Scholarly article on topic 'Policy Analysis for New Commuter Rail and Road Pricing Alternatives Using an SP Survey in Abidjan'

Policy Analysis for New Commuter Rail and Road Pricing Alternatives Using an SP Survey in Abidjan Academic research paper on "Civil engineering"

CC BY-NC-ND
0
0
Share paper
Academic journal
Transportation Research Procedia
OECD Field of science
Keywords
{"Joint Choice Model" / "Stated Preference" / "Revealed Preference" / "Côte d’Ivoire" / Abidjan}

Abstract of research paper on Civil engineering, author of scientific article — Sadayuki Yagi, Hideo Shiraishi

Abstract In Abidjan, a road pricing scheme is currently under review as a transportation control measure along with operation of a new commuter rail system. While this scheme may be effective for congestion reduction in central business district (CBD), provision of alternative means of transportation for the “pushed-out” auto users is of great importance to obtain public acceptance. Hence, it is necessary to simulate simultaneously the road pricing scheme and the commuter rail development which may serve as an alternative for assumed pushed-out auto users. Utilizing data from the available opinion survey, this paper studies how commuter rail and auto ridership are likely to vary as a function of travelers and system attributes. Additionally, the study attempts to evaluate the way this new travel mode is distinguished from other existing conventional transportation alternatives in Abidjan. The survey data contains socioeconomic information of over 4,000 respondents as well as details of to-work/school or to-other trips to CBD including the mode, travel cost, time, etc. Respondents were then asked about their willingness to shift from their current mode to commuter rail to make the same travel for different commuter rail fare levels. Modeling efforts suggest that a mixed logit model performs better in explaining choice behavior. Therefore, this model was used for policy simulation. The simulation results brought about many implications as to the tested policies. While the developed model may be applied only to future commuter rail corridors in which the survey was conducted, it captures the key variables that are significant to explain mode choice behavior and presents great potential for practical use in policy simulation and analysis in a large metropolitan area of the developing world.

Academic research paper on topic "Policy Analysis for New Commuter Rail and Road Pricing Alternatives Using an SP Survey in Abidjan"

Available online; at mwv.sciencedirect.com

Transportation

ScienceDirect Procedía

Transportation Research Procedía 25C (2017) 2524-2539 ■ ■ w «J «J

www.elsevier.com/locate/procedia

World Conference on Transport Research - WCTR dC16 Shanghai. 10-15 July dC16

Policy Analysis for New Commuter Rail and Road Prising Alternatives Using an SP Survey in Abidjan

Sadayuki Yagi a, Hideo Shiraishi b

i ALMEC Corporation, 5-3 Shi1juku 5-ctmome Shinjuku-ku Tokyo 160-0022, Japan b Oriental Consultants Global Co., Ltd., 12-1 Honmachi 3-chome Shibuya-ku Tokyo 151-0071, Japan

Abstract

In Abidjan, a road pric ing scheme is currently under review as a transportation control measure along with operation of a new commuter rail system. While this scheme may toe; effective for congestion reduction in central business districI (CBD), provi^on of al ternatiae moans of transportation uor the "pushed-oot" auto usere is of g reat importance to obtpin public acceptynce.Hence, a is neceesary to simul ate simultaneously the road pricing scheme and ahe commuter rail deaelopmept which may serve as an alternative oor assumed pushed- out auto users. Utilizing data from the avail able opinion survey, this paper studies how commuter rail and ahto ridership are like ly to vary as a function of travelers and system attributes. Additionally, the study attempts to e valuate the way this new tr^vol mode is distinguished from other exiting conventional transportation alternatives in Abidjan . The survey data contains socioe conomic information of ovor 4,000 respondnnts as well as details of to-wo rk/school or to-other trips to CBD including the modei travel cost, time, etc. Respondents were then asked about their willingness to shift from their xotrent mede to eommuter rail to meXp thn same travel for aiOferent eommutef rail fare levels. Modelino effosts suggest that a mmed ^git model berforms better in explaining chosce behavior. Therefore, this model was used for policy simoiation. The simulation results brought about many implications as to the tested pelicies. While the developed model may be applied only So future commuter rnil eorfidofs in which the survey was co nducted, it captures the key variables tnat are si gnificant to expeai n mcd e eVoice behavior and presents great potential for practical use in policy simulation and analysis in a large metropolitan area of the developing world.

© 2017 Tte Airttore. PiMidreU by EkCTkr B.V.

Prer-re^w imdCT rafpoofsbslsSУ of WORLD CONFERENCE ON TRANSPORT RESEARCH SOCIETY.

Keywords: Joint Choice Model; Stated Preference, Revealed Preference; Cote U'lvosra; Abidjan

1. Introduction

A master plan study called "The Project for the Development of the Urban Master Plan in Greater Abidjan" was conUucteU by Japan Ioternpts onal Cooperation Agency (JICA (2014)) in Greater Abidjan from March 201 3 to

2352-1465 © 2017 The Authors. Published by Elsevier B.V.

Peer-review under responsibility of WORLD CONFERENCE ON TRANSPORT RESEARCH SOCIETY. 10.1016/j.trpro.2017.05.285

December 2014. One of the overall objectives was to elaborate the urban transport master plan up to 2030 through identification of possible policy measures and solutions to develop sustainable transportation system in Greater Abidjan with a focus on encouraging public transport usage and improving mobility of people. As such, detailed transportation surveys and analyses were undertaken to prepare a comprehensive long-term transportation plan with the objective to develop and calibrate disaggregate travel demand models to simulate present and future interactions between land-use and transportation in the region.

The Household Travel Survey (HTS), among a variety of the surveys conducted, provides the largest and most comprehensive travel data in the region. The dataset covers as many as 20,000 households which correspond to 2% of the entire population, and provides daily travel patterns and detailed information on household socio-demographic characteristics. The ultimate goal of the authors' research on Abidjan data is to develop a novel activity-based microsimulation modeling system to test different transportation policy scenarios for Greater Abidjan. The study will simulate the way individuals schedule their daily activities and travel in an urban region of the developing world. It is hoped that the proposed new models will contribute to the improvement of the emerging travel demand forecasting techniques and provide an opportunity to evaluate urban transportation policy scenarios.

A Stated Preference opinion survey on transport system (SP survey) was also conducted in Abidjan to obtain the information regarding people's preference among the existing and future transport modes under different conditions and policies such as costs and service levels. Utilizing the SP data, the main purpose of this study was to determine how demand for a new commuter rail is likely to vary as a function of attributes that distinguish this new travel mode from other existing conventional alternatives in the region. In particular, the study explores several methods of joint modeling such as multinomial, nested, and mixed logit discrete choice models to predict the mode choice, using both RP and SP data and takes a further step into simulation analysis of some policies using the most appropriate mode choice model.

1.1. Stated preference data

Stated preferences (SP) refers to a wide array of possible ways of asking consumers about preferences, choices, ways of using options, frequencies of use, and so forth, while revealed preferences (RP) are associated only with actual choices (Louviere and Street (2000)). While the HTS has abundant RP data, the results of the SP survey contain both RP and SP data that can be used to establish discrete mode choice models for future forecasting.

An SP survey was conducted in the targeted quartiers along the future commuter rail corridor (Anyama - Abobo -Port Bouet), and interviewed transit users and auto users with regard to their preference of the commuter rail to the existing mode under different conditions. The area covered by the SP survey is shown in Figure 1. The survey targeted at the to-work, to-school, and to-other trips with zones along the planned commuter rail corridors (green zones in the communes of Abobo and Anyama) as the origin and with zones in downtown (i.e., communes of Plateau and Adjamé) as the destination. The survey was conducted by interviewing the persons who actually made trips from home to downtown regularly at least once a month.

The SP survey first asked information on the socioeconomic details of the respondents and their households. Then, as the current travel behavior constitutes the RP data, the survey asked the respondents about details of their trips to downtown that they made including the details such as mode(s) used, travel cost, and time. For the SP part, since there was no commuter rail in operation yet when the survey was conducted, detailed explanation and images of the planned commuter rail were presented to the respondents. Then, the survey asked respondents about their potential responses to different fare levels of the planned commuter rail. A total of six fare levels were prepared to ask the respondents whether they would be willing to shift from their current mode to the commuter rail to make the same travel. With regard to the trips that fulfill the above OD-zone criteria, effort was made to collect the samples so that the purpose and mode compositions would comply with those of the HTS database. In this way, the entire OD-zone pairs included in Figure 1 would be applied for mode choice modeling and policy simulation without using weights to the RP alternatives.

1.2. Modeling structure

Random utility based discrete choice models have found their ways in many disciplines including transportation, marketing, and other fields. Multinomial logit (MNL) model is the most popular form of discrete choice model in practical applications (Mohammadian and Doherty (2005)). It is based on several simplifying assumptions such as independent and identical Gumbel distribution (IID) of random components of the utilities and the absence of heteroscedasticity and autocorrelation in the model. As such, the MNL model belongs to a class of models that possesses the so-called independence of irrelevant alternatives (IIA) property, which is both one of the strengths of the MNL model and its major weakness (Meyer and Miller (2001)). Nested logit (NL) model is the model that has been developed in order to overcome the IIA limitation in the MNL model by modifying the choice structure into multiple tiers. NL models are very commonly used for modeling mode choice. The NL model permits covariance in random components among nests of alternatives. Alternatives in a nest exhibit an identical degree of increased sensitivity relative to alternatives in the nest (Williams (1977); McFadden (1978); Daly and Zachary (1987)). Each nest in the NL model has associated with it a log-sum parameter, or expected maximum utility associated with the lower-tier decision process, which determines the correlation in unobserved components among alternatives in the nest (Daganzo and Kusnic (1993)). The range of this log-sum parameter should be between 0 and 1 for all nests if the NL model is to remain globally consistent with the random utility maximizing principle.

Furthermore, recent research works contribute to the development of closed form models which relax some of the above-mentioned simplifying assumptions to provide a more realistic representation of choice probabilities. Mixed logit (ML) model is an example of these alternative structures (Bhat (2002)). The ML model has been introduced to bridge the gap between logit and probit models by combining the advantages of both techniques (Ben-Akiva and Bolduc (1996)). In ML models, heterogeneity can be accounted for by letting certain parameters of the utility function differ across individuals. It has been shown that this formulation can significantly improve both the explanatory power of models and the precision of parameter estimates (Ben-Akiva and Dolduc (1996); Bhat (2000)). There are a growing number of empirical studies implementing ML method.

1.3. Joint SP-RP models

Both RP and SP data have their strengths and weaknesses, namely that RP data are cognitively congruent with actual behavior while SP surveys can be collected in a tightly controlled choice environment and can provide richer information on preferences (Walker and Ben-Akiva (2002)). The strengths of both data sources could be exploited and weaknesses ameliorated by pooling both data sources as a joint SP-RP model (Louviere et al. (2000)). This "data enrichment" process should provide more robust parameter estimates and should increase confidence and accuracy in predictions (Verhoef and Franses (2002)).

Techniques of joint SP-RP models have been commonly used in different disciplines such as marketing, transportation, and environment for quite some time. In the context of activity-based modeling, current work includes development of an SP-RP combined mode choice model as the lowest-level model for the primary tour within the entire activity-based model system, drawing on the various available RP and SP data sources for the city of Tel-Aviv (Shiftan et al. (2003)).

2. Model estimation

2.1. Data preparation

The SP dataset comprises effective samples from 1,000 households (800 households in Abobo and 200 households in Anyama). Excluding data records presenting modes with few samples such as railways and focusing only on to-work, to-school, and to-other trips for the purpose of this study, a dataset containing 4,240 samples was established for mode choice analysis. Major characteristics of the respondents found to be:

• Workers and students constitute 33% and 46% of the total, respectively, and the remaining are homemakers, unemployed, retirees, etc.;

• The ratios of males and females are nearly the same (i.e., 50%);

• Only about 11% of the total respondents have a driver's license, whereas the rest have no license;

• Approximately 8% of their households own autos; and

• Among private mode users, about 98% of workers are provided with free parking by the employer, though there is no custom of transportation allowance being provided by the employer in Abidjan.

2.2. Alternative setting

For this study, three major motorized modes were included as existing modes: auto (AT), taxi (TX), and transit (TR). Furthermore, while auto trips are usually divided into two modes, namely, drive alone and shared ride, AT was placed as one alternative in the mode choice models. As for non-motorized mode of transport, it was omitted from the mode choice models because its share in the SP survey was nearly zero due to the long distance to downtown (i.e., at least 7km).

Thus, there are a total of three existing modes that have been set for this study: AT, TX, and TR. Shares of these representative modes in the dataset are 3%, 3%, and 94% respectively. For the SP part of the model, the commuter rail (CR) is added to these three existing modes. It is assumed that modes of TX, TR, and CR are available to all individuals, while availability of AT is limited.

2.3. Explanatory variable

The variables tested for modeling are attributes related to the travel as well as socioeconomic attributes of the household and the individual, and are listed as below.

• Travel related variables: travel cost, travel time, and travel distance

• Household related variables: household income, and vehicle ownership (i.e., number of autos in the household)

• Individual related variable: employment status (e.g., full-time, part-time, and student), school type, personal income, gender, age, vehicle availability, work/school location, and availability of free parking

• In addition, some composite variables such as travel cost divided by the household income and travel time multiplied by the household income were also tested as explanatory variables.

2.4. Modeling results

For developing MNL and NL models, an SP model and an RP model were first estimated, respectively. While the number of observations in the RP model is exactly the same as the number of sampled individuals, the number of observations in the SP model is nearly double because most of the individuals chose more than one alternatives (i.e., CR and TR modes) under different CR fare settings. Meanwhile, a joint SP-RP model was estimated in which one subset is labeled as the RP choice set and the other is labeled as the SP choice set. The RP choice set was placed just below the root of the tree with a scale (log-sum) parameter fixed at 1.0, while each alternative of the SP choice set was placed as a degenerate (single alternative) branch with a free but common scale parameter. It is an artificial nested logit model, and the scaling parameter does not need to lie in the unit interval but can be greater than 1.0 because individuals are not modeled as choosing from the full set of RP and SP alternatives (Greene (2002)). In fact, choice has to be made equally from the SP and RP choice sets. As such, each observation of the SP choice has the corresponding observation of the RP choice. Hence, total number of observations in the joint SP-RP model is double the number of observations in the SP model.

In the joint SP-RP model, effort was made to have common coefficients in both of the RP and SP utility functions for the same alternative. This means that the marginal rates of substitution among some of the variables are the same in the SP and RP models. On the other hand, only a joint SP-RP model was estimated for the ML model, in which both RP and SP choice sets were placed on the same tier.

2.4.1. MNL m/del

Results of the Multinomial Logit (MNL) models that are estimated separately for the RP and SP choice sets and the joint SP-RP MNL model are shown in Table 1.

The adjusted p2, as a measure of fit of the model, is 0.980 for the RP model, which is better by far than its SP counterpart, 0.698. This may be because RP data reflect actual choice behavior and hence have high reliability as to the choice among the existing alternatives as a current situation. In other words, under the current situation, few modes are actually available to people, and the mode is selected depending on their socioeconomic group. On the other hand, the adjusted p2 of the joint SP-RP model is 0.572. The scale parameter of the SP data is 1.387, indicating that the SP error variance is smaller than the RP error variance. It may be contributed to by the fact that estimated parameters in the RP model have generally larger absolute values than those in the SP model. Modeling outcomes are summarized and discussed as below.

All variables associated with the travel are included in the models as continuous variables. Travel cost per se did not work well in the models, but travel cost divided by household income was included in all the utility functions except for NM of the models. It implies that the magnitude of the negative impact of the travel cost on the choice of the motorized modes is larger for people in the lower-income household. Similarly, travel time did not work well in the models, but travel time multiplied by household income was included in some utility functions, that is, TR and CR of the SP model and TR of the joint SP-RP model. It implies that the travel time by the public mode is an important factor to select this mode, especially for people in the higher-income household. It also shows that travel time does not affect the choice of the private modes; it can be inferred that auto users are less likely to shift to the public mode whether there is serious congestion or not on the way to work/school.

Variables related to the household and individual are all included in the models as dummy variables. There are two household-related variables involved in the models. One is a high-income household dummy, which is positively significant in CR. CR is also regarded as a prospective alternative means of transport by people in the high-income household.

Similar tendencies can be found in one of the variables related to the individual, that is, a dummy of whether the individual has a driver's license. Though it is not included in the RP model, it reduces the utility to use transit and relatively increases the utility to select a private mode of transport.

2.4.2. NL m/del

Results of the Nested Logit (NL) models estimated separately for the RP and SP choice sets and those of the joint SP-RP NL model are shown in Table 2. As is the case with the MNL models, the adjusted p2 is 0.988 for the RP model, which is better by far than its SP counterpart, 0.704. As implied in the MNL model, people are captive to the dominant mode in their socio-economic group under the current situation. The adjusted p2 of the joint SP-RP model is 0.616, implying that the joint SP-RP model is better estimated than the SP only model. The scale parameter of the SP data is 3.846, again indicating that the SP error variance is smaller than the RP error variance. Furthermore, the logsum coefficients capturing the effect of expected utility from auto fall within the theoretically acceptable range between 0 and 1 in the RP, SP, and joint SP-RP models.

Table 1. MNL Models: Estimation of SP and RP Mode Choices_

Joint SP and RP Model SP RP ,. Coeff. Coeff.

SP Model

RP Model

Description

Coeff. (t-

Coeff.

.JkÊQÛ......

Continuous Vaiables

CI Travel cost (thousand FCFA)

C2 Travel time (minutes)

C4 C5 C6

Distance to the arterial road from home (km) Log of distance to Plateau/Adiame/bus(Gbaka) station Number of transfers (modes) in currently selected mode

Log of household income (thousand FCFA/month)

Number of household members (of age 12 or above)

C11 Log of age of individual

-0.201 (-2.32) -1.256 (-6.11) -5.150 (-12.64) -0.011 (-5.42)

-0.288 (-2.28) 0.867 (5.69)

0.060 (3.46) 1.794 (3.22)

-0.028 (-2.70)

-0.212 (-1.93) -4.214 (-1.83)

4.430 (3.39)

-0.192 -0.192

(-2.70) (-2.70)

-0.943 -0.943

(-6.04) (-6.04)

-0.008 -0.008

(-4.72) (-4.72)

-0.032 -0.032

(-6.65) (-6.65)

-0.291 -0.291

(-3.03) (-3.03)

0.537 0.537

(5.16) (5.16)

-0.106 -0.106

(-2.57) (-2.57)

0.048 0.048

(3.70) -0.713 -5.470 1.477 3.185

(3.70) -0.713 -5.470 1.477 3.185

D8 D9 D10

Dummy Variables

1, if low-income household (less than 60,000 FCFA/month); 0, otherwise 1, if middle-income household (between 60,000 and 400,000 1, if high-income household (over 400,000 FCFA/month); 0, otherwise 1, if low-class house type (5,6,7); 0, otherwise

1, if male; 0, otherwise

1, if age is below 20 years; 0, otherwise

1, if age is 60 years or above; 0, otherwise

1, if worker; 0, otherwise

Alternative-Specific Constants Auto

Taxi Transit

Commuter Rail

-1.058 (-1.64)

0.225 (2.72) -0.704 (-2.26) 0.312 (2.59)

-0.202 (-2.41)

-13.045 (-5.95)

-3.330 (-4.19)

1.357 (1.98)

-2.611 (-2.78)

-3.450 (-2.11) 1.498 (2.48) 1.361 (1.89)

-7.849 (-1.29)

-4.416 (-4.34)

2.862 (1.80) 0.142 (2.39) -0.099

2.862 (1.80) 0.142 (2.39) -0.099

(-1.66) (-1.66) 0.247 0.247 (2.77) (2.77)

1.710 1.710 (3.86) (3.86)

-0.153 -0.153 (-2.43) (-2.43)

-10.047 -10.495 (-5.57) (-5.63)

-8.229 -3.657 (-13.12) (-6.09)

Scale Parameters Auto

Transit

Commuter Rail Summary Statisics

4195 observations L (0) = -2084 L (ß) = -6899

AT TX TR CR

1.387 (16.77) 1.387 (16.77) 1.387 (16.77) 1.387 (16.77)

3258 observations 6516 observations L (0) = -76 L (0) = -6412

L (ß ) = -3884 L (ß ) = -15004

1.000 ( - > 1.000 ( - > 1.000 ( - > 1.000 ( - >

Table 2. MNL Models: Estimation of SP and RP Mode Choices_

Joint SP and RP Model SP RP Coeff. Coeff.

SP Model

Variable

Description

RP Model

Coeff.

Alternative Coeff. Alternative ........................................(t-stat).........................................(t-stat)

Alternative

(t-stat)......(t-stat).....

C2 C3 C4 C6 C7 C8

D2 D4 D5 D6 D7 D8 D9 D10 D11

Continuous Vaiables

Travel cost (thousand FCFA)

Travel time (minutes) Distance to

Plateau/Adjame/bus(Gbaka) station Distance to the arterial road from home (km)

Number of transfers (modes) in

currently selected mode

Number of autos (MC, PC, van) per

household member

Log of household income (thousand

FCFA/month)

Number of household members (of age 12 or above)

Log of individual income (thousand FCFA/month) C11 Log of age of individual

Logsum: expected maximum utility from Public Transport Summary Statisics

TR CR TR

TR TX,TR TX,CR AT TR TR TX

Dummy Variables

1, if middle-income household (between 60,000 and 400,000 1, if low-class house type (5,6,7); 0, otherwise

1, if middle-class house type (2, 3, 4);

0, otherwise

1, if middle-income individual (between 60,000 and 400,000 FCFA/month); 0,

1, if high-income individual (over 400,000 FCFA/month); 0, otherwise 1, if male; 0, otherwise 1, if age is below 20 years; 0, otherwise

1, if age is 60 years or above; 0, otherwise

1, if worker; 0, otherwise

Alternative-Specific Constants Auto

Taxi Transit

Commuter Rail

Scale Parameters Auto

Transit

Commuter Rail

TX TR CR

-0.185

(-2.03) -1.413 (-4.47) -5.822 (-8.98) -0.013 (-4.06)

-0.271 (-1.48) -0.424 (-1.87) 0.913 (3.84) 0.140 (1.74) 0.069 (2.45) -0.301 (-3.30) 1.572 (2.63)

-1.167 (-1.92)

1.190 (2.59) -1.014 (-1.99)

0.803 (2.07)

4.019 (4.22)

-9.654 (-3.83)

-0.511 (-0.34)

0.694 (3.20)

1946 observations L (0) = -998 L (ß ) = -3406 p 2 =0.704.....

AT,TR TX CR AT,TR

TX TR CR

0.359 (2.69)

3258 observations L (0) = -70 L (ß) = -5757 p 2=0.988

-0.029 (-2.37)

-0.182 (-1.53)

1.382 (1.93) -1.587 (-1.58) 1.276 (1.91) 5.794 (2.26)

1.046 (1.63) -7.151 (-3.41) 1.460 (2.06) 5.679 (2.42)

3.914 (2.47)

-4.456 (-3.90)

-0.083 -0.083

(-1.55) (-1.55)

-0.488 -0.488

(-1.96) (-1.96)

-0.004 -0.004

(-1.81) (-1.81)

-0.031 -0.031

(-1.88) (-1.88)

-0.107 -0.107 (-1.51) (-1.51)

0.255 (1.85) 0.019 (1.80)

0.255 (1.85) 0.019 (1.80)

-0.091 -0.091 (-1.86) (-1.86)

0.703 (1.67)

0.703 (1.67)

0.251 (1.51)

0.251 (1.51)

-2.285 -2.285 (-2.72) (-2.72)

-4.427 -4.419 (-2.11) (-1.94)

-7.739 -1.799 (-7.65) (-1.95)

3.846 (2.08) 3.846 (2.08) 3.846 (2.08) 3.846 (2.08)

0.354 (5.23)

2918 observations L (0) = -2891 L (ß) = -7553 p 2 = 0.616

1.000 ( - ) 1.000 ( - ) 1.000 ( - ) 1.000 ( - )

0.095 (1.83)

2.4.3. ML model

For the Mixed Logit (ML) modeling, 1,000 repetitions are used to estimate the unconditional probability by simulation. This improves the accuracy of the simulation of individual log-likelihood functions and reduces simulation variance of the maximum simulated log-likelihood estimator. Random parameters for this ML model are estimated as normally distributed parameters in order to allow parameters to get both negative and positive values. Both observed attributes associated with the mode alternative, individual, and household (explanatory variables) and the unobserved attributes (alternative specific constants) were tested by introducing random parameters.

Results of the estimated joint SP-RP ML model are shown in Table 3. The adjusted p2 is 0.510, presenting a good model fit with statistically significant parameters. Furthermore, estimated standard deviations of the random parameters of the variable representing the travel cost is statistically significant in the model at 95% confidence level or better. The significant t-statistics for these standard deviations indicate that these are likely to be statistically different from zero, confirming that parameters indeed vary across individuals. Features of the other variables and the implications are generally similar to those in the MNL and NL models. Since travel cost has been introduced with random parameters into the model, travel cost divided by household income is no more included in the model instead.

Table 3. ML Model: Joint Estimation of SP and RP Mode Choices

Joint SP and RP Model SP

Variable Description

Continuous Vaiables

C2 C3 C6 C8 C9 C10 C11

Travel cost (thousand FCFA)

standard deviation Travel time (minutes)

Distance to Plateau/Adjame/bus(Gbaka) station from home (km)

Number of transfers (modes) in currently selected mode

Log of household income (thousand FCFA/month)

Number of household members (of age 12 or above)

Log of individual income (thousand FCFA/month) Log of age of individual

Dummy Variables D9 1, if age is below 20 years; 0, otherwise

1, if worker; 0, otherwise

standard deviation

Alternative-Specific Constants Auto

Taxi Transit

Commuter Rail

Summary Statisics

standard deviation

Alternative

TR AT,TX AT,TR,CR CR TR TX,CR AT,TR TR TX

TR TR,CR TR,CR

2918 observations L(0)=-2969 L (ß ) = -6068 p2 = 0.510

Coeff. (t-stat)

-1.088 (-1.50) -1.313 (-3.51) 0.003 (1.81) -0.017 (-3.75) -0.099 (-3.19) -0.313 (-1.40) 1.243 (3.61) 0.104 (2.87) -0.177 (-1.70) 5.228 (2.02)

1.138 (2.35) -2.823

RP Coeff. (t-stat)

-1.088 (-1.50) -1.313 (-3.51) 0.003 (1.81) -0.017 (-3.75) -0.099 (-3.19) -0.313 (-1.40) 1.243 (3.61) 0.104 (2.87) -0.177 (-1.70) 5.228 (2.02)

1.138 (2.35) -2.823

(-1.39) (-1.39) 5.336 5.336 (1.90) (1.90)

-27.865 -29.226 (-2.85) (-3.00)

-11.875 -8.291 (-6.08) (-4.32)

0.261 (0.25)

1.619 (3.52)

3. Simulation of mode choice

3.1. Current policies under review

The Ministry of Transport of Cote d'lvoire has been trying to implement a commuter train system proposed by SDUGA (Figure 1). The first phase of the Blue Line commencing from Anyama includes 6 stations north of Central, the intersect station of the two metro lines and a further 9 stations to the south, terminating at the airport. The Red Line commencing from West Yopougon includes 4 stations west of Central and further 7 stations to the east of Central terminating at Bingerville. The stations on average are spaced at approximately every two kilometers. This commuter train system consisting of the Blue and Red Lines is scheduled to be completed by 2023.

On the other hand, it is obvious that future transport demand in Abidjan cannot be satisfied by private autos. To curb the growth of private auto use and encourage the shift to public transport, direct traffic demand control, through some kind of a pricing policy in the central business district (CBD), is required. In Europe, after Norway became the first country in the world to introduce automatic toll collection for road usage in 1986, many countries implemented traffic demand control through road pricing. It is one of the main road transportation control measures to alleviate traffic congestion and reduce air pollution. It mainly charges passenger auto users passing through designated roads, in order to minimize unnecessary utilization of passenger autos and divert users to public transport. It also has an important objective to specify the revenues collected from road pricing as the funds for transportation system improvement.

In the case of Abidjan, it is relatively easy to apply road pricing in the CBD (i.e., Plateau) as it is surrounded by the lagoon and there are a limited number of entry roads. The target area for pricing is presented in Figure 2 along with the proposed tollgates. Taking the entire area of Plateau as a restricted area, entering vehicles are charged a fee. The existing two bridges connecting Plateau and Treichville (another commune to the south of Plateau) will also be charged for road pricing.

While this road pricing scheme may be effective for congestion reduction in the CBD, provision of alternative means of transportation for the "pushed-out" users by the road pricing is of great importance to obtain public acceptance. Hence, it is necessary to consider simultaneously the road pricing scheme and the commuter rail development which may serve as an alternative for assumed pushed-out vehicle users. As such, in this study these two major policies are simulated in the mode choice model.

Fig. 2. Proposed target area and toll gates for road pricing.

3.2. M/del t/ be usedf/s simulation

The joint SP-RP ML model was applied for simulation due to the following reasons:

• Either SP or joint SP-RP models are necessary to simulate the future commuter rail mode share. Since SP data are hypothetical and experience difficulty taking into account certain types of real market constraints, estimating and applying a stand-alone SP model for simulation is not recommended but a joint SP-RP model is preferred (Louviere and Swait (2000));

• The existing mode shares were used as the benchmark to compare the three joint SP-RP models, and the mode shared derived from the ML model best matched with the observed data as shown in Table 4;

• NL models are essentially better than MNL models when the IIA property is suspected to be violated. However, a problem with the NL model is that it requires a priori specification of the nesting structure. The actual competition structure among alternatives may be a continuum that cannot be accurately represented by partitioning the alternatives into mutually exclusive nests (Bhat (2003)); and

• In combining SP and RP data, scaling differences and the correlation in unobserved attributes across repeated choices by the same decision makers are often an issue. Simple ML specifications easily incorporate unobserved correlation and scaling differences, and are statistically superior to the "standard" joint scaled logit models (Brownstone et al. (2000)).

Table 4. Comparison of Surveyed and Simulated Mode Shares

[unit: %]

Mode Survey Result MNL Model NL Model ML Model

Auto 2.7 1.3 5.7 3.9

Taxi 2.7 0.1 1.7 1.1

Transit 94.4 98.6 92.6 95.0

Total 100.0 100.0 100.0 100.0

3.3. Assumptions

The assumptions employed for the simulation are as follows:

• Since all the samples in the dataset are to-work, to-school, or to-other auto trips with CBD zones as the destination, all the trips are affected by the road pricing scheme;

• A variety of fare levels were tested for the commuter rail, ranging from 250 FCFA to 1,500 FCFA per ride with an interval of 250 FCFA. The commuter rail service frequency is every three minutes in all the cases;

• Four cases of levy rate were tested for road pricing, namely, 0 FCFA (i.e., no road pricing), 400 FCFA, 800 FCFA, 1,200 FCFA, 1,600 FCFA and 2,000 FCFA per trip;

• All vehicles passing/driving through the toll gates to the target area are to be charged under this road pricing scheme. It is a cordon pricing scheme in which only vehicles entering the area are to be charged; and

• Analyses are made as if changes took place now. Most of the transportation-related costs such as transit fares, expressway tolls, parking prices, and fuel cost are assumed to be the same as the time that the SP survey was conducted (August 2013).

The initial taxi fare is 150 FCFA and the conventional bus fare is 200 FCFA per ride. Prices of this range may not be so painful for high-income people but may be significant enough to low- and middle-income class people. On the other hand, the highest road pricing levy assumed, 2,000 FCFA, is something that high-income auto users can still afford, but it is not negligible even for them.

3.4. Simulation results

For each individual, the model simulates the mode choice decision to go to work/school in the CBD. As such, shares of the four representative modes, that is, auto, taxi, transit, and commuter rail are simulated under each combination of the commuter rail fare and the road pricing levy rate are shown in Table 5. No commuter rail and no road pricing cases are also included in the table. Major changes of the mode shares can be summarized as below.

When a new mode, that is, commuter rail is introduced, it is expected to play an important role in terms of the mode share. It will have the largest share in place of transit. Figures show that the majority of the prospective commuter rail passengers will come from the current transit users, that is, within the public modes. A significant portion of the current taxi users are also expected to shift to commuter rail, implying that taxi users are rather flexible in mode choice. Some of the current auto users are expected to shift to commuter rail as well; however, such portions are relatively small.

As the commuter rail fare rises, the share of commuter rail naturally decreases significantly. All the other transport modes will increase the shares accordingly, and it is especially remarkable in transit which is a mode previously used by many commuter rail passengers. Shares of auto users will also increase, but only marginally.

Table 5. Simulation Results Using Joint SP-RP ML Model

No Road Pricing [unit: %]

With CR: Fare Level (FCFA)

Mode Without CR 250 500 750 1,000 1,250 1,500

Auto 3.9 4.3 4.3 4.3 4.3 4.3 4.3

Taxi 1.1 1.3 1.2 1.3 1.3 1.3 1.3

Transit 70.4 71.4 71.4 71.3 71.4 71.3 71.3

Commuter Rail 24.6 23.0 23.0 23.1 23.0 23.0 23.1

r r r r r r

Total 100.0 100.0 100.0 100.0 100.0 100.0 100.0

Road Pricing: 400 FCFA per Entry [unit: %]

With CR: Fare Level (FCFA)

Mode Without CR 250 500 750 1,000 1,250 1,500

Auto 4.2 4.2 4.2 4.2 4.2 4.2 4.2

Taxi 1.3 1.3 1.3 1.3 1.3 1.3 1.3

Transit 71.3 71.4 71.4 71.3 71.4 71.4 71.4

Commuter Rail 23.2 23.2 23.1 23.3 23.2 23.2 23.2

r r r r r r

Total 100.0 100.0 100.0 100.0 100.0 100.0 100.0

Road Pricing: 800 FCFA per Entry [unit: %]

With CR: Fare Level (FCFA)

Mode Without CR 250 500 750 1,000 1,250 1,500

Auto 4.0 4.0 4.0 4.0 4.0 4.0 4.0

Taxi 1.3 1.3 1.3 1.3 1.3 1.3 1.3

Transit 71.4 71.3 71.3 71.4 71.3 71.4 71.4

Commuter Rail 23.4 23.4 23.4 23.3 23.4 23.4 23.3

r r r r r r

Total 100.0 100.0 100.0 100.0 100.0 100.0 100.0

Road Pricing: 1,200 FCFA per Entry [unit: %]

With CR: Fare Level (FCFA)

Mode Without CR 250 500 750 1,000 1,250 1,500

Auto 3.8 3.9 3.9 3.9 3.8 3.9 3.8

Taxi 1.3 1.3 1.3 1.3 1.3 1.3 1.2

Transit 71.4 71.3 71.3 71.4 71.4 71.3 71.3

Commuter Rail 23.5 23.6 23.6 23.5 23.5 23.6 23.6

r r r r r r

Total 100.0 100.0 100.0 100.0 100.0 100.0 100.0

Road Pricing: 1,600 FCFA per Entry [unit: %]

With CR: Fare Level (FCFA)

Mode Without CR 250 500 750 1,000 1,250 1,500

Auto 3.7 3.7 3.7 3.7 3.7 3.7 3.7

Taxi 1.3 1.3 1.3 1.3 1.3 1.3 1.3

Transit 71.3 71.3 71.3 71.5 71.4 71.4 71.3

Commuter Rail 23.7 23.7 23.7 23.5 23.7 23.7 23.7

Total 100.0 100.0 100.0 100.0 100.0 100.0 100.0

Road Pricing: 2,000 FCFA per Entry [unit: %]

With CR: Fare Level (FCFA)

Mode Without CR 250 500 750 1,000 1,250 1,500

Auto 3.6 3.5 3.6 3.5 3.6 3.6 3.6

Taxi 1.3 1.3 1.2 1.3 1.3 1.3 1.3

Transit 71.4 71.3 71.3 71.4 71.3 71.3 71.4

Commuter Rail 23.8 23.9 23.9 23.8 23.8 23.9 23.8

Total 100.0 100.0 100.0 100.0 100.0 100.0 100.0

All these mode share changes caused by the commuter rail fare increase are more striking under the road pricing scheme, especially with a higher levy rate.

If the road pricing scheme is applied to the CBD, the share of auto users are expected to drop as the levy rate increases. It seems that the road pricing with a higher levy rate makes auto users shift to commuter rail the most. Increase in shares of other modes is smaller. All these influences of the road pricing levy rates are more noticeable under the lower-fare cases of commuter rail. It implies that, as commuter rail becomes more affordable in terms of cost, it is expected to better serve as an alternative mode for autos under the road pricing scheme.

4. Summary

This study developed MNL, NL, and ML models to simulate the mode choice using the RP and SP data obtained in Abidjan, Cote d'lvoire. Although the models established in this study may be difficult to be incorporated into a practical activity-based micro-simulation modeling system due to the sample size limitation, purpose, and location, it captured the key variables that are significant for modeling mode choices in Abidjan and analyzed some of the policy scenarios that are currently under review, through the simulation using the best model derived.

Interpretation of the effects of each explanatory variable in the developed models led to several interesting insights. A wide variety of different types of variables contributed significantly to the models, including basic travel characteristics (cost, time, and distance), household characteristics (income, house type, auto ownership), and individual characteristics (gender, age, worker/non-worker, personal income). Thus, it appears that only the characteristics associated with the trip may not suffice to fully explain mode choice; rather, several household and individual factors play an important role. Moreover, as demonstrated in this paper, the choice model that has incorporated such factors enables better analysis of policy scenarios through the simulation.

Although the dataset consists of only from-home trips to the CBD, mode choice in this first segment of the individual's travel tour is important because these trips constrain the modes of the subsequent segments such as returning home trips and work-based sub-tours. Furthermore, according to the results of the SP survey, more than 80% of shopping tours are made by the same mode that are selected to go to work/school, whether it is on weekdays or weekends. As such, one extension of the study will be to develop a model that determines modes of the subsequent trips in a tour, focusing on the mode transition.

Furthermore, the authors' further effort includes establishment of a comprehensive mode choice model that is applicable to the activity-based micro-simulation modeling system. Although a variety of variables proved to be significant in this study, activity patterns were not included as explanatory variables in the mode choice model, because the SP survey data lacked such information. Using the abundant RP data source available from the HTS, a full-scale mode choice model that includes activity-related variables as input and returns full information to the upper-level choice of the modeling system should be developed. The prototype mode choice model presented in this study will still be useful in predicting the mode shift caused by some new policies such as commuter rail and road pricing.

Acknowledgements

The authors would like to express their gratitude to the Ministry of Construction, Sanitation and Urban Development (MCLAU) of Cote d'lvoire and Japan International Cooperation Agency (JICA) for permitting us to use the survey data in this study.

References

Ben-Akiva, M., and D. Bolduc (1996). "Multinomial Probit with a Logit Kernel and a General Parametric Specification of the Covariance Structure."

3rd International Choice Symposium, Columbia University. Bhat, C.R. (2000). "Incorporating Observed and Unobserved Heterogeneity in Urban Work Travel Mode Choice Modeling." Transportation Science, Vol. 34, No. 2: 228-238.

Bhat, C.R. (2002). "Recent Methodological Advances Relevant to Activity and Travel Behavior," in H.S. Mahmassani (ed.), In Perpetual Motion:

Travel Behavior Research Opportunities and Application Challenges, Oxford: Elsevier Science, 381-414. Bhat, C.R. (2003). "Random Utility-Based Discrete Choice Models for Travel Demand Analysis," in K.G. Goulias (ed.), Transportation Systems Planning: Methods and Applications, CRC Press, Chapter 10.

Brownstone, D., D.S. Bunch, and K. Train (2000). "Joint Mixed Logit Models of Stated and Revealed Preferences for Alternative-Fuel Vehicles." Transportation Research Part B, Vol. 34: 315-338.

Daganzo, C.F., and M. Kusnic (1993). "Two Properties of the Nested Logit Model." Transportation Science, Vol. 27: 395-400.

Daly, A.J., and S. Zachary (1978). "Improved Multiple Choice Models," in D.A. Hensher, and M.Q. Dalvi (eds.), Determinants of Travel Choice, Westmead: Saxon House.

Greene, W.H. (2002). "Nested Logit and Covariance Heterogeneity Models," in NLOGIT Version 3.0 Reference Guide, New York: Econometric Software, Inc., Chapter 8.

Japan International Cooperation Agency (JICA), and Ministry of Construction, Sanitation and Urban Development (MCLAU) of Cote d'lvoire (2014). The Project on the Development of the Urban Master Plan in Greater Abidjan (SDUGA), Final Report, prepared by Oriental Consultants Co., Ltd., Japan Development Institute, International Development Center of Japan, and Asia Air Survey Co., Ltd., Tokyo.

Louviere, J.J., and D. Street (2000). "Stated-Preference Methods," in D.A. Hensher, and K.J. Button (eds.), Handbook of Transport Modelling, Oxford: Elsevier Science, Chapter 8.

Louviere, J.J., D.A. Hensher, and J.D. Swait (2000). Stated Choice Methods: Analysis and Application, Cambridge: Cambridge University Press.

McFadden, D. (1978). "Modeling the Choice of Residential Location." Transportation Research Record, Vol. 672: 72-77.

Meyer, M.D., and E.J. Miller (2001). "Demand Analysis," in Urban Transportation Planning: A Decision-Oriented Approach, 2nd Edition, New York: McGraw-Hill, Chapter 5.

Mohammadian, A., and S.T. Doherty (2005). "A Mixed Logit Model of Activity Scheduling Time Horizon Incorporating Spatial-Temporal Flexibility Variables." Forthcoming in the Journal of Transportation Research Record, TRB.

Shiftan, Y., M. Ben-Akiva, K. Proussaloglou, G. Jong, Y. Popuri, K. Kasturirangan, and S. Bekhor (2003). "Activity-Based Modeling as a Tool for Better Understanding Travel Behaviour." Paper presented at the 10th International Conference for Travel Behaviour Research Conference, Lucerne, Switzerland.

Verhoef, P.C., and P.H. Franses (2002). "On Combining Revealed and Stated Preferences to Forecast Customer Behavior: Three Case Studies." Economic Institute Report, 257, Erasmus University Rotterdam, Econometric Institute.

Williams, H.C.W.L. (1977). "On the Formation of Travel Demand Models and Economic Evaluation Measures of User Benefit." Environmental Planning A, Vol. 9: 285-344.

Walker, J., and M. Ben-Akiva (2002). "Generalized Random Utility Model." Mathematical Social Sciences, 43(3): 303-343.