21
Mane, Sarkar, Arkatkar 1 STUDY OF MODAL SHIFT TO BUS RAPID TRANSIT (BRT) IN DEVELOPING 1 COUNTR Y: A CASE STUDY IN INDIA. 2 3 Authors: 4 5 Mr. Ajinkya S. Mane 6 Post Graduate Student 7 Civil Engineering Department, 8 Birla Institute of Technology and Science, Pilani, 9 VidyaVihar Campus, Rajasthan-333031 10 Email id: [email protected] 11 12 Dr.Ashoke Kumar Sarkar 13 Professor, 14 Civil Engineering Department, 15 Birla Institute of Technology and Science, Pilani, 16 VidyaVihar Campus, Rajasthan-333031 17 Email id: [email protected] 18 Ph: +911596-245073 19 Fax: +911596-244183 20 21 Dr.Shriniwas S. Arkatkar * 22 Assistant Professor, 23 Civil Engineering Department 24 Sardar Vallabhbhai National Institute of Technology, 25 Surat, Gujrat-395007 26 Email: [email protected] 27 Ph: +918058321357 28 Fax: +911596-244183 29 30 31 No of words: Text: 4232words + 3250 (8 tables + 5figures) =7482words 32 33 94 st Annual Transportation Research Board meeting, Washington D.C. 34 January 11-15, 2015 35 36 Date of Submission: 15 November 2014. 37 38 *Corresponding Author 39 40 41 42 43 44 45 46 47 48 49

Mane, Sarkar, Arkatkar 1 STUDY OF MODAL SHIFT TO BUS RAPID

Embed Size (px)

Citation preview

Mane, Sarkar, Arkatkar

1

STUDY OF MODAL SHIFT TO BUS RAPID TRANSIT (BRT) IN DEVELOPING 1 COUNTRY: A CASE STUDY IN INDIA. 2

3 Authors: 4

5 Mr. Ajinkya S. Mane 6 Post Graduate Student 7

Civil Engineering Department, 8 Birla Institute of Technology and Science, Pilani, 9

VidyaVihar Campus, Rajasthan-333031 10 Email id: [email protected] 11

12

Dr.Ashoke Kumar Sarkar 13 Professor, 14

Civil Engineering Department, 15 Birla Institute of Technology and Science, Pilani, 16

VidyaVihar Campus, Rajasthan-333031 17

Email id: [email protected] 18 Ph: +911596-245073 19

Fax: +911596-244183 20 21

Dr.Shriniwas S. Arkatkar* 22

Assistant Professor, 23 Civil Engineering Department 24

Sardar Vallabhbhai National Institute of Technology, 25 Surat, Gujrat-395007 26

Email: [email protected] 27

Ph: +918058321357 28 Fax: +911596-244183 29

30 31

No of words: Text: 4232words + 3250 (8 tables + 5figures) =7482words 32

33 94st Annual Transportation Research Board meeting, Washington D.C. 34

January 11-15, 2015 35 36

Date of Submission: 15 November 2014. 37

38 *Corresponding Author39

40 41 42

43 44

45 46 47

48 49

Mane, Sarkar, Arkatkar

2

STUDY OF MODAL SHIFT TO BUS RAPID TRANSIT (BRT) IN DEVELOPING 1 COUNTRY: A CASE STUDY IN INDIA. 2

3 ABSTRACT 4

5 In developing country like India, where overpopulation is a major problem, encouraging 6 public transportation is the need of hour. Keeping this in view, Government of India 7

implemented Bus Rapid Transit System (BRTS) in various metro cities in India. In May 8 2013, Government implemented BRTS with exclusive bus lane in rapidly growing city, 9

Indore (situated in the western part of India). After, six months of successful BRT service in 10 Indore city, judicial system ordered to allow private cars (Autos) in an exclusive BRT lane; 11 this was unique decision. The objective of this study is to assess the impact of BRTS service 12

on modal shift before-and-after introduction of cars in exclusive BRT lane For these two 13 cases, separate models are formulated and compared using binary logistic method and 14

Artificial Neural Network (ANN).The data on demographic and socioeconomic attributes 15 (Gender, Age, Occupation) and trip related attributes (Travel time details, Cost saving per 16 day) are collected using Revealed Preference (RP) survey. During this study, en-route survey 17

data is collected on this corridor. The data provides original analysis of changes in travel 18 behaviour in medium city due to BRT deployments. It also adds to the existing body of 19

research on BRT-induced modal shifts using a binary logistic and ANN approaches. All 20 considered variables are found to be statistically significant and influencing for shifting 21 behaviour of passengers to BRTS in both binary logistic models. In ANN, it is found that 22

travel time details, costs saving per day are the most influencing parameters for modal shift to 23 BRTS in both the situations. Due to the introduction of cars in exclusive BRT lane, 24

probability of passengers switching to BRT service is found to be decreased from 64.7% to 25 45.7% (i.e. 19% decrease).While, comparing both the techniques, it is found that ANN 26 provides more accurate results as compared to the binary logistic model in both the situations. 27

28 29

30 31 32

33 34

35 36 37

38 39

40 41 42

43 44

45 46 47

48 49

50

Mane, Sarkar, Arkatkar

3

1.0 INTRODUCTION 1 2

In developing countries like India, Bangladesh, Indonesia etc. overpopulation is a major 3 concern. To address this issue with respect to more sustainable and safe transportation, there 4

is urgent need to encourage public transportation like BRT, Light Rail Transit (LRT) and 5 METRO Despite of overpopulation, India is still rapidly growing economically and also 6 advancement in car manufacturing industries is seen, which resulted into increase in private 7

vehicle ownership. Moreover, due to the limited available road space in urban areas, 8 proliferation of personal vehicles results into traffic congestion high energy consumption, 9

proliferate delay and increased pollution levels in cities. Under such congested traffic, buses, 10 finds it difficult to maneuver through mixed traffic due to their relatively large size. So, it is 11 necessary to initiate bus priority measures like exclusive bus lanes, signal priority system for 12

buses at intersection, etc. to promote and improve the public transportation. 13 14

In India, bus is the main mode of public transport; however its image is deteriorating due to 15 its low level-of-service and inadequate capacity. Hence, Government has decided to improve 16 the public transportation system with Bus Rapid Transit System (BRTS). Currently, BRTS is 17

implemented in major metro cities across India covering about 141 km and is rapidly 18 increasing. Comprehensive analysis of these implemented systems as well as introduction of 19

new innovating ideas can help planners to upgrade the quality of transportation system. 20 Modal shift is the basic parameter for assessing the passengers shifting behaviour to newly 21 introduced public transportation system from other private or semi-private modes of 22

transport. 23 24

In May 2013, Indore (Population of 1.960 million in 2011), a medium size city located in 25 state of Madhya Pradesh, had introduced BRT system(i-Bus), which was implemented based 26 on Ahmedabad BRT model. Unfortunately, on 3rd October 2013, judicial system has made a 27

unique decision to allow private cars in an exclusive lane, which was designated for BRT 28 system. The study aims to assess the modal shift of passengers to newly implemented BRTS 29

from their earlier transport mode such as private cars, two wheelers, buses, paratransit called 30 as auto-rickshaw (three-wheeled motorized vehicle to carry passengers based on fare rates) 31 and also to assess the impact on modal shift after introduction of cars in an exclusive BRT 32

lane. In addition to this, the study also focuses on employing soft computational technique 33 like Artificial Neural Network (ANN) for modelling passengers shifting behaviour over 34

traditional discrete choice model. The results of this study will be very useful for the decision 35 makers and practitioners for improving public transportation policies. 36 37

2.0 LITERATURE REVIEW 38 39

Modal shift is among the basic parameters for predicting shifting behaviour of passengers 40 from their earlier transport mode (i.e. private transport or public transport) to the newly 41 introduced transport mode. As bus is the main public transport in Indian cities, Arasan and 42

Vedagiri (2011) have shown that modal shift from personal vehicles to bus will increase on 43 introduction of bus priority measure like an exclusive bus lane on Indian city roads. Binary 44

logit model was developed for predicting the modal shift from private cars to buses. 45 Variables such as gender age, walking time to bus stop and trip purpose are considered and 46 are found to be significant. In countries like China, Wang et al.(2012) study has shown that 47

due to implementation of BRT, modal shift to BRT in Chinese cities has increased 48 significantly. It was found that travel time saving, trip distances variables increases the 49

probability of modal shift. Similar studies like Nurdden et al. (2007) have shown that age, 50

Mane, Sarkar, Arkatkar

4

gender, car ownership, travel time, travel cost, household size and income are the significant 1 parameters for predicting modal shift from private car to public transportation in Malaysia. 2

Among the all variables, travel time, less distance from home to public transport and 3 subsidised fare are the most important variables for encouraging people to use public 4

transport. Andrade et al. (2006) found that increase in travel time of subway resulted into 5 modal shift from subway towards bus and automobile. However, on the other hand, Davsion 6 and Knowles (2006)have indicated that factors like subsidies for car travel and non-7

familiarity with BRT system has resulted into negative impact on modal shift to BRT. 8 Similarly Kittelson and Associates (2007) have shown that implementing BRT service will 9

have positive impact on modal shift of passengers to buses. The study has also underlined 10 that majority of these modal shift studies were done by Stated Preference (SP) survey. 11 12

In all these modal shift studies, disaggregate models were developed. Warner (1962) was a 13 pioneer of the binary choice model. The approach became popular in 1970’s and is now the 14

most commonly used methodology for modal shift study. However, Fagri and Hua(1991) 15 researched and explored the use of artificial intelligence system (ANN) as a frame work using 16 which many traffic and transport problems can be solved. Hensher and Ton (2000) have 17

compared ANN and nested logit model for mode choice analysis. Similar studies Nijkamp et 18 al. (1996), Shmueli (1996), Mozolin et al. (2000) have proved that both logit model and ANN 19

demonstrates a good predicative capability for modal shift study and hence, can be used. 20 From the literature review it may be noted that, both binary logistic model and ANN 21 technique are effective for modal shift study. Also, most of the studies conducted in India and 22

abroad are based on SP survey. Hence, at present to the best of the author’s knowledge there 23 is no any study conducted regarding modal shift using Revealed Preference (RP) survey in 24

India. In this study, RP survey is used which gives the perception of passenger's mode choice 25 behaviour. Moreover, very few models are developed for Indian cities and also impact of a 26 unique decision to allow introduction of cars in an exclusive BRT lane on modal shift has 27

never been studied in India. 28 29

3.0 OBJECTIVES AND SCOPE OF THE STUDY 30 31 The objective of the present study is to analyse shifting behaviour of passengers to newly 32

implemented BRT services as well as to analyse shifting behaviour of passenger to BRTS, 33 after judicial decision of allowing plying of private cares in exclusive BRT lane. For this 34

study, data was collected by Revealed Preference (RP) survey. In addition to this, the study 35 focuses on developing a shifting behaviour model using ANN technique as well as traditional 36 binary logistic model. By comparing the new technique with traditional technique, new 37

experiences in modal shift for modelling has been explored. The analysis is intended to 38 identify factors influencing modal shifts to BRT and also to quantify the extent for each 39

significant factor contributing to modal shifts. Figure 1, provides the flow chart for procedure 40 followed to study modal shift analysis in this study. 41 42

4.0 STUDY AREA 43 44

In May 2013, Indore (Population of 1.960 million in 2011), a medium size city located in 45 state of Madhya Pradesh, has introduced BRT system which was implemented based on the 46 experiences from Ahmedabad BRT model. This AB Road BRT pilot corridor in Indore runs 47

along 11.5 km and is currently having 21 bus stops placed at approximately 550m distance 48 without any grade separation at intersection. However, BRTS is equipped with signal priority 49

system controlled from system control centre. The trial run operation was launched in May, 50

Mane, Sarkar, Arkatkar

5

2013 and within a few months popularity of BRTS was witness which was resulted by 1 weekly increase in its ridership. Total nine buses were introduced for BRT corridor with 2

frequency of 10 min. Ticket fare is fixed to be varied with distance in kilometres and was 5, 3 10 and 15 rupees for 0-2 km, 2-8 km and 8-13 km respectively. Unfortunately, on 3rd October 4

2013, judicial authority, decided to allow the private cars to run in exclusive bus- lane 5 designated for BRT system. The decision was given in favour of the petition filed against 6 BRT system by a social activist. The social activist questioned the basis on which private cars 7

were not allowed in the exclusive BRT lane. Since, BRT system was introduced in 8 developing countries in the last decade and also due to the unavailability of particular 9

legalized documents (Codal Practice) to support it. Due to the lack of evidence, the 10 competent judicial authority allowed the private cars in the exclusive BRT lane without 11 charging a single penny for use of BRT lane. Private cars were allowed in BRT lane 12

throughout whole day-and-night with only restriction on overtaking, which resulted in 13 changing of exclusive BRT lane into mixed traffic flow condition. Hence, this unique policy 14

implementation resulted into drastic change on BRT ridership. BRTS route map and traffic 15 scenarios in exclusive bus lane before and after introduction of cars is depicted in figure 2. 16 17

During this study, en-route on board survey data was collected while the buses were running 18 in designed exclusive bus lanes in this corridor. The data provides original analysis of 19

changes in travel behaviour in the medium size city due to BRT deployments. The study also 20 attempts to contribute to develop more understanding related to the BRT-induced modal 21 shifts using binary logistic analysis and ANN. 22

23

24 25

FIGURE1Flow chart of procedure followed in modal shift analysis study 26

Mane, Sarkar, Arkatkar

6

1

NOTE: 1. Advanced Vehicle 2. Enhanced Bus Stop 3. Barrier 4. Exclusive Lane 5. Ramp 2

Figure 2(a) Traffic Scenario before and after introduction of cars in BRT lane 3

4

5 FIGURE 2 (b) BRTS route map 6

Figure 2 Traffic Scenarios and Route of BRTS 7 8

Mane, Sarkar, Arkatkar

7

5.0 DATA COLLECTION 1 Questionnaire survey format were administered to collect the relevant data and information 2

from BRT users and transit performance data from the operators. The questions were targeted 3 on passenger’s socioeconomic (Gender, Age, Occupation) and travel characteristics. Data on 4

travel characteristics; focused on information related to: (i) individual’s transport mode 5 choice before BRT implementation, (ii) travel time details before-and-after switching to 6 BRT, (iii) travel time details after introduction of cars in BRT lane and (iv) transport cost per 7

day before and after switching to BRT was collected. 8 9

The peak-hour passenger travel data along BRT corridor was collected before conducting 10 revealed preference (RP) survey. Peak hour passenger flow is identified as 2600 11 passengers/hr. The sample size was fixed at approximately 9% of peak-hour passenger flow, 12

which is considered as sufficient based on the studies by different researchers taking the 13 sample size ranging from1.25% to 2.5% of peak passenger flow (for example: Wang, 2012). 14

In this study, out of the total 300 questionnaires which were randomly distributed during peak 15 hours, 236 samples (78.67%) were found to be valid for further analysis. 16 17

5.1 Demographic Profile of respondents (N=236) 18 In this study, firstly the BRTS users were divided into two categories: captive users and the 19

users who have shifted from paratransit and personal vehicles like two-wheelers. Captive 20 users are the users who are confined to use particular public transportation, since they do not 21 own a personal vehicle. Majority of the passenger shifting to BRT are found to be from 22

paratransit mode contributing a percentage is more than 50%. However, captive users are 23 around 30%. It was also observed that car owners have not shifted to use BRTS. The reason 24

to this could be that the car ownership is considered as status symbol in India. Also, the 25 existing Level-of-Service (LoS) may not be acceptable to majority of car users. Hence, it may 26 be difficult to convince them to shift towards public transportation because of lack of 27

adequate LoS. 28 29

5.1.1 Socio-economic Parameters 30 31 Gender 32

Passengers are contacted randomly and are requested to participate in the survey with an 33 intention of getting samples keeping in view the factors like ratio of male-to- female ratio and 34

the balance of their age distribution. Out of the 236 samples, male and female respondents 35 are 75% and 25%, respectively.The male-female ratio of respondents (3.00) is considerably 36 higher than the male-female ratio of Indore city (1.08) and working class of male-female ratio 37

is 5.0 in 2011(Census of India, 2011). One possible reason for the higher ratio could be that 38 in developing countries the head of family is generally male and travels more. 39

40 Age Group and Occupation 41 As far as traveler’s age distribution is considered, approximately 38% and 42% of the 42

passengers are from the age group of 15 to 20 years and 21 to 30 years, respectively.It is also 43 observed that majority of the passengers are from students community. In this study, people 44

as young as 15years of age are included since in India, after 10th grade of education level i.e. 45 high school level, students are encouraged to travel on their own independently. Moreover, 46 BRT corridor alligned from residential area (scheme78,Shalimar Township,Vishnupuri, 47

Rajeev Gandhi as shown in figure 2) towards educational and commercial areas (Industrial 48 House, Palasia AICTSL,Intra Prtima Bhawar kua BRT stops as shown in figure 2). Major 49

schools and colleges are located near Bhawar kua,AICTSL,Satya sai, matahujri BRT stops. 50

Mane, Sarkar, Arkatkar

8

Hence,more students are attracted towards usage of BRTS due to its proximity, reliablity and 1 also low fare. Thus, BRTS corridor gives access to educational and commercial area from 2

both sides of corridor. It shows that BRTS is used by both working and non-working 3 (students) class. Hence, the sample size of 236 with respect to male-female ratio (between 4

5.0 and 1.08) also 9% of peak-hour ridership is considered to be significant as a 5 representative samples to model the shift before-and-after the implementation of BRTS. 6 7

Ridership 8 The distribution of sample size is based on the socio-economic characteristics and mode 9

choice (i) Paratransit, (ii) Bus, and (iii) Two wheeler is shown in figure 3. Ridership data of 10 BRTS before-and-after introduction of cars in BRT lane was collected from government 11 agency, which would be helpful for evaluating validity of the developed models. After 12

allowing introduction of cars in BRT lane by competent judiciary system, it was observed 13 that average daily ridership dropped to 24,000 passengers/day from 30,000 passengers/day 14

i.e. 20% decrease in ridership. 15 16 6.0 STUDY METHODOLOGY 17

18 In the present study two approaches, namely, binary logistic model and ANN were adopted. 19

Binary logistic model is widely used for transportation research and ANN is relatively 20 emerging technique for computation of modal shift model. Both these techniques are 21 discussed in the following sections. 22

23 6.1 Binary Logistic Model 24

25 The choice behaviour model developed in this study is based on the random utility theory. In 26 this theory it is assumed that people will select mode that maximizes the ir utility (U). In this 27

theory, utility function can be treated as a random variable and is defined calculated using 28 equation (1).Thus utility of any mode; m is comprised of two terms: deterministic term (V) 29

and an error component (E) as explained in equation (2) 30 31 Umi= β1xmi1+ β2xmi2 +........+ βkxmi (1) 32

Where Umi is the net utility function for mode m for individual i 33 Β1, .....,βk are k number of attributes of mode m for individual i and 34

xmi1,.......,xmik are the number of coefficients (or weights assigned to each attribute) 35 which need to be inferred from the survey data. 36 37

Umi = Vmi+Emi (2) 38 Where Vmi is the systematic component (observed) for utility of mode m for individual i; 39

Emi is the error component (unobserved) for utility of mode m for individual i. 40 41

42

43 44

Mane, Sarkar, Arkatkar

9

1 FIGURE3(a) Percentage distribution of age group 2

3 FIGURE3(b) Percentage distribution of occupation groups 4

5 FIGURE3(c) Various users diverted to BRT from other modes 6

FIGURE3 Socioeconomic and transport modes distribution of collected data. 7

0

10

20

30

40

50

60

70

15-20 21-30 31-50 50-+

Pe

rce

nta

ge,%

Age Distribution (Years)

%male

%female

0

10

20

30

40

50

60

70

80

Student Goverment employee

Private Sector Employee

Business Persons

Pe

rce

nta

ge,%

%male

%female

0.0

10.0

20.0

30.0

40.0

50.0

60.0

70.0

Paratransit Bus Two wheeler

Pe

rce

nta

ge.%

% Male

% Female

Mane, Sarkar, Arkatkar

10

Binary logistic model under discrete choice methods is commonly used for modal shift study 1 in the area of transportation planning, since they are able to model complex choice behaviour 2

of people with mathematical techniques. The mathematical framework of logit model is 3 based on the theory of utility maximisation and is discussed in detail by Ben-Akiva and 4

Lerman(1985). Briefly presenting the formulation, it is the probability of an individual i 5 selecting a mode n, out of M number of total available modes, is explained in equation (3), 6 7

Pin =exp(Uin)

exp (Uim )𝑚 ε𝑀 (3) 8

Uin is the utility function of mode n for individual i; 9

Uim is the utility function of any mode m in the choice set for an individual i; 10 Pin is the probability of individual i selecting mode n; and 11 M is the total number of available travelling modes in the choice set for individual i. 12

13 6.1.1 Calibration of Binary Logit Model 14

15 For the present study, IBM SPSS version 22.0 software was used for modelling of mode 16 choice and shifting behaviour analysis using binary logistic method. For the purpose of 17

analysis, socioeconomic parameters (age, gender, occupation), travel time saving, cost saving 18 and travel time increase were considered as attributes for developing utility equations. 19

Description of the variables considered for modelling are explained in Table 1. 20 Travel time saving, cost saving and increase travel time (both waiting time and in-vehicle 21 travel time) for individual passengers are calculated using the following formulations. 22

23 For model-1: BRT model 24

i. Travel Time Saving = (Travel Time in existing BRT service) - (Travel time in earlier 25 transport mode) 26

ii. Cost saving (Rs per day) = (Transport cost in existing BRT service) - ( Transport 27

cost in earlier transport mode i.e. before BRT service) 28 For model 2: BRT model after introduction of cars in BRT lane. 29

iii. Increase in Travel time = (Travel Time in existing condition of cars in BRT lane) - 30 (Travel Time during exclusive BRT service) 31

32

6.1.2 Dependent Variables 33 34

Model 1-BRT Model; Dependent variables is defined as dichotomous either as zero or one. 35 Captive users are assigned zero and all other passengers are assigned one, since they have 36 shifted from different transport modes (i.e. Paratransit, Two wheelers, etc.) to BRT. 37

Model 2-BRT Model after introduction of cars in BRT lane; It is assumed that if the travel 38 time in BRT buses after allowing cars in BRT lane is more than the travel time in earlier 39

transport mode i.e. before BRT service then passengers will not continue using BRT service. 40 Hence, dependent variable was assigned as zero to individual passenger who falls under this 41 category and for others dependent variable was assigned as one. 42

43 6.1.3 Significance of parameters 44

45 The p-value is a statistical measure for finding the significance of considered attributes in 46 binary logistic model. In other words, p-value signifies whether a particular variable is 47

contributing in model or not significantly by rejecting the null hypothesis at considered level 48

Mane, Sarkar, Arkatkar

11

of significance. For variable to be significant at 95% confidence interval p-value should be 1 less than 0.05. 2

TABLE 1 Variables and Types Defined 3

Variables Type of variable defined in IBM

SPSS Remark

Gender Dummy variable 1-Female,0-male

Age Categorical Variable

1 - (15-20 year) age group,

2 - (21-30 year) age group,

3 - (31-50 year) age group,

4 - (50plus year) age group

Occupation Categorical variable

1 for Student;

2-Goverment employee;

3-Private Employee; 4-Business Person

Travel Time (Saving

and Increase) Continuous variable Measured in minutes

Cost Saving Continuous variable Measured in Rupees

Note- For analysis, baseline has to be defined in categorical variables. In this study, 15-20 4 age group and students are defined as baseline. 5

6.1.4Goodness-of-fit of Model 6 7

The goodness-of-fit for the developed model can be assessed by likelihood ratio index (ρ2) 8 which is explained in equation (4). 9 10

ρ2 =𝐿𝐿 0 −𝐿𝐿 𝑃

𝐿𝐿 0 (4) 11

12 Where, LL (P) = Log-Likelihood of the estimated model; LL (0) = Log-Likelihood when the 13

coefficients are assumed to be zero. 14 15

6.2 Artificial neural network (ANN) 16 17 ANN is a computing system made up of a number of simple, highly interconnected nodes or 18

processing elements (neurons), which process information by its dynamic state response to 19 external inputs. It is an information-processing paradigm that is inspired by the way 20

biological nervous systems such as the brain, process information. ANNs, like people, learn 21 by example. An ANN is configured for a specific application through a learning process. The 22 goal of ANN is to map a set of input patterns onto a corresponding set of output patterns. The 23

network accomplishes this mapping by first learning from a series of past examples defining 24 sets of input (input layer) and output (output layer) correspondences for a given system. The 25

network then applies what it has learned to a new input pattern to predict the appropriate 26 output i.e. hidden layer. Thus neural network structure consists of a number of successive 27 layers of neurons: input, hidden and output which are shown in figure 4. 28

29

Mane, Sarkar, Arkatkar

12

1 FIGURE4 Details of Artificial Neural Network 2

3 6.2.1 Calibration Procedure of ANN 4

5 ANN learns using an algorithm called back-error propagation, which is explained in this 6

study (David, 2006). First Input Nodes, Hidden Nodes and Output Nodes are selected. After 7 that hidden layer and outer layer are activated using activation function. Activation Function 8 links the weighted sums of units in a layer to the values of units in the succeeding layer. In 9

this study, sigmoid function is used between input and hidden layer for forward network 10 activation (similar to binary logistic model), followed by softmax function between hidden 11

layer and outer layer since dependent variable should be defined as either 0 or 1. 12

Sigmoid function has the form: γ(c) =1

1+𝑒−𝑐 13

It takes real-valued arguments and transforms them to the range (0, 1).This is the similar 14 function used in binary logit model. 15

Softmax function has the form: γ(ck) =𝑒𝑐𝑘

𝑗𝑒 𝑐𝑗 16

It takes a vector of real-valued arguments and transforms it to a vector whose elements fall in 17 the range (0, 1). By using back error propagation algorithm, error signal and weights are 18 adjusted between outputs to hidden layer. Finally for minimization of error all the above steps 19

are repeated using new training patterns. For this study, IBM SPSS software is used. In this 20 study, Multilayer perceptron(MLP) isused.MLP is feed-forward ANN model that maps sets 21

of input layer onto a set of appropriate outputs.MLP utilizes a supervised learning technique 22 called back propagation for training of network which is explained in detail in (David,2006) 23 this study. Similar to binary logit model, due to the use of sigmoid and softmax as activation 24

function in ANN, probability of individual passenger is estimated at output layer. 25 Unlike binary logit model, after every iteration ANN develops a new model. To consider this 26

effect, number of iterations were done and the best model is selected based on the Mean 27 square error (MSE) (i.e. indirectly maximum accuracy) obtained during Testing of neural 28 network. In this study, while developing the ANN model accuracy of model was not 29

changing significantly in 20 iterations. 30 31

32 33

Mane, Sarkar, Arkatkar

13

7. RESULT AND DISCUSSION 1

7.1Binary Logistic Model 1: BRT model 2

3

The results using IBM SPSS 22.0 software for modelling a modal shift from private and 4 semi-private vehicles such as two wheelers, three wheelers, etc before introduction of cars in 5 exclusive bus lane are summarised in Table 2 and are detailed out in subsequent sections. 6

7 7.1.1 Result interpretation 8

In BRT model (before introduction of private cars), overall model and considered attributes 9 are quantified using statistical test like log-likelihood ratio index and p-value test 10 respectively, which is represented in Table1. All considered attributes (socioeconomic and 11

travel time) are found to be significant and influencing the modal shift to BRT form private 12 and semi-private vehicles. 13

Positive sign of Constant variable implies that passengers prefer to shift to BRT. The 14 Demographic variables show similar impact due to BRT service. The estimated coefficient 15 for gender using BRT model was observed to be negative, implying that male passengers are 16

more willing to shift to BRT service as compared with female passengers. The odds ratio, as 17 given in Column 5 of Table 2, is found to be increased by approximately 0.798 times for 18

female as compared with male. The odds of shifting to BRT are 20.2% lower for female than 19 that of male. The coefficients for age group variables are negative in the range of 21 to 30 20 and 50 plus year old passengers suggesting that they are less likely to shift to BRT as 21

compared to the passengers in the age group of 15 to 20 years. The odd ratio increased by 22 0.653 and 0.465 times for 21 to 30 and 50 plus year age group passengers, respectively as 23 compared to the passenger’s in the age group of 15 to 20 years. 24

25 TABLE2 Estimated Results for Binary Logit Model 1 26

Column 1 Column 2 Column 3 Column 4 Column 5

Variables Coefficient(B) S.E. Sig.(p-value) Exp(B)

Gender (X1) -.226 .084 .007 .798

Age

.000

21-30 (X2) -.426 .083 .000 .653

50 plus (X3) -.766 .281 .006 .465

Occupation

.021

Govt. Employee (X4) -.251 .135 .014 .778

Private employee (X5) -.304 .104 .004 .738

Travel time saving (X6)

.011 .003 .000 1.011

Cost saving (Rs/day) (X7)

.036 .003 .000 1.036

Constant .913 .067 .000 2.492

Likelihood Ratio Index (ρ2) =0.23

Note: (1 rupee = 0.016$ on 30-Nov-2013). 27

Mane, Sarkar, Arkatkar

14

However, the coefficients for government sector employee and private sector employee came 1 out to be negative compared with students group. The odd ratio increased by 0.778 and 0.738 2

times for government sector employee and private sector employee compared with student 3 group, suggesting, that students are more willing to shift to BRT service i.e. odds of shifting 4

to BRT lowered by 22.2% and 26.2% for government sector employee and private sector 5 employee compared to the students. 6 7

The coefficient of travel time saving is positive and significant indicating that the increase in 8 travel time saving would increase the probability of passengers shifting to BRT service when 9

all other variables are kept constant. The probability of shifting to BRT is increased by 1.1% 10 for every single minute travel time saving. Similarly, the coefficient of cost saving is positive 11 implying that increase in travel cost saving will increase the probability of passengers shifting 12

to BRT. The probability of passengers shifting to BRT is increased by 3.6 % for one rupee (1 13 rupee=0.016$ on 30-Nov-2013) saving in travel cost, when all other factors are kept constant. 14

Thus considering socioeconomic and travel detail attributes are found to be significant and 15 contributing in shifting behaviour of passengers from their earlier transport mode to BRT. For 16 overall of goodness-of-fit of model, likelihood ratio (ρ2) is generally used. The acceptable 17

ρ2value ranges from 0.2 to 0.4 (Alvinsyah et al., 2005).Thus this model is considered to be 18 satisfactory fit model. In addition to this, model as whole is considered to be significant by its 19

overall accuracy of predicting desired output, which in this case (67.8% ) found to be 20 satisfactory. 21 22

7.1.2 Probability: Utility function is defined for BRT model as given in equation 5. Using 23 the following equations (5) and (6), probability of individual passenger was found out. 24

Overall probability of switching to BRT service from other mode of transport (private and 25 semi-private) BRT model is 64.4%. 26

Utility Function=f(x) = 27

0.913-0.226*X1-0.426*X2-0.766*X3-0.251*X4- 0.304*X5+0.011*X6+0.036*X7 (5) 28

Probability equation,𝑃 =𝑒𝑓 (𝑥 )

1+𝑒𝑓 (𝑥 ) =1

1+𝑒−𝑓(𝑥 ) (6) 29

Where P=Probability, (variables are defined in Table 2). 30

7.2 Binary Logistic Model 2: BRT Model afte r Introduction of Car in BRT Lane 31 32 On 3rd October 2013, judicial system allowed private cars (Autos) in BRT lane with condition 33

of no overtaking. Due to this, average increase in waiting time and in-vehicle travel time was 34 observed by 5 minutes and 10 minutes respectively. In addition to this, the 20% reduction in 35

ridership was also observed. The results using IBM SPSS 22.0 software for modelling a 36 modal shift from private and semi-private vehicles such as two wheelers, three wheelers, etc 37 after introduction of cars in exclusive bus lane are summarised in Table 3 and are detailed out 38

in subsequent sections. 39 40

7.2.1 Result Interpretation (BRT model after introduction cars in BRT lane) 41 In this BRT model:2 (after introduction of private cars in exclusive BRT lane), overall model 42 and considered attributes are quantified using statistical test like log- likelihood ratio index 43

and p-value test respectively, which is represented in Table1. All considered socioeconomic 44

Mane, Sarkar, Arkatkar

15

and travel time attributes, except cost saving variable, are found to be significant and 1 influencing the modal shift to BRT form private and semi-private vehicles. 2

3 Positive sign of constant indicates that passengers are still willing to continue BRT service 4

after introduction of cars in BRT lane. Negative sign of gender indicates that male passengers 5 are still willing to continue BRT service. The coefficient for age group 21 to 30, 31 to 50, 50 6 plus were negative indicating that this age group is less willing to continue BRT service as 7

compared with the 15-20 year age group. However, Government sector employee and 8 business persons are more willing to continue BRT services as compared with students. This 9

can be attributed to more shifting of students to their earlier transport mode compared to 10 other occupational groups. Since major schools and colleges are located nearby this corridor 11 and hence, student percentage of using this BRT service out of the total passengers is more as 12

compared with other groups of passengers. Due to shifting of students to their earlier mode, 13 logit model shows the increase in government employee and business persons to continue 14

BRT service. This might be the limitation of logit model. 15 16

TABLE3 Estimated Results for Binary Logit Model 2 17

Column 1 Column 2 Column 3 Column 4 Column 5

Variables Coefficient(B) S.E. Sig.(p-value) Exp(B)

Gender (X1) -0.529 .093 .000 .589

Age

.000

21-30 (X2) -0.812 .096 .000 .444

31-50 (X3) -1.401 .171 .000 .246

50 plus (X4) -0.808 .308 .009 .446

Occupation

.000

Government Employee (X5) 1.638 .163 .000 5.143

Private Employee 0.201 .110 .068 1.223

Business (X6) 2.521 .251 .000 12.447

Increase In Waiting Time (min) (X7)

-0.159 .011 .000 .853

Increase in In-Vehicle Time

(min) (X8) -0.091 .005 .000 .913

Cost saving -0.001 .002 .734 .999

Constant 1.860 .103 .000 6.422

Likelihood Ratio Index (ρ2) =0.38

Note: (1 rupee = 0.016$ on 30-Nov-2013). 18 19

The coefficients of increased waiting time and increase in-vehicle time are negative implying 20 that increase in travel time parameter would reduce the probability of continuing the BRT 21 services. Probability of passengers continuing the BRT service will reduce by 14.7% and 22

8.7% for every minute increase in waiting time and in-vehicle time respectively. However, 23 cost saving parameter is not found to be significant since p-value is more than 0.05. Thus 24

except cost saving and private employee variable all other attributes are significant and 25 contributing in shifting behaviour of passenger from BRT to their earlier transport model due 26 to introduction of cars in BRT lane which resulted in increased travel time. The ρ2 values 27

around 0.4 may be considered as excellent fit (Ortuzar and Willumsen, 2004).Thus this model 28

Mane, Sarkar, Arkatkar

16

is considered to be good fit model. In addition to this, overall accuracy of this model came 1 out to be 72.9% which is acceptable. 2

3 7.2.2Probability: Utility function is defined for BRT model as given in equation 5. Using the 4

following equations (7) and (8), probability of individual passenger was found out. Overall 5 probability of passengers continuing the BRT service after introduction of cars was obtained 6 to be 46.6%. 7

8 Utility function =f(x) = 9

1.860–0.529*X1-.812*X2-1.401*X3-0.808*X4+1.638*X5+2.521*X6-0.159*X7-0.091*X8 (7) 10 11

Probability equation:𝑃 =𝑒𝑓 𝑥

1+𝑒𝑓 𝑥 =1

1+𝑒−𝑓 𝑥 (8) 12

Where P=Probability, the variables are defined in Table 3.While comparing Model 1 and 13

Model 2, it was observed that the probability of switching to BRT was decreased by 17.8% 14 when evaluated using binary logistic analysis. This decrease was attributed to increase in 15

waiting time and in-vehicle travel time to the introduction of cars in BRT lane. In addition to 16 this, 20% reduction in ridership was observed. Thus, both the models are realistic and able to 17 predict modal shift to BRT from other modes reasonably well. 18

7.3ANN Model 1: BRT Model 19

20 IBM SPSS 22.0 software is used to evaluate analysis of neural network. Model summary of 21 multilayer perceptron neural network is discussed in the subsequent section. Out of 236 22

samples 159 samples (approximately 70%) were assigned as training sample and remaining 23 77 (approximately 30%) cases were used as the testing sample. 24

25 TABLE4 ANN Information and Model Summary 26

Network

Information

Input Layer

Factors

1 Gender

2 Occupation

3 Age

Covariates 1 Cost saving

2 Travel Time saving

Number of Units 12

Rescaling Method for Covariates Standardized

Hidden Layer(s)

Number of Hidden Layers 1

Number of Units in Hidden Layer 1 8

Activation Function Sigmoid

Output Layer

Dependent Variables 1 BRT

Number of Units 2

Activation Function Softmax

Error Function Cross-entropy

Model

Summary

Training

Cross Entropy Error 71.922

Percent Incorrect Predictions 25.8%

Stopping Rule Used 1 consecutive step with no decrease in

error

Training Cross Entropy Error 32.577

Percent Incorrect Predictions 24.7%

Mane, Sarkar, Arkatkar

17

Table 4 provides information about the neural network model and is useful to ensure that the 1 specifications are appropriate. The number of units in the input layer is the sum of number of 2

covariates and the total number of factor levels. Sigmoidal function was selected for 3 activation function of hidden layer and softmax function was selected in output layer since 4

dependent variable has to be defined in the range of [0,1].The model summary displays 5 information about the results of training and applying the final network to the holdout 6 sample. Cross entropy error is displayed because the output layer uses the softmax activation 7

function. This is the error function that the network tries to minimize during training. The 8 percentage of incorrect predictions in testing samples helps to validate the neural network, 9

which is found to be 24.7% in this study. In other words, overall accuracy of this neural 10 network is found to be75.3%. 11 12

Figure 5 shows the input layer, hidden layer and output layer in neural network. Darker lines 13 indicate higher weights of corresponding input variables. This is quantitatively explained as 14

weights of variables as shown in Table 5.The importance of an independent variable is a 15 measure of predicting change in the network’s model predicted value for different values of 16 the independent variables. Normalized importance is simply the ratio of importance values to 17

largest importance value assigned to the variables in ANN modelling process and it is 18 expressed as percentage (%). From Table 5, it appears that variables related to cost saving 19

and travel time saving have the greatest impact on how the network classifies switching of 20 passengers to BRT service from their earlier transport mode. As discussed in ANN 21 calibration procedure, probability is estimated at output layer. Overall probability of 22

passengers switching to BRT service using neural network was obtained as 64.7% 23

24

FIGURE5 Flow chart of neural network: BRT model 25

26

Mane, Sarkar, Arkatkar

18

TABLE5 Independent Variable Importance 1

Variables Importance (weights) Normalized Importance

Gender .097 26.1%

Occupation .197 53.3%

Age .108 29.2%

Cost saving .370 100.0%

Travel Time saving .227 61.3%

2 3

7.4 ANN Model 2: BRT model when cars are introduced in BRT lane 4 5

Overall, BRT model using ANN analysis for modelling modal shift after cars are allowed to 6 ply in BRT lane is discussed below. In this analysis, out of 236 samples; 166 samples 7 (approximately 70%) were assigned as training sample and remaining 70 (approximately 8

30%) cases were hold aside as testing sample. Similar to earlier neural network model 1, all 9 parameters and activation functions were retained as it is in this network except the travel 10

time parameter as shown in Table 6. 11 12

TABLE6 ANN Information and Model Summary 13

Network

Information

Input Layer

Factors

1 Gender

2 Occupation

3 Age

Covariates

1 increase waiting time (ovtt)

2 Increase in-vehicle time

(invtt)

3 Cost saving

Number of Units 13

Rescaling Method for Covariates

Standardized

Hidden

Layer(s)

Number of Hidden Layers 1

Number of Units in Hidden

Layer 1 8

Activation Function Sigmoid

Output Layer

Dependent Variables 1 Bus PLUS CARS

Number of Units 2

Activation Function Softmax

Error Function Cross-entropy

Model

Summary

Training Cross Entropy Error 26.014

Percent Incorrect Predictions 7.8%

Stopping Rule Used 1 consecutive step(s) with no decrease in error

Testing Cross Entropy Error 17.927

Percent Incorrect Predictions 12.9%

Mane, Sarkar, Arkatkar

19

Similar to binary logit model 2(BRT Model after Introduction of Car in BRT Lane), in this 1 neural network increase in waiting and in-vehicle time variables are added as input layer. 2

Incorrect prediction percentage in the testing sample was found to be 18.9% which implies 3 that the overall accuracy observed for this network is 87.1% and which is fairly accurate. The 4

importance of independent variables is a measure of predicting change in the network’s 5 model-predicted value for different values of the independent variables. Normalized 6 importance is simply the ratio of importance values to largest importance value and is 7

expressed in percentage. Table 7 implies that in-vehicle travel time and cost saving 8 variables are the most influencing parameters for classifying the passengers who will 9

continue the BRT service even after introduction of cars in BRT lane. Table 6 and 7, overall 10 probability of passengers continuing the BRT service using neural network was found to be 11 45.7%. Thus, by comparing the both model: (i) BRT model (ii) BRT model- After 12

introduction of cars in BRT lane using ANN, it is observed that due to the introduction of 13 cars in exclusive BRT lane, modal shift is decreased from 64.7 to 45.7 ( i.e. 19% decrease) as 14

a result of increase in waiting time. 15 16

TABLE7 Independent Variable Importance 17

Variables Importance(weights) Normalized Importance

Gender .111 45.5%

Occupation .144 58.9%

Age .143 58.7%

Increase in waiting time

(inovtt) .172 70.4%

Increase in In-vehicle travel time(invtt)

.245 100.0%

Cost saving .185 75.5%

18

19 Table 8 provides the comparison of the accuracy percentage and probability derived using 20

both techniques. From table 8, it may be observed that for given situation, % accuracy of 21 different model is different but predicted probability and hence modal shift probability is 22 almost same with respect to the scenario. (Before and after introduction of cars in BRT lane) 23

24 TABLE8 Comparison of Accuracy Percentage and Probability 25

Parameters

Binary Logit Model Artificial Neural Network

BRT model Cars in BRT

lane model BRT model

Cars in BRT

lane model

Accuracy (%) 67.8 72.9 75.3 87.1

Significant

variables All

All Except cost

saving All All

Probability (%) 64.4 46.6 64.7 45.7

Likelihood Ratio Index (ρ2)

0.23 0.38 - -

Mane, Sarkar, Arkatkar

20

8.0 CONCLUSION 1 2

The study proves that, after implementation of BRT service in Indore, people are willing to 3 shift to BRT from paratransit and two wheelers. Travel time saving and cost saving are the 4

major factors that can be attributed for switching observed behaviour of passengers to BRT. 5 However, when the judicial advisory authority of honourable High court, decided to allow 6 autos (Private Cars) in BRT lane with full restrictions on overtaking; this has resulted in 7

increased travel time for BRT buses because of increased interactions between buses and 8 cars. Due to this increase, probability of passengers switching to BRT service is decreased 9

from 64.7% to 45.7%. Thus in order to enhance BRT system, Government may consider 10 reviewing the decision of allowing autos (Private cars) in BRT lane. In this study, data is 11 collected using RP survey. RP survey data gives the direct observations of choices made by 12

the individuals. Thus the results are pragmatic and reveal the perception of passenger’s mode 13 choice behaviour with respect to the change in policy. 14

15 The study also attempted to compare modal shift analysis using conventional binary logit 16 model and artificial neural networks technique. Both binary logit model and neural network 17

are able to predict same probability values in both the cases. However, accuracy of ANN 18 model is relatively higher as compared to the binary logit model in both the cases i.e. before 19

and after introduction of cars in exclusive BRT lane. Some limitations were observed with 20 logit model and were comprehensively discussed. The p-value for statistical testing of 21 variable significance for inclusion or exclusion for the model is usually set to 0.05. In this 22

study, the cost saving parameter in binary logit model was found to be insignificant. Since, 23 after introduction of cars in exclusive BRT lane, there was no change in cost saving variable 24

and also shifting behaviour of passengers was modelled based on the travel time assumption, 25 which was discussed earlier and validated by percentage decrease in ridership (20%). 26 However, using ANN this cost saving parameter was found to be significant and very 27

important for predicting the modal shift. Hence it can be concluded that, ANN can be used 28 over conventional binary logit model for modal shift study and it may provide more accurate 29

results as demonstrated in this study. 30 31 32

33 34

35 36 37

38 39

40 41 42

43 44

45 46 47

48 49

50

Mane, Sarkar, Arkatkar

21

References 1 1. Arasan and Vedagiri (2011). Modelling Modal Shift from Personal Vehicles to Bus 2

on Introduction of Bus Priority Measure. Asian Transport Studies, Vol. 1, Issue 3, 3 2011, pp. 288-302. 4

2. Yuanqing Wang, Zhicheng Wang, Zongzhi Li et al. A study of modal shift to Bus 5 Rapid Transit in Chinese cities. In Journal of Transportation Engineering, Vol.139, 6 Issue 5, 2012, pp. 515-523. 7

3. Abdullah Nurdden, Riza Atiq O.K. Rahmat and Amiruddin Ismail. Effect of 8 Transportation Policies on Modal shift from Private car to Public transportation in 9

Malaysia. Journal of Applied Sciences, Vol.7, Issue7, 2007, pp. 1014-1018. 10 4. Andrade, K., U. Kenetsu and K. Seichi, 2006. Development of Transport Mode 11

Choice Model by Using Adaptive Neuro-Fuzzy Inference System. In Transportation 12

Research Record: Journal of the Transportation Research Board, No. 1977, 13 Transportation Research Board of the National Academies, Washington, D.C., pp. 8-14

16. 15 5. Davsion and Knowles.Bus quality partnership, modal shift and traffic congestion. 16

Journal of Transport Geography, Vol. 14, issue 3, 2006, pp. 177-194. 17

6. Kittelson and Associates, Inc. (2007).TCRP Report 118: Bus Rapid Transit 18 Practitioner’s Guide. The National Academies Press, Washington, D.C. 19

7. Warner (1962).Stochastic Choice of mode in urban travel: A study of binary choice. 20 Northwestern university press, Evanston, Illinois, 1962. 21

8. Faghri A., Hua J.Evaluationof artificial neural network applications in transportation 22

engineering. In Transportation Research Record: Journal of the Transportation 23 Research Board, No. 1358, Transportation Research Board of the National 24

Academies, Washington, D.C., pp. 71-80. 25 9. Hensher, D.A. and Ton, T.T. A comparison of the predictive potential of artificial 26

neural networks and nested logit models for commuter mode choice. Transportation 27

ResearchPart E, 36, 2000, pp. 155-172. 28 10. Nijkamp, P., Reggiani, A. and Tritapepe, T. Modelling inter-urban transport flows in 29

Italy: a comparison between neural network analysis and Logit analysis. 30 Transportation Research Part C, 1996, pp. 323-338. 31

11. Shmueli, D., Salomon, I. and Shefer , D. Neural network analysis of travel behaviour: 32

evaluating tools for prediction. Transportation Research Part C, 4, 1996, pp.151-166. 33 12. Mozolin, M., Thill, J.C. and Lynn Usery. Trip distribution forecasting with 34

multiplayer perceptron neural networks; a critical evaluation. Transportation 35 ResearchPart B, 34, 2000, pp.53-73. 36

13. Ben Akiva, M. and Lerman, S. Discrete choice analysis theory and application to 37

travel demand. MIT press, 1985, Cambridge, MA. 38 14. IBM Corp. Released 2013.IBM SPSS Statistics for Windows, Version 22.0.Armonk, 39

NY: IBM Corp. 40 15. IBM SPSS. Neural Networks 21,user guide 41 16. Alvinsyah, Soehodho, S., Nainggolan, P.J. (2005) Public transport user attitude based 42

on choice model parameter characteristics, (case study: Jakarta bus way system). 43 Journal of Eastern Asia Society for Transportation Studies, 6, 480-491. 44

17. Ortuzar, J.D., Willumsen, L.G. (2001) Modelling Transport, 3rd Edition, John 45 Wiley& Sons Ltd. England. 46 47

48