Contributing Factors to the Changes in Public and Private Transportation Mode Choice after the COVID-19 Outbreak in Urban Areas of China

Abstract

The COVID-19 pandemic has tremendously affected the whole of human society worldwide. Travel patterns have greatly changed due to the increased risk perception and the governmental interventions regarding COVID-19. This study aimed to identify contributing factors to the changes in public and private transportation mode choice behavior in China after COVID-19 based on an online questionnaire survey. In the survey, travel behaviors in three periods were studied: before the outbreak (before 27 December 2019), the peak (from 20 January to 17 March 2020), and after the peak (from 18 March to the date of the survey). A series of random-parameter bivariate Probit models was developed to quantify the relationship between individual characteristics and the changes in travel mode choice. The key findings indicated that individual sociodemographic characteristics (e.g., gender, age, ownership, occupation, residence) have significant effects on the changes in mode choice behavior. Other key findings included (1) a higher propensity to use a taxi after the peak compared to urban public transportation (i.e., bus and subway); (2) a significant impact of age on the switch from public transit to private car and two-wheelers; (3) more obvious changes in private car and public transportation modes in more developed cities. The findings from this study are expected to be useful for establishing partial and resilient policies and ensuring sustainable mobility and travel equality in the post-pandemic era.

1. Introduction

1.1. Background

Since the outbreak of COVID-19 in December 2019, the disease has been spreading rapidly around the world and having distressing impacts on human societies. As of early February 2023, over 750 million confirmed cases of COVID-19 had been cumulatively reported worldwide [1]. In response to the pandemic, governments have been taking various measures to slow down the disease spread and to mitigate myriad adverse impacts on society, including investment in vaccine projects and all sorts of non-pharmaceutical interventions against COVID-19. The latter mainly refers to travel restrictions and lockdowns, which have reshaped activity patterns. For example, on 23 January 2020, the authority announced that the airport and railway station in Wuhan had been temporarily closed [2]. In addition to these coercive policies, the perception of personal safety is also a significant factor in minimizing social contact and unnecessary trips. Consequently, the flow pattern and people’s travel behavior have greatly changed. For example, the number of public transport passengers in 36 major cities in China decreased by 33.7% in 2020 compared to the previous year [3].

So far, multiple studies have been conducted in many countries to uncover the fact that COVID-19 has significantly affected travel behaviors, including but not limited to Germany [4], Sweden [5], China [6,7], the United States [8], Greece [9], and Canada [10]. These findings are not surprising as many researchers anticipated significant reductions in activities and travel. Public transport is a source of virus transmission [11,12,13]. Owing to the crowded conditions of buses and subways, passengers inevitably reduced the distance between themselves and other passengers. Moreover, the risk of infection among those who are highly dependent on public transportation is higher [14]. Despite this, many travelers still use public transportation as their primary or sole means of transportation, which leads to unequal risks of virus transmission [15]. Although there has been a sharp decline in public transport passenger flow worldwide, the reduction in the flow of people during the pandemic varies greatly across socioeconomic groups [9]. To manage the spread of the virus and related inequalities, we need to strengthen our understanding of which groups continue to use public transport. Recently, researchers have analyzed the socioeconomic characteristics of passengers and have found that the decline in the number of bus passengers among grassroots workers, lower-income persons, and people of color during this period was considerably smaller than that of higher-income persons and white people [16,17,18]. However, no study to date has explored how individual sociodemographic characteristics are related to the change in travel behavior, particularly mode choice, in China. Thus, this study aimed to fill this research gap. It is anticipated that better policy suggestions for slowing down the pandemic spread and ensuring the functioning of human societies could be proposed based on the understanding of the specific relationship between individuals’ travel behaviors and their characteristics.

On account of the necessity of both individuals’ travel behavior data and the sociodemographic characteristics data, a questionnaire survey was conducted. On the grounds of the severity of the pandemic in China, an online questionnaire was designed, and the study periods were divided into before the outbreak, the peak, and after the peak. In this way, changes in travel mode choice between the three periods could be captured, which emerged from the transitions in the severity of the pandemic. The questionnaire design and period division are elaborated on in the next section. Having collected the disaggregated data, statistical analysis was used to describe the changes in the choice of different modes. Furthermore, a series of random-parameter bivariate Probit models was employed to achieve the main objective.

1.2. Literature Review

There has been a large body of research on how COVID-19 changed travel behavior. From a macroscopic level, the transportation system and flow patterns have been significantly affected by COVID-19. Andreana et al. found that COVID-19 caused a reduction of above 80% for air transport in all of the world’s macro-regions in May 2020 [19]. Tian et al. evaluated the influence of the pandemic on urban transportation in Canada according to the consumption of motor gasoline and found that the consumption continuously decreased to the lowest level in April 2020 [10]. Cui et al. specified the shocks caused by COVID-19 to both the supply side and demand side and found that the passenger transport sectors had larger output decreases than the freight transport sectors [6]. Wen et al. figured that electric vehicle sales in China decreased in the short term, but later seemed to perform better under the stimulation of the pandemic [20]. Goenaga et al. found that the average reduction in traffic volume was 27% in the state of North Carolina and the Commonwealth of Virginia, US [21].

From a microscopic viewpoint, the COVID-19 pandemic can change individual travel behavior, including mode choice. Hara and Yamaguchi revealed empirical data illustrating a reduction in trips and a low population density index even without strong restrictions by the Japanese government in the early stages of COVID-19 [22]. Zhang et al. analyzed how the pandemic changed local travel behavior by using second-by-second smartcard data of adults, students, children, and seniors on the Metro Transit Railway in Hong Kong, China [23]. Zhang and Fricker estimated the causal impact of COVID-19 on non-motorized travel patterns with a Bayesian structural time-series model [24]. The result showed that COVID-19 led to reductions in non-motorized activities in densely populated cities; by contrast, walking and bicycle activities in less densely populated cities increased. Bucsky reported the modal share changes in Budapest, Hungary, and found that public transport has seen the greatest decline by far [25]. Jenelius and Cebecauer analyzed daily public transport ridership in Sweden by using ticket validation, sales, and passenger count data, and the decrease in public transport ridership was observed to be the most severe [26]. Teixeira and Lopes proved the better elasticity of the bike-share system than the subway system after the COVID-19 outbreak in New York City [27]. Brough et al. investigated the magnitudes of and mechanisms behind socioeconomic differences in travel behavior during the COVID-19 pandemic in King County, Washington [28]. The study implied that disparities in travel behavior across socioeconomic groups may become an enduring feature. As the transport authorities reported about a 95% reduction in users during peak COVID-19, Naveen and Gurtoo developed two frameworks to enable transport enterprises to open new windows of travel and efficiencies for passengers [29].

It can be concluded from the above-mentioned studies that the pandemic has influenced the transportation system and travel behavior in many ways. However, the impact on different groups in different areas varies greatly. The differences are thought to originate from population, industry, and transportation infrastructure. Therefore, survey questionnaires have also been commonly used for disaggregated data collection to analyze the effects of COVID-19. In developed countries, the public transportation system is well established and had better resiliency during the COVID-19 pandemic. Beck and Hensher identified the initial impacts on travel patterns associated with the restrictions imposed by the Australian government in the first month [16]. Anke et al. confirmed the shift away from public transport to private transport in Germany and found that the impact of the lockdown on travel behavior was minimal [4]. Politis et al. found that the number of daily trips decreased by 50% [9], and trips on foot were increased while trips on public transport modal shares were heavily reduced in Thessaloniki, Greece based on two questionnaire surveys. De Haas et al. investigated the willingness to use different modes of transportation and the activity changes of people with different sociodemographic characteristics through longitudinal data in the Netherlands [30]. An interesting finding indicated the possibility of structural behavioral changes in the Netherlands, including the intention to telecommute and to reduce fly travel. Downey et al. conducted an online questionnaire in Scotland and found that several factors, including pre-lockdown travel choices, perceived risk of infection, household size, and region significantly affected the anticipated use of public transport [31]. The findings implied the potential loss of demand for public transport and the consequences for future equitable and sustainable mobility. Palm et al. drew on a March 2021 follow-up survey of over 1900 people who used public transport regularly before COVID-19 in Canada, and who took part in a prior survey on the topic in May 2020 [32]. Survey results show that COVID-19 may have increased the attractiveness of auto ownership, and the participants are likely to eventually purchase cars anyway.

In some developed countries with higher population density and lower public transport service level, the situation might be different. Shakibaei et al. reported the extreme transition in travel activity patterns and transport mobility in Istanbul and provided several policy implications through the results [33]. Bhaduri et al. quantified the effects of the sociodemographic characteristics of travelers on the mode-specific trip frequencies in India with multiple discrete continuous extreme value models [34]. The estimation results indicated a high propensity to shift to virtual and private modes. China’s epidemic prevention policy is one of the strictest in the world, and with a large population and a large traffic demand, the travel behavior after COVID-19 is expected to change drastically compared to developed areas such as Greece and the Netherlands. Zhang and Lee analyzed the interactive effects between travel behavior and COVID-19 by using the spatial autoregressive model and logit model [35]. They found a positive correlation between passenger volume and the cumulative confirmed cases in China and revealed significant factors affecting the changes in travel behavior, including gender, occupation, and travel restriction. Luan et al. explored how COVID-19 impacted travel mode choice and the intention of car purchase in China [7]. They concluded with the propensity for long-distance travel on public transport and the great impact of COVID-19 on ride sharing. However, this paper did not consider (1) individual features, but only mode features (e.g., travel time, cost) along with congestion; (2) the changes in different periods. To sum up, in areas with a dense population and high dependence on public transportation, public transport travel is more severely affected.

Although the existing studies have proved the impact of COVID-19 on the transportation system and individual mobility, there are still several research gaps. First, few studies to date have quantified the exact effects of the sociodemographic characteristics on individuals’ changes in each mode of transportation in China. Second, there is a lack of a more detailed inquiry regarding the impacts of the severity of the transitions of the pandemic on the changes in mode choice. Therefore, data collection and model establishment were carried out in this study given the above deficiencies, which will be elaborated on in the following sections.

2. Data

2.1. Data Collection and Survey Design

The data used in this study were collected by an online survey questionnaire from 17 to 31 August 2020. The responses were collected from 753 individuals across different cities in China. Among them, 568 valid responses were obtained after screening the invalid responses that took too little time (less than 90 s) or contained unreasonable or inconsistent answers. The trend of new confirmed cases in China is presented in Figure 1. To compare the changes in travel behavior in different stages of COVID-19, three periods were defined based on the new confirmed cases in China, namely (1) before the outbreak (before 27 December 2019), (2) the peak (from 20 January to 17 March 2020), and (3) after the peak (from 17 March to the date of the survey).

The questionnaire was designed into two folds. The first fold investigated the sociodemographic characteristics and the second fold was structured to reveal the dynamics of respondents’ mobility, including travel frequency and mode choice in three periods.

2.2. Sociodemographic Characteristics

The sociodemographic characteristics of the sample of 568 respondents are summarized in

Table 1. Sociodemographic characteristics of the sample.

Independent Variables	Categories	Frequency	Percent
Gender	Male	218	38.4%
	Female	350	61.6%
Age	16–24 years	294	51.8%
	>24 years	274	48.2%
Residence 1	First-tier cities	105	18.5%
	New first-tier cities	89	15.6%
	Second-tier cities	55	9.7%
	Third-tier cities	222	39.1%
	Rural	97	17.1%
Ownership of a transportation mode 2	Car	184	32.4%
	Scooter	253	44.5%
	Bicycle	165	29.0%
	None	143	25.2%
Occupation	Student	232	40.8%
	Office worker	148	26.1%
	Non-office worker	64	11.3%
	Self-employed	34	6.0%
	Freelancer	39	6.9%
	Retired/unemployed	51	9.0%

¹ By the time we collect the questionnaire (i.e., 17–31 August 2020), first-tier cities in China include four cities: Beijing, Shanghai, Guangzhou, and Shenzhen. For the other cities, please refer to https://www.yicai.com/news/100648666.html (accessed on 17 August 2020). ² During the question of ownership of a transportation mode, respondents answered “none” or answered multiple modes that were available.

.

2.3. Travel Frequency

The same question about travel frequency per week was asked in the three periods. Five different weekly travel frequency intervals were set as choices: 0, 1–3, 3–7, 7–10, and more than 10 times per week. Before the outbreak of COVID-19, more than 50% of the respondents traveled more than seven times per week, while the proportion during the peak was 3.17%, and it rebounded to 25.53% after the peak. In order to visualize the changes in travel frequency, the weighted average travel frequency of the sample was calculated following the equation below:

$$\mathrm{Weighted} \mathrm{average} \mathrm{travel} \mathrm{frequency}={\displaystyle \sum }\mathrm{The} \mathrm{median} \mathrm{value} \mathrm{of} \mathrm{interval} \times \frac{n_i}{N}$$

where $N$ is the total sample size, and $n_i$ is the subset sample size of the respondent choosing interval $i$.

The weighted average implies the average travel frequency for the whole sample. Changes in weighted average travel frequency during the three periods are depicted in Figure 2. During the peak, the average travel frequency sharply dropped by 80.4%, and then recovered to 65.9% when the pandemic was well-controlled after mid-March (no new domestic confirmed cases were reported for the first time; the pandemic prevention and control strategy changed to guard against imported cases and prevent a resurgence of the outbreak at home; work resumption was fully advanced).

2.4. Mode Choice

The weekly use frequency of six common modes of transport was asked about in the same way as 2.3 in the survey questionnaire. Similarly, in order to more intuitively show the changes in the use of various modes of transportation, the weighted average travel frequency (per week) by different modes was calculated and is depicted in Figure 3 following the same equation.

It can be observed from Figure 3 that the use of all transport modes experienced a dramatic reduction and then recovered to some extent. While the restoration ratio of travel by taxi reached 252%, the counterparts of the other five transport modes were all less than 100%. Furthermore, each mode of transport changed to a different degree. The corresponding reduction and recovery rates were calculated by the following equations:

$$Reduction rat{e}_i=\frac{Travel frequenc{y}_{i,before}-Travel frequenc{y}_{i,peak}}{Travel frequenc{y}_{i,before}}$$

$$Recovery rat{e}_i=\frac{Travel frequenc{y}_{i,after}}{Travel frequenc{y}_{i,before}}$$

where $Reduction rat{e}_i$ implies the degree of reduction in travel by mode $i$ during the peak; $Recovery rat{e}_i$ implies to what extent the travel frequency of mode $i$ recovered to the situation before the pandemic. $Travel frequenc{y}_{i,before}$, $Travel frequenc{y}_{i,peak}$, and $Travel frequenc{y}_{i,after}$ are the average travel frequencies of mode $i$ before the pandemic, during the peak, and after the peak, respectively.

Table 2. The degree of reduction and recovery of travel frequency by different modes.

	Reduction Rate (before Peak)/before 1	Recovery Rate after/before 1
Bus and subway	81.3%	73.5%
Taxi and ride sharing	72.2%	252.0%
Private car	40.7%	92.3%
Scooter	61.1%	86.6%
Bicycle	57.6%	86.5%
Walking	52.5%	79.2%

¹ Before: before 27 December 2019; peak: from 20 January to 17 March 2020; after: from 18 March to the date of the survey (17 August).

summarizes the calculation results.

The proportion of transit and private mode of transport is depicted in Figure 4, in which the proportion of taxi and ride-sharing usage is isolated owing to its unique trend. The corresponding chi-squared test was calculated. The chi-squared statistic was 0.1781 (df = 4, p = 0.9593) and statistically insignificant, indicating a heterogeneity within the proportion of different modes.

3. Model and Methods

Past studies have utilized different methodological approaches to model travel behavior, including discrete choice models, discrete continuous choice models, spatial autoregressive models, time series, and so on [22,23,24,33]. Having confirmed the fact that COVID-19 significantly affected people’s travel behavior, the discrete choice model was adopted to explore the specific impact of the individuals’ sociodemographic characteristics on travel mode choice. Discrete choice models were built on the basis of the preference of decision makers, viewing the observed choice as the result of a random utility [36]. The most commonly used discrete choice models are the Probit model and logit model. The more frequently used one has been the logit model because it assumes a logistic distribution for the error component and is relatively easier to estimate. On the other hand, the Probit model has more statistical universality and rationality because of its normal distribution assumption [36,37].

3.1. The Definition of Variables

The random-parameter bivariate Probit model was chosen to predict travel behavior. The dependent variables were processed to be binary, as shown in

Table 3. Definition of dependent variables in binary choice model.

Categories	Definition of Dependent Variables
	before (B) → Peak (P) 1		Peak (P) → after (A) 1
Bus and subway	Decrease 2	P-B < 0: y = 1	Increase	A-P > 0: y = 1
		P-B ≥ 0: y = 0		A-P ≤ 0: y = 0
Private car	Increase	P-B > 0: y = 1	Decrease	A-P < 0: y = 1
		P-B ≤ 0: y = 0		A-P ≥ 0: y = 0
Walking	Increase	P-B > 0: y = 1	Decrease	A-P < 0: y = 1
		P-B ≤ 0: y = 0		A-P ≥ 0: y = 0
Two-wheelers	Increase	P-B > 0: y = 1	Decrease	A-P < 0: y = 1
		P-B ≤ 0: y = 0		A-P ≥ 0: y = 0

¹ Before: before 27 December 2019; peak: from 20 January to 17 March 2020; after: from 18 March to the date of the survey (17 August). ² The changes in decrease and increase were calculated by weekly percentage of the use of each mode.

. The dependent variables were defined to reflect changes in the weekly usage proportion of multiple modes.

Taking the model of the bus and subway as an example, there were two dependent variables in the bivariate Probit model: (1) the change in the travel frequency proportion from before to the peak; (2) the change in travel frequency proportion from the peak to after. For the first one, since the overall proportion of bus and subway was proved to decrease (Figure 4), y was defined as a binary variable.

$$y_1=\{\begin{array}{c}1, if the proportion decreased\\ 0, if the proportion not decreased\end{array}$$

For the second one, since the overall proportion bus and subway was proved to increase (Figure 4), y was defined as a binary variable

$$y_2=\{\begin{array}{c}1, if the proportion increased\\ 0, if the proportion not increased\end{array}$$

To sum up, for public transportation, the decrease (from before to the peak) and increase (from the peak to after) were set to be 1. For private modes of transportation, the dependent variables were processed in the opposite way; the increase (from before to the peak) and decrease (from the peak to after) were set to be 1.

Furthermore, the respondents’ personal characteristics were considered as independent variables, including gender, age, occupation, private vehicle ownership and residence. They were specified as follows:

(1)

Gender was a binary variable (1 = male, 0 = female).

(2)

Age was a binary variable indicating (relatively) older people (1 = older than 24, 0 = 24 and younger).

(3)

Ownership was clustered into a three-level variable: (a) two-wheelers (including scooter and bicycle in Table 1), (b) car, and (c) none. The reference level was none.

(4)

Occupation was clustered into a four-level variable: (a) office worker and non-office worker, (b) self-employed and freelancer, (c) retired and no job, and (d) student. The reference level was student.

(5)

Residence was clustered into a binary variable: (a) first-tier and new first-tier cities, and (b) others (including second-tier cities, third-tier cities and rural areas). The reference level was others.

Additionally, Pearson’s correlation test was conducted before modeling and no highly related independent variables were found.

3.2. Bivariate Probit Model

The observed individual choices related to travel behavior can be interpreted by a random utility model, revealing which choice between the two provides the better utility. The formulation of the random utility model is as follows [36],

$$y_i^{*}={\mathit{x}}_i^{\prime }\mathbf{\beta }+{\epsilon }_i,$$

where $\mathit{x}$ is a vector of the independent variables of the respondents’ sociodemographic characteristics, $\mathbf{\beta }$ is a vector of weight corresponding to $\mathit{x}$, ${\epsilon }_i$ represents the unobserved stochastic errors that derive from individual preference, and $y_i^{*}$ is a latent variable reflecting the utility of a specific choice. It is assumed that ${\epsilon }_i$ has a standard distribution in the Probit model.

A closer inspection of the two response variables revealed that both of them similarly captured respondents’ travel behavior changes in response to pandemic severity. Therefore, the bivariate Probit model was employed for modeling the joint determination of the two binary response variables subjected to the shared unobserved characteristics, which was defined as follows [36]:

$$\begin{array}{c}{y}_1^{*}={\mathit{x}}_1^{\prime }{\mathbf{\beta }}_1+{\epsilon }_1, y_1=1 \left(y_1^{*}>0\right),\\ y_2^{*}={\mathit{x}}_2^{\prime }{\mathbf{\beta }}_2+{\epsilon }_2, y_2=1 \left(y_2^{*}>0\right),\end{array}$$

with the errors defined as:

$$\left(\begin{array}{c}{\epsilon }_1\\ {\epsilon }_2\end{array}|{\mathit{x}}_1, {\mathit{x}}_2\right)\sim N\left[\left(\begin{array}{c}0\\ 0\end{array}\right)·\left(\begin{array}{cc}1& \rho \\ \rho & 1\end{array}\right)\right].$$

where $\rho$ is the cross-equation error correlation coefficient, and other terms are as previously defined.

The bivariate normal cumulative density function is as follows [36],

$$Prob\left(X_1<x_1, X_2<x_2\right)=\underset{-\infty }{\overset{x_2}{{\displaystyle \int }}}\underset{-\infty }{\overset{x_1}{{\displaystyle \int }}}{\varphi }_2\left(z_1,z_2, \rho \right)d{z}_{1}d{z}_2$$

To construct the log-likelihood, for $j=1,2$, define

$$q_{ij}=\{\begin{array}{c} 1, \mathrm{if} y_{ij}=1\\ -1, \mathrm{if} y_{ij}=0\end{array}$$

and let

$$\begin{array}{c}{z}_{ij}={\mathit{x}}_{ij}^{\prime }{\mathbf{\beta }}_j\\ w_{ij}=q_{ij}{z}_{ij}\\ {\rho }_{i^{*}}=q_{i1}{q}_{i2}\rho \end{array}$$

Thus, the log-likelihood function of bivariate Probit model is [36]

$$\ln L={\displaystyle \sum }_{i=1}^n\ln {\mathsf{\Phi }}_2\left(w_{i1},w_{i2},{\rho }_{i^{*}}\right).$$

Moreover, in order to accurately verify that the goodness-of-fit of the bivariate Probit framework is better than two separate Probit models, a log-likelihood ratio test was adopted:

$${\chi }^2=-2\left(\ln L\left(full\right)-\ln L\left(null\right)\right)$$

where $\ln L\left(full\right)$ denotes the log-likelihood of the bivariate Probit model, and $\ln L\left(null\right)$ is the sum of the log-likelihood of the two Probit models. The resulting test statistic is a chi-squared distribution, with degrees of freedom of

$$d.f=K\left(full\right)-K\left(null\right)$$

where $K\left(full\right)$ is the number of parameters in the bivariate model, and $K\left(null\right)$ is the sum of parameter numbers of the two Probit models.

3.3. Random-Parameter Approach

To account for the influences of unobserved heterogeneity across individuals, random parameters were incorporated in the bivariate Probit modeling framework. It assumes that the estimated parameters vary across respondents, usually according to some specified distribution [36]. The individual specific parameter vector ${\beta }_i$ was developed in this model as follows:

$${\beta }_i=\beta +u_i$$

where $u_i$ is a randomly distributed term. In this study, the random parameters were supposed to be normally distributed with a constant mean and variance. The maximum likelihood function was employed to estimate the random parameters.

4. Modeling Results

To explore and identify the individuals’ sociodemographic characteristics affecting travel behavior, four random-parameter bivariate Probit models were established, including an estimation of the changes in the weekly usage proportion of the bus and subway, private cars, two-wheelers and walking. All explanatory variables included in the final model specifications were statistically significant at 10% or lower. Additionally, chi-squared distribution likelihood ratio tests were conducted to verify the greater goodness-of-fit of the bivariate Probit model than two independent univariate models.

4.1. Bus and Subway

The weekly use of the bus and subway is depicted in Figure 5, the usage proportion change estimation results are summarized in

Table 4. Estimation results of the random-parameter bivariate Probit models of usage proportion of different modes.

Variables	Bus and Subway		Private Car		Two-Wheelers		Walking
	B-P (Decrease)	P-A (Increase)	B-P (Decrease)	P-A (Increase)	B-P (Decrease)	P-A (Increase)	B-P (Decrease)	P-A (Increase)
Intercept	0.560 ***	0.147	−0.441 ***	−0.575 ***	−1.038 ***	−1.294 ***	0.381 ***	0.124
Standard deviation	0.103 **		0.811 ***	0.137 ***	0.149 ***	0.475 ***	0.206 ***
Gender: male								0.195 ***
Standard deviation								0.464 ***
Age: >24 years		−0.467 ***		−0.051 ***
Standard deviation				0.319 ***
Ownership	Reference: none
Two-wheelers	0.164		−0.031	−0.046	0.684 ***	0.783 ***	0.039	0.263 **
Standard deviation					0.360 ***	0.166 ***		0.507 ***
Car	−0.204 *		0.505 ***	0.585 ***	−0.182	−0.174	−0.260 **	−0.179
Standard deviation	0.470 ***		0.424 ***	0.003 ***			0.708 ***
Occupation	Reference: student
Office worker and non-office worker	−0.583 **
Self-employed and freelancer	−0.888 ***
Retired and no job	−0.525 **
Standard deviation	1.478 ***
Residence	Reference: others
First-tier and new first-tier cities	0.569 ***	0.350 ***	0.226 *	0.288 **
Standard deviation	0.624 ***	0.177 **		0.081 ***
Correlation between the errors (ρ)	0.891 ***		0.99994 ***		0.982 ***		0.994 ***
Log-likelihood	−535.618		−460.812		−431.029		−534.561
K (number of parameters)	16		16		11		12
Comparison between Bivariate model and two independent univariate models
Log-likelihood (univariate)	−289.93	−318.092	−314.266	−305.428	−263.779	−245.177	−315.185	−316.661
K (number of parameters)	7	3	4	5	3	3	3	4
Likelihood ratio test	144.808 (d.f. = 6)		317.764 (d.f. = 7, p < 0.001)		155.855 (d.f. = 5, p < 0.001)		194.568 (d.f. = 5, p < 0.001)

***, **, * means significance at 1%, 5%, and 10%, respectively. B (before): before 27 December 2019; P (peak): from 20 January to 17 March 2020; A (after): from 18 March to the date of the survey (17 August).

, and the normal distribution curves of the coefficients of the random parameters are depicted in Figure 6A. As is clear from the results, travel by bus and subway of individuals older than 24 had a relatively lower rise after the peak. Compared to the others, the reduction in transit travel during the peak was less dramatic for more than half (~67%) of those who owned private cars. For office workers and non-office workers, the reduction in bus and subway usage was less obvious than for students, but more obvious than for self-employed individuals and freelancers. It is apparent from Figure 5 that the latter were less frequent users of transit before the outbreak of the pandemic, along with retired and jobless people. Transit travel of those living in developed cities fell more sharply during the peak (there was significant heterogeneity, with 18% of them not conforming to this finding), and they also recovered more.

The correlation between the errors $\rho$ was 0.891, which indicates that the two models had significant shared unobserved errors. The likelihood test also showed the better goodness-of-fit of the bivariate Probit model compared to two independent models.

4.2. Private Car

It is obvious from Figure 7 that changes in private car travel varied significantly among respondents with different sociodemographic characteristics. The results of the model estimating travel by private car are presented in Table 4. The normal distribution curves of the random parameters’ estimated coefficients are depicted in Figure 6B.

According to the results, age, vehicle ownership and residence had a significant influence on the change in the usage proportion of private cars. More than half of the older people’s (56.41%) usage proportion decline was less obvious than that of younger people after the peak. The majority of individuals who owned private cars underwent a more dramatic increase (88.35% of them show this trend) and a more evident reduction (~100%), indicating the significant shift from public transport to private cars after the outbreak of COVID-19. Compared to those who lived in underdeveloped cities, individuals living in first-tier and new first-tier cities also showed a similar trend.

The correlation between the errors, R (1, 2) was more than 0.999, illustrating the existence of a strong shared error. Additionally, the likelihood ratio test indicated that the bivariate structural model outperformed the two independent univariate models.

4.3. Two-Wheelers

Given in Figure 8 and Table 4, multiple sociodemographic characteristics were observed to affect the changes in travel by two-wheelers, including bicycles, electric bicycles and scooters. The probabilities below zero of the coefficients of the random parameters are also calculated in Table 4, and the corresponding curves are depicted in Figure 6C. Ownership of two-wheelers turned out to significantly impact changes in cycle travel. For individuals who owned two-wheelers, the proportion of travel frequency using two-wheelers increased more drastically and dropped more obviously. In addition, the bivariate Probit model outperformed the two independent univariate models according to likelihood ratio test (Table 4).

4.4. Walking

Weekly walking travel frequency and estimation results based on sociodemographic characteristics are shown in Figure 9 and Table 4 respectively, as well as the goodness-of-fit test, and Figure 6D offers additional insights into the normal distribution curves of the coefficients of the random parameters. It reveals that more than 66% of the male respondents underwent a greater degree of reduction in walking proportion after the peak. Meanwhile, ownership of cars and two-wheelers was also a statistically significant variable impacting individuals’ walking proportion (with observed heterogeneity), which reduced the shift to walking to some extent.

5. Discussion

The changes in travel behavior were affected by multiple factors, so it is difficult to directly measure the influence of each factor. Thus, the interpretation of the survey and modeling results need to consider the effect brought about by the stringent mandatory travel restrictions imposed by the government over the periods (

Table 5. The COVID-19 situation and corresponding restrictions in the three periods.

Period	Time Duration	The COVID-19 Situation	Restrictions [38,39]
Before	Before 27 December 2019	No cases.	No restrictions
Peak	From 20 January to 17 March 2020	The outbreak began in Wuhan, and quickly spread across the country.	Lockdown Stay-at-home order Screening and quarantine Traffic entrance management: establish checkpoints to inspect people entering the city, community and village Epidemiological investigations and prevention measures Mask-wearing requirement The “health code” regulation: Through big data and communication technologies, a quick response (QR) code was used to show the probability of a person having COVID-19 by displaying green, yellow, or red to indicate their health status Metro was stopped (Wuhan) Negative PCR test result when entering public places Social distancing School closures Teleworking Avoid/limit large gatherings and close public venues Non-essential businesses closure
After	From 18 March to the date of the survey (17 August)	Within the territory of the whole sporadic distribution, local areas appear small clusters of outbreaks.	Lockdown: only in areas with small clusters of outbreaks Stay-at-home order gradually lifted Traffic entrance management Metro reopens Public venues gradually reopen Schools gradually reopen Resume work gradually Non-essential businesses gradually reopen (Other restrictions remain the same as the peak period)

).

In other words, the changes were not solely affected by travelers’ concerns, but also by various travel restrictions imposed by the government.

5.1. Discussions of Changes in Mode Choice after the COVID-19 Outbreak

It is observed from Figure 3 and Table 2 that travel frequency by mode was substantially different in the three stages of COVID-19 within the research scope of this paper. First, there was a remarkable drop in travel frequency that occurred in every mode of transportation during the peak period, among which the largest decline was observed for public transport. Similar findings have been reported in multiple studies in Germany [4], Hungary [25], Greece [9,40], Turkey [33], Pakistan [41], the Netherlands [30], and Indonesia [42]. It is speculated that this reduction stemmed from both the increasing risk perception of public transportation and travel restriction policies. As shown in Table 5, there were, of course, no travel restrictions in the before period. During the peak period, there were very strict travel restrictions. Thus, travel frequency significantly decreased due to the restrictions as well as travelers’ concerns. Nevertheless, the proportion of private transport increased significantly, including private cars, two-wheelers, and walking, which is consistent with findings in the United States [27], Australia [16], Greece [9], Colombia [43], Brazil [44], Istanbul [45], and Scotland [31].

After the peak, the travel frequency of every other transport mode was observed to increase significantly, but not exceeding the counterpart before the outbreak, except for taxis. This is because the travel restrictions still remained effective, although they were relaxed compared to the peak period (Table 5). The taxi and ride-sharing modes had a recovery rate of more than 250% because of their lower perceived risk compared to the bus and subway. This figure partly reflects residents’ preference for taxis and ride-hailing after the peak of the pandemic, although it may be higher than the actual figure. According to the financial statement of DiDi, the largest ride-hailing company in China, the revenue in Q1 and Q2 of 2020 decreased by 40% and 18% year-over-year, while revenue in Q3 and Q4 increased by 6% and 12% year-over-year, respectively [46]. That is to say, people’s tendency to use ride-hailing has been demonstrated. It is noteworthy that the mandatory metro closure had already been lifted when the survey responses were collected (Table 5), indicating the mode shift caused by COVID-19 is presumably to be sustained in the medium to long term [44,47].

5.2. Discussion of Modeling Results

Although there has been a sharp decline in public transport passenger flow worldwide, the reduction in the flow of people during the pandemic varies greatly across socioeconomic groups. The results showed that (1) 66% of males recovered more than females in terms of walking travel after the peak, but generally speaking, gender was not a significant variable affecting travel mode choice. (2) For individuals older than 24 years, travel by bus and subway recovered less than for young people, and more than half of them (56%) exhibited less growth in the use of private cars after the peak. This finding indicates that even in the “new normal”, individuals’ mode shifts were still ongoing, affected by their sociodemographic characteristics. (3) Ownership of private cars and two-wheelers was a common significant variable affecting individuals’ travel by the corresponding mode of transportation. (4) Regarding occupation, students’ use of the bus and subway was impacted the most dramatically, as they are regular users of public transport in China, probably because they do not have a private vehicle or driver’s license. (5) Travel by both transit and private car of individuals living in developed cities (i.e., first-tier cities and new first-tier cities) declined more after the outbreak of COVID-19, and they also recovered more after the peak. The cities are evaluated based on five aspects, namely commercial resource concentration, urban hub, urban human activity, diversity of lifestyle, and future plasticity. Residents in developed cities have more diverse lifestyles and access to more convenient public transport services and business activities. The public transport system is well established in developed cities with a higher mode share, and the use of private cars is more common in developed cities. Additionally, the pandemic in developed cities is more severe and the travel restrictions are stricter. Therefore, their travel behavior changes are more sensitive to the pandemic. On the other hand, two-wheelers are more popular in underdeveloped cities, where scooters are a common means of transportation in many small cities in China.

5.3. Policy Implications about Effective Restrictions and Sustainable Transportation

The findings from the present study are expected to help policymakers establish effective transportation policies considering sociodemographic and local features. Considering the disparities between different groups, the main line is to enable transport enterprises and authorities to provide better services during the pandemic to passengers rather than restricting access and choices [29,48]. That is to say, travel equality and sustainability should be considered and emphasized in the COVID-19 era, as well as in another possible pandemic in the future.

Sustainable transportation refers to low- and zero-emission, energy-efficient, affordable modes of transport, including electric automobile, shared modes of transport and so on. Thus, walking, bicycle, and transit are encouraged for the purpose of sustainability. However, the first two are not suitable for long-distance travel. So, the bus and subway are encouraged. First, government authorities could design partial and resilient lockdowns and travel restrictions in different areas for different groups. Second, since travelers are reluctant to use public transportation or shared bikes because they are concerned about infection risk in the “new normal” [49], it is required to improve hygiene management (e.g., frequent disinfections) regarding transit and shared bikes.

(1)

Many travelers, especially young travelers, still use public transportation as their primary or sole means of transportation, which leads to unequal risks of virus transmission. Public transport needs to maintain strict anti-pandemic measures, especially in developed cities. Hygiene measures (such as sanitizations of vehicles and standard temperature checks for drivers and passengers) are expected to mitigate public perceived safety threats. Considering that students who do not own cars are the group most severely affected by COVID-19, the stakeholders of public transport are suggested to provide them with better information services related to the prevention of disease spreads (e.g., real-time number of passengers, last disinfection time).

(2)

It is necessary to provide facilities and hygiene measures for non-motorized vehicles (i.e., bicycles) to promote them, especially in underdeveloped cities and rural areas due to the potential safety hazard. In Toronto and London, it was recommended to construct physically separated bike lanes and connect existing bike networks to facilitate and maintain cycling safety [50].

(3)

It is also required to better connect bike networks to public transportation, namely multi-modal transportation [51,52,53]. Multi-modal transportation is recommended in developed cities where there are more younger citizens with diversified travel demands. Specifically, it is suggested to optimize the scheduling of shared bikes near public transport hubs in order to improve the efficiency of transport networks. At the same time, it is necessary to strengthen disinfection measures for shared bikes such as antiseptic wipes, masks, etc. Additionally, more intelligent recommendation and more various situational contexts (e.g., hygiene information and weather) could be integrated into navigation applications [51].

5.4. Limitations

Although the study revealed multiple key findings, there are still limitations to this study. First, the data were collected once after the peak and the responses were based on their recall. Second, the survey questionnaire was only collected online. Thus, it is possible that the data do not fully represent the general population (e.g., the age bias). Younger people were overrepresented in the sample. Based on the actual age structure from China’s National Bureau of Statistics, the SMOTE algorithm was employed to generate more samples of those over 24. It was found that there was little difference between the model results with the original sample, which is only reflected in the significance of some variables, while the coefficient signs of variables did not change. These limitations should be addressed in follow-up studies.

6. Conclusions

Due to the subjective risk perception towards COVID-19 and travel restriction policies imposed by governments, there is no doubt that the COVID-19 pandemic has exerted an unprecedented impact on social activities, including travel patterns and individuals’ mode choice. It is necessary to clearly understand the specific directions of these changes and their relationship with individuals’ characteristics. Thus, this study was devoted to clarifying the changes in travel mode choice and quantifying the effects of individuals’ sociodemographic characteristics (e.g., gender, age, occupation) on the mode choice changes in the three periods of the pandemic in China (i.e., before the outbreak, the peak, and after the peak periods). Random-parameter bivariate Probit models were developed to comprehend the significant factors of the switch in mode choice. The key findings were as follows:

(1)

Gender was generally not a significant factor affecting the change in travel mode choice;

(2)

Older people showed a trend of switching from transit to private cars or two-wheelers (heterogeneity was estimated because of heterogeneous ownership);

(3)

Ownership of vehicles or two-wheelers was a significant contributing factor to the changes in mode choice;

(4)

Students had the most drastic changes in the use of the bus and subway compared to other occupations;

(5)

Changes in transit and private car travel were more apparent for people in more developed cities (i.e., first-tier and new first-tier cities).

Analyzing the effects of individual characteristics on mode choice changes can provide important insights into policymaking and transportation planning in the “new normal” and a possible future epidemic. Government authorities and transportation service providers (e.g., bus, taxi, and ride-sharing companies) are required to incorporate individual heterogeneity and design more resilient policies to slow down the spread of disease and ensure travel equality and sustainability. Additionally, in the “new normal”, it is necessary to come up with effective solutions to mitigate the possible traffic safety problem, congestion, and air pollution caused by the observed changes in individuals’ willingness to use transit/private cars/two-wheelers. Thus, extra efforts are needed to encourage the use of transit as well as to improve the safety measures of the public transportation system based on understanding the effects of individuals’ characteristics, because there are lots of people with no access to a private car who need to commute by transit regularly. The equality and sustainability of urban public transportation are further required in the post-pandemic era, and the formalized management of bikes, electric bikes, and scooters requires more emphasis in undeveloped areas (i.e., third-tier cities).

Although this paper presented several important findings, it still has a few limitations, including the subjectivity of period division and the not full representation of the sample. In the future, more studies are needed to further explore the relationship between transportation and future disease spread, considering the differences within the epidemic prevention policies and transportation conditions in different countries and regions. In addition, it is also worth studying how to maintain the normal operation of the transportation system and society in the “new normal”, based on the premise of ensuring the COVID-19 pandemic can be controlled. Since multiple vaccines have been on the market, further studies should control for personal characteristics such as vaccination status, or personal attitudes about vaccination and COVID-19 severity as well [54].