.4 Choice Set Formation
The current study contributes to the existing body of literature by focusing on airport
choice in the San Francisco Bay Urban Area context. An important characteristic of the current
study is its recognition that travelers may not consider all the available airports when making the
choice of their departure airport. Earlier research on choice set generation has indicated the
important impact of consideration effects on consumer choice (see, for example, Roberts and
Lattin, 1991; Ben-Akiva and Boccara, 1995; Chiang et al., 1999).
Despite the importance of choice sets, all the airport choice models discussed earlier
assume that each traveler makes a choice from the full set of available airports, where an airport
is assumed to be available if there is at least one flight (direct or connecting) from the airport to
the destination city. Such an assumption is rather untenable because an individual’s choice set is
likely to depend on the traveler’s specific sociodemographic, informational, psychological, and
societal contexts as well as subjective criteria associated with individual attitudes/perceptions.
For example, an individual may consider a particular airport to be too far away to be even
considered, while another individual may consider this distance to be acceptable. Similarly, an
individual may eliminate from consideration any airport that does not have airline club lounges,
while another may include airports without airline club lounges in her/his choice set. Thus, it is
10
important to recognize that different travelers may, and in general will, consider different sets of
alternatives.
To be sure, considering the choice set formation process along with the actual choice
process is not merely an esoteric econometric issue. Earlier research in the transportation and
marketing fields has indicated that failure to properly specify the choice set considered by
consumers can lead to biased choice model parameters, a lack of robustness in parameter
estimates, and violations of the independence from irrelevant alternatives assumption (see
Shocker et al., 1991; Swait, 1984; and Williams and Ortuzar, 1982). On the other hand, the
explicit incorporation of consideration effects has both methodological and managerial benefits.
Methodologically, the incorporation of consideration effects can lead to a more accurate
prediction of the choice process being modeled (see Gensch, 1987; Chiang et al., 1999; and
Swait, 2001). Such prediction gains will result in improved forecasting of travel demand to/from
airports. Managerially, the recognition of consideration effects can help determine the relative
effects of policy relevant variables on consideration and choice, and thus aid in a comprehensive
understanding of the impacts of policy actions (discussed in sections 5 and 6). The important
point to note here is that regardless of the relative utility of an airport compared to other airports
in a traveler’s choice set, the airport will not be chosen if it is not first considered (see Andrews
and Srinivasan, 1995).
In addition to the methodological issue of modeling the choice set generation process and
airport choice from the choice set, the current study also considers the impact of
sociodemographic and trip characteristics of the traveler on airport choice. Harvey (1987) is one
of the only earlier studies that recognizes demographic impacts, but that study did not find any
statistically significant effects of personal characteristics on airport choice.
11
CHAPTER 3. MODEL STRUCTURE
3.1 Background
The model structure used in this study is based on Manski’s (1977) original two-stage
choice paradigm, which includes a probabilistic choice set generation model in the first stage
followed by the choice of airport from a given choice set.
The first stage uses a probabilistic choice set generation mechanism because the actual
choice set of travelers is unobserved to the analyst and, therefore, cannot be determined with
certainty by the analyst. Within the class of probabilistic choice set generation models, Swait
and Ben-Akiva’s (1987a) random constraint-based approach to choice formation is adopted (for
a detailed discussion of other approaches to probabilistic choice set generation, see Ben-Akiva
and Boccara, 1995). In the random constraint-based approach, an airport is excluded from the
choice set if the consideration utility for that airport is lower than some threshold consideration
utility level (the reader will note that the consideration of an airport is determined only by the
threshold level of that airport, not by any comparisons to the thresholds of other airports). Since
the threshold utility level is not observed to the analyst, the exclusion of an airport from the
choice set becomes probabilistic. In the current study, the consideration utility is allowed to vary
across individuals, so that the consideration probability of each airport varies across individuals.
Almost all earlier applications of probabilistic choice set generation have used the same
consideration probability across individuals (but see Andrews and Srinivasan, 1995). Swait and
Ben-Akiva (1987b) allow the consideration probabilities to vary across individuals, but their
parameterized logit captivity (PLC) model constrains consumers to be either captive to a single
alternative or to choose from the full set of alternatives. The parameterized choice set model in
the current study is more general, and allows consumers to choose from all possible choice set
sizes.
The second stage airport choice model, given the choice set, is based on the familiar
multinomial logit formulation. At this stage, the utilities of the airports in the choice set are
compared directly with each other in a utility maximizing process. The difference in the process
at the choice set generation and choice stages enables a change in an attribute associated with an
airport to have two separate effects: a consideration effect (i.e., the impact on the consideration
12
set of airports) and a choice effect (i.e., the impact on the choice of an airport, given that the
airport is considered by the individual).
3.2 Formulation
The model formulation in this section is developed assuming that all airports are feasible
for each traveler (though not all of the airports may be considered by each traveler). This
assumption simplifies the presentation and is consistent with the empirical context of the current
report, where each airport has at least one direct or connecting flight in the day to each traveler’s
destination airport.
Let the consideration utility of airport i (i=1,2,…,I) for individual q be
qi
U . The
alternative is included in the choice set if this consideration utility exceeds a certain threshold
and is eliminated if not. Since the threshold is not observed to the analyst, it is considered as a
random variable. In the current study, this random threshold is assumed to be standard
logistically distributed. Then, the probability that alternative i is considered by individual q can
be written as:
qi
w
qi
e
M
γ
′
−
+
=
1
1
, (1)
where
qi
w is a column vector of observed attributes for individual q and alternative i (including a
constant) and γ is a corresponding column vector of coefficients to be estimated (this coefficient
provides the impact of attributes on the consideration probability of alternative i).
Next, assume that the randomly-distributed threshold for each alternative is independent
of the threshold values of other alternatives. The overall probability of a choice set c for
individual q may then be written as:
∏
∏∏
=
∈∉
−−
−
=
I
i
qi
cicj
qjqi
q
M
MM
cP
1
)1(1
)1(
)( , (2)
where the denominator is a normalization to remove the choice set with no alternatives in it.
The choice of airport from a given choice set can be written, using a multinomial logit
formulation, as:
13
, if 0
if |
ci
ci
e
e
cP
cj
x
x
qi
qi
qi
∉=
∈=
∑
∈
β
′
β
′
(3)
where
qi
x is a column vector of exogenous variables and
β
is a column vector of coefficients
indicating the effect of variables at the choice stage.
Finally, the unconditional probability of choice of alternative i can be written as follows:
∑
∈
⋅=
Gc
qqiqi
cPcPP )()|( , (4)
where G is the set of all nonempty subsets of the master choice set of all airport alternatives. The
membership of G will include
1
2
−I
elements. For example, in a three airport case, denoted as
{A,B,C}, G includes the following choice sets:{A}, {B}, {C}, {A,B}, {B,C}, {A,C}, {A,B,C}.
The log-likelihood function for the estimation of the parameters
β
and
γ
is:
log ‹ ),(log),( γβ⋅=γβ
qiqi
Py , (5)
where
qi
y is a dummy variable taking the value 1 if individual q chooses airport i and 0
otherwise. Maximization of the log-likelihood function is accomplished using the GAUSS
matrix programming language.
3.3 Properties
The parameterized probabilistic choice set multinomial logit (PCMNL) model structure
presented in the previous section nests the multinomial logit structure as a special case. In
particular, the probability function of Equation (4) collapses to the MNL model if 1=
qi
M for all
alternatives i and all individuals q (also note that 1→
qi
M when +∞→γ
′
qi
w for all i and q). In
this situation, P
q
(c) = 0 for all choice sets c that are subsets of the master choice set and P
q
(c) = 1
for the master choice set, which is equivalent to assuming that all individuals consider all
airports.
The disaggregate-level elasticity effects in the PCMNL model can be computed from the
probability expression in Equation (4) in a straightforward manner (however, the author is not
aware of any earlier study presenting these expressions). In the following presentation of
elasticity expressions, the index q for individuals is suppressed for notation ease. Let
c
i
δ be a
dummy variable taking the value 1 if choice set c contains airport i and 0 otherwise, and let
c
ij
δ
14
be another dummy variable taking the value 1 if choice set c contains both airports i and j and 0
otherwise. Also, define
i
B as follows:
∑
∏
∈
−−
=δ=
Gc
k
k
i
c
ii
M
M
cPB
)1(1
)(
. (6)
Then the self- and cross-elasticities of a change in the m
th
attribute of an airport )(
im
zi that
appears at both the consideration stage and choice stage can be written as follows:
{}
{}
im
Gc
mji
j
m
Gc
i
c
ijj
j
P
z
im
Gc
mii
i
mi
P
z
zcPcPcP
P
BcPcP
P
zcPcPcP
P
B
j
im
i
im
β−+γ
−δ⋅=η
β−+γ−=η
∑∑
∑
∈∈
∈
)()|)(|(
1
)()|(
1
)()|1)(|(
1
)1(
(7)
The expression above comprises two terms. The first term represents the consideration elasticity
and captures the impact of a change in
im
z on the consideration of airport i in the self-elasticity
expression and on the consideration of airport j relative to airport i in the cross-elasticity
expression. The second term represents the substitution elasticity at the choice stage conditional
on the alternative being available in the choice set. Note that for a variable that does not appear
in the consideration stage, only the substitution elasticity applies in each of the expressions. On
the other hand, for a variable that does not appear at the choice stage, only the consideration
elasticity applies. In any case, the cross-elasticity expression is a function of the choice
probability for mode j. Thus, the PCMNL model does not exhibit the IIA property of the MNL
model. It is also easy to verify that the self- and cross-elasticity expressions collapse to those of
the MNL when all airports are considered.
15
CHAPTER 4. DATA SOURCES
4.1 Primary Data Source
The primary data source for this study is an air passenger survey conducted by the
Metropolitan Transportation Commission in the San Francisco Bay Area. This survey was
administered to randomly selected travelers in August and October of 1995 at four airports: San
Francisco International (SFO), San Jose International (SJC), Oakland International (OAK), and
Sonoma County (STS). The full data set included 21,124 samples, and was comprised of
twenty-one survey questions. Information collected in the survey included purpose of travel,
destination, size of the traveling party, mode of transport to the airport, airline carrier, and flight
details. Passengers were also asked how many flights they took from each of the six Bay Area
airports during the past twelve months. In addition, sociodemographic attributes of the traveler
such as gender and income were obtained
3
.
In the current research, the survey responses from the three major Bay Area airports;
SFO, SJC, and OAK; are used because of the very low share of travelers using the Sonoma
County airport
4
. For ease in data preparation and assembly, the top thirty domestic destinations
from these three Bay Area airports are identified from the sample and the airport choice of Bay
Area residents to these top destinations are considered
5
. These top thirty destinations are served
from each of the three Bay Area airports, either through direct flights and/or connecting flights.
Thus, all the three airports are available as potential choices, though not all of them may be
considered by travelers (please refer to Figure 1 on the following page for a diagram of the three
airports in this study).
The air travel market is segmented, for the purpose of this analysis, into business and
non-business trip purposes. To narrow the focus, only business trips are considered in this study.
The final business sample comprises 1,918 observations, of which 1,618 observations are used
for estimation and the remaining 300 observations are set aside as a validation sample for
evaluating the performance of an ordinary multinomial logit (MNL) model and the
parameterized probabilistic choice set multinomial logit (PCMNL) model of this report. The
sample shares and the market shares in the estimation sample are presented in Table 1.
3
For a listing of all questions asked in the survey, refer to Appendix C.
4
Please refer to Appendix D for a flowchart of the data screening/preparation process.
5
Please refer to Appendix E for a listing of the thirty destinations used in this study.
16
Figure 1. Study Area
17
Table 1. Estimation Sample Shares, Market Shares, and Weights
Airport
Estimation
sample shares
Market shares Weight
1
San Francisco International (SFO)
0.2559
0.6248
2.4420
San Jose International (SJC)
0.4932
0.1775
0.3596
Oakland International (OAK)
0.2509 0.1977 0.7882
1
The weight variable refers to the weight placed on individuals choosing each airport. Thus, for example,
each individual in the estimation sample choosing SFO is assigned a weight of 2.4420 during estimation.
As can be observed, there is an over sampling of travelers flying out of San Jose in the airport
survey (the actual shares of airport choice in the population are obtained from the Bureau of
Transportation Statistics). Since the sample is choice-based with known aggregate shares, the
Weighted Exogenous Sample Maximum Likelihood (WESML) method proposed by Manski and
Lerman (1977) is employed in estimation. This method weights the log-likelihood value for
each individual in Equation (5) by the ratio of the market share of the airport chosen by the
individual to the sample share of the airport chosen by the individual (the resulting weights are
presented in the final column of Table 1). Maximizing the resulting likelihood function provides
consistent estimates of the parameters. The asymptotic covariance matrix of parameters is
computed as
11 −−
∆
H
H
, where H is the hessian and ∆ is the cross-product matrix of the gradients
(H and
∆ are evaluated at the estimated parameter values). This provides consistent standard
errors of the parameters (Börsch-Supan, 1987).
4.2 Secondary Data Sources
In addition to the air passenger travel survey, three other secondary data sources are used
to develop the final sample. The first is a zone-to-zone ground access level of service file,
obtained from the Metropolitan Transportation Commission in Oakland. This information is
appropriately appended to the sample observations based on the originating zone of departure to
the airport and the zone that contains each airport. In the current analysis, level-of-service (time
18
and cost) values corresponding to the highway mode are used, since a majority of the trips to the
airport are pursued by a private or rental car
6
.
The second secondary data source used in the analysis is the daily flight frequency from
each Bay Area airport to the thirty destination airports, obtained from the 1995 Official Airline
Guide (Official Airline Guide Market Analysis, 1995)
7
. This information is appended to the
sample observations based on the origin-destination airport pair and the day of week of travel.
The third source of data is on-time flight statistics for nonstop flights from each airport to each
destination, obtained from the Bureau of Transportation Statistics (BTS). These data provide the
percentage of late flights, defined as the percentage of flights delayed beyond 15 minutes of the
scheduled departure time
8
.
The three secondary data sources discussed above provide measures of the quality of
service offered by each airport for the traveler’s trip.
6
Eighty-six percent of passengers in the estimation sample traveled to the airport by either private or rental car.
7
These include nonstop flights and flights with a stop but no change in equipment.
8
The BTS on-time flight statistics are for 1997, and its use in the current analysis assumes the absence of significant
changes between 1995 and 1997
19
CHAPTER 5. EMPIRICAL ANALYSIS
5.1 Variable Specification
The choice of variables for potential inclusion was guided by previous empirical work on
airport choice modeling, intuitive arguments regarding the effects of exogenous variables, and
data availability considerations. Three broad classes of variables were considered for inclusion:
(1) quality of service variables, (2) interactions of sociodemographics with quality of service,
and (3) interactions of trip characteristics with quality of service.
The quality of service variables, as discussed earlier, included ground-access level of
service variables (time and cost) and air travel level-of-service variables (flight frequency to
destination and percentage of late flights). Traveler sociodemographic variables considered in
the analysis included the gender, age, and household income of the traveler. Finally, the trip
characteristics explored in the specifications included the following dummy variables: (a) an
“alone” variable identifying whether or not the individual was traveling alone, (b) a “short trip”
variable representing if the traveler was away for fewer than 2 nights or 2 or more nights, (c) a
“car used to reach airport” variable indicating whether the traveler used a car (private or rented)
to reach the airport, (d) a “weekday” variable indicating if the trip was pursued on a weekday or
the weekend, and (e) a “left to airport from work” variable identifying if the traveler left to the
airport from work or from a nonwork location.
9
Additionally, although some earlier studies have
found nonstop flights to be a significant factor in airport choice, the variable was not included in
this study since almost all passengers in the estimation sample flew nonstop flights to their final
destination.
In the early stages of this study the significance of an airport loyalty variable was
explored in an attempt to capture airport desirability characteristics as well as a measure for
airport familiarity. The airport loyalty variable came out to be highly significant, showing that
the more a passenger flies out of one airport relative to the other airports in the area, the more
likely they are to fly out of this airport again. Although this variable came out to be statistically
significant, it was excluded from further estimation because of the fact that airport loyalty is
most likely a function of the other variables in the model. Likewise, though the “percentage of
late flights” variable came out to be significant, it was excluded from further estimation.
9
Please refer to Appendix F for a listing of those variables considered.
20
Preliminary results with respect to this variable indicated that passengers were likely to choose
airports with poor on-time flight performance records. The assumption was made that results
were opposite from what would have been expected either due to a) inaccurate data since the
only on-time flight data available was from 1997, while the air passenger survey was from 1995,
or b) the percentage of late flights variable was highly correlated with some other, desirable
characteristic of an airport such as size, number of airlines serving the airport, or simply overall
airport activity.
Several nonlinear forms for capturing the effect of access time and flight frequency were
explored in this analysis. But the simple linear functional form for access time and flight
frequency performed as well as the more complex functional forms. The arrival at the final
specification was based on a systematic process of eliminating variables found to be insignificant
in previous specifications and based on considerations of parsimony in representation.
5.2 Estimation Results
The results of the multinomial logit (MNL) model and the parameterized probabilistic
choice set multinomial logit (PCMNL) model are presented in Table 2 and discussed in the
subsequent two sections.
21
Table 2. Estimation Results
PCMNL Model
MNL Model
Consideration Stage Choice Stage
Variable
Parameter t-statistic Parameter t-statistic Parameter t-statistic
Access time-related variables
(access time is in 100s of minutes)
Access time -6.964 -13.43 -2.185 -1.91 -7.503 -11.13
Access time x traveling alone -0.825 -1.76 --- --- -2.169 -3.24
Access time x female -0.796 -1.92 1.748 2.02 -0.701 -1.11
Access time x weekday travel --- --- -3.788 -2.89 --- ---
Frequency-related variables
(frequency is in flights per day divided by 10)
Frequency 0.411 2.88 3.893 5.83 0.360 2.19
Frequency x traveling alone 0.271 1.87 --- --- 0.232 1.40
Frequency x female -0.173 -1.36 --- --- -0.092 -0.62
Frequency x high income indicator
(annual income > 150K)
-0.257 -2.09 --- --- -0.581 -2.85
Frequency x weekday travel --- --- 1.832 2.00 --- ---
Airport Constants
San Francisco International --- --- 3.826 3.08 --- ---
San Jose International -1.998 -12.30 -0.595 -1.90 -1.659 -8.44
Oakland International -2.162 -17.17 -1.531 -3.14 -1.522 -10.98
22
5.2.1 The MNL Model Results
The coefficients on the access time variable in the multinomial logit model indicate, as
one would expect, that business travelers are averse to traveling long durations to reach an
airport. This is particularly the case for individuals traveling alone and women travelers. The
coefficients on the frequency variable indicate a preference for airports that have frequent flight
service to the traveler’s destination. Individuals traveling alone, in particular, place a premium
on frequency. This result, along with the higher access time sensitivity of individuals traveling
alone, suggests that time is less onerous when traveling in a group (perhaps because of the
opportunity to socialize or conduct business when traveling together). The results also indicate
the lower sensitivity of women and high-income individuals to frequency of service. The latter
result is a little surprising, but may be a reflection of high-income individuals traveling at narrow
peak-period time windows of the day, and thus not being sensitive to the frequency of flights
over the entire day. Frequency of service does not impact airport choice for high-income women
travelers.
5.2.2 The PCMNL Model Results
The PCMNL model includes estimates of the probabilistic choice set generation model as
well as the airport choice model. The coefficients at the consideration stage provide estimates of
the
γ
vector in Equation (1). Table 2 shows that the coefficients on the access time and
frequency variables at the consideration stage are statistically significant, indicating variation in
the consideration of each airport across individuals. In particular, airports that are farther away
and/or that have a low frequency of flights are less likely to be considered by individuals. As
one would expect, these effects are magnified on weekdays compared to weekends.
Additionally, women appear to be more willing than men to consider airports that are distant
from their point of departure to the airport.
The coefficient estimates in the choice stage in the PCMNL model have interpretations
that are similar to those in the MNL model. However, there are differences in the magnitude of
the access time impacts. Specifically, the access time effects at the choice stage are higher than
the corresponding MNL estimates. The reason is that airports that are very far away are
“removed” from consideration in the PCMNL model. For example, consider an individual with
one close airport and two very distant airports, and assume that this individual considers only the
23
close airport. For this individual, access time has no impact (by definition) at the choice stage
(the probability of choice of the close airport is one, given that the choice set includes only that
airport). Thus, the sensitivity to access time at the choice stage in the PCMNL model is
automatically based on data from individuals who have a high probability of consideration of
two or more airports, and who are sensitive to access time at the choice stage. The MNL model,
on the other hand, includes relatively “captive” individuals in the choice model estimation,
despite these individuals not being sensitive to access time. The result is a dilution of the
sensitivity to access time in the MNL choice model. The impact of frequency at the choice stage
of the PCMNL model is not very different from the MNL model.
The combination of results at the consideration and choice stages shows that access time
is less important for women when developing the perception “space” of availability of airports,
but is more important for women when choosing an airport from the choice set of available
airports.
5.3 Trade-off Between Access Time and Frequency of Service
The coefficients on time and frequency can be used to examine the trade-offs between the
two determinants of airport choice. For example, the MNL model indicates that male, low-
income, individuals traveling in a group would be willing to travel about 6 minutes
[=0.411/(6.964/100)] longer if the frequency of flight service were to be increased by ten flights
per day. The corresponding values for other traveler subgroups are provided in Table 3 for both
the MNL and PCMNL models. In general, these results indicate that access time is the dominant
determinant of airport choice for business travelers, particularly for high-income group travelers.
In addition, the PCMNL values indicate that, at the choice stage, access time is an even more
dominant determinant than suggested by the MNL model.
The time values of frequency can also be computed for the consideration stage from the
PCMNL model. Interestingly, these values are very high. An additional flight per day from an
airport has the same impact on consideration utility as 18 less minutes to that airport for male
weekend travelers, 90 less minutes for female weekend travelers, 9.5 less minutes for male
weekday travelers, and 13.5 less minutes for female weekday travelers. These results show the
relatively dominant effect of frequency at the consideration stage, especially on weekends.
24
Table 3. Time Value of Frequency of Service at Choice Stage
Population Subgroup MNL
PCMNL
1, 2
Male, low-income, traveling in a group 5.9 4.8
Male, high-income, traveling in a group 2.3 --
Male, low-income, traveling alone 8.8 6.1
Male, high-income, traveling alone 5.5 ≈ 0
Female, low-income, traveling in a group 3.1 3.3
Female, high-income, traveling in a group ≈ 0 --
Female, low-income, traveling alone 5.9 4.8
Female, high-income, traveling alone 2.9 ≈ 0
1
The numbers indicate the additional access time travelers are willing to endure for an increase in ten flights per day
to their destination.
2
A “--” entry indicates that frequency has a negative effect at the choice stage for the corresponding population
group. While not intuitive, these negative frequency effects
are not significantly different from zero.
25
5.4 Substantive Policy Implications
The relative effects discussed above provide useful information about the effects of
access time and frequency on choice in the MNL model, and separately on consideration and
choice in the PCMNL model. However, these effects do not provide a measure of the absolute
magnitude of impacts. Further, in the PCMNL model, the overall effects of access time and
frequency are not directly discernible from the coefficients at the consideration and choice
stages.
To examine the overall effects of access time and frequency, we now compute the
aggregate self- and cross-elasticities. These aggregate elasticities provide the proportional
change in the expected market shares of each airport in response to a uniform percentage
improvement in access time and frequency across all individuals. The aggregate self- and cross-
elasticities can be obtained from the disaggregate-level elasticities presented in Equation (7).
Table 4 shows the elasticity effects for the MNL and PCMNL models.
Several common conclusions may be drawn from the elasticities of the MNL and
PCMNL models. First, in the overall, access time is a more important determinant of airport
choice than is air service frequency. This is consistent with several earlier studies on airport
choice. Second, the self-elasticities indicate that Oakland International is best positioned to
improve its market share through improvements in its quality of service (note the higher self-
elasticity effects for Oakland compared to the self-elasticity effects of the other two airports).
Third, San Francisco International has tremendous “clout” in the market, since it can easily
negate attempts by other airports to draw away share by making its own marginal service
improvements (see the much higher cross-elasticities corresponding to improvements in SFO’s
quality of service compared to the cross-elasticities corresponding to improvement in the quality
of service of other airports).
26
Table 4. Elasticity Effects of Quality of Service Improvements
Elasticity Impact on Market Share
MNL Model PCMNL Model
Improvement in Quality of Service
SFO SJC OAK SFO SJC OAK
San Francisco International (SFO)
Decrease in travel time 1.313 -1.597 -2.715 0.870 -1.150 -1.709
Increase in air frequency 0.277 -0.393 -0.524 0.169 -0.237 -0.322
San Jose International (SJC)
Decrease in travel time -0.220 1.111 -0.301 -0.205 0.971 -0.223
Increase in air frequency -0.054 0.227 -0.034 -0.096 0.400 -0.056
Oakland International (OAK)
Decrease in travel time -0.566 -0.306 2.063 -0.431 -0.251 1.582
Increase in air frequency -0.114 -0.053 0.409 -0.152 -0.079 0.549
27
The substantive policy implications from the MNL and PCMNL models, while similar in
some ways, are also quite different in others. First, compared to the MNL model, the PCMNL
model indicates substantially lower self- and cross-elasticities corresponding to access time. If
the PCMNL model is a more appropriate model (as we will clearly demonstrate in the next
section), use of the MNL model would overestimate the potential gain in an airport’s market
share due to an improvement in access time to that airport and would overestimate the reduction
in market share of other airports due to such an access time improvement. Second, the PCMNL
model shows higher self- and cross-elasticities corresponding to improvement in air frequency
from San Jose and Oakland airports. This can be attributed to the strong impact of air frequency
on consideration of an airport in the PCMNL model, as discussed in the previous section. The
reason why such an effect does not extend to San Francisco is that San Francisco already has a
very high consideration level in the market. In fact, the overall consideration level can be
estimated from the parameter estimates in Table 2. Defining
i
S as the share of individuals who
consider airport i when making a choice, we can write:
Q
cPw
S
qGc
q
c
iq
i
∑∑
∈
δ
=
)(
, (8)
where
q
w is the weight for individual q, Q is the total number of individuals in the sample, and
other quantities are as defined earlier. The estimated values of airport consideration are 99.4%
for SFO, 77.2% for SJC, and 70.7% for OAK. Clearly, there is little room to increase the
consideration level of SFO, which is the reason for the low self- and cross-elasticities
corresponding to air service frequency improvement for SFO.
To summarize, the substantive implications for policy analysis from the MNL and
PCMNL models are different in the current empirical context. These differences suggest the
need to apply formal statistical tests to determine the structure that is most consistent with the
data. This is the focus of the next section.
5.5 Measures of Data Fit
The fit of the MNL and PCMNL models is evaluated in both the estimation sample and a
validation sample. In the estimation sample, the standard measures of fit, including the log-
28
likelihood at convergence and the adjusted likelihood ratio index are computed. The adjusted
likelihood ratio index is defined with respect to the log-likelihood at market shares:
)(
)
ˆ
(
1
2
c
M
L
L −β
−=ρ , (9)
where
)
ˆ
(βL and )(cL are the log-likelihood functions at convergence and at market shares,
respectively, and M is the number of parameters estimated in the model (besides the alternative
specific constants of the choice model). In addition, the average probability of correct prediction
is computed. The average probability of correct prediction is computed as
∑∑
−
qi
qiqiq
PywQ
ˆ
1
,
where
qi
P
ˆ
is the estimated probability of individual q choosing airport i at the convergent values.
The results for the estimation sample are presented in the second main column of Table 5. The
adjusted likelihood ratio index and the average probability of correct prediction clearly favor the
PCMNL model (see the last two rows of the table). A formal statistical nested likelihood ratio
test between the convergent log-likelihood values of the two models indicates a value of 400.0,
which is larger than the corresponding chi-squared value with 8 degrees of freedom at any
reasonable level of significance.
The performance of the MNL and PCMNL models is also evaluated on a holdout
(validation) sample to verify that the results obtained from the estimation sample are not an
artifact of overfitting. Three hundred observations are set aside for validation such that the
shares in the validation sample are close to the actual market shares (this allows the direct
application of the estimated model results to the validation sample, without the need to adjust the
airport-specific constants). Two measures of fit are computed in the validation sample. The first
is the predictive adjusted likelihood ratio index, which is computed by calculating the predictive
log-likelihood function value at the parameter estimates obtained from estimation. The second is
the average probability of correct prediction, also computed at the parameter values obtained
from estimation. These disaggregate measures of fit are presented in the last two rows of the
third main column in Table 5. As can be observed, there is a drop in the adjusted likelihood ratio
index from the estimation sample for both the MNL and PCMNL models. But the PCMNL
model still provides a value that is higher than the MNL model. The average probability of
correct prediction in the validation sample also reflects this superior fit of the PCMNL model. In
summary, the PCMNL clearly outperforms the MNL model from a statistical standpoint.
29
Table 5. Measures of Fit in Estimation and Validation Sample
Estimation Sample Validation Sample
Summary Statistic
MNL PCMNL MNL PCMNL
Log-likelihood at zero
-1777.55
-1777.55
-329.58
-329.58
Log-likelihood at market shares
-1490.40
-1490.40
-275.69
-275.69
Log-likelihood at convergence (estimation) /
Predictive log-likelihood (validation)
-897.50 -697.04 -174.74 -151.15
Number of parameters
1
7
15
7
15
Number of observations
1618
1618
300
300
Adjusted likelihood ratio index (estimation) /
Predictive adjusted ratio index (validation)
0.393 0.522 0.340 0.397
Average probability of correct prediction 0.662 0.749 0.665 0.729
1
The number of parameters refers to the parameter on the exogenous variables; it does not include the alternative-
specific constants in the MNL model and the alternative-specific constants at the choice stage of the PCMNL model
30
Another more informal, but intuitive, way to compare the two models is to compute the
estimated distribution of consideration sets across resident air travelers in the Bay Area. This
can be computed as
∑
−
q
qq
cPwQ )(
ˆ
1
, where )(
ˆ
cP
q
is the predicted probability from the PCMNL
model of individual q having the consideration set c. The resulting distribution, providing the
percentage of individuals with each of the seven possible choice sets, is as follows: SFO only
(23.50%), SJC only (0.22%), OAK only (0.12%), SFO and SJC (13.46%), SJC and OAK
(0.07%), SFO and OAK (9.83%), and all airports (52.80%). These results indicate that about
half of all travelers do not choose from the universal choice set of all the three airports.
However, the MNL model assumes that all travelers choose from the universal choice set.
Another interesting observation is that about a quarter of all travelers consider only SFO. In
summary, these results again highlight the clout of SFO in the consideration perception map of
Bay Area air travelers.
31
CHAPTER 6. SUMMARY AND CONCLUSIONS
This report proposes the use of a probabilistic choice set multinomial logit model
(PCMNL) for airport choice analysis that generalizes the commonly used multinomial logit
(MNL) model. The PCMNL model takes the form of a random constraint-based approach to
choice formation in which an airport is excluded from the choice set if the consideration utility of
that airport is lower than a threshold utility level. The choice of airport from a given choice set is
based on the usual MNL structure. The properties of the PCMNL model are discussed, including
the presentation and interpretation of elasticity expressions.
The PCMNL model is applied to examine the airport choice of business travelers residing
in the San Francisco Bay Area. Several important conclusions may be drawn from the empirical
analysis. First, as found in earlier studies, access time to the airport and flight frequency are the
two primary determinants of airport choice. However, unlike earlier studies, this study indicates
variation in sensitivity to these two variables based on traveler demographics and trip
characteristics. Specifically, individuals traveling alone and women travelers are more sensitive
to access time, and individuals traveling alone are also more sensitive to flight frequency.
Further, women and high-income travelers are not very sensitive to flight frequency. In addition,
the results from the consideration stage of the PCMNL model indicate that access time and flight
frequency affect the consideration of an airport.
A second important conclusion of this study is that the access time parameter estimates of
the MNL model and the choice stage of the PCMNL model are quite different. This is because
the MNL model arbitrarily assumes that all airports are available to all individuals. A
comparison of the relative trade-off between access time and frequency from the two models
suggests the dominance of access time at the choice stage, particularly in the PCMNL model.
However, the PCMNL model also indicates that, in forming perceptions of the availability of
airports, flight frequency is the dominating factor. Interestingly, access time is less important to
women (relative to men) when forming the perception space of available airports, but is more
important to women when choosing an airport from the set of available airports. These results
have implications for the design of promotional marketing strategies. For instance, an airport
attempting to increase market share by improving access time to its terminals might consider
targeting informational campaigns within its traditional catchment area of travelers (i.e., areas in
32
close proximity to the airport) and by targeting women travelers (at airports, or by targeting
firms/occupations which are women-dominated). On the other hand, information campaigns
regarding frequency improvements are better positioned in areas that are not within the
traditional catchment area (i.e., in areas that are distant from the airport) and are likely to be
more productive if targeted toward weekend travelers. Clearly, only the PCMNL model is able
to offer such comprehensive insights into the effects of variables.
A third conclusion that may be drawn from this study is that the substantive elasticity
effects from the MNL and PCMNL models indicate that access time is the most important factor
in the choice of an airport. Also, in the San Francisco Bay Area market, San Francisco
International has tremendous clout, since it can easily compensate for service improvements at
other airports by making marginal improvements in its own service. Between the MNL and the
PCMNL model, the PCMNL model predicts a lower overall impact of access time, indicating
that the use of the MNL model overestimates the potential gain in airport market share due to an
improvement in access time to that airport. On the other hand, the PCMNL model predicts a
higher overall impact of flight frequency, suggesting an underestimation of the net gains from
improving frequency by the MNL model.
Lastly, the PCMNL model clearly outperforms the MNL model in statistical evaluation
of data fit in both an estimation sample and a validation sample.
In summary, the application of the PCMNL model to airport choice suggests that it is
important to model consideration sets of air travelers. Failure to recognize consideration effects
can lead to biased model parameters, misleading evaluations of the effects of policy actions, as
well as a considerably diminished data fit.
One future extension of this study would be to examine airport travel characteristics
using more recent data. The author originally planned to use an air traveler survey conducted by
the Metropolitan Transportation Commission in the fall of 2001, but surveying halted due to the
events of September 11, 2001. Once the newest survey is completed and released to the public,
and similar studies are conducted using the data, it would be interesting to compare results from
the current study with those done with post-September 11
th
data.
33
REFERENCES
Andrews, R.L., Srinivasan, T.C., 1995. Studying consideration effects in empirical choice
models using scanner panel data. Journal of Marketing Research, 32 (February), 30-41.
Ashford, N., Benchemam, M., 1987. Passengers’ choice of airport: an application of the
multinomial logit model. Transportation Research Record, 1147, 1-5.
Ben-Akiva, M., Boccara, B., 1995. Discrete choice models with latent choice sets. International
Journal of Research in Marketing, 12, 9-24.
Börsch-Supan, A., 1987. Econometric analysis of discrete choice. Lecture Notes in Economics
and Mathematical Systems, Springer-Verlag, Berlin, Germany, 34.
Bureau of Transportation Statistics, 1997. Airline On-time Flight Statistics.
Chiang, J., Chib, S., Narasimhan, C., 1999. Markov chain Monte Carlo and models of
consideration set and parameter heterogeneity. Journal of Econometrics, 89, 223-248.
Furuichi, M., Koppelman, F.S., 1994. An analysis of air travelers’ departure airport and
destination choice behavior. Transportation Research A, 28 (3), 187-195.
Gensch, D., 1987. A two-stage disaggregate attribute choice model. Marketing Science, 6 (3),
223-231.
Harvey, G., 1987. Airport choice in a multiple airport region. Transportation Research A, 21 (6),
439-449.
Innes, J.D., Doucet, D.H., 1990. Effects of access distance and level of service on airport choice.
Journal of Transportation Engineering, 116 (4), 507-516.
Manski, C., 1977. The structure of random utility models. Theory and Decision, 8, 229-254.
Manski, C., Lerman, S., 1977. The estimation of choice probabilities from choice-based samples.
Econometrica, 45 (8), 1977-1988.
Metropolitan Transportation Commission, 1995. Airline Passenger Survey. Oakland, CA.
Ndoh, N.N., Pitfield, D.E., Caves, R.E., 1990. Air transportation passenger route choice: a nested
multinomial logit analysis. Spatial Choices and Processes. M.M. Fischer, P. Nijkamp,
Y.Y. Papageorgiou (editors). Elsevier Science Publishers B.V., North Holland, 349-365.
Official Airline Guide, 1995. Market Analysis. OAG Worldwide, Oakbrook, IL.
Ozoka, A.I., Ashford, N., 1989. Application of disaggregate modeling in aviation systems
planning in Nigeria: a case study. Transportation Research Record, 1214, 10-20.
Pels, E., Nijkamp, P., Rietveld, P., 2001. Airport and airline choice in a multiple airport region:
an empirical analysis for the San Francisco Bay Area. Regional Studies, 35 (1), 1-9.
Pels, E., Nijkamp, P., Rietveld, P., 2003. Access to and competition between airports: a case
study for the San Francisco Bay Area. Transportation Research A, 37 (1), 71-83.
Roberts, J.H., Lattin, J.M., 1991. Development and testing of a model of consideration set
composition. Journal of Marketing Research, 28, 429-440.
Shocker, A.D., Ben-Akiva, M., Boccara, B., Nedungadi, P., 1991. Consideration set influences
on consumer decision-making and choice: issues, models, and suggestions. Marketing
Letters, 2 (3), 181-197.
Skinner, R.E., Jr., 1976. Airport choice: an empirical study. Transportation Engineering Journal,
102 (4), 871-883.
Swait, J., 1984. Probabilistic choice set formation in transportation demand models. Unpublished
Ph.D. Thesis, Department of Civil Engineering, MIT, Cambridge, MA.
Swait, J., 2001. Choice set generation within the generalized extreme value family of discrete
choice models. Transportation Research B, 35 (7), 643-666.
34
Swait, J., Ben-Akiva, M., 1987a. Incorporating random constraints in discrete models of choice
set generation. Transportation Research B, 21 (2), 91-102.
Swait, J., Ben-Akiva, M., 1987b. Empirical test of a constrained choice discrete model: mode
choice in S
āo Paulo, Brazil. Transportation Research B, 21 (2), 103-115.
Thompson, A., Caves, R., 1993. The projected market share for a new small airport in the north
of England. Regional Studies, 27 (2), 137-147.
Williams, H., Ortuzar, J., 1982. Behavioural theories of dispersion and the mis-specification of
travel demand models. Transportation Research B, 16 (3), 167-219.
Windle, R., Dresner, M., 1995. Airport choice in multiple-airport regions. Journal of
Transportation Engineering, 121 (4), 332-337.
35
APPENDIX A. Sample of Choices Involved in Air Travel
Mode of transport from destination
airport to final destination?
Whether to check baggage?
Mode of transport to airport?
From where to leave for the airport?
When to leave for the airport?
How to purchase tickets?
Desired departure time?
Desired arrival time?
Which price class?
Which airline?
Which origin airport?
Which destination airport?
By what mode? (assume air is chosen)
Duration of Stay?
When?
Where?
Whether to travel?
36
APPENDIX B. Literature Review Table
Study Author Year Dimensions Model
Structure
Empirical Context
Airport Choice: An Empirical
Study
Robert E. Skinner, Jr. 1976 Airport Choice MNL
Model
Baltimore-Washington region with three
airport alternatives: Reagan, Dulles, Baltimore
Airport Choice in a Multiple
Airport Region
Greig Harvey 1987 Airport choice MNL
Model
San Francisco Bay Area with three airport
alternatives: San Francisco, San Jose, Oakland
Passengers’ Choice of Airport:
An Application of the
Multinomial Logit Model
Norman Ashford,
Messaoud
Benchemam
1987 Airport Choice MNL
Model
Central England with five airport
alternatives: Manchester, Birmingham, East
Midlands, Luton, Heathrow
Air Transportation Passenger
Route Choice: A Nested
Multinomial Logit Analysis
Ngoe N. Ndoh,
David E. Pitfield,
Robert E. Caves
1990 Route (direct vs. connecting
flights), hub airport, and
departure airport choice
NMNL
Model
England with four airport alternatives: East
Midlands, Birmingham, Manchester, Liverpool
Effects of Access Distance and
Level of Service on Airport
Choice
J. David Innes,
Donald H. Doucet
1990 Airport Choice Binary
Logit
Model
New Brunswick, Canada with three airport
alternatives: Charlo, Chatham, St. Leonard
Application of Disaggregate
Modeling in Aviation Systems
Planning in Nigeria: A Case
Study
Angus Ifeanyi
Ozoka, Norman
Ashford
1989 Airport Choice MNL
Model
Nigeria with two airport alternatives: Enugu,
Benn
The Projected Market Share for
a New Small Airport in the
North of England
Amanda Thompson
and Robert Caves
1993 Airport choice MNL
Model
Northern England with three airport
alternatives: East Midlands, Manchester,
Birmingham
An Analysis of Air Travelers’
Departure Airport and
Destination Choice Behavior
Masahiko Furuichi,
Frank Koppelman
1994 Departure airport and
destination choice
NMNL
Model
Japan with four airport alternatives: Narita,
Osaka, Nagoya, Fukuoka
Airport Choice in Multiple-
Airport Regions
Robert Windle,
Martin Dresner
1995 Airport Choice MNL
Model
Baltimore-Washington, D.C. region with three
airport alternatives: Reagan, Dulles, Baltimore
Airport and Airline Competition
for Passengers Departing from a
Large Metropolitan Area
Eric Pels, Peter
Nijkamp, Piet
Rietveld
2000 Airport and airline choice NMNL
Model
San Francisco Bay Area with airport
alternatives: San Francisco, San Jose, Oakland,
Airport and Airline Choice in
the a Multiple Airport Region:
An Empirical Analysis for the
San Francisco Bay Area
Eric Pels, Peter
Nijkamp, Piet
Rietveld
2001 Airport and airline choice NMNL
Model
San Francisco Bay Area with four airport
alternatives: San Francisco, San Jose, Oakland,
Sonoma County
Airport and access mode choice
in the Bay Area
Eric Pels, Peter
Nijkamp, Piet
Rietveld
2002 Airport and airport access
mode choice
NMNL
Model
Three airport choices: San Francisco, San Jose,
Oakland
37
APPENDIX B. Literature Review Table (continued)
Author(s) continued Market Segment
Examined
Variables Considered Final Variables in
Model(s)
Important Results
Robert E. Skinner, Jr.,
1976
Business and
nonbusiness travelers
Air carrier level of service
measures, ground
accessibility measures
Weekday flight
frequency, airport
access utility
Improvements in airport access are the most
effective means of capturing more
passengers
Greig Harvey, 1987
Resident business and
resident nonbusiness
travelers
Airport access time, relative
and direct flight frequency
Airport access time,
flight frequency
Airport access time and flight frequency
provide good approximation of airport
choice in the Bay Area. Beyond a threshold
level, additional direct flights to a
destination do not make an airport more
attractive.
Norman Ashford,
Messaoud Benchemam,
1987
Domestic, international
business, international
leisure, international
inclusive tours travelers
Travel time to airport, number
of flights per day, air fare
Travel time and
flight frequency for
business and
inclusive tours, all
three variables for
remaining market
segments
Business travelers most sensitive to airport
access time, while leisure travelers are most
sensitive to air fare and airport access time.
Ngoe N. Ndoh, David E.
Pitfield, Robert E. Caves,
1990
Business travelers Airport access time, average
journey time, average
connection time to hub,
number of seats
Access time, journey
time, connection
time to hub, number
of seats, flight
frequency
Business travelers value access time the
most over any other variable
J. David Innes, Donald
H. Doucet, 1990
-----
Ticket type, length of stay,
who paid for the ticket, trip
purpose, aircraft type, flying
time, (direct vs. nonstop)
Same as those
considered
Type of aircraft plays significant role in
airport choice (air travelers are willing to
travel far for access to jet service).
Passengers prefer direct flights versus
connecting, and shorter flight routes.
Angus Ifeanyi Ozoka,
Norman Ashford, 1989
-----
Airport access time, flight
frequency, air fare
Airport access travel
time
Improving ground access to airport is the
best (and possibly only) means of
increasing an airport’s market share in
Nigeria. The catchment area concept does
not apply; airports compete.
Amanda Thompson and
Robert Caves, 1992
Business and
nonbusiness travelers
Airport access time, flight
frequency, air fare, number of
seats
Airport access time,
flight frequency, air
fare
Those departing from origins closer to the
airport are more sensitive to access time
than those living further away.
38
Masahiko Furuichi,
Frank Koppelman, 1994
Business and pleasure
travelers
Airport access travel time and
cost, line-haul travel time and
cost, relative flight frequency
Airport access travel
time and cost, line-
haul travel time and
cost, relative flight
frequency
Access travel cost valued more highly than
line-haul travel cost. Both business and
pleasure travelers have very high values of
access and line-haul time, as well as flight
frequency.
Robert Windle, Martin
Dresner, 1995
Resident business,
resident nonbusiness,
nonresident business,
nonresident
nonbusiness
Airport access time, weekly
flight frequency, airport
experience
Airport access time,
weekly flight
frequency, airport
experience
Airport access time and flight frequency
significant. Airport experience comes out
to be significant, but could be proxy for
omitted variables.
Eric Pels, Peter Nijkamp,
Piet Rietveld, 2001
Resident business and
resident leisure
travelers
Flight frequency, airport
access time, air fare
Flight frequency,
airport access time
Passengers first choose departure airport,
then choose airline is statistically favorable
to the opposite. Little difference between
business and leisure travelers.
Eric Pels, Peter Nijkamp,
Piet Rietveld, 2002
Resident business,
resident leisure,
summer and fall
Airport distance and access
time, average fare, daily flight
frequency
Air fare, flight
frequency, access
time
Access time most significant variable in
airport choice
39
APPENDIX C. Questions in the MTC Air Passenger Survey
(questions that were asked at all four airports, in both summer and fall)
• Residence status (Bay Area
10
resident or visitor)
• Final airport destination, including all flights
• Main trip purpose
• Number of people in the party
• Number of vehicles the party used to get to the airport
• Number of people in the vehicle in which the respondent traveled
• Number of pieces of luggage the party checked
• Mode of transportation used to get to the airport
• Among those who took a private car, how they would have traveled if the car had not
been available
• Among those who took a rental car, the company they rented it from
• Among those who took transit, how they got to the transit stop or station
• Mode of transportation used to get from the airport to the Bay Area destination the
last time the respondent flew into the airport
• Origin of departure for the airport
• Type of origin the respondent departed from
• Number of people who came into the terminal to see the respondent off
1
The Bay Area was defined as the nine greater San Francisco Bay Area counties: Alameda, Contra Costa, Marin,
Napa, San Francisco, San Mateo, Santa Clara, Solano, and Sonoma.
40
• Length of time prior to flight departure time that the respondent arrived at the airport
• Trip length (in nights away from home)
• Extent to which the respondent could have used another airport
• Individual who decided to use the departure airport
• Number of times the respondent had flown out of each of six area airports in the
twelve months preceding the survey
• Zip Code of the respondent’s residence
• Number of people in the respondent’s household
• Respondent’s household income before taxes in 1994
• Respondent’s gender (by observation)
• Date of interview, airline, flight number, departure time, and interview time (by
observation)
41
APPENDIX D. Data Screening Process
Full Sample = 21,124 cases
Focus only on residents
Sample = 9,510 cases
Focus on three airports: SFO, SJC, OAK (delete Sonoma County airport)
Sample = 9,476
Focus on top 30 domestic destinations for resident travelers
Sample = 7,336
Try to fill in observations missing critical elements
(date, flight departure time, number of connections can all be entered based on flight number and other
variables)
Remove observations missing critical items that cannot be filled
Add flight frequency variable, BTS on-time statistics, access times and costs to airport
Focus only on people who said they had the choice of flying from another airport
Sample = 3,795 surveys
Business trip Non-Business trip
Sample = 1,918 surveys Sample = 1,877 surveys
Estimation Validation
Sample = 1,618 surveys Sample = 300 surveys
Create weight variable Create weight variable
Base Market Share Models to check if weightings are accurate
42
APPENDIX E. Top Thirty Domestic Destinations
Estimation Sample Validation Sample
City
Airport
Code
Frequency % Frequency %
1 LOS ANGELES, CA LAX 248 15% 49 16%
2 SAN DIEGO, CA SAN 148 9% 22 7%
3 BURBANK, CA BUR 135 8% 22 7%
4 ORANGE COUNTY, CA SNA 134 8% 11 4%
5 SEATTLE, WA SEA 84 5% 18 6%
6 PORTLAND, OR PDX 78 5% 21 7%
7 LAS VEGAS, NV LAS 70 4% 11 4%
8 ONTARIO, CA ONT 70 4% 7 2%
9 DALLAS, FT. WORTH,TX DFW 60 4% 7 2%
10 PHOENIX, AZ PHX 59 4% 16 5%
11 DENVER, CO DEN 54 3% 11 4%
12 CHICAGO, IL(O'HARE) ORD 51 3% 9 3%
13 RENO, NV RNO 50 3% 7 2%
14 AUSTIN, TX AUS 48 3% 3 1%
15 SALT LAKE CITY, UT SLC 45 3% 13 4%
16 BOSTON, MA BOS 39 2% 15 5%
17 ATLANTA, GA ATL 31 2% 8 3%
18 NEW YORK, NY(JFK) JFK 30 2% 7 2%
19 WASHINGTON,DC(DULLES) IAD 27 2% 5 2%
20 ALBUQUERQUE, NM ABQ 25 2% 3 1%
21 NEWARK-NEW YORK, NJ EWR 23 1% 8 3%
22 HOUSTON, TX(INTERCON) IAH 23 1% 3 1%
23 BOISE, ID BOI 19 1% 3 1%
24 MINNEANAPOLIS/ST.PAUL MSP 18 1% 4 1%
25 COLORADO SPRINGS, CO COS 15 1% 1 0%
26 SPOKANE, WA GEG 9 1% 1 0%
27 HONOLULU, HI HNL 9 1% 4 1%
28 TUCSON, AZ TUS 8 0% 2 1%
29 ORLANDO, FL ORL 6 0% 8 3%
30 KAHULULI, HI OGG 2 0% 1 0%
Total 1618 100% 300 100%
43
APPENDIX F. Variables Used to Come to a Preferred Specification
Variable Description
Alt. Specific Constants
SFO Constant specific to SFO
SJC Constant specific to SJC
OAK Constant specific to OAK
Flight Frequency Daily flight frequency
Distance Distance to airport from origin point
Access time Access time to airport from point of origin
Access cost Access cost to airport from point of origin
On-time Statistics Percentage of late flights, specific to each O-D pair
Airport Loyalty
Proportion of flights from each airport
over a 12 month period
11
Weight
Weighting variable representing 1995 airport market shares
Income Total 1994 household income
Market Segmentation Variables
Peak Passenger is traveling during a peak period
12
Alone Passenger is traveling alone
Short trip Trip lasting 0 or 1 night
High income Income > $150,000 per year
Car Drove either a private or rental car to airport
Weekday Flight is on a weekday
Summer Flight is in summer
Nonstop Flight is a nonstop flight
Female Passenger is female
Work Passenger left straight from work for the airport
11
Loyalty Factor =
yearafromBayAreallflights
yearairportflights
n
1,
1,,#
∑
12
Passengers traveling during peak periods are those whose scheduled flight departure times are either 6AM -
9:59AM or 4PM - 8:59PM
44
