Mode Choice Analysis Using Generalized Nested Logit Model: Conference Paper

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/309321115

MODE CHOICE ANALYSIS USING GENERALIZED NESTED LOGIT MODEL

Conference Paper · May 2014

CITATIONS READS
0 210

2 authors:

Minal S. Ravi Sekhar Chalumuri


CSIR Central Road Research Institute CSIR Central Road Research Institute
11 PUBLICATIONS   15 CITATIONS    39 PUBLICATIONS   122 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

EVALUATION OF ECONOMIC LOSS DUE TO IDLING OF VEHICLES AT SIGNALIZED INTERSECTION AND MITIGATION MEASURES (ELSIM- ESC0204) View project

EVALUATION OF ECONOMIC LOSS DUE TO IDLING OF VEHICLES AT SIGNALIZED INTERSECTION AND MITIGATION MEASURES (ELSIM- ESC0204) View project

All content following this page was uploaded by Minal S. on 21 November 2016.

The user has requested enhancement of the downloaded file.


Colloquium onTransportation Systems Engineering and Management
NIT Calicut, India, May 12-13, 2014.

Paper Id: 155

MODE CHOICE ANALYSIS USING GENERALIZED NESTED LOGIT


MODEL
Minal1 and Ch.Ravi Sekhar2
1
M.Tech Student, Academy of Scientific and Innovative Research (AcSIR), CSIR-Central
Road Research Institute New Delhi-110025, minal.crri@gmail.com
2
Senior Scientist, Transportation Planning Division, CSIR-Central Road Research Institute,
New Delhi 10025, chalumuri.ravisekhar@gmail.com

Abstract

Dealing with the present bottlenecks as well as creating long lasting and sustainable transport
systems has been the greatest challenge of urban transport planning. Calibrating the present
need and forecasting the future demand is the underlying agenda of travel demand
forecasting. Mode choice forms an integral part of this process as it gives a complete insight
to the mode choice preferences of the commuters validating the introduction of new transport
systems to existing ones. This study aims at modelling the mode choice of commuters in
National Capital Region of Delhi in India. The data collected for the study was not through
the conventional household survey but through a technique of Web-based survey. This survey
was hosted at the site of Central Road Research Institute (CRRI) and reached out to people
from different walks of life and with varying socio-economic characters. These datawere
considered for the mode choice analysis. The present study uses discrete choice models for
the mode choice analysis. Logit models have been the most widely used and accepted discrete
choice model. Further Generalized Nested Logit (GNL) model has been explored in this study
and. A Generalized Nested Logit is an extension of Nested Logit model that allows for the
inclusion of a mode in one or more nest providing more flexible and heterogeneous structure
to the model. Alternatives that appear in more than one branch are probabilistically allocated
across branches. The result shows that the model developed by Generalized Nested Logit is
the superior with higher accuracy and better exploratory power than the Logit models.

Keywords Mode choice;Web-based survey; Logit models;Generalized Nested Logit

1. INTRODUCTION

Transportation community faces challenges that are very dynamic in nature. Amongst
the various confluences in Transportation system, congestion is by far the most common and
difficult factor to overcome. Congestion drastically affects the level of service of the transport
system leading to consequences like delay, accidents which lead to huge economic loss every
year. To help alleviate the situation an attempt to model the travel mode choice preferences of
commuters in Delhi has been done in this research. The ultimate interest lies in being able to
predict the decision making behavior of the commuters while taking under consideration the
attributes of different modes like cost, safety, convenience and travel time. The major
motorized modes of transport used in Delhi are private Cars, Two wheelers, Bus, Metro

1Corresponding author: minal.crri@gmail.com


andAuto Rickshaw. The objective behind mode choice model is to effectively manage the
demand and be able to provide for these demands by making changes in the existing system.

Since the transport planners aim for an integrated transport system where all the
components are rightly synchronised and closely co related Generalized Nested Logit models
can be a good option to explore. The objective of the present study is to develop Generalized
Nested Logit model for the commuter in Delhi. Generalized Nested Logit is an extension of
Nested Logit model that allows for the inclusion of a mode in one or more nest providing
more flexible and heterogeneous structure to the model discrete choice models for the mode
choice analysis. This model is compared with most widely used Logit models.

2. REVIEW ON MODE CHOICE ANALYSIS

2.1 Review on Model Frame Work

Logit model has the ability to model complex travel behaviors of any population with
simple mathematical techniques and thus proves to be the most widely used tool for mode
choice modeling. The mathematical framework of logit models is based on the theory of
utility maximization (Ben-Akiva and Lerman, 1985). Logit models can be categorized in three
types of logit models depending on whether the data or coefficients are chooser-specific or
choice-specific. MultinomialLogit(MNL) model has chooser specific data where coefficients
vary over the choices. Conditional logit model has choice-specific data where the coefficients
are equal for all choices. Mixed logit model involves both types of data and coefficients.

Probability of an individual "i" selecting a mode "n", out of "M" number of total
available modes, is given in Equation (1)
Eqn. (1)

Where,
is the utility function of mode "n" for individual "i"; is utility function of any
mode "m" in the choice set for an individual "i"; is the probability of individual "i"
selecting mode "n"; and "M" is the total number of available travelling modes in the choice
set for individual "i".

Nested Logit(NL) structure allows estimation of proportions among selected sub-


modes, prior to the estimation of proportions between modes. The distribution of the random
components in models which use identical, non-independent random components is generally
specified to be either normal or type I extreme value (Ben Akiva, 1985). In literature mostly
used distribution is type I extreme value since it nests the multinomial logit. The resulting
model, referred to as the nested logit model, allows partial relaxation of the assumption of
independence among random components of alternatives (McFadden, 1978). NL decomposes
the error component (which capture the unobserved elements) of the Utility into a portion that
related to every alternative and a portion related to nests of alternatives. The common error
term for the nested alternative represent a covariance structure that depicts increased
correlation and similarity amongst nested alternatives. This leads to higher cross elasticity
amongst the nested alternatives. Thus probability of selection of mode ‘n’ in a nest is given by
Equation (2). The probability of selection of mode ‘a’ which is non nested in nest ‘e’ is given

2
by Equation(3). where‘θ' is the nesting coefficient which estimates the scale parameter of the
Gumbel distribution for the nested alternatives. ‘τe’ is the measure of expected maximum
utility among the nested alternatives.

Eqn.(2)

Eqn.(3)

The MNL and NL model though widely used and accepted in choice modelling poses
restrictions over taste variation, repeated choices and patterns of substitution in the model.

Generalized Nested Logit(GNL) is one such member of generalized extreme value


(GEV) family of models which provides high degree of flexibility in the substitution of
choices. Each of the alternatives in the nested logit models appear only in one nest. In real
case scenarios alternatives may appear in more than one nest. Wen and Koppleman (2001)
have shown that a GNL model can be used to solve such problems where a mode appears in
more than one nest.

Let the nestsof alternatives be labeled B1,B2, etcBK. Each alternative can be a member
of more than onenest. In fact, an alternative can be in a nest to differing degrees. An
allocation parameterαjk reflects the extent to which alternative j is in nest k. This parameter
must be non-negative: αjk ∀ j; k.The allocation parameters should be constrained so that they
sum to one:

∑k αjk= 1 ∀ j. In this way, αjk reflects the portion of the alternative that is allocated to
each nest. A parameter ¸k is defined for each nestas in the standard nested logit model.

The probability that individual n chooses alternative i from the choice set is:

(4)

As before, this can be decomposed into two logits - the probability of choosing nest Bk
multiplied by the probability of choosing alternative i in nest k:

(5)
The relevant work done in mode choice analysis in India and abroad is briefly discussed in the
following section.

3
2.2 Review on Work done in Mode Choice Analysis
Bhat (1995) developed a Heteroscedastic Extreme Value model of intercity mode
choice. Estimation of the ridership share on a proposed new intercity travel service was done
and identification of the modes from which existing intercity travelers will be diverted to the
new or upgraded service was performed. The proposed model allowed a more flexible cross-
elasticity structure among alternatives than the nested logit model and was also free from the
IIA assumption of MNL. Five different models in the study were used: a multinomial logit
model, three possible nested logit models, and the heteroscedastic extreme value model. The
resulting heteroscedastic extreme value model has a number of advantages over other
commonly used discrete choice models

Abdelwahab and Abdel-Aty (2001) developed mode choice models for Florida, USA.
The mode choice model was estimated as three level nested Logit structure. The overall
model utilized full information maximum likelihood estimation. Among the significant
variables that entered into model are: transit access time, transit waiting time, number of
transfer, in-vehicle travel time, fare and household car ownership.Wen and Koppleman (2001)
explored Generalized nested logit model for analysis of intercity mode choice analysis. In the
study the statistical superiority of GNL model has been established over Paired combinatorial
logit, Cross nested logit and Nested logit models with indication of important differences in
cross elasticity relationships across pair of alternatives.

Al Ahmadi (2006) developed intercity mode choice models for Saudi Arabia.
Multinomial logit was used for the model development and data collection was done through
“revealed preference” surveys. These results indicated that in-vehicle travel time, out of
pocket cost, number of family members travelling together, monthly income, travel distance,
nationality of traveler, and number of cars owned by family played the major role in decision
related to intercity mode choice.

Khan (2007) estimated various nested logit models for different trip length and trip
purpose using data from stated preference (SP) survey. A unique computer assisted personal
interviewing (CAPI) instruments was designed using motorized and non-motorized travelling
modes in the SP choice set. Additionally a unique set of access modes for bus on bus way was
also generating containing hypothetical modes such as secure park and ride facilities and kiss
and ride drop-off zones. He found from the final model estimation that the travel behavior
forecasted for regional trip makers is considering different from that for local trip makers. The
regional travelers for work were found not to perceive the non- motorized modes as valid
alternatives to car, possibly due to longer trip length.

In India, most of the modeling approaches are oriented towards the use of economic
theory of Utility maximization. Many researchers have developed mode choice models based
on principles of utility maximization (Chari 1978). Whereasdisutility of minimization
employed by Rao (1988). Parida (1994) has employed stated and revealed preference
approaches for modeling home based work trips in Delhi.Subbaraoet. al (1997) developed
access mode choice model using ANN and compared the results with conventional
Multinomial Logit model (MNL). Ravi Sekhar(1999) developed mode choice model by using
ANN and MNL for Delhi data. In this study, data has been classified based on vehicle
ownership. Ravi sekhar et.al (2009) studied on Applications of Neural Networks in Mode

4
Choice Modelling for Second Order Metropolitan Cities of India for this they have considered
secon order cities travel behavior data in the cities of Visakhapatnam and Nagpur by Central
Road Research Institute (CRRI).Ashalata et al. (2013) attempted a Revealed preference study
of mode choice for Thiruvanathpuram city, India using Multinomial logistic regression. The
major modes included in the study were Car, Two wheeler and Bus. The analysis highlighted
the fact that preference to car increases with age while the preference to use of two wheeler
decreases in comparison to Bus. Also increase in Time per distance and Cost per distance
makes the commuters shift to private modes from the Public mode.

3. TRAVEL BEHAVIOR DATA

In this study an attempt has been made to collect travel behaviour data through Web
survey method. Travel Behaviour survey questionnaire was prepared and published CSIR-
Central Road Research Institute (CRRI).The questionnaire consisted of four sections namely
household Information, personal information, commute Information and mode choice
information. Household Information and personal information collected socio-demographic
data such as age, gender, household size, household income, number of vehicles in the
household, marital status, education status, and zone of residence in National Capital Region
(NCR) which includes New Delhi and the neighboring cities of, Gurgaon, Faridabad,
Ghaziabad, Noida and Greater Noida . Commute Information, and Mode Choice Information
were designed to gather information about the mode choice, the purpose of trip, the total cost
involved and the time taken in the commute of the respondent. The online survey form was
circulated on popular social media sites and e-mailing services targeted at random citizens
from Delhi and regions around Delhi. This survey was published after running a “Beta Test”
or a “Pilot Survey”, that returned acceptable results. A total of 98 responses were received till
the month of May out of which 96 samples were chosen after filtering some undesirable
responses. Six modes considered in the study were Car ,Two Wheeler ,Bus ,Metro , Auto
Rickshaw (Auto Rk) and Walk.

4. MODE CHOICE ANALYSIS USING MNL,NL AND GNL

In this study explanatory variables such as In vehicle Travel time(IVTT), Out of


Vehicle Travel Time (OVTT), Travel cost, House hold Income, House hold Size, Age of the
respondent and the Serviceability level provided by each modes. Variables of House hold
Income, House hold Size, Age and the Modal Serviceability level are calculated for each of
the modes to know the significance of each of these parameters mode wise. The walk mode
has been taken as the reference mode for estimation.The Multinomial logit model (MNL)
model was first calibrated using the data set. The result obtained by the MNL model is
presented in Table 1.

MNL suffers from the Independence from Irrelevant Alternative (IIA) assumption
which implies equal cross elasticity across the alternatives. To accommodate this lacking
mode choice analysis is carried out by considering Nested logit model to investigate the more
closely correlated alternatives of mode choice. Different nesting structures were implemented
for the mode choice analysis. Two of the most significant nests are included in Table1. In
NL1 (Fig 1) Car Two wheeler and walk modes are clubbed together under one branch while
Bus, Metro and Auto Rickshaw are clubbed into the second branch. The model shows better

5
overall fit than the MNL model also itdepicts that Metro and Auto are more closely related in
terms of Modal serviceability, which includes comfort, security, hygiene and convenience of
travel. In NL2, Private mode i.e. Car and Tw are clubbed under one branch while Bus, Metro,
Auto Rk and Walk are in the second branch. This nested structure gives a better result as well
as shows how walking is implicitly associated with the Transit facilities. The performance of
NL1 and NL2 are comparable but they definitely outperform the MNL model. Based upon the
log likelihood ratio test (-2(LLbase model – LLestimated model)), which is used to
statistically test the significance of relative improvements in model fits MNL model can be
rejected by the Nested logit model NL2.

Even though Nested Logit structures relax the IIA assumption they still do not
accommodate the situation in which modes appear in more than one nest. Such a scenario is
very rational and can be handled by Generalized Nested Logit structure (GNL). The result
obtained by the GNL model is compiled in Table1. The structure of the nest for the GNL
model is illustrated in Fig2. Three branches are formed where Car and Tw are grouped under
first branch, Bus, Auto Rickshaw and Walk mode under the second branch andMetro, Auto
Rickshaw and Walk modes under the third branch. The log sum parameter gives the weights
associated with each of the nest in respective models. While in Branch2, highest weight was
associated with the Transit modes, GNL model shows that private modes (i.e. Car and TW)
have higher weightages attached to it. The allocation parameter indicates how much a mode
belongs to a particular nest. It’s value ranges from 0 to 1.0 means absolute exclusion of that
mode from that nest while 1 means that a particular mode belongs entirely to that nest. A
value between 0 and 1 indicates the level of belongingness with the respective nests. From the
results it is evident that use of Auto Rickshaw and Walk is very well coordinated with Bus
facility. These feeder modes serve Bus users more at the present situation. While the same
feeder modes of Auto Rickshaw and Walk are not as well aligned with the Metro service. The
results explicitly reveal that Walking is not a feasible option with Metro.

The IVTT, OVTT and Travel Cost parameters have all negative signs as expected.
Commuters seem to be more sensitive regarding the In Vehicle travel time (IVTT) and an
increase in it will have creates higher dissatisfaction amongst the travellers than the same
increase in OVTT. On increasing the service quality of modes, all the modes will attract
travellers except that in Bus which implies that commuters ride buses despite of their poor
service quality and they do not have much expectation of it’s improvement. Based on
Household income, the higher income households prefer use of Car. Metro seems to be the
most preferred transit mode of travel for the higher income group. Buses are boarded by the
lower income group. Based on Age criteria Buses are mostly favourable for elderly while
Metro seems the most non preferred mode of travel for the old age people. The reason might
be the different interchanges that need to be made and also the poor accessibility of metro
stations from the residential areas.

6
AUTO BUS
CAR TW
WALK BUS
CAR
WALK AUTO METRO
TW
METRO

NL2 NL1
Fig.1. Different nesting structure of Nested Logit model

TW AUTO AUTO WALK


CAR BUS WALK METRO

Fig.2. Generalized Nested Logit model Structure


Table 1:Estimation of Coefficients for Various Mode choice models
Parameters MNL NL1(car,tw,walk), NL2(car,tw), GNL(car,tw),
(bus,metro,auto) (bus,metro, (bus,auto,walk),
auto,walk) (metro,walk,auto)
IVTT -.01780 -0.00317 -0.00381 -.01880
OVTT -.00132 -0.00214 -0.00033 -.00232
Travel Cost -.05788 -0.00839 -0.02777 -0.05553
Mode Constants - -12.176
Car -8.82619 - 0.70209 -4.959
Two-wheeler -3.02171 7.46865 4.20705 -20.481
Bus -.95297 - 1.50258 -17.230
Metro -3.73069 2.89925 - -12.647
Auto-Rickshaw 7.13485 - 1.33698
3.94998 7.05497
-
6.75282
4.81075
Modal Service 0.38882
0.03578 0.05858
Level 0.09518 0.71937
0.06103 0.08258
Car 0.11008 -
0.07169 0.09036
Two-wheeler - 1.08638
- -
Bus 0.04875 0.3181
0.06999 0.00381
Metro - 1.07309
- -
Auto-Rickshaw 0.13384
0.06317 0.08649
-
0.06871
HH Income 1.88689
0.51127 0.10747
Car 0.40912 1.32285
0.32991 -
Two-wheeler 0.17434 0.85362
0.29277 0.08009
Bus 0.12554 1.43253
0.4752 -
Metro 0.28599

7
Auto-Rickshaw 0.11062 0.19522 0.03401 1.19898
0.1525
-
0.16531
Age - -0.23247
0.06566
Car 0.0779 0.03959 -0.44293
-0.0127
Two-wheeler - - -0.15803
0.01665
Bus 0.00817 0.10706 -1.13879
0.05971
Metro - - -0.4853
-
Auto-Rickshaw 0.00338 0.06599
0.21474
- -
0.00254 0.03687
- -
0.27042 0.16876
HH Size
Car 0.37295 1.11246 0.55826 4.4744
Two-wheeler 0.26212 0.81457 0.23128 3.84948
Bus 0.60562 0.76 0.39328 4.6156
Metro 0.58496 0.89586 0.44971 3.66324
Auto-Rickshaw - 0.2851 - 4.66819
0.12348 0.05779

Logsum NA
Parameters 10.0265 14.3277 2.06970
Branch1 12.8712 17.3790 .07984
Branch2 - - .09719
Branch3
Allocation
parameter
Branch1 NA NA NA 1.0
Car 1.0
Tw
Branch2 1.0
Bus 0.929
Auto 0.947
Walk
Branch3 1.0
Metro .0709
Auto .0523
Walk
LL at Zero -172.01 -172.01 -172.01 -132.18
LL at Constants -131.98 -131.98 -131.98 -131.98
LL at -95.01 -88.212 -85.46 -83.84
Convergence 2.563 2.463 2.406 2.434
AIC
Likelihood Ratio - 13.6 19.1 22.34
Test
Prediction 48% 53.12% 53.12% 55.2%

8
Accuracy

5. CONCLUSIONS AND FINDINGS

The conventional method of data collection like Personal interview engages huge
resource and time. A novel technique of data collection through a web-based survey has been
tried in this study. The feasibility and accuracy provided by such unconventional data set has
been tested. The choice model developed for mode choice preference of commuters of
National Capital Region (NCR) of India has been attempted and its accuracy has been tested.
Multinomial Logit and Nested Logit models were developed to assess the mode choice. Since
the classification accuracy of logit models was too low to be considered appropriate, it was
also decided to test the applicability of advanced Discrete Choice model, Generalized Nested
Logit to model mode choice.

In the present study, it can be observed that the developed model utilizing online data
also demonstrate the similar type of capabilities in terms of prediction of mode choice;
therefore it is recommended that the web-based online survey is effectively managing the
resources and time which is otherwise hugely consumed in conventional data collection.

 A comparative evaluation between Multinomial logit (MNL), Nested Logit


andGeneralized Nested Logit for predicting the choice behavior for commuters in
Delhi shows that the performance Generalized Nested Logit model is much superior to
MNL and NL models.
 Even though the performance of GNL model is improving by 2%, Generalized Nested
Logit modelgives a better understanding of closely correlated alternative that can be
nested along with different alternatives in different nests can be developed.
 The result hints that Metro is still not as popular a mode as Bus in Delhi for elderly,
children and people from low income group. The reason can be the not so efficient
feeder service that can carry people to Metro stations. The accessibility if improved
then there might be a major shift in ridership from private vehicles to Metro. This
aspect can be explored in further studies.
 The low number of responses received posed difficulty in carrying out the analysis.
Models were difficult to fit due to the small sample size, so a Household survey was
conducted and large number of sample was collected and will be included in the
further analysis which would include Soft computing techniques like ANN and Fuzzy
logic along with the Statistical models.

Acknowledgements

We are very grateful to the Director, CSIR-Central Road Research Institute for
allowing us to publish this paper. Also this study acknowledges the resource of data collected
for SUSTRANS project, a 12th Five Year Plan project.

9
References
1. Ashalata, R., MAnju, V.S. and Zacharia, A. B.(2013), “ Mode Choice Behaviour of
Commuters in Thiruvananthapuram City ”, Journal of Transportation Engineering,
American Society of Civil Engineers, Volume 139, Issue 5, pg 494-502.
2. Abdel-Aty, M. and Abdelwahab, H. (2001), “Calibration of nested mode choice model
for Florida”, Final research report, University of central Florida.
3. Al-Ahmadi, H.M. (2006), “Development of Intercity Mode Choice Models for Saudi
Arabia”, JKAU: Eng.Sci, Vol. 17 No. 1 , pp. 3-12.
4. Ben-Akiva, M. E. and Lerman, S. R.,( 1985), Discrete Choice Analysis: Theory and
Application to Travel Demand, The MIT Press,Cambridge, Massachusetts, the USA.
5. Bhat, C.R. (1995), A Heteroscedastic Extreme Value Model of Intercity Mode
Choice, Transportation Research Part B, Vol. 29, No. 6, pp. 471-483.
6. Khan,O. (2007), Modelling Passenger Mode Choice Behavior Using Computer Aided
Stated Preference Data, P.H.D. thesis, Queensland University of technology.
7. McFadden, D. (1978), Modelling the choice of residential location, in A. Karlquist,
ed., ‘Spatial Interaction Theory and Planning Models’, North Holland, Amsterdam,
pp. 75–96
8. Parida,M.,(1994) Mode Choice Analyis Based on Stated and revealed Preferences for Home
Based Work Trips in Delhi, PHD Thesis Department of Civil Engineering, IITRoorkee,
Roorkee,India.
9. Ravi Sekhar Ch. (1999) Mode choice analysis using Neural Network, Department of Civil
Engineering, Master Dissertation, IITRoorkee (formally Known as Roorkee University),
Roorkee, India.
10. Ravi SekharCh.,Madhu,E, Durai,B.K and Gangopadyayay. SApplications of Neural
Networks in Mode Choice Modelling for Second Order Metropolitan Cities of
India,Proceedings of the Eastern Asia Society for Transportation Studies, Vol.7, 2009
11. SubbaRao, P. V., Dhingra, S. L., Sikdar, P. K. and Krishna Rao, K. V. (1997) Access mode
choice analysis using artificial neural networks, Proceedings of the Conference on Trends and
Techniques of Transportation, REC Warangal, India. 1997, 81-96.
12. Wen, Chieh-Hua& Frank S. Koppelman. 2001. The Generalized Nested
Logit.Transportation Research Part B 35:627–641.

10

View publication stats

You might also like