California statewide model for high-speed rail

Outwater, Maren et al. Article California statewide model for high-speed rail Journal of Choice Modelling Provided in Cooperation with: Journal of Choice Modelling Suggested Citation: Outwater, Maren et al. (2010) : California statewide model for high-speed rail, Journal of Choice Modelling, ISSN 1755-5345, University of Leeds, Institute for Transport Studies, Leeds, Vol. 3, Iss. 1, pp. 58-83 This Version is available at: https://hdl.handle.net/10419/66842 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by-nc/2.0/uk/ Journal of Choice Modelling, 3(1), pp. 58-83 www.jocm.org.uk California Statewide Model for High-Speed Rail Maren Outwater1,* Kevin Tierney2,† Mark Bradley3,Ŧ Elizabeth Sall4,§ Arun Kuppam5,¶ Vamsee Modugula6,β 1 Resource Systems Group, 917 E. Spooner Rd, Fox Point, WI 53217 2 206 Broad Meadow Rd, Needham, MA 02492 3 Mark Bradley Research and Consulting, 524 Arroyo Avenue, Santa Barbara, CA 93109 4 San Francisco County Transportation Auth, 100 Van Ness Ave, San Francisco, CA 94102 5 Cambridge Systematics, 9015 Mountain Ridge, Suite 210, Austin, TX 78759 6 Citilabs, A-2/14, Vanashree CHS, Plot No 1&2, Sector 58A, Nerul, Navi Mumbai, 400706 Received 19 March 2008, received version revised 22 December 2008, accepted 19 September 2009 Abstract The California High Speed Rail Authority (CHSRA) and the Metropolitan Transportation Commission (MTC) have developed a new statewide model to support evaluation of high-speed rail alternatives in the State of California. The approach to this statewide model explicitly recognizes the unique characteristics of intraregional travel demand and interregional travel demand. As a result, interregional travel models capture behavior important to longer distance travel, such as induced trips, business and commute decisions, recreational travel, attributes of destinations, reliability of travel, party size, and access and egress modal options. Intraregional travel models rely on local highway and transit characteristics and behavior associated with shorter distance trips (such as commuting and shopping). Keywords: discrete choice models, interregional, high-speed rail, induced demand, ridership forecasting, trip frequency, destination choice, mode choice * Corresponding author, T: +1 425 269 9684, F: + 011 802 295 1006, [email protected] †T: +1 617 839 5295, [email protected] Ŧ T: +1 805 564 3908, F: + 011 805 564 3927, mark_bradl[email protected] § T: +1 919 469 3609, F: + 011 919 302 0265, [email protected] ¶ T: +1 512 691 8503, F: + 011 512 691 3289, [email protected] β T: +91 022 27525025, F: + 011 510 523 9706, [email protected]m Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 1 Introduction By 2030, California’s population will grow to 50 million people, which will nearly double interregional travel to one billion trips per year. High-speed trains are being considered to alleviate the need to build – at a cost of nearly $100 billion – about 3,000 miles of new freeway, plus five airport runways, and 90 departure gates over the next two decades. Electric and fully separated from automobile traffic, California’s high-speed train would provide a new transportation option available to more than 90 percent of the residents of the state. At full build-out, the system will run from San Diego north to Sacramento and San Francisco. The project objectives were to develop a new ridership forecasting model that would serve a variety of statewide planning and operational purposes: • To evaluate high-speed rail ridership and revenue on a statewide basis; • To evaluate potential alternative alignments for high-speed rail into and out of the San Francisco Bay Area; and • To provide a foundation for other statewide planning purposes and for regional agencies to better understand interregional travel. The core model design feature is the recognition that interregional and urban area travel is distinct and should be modelled separately to capture these distinctions accurately. This led to our approach to develop separate, but integrated, interregional and intraregional models, as demonstrated in Figure 1. There are two primary reasons for developing separate models for interregional and urban area travel: first, the trip purposes are different and second, the interregional travel models need to explicitly estimate induced demand. These models are applied to both peak and off-peak conditions for an average weekday. Weekend travel demand and annual ridership estimates are developed using annualization factors developed from observed data on high-speed rail systems around the world. Figure 1. Integrated Modeling Process 59 Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 60 There are fourteen regions established in the state that define interregional and intraregional travel: • Interregional models estimate trip frequency, destination choice, and mode choice stratified by trip purpose (business, commute, recreation, and other) as well as by distance (trips greater than or less than 100 miles) and by trip type (trips made by residents of the four largest cities in California versus other trips). The interregional trip frequency models estimate induced travel based on improved accessibilities due to high speed rail options. The interregional models are similar in structure to models developed for the Australian Very Fast Train (VFT) project, except that there is more spatial detail (i.e. many more zones) and all models were estimated based on revealed preference with full and consistent nesting of the accessibility measures from access/egress models up through trip frequency models. There are also some similarities to the Norwegian and Swedish national models, which combine long-distance and short-distance travel and have similar modeling structures but have estimated simultaneous destination and mode choice with sampled destinations. • Intraregional models are based on trip tables generated from the Metropolitan Planning Organization (MPO) models and estimate mode choice of urban area trips. These mode choice models reflect local urban area highway and transit systems as well as options for high speed rail within the region. Urban travel is stratified by trip purpose (work, school, college, other, and non-home-based). The interregional and intraregional area models are based on travel survey data collected for these purposes. 2 Literature Review Although few intercity high speed ground transportation (HSGT) systems have been implemented in the United States, planners have been performing ridership forecasting analyses and benefits assessments for proposed systems for more than thirty years. Almost all of the ridership forecasts have been in support of analyses of one of the 11 designated intercity corridors authorized by the Federal Railroad Administration in the Intermodal Surface Transportation Efficiency Act (ISTEA) of 1991 and the Transportation Equity Act for the 21st Century (TEA-21) of 1998, including (Federal Railroad Administration 2006): • Northern New England; • Northeast Corridor; • Empire Corridor; • Keystone Corridor; • Southeast Corridor; • Gulf Coast Corridor; • Florida Corridor; • Chicago Hub Network; • South Central Corridor; • Pacific Northwest Corridor; and • California Corridor. Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 61 The first U.S. HSGT ridership forecasts were developed to support investments to improve the Northeast Corridor (Koppelman et al. 1984). Since then, a variety of forecasting techniques have been used in corridor feasibility analyses, ridership and revenue evaluations, and environmental impact studies. The literature describing the analytical techniques employed in these various studies and the current states of forecasting practice over time have been reviewed by Koppelman et al. (1984), Miller (1992, pp. 378-389), Forinash (1992), and Horowitz (2006). The initial forecasting efforts relied on aggregate direct demand models, such as the Quandt and Baumol abstract mode model (1966), the Rand Corporation model developed by DeFerranti et al. (1973), and the Peers and Bevilacqua structural intercity model (1976). As urban modelers adopted the four-step modeling process for regional forecasting, intercity models also became multi-step processes where the intercity trip tables are estimated and forecast, and separate mode choice models are used to determine modal shares under different service scenarios. Most of the more recent U.S. HSGT ridership forecasting efforts can be characterized by the specific analyses used to develop intercity trip tables, determine modal split, and the level to which the different model steps are integrated. Modelers have employed many trip generation and trip distribution techniques to forecast future year intercity trip tables. For an early analysis of the Florida high speed rail corridor, PRC Voorhees used trip purpose specific growth factor models to forecast trip tables (1983). Other modelers, such as Thakuriah et al. (1999) and Cohen et al. (1978), have relied on gravity models for trip table prediction. One researcher identified by Horowitz used a time series model to forecast future intercity travel demand (Yu 1970). Many others have used a cross-sectional regression-based total travel demand (direct demand) formulation, including for instance, Booz-Allen and Hamilton (1989), TMS/Benesch (1991), and Brand et al. (1992). The previous analysis of the California high speed rail corridor relied on this modeling approach (Charles River Associates 2000). Since the early 1980’s, several different forms of disaggregate mode choice models have been developed for forecasting U.S. HSGT ridership. Early efforts, such as those by Grayson (1981), employed multinomial logit models, but subsequent models have expanded the technique in different directions. Cohen et al. (1978, pp. 21-25), Brand et al.(1992, pp. 12-18), and Charles River Associates (2000) formulated mode choice as separate binary diversion models where percentages of auto, air, and bus passengers are diverted to HSGT individually through binary models. This approach seeks to eliminate the troublesome outcome of the multinomial logit model’s IIA property. Other researchers have relied on the use of nested logit models for intercity mode choice (TMS/Benesch 1991 and Forinash and Koppelman 1993). Researchers, most notably Chandra Bhat, have experimented with a wide variety of nested logit model formulations and variable combinations (1995) (1997) (1998). With advancements in effective ways for combining revealed preference and stated preference survey data (Ben-Akiva and Morikawa 1990) (Bradley and Daly 1997), intercity modelers have also begun to use stated preference survey methods to a greater extent for forecasting HSGT mode choice. Most intercity transportation demand models have treated traveler’s decisions regarding trip frequency, destination choice, mode choice, and route selection as separate sequential choices. The model system described in this paper seeks to connect the different model components by passing information from one choice component to the others during model development. When the multinomial logit and nested logit model components are applied, they are fully consistent with each other and are sensitive to each other’s changes. Integrated modeling approaches have been proposed Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 62 by and implemented by Koppelman and Hirsh (1986), Koppelman (1989), Proussaloglou and Tierney (1999) and, outside the U.S., by Algers (1993) and Gunn et al. (1992). The latter approach was developed for an Australian VFT study and uses a very similar approach to the models developed for California, including the use of both stated and revealed preference data and linking mode, destination and trip frequency models through the use of accessibility measures. 3 Interregional Models The interregional models are comprised of four sets of models: trip frequency, destination choice, main mode choice, and access/egress mode choice. The structure and contents of the interregional modelling system is presented in Figure 2. The trip frequency model component predicts the number of interregional trips that individuals in a household will make based on the household’s characteristics and location. The destination choice model component predicts the destinations of the trips generated in the trip frequency component based on zonal characteristics and travel impedances. The mode choice components predict the modes that the travelers would choose based on the mode service levels and characteristics of the travelers and trips. The mode choice models include a main mode choice, where the primary interregional mode is selected, and access/egress components, where the modes of access and egress for the air and rail trips are selected. A combined destination and mode choice model was initially considered, but the choice set for destination choice was all zones (4,667) and the combination of destinations and modes would have made this choice set too large. 3.1 Data for Estimation There were three types of data compiled for the study: travel surveys, networks, and socioeconomic data. Some of the travel surveys were collected specifically for this study, three were available from MPOs around the state (Southern California Association of Governments (SCAG), San Francisco Metropolitan Transportation Commission (MTC), and Sacramento Association of Governments (SACOG)), and there was a Caltrans statewide survey available. The interregional models were based on revealed and stated preference surveys, collected specifically for this study, of air and rail travelers, as well as additional households in the state to capture auto travelers. These new data were collected in fourteen regions in California. These were combined with revealed preference surveys of households across the state collected by Caltrans and interregional travel extracted from the MPO regional travel surveys (San Francisco, Sacramento, and Los Angeles). Intraregional mode choice models were based on urban area travel surveys in combination with a stated preference survey for high speed rail conducted in Los Angeles. By combining the various available data sources, we were able to provide more robust data sets for model estimation than was otherwise possible. A summary of these sources shows the trip records derived from each: Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 Trio Frequency/Day •Household Characteristics •Trip Purpose/Distance Class •Level of Service (Logsum & Accessibility •Region •Party Size (For Short Distance) Destination Choice •Level of Service (Logsum & Accessibility •Employment & Household Characteristic s •Region and Area Type •Trip Purpose/Distance Class •Party Size (For Long Distance) Main Mode Choice •Level of Service •Household Characteristics •Purpose/Distance Class •Party Size (For Long Distance) •Access & Egress (Logsum) Access Mode Choice •Level of Service •Household Characteristics •Purpose/Distance Class •Party Size (For Long Distance) •Main Mode (Rail/HSR/Air) Egress Mode Choice •Level of Service •Household Characteristics •Purpose/Distance Class •Party Size (For Long Distance) •Main Mode (Rail/HSR/Air) One Trip Two-Plus Trips No Trips Zone 1 Zone 2 Zone N-1 Zone N Car Rail HSR Air Drive and Park Drop Off Rental Car Transit WalkTaxi Transit WalkTaxi Rental CaPicked Up Unpark and Drive Figure 2. Interregional Model Structure Figure 2. Interregional Model Structure 63 Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 • Air, rail, auto passenger surveys 2678 trips • Air, rail, auto passenger surveys 2678 trips • Caltrans travel surveys 2820 trips • Caltrans travel surveys 2820 trips • SCAG travel surveys 343 trips • SCAG travel surveys 343 trips • MTC travel surveys 723 trips • MTC travel surveys 723 trips • SACOG travel surveys 318 trips • SACOG travel surveys 318 trips After combining these surveys, 6,882 completed surveys were available to use for model estimation, as shown in Table 1. There was different estimation datasets used for each model component, depending on the requirements for the model. This is described in more detail in the Interregional Model System Development Report (Cambridge Systematics 2006). After combining these surveys, 6,882 completed surveys were available to use for model estimation, as shown in Table 1. There was different estimation datasets used for each model component, depending on the requirements for the model. This is described in more detail in the Interregional Model System Development Report (Cambridge Systematics 2006). There are highway, air, rail, and local transit networks to support both the urban area and interregional travel models. The socioeconomic data includes household data in four classifications (household size, income groups, number of workers, and vehicle ownership) and employment data by type. There are highway, air, rail, and local transit networks to support both the urban area and interregional travel models. The socioeconomic data includes household data in four classifications (household size, income groups, number of workers, and vehicle ownership) and employment data by type. 3.2 Accessibility Measures 3.2 Accessibility Measures In the development of the trip frequency models, accessibility measures were estimated for all trips to approximate the destination choice logsum measure. In the final models, accessibility measures were retained for intraregional trips because the intraregional models maintained by the MPOs do not include destination choice models, which are necessary to produce logsum measures. Accessibility measures for interregional trips were replaced with logsum measures from the destination choice models in the final models, as described below. There were four accessibility measures calculated, as follows: In the development of the trip frequency models, accessibility measures were estimated for all trips to approximate the destination choice logsum measure. In the final models, accessibility measures were retained for intraregional trips because the intraregional models maintained by the MPOs do not include destination choice models, which are necessary to produce logsum measures. Accessibility measures for interregional trips were replaced with logsum measures from the destination choice models in the final models, as described below. There were four accessibility measures calculated, as follows: • Auto peak work trip accessibility • Auto peak work trip accessibility ⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎣ ⎡∑ ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ −+= d TimeTimeTotalEmp LN A mean p ea k auto p ea k d auto p ea k _ _ _ / × 2exp × 1 Table 1. Total of Survey Interregional Trips by Mode, Distance, and Purpose Table 1. Total of Survey Interregional Trips by Mode, Distance, and Purpose Drive Drive Air Air Rail Rail Bus Bus Other Other Total Total Long Trips Business 314 620 27 18 17 996 Commute 263 15 9 1 74 362 Recreation 1114 228 80 3 23 1448 Other 365 85 17 8 91 566 Short Trips Business 381 14 48 3 15 461 Commute 1136 0 168 9 108 1421 Recreation 873 2 29 3 52 959 Short Other 591 1 10 23 44 669 Total 5,037 965 388 68 424 6,882 Source: Bay Area/California High Speed Rail Ridership and Revenue Forecasting Study Interregional Model Systems Development Report, Table 2.6, August 2006. 64 Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 • Auto off-peak non-work trip accessibility • Auto off-peak non-work trip accessibility ( ) ⎥ ⎦ ⎤ ⎢ ⎣ ⎡−+++= ∑ d meanoffpea k autooffpea k d d d auto offpea k TimeTime ServiceEmp R etailEmp H ouseholds LN A _ _ _ / × 2exp × )(1 • Non-Auto peak work trip accessibility • Non-Auto peak work trip accessibility ⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎣ ⎡ ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ −+ =∑ d mean p ea k nonauto p ea k d nonauto p ea k TimeTimeTotalEmp LN A _ _ _ / × 2exp × 1 • Non-Auto off-peak non-work trip accessibility • Non-Auto off-peak non-work trip accessibility ( ) ⎥ ⎦ ⎤ ⎢ ⎣ ⎡−+++ =∑ d meanoffpea k nonautooffpea k d d d nonauto offpea k TimeTimeServiceEmp R etailEmp H ouseholds LN A _ _ _ / × 2exp*)( 1 where: where: TotalEmpd = total employment at the destination zone; TotalEmpd = total employment at the destination zone; Householdsd = total households at the destination zone; Householdsd = total households at the destination zone; RetailEmpd = retail employment at the destination zone; RetailEmpd = retail employment at the destination zone; ServiceEmpd = service employment at the destination zone; ServiceEmpd = service employment at the destination zone; Timepeak_auto = highway travel time during the peak (based on congested time) from the origin zone to the destination zone; Timepeak_auto = highway travel time during the peak (based on congested time) from the origin zone to the destination zone; Timepeak_nonauto = transit travel time during the peak (based on congested time) from the origin zone to the destination zone; Timepeak_nonauto = transit travel time during the peak (based on congested time) from the origin zone to the destination zone; Timeoffpeak_auto = highway travel time during the off-peak (based on free-flow travel time) from the origin zone to the destination zone; Timeoffpeak_auto = highway travel time during the off-peak (based on free-flow travel time) from the origin zone to the destination zone; Timeoffpeak_nonauto = transit travel time during the off-peak (based on free-flow travel time) from the origin zone to the destination zone; Timeoffpeak_nonauto = transit travel time during the off-peak (based on free-flow travel time) from the origin zone to the destination zone; Timepeak_mean = average travel time from the origin zone to all possible destination zones during the peak period, calculated from the average of survey respondents travel time based on peak network times; and Timepeak_mean = average travel time from the origin zone to all possible destination zones during the peak period, calculated from the average of survey respondents travel time based on peak network times; and Timeoffeak_mean = average travel time from the origin zone to all possible destination zones during the off-peak period, calculated from the average of survey respondents travel time based on off-peak network times. Timeoffeak_mean = average travel time from the origin zone to all possible destination zones during the off-peak period, calculated from the average of survey respondents travel time based on off-peak network times. 3.3 Logsum Measures 3.3 Logsum Measures Logsum measures are a means to estimate a weighted average of travel time and cost that can be fed back from one component to another. A summary of the logsum measures for each model component is as follows: Logsum measures are a means to estimate a weighted average of travel time and cost that can be fed back from one component to another. A summary of the logsum measures for each model component is as follows: • Trip frequency models use “logsum” measures from the destination choice models, which are intended to capture the fact that it is easier to make relevant interregional trips from some zones than from other zones. For initial model estimation, a synthesized network zone accessibility measure was used. • Trip frequency models use “logsum” measures from the destination choice models, which are intended to capture the fact that it is easier to make relevant interregional trips from some zones than from other zones. For initial model estimation, a synthesized network zone accessibility measure was used. • Destination choice models use logsum measures from the main mode choice models that are intended to provide measures of the composite impedance across • Destination choice models use logsum measures from the main mode choice models that are intended to provide measures of the composite impedance across 65 Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 72 Table 4. Access and Egress Mode Choice Models for Long Trips1 Access Models Egress Models Business/Commute Recreation/Other Business/Commute Recreation/Other Observations 1,500 2,724 1,466 2,668 Final log-likelihood -1,662.3 -2,519.4 -2,121 -3,066.6 ρ2(0) 0.276 0.365 0.075 0.231 ρ2(cons) 0.003 0.068 -0.023 0.053 Coeff. (t-stat) Coeff. (t-stat) Coeff. (t-stat) Coeff. (t-stat) Level of Service Cost ($) -0.075 constrained -0.120 constrained -0.075 constrained -0.120 constrained In-vehicle time (min) -0.060 constrained -0.030 constrained -0.060 constrained -0.030 constrained Out of vehicle time (min) -0.147 (-6.4) -0.083 (-2.5) -0.139 (-6.2) -0.060 constrained VOT IVT ($/hour) $48.00 $15.00 $48.00 $15.00 Ratio OVT/IVT 2.45 2.76 2.33 2.00 Drive and (un)park Travel alone -1.925 (-3.0) Fewer cars than persons -1.547 (-2.2) -1.903 (-2.8) Low income -2.741 (-1.8) -1.960 (-2.8) -18.006 (-2.5) -1.263 (-1.1) High income 0.709 (1.6) 0.339 (1.4) To/from conventional rail -9.490 (-2.5) To/from high-speed rail -2.251 (-1.8) Airport is LAX -3.128 (-3.8) -1.275 (-1.7) Airport is SFO -4.082 (-4.4) -3.036 (-2.6) Airport is SJC -1.479 (-2.1) Airport is SAN -1.410 (-2.3) -1.370 (-2.3) Rental car To/from conventional rail -5.0002 constrained -5.000 constrained -3.522 (-2.4) -1.176 (-3.1) To/from high speed rail -0.552 (-2.4) No cars in HH 5.110 (3.2) High income 2.953 (2.4) Travel alone -2.588 (-4.7) Low income -2.082 (-0.9) -1.891 (-3.7) Get dropped off/picked up In-vehicle time (min) -0.014 (-2.5) -0.031 (-3.1) -0.015 (-3.9) Household size 0.606 (2.9) 0.478 (2.8) 0.974 (2.8) Taxi Auto distance -0.084 (-4.8) -0.071 (-3.8) -0.126 (-7.9) -0.052 (-6.6) To/from conventional rail -2.827 (-2.6) -2.265 (-2.4) To/from high-speed rail -1.092 (-2.1) 2.507 (3.6) Travel alone -0.877 (-1.8) -2.768 (-4.6) Low income -3.010 (-1.9) -3.002 (-2.3) -1.038 (-2.3) High income 0.849 (1.9) Transit No walk egress -4.836 (-4.6) -1.807 (-1.9) Rail used in path 3.689 (5.2) 1.727 (2.4) 2.960 (5.0) To/from conventional rail 3.580 (5.2) 1.830 (2.8) To/from high-speed rail 0.592 (0.7) 1.032 (1.9) Travel alone 1.569 (2.3) No cars in HH 1.439 (1.7) Fewer cars than persons 1.480 (2.1) Low income 0.846 (1.0) 1.216 (1.9) Walk To/from airport -5.0002 constrained -2.634 (-1.0) -2.074 (-2.0) Nesting and scaling Nesttransit, walk, taxi 0.387 (5.9) 0.451 (3.3) 0.280 (6.9) 0.470 (5.3) Scale on hypothetical choices 0.682 (15.9) 1.000 constrained 0.516 (9.8) 1.000 constrained 1 Does not include alternative specific constants. 2 These were later reduced to -3.0 during model calibration. Source: Bay Area/California High Speed Rail Ridership and Revenue Forecasting Study Interregional Model Systems Development Report, Table 3.12 and 3.13, August 2006. Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 • The long segments, taxi, parking, and rental cars are generally less desirable to rail stations than to airports, while transit is more desirable from rail stations. Walking is very rare to or from airports, capturing accessibility affects that are not captured well in the zone system. • Drive and park access is less likely at the busiest airports – San Francisco (SFO), Los Angeles (LAX), and San Diego (SAN) – and somewhat at San Jose (SJC) as well. This may capture both cost and inconvenience effects at those airports. • For most segments, those in larger households are more likely to be dropped off. • In general, high income favors rental car, taxi, and drive and park, and low income slightly favors transit in some segments. • There is a logsum coefficient less than 1.0 on the nest that includes transit, walk, and taxi. Each of the other three alternatives is in its own “nest,” and scaled by the same logsum parameter to preserve equal scaling at the elemental level. • The scale (the inverse of the residual error variance) for the hypothetical choices relative to the actual choices was significantly lower than 1.0 for most of the Egress model segments. This result indicates that many respondents have difficulty making an accurate assessment of mode choice options in less familiar surroundings at the non-home end of their trip, so that hypothetical choices should be weighted less in estimation than actual ones. The main mode choice models produce probabilities that each trip will choose one of the main modes (auto, air, conventional rail, and high-speed rail). Several nesting structures were tested for the main mode choice models and the final nesting structure chosen is shown in Figure 4, with all the non-auto modes in a single nest. This structure provided the most logical and statistically sound nesting structure for the mode choice models. 73 Figure 4. Main Mode Choice Model Structure Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 74 The main mode choice models were based on stated preference (SP) survey data. The overall choice shares in the SP data were around 50% for high speed rail, with most of the other choices for the respondents’ actual chosen modes. The HSR choice share was highest for business trips and long trips, giving a first indication that HSR substitutes more closely with air than with car. To prepare the data for estimation, the access and egress mode choice models were first applied to calculate access and egress mode logsums for each alternative. Then, a nested logit model was estimated across the four main modes for each of the segments (only three alternatives for the Short segments, as air was not available for those segments). The estimation results for the Long Segments are shown in Table 5. Some results of note include the following: • The cost and in-vehicle time parameters were not constrained during model estimation and produce reasonable values of time. In general, the value of time for the longer, more expensive trips is higher than for the shorter, more frequent trips. This is a typical result. • The value of frequency (headway) is significant for all segments, but was only about 20 percent as large as the in-vehicle time coefficient. If wait time were half the headway and valued twice as highly as in-vehicle time, then we would expect the same coefficient on headway and in-vehicle time. For these modes, and particularly air, headway is less related to wait time than it is to scheduling convenience. Because none of the levels used in the SP had headways higher than a few hours, the implications for scheduling may not have been large enough to greatly influence mode choice. This coefficient was constrained to match in-vehicle time based on comments from the peer review panel. • The value of reliability is fairly low for all segments, although with the correct sign. It is very difficult to measure the effect of reliability in a large-scale mailout SP survey, so we decided to use a somewhat higher effect of reliability in application, based on evidence from other models that this was reasonable. • Those traveling with others are more likely to use car and less likely to use air. This effect was also tested on the cost coefficients and not found to be significant, so this relative mode preference appears to be related to more than just cost – such as the fact that people can share driving for long trips. Party size models were estimated to generate these data, but are not included here for brevity. • People in larger households are more likely to use car. Even though we already have the group/alone segmentation, people in larger households are likely to be in larger groups. • Higher income generally favors air and high-speed rail versus auto. • Low auto availability within the household is related to a lower chance of choosing the auto. • A nest with air, rail, and HSR, (with car in its own “nest”) produced a logsum coefficient below 1.0 for all segments, indicating that this was a reasonable nesting structure for interregional trips. • The access mode choice logsums were estimated with positive coefficients in the range of 0.14 to 0.46 for all segments. • For the long trips, the egress mode accessibility seems to have somewhat more influence on mode choice than does the access mode. Travelers may be less Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 75 constrained at the home end, where they know the options and can use their own auto, than they are at the destination end. Table 5. Main Mode Choice Models1 1 Does not include alternative specific constants. 2 These were later constrained during model calibration to match in-vehicle time based on comments from the peer review and the modeling team. 3After the headway coefficient was constrained, this ratio becomes 1. Source: The model was re-estimated when the headway measures were constrained and does not match previously published versions of this model - Bay Area/California High Speed Rail Ridership and Revenue Forecasting Study Interregional Model Systems Development Report, Table 3.15, August 2006. Long Trip Business/Commute Recreation/Other Observations 2,918 5,075 Final log-likelihood -1,998 -3,936 ρ2(0) 0.380 0.309 ρ2(cons) 0.151 0.154 Coeff. (t-stat) Coeff. (t-stat) Main Mode Characteristics Level of Service Cost ($) -0.017 (-12.8) -0.035 (-18.5) In-vehicle time (min) -0.018 (-13.4) -0.011 (-14.2) Service headway (min) -0.0042 (-3.9) -0.0032 (-3.6) Reliability (% on time) 0.023 constrained 0.005 (1.9) Access Mode Choice Logsum 0.136 (3.4) 0.204 (3.7) Egress Mode Choice Logsum 0.171 (3.9) 0.399 (7.1) Implied Value of Time IVT ($/hour) $63.64 $18.45 Ratio Headway/IVT3 0.21 0.24 Trip Characteristics Car – Travel in a Group 2+ 1.086 (4.6) 1.43 (9.1) Air– Travel in a Group 2+ -0.356 (-2.8) -0.505 (-3.7) Household Characteristics Car – Household Size 0.182 (1.2) 0.296 (4.4) Air – High Income 1.18 (4.6) Conventional rail – High Income 0.613 (1.4) High-speed rail – High Income 1.147 (4.8) Car – Less than 2 Cars per 2+ Household -0.308 (-2.3) Nesting Nest – air, rail, high-speed rail 0.692 (10.4) 0.738 (13.0) Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 76 4 Intraregional Models Intraregional models will be used to forecast high speed rail trips with both ends within an urban area that has more than one proposed high-speed rail (HSR) station. These areas are the San Francisco Bay Area, Greater Los Angeles, and San Diego regions. Regional travel forecasting models in these areas will be modified to forecast urban high-speed rail trips for the San Francisco and Los Angeles areas. The market segments for intraregional travel include typical trip purposes such as home-based work, school, university, shopping, social-recreational, and other trips as well as workand non-work-related non-home-based trips. San Diego is the only other region that contains the possibility of intraregional high-speed rail trips, but the estimate of these riders is very low relative to the other regions; and the level of effort to develop, calibrate, and apply the regional mode choice model is very high, so we decided to develop intraregional ridership for San Diego using a population-based estimate rather than a traditional mode choice model. To model intraregional trips, we relied on the trip generation and distribution models in each of the urban areas and modified existing mode choice models. The urban mode choice models include a variety of transit modes, but not specifically a high-speed rail mode. The MTC urban mode choice models were modified to insert a high-speed rail mode based on coefficients and constants from the commuter rail mode, as a conservative estimate. The SCAG urban mode choice model was built from the MTC framework. Following is a brief description of the model implementation for each of the urban areas: • San Francisco Bay Area - The San Francisco regional model was enhanced to include transit submodes (BART, commuter rail, light rail, ferry, local bus, and express bus) in the mode choice model. This allowed for easier inclusion of the high-speed rail mode in the model. The new mode choice model was validated at the regional level to match observed ridership numbers by operator. • Southern California Association of Governments Region - The Southern California Association of Governments (SCAG) mode choice models were developed using the parameters and structure of the MTC model in combination with the SCAG networks and trip tables. This model was validated at the regional level to match observed ridership numbers by operator. Urban trip tables from the MTC and SCAG metropolitan areas were added to the interregional trips for the assignment. 5 Model Application 5.1 Model Validation The validation of the combined interregional and intraregional (urban) models was completed for the year 2000, because the available observed data for 2000 was more robust than for any other year. This statewide model was estimated from a combination of existing and new household and intercept traveler surveys collected in California and combined with intraregional trips generated from regional and statewide sources. Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 77 The validation work included the calibration process, development of data used for observed travel behavior, and documentation of the resulting calibration parameters for the interregional trips. In addition, this work included summaries and reasonableness checks on the intraregional trips derived from the MPO trip tables. These were not separately validated or calibrated, because each MPO has provided assurances that these trip tables were validated. Trips by mode from the interregional models were combined with intraregional trips by mode to assign to the highway, air, and rail networks. Table 6 presents a summary of the 2000 interregional trips by mode and market. Highway trips were converted from person trips to vehicle trips using vehicle occupancy factors derived from the Caltrans Statewide Travel Survey. In addition, highway trips were separated into peak and off-peak time periods so that peak and offpeak trip tables could be assigned separately to the highway network. This ensures that peak-period travel times would more accurately reflect congestion that occurs in the peak period. Following the development of peak and off-peak auto vehicle interregional trips, these were combined with the auto vehicle intraregional trips. These intraregional trips come from four sources: MTC, SANDAG, SCAG, and Caltrans. The Caltrans Statewide Model is used to estimate intraregional trips for all the other regions (except MTC, SANDAG, and SCAG) so that the auto trip table will be representing all statewide travel. This ensures that congestion within each smaller urban area is adequately represented. Validation of the base year assignments by mode involved detailed review of observed and modeled volumes. For air, these reviews focused on assignments for the major markets. For rail, these reviews focused on assignments by operator. For highway, these reviews focused on assignments by gateway and by region. A summary of the assignments by mode is provided in Table 7. Table 6. 2000 Daily Interregional Trips by Mode Market Auto Air Rail Total Percent of Total LA to Sacramento 7,479 4,935 - 12,414 1% LA to San Diego 257,441 100 5,395 262,936 17% LA to SF 28,031 26,867 - 54,898 4% Sacramento to SF 137,739 25 1,816 139,580 9% Sacramento to San Diego 175 2,858 - 3,033 0% San Diego to SF 4,630 10,309 - 14,939 1% LA/SF to SJV 205,205 3,393 926 209,524 14% Other to SJV 281,750 243 344 282,337 19% To/From Monterey/ Central Coast 275,794 3,532 1,105 280,431 19% To/From Far North 184,506 3,005 16 187,527 12% To/From W. Sierra Nevada 59,192 668 11 59,871 4% Total 1,441,942 55,935 9,613 1,507,490 100% Percent of Total 95.7% 3.7% 0.6% 100% Source: Bay Area/California High Speed Rail Ridership and Revenue Forecasting Study Final Report, Table 5.1, July 2007. Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 78 Table 7. 2000 Daily Assignments by Mode Mode Units Observed Model4 Difference Percent Difference Air Boardings 54,2711 54,876 605 1% Rail Boardings 16,7102 17,743 1,033 6% Auto Vehicle Counts 27,145,3003 25,206,373 (1,938,927) -7% 1Source: U.S. Department of Transportation FAA O&D ten-percent sample database 2Source: Interregional rail operators and MTC 3Source: Caltrans, MTC and SCAG traffic count databases Source: Bay Area/California High Speed Rail Ridership and Revenue Forecasting Study Final Report, Table 5.2, July 2007. Even though the air and rail assignments were very small compared to auto, these were critical to the evaluation of high-speed rail, so a great attention to the validation of these modes was important. For the major markets and operators, these compared very well with observed numbers. Auto assignments were primarily validated based on gateways along the high-speed rail corridors. These compared very well to observed traffic counts. Additional validation effort to refine and improve the highway assignments is recommended if this model were to be used for highway planning purposes. Comparison of the 2030 forecast to a No-Build scenario was completed for validation to ensure that the 2030 forecasts are reasonable for each model component. Overall, there is a 42 percent increase in households and a 51 percent increase in employment, and there is a 62 percent increase in interregional trips. The 2030 interregional trip table is presented in Table 8. The higher percent of interregional trips compared to statewide household and employment growth is a reflection of the expansion of the regions beyond their regional borders, causing more travelers to make interregional travel instead of intraregional travel. The auto assignments (represented by total vehicle miles traveled) increase by 73 percent from 2000 to 2030, which is also caused by travelers having to go further to reach their destinations. These are presented in Table 9. Table 8. 2030 Daily Interregional Trips by Mode Market Auto Air Rail Total LA to Sacramento 12,636 8,105 – 20,741 LA to San Diego 340,862 96 25,898 366,856 LA to SF 30,253 25,351 – 55,604 Sacramento to SF 174,844 26 11,798 186,668 Sacramento to San Diego 164 5,258 – 5,422 San Diego to SF 5,038 18,259 – 23,297 LA/SF to SJV 360,177 9,609 6,237 376,023 Other to SJV 553,466 1,944 4,792 560,202 To/From Monterey/Central Coast 426,056 5,886 2,077 434,019 To/From Far North 320,667 5,957 962 327,586 To/From W. Sierra Nevada 96,404 1,177 335 97,916 Total 2,320,567 81,668 52,099 2,454,334 Source: Bay Area/California High Speed Rail Ridership and Revenue Forecasting Study Final Report, Table 5.3 July 2007. Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 79 Table 9. 2000 and 2030 Assignments by Mode Mode Units 2000 Model 2030 Model Difference Percent Difference Air Boardings 54,876 80,643 25,767 47% Rail Boardings 16,430 30,653 14,222 87% Auto Vehicle Miles Traveled 748,606,510 1,297,116,168 548,509,657 73% Source: Bay Area/California High Speed Rail Ridership and Revenue Forecasting Study Final Report, Table 5.4, July 2007. Rail boardings increase at a higher rate than auto, indicating that as congestion increases; more travelers are taking rail, as expected. Air boardings do not increase as fast as rail or auto because the air fares increased and frequencies decreased between 2000 and 2005, making air a less attractive option. The 2005 observed air level of service was kept constant through 2030. The primary reason for significant changes in air service from 2000 to 2005 was the September 11 terrorist attacks in 2001, which affected air travel more than other modes. 5.2 Forecast Results Table 10 presents a summary of the trips by mode and mode shares for the base year (2000) and the future year (2030) with and without the high-speed rail project. Highspeed rail captures over 7 percent of the trips and draws from all other modes. 5.3 Sensitivity Tests A series of sensitivity tests were conducted to test the impacts of changes in level of service on high-speed rail ridership and revenue. These tests were designed to assist in developing an improved operating plan, optimum fares, and to understand the impacts of potential changes in assumptions to the air and auto modes. The results of the sensitivity tests are provided in Table 11. Table 10. Summary of Trips and Mode Shares for Base and Future Conditions 2000 Base Year 2030 without HSR 2030 with HSR 2030 Difference Trips Mode Share Trips Mode Share Trips Mode Share Trips Pct of Total Auto 1,441,942 95.7% 2,320,567 94.5% 2,193,248 89.2% -127,319 -71% Air 55,935 3.7% 81,668 3.3% 53,823 2.2% -27,845 -16% Rail 9,613 0.6% 52,099 2.1% 31,790 1.3% -20,309 -11% HSR 179,482 179,482 100% Total 1,507,490 100.0% 2,454,334 100.0% 2,458,343 100.0% 4,009 Note: The 4,009 difference in 2030 trips with and without HSR demonstrates how much induced travel is a result of HSR. Source: Bay Area/California High Speed Rail Ridership and Revenue Forecasting Study Model Validation Report, Table 7.5, July 2007 and Bay Area/California High Speed Rail Ridership and Revenue Forecasting Study Ridership and Revenue Forecasts, Table 2.2, August 2007. Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 80 Table 11. Sensitivity Tests for High-Speed Rail Sensitivity Test Change in Level of Service Percent Change from Base Boardings Revenues High-Speed Rail Level of Service Tests Higher HSR Fares 25% increase -13% 2% Average Daily Headways HSR headways1 -15% -14% Higher HSR Freq 100% increase 15% 16% Express Service SF/LA Double Freq SF/LA to SJV, SD/SF to SAC 22% 24% Air and Auto Level of Service Tests Higher Air/Auto Times 6% increase2 6% 6% Higher Air/Auto Costs 50% increase 46% 53% Combined Level of Service Tests Higher HSR Fares and Higher Air/Auto Costs 25% increase in fares, 50% increase in costs 13% 19% Higher HSR Fares and Higher Air/Auto Costs 50% increase in both 31% 40% Higher HSR Fares and Higher Air/Auto Costs 100% increase in fares, 50% increase in costs -6% 1% 1 Average daily headways assume that the headway in the peak and off-peak periods are equal. This effectively increases peak headways and decreases off-peak headways. 2The 6 percent increase in travel time was based on a 30-minute increase in travel time from San Francisco to Los Angeles by car. Source: Bay Area/California High Speed Rail Ridership and Revenue Forecasting Study Final Report, Table 7.1, July 2007. The results show that improvements in high-speed rail frequencies can support much higher high-speed rail ridership; increased high-speed rail frequencies in the major corridors (San Francisco to Los Angeles, Los Angeles to San Joaquin Valley, San Diego to Sacramento, and San Francisco to Sacramento) were then retained for the alternatives analysis. These results also show that raising high-speed rail fares will not significantly increase revenues, unless this is combined with different assumptions of air and auto costs. Assumptions regarding air and auto cost increases remain a difficult issue, given the volatility in these costs in the past 5 years alone. The sensitivity tests do show that high-speed rail ridership is highly sensitive to the assumptions of air and auto costs and can increase as much as 46 percent with a 50 percent increase in air and auto costs, which seems quite reasonable compared to current trends in these costs. 6 Summary The travel forecasting models developed for predicting high-speed rail alternatives for the state of California have several immediate benefits over previous ridership forecasting methods used in the state: they are network-based and provide more accurate assessments of time and cost tradeoffs with other modes, modal choices are sensitive to reliability, party size, and detailed access and egress options, induced travel is assessed based on changes in level of service for all modes, and intraregional travel is estimated based on detailed urban area models where interregional travel is estimated based on statewide models estimated from observed travel behavior. The intraregional and interregional models are integrated to assess impacts of congestion on other modes and to reflect differences in peak and off-peak conditions. Outwater, Tierney, Bradley, Sall, Kuppam, Modugula, Journal of Choice Modelling, 3(1), pp. 58-83 81 The primary advancement in this model is the additional level of detail (4,600 zones used for all modeling components without sampling), the inclusion of peak and off-peak assignments, and the consistent use of logsum accessibility measures at all levels of the models (from access and egress models up to trip frequency models). These models were estimated using revealed preference data and by combining multiple survey datasets, a more robust estimation dataset was possible. There are some areas where these models may be improved for other statewide and regional planning activities. The trip frequency models could benefit from additional data on weekly or monthly long distance travel, because a one-day snapshot does not provide as strong a basis for travel decisions as longer-term data would provide. The destination choice models could also be improved by including data on special generators, such as Disneyland. Lastly, the mode choice models could benefit from a tour-based methodology, recognizing that decisions on mode are affected by both the outbound and return portions of the trip. In these cases, the models could benefit from additional data and resources that were beyond the original scope of the project. These integrated statewide models offer a comprehensive tool to forecast long and short distance travel in California. The separation of travel into market segments based on distance (short and long), purpose (business, commute, recreation and other) and travel markets (interand intraregional) provide a robust and accurate assessment of multimodal travel at the statewide level. 7 References Algers, S., 1993, Integrated Structure of Long-Distance Travel Behaviour Models in Sweden, Transportation Research Record 1413, Transportation Research Board, 141-149. Ben-Akiva, M. and Morikawa, T., 1990, Estimation of Switching Models from Revealed Preferences and Stated Intentions, Transportation Research Part A, 24(6), 485-495. Bhat, C.R., 1995, A Heteroscedastic Extreme Value Model of Intercity Travel Mode Choice, Transportation Research Part B, 29(6), 471-483. Bhat, C.R., 1998, Accomodating Variations in Responsiveness to Level-of-Service Measures in Travel Choice Modeling’, Transportation Research Part A, 32(7), 49-57. Bhat, C.R., 1997, Covariance Heterogeneity in Nested Logit Models: Econometric Structure and Application to Intercity Travel, Transportation Research Part B, 31(1), 11-21. Booz-Allen & Hamilton, 1989, Demand ‘Model Estimation: Final Report’, prepared for AMTRAK. Bradley, M.A. and Daly, A.J., 1997, Estimation of logit choice models using mixed stated preference and Revealed Preference Information’, in Stopher, P.R. and Lee-Gosselin, M. (Eds.) Understanding Travel Behaviour in an Era of Change, Oxford: Pergamon, 209-232. Brand, D., Parody, T.E., Hsu, P.S. and Tierney, K., 1992, Forecasting High-Speed Rail Ridership’, Transportation Research Record 1342, 12-18. Cambridge Systematics with Mark Bradley Research and Consulting, 2006, Bay Area/California High-Speed Rail Ridership and Revenue Forecasting Study: Interregional Model System Development Report’, prepared for the Metropolitan Transportation Commission and the California High-Speed Rail Authority.