See discussions, stats, and author profiles for this publication at: [Link]

net/publication/287431185

MODE CHOICE ANALYSIS: THE DATA, THE MODELS AND FUTURE AHEAD

Article in International Journal for Traffic and Transport Engineering (IJTTE) · September 2014
DOI: 10.7708/ijtte.2014.4(3).03

CITATIONS READS

67 6,244

2 authors:

Minal Chandra Ravi Sekhar Chalumuri

CSIR Central Road Research Institute CSIR Central Road Research Institute
29 PUBLICATIONS 228 CITATIONS 68 PUBLICATIONS 797 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ravi Sekhar Chalumuri on 05 January 2016.

The user has requested enhancement of the downloaded file.

 Minal et al. Mode Choice Analysis: The Data, the Models and Future Ahead
UDC: 656.02:519.863 DOI: [Link]

MODE CHOICE ANALYSIS: THE DATA, THE MODELS AND

FUTURE AHEAD
Minal1, Ch. Ravi Sekhar2
1
Academy of Scientific and Innovative Research (AcSIR), CSIR-Central Road Research Institute, New
Delhi-110025, India
2
Transportation Planning Division, CSIR-Central Road Research Institute, New Delhi-110025, India
Received 10 February 2014; accepted 7 May 2014

Abstract: Mode choice is one of the most vital stages in transportation planning process
and it has direct impact on the policy making decisions. Mode choice models deals very
closely with the human choice making behaviour and thus continues to attract researchers
for further exploration of commuter’s choice making process. The objective of this study
is to carryout detailed review on various modeling methods of mode choice analysis and
bottlenecks associated with the same. The factors that affect the psyche of the travelers
have been discussed; further various types of data required and their method of collection
has been briefed up. This paper particularly emphasizes on statistical mode choice models
such as multinomial logit and probit models as well as recent advanced soft computing
techniques such as Artificial Neural Network models (ANN) and Fuzzy approach model
that are employed for modal split analysis. Comparative analysis were made among various
modeling techniques for modeling the complex mode choice of behaviour of models carried
out by various researchers in the literature and a discussion on the need of future hybrid soft
computing models has been attempted.

Keywords: mode choice, data collection, logit, probit, soft-computing techniques.

1. Introduction the users of private modes to mass transport

modes seems to be a solution but is not
The choice of transport mode is probably easily attainable given the comfort factor
one of the most important classic models in of mass transport facilities. To alleviate such
transport planning. This is because of the deteriorating transportation conditions
key role played by public transport in policy researchers have carried out studies to
making (Ortuzar and Willumsen, 2002). understand the relationship between mode
Accelerated industrialization throughout choice and various factors affecting it.
the world has led to higher growth rates,
increased income and high demand for Modeling of mode choice is done by means
mobility. Increasing number of vehicles in of discrete choice model (Ben-Akiva and
city causes congestion and environmental Lerman, 1985); the different available
problems that lead to disrupted traffic alternatives in a discrete choice experiment
conditions like delay, accidents which cause are mutually exclusive and collectively
huge economic loss every year. Attracting exhaustive. Discrete model is based on

2
Corresponding author: [Link]@[Link]

269
International Journal for Traffic and Transport Engineering, 2014, 4(3): 269 - 285

selecting the alternative that provides highest travel time, travel cost, waiting time, number
utility to the choice maker. Predicting the and ease of transfers, comfort, etc. Over
correct alternative for an individual is not the years mode choice models have been
always possible as many unobserved and dealing with the general range of trade-
situational variables come into play for offs individuals are willing to make among
decision making, thus the concept of random these factors (Lerman, 1975; Ben-Akiva and
utility appeared (McFadden, 1980). Prior to Lerman, 1985; Koppleman and Wen, 2000;
1970s, discrete choice was used for binary Bhat, 2000). Zhao et al. (2002) supported
choice of travel mode (Lisco, 1967), later, the ability of transit systems with high-
its use spread to multi choice set as well as quality services to attract more users, as
in other stages of forecasting process. Since well as for poor services to encourage more
1990, application of soft computing models automobile use and classified them into the
such as Artificial Neural Network (ANN) four categories named as travel mode Level of
models (Cantarella and De Luca, 2003), Service (LOS), accessibility, land use/urban
Fuzzy logic applications (Rao and Sikdar, design and transit users, socioeconomic
1999) and Genetic Algorithm (GAs) was characteristics. Later, Racca and Ratledge
adopted. Study on mode choice analysis (2004) added characteristics of a trip as a
benefit engineers, transportation planner factor that affects choice of travel mode.
and policy makers to study the existing Researchers like Stratham and Dueker
transportation system and forecast the (1996) and Ye et al. (2007) have identified
future needs of the proposed transportation that tour complex it y inf luences mode
system as we get an insight to preferences choice substantially. Residential location,
and requirements of commuters. neighborhood type and urban form play a
prominent role in determining the favored
The philosophy behind mode choice model is travel mode for commute (van Wee et al.,
to effectively manage the transport demand 2003; Pinjari et al., 2007; Frank et al., 2008).
and be able to provide for these demands by
making changes in the existing system. The The factors discussed above clearly depict
objective of this paper is to provide a brief that travel time is one of the highly rated
review of the state of art of the different factors considered in mode choice and
stages and techniques of the mode choice is widely used concept in transportation
modeling process. Since mode choice analysis (Bhat and Sardesai, 2006). Recently,
closely affects the policy decisions, it needs various researchers are considering Value
to be understood and practiced with great of Travel Time (VOT) as an inf luencing
diligence. The paper aims at bringing to parameter in choosing the mode. VOT
forefront the critical issues that needs to be means how much a passenger is willing
dealt during mode choice analysis. to pay to avoid any increase in transport
time for an average or typical mode of
2. Factors Affecting Mode Choice transport. Estimation of VOT is crucial
Behaviour for cost benefit analysis of transportation
projects. It is also instrumental in developing
Mode choice of commuters is inf luenced congestion pricing policies. Algers et al.
by a whole panorama of social, economic, (1998) calculated value of time as a trade-off
cultural, and environmental factors like ratio between the in-vehicle time coefficient

270
 Minal et al. Mode Choice Analysis: The Data, the Models and Future Ahead

and the cost coefficient and they found that Assisted Telephone surveys and Commuter
estimation of VOT is sensitive to model Assisted mail surveys are increasingly being
specification and assumptions made on replaced by web based online (internet)
the coefficients. Koppeleman and Bhat sur veys. Online sur veys are becoming
(2006) stated that VOT can serve as an an essential research tool for a variety of
important informal test for evaluating the research fields, including marketing, social
reasonableness of the model. Researchers and official statistics research. Wang et al.
have found that the perceived VOT for work (2000) and Adler et al. (2002) stated from
trips is close to the wage rate of passengers. their study based on web-based travel survey
For non-work trips this value is found to that although this type of survey method is
be clustered around 25% of the wage rate highly capable in handling complex tasks
(Lesley, 2009). as stated preference experiments, it can be
highly unreliable. Thus, the model estimated
3. Methods of Collecting Travel Behaviour by web cannot be judged to give accurate
Data results. Secondly, the sampling frame for
internet surveys is often not available as it
The data required for modeling is collected cannot be known that the respondents may
through surveys like household survey, behave totally differently to the population
workplace survey, destination survey, and of interest. Recently, GPS data loggers
intercept survey. Sampling from the data are deployed for collecting second-by-
set is also a critical step and should be second location, position, and speed data
attempted with caution. Paper and pencil (Wolf, 2004). The first passive GPS study
interviewing (PAPI) is an orthodox method conducted as part of a major household travel
for data collecting. It represents a process survey occurred in Austin in 1997 (Casas and
of personal interviewing where the pollster Arce, 1999). In these studies, participants
holds a printed-out questionnaire, reads are provided with the GPS loggers for the
the question to the respondent and fills the duration of the study and participation is
answers into the questionnaire. It has higher the GPS component is completely voluntary.
chances of error compared to Computer The techniques of revealed preference and/or
A ssisted Inter v iew ing (CA PI) (Kalfs, stated preference are used as complementary
1995). CAPI is a computer assisted data tools to elicit the preferences of the decision
collection method for replacing paper-and- maker. Revealed choice data describes
pen methods of survey data collection and current observed travel patterns and costs
usually conducted at the home or business and hence gives very accurate picture of
of the respondent using a portable personal current modal choice. However, the use
computer. It allows interviewers to conduct of Stated Preference (SP) data had often
face-to-face interview using the computer. been and still is sometimes rejected due to
After the interviews, the interviewers send their unknown reliability. Morikawa (1989)
the data to a central computer. CAPI can also proved a milestone in the acceptance of SP
include Computer Assisted Self-Interview data and its major results were summarized
(CASI) session where the interviewer hands in Ben-Akiva and Morikawa study (1990a).
over the computer to the respondent for a SP responses cannot be used alone for
short period, but he/she remains available forecasting actual behavior because of their
for instructions and assistance. Computer unknown bias and error properties, but they

271
International Journal for Traffic and Transport Engineering, 2014, 4(3): 269 - 285

contain very useful information on trade-offs It has a more causal nature and is thus more
among attributes. Ben-Akiva and Morikawa transferable to a different point in time
(1990b) presented the data fusion method and to a different geographic context, very
that allows the combination of two or more well suited for proactive policy analysis.
complementary data sources into a single Efficiency of disaggregate approach is
data base. more than the aggregate approach in terms
of model reliability per unit cost of data
4. Review on Mode Choice Models collection.

4.1. Aggregate and Disaggregate Mode 4.2. Statistical Mode Choice Models
Choice Models
Discriminant model considers statistical
Aggregate models attempt to represent the classification technique to classify dependent
average behavior of a group of travelers variable between groups and to calculate
instead of a single individual. Different each respondent probability to get into
aggregate models devised and used over past one or another group. Zenina and Borisov
decades are a) Trend analysis where past (2011) used stepwise and forward stepwise
trends were extrapolated to estimate future discriminant analysis to model mode choice
travel (Daganzo, 1979); b) Mathematical and compared the results with Multinomial
models like the direct demand models and Logit (MNL) model and Decision Tree (DT)
sequential models are usually more difficult models. The main problem of discriminant
to implement, more time-consuming and analysis is the selection of discriminant
more costly but provide more accuracy; c) variables and the choice of discriminant
Trip-end modal split, applied immediately function. Discrete choice models based on
after trip generation, and d) Trip-Interchange random-utility maximization are widely used
modal split models when modal split is in transportation applications. They have
applied after the trip distribution. While the three different families of models depending
former preserves the various socio-economic upon the functional form of the error term
characteristic of the commuters the latter distribution namely Logit model, Probit
includes the characteristics of the journey Model and General Extreme Value (GEV)
and that of the alternative modes available to Model. Logit model has the ability to model
user. The aggregate transportation planning complex travel behaviors of any population
models have been severely criticized for their with simple mathematical techniques and
inflexibility and inaccuracy. These models thus proves to be the most widely used tool
at base attempt to represent the average for mode choice modeling. The mathematical
behavior of a group of travelers instead of framework of logit models is based on the
a single individual. Disaggregate models theory of utility maximization (Ben-Akiva
which appeared in 1980s offer substantial and Lerman, 1985). Logit models can be
advantage over its aggregate counterparts categorized in three types depending on
as it represents the behavior of individuals. whether the data or coefficients are chooser-
In disaggregate approach individual choice specific or choice-specific. Multinomial
responses as a function of the characteristics logit model has chooser specif ic data
of the alternatives available and socio- where coefficients vary over the choices.
demographic attributes of each individual. Conditional logit model has choice-specific

272
 Minal et al. Mode Choice Analysis: The Data, the Models and Future Ahead

data where the coefficients are equal for all logit models and mixed logit model were
choices. Mixed logit model involves both estimated. The results of this study indicate
types of data and coefficients. that the opening of high-speed rail services
or bullet trains would definitely increase the
4.2.1. Multinomial Logit Models ridership of railways provided the cost of
railways is pretty reasonable as Sensitivities
Multinomial Logit (MNL) model is the most of time and cost are interdependent on each
basic member of the family of GEV models. other. Abuhamoud et al. (2011) modeled
There are three basic assumptions which the mode choice preference of commuters
underlie the MNL formulation. The random towards car use and bus use using binary
components of the utilities of the different logit model, in Libya. The factors affecting
alternatives are independent and identically the choice pattern was also studied and the
distributed (IID) with a Type I extreme- major outcome of the study was that gender
value or Gumbel distribution. MNL model analysis needs to be incorporated into all
maintains homogeneity in responsiveness to transport planning, so that gender impacts
attributes of alternatives across individuals. are studied and considered before project
The error variance-covariance structure of implementation.
the alternatives is identical across individuals.
The three assumptions discussed above 4.2.2. Nested Logit Models (NL)
together lead to the simple closed-form
mat hemat ica l str uct ure of t he M N L . Nested L og it (N L) st r uc t u re a l low s
However, these assumptions also leave the estimation of proportions among selected
MNL model laden with the IIA property at sub-modes, prior to the estimation of
the individual level which proves to be the proportions between modes. The nested
greatest drawback of MNL model (Bhat, logit model have their random component
1995). IIA property implies that any changes identically, non-independently distributed
in the probability of a given alternative draw with type I extreme value distribution
equally from the probabilities of all the other allowing partial relaxation of the assumption
alternatives in the choice set (Hess, 2005). of independence among random components
A l A hmadi (2006) developed intercity of alternatives (McFadden, 1978). It has a
mode choice models for Saudi Arabia using closed form solution, is relatively simple
MNL and this study results indicated that to estimate and is more parsimonious than
in-vehicle travel time, out of pocket cost, the multinomial probit model. The major
number of family members travelling drawbacks of NL models are first, the number
together, monthly income, travel distance, of different structures in search for the best
nationality of traveler, and number of cars structure increases rapidly as the number of
owned by family played the major role in alternatives increases. Second, the actual
decision related to intercity mode choice. competitive structure among alternatives
Mukala and Chunchu (2011) through stated may be a continuum which cannot be
preference (SP) data modeled the mode choice accurately represented by partitioning
behavior of specific category of trip makers the alternatives into mutually exclusive
traveling from Guwahati to five metro cities subsets. Abdel-Aty and Abdelwahab (2001)
in India. The choice set included Railways developed mode choice models for Florida,
and Airways. Various forms of Standard USA. The mode choice model was estimated

273
International Journal for Traffic and Transport Engineering, 2014, 4(3): 269 - 285

as three level Nested Logit structure. The issues of interpretation in the case of counter-
overall model utilized full information intuitive results. Due to this complexity
maximum likelihood estimation. Among the transport planners generally prefer
the significant variables that entered into using logit models as they possess simple
model are transit access time, transit waiting mathematical framework and can accurately
time, number of transfer, in-vehicle travel model the travel behavior of a study area.
time, fare and household car ownership. Ghareib (1996) estimated the travel behavior
Khan (2007) designed Nested logit models for different cities of Saudi Arabia using
for different trip length and trip purpose for logit and probit models and concluded that
Redland Shire. Data was collected through a the logit models are superior to their probit
Stated Preference survey in which an entirely counterparts in terms of their goodness-
a new virtual travel environment was created of-fit measures and tractable calibration.
and the result of this study indicated that trip Dow and Endersby (2004) later supported
length affects the perception of alternatives his findings by concluding that the logit
to work for longer trips. models should always be preferred over
probit models and the latter should only be
4.2.3. Multinomial Probit Model (MNP) utilized if the travel behavior of the targeted
population to be determined is observed to
Multinomial Probit (MNP) model is the be complexly correlated. Bhat and Sardesai
main alternative to GEV-based model (2006) in their study pointed out that in
structures in discrete choice analysis. The Probit model due to increase in flexibility of
underlying assumption of Probit models is error structure, several additional parameters
that the error terms follow a joint normal are introduced in the covariance matrix. This
distribution with zero mean and covariance generates a number of conceptual, statistical
matrix with no priori restrictions on the and practical problems, including difficulty
correlation structure in the distribution. in interpretation, highly non-intuitive
The absence of the assumption of identically model behavior, low precision of covariance
distributed error terms means that taste parameter estimates, and increased difficulty
variation and repeated choices can also be in transferring models from one space-time
incorporated in Probit models. Thus, the sampling frame to another. Can (2011) used
cases where the utilities of some alternatives probit for tourist mode choice in Nha Trang.
are correlated in a complex way and use of The major findings regarding the factors
MNL models can make incorrect forecasts affecting the choice process are out-of-
regarding the probabilities of mode shares mode per kilometer, on-mode travel time
when the attributes associated to one or per kilometer, per-kilometer travel cost
more travelling alternatives are varied MNP to income ratio; safety, etc. affect choice
model can be used. MNP model imposes a decision very closely.
heavy computational burden due to the issue
regarding identification of an appropriate 4.2.4. Generalized Extreme Value (GEV)
covariance structure. Major disadvantage Models
of the Probit model is the requirement to
use a normal distribution for representing G e n e r a l i z e d e x t r e m e v a l u e (G E V)
random taste heterogeneity, leading to models were developed as an important
significant losses in terms of flexibility and simplification of multinomial logit models

274
 Minal et al. Mode Choice Analysis: The Data, the Models and Future Ahead

based on the stochastic utility maximization. from combination of ANN and Fuzzy logic.
It is a closed form distribution that allows This section describes these recent models.
for various levels of correlation among
the unobser ved par t of utilit y across 4.3.1. Artificial Neural Network Models
alternatives (Hess, 2005). Random utility
model is developed with independent but Cantarella and De Luca (2003) compared
non-identical error terms distributed with MNL and ANN models for mode choice
a type I extreme value distribution. Bhat modeling using disaggregate discrete choice
(1995) developed a Heteroscedastic Extreme data. Two types of neural networks were
Value Model of intercity mode choice. The trained and the results compared to the logit
proposed model allowed a more f lexible model. The results revealed that ANN model
cross-elasticity structure among alternatives adopted in this study outperformed the MNL
than the nested logit model and was also model. Xie et al. (2003) in their mode choice
free from the IIA assumption of MNL. Bhat studies used Decision tree (DT) and Artificial
(1995) made a comparative analysis among neural network (A NN). Datasets from
multinomial logit model, nested logit models, the San Francisco Bay Area Travel Survey
and heteroscedastic extreme value model and (BATS) 2000 were used. They compared
the results emphasing that heteroscedastic the results obtained by these techniques
extreme value model shows a number of with a traditional multinomial logit (MNL)
advantages over other commonly used model. Prediction results show that the two
discrete choice models. Yang et al. (2013) data mining models offer comparable but
developed a Cross-Nested Logit model slightly better performance than the MNL
for capturing a joint choice of residential model in terms of the modeling results,
location, travel mode, and departure time while the DT model demonstrates highest
using Beijing traffic data. The results were estimation efficiency and most explicit
compared to Logit model and three different interpretability and the ANN model gives
Nested logit models. Estimation results show a superior prediction performance in most
the Cross-Nested Logit model outperforms cases. Ravi Sekhar (1999) carried out a
the three kinds of NL model. It is also mode choice study for Delhi. In this study
found by estimation results that decision mode choice models based on ANN and
makers will change first their departure MNL were formulated and a comparative
times, then their travel modes, and finally analysis was done between both. The ANN
their residential locations, when exogenous model different models were developed
variables alter. based on the vehicle ownership and the
choice set available to the commuters. A
4.3. Soft Computing Mode Choice Models back propagation algorithm was used for the
ANN architecture. The relative importance
Alternatively researchers have sought out of input parameters was found out and Object
to more recent soft computing techniques Oriented Programming (OOP) was used
like Artificial Neural Network (ANN), to implement ANN network. ANN models
Fuzzy Logic and hybrid models resulting outperformed the MNL models.

275
International Journal for Traffic and Transport Engineering, 2014, 4(3): 269 - 285

4.3.2. Fuzzy Logic Based Mode Choice and hidden layers was sigmoid with Back-
Models Propagation algorithm. The output was the
weights assigned for each alternative mode.
Deb (1993) used Fuzzy set theory to study Rao and Sikdar (1999) carried out urban
the Mass transit mode choice pattern for mode choice analysis by calibration of Fuzzy
Calcutta. Different Mass transit alternatives functions from revealed preference survey in
like Bus, Car, Surface Railway, Metro and Mumbai. They used ANN for the calibration
Water Transport were taken into account and of Fuzz y membersh ip f u nc t ion. T he
fuzzy set theoretic approach was employed to membership function was modified by the
select the more preferable set of alternatives. back propagation of error. Modification was
Aggregate matrix was used to compare the in proportion to the error signal. The model
various alternatives included in this matrix to gave performance of 99.73% in calibration
select the best alternative(s). An alternative and 98.64% in validation suggesting accurate
was taken as superior to a second alternative result.
if it dominated the second-alternative in more
number of factors than the number of factors 5. Development of Mode Choice Models
in which the second dominates the first.
Seyedabrishami and Shafahi (2013) carried T h is sec t ion d isc usses t he d i f ferent
out a Trip destination and Mode choice joint psychological and economic theory that
model analysis by using Fuzzy set theory. is used in decision making process. Such
The model is structured as a decision tree in theories form the basis of understanding and
which the fuzzy and non-fuzzy classification developing mode choice models.
of influential variables regarding destination
selection and mode choice expand the 5.1. Attitude-Based Theories
tree for Shiraz, a large city in Iran. When
compared with a multinomial logit (MNL) Ajzen developed the Theory of Planned
model, the suggested models’ estimates are Behaviour (TPB), assuming that the choice is
more accurate than the traditional MNL also dependent on the individual’s perception
model. of his or her ability to execute certain
behaviour (Ajzen, 1985). According to TPB,
4.3.3. Hybrid Mode Choice Models the intention behind certain behaviour is
dependent on three factors namely, the
Fused Neuro Fuzzy (NF) architecture let attitude toward the behaviour, the social
ANN learning algorithms to determine the norm and the perceived behavioral control.
parameters of Fuzzy Inference System (FIS). Attitude-based theories have been criticized
Fused NF systems share data structures because it is difficult to know whether
and knowledge representations. Yaldi et attitudes control travel mode choice or
al. (2010) developed a NF model for trip vice versa. People do not always act as they
distribution and mode choice analysis. say they will. Persson et al. (1998) stated
The input nodes can be categorized into that attitudes’ value for predicting actual
two types representing the socio-economic behaviour is poor, because they are collected
characteristics of the individuals. The through interviews or questionnaires similar
activation function for nodes in both output to Stated Preference interviews.

276
 Minal et al. Mode Choice Analysis: The Data, the Models and Future Ahead

5.2. Utility Theory for Discrete Choice complete rational behaviour and perfect
Model information without taking into account the
taste and preferences of individual are called
Uti lit y is an indicator of va lue to an deterministic utility models. There are cases
individual. In a discrete choice experiment, when an individual chooses alternative i over
a decision-maker chooses a single alternative j even when Uj > Ui . The primary sources
from a choice set of finite number of mutually of error in the use of deterministic utility
exclusive alternatives where the choice set functions are: First, the individual may
is exhaustive. An individual is visualized have incomplete or incorrect information or
as selecting a mode which maximizes his misperceptions about the attributes of some
or her utility (Khan, 2007). The utility or all of the alternatives. Second, the analyst
of a travel ling mode is def ined as an or observer has different or incomplete
attraction associated to by an individual information about the same attributes
for a specific trip. This hypothesis is known relative to the individuals. Third, the analyst
as utility maximization. This can be stated is unlikely to know specific circumstances
as alternative, ‘i’, is chosen among a set of the individual’s travel decision, mostly
of alternatives, if and only if the utility of the personal situations of the traveler.
alternative, ‘i’, is greater than or equal to the These errors are considered while making
utility of all alternatives, ‘ j’, in the choice choice under Probabilistic Choice models
set, C. This can be mathematically in Eq. or Random Utility Model in which Utility
(1). If the utility of alternative i is greater is decomposed into two components. One
than or equal to the utility of all alternatives component of the utility function represents
j; alternative i will be preferred and chosen the portion of the utility observed by the
from the set of alternatives. analyst, often called the deterministic (or
observable) portion of the utility. The other
If (1) component is the difference between the
unknown utility used by the individual
Where, and the utility estimated by the analyst,
U( ) – mathematical utility function; represented by ε:
X i , X j – vectors of attributes describing
alternatives i and j, respectively, and (2)
S t – Socio-economic characteristics of
individual t. where Uit is the true utility of the alternative i
to the decision maker t, Vit is the deterministic
5.3. Deterministic and Probabilistic or observable portion of the utility estimated
Choice Theory by the analyst, and ε it is the error or the
portion of the utility unknown to the analyst.
According to the utility maximization
rule, an individual chooses the alternative 5.4. Logit Theory Model
with the highest utility. But there is always
uncertainty involved in the individual’s The mathematical framework of logit models
decision process. Utility models that assume is based on the theory of utility maximization

277
International Journal for Traffic and Transport Engineering, 2014, 4(3): 269 - 285

(Ben-A k iva and Lerman, 1985). Logit simple processing elements called neurons.
models can be categorized in three types Each neuron is connected to other neurons
of logit models depending on whether the by means of directed links and each directed
data or coefficients are chooser-specific or link has a weight associated with it. These are
choice-specific. Multinomial logit model has used to address problem that are intractable
chooser specific data where coefficients vary or cumbersome with traditional methods.
over the choices. Conditional logit model has The typical computation of Artificial Neuron
choice-specific data where the coefficients is presented in Fig. 1. It can be represented
are equal for all choices. Mixed logit model mathematically, if I1, ...Ii ...In are the input
involves both types of data and coefficients. values and synaptic weight values are W lj,
...Wj, Wnj then the summation, netj is over
Probability of an individual i selecting a all the incoming neurons of the product of
mode n, out of M number of total available the incoming neuron’s activation and the
modes, is given in Eq. (3): synaptic weight of the connection at the
typical jth neurode expressed as .A
threshold value θ1 is incorporated into the
(3) output. The mathematical representation of
netj is given in Eq. (4).
Vin – utility function of mode n for individual
i; (4)
Vim – utility function of any mode m in the
choice set for an individual i; Where n is the number of incoming neurons,
Pin – probability of individual i selecting mode I is the vector of incoming neurons, W is
n; and the vector of the synaptic weights and θ is
M – total number of available travelling node threshold, usually taken as the negative
modes in the choice set for individual i. weight from the bias unit (a unit whose
output is unity). The output at each neurode
The Logit model can be classified as Binary by considering Eq. (5):
Logit model, Multinomial Logit model and
Nested Logit model. OUTPUT = f(netj) (5)

5.5. Soft Computing Techniques Where f(netj) is activation function also called
as threshold function, transfer function or
5.5.1. Artificial Neural Network (ANN) squashing function used to map the input
pattern of neuron to the specified output
An Artificial Neural Network (ANN) is range. The most commonly used threshold
a parallel information-processing system functions are linear, non-linear, sigmoid,
that has certain performance characteristics hyperbolic, tangent, etc. Most widely used
similar to biological neural networks. A function in mode choice analysis is sigmoid
neural net consists of a large number of function.

278
 Minal et al. Mode Choice Analysis: The Data, the Models and Future Ahead

Fig. 1.
Artificial Neuron
µA(x) = 1 where x A and µA(x) = 0 where x A (6)
5.5.2. Fuzzy Logic (FL)
But fuzzy set contains certain degree of
Fuzzy set theory firstly introduced by Zadeh membership. The degree of membership
(1965) deals with propositions that can be has a value from zero to one, i.e. a value from
true to a certain degree (somewhere from 0 to the closed interval [0,1]. The greater µA(x),
1). Many of the influencing variables used as the greater the truth of the statement that
input to mode choice modeling do not follow element x belongs to set A. (illustrated in
definition of crisp set. Crisp sets allow either Fig. 2).
full membership or no membership at all
whereas fuzzy sets allow partial membership. 0 ≤ µA (x) ≤ 1, where x X.
Thus, fuzzy set is an extension of a crisp set
that capture vague or linguistic expression. After deciding the membership function
The performance of any system controlled the fuzzy rules are formed and then through
by Fuzzy Logic depends on the fuzzy rules, Fuzzy Inference System (FIS), the input is
which in turn depend on the membership mapped to output by using fuzzy set theory.
functions of the input variables. The most A FIS is usually designed by considering
tedious and crucial part of forming the Fuzzification, fuzzy interface through the
fuzzy logic is determining the membership bank of fuzzy rules and Defuzzification of
function. In a crisp set membership of ‘x’ the fuzzy output variables as depicted in
µ(x), in set A is defined in Eq. (6): Fig. 3.

a) Crisp membership function (b) Fuzzy membership function

Fig. 2.
Crisp and Fuzzy Membership Function

279
International Journal for Traffic and Transport Engineering, 2014, 4(3): 269 - 285

Fig. 3.
Schematic Diagram of a Fuzzy Inference System

5.5.3. Neuro-Fuzzy Models 5.6. Comparison of Various Mode Choice

Models
Hybrid models using ANN and Fuzzy set
theory were experimented to complement MNL model has been the most widely used in
the drawbacks of both the techniques. ANN mode choice analysis and its IIA assumption
has a learning capacity and can capture the is misleading as it is not rational. Nested logit
complex relationship between the variables model with its tree and branched structure
but needs good amount of data. Fuzzy set gives some relaxation from the IIA property
theory, on the other hand can capture the of MNL but the nested logit structures are
linguistic and vague expressions that is more either inconsistent with utility maximization
close to human behavior but has no learning principles or are not significantly better
ability and also finding out the best suited than the multinomial logit model. Probit
membership function is very challenging. mode choice model which provides greater
So on combining both these techniques, f lexibility, is rarely used in travel demand
can build a very powerful model that has modeling because these models poses the
a learning ability as well as can grasp the complexity of introducing several additional
vagueness involved in human decision parameters in the covariance matrix which
making procedure. The various fused NF require high dimensional integrals which is
models that can be used are: Cooperative, difficult. Neural Network based mode choice
Conc u r rent a nd Fused Neu ro-Fuzz y models gives higher accuracy than the logit and
Systems. In the Cooperative NF model, the probit models because these models are capable
ANN decides the membership function or in mapping the independent and explained
the fuzzy rules to be implemented based variable but its “black box” image does not
on the data and then go to background. In help when it comes to easy interpretation of the
concurrent NF model ANN dynamically results. Fuzzy logic based mode choice models
assists FIS to set its parameters. employed in modeling does associate closely

280
 Minal et al. Mode Choice Analysis: The Data, the Models and Future Ahead

to human linguistic expression but deciding provides basic hypothesis, major constraints,
on the best suited membership function is drawbacks and level of accuracy of various
very time taking and tiring process. Table 1 mode choice models in the literature.

Table 1
Comparison of Various Mode Choice Models
Model
Basic Hypothesis Major constraints Major Drawbacks Accuracy Examples
Name
Error term should
Extreme value Al Ahmadi (2006);
Logit be identically and Suffers from IIA
distribution (Type I High Mukala and
model independently assumption.
GEV). Chunchu (2011)
distributed (IID).
Error term is non- Difficult interpretation
Probit Normal identically and of results. Bhat and Sardesai
Low
model distribution. non-independently Computational (2006); Can (2013)
distributed. problems.
Independent, but
Multivariate Perceived utility of any
GEV not identically
extreme value alternative should not Low Bhat (1995)
model distributed error
distribution. exceed an upper bound.
terms.
Performance Edara (2003);
depends on Difficult to interpret Cantarella and De
ANN Inspired by Human
Architecture, Training the result in terms of High Luca (2003); Xie
model Neural system.
and Activation natural language. and Parkany (2003);
function. Ravi Sekhar (1999)
No learning ability and
Fuzzy Based on underlying Deb (1993); Kalic
Input Values should difficulty in deciding
Set assumption of High and Teodorovic
be fuzzy. Membership function
Theory Commensurability. (1999)
(Tuning).
Lack of common
Neuro ANN and FIS Pre-processing of framework, different Rao and Sikdar
Fuzzy models complement input training data is NF models can’t be High (1999); Yaldi et al.
model each other. required. compared. (2010)

6. Conclusions and Future Ahead characteristics of mode as well as the there

are many latent factors like comfort and
Review on Mode choice studies carried out convenience. Modeling of mode choice
in this paper shed light on various aspects can be approached in two ways: aggregate
of mode choice analysis as a transportation modeling and disaggregate modeling.
planning process. Mode choice modeling Disaggregate approach is widely used as it
directly deals with the behavioral aspect can capture the individual characteristics
of human nature thus it needs to closely in a much better way compared to aggregate
monitor and understand the factors that models that depend on zonal characteristics.
affect this decision making procedure. A Disaggregate mode choice models are of
number of factors come into play and can three types namely: Logit model, Probit
be broadly classified as characteristics model and General extreme value model.
of tr ip ma ker cha racter ist ics of tr ip, A mongst a l l t he t hree models, Log it

281
International Journal for Traffic and Transport Engineering, 2014, 4(3): 269 - 285

models have found the most application Abuhamoud, M.A.A.; Atiq, R.; Rahmat, O.K.; Ismail,
in choice models because of simplicity A. 2011. Modeling of Transport Mode in Libya: a
and easy interpretation of results. It also Binary Logit Model for Government Transportation
has comparatively reasonable accuracy. Encouragement, Australian Journal of Basic and Applied
Estimation of parameters of model is done Sciences, 5(5): 1291-1296.
by two techniques, Maximum likelihood
and Least squares methods. The maximum Adler, T.; R immer, L.; Carpenter, D. 2002. Use
likelihood is the most common technique of Internet-Based Household Travel Diary Survey
used in determining the estimators for simple Instrument, Transportation Research Record: Journal of
and nested logit model. The data required the Transportation Research Board. DOI: [Link]
for modeling is collected through surveys org/10.3141/1804-18, 1804: 134-143.
like household survey, workplace survey,
destination survey, and intercept survey. Ajzen, I. 1985. From intentions to actions: A theory of planned
Sampling from the data set is also a crucial behavior. In Kuhl, J., Beckman, J. (Eds.), Action-control:
step and should be done diligently. From cognition to behavior. Heidelberg, Germany:
Springer. 11-39.
The model specification for mode choice
should be done on case specific basis. Al-Ahmadi, H.M. 2006. Development of Intercity
Suitability of the model selected should Mode Choice Models for Saudi Arabia, JKAU: [Link],
be decided first so that the entire process 17(1): 3-21.
starting right from data collection to analysis
and interpretation can be carried out in a Algers, S.; Bergström, P.; Dahlberg, M.; Dillén, J.L.
very planned and streamlined manner. 1998. Mixed Logit Estimation of the Value of Travel Time,
Indian society due to its disaggregated Swedish Institute for Transport and Communications
cha rac ter ist ics needs to be modeled Analysis (SIKA).
differently than the developed nation
where the societies are more of aggregated Ben-Akiva, M.E.; Lerman, S.R. 1985. Discrete Choice
nature. Due to the complexity involved in the Analysis: Theory and Application to Travel Demand, The MIT
Indian travel characteristics a model highly Press, Cambridge, Massachusetts, the USA.
flexible as well as compatible to handle such
heterogeneity should be used. Hybrid models Ben-Akiva, M.E.; Morikawa, T. 1990a. Estimation
seem to be more promising in this regard. of switching models from revealed preferences and
Hybrid mode choice models such as Neuro- stated intentions, Transportation Research Part A: General.
Fuzzy models gives better results than the DOI: [Link]
individual models. 7, 24(6): 485-495.

References Ben-Akiva, M.E.; Morikawa, T. 1990b. Estimation of

Travel Demand Models from Multiple Data Sources. In
Abdel-Aty, M.; Abdelwahab, H. 2001. Calibration of M. Koshi (Ed.), Transportation and Traffic Theory. Oxford:
nested mode choice model for Florida, Final research report, Elsevier Science Ltd. 461-476.
University of central Florida.

282
 Minal et al. Mode Choice Analysis: The Data, the Models and Future Ahead

Bhat, C.R. 1995. A Heteroscedastic Extreme Value Dow, J.K.; Endersby, J.W. 2004. Multinomial Probit and
Model of Intercit y Mode Choice, Transportation Multinomial Logit: A Comparison of Choice Models for Voting
Research Part B: Methodological. DOI: [Link] Research. Electoral Studies. 23 p.
org/10.1016/0191-2615(95)00015-6, 29(6): 471-483.
Edara, P.K. 2003. Mode Choice Modeling Using Artificial
Bhat, C.R . 20 0 0. Incor porating Obser ved and Neural Networks, M.S. Thesis, Virginia Polytechnic
Unobserved Heterogeneity in Urban Work Mode Choice Institute, Virginia.
Modeling, Transportation Science. DOI: [Link]
org/10.1287/trsc.34.2.228.12306, 34(2): 228-238. Frank, L.; Bradley, M.; Kavage, S.; Chapman, J.;
Lawton, T.K. 2008. Urban form, travel time, and cost
Bhat, C.R.; Sardesai, R. 2006. The Impact of Stop- relationships with tour complexity and mode choice,
Making and Travel Time Reliability on Commute Mode Transportation. DOI: [Link]
Choice, Transportation Research Part B: Methodological. 007-9136-6, 35(1): 37-54.
DOI: [Link]
40(9): 709-730. Gha reib, A .H. 19 96. Est i mat ion of L og it a nd
Probit Models in a Mode Choice Situation, Journal
Can, V.V. 2013. Estimation of travel mode choice of Transportation Engineering. DOI: [Link]
for domest ic tou r ists to N ha Tra ng usi ng t he org/10.1061/(ASCE)0733-947X(1996)122:4(282),
multinomial probit model, Transportation Research Part 122(4): 282-290.
A: Policy and Practice. DOI: [Link]
tra.2013.01.025, 49: 149-159. Hess, S. 2005. Advanced discrete choice models with
applications to transport demand, Ph.D Thesis, Centre
Cantarella, G.E.; De Luca, S. 2003. Modeling for Transport Studies Imperial College London.
Transportation Mode Choice through A rtificial
Neural Networks. In Proceedings of the Fourth International Kalfs, N. 1995. The Effects of Different Data Collection
Symposium on Uncertainty Modeling and Analysis. DOI: Procedures in Time Use Research. Paper presented
[Link] at the Transportation Research Board Annual Meeting,
84-90. Washington, D.C.

Casas, J.; Arce, C. 1999. Trip Reporting in Household Kalic, M.; Teodorovic, D. 1999. Modal Split Modeling
Travel Diaries: A Comparison to GPSCollected Data. Using Fuzzy Logic. In Proceedings of the Conference
Presented at the 78th Annual Meeting of the Transportation “Modelling and Management in Transportation”, Poznan-
Research Board, Washington, D.C. Cracow, Poland. 91-96.

Daganzo, C. 1979. Multinomial Probit: The Theory and Khan, O. 2007. Modelling Passenger Mode Choice Behavior
its Application to Demand Forecasting, Academic Press, Using Computer Aided Stated Preference Data, Ph.D Thesis,
New York. Queensland University of Technology.

Deb, S.K. 1993. Fuzzy set approach in mass transit mode Koppeleman, F.; Bhat, C. 2006. A Self Instructing Course
choice. In Proceedings of the ISUMA ‘93, Second International in Mode Choice Modeling: Multinomial and Nested Logit
Symposium on Uncertainty Modeling and Analysis. IEEE Models, U.S. Department of Transportation, Federal
Computer Press, College Park, Maryland. 262-268. Transit Administration.

283
International Journal for Traffic and Transport Engineering, 2014, 4(3): 269 - 285

Koppleman, F.S.; Wen, C-H. 2000. The Paired Pinjari, A.R.; Pendyala, R.M.; Bhat, C.R.; Waddell, P.A.
Combinatorial Logit Model: Properties, Estimation and 2007. Modeling residential sorting effects to understand
Application, Transportation Research Part B: Methodological. the impact of the built environment on commute mode
DOI: [Link] choice, Transportation. DOI: [Link]
0, 32(2): 75-89. s11116-007-9127-7, 34(5): 557-573.

Lerman, S. 1975. A Disaggregate Behavioral Model of Urban Racca, D.P.; Ratledge, E.C. 2004. Factors That Affect
Mobility Decisions, Ph.D. Thesis, Department of Civil and/or Can Alter Mode Choice, Delaware Transportation
Engineering, Massachusetts Institute of Technology, Institute and The State of Delaware Department of
Cambridge, Massachusetts. Transportation, University of Delaware, Newark.

Lesley, L. 2009. Generalized Cost and Value of Time Rao, S.P.V.; Sikdar, P.K. 1999. Calibration of fuzzy
As a Method of Patronage Forecasting, Acta Technica functions from revealed preference pattern in urban
Jaurinensis, 2(1): 57-68. mode choice. CUPUM ‘99 Computers in Urban Planning
and Urban Management On the edge of the millennium.
Lisco, T.E. 1967. Value of Commuters time-A Study in In Proceedings of the 6th International Conference, Venice.
Urban Transportation, PhD. Thesis, University of
Chicago. Ravi Sekhar, Ch. 1999. Mode choice Analysis using Artificial
Neural Network, Master Thesis, Indian Institute of
McFadden, D. 1978. Modelling the choice of residential Technology, Roorkee.
location. In Karlquist A., ed., Spatial Interaction Theory
and Planning Models, North Holland, Amsterdam. 75-96. Seyedabrishami, S.; Shafahi, Y. 2013. A joint model
of destination and mode choice for urban trips: a
McFadden, D. 198 0. Economet r ic models for disaggregate approach, Transportation Planning and
Probabilistic choice among Products, The Journal of Technology. DOI: [Link]
Business, 53(3): 13-29. 0.2013.851507, 36(8): 703-721.

Mukala, P.K.; Chunchu, M. 2011. Mode choice Stratham, J.; Dueker, K. 1996. Transit service, parking
modelling for intercity transportation in India: A case charges, and mode choice for the journey to work: an analysis
of Guwahati to five metro cities, International Journal of of the 1990 NPTS, Journal of Public Transportation, 1(1).
Earth Sciences and Engineering, 4(6): 364-374.
Van Wee, B.; Holwerda, H.; van Baren, R . 2003.
Ortúzar, J. de D.; Willumsen, L.G. 2002. Modeling Preferences for modes, residential location and travel
Transport, 3rd Edition, JohnWiley & Sons, Chichester, behaviour: the relevance of land-use impacts on mobility,
Sussex, England. European Journal of Transport and Infrastructure Research,
2(3-4): 305-315.
Persson, U.; Nilsson, K.; Hjalte, K.; Norinder, A. 1998.
Beräkningav Vägverkets riskvärden. En kombination Wang, D.; Borgers, A.; Oppewal, H.; Timmermans, H.
av “contingent valuation” – skattningar och uppmätta 2000. A stated choice approach to developing multi-
hälsoförluster hos vägtrafikskadade personer behandlade faceted models of activity behavior, Transportation
vid fyra sjukhus. The Swedish Institute of Health Research Part A: Policy and Practice. DOI: [Link]
Economics (IHE), Mimeo. org/10.1016/S0965-8564(99)00045-2, 34(8): 625-643.

284
 Minal et al. Mode Choice Analysis: The Data, the Models and Future Ahead

Wolf, J. 2004. Defining GPS and GPS Capabilities. In Ye, X.; Pendyala, R .M.; Gottardi, G. 2007. A n
Hensher, D., Button, K., Haynes, K., Stopher, P. exploration of the relationship between mode choice
(editors), Handbook on Transport Geography and and complexity of trip chaining patterns, Transportation
Spatial Systems, Elsevier, Handbook No. 5, Chapter 20. Research Part B: Methodological. DOI: [Link]
org/10.1016/[Link].2006.03.004, 41: 96-112.
Xie, C.; Lu, J.; Parkany, E. 2003. Work Travel Mode
Choice Modeling Using Data Mining: Decision Trees Zadeh, L. 1965. Fuzzy sets, Information and Control, 8:
And Neural Networks, Transportation Research Record: 338-353.
Journal of the Transportation Research Board. DOI: http://
[Link]/10.3141/1854-06, 1854: 50-61. Zenina, N.; Borisov, A. 2011. Transportation Mode
Choice Analysis Based on Classification Methods, IT
Yaldi, G.; Taylor, A.P.; Yue, W.L. 2010. Examining the and Management Science, 49: 49-53.
Possibility of Fuzzy Set Theory Application in Travel
Demand Modelling, Journal of the Eastern Asia Society for Zhao, F.; Li, M-T.; Chow, L-F.; Gan, A.; Shen, L.D. 2002.
Transportation Studies, 8: 579-592. FSUTMS Mode Choice Modeling: Factors Affecting Transit
Use and Access, National Center for Transit Research,
Yang, L.; Zheng, G.; Zhu, X. 2013. Cross-nested logit University of South Florida, Tampa, Florida.
model for the joint choice of residential location, travel
mode, and departure time, Habitat International. DOI:
[Link]
38: 157-166.

285

View publication stats