Airport Ground Access Mode Choice Models (2008)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

This chapter provides a short introduction to the motivation and methodology for modeling airport ground access mode choice. It is primarily intended to give airport management and planning staff some background on how these models can be used in the planning and decision-making process, and provide an overview of the technical issues involved in developing and applying these models so that they can inter- act with more technical specialists in an informed way and properly supervise contracts for the development and use of such models. It is also hoped that it will be helpful to airport planning consultants and other planning specialists, includ- ing transportation planners working for MPOs and other planning agencies, who do not have particular expertise in airport ground access modeling, but become involved in air- port ground access modeling issues. As will be apparent from the detailed discussion in this report, airport ground access mode choice modeling is a highly specialized field, with many complex aspects that make this a particularly challenging problem. This has two implications. The first is that the development of these models typically requires the use of specialists with prior experience in model- ing airport ground access mode choice. The second is that it is therefore helpful if other planners and managers involved in the larger planning process have some idea of what is involved in developing and using these models, so that they have appropriate expectations of both the resources required and what it is reasonable to be able to accomplish with a given level of resources. MOTIVATION FOR MODELING AIRPORT GROUND ACCESS MODE CHOICE Airport travelers make use of a wide variety of different modes for their ground access trips to and from the airport, including private vehicles, rental vehicles, taxis, and multi- ple private and public transportation services. Large airports in particular are served by a large number of different ground transportation modes and services. Planning landside and air- port ground access facilities, as well as accounting for the environmental impacts of airport ground transportation activities, requires the ability to predict how airport users will change their ground access and egress decisions in response to changes in the array of options that they face. It would be impractical to imagine that the proportion of airport users choosing a particular mode will remain constant when the factors influencing their choices are continuously changing. 14 In some cases, an airport may want to understand the consequence of decisions that are largely outside its con- trol, such as changes in the price or service pattern of ground transportation services operated by other entities or changing congestion levels on the regional transportation system. In other cases, it may want to understand how deci- sions that it is considering will affect the airport ground transportation system or be affected by it. Examples of such decisions could include proposed changes in parking rates for airport-operated parking lots, the introduction of a shut- tle bus link to a nearby rail station, or the construction of a consolidated rental car facility some distance from the air- port terminal. Each of these decisions is likely to result in changes in the relative attractiveness of the different access and egress modes (including the different on-airport and off-airport parking facilities) and resulting shifts in the pro- portion of airport travelers using each mode, as well as the effect of this on the revenue generated by each facility or service. Finally, the requirement to predict the environ- mental impacts of decisions that change or affect the airport ground transportation system, particularly the need to pre- dict air quality impacts, demands an ability to assess how those decisions will affect the use of the various modes (Gosling 2005). This can be a particularly critical issue for airports in regions that are not in attainment of the National Ambient Air Quality Standards, where it may be necessary to be able to demonstrate that a proposed project will not increase total emissions or that appropriate mitigation mea- sures will offset any increase in emissions from the project. Indeed, it may be necessary to be able to demonstrate a reduction in total emissions. Prediction of how changes in the airport ground trans- portation system will influence the mode choice decisions of airport travelers is complicated because those decisions depend not only on the price and level of service of the alter- native modes but also on the characteristics of the individual travelers. In the case of air passengers, these characteristics include their trip purpose, whether they are residents of the region or visitors to it, how long residents of the region will be away from home on their trip, and whether visitors to the region will need a rental car for local travel during their visit. The distribution of these characteristics across the population of airport travelers not only varies seasonally but also in response to external influences, such as currency exchange rates and the state of the regional economy, and changes in the air services offered at the airport. CHAPTER TWO AIRPORT GROUND ACCESS MODE CHOICE MODELING PROCESS

15 Given the many different factors influencing the propor- tion of airport travelers using each mode, it is unrealistic to expect planners and decision makers to be able to make quantitative estimates of the effect on mode use of any given change in the system without the use of formal analytical tools that model how airport users respond to changes in the available airport ground transportation services. Thus, air- port ground access mode choice models provide the basic input to other analysis tools that are used to support airport landside and ground access planning and decision making, such as traffic flow models, simulation models, or financial planning tools. Representative Applications To illustrate the type of decision that can benefit from the availability of an airport access mode choice model, this sec- tion describes three representative situations where such a model could be applied. Airport A is experiencing steady growth in air traffic and is planning to construct a new passenger terminal alongside the existing terminal. The existing surface parking lot in front of the current terminal is already reaching capacity at peak periods and there is no space between the terminal roadways and the site of the new terminal to significantly expand the parking lot. The terminal redevelopment plan envisages con- structing a multi-level parking structure on the site of the existing surface lot. To pay for the parking structure, the air- port authority is considering raising the daily parking rate once the new structure is in operation. However, it is con- cerned that this will reduce the parking demand for the new structure and divert air passengers to the surface long-term parking lot some distance away, privately operated off- airport parking facilities, or even convince some air passen- gers who might otherwise park at the airport to be dropped off and picked up instead, or make use of other modes such as a shared-ride van. It recognizes that the extent of any such diversion will depend on the new rate charged for the park- ing structure, any change in the rate for the use of the long- term lot, and whether the off-airport parking lot operators also adjust their rates. To study the effect of different rate structures on the demand for parking at both the new parking structure and the long-term lot, and the implications for the financial feasibility of the new parking structure, the airport retains a consultant to develop an airport access mode choice model that can be used to analyze the impact of different rate structures on parking demand and identify the optimal pric- ing strategy and resulting parking demand in both facilities. Airport B is in a non-attainment region for several ambi- ent air quality standards and is under pressure from the local Air Resources Board to reduce the emissions from airport ground access and egress travel. The airport authority rec- ognizes that unless it can show significant progress at reducing emissions, it is unlikely to be able to undertake a much-needed terminal modernization and expansion pro- gram. The airport planning staff has suggested that as part of the terminal modernization program the airport construct an automated people mover (APM) link to the nearest station on the regional light-rail system, located approximately one mile from the airport, to replace the existing infrequent tran- sit bus service between the airport and the light rail system. However, there are concerns about the capital and operating cost of the proposed people mover link and the likely use it will attract. As part of the landside planning for the modern- ization program, the airport decides to undertake a feasibil- ity study of the APM link in cooperation with the regional transit authority. The scope of work for the feasibility study includes the development of a ground access mode choice model that can be used to evaluate the likely increase in the use of the light rail system for airport trips if the APM is constructed, as well as explore alternatives such as an enhanced shuttle bus connection to the light rail system or an express bus service to several off-airport terminals with remote parking. The MPO for the region served by Airport C has recog- nized for some time that its regional travel demand model significantly underestimates vehicle trips to and from the air- port when compared with the vehicle traffic volumes pro- jected in landside planning studies undertaken by the airport authority and environmental documentation prepared for air- port projects. A detailed review of the travel demand model and discussions with airport planning staff reveal that the underestimate is the result of two factors. The first is that the standard trip generation and attraction relationships in the model substantially underestimate the number of air passen- ger trips to and from the airport owing to the absence of relevant variables in the model and that the modeling process does not explicitly consider visitor trips to the region. Sec- ond, the mode choice model in the overall modeling process underestimates vehicle trips and produces poor estimates of the traffic composition by not accounting for several key characteristics of airport ground access and egress travel. These issues include not accounting for the two-way trips by private vehicles picking up or dropping off air passengers, underestimating the proportion of shared-ride automobile trips by ignoring that many air travel parties have more than one person, and those dropped off or picked up by private vehicles necessarily involve a shared-ride trip in one direc- tion, not considering taxi and limousine trips, and treating vehicle trips by public modes such as a shared-ride van as part of regular public transit use. To improve the ability of the regional travel demand model to estimate airport trips the MPO decides to develop a special-generator sub-model for airport access and egress trips. This sub-model combines a trip generation module and a mode choice module and generates vehicle trip tables for each category of vehicle trips included in the regional model that are combined with the trip tables for general regional travel before the traffic assignment stage of the regional

model. The trip generation module converts airport passenger traffic forecasts into estimates of person-trips between the airport and each regional travel analysis zone, whereas the mode choice module converts these estimated person-trips into the corresponding vehicle trips for use in the regional travel demand model. Airport Choice The need to model air traveler choice of airport in multi- airport regions or situations where air service competition exists between local and more distant airports is becoming increasingly prevalent in airport system planning studies, par- ticularly as low-fare carriers introduce service at secondary airports often some distance from congested major hubs. It is recognized that airport accessibility is a major determinant of airport choice, because the availability and use of different ac- cess modes affects the perceived accessibility of different air- ports. In consequence, the airport access mode choice process is often embedded within models of airport choice and im- proved representation of airport access mode choice behavior should benefit those applications as well. OVERALL MODE CHOICE MODEL DEVELOPMENT PROCESS This section presents an overview of the process of develop- ing an airport access mode choice model. The application of such a model to support planning or decision-making activi- ties is discussed in a subsequent section. In broad terms, the general approach to developing a mode choice model is no different from the development of any other mathematical model of a physical or behavioral process. A set of data is assembled that describes the process being modeled. Then, a suitable functional form for the mathematical model is defined that expresses the value of the variable that the model is intended to predict in terms of some other explanatory variables and model coefficients (sometimes referred to as parameters), the values of which are to be determined from the data. Statistical model estimation software is then used to estimate the values of model coefficients that best explain the observed values of the variable that the model is intended to predict, termed the dependent variable, using the observed values of the explanatory variables, termed the independent variables. The extent to which the model is able to reproduce the observed values of the dependent variable for any given set of values of the independent variables is referred to as the goodness-of-fit of the model and is an important measure of the usefulness of the model. As with most model development efforts, the process is usually iterative. An initial functional form for the model is proposed. The values of the model coefficients are estimated and the fit of the model to the data examined. Then, based on how well the model appears to fit the estimation dataset and the reasonableness of the estimated coefficient values, the 16 functional form is modified and new coefficients are esti- mated, hopefully improving the fit of the model. The process continues until a satisfactory model is obtained. Changes to the functional form of the model that are typically explored include adding or dropping independent variables, changing the way that an independent variable is defined or appears in the model, or segmenting the data so that different coefficient values are obtained for different subsets of the data, or dif- ferent independent variables or model functional forms are used for different subsets of the data (e.g., developing sepa- rate models for business and non-business travel). In the case of airport access mode choice models, there are two types of data that are required to estimate a model. The first type of data consists of the access mode choices made by a representative sample of airport users, together with ex- planatory data about their characteristics (such as their household income) or the characteristics of their trip (such as where they commenced their journey to the airport or the purpose of their air trip). These data are typically obtained from surveys of air passengers or airport employees, as the case may be. The second type of data consists of the trans- portation characteristics (such as travel time and cost) of the various access modes between which the mode choice deci- sion was made. Because the mode choice decision depends not only on the transportation characteristics of the mode actually chosen, but also on the characteristics of the modes that were not chosen, it is generally necessary to obtain the transportation characteristics for all the modes included in the model corresponding to each airport user included in the estimation dataset. This can be a significant amount of work and is discussed further later. The overall process of developing an airport access mode choice model is summarized in Figure 1 and discussed in more detail in the following sections. Aggregate Versus Disaggregate Models There are two broad types of behavioral models such as a transportation mode choice model. An aggregate model attempts to predict the value of an attribute of a group of decision makers, such as the percentage of air passengers from a given origin zone choosing a particular mode. In con- trast, a disaggregate model attempts to predict how an indi- vidual decision maker with a given set of characteristics will behave. In the context of airport access mode choice, such a model typically attempts to predict the probability that a given airport user will choose a particular mode, because two airport users with apparently identical characteristics may well choose different access modes. If the probability of choosing each access mode is estimated for each airport user in a given sample of users, then the percentage of any group of such users choosing a given mode can be calculated. Because a disaggregate mode choice model predicts the probability of a decision maker choosing a given mode from

17 among a defined set of alternatives, these models are also referred to as discrete choice models. Because the airport access mode choice decision depends on the characteristics of each individual travel party as well as the transportation characteristics of the different modes faced by that travel party, which in turn vary with the travel party characteristics (e.g., the travel costs typically vary with the access trip origin and the travel party size), most airport access mode choice models take a disaggregate approach. Indeed, it arguably impossible to develop a reasonable aggregate airport access mode choice model that adequately reflects all the relevant variables. However, the decision to develop a disaggregate model imposes a number of constraints on the form of the model and the approach used to estimate the model coefficients. A mathematical form must be chosen for the model that pre- dicts the probability of a given air travel party or airport employee choosing a given access mode. Suitable mathe- matical forms for the model are discussed in the next section. Because the model predicts the probability of a given deci- sion maker choosing each of the available access modes but the estimation data give the access mode actually chosen, the statistical approach to estimating the model coefficients typ- ically uses a maximum likelihood method rather than the more common statistical regression methods. The maximum likelihood method determines the value of the model coeffi- cients that maximize the likelihood that the model will pre- dict the mode actually chosen by each decision maker in the sample, where this likelihood is defined as the overall proba- bility that the model predicts the actual mode choices (this is simply the product of the predicted probabilities of choosing the mode actually chosen across all the decision makers in the sample). Readers interested in more background on the mathematical details of estimating disaggregate choice mod- els are referred to standard texts on the subject, such as Ben-Akiva and Lerman (1985) or Train (2003). It should be noted that each data point (or record) in a dataset used to estimate a disaggregate choice model repre- sents a single choice decision. In the case of airport access mode choice, air passengers often travel as a party of more than one person and the data point is the travel party, whether composed of one or more air passengers. Each travel party is considered to make the mode choice decision as a single unit, Model Specification Model Estimation Model Calibration Model Validation Review Goodness of Fit Model Application Airport-User Survey Assemble Transportation Service Data Airport-User Survey Data Transportation Service Data Model Estimation Dataset Model Calibration Dataset Assemble Calibration Data Assemble Validation Data Model Validation Dataset FIGURE 1 Mode choice model development process.

whether or not the decision is made jointly by the members of the party or by one individual within the party. In contrast, airport employees are usually assumed to make airport ac- cess mode choice decisions individually, even if they decide to travel to the airport in a group (e.g., a car pool). A further complication with airport ground access travel by air passengers that is commonly ignored in developing mode choice models is that the ground access travel party and the air travel party may be different. For example, two colleagues making a business trip together may travel sepa- rately to the airport and then meet up at the airport and fly together to their destination. Conversely, situations may arise in which the members of a ground access travel party take different flights after they reach the airport, such as attendees at a conference who decide to share a taxi to the airport. In these cases the unit of decision is the ground travel party, not the air travel party. Therefore, the access mode choice model should strictly predict the mode choice decision of each ground travel party, rather than each air travel party. Shared- ride modes such as a door-to-door van will typically combine more than one ground travel party in a single vehicle, al- though each ground travel party is considered to make a sep- arate decision, because its members could have chosen a different access mode without affecting the decisions of the others on the vehicle. In practice, any error that would be introduced in an airport access mode choice model by ignor- ing the distinction between the ground access travel party and the air travel party is likely to be very small, because the dis- tinction is only relevant for a small proportion of air travel- ers. However, it is a point that should be borne in mind when designing air passenger surveys that will collect data for use in developing airport access mode choice models, because it can affect the way that questions are worded. Model Estimation Data As noted earlier, model estimation data are obtained in two different ways: 1. Surveys of air passengers or airport employees, as the case may be; and 2. Assembly of transportation service data for each mode from operator and agency records and data files, pub- lished information, and other sources. Survey Data The air passenger or airport employee surveys obtain informa- tion on the access mode used for a specific trip, typically the most recent one. In the case of air passengers who are surveyed at the airport, this will usually be the access trip that they have just completed. In the case of airport employees, the survey may ask about access trips over a recent period, such as the previous week, to identify variation in access mode use from day to day or to reduce the effect of some special circumstance 18 on the day of the survey. In addition, the survey will obtain information on respondent characteristics that may be used as independent variables in the mode choice model or to segment the data for model estimation. The survey will also need to identify where the access trip began to determine the transportation service characteristics of the different modes that were faced by the respondent. The more accurately the trip origin location can be determined, the more precisely the corresponding transportation service levels for the various modes can be estimated. Because the data on transportation service levels are typically assembled on a zonal basis, the trip origin locations need to be expressed in terms of the system of analysis zones used for the trans- portation service data. The issues involved in selecting an appropriate system of analysis zones are discussed further later. However, with the possible exception of U.S. Postal Service zip code areas, most practical systems of analysis zones will not be such that respondents can be expected to know which zone their trip began in. Some respondents may even have difficulty with zip codes. Therefore, the usual approach is to attempt to obtain the street address of the trip origin or the name of a specific origin location, such as a hotel, for which the address can be obtained later. For pri- vacy reasons, the exact street address is not required and typically the block number or a nearby street intersection is considered adequate. These addresses can then be geocoded and later assigned to the appropriate analysis zone. The respondent characteristics that are used as indepen- dent variables or to segment the data will depend on the func- tional form of the model. Because this is typically not known at the time that the survey is conducted, but evolves during the model estimation process, the survey should attempt to collect information on those respondent characteristics that are believed to influence the mode choice process, even if some end up not being used in the model. Although there is a small cost to collecting data that are not used, if the infor- mation is not collected, one will never know how important it might have been in the model. The following is list of air passenger or airport employee characteristics that have either been shown to influence airport access mode choice or might reasonably be expected to do so. • Air Passengers – Essential  Trip purpose (business vs. personal)  Resident of region or visitor  Primary ground access mode  How accessed primary mode (where relevant)  Where parked (if relevant)  Trip origin location (address, hotel name, etc.)  Trip origin type (residence, hotel, etc.)  Number of air passengers in travel party  Air trip duration (nights away on trip)  Household income  Household size.

19 – Potentially useful (used in some models)  Amount of checked baggage  Number of air trips from airport in past year  Whether trip costs paid by employer or client  Time arrived at airport  Gender of respondent. • Airport Employees – Essential  Primary ground access mode  How accessed primary mode (where relevant)  Where parked (if relevant)  Monthly parking cost (if any)  Whether any travel costs paid by employer Δ Trip origin location (home address)  Work location on airport  Times shift starts and ends  Variability of shift times  Household income  Household size. – Potentially useful  Job type/classification  Employer  Number of automobiles owned by household. Some of the characteristics may appear directly in the mode choice model, whereas others are needed to deter- mine the appropriate travel costs of the various modes. In the case of air passengers, the number of air passengers in the ground access travel party affects the relative cost of different modes, whereas the duration of the air trip affects the cost of parking for residents of the area. For those modes (such as rail transit or express airport bus) where a secondary access mode is needed to reach the primary mode, the access mode used will affect the cost and time of the trip. Some mode choice models may even model this access mode choice. Accounting for income effects in mode choice models is problematic, as discussed elsewhere in this report, and many past models have ignored the issue entirely. However, clearly income must have some effect on airport access mode choice. Household income may be a more appropriate mea- sure than individual income, because this reflects the contri- bution of other members of the household in covering basic household costs; however, at the same time the discretionary income for a given level of household income will also depend on the size of the household. As important as deciding what information should be col- lected in the survey is deciding how to word the questions. Poorly worded questions will produce unreliable data, be- cause respondents may misunderstand the question and give an incorrect answer. In the case of self-completed surveys, this applies not only to the questions themselves, but also to any predefined response options that are provided. The de- sign of survey questionnaires and question wording is a topic in its own right and beyond the scope of this report. However, an ACRP study currently underway (Project ACRP 03-04) is developing a “Guidebook for Airport-User Surveys,” which will provide detailed guidance on these and other related issues. Stratified Sampling Because the use of different transportation modes for airport access varies widely, and in particular at many airports pub- lic transportation has a fairly small market share, surveying a random sample of air travelers (e.g., in airport terminal departure lounges) will result in relatively few respondents who used the less common modes. This will in turn adversely affect the ability of a model estimated on that data to explain the choice behavior of the respondents using those modes. One approach to overcoming this problem is to perform stratified sampling, in which the survey is done in a way that obtains a greater number of responses from a particular sub- set of travelers than would be expected from a truly random sample. For example, surveys could be performed where pas- sengers alight from transit vehicles. Alternatively, the survey could be done in a location where respondents selected at random will intercept all types of travelers, but a screening question asked at the start of the survey will identify respon- dents in the subset of interest and the survey of those respon- dents will be performed in more depth. Estimating mode choice models on the basis of data ob- tained from a stratified sample requires some adjustments to the standard model estimation techniques to weight the response data in the estimation process; however, these ad- justments are discussed in most standard textbooks on dis- crete choice modeling. Similarly, in presenting the results of a stratified sample, it is important to weight the responses appropriately to reflect the correct distribution of respondent characteristics across the population of airport users as a whole. Developing appropriate response weights for both model estimation and presenting survey results requires data on the frequency of occurrence of the different categories of respondent. These weights can be obtained from actual traf- fic counts of the different respondent categories where these are available or by comparing the results of the stratified sample with that of a random sample of the larger population of air passengers or airport employees as the case may be. Transportation Service Data To calculate the travel times and costs faced by each travel party in using the different airport ground transportation ser- vices, it is necessary to assemble the transportation service data for each mode on the basis of a defined system of analysis zones so that the appropriate value of any particular service charac- teristic for a given travel party can be determined from the trip origin zone for that party. Some data, such as the highway travel time to the airport, will vary across the zones. Other data, such as daily airport parking rates, are independent of the trip

origin zone and thus constant for every zone, although the cost for parking will depend on how long the vehicle is parked. Some costs, such as fares for some public transportation ser- vices will depend on both the origin zone and the number of people in the travel party. Highway travel times, highway distances, and transit travel times and fares are typically available from the re- gional transportation planning agency. However, modes such as taxi, shared-ride van, and scheduled airport bus, which are not usually included in the regional travel demand model, will require some work to assemble the necessary data. The ground transportation information pages of the airport web- site may have current information on many of the ground transportation services and parking rates. Otherwise, it may be necessary to contact the individual operators to obtain fare and schedule information. In the case of scheduled services with defined stops or stations, the schedule will generally provide headways and travel times from each stop or station. It will be necessary to determine the analysis zone for each stop, as well as the analysis zones served by that stop and the transportation service characteristics (travel times and costs) for the secondary access trip to reach the stop from each zone served by the stop. Because there may be several alternative secondary access modes (such as walk, drop off by private vehicle, taxi, or public transit), the transportation services characteristics will have to be determined for each mode and each analysis zone. The result of this process is a large data table with a row (or record) for each analysis zone and the transportation ser- vice variables for each mode and sub-mode forming the columns (or position in the record). The appropriate trans- portation service values can then be assigned to each travel party in the model estimation dataset by looking up the rele- vant data for the trip origin analysis zone for that party and computing the values where necessary to account for the travel party size, air trip duration, or other travel party char- acteristics that affect the value of the transportation service variable for that party. Analysis Zones Selection of a suitable system of analysis zones involves a tradeoff between the precision of the transportation service data used for each travel party in the sample and the work involved in assembling the necessary data. Some of the data may already be available in a particular system of analysis zones. For exam- ple, regional transportation planning agencies will generally have computer files with highway travel times between each transportation analysis zone (TAZ) used in their regional travel demand model and the TAZ containing the airport. Indeed it may be desirable or even required that an airport access mode choice model is based on the regional travel demand model TAZ system so that the results of the modeling can be inte- grated with the regional travel demand modeling process. 20 However, a large metropolitan region may easily have more than 1,000 TAZs, and some larger regions have signif- icantly more than that. The current regional travel demand model for the San Francisco Bay Area utilizes 1,454 TAZs, whereas that for the Washington metropolitan area utilizes 2,191. For those travel variables that can be obtained directly from the regional transportation modeling datasets, such as highway times or transit fares, this is not a particular prob- lem. However, for other modes, such as shared-ride van, where fares are not typically available from the regional datasets, or are usually expressed in terms of the TAZs, obtaining the relevant data and converting it to the TAZs can be a major task. It might appear that the data management problem can be reduced somewhat by using larger zones, such as U.S. Postal Service zip code areas. However, apart from the loss of pre- cision involved in using larger zones, most transportation service data are not available on a zip code basis anyway, so the work involved in converting the data to a zip code-based system may not be significantly less than using TAZs. Where data are available by zip code (e.g., some shared-ride van operators base their fares on zip codes), it is fairly easy to develop a mapping from zip code areas to the corresponding TAZs and convert the data to a TAZ basis. Most geographi- cal information systems have functions that can do this auto- matically provided that the TAZ boundaries are available as a geographical information systems file. Revealed Versus Stated Preference There are two different approaches to assembling the neces- sary data on airport access mode choice behavior to develop mode choice models: revealed preference and stated prefer- ence surveys. Revealed preference surveys identify the travel choices actually made by airport travelers, as well as collect information on other traveler characteristics and details of the trip that are believed to influence the choice. The model estimation process attempts to develop a model that explains the mode choice decisions in terms of the traveler character- istics and the service characteristics of the different airport ground access modes available to the traveler. Stated preference surveys follow a similar process, except that the respondent is presented with a set of hypothetical choices and asked to select from them. For realism, the stated preference experiment is usually structured so that the choices presented to the respondent correspond to their current trip or a recent actual trip but change the characteristics of the ground transportation options available, such as different prices or travel times or the introduction of a new service or mode. Estimating a mode choice model on such data allows the model to incorporate ground access options that do not currently exist or to explore the effect of changing factors that do not exhibit much variation in the real world. Although this is a powerful capability, there are concerns about how well

21 travelers’ expressed choices between hypothetical situations that they have not actually encountered correspond to how they would really behave if faced with those situations in practice. To attempt to address this concern, stated preference studies are often combined with analysis of revealed prefer- ence to at least ensure that the stated choice behavior is consistent with the actual behavior when applied to situations that have actually been experienced. Even so, there remains the concern that survey respon- dents may overstate their willingness to use new alterna- tives that do not yet exist, whether out of a desire to appear responsive to the survey or because they misinterpret how attractive the new service will be relative to their current choices. There is a limit to how detailed the description of a new service can be in the context of a stated preference survey that has to be completed in a fairly short time period, and respondents may not fully consider all the factors that would arise in using the new service, such as how they would get to it at the time that they need to travel. A related factor is that a stated preference survey necessarily informs the respondents about the options that are hypothetically available so that they can make a choice. However, in prac- tice many travelers may not be aware of the existence of the service or may have a misperception of the nature of the service offered. Habit and Information An important aspect of air passenger airport access mode choice is the role of travel habits and the awareness of travel alternatives. In contrast to most urban trips, such as the journey- to-work or shopping trips, most air passengers do not make air trips that often, and many visitors to a region may be vis- iting for the first time. It is self-evident that travelers will not choose to use modes that they do not know exist, or even if they have significant misperceptions about the nature of the service offered, such as the travel time involved or the cost. However, most air passenger surveys ignore questions of travelers’ awareness of difference services and their percep- tions of the service offered by modes that they did not use. As a result, the estimation process of most mode choice models has implicitly assumed that travelers have full information about each mode. The issue of how to address habit and in- formation in airport access mode choice models (or any mod- els of travel behavior for that matter) is not well understood and is one that could benefit from further research. This is an important consideration because it directly ad- dresses the role of marketing in the provision of airport access services. Airport authorities or transportation operators can choose to spend resources improving services or more inten- sively marketing the services that they already provide. It would be extremely helpful if airport access mode choice models could help shed some light on how best to allocate resources between service improvements and marketing. However, for this to happen it will be necessary to incorporate traveler information more explicitly in the models. Model Specification, Estimation, Calibration, and Validation The terms model estimation and model calibration are often loosely used interchangeably. However, strictly speaking they are two different steps in developing an airport access mode choice model (or indeed any model). Model estimation refers to the use of statistical procedures to determine the val- ues of the model coefficients that best fit the data from which the model is being developed. Typically, this will be derived from a sample of air passenger or airport employee trips obtained from a survey. Once an estimation dataset has been assembled, the first step in model estimation is model specification. This involves selecting an appropriate functional form and market seg- mentation for the model, and defining relevant explanatory variables. Model estimation software is then used to obtain es- timated values of the model coefficients. The statistical sig- nificance of these estimated values and the overall goodness- of-fit of the model is examined and the model specification revised as necessary to address problems with the resulting model coefficients or statistical fit. Model development typi- cally proceeds iteratively. A fairly simple functional form with relatively few explanatory variables is initially estimated. Then the model is improved progressively by adding variables or modifying the functional form, such as changing the struc- ture of the way that modes are grouped within the model to im- prove the statistical fit of the model to the estimation data. However, statistical fit is not everything. A model must also make sound behavioral sense. A model that reflects a plausible structure of behavioral causality is generally preferred to one that contains counterintuitive features, even if the latter has a better statistical fit. Once a satisfactory model has been estimated, model cal- ibration should be undertaken to make any necessary adjust- ments to the model so that the model predictions agree with observations. If the model estimation has been done cor- rectly, the model predictions will agree with the observed data in the model estimation dataset. However, because these data may not be a truly representative sample of the larger population being modeled, the model may need to be ad- justed to produce satisfactory predictions. In the case of an airport access mode choice model, where the composition of the air travel market will vary seasonally or even from day to day and assumptions will need to be made about average vehicle occupancy of some modes to convert the number of person-trips using those modes to the equivalent number of vehicle trips, the total volume of vehicle trips by each mode predicted by the model using the estimated coefficients may differ from the observed volume of vehicle trips for any given period. This is particularly the case when a model has been estimated on survey data collected at one point in time,

typically a few weeks or less, but is being used to predict mode use for a different time period, such as a year. Even if the survey data used for model estimation is an ac- curate representation of the larger population, the conversion of model predictions from travel party trips to vehicle trips generally will require some model calibration. Because access mode choice models are estimated on data for a sample of air passenger travel parties or airport employees, they generate predictions of travel party or employee trips using each mode, which typically need to be converted to vehicle trips for use in planning studies or other applications. For most private vehi- cle modes the conversion is straightforward, because each travel party generates one vehicle trip in each direction (or two in the case of drop off or pick up by private vehicle). However, for most other modes the ratio of vehicle trips to travel party or employee trips depends on assumptions about average vehicle occupancy or (in the case of modes such as taxi or limousine) the proportion of deadhead trips to revenue trips. For sched- uled modes, the number of vehicle trips is determined by the schedule rather than the mode use. Where observed counts of vehicle trips are available, it will generally be necessary to adjust the assumptions about average vehicle occupancy or proportion of deadhead trips to match the predicted values of vehicle trips to the observed values. Model validation is the final step in model development and involves comparing the predictions of the model under different conditions from those under which it was estimated and calibrated, usually after some change has occurred in the system being modeled, with the actual values of the phe- nomena being modeled. In the case of an airport access mode choice model, this could involve comparing projections of mode use in subsequent years or after changes have occurred in the ground transportation services available at the airport with the observed mode use under those different conditions. Although limited model validation can be done using partial data on the use of certain modes, such as comparing the num- ber of private vehicles parked at the airport with the number projected by the model, a more thorough model validation will require an extensive effort to collect comprehensive data on mode use for the validation period. Because the pattern of access mode use at an airport will de- pend on the composition of the air travel market or of the employee workforce, as well as the transportation services available, a true validation of an access mode choice model should include a new air passenger or airport employee survey to ensure that the market composition assumptions being used in the modeling are correct. Otherwise, it is unclear whether differences between the predicted mode use and the observed mode use are the result of problems with the model or invalid assumptions about the market composition. Similarly, it is important that the transportation service assumptions for the various modes that are used in the modeling are updated to reflect changes in costs and service levels since the model was originally developed. 22 Because of the effort and resources required to assemble the necessary data to perform a proper validation of an air- port ground access mode choice model, such validations are rarely if ever done and it is simply assumed that a calibrated model that has been developed from data on mode use pat- terns at one point in time will remain valid when used to model airport access mode use at other points in time or under different conditions. Also, because model validation involves comparing model predictions with observed condi- tions under different conditions from those from which it has been calibrated, this often cannot be done until some time after the initial model estimation effort, by which time it may already have been used for the application for which it was developed. However, airport access mode choice model de- velopment should not be viewed as a one-time effort, any more than any other aspect of travel demand analysis. Rather, opportunities to validate a model should be sought and pur- sued following its initial development, and the model refined and improved over time. In this way the model will be avail- able for subsequent applications with a growing level of con- fidence in its predictive reliability. MATHEMATICAL FORM OF TYPICAL MODE CHOICE MODELS The challenge of developing mathematical models of dis- crete choice behavior has attracted the interest of statisti- cians, economists, social scientists, and transportation plan- ners over a long period of time and thus not surprisingly there is an extensive literature on the subject that is beyond the scope of this report to summarize. Some of the earliest appli- cations of discrete choice behavior models to transportation travel demand were undertaken by Daniel McFadden and colleagues (described in Domencich and McFadden 1996), as part of work for which McFadden was awarded the Nobel Prize in economics. Readers interested in the theoretical background and evolution of the current state of practice of transportation mode choice models can refer to standard texts such as Ben-Akiva and Lerman (1985) or Hensher et al. (2005). However, to help readers who have limited or no prior familiarity with these techniques understand the gen- eral approach; this section will attempt to provide a simpli- fied introduction to the current state of practice. Those with some familiarity with transportation mode choice modeling may choose to skip this discussion. The basic concept underlying most disaggregate discrete choice analysis is that each alternative in the choice set pro- vides the decision maker with some utility that can be expressed in terms of measurable or observable characteris- tics of both the decision maker and the alternative (e.g., the travel time involved or the income level of the decision maker). The larger the difference in the utility between two alternatives, the more likely the decision maker is to choose the alternative with the higher utility. This can be illustrated by the relationship shown in Figure 2, which shows the prob- ability of choosing alternative 1, P(1), as a function of the

23 difference between the utility of alternative 1, U1, and that of some other alternative 2, U2, (termed a binary choice be- cause it involves only two alternatives). As this difference becomes large and positive (i.e., the utility of alternative 1 is much greater than that of alternative 2), the probability of choosing alternative 1 approaches one (a probability by def- inition must lie between zero and one). Conversely, if the difference in utility is large and negative (i.e., the utility of alternative 2 is much greater than that of alternative 1), the probability of choosing alternative 1 approaches zero. If the difference in the utilities is zero, the two alternative are equally attractive and the probability of choosing either is equal to 0.5 (50%). Multinomial Logit Model As shown in Figure 2, the functional relationship between the difference in utility and the probability of choosing a particular alternative will be an S-curve, because it must be asymptotic to one and zero as the utility difference becomes very large in either the positive or negative direction. Early work on choice modeling identified the logistic function, defined as: as providing a suitable S-shaped curve, where ex is the expo- nential function of x. It can be seen that as x becomes very large and positive, f(x) approaches one and as x becomes very large and negative, f(x) approaches zero. In the context of a choice model between two alternatives (termed a binary or binomial choice) the logistic function can be restated as: P i e U Ui j( ) ( )= +( )− −1 1 f x e e ex x x( ) ( ) ( )= + = + −1 1 1 where Ui and Uj are the utilities of alternatives i and j, respectively, and P(i) is the probability of choosing alterna- tive i. This in turn can be reexpressed as: which became known as the logit model (strictly the bino- mial logit model). It can be shown fairly easily that with more than two alternatives the model can be extended as follows: where J is the number of alternatives. This is termed the multinomial logit (MNL) model and has been widely used for airport access mode choice models, because most air- port access mode choice situations involve more than two alternatives. Although the logit model was initially simply a convenient way of generating the required S-shaped relationship, later work showed that under certain assumptions regarding the form of the utility terms the logit model can be derived from theoretical principles of utility maximization, whereby the decision maker chooses the alternative that offers the highest utility, and the probability of choosing a given alternative is thus given by the probability that that alternative offers the highest utility of all the alternatives. Utility Function In developing discrete choice models, the utility of a given alternative is assumed to comprise two parts; a deterministic part that consists of a function of measured and observed P i e e U U j J i j( ) = ∈ ∑ P i e e eU U Ui i j( ) = +( ) 0.0 0.5 1.0 -4 0 4 Difference in Utilities (U1-U2) Pr ob ab ili ty o f C ho os in g A lte rn at iv e 1 FIGURE 2 Illustrative binary choice relationship.

variables and an error term that accounts for unobserved char- acteristics and variability in the perceived utility of a given set of characteristics across different individuals, therefore: Ui = Vi + ε where Vi is the deterministic part of the utility and ε is the error term. In logit choice models, the error term is assumed to be a random variable with values that are independent and identically distributed with a Gumbel (double exponential) distribution with a zero mean. (This assumption allows the logit model to be derived from utility maximization theory.) The variance of the error term reflects the goodness-of-fit of the model. The deterministic part of the utility function typically con- sists of a linear combination of explanatory variables with their associated model coefficients, the values of which are determined in the model estimation process, therefore: Vi = ai + b1x1 + b2x2 + . . . + bnxn where ai and the b’s are the model coefficients and the x’s are the values of the explanatory variables, such as travel time and cost. In general, the utility function for each alternative will have a constant term ai, known as the alternative-specific constant (ASC), which reflects attributes of the alternative that are not accounted for by the other variables. Therefore, a fairly simple utility function might comprise: Vi = ai + b1(travel time) + b2(waiting time) + b3 (walk distance) + b4(cost/income) Note that in this example travel cost is divided by house- hold income in the fourth explanatory variable so that the choice process becomes less sensitive to cost for higher income travelers. This is included as an illustration that all explanatory variables do not have to enter the utility function as separate terms and may not be the best way to reflect the effect of household income. It should also be noted that changing the utility functions of each alternative in a logit choice model by the same amount will not affect the resulting probabilities (because the change will factor the numerator and denominator of the logit expression for each alternative up or down by a constant amount that will cancel out). Therefore, it is usual practice to set the ASC for one of the alternatives to zero, so that the ASCs for the other alternatives then reflect the differences in the constant part of the utility for those alternatives relative to the alternative without an ASC. The estimated coefficients of the utility function can be thought of as weighting factors that convert the units of the explanatory variable (e.g., minutes of travel time) to a mea- sure of perceived utility. Because perceived utility is an abstract concept that has no intrinsic units of measurement, 24 the estimated values of the coefficients have no direct inter- pretation. However, the ratio of the coefficients for two variables (or the ratio of the ASC to a coefficient of an ex- planatory variable) is another matter. This ratio expresses how an increase in one variable (or the ASC) will offset a decrease in the other variable and thus can be expressed as implied values. The ratio of a given coefficient or ASC to the coefficient for the cost term gives the implied value of that variable or constant in the units of the cost term. With appro- priate adjustments for the units of two variables, this can give implied values of time in dollars per hour. Where the travel cost variable incorporates some function of household or other income the resulting implied values will be expressed in terms of this income measure. This has an important implication for the specification of the utility function. Although different components of travel time can (and often should) be expressed with sepa- rate variables, cost terms should be combined into a single variable. This will avoid problems with the model giving different implied values of a given variable depending on which cost term coefficient is used. This is also conceptu- ally sound. Although travelers may (and usually do) perceive different components of travel time as having a different disutility per unit time (i.e., different implied val- ues), it would be surprising if they view a dollar spent on one aspect of the airport access journey any differently from a dollar spent on another aspect, because the money involved is completely interchangeable. (The one exception to this principle would be if some travel costs are reim- bursable and others not. In this case, it would make sense to use separate variables for the two types of cost so that the ratio of the two cost coefficients reflect the relative impor- tance of reimbursable to nonreimbursable costs. This would still allow values of time to be expressed consistently in terms of nonreimbursable costs.) Limitations of Multinominal Logit Model Although the MNL model has been widely used, it is vulner- able to problems that arise from a property of the model termed the “Independence from Irrelevant Alternatives.” This states that including a new alternative in the choice set (or changing the perceived value of one of the alternatives) should not affect the relative probabilities of choosing any of the other alternatives. It can be seen from the previous equa- tion for the MNL model that the ratio of the probability of choosing any two alternatives is determined only by the per- ceived utilities of those alternatives. However, in many situations in airport access mode choice it is quite unlikely that changing the characteristics of one mode or sub-mode will leave the relative probabili- ties of choosing all the other modes and sub-modes unchanged. For example, increasing the parking rates in the short-term parking lot is likely to have a greater effect on

25 the probability of an air party choosing to park in the long- term parking lot than on the probability of choosing to use a shared-ride van, because those who would have parked in the short-term lot at the former rates are much more likely to choose to park in the long-term lot instead than to decide to use a shared-ride van. Similarly, changes in one public transportation service are likely to affect the use of other public transportation services to a greater extent than the use of private vehicles. This is not usually a significant problem with a model that only includes a limited number of well-differentiated modes, but becomes increasing problematical with models that at- tempt to distinguish between the use of similar modes (such as taxi and limousine) or to account for sub-modes (such as different parking facilities). Nested Logit Model The limitations of the MNL model can be addressed by grouping similar modes or sub-modes into separate groups or nests in a choice structure referred to as a nested logit (NL) model, as illustrated by Figure 3. In the nested model shown in the figure, alternative b consists of a second-level nest of two sub-alternatives, b1 and b2, the second of which consists of a third-level nest of two further sub-alternatives, b21 and b22. For example, alternative b might represent the use of a private vehicle, with alternative b1 representing the air party being dropped off at the airport and b2 the use a private vehicle that is parked at the airport for the duration of the air trip, where b21 represents the use of the short-term parking lot and b22 the use of the long-term parking lot. The general form of the NL model is similar to the MNL model, with the addition of a scaling parameter μm for each nest m, as follows: P m e e U j j N U k k N m m m l l ( ) ( ) ( ) = ( )⎛⎝⎜ ⎞ ⎠⎟ ( ) ∈ ∈ ∑ ∑ 1 1 μ μ μ ⎛ ⎝⎜ ⎞ ⎠⎟∈∑ μl l S P i m e e U i U j j N m m m ( ) ( ) ( ) = ( ) ( ) ∈ ∑ 1 1 μ μ where P(i|m) is the probability of choosing mode i in nest m from among the set Nm of modes in nest m, given that nest m is chosen (strictly that a mode in nest m is chosen) and P(m) is the probability of choosing nest m from the set S of nests at the same level as nest m. If one branch of a nest consists of a discrete mode m* rather than a lower-level nest, the value for the scaling parameter for that mode μm* = 1. Therefore, if there is only one nest, the previous equations reduce to the MNL model. Estimation of NL models involves estimating values for the scaling parameters as well as the utility function coefficients. Other Model Specifications Although the NL model overcomes some of the inherent limitations of the MNL model, there remain a number of other limitations to the use of NL models for modeling air- port access mode choice. Perhaps the most significant of these is the assumption that the variance of the error term in the utility function is the same for all air parties and all alternatives. Another limitation can arise where the same alternative appears in different nests; for example, if several public transportation alternatives have station or stop access sub-mode nests, because these will typically involve the same sub-modes. Efforts to explore alternative model formulations to NL models for transportation mode choice applications have taken two approaches. One is to use more advanced logit model formulations that address some of the limitations in the standard model. The other is to use an entirely different conceptual approach to representing the mode choice process. However, to date neither of these approaches has been applied to stand-alone airport ground access mode choice mod- els, although some work has been done on modeling airport choice in which ground access mode choice forms part of the choice process (Hess and Polak 2005, 2006). Although these alternative approaches may be a promising area for future research, given the limited experience applying approaches other than MNL and NL models to airport ground access, the details of these approaches are not discussed further here. Read- ers interested in more information will find a discussion of the potential application of alternative mode choice model approaches to airport access travel in Lu et al. (2006). MODEL ESTIMATION CONSIDERATIONS Apart from questions of selecting an appropriate functional form and market segmentation for the model and defining relevant explanatory variables, careful consideration needs to be given to developing the necessary estimation dataset. The reliability of the resulting model is critically dependent on the accuracy of the data on which it is estimated. If the data contains errors or biased values of explanatory vari- ables, the model will attempt to explain the traveler behavior b22 a b c d alternatives Multinomial Model a b1 b21 alternatives Nested Logit Model FIGURE 3 Multinomial and nested logit models.

in terms of those values rather than the real values. This may not be a problem in obtaining a good fit to the data, but will produce biased predictions when the model is later applied to other datasets that do not have the errors or biases. Model Estimation Software The process of estimating airport access mode choice mod- els, such as mode choice models for other transportation applications, requires specialized software that is designed to perform maximum likelihood estimation with fairly com- plex model specifications. This software falls into three broad categories: 1. Special-purpose commercial software that has been developed specifically to estimate logit-type choice models or the mode choice component of transporta- tion planning models. 2. General purpose commercial statistical packages that include modules for performing maximum likelihood estimation of discrete choice models. 3. Software developed by academic researchers to esti- mate a fairly broad class of econometric models that includes discrete choice models. As might be expected, there is typically a tradeoff be- tween the ease of use and cost of the different categories of software. Special-purpose commercial software for estimat- ing logit choice models or transportation mode choice mod- els is often the easiest to use, because considerable attention has been given to the design of the user interface and the soft- ware has been specifically developed for this purpose. How- ever, the software is often more expensive than the other options, although the cost of acquiring software is usually small compared with the overall cost of the model develop- ment process. A number of general purpose statistical software pack- ages provide the capability to estimate discrete choice mod- els using maximum likelihood techniques. Sometimes this requires the purchase of an additional software module. Although the cost of this additional capability (if required) is usually less than acquiring special-purpose software, esti- mating a model with this software may require somewhat more effort than using special-purpose software, although the interface between the model estimation function and the data management capabilities of the package is usually straightforward. This may be an attractive option for organi- zations that already are using one of these software packages for general statistical analysis. Software developed by academic researchers to estimate various types of econometric models, often in support of their research or teaching activities, is often available at little or no cost. Depending on the scope of the software, this may require some customization to use to estimate a particular model specification. Documentation, user interface, and data 26 management features are often fairly basic, and user support may be limited or nonexistent. On the other hand, the soft- ware may contain features or provide capabilities that are not available in commercial packages. A good example of this class of software is BIOGEME, developed by Michel Bierlaire at the Ecole Polytechnique Fédéral de Lausanne, Switzerland (Bierlaire 2003), and available on the Internet at http://biogeme.epfl.ch. Unfortunately, obtaining comparative information on dif- ferent software options is not as easy as might be thought. With some persistence, Internet searches can generally locate information on the principal software packages. However, some creativity is required in defining the search expressions to avoid being swamped by links to articles about choice model estimation methodology rather than the software in- volved or econometric software in general. Market Segmentation The air passenger market is not homogeneous and different market segments have different airport access needs and available options. The most obvious distinction is between residents of the local area and visitors. Residents typically have access to a private vehicle and often someone who can take them to the airport or pick them up. Visitors on the other hand may need to rent a car to meet their transportation needs while in the area or may be staying at a hotel that does not provide a courtesy shuttle service to the airport. Another important distinction is between those traveling on business trips, whose travel costs may be reimbursed by their em- ployer or client, and those making trips for non-business pur- poses. These distinctions are typically addressed by defining different air passenger market segments and estimating a separate sub-model for each segment. The market segment sub-models may include different modes, may use different explanatory variables, and will generally have different esti- mated coefficients for a given variable. Although a common market segmentation approach, dis- cussed in more detail in subsequent chapters, is a four-way division into resident business trips, resident non-business trips, visitor business trips, and visitor non-business trips, other market segments may be worth considering. One is to differentiate visitors staying in a hotel from those staying with relatives or friends, because the latter may have access to pri- vate vehicles owned by the people they are staying with as well as people who can pick them up or drop them off at the airport. Changes in the Airline Industry The terrorist attacks of September 11, 2001 (9/11) have led to dramatic changes in the airline industry, in particular pas- senger security processing at airports. No longer can greeters and well-wishers enter the secure side of airport terminals,

27 and passengers need to arrive at the airport earlier than before to ensure being able to clear security in time to make their flight. In addition, airports rigorously enforce the prohibition on leaving vehicles unattended at the terminal curb front or even waiting when not actively loading or unloading passen- gers. The combined effect of these measures has been a significant change in airport ground access mode use. Fewer well-wishers accompanying passengers to the airport (or driv- ers dropping passengers off at the airport) park for a short time to accompany the passengers into the terminal, because the passengers are usually anxious to get through security as quickly as possible and the well-wishers cannot accompany them. Similarly, greeters can no longer meet arriving pas- sengers at the gate, and the increased use in cell phones over the past five years has simplified the process of picking up arriving passengers at the terminal curb front. As a result, many airports have introduced cell phone lots where drivers picking up air passengers can wait until the passengers are ready to be picked up. Another significant change in the airline industry since 9/11 is a reduction in the percentage of short-haul trips. This is widely believed to be at least in part a result of the need to arrive at the airport for a flight earlier than before, which increases the time required to make an air trip and in turn makes driving or other surface transportation modes rela- tively more attractive. There may also be a heightened con- cern over aviation security that leads some who do have a surface transportation alternative to take this option even if it involves more time than flying. A third consideration is the increasing market share of low-fare airlines and the compet- itive response of the network carriers. This has led to a much higher proportion of air travelers using low-fare or heavily discounted tickets. It would seem likely that passengers attracted to air travel by cheaper fares would also be more cost-sensitive when making their airport access decisions. It is also likely that a higher proportion of air travelers are fly- ing for personal rather than business reasons, although that has also been affected by cyclical changes in the economy. Although these changes in the airline industry do not fun- damentally affect the basic approach to modeling airport access mode use, they do affect the relative attractiveness of the different transportation services represented in the model and hence the estimated model coefficients, as well as the mar- ket composition. Any airport access mode choice model estimated on data from before 9/11 is likely to produce biased predictions of current or future air passenger access behavior. Similarly, current data on air passenger market composition, particularly the split between business and non-business travelers and the distribution of trip durations, is likely to have changed significantly from data collected before 9/11. Level of Effort Required It is clear from the foregoing discussion that developing an airport ground access mode choice model is not a simple or inexpensive matter. This raises the questions of how large an airport needs to be to justify the effort and what such an effort might cost, not easy questions to answer. For most large air- ports, with complex ground access systems to plan and man- age, it is very difficult to conceive how this can be done well without such a model (although many airports try to make do without one). For smaller airports, perhaps with traffic levels in the range of 5 to 10 million annual passengers, the need may depend on the type of planning issue being faced. The resources needed to develop an airport access mode choice model depend in part on the availability of air passen- ger survey data (or airport employee survey data in the case of employee access mode choice models). As discussed ear- lier, good air passenger survey data are critical to model development. Experience suggests that for adequate model development such a survey should have at least 3,000 re- sponses, although models have been developed with smaller sample sizes. Stated preference surveys have been conducted with fewer respondents, approximately 800 and 1,100 in two recent cases; however, each respondent typically answers several choice experiments, each of which provides a data point, and such surveys cost more to perform per respondent than a revealed preference survey. Although the cost per completed survey of a revealed preference survey can vary significantly with local circumstances as well as the number of questions asked, typical costs identified in an on-going ACRP study of airport-user survey methodology (Project ACRP 03-04: “Guidebook for Airport-User Survey Method- ology”) are in the range of approximately $30 to $50 per completed response. This suggests that developing the necessary survey data might cost somewhere in the range of $90,000 to $150,000. Assembling the corresponding transportation service data is also not trivial, although the local MPO may be able to provide some of these data in electronic format from the regional travel demand model network data. The amount of work involved in assembling the remaining data required will depend in part on how much information the airport au- thority already has available on the ground transportation pages of its website or in other files. Overall, this task might require between one and two person-months of effort and cost somewhere between $20,000 and $50,000 at typical consultant rates. Once the necessary data have been assem- bled, estimating, calibrating, and validating the model might require two to three person-months of effort, or perhaps be- tween $40,000 and $75,000. Therefore, a complete study, including an air passenger survey, might cost between $150,000 and $275,000. However, the largest part of this cost is performing the air passenger survey, which has other value for airport planning purposes and should be done periodically anyway. Whether such an investment is worthwhile depends on what the resulting model would be used for and the likely cost of making a bad decision. Certainly, in the case of evaluating

the feasibility of a major infrastructure investment such as an airport rail link or even an automated people-mover link to a nearby rail station that might cost several hundred million dollars or more, the cost of having a good modeling capabil- ity is trivial and would easily be justified by avoiding a poor decision that results in an increase in the cost of the project by even 1%. Indeed, having such a modeling capability is prob- ably a necessary requirement for obtaining environmental approval and funding. In the case of a smaller airport where the issues being addressed involve less money, the justifica- tion for such a modeling capability is less clear. Even so, an airport handing 5 million annual passengers, 20% of whom park at the airport for their trip duration, could be generating as much as $10 million per year in parking revenue. An increase in revenue of only 3% as a result of better pricing decisions could pay for the cost of developing the modeling capability in less than year. MODEL APPLICATION Once a mode choice model has been developed, to apply it to support planning and decision making it is necessary to have the ability to use it to analyze specific scenarios. This typi- cally involves a significant amount of data management and model configuration to define the scenarios to be analyzed. One approach is to use the model in conjunction with stan- dard transportation planning software or proprietary airport landside modeling software. These software tools are de- signed to facilitate the data management involved in model- ing transportation network flows and typically provide users with the flexibility to define the structure and coefficients of the mode choice model incorporated in the analysis. If such models are not available, or the models that are available do not have the necessary capabilities, it will then be necessary to develop custom software to apply the model to analyze any given scenario. Many discrete choice model estimation tools also provide the capability to apply a defined choice model to any suitably configured dataset of decision-maker characteristics and as- sociated properties of the choice alternatives to estimate the resulting choice probabilities. The result of this process is typically a table of the probability of choosing each alterna- tive for each decision maker (air passenger travel party or airport employee) in the dataset. Converting this table to a projection of passenger trips or vehicle trips by mode is then a matter of factoring up the results to correspond to the total airport activity for the period in question and applying the ap- propriate ratios of vehicle trips to travel party or employee trips. These estimated vehicle trips can in turn be allocated to the transportation network by segmenting the results by the trip origin zone and creating a zonal trip table. Although con- ceptually this is not difficult and the required calculations can generally be performed fairly easily using spreadsheet or database management software, because the process typi- cally has to be repeated multiple times to analyze a number 28 of scenarios it may be helpful to develop utility routines to perform the various steps. Model Application Considerations Although the process of applying an airport access mode choice model is technically fairly straightforward, there are a number of aspects that need to be carefully considered in de- veloping the required input data and interpreting the results. The first is that whereas a relatively small survey sample size (a few thousand respondents in the case of air passengers and perhaps even fewer for airport employees) may be ade- quate to estimate an airport access mode choice model, a much larger sample may be required for a given application of the model, depending on the issues of interest and the desired level of geographic resolution of the results. For example, a study to estimate the likely use of a proposed off-airport terminal in a particular location in the region served by the airport needs to have enough data points in the application dataset in the vicin- ity of the proposed terminal that the resulting estimate of the likely use of the facility is sufficiently accurate. To create a large enough application dataset, it may be necessary to de- velop a synthetic sample using Monte Carlo simulation meth- ods based on the distribution of travel party characteristics in the original survey. This is more accurate than simply dupli- cating the survey records, because that cannot create a record with different characteristics from those that appear in the sur- vey. Therefore, record duplication will simply create multiple records with identical characteristics, and combinations of traveler characteristics that do not appear in the original survey sample will never appear in the expanded sample no matter how large it is, whereas a synthetic sample will create records with characteristics that do not appear in the original sample. This will ensure that analysis zones that were thinly populated with survey respondents in the original sample will have a much more representative mix of traveler characteristics in the expanded sample and not be biased toward the characteristics of those travelers that happened to appear in the original sur- vey sample. Two other considerations relate to the application of a model estimated on data at one point in time to predict behavior in future years, as is typically done in planning studies. The first consideration is how to adjust travel times and costs for the various modes to correspond to future con- ditions. Highway travel times for future years should reflect any anticipated changes in highway congestion. Future costs are more problematical, because their effect on traveler decision making will depend on their value relative to the overall cost of living as well as changes in the income level of travelers over time. Typical practice is to consider that the cost and income variables in the model are expressed in real dollars (i.e., do not need to be increased to account for infla- tion). However, that is not to say that the cost of different modes will not change over time in real dollars. If general

29 income levels rise or fuel prices increase in real terms, then transportation operators will need to raise their prices to cover their higher costs of doing business and the cost of operating private vehicles will rise. Conversely, if trans- portation operators are able to achieve productivity gains or rising levels of air travel increase the traffic that they carry, allowing them to be more efficient, then they may be able to reduce their prices in real terms. Although these effects may tend to offset each other, the net effect is likely to vary by mode. Therefore, some thought should be given to assump- tions about future travel costs and not simply assume that they will remain unchanged in real terms. Historically, household incomes have increased in real terms over time, although over the past 30 years the distribu- tion of household incomes has also changed, with the incomes of higher-income households increasing faster than that of lower-income households (U.S. Census Bureau 2006, Table A-3). Whether the effect of household income is explicitly included in the mode choice model or implicitly included in the coefficients of cost terms, some thought should be given to how to incorporate future growth in household income in the application of the model for future years. This is not a triv- ial matter. From 1995 to 2005 the average household income in the United States increased by almost 11% in real terms (U.S. Census Bureau 2006, Table A-1). This is equivalent to a reduction in the relative costs of different modes by at least that amount, which could easily have a greater effect on mode use than the type of transportation system enhancement that mode choice models are used to evaluate. Because higher income households use air travel more than lower income households, the real increase in the average household income of air travelers and the corresponding effect on airport access mode choice is probably even greater. Pivot Point Analysis Rather than use a disaggregate mode choice model directly to predict the change in mode shares as a result of some change in transportation service levels, it is possible to use the relevant model coefficients and the existing mode shares to predict the change in mode share for a given change in the explanatory variables. This has come to be known as the incremental logit model or pivot point analysis (the predicted mode shares are considered to pivot about the existing mode shares) (Kumar 1980). Another way to think about this type of analysis is to use the model coefficients and existing mode shares to deter- mine the slope of the demand curve at the current values of the transportation service variables and mode use, and then calcu- late the change in mode share for a given movement along the demand curve corresponding to the change in the value of the transportation service variable (Meyer and Miller 1984). In the case of MNL models (or binary logit models), the incremental logit equations are fairly straightforward and are given by Kumar (1980). The derivation of the corresponding equations for a NL model is somewhat more involved, owing to the existence of the scaling parameters and the more com- plex structural form of the model, but the same approach can be applied. The advantages of this approach are two-fold. The first is that it eliminates any error in the predictions resulting from differences between the mode use predicted by the model at current values of the transportation service variables and the actual mode use. The second is that it only requires values for the model coefficients of the variables that change in value and not for the other variables or the ASCs. However, this is also a potential weakness of the approach. If the ASCs reflect some of the contribution of the transportation service vari- ables to the predicted mode shares (perhaps owing to speci- fication errors in the model or errors in measuring the trans- portation service levels), then those effects will be ignored in the analysis. Therefore, an alternative approach would use the full model, but apply the predicted change in mode use to the existing observed mode use. Demand Elasticity The concept of demand elasticity refers to the percent change in demand for a 1% change in some variable. There- fore, the demand elasticity with respect to the price for taxi use at an airport would express the percent change in taxi use for each percent change in fare levels. In the case of a MNL model with a linear utility function, it can be shown that the elasticity of the probability of choosing a particular mode with respect to a given variable in the utility function is given by: εik = αki • Xki • (1 − Pi) where εik is the elasticity of the probability of choosing mode i with respect to changes in the value of the explana- tory variable Xk for mode i (Xki), αki is the coefficient of Xk in the utility function for mode i, and Pi is the current prob- ability of choosing mode i. In simple terms, this is saying that the elasticity is given by the product of the coefficient of the variable, the current value of the variable, and the probability of not choosing the mode. As the probability of choosing a particular mode increases, the elasticity of the probability of choosing that mode with respect to any vari- able becomes less. However, this is of limited use because it is clear from the previous equation that the elasticity of demand varies with the value of the variable in question for every air party and with the probability of each air party selecting the mode in question at current values of all the explanatory variables. Therefore, elasticity is not a constant property of a given mode or a given situation, but varies with the values of the transportation service variables and with the market shares of the different modes. Nonetheless, for any given situation the

elasticity of demand for a given mode with respect to a given service variable can be calculated numerically, and this may be a useful thing to do to give planners and managers an eas- ily understood tool to make quick assessments of the likely effect of any proposed change. The important caveat to en- sure this is clearly understood is that any given elasticity 30 value is only valid for the particular situation for which it has been calculated and will change as the situation changes. In particular, if changes occur in other modes the elasticity for the mode in question will change because changes in other modes will change the probability of choosing the mode for which the elasticity has been calculated.