|
Short-term Travel Model Improvements
Click HERE for graphic. Travel Model Improvement Program The Department of Transportation, in cooperation with the Environmental Protection Agency and the Department of Energy, has embarked on a research program to respond to the requirements of the Clean Air Act Amendments of 1990 and the Intermodal Surface Transportation Efficiency Act of 1991. This program addresses the linkage of transportation to air quality, energy, economic growth, land use and the overall quality of life. The program addresses both analytic tools and the integration of these tools into the planning process to better support decision makers. The program has the following objectives: 1. To increase the ability of existing travel forecasting procedures to respond to emerging issues including; environmental concerns, growth management, and lifestyle along with traditional transportation issues, 2. To redesign the travel forecasting process to reflect changes in behavior, to respond to greater information needs placed on the forecasting process and to take advantage of changes in data collection technology, and 3. To integrate the forecasting techniques into the decision making process, providing better understanding of the effects of transportation improvements and allowing decisionmakers in state governments, local governments, transit operators, metropolitan planning organizations and environmental agencies the capability of making improved transportation decisions. This program was funded through the Travel Model Improvement Program. Further information about the Travel Model Improvement Program may be obtained by writing to: Planning Support Branch (HEP-22) Federal Highway Administration U.S. Department of Transportation 400 Seventh Street, SW Washington, D.C. 20590 Short-Term Travel Model Improvements Final Report October 1994 Prepared by Cambridge Systematics, Inc. with Barton Aschman Associates Prepared for U.S. Department of Transportation Federal Highway Administration Federal Transit Administration Office of the Secretary U.S. Environmental Protection Agency Distributed in Cooperation with Technology Sharing Program U.S. Department of Transportation Washington, D.C. 20590 DOT-T-95-05 Table of Contents Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . .v Introduction . . . . . . . . . . . . . . . . . . . . . . . . .vii 1.0 Travel Surveys. . . . . . . . . . . . . . . . . . . . . .1-1 1.1 Household Travel Surveys . . . . . . . . . . . . . .1-2 1.2 Transit On-Board Surveys . . . . . . . . . . . . . . 14 1.3 External Station Surveys . . . . . . . . . . . . . .1-5 1.4 Commercial Vehicle Surveys,. . . . . . . . . . . . .1-6 1.5 Work Place Surveys . . . . . . . . . . . . . . . . .1-6 1.6 Stated-Preference Surveys. . . . . . . . . . . . . .1-7 1.7 Longitudinal or Panel Surveys. . . . . . . . . . . .1-9 1.8 Geocoding. . . . . . . . . . . . . . . . . . . . . 1-10 1.9 Expansion Factors. . . . . . . . . . . . . . . . . 1-11 1.10 References . . . . . . . . . . . . . . . . . . . . 1-14 2.0 Modeling Non-Motorized Travel . . . . . . . . . . . . . .2-1 2.1 A Typical Mode Choice Model Including Non-Motorized Travel . . . . . . . . . . . . . . . . . . . . . . .2-1 2.2 Issues in Modeling Non-Motorized Travel. . . . . . .2-2 2.3 Incorporating Measures of Pedestrian/Bicycle Environment in Travel Models . . . . . . . . . . . .2-3 2.4 Analysis of the Use of Pedestrian Environment Variables. . . . . . . . . . . . . . . . . . . . . .2-6 2.5 Summary. . . . . . . . . . . . . . . . . . . . . . .2-7 2.6 References . . . . . . . . . . . . . . . . . . . . .2-7 3.0 Land Use Allocation Models. . . . . . . . . . . . . . . .3-1 3.1 Available Land Use Allocation Models . . . . . . . .371 3.2 Resources Necessary for Land Use Allocation Models .3-3 3.3 Drawbacks of Land Use Allocation Models. . . . . . .3-4 3.4 Most Favorable Situations for Land Use Allocation Models . . . . . . . . . . . . . . . . . . . . . . .3-5 3.5 Alternatives to Land Use Allocation Models . . . . .3-5 3.6 Summary. . . . . . . . . . . . . . . . . . . . . . .3-6 3.7 References . . . . . . . . . . . . . . . . . . . . .3-6 4.0 Dynamic Assignment. . . . . . . . . . . . . . . . . . . .4-1 4.1 Description of Dynamic Assignment. . . . . . . . . .4-2 4.2 Available Software . . . . . . . . . . . . . . . . .4-3 4.3 Advantages of Dynamic Assignment . . . . . . . . . .4-4 4.4 Disadvantages of Dynamic Assignment. . . . . . . . .4-5 4.5 Summary. . . . . . . . . . . . . . . . . . . . . . .4-7 4.6 References . . . . . . . . . . . . . . . . . . . . .4-7 5.0 Air Quality Analysis Methods. . . . . . . . . . . . . . .5-1 5.1 Prediction-of Trips by Vehicle Operating Mode. . . .5-1 5.2 Improved Speed Models. . . . . . . . . . . . . . . .5-3 5.3 Assignment Post-Processors . . . . . . . . . . . . .5-4 5.4 Resources Necessary for Air Quality Analyses . . . .5-7 5.5 Drawbacks of Air Quality Analysis-Procedures . . . .5-8 5.6 Summary. . . . . . . . . . . . . . . . . . . . . . .5-9 5.7 References . . . . . . . . . . . . . . . . . . . . .5-9 Table of Contents (continued) 6.0 Modeling Trip Chaining Behavior . . . . . . . . . . . . .6-1 6.1 Recent Trip Chain Modeling Work. . . . . . . . . . .6-2 6.2 Data Resources Needed for the Incorporation of Trip Chaining into the Four-Step Modeling Process . . . .6-4 6.3 References . . . . . . . . . . . . . . . . . . . . .6-4 7.0 Mode Choice Modeling Improvements . . . . . . . . . . . .7-1 7.1 Incremental Logit Modeling . . . . . . . . . . . . .7-1 7.2 HOV Modeling . . . . . . . . . . . . . . . . . . . .7-3 7.3 Transit Captivity. . . . . . . . . . . . . . . . . .7-5 7.4 Transit Transfers. . . . . . . . . . . . . . . . . .7-6 7.5 Integrating Mode Choice Models with Trip Distribution, Trip Generation and Land Use Models. . . . . . . . .7-9 7.6 Model Transferability, Monte Carlo Simulation, and the Choice Between Toll and Non-Tolled Facilities. . . 7-11 7.7 Summary. . . . . . . . . . . . . . . . . . . . . . 7-14 7.8 References . . . . . . . . . . . . . . . . . . . . 7-15 8.0 Parking Analysis Procedures . . . . . . . . . . . . . . .8-1 8.1 Reallocation of Trip Ends to Parking Locations . . .8-1 8.2 Parking Cost Modeling. . . . . . . . . . . . . . . .8-3 8.3 Summary. . . . . . . . . . . . . . . . . . . . . . .8-4 9.0 Time-of-Day Models. . . . . . . . . . . . . . . . . . . .9-1 9.1 Hourly Factoring of Daily Trip Tables. . . . . . . .9-2 9.2 Peak-flour Trip Table Reduction to Reflect Network Capacity Constraints . . . . . . . . . . . . . . . .9-3 9.3 Traffic Assignment with Peak Spreading . . . . . . .9-4 9.4 Pre-Distribution Time-of-Day Models. . . . . . . . .9-5 9.5 Summary. . . . . . . . . . . . . . . . . . . . . . .9-6 9.6 References . . . . . . . . . . . . . . . . . . . . .9-6 10.0 Trip Table Estimation . . . . . . . . . . . . . . . . . 10-1 10.1 Available Trip Table Estimation Procedures . . . . 10-2 10.2 Resources Needed for Trip Table Estimation . . . . 10-3 10.3 References . . . . . . . . . . . . . . . . . . . . 10-4 11.0 Modeling of Trip Generation Input Variables . . . . . . 11-1 11.1 Use of Existing Data . . . . . . . . . . . . . . . 11-1 11.2 Use of Separate Models . . . . . . . . . . . . . . 11-2 11.3 Household Simulation . . . . . . . . . . . . . . . 11-4 11.4 References . . . . . . . . . . . . . . . . . . . . 11-5 12.0 Trip Assignment Issues. . . . . . . . . . . . . . . . . 12-1 12.1 Coding Transit Access Links Using GIS, . . . . . . 12-1 12.2 Toll Analysis - Using Link Time Penalties and Path Choice Models. . . . . . . . . . . . . . . . . . . 12-2 12.3 Instability of Highway Assignments in Saturated Networks . . . . . . . . . . . . . . . . . . . . . 12-3 List of Tables 7.1 Home-Based Work Mode Choice Model Coefficients From Selected Cities. . . . . . . . . . . . . . . . . . . . . . . . . .7-7 7.2 Home-Based Non-Work Mode Choice Model Coefficients From Selected Cities . . . . . . . . . . . . . . . . . . . . 7-12 7.3 Non-Home-Based Mode Choice Model Coefficients From Selected Cities. . . . . . . . . . . . . . . . . . . . . . . . . 7-13 iii Acknowledgements A number of travel demand modeling experts and practitioners were interviewed for this project. The authors would like to thank the following individuals for providing information for use in this report: Bernie Alpern, URS Consultants Jeff Bruggeman, KMPG Peat Marwick Kuo-Ann Chiao, New York Metropolitan Transportation Council Gary Davies, Garmen Associates Rick Dowling, Dowling Associates Jim Fennessy, Urban Analysis Group (TRANPLAN) Murray Goldman, Southern California Association of Governments Tom Golob, University of California, Irvine Paul Hamilton, Tri-County Regional Planning Commission (Lansing) Paul Hershkowitz, Michigan Department of Transportation Jim Hogan, Metropolitan Washington Council of Governments George Hoyt, George Hoyt & Associates (TRIPS) Keith Lawton, Metropolitan Service District (Portland) Hugh Miller, URS Consultants Elaine Murakami, FHWA Adiele Nwanko, Southeastern Michigan Council of Governments (Detroit) Bill Olson, URS Consultants Bob Parrott, San Diego Association of Governments Eric Pas, Duke University Chuck Purvis, Metropolitan Transportation Commission (San Francisco) Karl Quackenbush, Central Transportation Planning Staff (Boston) David Reinke, Research Decision Consultants Gordon Schultz, Parsons Brinckerhoff Robert Sicko, Puget Sound Regional Council Peter Stopher, Louisiana State University Theodore Treadway, Southwestern Pennsylvania Regional Planning Commission (Pittsburgh) Thomas Walker, Delaware Valley Regional Planning Commission (Philadelphia) Ken Yunker, Southeastern Wisconsin Regional Planning Commission (Milwaukee) v Introduction This document is a product of Phase II of a project to document short- term improvements to urban travel demand models. This study has been performed for the Federal Highway Administration by the consultant team of Cambridge Systematics, Inc. and Barton Aschman Associates. This effort is part of Track B of the Travel Model Improvement Program of the U.S. Department of Transportation. This report summarizes several potential improvements to the traditional urban travel demand modeling process. These improvements generally could be implemented in the short term in most urban areas, and many have been tested or are in use. This work reflects not only work which the consultants performed or with which they are familiar, but also the results of canvassing many travel demand modeling experts and practitioners. These included staff members of Metropolitan Planning Organizations and state DOT's, private consultants, software proprietors, and researchers. The work reflects the consultant team's interpretation of the information provided to us by these individuals. This report was edited by Thomas Rossi of Cambridge Systematics and was written by John Bowman, Thomas Rossi, Earl Ruiter, and Kevin Tierney of Cambridge Systematics and David Kurth and William Martin of Barton Aschman. vii 1.0 Travel Surveys Travel surveys are the basic tools used to gather travel information necessary to estimate and calibrate travel models. Large scale, regional travel surveys have been performed since the 1950s in most major cities. Four basic types of travel surveys have traditionally been performed for urban areas: - Household Travel Surveys; - Transit On-Board Ridership Surveys; - Commercial Vehicle (Truck) Surveys; and - External Station Surveys. There have been various levels of enhancements to each of the above types of surveys in recent years with, perhaps, the most activity being associated with household travel surveys. In addition, several other types of surveys have been used in urban areas to collect information on various aspects of travel in recent years. These surveys include: - Work Place Surveys; - Stated-Preference Surveys; and - Longitudinal or Panel Surveys. For each of the types of survey, Sections 1.1 through 1.7 briefly discuss the main uses of the data collected in the survey effort. Recent enhancements to the survey process, the effect of the enhancements on the modeling process, the positive and negative effects of the enhancements, and the level of effort to implement the enhancements are also discussed, as appropriate. There have also been significant enhancements in processing travel surveys recently - in preparing the surveys for use in estimating travel models and in obtaining descriptive statistics. These enhancements include improved geocoding procedures and more accurate determination of survey expansion factors. Sections 1.8 and 1.9 discuss these survey-related developments. 1-1 1.1 Household Travel Surveys Household travel surveys have been the most popular tool for collecting household-based travel by residents in a region. These surveys have provided the basis for the development of most aspects of the traditional four-step modeling process including trip generation (both trip production and trip attraction models), trip distribution, mode choice, and time-of-day/direction split factors. In the 1950s and 1960s, it was common for household travel surveys to include 10,000 to 20,000 or more households. The surveys were typically in- home interviews by trained surveyors. The tremendous cost of household travel surveys resulted in their evolution from large scale, in-home interviews of sampled households to "mail out-telephone collection" or self-administered, "mail out- mail back" surveys. Most surveys in the past 15-20 years have been administered using one of these two techniques. Current sample sizes tend to be much smaller than the original sample sizes, ranging from 1,500 to 2,500 households, although some cities (e.g., Minneapolis-St. Paul, Los Angeles, and New York City) have collected or will collect survey data from 10,000 or more households. The specification of small sample sizes have been based, in part, on information gathered from the early, large-scale travel surveys. Those surveys showed that the coefficient of variation for home-based trip rates tended to be around 1.0. This means that a sample size of about 1,600 is sufficient to estimate the regional average trip rate within five percent at the 95 percent confidence level. Larger sample sizes are required when specified levels of statistical confidence are desired for specific subareas or specific socioeconomic groups within the region. While the small sample sizes have proved generally acceptable for calibrating regional trip generation and trip distribution models, they are generally inadequate for producing sufficient data to calibrate mode choice models. This is because usually there are too few transit trips reported in areas with low transit mode shares. Both mail out-telephone collection and self-administered, mail out- mail back surveys have been used to keep the cost of data collection to approximately $100 per survey. Both techniques tend to be applied in a similar manner: - A household is recruited for the survey from a list of random telephone numbers or using random digit dialing; - Households agreeing to participate in the survey are assigned a travel day and sent travel diaries for all members of their household age five or older; and - For telephone collection surveys, the household is called one or two days after the travel day and the household and travel data are collected via the telephone; for 'mail back surveys, the households are called and reminded to return their completed travel diaries via a postage-paid return envelope. The data collection step in the mail out-telephone collection procedure is more costly than the self administered mail back survey technique and places a practical limit on the length 1-2 of the survey. The average telephone collection time per household is about one-half hour for typical travel surveys. However, the additional data collection cost can offset increased data editing costs with the mail back process. With the use of computer assisted telephone interviewing (CATI) techniques, the travel data are collected by a trained surveyor when the travel is still "fresh" in the respondent's mind. The surveyor can clarify and correct illogical responses while talking to the respondent and probe for easily forgotten trips. The ability to clarify responses interactively is especially important for collecting address information for geocoding. Response rates for the survey methods vary. The response rate for the recruiting step should be similar for both types of surveys. Based on a number of surveys performed in the past five years, approximate 55 to 65 percent of the qualified households contacted agree to participate in the survey. A household is "qualified" if it satisfies all criteria established for participation in the survey (e.g., the household is in the survey area, the residence is not a group quarters, dormitory, or barracks, and a responsible adult member of the household has been contacted). The second portion of the response rate, the percent of households originally agreeing to participate in the survey that result in being completed, usable surveys depends on a number of factors. These factors include the type of area being surveyed (very large cities will typically result in lower response rates), the actual data collection method, and the criteria used to determine whether or not a survey is deemed to be complete and usable. Several mail out- telephone collection surveys using paper and pencil to record responses have reported response rates as high as 80 to 85 percent. In Los Angeles, the response rate for the 1991 Southern California Association of Governments household survey using CATI was 55 percent. For mail out-mail back surveys, response rates are typically lower. One market research firm has reported that 40 to 45 percent of the agreeing households result in completed, usable surveys for this type of data collection. Household travel surveys have traditionally focused on trips made by household members. In recent years, some researchers and modelers have advocated changing the focus of household travel surveys from surveys of "trips" to surveys of activities of household members. This shift in focus has been driven by two major concerns. First, a "trip" is an abstract term used by travel modelers to describe travel from one point to another. As such, it is not always well understood by the population being surveyed and, as a result, trips are unreported because they are perceived to be unimportant or are simply forgotten. On the other hand, people understand their activities during the day they are at home, they work, they attend school, they shop, etc, and are more likely to recall all of their important activities during the day. Once all the activities performed during the day are recorded, it is easy to collect the travel that was necessary to get from each activity to the subsequent activity. Thus, it has been hypothesized that activity surveys minimize the under reporting problem. A second reason for developing activity-based surveys as opposed to trip-based surveys is to gather more information on the reasons for trip making. Few people travel for the sake of traveling. Most travel is necessary to link meaningful activities (home activities, work, school, shop, etc.). Thus, in order to properly understand and model the effects of the changing transportation supply and socioeconomic pressures on travel, we need to 1-3 understand the activities being performed and the decision processes a household uses in determining the activities that are performed during a day. While activity diaries have the potential of collecting more information relating to activity choice and its ancillary effect on travel decisions, they suffer from being longer. While trip-based travel surveys typically strive for a one or two page travel diary, many activity diaries tend to be small booklets, 20 to 30 pages in length. The length of the activity diaries typically means that they must be self-administered using mail back collection techniques. Another problem is that the time required to collect the data from the household may be increased significantly. For example, for an upcoming survey in Detroit for the Southeast Michigan Council of Governments, a 45-minute period for retrieving responses from households using CATI is estimated. Two recent enhancements to household-based travel surveys are the collection of travel by all modes and the collection of specific vehicle use for each trip made in an auto, van, or pick-up. The passage of ISTEA and the CAAA focused attention on alternative, non- motorized travel modes. Before the passage of these pieces of legislation, most travel surveys (especially trip-based surveys) focused on travel made by motorized modes. The incorporation of non- motorized travel into travel surveys has been relatively straight- forward and simple. Preliminary results from several southwestern cities have shown nonmotorized trips (i.e., walk and bicycle trips) to be about six to eight percent of the total trips made per day. The second enhancement, the collection of automobile information, has recently been implemented in several surveys in Texas and Arizona, for the city of El Paso and Tucson's Pima Association of Governments. Households are asked to enumerate and describe (make, model, fuel type, odometer reading) the vehicles available to the household. Vehicle information is also requested for each trip made during the day. The resulting information can be used to summarize a region specific estimate of the vehicle fleet -and information on cold starts and hot starts for air quality modeling. Other potential uses are tying vehicle type use to type of trip. For example, it might show that older vehicles in households tend to be used for work trips in two vehicle, one worker households. 1.2 Transit On-Board Surveys Transit on-board surveys have traditionally been used by transit operators to gain an understanding of transit users (ridership "profiles"). Such surveys have also been used by travel demand modelers to develop transit trip tables for travel model validation and to enhance household survey data for development of mode choice models. Survey techniques have remained relatively consistent over the last 30 years. Surveys tend to be self-administered and sufficiently short for a transit rider to complete while on the transit vehicle. A surveyor is typically on each surveyed transit run to distribute and collect survey instruments, answer questions, and take boarding counts for survey expansion. Some surveys have experimented with collection of data via lap-top computers. 1-4 Perhaps more improvement has been made in the use of on-board surveys. Recently, the results of on-board surveys have been combined with the results of household travel surveys to develop "choice-based" calibration data files for mode choice model estimation. This has been particularly important in cities that have collected small sample household travel surveys. In addition, the on-board surveys have been used in some cities.(for example, by the city and county of Honolulu) as the basis of incremental mode choice models for alternatives analyses. 1.3 External Station Surveys External station surveys have been used to provide information for trips traveling into and out of a region, and for trips traveling through the region. Survey techniques have included roadside interviews, postcard handout/mailback surveys, and license plate recording/survey mailing1. Roadside interviews consist of stopping some or all vehicles at the external station and interviewing the drivers. Advantages include high response rates from a captive group of respondents, and quick provision of the collected data. Disadvantages include delays and disruptions to traffic and the need to include many organizations (such as police). Postcard handout/mailback surveys consist of stopping vehicles and handing out postcard survey forms to be completed and mailed back. This is less disruptive than the roadside interviews since the vehicles are stopped only for a few seconds. The response rate is much lower, however, than for a roadside interview. The license recording/mailing method involves recording license pate numbers as vehicles go by, matching the numbers against motor vehicle registrations, and mailing surveys to vehicle owners. While the main advantage is that traffic is not disrupted, there are several disadvantages, including low response rates, recollection error by respondents, and license plate recording errors. Privacy issues are also a concern. Recording can be done by hand, audio tape recorder, portable computer, or video camera. In Boston, the license recording/mailback method was used in a 1991 external station survey performed by the Central Transportation Planning Staff. Stations with average daily traffic of greater than 10,000 were recorded using a relatively inexpensive ($2,300) video camcorder; lower volume stations were recorded manually. The estimated unit cost was $9.18 per completed survey. While the Boston survey had to use a fairly expensive method of manual transcription of plate numbers from the videotape, the report's authors note that computer software is now available for automated transcription. Costs vary among the methods described above, with the mailing costs of the latter two methods being offset to a degree by the higher labor costs of the roadside interview. In San Antonio, a study concluded that the mailback surveys cost much more than the roadside interviews2. This reflects the much lower response rates associated with mailback surveys. 1-5 1.4 Commercial Vehicle Surveys Commercial vehicle, or truck, surveys have traditionally been used to collect information on truck trips made in a region. These surveys have provided information on a significant amount of travel made within a region not captured by any of the other survey types. Some recent work has been focused on commodities movements, rather than truck trips (e.g., analogous to "activity"-based household travel surveys as opposed to trip-based household surveys). There are several problems inherent to the collection of data on truck travel. The cooperation of both shippers and drivers is needed. This is more problematic than for household surveys since the information on the amount and locations of truck trips may need to be kept confidential for competitive reasons. Another issue is determining the population to be surveyed. Truck registration information is not a complete source of data, and a disproportionate share of truck trips are external, even out of state. Because of these problems and budgetary constraints, few comprehensive truck surveys have been performed in recent years. However, the Texas Department of Transportation has been funding travel surveys in many of metropolitan areas of the state, including comprehensive truck surveys. Such a survey has recently been collected in the Beaumont- Port Arthur area for the Southeast Texas Regional Planning Commission, and a truck survey is currently being performed for the City of El Paso. 1.5 Work Place Surveys Work place surveys have been performed in a number of cities since 1984. Work place surveys have been used to gather detailed information on trips at the locations attracting those trips. Their main use has been to collect data for the calibration of improved trip attraction models, including the separation of attraction purposes (e.g., visitor, customer, employee). In addition, they can be an important source of information used in the estimation of other models in the traditional four-step process such as parking cost-walk distance information for mode choice model estimation. Trip attraction models have, typically, been calibrated based on the results of a household travel survey. Several techniques have been used, depending on the size and quality of the survey data. These techniques include zonal or district level regression analysis using the expanded trips summarized by attraction zone and the calculation of regionwide or area type trip attraction rates by dividing the expanded trips by the total land use or employees in the region. The calibration techniques are aggregate in nature and dependent on the combination of expanded data from a travel survey with independently generated land use or employment data. 1-6 Work place surveys provide disaggregate data that can be used to estimate trip attraction rates. The surveys collect information on the employer, the employees, and the visitors to the work place for each individual work place surveyed. Typically, the sample of work places surveyed is stratified by area type and industry type (basic, retail, and service) and the sample is drawn using techniques that ensure that both large and small establishments are represented in the survey. The resulting data can be used at a disaggregate level to estimate trip attraction models. Since individual work sites have been surveyed, it is possible to include explanatory values in the model not generally included in trip attraction models calibrated using aggregate means. For example, parking availability or transit accessibility could be considered in the estimation of the trip attraction rates. Work place surveys being performed in Texas, both by the Southeast Texas Regional Planning Commission and the City of El Paso, are also collecting information on whether a site is free- standing or non-freestanding with the hypothesis that vehicle trip attraction rates to non-freestanding locations will be different than the vehicle trip attraction rates to freestanding sites. While work place surveys offer the potential of providing detailed trip attraction data, they tend to be expensive and difficult to perform. There are at least four surveys or tasks that must be performed for each site: completion of an employer questionnaire including questions on total employment and attendance on the survey day; a survey of employees including travel questions on all trips made to or from the work site on the survey day; a survey of visitors to the work site on the travel day; and counts of all persons and/or vehicles entering the work site on the survey day. The average cost of surveying each work site ranges from $1,000 to $2,000. Due to cell quotas and the number of strata for a region (e.g., basic, retail, and service employment types for three area types results in nine strata), many of the surveys include 200-300 sites. Thus, the total cost for a work place survey can easily cost from $200,000 to $500,000. 1.6 Stated-Preference Surveys Travel demand models, especially urban area models, have traditionally relied on revealed-preference data collected from household and other surveys. These data represent individuals' reporting of their travel behavior during the survey period. Some analysts have argued that such data is insufficient for travel demand model estimation. The behavior that can be modeled is limited to the conditions observed during the survey, and models are not necessarily accurate when extrapolated to conditions beyond the observations. The hypothetical conditions may include new modes, highways, or policies, or even greatly increased congestion beyond what was experienced during the survey period. One technique that has been available for a number of years to analyze hypothetical travel conditions is stated-preference surveying3. This type of survey has been used extensively in market research and in long-distance travel demand modeling, but is just now beginning to be used in U.S. urban area travel modeling. In a stated-preference survey, each respondent is asked to make a travel decision for a scenario describing the available alternatives and their characteristics. For example, a respondent might be asked to choose, for a 1-7 trip from one location to another, among a set of travel modes, with each mode having associated cost, travel time, frequency of service, and other relevant characteristics. The choice set of modes might include one or more new modes, with realistic service characteristics, so that information on the demand for such modes could be obtained. An advantageous feature of stated-preference modeling is that one respondent can be asked to make a choice for many scenarios, not only greatly reducing survey costs, but yielding a better sense of a particular individual's sensitivities of the travel choice to changes in the different service characteristics. Data collection for stated-preference surveys is much more efficient than for revealed-preference surveys. Besides the ability to ask a single respondent to reply to many scenarios, the surveys do not require repeated contacts with the respondent for recruitment, information, reminders, and data retrieval. In theory, a respondent could be contacted and surveyed immediately; for complex scenarios, however, a written or computerized survey would probably be necessary. In addition, stated-preference surveys do not have to be taken at a specific time as do revealed-preference surveys. The latter should be taken during "average" conditions, usually fall or spring weekdays away from holidays or special events. Stated-preference surveys can be conducted at any convenient time for the surveyors and respondents. The major drawback to stated-preference surveys is the obvious: they do not represent actual travel behavior. Travel models must reflect how individuals would actually behave, not how they say they would act. Often, respondents will respond the way they would prefer to behave; for example, they would like to take transit or would like to try a new mode. There is a lack of information on how to correct for such problems. Another issue is how to combine revealed and stated- preference data for model development4. Stated-preference survey respondents must be presented with more information than they would actually have in making their travel decisions. For example, travel times would be specified precisely, as would auto operating costs for a single trip. So models developed using stated-preference data would not reflect the uncertainty travelers experience in reality. On the other hand, the more complete information makes model development easier since all relevant information is obtained from the definition of the scenarios and from the respondents. In models based on revealed-preference data, information must be obtained from external sources, such as transportation network models. Stated-preference surveys have been used extensively in travel demand modeling, but mainly for intercity travel. For example, Cambridge Systematics has used them for development of models to estimate ridership for proposed intercity high speed rail services in several locations5,6. However, urban area modeling based on stated- preference data has not been done, to the knowledge of the consultant team. There are a few areas where such efforts are underway or proposed, including an ongoing survey effort of the Metropolitan Service District and the Oregon Department of Transportation for Oregon's four urban areas and an ongoing survey to be used for the development of New Hampshire's statewide travel model for the New Hampshire Department of Transportation. 1-8 1.7 Longitudinal or Panel Surveys One drawback to household travel surveys is that they only provide a snapshot of travel for each household surveyed. The travel behavior of households is inferred by summarizing and analyzing on a cross- sectional basis many households with similar and dissimilar socioeconomic characteristics. Although a great deal of information can be inferred from household travel surveys, such snapshots do not provide true information on changes in a household's travel behavior over time based on changing influences. For example, a household's travel patterns change substantially when a child is born or adopted, when an child receives a driver's license, or when a second (or third) vehicle is added to the household. Average differences in trip rates, trip types, or mode shares between households with the different characteristics can be measured using household travel survey data. However, the changes occurring within an individual household are not measured. Snapshot survey data do not show the change in household and individual travel characteristics due to changes in transportation supply. Again, changes in travel behavior are inferred by comparing trip characteristics of disparate households and travelers. Mode choice models are, in effect, estimated by comparing the travel choices of, for example, traveler "A" who fives in a high density area and is well served by transit with traveler "B" who lives in a low density suburban area poorly served by transit. The sensitivity of a mode choice model might be substantially different if we could calibrate the models using data from both travelers "A" and "B" at two different times, say before and after transit service changes. Panel surveys are designed to provide data to answer the above types of questions. A sample of households is surveyed over time to determine changes in travel behavior of the same individual households under different socioeconomic and transportation supply conditions. Typically, panel surveys take place in "waves" which are two or three years apart. Panel surveys can provide a wealth of information not available from normal snapshot surveys. However, panel surveys are more difficult to control. The surveyor needs to maintain contact with a number of households over time - possibly for years. Since households cannot be compelled to participate in the survey, there can be a problem with dropouts over time. Finally, panel surveys are, by definition, long-term efforts measured in years rather than weeks or months for the collection of useful data. Many aspects of travel behavior can be studied only through the use of longitudinal data. Examples include: the process of information acquisition (e.g., becoming aware of a new transit service), experience and learning (e.g., trying a carpool), and behavioral turnover (e.g., switching between travel modes). Locational decisions - at the household (where to live, work, and shop), developer (where to build homes and activity centers), and government (zoning) levels - have impacts on travel that are often not realized for years. In such cases, lagged variables, which cannot be measured from cross-sectional data, would be needed to analyze the effects of such decisions on travel demand. 1-9 There are also potential drawbacks associated with panel surveys. One of the best known problems is the issue of attrition bias. For example, in the Puget Sound Regional Council's (Seattle area) panel survey, there was a 30 percent dropout of respondents between the first wave in fall 1989 and the last wave in fall 1992. Whatever the reasons for attrition whether related to socioeconomic or demographic characteristics, mobility issues, or something else - it may imply bias in model estimation. The use of panel data in the estimation of travel demand models presents some extremely difficult estimation problems. Correlation among unobserved error components ("heterogeneity") is likely to exist in such data sets. The apparent dependence of current choices on past choices, as shown through simple lagged dependent variable models, may actually be due to heterogeneity. 1.8 Geocoding Geocoding is an integral part of collecting travel survey data. Home locations, trip origins, and trip destinations must be assigned to geographic locations in the survey region in order to make the data useful for the estimation of trip distribution and mode choice models. In the past, geocoding meant assigning the home, origin, or destination to a traffic zone. However, with the growing use of geographic information systems (GIS), geocoding is defined by assigning a home, origin, or destination to an X and Y coordinate or to a location defined by longitude and latitude. Point specific geocoding provides several important capabilities. First, the geocoding is not zone specific and can easily be aggregated to varying zone structures. This provides important flexibility, especially if the results of two surveys are compared. For example, if a survey taken by one agency is coded to census tract and a second survey is coded to traffic analysis zones, it might be difficult to compare the results of the surveys at a common geographic level. However, if each is coded to X and Y coordinates, data from each survey can easily be aggregated to a common geography. Second, point specific geocoding provides additional accuracy for modeling purposes. For example, transit access distance can be very important to the estimation of mode choice models. If travel data are geocoded to points, it would be possible to determine the actual access distance from a home to a bus stop rather than relying on reported distances or average distances based on the location of zone centroids. Finally, the use of a GIS to perform geocoding can simplify the geocoding process through using automated, rather than manual, methods. This can provide a cost and time savings for geocoding. However, several items are crucial to the success of geocoding using GIS. These items include: 1-10 - A good address coverage file; and - Good survey address data. The Census Bureau TIGER files provide a good starting point for the coverage files; however, substantially improved match rates of addresses to points can be obtained if the TIGER files are locally updated and maintained on a routine basis, especially in rapidly developing areas. The success of matching address data coded on the survey records will be directly related to the quality of the address data. Quite frequently, travelers do not know exact addresses of their origins or destinations and provide only intersecting streets close to the location or general place names (e.g., City Hall). The match rate can be improved substantially by ensuring that high quality address data are collected during the survey. This includes getting street directions where important (e.g., 100 N. Main, rather than just 100 Main) and always obtaining the establishment name or place name as part of the address. The p lace name can be used along with the address information and a telephone directory to clarify ambiguous addresses. For example, McDonald's at Main and Mesa could be recoded to 100 N. Main prior to address matching with the GIS. 1.9 Expansion Factors Since the earliest travel surveys, the responses have been 'expanded' to the entire population by applying an expansion factor to each data record, whether these records represent households, persons, vehicles or trips. When these factors are appended to the survey data, they can be summed to provide expanded totals and subtotals for various subsets of the survey data. These totals and subtotals can then be used to develop aggregate travel models and to obtain descriptive statistics on the most likely socioeconomic characteristics and travel behavior of the surveyed population. The methods traditionally used to determine expansion factors for a survey involve dividing a population estimate by the number of responses obtained from this population. The population can be defined as the total number of units represented in the survey (in which case a single expansion factor, to be applied to all responses, would be determined) or as the number of units in any subset of this total. When subsets of the total are used, information must be available on the size of the subset both within the total population- and in the survey responses. For the size of the subset in the survey responses to be known, information on the variables defining the subset must have been collected in the survey. An example would be a plan to determine expansion factors for a household survey for subsets characterized by categories based on two variables: household size and auto ownership. This plan is feasible if information is available, from the decennial census for example, on the joint distribution of households by these two variables; and if each household was asked to report on its size and auto ownership. 1-11 In recent years, enhanced survey expansion methods have been applied by a number of planning agencies. These methods involve one or more of the following strategies, each of which is discussed in the paragraphs which follow: - Using an iterative adjustment process to match two or more marginal distributions of demographic, economic, locational, and/or travel behavior characteristics; - Using a sequential process to refine an existing survey expansion strategy to match additional characteristics of the population; and - Using mathematical programming methods to ensure that the resulting expanded data is as dose as possible to a large number of measurements of the total population. Iterative Proportional Fitting The most common iterative proportional fitting (IPF) process in travel forecasting is the Fratar trip table growth or adjustment process. In this process, a seed trip table is adjusted to match trip origin and destination targets by zone. These targets represent two marginal distributions which are matched by iteratively adjusting all rows and columns of the trip table. This process can be extended to any number of dimensions, adjusting a multidimensional 'table' of survey observations, for example, to match observed characteristics of the population such as total population by household size, auto ownership, residence zone, and mode of travel for a particular trip purpose. After enough iterations have been performed to match each of the target marginal distributions, the quantities in each cell of the table can be divided by the corresponding numbers of observations to provide expansion factors which, generally, will be different for each cell in the table. IPF has a number of advantages over the usual process of determining expansion factors directly by cell. Because the complete distribution in the population over all dimensions or variables is not required, it is usually possible to increase the number of variables to be matched by the expansion factors. For example, the expanded survey can match both work trips by mode of travel and by housing type of the tripmaker, without requiring a joint distribution for the population of work trips by these two variables. A second advantage is that IPF automatically compensates for cells in the multi-dimensional table which may have non-zero values in the population, but have no observations in the survey data. Sequential Refinements of Expansion Factors Often, transportation analysts find, after beginning to use an expanded survey, that it poorly matches some variable observed in the population which was not used in the expansion process. For example, a survey expanded by residence district, household size, and household income may be found to poorly match total vehicle registration data. Frequently, these types of problems are addressed by adjusting the original factors based on auto ownership to obtain a better match to this new variable. Usually, this second round of survey expansion results in obtaining a better fit to the new variable at the cost of 1-12 a poorer match of the original variables. The latter results are typically minimized by using a variation of the IPF process in which the original cells with unique expansion factors form one dimension, and the new variable becomes the second dimension. The target values for the first dimension retain the levels provided by the original expansion factors, while the target values for the new dimension are obtained from an external source. This approach has been used recently in both Chicago7 and Boston8 to expand travel surveys conducted for the Chicago Area Transportation Study and the Central Transportation Planning Staff, respectively. Although this process provides a reasonable means of improving a travel survey which was previously expanded, travel forecasters should consider carefully if a completely new expansion effort should be carried out rather than an adjustment process. The former may be preferred if any joint distributions not used before are to be employed or if, by redesigning the complete expansion strategy, it will be possible to avoid having differing household and trip factors or differing trip factors by purpose. Applying Mathematical Programming Methods As survey expansion procedures become more complex and involve more types of observed data, differences are often detected in information obtained from different sources. For example, the total passenger vehicles registered in an area may not match the total vehicles reported in the decennial census. If both sources are to be used - one to provide vehicles by type and the other households by auto ownership level, for example then the analyst must decide how to reconcile the differences, if possible, considering variations in definitions and in reporting accuracy. Even after an attempt at reconciliation, however, some level of difference may remain. By using mathematical programming methods, these differences need not be resolved, because the objective used in these methods is typically the minimization of differences between survey and observed data, rather than the exact matching of the observed information. Beyond this advantage, programming methods allow a much greater flexibility in the structure of the observed data, and allow the error ranges of the observed data to be considered in determining expansion factors. One example of the flexibility possible when mathematical programming methods are used is the possibility of using observed locational information recorded using one zone system and survey data based on another set of zones; another would be the use of observed bridge crossings by time of day, in addition to demographic data, to expand a household-based survey. Error ranges, even simple relative indices of observed data accuracy, can be incorporated into the objective functions of mathematical programming approaches to ensure that, if inconsistencies in the observed data exist, more weight will be assigned to the more accurate data. Mathematical programming methods applicable to the survey expansion problem have been developed by Ben-Akiva,9 who considers the general problem of adjusting data from multiple sources to achieve the maximum level of consistency. List and Tumquist10 have applied these methods to the problem of estimating truck trip tables; their approach is readily adaptable to the survey expansion problem. Finally, the matrix estimation function in the TRIPS travel forecasting package11 is a mathematical programming tool which can be 1-13 adapted to the survey expansion problem in which trip ends, trip interchanges, and screenline counts are available as observed data. 1.10 References 1. Miller, K., T. Harvey, P. Shuldiner, and C. Ho. "Using Video Technology to Conduct the 1991 Boston Region External Cordon Survey." Presented at the 72nd Annual Meeting of the Transportation Research Board, Washington, D.C., January 1993. 2. McKinistry, D. and L. Nungesser. "An Evaluation of On-Site Administered Origin-Destination Survey Methodologies: Postcard Mailback Vs. Interview." Proceedings of the Third National Conference on Transportation Planning Methods Applications, April 1991. 3. Ben-Akiva, M. and S. Lerman. Discrete Choice Analysis: Theory and Application to Travel Demand. MIT Press, Cambridge, Massachusetts, 1985. 4. Ben-Akiva et al. "Combining Revealed and Stated-Preference Data." Prepared for publication in Marketing Better. 5. Cambridge Systematics, Inc. and Hague Consulting Group. VFT Feasibility Study. Market Analysis, Final Report, July 1988. 6. Cambridge Systematics, Inc. Memos on the Boston-Albany-New York MagLev Feasibility Study, 1994. 7. Kim, H. "Factoring Household Travel Surveys," 8. Wang, C.Y. and I. Harrington. "Revised Expansion of 1991 Regional Household-Based Travel Survey," Central Transportation Planning Staff, 1994. 9. Ben-Akiva, M. "Methods to Combine Different Data Sources and Estimate Origin-Destination Matrices," Transportation and Traffic Theory: 459-481, ed. N.H. Gartner and N.H.M. Wilson, Elsevier, New York, 1987. 10. List, G.F. and M.A. Tumquist. "Estimating Multi-Class Truck Flow Matrices in Urban Areas," presented at the 73rd Annual Meeting of the Transportation Research Board, Washington, D.C., January 1994. 11. MVA Systematica. TRIPS Documentation, Working Survey, England, October 1990. 1-14 2.0 Modeling Non-Motorized Travel Traditionally, urban travel demand models have focused on travel by highway and transit modes. Few U.S. urban models analyze non- motorized travel, including walking and bicycling. Many models analyze only auto travel, or analyze transit only for work trips. Recent interest in reducing congestion and improving air quality, due in part to federal legislation, has stimulated interest into analyzing non-motorized modes. Several urban areas, including Los Angeles, Portland, and San Francisco, have been analyzing walking and bicycling modes. The general process is to incorporate these modes (sometimes as a single mode) into traditional mode choice models. This process requires good data on non-motorized travel in the urban area, usually from household travel surveys. It is important to recognize that simply adding these modes into models of other modes without recognizing the different factors affecting their use is unlikely to succeed. Another issue concerning non-motorized travel is transit access modes, (i.e., whether transit users walk (or bicycle) or use an automobile to get to the transit station). While many travel models do not consider access modes to transit in the mode choice process, it is becoming increasingly common to do so, particularly for work trips. This is often done using a nested logit formulation. Many models consider access mode choice to transit, but not walking as a separate mode. 2.1 A Typical Mode Choice Model Including Non-Motorized Travel Portland has one of the more sophisticated travel demand modeling systems in the U.S.1 This model system was developed by the Metropolitan Service District ("Metro"), which serves as the Metropolitan Planning Organization for Portland. The mode choice model is a two-step process using a logit formulation. First, the choice between using a motorized or non-motorized mode (walk or bicycle) is determined. Then, for motorized travel, the choice between auto and transit (including, for work trips, single versus multiple occupant auto and walk access versus auto access to transit) is determined. While the model is estimated as a sequential choice model, not as a nested logit model, it behaves in a similar manner. 2-1 The choice between motorized and non-motorized travel is estimated as a binary logit model using data from a 1985 household travel survey. Walking and bicycling are considered together as a single mode since the survey yielded an insufficient number of trips to estimate bicycle as a separate mode. Models were developed for five purposes: home- based work, home-based college, home based other, non-hon-d--based work, and non-homebased non-work. (Mode splits for a sixth purpose, home-based school, were determined using a separate procedure not involving a mode choice model.) For home-based trips, households were stratified by whether or not they owned an auto. The models were estimated only for households which owned one or more autos. For households without autos, simple percentages were used in applying mode splits. As an example, the home-based work non-motorized mode choice model estimated by Metro in 1988 for households with autos is presented. The utility function for the motorized mode is: U = 1.299 + 0.718 TDIST - 1.347 VALCAR1 where: U = utility TDIST = trip distance in miles VALCAR1 = 1 if household has fewer cars than workers 0 otherwise The utility is zero for the walk/bike mode. In addition, the model is applied only for trips that are less than a certain length, based on the survey data. The model implies that the probability of choosing a motorized mode increases with the trip distance and is inversely related to auto availability. This model fairly accurately reflected the survey data in terms of the percentage of non-motorized mode users on an aggregate basis. The project "Making the Land Use, Transportation, Air Quality Connection" (LUTRAQ) later enhanced this model to better reflect the quality of the pedestrian environment and the effect of development density on mode choice2. This effort is described later. 2.2 Issues in Modeling Non-Motorized Travel Probably the most critical issue in analyzing pedestrian and bicycle travel is defining what constitutes a trip by these modes. Many areas have no data with which to estimate nonmotorized mode choice models because their travel surveys did not ask about such trips. In addition, non-motorized trips, especially bicycle trips, are much less numerous than auto trips. Even when they are asked about in household travel surveys, they may not be 2-2 numerous enough to estimate detailed models. While the same is true for transit trips in many areas, transit riders can be surveyed separately; this would be impossible for pedestrians or cyclists. Household travel surveys which do include questions about non- motorized tripmaking must be carefully worded to obtain information on all such trips. Many respondents apparently do not consider short walk trips worth reporting, especially nonhome-based trips. This is demonstrated by a comparison of survey data between the Portland and San Francisco areas3, which showed for higher walk trip rates in similar areas for San Francisco. This implies under-reporting of walk trips in Portland. The conclusion is that great care must be taken to obtain sufficiently accurate information to estimate models of non-motorized mode choice. Since many areas do not have the survey data necessary to estimate non-motorized mode choice models, it might seem attractive to attempt to transfer an existing model as is frequently done for other mode choice models. Although the consultant team is unaware of any attempts to transfer models of non-motorized mode choice, it is clear that even greater caution must be taken than when transferring more traditional mode choice models. Many of the factors that affect the choice of whether to use the walk/bicycle mode are unmeasured in most models, and can vary greatly between cities. These factors include climate, terrain, age of population, and density of development. 2.3 Incorporating Measures of Pedestrian/Bicycle Environment in Travel Models There have been a few studies on how the quality of the pedestrian/bicycle environment can be quantified for use in travel demand models. Two are documented here: one by the Maryland-National Capital Park and planning Commission (M-NCPPC), the other by the LUTRAQ project. The M-NCPPC developed a nested logit mode choice model for home-to- work trips4. While the model did not include walk or bicycle as a separate mode, it did include walk/ bike access to transit as a submode to the transit mode. To better estimate this choice, the model includes a variable which is an index of pedestrian and bicycle friendliness. This variable is included at the auto vs. transit and walk vs. auto access to transit levels of the model. 2-3 The index consists of five factors as follows: Amount of Sidewalks .00 No sidewalks .05 Discontinuous, narrow .15 Narrow sidewalks along all major streets .25 Adequate sidewalks along all major streets .35 Adequate sidewalks along most streets with some off-street paths .45 Pedestrian district with sidewalks everywhere, pedestrian streets, and auto restraints Land Use Mix .00 Homogeneous land use within easy walking distance .10 Some walk accessible lunch time service retail in employment centers .20 Mixed land use at moderate density .25 Mixed land use at high density Building Setbacks .00 Mostly set-back sprawled campus style .05 Mixed campus style but clustered with bus stops within walking distance .10 Few or no building setbacks from transit-accessed street Transit Stop Conditions .00 No shelters .05 Some bus stop shelters .10 Widely available bus stop shelters Bicycle Infrastructure .00 Little or none .05 Some cycle paths or routes .10 Many cycle paths, lanes, or routes forming network Thus the index can range from zero to one, with higher numbers representing more pedestrian friendly areas. Note that some of the components are specific to a transit access mode choice and would not be appropriate in estimating the probability of walking all the way to a destination. The value of the index was estimated for each zone in the M-NCPPC model by an independent consultant to M-NCPPC. Since the measures are somewhat subjective, they are subject to the judgment of M-NCPPC and its consultant. However, a more formal process involving more participants would have been both time-consuming and costly. This index, given the variable names TSI_ORG and TSI_DES for the origin and destination zones, respectively, for a trip, was included in the transit access submodel as specific to the walk access mode. The coefficient for TSI_ORG was estimated at 2.157, which has the correct sign, indicating a positive correlation between the index and the probability of 2-4 choosing the walk access submode. This and other coefficients imply that a 0.1 increase in the index is equivalent to about a five-minute increase in travel time or a 50 cent increase in cost for the auto access submode. This seems highly significant. For the main mode choice model (auto vs. transit), the coefficient for TSI_ORG is 1.597 (specific to transit), which again has the correct sign. In this case, a 0.1 increase in the index is equivalent to about a one-minute increase in auto out-of-vehicle time, which seems reasonable. The LUTRAQ project introduced another type of pedestrian environment variable into the mode choice and auto ownership models in Metro's Portland travel demand forecasting system5. The pedestrian environment factor (PEF) includes four components: - Sidewalk availability; - Ease of street crossing; - Connectivity of street/ sidewalk system; and - Terrain. Each zone was ranked for each component on a scale of one to three, with higher numbers representing higher quality pedestrian environments. So the PEF can range from four to 12. As described above, the Portland mode choice model is a two-step process, with the nonmotorized/motorized mode choice being performed first. The PEF was introduced into the initial model and was found to be a significant indicator for four trip purposes: homebased work, home-based other, nonhome-based work, and nonhome-based non-work. As an example, the revised home-based work model is shown here. (Other variables were also introduced, including auto availability and development density variables, and so the model differs somewhat from the original model shown above.) The utility function for the motorized mode for the revised model is: U = 1.717 + 0.705 TDIST - 0.954 VALCAR1 + 0.408 VALCAR2 - 0.0000191 TOT1M - 0.0632 PEF where: U = utility TDIST = trip distance in miles VALCAR1 = 1 if household has fewer cars than workers 0 otherwise. VALCAR2 = 1 if household has the same number of cars as workers 0 otherwise. 2-5 TOT1M = total employment within one mile of zone PEF = pedestrian environment factor as described above The utility is zero for the walk/bike mode. Note that a measure of density, which is part of the M-NCPPC pedestrian friendliness index, is used as a separate variable in the Portland model. As in the M-NCPPC model, a subjective evaluation of each zone, in this case done by Metro, was performed to provide the zonal PEF values. In the home-based work model, an increase of one in the PEF is equivalent to a decrease of about 500 feet in walking distance. This relationship seems reasonable. An evaluation of the effectiveness of the PEF (and of the other model revisions) was performed5. While the original model performed quite well in terms of estimating walk/bike trips over the entire region (compared to the 1985 survey results), the model under-predicted walk/bike trips in the most pedestrian friendly and highest density areas by 20 to 25 percent. The revised model corrected the error to within seven percent. Similar results were noted for other trip purposes. The LUTRAQ project also updated the main mode choice model (auto vs. transit) and the auto ownership model. The PEF was found to be a significant indicator of travel behavior in both models. 2.4 Analysis of the Use of Pedestrian Environment Variables Both pedestrian environment variables described above are subjective numeric indices (with the possible exception of the terrain component of the. LUTRAQ model). Use of such variables requires an assessment of every zone in the analysis area for each component of the variable. This requires not only a significant amount of time (though not unreasonable given the overall effort required to estimate an urban travel demand model), but a detailed knowledge of the entire region. Although a consultant was used in the Maryland example, it would seem likely that in most areas the MPO staff or other local officials would possess the best knowledge for this task. In addition, it might make sense to have several people perform the evaluations separately and average the results so that the biases or uneven regional familiarity of a single analyst would not skew the values of such a subjective variable. Another issue dealing with the use of pedestrian environment variables as discussed above is that they are zone-based; they do not consider variations within a zone. This means that the values are totally dependent on the zone boundaries; combining two zones, for example, would likely change the pedestrian variable values in both. This level of aggregation is very common in existing U.S. urban area models, especially regarding variables that are obtained from the model's transportation networks. However, the need for zone 2-6 level aggregation is being reduced greatly as geographic information systems (GIS) technology is becoming integrated with travel modeling. With such a system in place, all network-related variables could be determined using the geocoding of each trip's origin and destination. In the Portland model cited above, the socioeconomic variables are already disaggregate (based on the household survey), and so only the density variable (TOT1M) and the PEF would be zone-based. Variables such as TOT1M, which have already been normalized, in this case to employees within one mile, could also be made location-specific by using GIS. It might be possible to use GIS to help determine values for a pedestrian environment index although this has never been tried. Data on sidewalk locations and widths could be added to the transportation network and read from the GIS. Even more subjective data such as "ease of street crossings" could be entered into a GIS using overlays. This would require a significant amount of work, but even the manual applications that have been used to date required time-consuming zone by zone evaluation of several factors. Aggregation error is already a concern in urban travel demand modeling. It is worth researching whether the need to have an aggregate (zone level) measure of the pedestrian environment can be eliminated. 2.5 Summary The results of the Portland model show that non-motorized travel can be estimated using traditional travel models. Modeling such travel can be improved by incorporating Measures of the quality of the pedestrian/bicycle environment, However, great caution must be taken in estimating such models, particularly regarding the use of travel survey data. 2.6 References 1. Metropolitan Service District. "Travel Forecasting Methodology Report, Westside Light Rail Project," September 29, 1989. 2. Cambridge Systematics, Inc., S.H. Putman Associates, Calthorpe Associates, and Parsons Brinckerhoff Quade and Douglas, Inc. "Making the Land Use, Transportation, Air Quality Connection, Volume 4: Model Modifications," prepared for 1000 Friends of Oregon, November 1992, pp. 7-25. 3. Cambridge Systematics, Inc., Calthorpe Associates, and Parsons Brinckerhoff Quade and Douglas, Inc. "Making the Land Use, Transportation, Air Quality Connection: The 2-7 LUTRAQ Alternative/Analysis of Alternatives - An Interim Report," prepared for 1000 Friends of Oregon, October 1992, pp. 97-101. 4. Replogle, M. "M-NCPPC 1988 Logit Mode Choice Model for Home-to-Work Trips," April 9,1991. 5. Rossi, T., K. Lawton and K. Kim. "Revision of Travel Demand Models to Enable Analysis of Atypical Land Use Patterns," proceedings of the Fourth National Conference on Transportation Planning Methods Applications, Transportation Research Board, September 1993. 2-8 3.0 Land Use Allocation Models Land use allocation models improve the traditional transportation planning modeling process by adding the ability to reflect the effects of transportation accessibility and other measures on the locations of future development. The traditional four-step travel modeling process is sequential and ignores the effects of transportation access (which can be measured in the outputs of trip assignment) on land use, and, therefore, trip generation. Transportation professionals are becoming increasingly aware of the need to incorporate such relationships into travel demand models, and legislation such as the 1990 Clean Air Act Amendments requires that such factors be considered. Some general questions related to incorporating land use allocation models into the travel demand forecasting process include: 1. What land use allocation modeling procedures are available to urban areas? 2. What capabilities do the available models provide to improve the travel modeling process? 3. What resources are needed to develop a land use allocation modeling capability for an urban area? 4. What are the drawbacks of using a land use allocation model? 5. Where are such models useful, practical, and accurate? 6. What alternatives to land use allocation models exist? This document does not describe the existing land use allocation models in detail. For a more thorough discussion, refer to Volume 1 of the project report for "Making the Land Use, Transportation, Air Quality Connection" (LUTRAQ)@ 3.1 Available Land Use Allocation Models There have been few attempts to document land use modeling procedures. The most recent of which the consultant team is aware was conducted in 1991 by Cambridge Systematics and the Hague Consulting Group for the LUTRAQ project for 1000 Friends of Oregon1. This document drew on the findings of the ISGLUTI (International Study Group on Land Use/Transport Interaction) study in the United Kingdom, which were published 3-1 in 19882. Since these documents were published, there have been a few additional models that have been developed or documented in the U.S. The ISGLUTI study included nine models, only two of which were commercially available: MEP, now known as MEPLAN, and ITLUP (DRAM/EMPAL). The LUTRAQ report identified three other models that are not available, and two optimization models (TOPAZ and TOPMET) which were not being sold at the time of the report, but for which marketing plans existed. The LUTRAQ report classified the models into three groups as follows: 1. ITLUP, TRANSTEP, TRACKS; 2. MEPLAN, TRANUS; and 3. TOPAZ, TOPMET. The first group is dominated by ITLUP, developed by Dr. S.H. Putman. ITLUP consists of the submodels EMPAL and DRAM. These models allocate employment and households, respectively, to zones. The allocations are based on such factors as the accessibility (travel time/cost) to other trip generators and available land. Accessibility information comes from a transportation (generally highway) network. The basic formulation is an improvement to that put forth originally by Lowry3 in 1964, which is similar to the gravity model. TRANSTEP and TRACKS are less well, developed models from Australia based on the Lowry approach. ITLUP is the only commercial land use allocation modeling program used in the U.S. It has been used in a number of U.S. cities, including Dallas, Kansas City, Houston, Los Angeles, Portland, and San Francisco, and a few locations abroad. In general, the results have been satisfactory. The major criticisms of the ITLUP/Lowry approach is that little attention is paid to the internal economics of the land market, in particular to how land values are affected by factors other than accessibility. On the other hand, this simplification makes the models much more practical and less data intensive. MEPLAN5 has been used exclusively outside the U.S. to the knowledge of the consultant team. It was developed in the U.K. by Marcial Echenique and Partners. MEPLAN focuses directly on the competition and resulting rents as a means to analyze the available supply of land and the demands of various activities. Network-based accessibility measures are incorporated, as in ITLUP. The economic interactions increase the ability of the model to accurately forecast land use, but make it very data intensive and time-consuming to calibrate (though not necessarily to apply). Calibration is a trial and error process that requires a significant amount of time from experienced personnel. TOPAZ and TOPMET1 (a derivative of TOPAZ), in their most widely used forms, have a different orientation than the other models described above. Where the other systems attempt to predict what will happen, TOPAZ and TOPMET attempt to determine what should happen, given the objectives of the user. Thus TOPAZ would be difficult to use as a forecasting model, but might be useful as a planning tool. 3-2 There have been a few other land use allocation modeling efforts in the U.S. that do not use commercially available software. In New York, the NYREG model5 allocates household locations based on accessibility to other activities (measured by employment) and economic factors. Unlike ITLUP, NYREG incorporates supply side variables in the allocation of land uses. However, it allocates only residential land use; employment locations must be specified as inputs. Resource Systems Group of Vermont has recently applied land use allocation in both New Hampshire (for the New Hampshire Department of Transportation and the Seacoast MPO) and Florida6. RSG improves on the Lowry/ITLUP procedure by using a logit formulation to compute a generalized accessibility function (as opposed to accessibility to the primary workplace) and a composite multimodal impedance function rather than a highway-based function. These types of improvements were first implemented as improvements to Seattle's DRAM/EMPAL model by the Puget Sound Council of Governments7. To summarize, the ITLUP model is the only widely used land use allocation procedure in the U.S. MEPLAN has been used in many foreign cities. The Lowry/ITLUP formulation has been improved upon by a limited number of MTO's and consultants. 3.2 Resources Necessary for Land Use Allocation Models For the most part, the amount of data needed for the application of land use allocation models is not significantly greater than for a complete four-step transportation planning model. For ITLUP, the main additional requirements would be information on existing land use and available land, regional population and employment forecasts by category, and possibly more detailed employment information, depending on how detailed the transportation model is. This type of information is generally available in most areas. For MEPLAN, the same information as for ITLUP would be needed, plus floor space by activity for each zone; trip tables for total, work, and shopping trips; and elasticities of household consumption of space, transportation, and goods, which might require special surveys. For calibration, however, additional data would be needed. Specifically, since the effects of transportation and other variables on land use is a lagged effect, the models would require data collected over a period of time. So where a traditional transportation model would require data for a single-base year, a land use allocation model would require data for at least two years, usually about five years apart to simulate the lagged effects. In this sense, the land use allocation model requires twice as much data as a transportation model. This may be problematic for many areas without good historical data on land use, employment, or transportation use. 3-3 3.3 Drawbacks of Land Use Allocation Models The major drawback to using land use allocation models is the time needed to calibrate them. Because of the large number of variables, and the fact that land use models require calibration over two-base years, it usually takes a year or two to calibrate such a model, sometimes longer. In addition, only the model developers, a few consultants, and some local planners have the expertise necessary to calibrate a land use allocation model. Other than in areas which already have land use allocation models, it is rare to find someone on an MTO staff with such experience. An important issue that is not dealt with well in land use allocation models is the effects of areas external to the model area. While this is a problem in transportation models, it is much more critical in land use models. In a transportation model, knowing the number and type of trips crossing the area cordon is usually sufficient; it is unimportant how these travelers behave outside the model area. But in a land use allocation model, development decisions made outside the model area can have a substantial effect on travel within the area, and the effects of development inside the cordon are not limited to the model area. This is particularly important in smaller areas located close to larger ones, such as Providence or Trenton, but could even be a major issue in large areas such as Baltimore and San Diego. It is not well known how the various land use allocation modeling efforts have dealt with the issue of external areas. Since most land use allocation modeling efforts in the U.S. are relatively recent, there is little information on the long-term accuracy of such models. It would be logical to assume that the models should be updated with more recent information as it becomes available. For example, a model first developed in 1988 whose first forecast years were 1993 and 1998 could be updated with observed 1993 data to revise the 1998 forecast. Is such an update valid without model recalibration? How often should a land use model be recalibrated? It is unclear what experience is available to answer such questions. It must be recognized that the regional forecasts are inputs to the land use allocation modeling process, and the results can be only as good as the regional inputs. Regional forecasts are not necessarily accurate, and can be a significant source of error. Finally, the consultant team knows of no study to determine whether land use allocation models produce significantly better results than do non-quantitative methods which are commonly employed in many urban areas. While most would agree that transportation accessibility is an important factor in locational decisions, it is unclear to what extent it is overlooked in manual forecasts. Since land use allocation models are much more expensive and time-consuming to implement, it would be worthwhile to know what increase in accuracy is being bought through the use of land use models. 3-4 3.4 Most Favorable Situations for Land Use Allocation Models Given the issue of external effects, it would seem that the ideal location for implementation of a land use allocation model would be a relatively isolated urban area such as Portland, Oregon, or Columbus, Ohio. Such an area would not only have few external impacts, but would also be capable of providing more accurate regional control forecasts. The other requirements for land use allocation modeling would be the existence of a good transportation model capable of providing the necessary accessibility data and good historical land use, employment, and population data by sector. 3.5 Alternatives to Land Use Allocation Models For most areas, the current alternative to land use allocation modeling is to use manual forecasts of land use, employment, and/or population by zone. While local knowledge can be quite valuable in such forecasts, it is impossible to consider, without some basis for quantifying such effects, the impacts of changes in transportation accessibility throughout the area, due to both infrastructure improvements and increased congestion. However, if such a process incorporates participation from local and private groups.. it may often be accepted more readily than model outputs as a basis for assigning growth. Some areas have used a Delphi process8 to allocate expected growth. In such a process, a panel of experts provides input into an iterative process, which yields a consensus on growth allocations. The advantages include a wider range of opinions and expertise than a manual forecast by MPO staff, a somewhat more quantitative process than a manual procedure (though much less so than a model), and local support for the process gained by incorporating a variety of expert local opinions, which may include some developers themselves. The Delphi process may be more expensive and time-consuming than a manual forecast, but should be considerably less so than use of a land use allocation model. The main disadvantage of such a process would be the lack of any technical basis for the results, particularly the impacts of changes in accessibility and other factors on development patterns. There have been proposals to incorporate some features of land use allocation modeling into transportation models. For example, workplace choice models for residents, or residential location choice models for workers, may be used as part of the trip distribution step. As far as the consultant team knows, this has not been tried as part of any U.S. urban area model, but has been discussed for the New York area. 3-5 3.6 Summary ITLUP (DRAM/EMPAL), from S.H. Putman Associates, is the only land use allocation model widely used in the U.S. It is based on the Lowry formulation and has been used successfully in many cities. Some improvements to the Lowry/Putman formulation have been made in other areas, using ITLUP or specially developed software. ITLUP allows the consideration of transportation accessibility in determining future land use development, but does not explicitly consider economic factors, such as land prices, in location decisions. The installation of ITLUP or a similar model would require, in most areas, the participation on S.H. Putman Associates or another consultant familiar with such models. MEPLAN is a commercially available model that has been used in many cities abroad. It has the advantage over ITLUP of explicit consideration of economic factors other than transportation accessibility and land availability in location choice. However, it requires a great deal of data and a long time to calibrate. The use of MEPLAN would likely require the participation of the developer or another foreign consultant to install. 3.7 References 1. Cambridge Systematics, Inc. and Hague Consulting Group. "Making the Land Use, Transportation, Air Quality Connection, Volume 1: Modeling Practices," prepared for 1000 Friends of Oregon, October 1991, pp. 9-38. 2. Webster, F.V., P.H. Bly, and N.J. Paulley. "Urban Land Use and Transportation Interaction," Avebury 1988. 3. Lowry, I.S., A Model of Metropolis, RM-4035-RC, Santa Monica, California, The Rand Corporation, 1964. 4. Hunt, J.D. and M.H. Echenique. "Experience in the Application of the MEPLAN Framework for Land Use and Transport Interaction Modeling," proceedings of the 4th National Conference on Transportation Planning Methods Applications, September 1993. 5. Anas, A. and R. Armstrong. "Land Values and Transit Access: Modeling the Relationship in the New York Metropolitan Area, An Implementation Handbook," final report, prepared for Capital Development Division, Urban Mass Transportation Program, 1992. 6. Marshall, N.L. and S.J.C. Lawe. "Land Use Allocation Models for Multi-County Urban and Suburban Areas," proceedings of the 4th National Conference on Transportation Planning Methods Applications, September 1993. 3-6 7. Watterson, W.T. "Adapting and Applying Existing Models: DRAM and EMPAL in the Seattle Region," Journal of the Urban and Regional Systems Association, Fall 1990. 8. Gamble, T. and D. Pearson. "Growth Allocation Using the Delphi Process," proceedings of the 4th National Conference on Transportation Planning Methods Applications, September 1993. 3-7 4.0 Dynamic Assignment Highway trip assignment procedures in existing U.S. urban area travel demand models range from all-or-nothing assignment, where travel times do not vary to reflect congestion levels, to fairly detailed equilibrium assignment procedures which assume travelers wish to minimize their travel times on congested networks. While there is much variation within this range, the procedures used have many elements in common, including: - They are based on minimum impedance path-finding procedures. - They use travel time as the primary - often only - component of impedance. Cost is sometimes considered, particularly where tolls are present. - They are link-based network assignment procedures. Generally, if link travel times vary according to congestion levels, the travel time on each link is solely a function of the volume and capacity on that link. The most common volume-time, or link performance, function is the BPR equation: b T = T x (1 + a (v/c) ) O where: T = travel time TO = free flow travel time v = link volume c = link capacity a,b = parameters (often, a = 0.15, b = 4) - Each assignment uses as inputs a fixed origin-destination trip table and a highway network. - They estimate link volumes for an entire individual time period, such as the entire a.m. peak hour, p.m. peak period, or average weekday, based on the fixed O-D table for the period. While assignment procedures of this type can produce satisfactory results in well-calibrated models, there are a number of shortcomings, including: - Most volume-time functions, such as the BPR function, do not take into consideration intersection-related factors such as traffic signal timing and phasing and the presence and adequacy of turning lanes. 4-1 - Interactions between links are not considered; the travel time on one link is independent of the volumes on other links. This is an obvious oversimplification. At intersections, link travel times are affected by volumes on other approaches and opposing left turns. On freeways, merging and weaving conditions can greatly affect travel times. Queuing caused by bottlenecks on other links can also be a factor. - There is no temporal dimension to traffic assignment. Even within short time periods such as a single hour, traffic flows can vary significantly. In addition, such phenomena as queuing have a temporal dimension that cannot be modeled by such procedures. Queues build as volumes exceed the bottleneck capacity and dissipate as the demand declines. - Because the trip table is fixed, the entire table must be assigned from origin to destination, during the analysis period regardless of whether sufficient capacity exists. This leads not only to links having assigned volumes exceeding what they can carry in reality, but also a lack of understanding of how the number of vehicles on the network varies during the period. Some of these problems can be addressed by changes in the way networks are coded and assignments are performed. For example, volume-time functions can be improved to better represent the effects of congestion in urban settings. Some software packages allow nodebased capacities, delays, or performance functions. This allows for better modeling of intersection dynamics. But many of the problems described above cannot be eliminated through network solutions. Solving all of them would require relaxing the fixed trip table assumption and allowing for fink performance functions to consider what is happening on other links. A procedure that allows for these assumptions is dynamic assignment. 4.1 Description of Dynamic Assignment There are several algorithms that have been developed to perform dynamic assignment. In general, dynamic assignment has the following properties: - The analysis period is divided into several intervals, or "time slices," generally of equal length. - The trip table is divided into subsets corresponding to the time slices. The demand during each time slice varies according to observed patterns. - Trips are assigned during each time slice from their origins toward their destinations. Each trip traverses the network only as far as the vehicle could travel during the time slice, as determined through the network travel times. Trips which did not reach their destinations during the previous time slice continue from the points reached previously. 4-2 - Capacities can be treated as limits on flow rates that cannot be exceeded. Demand on a link that exceeds capacity creates a queue, which can spill backward onto upstream links. Link performance functions specify how quickly vehicles pass through the bottleneck. Intersection dynamics could also be simulated using node performance functions. - Many trips, including vehicles still waiting in queues, may not be completed by the end of the analysis period. This represents congestion that spills over from a peak period to a subsequent period. (This implies that the beginning of the analysis period should be at an uncongested time on the network, since no queues are assumed to exist.) It should be noted that, to the consultant team's knowledge, no U.S. urban area has used dynamic assignment as the specified procedure in its travel demand model. There have, however, been some applications abroad. In addition, several researchers have applied dynamic assignment using data from U.S. urban areas. The usefulness of dynamic assignment goes beyond traditional urban transportation planning needs. As IVHS and motorist information systems gain in popularity, the need for real time information on vehicle routings; and how drivers react to changing conditions is growing. This type of information would require a dynamic modeling approach. This document, however, deals only with dynamic assignment in the context of urban transportation planning models. 4.2 Available Software One commonly used transportation modeling software package that currently offers dynamic assignment procedures is TRIPS1. TRIPS performs dynamic assignment by modeling varying flow rates throughout the network for each time slice. Individual vehicles are not simulated. Input requirements, besides the total trip table and the highway network, include: - Flow profiles for each origin zone, which determine how the trip table is divided among the time slices. It is recommended that representative profiles for regions, such as CBD, inner suburbs, outer suburbs, etc., be used. - High quality intersection and queuing capacity data. - "Blocking back" curves, which must be calibrated by the user, representing the extent of queuing effects on upstream links. Logie notes that dynamic assignment using TRIPS was just beginning to be used in the United Kingdom in summer 1992. 4-3 Another feature of TRIPS that is related to dynamic assignment is called "intersection modeling2." A user can elect to define intersections to be modeled dynamically, with capacities computed for the various movements at the intersections. As in the dynamic assignment procedure, time slices and flow profiles are used to reflect changes in demand over the analysis period. At the end of each time slice, queues are calculated, and queuing delay is included in the analysis of the next slice. Information required for intersection modeling goes beyond what is needed for traditional static link-based traffic assignment. This information includes geometric information (number of lanes, etc.) for each approach, signal timing information for each intersection, and the flow profiles. TransCAD is reported to be developing dynamic assignment capabilities. The current version of TransCAD does not offer this capability3. There are also programs that offer dynamic assignment capabilities outside the traditional travel demand modeling framework. An example is INTEGRATION4, developed by IBI Group. INTEGRATION performs a dynamic traffic assignment by tracking individual vehicles through the network. It is designed as a simulation tool which incorporates such items as queuing, traffic control, ramp meters, and IVHS-type actions. INTEGRATION is not a complete travel demand modeling package although it has trip table estimation procedures as well as the dynamic assignment and simulation features. 4.3 Advantages of Dynamic Assignment One of the most common criticisms of traditional "static" assignment procedures is that they do not model congestion in a realistic manner. While queuing occurs in real life congested locations, static assignment procedures are incapable of simulating the buildup and dissipation of queues and the associated effects on travel time. As a result, capacities may be exceeded to an unrealistic extent. Dynamic assignment addresses this issue directly. Another problem associated with static assignment methods is that variations in demand within the analysis period are ignored. This is most noticeable for assignments made for a 24-hour (average weekday) analysis period, but is also a problem for shorter peak periods. In fact many areas use simple factors to obtain peak volumes from the results of assignments for longer periods. This problem is also solved using dynamic assignment. In theory, static assignments could be performed for shorter periods, equivalent to the time slices used in dynamic assignment, and aggregated to the larger periods. However, the time slices are usually no longer than 15 minutes, and a large portion of trips in most urban areas are significantly longer than that. The assignment of trips all the way from origin to destination would become problematic for such a short analysis period. 4-4 A general problem with travel demand models is "aggregation error." This most commonly applies to the lack of variation when trip generation units are aggregated to zones. But it also applies to the aggregation of trips within an analysis period and to the aggregation of all trips from an origin to a destination assigned in one assignment iteration. The use of time slices in dynamic assignment allows demand to vary, in theory, for every origin-destination pair in a different way within the analysis period. In addition, dynamic assignment can (but need not) be applied to simulate each vehicle as it traverses the network, which would represent the most disaggregate process possible at this level. 4.4 Disadvantages of Dynamic Assignment The level of detail implied by dynamic assignment is finer than that assumed in most U.S. urban transportation models. Since more detailed information about how roadways operate and interact with one another is needed, most roads other than those serving local traffic almost exclusively should be included in the highway network. While it can be argued that this should be the case even in multipath static assignment procedures, it can be impractical in large areas, both from computing and data collection standpoints. The computing problem, however, is becoming less of an issue as computers become more powerful. Even at the same level of detail, the data requirements for dynamic assignment significantly exceed those for static methods. Information is needed on departure time profiles within the analysis period for all internal and external zones - actually for each origin-destination pair - so that the trip table can be divided into time slices. While any practical application would undoubtedly rely on a few default profiles, perhaps based on area type or land use for the zone, data might be difficult to come by for even these few defaults. For internal zones, traffic counts would not be reliable data sources for departure time profiles. Counts would not take into consideration the actual demand and would be controlled by capacity considerations at congested locations. Survey data would provide more reliable information, but they would probably be useful for time slices of no shorter than 15 minutes, given the propensity of survey respondents to "round off' time responses. In addition, household survey data could yield departure profiles for trips from nonresidential generators that would be less accurate than those for households since the former would consist of a set of random departures from all generators rather than all departures from a subset of all generators. For external zones, traffic counts could provide reasonable departure profiles for base-year conditions. But future year profiles would have to be based on changes in land use outside the analysis area as well as the effects of conditions both upstream and downstream of the area cordon. Information from the internal zones would likely have to be incorporated. 4-5 A final issue concerning departure time profiles is that they are affected by traffic conditions. To address such effects, a departure time choice model would be required. This shortcoming is not characteristic only of dynamic assignment; it is true of all demand models. Another area where more data would be needed than for static assignment methods is model validation. In static assignment there is a single analysis period for which traffic count and other information is needed. In dynamic assignment, however, there are as many periods to be validated as there are time slices of the analysis period. Not only are more data needed, but the effort required to validate the assignments are multiplied several times. In dynamic assignment, more accurate information on capacities and travel times is needed than for static assignment. Even using the most advanced static assignment methods, travel time information is often considered unreliable, and time/speed outputs are often "post- processed" because of their inaccuracy. But for dynamic assignment, the need for accurate information is increased since capacities determine at what point queues form and dissipate, and errors occurring during a time slice can be compounded in subsequent intervals. In general, dynamic assignment requires the same level of information for links and intersections as do simulation programs such as TRANSYT and NETSIM. Link capacity can depend on a number of factors in urban areas, including signal timing, availability/adequacy of turning lanes, opposing traffic flows, and merging and weaving considerations. These factors are rarely considered in typical link-based urban area highway networks. The result, as discussed in the section on speed post- processing, is that speeds used by the traffic assignment model usually are not realistic and require adjustment for use in other analyses. While, in dynamic assignment, data on these factors can remain exogenous to the network itself, the implication is that detailed capacity computations for possibly thousands of links must be undertaken, and significant additional data not typically collected for urban travel models (e.g., signal timing, turning lanes) would have to be obtained. The practical alternative of using generalized capacities based on link type, similar to static assignment models, would result in much less accuracy than the theory of dynamic assignment implies. One final point to be made concerns the concept of "disaggregation error." One of the main criticisms of traditional urban transportation planning models is that they rely on information that is too aggregate (i.e., at the zone level and for single analysis periods) and that variations within zones or time periods tend to be lost. In this sense, dynamic assignment is desirable since it is more disaggregate. However, as more disaggregate information is required, random variations can become more pronounced, skewing the results of models. For example, if it is assumed that a traffic volume for an hour-long analysis period has an error (or even an observed fluctuation) of ten percent, the error for each of four 15-minute time slices within that hour can be much larger. Given that dynamic assignment would compound errors from the first time slice, the error in the final slice could be substantial. The issue of disaggregation error must be addressed in general terms as travel demand modeling becomes a more disaggregate process. 4-6 4.5 Summary Dynamic assignment overcomes many of the problems characteristic of static assignment procedures, such as variations in demand during the analysis period and accounting for queuing. While dynamic assignment has not been used as the basis for any U.S. urban area model, it has been applied abroad and to U.S. urban areas in research projects. The software to apply it within a traditional transportation planning modeling framework exists. However, there are several drawbacks to using dynamic assignment, including increased data and resource requirements, the need for more accurate capacity computations, and disaggregation error. 4.6 References 1. Logie, M. "Assignment Modeling with Dynamic Traffic Effects." Proceedings of the Fourth International Conference, Microcomputers in Transportation. Published by the American Society of Civil Engineers, 1992. 2. MVA Systematica. Trips Documentation, Woking, Surrey, England, October 1990. 3. Caliper Corporation. TransCAD Reference Manual Version 2.0 1990. 4. IBI Group. "INTEGRATION" informational brochure, 1993. 4-7 5.0 Air Quality Analysis Methods Since the passage of the Clean Air Act Amendments of 1990, much higher levels of importance have been placed on the interface between travel models and air quality models which predict vehicular emissions and pollution dispersion. To support the needs of air quality analysis, transportation planners have been faced with the challenge of estimating the pollutant emissions associated with the vehicular travel estimated using the highway assignment process. To meet the preferred level of detail of the emissions and dispersion models, these estimates require the breakdown of vehicular travel forecasts by facility, by vehicle type (for example, for light duty gasoline and diesel passenger cars; light, medium and heavy duty gasoline and diesel trucks; motorcycles; and other vehicles), by hour of the day, and by vehicle operating mode (for vehicles in cold start and normal running modes). Furthermore, accurate forecasts of speeds by hour of the day and by vehicle type are required due to the wide variation of emissions levels as vehicle speeds change. Travel planners have enhanced their forecasting procedures in a number of ways to bridge the gap between the traditional outputs - daily and, often, peak period vehicle volumes and those desired for air quality analysis. Generally, these enhancements fall into three groups: - Revisions of the traffic assignment process to provide link volumes by vehicle operating mode; - Refinements of the volume-speed functions used in assignment procedures to provide improved estimates of travel speeds by highway facility; and - Assignment post-processors which provide either more accurate speed estimates, or breakdowns of link travel by vehicle type and hour, or both. These groups of enhancements are discussed in Sections 5.1-5.3 of this chapter. The remaining sections discuss the resources necessary to apply each type of air quality analysis enhancement (5.4) and the drawbacks of their use (5.5). Finally, Section 5.6 provides a summary of the material contained in this chapter. 5.1 Prediction of Trips by Vehicle Operating Mode When vehicles begin operation after.having completely cooled down following their former use, they are said to be in the "cold start" running mode. EPA defines the length of time required to progress from cold start to normal running mode as 505 seconds or 8.4 minutes. 5-1 During this time, vehicular emissions rates begin at a much higher than normal level, followed by a rapid decline to the rate typical for normal running conditions. If vehicular emissions are to be estimated accurately by location, the estimation process used must provide some means of predicting variations in emissions rates by facility which reflect the varying fractions of cold start vehicles using the facility. The simplest enhancement which accomplishes this objective is to use EPA's MOBILE5 program to determine emissions rates which, for a given speed, vary by link type. These variations exist because the fractions of cold start vehicles are much higher than average on local streets and much lower than average on freeways. These facility type- specific variations can be obtained from MOBILE5 by specifying different cold start fractions depending on facility type. The appropriate fractions to use can be based on roadside surveys or on the results of the more detailed approaches discussed below. Typical values for these fractions are provided in References 1, 2 and 3. A second approximation which accounts for cold start emissions somewhat more accurately than the use of overall average emission rates involves allocating the difference between emissions due to cold starts and those due to normal running entirely to the zone centroid links in the highway network. This strategy tends to focus the added emissions too heavily near the origins of trips but does remove cold start emissions increments from freeways and major arterials where they would otherwise be overestimated. The most accurate means of dealing with cold start emissions involves modifications to standard highway assignment algorithms to allow the separation of link volumes into cold start and normal running components as the assignment process is carried out. Basically, this involves simultaneously assigning two trip tables, one of all trips which begin in normal running mode, and the other of all cold start trips. The first trip table is assigned normally, with all link volumes accumulated as normal running volumes. The second trip table is assigned with special features applied so that the first 505 seconds of the trips between each zone pair are allocated to the appropriate links as cold start volumes. These trips are then assigned to the remainder of their paths with volumes allocated as normal running volumes. At the end of the assignment process, each link has two assigned volumes, one for cold start vehicles and one for normal running vehicles. By applying the appropriate emissions rates to these two volume totals, the effects of link-specific cold start travel characteristics on the total emissions on each link can be estimated very accurately. A number of transportation planning packages, including MINUTP, TRANPLAN and EMME/2, now incorporate the enhancements discussed above. The additional trip making information required to use these procedures for emissions predictions consists of the fractions of total vehicular trips which begin in the cold start mode. These fractions are normally specified by trip purpose, ideally based on a travel survey which obtains information on which vehicles are used (if any) for all household trips. 5-2 5.2 Improved Speed Models Highway traffic assignment procedures provide two major outputs for each link in the highway network: predicted volumes and speeds. Traditionally, the major purpose of the assignment process for transportation planners has been to obtain volumes for each highway facility; the corresponding speeds tend to be treated only as intermediate variables required to obtain realistic volumes. More recently, these speeds have attained a high level of importance because they are needed to estimate vehicle emissions. Also, the importance of using consistent travel speeds in all forecasting steps, achieved by 'feeding back' speeds after traffic assignment to subsequent iterations of trip distribution and mode choice, is now more widely recognized. Faced with these new levels of importance for travel speeds, planners have often assessed the speeds predicted in the highway assignment process and found them to have unacceptable error ranges. Typically, planners have adopted one of two strategies to improve highway speed estimates: - Begin by carrying out a thorough improvement of the link speed prediction process incorporated in the traffic assignment procedure, including better estimates of capacities, free-flow speeds, and speed-volume functions. Follow this with a recalibration of the base-year assignment involving changes as required to trip tables, network geometry, and link characteristics, but avoiding arbitrary changes affecting predicted speeds which can only be justified in terms of improvements in link volumes. - Alternatively, retain the previously calibrated (with respect to volumes only) highway assignment process, supplemented by a post- processing capability which provides improved link speed estimates. Significant progress has been made on both of these enhancement strategies in recent years. This section deals with the work done to improve the volume-speed functions used in the highway assignment process; the next deals with speed post-processors. Since both require good volume-speed functions, the advances discussed in this section also are important components of many of the post-processors documented in the next section. The standard traffic assignment process places significant limitations on the level of enhancement that is possible in volume-speed functions to be used in these processes. These limitations include: - Because trips are assigned over distance but not over time, the dynamics of traffic queues - their build-up and dissipation over the peak period - cannot be reflected in the required functions. - Assignment procedures are set up to use only the volume and characteristics of a given link to estimate its speed or travel time. Because volumes on other links are not available, the resulting procedures can be only general approximations of intersection and weaving section delays, and of back-ups caused by downstream congestion. 5-3 - The number of link descriptors is typically limited to no more than four to eight. This number typically places a significant restriction on the ability to represent important variables such as trucks as a percentage of total travel, grade, green time at inter- sections, parking, and pedestrian cross-traffic. Within these limitations, however, many agencies have found improved parameters or functions for use in volume-speed relationships. In most cases, the parameters of the "standard" BPR function are changed while retaining its basic functional form. Generally, travel forecasters have found that larger exponents of the volume to capacity ratio (V/C) than the standard value of 4 provide more accurate speeds. A recent report prepared by the Metropolitan Transportation Commission in the San Francisco Bay area reviews a number of these studies, as well as the data in the most recent chapters of the Highway Capacity Manual (HCM)4, and reports on tests which show that the best function for its regional network is one with an exponent of V/C of 105. Guidance for changing the functional form of volume-speed functions and for reflecting the data contained in the HCM is provided in a recent EPA document6. 5.3 Assignment Post-Processors As discussed in the previous section, an alternate to changing the volume-speed functions used in traffic assignment is to develop an assignment post-processor to refine the speed predicted during assignment. This strategy can provide greater accuracy, typically using additional link variables not available within the assignment program, combined with more complex relationships between volumes and speeds. Another advantage of this approach is that it avoids the necessity of recalibrating an existing highway assignment process. In addition, the post-processor can be expanded to convert assignment outputs to the expanded level of detail desired for emissions estimation. This section discusses recent work done to develop and apply both speed and emissions post-processors; procedures providing enhanced speed estimates are reviewed first, followed by a discussion of procedures which include both speed and emissions estimation features. Speed Post-Processors A number of speed post-processors have been documented in recent TRB papers, in the NARC Manual of Regional Transportation Modeling Practice for Air Quality Analysis7, and in EPA documents such as Highway Vehicle Speed Estimation Procedures6. In general, these sources describe procedures which include improved speed-volume functions (compared with the standard BPR function); considerations of a wider range of link characteristics than simply link types, area types, and numbers of lanes; and considerations of the effects of traffic conditions such as queues on adjacent links. The available postprocessing methods exist in various forms, ranging from simply conceptual and/or mathematical models to computerized capabilities designed for use in conjunction with 5-4 particular transportation software. Some of the specific post- processors now available are the following: - The Highway Performance System Analytic Process (HPMS-AP), which estimates link speeds as functions of many more link characteristics than is typical in assignment programs. These additional characteristics can include, if available, pavement condition, curves and gradients, speed change cycles and their minimum speeds, stop cycles, acceleration and deceleration rates, and the fraction of time spent idling. - The Dowling and Skabardonis post-processor8 combines considerations of speed changes due to congestion with delays due to queuing to provide total link travel times and average link speeds. For the congestion component, modifications of the BPR function are used. Queue delays are predicted for all links on which capacity flows occur; back-ups onto upstream links are approximated as occurring on the capacitated links. In recent work, a later version of this model has been implemented as part of the DTIM2 package used in California for emissions estimation9. - For the Central Artery planning effort in Boston, the management consultant/travel forecasting subconsultant team developed speed post-processing programs which use TRANPLAN assignment outputs plus additional data such as a revised capacity value, if necessary, observed volumes and speeds for existing links, and the facility type. The possible facility types, each of which has a unique speed estimation relationship, are links with travel time constrained by signalization, by geometries, freeway and ramp links with low and high-volume/capacity ratios, links where the speed is unconstrained, and links which experience queues due to capacitated flows, either downstream or locally. For the first two facility, types (constrained by signalization -and geometries), the Highway Capacity Manual4 relationships for signalized intersections are used with varying parameters depending on the constraining factor. For low-volume freeways and ramps, the HCM relationships for freely flowing highways are used. For high-volume freeways and ramps and for unconstrained facilities, a modification of the BPR function is used. Links subject to a single queue combined into 'facilities' composed of 'sections.' These facilities are then analyzed as a unit to determine queues by section and hour, queue lengths, delays and speeds. When the post-processor is applied to existing facilities, the analyst has the option of calibrating the speed estimation function used to match the observed speed at the observed volume level. - The EPA-distributed document referenced above6 discusses how both traffic assignment and post-processing speed estimation methods can be made consistent with observed link delay information, obtained either locally or from sources such as the HCM. The basic approach makes maximum use of the facility-specific information normally obtained by MPOs along with the HCM relationships. Since the speed estimation methods in the HCM are typically based on a richer set of link characteristics than is generally available to MPOs, the methods presented focus on procedures which can be used to generalize these relationships using average or typical values for the additional variables. Extensions of the basic methods are also presented to demonstrate the value of committing additional analysis time and effort to provide improved travel speed estimates. 5-5 To summarize, a number of speed post-processors have been developed, mainly in response to the passage of the 1990 Clean Air Act. As planners in a number of urban areas have seen the need to bridge the gap between transportation assignment programs and the needs of emissions estimating procedures, they have developed these post- processors, largely on an ad hoc basis. As a result, the existing post-processors lack uniformity and, often, a high degree of integration with the traffic assignment programs on which they depend for a significant portion of their input data. Combined Speed and Emissions Post-Processors In addition to improved speed estimates, the prediction of vehicular emissions requires a number of refinements of the outputs available from traffic assignments. Typically, only total vehicle trips by link are available, or possibly broken down into auto, truck and bus estimates. These trips may be available only as daily totals, or for a number of periods such as a.m. and p.m. peak and off-peak. Since emissions vary significantly not only by vehicle speed but also by vehicle type, it is desirable to disaggregate these assignment outputs as follows: - Vehicles by hour of the day and by type, where the types of interest may be light duty gasoline and diesel passenger cars; light, medium and heavy duty gasoline and diesel trucks; motorcycles; and other vehicles; and - Speed by hour of the day and by vehicle type. Emissions post-processors generally apply factors which vary by link type and area type to divide the input vehicular volumes into estimates of vehicles by hour and by type. These factors are obtained from vehicle classification counts which provide observations by hour and by vehicle type of the vehicles using typical links within each of the relevant link type and area type categories. After these factors are applied, speed estimation procedures such as those discussed in this section are used to provide speeds by hour and by vehicle type. Finally, the products of emissions rates for each combination of vehicle type and hour of the day and vehicular volumes are summed to provide total emissions by link. These link-based emissions can then be added to trip end emissions (cold start and evaporative) to provide vehicular emissions totals for the study area. In general, the speed estimation procedures discussed in this section are sufficiently flexible to fit directly into emissions post- processors, although additional effort may be required to combine the full set of functions into an integrated process. Typically, the speed estimation procedures can be applied to any subset of vehicles representing any time period during which travel conditions remain stable; thus they may be applied either to total vehicle flows in a multi-hour peak or off-peak period, or to vehicle flows of a specific type during a specified hour of the day. 5-6 Examples of the combined speed and emissions post-processors developed recently are the following procedures which have been developed for use either with specific transportation analysis packages (i.e., TRANPLAN, MINUTP, EMME/2, SYSTEM II), or with generically-defined traffic assignment Outputs: - In work done for the California Air Resources Board, DHS has combined revised volume-speed functions, a queuing analysis model, and vehicle activity data (vehicle miles of travel by speed and acceleration/deceleration classes obtained in vehicle simulation runs made in INTRAS and TRAF-NETSIM) by link type to provide more detailed inputs to emissions modeling components10. - JHK has developed a post-processor which estimates peak spreading due to congestion on a link-by-link basis, applies refined volume- speed functions, and interfaces with the simulation package FREQ to conduct queuing analyses on congested freeway facilities11. - In another Caltrans project, Dowling Associates is updating the Direct Travel Impact Model (DTIM) which is used for emissions estimation to include updated volume-speed functions, queuing analysis, and recent distributions of travel by time of day for a range of link types9. 5.4 Resources Necessary for Air Quality Analyses Nearly matching the range of capabilities in the available procedures is their range of data requirements. At the low end, some procedures simply involve improved speed-volume functions with no change of independent variables. In this case, no additional inputs, beyond those used in the traffic assignment process, are required. More typically, additional information is required in order to estimate hourly volumes by vehicle class and travel times which reflect link characteristics - beyond free-flow travel speeds, capacity, fink type, and area type. This additional information can vary from regionwide averages to link-specific observations. The additional variables may include the entire range of link characteristics used to determine highway capacities: examples are cycle times and green times at signalized intersections, capacities within particular segments of weaving sections on freeways, bus and truck flows as fractions of total volumes, grades, and parking availability. Also included may be vehicle type distributions and volume distributions over hours of the day. In some cases, MPOs currently maintain highway facility files as parts of their management systems which include many of these types of information on a link by link basis. When this is the case, much of the information required by post-processors, beyond that used in the assignment process, may be obtained from these systems. Some methods also require the identification of upstream links, so queuing which extends over multiple links can be considered. Finally, all of the methods can be adapted to reflect locally collected data such as vehicle type distributions, capacities and speed-volume relationships, if these data are found to differ from the national averages. 5-7 Due to the wide range of data requirements, there is also a large variation in resource needs for analysis. For MPOs which can implement these systems without collecting new link-specific data, existing (generally as implemented in another urban area) air quality analysis components can be added to their forecasting process in a matter of weeks. This level of effort may be sufficient for regional planning purposes in many urban areas. For more detailed subarea or project-level studies, additional effort will probably be required to collect link-specific data and to deal on a case-by-case basis with critical links such as freeway weaving sections and approaches to complex signalized intersections. In these cases, the resource requirements related to the air quality analysis function can become as high as 10 to 15 percent of the total analysis process. Higher resource levels will be required if an extensive program of data collection on vehicle and link characteristics and speed-volume relationships is put in place in support of a more detailed emissions estimation procedure. 5.5 Drawbacks of Air Quality Analysis Procedures Given the need for accurate data to estimate vehicular emissions and to provide the basis for feedbacks in the trip distribution/mode choice/traffic assignment process, most MpOs require improvements in their speed and emissions estimation processes. Thus, the drawbacks of additional development and implementation work, and possibly additional data collection effort, are generally inescapable. The major drawback of attaining this required improvement using only post- processors rather than by building improved speed and operating mode procedures directly into traffic assignment is that discrepancies will continue to exist between the final estimates of highway speeds and the values used to predict route choices. However, the impedance used for route choice may be considered as a generalized cost which is related, but not exactly equal, to link travel times. In any case, some type of justification is required when only post-processors are used. The second drawback of the air quality analysis procedures developed to date is their lack of consistency and general availability. Most have not been fully integrated into. the commercially available transportation planning packages and thus are likely to be obtained as 'shareware' provided by other MPOs or in connection with a consultant contract. The development of standard methods for air quality analysis would address the consistency issue. The implementation of this method as a program in the public domain would address the availability issue, as would the incorporation of this method into each of the generally available packages. California, in its current support of DTIM updating, appears to be in the lead in providing a standardized, generally available, process. 5-8 5.6 Summary Air quality analysis procedures which include methods to determine VMT by operating mode, accurate link travel speed, and vehicle emissions, are coming into usage by a significant number of MPOs. These procedures help to fill the gap between traffic assignment results and air quality forecasts, and can also be used as part of a feedback process to ensure that the final estimated speeds match those used in trip distribution and mode choice models. The data and analysis resource requirements of these procedures can vary widely. Drawbacks of using these procedures, in addition to their resource requirements, are problems of consistency - both within the forecasting process in the case of speed post-processing and generally from one processor to another - and availability. 5.7 References 1. Midurski, T. and A. Castaline. "Determination of Percentages of Vehicles Operating in the Cold Start Mode," prepared by GCA Corporation, Bedford, MA, for U.S. EPA, Research Triangle Park, N.C., report EPA-450/3-77-023, August 1977. 2. Brodtman, K. and T. Fuca. "Determination of Hot and Cold Start Percentages in New Jersey," prepared by NJ DOT, report 84-001-7792, July 1984. 3. Benson, P. "Corrections to Transient Vehicle Fraction for Microscale Modeling," presented at 66th Annual Meeting, TRB, January 1987. 4. Transportation Research Board. Highway Capacity Manual, 1985. 5. Singh, R. "Updating Speed-Flow and Speed-Capacity Relationships in Traffic Assignment," MTC, Oakland, California, April 1994. 6. Ruiter, E. "Highway Vehicle Speed Estimation Procedures." Prepared for the Environmental Protection Agency, 1991. 7. Harvey, G. and E. Deakin. A Manual of Regional Transportation Modeling Practice for Air Quality Analysis. Prepared for the National Association of Regional Councils, July 1993. 8. Dowling, R. and A. Skabardonis. "Improving the Average Travel Speeds Estimated by Planning Models." Presented at the 71st Annual Meeting of the Transportation Research Board, January 1992. 5-9 9. Dowling, R. "Update of the Direct Travel Impact Model (DTIM)," Technical Memorandum 8-1, prepared for California Department of Transportation, June 14, 1994. 10. Skabardonis, A. "Feasibility and Demonstration of Network Simulation Techniques for Estimation of Emissions in a Large Urban Area," Draft Final Report prepared for California Air Resources Board by DHS Inc, Berkeley, California, January 1994. 11. JHK and Associates. 'Travel Demand and Simulation Modeling Contract - Draft Final Report," prepared for California Department of Transportation, Emeryville, California, January 1994. 5-10 6.0 Modeling Trip Chaining Behavior "Trip chaining" is one of the travel behaviors in which conventional travel models fail to reflect the actual travel behavior of residents in metropolitan areas. People often combine travel to several destinations into a single trip circuit. Rather than making separate trips, often with different purposes (e.g., home-to-work-to-home then home-to-shopping-to-home), people often "chain" trips (e.g., home-to- work-to-shopping-to-home). Since the inception of disaggregate travel demand modeling, analysts have recognized this behavior and studied it, but to date little has been done to include these complex activities in conventional metropolitan travel models. The "standard" modeling procedure is to analyze home-based work, home-based non-work, and nonhome-based trips as separate independent trips. In reality, the trip generation, destination choice, and mode choice of each trip in a chain is related to the other trips in the chain. For the most part, statistical analyses of trip chains has been restricted to academic research. However, a number of factors make it desirable to assess the value of including trip chaining behavior in urban travel demand models. First, a number of prominent modelers (including academics, consultants, and urban area planners) seem to have concluded that activity-based travel models represent the "next generation" of modeling. The decisions leading to the formation of trip chains are essential elements of activity-based models. The development of procedures to include trip chaining analyses into existing modeling systems is likely to be valuable in the eventual of new activity-based models. Second, many urban area planning agencies have recently completed household travel surveys, with trip diaries, so a great deal of fairly recent trip chain data are available. In addition, the inclusion of trip chaining analyses in urban area models may help improve the overall quality of non-work models. Non- work travel accounts for the largest share of urban area travel. A recent study estimated that in large urban areas over one-half of the person trips in the a.m. peak and over two-thirds of the person trips in the p.m. peak are made for non-work purposes1. But despite their importance in urban areas, non-work trips (and nonhome-based trips, in particular) have proven to be quite difficult to model. Generally speaking, non-work travel models perform far worse than the work models that have been developed for urban areas. 6-1 6.1 Recent Trip Chain Modeling Work The trip chain modeling work of which we are aware falls into three categories: - Descriptive research; - Development of trip generation models that incorporate trip chains; and - Development of nonhome-based destination choice models which incorporate trip chaining behavior. Descriptive Research A number of researchers and planners have studied the basic characteristics of people's trip chains, including: - The distribution of the number of links in trip chains; - The combinations of trip purposes which are being chained; - The household characteristics which affect the types of trip chains that are being formed; - The travel modes used for various types of trip chains; and - The time-of-day characteristics of trip chains. Strathman, Dueker, and Davis developed two descriptive trip chaining models based on data for the Portland, Oregon metropolitan area2. The models relate the probability of trip chaining to household and trip characteristics. For their first model, the authors developed a binary logit model to predict the probability that non-work trips will chained with a given work trip. The two choices in the logit model are a simple commuting trip versus a complex trip chain. The utility functions for the two choices include: - Traveler's characteristics (specifically gender); - Household characteristics (total number of non-work trips made by a household, household income, household structure); and - Home-to-work travel characteristics (travel mode, time of travel, traffic congestion measures, residential location and place of employment variables, distance from home-to-work). The second model developed by Strathman, Dueker, and Davis is a simultaneous equations model which seeks to explain more generally the allocation of a household's non-work 6-2 travel into trips of three types: simple unlinked non-work trips, multi-stop non-work journeys, and non-work trips chained to commuting trips. The independent variables in this model are similar to those in the binary logit model. Research studies, like that of Strathman, Dueker, and Davis, are valuable because they shed light on how trip chains are formed and why. However, the models that have come out of this work are not easily applicable to standard forecasting models without the development of a feedback mechanism. The models use information on survey respondents' trip distribution and mode choice to explain their level of trip generation. This conflicts with the standard sequencing of the models in the four-step process. Trip Generation Models That Incorporate Trip Chaining A great deal of trip-chaining research has been conducted by researchers who are studying activity-based modeling. These researchers are attempting to develop models which make use of the fact that transportation is a derived demand. They believe that models of household activities, rather than simple travel models, will ultimately prove to describe travel behavior more accurately. Some of the most important activity-based modeling is underway at the University of California at Davis, where Ryuichi Kitamura and others are working to develop practical activity-based models. Among their activities is the development of trip generation procedures that allow for the analysis of trip chaining3,4,5. The trip generation approach proposed by these researchers is a recursive set of regression equations. First, trip generation relationships are developed for the "mandatory" trip purposes - work trips and school trips. These equations are standard regression-type trip generation equations using household and zone characteristics as explanatory variables. Then, trip generation equations are developed for "discretionary" trip purposes (social trips, shopping trips, personal business trips, and serving passenger trips). These equations use as explanatory variables the household and zone characteristics, as well as the number of household mandatory trips. Next, the number of trip chains for a household is modeled as a function of the number of predicted trips by each purpose. The expected number of home-based and nonhome-based trips made by a household can be calculated from the estimated numbers of trip chains and the total numbers of trips. The UC-Davis trip generation models, as described in the referenced papers, are still not totally applicable to the standard four-step travel modeling process yet because they do not address the question of the sequence of the trip chains. However, with relatively little work, models like these could be estimated and applied as part of a four-step modeling system. Trip Chaining and Destination Choice Cambridge Systematics' 1980 MTC modeling system addresses the trip chaining phenomenon in its destination choice model6. CS' nonhome- based models estimate the probability that a traveler at a trip end other than his or home will make a trip to a destination 6-3 other than his or her home, then, if the traveler is not returning to home, they estimate the probability that the trip will be made to each particular zone. For each traveler, the sequence of models is repeated until the traveler reaches home. The MTC nonhome-based models assume that each traveler's destination choice decision is independent of previous decisions. This assumption obviates the need to represent several alternative trip chains as explicit choice alternatives. In addition, the non-home-based models assume that no mode switching occurs (i.e., the mode used for the non- homebased trip is the same as the mode used in the home-based trip to get to or from the trip end location). The second assumption makes it impossible to develop a nonhome-based mode choice model, however the assumption probably describes people's behavior more accurately than nonhome-based models estimated independently from the home-based models. 6.2 Data Resources Needed for the Incorporation of Trip Chaining into the Four-Step Modeling Process As we have mentioned, many local planning agencies have assembled the revealed preference data necessary for developing trip generation, destination choice, and/or mode choice models which account for trip chaining. Home interview travel diary surveys provide records of how travelers link trips into trip chains. These survey data are probably adequate for the agencies to consider the effects of trip chaining in a limited way, such as developing improved trip generation relationships as proposed by Kitamura et al., or developing nonhome- based destination choice models similar to the CS MTC models. To develop models that analyze the full effects of trip chaining on travel decisions, a planning agency would probably need an enhanced travel diary survey with increased detail consistent with the development of activity-based models. 6.3 References 1. Gordon, P., A. Kumar and H. Richardson. "Beyond the Journey to Work," Transportation Research - Part A 22A (1988): pp. 419-426. 2. Strathman, J., K. Dueker, and J. Davis. "Effects of Travel Conditions and Household Structure on Trip Chaining," Transportation Research Record (forthcoming). 3. Nishii, K., K. Kondo, and R. Kitamura. "Empirical Analysis of Trip Chaining Behavior," Transportation Research Record 1203 (1989): pp. 48-59. 6-4 4. Goulias, K. and R. Kitamura. "Recursive Model System for Trip Generation and Trip Chaining," Transportation Research Record 1236 (1990): pp. 59-66. 5. Goulias, K., R. Pendyala, and R. Kitamura. "Practical Method for the Estimation of Trip Generation and Trip Chaining," Transportation Research Record 1285 (1990): pp. 47-56. 6. Cambridge Systematics, Inc. Travel Model Development Project: Phase II Final Report, two volumes (1980), Volume 2: Detailed Model Descriptions. 6-5 7.0 Mode Choice Modeling Improvements There is a wide variety of mode choice modeling procedures used in U.S. urban areas. They range from no procedure (modeling only highway vehicle trips) to detailed nested logit modeling procedures. Because of the additional requirements now being imposed on MPOs by the Clean Air Act Amendments and ISTEA, the simpler procedures are being questioned as to their adequacy to meet the new requirements. Rather than attempting to identify all of the shortcomings with current mode choice modeling procedures, this document focuses on five advanced procedures (relative to the state of the practice): - Incremental logit models; - Modeling of high occupancy vehicle (HOV) trips; - Dealing with transit captivity; - Dealing with transit transfers in mode choice models; and - Integrating mode choice models with trip distribution, trip generation and land use models. In addition, discussions of model transferability from one urban area to another, the use of Monte Carlo simulation and modeling the choice between tolled and non-tolled travel paths are included. 7.1 Incremental Logit Modeling The logit formulation is the most popular mode choice model formulation in the U.S. This is due to its computational ease, the availability of software to estimate such models, and the large number of areas which have models that could be transferred to other areas. In a logit model, the probabilities of the various modes being chosen are based on the modes' relative utilities, which are functions of the various service characteristics of the modes (e.g., travel times, costs), characteristics of the travelers (e.g., auto availability), and perhaps other variables1. 7-1 Another advantage of the logit formulation is that it can be applied incrementally. In this case, the changes in mode shares are computed using the base mode shares and the changes in the variables in the utility functions. This can be much more efficient and less data intensive than applying the full logit model because: - Data are needed only for those variables that differ between the base and alternative scenarios. For example, if the alternative is simply to change transit fares, the only variable needed is the change in the fare; the only model coefficient needed is that for the transit fare variable. - Information on the "unobserved attributes" in each mode's utility function, (i.e., the constant term), is not needed. - It is easier to transfer a model from another area since only those coefficients representing sensitivity of mode choice to specific variables are needed to apply an incremental model. - It is possible to perform mode choice modeling without expensive survey efforts. Base mode shares can be determined through use of data on existing travel. This information might include U.S. Census Journey to Work data, transit ridership counts and surveys, and trip tables estimated from observed counts. The incremental model can be applied in a multinomial or nested logit format. There are also disadvantages associated with incremental mode choice modeling including the following: - It is difficult to determine demand for non-existing modes, which have no base mode share. A logit model cannot yield a share (choice probability) of zero for an available mode, and an incremental model will have no basis for determining a change in mode share without a positive base mode share. - The incremental model cannot be calibrated to reflect existing conditions unless it is estimated as a full logit model, which of course, would negate much of the efficiency advantage (in model estimation) over full models as described above. Since base conditions are pre-specified as model inputs, and coefficients unnecessary for model application (including all unobserved attributes) are not determined, there is no way to tell if the (partially specified) model matches observed behavior. The model can be validated to a certain extent, by determining whether the complete version of the incremental model could produce the observed base data and by performing sensitivity tests. A good example of incremental logit mode choice modeling is the set of models recently developed for Seattle Metro. Nested logit models were developed for three purposes as follows: 7-2 - Home-based work: - First level: highway vs. transit - Second level (highway): drive alone vs. shared ride - Second level (transit): walk vs. drive access - Third level (shared ride): 2, 3, or 4+ occupancy - Home-based non-work: - First level: highway vs. transit - Second level (transit): walk vs. drive access - Non-home-based: - Same as home-based non-work Base-year modal trip tables were obtained from two main sources: the 1980 U.S. Census Journey to Work data and a 1985 on-board transit survey. Model coefficients were obtained from a review of 13 models from other U.S. urban areas. 7.2 HOV Modeling There are two main issues involved with modeling HOV modes: - How to incorporate modes for autos with different occupancy levels into mode choice models; and - How to model the effects of carpooling incentives, including exclusive lanes, preferential parking, and other policies. The easy answer to the first question is to use a nested logit model, with the various occupancy levels corresponding to modes in the model. In general, carpool occupancy levels of 2, 3, and 4 or more have been considered sufficient. Often, especially in smaller areas, two levels (2 and 3+) or even a single level of carpool occupancies have been deemed sufficient. The general rule is to separate occupancy levels which are, or are likely to be, treated differently in terms of qualifying to use HOV lanes or to obtain preferential treatment, such as in parking. The method of nesting for HOV modes varies in different areas. Usually, auto modes are separated from others at one level of nesting, and the auto mode is divided into the occupancy levels. Or, as is done in the Seattle home-based work model described above, the auto mode may first be divided into drive alone and shared ride modes, and the shared ride mode then divided into occupancy levels at the next nesting level. Some models separate the auto mode into auto driver and auto passenger modes. This has the advantage of better taking into account factors affecting whether or not a traveler drives, such as possession of a driver's license and auto availability. However, each of the 7-3 two submodes (driver and passenger) would have to be further subdivided into occupancy levels to estimate the effects of carpool incentives, which creates more modes than would be present in a model without the distinction between drivers and passengers. It is important to recognize that travel surveys need to ask the number of occupants in the car for each auto trip, whether the trip maker is a driver or passenger. Without such information, it is impossible to determine whether the trip qualified for carpool incentives. As far as modeling incentives for carpooling is concerned, the most detailed study in the U.S. was conducted by Comsis Corporation for the Shirley Highway in the Washington, D.C. area , and later revised and transferred to Houston4. The Shirley Highway had had HOV lanes in place for many years when the models were first developed. These lanes carry a substantial number of carpools and transit riders who enjoy a significant travel time savings over travelers in the general purpose lanes. The Shirley and Houston models are nested logit models for the home- based work purpose. The Houston model has the same structure as the Seattle home-based work model described above. The Shirley model differs only in that it does not estimate transit access mode choice. The original Shirley model included several variables that reflect the decision of whether or not to use an HOV mode. Besides the traditional service characteristics variables representing time and cost and some socioeconomic variables, these included: - Whether or not preferred parking was offered to carpools (a binary dummy variable); - Whether the traveler was a government employee (a binary dummy variable); - Whether the traveler worked for a large employer with over 500 employees at the traveler's work site (a binary dummy variable); - Whether the traveler's employer offered flexible working hours (a binary dummy variable); - Whether the trip could feasibly use the HOV path, i.e., if the path using the HOV lanes was faster than the non-HOV path (a binary dummy variable); - The distance for the HOV path, in miles; and - The time savings using the HOV lanes, in minutes. Note that the last variable implies that travelers perceive the HOV time savings differently than the overall travel time for the HOV mode alone. This implies that this variable should be considered in areas where HOV lanes offer significant travel time savings. Some of the variables noted above (e.g., government worker) might not be applicable in other areas. In addition, coefficients might differ significantly in other areas, where conditions such as parking are unlike Washington. So caution should be used when 7-4 attempting to transfer this model to other areas. In fact, not only did Comsis recommend different coefficients for the Houston model, but redefined a number of variables. In addition to including important new variables which are specific to the HOV mode alternatives as exemplified in the Shirley model, mode choice models with HOV alternatives must include the traditional time and cost variables. Their estimation and application requires the input of estimated time and cost variables which accurately reflect the differences associated with different vehicle occupancy levels. These differences include, for example, the time savings from higher HOV lane speeds, additional time required for HOV lane access and egress, and time required for collecting and distributing passengers at trip ends. This requires network representations which include links for these distinct impedances, and network assignment methods which yield separate assignment times for SOV and HOV trips. Boston's Central Transportation Planning Staff (CTPS) provides an example of methods used and problems encountered in modeling HOV, especially related to the generation of the time and cost variables required by the model. CTPS was able to generate separate HOV and SOV link speeds by assigning SOV and HOV trips separately in a sequential process, but their assignment software was not flexible enough to generate separate OD trip speeds for HOV and SOV trips. Therefore, in order to supply OD travel times to the mode choice model they used a linear regression of HOV in-vehicle travel times on SOV in-vehicle travel times from a travel survey. They used the resulting equation to transform the network generated OD travel times, which were assumed to be SOV times, into HOV OD travel times for use in the mode choice model5. 7.3 Transit Captivity It is clear that many travelers in urban areas are "transit captives," (i.e., that the auto mode, at least the auto driver mode, is unavailable). Transit captives likely make up a substantial proportion of transit riders. This implies that simply estimating and applying mode choice models over the entire population will overestimate auto use by transit captives and overestimate transit use among the rest of the population. Most models ignore this issue. A few models, however, have attempted to address it. In Portland6, the Metropolitan Service District models mode choice separately for households which do not own autos and those that do. For the latter, variables indicating whether the household has fewer, the same number as, or more autos than workers are included in the mode choice model. These variables are proxies for auto availability, which could have been obtained directly from household survey data to use in model estimation but is nearly impossible to forecast. In theory, it would be possible to estimate separate mode choice models for each level of auto ownership. This might prove impractical for most areas, not only because of the substantial amount of time required to estimate three or four times as many mode choice 7-5 models, but because of the lack of sufficient data for each model estimation as the amount of stratification increases. Transit captivity and stratification of models are issues that go beyond auto ownership and mode choice. For example, trip distribution models are generally based on highway travel times. This ignores the fact that transit captives do not base their destination choices on highway time, but on transit travel time. While the vast majority of trips in most regions are by auto, there are many zones within a region where transit accessibility is much more important. In Atlanta, the Atlanta Regional Commission (ARC), and the Metropolitan Atlanta Regional Transit Authority (MARTA) have dealt with this problem by using an accessibility measure that considered both auto and transit attributes, including both time and cost7. The measure was the "logsum" variable from the mode choice model, which is the denominator of the logit formula. In areas where transit is not prevalent, the highway attributes dominate; for transit-dependent zones, transit attributes are a significant component of accessibility. The Minneapolis/St. Paul model system also uses the logsum variable in the trip distribution model. In addition, it is also stratified by car ownership for generation, distribution and mode choice, capturing the effects of varying auto availability in all three models. Another innovation used in Atlanta's travel model was the stratification of the trip distribution (gravity) model by income classes for home-based work trips. This was necessary because the traditional gravity model considers only travel time and the number of work trips and not whether the residents of a production zone are likely to have jobs in the attraction zone (which could be determined using sources such as census journey to work data). For example, many relatively high-income jobs are located in the CBD while many lower income residential areas are located closer to downtown. In Atlanta, separate home-based work gravity models were estimated for four income classes, resulting in a much better match of residents to jobs. The major difficulty in such an undertaking is the lack of data on income at the job (attraction) zone level. 7.4 Transit Transfers It is generally accepted that the requirement of transit transfers, either between transit modes or between routes on one mode (e.g., bus to bus) generally discourage transit use. Mode choice models in use today have a variety of ways of dealing with transfers. Most models have one or more variables representing out-of-vehicle time, including access, wait, and transfer time. Components of out-of-vehicle time are sometimes broken out as separate variables; these may include: walk access time, wait times (first link, second link, etc.), and transfer time. Some models include the number of transfers as a model variable. Two efforts that summarized the way in which transfers (and other service characteristics) are dealt with in many mode choice models were prepared by Schultz2 and Charles River Associates8. As described below, Schultz reviewed models from several cities. For homebased work trips, four models had separate coefficients for transfer time. As Table 7.1 shows, all of these models had similar coefficients for transfer time and walk time, implying that separate coefficients for each individual variable were not estimated. Three 7-6 Click HERE for graphic. 7-7 Click HERE for graphic. 7-8 of the four models, however, had coefficients for transfer (and walk) time higher than for wait time. Unfortunately, the effect of the act of transferring is unknown since the amount of time needed for a transfer is not identified. The CRA memo dealt specifically with the issue of transfers. Seven cities were examined, including three U.S. cities (Honolulu, Chicago, and Boston), two Canadian cities, and two cities abroad. Transfer penalties were estimated in units of in-vehicle or out-of-vehicle time and ranged from 15 to 37 minutes of in-vehicle time and from 2.3 to 15 minutes of out-of-vehicle time. CRA recommended penalties of 8 minutes of out-of-vehicle time for peak/work trips and six minutes for off-peak/non-work trips. The memo also included a comparison of the ratio of transfer time to in-vehicle time coefficients in 19 models in 15 U.S. metropolitan areas. (Some of these were also included in the Schultz memo.) The ratios ranged from 0.825 to 6.87, with most ranging from 1.8 to 2.8. It would seem that in any area where there are a significant number of transit transfers, the transfer should be treated separately from other out-of-vehicle time. If survey data are sufficient to estimate a coefficient separately, this should be done; if not, information such as that presented in the CRA memo could be used to develop appropriate transfer penalties. Finally, the CRA memo touches on the issue of transfers on transit assignments. In many models, the transfer is instantaneous; no penalty is applied. It would seem that if transfers are not considered. (with both time and fare considerations), transit assignment models will estimate too many transferring transit trips. If transit networks are skimmed to provide information to mode choice models, the number of transit trips could be overestimated as well. 7.5 Integrating Mode Choice Models with Trip Distribution, Trip Generation and Land Use Models In some areas, including Minneapolis/St Paul, Atlanta, (ARC/MARTA), the New Hampshire/Maine seacoast region (New Hampshire Department of Transportation (NHDOT) and Seacoast MPO) and Chicago (Chicago Area Transportation Study), output from the logit mode choice model is being used as an explanatory variable in the distribution model. An accessibility variable, commonly known as the logsum variable, is calculated for each origin-destination pair as the natural logarithm of the denominator of the logit mode choice model. This variable represents the expected maximum utility derived from all the possible modes of travel between the origin and destination, taking into consideration, among other things, the time and cost variables. Thus the logsum variable can be interpreted and used as a composite impedance or generalized cost. It provides a much better measure of the costs a traveler faces in making the trip and destination decision because it factors in the various transport options and their associated costs, rather than using only highway travel times. The use of the logsum variable in these regions stands in contrast to the current practice in most urban areas today of using travel 7-9 time, usually highway travel time, as the only impedance variable in a gravity model of trip distribution. The Chicago Area Transportation Study (CATS) uses the logsum variable in an intervening opportunities trip distribution model. The model relies on a parameter called the L value, which is a measure of the probability a destination will be chosen if it is considered. The size of the L value is fixed for a particular origin, but varies across origins. The L value of an origin depends on the accessibility between it and all possible destinations. Accessible origins have low L values, because many destinations can be considered and the proba- bility of any single destination being chosen is relatively low. CATS bases this accessibility measure on the logsum variable, by defining the accessibility variable in the trip distribution model as the number of destinations with logsum values exceeding a specified threshold. Although CATS uses the logsum variable in an intervening opportunities model, it can be just as easily used in the more familiar gravity model of trip distribution. This was done, for example, for the New Hampshire/Maine seacoast region9. Instead of using highway travel time as the impedance variable in the gravity model, or even a cost function which represents some composite of highway and transit times, a form of the logsum variable can be used as the impedance variable, capturing the effect of a broader spectrum of costs.and benefits on the trip destination decision. An even more ambitious modeling approach, which has rarely been implemented in the United States, although it has been successfully implemented in the Netherlands and Sweden, and is being considered more and more in the U.S., is to extend the nested logit structure beyond the mode choice decision to include the decision to travel (generation) and where (distribution). It is here, in the context of a nested logit representation of interrelated travel decisions, that the theory of the logsum variable was developed and first applied. The nested logit model is based on random utility theory, in which people choose the alternative which maximizes their utility. Lower level decisions in the nest take as given the results of upper level decisions. Upper level decisions take into account the expected maximum utility of all lower level alternatives available to the decision maker. The logsum variable is the mechanism by which the model incorporates this last feature. Thus the logsum is an important element in an integrated model system architecture with a strong theoretical and mathematical foundation. In theory, this structure could be extended even beyond the trip generation and distribution models, to include other important decisions such as residential location, travel time-of-day, and trip chaining. The logsum variable has also been used in Seattle (by the Puget Sound Regional Council) and in the New Hampshire/Maine seacoast region (by NHDOT and Seacoast MPO in models developed by Resource Systems Group) to integrate the mode choice model with a land use allocation model, thereby improving the accessibility information used in the land use model. The land use model uses generalized accessibility variables, one for each land use type (residential, retail and non-retail) in each zone, which are derived from the mode choice logsum variable. Each accessibility variable is a weighted sum of the mode choice logsum variables for all the origin-destination pairs associated with the zone. Included in 7-10 the sum are logsums, for home-based work, home-based non-work and nonhome-based trips. The weights in the sum depend on the land use type of the accessibility variable being calculated, and are estimated using multiple linear regression methods10. Integrating the mode choice models with other parts of the travel demand model system comes at a cost. Estimation and application of the models becomes more complex. For example, the use of the logsum variable in the trip distribution model requires an iterative equilibration procedure, because it introduces an interdependence between models. The mode choice model parameters are generally adjusted to achieve a reasonable match with trip counts estimated by the trip distribution model. This adjustment changes the logsum variable which, when used in the trip distribution model, changes the trip counts. Thus, the mode choice and trip distribution models must be adjusted iteratively until the trip counts and logsum variables match in both models. Furthermore, the standard software packages available today for travel forecasting were not designed for his kind of model integration. This means that implementing the enhancements will require the replacement of currently used packages, or a substantial amount of custom programming. However, the advantages of model integration are substantial. These include a better behavioral representation, the incorporation of important variables in the models, and consistency across related models. Together, these advantages should yield more accurate forecasts for intermodal and clean air policy decisions. 7.6 Model Transferability, Monte Carlo Simulation, and the Choice Between Tolled and Non-Tolled Facilities Because of the significant costs of collecting data and estimating travel demand models, particularly disaggregate models such as mode choice models, some areas choose to transfer models or parameters from other areas. While this is generally accepted practice and is often the only feasible alternative, it is considered more desirable to develop models using local data. While determining parameters for transferring mode choice models, Schultz2 compiled a comparison of coefficients for service variables from a set of mode choice models developed in the US using data between 1970 (except for one 1960 survey) and 1984. There were a total of eleven home-based work models and six models each for home- based nonwork and nonhome-based trips. This comparison showed that the range of coefficients for each variable was, generally, fairly tight. For example, for in-vehicle time (in minutes) inhome-based work models, the coefficients ranged from -0.015 to -0.040. The complete comparison is shown in Tables 7.1, 7.2, and 7.3 for the home- based work, home-based nonwork, and nonhome-based purposes respectively. Schultz used this comparison to determine "consensus" parameters for mode choice models to be applied in other cities. The comparison is useful not only for the purpose of model 7-11 Click HERE for graphic. 7-12 Click HERE for graphic. 7-13 transfer, but also to use a guideline to check parameters estimated directly from survey data in other cities. In Chicago, CATS is using a Monte Carlo technique to forecast mode choice by simulating, one by one, all the mode choice decisions in the metropolitan area. For each decision, a transit accessibility variable is simulated using a random number and a probability distri- bution. The probability distribution is based on the route miles of transit service in the zone and a locational variable describing the dispersion of population and employment relative to the transit stops. The probabilities for each mode are then calculated by using the transit accessibility variable along with the other variables in the logit model. Finally, another random number is drawn and its value, in conjunction with the logit probabilities, determines the mode which is chosen for this individual decision. This process is repeated for all trips to simulate the mode choices throughout the region. The advantage of this method is that it helps overcome prediction errors associated with the fact that transit accessibility, an important explanatory variable in the mode choice model, is not uniformly distributed throughout the zone. The errors can be severe if the zones are large and have a lot of variability in transit service, population density and employment density. The method is also easier and less costly than subdividing the traffic analysis zones into smaller homogeneous zones. Finally, it is also easier and more flexible than segmenting the population within a zone by proximity to transit service. In making forecasts, if uneven growth is expected in a zone, it can be handled easily by adjusting the population dispersion variable on which the simulated transit accessibility variable is based. Another model improvement closely related to the mode choice has been developed to represent, for the auto travel mode, the choice between tolled and non-tolled travel paths. This has been done in Orlando for toll road demand studies with a binary logit model. During the assignment process the binary logit model is applied to all zone pairs for which the minimum time path involves the payment of a toll. The probability of a trip using the tolled path is calculated using the logit formula, which includes travel time and cost variables. The number of trips between the zone pair is then split in proportion to the logit probability. Finally, the non-tolled trips are assigned to the fastest non-tolled path11. A similar splitting of tolled and non-tolled trips, as the lowest level in a nested logit mode choice model, has been proposed for the Washington D.C. area. 7.7 Summary This section documents a variety of model improvements, focusing on mode choice modeling. All of the methods described above have been tried in U.S. urban areas. Most seem to be worthwhile to incorporate into the demand models for other cities. Many model parameters appear to be transferable to other areas. 7-14 7.8 References 1. Ben-Akiva, M. and S. Lerman. Discrete Choice Analysis: Theory and Application to Travel Demand. MIT Press, Cambridge, Massachusetts, 1985. 2. Schultz, G. Memorandum to Seattle Metro Files, March 5,1991. 3. COMSIS Corporation. "Models of Mode and Occupancy Choice in the Shirley Highway Corridor," July 7,1988. 4. COMSIS Corporation. "Review of the Shirley Highway Corridor Mode Choice Analysis," October 17,1990. 5. Quackenbush, Karl. "HOV Modeling in Eastern Massachusetts." Paper prepared for presentation at the 7th International TRB Conference on High-Occupancy Vehicle Systems, June 1994. 6. Metropolitan Service District. "Travel Forecasting Methodology Report, Westside Light Rail Project," September 29, 1989. 7. Metropolitan Atlanta Rapid Transit Authority. North Line Corridor Alternatives Analysis/DEIS Environmental Impact Statement, Atlanta, Georgia, Deliverable 10: Methodology for Analysis of Service and Patronage Impacts. Prepared for Urban Mass Transportation Administration, April 1989. 8. Charles River Associates. "Development of a Consensus Paper on How Transit Transfers Affect Ridership." Draft Memorandum to Houston Metro, September 15, 1989. 9. Vanasse Hangen Brustlin, Inc., et al. "Pease Surface Transportation Master Plan," April 1994. 10. Marshall, Norman L., and S.J.C. Lawe. "Land Use Allocation Models for Multi-County Urban and Suburban Areas," in 4th National Conference on Transportation Planning Methods Applications, Volume II, Transportation Research Board, Washington, D.C., September 1993. 11. URS Consultants, Inc. Unpublished project file memo from the Orlando, Three-Model Development project, January 1994. 7-15 8.0 Parking Analysis Procedures Parking is a serious issue affecting many aspects of travel models. One problem is that many travelers do not park at their actual trip destinations, and travel models do not reflect this, either in terms of travel times for mode and destination choice models, or for traffic assignment. Parking cost is another difficult issue for mode choice modeling as well as parking location choice. Further complicating this problem is the existence of employer subsidies and prepaid parking fees. Another issue involves parking at transit stations and park-and-ride facilities. Many transit stations have insufficient parking to accommodate the park-and-ride demand. These real limits on the demand for the park-and-ride mode which are also complicated by the time-of-day issue - should be incorporated into mode choice models. 8.1 Reallocation of Trip Ends to Parking Locations The Traffic Analysis Zone (TAZ) system used by most travel demand models for central business districts typically reflects small units of geography, even down to the block level in some cases. While trip attractions are specific to a TAZ, parking locations for auto trips may be in a different TAZ. Major determinants in CBD traffic assignment include the location and capacities of available parking facilities. The typical vehicle trip to the CBD parks in a lot or garage, and the drivers and passengers walk to the final destination, often in a TAZ different than the parking TAZ. Additionally, in most cities, there is a high percentage of employer-provided free, reserved parking, which complicates the analysis of choice of parking location. There have been a few attempts to assign vehicle trip ends to parking locations rather than directly to destinations. For Central Artery project planning in Boston, Cambridge Systematics used a simple Fratar procedure to reallocate trip ends in a daily trip table to parking locations, based on parking capacities. In a more detailed analysis, Barton Aschman conducted an Uptown Traffic Circulation study for the City of Charlotte, N.C. in which the travel demand model was required to estimated link specific, peak-hour traffic volumes and peak-hour intersection turning movements. In the Charlotte application, the proportion of trips associated with the reserved and free parking received priority in the assignment to the nearest parking facility. The capacity and costs of the parking facility were considered in the assignment algorithm. Barton Aschman developed network coding and assignment procedures that produced CBD peak-hour traffic assignments. 8-1 The development of the CBD parking assignment procedures required the following steps: - The identification of the capacity and cost of each off-street parking facility in the subarea (in most cases the CBD); - The identification of the proportion of employment in the subarea that receive free and/or reserved parking; - The detailed coding of the highway network in the study area that reflects the complete roadway system and the actual access points of the parking facilities to the network (this could be problematic for on-street parking); - The systematic coding of a walk network that connects the parking facility to the ultimate zone of destination; - The coding of "dummy" connector links between the highway network and the walk network that reflect the capacity and costs of the parking facility; and - The "calibration" of the parking demand and capacity/costs relationship for the parking link. The choice of the parking location can be done by two modeling methods; the use of a logit-based choice model or the use of the capacity restrained equilibrium assignment model. With the logit model, the probability of the trips using a given parking facility is computed based on cost, distance and capacity considerations, and then the trip ends are reallocated to the appropriate parking TAZ. The connection from the parking location to the TAZ of destination is used to compute the input variables to the model. The consultant team knows of no application using the logit formulation. A simpler approach, and one that worked in Charlotte, was to let the minimum path, equilibrium assignment algorithm determine the parking location. This approach required a travel demand modeling platform that allowed for great flexibility in the definition of the speed- capacity-delay relationships. Many of the available programs have this capability. The parking link added equivalent minutes (or costs if travel time is converted to cost units for the highway network) that reflected the parking costs and when the link reached the capacity of the facility the link time increased dramatically, effectively shutting down the parking link and requiring the minimum path to be found through other parking "links." As with time-of-day traffic assignments, the accuracy of the results is highly dependent on the accuracy of the coded highway network. In particular, for analyses such as CBD circulation studies, virtually each block in the CBD would have to be coded as a separate TAZ and nearly every street coded in the network. Further refinements of the procedure could include the use of traffic signal delay in the estimation of travel times on the highway network. At that point the procedure starts to become a microsimulation model of real time traffic flow. There are other traffic simulation programs available, such as NETSIM, that are used to model traffic signal systems, or a dynamic assignment procedure (see section on Dynamic Assignment) could be used. The CBD parking model described above 8-2 is capable of forecasting traffic circulation in the CBD and evaluating the adequacy of parking supply and locations within the CBD-The most positive effect the parking allocation procedure has on travel demand modeling is the ability to forecast traffic patterns in the CBD that are more reflective of the actual highway and parking supply. Regional travel demand models can not provide the same information. As with time-of-day assignment models, a serious negative effect of the application of the procedure is the direct use of the results for traffic signal design or detailed project design. Even with the best of modeling procedures, some review and post processing of results may be required. The user of the model must be advised of the possible misapplication and representation of the model results. Another problem with reallocation of trip ends to parking locations is the time-of-day issue; at many times of day, capacity is not an issue. Another problem is the variations in cost for different facilities, trip durations (e.g., all day work trips versus short shopping trips), and times of day. A good deal of additional work is required to apply this procedure. The parking capacity of each lot must be determined, and cost information obtained. The highway network must be recoded to include the walk links and the parking connector links as well as the cost/ time conversions for parking links. The calibration effort is also significant. The CBD parking model would be applicable in all cities where traffic circulation and parking in the CBD are of major concern. This procedure is best applied in medium to larger sized cities (200,000 to 1,000,000 population). For very large cities (population > 1,000,000) the development of the zone system and the required highway network may not be feasible; however for cities similar in size to Charlotte, the procedure works quite well. 8.2 Parking Cost Modeling An important variable for mode choice models is parking cost. Usually data on the parking cost for travelers using modes other than auto (i.e., how much would they pay to park if they chose the auto mode) are unavailable from travel surveys. The usual solution is to assume the same fixed parking cost for all travelers to a zone, based on observed parking prices in the zone. This simplification is obviously flawed since not only does the parking cost vary within a zone, but many travelers, especially commuters to work, have subsidized parking available and may have no cost. The purpose of developing a parking cost model is to be able to estimate parking cost for all travelers based on characteristics of the tripmaker and the trip. 8-3 A parking cost model is under development for the Los Angeles area mode choice models being estimated by Cambridge Systematics for the Southern California Association of Governments. The parking cost model is a discrete/continuous model. This model is estimated based only on the auto users who pay for parking and is therefore expected to be biased. The bias occurs when the unobserved factors that cause high parking cost increase the probability of paying for parking (and, in fact the probability of parking). The bias is corrected using a "selectivity correction" variable which is a function of the probability of choosing auto. Since the household travel survey for Los Angeles indicated that most auto users did not pay for parking, it was decided that the choice probability model should be between auto users who pay to park and all others. In this case the logit model estimates the probability of a traveler to use auto and to pay for parking, and the continuous model shows, given that the traveler chose to pay for parking, how much he pays for parking. Inclusion of the ft selectivity correction" variable in a linear regression model of parking cost (for those who park and pay) enables the use of the estimated model for all travelers. When the model is later applied to all travelers, the selectivity correction variable is omitted. This model predicts the parking cost for all travelers unrelated to their choice probabilities and can be used as a variable in the development of the min choice mode model. 8.3 Summary The accurate peak-hour (or period) assignment of vehicle trips in a CBD is highly dependent on the location and availability of parking. Conventional highway networks and TAZ systems do not address the likelihood that the location of available parking is often not in the actual destination zone. The procedure described above has proven to be effective in estimating CBD traffic patterns. Parking costs are also a major input to mode choice. A discrete/continuous model has been proposed that estimates parking costs based on the probability of using the auto mode and paying for parking. Both of these procedures improve what has been a set of submodels that has received minimal attention in the overall travel demand modeling process. 8-4 9.0 Time-of-Day Models The purpose of time-of-day travel demand models is to produce traffic assignment results that more accurately reflect the capacity restraining impact of the highway network on the traffic volumes. In highly congested areas, particularly large urban areas, the finite amount of physical highway capacity results in the spreading of the peak periods. While it is not possible for a roadway to carry an hourly volume of traffic that is greater than its (or its intersections') theoretical capacity, the highway assignment algorithms commonly used can (and often do) produce traffic volumes on roadways that exceed the capacity. In these cases, the volume of traffic assigned during the peak periods must be dynamic and change as the capacity of the highway system is reached. This can be done by using a dynamic assignment procedure (see section on Dynamic Assignment) or by increasing the time period over which the volume can be assigned. Methods have been developed that account for this spreading out of the peak volumes. In smaller to medium sized urban areas the peak periods have not spread to the extent as those in the larger areas. While there are capacity restraints at some localized points in the highway system, the overall highway system has not reached capacity during the peak period, and the capacity restraint assignment procedures can adequately reflect highway capacity. Rather than shifting to another time period, the vehicles shift to alternative routes that are uncongested. For these smaller to medium sized areas (and even for the large areas), historically the method for obtaining daily capacity restrained traffic assignments has been to multiply the hourly capacity by 10 to reflect. the "daily" highway capacity. This was based on the assumption that the peak-hour traffic represented about 10 percent of the daily volumes. The UTPS programs (and some of the microcomputer-based packages) contain the CONFAC parameter which is used to adjust for daily capacity restraint assignments. There are major problems with this simplistic approach. This type of factoring does not account for the differences in peaking characteristics among different locations in the network. The directional imbalance of traffic volumes during the a.m. and p.m. peak periods is not considered. This approach has been adequate for the application of regional travel demand models for preparation of the long range transportation plan. However, models are being asked to provide more information, focusing on directional traffic assignments and turning movements. Newer models are being used for short-term analysis. 9-1 9.1 Hourly Factoring of Daily Trip Tables A procedure that is widely used (but not well documented) is to factor the daily trip tables by purpose and produce peak-hour (or period) directional origin-destination trip tables. These trip tables are static and are not dynamically adjusted during the assignment process as are those that result from peak spreading algorithms mentioned above. The daily volumes are produced by adding up the results of the a.m., p.m. and off-peak traffic assignments. An added benefit of using this technique is that assignments by time of day can be produced for input to the air-quality analysis and for the better estimation of congested speeds for use in the trip distribution and mode choice models. The process for preparing peak-hour directional trip tables requires the factoring of the person or vehicle production-attraction formatted trip tables to peak-hour (or period) origin-destination formatted vehicle trip tables. The data required is an hourly distribution of trips across the day. These should be by purpose, usually grouped into home-based work, home-based non-work, and nonhome-based. From the this diurnal distribution of trips, factors are developed which represent the percentages of the trips (by purpose) during each hour and for each direction, production to attraction or attraction to production. The hourly distribution is developed from local travel survey data. The production-attraction formatted trip tables are multiplied by the appropriate factors to produce origin-destination trip tables. The ability to accurately forecast travel by time of day and direction is dependent on the accuracy and detail of the coded highway network and the validity of the diurnal factors. Historically, mainframe travel demand modeling packages (UTPS) provided little graphical input/output with which to code and check the highway network. Interchanges were coded as single nodes and symmetrical two-way links dominated the network. With the widespread use of the microcomputer- based travel demand software packages that incorporate on-screen graphical display and editing of the networks, the accuracy of the highway network is greatly improved. Additionally, the high cost of mainframe computer time is no longer a determining factor in the size of the network (number of links and zones) which in turn determined the run times. The diurnal factors are best derived from home interview survey data. Person trips by time of day and by trip purpose are required. Also, a good estimate of auto occupancy rates by purpose and time of day are also required. If the region is using a mode choice model to produce the auto vehicle trips then the model results should be compared with observed auto occupancy rates. Even with best of network coding and estimation of the diurnal factors, the resulting peak period traffic assignment will still require review and in some cases post-processing to produce the final predicted traffic volumes. An important TRB publication that addressed the use of the regional travel demand models for project planning and analysis was NCHRP 2551. 9-2 The most positive impact of this procedure on travel models is the use of more realistic peak-hour or period traffic volumes in the development of capacity restraint assignments as opposed to using the pseudo peak volumes used in 24-hour capacity restraint assignments. Another positive impact is the direct use of traffic assignment results in the air quality analysis. The most negative impact may be the false security that can be associated with a model that produces peak-hour directional specific volumes. Forecast volumes can be assumed to be more precise than can be reasonably assumed and can be improperly used for such purposes as traffic signal design and freeway weaving analysis. As with the historical regional model results, the time of day, directional results must be carefully reviewed and applied properly. Time-of-day factors could be borrowed from other urbanized areas if original O-D survey data are not available. If the highway network is coded properly then borrowed factors by purpose can be used. 9.2 Peak-Hour Trip Table Reduction to Reflect Network Capacity Constraints When forecast year peak-hour vehicle trip tables are assigned to highway networks which are at capacity or congested in the base year, the resulting forecast year traffic volume estimates can exceed capacities by unrealistic amounts. This is because, typically, growth rates are applied during the trip generation phase of the modeling system, without consideration for traffic conditions. Trip distribution models and mode choice models reflect the highway capacity constraints by shortening trip lengths and increasing HOV and transit shares, but the effect of peak spreading - where tripmakers who would prefer to travel during peak hours make their trips earlier or later to avoid congestion - are not captured in peak-hour analyses. To combat this problem, Cambridge Systematics has developed a technique to reduce a trip table selectively. In this procedure, individual origin-destination cells of the trip table are reduced according to how congested the corridor corresponding to the origin- destination pair is. Selective reduction, which is accomplished using "selected link analysis," is superior to global reduction (which implies a general decrease in trip generation rates) because predicted traffic volumes in uncongested corridors are not changed by unrealistic amounts. 9-3 The trip table reduction process consists of five major steps: 1. Unconstrained assignment; 2. Selection of links to be examined; 3. Sequential adjustment of volumes for origin-destination pairs in the selected link trip tables of congested links; 4. Reassignment using adjusted trip table, and 5. Comparison of final link volumes with link capacities. If the link volumes are sufficiently close to the target capacity, the process is complete. If not, the trip table reduction process is repeated using the new selected fink trip tables. Cambridge Systematics has applied this procedure in their traffic forecasting work for Boston's Central Artery/Tunnel Project, for the Massachusetts Highway Department. 9.3 Traffic Assignment with Peak Spreading As peak-hour congestion increases on urban highways, drivers wishing to avoid the added delay have a number of options, including: - Seek alternative routes to bypass the congestion; - Switch from auto to transit; - Choose a different, more accessible, destination; - Stop making the trip; and - Make the trip at a different time of day. Existing travel models can predict the extent to which some of these options (rerouting, mode shifts, destination shifts) will be chosen, but not the complete set of possible responses. Cambridge Systematics has developed a procedure which provides a method to reflect the net effect of all of the possible responses. The procedure, which has been applied in a TRANPLAN setting, represents a shortcut estimate of the full set of options without identifying the magnitude of each individual type of response. However, by providing an estimate of the total impact, the procedure provides more realistic estimates of future travel volumes on facilities where congestion is expected to worsen. 9-4 The peak spreading procedure is applied as part of a peak period (typically, three-hour) equilibrium assignment. As each link is considered, in turn, during the equilibrium assignment's travel time updating, peaking factors representing the ratio of peak-hour volumes to peak period volumes are computed using a decreasing function of the link's three-hour volume-capacity ratio. The peaking factor function can be estimated with time series and/or cross-sectional vehicle count data. The peak-hour volume corresponding to this peaking function is used to estimate revised travel times during the assignment procedure. Cambridge Systematics applied the approach in Phoenix, in model enhancement work done for the Maricopa Association of Governments, with some success. The accuracy levels of the modeling system's VMT and speed estimates were increased significantly. 9.4 Pre-Distribution Time-of-Day Models The previously discussed time-of-day models were either post- distribution or post-assignment techniques. Many travel models use peak period level-of-service characteristics (times and costs) for distribution and mode choice analysis of home-based work trips and off-peak characteristics for non-work trips. However, there are trips of all purposes during each of these periods. In models developed for the Metropolitan Transportation Authority's Red Line East Side Extension project in Los Angeles, a pre-distribution time-of-day model was developed. In this technique, the trip ends are split by time period for each trip purpose. The same technique will be applied in the model developed for the Dulles corridor alternatives study1. The time-of-day model used is a two-step model. The initial step is the pre-distribution model, in which a set of factors is used to calculate trips by time of day, usually for multi-hour peak and off- peak periods, and by trip purpose. The factors are based on peaking characteristics such as trip purpose, jurisdiction, area type, and socioeconomic stratification. These factors are applied to the trip ends from the trip generation model and produce trip ends by peak and off-peak periods for each of the trip purposes. The peak and off-peak trip ends are used in the trip distribution and mode choice models. The resulting trip tables, by mode, are then factored in the second, or final, time-of-day model. This second time-of-day model is similar to the post-distribution model discussed previously in this chapter. The user can specify the time period desired and factors based on trip purposes and mode are applied to produce the desired trip tables, usually representing peak and off- peak hours rather than multi-hour periods. Secondary factors which may be input to the model include length and location of the trip. 9-5 9.5 Summary The time-of-day highway assignment procedures have become standard practice in new model development studies. The procedures provide for more realistic travel, congested speeds, and daily capacity restrained traffic assignments that those which use the assignment of a single 24-hour capacity restrained assignment. The procedures require detailed coding of the highway networks and the availability of diurnal travel factors. 9.6 References 1. Parsons Brinckerhoff. "Technical Methods: Specifications for Travel Forecasting Models," report for the Dulles Corridor Alternatives Study, May 18,1994. 9-6 10.0 Trip Table Estimation The development and maintenance of accurate trip tables is a key element of urban travel demand modeling. Traditionally, travel patterns are represented in trip tables by using a single source of data, such as a household survey, and a trip distribution model, such as a gravity model or a Fratar model, but in recent years, modelers in several urban areas have attempted to improve their ability to generate and update current year trip tables by making the maximum use of available travel data, including existing trip tables, household survey information, traffic counts for specific locations, limited origin-destination surveys for specific locations, and total zonal origin and destination volumes (zonal trip generation estimates). In many cases, the synthesized trip tables are believed to be more accurate than traditional tables since they reflect many separate data sources. Trip table estimation can provide a more efficient use of a wide range of data. In addition, it can provide analysts the ability to adjust old or subarea trip tables to make them more accurate and timely. The use of the additional data sources may help an agency interested in updating or geographically focusing trip tables avoid expensive survey work. The improvements in the current year trip tables can also help enhance future-year trip tables. Using the more accurate current year data, analysts can determine correction factors to apply on future trip tables. Trip table estimation also provides us with the ability to estimate and apply incremental modeling approaches which predict changes from existing travel patterns. There is reason to believe that these approaches can be more accurate than fully synthetic models for near-term forecasts and for stable regions. On the negative side, as more data sources are included in the estimation of trip tables, inconsistencies and conflicts between sources become common. These problems can be perplexing, and some subjective judgment may be necessary to develop the trip tables. In addition, trip table estimation relies on the accuracy of the transportation network and assignment techniques. For these reasons, the results of any trip table estimation procedure need to be reviewed carefully for reasonableness. In addition, the most appropriate trip table estimation procedures are specific to individual cases, so all trip table estimation efforts will involve new considerations. Flexible software and highly skilled staff are probably prerequisites. 10-1 10.1 Available Trip Table Estimation Procedures Two general approaches have been followed in estimating trip tables; procedures which factor existing trip tables (and are essentially extensions to the Fratar trip distribution process) and procedures involving mathematical programming and statistical techniques. The trip table factoring approaches iteratively adjust the rows and columns of an existing trip table until a pre-specified variance between actual and predicted values is reached. The mathematical programming and statistical procedures attempt to develop trip tables which minimize the variance between predicted and observed counts. This is done either by the formulation of a linear (or nonlinear) programming problem or through advanced regression techniques. Trip Table Factoring Procedures Several commercial transportation planning software packages have limited trip table estimation capabilities that rely on various forms of iterative proportional fitting - a technique for factoring matrices similar to the way a Fratar model works. EMME/2's procedure allows users to make adjustments to existing trip tables in three dimensions - origin totals, destination totals, and trip length categories. The EMME/2 trip table estimation procedure is accessible to users and is efficient in terms of computer memory requirements, but its ability to improve upon existing trip tables is limited, particularly in areas with inconsistent data sources. Two other software packages TMODEL/2 and THE use iterative proportional fitting techniques to match data on origin volumes, destination volumes, and specific link counts. These procedures are not difficult to apply, but they can estimate unrealistic trip tables if the traffic count data are too sparse or if count stations are not well placed. In addition, trip table estimation procedures that use link volume counts are susceptible to problems involving data inconsistencies. Cambridge Systematics has developed the TTE program, which uses a modified iterative proportional fitting algorithm focusing on matching screenline volumes, rather than specific link volume estimates. Since the screenline counts are more stable than individual counts, TTE avoids many of the path data problems of the other trip table estimation algorithms. TTE can be used with either TRANPLAN or EMME/2. Cambridge Systematics has applied TTE on Boston's Central Artery/Tunnel project, on a corridor study in Staten Island, New York, and on intercity analyses in southern Indiana and Wisconsin. In most of the applications, separate vehicle type trip tables were developed. 10-2 Mathematical Programming and Statistical Procedures for Trip Table Estimation At least one major transportation modeling software package, TRIPS, uses a mathematical programming approach to estimate trip tables. MVESTM, a module of TRIPS, is much more flexible than the procedures described above in terms of the types of travel data that can be included in the development of trip tables. However, analysts are not likely to be very familiar with the mathematical procedures which underlie MVESTM, so the module is most likely treated as a 'black-box" in most instances. In addition, like most mathematical programming approaches, the procedure requires a great deal of computer running time and memory. MVESTM has not yet been adapted to multi-class or multi-commodity problems as the approach described below has. List/Turnquist1 have applied another mathematical programming approach for estimating trip tables in their recent study of multi- class truck trips in New York City. In this approach, a wide variety of incomplete and fragmentary data sources, including: - Link volume data; - Classification counts; - Partial origin-destination estimates for various zones, time periods, and truck classes; and - Originating/terminating data by class for internal and external zones. In the List/Tumquist procedure, a large-scale linear programming approach is formulated which seeks to minimize the weighted sum of deviations from the observed data. Because the deviations are weighted, analysts can subjectively account for the differences in the perceived accuracy of the data sources. Because this approach is more flexible in terms of data inputs and because it develops multi-class trip tables, it is more difficult to set up than the other approaches. Another problem is that the approach tends to yield sparse matrices. The authors have not yet developed provisions for filling empty cells. (This capability is available in TRIPS) 10.2 Resources Needed for Trip Table Estimation The data requirements for the different trip table estimation procedures vary slightly, but each is oriented to increase the use of available data. In general, the more available travel data, the better - of course, this assumes that the data sources are reliable and accurate. The procedures do not require new data collection - in fact, the trip table estimation procedures grew out of desire to avoid new expensive data collection efforts. However, an 10-3 output of the trip table estimation procedures may be the identification of specific data collection needs which could most improve the trip tables. It is likely that an agency interested in trip table estimation will need to invest in extensive staff training and/or consultant agreements to develop the ability to update and maintain its trip tables. In addition, the agency will need to purchase proprietary software, and perhaps may also need additional computer resources. 10.3 References 1. List, G. and M. Turnquist. "Estimating Multi-Class Truck Flow Matrices in Urban Areas," presented at the Transportation Research Board Annual Meeting,-Washington, D.C., January 1994. 10-4 11.0 Modeling of Trip Generation Input Variables Most U.S. urban area models compute trip productions through either cross classification tables or linear equations. These models usually estimate trips produced by a household. Variables affecting a household's trip generation depend on the trip purpose being generated and may include the number of persons, the number of vehicles available, number of workers, income, age of household head, and number of children. Information from which households can be classified by these variables is generally available from census data. To apply trip production models, however, forecasts of households disaggregated by the variables are needed. In most areas, such forecasts would not be available from other sources since their usefulness for other applications is limited. Metropolitan areas have dealt with this problem in a number of ways, ranging from having relatively unsophisticated trip generation models, which require little in the way of socioeconomic forecasts to use, to developing separate models to estimate the necessary variables. This section focuses on two methods: - Use of data on existing households to estimate forecasts; and - Use of separate models to estimate the forecast variables. In addition, a third method, simulation of households in terms of formation, location, and other decisions is briefly discussed. Although this method is not currently used in any U.S. urban area, it has been documented in academic research literature. 11.1 Use of Existing Data Few U.S. urban areas develop models to estimate forecasts of variables needed for trip generation models. Commonly, MPO's use one or more of the following methods: - Disaggregation of totals using percentages derived from base year conditions - Often, an urban area will have forecasts of total households by zone, but no information on how to disaggregate by income level, auto ownership, household size, etc. In many cases, the MPO will estimate percentages based on the base year disaggregation, which may be derived from household survey or census data. 11-1 - Use of regional economic or demographic models - Forecasts of population, households, or employment may be obtained from such models, which may be used in the urban area for other planning purposes. Usually, the forecasts from these models will be at a much more aggregate level than the zone system for the transportation model, and some form of allocation must take place. - Use of zonal averages - Rather than estimate the number of households at different levels of a variable such as income, trip generation models (or other models such as auto ownership) may be developed so that only a zonal average is needed. The observed base year average may be used, or it may be modified to reflect regional forecast conditions. 11.2 Use of Separate Models Only a few U.S. metropolitan areas have developed separate models of household characteristics to forecast input variables to trip generation procedures. Auto ownership models have been developed in a few areas, including Portland1, San Francisco2, and Milwaukee3. Number of children and number of workers models have also been estimated in Portland. The most common forms of household characteristics models in U.S. urban transportation modeling processes are multinomial logit and linear regression. As a general example, the Portland models for workers, autos, and children per household are presented. Though the models estimated in other areas may use different forms or data sources, the underlying philosophy and objectives are the same. In Portland, the number of households in each zone is estimated exogenously to the transportation modeling process. These household estimates are disaggregated by household size (1, 2, 3, or 4 or more), income (4 categories), and age of householder (4 categories). This 4x4x4 cross classification is referred to as the HIA classification. These HIA classifications are estimated for each zone for base and forecast years. The workers, autos, and children per household models are multinomial logit models. While logit models in transportation planning settings are usually choice models, such as mode or destination choice, the logit formulation can be used to model any set of probabilities for which a utility function based on a data set of independent variables can be derived. (Each of these models does, in fact, have an element of household choice; however, the decision process is not really being modeled in this case.) The workers per household model is specified as follows: U = 5.85 - 1.98*HHSIZE - 1-12*INCOMECL + 1.37*AGECAT for 0-worker households U = 8.92 - 1.55*HHSIZE - 0.55*INCOMECL - 0.30*AGECAT for 1-worker households 11-2 U = 7.71 - 1.20*HHSIZE - 0.16*INCOMECL - 0.76*AGECAT for 2-worker households U = 0 for 3 or more worker households where: U = utility HHSIZE = household size (from HIA distribution) INCOMECL = income classification (roughly in units of $10,000 in 1985 dollars from HIA distribution) AGECAT = age category (from HIA distribution) The auto ownership model is specified as follows4: U = -1.684 - 0.881*HHSIZE - 1.452*WORKERCL + 3.255*INCOM1 + 1.942*INCOM2 + 0.000220*RET1M + 0.00001063*TOTAL30T + 0.2095*PEF for O-car households U = 1.497 - 0.720*HHSIZE - 1.065*WORKERCL + 2.259*INCOM1 + 1.944*INCOM2 + 1.033*INCOM3 + 0.000132*RET1M + 0.00000615*TOTAL30T + 0.0902*PEF for 1-car households U = 1.619 - 0.141*HHSIZE - 0.660*WORKERCL + 0.377*INCOM1 + 0.555*INCOM2 + 0.478*INCOM3 + 0.000060*RETlM + 0.00000334*TOTAL30T + 0.0337*PEF for 2-car households U = 0 for 3 or more car households where: U = utility HHSIZE = household size (from HIA distribution) WORKERCL = number of workers (from workers model) INCOM1 = 1 if income class = 1 (from FRA distribution) INCOM2 = 1 if income class = 2 (from HIA distribution) INCOM3 = 1 if income class = 3 (from HIA distribution) RET1M = number of retail employees located within one mile TOTAL30T = number of employees located within 30 minutes via transit PEF = pedestrian environment factor (see Section 2.3 for definition) The children per household model is specified as follows: U = - 3.24*HHSIZE + 5.54*AGECAT for 0-child households U = - 1.82*HHSIZE + 3.46*AGECAT for 1-child households U = 0.01*HHSIZE + 0.32*AGECAT for 2-child households U = 0 for 3 or more child households 11-3 where: U = utility HHSIZE = household size (from HIA distribution) AGECAT = age category (from IRA distribution) The models for Portland were estimated using household survey data, and similar models could be estimated for any area with such data available. Even if local survey data were not available, other data sources such as census data can be used. Purvis2 describes a procedure where the census Public Use Microdata Sample (PUMS) was used by the Metropolitan Transportation Commission to develop auto ownership models for the San Francisco Bay Area. The estimated models were statistically significant, but the variables that can be tested are limited to the demographic variables available in the census data. Since the PUMS samples were for fairly large areas, such area-specific variables such as those used in the Portland model cannot be tested. 11.3 Household Simulation Some researchers contend that the best way to model household characteristics is by simulating households. Simulation is performed in terms of such features as formation and location of households, changes in household structure (such as births, deaths, and persons moving in or out), as well as choices made by households. This type of simulation requires that households are "tracked" over time. In one example of a household simulation model, Brownstone et al.5 developed simulation models of auto ownership by type and number. In the personal vehicle model, the inputs are the current household structure and the vehicle holdings, classified by 14 types. Household characteristics are modeled in terms of household size, age, etc. This includes simulation of births, deaths, marriages, children leaving home, and other changes in household structure. From this information, whether a vehicle transaction (disposal, replacement, or addition of a vehicle) occurs is modeled. If a transaction is simulated, then the vehicle characteristics of the household are updated. Finally, a vehicle utilization model estimates the use (VMT) for each vehicle. The vehicle types include various conventional and clean fuel vehicles so that choices of vehicle types which affect fuel consumption and air quality can be modeled. Another example of a detailed microsimulation of households is described by Goulias and Kitamura6. This includes a more detailed simulation of household socioeconomics and demographics, using data from the Dutch National Mobility Panel survey. The simulation involves household type (singles, childless couples, families, single parents, and others), birth and death, formation of new households, employment, income, and other household characteristics. The mobility component of the model includes simulation of auto ownership, trip generation, mode split, and trip distance (by mode) models. 11-4 Household simulation provides substantial benefits to travel modelers, reducing the error of simple forecasting and allocation methods and better modeling the choices and attributes that affect household structure and auto ownership. The biggest problem with simulation models is that their estimation requires time series (panel survey) data. Such data are available in only one U.S. urban area (Seattle) at present, and for only three waves so far. Obviously, even if such data collection efforts are undertaken in other areas, the data collection will take years, and the data will therefore be unavailable for short-term modeling. The ability to transfer such models has not been widely examined, but how to deal with area specific factors such as income differences, employment and land use composition, and area types (density, pedestrian friendliness, etc.) implies that such transfers would be questionable. 11.4 References 1. Metropolitan Service District, "Travel Forecasting Methodology Report, Westside Light Rail Project." September 29, 1989. 2. Purvis, C. "Estimating Regional Auto Ownership Models Using 1990 Census PUMS." Proceedings of the Fourth National Conference on Transportation Planning Applications, Transportation Research Board, September 1993. 3. Southeastern Wisconsin Regional Planning Commission. "Travel Simulation Models for the Milwaukee East-West Corridor Transit Study." May 1993. 4. Rossi, T., K. Lawton, and K. Kim. "Revision of Travel Demand Models to Enable Analysis of Atypical Land Use Patterns." Proceedings of the Fourth National Conference on Transportation Planning Applications, Transportation Research Board, September 1993. 5. Brownstone, D., D. Bunch, and T. Golob. "A Demand Forecasting System for Clean-Fuel Vehicles." Presented at the OECD Conference on Fuel Efficient and Clean Motor Vehicles, Mexico City, March 28- 30, 1994. 6. Goulias, K. and R. Kitamura. "Travel Demand Forecasting with Dynamic Microsimulation." Presented at the 71st Annual Meeting of the Transportation Research Board, January 1992. 11-5 12.0 Trip Assignment Issues There are several short-term model improvements identified that can be grouped under the single topic of trip assignment issues. These range from network coding techniques to the interpretation of the model results. This chapter discusses several assignment issues and techniques: - Coding transit access links using GIS; - Toll analysis - using link time penalties and path choice models; and - Instability of highway assignment in saturated networks. 12.1 Coding Transit Access Links Using GIS Transit modeling is usually performed using a series of path and access mode choices. Additionally, market segmentation of potential riders (percentages of zone within prescribed walking distance of transit stops) has been used in the mode choice model. Advances in the mode choice models using the nested logit structure have produced probabilities of transit paths and modes of access; however, travel time and cost characteristics must be created for each path (local versus express bus) choice and each mode of access (walk versus drive). The path for each combination must also be saved for the eventual assignment of the transit trips to the matching combination of path and access mode. Key to the identification of these choices is the coding of the transit access links, which connect the zone to the transit routes. The coding of transit access links has followed a set of rules that have changed little over the last several years. The basic rule was that if all of the zone is within walking distance (usually 1/4 to 1/3 mile), then only a walk access link(s) would be coded connecting the zone to the nearest transit route stop(s). If the entire zone is outside the walking distance, only an auto access link would be coded connecting the zone to the nearest park-and-ride location. Zones that have partial coverage within the prescribed walking distance would be coded with both link types. A variation to this dual access link convention is to code auto access links from all zones, regardless of the walk potential. This represents the availability of park-and-ride to tripmakers who could also walk to transit. The coding of these access links is a manual process using zone maps overlaid on the highway and transit networks. The percentage of each zone within the prescribed walking distance was estimated by visual inspection, or possibly by using a planimeter to measure the area. The nearest park- and-ride lot to each zone was also manually identified and coded. 12-1 The extensive implementation of geographic information systems (GISs) provides the potential for automating the coding of transit access links. The Central Transportation Planning Staff in Boston uses the ARC/INFO GIS for the identification of the closest park-and-ride nodes from each centroid. Node selection is based on the minimum time path from the highway network. The auto access link is then coded. The potential park-and-ride locations are not limited; therefore a zone could be connected to several available park-and-ride nodes. The GIS is also used to code the walk access links and the transit coverage of zones provided by the walk access mode. Using the buffer feature of the GIS, the percentage of the zone within the prescribed walk distance of transit stops or routes is easily calculated. 12.2 Toll Analysis - Using Link Time Penalties and Path Choice Models Toll roadways and bridges are becoming an attractive alternative for financing highway improvements. As the toll facilities become more prevalent in a region, they must be incorporated into the travel demand models. The estimation of future traffic for a toll facility is important in estimating the future revenues required to retire any bonds used for construction and to finance the operation and maintenance of the facility. The toll costs (both the actual dollar costs and the travel time delay resulting from the collection of tolls) can be incorporated into the travel demand model in one of two manners: - Adding a toll link with an impedance that represents the time delay and costs of the toll and allowing the highway assignment algorithm to produce the traffic demand for the toll facility; or - Creating a highway path logit model that estimates the shares of trips that would use the toll path versus the free path. URS Consultants has developed a two-parameter binary logit model that can estimate the shares of the toll and free paths. The advantage of incorporating a logit model into the toll analysis is that the highway travel market can be segmented (by income group for example) prior to highway assignment. Similar to transit mode choice models, the sensitivity to toll costs and travel time savings should vary according to the income of the tripmaker the higher the income, the less sensitivity to costs and more sensitivity to time savings. In this logit model application, the skims for each path choice would be created and used as input to the logit model. In addition to the highway assignment model, toll link impedances should also be used in trip distribution models. If a logit model is used to allocate trips to toll and free highway paths, a composite impedance of these paths should be used in distribution model calibration and estimation. 12-2 12.3 Instability of Highway Assignments in Saturated Networks Equilibrium highway assignment and other capacity restrained procedures have become standard in most travel forecasting systems. Because the algorithms used are iterative, assignments are often characterized by "flip-flopping" of traffic volumes between competing facilities (paths) with each iteration of the model. In highway networks that are extremely saturated (a high percentage of links operate near, at, or over capacity), a small change in a single link can have a ripple effect throughout the network. This problem becomes readily apparent when the transportation planner attempts to decipher the mode results after a change in the network. It is difficult to determine if the forecasted travel is consistent with, and directly attributed to, the proposed network change. In many cases, changes in the assigned volumes occur on links far away from the changed link - a result of the minor impedance change that changes many of the equilibrium paths. Both Barton Aschman and KMPG Peat Marwick have encountered this problem in travel forecasting work for the Florida Department of Transportation and Metro-Dade in the Miami area. The highway network in the region is a basic grid with major arterials spaced about one- half mile apart in both the north-south and east-west directions. To travel diagonally through the region, any number of alternative paths are available, with the differentiation of impedances between competing paths being very small. Not only does this instability of the minimum paths impact the highway assignment; it can also have significant impacts on the mode choice and transit assignment models. A highway link used by transit routes that "shuts down" due to the high assigned volumes produces unrealistic time data for the mode choice model and will impact the estimated transit demand. In the Miami model, it was found that transit routes using extremely congested highway links simply do not receive any ridership. The major problem associated with the above characteristics of highway assignment procedures is the correct interpretation of results used in decision making. The transportation planner must be able to review the results of the model and determine if they are reasonable. In addition, assumptions and procedures must be consistent among all model runs. 12-3 NOTICE This document is disseminated under the sponsorship of the U.S. Department of Transportation in the interest of information exchange. The United States Government assumes no liability for its contents or use thereof. The United States Government does not endorse manufacturers or products. Trade names appear in the document only because they are essential to the content of the report. This report is being distributed through the U.S. Department of Transportation's Technology Sharing Program. DOT-T-95-05 DOT-T-95-05 TECHNOLOGY SHARING A Program of the U.S. Department of Transportation
