Impact of COVID-19 restrictions on mode use and mode captivity the city of Santo Domingo in Latin America

This paper explores transportation mode choice patterns of a population in Latin America before and during the COVID-19 pandemic based on survey data from the city of Santo Domingo (Dominican Republic). An online survey consisting of two waves was conducted between April and June 2020 among randomly chosen transport system users. The survey consisted of both a stated choice experiment, revealed preference, and psychometric indicators. Hybrid choice models were developed and included two latent variables (1) satisfaction with public transport and (2) the response to COVID-19. The results indicate that there is a strong habit of private vehicle use in the Dominican population. The main deterrents for public transport use are limited capacity and reduced safety (including regarding social distancing). The findings particularly highlight the success of the metro as a public transport mode in the city of Santo Domingo. Both cost and time are significant factors in mode choice, before and during the pandemic, but more so for OMSA bus use than for metro use. One of the conclusions from this study is that ignorin covid-related latent effects on public transport service during the pandemic may yield biased (lower) estimations of Value of Travel Time Savings. Finally, the results also show that mode captivity for public transport users plays an important role during the pandemic. These users have no access to alternative transportation and are obliged to continue using public transport. This implies that governments would be wise to urge transport operators not to reduce capacity and frequency during the pandemic.


Travel behavior and mode choice
Transport-related research is becoming increasingly important, which is partly related to newly emerging transportation modes and services such as shared services, automated scheduling systems, and selfdriving cars. In addition to the fast-paced advance of technology and the urgent need for flexibility towards more energy-efficient services, citizens are becoming better-informed as well as more dynamic and flexible. This is changing our approach towards analyzing and understanding travel behavior. The arrival of the COVID-19 pandemic has further complicated the way users choose their mode of transportation. As transport interactions could play a major role in the spread of the virus, it is vital to understand how the pandemic impacts users' transport choices.
Several studies have focused on mode choice considering its multidimensional nature, as it incorporates both spatial and temporal correlations and is influenced by a multitude of factors. However, it tends to become habitual, and its proper characterization is often sidelined in research. The traditional approach does not work very well for assessing how the pandemic is changing users' travel behaviour. It is also known that for a long time, our understanding of people's travel behavior has been based on cross-sectional surveys in which only one day is surveyed for each respondent, which at the same time infers periods of maximal traffic flows (see for example Ortúzar et al., 2011). This means that there is a substantial gap in our knowledge, which hinders the design and implementation of effective transportation strategies, policies and measures. Innovative travel data collection methods would allow researchers to derive extensive origin-destination matrices for traffic modeling (Castaigne et al., 2009), advance modelling and more detailed scale of travel data. Schmid et al. (2019) analyzed travel time savings per different transport user type (e.g. transport mode), highlighting that residential location has the strongest impact on Value of Travel Time Saving (VTTS). The effect of user type differences should be taken into account in the design of transport pricing. For example, car users' value of time can be substantially higher than that of current public transport users. Furthermore, Hu (2016) suggested that transportation and land use policies need to address the specific needs of distinct population groups and underscored the importance of spatial access for the middle class, which tends to be overlooked in the literature on transportation equity. The extent to which user type and mode captivity affect mode choice in low-income populations requires more attention as well, given the increasing role of IT (i.e. Uber, micro-mobility, shared mobility) in transport. Policymakers have to guarantee that all areas and population groups are served and their transportation needs met. In general, the VTTS is used to assess the costs and benefits of infrastructure investments and assist in decision-making. However, it incorporates many variables that are difficult to compute accurately, such as variations in fuel costs, subsidies, how travel time is perceived by drivers, and the desire to drive due to social pressure. Shao et al. (2014) analyzed the willingness-to-accept (WTA) alternative transport of the private vehicle owner to investigate difficulties in quantifying the VTTS. Research to analyze the social inclusion role of transport in latin America and Caribbean is very scarse, despite the challenges faced by these countries in terms of rapid urban population growth, congestion and transport fares. Yañez-Pagans et al. (2019) conducted a review of effectiveness of transport interventions and highlighted the spatial mismatch of low-income minorities from skill-appropriate jobs and the possibilities to commute to the desired jobs. In practice, other authors have investigated how mode captivity (e.g. motorcycle) can introduce more benefits to higher income population when specific transport measures are implemented (Andani et al., 2021).
Similarly, Line et al. (2010) found that the transport intentions of young people are fueled by a desire to drive. In countries like the Dominican Republic, low-income groups cannot afford to drive a car, but there is still among all population groups a strong preference for car use. Car Pride has been defined by previous researchers as "car pride is defined as the self-conscious emotion derived from the appraisal of owning and using cars as a positive self-representation". Furthermore, this concept has been analyzed within a latent variable framework, utilizing liker scales (Zhao and Zhao, 2020). In Latin American countries, car pride may also impact mode choice decisions, and create a need to own a car even when it cannot be afforded.
Therefore, quantifying the impact of unobserved effects during this unprecedented situation of the COVID-19 pandemic (e.g. attitudes towards public transport and perceptions of social distancing measures) on VTTS variables will also generate solutions for improving public and private transport services in the future. This paper investigates the impacts of COVID-19 protocols on mode choice in a predominantly lowincome population, using the city of Santo Domingo as a case study. The results should apply to comparable situations in other developing countries.

Impact of COVID-19 on mode choice
The first COVID-19 infection outside of China was reported in Thailand on January 13, 2020. Since then, the pandemic has turned the world upside down, forcing it to cope with social distancing measures while also still accommodating the needs of the economy and preserving a satisfactory level of quality of life. In European countries with a "smart" lock-down, such as the Netherlands, the impact of COVID-19 generated more bicycle and walking trips (de Haas et al., 2020). Perhaps surprisingly, an analysis of ticket validations, sales, and passenger counts in Sweden shows no strong correlation between ridership and transmission of COVID-19 at an aggregated level (Jenelius and Cebecauer, 2020). However, it was beyond the Swedish study to disentangle a strong correlation between ridership and transmission of COVID. Similarly, Jenelius and Cebecauer (2020) argued that indicators based on human proximity to public transport stations and travel planner queries have overestimated the post-pandemic recovery of public transport ridership in the developed world.
In developing countries, by contrast, sociodemographic factors play a crucial role in mode choice. People in low-income groups are dependent on public transport and therefore may still use it frequently even when they are not satisfied with it and would rather avoid it. In India, only about 41% of commuters stopped traveling during the transition to lockdown (Pawar et al., 2020). Increases in transport fares, as well as reductions in capacity and frequency, were common in developing countries during the COVID-19 pandemic (see, for example, Mogaji, 2020). In addition, job losses reduced family's budgets, making affordable transport options a multidimensional social issue. Le et al. (2020) found that satisfaction with operations significantly affects the level of transit use. They also highlighted that the study of causal connections with satisfaction requires longitudinal surveys, but has been forced to rely on cross-sectional surveys. Khaddar and Fatmi (2021) analyzed how the pandemic has impacted satisfaction with traveling and found that sociodemographic attributes are significant. For example, traveling for work has a negative relationship with satisfaction, while traveling for shopping or household errands has a positive association.
De Vos (2020) published a European perspective on how COVID-19 has affected mobility; active modes such as cycling became more popular than public transport. In the case of the Dominican Republic, we might be dealing with mode captivity. This would mean that mode choice will not change as these public transport users have no suitable alternatives, even though they may find for example taxi or Uber use more attractive from the viewpoint of social distancing.
Working from home can bring changes in both the distance and frequency of non-mandatory trips; see for example Moeckel (2017). Income levels impact the mobility dynamics of population groups. For example, research has confirmed that online shopping has increased across countries, recreational trips decreased, but telemedicine for lowincome groups has not increased (Anwari et al., 2021). Indeed, a substantial asymmetry among income segments has been revealed in potential changes and the ability to adapt. For low-income groups, there simply are no other options for public transport or work on location and appointments for medical care. Anwari et al. (2021) found that the COVID-19 pandemic predominantly caused changes in trip frequency and mode preference for various trip purposes; this has implications for transport planning and policymaking. Thomas et al. (2021) analyzed the attitudes and intentions for commuting modes before and after the start of the pandemic in Australia and New Zealand. The authors identified a tendency to reduce public transport use and increase car use. They also found that regular users of public transport had more positive attitudes towards public transport before COVID-19 than those who were not regular users. It is important to quantify the impact of these attitudes in terms of travel time valuations. Travel time valuations are the main component of transport models and feature prominently in the design of transport services (e.g. fares and schedules).
Most of the studies that looked at attitudinal (intention) versus revealed (actual) travel behavior at later time points during the pandemic focus on Europe, whereas the behavior of Latin Americans has been less explored. However, it is imperative to understand the impact of COVID-19 on travel behavior to counteract any negative effects on travelers everywhere. Although there are travel restrictions and social distancing measures, essential trips must still be completed. Also in the Dominican Republic, the government's response to the pandemic affects traveler attitudes and perceptions, and influences mode choice and travel behavior. Vaccination programs work at a slower pace in developing countries, which means that particularly poorer groups of the population remain at risk and new infection waves may occur.

Methodology and purpose of this paper
This paper aims to explore the travel needs of a low-income population in the city of Santo Domingo before and during the COVID-19 pandemic and provide recommendations to improve travel policy during the evolution of the pandemic. More in detail, the present paper aims to fill a knowledge gap by analyzing the following emerging transport-related aspects of the COVID-19 pandemic: Mode choice by user type, including young and low-income populations, as descriptors of user type and public transport captivity; Changes in activity patterns (e.g. online versus physical); Satisfaction with transport services concerning social distancing measures.
Furthermore, we explore whether the VTTS has changed (i.e. decreased) due to the mode captivity of users during the various phases of the pandemic so far.

Use of a hybrid choice model to show Covid-19 impacts on travel behavior
Understanding the dynamics in sociodemographic and attitudinal factors can be considered crucial in the analysis of mobility. There is evidence that at least 50% of the improvements in overall model statistics are due to the presence of repeated observations (Cherchi and Cirillo, 2008). Concerning public transport chains, La Paix Puello and Geurs (2015) have shown that both attitudes and observable travelrelated elements are important in people's decisions whether or not to use the bicycle to travel to a train station. Variations in these perceptions and attitudes significantly affect the bicycle-train share. However, there are almost no studies that have analyzed transit attitudes and perceptions based on big data.
Discrete choice models are very popular in transport research for analyzing travel behavior. Any behavioral process is informed by people's perceptions and beliefs based on the available information and influenced by effect, attitudes, motives, and preferences (Ben-Akiva et al., 1999). The information about available alternatives is contained in observable elements, whereas attitudes, motives, and preferences form unobservable or latent elements. In this context, latent means potentially existing but not evident or realized. As explained by Ben-Akiva et al. (1999), perceptions refer to the cognition of sensation, while attitudes are enduring psychological tendencies to like or dislike certain outcomes or activities. Travel-related attitudes and behaviors influence each other over time (Kroesen et al., 2017).
Accurately estimating travel behavior presents a high degree of difficulty due to the multitude of factors that influence the traveler, such as sociodemographic factors and social norms as well as the aforementioned perceptions, attitudes, and motives. Hybrid choice models (HCMs) are utilized to detect these latent variables and incorporate them into discrete choice models to improve the modeling precision. Kim et al. (2014) explain that the latent attitudes must be identified through attitudinal factors as they are not directly observed from revealed choices and that the latent variable model uses indicators to demonstrate the correlation between external explanatory variables and the latent variables. Also, through the concurrent integration of the discrete choice and latent variable models, the latent variables can be handled as explanatory variables in the utility functions of the choice alternatives (Kim et al., 2014).
A body of literature is now developing on the inclusion of these unobserved or latent variables in choice models to capture attitudes and preferences that co-determine travel behavior; see, e.g., applications of HCMs by Paulssen et al. (2014), Espino et al. (2006), andGeurs (2015). The model function can incorporate zonal elements, level of service, and attitudes of users (so-called soft elements). Psychometric indicators are a well-known representation of soft elements. In discrete choice models, psychometric indicators are manifestations of latent variables (Walker and Ben-Akiva, 2002). Therefore, several studies have used psychometric indicators to represent attitudes and perceptions in the field of transport modeling (Hurtubia et al., 2014;Jing et al., 2014;La Paix Puello and Geurs, 2016;La Paix Puello et al., 2017;Yáñez et al., 2010), in addition to semi-open questions , for example, using individual replies on a Likert scale in which zero means 'cannot be worse' and 10 means 'excellent. Aligning big data outputs with the traditional data collection of soft elements continues to be a challenge for modeling purposes, however.
HCMs, with longitudinal series of latent variables, are expected to be able to support the development of policies over time (Chorus and Kroesen, 2014), but finding efficient and realistic measurements of latent variable indicators (over time) remains a major challenge for hybrid choice modeling. Using HCMs, La Paix Puello et al. (2017) have confirmed the importance of conducting repeated multi-day panel studies to understand the temporal dynamics of travel behavior; which simply are not captured in the one-day cross-sectional travel surveys that are currently still used all over the world.
In the present paper, a latent variable model is developed to represent user types and satisfaction. We analyze the preferences pre-COVID and during COVID for public and private transport modes and measure satisfaction via psychometric indicators over two waves of an online survey.
The remainder of this paper is structured as follows. Section 3 presents the data collection and sample statistics. Section 4 describes the model formulation, while the model results are discussed in Section 5 and Section 6 presents the conclusions of the paper.

Data collection and sample statistics
The research we present here was conducted in the city of Santo Domingo in the Dominican Republic. A combined panel/retrospective survey was conducted online for a random sample recruited by university students. We consider the present study a pilot. In the future, data collection will take place in two waves of GPS smartphone tracking. An app is being designed to collect mobility data. This app will track user movements over two weeks per wave. A household survey will be conducted to identify mobility patterns and invite respondents to use the app (mapping).
The data were collected in two waves, the first at the beginning of April 2020 -the beginning of the COVID-19 pandemicand the second at the end of June 2020. The first wave of the online survey contained 635 respondents and the second wave 629. Unfortunately, only 100 respondents over 600 completed the survey in the two "live" stages. Therefore, the drop-out between waves is 83%. The survey was composed of three parts, namely revealed preference, stated preference (stated choice), and satisfaction with the public transport service.
In the revealed preference (RP) section, the respondents were asked about their most frequent trips. The following variables were collected: sociodemographic characteristics, travel time, main transport mode, access and egress modes, origin and destination, frequency of public transport use, as well as the station of departure and destination in the case of public transport use.
In the satisfaction part, respondents were asked about their level of agreement with specific statements about capacity, safety, and parking at the public transport stop or station, as follows: • The frequency of transportation is appropriate to the needs of my trip.
• The capacity of the mode allows me comfortable and safe transport. • The waiting time is reasonable.
• Access modes are tailored to the needs of travelers.
• Available parking makes transfer easy.

Data source
The transport network data used in this research was retrieved from the National Institute of Ground Transport and Traffic (INTRANT). The database comprises the bus stops (OMSA), metro and cable car (Teleférico) stations. The stations were modeled in QGIS; Fig. 1 presents the complete public transport network for the city of Santo Domingo.

Sample characteristics
The survey was completed mostly by students (53.7% in Wave 1, 58.35% in Wave 2), between 16 and 24 years old (75.3% in Wave 1, 80.1% in Wave 2), with a medium to low-income level (e.g. less than 30,000 pesos a month; 54.3% in Wave 1, 61.1% in Wave 2). The majority of the respondents had a driving license (60% in Wave 1, 65% in Wave 2), and the main trip purposes are related to work and studies (75.9% in Wave 1, 71.4% in Wave 2).
The mode of transport most used by this sample is the private vehicle (51.3% in Wave 1, 60.6% in Wave 2), followed by metro (20% in Wave 1, 14% in Wave 2) and public cars and buses (18.7% in Wave 1, 13.5% in Wave 2). The respondents reported frequency and capacity as the main drawbacks of the public transport system. When receiving scenarios with changes in service levels (time and cost) of public transport and private vehicle, respondents showed a great tendency to choose a private vehicle. Based on this data, we then carried out the first demand models, with and without inertia effects.  The survey shows that the students travel both in a private vehicle (46%) and single (41%) or combined public transport (12%). That 75% of this sample was between 16 and 24 years old, is an important component for the interpretation of the results, as it can be seen that older travelers use more often private vehicles.
Regarding the spatial distribution of the trips, Fig. 3 shows a cluster map of the stations of departure for metro and cable cars, revealing that the most used stations of departure are located in the outer boundary areas of the metro/cable car network. This includes the stations Mamá Tingó and Gregorio Urbano Gilbert at the Northern boundary, Concepción Bona in the East, Maria Montez in the West, and Centro de Los Heroes at the Southwest boundary. The survey data also show that a large proportion of the trips are to universities such as INTEC, UNPHU, and UASD, as is not surprising, since the survey was carried out among mostly INTEC students and their peers. It is interesting to observe that many trips originate in areas near metro stations, but not near the cable car stations. Thus, the cable car is understood as a feeder transport mode for other main modes such as the metro.
The most frequent access and egress stations of the respondents were obtained via the Revealed Preference portion of the survey; this data can be used in the mode choice analysis. Multiple factors influence the use of the stations such as sociodemographic variables, travel distance, travel times, acceptable multimodality (bikes, parking, etc.), availability of feeder services and public transport connectivity.

Statistics of change between waves
Wave 1 of the survey was completed in April 2020. COVID-19 had been detected in the country over 1 month previously, but infections were still low and the government was just beginning to implement measures to reduce the propagation of the virus. Wave 2 was completed in June 2020; by then COVID-19 countermeasures were already in place and respondents were asked retrospectively how their journeys were before COVID-19 as well as how they were after implementation of the new measures. To differentiate between these two responses for Wave 2, we refer to them as "Wave 2 pre-COVID retrospective" and "Wave 2 during COVID". The following table (Table 1 shows the impact of COVID-19 measures on multiple trip characteristics such as trip purpose, frequency of public transport use, mode of transport, and satisfaction with public transport. As can be observed in Table 1, before COVID-19, 42% of the trips were for studies and this reduced to 14.94% during the pandemic. Shopping represented 6.77% of trips for Wave 1 pre-COVID), 3.97% for Wave 2 pre-COVID retrospective and increased to 29.25% for Wave 2 during COVID. Work represented 31.18% of trips during Wave 1, was estimated as 25.44% for Wave 2 pre-COVID retrospective, and reported as 27.5% for Wave 2 during COVID. Business trips showed a variation of less than 2%, thus work and business trips were not highly affected by the COVID-19 measures and seemed to stabilize soon after their implementation. The reliability of some of the information on recreational trips is debatable since the retrospective information does not resemble the data for Wave 1 and is closer to the data reported for Wave 2 during COVID, complicating the interpretation of the data. Lastly, trips for services/personal care increased from around 2% to 8.43% and trips for other purposes increased to nearly 4% from around 1 or 2%. Therefore, during COVID-19 measures, the main trip purpose changed from studies to shopping, while work and business trips were not highly affected by the measures, whereas trips for services/personal care demonstrated a notable increase.
The frequency of public transport use was highly affected by COVID-19 measures; the response of "never using public transport" increased from 37.32% (Wave 1) and 43.4% (Wave 2 pre-COVID retrospective) to  73.13% for Wave 2 during COVID. Infrequent public transport use (1-2 times a month) decreased from 19.53% to 4.13%. Additionally, 3-4 times a week use decreased from 11.50% to 2.86%, and daily PT users reduced from 19.69% to 6.20%. Hence, public transport experienced a significant reduction in ridership after the implementation of COVID-19 measures; some of this can be attributed to the temporary closures of the metro and cable car services in the beginning of the pandemic. The use of the private vehicle as the main mode of transport increased by 30.22% after the implementation of COVID-19 measures. Metro use dropped from a high of 20% to just 2.38%, while the OMSA buses and cable cars maintained a steady use share throughout. Public car and autobus services (non OMSA) bus use fell from 18.74% to 4.13%, while taxi use was mostly unchanged. Lastly, motorcycle and walking/ biking increased by under 2%. The private vehicle became the preferred mode of transport after the start of the COVID-19 measures for 81.56% of the respondents.

Satisfaction with PT services
Table 1 also shows the average score for the satisfaction questions included in the survey. In general, we can notably observe a substantial increase of the "neutral" answers for Wave 2 during COVID, likely reflecting that most of these respondents did not recently use PT.
The respondents' dissatisfaction with the headway of public transport remained mostly stable during COVID-19 with a less-than-1% increase and ending at nearly 31%. The "neutral/don't know" response increased by nearly 15%, likely due to the reduction in PT use during the pandemic. The percentage of those who were satisfied with the PT frequency reduced from 30.55% to 18.44%, indicating that the satisfaction with the frequency of PT was moderately affected by the COVID-19 measures.
As can be observed, the number of people satisfied with the capacity of PT slightly reduced during the COVID-19 measures; this concurs with the fact that the capacity of PT vehicles was reduced to improve social distancing within the vehicles, causing less availability for passengers. One might also be tempted to conclude that satisfaction with capacity has improved since fewer people indicated that they disagreed with the statement "The capacity of the mode allows me a comfortable and safe transport during COVID. However, as the number of "neutral" responses increased, this simply reflects that many people were not using PT in the second wave (during COVID).
Satisfaction with the accessibility of public transport decreased during the COVID-19 measures, possibly due to route closures and schedule modifications to overcome the lack of drivers and vehicles during the beginning of the pandemic. Finally, the level of satisfaction with the availability of parking places does not show significant changes. Table 2 shows the measurements for each statement shown to the respondents. The questions in Table 2 are meant to ascertain the public transport users' perceptions of how COVID-19 has changed their trips.

Psychometric indicators for PT
The results also show that 60.09% of respondents agreed that public transport was unsafein terms of COVID riskand preferred not to use it, while 19.72% disagreed and the remaining respondents were neutral on the matter. In addition, 55.17% disagreed when asked if they had no other options to conduct their journeys, whereas 18.76% confirmed that they had no other options. There was no use of new public transport modes for 37.84% of the participants, but 36.04% did use new (different) modes. 62.48% responded that they were avoiding public transport and using it only when extremely necessary, but 17.65% stated that they did not. This matches that 25.44% considered it possible to maintain proper distancing in public transport, while 52.94% did not. 32.27% of the surveyed answered that they completed more journeys by walking or biking, while 40.86% stated that they did not. Lastly, 59.14% preferred avoiding crowded places and traveling farther to shop in less crowded places; 20.35% did not.
Most respondents felt unsafe using public transport, considered it difficult to maintain social distancing, and avoid its use unless extremely necessary. In addition, the majority had other travel options available, which can explain the considerable increase in private vehicle use during the COVID-19 pandemic. Meaning that many private vehicle owners were willingly to use public transport but due to the reduction in service quality opted to return to their private vehicle. Also, there was an uptake in biking or walking to complete trips, and many respondents preferred to avoid crowded places and traveled farther to accomplish this. But these two options are not feasible for everyone due to various sociodemographic and/or geographic reasons. Hence, these results highlight the importance of traveler satisfaction and access to the public transport network.

Modeling framework
In the hybrid choice model, mode choice takes up the discrete choice part and the latent variable part accommodates constructs from the satisfaction survey. We developed the following five models: Starters Wave 1: A model for Wave 1 with the first group of respondents. This sample completed the survey in April, before the pandemic took hold of the Dominican Republic, and the context question was: "How was your most frequent trip".
Stayers Wave 2: A model for Wave 2 with stayers from the first group of respondents. This group had completed the survey in April and completed it again in June (Wave 2). The context question was: "How was your most frequent trip during COVID-19". Refreshment: A model for the refreshment sample, with Wave-2 respondents who completed answers with regard to the Wave-1 period retrospectively, as "how was your most frequent trip before COVID-19".
Refreshment Wave 2 with LV1: Model for Wave 2 of the refreshment sample, using the following context question: "How was your most frequent trip during COVID-19" and latent variable 1 (satisfaction PT).
Refreshment Wave 2 with LV2: Model for Wave 2 of the refreshment sample, using the following context question: "How was your most frequent trip during COVID-19" and latent variable 2 (COVID attitudes).
DCM model: a discrete choice model without latent variables for both Wave 1 and Wave 2.
Two latent variables were constructed based on the answers to the satisfaction questions and COVID-19 affected travel behavior. For models 1 to 4, we included the latent variable satisfaction, whereas the COVID-19 latent variable was added for model 5. The following sections Fig. 4. Model framework. explain the structure of both discrete choice and latent variable model.

Discrete choice model
The approach used consists of a multinomial logit model for mode choice. The discrete choice model was based on the following alternatives: Metro, OMSA, car, no-choice. The unit of analysis is trip level. The choice model, in this case, U jn , is the utility faced by individual n, taking j mode (choice) of transport: where the utility function is expressed as a function of a vector of sociodemographic characteristics (SE n ) with size l, and revealed preferences Mn with vector size m of individual n; t means the period of time, β LOS j is the vector of parameters associated with the level of service (LOS) variables (time, cost, distance). μ jn is the alternative-specific error component that captures the individual and household correlation with zero mean and standard deviation σ μ . ε jn is the generalized extreme value (GEV) error term, identically and independently distributed. As Fig. 4 shows, the latent class model is composed of a membership part explained by sociodemographic (income) and travel related variables. The discrete choice model is composed of a choice set (car, metro, OMSA and no-choice) explained by socioeconomic variables (age, income and gender), level of service variables (time, cost and transfers) and neighborhood accessibility (public transport density). Fig. 5.

Latent variable model (LV)
Factor analysis will be completed to identify the latent variables, among other elements, trip frequency and distance traveled as inertia measurements. The latent variables (LV) are explained via a set of explanatory variables, such as sociodemographic characteristics, household and individual preferences. Since both the latent variable and the indicators of (In) are assumed to be normally and independently distributed, their distribution indicators (fLV and fI) are given respectively by: (2) I n is the indicator; λ is the associated parameter to be estimated and σ is the error term, normally distributed with zero mean and standard deviation. S is a vector of parameters associated with the sociodemographic characteristics with k elements. Simulation is usually applied to estimate the mixed logit. Given that the values that describe the population parameter of the individual parameters as R values of μ jn ω are drawn from its distribution and the probability is calculated conditional on each realization, the simulated probability is the average of the conditional probabilities over R draws: The simulated log-likelihood (SLL) function can then be constructed as SLL ( μ jn ) = ∑ n,j ln(SP jnt ) and the estimated parameters are those that maximise the SLL. The SLL decreases as the number of repetitions increases (Train, 2000). The number of draws to use is a trade-off between computational time and accuracy (Hensher and Greene, 2002). The models were estimated with the BIOGEME extended package (Bierlaire and Fetiarison, 2009).
The (logit) probability is then estimated as a joint probability product of the probabilities between waves. It represents the probability of a sequence of (wave) modal choices evaluated at the parameters. An integral is computed over the distribution of respondents for each period (t) and wave (w).
Since with non-zero error components, utility is correlated over alternatives (Train, 2003), the estimation involves a covariance matrix of the random portions. The estimation is a function of the conditional choice probabilities.

Factor analysis for latent variable model
The first step to identify the latent variables was to conduct a factor analysis; Table 3 shows the results for the 12 selected statements. The analysis presents the association between variables, grouping them in two factors composed of those variables with communalities larger than 0.50. We used principal component to extract the factors, orthogonal rotation method and the communality index to determine which variables belong to each factor. Further, we used eigenvalues to analyze the total amount of variance that can be explained by a given component. Eigenvalues above 1 are considered more significant, and this value will be used in determining the final number of factors.
The factor analysis shows the two identified potential latent variables: Satisfaction with public transport service; (2) COVID-modified perceptions. The two components explain 62.89% of the total variance and two clear factor loadings can be determined from the rotated component matrix in Table 3. Questions 1 to 5 show high loadings for factor 1, with the questions pertaining to the users' satisfaction levels with PT. Questions 6 to 12 demonstrate high loadings for factor 2; those questions relate to the perceptions of change in PT during COVID. Thus, latent variable 1 (LV1) is identified as Satisfaction with PT and latent variable 2 (LV2) as the COVID effect.

Latent variable model
We developed a hybrid choice model where the latent variables were built over the Likert constructs of the survey. The model results show that sociodemographic characteristics become less significant when panel effects (error components) are added. This is a reasonable result because the error components represent the individual correlation.
The latent variable is significant for both car and metro users, although more significant for car users. It shows that there is a substantial amount of explanation of behavior related to the satisfaction with the current public transport service. The parameter of the latent variable is shared for both OMSA and metro, which means that the effect of satisfaction is similar among public transport users of different modes. (Car users in the RP did not complete the satisfaction survey.) Satisfaction is negative for both car and PT users, but less negative for car use. It means that PT users are unsatisfied, but still use the service, probably because they do not have any other options (captivity). Similarly, in a developing country context, Pawar et al. (2020) found that though people perceived public transportation as unsafe relative to personal vehicle use, the actual commute patterns did not validate this, with a possible reason that these commuters have no alternatives.
A share of respondents (23%) did not complete the satisfaction questions, indicating 'neutral' due to lack of familiarity with the public transport system. Therefore, the levels of agreement were coded with neutral as equal to zero, agree and strongly agree as positive (+1, +2), and disagree and strongly disagree as negative (− 1, − 2). The estimated parameters in the latent variable model show that people with kids are less satisfied with public transport. Similarly, car owners feel negative about a (potential) trip by PT. The results of the LV model show that less frequent PT users have a more negative satisfaction level concerning PT. Also, adults have a more positive impression than youngsters. Consistent with the work of Khaddar and Fatmi (2021), our results show that access to transport resources affects the level of satisfaction; when people cannot travel by their preferred mode, this influences their satisfaction.
The model results for the refreshment sample, including both Wave 1 and Wave 2, show that patterns are very similar for starters (respondents who answered the survey in April) and the refreshment sample (respondents who answered the survey in June). The latent variable is significant and positive, which shows that people with low satisfaction with public transport tend to choose the metro and OMSA less. Furthermore, the results of the stayers model show that the latent variable is statistically significant for the stayers during the pandemic. The negative sign of the latent variable shows that people were not using public transport during COVID even when they had a positive image of it, whereas people who had a low level of satisfaction with PT (negative image) did use it. The second latent variable, COVID-affected perceptions, shows a negative sign for both car and PT alternatives. It means that there were changes in perception and habits related to PT for all respondents during COVID. Also, the perceptions of PT travelers display a stronger, negative, influence of COVID than car users. Table 4 shows the estimation results of six models based on the three samples (starters, stayers and refreshment) and the two waves of data collection (Wave 1 -Pre Covid and Wave 2 -During Covid). Five HCMs were estimated and one DCM (without latent variables). The model results show that both cost and time are significant factors for mode choice, where time and cost are most strongly penalized for the OMSA, followed by the car and least strongly for metro use.

The choice models
We can observe that the LOS parameters remain significant after the addition of the latent variables. The sign of the cost and time parameters is negative, which means that increasing cost and time for any of the modes discourages their use. The most significant LOS parameter is time for OMSA, indicating a high sensitivity of OMSA bus users to travel time. Similarly, several transfers harm metro use, but it surprisingly has a positive impact among OMSA users of the Refreshment sample. This can be credited to the OMSA's extensive network that allows longer trajectories and facilitates transfers with other routes. The results show that the willingness to pay for transport when using the metro is higher than when using the OMSA. Also, the parameter of cost presents a lower sensitivity for the metro, both before and during COVID-19. It can be associated with either the (lack of) availability of a car or the possibility     to undertake parallel activities while traveling on the metro (multitasking). Younger transport users have a stronger tendency to use OMSA buses. Certain variables such as household size and work purpose were not significant or unstably significant (due to correlations). Finally, the RP choices affected the SC choices. People who were already using the car were more willing to select the car in the experiment (RP_car parameter).
Several sociodemographic characteristics were significant in the differentiation between public transport and private vehicle users, such as income level, gender, availability of driver's licenses, and car ownership. Consistent with our results, Abdullah et al. (2020) found that gender, car ownership, employment status, travel distance, and the primary purpose of traveling are significant predictors of mode choice. Furthermore, having a low income and being male makes people more willing to take PT, while high income is associated with a willingness to use a car, in line with Abdullah et al. (2020). Travel distance was a relevant factor, in agreement with other studies, meaning that short travel distances are more likely to be covered by a car. Abdullah et al. (2020) found that respondents traveling over longer distances were less likely to choose private transport than public transport, relative to those traveling shorter distances. As expected, infrequent users of public transport were more willing to select the car and this is the case in both waves and all samples (starters, refreshment, and stayers). We also found that frequent PT users were not willing to choose the car in our experiment. It is probably associated with the (lack of) availability of a car at home. To identify the desire to drive a car when no car is available, the stated choice experiment included all alternatives (metro, bus, car, and no-choice) as available at all times. Respondents who were taking a multimodal trip were more willing to select the OMSA bus, reflecting the impact of the number of transfers, and respondents traveling for work purposes were more willing to use the metro (possibly reflecting the ability to be able to read or make calls during their trips).

Value of travel time (VVTS)
In this section, we analyze to what extent the value of travel time (VTTS) is affected by the latent variables (LVs). It turns out that when LVs are incorporated into the model, the VTTS becomes higher. It means that willingness to pay for transport cost is affected by unobserved elements, such as satisfaction and COVID policies, and not only by level of service variables (cost, time, and transfers). These results show that ignoring satisfaction with the public transport service during the COVID pandemic may bring biased estimations of VTTS.
The VTTS pre-COVID (Wave 1) is lower than the VTTS during COVID (Wave 2). We observe that the LV1 induces a lower VTTS than LV2. It means that COVID-related latent effects (LV2) increase the VTTS.
Regarding mode differences, the VTTS for car users is the highest, as expected. Car users are willing to pay higher prices for transport to ensure their comfort. Second, metro users have a more expensive VTTS than OMSA (public bus) users as metro services are more expensive (e.g. 20 RD$ vs. 10 RD$ per distance within the metropolitan area) and sometimes more comfortable than OMSA in Santo Domingo.

Conclusions
The objective of this study was to analyze the impact of two latent variables, satisfaction with public transport and impacts of COVID-19 on travel attitudes on mode choice patterns of a low-income population in Santo Domingo in the Dominican Republic. We identified the impacts of unobserved factors related to the COVID-19 pandemic on travel time valuations, which impacts transport fares and estimations of travel costs for any transport model. The novelty of this paper lies in both the used methods and the conclusions. We utilized an online combined panel/ retrospective survey, the results of which were later fed into a hybrid choice model, eventually leading us to the following main conclusions. Firstly, the model results demonstrate changes in travel time valuations concerning satisfaction and health aspects (COVID risk). In our study, people were more tolerant of delays in private vehicles, likely due to the lower COVID risk when traveling alone instead of in busy public transport. The pandemic resulted in a higher sensitivity to time, however, possibly due to the strict curfews imposed by the government. It is important to note that the metro transport service has a limited networkmostly serves the metropolitan area, with anaccess/egress connectivity, that can be improved. The VTTS results show higher valuations of travel time during the pandemic than before the pandemic. It means that people are willing to pay more for a better transport service during the pandemic.
Secondly, from a transport policy perspective, the results indicate that satisfaction plays an important role in mode choice in addition to the level of service, and the perceived public satisfaction levels are quite low. Therefore, if satisfaction levels are increased, it is possible to boost the use of public transport. The metro in Santo Domingo showed a lower sensitivity to time and cost than the use of OMSA and private vehicles. It means that connectivity of access and egress modes is a major factor for increasing ridership and enhancing multimodality and accessibility of the public transport network. In the city of Santo Domingo, as in many other developing countries, this can be improved by the centralization of public transport authorities, route network optimization, and improvements in the provision of information to users. Overall satisfaction levels were low for both users of private vehicles and of public transport, but there was still a preference for private vehicle use. This means that lacking a better alternative, public transport users are generally obliged to continue using the public transport service.
Finally, the findings of this study emphasize the lack of resilience of transport services concerning overcoming natural disasters or other natural phenomena, such as a pandemic, in developing countries. As a result, vulnerable populations have been substantially and multidimensionally affected by the pandemic. We advocate for greater equity in transport use, providing safe and healthy travel options for all income levels. Regarding social distancing, public transport operators should focus on making public transport a safer choice during the current COVID-19 pandemic. To that end, governments should encourage public transport operators to maintain the normal frequencies and not reduce capacity as a result of lower ridership. Several limitations can be acknowledged for this study. First, this study focuses on the low-income population, but a substantial amount of respondents are students (approximately 53%), which can be a correlation between income level and working stage of life. Secondly, the study does not include built environment measures to characterize residential zones. The authors expect to include these measures in the upcoming research steps of the current project. Third, informal transportation has not been included in the stated choice experiment, but it has been included in the revealed preference part. Other stated choice experiments could include the informal bus transport as part of the alternative set, but this would add complexity and increase the number of profiles.
In future research, travel data will be collected via a GPS smartphone application over multiple waves, with an automatic trip and mode detection during and, presumably, after COVID. This will provide more accurate trip data since it will not depend on the respondents' reporting. The use of big data can be invaluable for mode choice modeling and analyzing the temporal and spatial distribution of individual travel patterns. For example, Wang et al. (2011) successfully correlated human mobility and social interactions using big data from a database with six million telephone users. Chen et al. (2016) highlighted the synergies between big data and small data for travel behavior analysis and indicated that the discovery of mobility patterns from big data offers the opportunity to identify and create models of microscopic individual choices. Unfortunately, progress is stalling. The longitudinal data collection will allow the identification of intra and interpersonal variations of transport modes and destinations. This will enable the government to assess the effects of the taken measures as well as develop bettertailored measures during the current pandemic or in the event of a new pandemic. This research will provide public transportation authorities of developing countries with lessons learned from public transportation dynamics during the pandemic. It will contribute to the design of more equitable transportation policies during emergency times, how to create a resilient public transport system and how to reflect the effects of users' needs in terms of the value of time and transport fees.
Annex Considering the uniqueness of our sample, the following tables have been added to show the descriptive statistics for the variables income level and household car ownership. Unfortunately, the variable car ownership per household does not reflect individual car ownership. However, we can observe in the next table that 53% of the sample is students, and among these, 76% belongs to the low-income category. Within the low-income category, 22% have no car at home. And 35% has only one car at home. We can observe from the table that 66% of the students do not own a driver's license.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.