TRB ID: 01- 3271

Combined RP/SP Nested Logit Access/Mode Choice Model for Multiple Mass Transit Technologies

Amalia Polydoropoulou Department of Shipping, Trade, and Transport University of the Aegean Korai 2a, Chios, GREECE 82100 Tel/Fax: +30-1-895-0341 Email: [email protected]

Moshe Ben-Akiva Department of Civil and Environmental Engineering Massachusetts Institute of Technology 77 Massachusetts Avenue, Room 1-181 Cambridge, MA Tel: (617) 253-5324 Fax: (617) 253-0082 Email: [email protected]

TRB 2001 Annual Meeting

ABSTRACT

This paper presents a combined Stated Preference (SP) and Revealed Preference (RP) Nested Logit model of access and main mode choice developed for the Tel-Aviv Metropolitan area. The data used for model estimation include a conventional travel/activity survey and a customized and computerized SP survey. The model includes a generic Mass Transit alternative that represents any possible combination of existing and new technologies such as heavy rail, enhanced bus, light rail, and suburban rail. The lower level of the nested logit model represents the access choice to the bus or mass transit alternatives, including walk, park and ride, kiss and ride, and bus access alternatives; and the upper level represents the mode choice among bus, mass transit, car driver and car passenger alternatives. Travel time reliability and seat availability are among the attributes of the public transportation alternatives. Keywords: Mass Transit, Revealed and Stated Preferences, Nested Logit, Mode and Access Choice Model

Acknowledgments: This paper is based on an HCG study performed for NTA, the organization charged with the planning and development of a mass transit system for the Tel-Aviv region. The authors benefited greatly from contributions by Nick Cohn, Marcos Seinuk, Shlomo Bekhor, Amir Freund, Yoram Shiftan, Yossi Prashker and several others at NTA and their technical committees.

INTRODUCTION

The mode choice models developed in this study are suitable for testing travelers choices among bus and car alternatives, and mass transit services that may include technologies such as light rail, metro, enhanced bus, suburban rail systems, as well as combinations of those. The main mode and access choice models were estimated simultaneously. Alternative access modes to public transportation included walk, park and ride, kiss and ride, and bus access to mass transit. The models have three unique features: (1) Joint stated preferences (SP) and revealed preferences (RP) estimation of a nested logit model structure. SP data are often collected to study travelers' acceptance of new modes. It captures the mode choice behavior of travelers under hypothetical situations where new modes are present. However, SP data have inherent biases stemming, among other things, from the respondents lack of familiarity with the new modes. A combination of RP and SP data is used in model estimation to overcome these potential biases and to provide more reliable estimates (Ben-Akiva and Morikawa 1990a and 1990b; Hensher and Bradley, 1993; BenAkiva et al. 1994; Ben-Akiva and Morikawa, 1997; Bradley and Daly, 1997; Kuniaku et al., 1999). The present paper contributes to the growing literature on this topic by extending previous work with multinomial logit models to the combined estimation of nested logit models. The approach developed in this paper explicitly tests differences among structural parameters of the SP and RP nested logit models. (2) Generic mass transit alternative for all the possible combinations of new and existing technologies. Mode choice models usually consider the choice between bus and rail transit. The model developed in this paper extends conventional models by representing travelers preferences towards multiple technologies and combinations of technologies. The generic mass transit alternative includes explicit dummy variables of the technologies as additional attributes. The inclusion of such mass transit alternative required the development of a new SP survey. (3) Successful inclusion of travel time reliability and probability of finding a seat among the attributes of the public transportation alternatives. The importance of these attributes was demonstrated in previous studies by Prashker, 1978a and 1978b; Ben-Akiva et al. 1996; and Benjamin et al., 1999. The models were developed for the Tel-Aviv metropolitan area and are being employed in feasibility and planning studies of a new mass transit system. The collection of the SP data was undertaken during the Fall of 1999 using the HCG MINT computer system. 10 different combinations of mass transit technologies were considered in the SP experiments. The RP data come from the 1995 National Travel Habits Survey (NTHS), a conventional travel/activity survey. Separate models were estimated for five trip purposes. In this paper we present results for the overall model estimated with all the available data. Section 2 describes the model design. Section 3 presents the data collection methodology. Section 4 presents descriptive statistics. Section 5 discusses the model estimation methodology and results. Section 6 discusses the application methodology; and Section 7 concludes the paper.

MODEL DESIGN

Figure 1, presents the structure of the Nested Logit model of main and access mode choice. In this model structure, the usage of a generic mass transit (MT) alternative allowed the estimation of model parameters for different combinations of alternative technologies. There are 4 MT technologies (metro, lrt, suburban rail and enhanced bus) and 10 possible MT concepts that involve single or combinations of technologies (four concepts with a single MT technology and 6 combinations of two MT technologies). These 10 MT concepts were considered to be the reasonable combinations of mass transit technologies that could be implemented in the study area.

Figure 1: SP Main Mode/Access Choice Model Design

Public Transport

Private Means

Bus

Mass Transit

Car Driver

Car Passenger

Walk K&R P&R Access Access Access

Bus Walk K&R P&R Access Access Access Access

The SP questionnaire introduced each respondent to only one mass transit concept and then presented multiple choice experiments with a choice among mass transit, bus and (if a car is available) car. The mass transit concept was Origin-Destination specific. In the computer-aided survey, an availability matrix indicated the reasonable mass transit concepts that could be available for every O-D pair. The individual had to choose among car, bus and the mass transit concept that was randomly generated. For the application of the model system, the network assignment model determines the attributes of the mass transit alternative including the combination of technologies that serves an O-D pair. In order to account for travelers perceptions on travel time reliability and of seat availability, two variables were included in the SP experiments, defined as follows: 1. Probability (or the number of times out of 10 trips) of a travel time delay greater than 10 minutes; and

2.

Probability (or the number of times out of 10 trips) of finding a seat when boarding a bus or a MT vehicle.

DATA COLLECTION METHODOLOGY

The key aspects of the data collection effort are the SP questionnaire and the sampling strategy.

3.1

Questionnaire Design

The questionnaire included five parts:

(1) Household and Respondent Characteristics, and Trip Qualification The first part briefly explained the survey and included questions on (1) household composition; (2) respondent characteristics, such as car availability and availability of Employer Provided Transportation (EPT); and (3) a simple trip diary used to select a recent trip for the remainder of the questionnaire.

(2) Attributes of Selected Trip The second part included questions on the attributes of the selected trip. These responses were used in the SP experimental design to generate the levels of the attributes for the choice experiments.

(3) Presentation of the MT Mode Based on the origin-destination pair of the selected trip, the SP questionnaire program (MINT) randomly selected one mass transit system concept. Then, a PowerPoint presentation on the selected MT system concept was shown. The aim of this presentation was (1) to explain in detail the technologies that the MT concept involved, (2) to present pictures of the vehicles and stations, and (3) to answer any questions the respondent may have about the MT service. Thus, it was assured that the respondent understood the new mode and was able to provide reliable responses to the SP experiments. Efforts were made to make the descriptions of the MT concepts objective; showing MT services that some respondents may have already seen in other cities. Every respondent was presented with only one MT system description and the selected mass transit concept was included in all the choice experiments presented to a respondent. The attributes of the alternatives changed among experiments but the mass transit mode did not. The mass transit mode varied across respondents in order to obtain respondents' choices for all the 10 concepts that were included in the experimental design.

(4) Stated Preference Experiments An SP choice experiment consisted of the following sequence of choices: 1. choice of access mode to bus, among walk, park and ride, kiss and ride; 2. choice of access mode to mass transit, among walk, park and ride, kiss and ride, and bus; 3. main mode choice among bus, mass transit and (if car available) car alternatives (car driver or car passenger); and 4. if Employer Provided Transportation (EPT) was available, choice between the preferred main mode alternative and EPT. Six (6) SP experiments were administered to each respondent, varying the attribute levels of the access and main mode alternatives. Table 1 presents the attributes of the main mode and access mode alternatives.

Table Error! Unknown switch argument.: Attributes of Main /Access Mode Alternatives Attributes of Main Mode Alternatives Bus and Mass Transit Concept Car Wait time Driving time Transfer time Fuel price Time riding the vehicle Parking cost Number of transfers Parking search time Fare Walk time to/from car Probability of delay Probability of getting a seat Attributes of Access Mode Alternatives Access Mode to Bus Access Mode to Mass Transit Walk time Walk time Drive time Drive time Parking cost Parking cost Parking search time Parking search time Walk time to/from car Walk time to/from car OR BUS

Figure 1 presents an example of the structure of a main mode choice experiment, as it was presented on the computer screen. The levels of the attributes presented to each respondent were randomly generated within the MINT program.

(5)

Respondent Characteristics

The last part of questionnaire included questions about the respondents socio-economic characteristics.

Figure 1: Example Choice between Bus, Mass Transit System, and Car Alternatives

3.2

Sampling

The sampling strategy was based on a geographic stratification and on households car availability status (0, 1, and 2+ cars). The geographic stratification aimed at areas that had a higher expected number of transit trips and were expected to be served by the mass transit system under consideration. Therefore, as indicated in Table 2, 998 out of 1830 surveys involved a recent trip made by public transportation. The computerized interviews took place at the respondents home. For a sampled household all adult members (16 or older) were recruited as respondents. A small gift was given to the respondents by the fieldworker at the end of the interview. More than 70% of the interviewees received a call back for verification. During the period of the survey in 1999, there was is a very limited commuter rail service in the Tel-Aviv area. Therefore the stratified household sample contained very few rail users. Therefore, the sample was enriched with 150 interviews of intercepted rail users. The rail sample was distributed across 5 boarding stations, and 3 car availability classes.

A total of 1830 valid surveys were collected. Table 2 presents summary statistics of the major variables related to the sampling plan. The sample collected was representative of the population in terms of gender and employment status. Table 2: Summary SP Statistics Gender Male - 1 Female 2 Transit Trips No 0 Yes 1 Car Ownership 0 car 1 car 2+ cars Employment No 0 Yes 1 Purpose of Trip Home-Based Work (HBW) - 1 Home-Based Education (HBE) - 2 Home-Based Shopping (HBS) - 3 Home-Based Other (HBO) 4 Non Home-Based NHB - 5 Total Count 883 947 Count 832 998 Count 698 776 356 Count 614 1216 Count 603 184 280 445 318 1830

DESCRIPTIVE STATISTICS

Descriptive statistics of some key data items for the SP and RP surveys are presented below.

4.1

SP Survey

The SP survey has a total of 10,980 choice experiments, which correspond to six choice experiments per respondent. The mass transit (MT) mode assignment to the respondents is presented in Table 3. Note that the mass transit concepts were evenly distributed among respondents.

Table 3: Percent of Observations per Mass Transit Concept Mass Transit Concept 1=Light Rail Train (LRT) 2=Suburban Rail (SR) 3=Metro (M) 4=Enhanced Bus (EB) 5=LRT/SR 6=LRT/M 7=LRT/EB 8=SR/EB 9=SR/M 10=M/EB Total Percent 9 11 10 10 9 12 10 10 10 9 100%

In the estimation process some choice experiments were rejected, based on incomplete data. The number of observations used for the estimations was 10,867. Table 4 presents the number of observations that chose each alternative and the number of observations that had available this alternative, for the overall sample.

Table 4: SP Data- Mode Choice and Choice Set Availability Alternative Alternative 1: Bus Walk Observations 3411 10867 1438 10867 97 4383 2496 10867 976 10867 89 4383 424 10867 1433 4383 503 5657

Chosen Available Alternative 2: Bus K&R Chosen Available Alternative 3: Bus P&R Chosen Available Alternative 4: MT Walk Chosen Available Alternative 5: MT K&R Chosen Available Alternative 6: MT P&R Chosen Available Alternative 7: MT Bus Chosen Available Alternative 8: Car Driver Chosen Available Alternative 9: Car Passenger Chosen Available

4.2

RP Survey

The RP data used in this study comes from the Israeli National Travel Habit Survey (NTHS) conducted during 1995. This is a conventional travel/activity diary. The RP estimation data set was prepared to be compatible with the SP data and the model structure presented in Figure 1. However, the RP data set included too few K&R and P&R observations. Moreover, the available EMME/2 network model provided level of service data only for the bus network (with walk access) and the car network. Therefore, only a trinomial model, with the following alternatives: (1) Bus with walk access, (2) Car driver, and (3) Car passenger, could be estimated from the RP data. Table 5 presents the number of observations that chose each alternative and the number of observations that had available this alternative, for the overall RP sample. Note that the final number of observations used for model estimations was slightly less than the original data provided due to illogical or missing data.

Table 5: RP Data - Mode Choice and Availability for All Trip Purposes Alternative Alternative 1: Bus Walk Observations 10239 42189 22783 28599 9167 42189

Chosen Available Alternative 2: Car Driver Chosen Available Alternative 3: Car Passenger Chosen Available

COMBINED RP/SP ESTIMATION: METHODOLOGY AND RESULTS

In the combined RP/SP model estimation, the RP and SP data set are pooled together. Two separate models are specified for the RP and SP data. The SP part is estimated using the SP data set. However, the utility functions share some common coefficients. This allows for the trade-offs among the basic modal attributes to be the same in the RP and SP models. Therefore, this methodology has the following major advantages: 1) it provides more reliable estimation results, since the RP data which represents the actual decision making of the individuals counter acts the SP-related biases; and 2) it provides the capability to estimate the demand for new mass transit technologies.

This combined RP/SP estimation methodology and the estimation results are presented below.

10

5.1

Combined RP / SP Model Structure

In the models developed, the dependent variable was the joint choice of main and access modes. The SP model includes nine alternatives: Alternative 1: bus with walk access Alternative 2: bus with kiss and ride (k&r) Alternative 3: bus with park and ride (p&r) Alternative 4: mass transit (mt) with walk access Alternative 5: mt with k&r Alternative 6: mt with p&r Alternative 7: mt with bus access Alternative 8: car driver Alternative 9: car passenger The RP model includes three alternatives: Alternative 1: bus with walk access Alternative 2: car driver Alternative 3: car passenger Figure 2 presents the tree structure of the combined SP/RP estimator consisting of two sub-trees. The SP sub-tree on the right side of the figure is augmented with 9 dummy nodes at the bottom. These nine dummy nodes permit us to include a different scale for the SP model. The structure of the model on the RP side is the same as the SP model structure, except that 6 of the 9 alternatives are unavailable in the RP data. Thus, when omitting these alternatives we have an RP sub-tree that appears to have redundant nodes. The structural parameters of these nodes are unidentified in a separate RP estimation but are estimable in the combined RP/SP estimation. In the combined estimation, the intermediate nodes are kept and their structural, or LOGSUM parameters are identified by the SP data. Thus, in the RP sub-tree there are two nodes above the bus with walk alternative. The combined RP/SP estimation includes 3 LOGSUM coefficients. The first is the scale of the SP model. For level one the structural parameter of RP is equal to the structural parameter of SP because choices at this level only exist in the SP data. For level 2 the log sum coefficient of the RP model is different from that of the SP model. This RP coefficient exceeded 1, and therefore was constraint to be equal to 1 in the final model. The Nested Logit Structure presented above was the outcome of several model runs and tests of model structure. Initial IIA tests of estimation results of the SP model showed that an MNL structure was not appropriate. In the combined estimation, the utility functions of the RP and SP models share some common coefficients, implying that the trade-offs (or marginal rates of substitution) among some of the attributes, are the same in the RP and SP data. Other coefficients are specific to either the RP or to the SP utility functions. The relative levels of randomness in the RP and SP utilities is captured by the SP scale parameter. One of the advantages of the modeling methodology is that it employs of-the-shelve software tools both for data collection (MINT) and for model estimation (ALOGIT). Thus the methodology is readily available to both researchers and practitioners that would like to estimate

11

combined SP/RP mode choice models. The challenging part of the modeling task is the specification and testing of alternative model structures and utility functions.

Figure 2: Combined SP/RP ALOGIT Structure

33 RP alternatives SP alternatives

Level 2 29 Public Transport Level 1 22 23

30 Private Means

31 Public Transport 27

32 Private Means

24

25 Bus

26 MT

28

13 1 Bus W 2 CD 3 CP

14

15

16

17

18 19

20

21 SP Scale

4 W

5 K & R

6 P & R

7 W

8 K & R

9 10 P B & R

11 CD

12 CP

5.2

Model Estimation Results

Selective model estimation results using all the available data are presented in Table 5. Note that the total number of estimated coefficients for that model is 52. The table includes the coefficient names, the definition of the corresponding variables, and the model (RP, SP, or RP/SP) for which each coefficient is specified. The important properties of the estimated coefficients presented in Table 6 are: (1) Nine mass transit concept constants, all negative relative to the base case, which is metro. This shows that, everything else being equal, metro is the most preferred MT concept. The

12

(2)

(3) (4)

(5) (6)

(7)

(8)

(9)

enhanced bus concept is the least attractive MT concept based on the negative magnitude of its estimated coefficient. The number of transfers variable is generic among the bus and mass transit alternatives: (a) Number of transfers for the RP model (transfrp). This coefficient is negative and very significant as expected; (b) Number of transfers for the SP model (transfsp). The coefficient is negative but statistically insignificant as demonstrated by the t-stat (less than |2|). This variable captures an extra penalty reflecting the inconvenience and uncertainty associated with a transfer. The results indicate significant differences between the RP and SP data sets. Clearly for existing public transport systems, such as bus, the inconvenience of a transfer is very high. However, for the new mass transit systems, transfer inconvenience is expected to be much less, due to the implementation of new technologies and the better service coordination expected at MT stations. The public transport in-vehicle time (invt) is generic for the public transport alternatives and common to both RP and SP models. The coefficient is negative and significant as expected. The out-of-vehicle travel time has a piece-wise linear specification. This accounts for different sensitivity to out-of-vehicle time of less than 6 minutes and of more than 6 minutes. The out-of-vehicle-time includes transfer walk, wait time, and egress time. Ovt1 and ovt2 are generic for the public transport alternatives and common among the RP and SP models. These coefficients are negative as expected. Note that the magnitude of coefficient ovt2 is less than ovt1, meaning that individuals are more sensitive to an increase of out-of-vehicle time when it is less than 6 minutes and less sensitive to an increase of out-of-vehicle time when it is greater than 6 minutes. Cost of parking (cstpark). The variable is specific to the car driver and car passenger alternatives, for both the RP and SP models. Its coefficient is negative and significant. The coefficients of fare, and car in-vehicle travel time vary among the two segments of 0 car and 1+ car available households. (a) Public transport fare for 0 car (fare0) and 1+car (fare1). These fare coefficients are generic among bus and mass transit alternatives and common to both RP and SP models. In general car owners are less sensitive to fare than non-car owners. (b) Car in-vehicle time for car driver for 1+ car ownership (ctimecd1). This variable is specific to the car driver RP and SP alternatives. It is negative and highly significant. (c) Car in-vehicle time for car passenger for 0 car (ctimecp0) and 1+ car (Ctimecp1). The coefficients are specific to the RP and SP car passenger alternatives. They are negative and significant. Passengers who have car available are more sensitive to in-vehicle time than passenger with car unavailable. The coefficient ctimecp1 is greater than the coefficient ctimecp0. There are several access-related coefficients, such as public transport walk access time (accwlkt); public transport access drive time (accdrvt); and public transport access parking cost (accprkc). All coefficients are negative and significant. Probability of delay later than 10 minutes (delay). Delay is generic among the public transportation alternatives. The negative sign is expected and is statistically very significant. The inclusion of this variable provides the opportunity to test individuals responses to various levels of reliability of public transportation, a factor acknowledged by researchers and practitioners to have a significant effect on mode choice. Probability of finding a seat (seat). The variable is generic among SP bus and mass transit alternatives. The coefficient is positive and very significant as expected. This variable allows the users of this model to test how different levels of services affect the ridership, and decide on the appropriate level of service to offer (assisting decisions about service frequency and fleet management.)

13

Table 5: Combined RP/SP Model Selected Estimation Results

Coefficient Name (Specific Alternative) (2) Lrt: Constant for light rail train SP concept (Alt. 7,8,9,10) Subrail: Constant for Suburban rail SP concept (Alt. 7,8,9,10) Enhbus: Constant for enhanced bus SP concept (Alt. 7,8,9,10) Lrtsubrail: Constant for light rail and suburban rail SP concept (Alt. 7,8,9,10) Lrtmetro: Constant for light rail and metro SP concept (Alt. 7,8,9,10) Lrtenhbus: Constant for light rail and enhanced bus SP concept (Alt. 7,8,9,10) Subrenhbus: Constant for suburban rail and enhanced bus SP concept (Alt. 7,8,9,10) 8 Subrmetro: Constant for suburban rail and metro SP concept (Alt. 7,8,9,10) 9 Metrenhubs: Constant for metro and enhanced bus SP concept (Alt. 7,8,9,10) 10 Transfrp: Number of transfers for RP (Alt.1) 11 Transfsp: Number of transfers for SP (Alt. 4,5,6,7,8,9,10) 12 Invt: Public transport in-vehicle time RP/SP (Alt. 1,4,5,6,7,8,9,10) 13 ovt1: Public Transport: Piece-wise linear out-of vehicle time <=6 minutes RP/SP (Alt. 1,4,5,6,7,8,9,10) 14 ovt2: Public Transport: Piece-wise linear out-of vehicle time > 6 minutes RP/SP (Alt. 1,4,5, 6,7,8,9,10) 15 Cstpark: Cost of parking RP/SP (Alt. 2,3,11,12) Public Transport Fare for 0 car ownership RP/SP 16 Fare0: (Alt.1,4,5,6,7,8,9,10) RP/SP 17 Fare1: Public Transport Fare for 1+ car ownership (Alt.1,4,5,6,7,8,9,10) 18 ctimecd1: In-vehicle time for car driver RP/SP 1+ car own (Alt. 2,11) 19 Ctimecp0: In-vehicle time for car pass 0 car own RP/SP (Alt.3,12) 20 ctimecp1: In-vehicle time for car pass 1+ car own RP/SP (Alt. 3,12) 21 Accwlkt: Access walk time for RP and SP (Alt. 1,4,9) 22 Accdrvt: Access Drive time for SP (Alt.5,6,8,9) 23 Accprkc: Access parking cost for RP and SP (Alt.3,6) 24 Delay: Probability of delay SP (Alt.4,5,6,7,8,9,10) 25 Seat: Probability of Seat Availability SP (Alt.4,5,6,7,8,9,10) 26 Scalesp 27 lglv2sp: Level 1 logsum coefficient for SP/RP 28 lglv2sp: Level 2 logsum coefficient for SP 29 lglv2sp: Level 2 logsum coefficient for RP Summary Statistics Observations Likelihood Coefficients
5KR  5KR F

No (1) 1 2 3 4 5 6 7

Estimation (3) -0,1629 (-2,4) -0,3244 (-4,6) -0,3667 (-5,1) -0,2680 (-3,8) -0,2115 (-3,2) -0,1640 (-2,5) -0,1607 (-2,4) -0,1081 (-1,6) -0,2231 (-3,3) -0,4691 (-8,9) -0,03429 (-1,4) -0,03810 (-13,9) -0,05025 (-5,2) -0,04014(-11,2) -2,883e-4(-3,4) -0,001149(-13,4) -8,971e-4 (-11,8) -0,05697(-11,8) -0,01720 (-2,7) -0,04525 (-9,3) -0,07202(-17,3) -0,03729(-10,9) -4,364e-4(-5,1) -0,01202 (-3,4) 0,04181 (7,8) 3,001 (22,7) 0,5349 (19,7) 0,5211 (16,6) 1

53056 -40992,4 52 0.33 0.20

14

(10) The SP scale coefficient (Scalesp) is around 3, indicating that the RP data contains more random noise than SP data. The scale coefficient is included in the combined RP/SP estimation in order to account for the difference in the variance of the error terms and to allow the combination of the two sets of data. (11) Level 1 logsum coefficient (lglv1b). This logsum coefficient is common between RP and SP and corresponds to the access choice of the public transport alternatives. (12) Level 2 RP logsum coefficient (lglv2rp). This logsum coefficient is specific to the RP alternatives and corresponds to the choice between the main mode alternatives. This coefficient is almost equal to 1, as expected from the separate RP estimation results. Note that the lglv2rp with value greater than 1, in previous runs was constrained to be equal to 1. (13) Level 2 SP logsum coefficient (lglv2rp). This logsum coefficient is specific to the SP alternatives and corresponds to the choice between the main mode alternatives. For the estimated model this coefficient takes the value 0.5. The logsum coefficients at the main mode levels were allowed to differ between the RP and the SP data. This was based on the separate RP and SP estimation results that indicated that this logsum coefficient is about one for the RP data and significantly less than one for the SP data. Hence, the joint SP/RP estimation includes separate main mode logsum coefficients for SP and RP. The RP coefficient is used for the application model. The estimation results demonstrate the viability of the combined SP/RP nested logit model. Particularly noteworthy are the significant differences among the nested logit structural parameters between the SP and the RP data. Therefore, the combination of the two data sets provided significant insights that the separate data sets would not have captured. The estimated coefficients of the variables related to the reliability of travel time and the travelers perceptions of delays of public transportation modes and the probability of finding a seat available when boarding, were found to be highly significant. Thus the model is sensitive to a rich set of level of service attributes and is therefore suitable for an analysis of the ridership and users benefits of alternative mass transit system.

MODEL APPLICATION

The application model is a three-level nested logit model with the same structure as the one shown in Figure 1. The availability of the alternative private modes depends on auto availability. The availability of the transit modes also depends on the location of transit facilities and services as coded in the network model. Thus, the availability of the transit modes varies among O-D pairs according to the mass transit scheme being evaluated. It is the role of the network model to determine which of alternative modes are available for a given O-D pair. The mass transit concept type is an attribute of the mass transit alternative. Hence, for any given mass transit network, one of the inputs to the demand model is a matrix that specifies the mass transit concept, or type, for every OD pair. This matrix needs to be generated from the network as a function of the link types forming the mass transit path. Only one MT concept may be assigned to an OD pair. For a given OD pair the model may contain up to 4 MT alternatives for the 4 possible access modes (walk, k&r, p&r and bus). These MT alternatives for a given OD pair have the same MT concept.

15

The EMME/2 network model is used to find the best MT path for a given OD and a given access mode. Thus, for a given OD there are up to 4 MT paths. In other words, the nested logit model is being applied together with the transit path-finding model. The implementation of the application model required the development of an algorithm that transformed the EMME/2 network into a hyper-network containing 600 traffic zones and up to 200 mass transit stations (in future years). This algorithm identified the best mass transit combination that served each origin-destination pair. This required the evaluation of alternative mass transit paths between each origin, boarding station, disembarking station, and final destination. This model system provides a useful policy analysis and forecasting tool for predicting travelers mode choices in the Tel-Aviv area. It is currently being used to evaluate the demand of alternative mass transit and pricing scenarios.

CONCLUSION

This paper provided an overview of a methodology to estimate a Nested Logit model using both RP and SP data. The use of the generic mass transit concept and a customized SP questionnaire proved to be successful in terms of estimating different coefficients for alternative mass transit modes and combinations of new mass transit technologies, not currently available in the study area. The model may also be used to conduct sensitivity/policy analysis and identify the preferences, price elasticity, and value of time of different types of users. The results demonstrated the necessity of the nested logit structure, as well as the differences among the RP and SP data sets, and the advantages of simultaneously estimating models using different data sets and sharing common coefficients.

REFERENCES

Ben-Akiva, M., J. Benjamin, G. Lauprete, and A. Polydoropoulou (1996). Evaluation of Advanced Public Transportation Systems (APTS) Impact on Travel by Dial-a-Ride. Transportation Research Record, No. 1557. Ben-Akiva, M., N. Bradley, T. Morikawa, J. Benjamin, T. Novak, H. Oppewal, and V. Rao (1994). Combining Revealed and Stated Preferences Data. Marketing Letters 5, 4: pp. 335-350. Ben-Akiva, M. and S. Lerman (1985). Discrete Choice Analysis: Theory and Application to Travel Demand. The MIT Press. Cambridge, MA. Ben-Akiva, M. and T. Morikawa (1990a). Estimation of Switching Models from Revealed Preferences and Stated Intentions. Transportation Research A, 24A(6), 485-495. Ben-Akiva, M. and T. Morikawa (1990b). Estimation of Travel Demand Models from Multiple Data Sources. In M. Koshi (ed.), Transportation and Traffic Theory. New York: Elsevier, pp. 461-476.

16

Ben-Akiva, M., Morrikawa, T., 1997. Methods to estimate discrete choice models with stated and revealed preferences. Paper presented at 1997 NSF Symposium on Eliciting Preferences, Berkeley, California, July 1997. Benjamin, J., S. Kurauchi T. Morikawa, A. Polydoropoulou, K. Sasaki, and M. Ben-Akiva (1999). Forecasting Paratransit Ridership Using Discrete Choice Models With Explicit Consideration of Availability. Transportation Research Record 1618, pp. 60-65. Bradley, M. A. and Daly, A. J. (1997) Estimation of logit choice models using mixed stated preference and revealed preference information, Stopher, P.R. and Lee-Gosselin, M. (Eds.) Understanding Travel Behaviour in an Era of Change, Pergamon, Oxford, 209-232. Hensher, D.A. and Bradley, M. (1993) Using stated response data to revealed preference discrete choice models. Marketing Letters, 4(2), 139-152. Kuniaki Sasaki, Taka Morikawa, Shogo Kawakami (1999). A Discrete Choice Model with Taste Heterogeneity Using SP, RP and Attribute Importance Ratings. Proc. of 8th World Conf. on Transport Research, Vol.3, 39-49, Elsevier. Prashker, J. (1979a) Mode Choice Models with Perceived Reliability Measures. Transportation Engineering Journal. 105 TE3: pp. 251-262. Prashker, J. (1979b) Scaling Perceptions of Reliability of Urban Travel Modes Using INDSCAL and Factor Analysis Methods. Transportation Research A 13: pp .203-212

17