Discrete Choice Modelling and Air Travel Demand

Theory and Applications

Laurie A. Garrow

• Publisher: Routledge
• Year: 2016
• ISBN:
  9781317149705
• Edition: 1
• Topic: Technology & Engineering
• Subtopic: Transportation & Navigation
• Index: Technology & Engineering

In recent years, airline practitioners and academics have started to explore new ways to model airline passenger demand using discrete choice methods. This book provides an introduction to discrete choice models and uses extensive examples to illustrate how these models have been used in the airline industry. These examples span network planning, revenue management, and pricing applications. Numerous examples of fundamental logit modeling concepts are covered in the text, including probability calculations, value of time calculations, elasticity calculations, nested and non-nested likelihood ratio tests, etc. The core chapters of the book are written at a level appropriate for airline practitioners and graduate students with operations research or travel demand modeling backgrounds. Given the majority of discrete choice modeling advancements in transportation evolved from urban travel demand studies, the introduction first orients readers from different backgrounds by highlighting major distinctions between aviation and urban travel demand studies. This is followed by an in-depth treatment of two of the most common discrete choice models, namely the multinomial and nested logit models. More advanced discrete choice models are covered, including mixed logit models and generalized extreme value models that belong to the generalized nested logit class and/or the network generalized extreme value class. An emphasis is placed on highlighting open research questions associated with these models that will be of particular interest to operations research students. Practical modeling issues related to data and estimation software are also addressed, and an extensive modeling exercise focused on the interpretation and application of statistical tests used to guide the selection of a preferred model specification is included; the modeling exercise uses itinerary choice data from a major airline. The text concludes with a discussion of on-going customer modeling research in aviation. Discrete Choice Modelling and Air Travel Demand is enriched by a comprehensive set of technical appendices that will be of particular interest to advanced students of discrete choice modeling theory. The appendices also include detailed proofs of the multinomial and nested logit models and derivations of measures used to represent competition among alternatives, namely correlation, direct-elasticities, and cross-elasticities.

Chapter 1
Introduction

Introduction and Background Context

In Daniel McFadden’s acceptance speech of the Nobel Prize in Economics, he describes how in 1972 he used a multinomial logit model based on approximately 600 responses from individual commuters in the San Francisco Bay Area to forecast ridership for a new BART line (McFadden 2001). This study, typically considered the first application of a discrete choice model in transportation, provided a strong foundation and motivation for urban travel demand researchers to transition from modeling demand using aggregate data to modeling demand as the collection of individuals’ choices. These choices varied by socio-demographic and socioeconomic characteristics, as well as by attributes of the alternatives available to the individual.

At the same time that McFadden and other researchers were investigating forecasting benefits associated with modeling individual choice behavior to support transit investment decisions, the U.S. airline industry was predicting demand for air travel using Quality of Service (QSI) indices. QSI indices were developed in 1957 and predicted how demand would shift among carriers as a function of flight frequency, level of service (e.g., nonstop, single-connection, double-connection) and equipment type (Civil Aeronautics Board 1970). At the time, the airline industry was regulated, fares and service levels were set by the government, and load factors were about 50 percent (e.g., see Ben-Yosef 2005). Competition was based primarily on marketing promotion and image.

The airline industry changed dramatically in 1978 when it became deregulated and airlines could decide where and when to fly, as well as how much to charge passengers (Airline Deregulation Act 1978). Operations research analysts played a critical role after deregulation, helping to design algorithms and decision-support systems to optimize where and when to fly, subject to minimizing costs associated with assigning pilots and flight attendant crews to each flight while ensuring each plane visited a maintenance station in time for required checks and service. A second milestone event happened in 1985, when American Airlines implemented a revenue management system that offered a limited set of substantially discounted fares with advance purchase restrictions as a way to compete with low fares offered by People’s Express Airlines; the strategy worked, and People’s Express went out of business shortly thereafter (e.g., see Ben-Yosef 2005). A role for operations research had emerged in the revenue management area, with the primary objective of maximizing revenue (or profit) under uncertain demand forecasts, passenger cancellations, and no shows.

The “birth” of operations research in a deregulated airline industry occurred in an era in which computational power was much more limited than it is today. A major airline faced with optimizing schedules that involved coordinating arrivals and departures for thousands of daily take-offs and landings, assigning tens of thousands of pilots and flight attendants to all of these flights (while ensuring all work rules were adhered to), and keeping track of millions of monthly booking transactions, was clearly facing a different problem context than Daniel McFadden and other travel demand modelers. The latter were making demand predictions to help support investment decisions and evaluation of transportation policies for major metropolitan areas. In this context, the use of discrete choice models to help rank different alternatives and assess short-term and long-term forecast variation across different scenarios was of primary importance to decision-makers.

However, from an airline perspective, it would have been computationally impractical to model the choice of every individual passenger (which would require keeping track of all alternatives considered by passengers). Instead, in the U.S. it was (and still is) common to model market-level itinerary share demand forecasts using ticket information compiled by the U.S. Department of Transportation (Bureau of Transportation Statistics 2009; Data Base Products Inc. 2008) and to use time-series and/or simplistic probability models based on product-level booking or flight-level data to forecast demand for flights, passenger cancellation rates, passenger no show rates, etc.

More than thirty years after deregulation, the airline industry is faced with intense competition and ever-increasing pressures to control costs and generate more revenues. Multiple factors have contributed to the current state of the industry, including the increased use of the Internet as a major distribution channel and the increased market penetration of low cost carriers. It is clear that the Internet has transformed the travel industry. For example, in 2007, approximately 55 million (or one in four) U.S. adults traveled by commercial air and were Internet users (PhoCusWright 2008). As of 2004, more than half of all leisure travel purchases were made online (Aaron 2007). In 2006, more than 365 million U.S. households spent a total of $74.4 billion booking leisure travel online (Harteveldt Johnson Stromberg and Tesch 2006).

The market penetration of low cost carriers has also steadily and dramatically grown since the early 1990’s. For example, in 2004, approximately 25 percent of all passengers in the U.S. flew on low cost carriers, and 11 percent of all passengers in Europe flew on low cost carriers (IBM Consulting Services 2004). Importantly, the majority of low cost carriers in the U.S. use one-way pricing, which results in separate price quotes for the departing and returning portions of a trip. One-way pricing effectively eliminates the ability to segment business and leisure travelers based on a Saturday night stay requirement (i.e., business travelers are less likely to have a trip that involves a Saturday night stay). Combine the use of one-way pricing with the fact that the Internet has increased the transparency of prices for consumers and the result is that today, approximately 60 percent of online leisure travelers purchase the lowest fare they can find (Harteveldt Wilson and Johnson 2004; PhoCusWright 2004).

Within the operations research community, these and other factors have led to an increasing interest in using discrete choice models to model demand as the collection of individuals’ decisions, thereby more accurately capturing how individuals are making decisions and trade-offs among carriers, price, level of service, time of day, and other factors. To date, much of the research in using discrete choice models for aviation applications has focused in areas where it has been relatively straightforward to identify the alternatives that individuals consider during the choice process (e.g., airlines have itinerary-generation algorithms that build the set of itineraries or paths between origin-destination pairs). In addition, this research has focused on areas in which it would be relatively easy for airlines to replace an existing module (e.g., a no show forecast) that is part of a much larger decision-support system (e.g., a revenue management system). Itinerary share predictions, customer no show behavior, customer cancellation behavior, and recapture rate modeling all belong to this stream of research (e.g., see Coldren and Koppelman 2005a, 2005b; Coldren Koppelman Kasturirangan and Mukherjee 2003; Garrow and Koppelman 2004a, 2004b; Iliescu Garrow and Parker 2008; Koppelman Coldren and Parker 2008; Ratliff 2006; Ratliff Venkateshwara Narayan and Yellepeddi 2008).

More recently, researchers have also begun to investigate how discrete choice models and passenger-level data can be integrated with optimization models at a systems level. Advancements in computing power combined with the ability to track individual consumers through the booking process have spawned a new era of revenue management (RM), commonly referred to as “choice-based” RM. Conceptually, choice-based RM methods use data that effectively track individuals’ purchase decisions, as well as the menus of choices they viewed prior to purchase. That is, in contrast to traditional booking data, on-line shopping data provide a detailed snapshot of the products available for sale at the time an individual was searching for fares, as well as information on whether the search resulted in a purchase (or booking). These data effectively enable firms to replace RM demand models based on probability and time-series models with models grounded in discrete choice theory. To date, several theoretical papers on choice-based RM techniques have appeared in the research community and a few empirical studies based on a limited number of markets and/or departure dates have also been reported (e.g., see Besbes and Zeevi 2006; Bodea Ferguson and Garrow 2009; Bront Mendez-Diaz and Vulcano 2007; Gallego and Sahin 2006; Hu and Gallego 2007; Talluri and van Ryzin 2004; van Ryzin and Liu 2004; van Ryzin and Vulcano 2008a, 2008b; Vulcano van Ryzin and Chaar 2008; Zhang and Cooper 2005).

To summarize, it is clear that the momentum for using discrete choice models to forecast airline demand as the collection of individuals’ choices is building, and most importantly, this momentum is building both in the travel demand modeling/discrete choice modeling community as well as in the operations research community.

Primary Objectives of the Text

Although the interest in using discrete choice models for aviation applications is building, there has been limited collaboration between discrete choice modelers and optimization and operations researchers. Part of the challenge is that many operations research departments have provided students with a limited exposure to discrete choice models. This is due in part to the fact that the primary affiliation of most discrete choice modeling experts is not with operations research departments, but rather with transportation engineering, marketing, and/or economics departments. The distinct evolution of the discrete choice modeling and operations research fields has resulted in researchers from these fields having different perspectives, research priorities, and publication outlets.

One of the primary objectives of this text is to help bridge the gap between the discrete choice modeling and operations research communities by providing a comprehensive, introductory-level overview of discrete choice models. This overview synthesizes major developments in the discrete choice modeling field that are relevant to the aviation industry and the challenges this industry is currently facing. An emphasis has been placed on discussing the properties of discrete choice models using terminology that is accessible to both the discrete choice modeling and operations research communities, and complementing these discussions with numerous examples. The discrete choice modeling topics covered in the text (that represent only a small fraction of work that has been developed since the early 1970s), provide a fundamental base of knowledge that analysts will need in order to successfully estimate, interpret, and apply discrete choice models in practice. Consequently, it is envisioned that this text will be useful to aviation practitioners, researchers and graduate students in operations research departments, and researchers and graduate students in travel demand modeling.

Important Distinctions Between Aviation and Urban Travel Demand Studies

Given the different backgrounds and perspectives of aviation operations research analysts and urban travel demand analysts, it is helpful to highlight some of the key distinctions between these two areas.

Objectives of Aviation and Urban Transportation Studies

The overall objectives driving demand forecasting studies conducted for aviation firms and studies conducted for government agencies evaluating transportation alternatives in urban areas tend to be quite distinct. Deregulated airlines, such as those in the U.S. that are private firms and are not owned by governments, are generally focused on maximizing net revenue through attracting new customers and retaining current customers while ensuring safe and efficient operations. Many of the problems investigated by operations research analysts reflect this strong focus on maintaining safe and efficient operations throughout the airline’s network (or system). These problems include building robust network schedules and assigning pilots and flight attendants to aircraft in ways that result in fewer aircraft delays and cancellations and fewer passenger misconnections; assigning aircraft to specific airport gates to ensure transfer passengers have sufficient time to connect to their next flight while considering secondary objectives, such as minimizing the average distance that premium passengers need to walk between a loyalty lounge and the departing gate; scheduling multiple flights into a hub to achieve one or more objectives, such as maximizing passenger connection possibilities, minimizing passenger connection times, and/or flattening peak airport staffing requirements; developing efficient processes to screen baggage and minimize the number of bags that are lost or delayed; creating rules that minimize average boarding time for different aircraft types; developing processes that help airlines quickly recover from irregular operations; overbooking flights to maximize revenue while minimizing the number of voluntary and involuntary denied passengers, etc.

Government agencies, in contrast to airlines, are generally focused on predicting demand for existing and proposed transportation alternatives. A broad range of alternatives may be considered and include infrastructure improvements, operational improvements, new tax fees, credits or other policy instruments, etc. Thus, the primary focus of urban transportation studies is centered on supporting policy analysis, which includes gaining a richer understanding of how individuals, households, employers and other institutions will react to different alternatives. Urban travel demand analyses are also often conducted within a systems-level framework (i.e., examined within the entire urban area), in part to ensure equitable allocation of resources and services across different socio-economic and socio-demographic groups.

Data Characteristics of Aviation and Urban Travel Demand Studies

Given the different objectives of aviation firms and government agencies, it is not surprising that the data used for analysis also differ. Within aviation, the strong operational focus within a relatively large system has resulted in decision-support models based almost exclusively on revealed preference data that contain limited customer information. Revealed preference data capture actual passenger choices under current and prior market conditions. The airline industry is characterized by flexible capacity which results in a large number of observations that tend to vary “naturally” or “randomly” within a market or across different markets. For example, in itinerary share models, frequent schedule changes create “natural” variation in the itineraries available to customers; that is, over the course of a year (or even from month to month), individuals are faced with alternatives that vary by level of service, departure and/or arrival times, connection times, operating carriers, prices, etc. In turn, given the dynamic nature of the airline industry and the need for carriers to identify and respond quickly to changes in competitive conditions, it is highly desirable to design decision-support models that rely heavily on recently observed revealed preference data.

In addition, due to the large number of flights major carriers manage, any customer information stored in databases tends to be limited to that needed to support operations. For example, from an operations perspective, it is important for gate agents to know how many individuals on an arriving flight need wheelchairs; however, knowing the individual’s age, gender, and household income level is irrelevant to the ability of the gate agent to make sure a wheelchair is available for the customer, and is thus not typically collected as part of the booking process. Similarly, although algorithms have been developed to reaccomodate passengers automatically to different flights when their original flight experiences a long delay or cancellation, the prioritization of customers is typically based on prior and current travel information. Archival travel information may include the customer’s current status in the airline’s frequent flyer program and/or the customer’s “value” to the airline that considers both the number of trips the customer has purchased on the carrier as well as how much the customer paid for these trips. Current travel information may include the amount the customer paid for the trip, whether the trip is in a market that has a low flight frequency (resulting in fewer reaccommodation opportunities), and whether the cost of reaccommodating the passenger on a different carrier is high (as in the case for an international itinerary).

In contrast to airline applications with an operations focus, urban travel demand studies rely heavily on socio-economic and socio-demographic information, such as an individual’s age, gender, ethnicity, employment status, marital status, number and ages of children in the household, residence ownership status and type (owned or rented; single family home, multi-family residence, etc.), household income, etc. These and other variables (such as the make, model, and age of each automobile owned by the household) are inputs to the travel demand forecasts for an urban area. Conceptually, these models create a simulated population that represents characteristics of the existing population in an urban area. Different transportation alternatives and/or combinations of different transportation alternatives are evaluated by testing how different segments of the population respond, assessing system-level benefits (such as reductions in emissions due to shifting trips from automobile to transit or due to modernizing vehicle fleets over time), and identifying any impacts that are disproportionately allocated across different socio-economic groups.

Urban travel demand studies use a wide range of revealed preference, stated preference data, and combinations of revealed and stated preference data. Revealed preference data sources include observed boarding counts on buses and other modes of transportation, observed screen-line counts (or the number of vehicles passing by a certain “screen-line” in a specified time period), travel survey diaries that ask individuals to record every trip made by members of the household over a short period of time (typically two days), intercept surveys that interview current transit users to collect information about their current trip, etc. From a demand forecasting perspective, the socio-demographic and socio-economic variables that are inputs to urban travel demand models are available, often at a detailed census tract or census block level, from government agencies. Moreover, for many major infrastructure projects (such as a proposed transit project in the U.S. that requests federal funding support), it is expected that demand forecasts will be based on “recent” customer surveys.

Whereas revealed preference data reflect the actual choices made by individuals under current or previous market conditions, stated preference data are collected via surveys that ask individuals to make hypothetical choices by making tradeoffs among the attributes of the choice set (such as time, cost, and reliability measures) determined by the analyst. Stated preference data are particularly useful when investigating customer response to new products or transportation alternatives, or when existing and past market conditions do not exhibit sufficient “natural variation” to allow the analyst to estimate how individuals are making tradeoffs (because the number of distinct trade-off combinations is limited). For example, time-of-day congestion pricing is a relatively new concept that has been implemented in different forms throughout the world. Stated preference surveys designed to investigate how commuters and shippers would potentially change their behavior under different congestion pricing alternatives in a major metropolitan area would be valuable for assessing likely outcomes associated with implementing a similar policy in a new area.

Whereas many aviation studies with an operational focus tend to rely heavily on revealed preference data, stated preference data are also used within the airline industry, albeit primarily in marketing departments where new product designs are of primary interest. For example, Resource Systems Group, Inc., a firm located in Vermont, has been conducting an annual survey of air travelers since 2000. This annual stated preference survey has been supported by a wide variety of airlines and government agencies. Consistent with the use of stated preference data seen in the context of urban travel demand studies, these stated preference surveys have supported a range of new product development studies for airlines (e.g., cabin service amenities, unbundling product strategies, passenger preferences for connection times, etc.) Government agencies have also used this panel to...

Table of contents

• Cover Page
• Dedication
• Title Page
• Copyright Page
• Contents
• List of Figures
• List of Tables
• List of Abbreviations
• List of Contributors
• Acknowledgements
• Preface
• 1 Introduction
• 2 Binary Logit and Multinomial Logit Models
• 3 Nested Logit Model
• 4 Structured Extensions of MNL and NL Discrete Choice Models
• 5 Network GEV Models
• 6 Mixed Logit
• 7 MNL, NL, and OGEV Models of Itinerary Choice
• 8 Conclusions and Directions for Future Research
• References
• Index
• Author Index