Development of New Toll Mode Choice Modeling ...

Dehghani, Adler, Doherty & Fox

1

TRB PAPER # 03-3710 DEVELOPMENT OF A NEW TOLL MODE CHOICE MODELING SYSTEM FOR FLORIDS’S TURNPIKE ENTERPRISE BY Youssef Dehghani, Ph.D., P.E. Parsons Brinckerhoff Quade & Douglas, Inc. 999 Third Avenue, Suite 2200 Seattle, WA 98104 Tel: (206) 382-5251 Fax: (206) 382-5222 E-mail: [email protected] Thomas Adler, Ph.D. Resource Systems Group 331 Olcott Drive White River Junction, VT 05001 Tel: (802) 295-4999 Fax: (802) 295-1006 Email: [email protected] Michael W. Doherty URS Corporation 3676 Hartsfield Road Tallahassee, Florida 32303 Tel: (850) 574-3197 (Ext. 366) Fax: (850) 205-3229 Email: [email protected] And Randy Fox, AICP Florida Turnpike Enterprise Milepost 263, Florida’s Turnpike Bldg 5315, Turkey Lake Service Plaza Ocoee, FL 34761 Tel: (407) 532-3999 Fax: (407 822-6612 E-mail: [email protected]

November 13, 2002

Paper for Presentation at the Annual Meeting of the Transportation Research Board in January 2003

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.


2

ABSTRACT The Florida Department of Transportation (FDOT) employs a statewide system of travel forecasting software known as the Florida Standard Urban Transportation Modeling Structure or FSUTMS. First conceived and implemented over 20 years ago, the FSUTMS has been subject to numerous updates and improvements designed to improve its effectiveness as the main transportation-planning tool for all urban and regional study areas within the State of Florida. This paper describes recent toll mode choice model development activities undertaken by the FDOT Turnpike Enterprise. Within the FSUTMS, toll modeling originated by establishing specific toll amounts for appropriate network links together with a coefficient of toll to convert tolls to travel time impedance. That simple beginning has transitioned to the current method that provides for multi-channel queue simulations of toll plaza operation as part of the traffic assignment process. In order to address contemporary toll study issues, however, tollmodeling innovations were desired that addressed trip makers toll route decisions as a mode choice step sensitive to changes in service levels by time of day (such as change in toll and congestion levels during peak and off-peak periods), trip purpose, and socio-economic attributes of trip makers (such as income). Innovations developed for Florida’s Turnpike began with the data collection efforts and toll model development activities performed for the Central Florida (Orlando) Region. Similar efforts are underway for Florida’s Southeastern (Miami/Ft. Lauderdale) Region. The Orlando Region Toll Mode Choice Model, which is going through a final validation phase, includes a statistically estimated nested mode-choice modeling system that has a discrete choice for toll travel. The models have been developed for a combination of four time periods and four trip purposes, including visitor trips. Other key features incorporated into this next generation of Florida’s Turnpike modeling system include: (a) a pre-mode choice time-of–day process; (b) a generalized costs assignment procedure that uses both travel time and costs by time of day (rather than travel time alone) in the highway equilibrium assignment process; (c) production of zone-to-zone travel time and costs consistent with travel paths, based on using generalized costs; and (d) a feed-back loop process that employs an iterative successive averaging procedure for estimating travel times. Initial application of the new toll models is to test alternatives and develop traffic and revenue estimates for variable pricing projects in the State. Toll modeling sensitivity to time of day attributes, trip purpose and income is expected to yield benefits of increased accuracy and reliability of toll project forecasts.




1.0

3

OVERVIEW

The Florida Department of Transportation (FDOT) Turnpike Enterprise has relied on “best practice” toll-forecasting procedures for its periodic update of traffic and revenue forecasting analysis. This analysis is required for the existing toll facilities and for the planning, design, and the economic feasibility assessment of proposed new toll facilities. In keeping with this tradition, a panel of travel demand modeling experts was formed in 1998 to advise the Florida’s Turnpike Enterprise on the short and long-term “best practice” improvements that could be developed within three to ten years to enhance the existing toll modeling capabilities. The cornerstone of the expert panel’s recommendations was the development of a multi-modal modeling system that has a discrete choice component for toll travel, based on using surveys and other observed data unique to each metropolitan area. As a result of this recommendation and because of the presence of an extensive network of toll facilities in the Orlando area, it was logical to first launch such an undertaking in this region. Pertinent data were collected in 2000 to provide the information required to support development of a toll mode choice modeling system. A similar effort is underway for the Miami/Ft. Lauderdale region. Key features in the toll mode choice modeling system are: •

A post-trip distribution time-of-day modeling capability to reflect variation in travel times during four time periods (i.e., AM peak, mid-day, PM peak, and night).

•

A multi-modal modeling system encompassing 16 nested mode-choice models that are statistically estimated, using survey data and reflecting four time periods and four trip purposes (i.e., home-based work, home-based non-work, non-home-based, and tourists/visitors). These models include specific decision-tree hierarchies for transit submodes (i.e., primary transit versus walk or auto access) and different auto occupancy classes (i.e., SOV versus HOV2 and HOV3+) for toll and toll-free choices.

•

A generalized assignment procedure that uses both travel time and costs by time of day (rather than travel time alone) in the highway equilibrium assignment process.

•

A feed-back-loop process that uses a successive method of averaging highway travel times. This iterative process involves updating highway travel times via toll and free roads, and feeding them back into the mode choice modeling process.

This paper is organized in six sections. The overview section is followed by brief descriptions of the surveys and other data collection activities undertaken. The toll mode choice model estimation analysis results are discussed in Section 3. Time-of-day highway assignments using generalized costs, and feedback procedures are described in Section 4. The model implementation process is discussed in Section 5. Finally, some summary findings and additional refinement areas are outlined in Section 6. 2.0

DATA COLLECTION

An extensive effort was undertaken to collect the key data required to support the toll modeling improvement program. In addition to securing readily available data from various public agencies (such as traffic counts, surveys data, land use and network data), speed measurements, traffic counts and vehicle occupancy data were collected. The data collected were for the following time periods: • • •

AM peak period: 6:00 to 9:00 AM Off-peak (mid-day) period: 9:00 to 4:00 PM PM peak period: 4:00 to 7:00 PM

To estimate toll mode choice models, a number of surveys were conducted and analyzed in spring and summer of 2000 (Resource Systems Group, 2002), including: • • •

Focus Group Survey Stated Preference Survey Origin-Destination Survey

Brief descriptions of these surveys follow.




4

Focus Group Survey A series of ten focus groups was conducted in late May to early June 2000, to determine how Orlando-area travelers who use one of the most congested corridors in the region (the I-4 corridor) feel about transportation improvement alternatives and current travel conditions in the corridor. The focus groups were intended both to support planning studies related to development of express lane concepts for the I-4 and to identify key issues to be addressed in modeling their current and future behavior. Each group consisted of approximately ten participants, led by a moderator. Discussion was based on the following questions: • • • •

How do Orlando-area I-4 travelers perceive current I-4 travel conditions? What do they perceive are their current travel alternatives? What do they know about current I-4 improvement plans or proposals? How would they change their behavior in response to changes in the I-4 corridor (including express lanes, toll pricing and transit improvements)?

The groups included randomly identified regular peak-period commuters, other residents who use I-4, and Orlando-area visitors. Results from this survey provided useful insights on the travel behavior of these groups that were used in designing the stated preference survey questionnaires. For example, participants indicated that they actively manage the time of day of their travel to avoid peak periods wherever possible; when they do travel in the peak period, they build in substantial time buffers to allow for delays caused by frequent I-4 incidents. They also use toll routes that involve substantially longer travel distances to avoid potential I-4 delays. These current behavioral adaptations were reflected in the design of the stated preference surveys. Stated Preference Survey The Stated Preference Survey provided quantitative information on the trade-offs that travelers make among travel time, cost, and other trip characteristics when they choose their mode, route, and time of travel. These data were used to statistically estimate the coefficients used to compute the share of travelers who choose a particular travel alternative, given the characteristics of all of the available alternatives. The alternatives includes auto drive-alone (DA), auto shared-ride, and transit modes. For the auto modes, the choice between tolled and non-tolled routes and between different travel time periods (i.e., AM peak, mid-day, PM peak, and night) is represented in the model. The Stated Preference Survey was administered over a ten-day period from Monday, June 12th through Wednesday, June 21st, 2000. During this period, a total of 1044 respondents consisting of residents and visitors in the Orlando area completed the survey. During the field intercept, the Stated Preference Survey was administered on laptop computers using the IVIS (Interactive Video Interview Stations) survey technique developed by Resource Systems Group. Each day, two interview stations were set up with five to six laptops, a staff of four to five attendants, and a field manager. Survey sites were chosen to provide exposure to Orlando residents and visitors and to members of a range of socioeconomic groups. The sites included shopping malls, theme parks, motor vehicle licensing offices, and offices of major local employers. During the survey administration period, several area businesses/organizations were directly recruited to a website to complete the Stated Preference Survey. In addition, 520 respondents who completed the origindestination survey online were invited to participate in the Stated Preference Survey.



Origin-Destination Survey The Origin-Destination Survey provides details about travel times, purposes, vehicle occupancies, routes, and trip start/end locations for trips within the I-4 corridor and for regional trips. These data provide trip purpose by time-ofday factors and give a profile of the locations of internal trips (for use in the trip distribution). They were used, along with the current travel information from the stated preference survey as “revealed preference” data for estimating route, timeof-day, and choice of mode. Two versions of the paper-based Origin-Destination Survey form were developed. The first version collected information on trips that used (or could have used) I-4 or a Central Florida toll highway. This version was sent to three quarters of the survey recipients. The second version asked about all trips in the Central Florida region, and was sent to the remaining quarter of the households on the mailing list. Both questionnaire versions were printed on a full color trifold card that included a letter and instructions. Respondents could complete the paper-based survey form and mail it back postage-paid, or log on to the Internet to answer the questions online using a unique password.




5

Mailing of the Origin-Destination Survey forms to 100,000 households in the Central Florida region began during the last week of July 2000, and delivery was staggered over a period of about three weeks. A second copy of the survey form with a modified cover panel was mailed to the full mailing list beginning the third week of August. Again, delivery of the forms was staggered over a period of several weeks. The second mailing served as both a reminder and a replacement copy of the original questionnaire form. 3.0

TOLL MODE CHOICE ESTIMATION ANALYSIS RESULTS

This section describes the toll mode choice model estimation analyses that were performed. The model estimation used the Origin-Destination and Stated Preference surveys that were conducted as part of the study and a 1996 transit (Lynx) on-board survey that provided revealed preference data for transit passengers. Travel time and cost data from the Metroplan Orlando’s regional four-step travel model were used to provide information about chosen and unchosen alternatives for the revealed preference modeling. The basic nested logit modeling approach focused on capturing travel behavior for choice of mode, route and timeof-day: • • •

Mode: auto drive-alone, auto with two occupants, auto with three or more occupants, bus, and rail Route: travel via a toll road and travel via a non-toll (free) road Time-of-day: choice between desired time of travel and time-shifted trip.

The statistical analysis of estimating logit choice models was performed using commercially available ALOGIT software. The statistical estimation was conducted in several stages, beginning with simple specifications and successively testing a wide range of specifications, segmentation schemes and model structures. The initial modeling divided all travel time and cost variables into alternative-specific effects (assuming for example that transit travel time might be considered by travelers to be more or less onerous than auto travel times). In addition, all of the available demographic variables were included as alternative-specific variables in the models. These analyses were used to determine which mode-specific effects should be considered in the model specifications. The second estimation stage explored different segmentation schemes. Initially, two different segmentation approaches were evaluated: time-of-day and trip purpose (each having four segments): • •

Time-of-day segments: AM Peak, PM Peak, midday, and night. Trip purpose segments: Home-based work (HBW), home-based non-work (HBNW), non-home-based (NHB), and visitors.

The third estimation stage included tests of a wide range of nesting structures. The nesting structures were initially tested with the time-of-day segmentation and later tested with alternative segmentations. The fourth estimation stage incorporated revealed preference data into the estimation process. These data are useful for testing the match between the stated preference responses and actual observed behavior. In the final segmentation, combinations of time-of-day and trip purpose were evaluated. Model Estimation Analysis Results Statistical Chi-square tests were constructed to determine the most appropriate model segmentation scheme (see Ben-Akiva and Lerman, 1985 for a description of this approach for evaluating model segmentation). Initial analysis showed the time-of-day groupings to be slightly better statistically, though both segmentations were statistically significantly (at ∝=0.05) better than a simple pooled model. With a more refined specification, however, the trippurpose segmentation was found to be slightly better than the time-of-day models. In addition, a combination of trip-purpose and time-of-day segmentations provides more explanatory power than either segmentation by itself. For this reason, segmented models by both trip purpose and time of day were developed. With this segmentation approach, the models for night and midday were not statistically different, so those time periods were combined into a single off-peak period. The model segments are:



Dehghani, Adler, Doherty & Fox • • • • • •

6

Home-based work (HBW) peak trips HBW off-peak trips Home-based non-work (HBNW) peak trips HBNW off-peak trips Non-home-based (NHB) trips Visitor trips

The initial model specifications for each of the six new segments were modified as necessary, to exclude variables whose coefficients were statistically insignificant and to re-specify effects so that the models corresponded with behavioral expectations. Several tests were performed to determine the most appropriate structure and specification. The initial tests of nest structure were conducted using models segmented by times of day only. As different model specifications and segmentations were tested, the nesting structure evolved to the final adopted form. Table 1 shows estimated coefficients for mode-choice models, for four trip purposes and four time periods. All of these coefficients have the expected sign and were found to be statistically significant and within an acceptable confidence level using t-statistic test (at ∝=0.05). The mode coefficients for AM and PM peak periods are similar. Figure 1 illustrates a common structure estimated for these models. The nesting coefficients for the adopted mode choice models are shown in Table 2. Some key findings from the model estimation analysis are: •

Trip length was found to be a statistically significant factor in the choice between toll routes and non-tolled routes. This reflects the propensity of travelers to use toll roads for long-haul trips, which is in part due to the longer spacing between toll road interchanges. This variable is alternative-specific to the toll drive alone mode only.

•

Household income was a significant variable in explaining sensitivity to toll prices. Travel costs divided by the natural logarithm of average household income for HBW, HBNW, and NHB are used as generic variables to all modes. In addition, average household income is used as an alternative-specific variable to all transit submodes.

•

Extensive testing was conducted to determine how travel costs affect utility by vehicle occupancy. It was found that vehicle occupancy does not have any significant effect on the disutility of travel costs, except for home-based work trips, where the relationship was best represented as costs divided by ln(vehicle occupancy+1). This has an important implication for forecasting models in that dividing costs by vehicle occupancy for non-work trips would likely under represent the effects of cost changes on these trips.

•

The values of time implied by the model coefficients range from about $3.00/hour to $13.50/hr across the segments. For all except the visitor segment, the values of time vary by household income; higher values were estimated for higher income households. The values of time by segment are: for home-based peak work trips – $4.50 to $9.50/hour; for home-based off-peak work trips – $4.00 to $13.50/hour; for homebased peak nonwork trips – $4.00 to $7.50/hour; for home-based off-peak nonwork trips – $3.00 to $8.00/hour and for visitor trips – $5.00/hour.

•

The time-of-day choice was modeled using a variable representing the amount of time shift away from the time that the travelers said that they would most like to make the trip. The marginal utility of this “shift time” was estimated to be about 40% of the marginal utility of travel time for both work and non-work trips. This factor can be used in an incremental modeling approach to shift trips among time periods based on relative travel times, time shift “distances” and costs between adjacent time periods.

•

The model estimation work used both stated preference and revealed preference data. Since the stated preference experiments did not include transit submode or out-of-vehicle travel time trade-offs, the revealed preference data were used exclusively to estimate those effects. And, since the revealed preference data did not include time shifting effects, the stated preference data were used for those variables. Where both stated and revealed preference data were available to measure the same variables, it was found that those measurements were reasonably consistent; within 20% to 30% of each other. The scaling factor for stated preference data in the combined model is 0.86, indicating that the data have comparable scales. The




7

scaling factor is theoretically unbounded in range and a value of 1.0 means that the two data sources have exactly the same scale.

4.0

TIME-OF-DAY, GENERALIZED COSTS, AND FEEDBACK PROCEDURES

This section provides brief descriptions of the time-of-day, generalized costs, and feedback procedures that were developed and incorporated into the toll mode choice modeling system. Time-of-Day Analysis It seems safe to assume that prevalent congestion on competing roadway facilities plays an important role in the amount of patronage experienced on a given toll facility within a corridor. In recognition of this proposition, a process was developed to allow better representation of congestion over various time periods. Traffic counts collected on major arterials and freeways were analyzed to determine the most appropriate time-of-day categorizations. This analysis resulted in using four time periods: • • • •

AM peak period, between 6:00 to 9:00 AM Mid-day period, between 9:00 AM to 4:00 PM PM peak period, between 4:00 PM to 7:00 PM Night period, between 7:00 PM to 6:00 AM.

The Origin-Destination Survey data collected in Summer 2000 were analyzed for the purpose of developing appropriate time-of-day factors. These factors (shown in Table 3) were derived by trip purpose, trip direction (i.e., Production-to-Attraction (P-A) and Attraction-to Production (A-P)), and the four time-of-day periods indicated above. The time-of-day factors were applied to person-trip tables that correspond to seven trip categories (shown in Table 3). The resulting person-trip tables for HBW, HBNW, and Visitor purposes classified by time-ofday were used in the toll mode choice modeling process discussed in the previous section. Supplementary survey data from the Metroplan Orlando were used to partition the other resulting trip tables by time of day (i.e., Internal/External (I/E), Truck/Taxi, and External/External (E/E) trips) into different auto vehicle occupancies for toll and non-toll trips. Highway Assignments Procedure Using Generalized Costs A generalized cost assignment procedure was developed for the purpose of achieving more realistic highway loading travel times. This process involved constructing two separate functions to reflect the generalized costs of traveling between two zones via a toll road (GCtp), and via a free road (GCfp) for time period (p). These two functions are mathematically expressed as follows: GCtp = ap * [Time via toll road (minutes) ] + bp * [Distance via toll road (miles) x ( Tolls (cents/mile) + Operating Costs (cents/mile))] GCfp = ap * [Time via free road (minutes) ] + bp * [Distance via free road (miles) x ( Operating Costs (cents/mile)] For the purpose of determining travel time and cost coefficients (represented, respectively, by a and b in the equations above for each time period, p), the time-period-specific route choice model were statistically estimated using the survey data. The resulting travel time and cost coefficients are somewhat different than those shown in Table 1. The travel time and cost coefficients resulted from the route choice modeling analysis are: • •

For travel time (a), -0.047 for peak and –0.06 for mid-day and night periods for travel time; and For travel cost (b), -0.006 for peak and mid-day periods and –0.003 for night period.

The TRANPLAN software program was modified to produce appropriate highway travel-time skim matrices. This permits the determination of congested travel-time matrices along the generalized cost paths for trip distribution and mode-choice applications. The initial results from using generalized costs in the highway equilibrium assignment process have shown improved loadings on toll facilities.




8

Feedback Procedure The toll mode choice model application involves updating highway link times via toll and toll-free roads, and feeding them back into the toll mode choice model through an iterative process. The purpose of this process is to produce more realistic modal travel demand estimation in each time period, by using a more accurate representation of service levels for each time period. Highway travel times are calculated through a Mean Successive Averages COMbinations (MSACOM) program to achieve convergence on congested link times via toll and free roads. This program was implemented in the toll mode choice model application process. The process involves using loaded link volumes from successive model iterations “n” and “n+1” to calculate link volumes for iteration n+1, based on the following relationship: vol(n+1) = (1.0 – 1.0/n) x vol(n-1) + (1.0/n) x vol(n) The resulting link volume, vol(n+1) , is used in the volume-delay equation to determine the link time for cycle n+1. A generalized cost highway assignment process is used to produce loaded link volumes at each iteration. Four cycles (iterations) are used in the toll mode choice model. Corresponding weights (1.0 – 1.0/n) and (1.0/n) used in the above formula for these four cycles are (respectively) 1/2 and 1/2; 2/3 and 1/3; 3/4 and 1/4; and 4/5 and 1/5. 5.0

TOLL MODE CHOICE MODEL IMPLEMENTATION AND VALIDATION ANALYSIS PROCESS

The toll mode choice modeling methods and procedures discussed in the previous sections have been implemented within the FSUTMS framework of the existing Metroplan Orlando Regional Model. The implementation of all toll mode choice models was achieved efficiently, using a general program called Generalized Nested Logit (GNL). This program, designed by Fennessy Associates, performed two key functions: • •

Mode choice model implementation – it was used to implement the utility equations describing the statistically estimated toll mode choice models as well as to specify the required input data files and parameters. Modal constants calibration – the process of modal constants calibration was automated and incorporated into the GNL program. The modal constants produced from the statistical estimation analysis reflect relative modal shares of the surveyed trip makers and not necessarily of the entire population. For this reason, the estimated modal constants usually need to be calibrated to reflect the aggregate total person-trip target values for each auto and transit submode, trip purpose, and time of day. The Origin-Destination and Transit On-Board surveys data were used to establish trip target values. The modal constants were considered calibrated when the difference between estimated and person-trip target values for each mode by trip type and time of day combination came within one percent. On the average, it required about 35 to 55 iterations to achieve the designated one percent convergence level.

The model validation analysis is an iterative, evolving process. Initially, the resulting modal constants from the calibration process are used to perform a model run through mode-choice and assignment-model components. Pertinent output such as average trip length, speed, and link volumes are compared against their observed counterpart values. This process highlights the apparent adjustments required, such as modifying target trip values (which necessitates recalibration of modal constants). This step also involves possible adjustments to other attributes that may need to be made. The iterative process is continued until the model overall performance is deemed acceptable in comparison to observed measures, including screenline volumes, patronage at key toll plazas and on/off ramps, speed, and average trip length. Preliminary sensitivity analysis results with respect to change in toll rate are graphically shown in Figures 2a and 2b, respectively, for each time period and trip purpose. As shown in Figure 2a, reduction in toll usage is most pronounced during the night period as a result of the increase in toll rate. Non-home-base (NHB) trips have shown lower sensitivity to toll increase. This is primarily due to having lower travel cost coefficient (or higher value of travel time) in the estimated mode choice models for NHB trips. The implied elasticity estimates from these results are:




o o o o o

9

33 percent for AM peak period 36 percent for midday period 30 percent for PM peak period 43 percent for night period 35 percent for all periods combined

These elasticities are reasonable and seem to indicate that peak toll users are relatively less sensitive to changes in toll rates than non-peak toll users. For example, the existing regional toll model estimates that a doubling of toll rates for AM peak toll users could result in a 33 percent reduction in toll patronage for that period. Upon completion of the ongoing model validation analysis additional results will be included in the next version of this paper. 6.0

SOME SUMMARY FINDINGS AND DIRECTION FOR FURTHER REFINEMENTS

Although the model has not yet been fully tested, a milestone has been achieved for successfully developing and implementing the “best practice” toll mode choice modeling procedures. Key features of the current toll modeling system are: •

The use of locally collected revealed and stated preference survey data, to allow statistical estimation analysis of mode choice models (including discrete choice for toll travel).

•

The availability of pertinent survey data to allow determination of target trip values.

•

The time-of-day modeling capability, reflecting variation in service levels by time of day. In addition, an incremental post modeling procedure will be implemented to allow time-shifting analysis of toll trips. Pertinent coefficients have been estimated using survey data – these indicate the trade-offs that travelers make among costs, travel times and shifts away from their “ideal” time-of-day for the trip.

•

The use of generalized costs in the highway assignment process. This has allowed a more realistic loading on roadway facilities.

•

The use of a Method of a mean Successive Averages, allowing more accurate representation of service levels in the mode choice modeling process;

•

The option of being able to use a feedback process between the mode-choice and assignmentmodeling components.

•

The capability for using an automated process to calibrate modal constants.

The current toll mode choice model could be further enhanced in a number of areas, including but not limited to: •

Extension of the time-of-day and feedback processes through trip distribution. Currently, the time-of-day analysis procedure is applied in the post-trip distribution-modeling step. This process could be altered so that the time-of-day analysis is performed prior to the trip-distribution modeling step. Traditionally, congested travel times are used to distribute work trips only. This may be a realistic assumption for AM peak period when work trips are dominate but not for PM peak period when non-work trips are dominant. In addition, the time-of-day extension into the trip distribution process would make it internally more consistent with the proposed toll mode choice modeling system.

•

Explore the possible merits of using composite impedance in the trip-distribution modeling step. This is theoretically more appealing and correct to use the overall mobility provided by all modes to distribute trips rather than highway travel time alone.




10

•

Refinements to trip-generation/distribution for modeling Disney-related trips. Primarily, the existing toll mode choice model can be further refined regarding trip distribution. This includes the distribution of resident trips and visitors that are Disney-related. The trip-generation part of the Disney-related trips could also be improved using recently collected survey data.

•

Develop procedures to allow modeling of toll users who pay electronically versus manually. Surveys conducted as part of other Turnpike Enterprise studies indicate that those who use electronic toll payment have shorter plaza delays and appear to have lower toll disutilities. Implementation of such a procedure – either internal to the existing toll mode choice modeling system or as a post modeling process – will provide the Florida’s Turnpike Enterprise with a capability to model more accurately toll patronage by payment method.

In summary, it should be concluded that the toll mode choice model innovations developed for the Florida’s Turnpike Enterprise have embodied the “best practice” toll mode choice modeling procedures that are supported by survey data and rigorous model estimation analysis. This is expected to yield benefits of increased accuracy and reliability of toll project forecasts.

ACKNOWLEDGMENTS The authors would like to thank the Peer Review Panel and Model Development Team members, including Jim Fennessy of Fennessy Associates. Thanks also to Edward “Terry” Denham, AICP, Florida Turnpike’s Director of Planning and Programming (Retired), Hugh Miller, Jr., Ph.D., P.E., and William Olsen, Ph.D., P.E., respectively, Turnpike’s Traffic Engineering Program Director and Travel Forecasting Program Manager for their encouragement and support of the toll model development effort. The opinions and views presented in this paper are solely the authors and do not necessarily represent the official policy of Florida’s Turnpike Enterprise and its staff.




11

REFERENCES 1. 2.

Resource Systems Group, Orlando Survey and Mode Choice Model Documentation, January, 2002. Ben-Akiva, M. E. and S. R. Lerman, Discrete Choice Analysis: Theory and Application to Travel Demand, MIT Press, 1985.




12

LIST OF TABLES AND FIGURES TABLE 1

Statistically Estimated Toll Mode Choice Model Coefficients by Trip Purpose and Time-of-Day

TABLE 2 Statistically Estimated Nested Coefficients for the Toll Mode Choice Models by Trip Purpose and Time-of-Day TABLE 3 Time of Day Factors by Four Time Periods

FIGURE 1 Toll Mode Choice Model Structure

FIGURES 2a & 2b Toll Mode Choice Model Sensitivity Analysis Results




13

Table 1 Statistically Estimated Toll Mode Choice Model Coefficients by Trip Purpose and Time-of-Day Mid-Day Period

AM or PM Peak Period Explanatory Variables' Definition Auto In-Vehicle Travel Time (Minutes) Auto Terminal Time (Minutes) Transit In-Vehicle Travel Time (Minutes) Walk Time to Transit (Minutes) First Transit Wait Time ( Less or Equal to 7 Minutes ) First Transit Wait Time (Greater than 7 Minutes) Transit Transfer Wait Time (Minutes) Travel Cost / Ln(Zonal Avg. Annual Income in 1000s$) Transit Fares (Cents)

HBW

HBNW

NHB

Visitors

HBW

HBNW

NHB

Visitors

HBW

HBNW

NHB

Visitors

-0.01500 -0.02400 -0.01500 -0.02400 -0.02400 -0.02400 -0.02400 -0.00711 -0.00711

-0.01200 -0.01920 -0.01200 -0.01920 -0.01920 -0.01920 -0.01920 -0.00698 -0.00698

-0.00400 -0.00640 -0.00400 -0.00640 -0.00640 -0.00640 -0.00640 -0.00091 -0.00091

-0.00700 -0.01120 -0.00700 -0.01120 -0.01120 -0.01120 -0.01120

-0.01000 -0.01600 -0.01000 -0.01600 -0.01600 -0.01600 -0.01600 -0.00482 -0.00482

-0.00500 -0.00800 -0.00500 -0.00800 -0.00800 -0.00800 -0.00800 -0.00310 -0.00310

-0.00400 -0.00640 -0.00400 -0.00640 -0.00640 -0.00640 -0.00640 -0.00091 -0.00091

-0.00700 -0.01120 -0.00700 -0.01120 -0.01120 -0.01120 -0.01120

-0.01000 -0.01600 -0.01000 -0.01600 -0.01600 -0.01600 -0.01600 -0.00482 -0.00482

-0.00500 -0.00800 -0.00500 -0.00800 -0.00800 -0.00800 -0.00800 -0.00310 -0.00310

-0.00400 -0.00640 -0.00400 -0.00640 -0.00640 -0.00640 -0.00640 -0.00091 -0.00091

-0.00700 -0.01120 -0.00700 -0.01120 -0.01120 -0.01120 -0.01120

0.00000 0.00100

-0.00600 0.00200

0.00000 0.00100

-0.00600 0.00200

Travel Cost (Cents) Zonal Avg. Annual Income in 1000s$ (Transit Modes Only) Total Trip Distance via Toll Road for Drive Alone (Miles)

Night Period

-0.00800 0.00700

-0.00108 -0.00108 0.00100

-0.00108 -0.00108 0.00100

NOTES:

- The mode choice estimation analysis resulted in relative ratio of 1.60 between out-of-vehicle travel time and in-vehicle travel time coefficients. This ratio is implied in travel time coefficient used for variables representing out-of-vehicle travel time components. - For HBW trips in each time period, travel costs for auto modes are divided by natural logarithm of occupancy level plus 1. For other trip purposes, travel costs are not divided by any occupancy level.



-0.00108 -0.00108 0.00100


14

Table 2 Statistically Estimated Nested Coefficients for the Toll Mode Choice Models by Trip Purpose and Time-of-Day AM or PM Peak Period

Mid-Day Period

Night Period

Nest Definition

HBW

HBNW

NHB

Visitors

HBW

HBNW

NHB

Visitors

HBW

HBNW

NHB

Visitors

Auto Nest Auto Toll Nest Auto Toll Nest - Shared Ride Auto Free Nest Auto Free Nest - Shared Ride Transit Nest Transit - Walk-Access Nest Transit - Drive-Access Nest

0.308 1.000 1.000 0.680 1.000 0.990 0.990 0.310

0.381 1.000 1.000 0.812 1.000 0.990 0.990 0.310

0.076 0.384 1.000 1.000 1.000 0.990 0.990 0.310

0.291 1.000 1.000 0.843 1.000 0.990 0.990 0.310

0.148 1.000 1.000 1.000 1.000 0.990 0.990 0.310

0.093 1.000 1.000 0.890 1.000 0.990 0.990 0.310

0.076 0.384 1.000 1.000 1.000 0.990 0.990 0.310

0.291 1.000 1.000 0.843 1.000 0.990 0.990 0.310

0.148 1.000 1.000 1.000 1.000 0.990 0.990 0.310

0.093 1.000 1.000 0.890 1.000 0.990 0.990 0.310

0.076 0.384 1.000 1.000 1.000 0.990 0.990 0.310

0.291 1.000 1.000 0.843 1.000 0.990 0.990 0.310




15

Table 3 Time of Day Factors by Four Time Periods AM Peak (6-9) Trip Purpose/Type Home-Based Work (HBW) Home-Based Non-Work (HBNW) Non-Home-Based (NHB) Visitors Internal External (I/E) Truck/Taxi External/External (E/E)

Mid-Day Period (9-4)

PM Peak (4-7)

Night Period (7-6)

All Periods

P-A

A-P

P-A

A-P

P-A

A-P

P-A

A-P

Combined

0.2556 0.0738 0.0875 0.1874 0.0930 0.0945 0.0930

0.0860 0.0524 0.0875 0.0047 0.0930 0.0945 0.0930

0.1242 0.2921 0.3010 0.1918 0.1984 0.1860 0.1981

0.0855 0.2631 0.3010 0.0524 0.1984 0.1860 0.1981

0.0453 0.1013 0.0730 0.0609 0.1100 0.1270 0.1105

0.2890 0.1286 0.0730 0.1823 0.1100 0.1270 0.1105

0.0748 0.0328 0.0385 0.0341 0.0986 0.0925 0.0984

0.0396 0.0559 0.0385 0.2863 0.0986 0.0925 0.0984

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000




16

Figure 1 Toll Mode Choice Model Structure Person Trips

Auto

Transit

Toll

Drive Alone (SOV) HOV2

Toll Free

Shared Ride

Drive Alone (SOV) HOV3+

Walk-Access

Shared Ride HOV2

HOV3+

Local Bus

Drive-Access Express Bus

NOTE: Structure is common to all 16 mode choice models reflecting four trip purposes and four time periods.



P&R

Kiss&Ride


17

% Reduction in Toll Person Trips

Figure 2a: Toll Mode Choice Model Sensitivity Analysis 70% 60% 50% 40% 30% 20% 10% 0%

50% Toll Increase 100% Toll Increase 200% Toll Increase AM Peak

Mid-Day

PM Peak

Night

24-Hour

Time Period

% Reduction in Toll Person Trips

Figure 2b: Toll Mode Choice Model Sensitivity Analysis 80% 60%

50% Toll Increase 100% Toll Increase 200% Toll Increase

40% 20% 0% HBW

HBO

NHB

Total

Trip Purpose

