Investigating Anticipated Changes in Post-Pandemic Travel Behavior: Latent Segmentation-Based Logit Modeling Approach Using Data From COVID-19 Era

Abstract

The unprecedented situation created by the COVID-19 pandemic in the year 2020 has drastically changed daily mobility patterns around the world. Various measures were implemented to prevent the transmission of the virus, which have resulted in short- and long-term impacts on the activity systems and daily travel. To capture the impacts of the pandemic on travel behaviors and activity systems, a web-based survey was designed and administered in April–May 2020 in Montreal, Canada. In addition to questioning on pre- and during COVID-19 behaviors, it included a section on how people expected to travel, telework, shop online, and so forth in the post-pandemic era. Using data from this survey, which gathered 1,620 completed questionnaires, this paper proposes insights into how people are planning to travel in a post-COVID-19 world using latent segmentation-based logit modeling technique. Three models are estimated to identify factors related to expected trip frequency, expected transit usage, and expected bike usage. Undertaking such modeling approach provides opportunity to understand different types of individuals’ preferential behaviors. This study probabilistically identifies two latent segments, suburbanite and urbanite people, and finds considerable heterogeneity across sample individuals. For example, urbanite people tend to increase their expected number of trips after COVID-19 if they have at least one bike in their household. Suburbanite people exhibit an opposite relationship, and they are more likely to keep their trip frequency the same as before. Findings of this study will assist decision makers in developing effective policy measures to better prepare for the changes in travel behaviors after COVID-19.

The unprecedented situation created by the COVID-19 pandemic in the year 2020 has drastically changed daily mobility patterns around the world. COVID-19 is an infectious respiratory-based disease, which is caused by a novel virus called severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) (1). The virus first appeared in Wuhan, China at the end of the year 2019 (2). It is highly transmissible, and the transmission primarily occurs by breathing droplets from coughs and sneezes, person-to-person direct contact, and via infected surfaces (3). It has a significant fatality rate and can cause societal and financial disruptions worldwide (3). COVID-19 was pronounced a worldwide pandemic by the World Health Organization (WHO) on March 11, 2020 (2). This resulted in strict lockdown, social distancing, quarantine measures, and so forth to avoid the spread of the virus and to protect public health, combined with restrictions on travel (4). The pandemic has influenced long-term decisions in households, but short-term activities and travel decisions have experienced even more impacts across the world. Measures taken to stop the transmission ultimately decreased people’s overall mobility and had significant impacts on typical travel behaviors. To understand the changes in individuals’ daily mobility related to COVID-19, it is important to explore the relationships between pandemic and transportation. A transportation system works as a connecting hub of travelers’ activities and destinations, which makes it a potential source of infectious disease outbreak (5). Therefore, during the outbreak, people started to avoid going out to participate in activities and restricted their mobility to reduce the chances of becoming infected by the virus, which influenced the overall travel patterns of whole regions (6). The number of daily trips undertaken was reduced significantly (7, 8). This subsequently yielded less traffic in the network and lower congestion during peak hours (9). Although distance traveled was found to decrease at the initial stage of the outbreak, it started to increase with the reopening phases—mostly by active modes such as cycling and walking (10). In particular, bike usage has seen an exceptional surge during the COVID-19 pandemic (11). People were also more inclined toward private auto usage since it enabled them to have limited contact with other people (12). The public transport and active transportation sectors were more affected, with a considerable decrease in public transport usage (9). De Vos (13) confirms that various measures implemented during COVID-19 resulted in overall reduction in travel demand, where public transportation was the most affected. Zhang and Lee (7) reported similar behaviors—the usage of public transport decreased significantly and was mostly replaced by walking and cycling. Shamshiripour et al. (14) found that usage of public transit, taxi, and ridehailing services decreased significantly during COVID-19 because of medium to extremely high risk of exposure to the virus in these modes. Similar findings were observed by Nian et al. (15), where they explored lower share of ridehailing services and transit during COVID-19. The primary reason for such changes in regular travel behavior during COVID-19 was fear of being infected with the virus, which ultimately led to increasing telecommuting and online education, and lower participation in public activities and events during the pandemic (7, 13).

Most existing studies on COVID-19 and transportation have explored the impacts of COVID-19 on current travel behavior. It is not clear in the existing studies how people anticipate their daily travel to be in the post-pandemic period. With the ongoing vaccination process across many countries, it is of high importance to assess how people expect to accommodate their travel behavior in the post-pandemic time. This has not yet been explored adequately in the existing literature. It is obvious that with the changes in people’s travel behavior during the COVID-19 outbreak, there will be prevalent impacts on the way people will interact and travel in the post-COVID time. There could be radical consequences from the measures taken during COVID-19. Few studies have attempted to investigate how people expect to travel in the post-pandemic world. It has been found that even after the pandemic, people may travel less, usage of public transport may remain lower than before COVID-19, and a significant reduction in shared transportation services may occur because of the fear of virus exposure (13, 16). In contrast, walking and cycling may gain considerable importance (17). De Haas et al. (18) presented a descriptive analysis of expected travel behavior based on the Netherlands Mobility Panel survey. They found that most respondents expect to use all travel modes just as much as they did before COVID-19; however, travelers’ attitudes toward choosing private cars is expected to improve while public transport would become notably less favorable in the post-COVID time. Conway et al. (19) conducted a survey on highly-educated adults in the United States in spring 2020, which found that public transport may not completely recover to pre-COVID ridership levels, and people are more likely to walk and cycle than before in the post-pandemic time. However, most of these studies until now have focused on the descriptive analysis of the impacts of COVID-19 on future travel behaviors.

Contribution of the Study

The future of travel behavior during the post-pandemic era is uncertain. Different studies have explored different outcomes based on travelers’ opinions using descriptive statistics. However, there is a gap in understanding the behavioral insights such as whether travel behaviors return to normal or transform into a new normal with significant changes in trip frequency and mode choices after a massive pandemic event like COVID-19. Using a descriptive analysis approach, it may not be entirely possible to statistically demonstrate the relationships between choices made by a person and the attributes of the person and/or the choices. In this regard, a discrete choice analysis, that is, an econometric modeling-based micro-behavioral analysis that explores the critical determinants to understand different types of individuals’ expected inherent travel behavior, is warranted. This may assist decision makers to develop effective and long-term plans and policies to accommodate changes in individuals’ travel preferences. Therefore, based on the needs and gaps, this study contributes to the existing literature in two ways while utilizing data from the COVID-19 era: (i) develop econometric micro-behavioral models of anticipated travel behaviors after COVID-19, and identify factors affecting such behaviors, and (ii) explore behavioral differences among different types of individuals based on their characteristics. Since the data used in this study were collected during the pandemic, and individuals’ daily trip frequency and usage of public transport and active modes have experienced considerable changes during COVID-19 (7, 8, 10), this study focuses on developing a trip frequency model, a transit usage model, and a bike usage model to explore individuals’ anticipated changes in travel behaviors. The models are developed following a latent segmentation-based logit (LSL) modeling technique. Undertaking this modeling approach provides an opportunity to understand different types of individuals’ preferential behavior (20). The LSL model estimation process follows a segmentation approach. The model develops a flexible latent segment allocation model endogenously based on individuals’ characteristics to capture the unobserved heterogeneity across sample individuals within the modeling process. It is probabilistically assumed in this study that individuals’ expected changes in trip frequency, transit usage, and bike usage may vary based on the socio-demographic and built environment characteristics of their home location. The models provide opportunity to test individuals’ discrete latent segment-specific preference of expected travel behavior changes in activity-travel characteristics before COVID-19, expected activity attributes after COVID-19, socio-demographic characteristics, and built environment. This study develops all three travel behavior models by utilizing data from the “COVID-19 Travel Study” survey conducted in Montreal, Canada. Insights from this study will inform decision makers and planners to develop policy interventions and prioritize effective measures to accommodate changes in future travel behavior in the post-pandemic era.

Data

Data Sources

In April 2020, a survey was designed to collect data on the impacts of the COVID-19 pandemic on travel behaviors and activity systems. It had a 60% completion rate which is similar to comparable web surveys on transportation. The average duration of completing the survey was 19.5 min (excluding durations over 40 min) or 21.3 min (excluding durations over 60 min). The median duration was 20 min. The survey was designed in-house and administered online using a platform developed by the research team (21) and used for more than 15 online surveys in the last decade. A lot of attention has been paid throughout the years to the improvement of the web-based survey tool; this is why the research team did not rely on a generic tool (such as Survey Monkey) for questions as complex as travel behavior. Also, self-administered surveys such as web-based surveys have become mainstream since the Canadian census has also moved to online questionnaire. The questionnaire was segmented in four main sections: home location and household attributes; people attributes; travel behaviors before COVID-19 and during COVID-19; and expected travel behaviors after COVID-19 (mode, frequency, etc.) including questions on teleworking, online shopping, and other changes in activities. The survey was person-based, meaning that a single respondent per household completed the questionnaire on their travel behaviors and COVID-19 impacts. However, it also collected basic information for each member of the household. The survey questionnaire included clear interpretation and definition of the questions. For instance, the questions on carsharing were asked to everyone, and the names of the carsharing companies and services were included in the questionnaire when people were asked if they were members or if they were using such service.The survey was launched on April 28, 2020, and distributed using social networks of partners (transit operators), sent to an in-house panel of respondents containing email addresses of people who participated in previous surveys and agreed to be contacted again, and advertised in a newspaper. In total, 1,620 respondents completed the questionnaire. From the way it was constituted, this sample cannot be considered representative of the population. To assess the composition of the sample, some reference indicators for the COVID-19 survey sample were estimated using the more recent large-scale “Origin-Destination” survey held in Montreal in 2018 (Table 1). This allows us to conclude that the sample covers the eight main zones of the region with fewer respondents from the eastern part of Montreal Island. Average age is within similar range (40.05 years for the reference value); the 25 to 44 year old age group are over-represented to the detriment of those 65 years and older but not drastically. Men are under-represented since they represent 48.77% of the population, and driving license ownership is higher in the COVID-19 survey sample than in the Montreal population (82.58% is the reference), which is partly because of lower representation of elderly people in the COVID-19 survey. Household size is within the 2.34 person average for Montreal.

Table 1.

Summary Statistics

		Expected changes in trip frequency		Expected changes in transit usage		Expected changes in bike usage
		Distribution of dependent variables
Number of respondents		1,580		1,357		1,098
Alternatives
Less usage than before (decrease)		36.58%		41.60%		2.80%
Same usage as before (same)		60.06%		57.00%		65.10%
More usage than before (increase)		3.36%		1.40%		32.10%

Note: $ = Canadian dollars (CAD); na = not applicable.

Data Preparation and Variables Considered

Preparation of the dataset in this study included various processing steps. After the survey data collection, all data were cleaned carefully, and all identifiable information was replaced with anonymous codes. During the survey, respondents were asked how they anticipated changes in their travel behaviors, that is, daily trip frequency, transit usage, and bike usage after COVID-19 (among others). Four alternatives were provided for each travel behavior dimension: never used, less usage than before the pandemic (decrease), same usage as before, and more usage than before (increase). Respondents who answered “never used” were omitted during the analysis. The alternatives are used as the dependent variables during the model estimation processes. Socio-demographic information, activity-travel characteristics before COVID-19, and expected activities after COVID-19 were extracted directly from the “COVID-19 Travel Study” database. Built environment variables were obtained from the 2016 Canadian census and 2020 Proximity Measure Database (PMD). The Canadian census database provides various types of neighborhood information at the dissemination area (DA) level, and PMD provides measures of proximity to various services and amenities at the dissemination block (DB) level. The measurements are represented as an index value ranging from zero to one, where zero indicates the least proximity and one the greatest proximity. To obtain built environment characteristics, the spatial join function in ArcGIS was utilized to join home locations with DA-level data from the census and DB-level data from the PMD. Finally, a complete database was built by joining the survey data with corresponding neighborhood and proximity measure data. Three separate choicesets were extracted from this primary database. Table 1 reports the descriptive statistics of the variables retained in the final models. It includes the distribution of both dependent and independent variables. The table presents the mean and standard deviation values for the continuous variables, and proportion values for the categorical/dummy variables. Note that the number of respondents (n) for all three models are not same. This is because not all respondents provided responses for all the questions in the survey. For example, of the 1,620 respondents, 1,580 reported their opinion on “expected changes in trip frequency,” 1,357 reported their opinion on “expected changes in transit usage,” and 1,098 reported their opinion on “expected changes in bike usage.”

Modeling Approach

This study utilizes a random utility-based discrete choice modeling technique to estimate individuals’ expected travel behavior. The multinomial logit (MNL) model is one of the suitable approaches for discrete choice modeling. However, such traditional models generally have an inherent monotonic assumption of independence from irrelevant alternatives, which overlooks the unobserved preference heterogeneity that may occur when choices are made. This may result in inconsistent and biased estimations (22). To overcome this issue, researchers have developed flexible alternative approaches such as random parameter logit (RPL) model and LSL model. The RPL model can capture heterogeneity by allowing random parameters to vary with a continuous joint parametric distribution across a population, which is required to be assumed by the analysts (23). In contrast, the LSL model provides a semiparametric specification that does not require any strong or unseemly parametric distribution about preference heterogeneity to be assumed (20); instead, it captures unobserved heterogeneity by implicitly sorting individuals into discrete latent segments. Unlike multifactor analysis or structural equation models, where differences in individuals’ behavior are measured based on latent factors, LSL models have the capacity to explain individuals’ behavioral differences through latent segments which are estimated based on individuals’ characteristics (Gibson, 1959). This study attempts to explore behavioral differences among different types of individuals based on their characteristics (refer to the section “Contribution of the Study”) in the case of anticipated travel behavior after COVID-19. Therefore, it develops LSL models to estimate individuals’ behavior with regard to anticipated changes in trip frequency, transit usage, and bike usage after COVID-19. Individuals’ probability of being included in different latent segments is determined in this study by developing a latent segment allocation model, which is defined using individuals’ observed socio-demographic and built environment characteristics. The probability of an individual n being allocated to latent segment s can be written as:

$${\lambda }_{\mathrm{ns}}=\frac{e^{{\delta }_s+{\omega }_s{\mathbf{Y}}_n}}{\sum _{s=1}^S{e}^{{\delta }_s+{\omega }_s{\mathbf{Y}}_n}}$$ (1)

where δ is the segment membership constant, Y is the observed attributes of the individuals, and ω is the segment membership vector coefficients. One of the segments is assumed to be the reference segment for model identification by considering δ and ω fixed for the segment. Assuming that individual n allocated to segment s chooses an alternative j from a set of alternatives K, the choice probability can be expressed as:

$$P_{\mathrm{jn}|s}=\frac{e^{{\beta }_s{\mathbf{X}}_{\mathrm{jn}}}}{\sum _{j=1}^K{e}^{{\beta }_s{\mathbf{X}}_{\mathrm{jn}}}}$$ (2)

where X is the observed vector parameters and β is the segment-specific vector coefficients. The likelihood of an individual n choosing an alternative j is the expectation (over segments) of the segment-specific contributions:

$$P_n=\sum _{s=1}^S{\lambda }_{\mathrm{ns}}{P}_{\mathrm{jn}|s}$$ (3)

The LSL model developed in this study estimates the parameters by maximizing the likelihood using an expectation maximization algorithm. The likelihood function to estimate the parameters can be written as:

$$L{L}_{\max }=\sum _{r=1}^R\ln P_n^{\mu }$$ (4)

where R is the total number of observations and μ is a dummy variable representing one if an alternative j is chosen by individual n, and zero otherwise. Each model estimates segment-specific parameter vector β for S segments and segment membership parameter vector δ and ω for S-1 segments. The goodness-of-fit measure of each model is evaluated based on the log-likelihood value at convergence, McFadden’s pseudo-R², Akaike information criteria (AIC), and Bayesian information criteria (BIC).

Discussion of Results

Determination of the Number of Latent Segments

Estimation of the LSL model starts with identifying the number of latent segments, S. Since the number of segments is not a parameter, a hypothesis on this cannot be tested directly. Based on the previous studies, this study determines the number of segments based on the AIC and BIC values (24). Comparatively, the value of S that minimizes the AIC and BIC values indicates which model should be considered. Results suggest that AIC and BIC values are minimum for the models with two segments. Therefore, the final models are assumed to have two latent segments (see the top rows of Tables 2–4). Note that, although most of the parameters retained in the final models are statistically significant at least at 90% confidence level (t-stat ≥ 1.64), a few parameters with statistical significance below 90% confidence level are kept during analysis for their critical insights and for guiding future efforts.

Table 2.

Expected Changes in Trip Frequency Model Estimation Results

Determination of number of segments
Goodness-of-fit measures	No. of segments = 2			No. of segments = 3
Number of parameters	63			57
Log-likelihood at convergence	−809.78			−1156.78
McFadden’s pseudo-R2	0.3601			0.2846
AIC	1745.56			2391.56
BIC	2083.57			2600.80

Note: AIC = Akaike information criteria; BIC = Bayesian information criteria; $ = Canadian dollars (CAD); na = not applicable.

Table 3.

Expected Changes in Transit Usage Model Estimation Results

Determination of number of segments
Goodness-of-fit measures	No. of segments = 2			No. of segments = 3
Number of parameters	65			63
Log-likelihood at convergence	−730.71			−986.55
McFadden’s pseudo-R2	0.3801			0.3117
AIC	1595.42			2069.10
BIC	1944.69			2319.33

Note: AIC = Akaike information criteria; BIC = Bayesian information criteria; $ = Canadian dollars (CAD); na = not applicable.

Table 4.

Expected Changes in Bike Usage Model Estimation Results

Determination of number of segments
Goodness-of-fit measures	No. of segments = 2			No. of segments = 3
Number of parameters	65			65
Log-likelihood at convergence	−602.06			−1006.47
McFadden’s pseudo-R2	0.3957			0.3007
AIC	1338.12			2102.94
BIC	1673.20			2328.00

Note: AIC = Akaike information criteria; BIC = Bayesian information criteria; $ = Canadian dollars (CAD); na = not applicable.

Expected Changes in Trip Frequency Model Results

Table 2 presents the parameter estimation results of the expected changes in trip frequency model. The latent segment allocation model is estimated considering segment 2 as the reference segment. Results exhibit a negative sign for the variable representing household income below $120,000 Canadian dollars (CAD), which demonstrates a lower probability of individuals living in such households being allocated to segment 1. The positive parametric value of the age variable indicates that older individuals are more likely to belong to segment 1. Among the built environment characteristics, the negative sign of the variable “percentage of rented houses” and the positive sign of the variable “percentage of single-detached houses in the neighborhood” reveal that suburban dwellers have a higher likelihood of being allocated to segment 1. In summary, segment 1 has a higher propensity to include suburban dwellers with higher household income and older individuals. Therefore, segment 1 can be probabilistically identified as a segment for “suburbanite people.” In contrast, segment 2 can be identified as a segment for “urbanite people”.

Several socio-demographic variables, built environment attributes, activity-travel characteristics before COVID-19, and expected online activities after COVID-19 are found to influence individuals’ expected travel usages. For example, results suggest that male individuals are more likely to decrease their trips in segment 1. In contrast, male individuals in segment 2 exhibit lower likelihood of decreasing their expected trips. On the contrary, they tend to increase their trip frequency. This could be because of the employment status and type of occupation of male urbanite people. Bike ownership shows heterogeneous relationships across segments. Results suggest that individuals in segment 2 have a higher likelihood of increasing their expected trip frequency if they own a bike. This might be because cities have already taken steps to improve the active transportation opportunities from the travel pattern experience during COVID-19 (25), and the “15-minute city” concept has been gaining popularity based on travel patterns during the pandemic (26). Suburbanite people in segment 1 exhibit a negative parametric value. Interesting heterogeneous outcomes are observed with regard to driving license ownership. Older and higher income suburbanite individuals exhibit a higher tendency to decrease anticipated trip frequency, whereas urbanite people who are younger with lower income are less likely to decrease trip frequency after COVID-19. This is perhaps because of the differences in telecommuting between urban and suburban areas. COVID-19 has increased telecommuting opportunities among suburban dwellers (27), so such people might decrease their number of trips even if they possess a driver’s license. Furthermore, both urbanite and suburbanite full-time employed individuals who plan to relocate their residence after COVID-19 tend to expect a decrease in their trip frequency in the post-COVID-19 period, probably indicating such individuals’ increasing inclination toward online activities. This parameter in the suburbanite segment is statistically significant, but insignificant in urbanite segment. However, this result is critical in the sense that it would be beneficial to planners and policy makers to focus on efficient residential developments in urban and suburban areas, and reshaping the land uses in commercial and industrial areas. With a larger sample in future, the variable may provide a statistically significant parametric value. However, for now, caution should be exercised when applying the urbanite segment result to the general population of the study area. As to pre-COVID-19 activity-travel characteristics, individuals in both segments exhibit lower probability of increasing their expected trip frequency if the number of places visited weekly was higher before COVID-19. This is probably the result of individuals’ growing habits of higher online activity participation such as teleworking, online education, ordering food online, and online grocery shopping, among others. Interestingly, individuals who primarily participated in leisure activities before COVID-19 exhibit positive relationships with expected reduction in trip frequency in both segments. Despite this parameter in the urbanite segment being statistically insignificant, it shows intuitive insights. Possibly, social distance measures and limited number of people gathering in leisure activity places contribute to such anticipation. This insight could be useful for the urban planners to develop strategies to create safer environments for leisure activities, however, they should be cautious to apply this finding to the general population. Similar outcomes are observed in the case of public transportation as the primary travel mode before COVID-19. In both segments, individuals are less likely to increase their expected trips, rather they tend to decrease their expected trip frequency if public transport was their primary travel mode before COVID-19. This is probably because of the fear of being infected with the virus, since public transportation offers collective mobility that makes it vulnerable to disruption and shocks from pandemics (28), and since transit is often used for commuting trips which are also more likely to be less frequent after COVID-19. In both segments, car drivers tend keep their anticipated trip frequency the same after COVID-19. As expected, people who foresee never participating in online activities after the pandemic are found to increase trip frequency in the post-COVID-19 period. For instance, in both segments, variables representing no online grocery order after COVID-19 demonstrate positive parametric values in the case of increase in expected trip frequency. Moreover, individuals who reported not to do telecommute and online food order ever, show negative signs in case of the alternative ‘decrease in expected trip frequency’ in suburbanite and urbanite segments. The built environment variables retained in the final model exhibit similar behavior across segments. For instance, individuals are more likely to decrease their trip frequency after COVID-19 as proximity to parks increases. Moreover, higher proximity index representing closeness to health care facilities demonstrates a negative relationship with the increase in expected trip frequency, which means individuals are less likely to increase their trip frequency when they live farther away from health care facilities.

Expected Changes in Transit Usage Model Results

The latent segment allocation component results of the expected changes in transit usage model (Table 3) indicates negative signs for the socio-demographic indicators: age, full-time employment, and household annual income above $120,000 CAD. This suggests that segment 1 has a higher likelihood of including urban dwellers with lower household income, no full-time workers and younger age. Presumably, segment 1 can be identified as a segment for “urbanite people.” On the other hand, segment 2 can be probabilistically identified as a segment for “suburbanite people.”

Model results suggest that individuals across both segments are less likely to keep their expected transit usage the same in the case of higher number of cars in the household and having carshare membership. As expected, ownership of a monthly transit pass increases urbanite people’s probability to anticipate increased transit usage after COVID-19. In contrast, suburbanite people who belong to segment 2 show an opposite relationship. When transit pass ownership interacts with the pre-COVID-19 work activity, individuals from both segments demonstrate higher propensity to decrease their expected transit usage. This parameter in the suburbanite segment is not statistically significant, yet the variable is considered in both segments in the final model estimation for its critical insight. The estimation results perhaps indicate such individuals’ greater inclination toward telecommuting after COVID-19, which may in turn decrease their transit usage even if they have a transit pass. However, readers should be cautious in applying this finding to the general population. Driving license ownership exhibits the expected outcomes. Individuals are less likely to increase their transit usage after COVID-19 if they hold a driving license. Most of the activity-travel characteristics before COVID-19 exhibit similar outcomes across segments. Results suggest that individuals whose number of places visited weekly was relatively high before COVID-19 are less likely to increase their expected transit usage in both segments. This is probably suggesting such individuals’ inclination toward switching to more online activities after COVID-19. As expected, individuals with lower visiting frequency (≤1/week) to a primary activity location before COVID-19 have a higher tendency to decrease their expected transit usage during the post-pandemic period. Individuals in the urbanite segment who used transit for shopping as their primary activity before COVID-19 exhibit lower probability to decrease transit usage after COVID-19. Since shopping places are closer together in urban areas, people may not reduce transit usage for short-distance travel. Although the parameter demonstrates a statistically insignificant relationship, the variable is considered for use in the urban segment for such intuitive understanding. Therefore, caution is required to apply this result to the general population. On the other hand, suburbanite people tend to decrease their expected transit usage, perhaps suggesting that such individuals have less inclination toward transit usage for long distances after COVID-19, since shopping places in suburban areas are generally further away. In the case of primary travel mode, individuals who used car driving before COVID-19 exhibit a lower probability of keeping their expected public transit usage the same; instead, they are more likely to decrease their transit usage after COVID-19 in both segments, which is expected. In the expected activities after COVID-19, both urbanite and suburbanite people exhibit higher propensity to decrease expected transit usage if they plan to relocate their residences after COVID-19. This is plausible because telecommuting, which has become popular since COVID-19 struck, might allow people to fulfill their desire to move to suburban areas and small towns, where transit infrastructures are less developed. When the variable representing “plan to relocate residences” interacts with the driving license ownership, higher coefficient values are observed in both segments (2.12 in segment 1, and 3.19 in segment 2); suggesting that holding a driving license while planning to relocate after COVID-19 has higher effects on reducing the expected transit usage. Interestingly, individuals in both segments who expect not to telecommute after the pandemic are less likely to increase their expected transit usage, rather they tend to anticipate decreasing their use of transit after COVID-19. This might be attributed to fear of using transit after COVID-19 and/or by expected reduced congestion which may improve expected driving conditions. Individuals who expect not to order groceries online in post-pandemic period, exhibit lower probability to decrease their transit usage. Perhaps, such individuals regularly used transit for grocery shopping before COVID-19 and they are planning to go back to normal once the pandemic is over. In the case of built environment, results suggest that higher employment rate in the neighborhood reduces the probability of increased expected transit usage in both segments. This is probably because of the increased reliance on teleworking, which may undermine transit usage during the post-COVID period. When residing close to their employment, individuals in segment 1 (characterized as the segment of urbanite people) are less likely to decrease their transit usage; instead, they tend to increase their expected transit usage after COVID-19. Suburbanite people exhibit opposite relationships. With greater distance to public transport station, individuals are more likely to decrease their transit usage after COVID-19—which is expected.

Expected Changes in Bike Usage Model Results

Table 4 presents the latent segment allocation component results of the expected bike usage model. A positive parametric value is observed in segment 1 for the age indicator. Household income below $120,000 CAD exhibits a negative sign in segment 1. Built environment indicators such as percentage of single-detached houses in the neighborhood show a positive coefficient value, and percentage of rental houses in the neighborhood shows a negative value in segment 1. Based on these segmentation coefficients, segment 1 is assumed to be a segment for “suburbanite people.” In contrast, segment 2 is probabilistically specified as the segment for “urbanite people.”

Model estimation results suggest that expected changes in bike usage are significantly influenced by individuals’ socio-demographic characteristics. For example, female individuals are less likely to increase their expected bike usage in both suburbanite and urbanite segments; instead, they are more likely to expect to keep their bike usage similar after COVID-19. With the increase in number of cars in the household, urbanite people demonstrate higher probability to increase expected bike usage in segment 2, the urbanite segment. Although an effect of the COVID-19 pandemic was to decrease cycling among the higher car ownership households (29), urbanite individuals’ higher bike usage probably indicates the presence of safer and improved cycling infrastructure in urban areas. In contrast, suburbanite people show lower propensity to anticipate increased bike usage after COVID-19. As expected, full-time employed individuals who expect to use cars more after COVID-19, have a higher tendency to decrease bike usage in the post-pandemic period. Mobility tool ownership indicators exhibit interesting results. For instance, monthly transit pass owners demonstrate a higher probability of increase in expected bike usage in both urbanite and suburbanite segments; perhaps indicating inclination toward avoiding public transport for fear of virus transmission, or changes in destinations which are more accessible by bike. Individuals who do not have any carshare membership exhibit a higher probability to increase their expected bike usage. Similarly, those with bikeshare membership show lower propensity to decrease expected bike usage in both segments; instead, they are more likely to increase expected bike usage. With regard to activity-travel characteristics before COVID-19, people tend to keep their expected bike usage the same as before if their primary activity type in a day, before the pandemic, was work. In the case of the primary mode of travel before COVID-19, both cycling and public transport demonstrate positive parametric relationships in both segments for the alternative representing more expected bike usage. Public transport users in both segments also exhibit lower likelihood to decrease their expected bike usage after COVID-19. The parameters in both segments for the “decrease” alternative are not statistically significant, yet they are kept in the final model estimation since they correctly capture the expected behaviors. The parametric results are reasonable because such users may want to avoid public transport after COVID-19 to avoid possible virus transmission. Significant heterogeneity is observed in the case of variables representing expected activities after COVID-19. For instance, urbanite people in segment 2 exhibit a lower likelihood of decreasing their anticipated bike usage if they plan to perform no telecommuting and online grocery orders during post-COVID-19 times, which probably correlates with a built environment favoring active travel. In contrast, suburbanite people tend to decrease their expected bike usage. With the increase in trip frequency after COVID-19, individuals from both segments demonstrate higher inclination to increase their expected bike usage, perhaps indicating substitutions of their previous mode of travel with cycling since cycling is a safe (and becoming safer with improved infrastructure) and socially-distanced alternative mode. Furthermore, suburbanite individuals who plan to increase their car usage after COVID-19 are less likely to keep their expected bike usage the same. In contrast, urbanite individuals in segment 2 expect to keep their bike usage the same as during post-COVID-19 times. This might be attributed to the bike-friendly transportation infrastructures in urban areas that may encourage urban dwellers to keep using bikes alongside auto modes. As a built environment attribute, proximity index represents closeness to a facility in the neighborhood. Model results suggest that individuals have lower propensity to increase their expected bike usage as proximity to employment increases. Parametric value in the suburbanite segment is found to be statistically insignificant, however, it exhibits important insight. Perhaps, when everything starts to go back to normal during the post-pandemic era, the amount of motorized traffic will increase, which might make some bike users uncomfortable while traveling long distances from home to employment. Therefore, planners and policy makers could use this information to focus on encouraging road users to maintain “share the road” policies, developing dedicated and/or protected bike lanes, increasing the width and length of bike lanes, and so forth. However, caution is important in applying this finding to the general population of the study area because of the statistical insignificance of the parameter. Proximity index of grocery store exhibits negative relationship with higher expected bike usage in both segment 1 (suburbanite) and segment 2 (urbanite). This means that suburbanite individuals are less likely to increase their expected bike usage, whereas urbanite people tend to increase it after COVID-19. This outcome is probably related to the presence of active mode-friendly infrastructure in the neighborhood.

Marginal Effects

The parameter estimation results are unable to provide the magnitude of the impacts of the determinants on individuals’ anticipated travel behavior. To determine such magnitude of impacts, this study estimates the marginal effects (in percentage points) of the determinants across two segments. This study presents marginal effects of the five most impactful factors that affect the expected travel behavior after COVID-19 (see at the end of Tables 2–4). For example, in the expected changes in trip frequency model, among all the variables retained in the final specification, “primary travel mode: public transport” exhibits the highest impact (−13.5%) in the suburbanite people segment (for the “increase” alternative). This suggests that the probability of “increased trip frequency” after COVID-19 reduces by 13.59% in case of the suburbanite people who used public transport as the primary travel mode during the pre-pandemic period (assuming that other variables are constant). In the urbanite segment, bike ownership reveals the highest impact, 24.43% (for the alternative “increase”), which means that the likelihood of “increased trip frequency” in the post-pandemic time increases by 24.43% in the case of urbanite bike owners.

Conclusion

This study presents the findings of an investigation into anticipated changes in travel behaviors in the post-pandemic period using data from the time of the COVID-19 pandemic. It develops three LSL models to explore the expected changes in trip frequency, transit usage, and bike usage after COVID-19. Anticipated changes in such behaviors are estimated considering three alternatives for each dimension: less than before (decrease), same as before, and more than before (increase). The study utilizes data from the COVID-19 Travel Study survey that was administered in Montreal, Canada. The survey was used to create datasets that include information on individuals’ socio-demographic characteristics, built environment, and activity-travel characteristics before and after COVID-19. Within the LSL modeling framework, the sample individuals are implicitly allocated into discrete latent segments based on their socio-demographic and built environment characteristics. The estimation process suggests that models with two latent segments are the best models to explain individuals’ anticipated travel behavior. These latent segments are probabilistically identified as the segments of suburbanite people and urbanite people. All three LSL models examine the effects of a variety of factors affecting the anticipated changes in travel behaviors after COVID-19.

The study explores critical impacts of several socio-demographic characteristics, built environment, and activity-travel characteristics during pre- and post-pandemic times on individuals’ anticipated travel behaviors after COVID-19. It discovers considerable behavioral heterogeneity across the sample individuals. Results suggest that ownership of a monthly transit pass is more likely to increase individuals’ transit usage during the post-pandemic period in the urbanite segment, whereas individuals in the suburbanite segment have a lower likelihood of increase in their expected transit usage. Full-time employed people who plan to relocate their residence after COVID-19 demonstrate higher propensity to decrease their expected trip frequency. As to pre-COVID activity-travel characteristics, results suggest that suburbanite people who used cycling as their primary travel mode before COVID-19 tend to decrease their expected number of trips, however, urbanite people are less likely to reduce their trip frequency. In the case of public transport, individuals who used it as a primary travel mode before the pandemic show higher likelihood to decrease their expected trip frequency in both segments. It is found that such individuals are more likely to increase their expected bike usage after COVID-19. Expected results are found in the case of post-pandemic anticipated activities. Individuals who expect not to telecommute or order food online after the pandemic are less likely to decrease their anticipated number of trips. Plan to relocate residence after the pandemic increases the probability of reduced expected transit usage in both urbanite and suburbanite segments. Furthermore, as the proximity to employment increases, results suggest that individuals from both latent segments are less likely to increase their expected bike usage. However, heterogenous behaviors are observed in case of anticipated transit usage. Individuals belonging to the urbanite segment expect to increase their transit usage with proximity to employment; however, suburbanite people demonstrate a lower probability to increase their expected transit usage after the pandemic.

This study has certain limitations. It presents the estimation of expected changes in trip frequency, transit usage, and bike usage. However, to clearly understand the travel behaviors after COVID-19, it is important to estimate the changes in activity and travel attributes as well. One of the immediate future works of this study, therefore, is to estimate the changes in expected activity participation, private car usage, walk mode usage, and so forth using a similar modeling framework, and compare the estimated outcomes to gain a clear understanding of expected activity and travel behavior during the post-pandemic period. This study also anticipates heterogeneity across sample individuals based on their characteristics, but it is unlikely that all individuals with similar characteristics will have similar behavioral preferences. There could be behavioral differences across individuals with similar characteristics. Therefore, another immediate future work of this study is to include random parameters in the LSL modeling framework to capture the heterogeneity within each latent segment, so that the models can predict more behavioraly realistic outcomes. In addition, data used in this study were collected during the first wave of COVID-19 in April 2020. There is a possibility that travelers’ perspectives, attitudinal preferences, and perceived risks toward traveling may have changed during the next waves of the pandemic. COVID-19 vaccine availability, increasing awareness toward virus transmission, vaccination rates, and emergence of new variants of the coronavirus may change individuals’ travel preferences during the post-pandemic period compared with their opinions during the first wave of coronavirus. To capture such preferential changes, another wave of the COVID-19 Travel Study survey is currently being administered in the Greater Montreal Area, which will be used to re-estimate the travel behavior models. Furthermore, the survey used in this study did not collect respondents’ personal attitudes and perceptions toward risks during the post-COVID-19 period. Inclusion of such questions in the future survey could be useful to develop advanced behavioral models such as integrated choice and latent variable models, which would examine how individuals’ underlying latent behavior constructed from their attitudes and perceived risks might affect their travel behavior during the post-pandemic period. Another limitation of this study is not to accommodate the inherent ordering in the models. Future research should address this issue and develop models that can capture both the ordinal nature and unobserved heterogeneity across the population within the modeling framework, for example, latent segmentation-based ordered logit model (30) and latent segmentation-based generalized ordered logit model (31), among others. Finally, jointly estimating the changes in trip frequency, transit usage, and bike usage with the same class membership using copula-based Bayesian modeling approaches could be an interesting future work to understand individuals’ anticipated travel behavior. Nevertheless, this study provides important insights on expected changes in trip frequency, transit usage, and bike usage during the post-pandemic period. Models presented in this study will be used to develop forecasting techniques to predict individual-level future travel behaviors in the Greater Montreal Area. Also, results of this study can assist in the development of effective policies to support active travel and improve the transit ridership experience. For example, both urbanite and suburbanite full-time employed individuals who plan to relocate their residence after COVID-19 tend to expect a decrease in their trip frequency in the post-COVID-19 period. This may be because of their increasing inclination toward online activities. Planners and policy makers could use this information to develop interventions for efficient residential developments in urban and suburban areas, and reshaping the land uses in commercial and industrial areas. Also, leisure activities are found to have a positive impact on expected reduction in trip frequency during the post-pandemic period. Social distance measures and limited gatherings might lead the cause of this outcome. Therefore, urban planners could utilize this information to develop strategies for creating safer environments for leisure activities. Furthermore, probability of bike usage is less likely to increase after COVID-19 as proximity to employment increases. After the pandemic, when everything starts to go back to normal, the volume of motorized traffic will probably increase which might put the bike users at risk while traveling long distances from home to employment. Therefore, planners and policy makers could use this information to increase their focus on developing policy interventions such as increasing the length and width of bike lanes, establishing more dedicated and/or protected bike lanes, encouraging all road users to maintain “share the road” strategies, and so forth. Finally, since the results of the study demonstrate the existence of heterogeneity across individuals, target marketing strategies could be advantageous to develop various policy interventions. In summary, the findings of the study will assist the policy makers and planners to better prepare for the post-pandemic era by developing policies to accommodate the anticipated changes in travel behaviors.