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.
67 Given the considerable cost and effort required to develop models of airport ground access mode choice, it is natural to ask whether separate models need to be developed for every airport or whether it would be possible to adapt or apply a model developed for one airport at another. This would depend on how transferable such models are. In general, experience applying models of transportation behavior in sit- uations that are different from the one for which they were de- veloped has been mixed. An early study of the transferability of work-trip mode choice models (Atherton and Ben-Akiva 1976) suggested that these models might transfer fairly well. However, further work that examined model transferability issues in more detail (McFadden et al. 1977) found that whereas the evidence for transferability of work-trip mode choice model coefficients between different market segments within a region was mixed, their transferability between regions did not appear very encouraging. Subsequent work (Pas and Koppelman 1984; Badoe and Miller 1995) showed that although mode choice models appeared reasonably trans- ferable within the same metropolitan region, the model parameters were not temporally stable. McCoomb (1986) es- timated journey-to-work models for the ten largest Canadian cities using the same mode choice model specification and consistent travel data that was collected for each city using the same survey for the same day. Although the model coeffi- cients were found to differ across the cities, some models were similar enough to lead the author to conclude that they could be transferred when the cities are reasonably similar in size, urban structure, and transportation system. The impor- tance of consistent data as a requirement for model transfer- ability has also been noted by Galbraith and Hensher (1982). More recently, Rossi and Outwater (1999) undertook a comparative analysis of mode choice model parameters for home-based work, home-based non-work, and non-home- based segments of regional travel demand models for 11 U.S. metropolitan areas. The values of the parameters for a given variable were found to vary quite widely, particularly for home-based non-work and non-home-based trips. The au- thors undertook a simple experiment of applying the param- eters for each of the regional models to a hypothetical origin- destination zone pair in some other region after obtaining ASCs for transit for the other region by fitting each model to an assumed transit share in a different origin-destination zone pair. The results showed that the estimated transit share in the test market varied from 10.2% to 19.8%, or by a factor of al- most two. The authors concluded that whereas transferring mode choice model parameters may sometimes be necessary where adequate data are not available to estimate a model on local conditions, the results should be used with caution. They also stress that the complete model should always be used and not just selected coefficients, because variables may be corre- lated with each other. For readers interested in more details on past work on model transferability, there is a good summary of the litera- ture on the transferability of regional travel demand models by Karasmaa (2003). Past studies of travel demand model transferability have fo- cused on models of general urban travel behavior, which includes a wide range of trip purposes that are likely to be heav- ily influenced by the local characteristics of the transportation system, and airport ground access travel behavior may be more consistent. In principle, one would expect that air travelers would behave similarly when faced with a similar choice situ- ation, controlling for differences in ground access service char- acteristics (e.g., fares or travel times) and differences in traveler characteristics (e.g., trip purpose and duration, household com- position, and income). Airport access mode choice models at- tempt to account for the effect of these variables on the choices made by a given air party. Therefore, to the extent that a model accurately reflects the effect of these variables, it should explain the behavior of air parties in other geographic regions. How- ever, this is a significant caveat that may well not hold for rea- sons discussed here. In particular, there may be regional differences in atti- tudes toward the service characteristics of different modes, as well as differences in the nature of the services offered. This is likely to be a significant consideration with the use of pub- lic transit services. Air travelers living in large metropolitan areas with well-developed transit systems are more likely to consider using public transportation for a trip to the airport than those from areas with less extensive transit services and less use of transit, partly owing to familiarity and comfort with using public transportation and partly the result of dif- ferences in automobile availability. Whereas differences in service characteristics of different modes should in principle be accounted for by the explanatory variables in the model, in practice this is not necessarily the case. Misspecification of utility functions or inaccuracies in the values of service char- acteristics used for model estimation can result in biased values of the ASCs or coefficients of other variables that will result in biased estimates of the utilities of each mode when these coefficients are applied in different situations. CHAPTER SEVEN TRANSFERABILITY OF AIRPORT GROUND ACCESS MODE CHOICE MODELS
ISSUES IN MODEL TRANSFERABILITY Because the mix of air passenger or airport employee char- acteristics and the details of the airport ground transporta- tion services that are available differ widely from airport to airport, for an airport access mode choice model to be trans- ferable it must correctly reflect the influence of the different factors that determine airport traveler ground access mode choice. If model coefficients are partly accounting for fac- tors that are not directly associated with the variable in ques- tion (or are inherent to the mode in the case of the ASCs), then differences in these factors when the model is applied in other situations will result in biased estimates of the modal utilities and errors in the predicted probabilities of using each mode. Although this applies to the transferability of a model from one region to another, it also applies when a model is used to analyze a significant change in the ground trans- portation system serving the airport for which it was devel- oped, such as the addition of a new mode or service, as well as the application of a model to forecast future mode use, as discussed by Gosling (2006). Therefore, ensuring that mod- els correctly reflect the factors that influence mode choice decisions not only improves their transferability to other regions, but also their use in planning studies at the airport or airports for which they were developed. This is a lot easier said than done. In particular, model transferability is likely to be influ- enced by the following factors: ⢠Incorrect or incomplete model specification, ⢠Missing explanatory variables, ⢠Incorrect market segmentation, and ⢠Problems with the data used to estimate the model coefficients. VARIABILITY IN MODEL SPECIFICATION Although there are a fairly limited number of different airport access mode choice models that have been documented in the review of recent literature and professional practice, it is relevant to ask how consistent these models are in terms of their functional specification and use of explanatory vari- ables. Wide variation in model specification would suggest that either the underlying behavior that they are attempting to explain varies widely from region to region or the models do not correctly reflect that behavior in their technical specifica- tion. Either way, this would suggest that they are not likely to be reliably transferable to other situations. Functional Form Although a number of recent models have continued to use a MNL structure, the more advanced models have used a NL 68 structure. There are sound theoretical reasons why the NL structure should be superior to the MNL, particularly if the model attempts to include modes that are likely to be per- ceived as having a stronger substitutability with some alter- native modes than others or to distinguish between different secondary modes that are used to access a primary mode, such as parking at an off-airport terminal compared with being dropped off by private vehicle. Because the nesting structure of the limited number of models that have used a NL structure is very dependent on the particular modes that exist in the region being modeled, it is difficult to generalize about how the various models have grouped the modes. The model for Boston Logan Interna- tional Airport developed by the Central Transportation Plan- ning Staff (Harrington et al. 1996; Harrington 2003) adopted a hybrid structure, in which the resident models were nested but the nonresident models were not. The resident models included an automobile nest (drop off, short-term parking, and long-term parking on and off airport) and a door-to-door nest (taxi and limousine). The public transport modes (MBTA transit, scheduled bus/limousine, Logan Express, and Water Shuttle) were not nested, although that would have been redundant because there were no other modes at the top level of the tree. In contrast, the nesting structure of the SERAS Air Passenger Surface Access Model (Halcrow Group 2002b) involves up to six levels of nest, including a three-level nest of different rail options. The decision of which modes are nested together appears to have been made on the basis of what gave the model the best fit. For some market segments the use of taxi and drop off by private vehi- cle (termed kiss and fly) are grouped in the same second- level nest, with private vehicle parked for the trip duration (termed park and fly) at the top level, whereas for other mar- ket segments all three modes are at the top level or are grouped differently at lower levels. However, the discussion of household income brings up an interesting issue with respect to model transferability, namely how to reflect differences in the cost of living in different regions. Although including household income in an airport access mode choice model in some way is clearly necessary if the model is to correctly reflect traveler behav- ior, a given income level may provide more disposable in- come in one region than in another. The extent to which this will influence traveler sensitivity to time and cost tradeoffs in airport access trips is unclear. Explanatory Variables Differences in the explanatory variables included in different airport access mode choice models, although not necessarily a problem for model transferability, do raise questions about the extent to which the model coefficients may be explaining factors not directly measured by the variables in question. The absence of studies that have taken a model estimated in
69 one region and applied in another to compare the model pre- dictions with the observed mode choices in the second region makes it difficult to assess how sensitive the model predic- tions might be in such a situation to the choice of variables included in the model. In particular, the absence of a house- hold income variable (or equivalent measure of the perceived value of time) in many models means that such models would most likely not work well at all if applied in a situation where the distribution of household income is very different. An ideal model would not have ASCs in the utility function because these imply that the mode will have some probability of being selected even if the values of all the continuous vari- ables are zero. Rather, the intrinsic attributes of the mode would be expressed through appropriate continuous variables that could be adjusted for different situations. However, in practice the ASCs account for missing variables and other data problems and frequently take values that are quite large com- pared with the effect of the continuous variables. This is likely to be a significant problem for the transferability of these mod- els, because the factors represented (at least in part) by the values of the ASCs are not likely to occur in the same way in another situation. Market Segmentation This is the aspect in which there is the most consistency in re- cent airport access mode choice models. It has become general practice to develop separate sub-models for four market seg- ments (although terminology may vary slightly) comprising: ⢠Resident business trips, ⢠Resident non-business trips, ⢠Non-resident business trips, and ⢠Non-resident non-business trips. This market segmentation recognizes that residents of the area have different airport access mode options than visitors to the region (non-residents) and that travelers on business trips may value their time differently from those on personal trips. There is also the consideration that many business trav- elers are able to charge their travel expenses to their employer or client. The need of some visitors to rent a car for their local travel while in the area is another important difference be- tween residents and visitors. Once a decision is made to rent a car for other reasons, then that typically predetermines the airport access mode choice. However, examination of recent airport access mode choice models suggests that this four-way market segmenta- tion may be too simplistic to fully reflect the factors that shape airport access mode choice behavior. Visitors staying with friends or relatives often have access to a private vehi- cle for local travel (indeed, the local travel may be under- taken together with those they are visiting) as well as people who can pick them up from the airport and drop them off, whereas those staying in a hotel or visiting a business typi- cally do not. Similarly, travelers on business who can charge their travel expenses to an employer or client may behave very differently from those who have to meet their travel expenses out of their own pocket or out of a limited travel budget. Thus, there may be significantly different behavior by someone traveling to a professional conference at their own (or their organizationâs) expense compared with some- one attending a meeting the costs of which they can charge to a client. Although these issues have typically been handled (if at all) through the specification of the utility functions, this imposes limitations on the ability of the model to reflect the relevant behavior and it may be better to address these issues through more detailed market segmentation. COMPARATIVE ANALYSIS OF MODEL PARAMETERS Although the numerical value of the model coefficients in a nested or MNL model have no direct significance because they depend in part on the variance in the dataset from which they have been estimated, the ratio of two coefficients does have an interpretable meaning that can be compared across different models. The ratio of the travel time coefficient to the cost coefficient represents the implied value of travel time. Similarly, the ratio of coefficients of other variables to the travel time coefficient can be interpreted as a factor that converts the values of the other variable to equivalent min- utes of travel time. Even if the implied value of travel and other continuous variables is reasonably consistent between two models, the values of the ASCs are likely to be problematic from the per- spective of the transferability of a given model. If indeed these coefficients represent intrinsic characteristics of the mode that will be similar in other situations, then using them in another situation should produce reasonable results. How- ever, in practice they are likely to account for missing variables, model specification errors, inaccuracies in the estimation dataset, and other issues. To the extent that these are likely to be different in different situations, the use of the ASCs in the second situation would most likely gener- ate incorrect results. DATA CONSIDERATIONS It is a truism that a model is only as good as the data from which it was developed and with which it is applied. This is particularly relevant to airport access mode choice models owing to the extensive amount of data that are needed to de- velop and apply them and because much of these data are not readily available. Errors in the data used for model estimation will result in biased model coefficients, as those coefficients attempt to explain behavior in terms of data that are incorrect. Errors in the estimation data will not typically be detected by the usual tests of model goodness-of-fit. Similarly, even if the
model is estimated on valid data and the model coefficients are not biased by data problems, applying the model with incorrect data will produce incorrect results. Potential data problems arise in both of the two broad cat- egories of data required for their development: survey data of air passenger or airport employee characteristics and data on the transportation service levels in the airport ground access system. Air passenger and airport employee surveys are rarely truly random samples of the underlying population, although this is often overlooked in their analysis. In particular, the logistics of performing air passenger surveys necessarily restricts the sam- ple to a subset of the total population. Surveys are typically performed for a limited time period and often for only part of each day on which the survey takes place, owing to staffing considerations. Air passenger surveys are often done in airline gate lounges, because that is a location in which passengers are most likely to be willing to take the time to answer a survey. However, this limits the sample to passengers on specific flights traveling to a specific destination. Although an effort is typically made to survey a reasonably representative sample of flights, budget limitations on the number of flights that can be surveyed will usually result in a biased sample. This can be partly corrected by appropriate weighting techniques; how- ever, careful thought needs to be given to how this is done. Weighting responses to correct for one issue such as the distri- bution of responses by flight destination can exacerbate any bias in the responses with respect to a different issue, such as the distribution of responses by time of day. The second set of problems that can arise with air passen- ger and airport employee surveys relates to question wording or which questions are asked. Omitting a topic in the survey, such as the household income of the respondents, necessar- ily precludes including that factor in the specification of a mode choice model estimated from the data. Poorly worded questions can also introduce errors in the data. Air passen- gers in particular may not use the same terms to describe their trip characteristics or ground access modes used as the de- signers of the survey expected. For example, an air passenger may refer to a shared-ride van as a limousine (indeed some shared-ride van operators use the word limousine in the name of the service). Although these errors in individual responses may not be too serious in presenting the aggregate results of the survey, particularly where there may be offsetting errors by other respondents that reduce the overall error in the results, they can have a significant impact on the use of sur- vey results to estimate disaggregate mode choice models that attempt to explain why individual survey respondents chose the access mode that they did. Whereas in principle it should be fairly easy to determine the travel times and costs of using different access modes from existing data sources, such as the regional transporta- tion planning process or airport ground transportation infor- 70 mation systems, in practice this is often problematic and prone to error. Highway travel times vary by time of day and day of the week. Although regional transportation planning agencies are increasingly developing datasets that better re- flect this variation, it is still common for planning agencies to divide the day into a limited number of time periods (e.g., the a.m. peak period, p.m. peak period, and off-peak period) and determine average weekday travel times for each period. However, airport travelers will base their ground access de- cisions on how long they think their trip will take at a partic- ular time of day, which may be quite different from the travel times given by the regional transportation network datasets. Airports, particularly large airports, are typically served by a large number of private transportation providers, such as shared-ride van or limousine (black car) services, that may charge a wide range of fares that vary by the location of the trip origin as well as from provider to provider. The travel time involved in taking a shared-ride van is also influenced by how many other travel parties have to be picked up or dropped off. Basing the mode choice model estimation on travel time and cost assumptions for each mode that is dif- ferent from what the travelers perceived those values to be when they made their decision will inevitably bias the model. SUMMARY Given the current lack of consensus over model specification and typical coefficient values between different airport ac- cess mode choice models, it can be assumed that the trans- ferability of these models is highly suspect. Although it seems plausible that the underlying airport traveler behavior may not differ that much from region to region, after taking into account differences in air passenger or airport employee characteristics and transportation system service levels, it appears unlikely that current airport access mode choice models do this correctly, based on the significant differences between the different models. Both because of the obvious value of being able to apply airport access mode choice models in different situations from those for which they were originally developed, as well as the concerns about the reliability of even applying them to different situations at the airports for which they were devel- oped, there is a pressing need to better understand how well current models reflect the factors influencing the underlying travel behavior and how they can be improved to better reflect this behavior. It is not sufficient to simply say that existing models are unlikely to be transferable and that a new model must be developed for every application. This is to admit that the existing models do not properly reflect the causal structure of the underlying traveler behavior and by implication calls into question the reliability of the predic- tions generated by these models. In the event that a study sponsor does not have the re- sources or time to develop a new mode choice model for a
71 particular planning study where an existing model is not available and wishes to try using an existing model that was developed for another situation elsewhere, at a minimum it would be prudent to conduct validation tests of a number of candidate models by examining how well they predict recent travel behavior in the new location. Although the predictions from these models can usually be improved by adjusting the ASCs, it should be recognized that this may well be creating a false sense of confidence. What is more important is to ensure that the effects of differences in the continuous vari- ables are properly reflected in the model predictions. If the prediction errors are reasonably consistent for each mode across all subsets of the data (e.g., by distance from the airport or across different income levels or air trip duration), then it is more likely that the model is reflecting the underlying behav- ior. On the other hand, even if the model correctly predicts the total number of trips using each mode but predictions for sub- sets of the data are wildly off, then it is very unlikely that the model will produce reasonable predictions of the likely effect of changes in the system.
