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.
43 Although the details of the different air passenger airport ac- cess mode choice models discussed in the literature review in chapter three vary widely, it is clear that a standard of best practice has evolved, although by no means is it always fol- lowed. This chapter summarizes that standard and identifies aspects where further improvement is needed. MODEL APPLICATION The airport access mode choice models reported in the liter- ature fall into four broad categories. The first category con- sists of academic studies where the primary objective is to develop a model that explains air passenger access travel behavior. These studies address such issues as the appropri- ate functional form of the model to use, variables to include, and market segmentation issues. Some recent studies have also explored alternative approaches to the traditional MNL and NL choice models. The second category consists of models that have been developed in support of specific airport ground access plan- ning studies, such as an evaluation of the feasibility of con- structing an APM link between the airport and a nearby rail station or of extending an urban or intercity rail system to an airport. These models are often less detailed than the first cat- egory of models, possibly because the primary focus of the study is not on model development, and the time and resources available for model development are limited. The third category consists of airport access mode choice components of models developed to explain or predict air passenger airport choice. These too often tend to be fairly simple in terms of the number of modes included in the model, the variables used, and the structure of the choice process, in part because the primary focus of the study is on airport choice rather than ground access mode choice, and many of the more detailed aspects affecting ground access mode choice (such as differences in fare and travel time for different public transportation services) have only a second- order effect on airport choice. Finally, the fourth category of airport access mode choice models consists of the components of regional transportation planning models that are used to generate estimates of vehi- cle trips to and from the airports in the region. In many cases these are not true airport access models, in that they are the result of a specific effort to model these trips, but rather are simply the application of models that have been developed to predict urban travel behavior in general to account for airport trips. Given the differences between the characteristics of airport trips and other types of regional travel, it can be ex- pected that the application of general regional travel models to predict airport trips will not produce very good results, and indeed there is a growing interest in developing airport- specific sub-models to account for these trips within the over- all regional transportation modeling framework. TECHNICAL APPROACH Current best practice uses NL choice models with separate coefficients (and possibly including different variables) for at least four market segments: ⢠Resident business trips, ⢠Resident non-business trips, ⢠Non-resident business trips, and ⢠Non-resident non-business trips. The modes available for resident and non-resident trips will generally be different, because non-residents do not have the option of parking a private vehicle at the airport (indeed this would make no sense because their visit to the region begins at the airport and they return to the airport at the end of their stay). On the other hand, many non-residents rent a car on their arrival at the airport to provide local transportation during their stay in the region, whereas most residents already have a car that they can use for airport trips and do not con- sider renting a car for their airport access or egress trip. Revealed or Stated Preference As discussed in chapter two, there are two different ap- proaches to assembling the necessary data to develop airport ground access mode choice models: revealed preference and stated preference surveys. The majority of past airport access mode choice models have been developed using revealed preference techniques. However, in a number of recent stud- ies, including ridership studies for proposed airport rail links in Chicago and Toronto, demand analysis for the planned Airport MAX Red Line extension to Portland International Airport, and a feasibility evaluation of a proposed APM link to a light rail stop at San José International Airport, stated CHAPTER FIVE STATE OF PRACTICE OF AIR PASSENGER MODE CHOICE MODELS
preference studies have been done to develop model coeffi- cients for the planned modes or services that do not yet exist, and thus are not reflected in surveys of actual mode choice. Functional Form of the Model Although the MNL model continues to be used in airport access mode choice models, it is becoming accepted that the NL model is generally a more appropriate formulation. The airport ground transportation system consists of a large number of different modes and sub-modes, which are likely to be viewed by a traveler as having very different substi- tutability. Therefore, a traveler trying to decide between parking in a close-in parking lot or a remote parking lot with a shuttle bus connection to the airport terminal is not likely to care very much about differences between bus fares and taxi rates. Conversely, a traveler trying to decide between the lower cost but longer travel time of taking transit to the air- port and the higher cost but greater convenience of a taxi or shared-ride van is unlikely to care very much about differ- ences in parking rates between on-airport and off-airport lots. It is precisely the effect of these different substitution rates that the NL model is designed to reflect. Because the private vehicle generally accounts for the majority of airport ground access trips at most airports, at a minimum an air passenger mode choice model should nest the private vehicle choice alternatives for residents of the region (drop off and park for the duration of the trip). Although no generally accepted practice has yet emerged for how to structure the nests of a NL model, this is largely determined by the characteristics of the different modes be- cause the primary purpose of using a NL model is to allow higher rates of substitution between modes that have similar characteristics. Therefore, it would appear logical to group private vehicle modes in one nest, with different parking op- tions as a second-level nest, group exclusive ride on-demand modes (taxi and limousine) together in a second first-level nest, and group shared-ride scheduled modes (public transit, scheduled airport bus) together in a third first-level nest, pos- sibly with different transit options (e.g., rail and bus) as a second-level nest. It is not clear where the door-to-door shared-ride van would best fit in this structure, as a separate mode at the top level, in the on-demand nest with taxi and limousine, or in the shared-ride nest with the scheduled modes. This may be an issue to resolve empirically by ex- ploring which option gives the best fit to the data. Alternative access modes to scheduled services can also be included as lower-level nests to each mode. Rental car and hotel shuttle use by non-residents is best modeled outside this choice process, because use of both modes is determined by factors that are largely independent of the service levels of other modes. Rental car use is often determined by local travel needs other than the airport egress and access trip. Therefore, visitors to the region may rent a 44 car even if they are staying at a nearby hotel that has a free shuttle to and from the airport. The form of the utility functions for each choice alternative will generally be a linear combination of explanatory vari- ables with their associated coefficients. However, some vari- ables are best entered in the utility function as an inverse or ratio. For example, the service headway of scheduled modes, which is a direct measure of average waiting time, is the inverse of the service frequency. The effect of household income may best be entered in the utility function by express- ing direct travel costs as a ratio of the cost to some function of the per capita or total household income. Therefore, higher- income travelers will be less influenced by cost than lower- income travelers. Some model developers have attempted to reflect nonlin- earity in the effect of some variables (such as service frequency or income) through the use of the logarithm of the variable value in the utility function. However, there are the- oretical problems with this approach and great care is needed when introducing logarithms into the specification of terms in the utility function. Although a logarithmic transformation may approximate some other function (such as the inverse) over part of the data range (with appropriate coefficients), it can differ significantly at lower and higher values. However, it can be precisely at these values that the effect of the vari- able on traveler behavior is most important. Therefore, it is critical to consider whether the logarithmic transformation of the variable best reflects the nonlinear effect that is desired compared with some other function. Survey Data Used in Model Development Although air passenger or airport employee surveys are some- times performed primarily (or entirely) to support develop- ment of airport access mode choice models, more commonly model development makes use of surveys conducted for other purposes. However, because such models, particularly air passenger models, require detailed information on air traveler characteristics and factors likely to influence their choice of access mode, it is highly desirable that if a survey will be (or even might be) used later to support development of an airport access mode choice model, the survey questions are reviewed by modeling specialists before the survey is performed to ensure that key questions are not omitted that would later compromise the ability to develop a reasonable model. Among the factors that are often omitted from air passen- ger and employee surveys, but are critical for model devel- opment, is information on household income and household composition. It is self-evident that traveler decisions between alternative access modes that offer a trade-off between travel time and cost will be influenced by their perceived value of time (even if they do not think of the decision in those terms). Because one cannot meaningfully ask survey respondents to
45 state their value of time, the functional form of the model needs to include some measure of household income so that the values of time implied by the model coefficients will vary by income. However, as noted elsewhere in this report, will- ingness to pay to save time is influenced not only by the total household income, but also by how many people that income has to support. Although these questions are often omitted as intrusive or even of little relevance for other purposes to which the survey data will be put, experience has indicated that reasonable responses can be obtained if the questions are asked in the right way. Another factor that affects airport access travel decisions is the time that a traveler begins a trip to or from the airport, both in terms of the time of day and (in the case of departing air pas- sengers) the time remaining before their scheduled flight departure. This information can also be useful in the case of shared-ride van services to determine how much lead time the operator required to schedule other pickups, because this can affect the overall access time experienced by the traveler. Similarly, with scheduled modes, this information can provide an indication of how much margin of error travelers allow in their travel plans to avoid missing scheduled services. Transportation Service Data In addition to data on the characteristics of the airport travel- ers, mode choice model development and application requires data on the transportation service levels of the dif- ferent modes. These will include travel times, fares or costs, service frequencies of scheduled modes, any walking dis- tances involved, and network connectivity issues such as the number and type of anticipated transfers required between different services. The extent to which different explanatory variables have been included in various existing models is discussed later in this chapter. Generally, these data are obtained from transportation net- work models or from published information by transportation providers, rather than by asking survey respondents. There are two reasons for this. The first is that transportation service- level data are required for all the modes included in the model and it would be too cumbersome to ask each survey respon- dent for their perception of the value of these service levels for every mode, and indeed they may have no idea of the details of those modes that they did not use. The second reason is that for any application of the models, different values of at least some of the transportation service levels are typically assumed, whether to reflect expected future values or to model the effect of changing the ground access system in some way, and therefore the values used in the model appli- cation can be established in a consistent way to those used in the model development. To the extent that the traveler perception of the service lev- els of the various modes differs from the values used in model estimation, it is implicitly assumed in the modeling that the estimated values of the model coefficients convert the values of the service levels used in the modeling to the values per- ceived by the travelers. Therefore, the model coefficients serve two purposes: to adjust the values of the service levels used in the modeling to the values perceived by the travelers and to express those perceived values in terms of their contri- bution to the overall perceived utility of each alternative. Because it is highly likely that many airport travelers have only limited information about many of the alternatives in- cluded in the modeling, and the perception of the transporta- tion service levels relative to the values used in the modeling may vary widely across the respondents to surveys used to develop the models, this is likely to have a significant impact on the overall ability of the model to explain the observed choices. Unfortunately, the question of how well airport trav- eler perceptions of the service levels of the different modes conform to the values used in modeling is one that has been largely ignored in the literature on modeling practice. Because it is necessary to determine the transportation service levels faced by each travel party included in the mod- eling process, it is normal practice to define a system of analysis zones and assign the ground access trip origin of each respondent in an airport traveler survey to the appropri- ate analysis zone. To provide a reasonable level of precision for the ground access service levels, in particular travel times, these analysis zones cannot be too large. A fairly com- mon practice is to use the TAZs defined for the regional travel demand model developed for general transportation planning in the region. The relevant regional transportation planning agency will generally be able to provide the highway travel times and transit service levels (travel times and fares) from each TAZ to the airport from the regional transportation network model. The extent to which these travel times can be expressed in terms of their components (e.g., walking time, waiting time, and in-vehicle time) or different values can be provided for different times of day or different days of the week is likely to vary from agency to agency. In general, regional travel demand modeling tends to be based on average weekday con- ditions and distinguish between peak and off-peak travel times. However, it is not uncommon to only model one peak period (typically the morning peak). To develop the appropriate transportation service data for the other airport access modes, such as shared-ride van ser- vices, taxis, and any scheduled airport bus services, it will be necessary to obtain the relevant service information (travel times, fares, and schedules) from the transportation pro- viders and convert these to the analysis zone system being used. Given the number of analysis zones typically used, this can be a significant task. This task can become more complicated if some time has passed since the airport trav- eler survey that is being used to develop the model was per- formed, because transportation providers may have changed
their service levels in the interval and do not always main- tain good records of past service levels. One potentially use- ful source of information on transportation provider service levels may be the ground transportation section of the air- port website, where this provides the relevant level of de- tail, or websites of the transportation providers. Ideally, this information would be archived whenever an airport trav- eler survey is done, so that it can be referred to later if the data changes; however, unfortunately this is not often done. An historical archive of some website content is available through the Internet Archive, a nonprofit organization that maintains an online web archive called the Wayback Ma- chine (http://www.archive.org). However, this only provides the content of web pages and cannot recover information in underlying databases that may have been accessed by web pages. Travel times by services such as taxi and limousine will generally be the same as other highway modes. Taxi fares that are distance-based can be calculated from the meter rates and highway distances obtained from the regional travel demand model network data. Shared-ride van travel times will typically depend on the number of other parties that have to be picked up after picking up the party in question and the additional travel involved. Because this will vary from air party to air party, it will be necessary to make some assump- tions about the average travel times experienced from a given analysis zone. In the case of fixed-route services, such as scheduled airport bus or rail modes, consideration needs to be given to how airport travelers will access the relevant stop or station where they will board the service. In the case of nested mode choice models, the access mode to the fixed- route service may be treated as part of the choice process and will require separate travel times and costs for the different access alternatives (e.g., walk, drop off by private vehicle, taxi, and public transit). Where this secondary access choice process is not explicitly modeled, assumptions will need to be made about the access time and cost to be included in the utility function of the primary mode. MODES INCLUDED IN MODEL The modes to be included in an airport access model are largely determined by the modes available at the airport, although not all the modes available at an airport may be in- cluded in a given model. In particular, some modes may be combined in the model or excluded from the analysis with users assumed to be captive to the mode. The modes included in nine recent airport access mode choice models described in chapter three are shown in Table 2, together with the type of model used and the nature of the data from which the model was estimated. Rental car and taxi are generally available at all airports. Most of the larger airports will also have some form of lim- ousine (black car) service and some door-to-door shared-ride 46 van service. Public transit service is likely to be very location- specific. Representation of the use of private vehicles needs to distinguish between air passengers who are dropped off at the airport and those who park the vehicle at or near the airport for the duration of the air trip, because there are sig- nificantly different costs involved and being dropped off at the airport generates additional vehicle travel for the return trip. Whether or not those dropping off or picking up air passengers park the vehicle for a time at the airport is less important and generally not addressed in mode choice mod- els. In any event, the decision of whether or not to park is likely to be determined by factors not typically included in ground access mode choice models, such as whether the ve- hicle driver arrives at the airport earlier than intended. Whether or not to model the choice of parking location by those parking for the duration of their air trip is a more diffi- cult question. Most airports offer a choice of parking facili- ties, with less expensive options often involving a shuttle bus ride to reach the terminal. In many cases there are also pri- vately operated off-airport parking lots that provide a shuttle bus service to the airport. These various facilities typically charge different daily rates. Thus, the choice of which facil- ity to use affects the cost of the ground access trip as well as the travel time involved in the access trip. Where these choices are not explicitly included in the model, the travel time and cost assumed for parking for the duration of the air trip needs to reflect the proportional use of the different park- ing options, which will vary with the air trip duration. The degree to which it is necessary to identify specific public transportation services depends to an extent on the purpose of the model. Because in general public transit tends to attract a fairly small mode share at most U.S. airports, except at those with an extensive fixed-rail system serving the airport (e.g., the Washington Metro at Ronald Reagan National Airport), it may be sufficient to consider this a sin- gle mode and assume that travelers will use the best path through the network. However, where the model will be used to evaluate specific services or analyze the market for the introduction of a new mode, it will then be necessary to iden- tify the alternative services in more detail so that ridership on specific services can be calculated. The model developed for Boston Logan International Airport discussed earlier in- cluded rail transit, the Logan Express scheduled airport bus service from off-airport terminals in the region, and a Water Shuttle ferry that ran between the airport and downtown Boston. However, the model does not distinguish between the different Logan Express terminals. In contrast, the model developed for the SERAS study in the United Kingdom in- cludes an explicit representation of several different types of rail services to be able to model how changes in specific rail services affect not only the overall mode share of rail but how those using rail would choose between the different services. Two important issues to be addressed are how to incorpo- rate rental car use and the use of a courtesy shuttle bus service
47 Airport or Study ATL BOS CHI MIA OAK PDX SJC YYZ UK Year Model Developed 2002 1996 2004 1995 2001 1997 2002 2002 2002 Model Structure Binomial logit Multinomial logit Nested logit Estimation Data RP RP RP/SP RP RP RP/SP RP/SP RP/SP RP Modes in Model Private vehicleâdrop off Private vehicleâpark Private vehicleâpark ST Private vehicleâpark LT Private vehicleâpark OA Private vehicleâcombined Rental car Taxi Limousine Shared-ride van Scheduled airport bus Express bus from OAT Transit (all services) Transit (rail) Transit (bus) Intercity coach Airport express train Intercity rail/coach links Ferry (water shuttle) Hotel shuttle bus/van Charter coach Other private modes Other public modes Model: ATL = HartsfieldâJackson Atlanta International Airport (Travel Demand . . . 2005). BOS = Boston Logan International Airport (Harrington 2003). CHI = Chicago OâHare International Airport, Chicago Midway Airport (Resource Systems Group 2004). MIA = Miami International Airport (ICF Kaiser Engineers 1995). OAK = Oakland International Airport (BARTâOakland . . . 2002). PDX = Portland International Airport (PDX Ground Access . . . 1998). SJC = San José International Airport (Dowling Associates 2002). YYZ = Toronto Lester B. Pearson International Airport (Halcrow Group 2002a). UK = United Kingdom SERAS study (Halcrow Group 2002b). Notes: RP = revealed preference data; SP = stated preference data; ST = short-term (i.e., associated with air passenger drop off); LT = long-term (i.e., park on airport for duration of air trip); OA = off-airport; OAT = off-airport terminal (e.g., Logan Express service at Boston)âincludes inter-airport transfer coach (London). TABLE 2 MODEL TYPE AND MODES INCLUDED IN RECENT AIRPORT ACCESS MODE CHOICE MODELS
from nearby hotels in the model. Visitors to the region may rent a car for local travel quite independently of their airport egress and access travel decisions. Those staying at hotels with courtesy shuttle bus service are likely to use that service, because it is usually free, unless they rent a car for other rea- sons. Therefore, it would appear appropriate to model rental car use by visitors first and then assign visitors staying at hotels with courtesy bus service to that mode if they are not assigned to a rental car. In some cases residents of the region may also rent cars for their airport access trip, particularly if they live some dis- tance from the airport and will be away for some time on their air trip. Local rental car agencies may not charge a drop- off penalty for returning the car to the airport, and two one- way rentals may be much less expensive than parking their own car at the airport for the duration of their trip or using other modes. Therefore, a rental car can be included in the regular choice set for resident travelers. In general, it will have a very low probability of being chosen owing to the high fixed cost, but will become more attractive for the type of trip described earlier. EXPLANATORY VARIABLES Existing models differ widely in terms of the explanatory variables used in the utility functions, as well as how those variables are defined. The variables included in the nine air- port access mode choice models discussed in the previous section are shown in Table 3. All existing models include some measures of travel time and cost. However, the extent to which travel time is sep- arated into its components (walk time, waiting time, in-vehicle travel time, and time involved in private vehicle access to pub- lic transportation services) varies. The model developed for Boston Logan International Airport estimated separate coeffi- cients for in-vehicle time, walk and wait time (combined as out-of-vehicle travel time), and automobile access time to pub- lic transportation. The models developed for Oakland Interna- tional Airport and San José International Airport to evaluate proposed APM links (BARTâOakland . . . 2002; Dowling Associates 2002) used separate coefficients for walk distance, waiting time, and in-vehicle time in private vehicles, buses, and light rail, although these were adopted directly from Harvey (1988). The use of separate coefficients for in-vehicle time in the different modes was intended to capture differences in the perceived comfort and convenience of rail compared with bus, as well as between private vehicle and transit. The models developed to estimate ridership on the proposed Chicago Airport Express train (Resource Systems Group 2004) used a weighted travel time, with the weight for access time, waiting time, and transfer time adjusted incrementally to obtain the best model fit. Both the Boston Logan model and the U.K. SERAS model included fixed penalties for transfers on the public transit system. The coefficient of the variable for the 48 number of transfers in both models can be thought of as measuring the perceived disutility of a transfer in terms of the equivalent additional riding time on a non-stop service. The SERAS model distinguished between different types of transfer, such as cross-platform or those involving a level change. In contrast, the model developed for the Portland ground access study (Bowman 1997; Cambridge Systematics 1998; PDX Ground Access . . . 1998) combined the different com- ponents of travel time into a single variable for total travel time without weighting any of the components. The model developed by the Atlanta Regional Commission included sep- arate coefficients for highway travel time, transit in-vehicle time, waiting time, and walk time, although the values of these coefficients were adopted from other models and not es- timated for the Atlanta region. Similarly, the model developed for the Miami Intermodal Connector project distinguished between in-vehicle time on all modes and the combination of waiting and terminal time (the time required to access the zone centroid in the transportation network), although this model also adopted coefficients from another model for a dif- ferent region. Cost and Income Not unreasonably, most models combine costs into a single variable, although the Boston Logan model defined different cost variables depending on the income of the traveler and whether the employer was paying the cost of a business trip. However, this was simply a way to obtain different cost coefficients for each case. Several models have divided the cost by a function of the household income, although there is no consistency in what function was used. The Boston Logan model (Harrington et al. 1996; Harring- ton 2003) divided respondents who paid their own travel expenses into two income categories (low and high) and esti- mated separate travel cost coefficients for each income cate- gory as well as for a third category of respondents whose travel expenses were paid by their employer. The Portland ground access study model (Bowman 1997; Cambridge Systematics 1998; PDX Ground Access . . . 1998) divided the travel costs incurred by air passengers by the natural logarithm of the household income, although this was not done for the automo- bile operating cost for drop-off trips. The Oakland Airport Connector model (BARTâOakland . . . 2002) and the San José International Airport Model (Dowling Associates 2002) divided the travel costs for per- sonal trips by the household income raised to the power 1.5, but did not do this for business trips. Although each of these approaches will ensure that higher-income travelers (at least those paying their own expenses in the case of the Boston Logan model and those
49 Airport or Study ATL BOS CHI MIA OAK PDX SJC YYZ UK Continuous Variables Off-peak highway time Travel time (private vehicle) In-vehicle time (all modes) In-vehicle time (transit) In-vehicle time (rail transit) In-vehicle time (bus transit) In-vehicle time + walk time Out-of-vehicle timea Total travel time (all modes) Weighted travel timeb Wait time + terminal timec Wait time/headway Walk time Walk distance Auto access time to transit Number of transfers Interchange penalties Drop-off driver time Driving distance Travel cost (all modes) Private vehicle cost Drop off vehicle operating cost Transit fare Taxi fare Household income Dummy variables Air party size Baggage check-in Em ployer pays cost Flights/year Household income Luggage/checked bags Non-residence trip origin One-person travel party Use of intermediate station Model: ATL = HartsfieldâJackson Atlanta International Airport (Travel Demand . . . 2005). BOS = Boston Logan International Airport (Harrington 2003). CHI = Chicago OâHare International Airport, Chicago Midway Airport (Resource Systems Group 2004). MIA = Miami International Airport (ICF Kaiser Engineers 1995). OAK = Oakland International Airport (BARTâOakland . . . 2002). PDX = Portland International Airport (PDX Ground Access . . . 1998). SJC = San José International Airport (Dowling Associates 2002). YYZ = Toronto Lester B. Pearson International Airport (Halcrow Group 2002a). UK = United Kingdom SERAS study (Halcrow Group 2002b). a Out-of-vehicle time on Boston model combined waiting time and walking time (access, transfer, and egress) as appropriate for the mode. b Weighted travel time in Chicago model combined in-vehicle time and egress time with a weighted sum of access time, transfer time, and waiting time. c Terminal time in Miami model used the regional travel demand model zonal terminal times that represent the additional travel time required to reach the TAZ centroid from the actual trip origin. TABLE 3 EXPLANATORY VARIABLES INCLUDED IN RECENT AIRPORT ACCESS MODE CHOICE MODELS
making personal trips in the case of the San José Interna- tional Airport model) will be less sensitive to travel cost than lower-income travelers, clearly the effect on travelers of a given income level varies significantly. It is most unlikely that they all correctly reflect sensitivity to income, and those that simply classify travelers into two income categories obviously can only approximate the effect. Indeed, common sense suggests that total household income is only part of the story and the effect of household income must depend in part on the size of the household. There is clearly a huge differ- ence between a single person making $100,000 per year and a family of six making the same annual income. A related issue that is generally ignored is the difference of purchasing power of a given income level between resi- dents of the area in question and visitors. Perceived values of time are likely to depend on the level of discretionary income of individual travelers, which in turn is affected by the cost of living in their home region rather than that of the region where the airport is located. Therefore, it can be expected that visitors to a region from another region where the cost of living is significantly different are likely to have different perceived values of time for a given level of per capita house- hold income than residents of the region where the airport is located. This issue has important implications for the trans- ferability of mode choice models from one region to another, as discussed further in chapter seven. In view of the lack of any well-established practice re- garding how best to incorporate traveler income into airport ground access mode choice models, this would seem to be a useful topic for future research. Travel Time Components As discussed earlier, it is widely recognized that the different components of travel time are likely to have a different dis- utility per minute. This is particularly true for waiting time and walking time (although this may be measured by distance rather than time). There are two approaches to addressing this. The first is to weight the various components of travel time differently before combining them into a single travel time variable for the purpose of model estimation. The sec- ond is to estimate separate coefficients for each component. The latter is preferable, because assumptions for the relative weights of different travel time components based on the experience of other models may not be appropriate for the particular situation being modeled. Some models estimate separate coefficients for travel time on different modes. This has the advantage over the use of a single travel time variable for all modes that if the relative perceived disutility of differ- ent modes varies by length of trip, this will be reflected in the estimated values of the coefficients. Only relying on ASCs to reflect differences in the perceived disutility of different modes implicitly assumes that any differences are indepen- dent of the length of the trip, which is highly unlikely. 50 One issue that arises with scheduled modes is how to han- dle the waiting time involved. There is an extensive literature on how travelers value the disutility of waiting time relative to travel time, particularly in the context of transfers between different transportation modes or services (e.g., Mohring et al. 1987; Moreau 1992; Small 1992; Chang and Hsu 2001; Hensher 2001; Lam and Small 2001). The general consensus is that waiting time has a higher disutility (higher perceived value of time savings) than in-vehicle travel time, although the ratio between the two appears to vary with circumstances, typically in the range between 1.5 and 2.5. Some authors have argued that because the schedule is published and people will arrive a certain amount of time be- fore departure, waiting time is effectively independent of ser- vice headway. Others have used half the service headway as the average waiting time. This is commonly referred to as schedule delay, and it is argued that although the traveler may only wait at the bus stop or station for few minutes, they have a time that they would prefer to have departed and the difference between that time and the time that they actually depart is part of the disutility of travel, even if they are actu- ally doing something else. In the case of airport access and egress trips, this argument is quite persuasive. Airport trav- elers will generally take an earlier bus or train if one is not leaving at the time they would prefer to have departed, be- cause taking a later service may cause them to miss their flight or, in the case of airport employees, be late for work. However, arriving early at the airport has relatively little value, because they are likely to just spend longer waiting in the gate lounge or for their shift to start. Similarly, in the re- verse direction, any time spent waiting at the airport for the next departure from the airport is generally not usable for other activities apart perhaps from reading. The changes in the amount of time that air travelers have to allow for airport security screening since the events of 9/11 adds a further complication to how passengers may per- ceive the disutility of arriving at the airport earlier than they would prefer. On the one hand, any additional time provides something of a buffer against unexpected delays in clearing security. On the other hand, if passengers are already con- cerned about the amount of time that they have to spend at the airport because of the need to allow enough time for security delays, they may be even more intolerant of having to arrive at the airport earlier than they would prefer owing to the waiting time while traveling to the airport. A related issue arises from the schedule dependability when a connection is involved. If the trip to the airport involves connecting between two (or more) scheduled ser- vices, then the traveler needs to consider the possibility that if one service runs late, they may miss their planned connec- tion and have to wait for the next departure on the subsequent segments of the trip. Therefore, the perceived schedule relia- bility may play an even more important role than the actual waiting time involved according to the schedule. Ironically,
51 the less the waiting time involved in making connections according to the published schedule, the higher the probabil- ity that the traveler may miss the connection, unless the out- bound services at the connecting point wait for the inbound services to arrive. This issue is not unique to airport ground access trips and arises with all scheduled transportation services. The issue of travel time dependability is not restricted to waiting time. One advantage of rail access modes is that the travel times involved are not subject to the variability that can affect highway travel times resulting from accidents or unex- pected congestion. Thus, it might be expected that the per- ceived disutility of each minute of travel time on a rail mode might be different from that of each minute of travel time on highway modes. Because the disutility of time spent on dif- ferent modes is also affected by the level of comfort offered by the mode and other factors, such as perceptions of personal safety, it is not obvious whether on balance the higher relia- bility of rail modes could be offset by other factors. The use of separate coefficients for travel time on each mode allows the model estimation process to determine the relative per- ceived disutility. However, the more model coefficients that have to be estimated, the larger the dataset needs to be to ob- tain statistically significant estimates of the coefficients. Other Variables In addition to travel time and cost, several models have at- tempted to include other variables believed to influence air passenger ground access mode choice, including the amount of luggage that an air party has, familiarity with the airport, and whether the traveler or someone else is paying for the trip. Some models, such as the Boston Logan model (Harrington et al. 1996; Harrington 2003), have also included air party size as a separate variable. However, to the extent that some costs, such as transit fares, vary with the number of people in the travel party, it is generally better to address the effect of air party size through the calculation of the travel costs of the different modes, rather than try to handle this through a separate variable, which simply adds a fixed amount to the disutility of the affected modes irrespective of the actual value of the travel costs involved. Those models that have included variables reflecting the amount of checked baggage, such as the Boston Logan model (Harrington et al. 1996; Harrington 2003), the Toronto Air Rail Link model (Halcrow Group 2002a), and the earlier work by Harvey (1988) on which the models for Oakland International Airport and San José International Airport were based, have generally found this to be a significant factor in explaining model choice. It would therefore seem desirable to include this in any airport access mode choice model, par- ticularly if the model is to be used to study proposed services that include off-airport baggage check-in, where the attrac- tiveness of that capability will depend on the amount of bag- gage that an air party has. However, some thought needs to be given to the appropriate specification of a baggage vari- able in the post-9/11 environment, in which there are strict limits on carry-on bags. As a result, passengers may well check a bag, even though it is not large enough to present any limitation to their use of public transportation. The wide- spread adoption of wheeled bags over the past decade has also reduced the difficulty of carrying bags on public trans- portation services. Not surprisingly, the issue of whether any of the travel costs will be paid by someone other than the traveler was found to be a significant factor in mode choice in the Boston Logan model (Harrington et al. 1996; Harrington 2003), the only model where this has been included. Although this is not a factor that can be independently fore- cast and simply applies to a particular subset of respondents in a given survey sample (generally those on business trips), its inclusion in a model improves the model fit to the data and helps reduce any bias in the coefficients of other variables that might result from not considering this issue. Similarly, the Boston Logan model found that familiarity with the airport and available ground transportation ser- vices, as measured by the number of flights from the airport in the past year by the survey respondents, was a significant factor in explaining mode choice of residents of the region. Although frequency of use of an airport may be a useful indicator of familiarity with alternative ground transporta- tion services, it is likely to be a fairly inaccurate one. Some frequent travelers may make no effort to learn about the al- ternative transportation options, whereas some infrequent (or even first-time) users of the airport may have researched their different options or sought advice from more frequent users. Therefore, ideally, air passenger surveys would attempt to obtain a better indication of the familiarity of respondents with alternative ground transportation ser- vices. However, it is far from clear how to do this effec- tively and this is an important aspect for further research, because it is self-evident that travelers will not choose ser- vices that they do not know exist. Implied Values of Time The ratio of the travel time coefficient to the cost coefficient can be interpreted as the implied value of time (strictly of travel time savings). These values for eight of the mode choice models described in Tables 2 and 3 and discussed in more detail in Appendix D (web only version) are summa- rized in Table 4. The implied travel time values for the U.K. SERAS model are not included in the table because they are expressed in pounds and were derived using a different methodology; therefore, they are not directly comparable to the U.S. experience. Because separate coefficients were esti- mated for the two Chicago airports in the Chicago Airport Express study, Table 4 presents the implied values of time for each airport separately.
52 Airport or Study ATL BOS ORD MD W MIA OAK PDX SJC YYZ Year of Cost Data a 1993 2003 2003 b c 1996 c 2002 Travel Times ($/hour) d e f Highway time g h h i Resident business trips 15 11 33 63 78 15 19 15 53 Resident non-business trips 13 17 25 22 78 16 29 10 29 Non-resident business trips 16 40 33 63 78 15 19 15 71 Non-resident non-business trips 12 13 25 22 78 16 30 10 34 Transit in-vehicle time j k k Resident business trips 11 26 33 63 78 11 19 11 53 Resident non-business trips 9 7 25 22 78 12 29 7 29 Non-resident business trips 12 15 33 63 78 11 19 11 71 Non-resident non-business trips 9 9 25 22 78 12 30 7 34 Travel time (other cases) l m m n o n Resident business trips 22 92 82 20 24 20 Resident non-business trips 38 55 57 19 37 12 Non-resident business trips 40 92 82 20 24 19 Non-resident non-business trips 13 55 57 19 39 11 Model: ATL = HartsfieldâJackson Atlanta International Airport (Travel Demand . . . 2005). BOS = Boston Logan International Airport (Harrington 2003). ORD = Chicago OâHare International Airport (Resource Systems Group 2004). MDW = Chicago Midway Airport (Resource Systems Group 2004). MIA = Miami International Airport (ICF Kaiser Engineers 1995). OAK = Oakland International Airport (BARTâOakland . . . 2002). PDX = Portland International Airport (PDX Ground Access . . . 1998). SJC = San José International Airport (Dowling Associates 2002). YYZ = Toronto Lester B. Pearson International Airport (Halcrow Group 2002a). a Coefficients for time and cost adopted from other models (date unspecified). b Coefficients for time and cost adopted from another model (date unspecified). c Coefficients for time and cost adopted from Harvey (1988), 1995 cost data. d For household income (only applies to non-business trips) of $75,000/year. e For household income (only applies to non-business trips) of $55,000/year. f Canadian dollars. g Private vehicle drop off, costs not reimbursed, low income (undefined). h Weighted travel time, low household income (less than $100,000/year). i For household income of $50,000/year. j Costs not reimbursed. k Rail transit. l Private vehicle drop off, costs not reimbursed, high income (undefined). m Weighted travel time, all modes, high household income ($100,000/year or more). n Bus transit. o For household income of $150,000/year, all modes. TABLE 4 IMPLIED VALUES OF TIME FROM RECENT AIRPORT ACCESS MODE CHOICE MODELS
53 In addition to the implied values of time, Table 4 shows the year for which the cost data were obtained where this is known. Because the implied value of time is obtained from the ratio of the travel time coefficient to the cost coefficient, it is expressed in dollars of the year for which the cost data were assembled from which the model was estimated. In cases where the travel time and cost coefficients were adopted directly from other models, the value of time is expressed in dollars of the year for which the cost data used in estimating those models were assembled. For those models where household income was included in the cost term of the utility function or different cost coef- ficients were estimated for different income categories, the implied value of travel time is a function of income (or varies by income category). In those cases where the values of time vary with income the values shown in Table 4 have been cal- culated for specific household income levels. Because the different models were estimated using cost data for different years, the values of time are not directly comparable. Furthermore, for those models where the implied value of time is a function of household income, the values shown in Table 4 have not been calculated for comparable household income levels. Even so, it is clear from Table 4 that the variation in implied value of time across the different models is much greater than can be explained by either the different year for which the cost data were obtained or differ- ences in the household income used to calculate the implied value of time. Given the differences in model structure and explanatory variables included in the model, this is hardly sur- prising. Including other terms than travel time and cost in the utility functions, or changing the way that travel time is de- fined in the model, will change the estimated values of the travel time and cost coefficients and hence the implied values of time. However, this has important implications for the abil- ity of the models to predict how airport travelers will respond to changes in travel time and cost of airport access modes. If other terms in the model utility functions (such as dummy variables or ASCs) are accounting for part of the perceived disutility of travel time or cost for a particular mode, the con- tribution of those terms to the overall disutility of the mode will not change when the values of travel time or cost are changed in a given application of the model, resulting in an in- correct prediction of the impact of the change on perceived disutility and hence on predicted mode use. MARKET SEGMENTATION As noted earlier, although some early models considered more limited market segmentation, it is now widely recog- nized that at a minimum separate airport access mode choice models should be developed for at least four market segments corresponding to residents of the region being modeled ver- sus visitors to the region, with each segment further split into business and non-business (or personal) trips. Almost all of the recent models developed in the United States have used this approach, although the model developed for the United Kingdom the SERAS Study (Halcrow Group 2002b) to ana- lyze surface access to airports in the London region included two additional market segments to distinguish between do- mestic and international trips by U.K. residents. Although this model distinguished between U.K. residents and non-U.K. residents, rather than residents of the London region and vis- itors to the region, the distinction is less critical than it might at first appear. Distances in the United Kingdom are such that many U.K. residents will travel to the London airports by sur- face modes for international trips, whereas visitors to the United Kingdom will fly in to the London airports, even if their final destination is elsewhere in the country. The inclu- sion of a U.K. domestic market segment accounts for travel between London and the rest of the United Kingdom. Even so, this segmentation is problematical on a number of counts, and if the distinction between domestic and international travel is considered to be important (as it might well be in the United States as well as the United Kingdom), then it would probably be desirable to distinguish between international and domestic visitors to a region as well as international and do- mestic travel by residents of the region, giving eight market segments rather than the six used in the SERAS study. Although the distinction between business and personal trip purposes is widely followed, air travel forecasting practice often also makes a distinction between what is usually termed vacation travel and visiting friends and relatives travel. The factors that influence the choice of where to stay (a hotel ver- sus a local residence) as well as the airport egress and access travel options are quite different for visitors in the two cate- gories. This distinction is less important for residents of the region, because their airport access trip is generally unaffected by what they are going to do at the other end of their trip. The distinction between business and personal trips is largely a reflection that travelers on business are typically not paying their own travel costs. However, it does not always fol- low that business travelers are less concerned about cost than non-business travelers. The airport access mode choice model developed for Boston Logan International Airport by the CTPS (Harrington et al. 1996; Harrington 2003) distinguished between business travelers paying their own out-of-pocket costs from those whose travel costs were paid by their em- ployer. This was handled by including a separate variable in the model utility function, but it could equally well (perhaps better) be handled by defining a separate market segment. One important consideration in choosing the number of market segments to include in a model is the implications for the number of survey responses required to obtain reliable es- timates of the model coefficients. Somewhat oversimplified, the more market segments that are defined the larger the sam- ple size needed to estimate the model. This is not always true, because a poorly specified model may need a larger data sample size to achieve a desired level of confidence for the
model coefficients; therefore, adding another market seg- ment may actually improve the model fit with a given data sample size if indeed that better reflects differences in the un- derlying behavior of the travelers. However, consideration also has to be given to the number of travelers in each mar- ket segment. A market segment with very few travelers will be difficult to model with any degree of confidence without a very large data sample, unless a stratified sample is some- how obtained that ensures an adequate number of responses in the data sample for each defined market segment. MODEL PERFORMANCE The issue of model performance is one that has largely been ignored in the literature, apart from the simplistic approach of reporting the level of significance (t-statistics) of the model coefficients and the overall goodness-of-fit of the model. The latter is usually expressed by the likelihood ratio, a statistical measure of overall fit of the model to the data that is quite meaningless to most users of the models. However, as suggested by Gosling (2006), what really matters is not how well the model fits the data from which it has been esti- mated, because the nature of the model estimation process largely forces this to be fairly good (that after all is the whole point of model estimation). Rather, what matters is how well the model predicts the airport ground access mode choice be- havior when conditions change, such as the transportation services levels are improved (or degraded) or a new service is introduced. Even the stability of the model coefficients over time is an important concern if the model is to be used for forecasting (as they almost always are). Unfortunately, there is very little discussion of these as- pects in either the general literature on airport access mode choice modeling or the technical documentation of existing models. Some academic studies have selected a holdout sub- set of the data that they can then use to test how well the es- timated model explains this subset. However, although this is better than no test at all, it still does not address many of the concerns identified previously, such as the effect of changing the transportation system or the stability of the model coeffi- cients over time. This issue is not unique to airport access mode choice models and there is very little literature on how well most travel demand models perform over time or when conditions change. There is a growing interest in this issue by the FTA in the context of the modeling used to support the Federal New Starts capital grant program (U.S. Federal Transit Administration 2006) and the term âreference-class forecast- ingâ has started to come into use to describe a forecasting method that is based on the development of a so-called âout- side viewâ of a particular project on the basis of information from a class of similar projects (Flyvberg 2005, 2007; 54 Flyvberg et al. 2005). This approach involves placing the project in a statistical distribution of outcomes from a class of similar projects. Although the concerns that have moti- vated the development of reference-class forecasting are broader than simply the performance of travel demand mod- els themselves and have as much to do with how those models are used, including the input assumptions on which forecasts are based, the approach is directly applicable to air- port ground access mode choice modeling, particularly where the motivation for the development of these models is to forecast ridership on proposed new airport access links or services. The benchmark comparison analysis that was un- dertaken as part of the modeling for the Toronto Air Rail Link described in chapter three represents an example of this approach. This suggests that monitoring the performance of the air- port ground access system after the introduction of new ser- vices or transportation links should be a part of the project funding, so that adequate resources are available to do the necessary ongoing data collection and analysis after the proj- ect is complete, and indeed that the requirement to undertake this monitoring should be a condition of the project funding. Only by collecting data on the use of all of the ground access modes on an ongoing basis after a new service or link is opened can the changes in mode use be compared with the predictions of the mode choice models during the project planning stage and the performance of the models assessed. However, although improved models are clearly desirable, as is a better understanding of the likely reliability of existing models; even an inaccurate model may be better than no model at all. The range of transportation alternatives that comprise the airport ground access system at most airports and the myr- iad traveler decisions that result in the observed pattern of access mode use are sufficiently complex that it is highly unrealistic to attempt to determine the likely effect of any sig- nificant change in the system by intuition alone. A well- constructed model is better than a guess. Even so, it is impor- tant to give appropriate consideration to likely sources of forecast error in any model application and undertake sensitive analysis to explore the possible effect of potential error in both the modeling and forecast assumptions on the model predic- tions. Potential sources of error include omission or incorrect representation of important explanatory variables, incorrect representation of the mode choice process implicit in the model structure, and optimism bias in selecting future values of explanatory variables. One useful approach to assessing the likely reliability of a model is to undertake a back-casting analysis that applies the model to explain observed behavior in the period before that for which the model was developed. This is particularly useful if significant changes occurred in the air- port ground transportation system during that period, but may also shed light on the ability of the model to reflect the effect of changes in real costs and income levels over time.
