key: cord-0920453-czzljsk1
authors: Pawar, Digvijay S.; Yadav, Ankit Kumar; Akolekar, Ninad; Velaga, Nagendra R.
title: Impact of Physical Distancing due to Novel Coronavirus (SARS-CoV-2) on Daily Travel for Work during Transition to Lockdown
date: 2020-08-19
journal: Transportation research interdisciplinary perspectives
DOI: 10.1016/j.trip.2020.100203
sha: 384ffc98895b80d69b900ba089f1d0a263215c69
doc_id: 920453
cord_uid: czzljsk1

Abstract The outbreak of COVID-19 pandemic has resulted in change in both commute and personal travel patterns. Though, in India, lockdown was implemented from 25th March 2020, due to self-awareness and pandemic risk perception, change in commuter behavior was observed from the beginning of March 2020. The period from 15th to 24th March 2020 is considered as the transition phase of COVID-19 outbreak in India (i.e., between no lockdown and lockdown period). This study attempts to use a decision tree approach to investigate the modal preference of 1542 commuters in association with socio-economic and travel characteristics, and safety perceptions with respect to public and private modes during transition to lockdown due to COVID-19 in India. About 41% of commuters stopped travelling during the transition to lockdown phase, 51.3% were using the same mode of transport and 5.3% of commuters shifted from public to private mode. The study findings reported different interactions of factors influencing the decision to use public or private modes of transport for daily commuting during pandemic situations like COVID-19. Interestingly, safety perceptions (associated with personal health) of commuters did not play a significant role in their mode choice behavior during the transition phase. Though people perceived public transportation as unsafe over personal vehicle use, the actual commute patterns did not validate this due to a possible reason that commuters do not have enough alternative modes. Given the uncertainties in the decision making of the commuters regarding their travel behavior due to physical distancing, the insights from this study are important to policymakers and local transport authorities to understand the change in travel patterns.

1 As on 8 th August 2020, 19.5 million people around the world are affected due to the Novel 2 Coronavirus (SARS-CoV-2) pandemic that started in Wuhan city, China, in early December 3 2019 (Lipsitch et al., 2020; Sohrabi et al., 2020; Jiang et al., 2020; Huang et al., 2020) . In 4 addition to its severe effects on human health and life, the virus has potentially impacted the 5 transportation system (Rodríguez-Morales et al., 2020; Bogoch et al., 2020; Sobieralski, 2020) . 6 Avoiding physical contacts and reducing interaction between individuals (i.e., physical 7 distancing) became a compulsory norm in most of the countries (Vos, 2020; Wilder-Smith and 8 Freedman, 2020). India has restricted human movement by incorporating lockdown to ensure 9 physical distancing and promote self-isolation. In Indian context, lockdown was defined as the 10 situation where complete travel restrictions were imposed on general public except for the 11 supplies of essential goods and services (such as daily groceries), essential personnel (such as 12 police, medical professionals and media reporters) and for people facing medical emergencies. 13 In India, the first case of COVID-19 was detected on 30 th January 2020 in a university student An online questionnaire survey was designed to capture travel behavior information of 6 commuters before and during the transition period of COVID-19 outbreak. The responses were 7 collected from 18 th to 28 th March and the participants were specifically asked to fill data 8 corresponding to the transition to lockdown period (i.e., 15 th March to 24 th March 2020). The 9 questionnaire enquired about the commuters' socio-economic characteristics (such as age, 10 income, and city of residence) along with work-related travel characteristics (such as preferred 11 mode of transport, distance from home to work, travel time and frequency of traveling to work) 12 and health related safety perceptions corresponding to public and private modes. A detailed 13 description of the research organizations involved and motivation of this study was provided to 14 the respondents at the beginning of the questionnaire. Moreover, the participants were informed 15 to provide responses related to COVID-19 impacts on their travel behavior. Figure 2 shows the research methodology adopted in this study. The first step was to design the 1 questionnaire survey, followed by its distribution and data collection. After the survey, the 2 responses were filtered to remove the erroneous/incomplete data. Thereafter, preliminary 3 analysis was conducted to analyze the mode choice behavior of commuters and their safety 4 perceptions related to public and private modes. Further, the decision tree was developed by 5 adopting a 5-fold cross validation technique to identify the potential interactions of factors 6 influencing the mode choice. Finally, interpretation of the decision tree was discussed and 7 conclusions were outlined. In total, 1945 responses were obtained from the survey. Among these responses, 74 were found 10 to be erroneous and 118 were incomplete, which were filtered. As the present study focused on 11 examining the modal preference between public and private modes of transport, the survey data 12 for respondents who used transport modes other than public and private modes to reach the 13 workplace were excluded from the analysis. The present study analyzed the decision making of commuters related to selecting the public or 8 private modes during the transition period of COVID-19 outbreak. Previous research had shown 9 that the estimation efficiency of decision tree (DT) models is better than the multinomial logit 10 J o u r n a l P r e -p r o o f Journal Pre-proof models (MNL) and neural network (NN) models (Rudd and Priestley, 2017) . Therefore, a 1 decision tree approach was used to understand the underlying interactions among various 2 explanatory variables (shown in Table 1 ). Among these variables, distance and travel time might 3 be correlated with each other but may not necessarily be associated with a direct/linear 4 relationship. This can happen due to the level of traffic congestion faced by the commuters. To 5 account for such samples and cases, it is important to consider both the features simultaneously 6 rather than considering only one of them. During the analysis, we empirically tested all the three 7 possibilities: using only distance, travel-time and both features simultaneously, and found that 8 the model using both distance and travel time along with other features produced the most 9 promising predictive model.

Decision trees are constructed using a top-down approach in a greedy fashion. Specifically, at 11 each node in the decision tree, the training set is split into smaller subsets until the purity of the 12 nodes can no longer be increased. In this work, the Scikit-learn's implementation of 13 Classification and Regression Trees (CART) algorithm were adopted to build a binary decision 14 tree. We also experimented with other popular decision-tree training algorithms such as the Chi-15 square Automatic Interaction Detector (CHAID) algorithm and the Iterative Dichotomiser 3 16 (ID3) algorithm. However, analysis results using CART algorithm were found more promising 17 than its counterparts. CART algorithm uses the GINI impurity as the measure of diversity at a 18 particular node (Pedregosa et al., 2011) . The GINI impurity at a particular node k is given by the 19 following expression in Equation 1 (Breiman et al., 1984; Yao et al., 2019) :

Where, c denotes the number of classes and denotes the fraction of data points at node 22 belonging to class .

At a particular node, if the GINI index is zero, it implies that there is zero diversity in samples at 24 that node (i.e., all samples at that node belong to only one class). Therefore, an absolute split (0% 25 and 100%) is expected wherever the GINI index is zero. The absolute split at a particular node in 26 the decision tree indicates that all the samples (individuals) reaching that node (i.e. satisfying 27 decision rules from the root node to that particular node) made the same decision of choosing 28 that particular mode of transport.

To prevent overfitting the training data, the decision trees were pruned by setting a maximum implemented on the questionnaire dataset to obtain the best possible decision tree. The cross-1 validation technique used data more efficiently as every observation is used for both training and 2 testing. It was found that the decision tree with depth 5 and geometric mean score (G-mean) as 3 0.57 outperformed. G-mean score tries to maximize the accuracy on each of the classes while 4 keeping them balanced. The questionnaire enquired the travel behavior of commuters before and during the transition to 8 lockdown period. The respondents' preferred mode of travel between their place of residence and 9 workplace before the transition to lockdown period of COVID-19 outbreak is presented in Figure   10 3a. The figure indicates that motorized two wheelers account for the highest share (32.1%) 11 followed by personal cars (23%), buses (16.7%) and walk (12.8%) among the others.

In response to a question about respondents' travel mode choice during the transition to 13 lockdown phase of COVID-19 outbreak, 41.65% said that they stopped traveling during that 14 phase, 51.31% reported that they were using the same mode of transport as before, 5.3% shifted 15 from public to private mode, and remaining 1.74% switched from one public mode to another 16 ( Figure 3b ). People who shifted from one public transport mode to another comprised 44.4%, 17 38.9%, and 16.7% bus, train and metro-rail users respectively. showed that 75.5% feel unsafe, 16.3% slightly unsafe, 6.1% moderately safe, 1.7% safe and 2 remaining perceived it as very safe. Whereas, safety perception of private modes of transport 3 revealed that 12% of the commuters perceive it as very safe, 41% as safe, 34% as moderately 4 safe, 10% as slightly unsafe and remaining 3% as very unsafe (Figure 4 ). 

The developed decision tree is shown in Figure 5 . The nodes with shades of blue color represent 1 the public mode selection and the nodes with shades of orange represent the private mode 2 selection by commuters. In each node of the decision tree, a particular condition is defined based 3 on which the subsequent possible node is decided. For example, the topmost node defines the 4 condition 'travel time ≤ 30 minutes' (Figure 5 ). If the condition is true, it will move to the 5 immediate left node, otherwise it will move to the immediate right node. The combinations of various factors influencing the selection of private mode of transport during 8 the transition to lockdown period due to COVID-19 outbreak are presented in Table 2 . The first 9 three interaction terms focused on the individuals whose travel time is less than or equal to 15 10 minutes. It further categorized them into two groups with income less than or equal to 12 lakh 11 rupees (1.2 million) per annum (Interaction terms 1 and 2) and income more than 12 lakh rupees 12 per annum (Interaction term 3 and 4). Among these commuters, the first interaction term 13 suggests that people with income less than or equal to 12 lakh rupees, who have to travel for less 14 than 5 days a week, are expected to use private mode of transport even if their perceived safety 15 shift is less (i.e. less perceived safety of private mode over public mode). The second interaction 16 term indicates that private mode of transport is preferred by people with age more than 32.5 17 years whose income is less than or equal to 12 lakh rupees and higher perceived safety shift (i.e.,

higher perceived safety of private mode over public mode). Under the same category of 19 commute time less than 15 minutes, and annual income higher than 12 lakhs, it was observed 20 that the use of private transport is preferred by younger people with age less than or equal to 27.5 21 years (as indicated by the third interaction term), and by individuals aged above 32.5 years as to 30 minutes who tend to use private mode of transport. Further, they are categorized on the 2 basis of distance travelled during commuting ( Figure 5 ). In this, the first category is of young 3 people (less than or equal to 32.5 years) who have to travel for less than 10 kilometers and less 4 than 5 days a week (Interaction term 5). The second category is of people with age more than 5 32.5 years earning high income (greater than 18 lakhs per annum) and need to travel for less than 6 10 kilometers (Interaction term 6). The third category is of commuters who have to travel for 7 longer distances (more than 10 kilometers) and they prefer private mode even after having 8 reduced perceived safety shift (Interaction term 7). The last category reports that the commuters 9 of age less than or equal to 27.5 years with higher perceived safety shift, and have to travel for 10 more than 10 kilometers for work, prefer private mode of transport over the public mode 11 (Interaction term 8). 

Interaction term 1 Commuters with travel time less than or equal to 15 minutes, income less than or equal to 12 lakh rupees per annum, safety shift less than or equal to 2.5, and frequency of travel less than 5 days per week Interaction term 2

Commuters with age more than 32.5 years, travel time less than or equal to 15 minutes, income less than or equal to 12 lakh rupees per annum, and safety shift higher than 2.5, Interaction term 3

Young commuters with age less than or equal to 27.5 years, income higher than 12 lakh rupees per annum, and travel time less than or equal to 15 minutes Interaction term 4

Commuters with age greater than 32.5 years, income higher than 12 lakh rupees per annum, and travel time less than or equal to 15 minutes Interaction term 5

Commuters with age less than or equal to 32.5 years, travel time between 15 to 30 minutes, travel distance less than or equal to 10 kilometers, and frequency of travel less than 5 days per week Interaction term 6

Commuters with age more than 32.5 years, income greater than 18 lakh rupees per annum, travel time between 15 and 30 minutes, distance less than or equal to 10 kilometers Interaction term 7

Commuters with travel time between 15 and 30 minutes, travel distance higher than 10 kilometers and perceived safety shift less than or equal to 1.5

Interaction term 8

Young commuters with age less than or equal to 27.5 years, travel time between 15 and 30 minutes, travel distance higher than 10 kilometers and perceived safety shift greater than 1.5 J o u r n a l P r e -p r o o f

1 Various interactions were observed from the decision tree designating the categorization of 2 commuters preferring public mode of transport during the transition period of COVID-19 3 pandemic. These interactions are illustrated in Table 3 . The ninth interaction term suggests that 4 young commuters with age less than or equal to 27.5 years living in Tier-3 cities, with travel 5 time more than 30 minutes, travel distance less than or equal to 30 kilometers and frequency of 6 travel less than 5 days a week preferred public transport. In the tenth interaction, within the same 7 distance category, commuters staying in Tier-1 and Tier-2 cities, with travel time between 30 8 minutes to 2 hours, and frequency of travel more than or equal to 5 days a week, are more likely 9 to use public transport ( Figure 5 ).

The eleventh interaction term suggests that individuals with age less than or equal to 32.5 years, 11 annual income less than or equal to 12 lakh rupees (1.2 million), travel time between 30 minutes 12 to 1 hour, and need to travel for more than 30 kilometers to their workplace, consider public 13 mode as favorable. The next two interactions involve the perceived safety shift of commuters. In 14 the twelfth interaction, the use of public transport was preferred by the commuters with travel 15 time between 1 to 2 hours, travel distance more than 30 kilometers, and reduced perceived safety Table 3 Interactions of explanatory variables for preferring public transport during COVID-19 1 transition to lockdown period 2

Interaction term 9

Young commuters with age less than or equal to 27.5 years living in Tier-3 cities, with travel time more than 30 minutes, travel distance less than or equal to 30 kilometers and frequency of travel less than 5 days a week Interaction term 10

Commuters staying in Tier-1 and Tier-2 cities, with travel time between 30 minutes to 2 hours, travel distance less than or equal to 30 kilometers, and frequency of travel 5 or more days a week Interaction term 11

Commuters with age less than or equal to 32.5 years, income less than or equal to 12 lakh rupees per annum, travel time between 30 minutes to 1 hour, and travel distance more than 30 kilometers

Interaction term 12 Commuters with travel time between 1 to 2 hours, travel distance more than 30 kilometers, and perceived safety shift less than or equal to 1.5

Commuters with age less than or equal to 42.5 years, travel time more than 1 hour, travel distance more than 30 kilometers, and perceived safety shift greater than 1.5

The relative importance of various interaction variables used in the decision tree analysis is 4 shown in Figure 6 . The importance (value) of a particular feature is computed as the 5 (normalized) total reduction of the criterion, in our case the GINI impurity, particular to that 6 feature. This is also known as the GINI importance. To capture the total reduction, the weighted 7 impurity decrease at a given node is first defined using the following Equation 2 (Pedregosa et 8 al., 2011) .

(2) 10 where, is the total number of samples, is the number of samples at the current node , 11 denotes the value of GINI impurity at node , and & denote the left and right 12 child nodes of node respectively.

The feature importance of a feature k, would then be given by Equation 3 as shown below 14 (Pedregosa et al., 2011) .

The higher the value of feature importance, the more important is the corresponding feature.

It can be observed that people give the highest importance to travel time followed by the distance 1 covered between the residence and workplace. The next deciding factors are age and income of 2 commuters, followed by the frequency of going to work, safety shift and the city of residence. commuters during the initial phase of COVID19 (John, 2020) . to other areas of mainland China (Chinazzi et al., 2020) . In South Korea, the transport ministry 5 created a national level integrated surveillance system to monitor contact tracing on a large-scale. 6 This strategy has been found effective in slowing down the spread of COVID-19 virus (Lee and 7 Lee, 2020). Based on these studies, it can be concluded that lockdown strategies are beneficial in 8 minimizing the spread of COVID-19 pandemic. This study is quite useful in understanding the decision-making behavior of commuters while 27 selecting their preferred mode of transport during a pandemic like COVID-19, which is a threat 28 to public health as well as economy of the world. The rapidly changing diaspora of the pandemic 29 is making human life more challenging. More research is required to investigate the public 30 concern over mode choice during disastrous situations like COVID-19.

J o u r n a l P r e -p r o o f

The study faced certain limitations. The survey data was collected using an online questionnaire. 

VNIT, IIIT, IIM ask over 4,000 students to vacate hostels due to Coronavirus. Times of India

Transit Has Been Battered by Coronavirus. What's Ahead May Be Worse. The New York Times

Use of public transport by older women in semi-urban West Bengal

Potential for global spread of a novel coronavirus from China

Classification and Regression Trees

Understanding Socioeconomic Disparities in Travel Behavior during the COVID-19 Pandemic

Modal share changes due to COVID-19: the case of Budapest

UK road travel falls to 1955 levels as Covid-19 lockdown takes hold. The Guardian

The effect of travel restrictions on the spread of the 2019 novel coronavirus (COVID-19) outbreak

How COVID-19 and the Dutch 'intelligent lockdown' change activities, work and travel behaviour: Evidence from longitudinal data in the Netherlands

Clinical features of patients infected with 2019 novel coronavirus in Wuhan

Coronavirus: Is your organization ready for work from home in India? The Hindu

What is Janata Curfew: A curfew of the people, by the people, for the people to fight coronavirus

How Prepared Is India to Control the COVID-19 Pandemic?

Review of the Clinical Characteristics of Coronavirus Disease 2019 (COVID-19)

World cities turn their streets over to walkers and cyclists. The Guardian

Testing on the move: South Korea's rapid response to the COVID-19 pandemic

Defining the Epidemiology of Covid-19 -Studies Needed

Coronavirus: Allow employees to work from home, states tell firms. Live mint

Recommendations of the Sixth Central Pay Commission -Decision of Government relating to grant of Dearness Allowance to Central Government servants

Ministry of Health and Family Welfare, Government of India

Impact of COVID-19 on transportation in

Changes in subway ridership in response to COVID-19 in Seoul, South Korea: Implications for social distancing

Scikit-learn: Machine learning in Python

The effect of control strategies to reduce social mixing on outcomes of the COVID-19 epidemic in Wuhan, China: a modelling study. The Lancet Public Health

Going global-Travel and the 2019 novel coronavirus. Travel medicine and infectious disease 33

A comparison of decision tree with logistic regression model for prediction of worst non-financial payment status in commercial credit

Coronavirus: Number of confirmed cases in India crosses 100. The Hindu

Comparison of life cycle environmental performance of public road transport modes in metropolitan regions

Devotees banned at Guruvayur temple in Kerala from

COVID-19 and airline employment: Insights from historical uncertainty shocks to the industry

World Health Organization declares global emergency: A review of the 2019 novel coronavirus (COVID-19)

The effect of COVID-19 and subsequent social distancing on travel behavior

Isolation, quarantine, social distancing and community containment: pivotal role for old-style public health measures in the novel coronavirus (2019-nCoV) outbreak

Classification of Fatigued and Drunk Driving Based on Decision Tree Methods: A Simulator Study