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.

55 As discussed in the literature review in chapter two, the topic of airport employee ground access and egress mode choice has re- ceived much less attention in the literature than that of air pas- sengers, and only one study was identified that described an air- port access mode choice model developed to account for airport employee access mode choice behavior. In the case of airport employees, airport access trips are more commonly referred to as journey-to-work trips. However, these trips will be referred to as airport access trips in the following discussion for consistency with the treatment of airport access travel by air passengers. As with air passengers, discussion of airport employee ac- cess mode choice behavior should really address both access and egress travel. However, unlike air passengers, airport employee egress mode use is likely to be the same as their ac- cess mode use. If they drive to work by private vehicle, then that constrains their choice mode for the return trip. Simi- larly, if they use public transportation for the journey to work, they do not generally have a private vehicle available for the return trip. Although situations may arise on an indi- vidual basis from time to time in which airport employees travel to work one way and return home another (e.g., they may come to work by public transportation but get a ride home from a co-worker), it can be expected that the effects of these irregular commute patterns will generally cancel each other out in the overall mode split. Airport employee trips can account for a large share of the ground access trips generated by an airport, particularly for large airports that have a wide range of support activities, in- cluding air cargo handling, aircraft maintenance, and airline crew bases. An earlier TCRP study (Leigh Fisher Associates 2000) assembled data for a sample of airports on air pas- senger traffic and the estimated average number of daily employees working at the airport. These data are shown in Table 5, together with the ratios of average daily employees to average daily enplaned passengers and average daily orig- inating passengers. The first ratio indicates the relationship between the number of airport employees and the overall level of airport activity, whereas the second ratio indicates the relationship between the number of airport employee airport access person-trips per day and the number of air pas- senger ground access person-trips per day, assuming that each employee makes one access trip per day. The data in Table 5 show that the ratio of daily employees to enplaned passengers varies widely, from 0.13 to 0.83, with the values for major airline hub airports in the range from 0.3 to 0.6; whereas airports with predominantly origin and desti- nation traffic generally fall in the range from 0.2 to 0.3. The high value for Oakland International is most likely the result of the presence of an on-site major sorting facility for Federal Express, which had relatively low passenger traffic com- pared with the other airports in the sample. The very low value for San Diego International could be the result of part of the very constrained site, which limits the opportunities for ancillary activities at the airport. The higher values for the major airline hub airports reflect the additional airline activ- ities that occur at those airports, including maintenance ac- tivities and crew bases. The more interesting statistics from the perspective of airport ground access are the ratios of average daily em- ployees to average daily originating passengers. Apart from San Diego International, these range from about 0.25 in the case of predominantly origin and destination airports to a high of 1.55 for Dallas/Fort Worth International. Most of the major airline hub airports have values between 0.7 and 1.0, with Lambert–St. Louis International (which at the time was the main hub for TWA) having a ratio of 1.27. Although person-trips do not translate directly to vehicle trips owing to differing use of high-occupancy vehicles (HOV) and the additional vehicle trips resulting from air passenger drop-off and pick-up trips by private vehicle, this suggests that daily employee vehicle trips at major airline hub airports are of the same order of magnitude as vehicle trips generated by air passengers, and could be significantly higher. The statistics given in Table 5 also provide a perspective on the relative importance of airport employee trips in the overall pattern of regional travel. A flow of 10,000 to 40,000 trips per day each way clearly has a significant im- pact on the regional transportation system in the vicinity of the airport, even though it may be only about 1% of the total regional journey-to-work travel (the number of average daily employee trips shown in Table 5 averages approxi- mately 1.1% of the corresponding metropolitan area aver- age weekday journey-to-work trips, although the ratios vary from 0.2% to 2.1% across the different metropolitan areas). Airport employee travel to and from work differs in a number of ways from typical journey-to-work travel. Major CHAPTER SIX AIRPORT EMPLOYEE MODE CHOICE

airports operate 24 hours a day, 7 days a week, and as a result many employees work shifts that are significantly different from the conventional workday. In addition, airline flight and cabin crews often have multi-day duty cycles during which they are away from their crew base. It is unclear how these characteristics cause employee access mode use patterns to be different from more general journey-to-work trips repre- sented in regional travel demand models. Although there may be other trip generators in a region that have similar work shift patterns to those at the airport, these trips are prob- ably not well handled by standard regional travel demand models either. The survey of MPOs described in chapter three exam- ined how airport employee access travel is handled in re- gional travel demand models. Although no MPOs reported the use of a special-purpose airport employee mode choice model, six reported the use of a special-generator sub- model for airport employee trips, with the majority treating airport employees the same way as other regional journey- to-work trips. This chapter therefore examines the use of special-generator sub-models for airport travel, as well as a selection of general urban journey-to-work mode choice 56 models, to give an indication of the typical structure of such models. MODEL APPLICATION The application of airport employee mode choice models, whether special-purpose models, adapted special-generator sub-models, or general-purpose journey-to-work models, falls into two broad categories. The first is to predict airport employee travel patterns for airport planning studies, whether specifically oriented to employee travel or more general airport ground transportation issues. Airport em- ployee travel accounts for a significant proportion of the vehicle trips generated by airports, particularly large air- ports, as well as of the ridership on transit services to and from airports. Therefore, modeling these trips forms an im- portant part of feasibility studies for improved airport ground transportation services or infrastructure, as well as studies directed at reducing the environmental impacts of airport activities. In particular, environmental impact docu- mentation for airport projects needs to address airport em- ployee travel as part of considering transportation impacts. Although this can be (and often is) addressed by simply Avg. Daily Employees to Avg. Pax per Day Airport Enplaning Passengers 1998 Originating Passengers 1998 Average Daily Employees Enplaned Originating Los Angeles International 30,826,859 18,313,990 40,000 0.47 0.80 Chicago O’Hare International 35,841,551 16,127,120 40,000 0.41 0.91 San Francisco International 19,658,626 12,531,590 31,000 0.58 0.90 Dallas/Fort Worth International 30,121,523 11,279,510 48,000 0.58 1.55 Las Vegas McCarran International 15,132,220 11,114,050 7,500 0.18 0.25 Boston Logan International 13,250,754 10,364,870 14,500 0.40 0.51 Phoenix Sky Harbor International 15,984,620 10,323,330 23,655 0.54 0.84 Seattle–Tacoma International 12,867,830 9,592,000 11,375 0.32 0.43 Denver International 18,444,540 8,956,900 17,400 0.34 0.71 San Diego International 7,453,186 6,980,960 2,600 0.13 0.14 Houston Bush Intercontinental 15,492,252 6,356,870 14,406 0.34 0.83 Tampa International 6,955,805 6,289,530 8,219 0.43 0.48 Lambert–St. Louis International 14,334,844 5,442,800 19,000 0.48 1.27 Portland International 6,487,226 5,353,170 5,000 0.28 0.34 San Jose International 5,202,502 4,861,650 3,500 0.25 0.26 Salt Lake City International 10,097,036 4,762,760 13,026 0.47 1.00 Metropolitan Oakland International 4,612,614 4,183,300 10,500 0.83 0.92 Sacramento International 3,593,647 3,521,200 2,300 0.23 0.24 Source: Leigh Fisher Associates (2000, Table 1-1). Average daily employees estimated for 1998 based on data reported by airport operators. Passenger ratios calculated by author. Notes: Avg. = average; Pax = passengers. TABLE 5 AIRPORT EMPLOYEE ACCESS TRIP GENERATION

57 using existing mode split data from employee surveys or similar sources, to the extent that proposed projects will change the factors affecting employee choices or that pro- posed mitigation measures have the goal of modifying employee access mode use, it becomes necessary to have tools that can predict the resulting changes in airport em- ployee access mode behavior. The second category of model application is the represen- tation of airport employee travel in general urban or regional travel modeling. Where airport employee travel is simply treated the same as any other journey-to-work trip then the only question becomes the number of employee trip ends that are assigned to the airport TAZs in the trip generation process. However, where special-purpose airport employee mode choice models or adapted special-generator sub-models for airport employee trips are used, then in addition to the models themselves consideration needs to be given to how the resulting airport employee trip patterns are integrated into the overall trip assignment process. TECHNICAL APPROACH In general, the airport ground access modes available to air- port employees are the same as those available to air passen- gers, although some modes, such as rental car and drop off by private vehicle, are not likely to be used by airport employ- ees on a regular basis, whereas other modes that may be available to airport employees, such as car pool and van pool, are not applicable to air passenger travel. This suggests that a similar modeling approach would be appropriate. Given the number of alternative modes that would need to be included in the model, as will be discussed later in this chapter, it would seem likely that a NL model would be the most appropriate structural form (at least at the current state of practice of airport access mode choice modeling). This also corresponds to the current state of best practice for general regional travel mode choice models. Because different fac- tors influence air passenger and airport employee mode choice decisions, the explanatory variables included in the modal utility functions and the market segmentation (if any) used in airport employee mode choice models will be differ- ent from that used in models of air passenger mode choice. Pivot-point analysis, as discussed in chapter two, may be an appropriate approach for modeling employee travel where data on existing employee mode use are available. Airport Employee Access Mode Choice Models The literature review only identified three studies in which mode choice models were specifically developed to predict airport employee access mode use. All three studies adapted other journey-to-work mode choice models to predict airport employee mode use rather than developing an entirely new model from airport employee travel data. SERAS Model A special-purpose airport employee mode choice model was developed for the Greater London region as part of the U.K. SERAS study (Halcrow Group 2002b). This model was a fairly simple binary (two-mode) logit model that predicted the percentage use of private vehicle and public transport. The model used an incremental logit formulation, sometimes described as a pivot-point analysis. Rather than directly pre- dicting the percentage of employees using private vehicles, it predicted the change in the percentage use in terms of the change in the explanatory variables. This change was then applied to the observed mode split in the base case. Mathe- matically there is no difference between this approach and the use of a regular logit choice model as long as the utility func- tions in the regular logit model have ASCs to ensure that the mode splits in the base case correspond to the observed data. The SERAS employee mode choice model was based di- rectly on one developed for the South and West London Transport Conference for a study covering the area to the south and west of London Heathrow Airport. It was there- fore felt that this provided a good basis for the SERAS work. However, the Conference study covered all journey-to-work travel, not just airport employees, and thus the model does not include factors specific to airport employee travel (such as irregular shift times or multi-day duty cycles). The param- eters used in the SERAS model were derived from an earlier revealed preference study of multi-modal travel, the details of which were not reported. It is unclear whether this earlier study comprised (or even included) airport employee travel. Because the SERAS model included only two modes, pri- vate vehicle and public transport, no account was taken of factors influencing shared-ride behavior. However, the costs and travel times involved in using public transport were ob- tained from a public transport network model that took into account interchanges between different services (such as mainline rail, London Underground, and local bus). There- fore, the use of different public transport modes is included in the model by implication. The model used a generalized cost approach with separate travel time coefficients for access and egress time, waiting time, and in-vehicle time. In addition, public transport travel times included a constant for travel on each mode, termed a boarding penalty, which reflected relative preferences for different modes. Each bus journey incurred a boarding penalty of 7.5 min, whereas each journey by rail, light rail transit, or guided bus incurred a boarding penalty of 2.5 min. The boarding penalties were added together if a trip involved a change of mode. However, there were no penalties for an interchange within a mode, apart from any waiting time involved. Costs were combined into a single variable. Costs and times were converted to a generalized cost by multiplying travel times by an assumed average value of time. The same value was used for all

employees, because the model was applied at the level of an analysis zone rather than individual employee. Oakland Airport Connector Model As part of the analysis for the preparation of an Environmen- tal Impact Report/Environmental Impact Statement for a planned APM connection between the Oakland International Airport and the Coliseum station of the BART system, located approximately 2.5 miles from the airport, an airport access mode choice model was developed and applied to generate ridership projections for the APM connection, termed the Oakland Airport Connector (BART–Oakland . . . 2002, Appendix B: Transit Ridership Procedures and Inputs). The mode choice model comprised five market seg- ments covering four air passenger market segments and an airport employee segment. The airport employee segment of the mode choice model considered only two modes, private vehicle and transit, and used the same coefficient value for both highway and transit travel time, including rail and bus. In addition to travel time, the utility functions included walking distance, waiting time, and cost. The coefficient values were not estimated from em- ployee survey data, but were adapted from the regional travel demand model developed by the MTC, using the mode choice model coefficients for home-based work trips. The ASC for transit does not correspond to the MTC model coef- ficients and appears to have been estimated to make the tran- sit mode share match observed data for airport employees. However, the MTC model includes seven modes and sev- eral other explanatory variables, including household income and vehicle ownership. It could be expected that excluding these variables from the model would change the appropriate values of the other coefficients, because the behavioral explanation that they provide in the MTC model would now have to be accounted for by the remaining variables. The likely impact of this on the performance of the model is unclear. Furthermore, the MTC mode choice model was estimated using constant 1990 dollars for travel costs. Apply- ing these model coefficients to costs in 1999 dollars, as appears to have been done in the analysis, without making any adjustments for inflation would overstate the effect of cost in mode choice decisions. A more detailed discussion of this model is provided in Appendix D (web version only) and additional information on the MTC home-based work mode choice model is given later in this chapter. San José International Airport Model A very similar approach to that used to model airport em- ployee travel in the Oakland Airport Connector study was followed in a similar model that was developed about a year 58 later for a study of the potential ridership on a planned APM that would connect San José International Airport with a nearby station of the Santa Clara Valley Transportation Authority light rail system (Dowling Associates 2002). As with the Oakland Airport Connector model, coefficients for travel time and cost were adopted from those for home-based work trips in a regional travel demand model, in this case the Santa Clara countywide transportation model, and the ASC for transit was estimated to fit the model predictions of mode use to airport employee survey data. Although details of the Santa Clara countywide model were not reviewed, it appears from the values of the coefficients that the inconsistency in the years in which the cost data was expressed in the two models may be even greater than for the Oakland Airport Connector model. The application of the model allowed for up to four differ- ent connecting transit routes between any given analysis zone and the airport. For those routes involving the use of the APM, the ASC for transit was adjusted on the basis of the results of a stated preference survey performed as part of the study to reflect the greater reported likelihood of using transit if the APM link was available. However, the adjustment was not particularly large and the difference would be offset by a dif- ference in fare of only 20 cents. A more detailed discussion of this model is provided in Appendix D (web version only). Special-Generator Models Special-generator models form part of the regional travel demand model and make use of customized techniques to predict travel to zones such as a sports stadium or airport that is not likely to be properly represented by the regular modeling process. Six of the MPOs responding to the survey described in chapter three indicated that they used a special- generator model for airport employee trips, although only one of these, the Houston–Galveston Area Council, provides any relevant information on its website. The Council com- missioned a travel survey at Houston Intercontinental Airport in November 1995 to support the development of a special- generator model. This survey assembled data on air passen- ger ground access travel, airport employee travel, and com- mercial vehicle trips to and from the airport. In addition to surveys of air passengers, airport employees, and commercial vehicle drivers, traffic counts were undertaken on airport ac- cess roadways to develop expansion factors to convert the survey results to daily trips. The airport employee survey obtained data on household and travel characteristics for 193 employees, including the number of people and workers in the household, the number and age of vehicles available, and the household income. The survey respondents reported an average of 2.4 trips to and from the airport on the travel survey day. The majority were by private vehicle, with 88% auto driver, 10% auto passen-

59 ger, and 1% each transit and other. These trips were ex- pressed in terms of the trip types used in the regional travel demand model and factored up on the basis of total airport employment to give average weekday trips by trip type. General Regional Travel Models The most common way to model employee travel to and from airports is to treat the airport in exactly the same way as any other TAZ in a regional travel demand model and use the trip generation, trip distribution, and mode choice sub-models to generate the number of person and vehicle trips associated with airport employee travel. Typically, an airport will be rep- resented as single TAZ within the regional travel analysis zone system. Most regional travel demand models include separate sub-models for different types of trips, such as travel to and from work or for shopping or recreation. Most models consider journey-to-work, commonly termed home-based work (HBW), trips and assume that the reverse trip is sym- metrical (although at a different time of day). Because the focus of this report is on airport access mode choice, this section provides a brief review of current prac- tice with HBW mode choice models in regional travel demand modeling. However, it should be noted in passing that the total number of employee trips and the resulting travel patterns on the regional transportation network depend on the trip generation and trip distribution components of the regional models as much as on the mode choice component. The extent to which the trip generation relationships incor- porated in the regional model are representative of the employment at an airport is an issue that should be consid- ered in interpreting the predictions of airport employee travel produced by a regional model. There are two aspects that may need to be addressed. The first is how well the base year employment levels implicit in the model correspond to ac- tual employment at the airport, whereas the second is how well the model predicts the expected growth in employment at the airport in the future. The first aspect is primarily a ques- tion of how well the regional model is calibrated to the air- port employment data at the TAZ level. The second aspect relates to how well the way that the regional model predicts future growth in employment corresponds to the expected growth in airport activity and employment. Where there are exogenous forecasts of future airport employment levels, such as forecasts produced by the airport authority, it may be possible to improve the predictions of future airport em- ployee travel produced by the regional travel model by fac- toring the airport employee trips predicted by the regional model for future years so that the number of employee trip ends in the airport TAZ corresponds to the exogenous fore- cast of airport employment. Although trip distribution issues will generally be of lesser concern than trip generation and mode choice, the air- port authority will typically have data on the distribution of employee residences that can be compared with the distribu- tion of employee trip origins predicted by the regional model trip distribution process. If the differences are significant, and this is important for a particular application such as pre- dicting the likely demand for a new airport access service, then further adjustments to the employee travel pattern pre- dicted by the regional model may be necessary. A survey of MPOs recently completed by Vanasse Hangen Brustlin for TRB explored the state of practice in metropoli- tan area travel forecasting (Vanasse Hangen Brustlin 2007). The survey found that the great majority of the 228 MPOs re- sponding use a trip-based four-step or similar travel forecast- ing process. However, although approximately 95% of the large MPOs reported that they used a mode choice model for HBW trips, overall slightly less than 50% of the MPOs mod- eled HBW mode choice. About three-quarters of the large MPOs use a NL models for HBW trips. The study presented data on the frequency of use of different modes in the mode choice models, but did not indicate which of these were used for HBW trips or provide any information on the explanatory variables used or how the modes were nested. Home-Based Work Mode Choice To illustrate the variety of model structures and explanatory variables included in typical HBW mode choice models, this section presents some technical details on five representative models for the following regions: Atlanta, Dallas/Fort Worth, the San Francisco Bay Area, Seattle, and Metropoli- tan Washington (the Baltimore–Washington region). Each of these models differed in a number of ways, including the modes represented, the structural form of the model, and the explanatory variables. Because of the complexity of the models, their need to reflect the different transportation facilities and geographical aspects of each region, and differ- ences in the underlying transportation system data used to apply the models, this discussion does not cover all the de- tails of the models, for which the interested reader is referred to the technical documentation referenced in the discussion, but rather attempts to provide a comparative overview to give a broad sense of how these models function and differ. The principal characteristics of the five models are sum- marized in Table 6, which shows the model structure, the modes included in the model, and the explanatory variables used in each of the models. Four of the models were true NL models with two or more levels, whereas the fifth consisted of two linked MNL models that were estimated sequentially. The terminology varied in the documentation for the differ- ent models (e.g., some models refer to shared-ride auto trips, whereas others use the term group ride), but consistent terminology has been used in the following discussion for clarity. Therefore, the terminology in the discussion of a spe- cific model may differ slightly from that used in the model documentation.

60 Regional Model ARC MTC MWCOG NCTCOG PSRC Model Structure Sequential multinomial logit Nested logit Modes Included Drive alone Shared ride 2 Shared ride 3 Shared ride 3+ Shared ride 4+ Transit auto access Transit walk access Bus transit walk access Rail transit walk access Bicycle Walk Nests Motorized (auto, transit) Drive alone/shared ride Drive alone/drive to transit Shared ride a Transit access Bus vs. rail transit Variables In-vehicle travel time Waiting time b Walking time b Bicycle time Trip cost b Persons/household Workers/household Vehicles/household Vehicles/worker Households w/ vehicles < people Bus miles in peak hour Household income Employment density Land use mix index Sub-area dummy variables Model: ARC = Atlanta Regional Commission (Travel Demand . . . 2005). MTC = Metropolitan Transportation Commission (Purvis 1997). MWCOG = Metropolitan Washington Council of Governments (COG/TPB Travel . . . 2007). NCTCOG = North Central Texas Council of Governments (Dallas–Fort Worth Regional Travel Model . . . 2007). PSRC = Puget Sound Regional Council (Cambridge Systematics 2003). aEstimated as separate multinomial logit model. bCoefficients adopted from other models. TABLE 6 CHARACTERISTICS OF SELECTED HOME-BASED WORK MODE CHOICE MODELS

61 All five models distinguish between single-occupant auto trips (drive alone) and at least two categories of shared-ride auto trips. Three models have three vehicle oc- cupancy categories (drive alone, two people, and three or more people), whereas two of the models add a fourth cat- egory and distinguish between three people and four or more. The representation of transit in the models is more varied. All the models distinguish between walk access and automobile access to transit in some way; however, the de- tails differ. Some models represent the access mode as a secondary choice, whereas others treat walk access to tran- sit and auto access to transit as separate modes. The Metro- politan Washington model (FY-2003 Models . . . 2007) as- sumes either walk access or automobile access to transit based on the distance from the trip origin to the nearest bus stop or rail station. Transit trips beyond the assumed maxi- mum walking distance from the nearest rail station but within walking distance of a bus stop are assumed to walk to the bus stop. Use of rail or bus then depends on the quick- est path through the transit network. Two of the models in- cluded bicycle and walk the entire trip as alternatives in the choice set for some trips. Not only did the models differ in which modes they in- cluded but the way in which the modes were nested dif- fered across every model. The Atlanta HBW regional model (Travel Demand . . . 2005) divides person-trips into highway trips and transit trips at the top level and then at the second level splits highway trips into drive alone and shared-ride and transit trips into walk access and automo- bile access. Shared-ride auto trips are split into three vehi- cle occupancy categories at the third level, whereas walk access transit trips are split into local bus and rail rapid transit service (termed premium transit) trips, as shown in Figure 6. This structure implicitly assumes that the disutil- ity involved in choosing to walk to transit depends on a lower-level choice of whether to walk to a local bus stop or walk directly to a rail station. Although some travelers in- deed face this choice, most do not and have a trip end far enough from the nearest rail station that their options are to drive to the station or walk to a local bus stop. Those who choose to walk to a local bus stop may subsequently use the rail rapid transit system depending on their route through the transit network. In contrast, the HBW mode choice model for the Seattle region (Cambridge Systematics 2003) has five modes and one composite alternative at the top level: auto, two-person shared-ride, three or more person shared-ride, walk access to transit, walk to destination, and bicycle, as shown in Figure 7. The auto composite alternative is split into drive alone and auto access to transit at the second level. This structure im- plicitly assumes that the disutility involved in choosing to make an auto trip depends on a lower-level choice of whether to drive alone to the destination or to an interme- diate transit stop or station. In the case of automobile ac- cess to transit, the perceived disutility of the alternative will include the effect of the travel time and cost involved in riding transit. The HBW mode choice model for the San Francisco Bay Area (Purvis 1997) lies somewhere between the Atlanta and Seattle models, as shown in Figure 8. At the top level there is a choice between motorized modes (both automobile and transit), bicycle, and walking to the destination. The motor- ized modes are then split into three auto modes (drive alone, FIGURE 6 Home-based work mode choice structure—Atlanta region. (Source: Travel Demand . . . 2005).

two-person shared-ride, and three or more person shared- ride) and transit. At the third level, transit trips are split into automobile access and walk access. There is no explicit dis- tinction between the use of local bus or regional rail services (BART or commuter rail); however, the use of rail depends on the path taken through the transit network. The HBW models for the Dallas/Fort Worth region (Cam- bridge Systematics n.d.; Dallas–Fort Worth . . . 2007) and for the Metropolitan Washington region (COG/TPB . . . 2007) both have a somewhat simpler structure than the other three models, with only three alternatives at the top level, one of which is split into three modes at the second level. However, the modes at each level are different. The NCTCOG model has an auto composite alternative at the top level, together with walk access to transit and auto access to transit as sepa- rate modes. The auto composite alternative is then split into three modes at the second level: drive alone, two-person shared-ride, and three or more person shared-ride. The Met- ropolitan Washington Council of Governments (MWCOG) model has the following three alternatives at the top level: transit, drive alone, and shared-ride auto (termed group ride). As described previously, access to transit is not explicitly modeled as a mode choice, but transit trips are assumed to in- volve either walk access or automobile access depending on the distance to the nearest bus stop or station. The shared-ride alternative is then split into three carpool/vanpool modes at the second level for different vehicle occupancy: two-person shared-ride, three-person shared ride, and four or more per- 62 son shared-ride. The split into three carpool/vanpool modes is done with a separate carpool occupancy model that is esti- mated independently from the top-level model. The resulting vehicle occupancy information is then used in the main model to determine the average costs for the shared-ride mode for each person. Variables Used in Models The range of explanatory variables included in the utility functions of the HBW models varies from region to region, as indicated by Table 6. Furthermore, not all variables used in a given model are used in the utility function for each mode because many of the variables, such as waiting time or bicycle travel time, are only relevant for some modes. Other variables that are not mode-specific, such as measures of ve- hicle ownership, have been found to provide a statistically significant contribution to the explanatory power of the model when included in the utility function for some modes but not others. All five models include measures of in-vehicle travel time, waiting time, walking time (or walking distance), and travel cost, although the coefficients for travel time and cost in the Atlanta model were adopted from other regional mod- els rather than being estimated from local data. Some models used the same travel time or travel cost coefficients for each mode, whereas others estimated different coefficient values HBW Person Trips Auto Drive Alone Auto Access to Transit Shared Ride 2 Shared Ride 3+ Transit-Walk Access Walk Bicycle FIGURE 7 Home-based work mode choice structure—Seattle region. (Source: Cambridge Systematics 2003.) Drive Alone Shared Ride 2 Shared Ride 3+ Transit Auto Acc Transit Walk Acc Bicycle Walk FIGURE 8 Home-based work mode choice structure—San Francisco Bay area. (Source: Purvis 1997.)

63 for different modes, typically distinguishing between transit and automobile use. Similarly, some models estimated sepa- rate model coefficients for different travel time components, such as waiting or walking, whereas others assumed that these coefficients were a fixed multiple of the in-vehicle travel time coefficient and only estimated a single coefficient for the weighted travel time. If the assumed weighting factor for the out-of-vehicle time components is incorrect, this will bias the estimated values of the coefficients for each of the travel time components. Only three of the regional models included household in- come in the model in some form, and the approach varied across the models. The Bay Area model used a continuous variable for household income up to a maximum value of $25,000 in 1989 dollars. The NCTCOG model used dummy variables for low- and high-income households, defined as those with incomes of less than $30,000 or more than $75,000 in 1996 (when the household travel survey used to estimate the model was performed). Separate values of the dummy variable coefficients were estimated for each mode. The Atlanta model stratified the model estimation sample into four household income groups, with household incomes in 2000 dollars of less than $20,000, $20,000 to $49,999, $50,000 to $99,999, or $100,000 or more. Separate ASCs were estimated for each income group for each mode apart from local bus and rail rapid transit. Estimating separate ASCs is effectively the same as using a dummy variable for those income groups. Two of the models considered employment density in the destination zone, whereas a third model used an index of land-use mix in both the origin and destination zone to explain differences in mode use behavior in different areas of the region. Four of the models used dummy variables for certain sub-areas within the region, although the way that these were defined and how extensively they were applied varied widely. The use of sub-area dummy variables pro- vides a way to adjust the model predictions to better match observed mode shares in specific areas. Although this im- proves the fit of the model to the estimation data, it also im- plies that the other variables in the model are not adequately explaining the observed choice behavior for those areas. This could in turn lead to forecasting errors when the model is used to predict travel patterns in future years when the conditions or influences reflected by the other variables are expected to change. Because of the differences in the way that the variables are defined in each model and of the differences in func- tional form, the specific values of the coefficients are not directly comparable. Similarly, differences in the years when the data were collected from which the models were estimated, as well as differences in the way that household income was (or was not) included in the models, means that implied values of travel time are also not directly com- parable. The work required to reexpress this information on a consistent basis is beyond the scope of the current study. Summary Although the current state of practice of regional travel demand modeling is fairly consistent in broad terms among large metropolitan planning agencies, the details of how the components of this process are implemented vary widely from region to region. Therefore, the use of these models to study airport employee mode choice in a specific region should be approached with caution. There is a need for sig- nificantly more work to determine how the differences in the models could affect their ability to predict journey-to-work travel behavior of specific sub-groups of the larger regional population, such as airport employees. MODES INCLUDED IN MODEL Generally, the modes included in airport employee mode choice models are more limited than those for air passenger mode choice models, because many of the modes used by air passengers are not relevant to airport employee journey-to- work trips. Airport employees will not usually consider rent- ing a car to get to work (except perhaps in the rare case where their own car is not available), whereas modes such as taxi and limousine are too expensive for use on a regular basis. Although modes such as scheduled airport bus or shared-ride door-to-door van will often be too expensive to be a viable option, the cost of using these modes may be affordable if discounted fares are available to airport employees. On the other hand, airport employees may have access ser- vices available to them that are not available to air passengers, such as dedicated employee buses or van pools. Unlike air passengers, who generally cannot arrange to share a ride to the airport with another air party other than by the use of com- mercial shared-ride services, airports or airport employers often establish ride-matching services to arrange car or van pools, or airport employees may be able to make use of local ride-sharing organizations. Given the interest of many regional transportation plan- ning agencies in modeling shared-ride trips, as well as that in many regions the number of vehicle occupants affects the ability to use HOV lanes or avoid paying tolls, it is desirable that airport employee mode choice models identify the num- ber of occupants in car or van pools. This is necessary any- way to be able to convert from employee trips to vehicle trips. However, where the number of occupants that are re- quired to use HOV lanes varies for different facilities in the region, then determining an average vehicle occupancy for shared-ride trips is not sufficient and it will be necessary to distinguish between the occupancy categories that allow the

use of each facility (typically shared-ride occupancy criteria are either two or more people or three or more). Although there may be no reason to distinguish between formal van pools and car pools with three or more people from the per- spective of HOV lane use, it may be desirable for other reasons, such as evaluating the effectiveness of ride-sharing programs or because van pool vehicles are provided by the airport or employer. Although airports are typically located sufficiently far from residential areas that walking is not usually a feasible access mode for the entire journey to work (as distinct from the access or egress trip from other modes), some employees may cycle to work, particularly if the airport provides facili- ties to encourage this. Indeed, the provision of such facilities may form part of airport ground access trip reduction or mit- igation programs. Therefore, a fairly comprehensive airport employee mode choice model might include the following modes: • Single-occupant private vehicle • Shared-ride private vehicle—two occupants • Shared-ride private vehicle—three or more occupants • Van pool • Charter bus or van • Public transit • Scheduled airport bus service • Bicycle. Depending on the application, it may also be desirable (or even necessary to estimate the model coefficients ade- quately) to include sub-modes such as different parking lo- cations, different transit services, and different access or egress modes to or from the stations or stops used for public transit or other fixed-route services, as with air passenger mode choice models. Although mode choice models devel- oped for general urban travel are increasingly incorporating a wider range of modes and many explicitly considering how transit users will access transit services, very few cur- rent models include the range of modes that are typically available to employees at a large airport. Even so, as a wider range of modes is included in the more sophisticated re- gional travel demand models, including different types of shared-ride trip and nonmotorized modes such as cycling, these may have an adequate degree of resolution to model airport employee trips. As the number of different modes and sub-modes in- cluded in the model increases, so the structure of the model becomes an important issue. The MNL model is predicated on the assumption that the utility of each mode is indepen- dent of that of other modes, which is clearly not the case with many of the modes discussed earlier. This difficulty can be overcome with the use of a NL model. However, this raises the question of how best to group the modes into different nests. This is an issue that deserves further research, as evi- 64 denced by the widely different model structures used in dif- ferent regional travel demand models, which is discussed later in this chapter. Because of the timing of shift start and end times, air- port employee travel may well occur at times when some modes, particularly public transportation services, are not available. This will need to be reflected in the mode choice model through constraints on mode availability. Shift times will also constrain the ability of employees to find a car or van pool at the appropriate time, and this needs to be reflected in how these modes are represented in the model. EXPLANATORY VARIABLES As with air passenger airport access mode choice models, airport employee mode choice models will need to include travel time and cost variables. Although costs can be aggre- gated into a single variable, the different components of travel time (walking, waiting, and in-vehicle time) should be expressed as separate variables to allow for the possibility of different perceived disutility of each component. An important consideration in determining the values of explanatory variables is how to represent travel constraints arising from the timing of shifts. Where highway travel times or transit travel times, service frequency, or fares vary sig- nificantly by time of day, it will be necessary to ensure that the appropriate values are used for each employee. It may be desirable to have separate travel time variables for use of public transportation late at night to reflect differences in per- ceived disutility owing to concerns for personal security, in addition to greater travel time as a result of infrequent ser- vice. However, the majority of regional travel demand mod- els provide limited resolution for differences in travel times by time of day. It is not uncommon to only model the morn- ing commute peak and average off-peak conditions, which will not distinguish between mid-day and late night condi- tions. In this situation, it will not be possible to obtain the appropriate travel times directly from the regional travel demand model data files and significant adjustments may be needed to appropriately represent late night travel times on public transportation. Where employees pay for parking (or a transit pass) through a payroll deduction, it may be desirable to use a separate variable for this cost component to allow for the possibility that employees may perceive this cost differ- ently from a regular out-of-pocket expense. In the case of a transit pass, there is also the benefit that the pass can usu- ally be used for other transit trips in addition to the journey to work. Employee income level is likely to be a significant factor in mode choice and an appropriate variable should be included in

65 the model. Employee wage rate is likely to be a more relevant measure than household income, because the latter may be unduly influenced by the income of other workers in the household. MODELING CONSIDERATIONS Development of airport employee access mode choice mod- els requires both survey data on existing employee mode use patterns and service data on the costs and travel times in- volved in using the alternative modes available. Much of the modal service data, such as highway travel times and transit travel times and frequencies, are the same as that required for developing air passenger airport access mode choice models. However, additional data will have to be collected on employee-specific costs, such as parking fees, transit passes, and employee discounts on scheduled airport bus services, as well as information on employer-provided services such as free parking or employee charter buses. As with air passenger airport access mode choice, the range of factors that influence the mode choice decisions of a given employee are sufficiently individual-specific that any reasonable model will have to be disaggregate in nature and predict the probability of a given employee choosing a par- ticular mode. This implies that such a model can only be developed from fairly detailed survey data of a large enough sample of employees. In addition to information on the existing mode use, the survey needs to collect data on all the factors that might reasonably influence that choice, including the respondent’s shift patterns, type of work, residence loca- tion, wage or salary level, and eligibility for any special benefits, such as free parking or transit passes, even if the respondent is not actually using that benefit. It will also be necessary to determine the respondent’s employer (or at least the type of employer), because this is important in expanding the model to predict airport access mode use patterns for all airport employees, not just the re- spondents to the survey. Because the survey response rate is likely to vary across employers or even departments within an organization, reflecting different managerial support for performing the survey or even willingness to participate, it will be necessary to develop expansion factors for each respondent based on the total employment within each orga- nization or unit. Fortunately, the requirement for most airport employees to have security badges means that the airport authority usually has good data on the number of employees in each category. Given the very limited experience developing employee access mode choice models, it is unclear how large a sample size would be necessary to estimate a reasonably accurate model. This is yet another aspect that would benefit from fur- ther research. Until this issue is better understood, based on experience with air passenger access mode choice models, an airport undertaking an employee survey for use in develop- ing such a model should attempt to obtain as large a sample as possible, in the range of 1,000 to 3,000 responses. Although not strictly an airport employee mode choice modeling issue, the application of an airport employee mode choice model requires an airport employee trip generation model that predicts the number of airport employees living in each TAZ and generates a representative sample of airport employee trips with their associated characteristics. Unless the employee survey responses are a very large sample of total employment, simply factoring up the employee charac- teristics from the survey by the ratio of the total airport em- ployment to the survey responses will result in a trip genera- tion pattern that has a large number of employees with identical characteristics from some TAZs and no employee trips from other zones. This will tend to overstate the poten- tial for ridesharing, as well as bias the predicted mode use from any given zone. This issue is not that different from regional travel de- mand models, which typically include a trip generation model component. However, the journey-to-work trip gener- ation models in regional travel demand models cover all types of employment and typically do not generate individ- ual traveler characteristics, much less those appropriate to airport employees (such as shift start and end times). There- fore, it may be desirable to generate a synthetic sample of trips using Monte Carlo simulation methods based on the distribution of airport employee characteristics obtained from an employee travel survey. It is common practice in regional travel demand modeling to treat the airport as a single TAZ. This is reasonable within the context of the overall regional travel demand, but may not be adequate for airport planning studies, particularly where different areas of the airport may be served by different ac- cess roads. Therefore, it may be desirable to subdivide the airport into different zones and identify the level of airport employment in each zone as well as use a more refined rep- resentation of the transportation network in the vicinity of the airport. Although this will not significantly change travel times to trip origin TAZs at some distance from the airport, it may affect travel times for those employees who live closer to the airport and will be necessary to properly assign to trips to the different access routes. MODEL PERFORMANCE Given the limited experience with special-purpose airport employee mode choice models, there is no basis for assess- ing the likely performance of such models. Assessments of the overall performance of general regional travel models shed little light on their performance at predicting airport employee mode choice, because airport employees form such a small percentage of the total regional journey-to-work

travel. It would of course be possible to compare the regional travel model predictions of journey-to-work mode use for the airport TAZs with data from airport employee surveys. How- ever, no such studies have been identified. This would appear to be a promising area for further research, because it is fairly simple to do and would provide valuable information on the extent to which existing regional travel models can be used to predict airport employee mode use. 66 As with air passenger airport access mode choice models, assessment of the performance of existing models is critical to understanding how much reliance can be placed on their predictions and identifying the need for further improvement in these models. However, before the performance of special- purpose airport employee mode choice models can be as- sessed, it is first necessary to develop the models to assess. This is another area that could benefit from further research.