«Technical Report Documentation Page 1. Project No. 2. Government Accession No. 3. Recipient's Catalog No. SWUTC/11/161127-1 4. Title and Subtitle 5. ...»
The primary difference between the NL model and the PCL model is how they represent similarity between pairs of alternatives: in the NL model all pairs of alternatives in a common group have the same similarity as all other pairs, while in the PCL model each pair of alternatives can take on a similarity relationship that is independent of the similarity relationship between other pairs of alternatives (Koppelman & Wen, 2000). Koppelman & Wen (2000) calibrated the MNL, NL, and PCL models for the same study areas. The study is to estimate the demand for HSR in the Toronto-Montreal corridor, and used data from observations of 2769 travelers who chose air, train or car to travel in the corridor. The NL model contains train-car nest and air-car nest; the PCL model contains train-car similarity, air-car similarity, as well as train-car and air-car similarity. The model containing similarity parameters for air and car and for train and car represents the specialty of the PCL model structure which cannot be realized by the MNL or the NL models. It also yields the highest loglikelihood, and hence demonstrates the statistical and structural superiority of the PCL model (Koppelman & Wen, 2000).
Figure 2: Overview of different random utility models (Source: Koppelman & Sethi, 2005) Koppelman & Sethi (2005) developed a Heterogeous GNL model which combined the GNL model that allows for non-independent errors, the Heteroscedastic MNL which allows nonconstant errors across observations, and the Covariance Heterogeneous NL model which allows for non-constant correlation structure across observations. This combination of three model structure enhances the model’s ability to represent the complexity of intercity travel choice decision making. This study used data from a Stated Preference survey of both existing rail users and travelers using other intercity travel modes but not rail to analyze rail service class choice.
Koppelmand & Sethi developed the model in four stages starting from the simple MNL structure and sequentially relaxed some of its restrictive assumptions to realize more flexible models.
Both NL and GNL models relax the identical distribution assumption of the IIA, another model structure, mixed logit, was recently developed to relax both the identical and independence assumption simultaneously. Ashiabor et al. (2007) compared a NL model structure and a mixed logit model structure. The two models were developed to study national-level intercity travel market share among automobile, commercial airline and a possible new mode – small aircraft transportation system (SATS) in the U.S. In this analysis, the 1995 ATS data served as the source of traveler information supplemented with a random survey of 2000 samples conducted by the authors. The automobile trips included all trips between any pair of counties and the air trips included all trips between any pair of airports in the U.S. The models were separately calibrated for business and non-business trip purpose. The NL and the mixed logit model included the same variables which are travel time, travel cost, household income and location of the trip origin or destination (whether it is inside MSA or not). The differences between the two model structures are the time coefficient is no longer fixed in the mixed logit model which also does not have nests. Ashiabor et al. found that all variables in the models were significant, and the mixed logit model always had a better fit than the NL model.
Srinivasan et al. (2006) calibrated a rank-ordered mixed logit model to evaluate the impact of security perceptions on intercity mode choice is the aftermath of the event of September 11, 2001.
The study used data collected in New York City from October 2003 to May 2004. The survey asked respondents to rank-order four travel modes for a business trip under different scenarios for one of six intercity corridors in the Northeast and Midwest region. Ten scenarios were defined by varying the values of variables including time-of-day departure, travel time, airplane inspection and boarding time, and travel cost. In addition, the survey asked questions on the individual’s security perceptions and the travel characteristics of respondents in their assigned travel corridor. Srinivasan et al. concluded from the model results that the success of the strategies applied to improve aviation security to sustain air users depends on the passengers’ perception of the measures implemented and the inspection times.
4.3. Application Example: Intercity Bus Transportation
Intercity bus represents a very small intercity travel market segment. Yet, intercity bus transport provides a critical role for smaller communities or rural areas where air or passenger rail options are not available, and for population who cannot afford other higher-cost transport modes (TRB, 2002). The Texas Department of Transportation conducted an on-board intercity bus survey in the state of Texas in the early 1980s (Urbanik, et al., 1982). Urbanik et al.(1982) summarized the survey results and compared the results to an on-board survey conducted in Michigan in 1977.
They found that low-income persons were a significant part of intercity bus riders, and having no automobile was an important reason why people chose riding intercity bus. However, they also found that the loss of bus service would appear to leave only a small number of persons without an alternative travel mode. In addition, Urbanik et al discovered that the younger passengers represented more of a captive market than the elderly. Overall, they concluded the Texas intercity bus rider did not appear to be significantly different from those in other parts of the U.S;
they also suggested improvement in intercity bus service should be focused on safety, on-time performance and comfort based on respondent’s attitudes.
After the regulatory reform on the intercity bus industry, another study was conducted by Fitzpatrick et al (1996) to produce data necessary to define the status of the intercity bus industry in Texas in the 1990s. The study revealed that the number of communities served by the intercity bus in Texas decreased from 1,106 in 1970 to 596 in 1992. The study included a household survey mailed to Texas residents to gather data about demographics, information about intercity bus use and people’s attitude to the intercity bus system. Based on the responses of 545 households, it appears that the intercity travel users had similar characteristics to that in the early 1980s in spite of the service drop: intercity bus riders were generally a lower-income group and to visit friends or relatives; express bus service, better bus station locations, increase in air and train fares, as well as bus safety and comfort are major factors that affected people’s decision in taking intercity bus.
4.4. Application Example: Forecast Ridership for New Mode (HSR)
Decision in investment of HSR is an important concern of intercity transportation policy. The approaches used to forecast ridership of the HSR can be classified into two major groups. The first approach starts the forecast from projecting total travel and then uses the discrete choice model to determine the share, based on which predict the future HSR ridership. The second approach starts from projecting trips that would be made by each existing mode and then determining with separate mode choice model to shift to the new HSR mode (or other new mode) as a function of relative factors (Brand, et al., 1992; Peeta, et al., 2008). The argument for the second approach lies in that different mode users exhibit different behavior when confronted with the choice to use HSR.
Brand et al. (1992) applied the second approach to forecast the HSR ridership for the proposed Texas TGV. The forecasting process was divided into two steps. In the first step, the total trip volumes by each existing modes were estimated based on population in origin/destination cities, income of intercity travelers and the level of service of each mode between the origins and destinations. In the second step, the share of total trips by trip purpose (business and nonbusiness) that were expected to divert to HSR was estimated for each existing mode based on the comparison between the access, egress and line-haul time/cost and service frequency of the existing mode and those of HSR. The third step estimated induced travel by the new HSR service by incorporating the mode choice model utility function into the total demand model of the first step.
Because HSR does not exist in the U.S., it is not possible to use revealed preference (RP) data to estimate the HSR market share. Then, stated preference (SP) data are often utilized by researchers in studying a non-exist travel mode. The SP method was originally developed in marketing research in the early 1970s and received increasing attention in transport research since the 1980s (Kroes & Sheldon, 1988). Compared to RP method, the SP method has advantages of easier control of variables, more flexibility and capability to explore non-existing situation. Yet, it also has several limitations for travel demand modeling: respondents may respond differently in a real situation from what they indicate in the survey; they may not understand the explained scenarios in the survey, or too detailed explained scenarios will lead to a very long questionnaire (Peeta, et al., 2008). The uncertain reliability of SP data often leads to the question about the validity of the forecasting results only based on the SP data. Thus, to forecast HSR ridership, researchers often choose to combine RP data and SP data. In their HSR forecasting study, Morikawa et al (1991) and Yao & Morikawa (2005) developed RP and SP models separately and then combined RP/SP estimator by maximizing the joint log-likelihood function.
The Bay Area HSR ridership and revenue study (Cambridge Systematics Inc., 2006) was reviewed at the end of this report to serve as a practical case of studying intercity travel demand.
The HSR ridership forecasting model followed the framework of the California Statewide model, which contains four components: urban travel, interregional travel, external travel and trip assignment. This report will only focus on the interregional travel model component. The interregional trips were defined “as all trips with both ends in California and whose origin and destination are in different urban areas (or different counties outside the urban areas) having proposed HSR stations” (Cambridge Systematics Inc., 2006, pp. 1-3). The interregional trips were further segmented into short trips that were less than 100 miles and long trips that were longer than 100 miles.
In order to calibrate the models, intercept survey and household surveys were conducted to obtain RP and SP mode choice data from air, rail and auto trip passengers. The air passenger survey (1,234 samples) was conducted at six key airports throughout California. The rail passenger survey (430 samples) was conducted both as an on-board self-administered survey and as a telephone survey among qualified existing rail users. The auto passenger survey (1,508 samples) was conducted as a household telephone survey using a stratified sampling approach.
The interregional models consist of four model components: trip frequency, destination choice, main mode choice, and access/egress mode choice. The market segmentations in the models were defined by trip purposes as business, commute, recreation, and other, as well as trip length as short trip (less than 100 miles) and long trip (longer than 100 miles). The trip frequency model treated “person-day” as the decision unit and applied the MNL structure. The model results showed that intraregion accessibility, travelers’ living location, and travelers’ household characteristics including income, auto ownership, number of workers and household size all affected the frequency of trip making. The destination choice model also applied the MNL structure. The model considered travel impedance, distance, area type, region, location interaction, and the amount of activities that occurs at destinations, yet the location interaction variables were tested as insignificant.
The access/egress mode choice model and the main mode choice mode all applied the NL structure. The access/egress model produced probabilities that each access and egress mode will be chosen for each origin-destination pair given the specific transportation characteristics and demographic characteristics of that travelers (Cambridge Systematics Inc., 2006, pp. 3-27).
Driving including drive/park, drop off and rental car, and non-driving are the fist level of choices;
taxi, transit and walk/bike are nested in the non-driving option. The main mode choice model contains auto and no auto options while air, conventional rail and HSR are nested in the non-auto option. The final model results indicate a higher HSR share due to the attractiveness of the time and cost. It should be noted that the trip frequency, destination choice and mode choice models all utilize accessibility or impedance measure as inputs.
5.1. The 1995 American Travel Survey (ATS) The 1995 ATS represents the most current comprehensive survey on the long-distance travel of persons living in the United States. The ATS interviewed approximately 80,000 households beginning in April 1995 and ending in March 1996. Sample households were interviewed four times during this period at about three-month interval. The survey population consisted of persons resident in households and in group dwellings such as dormitories. The ATS data include basic social and economic characteristics of travelers including age, sex, education, income, etc, and detailed information about each trip including trip purpose, means of transportation, origin, destination, intermediate stops, travel dates and duration, number of nights away, and types of lodging used.