Advanced modelling of commuter choice model and work from home during COVID-19 restrictions in Australia.

The decision to work from home (WFH) or to commute during COVID-19 is having a major structural impact on individuals' travel, work and lifestyle. There are many possible factors influencing this non-marginal change, some of which are captured by objective variables while others are best represented by a number of underlying latent traits captured by attitudes towards WFH and the use of specific modes of transport for the commute that have a bio-security risk such as public transport (PT). We develop and implement a hybrid choice model to investigate the sources of influence, accounting for the endogenous nature of latent soft variables for workers in metropolitan areas in New South Wales and Queensland. The data was collected between September-October 2020, during a period of no lockdown and relatively minor restrictions on workplaces and public gatherings. The results show that one of the most important attributes defining the WFH loving attitude is the workplace policy towards WFH, with workers that can decide where to work having a higher probability of WFH, followed by those that are being directed to, relative to other workplace policies. The bio-security concern with using shared modes such as public transport is a key driver of WFH and choosing to commute via the safer environment of the private car.

Introduction

The COVID-19 pandemic has had serious global implications, particularly for health matters, which have resulted in significant changes in the way businesses and people operate on a daily basis. Businesses around the world have adapted quickly to the changing circumstances, allowing their workers to work from home (WFH) when possible – which has had a major influence on not only the nature of where and how work is done, but on the performance of the transport network. During these uncertain times, attitudes, perceptions, and beliefs are likely to be playing, and continue to play, a very important role in individual choice making and behaviour related to working from home and/or commuting, and most notably on the use of shared modes such as public transport (PT).

In this paper, a hybrid choice model will be used to identify the nature and role of underlying attitudes, perceptions and beliefs that influence the decision to work from home for a specified number of days per week, and how this relates to the incidence of commuting by day of the week and time of day. The hybrid choice model is estimated to account for the latent structure associated with a number of important soft variables related to perceived productivity and overall advantages of working from home, and concern towards the use of public transport due to COVID-19. These soft variables cannot be directly measured (i.e., asked to respondents), so they are estimated using individual responses to multiple attitudinal questions included in the survey (e.g., I have felt productive this last week while working from home, or I am able to find a balance between paid and unpaid work). The choice model considers 12 different alternatives for each day of the week: not to work, WFH, and to commute by up to 10 different modes of transport (depending on which modes are available to each individual). The latent variables feed into this model in an endogenous way to understand how the attitude towards WFH and the concern towards PT influence the probability to not work, WFH or commute by a specific mode each day of the week. The data was collected during late 2020 as part of a larger ongoing study to understand the implications of COVID-19 on the transport network around Australia (see Hensher et al., 2021a, Hensher et al., 2021b; (Beck and Hensher, 2020). Data from workers in two metropolitan areas will be used, the Greater Sydney Metropolitan Area (GSMA) and Southeast Queensland area (SEQ).

This paper is organised as follows. The next section presents a brief background of literature on the impacts of COVID-19 on travel behaviour and how the pandemic has been considered to date within the setting of a hybrid choice model. Section 3 presents the data used in this study. The next section describes the methodology used in this study, followed by the model results and elasticities. Section 7 presents simulated scenarios, and the final section discusses policy implications and conclusions.

Literature review

COVID-19 and working from home

The earliest studies on working from home, termed telecommuting, focused on how information technology was changing the nature of white-collar work and how a progression towards higher levels of remote working might create negative externalities with respect to social interaction and the work-life balance, but also how it could generate benefits associated with greater flexibility in time management and positive social externalities such as reduced congestion (see pioneering studies such as (Salomon and Salomon, 1984, Nilles, 1988, Hall, 1989, Kitamura et al., 1990, Mokhtarian, 1991); and (Maynard, 1994). While telecommuting did not quite grow at the rate first envisaged, indeed in Australia only 8% of people had a formal WFH arrangement (Pc, 2021) and only 5% of people were working from home on the day of the most recent census (Abs, 2016), studies in transportation continued to examine factors affecting the uptake of telecommuting (Mokhtarian et al., 1998, Choo et al., 2005, Fu et al., 2012, Caulfield, 2015).

However, as a result of the COVID-19 pandemic there have been many studies that examine the impact of the virus and associated measures in different jurisdictions to reduce the spread. For example, Transport Policy launched a special issue on the impact of COVID-19 on passenger transit (Transport (Policy, 2021). There has been great interest in the way in which travel restrictions have helped stop the spread of the virus (Chinazzi et al., 2020) or how travel can spread the virus (Böhmer et al., 2020), and how transport policy makers might need to react to these changes (Zhang, 2020, Gkiotsalitis and Cats, 2021). Similarly, there has been interest in the role of working from home, mostly in literature outside of transportation examining the impacts of working from home on work-life balance (Chu et al., 2022), the strengths and weaknesses of WFH and associate transition policies (Vyas and Butakhieo, 2020), however research has examined the intersection of lockdowns, working from home, transportation and mental health (Dam et al., 2020). In interesting work conducted prior to COVID-19, an Australian study found that people’s desire for greater work location flexibility increased once experiencing such freedoms (Hopkins and McKay, 2019). In the US, several studies have shown that as with the earlier research in transportation, that occupation is often the best predictor of working from home both before COVID-19 and currently (Bick et al., 2021) and in echoing themes from (Beck and Hensher, 2021) that working from home is likely to stick due to better-than-expected experiences (Barrero et al., 2021).

Given that it is likely that working from home will be a greater feature of work moving forward from the pandemic, it is important to begin to understand how choices are made regarding working from home beyond the aggregate level, to look more closely at what might motivate choice at a disaggregate, individual level.

Hybrid models of choice behaviour

Hybrid choice models which integrate discrete choice models with latent variables have been developed over a number of years, with early examples by (Walker, 2001, Walker and Ben-Akiva, 2002), and are reviewed in a number of sources such as (Hensher et al., 2015). These models emphasise the importance of amorphous influences on behaviour such as knowledge and attitudes. Several articles have found interesting results when incorporating latent variables, that allow for a better understanding of individuals preferences and how these underlying attitudes affect their choice making (Daly et al., 2012, Prato et al., 2012, Morikawa et al., 2015, Beck and Hess, 2017). COVID-19 as an extreme event has influenced in a significant way how we work, travel and live, resulting in changes that have impacted on the transport systems, notably attributed to working from home and bio-security concerns over using public transport and other shared modes. Our interest in this paper is how we might integrate revealed preference data on actual changes in travel behaviour and the growth in WFH with a number of soft latent variables that represent underlying attitudes and opinions that condition observed travel and non-travel activities. Before presenting new evidence, we review a number of existing studies that are relevant to the focus of this paper and that also used hybrid choice models in this context.

(Beck et al., 2021) use data collected across three waves in Australia throughout 2020 to study the impacts of COVID-19 on the number of trips by public transport. They used a zero inflated Poisson regression model to explain the number of trips by public transport by different socioeconomic characteristics and attitudes towards the use of public transport. They incorporate latent variables as explanatory variables in the utility function (estimated through factor analysis). The findings suggest that individuals concerned about public transport before COVID-19 and during the first (March 2020) and second (June 2020) data collection periods, usually make more weekly trips, suggesting that greater exposure is driving the attitudes towards hygiene and risk. Those that are more concerned about the hygiene in public transport tend to have higher odds of making zero public transport trips in the data collected after COVID-19, switching to greater use of the private car.

Balbontin et al. (2021) undertook a similar study using data collected in Australia in late 2020. They incorporate latent variables as additional explanatory variables in a model explaining weekly commuting and non-commuting trips. The results in metropolitan areas suggest that people that love WFH tend to commute more than those that do not love WFH, suggesting that people prefer a balance between WFH and going to the office. In terms of concern towards public transport, these results suggest that public transport commuters also tend to do less work-related travel and fewer social/recreation trips. The focus of this study was understanding the influences on the number of weekly commuting and non-commuting trips, as opposed to the current paper, which focuses on identifying the influences on the daily probability to not work, work from home or commute.

(Hurtubia et al., 2021) estimate a hybrid choice model that integrates a discrete choice model with latent variables to identify the probability of working from home versus going to the office. They use data collected in Chile at the beginning of the pandemic, during March 2020 and consider two latent variables: one associated with concern towards health and the other one concern towards the economy. Their results show that female respondents, respondents with a lower income, and older residents, are more concerned about health, which has a positive influence on the probability to WFH, possibly due to greater transmission risk outside the home and the severity of COVID-19 if they catch it. Their results also suggest that respondents with a lower income, without secondary education, or independent workers, are more concerned about the economy, resulting in a negative influence on the probability to WFH.

(Aaditya and Rahul, 2021) also develop a hybrid choice model to identify the influence of awareness of COVID-19 on the modal shift in India. Their results suggest a significant modal shift from public transport to personal vehicle use, attributed to the increased awareness of COVID-19 and a perception of the deterioration in public transport safety. Their findings suggest that respondents were willing to reconsider using public transport in a post-lockdown scenario if several preventive measures towards COVID-19 were implemented (e.g., social distancing, crowd management, and sanitisation).

(Hensher et al., 2022), developed a model system to identify the choice between working from home as opposed to commuting and not working by day of the week and time of day in Australia. They estimated a mixed multinomial logit model to determine and map the probability to WFH using individuals’ socioeconomic characteristics, modal attributes such as time and cost, day of week and time of day and some attitudinal variables such as concern about using public transport. The model outputs provide evidence of the key drivers of the probability of WFH compared to commuting over a week and form the basis of a mapping equation in the construction of a full origin–destination (OD) matrix for the study area to identify the varying spatial incidence of WFH across the OD pairs.

Since the start of COVID-19 in March 2020, there has been an accumulating body of research undertaken to gain an understanding of the impact that the pandemic has had on some major consequences, notably WFH and reduced use of public transport. These studies have identified a range of key influences, both of a quantitative nature such as socioeconomic descriptors, but also of a qualitative nature such as attitudes and opinions. These latter latent variables require an endogenous treatment when mixed with the other explanatory variables, resulting in a preference for a hybrid choice model. To the best of our knowledge, there are no current studies that have investigated attitudes towards public transport and work from home in the context of daily mode choice into a hybrid choice model, including not to work and work from home as possible alternatives. In this paper we focus on gaining a greater understanding on how attitudes towards work from home and concerns about COVID-19 are influencing individuals’ decision not to work, to work from home or to commute by different modes of transport.

Data Description

The data was collected during late 2020 as part of a larger ongoing study to understand the implications of COVID-19 on the transport network around Australia (Beck and Hensher, 2020); Hensher, Beck, & Wei, 2021). Data from two metropolitan areas will be used in this paper, the Greater Sydney Metropolitan Area (GSMA) comprising the Sydney Metropolitan Area, Newcastle and the Illawarra/Wollongong, and Southeast Queensland area (SEQ), the latter comprising the Sunshine Coast, Brisbane metropolitan area and the Gold Coast. During the time period being analysed, Australia had pursued an elimination strategy with relative success, having emerged from lockdown in June 2021 and (outside of Victoria) having had a sustained period of zero community transmission, with COVID-19 cases in the GSMA and SEQ being almost exclusively within the hotel quarantine system. As such, in October 2020 both metropolitan areas had returned to minimal government-imposed restrictions on travel, activities, and work.

Respondents answered questions about work behaviour prior to COVID-19, for example: which days they worked (Monday, Tuesday, etc., including weekends), where did they do work each day, and the modes of transport available to them for commuting and non-commuting. A number of attitudinal statements were included, with a specific focus on attitudes towards WFH and public transport use (PT), which will be used as indicators in the hybrid choice model. 650 respondents completed the survey, indicating where they worked from each day of the week. In this paper, the decision to not work/WFH/commute will be modelled for each day of the week separately, which will be detailed in Section 4. For each respondent we have up to1 7 observations representing each day of the week and, in total, we have 4,518 observations for modelling. An overview of the sample is presented in Table 1 . 63% of respondents are located in the GSMA and the rest in SEQ. 39% of respondents have their own place or room to WFH. In terms of WFH, prior to COVID-19, on average, respondents worked 0.86 days from home, while they worked on average 1.64 days last week (i.e., at the time of the survey). The total number of days worked is relatively similar, with an average of 4.59 days worked prior to COVID-19 and 4.51 last week. We note that in comparison to the population level statistics there is an overrepresentation of females in the sample data, and white-collar workers are also disproportionately represented (which is to be expected as they are the occupation groups more likely to be able to WFH; particularly WFH more often).

Table 1

General sample characteristics.

Variable	Sample Mean (std deviation)	GSMA & SEQ Averages (Abs, 2016)
Age (years old)	40.10 (13.40)	42.29 (13.81)
Gender female (1,0)	0.64	0.51
Income ('00AUD$) personal	78.13 (51.81)	94.28
Number of adults in household	2.79 (1.32)	2.8
Number of cars per adult in household	0.65 (0.36)
Occupation labour and machine operators (1,0)	0.06	0.14
Occupation white collar (1,0)	0.84	0.66
Workplace located in CBD (1,0)	0.21
Has their own space to WFH (1,0)	0.39
Located in the GSMA in New South Wales (1,0)	0.63
Located in Brisbane (1,0)	0.21
Located in the Sunshine Coast (1,0)	0.05 (0.22)
Number of days WFH last week	1.64 (2.11)
Number of days worked last week	4.51 (1.28)
Number of days WFH prior to COVID-19	0.86 (1.60)
Number of days worked last week prior to COVID-19	4.59 (1.06)
Sample size	650

For those respondents who made a commuting trip(s) over the last week, the trip characteristics reported are presented in Table 2 . It is important to note that, for each day of the week, respondents said if they worked and where they worked from. However, respondents identified once which modes they had available to go to work and which one they used last week and prior to COVID-19, regardless of how many days of the week they commuted. For each mode available, respondents identified their service levels such as travel time and costs. The majority of respondents (68%) used a private car to go to work last week (prior to COVID-19 it was 62%), while 12% used active modes (walk, bicycle including e-scooters) (prior to COVID-19 it was 8%). In terms of mode changes, 27% of the respondents that used public transport to go to work prior to COVID-19 used car last week, 11% used active modes, while the rest are still using public transport. As to respondents that used car to go to work prior to COVID-19, 95% are still using car, 3% used active modes last week, and 2% used public transport. The average travel time by active modes is of 33.3 min, the average in-vehicle time in public transport is 32.9 min, taxi/rideshare averages 25.2 min, and in private motorised vehicles, the average time is 28.6 min.

Table 2

Commuting trip characteristics.

Variable	Mean (std deviation)
Used car to go to work last week (1,0)	0.68
Used public transport to go to work last week (0,1)	0.19
Used bicycle or walked to work last week (1,0)	0.12
Used car to go to work prior to COVID-19 (1,0)	0.62
Used public transport to go to work prior to COVID-19 (0,1)	0.29
Used bicycle or walked to work prior to COVID-19 (1,0)	0.08
Walking or bicycle available to go to work (1,0)	0.31
Public transport available to go to work (1,0)	0.66
Rideshare/taxi available to go to work (1,0)	0.32
Car driver, passenger or motorcycle available to go to work (1,0)	0.83
Walking or bicycle travel time (minutes)	33.31 (27.63)
Public transport in-vehicle travel time (minutes)	32.86 (23.97)
Rideshare/taxi travel time (minutes)	25.15 (22.26)
Car driver, passenger or motorcycle travel time (minutes)	28.64 (30.96)
Public transport fare (AUD$)	6.83 (8.91)
Rideshare/taxi fare (AUD$)	39.05 (56.33)
Car driver, passenger or motorcycle cost (AUD$)	6.22 (15.35)
Public transport access, egress and waiting time (minutes)	42.89 (33.02)

Fig. 1 summarises the work from home policy of participants’ place of employment at the time of the survey, i.e., end of 20202 . Over 30% of respondents said their place of employment did not have plans to allow them to WFH, 29% of respondents said they were given the choice to WFH when they chose to, 21% of respondents said it is not possible for them to WFH as they need to be onsite to do their job, and over 19% of respondents said their place of employment is directing them to WFH.

Fig. 1

Work from home policy of their place of employment as it stands today.

Fig. 2 presents the commuting, work from home and not work partition in the sample, by day of week. Work from home is relatively stable across weekdays, varying between 35% and 39%. The variation in behaviour is observed mainly in commuting and not working on any given day, perhaps indicating that those who are required to travel to work are working less days per week compared to those who are able to work from home. On Fridays, the percentage of participants not working is the highest across weekdays, where 26% of participants reported not working; the lowest is on Wednesdays where less than 15% of participants reported not working. On weekends, the majority of participants do not work (almost 84% Saturday and over 90% on Sunday), but the existence of a reasonable amount of weekend work means that all seven days should be included in the analysis in WFH opens up greater flexibility for many workers as to when work is undertaken. There seem to be some (see (Hensher et al., 2022, Beck and Hensher, 2020) differences across days of the week, although the main change is between weekdays and weekends, as expected.

Fig. 2

Commuting, WFH and no work behaviour.

Fig. 3 summarises the overall modal share prior to COVID-19 compared to the situation at the time of the survey. Results show a significant increase in the use of motorised private vehicles (car driver, passenger and motorcycle), with a particular increase in car driver, as is expected given the health concerns associated with shared modes. The results indicate a significant decrease in the use of train from almost 19% to slightly more than 12%; and in the use of bus from almost 10% to almost 7% of participants. The modal share of light rail and ferry is not significant compared to the other two modes of public transport: bus and train. In terms of active transport, results show an increase from 5.9 to 9.7% in walking, whereas the use of bicycle has remained almost the same.

Fig. 3

Modal share prior to COVID-19 and currently.

A particularly interesting finding related to the modal services is shown in Fig. 4 , summarising the sample’s experience when waiting for public transport. Around 33% of participants do not have public transport available to go to work (which is accounted for when modelling mode choice), so they did not answer this question; approximately 59% of participants said they entered a PT mode when they wanted to without delay, while 7% said they had to wait longer than normal, and less than 1% said they had to wait so long that they gave up using PT. In this study, we are interested in understanding if the experience in waiting for PT, which is related to crowding and frequency, has any impact on participants’ concern towards using public transport.

Fig. 4

Waiting time experience last time they used public transport.

The model framework

The hybrid choice model is formed by two models that are estimated simultaneously: (1) an ordered probit model that represent the latent variables, measured using attitudinal questions in the survey; and, (2) a mixed multinomial logit model that represents the WFH/commute decision for each day of the week. The WFH/commute model considers three alternatives for each day of the week: not work, work from home, or work outside home. If someone decided to work outside home, the mode used is relevant in understanding individual commuting behaviour. The daily alternatives’ structure is presented in Fig. 5 .

Fig. 5

Individuals’ daily alternatives structure.

For each day of the week, respondents can have up to 12 alternatives available, which are presented in Table 3 . The alternatives available will depend on which modes of transport are available to the respondent for commuting, and if they can work from home.

Table 3

Alternative numbers per DoW.

Monday - Sunday
Altij	Description
1	Not work
2	Work from home only
3	Work outside home - car driver
4	Work outside home - car passenger
5	Work outside home - taxi/rideshare
6	Work outside home - train
7	Work outside home - bus
8	Work outside home - light rail
9	Work outside home - ferry
10	Work outside home - walk
11	Work outside home - bicycle
12	Work outside home - motorcycle

The overall modelling framework is presented in Fig. 6 , showing the latent variable and the WFH/commute model. The structural and measurement equations of the latent variable model, as well as the WFH/commute choice model each have their own associated error term:, and, respectively. The proposed WFH/commute choice model accounts for preference heterogeneity through random parameters (error components) and allows for the panel effect across the observations related to the same individual for different days of the week. This error component is alternative-specific, that is, there is a for the no work alternative, and one common for all the commuting alternatives. The second one creates a hierarchical structure allowing for a correlation between all commuting alternatives.

Fig. 6

Hybrid model framework.

There is an additional error component considered,, which takes into account the relationship between the structural equations and the WFH/commute choice model derived from using simultaneous estimation of the hybrid choice model, referred to as serial correlation (Bierlaire, 2016, Sottile et al., 2019). If this error term was not included, the simultaneous estimation would be assuming that the error terms involved in these models are independent. Serial correlation is taken into consideration by including an agent effect in the model specification, which is an error component in all the models involved (i.e., structural equations and mode choice).

The latent variables refer to variables that cannot be directly observed but are explained by certain indicators. A factor analysis was carried out using all the attitudinal variables included in the survey (Balbontin et al., 2022). The results suggested five different factors: WFH loving attitude, authorities and community’s response supporters, social meeting lovers, massive meeting lovers, concerned about health, and concerned about public transport. (Balbontin et al., 2022) use these factors to explain the number of commuting and non-commuting weekly trips. In the exploratory phase of this study, these factors were included as explanatory variables in the mode choice model, and only the WFH lover and concerned about public transport factors were statistically significant. Therefore, two latent variables will be considered: (1) WFH lovers, and (2) Individuals concerned about using public transport (PT) to go to workplace due to COVID-19,. The linear structural equations of the latent variables are expressed as follows:

Where represents attribute j of respondent q (e.g., age, gender, income, occupation); represents attribute i of the home or work of respondent q (e.g., distance to work, travel time, has their own WFH space, mode used to go to work, location); and are the estimated parameters associated with each attribute. These represent the deterministic part of the linear structural equations, which allow for deterministic heterogeneity through the inclusion of socio-demographics and work/home characteristics. The random part (unobserved) of the linear structural equations is defined by, which are the error terms associated to the latent variable n; and is a part of the error term that takes into account serial correlation. The error terms and are normally distributed with a mean of 0 and a standard deviation equal to 1, but they defer in that the second one will also be included in the choice model explained below (representing serial correlation) and that is why both can be estimated simultaneously.

The measurement equations of the latent variables are linear additive, as follows:

where represents an indicator associated with the latent variable; are the parameters to be estimated; and the error term. The indicators are attitudinal questions asked in the survey, as shown in Table 4 and Table 5 3 . Given the high complexity in the estimation of hybrid choice models, a thorough process was carried out to analyse which attitudinal questions should be used as indicators for each latent variable. The attitudinal questions were chosen based on the results from the factor analysis explained earlier and a correlation analysis between the responses to the different attitudinal questions. If the responses to two attitudinal questions were highly correlated (correlation over 0.9) then only one of them was included. This was an iterative process where different simple hybrid choice models to ensure that the results were not statistically different when removing one attitudinal question as an indicator.

Table 4

Indicators associated with the latent variable WFH lovers.

Acronym	Question
WFHPrdM	How productive do you think you have been in the last week whilst working from home?*
BalPdUnP	I am able to find a balance between paid work and unpaid work (e.g., housework, yard work, childcare)**
ReqEqu	I still require equipment / technology to be able to complete work from home as well as I would like**
WFHlFlex	I would like to have more flexible starting and finishing times in the future**

*Scale: A lot less productive (1), A little less productive (2), About the same (3), A little more productive (4), A lot more productive (5).

**Scale: Strongly disagree or disagree (1), Somewhat disagree (2), Neither agree nor disagree (3), Somewhat agree (4), Agree or strongly agree (5).

Table 5

Indicators associated with the latent variable concerned about PT and workplace*.

Acronym	Question
ACvConc	Imagine you had to catch public transport tomorrow, what would be your level of concern about hygiene be?
ACvCoNUs	Imagine you had to catch public transport tomorrow, what would be your level of concern about the number of people using public transport?
WkEnvCnc	How concerned are you today about Covid-19 and work, given the environment that you normally work in (i.e., before Covid-19)?

*Scale: Not at all concerned (1), Slightly concerned (2), Somewhat concerned (3), Moderately concerned (4), Extremely concerned (5).

These indicators were measured on a Likert scale with 5 levels and for model estimation we define four parameters,. We assumed symmetry in the indicators, using two positive parameters as follows:

The probability of a given response is given by an ordered probit model (Greene and Hensher, 2010), where the observed responses to the attitudinal questions and the measurement equations are related as follows:

The latent variable that represents attitudes towards WFH will be included in the utility function of the WFH alternative, which can be expressed as follows:

The latent variable that represents concern towards the use of public transport is included in the commuting alternatives using a mode m-specific parameter estimate, 4 . The utility function to commute in mode m is given by:

The utility function of the no work alternative is expressed in equation (7):

It is important to note that the error term associated with the WFH alternative,, is different to the error term associated with the WFH latent variable structural equation,. Similarly, with the error term associated with the alternative to commute by mode m, , and the error term associated with the PT concern latent variable structural equation, . Respondents provided responses on the choice made each day of the week, and hence there are 7 choice sets per respondent. To recognise this, the error terms account for the panel structure of the data, i.e., varying across individuals but the same within individuals. The hybrid model was estimated simultaneously using the Apollo Software (Hess and Palma, 2019) and using a high-speed computer with 24 nodes.

Results

The final model includes the structural equations for the WFH and PT variables, as well as the choice model between the alternatives of no work, WFH, and commute by each mode. All the parameter estimates in the final model are statistically significant at a 90% confidence level5 , with the majority being so at the 95% confidence level.

Structural equations

The model results for the structural equations for the WFH lover and PT concern latent variables are presented in Table 6 and Table 7 , respectively. The results show that respondents between 25 and 40 years old are the ones that feel more positive towards WFH, followed by those older than 40 years. Our broader research on WFH (Beck and Hensher, 2021) suggest that younger employees are keener to return to the office for a number of reasons including social interaction and building networks for career progression. Respondents with a personal annual income above AUD$200,000 tend to be less positive towards WFH, as well as people that work as labourers. There are only a few respondents in the sample with an income level over AUD$200,000, and most of them are either managers or employers, who might be more inclined to attend the office perhaps preferring to manage people in a non-remote environment6 . If the respondent has their own space or room to WFH, they are more positive towards doing so. There are some location-specific dummy variables that were statistically significant in the WFH lover latent variable, which shows that respondents whose work is located in the Sunshine Coast are more positive towards WFH, followed by those who work in Brisbane, relative to the rest of the study locations (including the GSMA). The workplace attitudes towards WFH have a statistically significant influence on the attitude towards WFH, showing that people that are being directed to work from home or are given the choice to do so, are more positive towards WFH than those who are not.

Table 6

Structural equation model estimates for WFH lover latent variable.

Description	Mean	T-Value
Intercept	0.790	0.459
Personal income above AUD$200,000 (1,0)	−4.835	−2.002
Age between 25 and 40 years old (1,0)	2.324	2.114
Age older than 40 years old (1,0)	2.046	1.874
Has own space or room to work from home (1,0)	2.018	2.789
Occupation labourer (1,0)	−4.080	−1.981
Workplace located in Brisbane (1,0)	2.188	2.261
Workplace located in Sunshine Coast (1,0)	2.875	1.874
My workplace is directing me to work from home (1,0)	6.785	4.090
My workplace gives me the choice to work from home (1,0)	4.384	3.753

Table 7

Structural equation model estimates for PT concern latent variable.

Description	Mean	T-Value
Intercept	−0.704	−3.231
Occupation white collar (1,0)	0.542	3.670
Workplace located in CBD (1,0)	0.343	2.262
Workplace located in New South Wales (1,0)	0.435	3.771
Last week used car to go to work (1,0)	1.020	6.314
Last week used bicycle or walked to go to work (1,0)	0.667	2.924
Last time I used public transport I got on when I wanted (1,0)	−0.305	−2.457

The structural equation results representing the level of concern towards the use of public transport (PT) show that those respondents in white collar occupations tend to be more concerned about PT. One possible explanation for these results could be because white collar workers tended to use PT more pre-COVID-19 than blue collar workers7 . People that used the car to commute to work in the last week are the most concerned about the use of PT, somehow explaining the amount of PT users prior to COVID-19 who are now driving, followed by those that used active modes to go to work last week. Respondents that work in central business districts (CBD) areas tend to feel more concerned about the use of PT (because they are mainly office workers travelling in relatively high-density PT settings pre-COVID-19), followed by those who work in the GSMA in NSW, relative to the rest of the study locations (including SEQ in QLD). Results suggest that people that could board the bus/train/light rail without delay in waiting are less concerned about the use of PT than those that had to queue longer than prior to COVID-19, presumably linked to crowding and its associated transmission risk. An alternative way of interpreting these findings could be that time spent waiting could be a proxy for reliability, which impacts their concern towards using PT.

Choice model

The choice model parameter estimates are presented in Table 8 , together with the results of a stand-alone mixed logit model (MML). This MML model includes an error component in the no work and the commuting alternatives that takes into account the panel nature of the data, which is normally distributed with mean 0 and an estimated standard deviation, equivalent to the hybrid choice model. The overall goodness of fit of the choice model component of the hybrid model8 is statistically superior to the MML model, reinforcing the position that taking into account respondents’ underlying attitudes towards WFH and the concern towards the use of PT as latent variables significantly improves the statistical fit of the model and provides an improved understanding of individual preferences. The running time of the hybrid choice model is considerably higher than the MML, which is expected considering it is estimating two latent variables’ models and the MML model simultaneously.

Table 8

Choice model parameter estimates.

Description	Alternative	MML	Hybrid model
Alternative specific constant no work (base)	No Work	–	–
Alternative specific constant WFH	WFH	−2.24 (9.52)	−7.39 (8.97)
Alternative specific constant commute by car driver	Car driver	0.20 (1.57)	0.17 (1.21)
Alternative specific constant commute by car pax	Car pax	−1.37 (10.73)	−1.39 (10.29)
Alternative specific constant commute by taxi/rideshare	Taxi/Rideshare	−3.05 (8.96)	−3.14 (9.10)
Alternative specific constant commute by train	Train	−2.58 (8.19)	−1.73 (6.24)
Alternative specific constant commute by bus	Bus	−3.04 (9.67)	−2.29 (8.29)
Alternative specific constant commute by light rail	Light rail	−2.76 (6.11)	−1.89 (4.54)
Alternative specific constant commute by ferry	Ferry	−3.78 (5.33)	−2.90 (3.99)
Alternative specific constant commute walking	Walking	−0.17 (0.84)	−0.26 (1.27)
Alternative specific constant commute by bicycle	Bicycle	−1.17 (4.98)	−1.16 (4.82)
Alternative specific constant commute by motorcycle	Motorcycle	−1.21 (4.38)	−1.22 (4.31)
In-vehicle travel time (mins)	All modes	−0.003 (1.83)	−0.003 (1.93)
Travel time active modes (mins)	Walking and bicycle	−0.02 (5.46)	−0.02 (5.22)
Access, egress and waiting time (mins)	Train, Bus, Light Rail and Ferry	−0.01 (2.38)	−0.01 (2.30)
Cost (AUD$)	All modes except walking and bicycle	−0.02 (5.20)	−0.02 (4.58)
Female (1,0)	No Work	0.26 (2.96)	0.29 (2.90)
Personal income ('000$AUD)	WFH	0.00 (4.11)	0.01 (2.83)
Number of individuals per household	WFH	0.09 (2.15)	0.13 (1.73)
Monday (1,0)	WFH	3.20 (17.07)	4.39 (17.87)
Tuesday (1,0)	WFH	3.09 (16.58)	4.21 (17.42)
Wednesday (1,0)	WFH	2.95 (15.92)	4.01 (16.83)
Thursday (1,0)	WFH	2.95 (15.91)	3.99 (16.78)
Friday (1,0)	WFH	2.89 (15.67)	3.91 (16.57)
Workplace located in New South Wales (1,0)	WFH	0.07 (0.67)	0.79 (2.54)
Number of cars per person in household	Car driver	0.45 (3.49)	0.47 (3.32)
Latent variable PT concern	Train	–	−0.69 (5.47)
Latent variable PT concern	Bus	–	−0.54 (4.40)
Latent variable PT concern	Light rail	–	−0.69 (2.24)
Latent variable WFH lovers	WFH	–	0.37 (4.28)
Standard deviation error component ε_NoWork	No Work	2.53 (10.82)	−1.74 (8.69)
Standard deviation error component ε_Commute_m	Train, Bus, Light Rail and Ferry	−0.52 (8.86)	0.63 (9.58)
Sample size	650 respondents and 4,518 observations
Choice model
Number of parameters		27	31
Log-likelihood		−4521.05	−4,327.04
AIC/n		2.013	1.929
McFadden’s Pseudo-R2		0.323	0.352
Hybrid model (full)
Number of parameters		–	691
Log-likelihood		–	−8,537.28
AIC/n		–	3.810
Running time (minutes)		8	217

The total number of parameters in the hybrid model include the parameter estimates of the choice model (31 parameters), of the structural equations which are presented in Table 6 (10 parameters) and Table 7 (7 parameters), and the measurement equations parameters which are presented in Table 11 (21 parameters) in the Appendix.

Table 11

Parameter Estimates Hybrid Model Measurement Equations. Note: The role of the delta parameters in the measurement equations is explained in more detail in Equation (3), and of the alpha parameters in Equation (2). The names used to describe the indicators are defined in Table 4 and Table 5.

Description	Mean	T value
Alpha parameter ACvConc	2.645	4.34
Delta parameter 1 ACvConc	1.254	4.53
Delta parameter 2 ACvConc	3.121	4.56
Alpha parameter ACvCoNUs	2.340	5.56
Delta parameter 1 ACvCoNUs	1.041	5.52
Delta parameter 2 ACvCoNUs	2.967	5.82
Alpha parameter WkEnvCnc	0.445	12.92
Delta parameter 1 WkEnvCnc	0.354	13.41
Delta parameter 2 WkEnvCnc	0.994	20.56
Alpha parameter WFHPrdM	0.038	3.11
Delta parameter 1 WFHPrdM	0.539	13.21
Delta parameter 2 WFHPrdM	0.739	12.54
Alpha parameter BalPdUnP	0.091	3.59
Delta parameter 1 BalPdUnP	0.227	6.77
Delta parameter 2 BalPdUnP	0.671	12.06
Alpha parameter ReqEqu	0.016	2.24
Delta parameter 1 ReqEqu	0.187	7.41
Delta parameter 2 ReqEqu	0.468	11.88
Alpha parameter WFHlFlex	0.108	3.76
Delta parameter 1 WFHlFlex	0.323	7.37
Delta parameter 2 WFHlFlex	0.558	10.21

The results suggest that female respondents are more likely to not work any given day. If the respondent has a higher personal income, then they are more likely to WFH, same if they live in a household with more people, if they have more cars per person in their household or if their occupation is clerical and administration. Similarly, if they work in the GSMA of NSW they are more likely to WFH. This is reinforced by evidence in (Hensher et al., 2022) where SEQ displays a lower incidence of WFH than the GSMA. In terms of the days, respondents seem more likely to WFH on Mondays, followed by Tuesday, then Wednesday, Thursday and Friday, relative to weekends. Even though the parameter estimates for each day of the week are relatively similar, we decided to estimate them as specific since even minor changes in the probability to WFH on weekdays particularly are of interest for policymakers. As expected, the travel time, access, egress and waiting time and cost have a negative marginal utility influence in the choosing of a specific mode of transport, with separate parameter estimates for motorised and non-motorised (i.e., walk and bicycle) modes. In this study we are specifically interested in the effect of WFH versus commute or not work, and that is why most of the explanatory variables are included in the WFH alternative. We did test for different socio-demographics in the modal alternatives (all those presented in Table 1 except for the ones that refer to the number of days worked or WFH), but the number of cars in a household was the only one statistically significant in the car driver alternative. We also tested for some non-linear effects in travel time and costs (i.e., liner plus quadratic) but they were not statistically significant.

The WFH lover latent variable has a positive influence on the probability to WFH, while the level of concern towards public transport has a negative influence on the probability to commute by train, followed by light rail and bus, relative to the other modes of transport.

Given that we are interested in understanding the choice to WFH versus not, we are not as interested in the valuation of travel time vis a vis the commute. Additionally, previous studies have suggested that the value of travel time savings is in a state of flux and will continue to be until there is some equilibrium with regards to WFH and the factors that influence the probability of doing so (Hensher, Beck, & Balbontin, 2021). Therefore, the next section will analyse the sensitivity in the probability to WFH given variations in the explanatory variables.

Elasticities

The sensitivity of the probability to WFH or to commute to changes in the attributes can be analysed through the elasticities, which are the preferred indicators of the nature and extent of behavioural response of most interest. The elasticities are calculated as follows:

For each individual we use the measurement equations to estimate the value of the latent variables, and then use them to calculate the utility functions and probability of choosing each alternative (i.e., WFH, no work, and commute by different modes). The results are summarised in Fig. 7 ordered from highest to lowest elasticity (mean and standard errors are presented in Table 10 in the Appendix). In terms of WFH, results show that a person that has their own space to WFH, ceteris paribus, is 13.5% more likely to do so than a person that has to share a space or does not have such space at all. The workplace policy towards WFH is logically a very relevant influence, incorporated through the WFH loving latent variable. It suggests, ceteris paribus, that people directed to WFH are 18.0% more likely to do so, while respondents that are given the choice to WFH are 26.5% more likely to do so, relative to other respondents. Since this influence is through the WFH loving attitude, these results are showing that people that are given the choice to WFH have a more positive attitude towards WFH, followed by those that are being directed to WFH. This is an interesting finding. Given that respondents who are able to choose when to work from home and when not to, thus presumably having flexibility to work between home and the regular work environment, are more positive than those who are being directed to WFH and thus presumably must only work from home, we hypothesis that a balance between WFH and going to the office leads to better WFH experiences.

Fig. 7

Direct Mean Elasticity Effects Hybrid Choice Model.

Table 10

Elasticities Hybrid Choice Model.

Alternative probability	Variable	Mean	Std error
WFH	Has own space or room to work from home (1,0)	0.133	0.003
WFH	My workplace is directing me to work from home (1,0)	0.184	0.008
WFH	My workplace gives me the choice to work from home (1,0)	0.270	0.008
WFH	Female (1,0)	−0.073	0.001
WFH	Personal income ('000$AUD)	0.327	0.004
WFH	Number of individuals per household	0.265	0.003
WFH	Monday (1,0)	0.392	0.017
WFH	Tuesday (1,0)	0.385	0.017
WFH	Wednesday (1,0)	0.374	0.016
WFH	Thursday (1,0)	0.376	0.016
WFH	Friday (1,0)	0.374	0.016
WFH	Workplace located in New South Wales (1,0)	0.327	0.004
Car driver	Number of cars per person in household	−0.102	0.002
Car driver	In-vehicle travel time (mins)	−0.033	0.001
Car driver	Cost (AUD$)	−0.092	0.004
Taxi/rideshare	In-vehicle travel time (mins)	−0.026	0.001
Taxi/rideshare	Cost (AUD$)	−0.228	0.010
Train	Last time I used public transport I got on when I wanted (1,0)	0.123	0.002
Train	In-vehicle travel time (mins)	−0.050	0.001
Train	Cost (AUD$)	−0.042	0.001
Train	Access, egress and waiting time (mins)	−0.123	0.003
Bus	Last time I used public transport I got on when I wanted (1,0)	0.097	0.001
Bus	In-vehicle travel time (mins)	−0.070	0.002
Bus	Cost (AUD$)	−0.046	0.004
Bus	Access, egress and waiting time (mins)	−0.132	0.004
Light rail	Last time I used public transport I got on when I wanted (1,0)	0.124	0.002
Light rail	In-vehicle travel time (mins)	−0.006	0.000
Light rail	Cost (AUD$)	−0.005	0.000
Light rail	Access, egress and waiting time (mins)	−0.017	0.001
Ferry	In-vehicle travel time (mins)	−0.002	0.000
Ferry	Cost (AUD$)	−0.002	0.000
Ferry	Access, egress and waiting time (mins)	−0.008	0.001
Walk	Travel time active modes (mins)	−0.280	0.010
Bicycle	Travel time active modes (mins)	−0.150	0.009
Motorcycle	In-vehicle travel time (mins)	−0.005	0.000
Motorcycle	Cost (AUD$)	−0.006	0.001

Specifically for public transport, a person that does not have to queue when waiting for their bus, linked we suggest to exposure to others, ceteris paribus, is 9.6% more likely to use the bus than a person who has to wait longer than they did prior to COVID-19; and 12.2% more likely in the case of train and light rail. Mask wearing is mandated on-board PT but not at the stops and station, so the health concerns in the stops and stations might be higher than inside the vehicles. The higher influence associated with train and light rail probably could be related to the fact that if there are more people waiting, the vehicles will be more crowded – and considering the train and light rail have a higher capacity of compartments relative to buses, the biosecurity risk associated with them might be higher.

Probability of working from home simulated scenarios

Five scenarios were simulated to show the sensitivity of the probability to WFH variable due to variations in the explanatory variables. The observed versus estimated results for the hybrid model probability to WFH are presented in Fig. 8 . The observed probabilities to WFH are lower than the estimated ones using the hybrid model, and they decrease from Monday to Friday. That is, the probability to WFH on Monday is higher than on Tuesday, and so on. The probability to WFH on weekends is much lower which, as was presented in Fig. 2, is due to a high proportion of people not working on weekends. These probabilities are the base scenarios, and the simulations represent variations in these base scenarios due to changes in the explanatory variables, which are described in Table 9 with the results presented in Fig. 9 .

Fig. 8

Observed versus estimated (base scenario for simulation) probability to WFH.

Table 9

Simulated scenarios description.

Scenarios	Description
1	Everyone has their own space or room to work from home
2	Everyone is a blue-collar worker
3	Everyone has an income level above AUD$200,000 with a population average of AUD$201,000
4	The number of cars per person in the household increase by 1 on average
5	Travel time in all modes of transport increases by 50%

Fig. 9

Simulated scenarios results.

The first scenario represents a situation where everyone has their own space or room to work from home which currently is 39% (Table 1), which would generate an increase between 1.4% and 1.6% in the probability to WFH on weekdays and 0.4% on weekends.

Scenarios 2 to 4 represent subgroups of the population and their behaviour towards WFH. Scenario 2 represents a situation where everyone was a blue-collar worker (i.e., technicians and trades, machine operators/drivers and labourer) where the probability to WFH would decrease between 14.0% and 14.7% on weekdays and 2.3% on weekends.

If everyone in the sample would have a high income of above $200,000 with a population average of $201,000 (scenario 3) then this would represent a decrease in the probability to WFH between 9.5% and 10.7% on weekdays, and 2.1% on weekends. A high income of above $200,000 has a negative influence in the latent variable WFH lover (which has a positive impact in the probability to WFH), but a higher income has a positive influence in the probability to WFH. This shows that a higher income implies a higher probability to WFH, unless the income is very high (above $200,000), in which case individuals seem to be less positive towards WFH. As previously discussed, this might be due to the fact that only a few individuals in the sample have a very high income (above $200,000) and most of them are managers or employers, who in turn may still prefer to manage or supervise staff in a more traditional framework for reasons of trust and authority, but also that managing during the pandemic has its own set of challenges that may create greater burden (Teodorovicz et al., 2021).

Scenario 4 represents a situation where vehicle ownership per adult increased by 1 on average, which has a negative influence in the probability to WFH. The results suggest that if vehicle ownership per adult increased by 1, then there would be between a 1.7 to 1.8% decrease in the probability to WFH on weekdays and 0.3% on weekends.

The last scenario shows the situation where all travel times increase by 50%, which would imply an increase of 0.3% in the probability to WFH on weekdays and 0.1% on weekends. This last scenario is interesting as it shows that travel times are not the most important factors when deciding to WFH (as shown also in (Hensher et al., 2022), compared to having their own space to work from home, workplace attitudes towards WFH or even having to wait longer than prior to COVID-19 when waiting for the bus or train/light rail.

Policy implications and conclusion

The main focus of this study is to understand the influence of underlying attitudes towards COVID-19, namely in relation to WFH and the use of PT, in the probability to WFH, to commute or not to work for each day of the week. A hybrid choice model was developed and estimated, integrating a choice model with two latent variables; - WFH loving attitude and PT concerned attitude.

It is not surprising that a key indicator of working from home, is a positive attitude towards the experience. This positive attitude is highest among those 20–40 years of age who likely appreciate the flexibility that WFH brings, and those whose workspace at home is more conducive to doing so. In the longer term we may see further pressure on suburban real estate markets where homes are typically larger, or a rethink in the design of higher density homes to facilitate work from home. It is interesting that those on very high incomes are less positive towards WFH: these people may be managers who are not as positive towards managing people who are WFH, but equally they may also have sufficient income to overcome household constraints more easily, thus are not as appreciative of greater work flexibility. The simulations also show that socioeconomic characteristics such as occupation or income have a high influence on the probability to WFH, as compared to the travel times exhibited by the different modes or the number of cars per adult in the household. These are structural issues that are difficult for transport policy makers to directly influence.

However, from a transport policy perspective, it is key to note that respondents that used the car to go to work the week prior to the survey tend to be more concerned about PT, followed by respondents that used active modes, relative to other modes of transport. We see the implications of this result playing out in many jurisdictions outside of just the GSMA and SEQ, where car use has rebounded strongly and is above pre-pandemic levels in many instances. This should be an alarm for transport policy makers. If the attitudes towards public transport are not improved rapidly, the models suggest that the car will become even more dominant than ever. The elasticities’ results showed that people that did not have to queue were 12.2% more likely to use train and light rail, and 9.6% more likely to use the bus, compared to those that had to queue before getting on the bus or train. These results highlight the bio-security risk associated with COVID-19 where mask-wearing and social distance was enforced on public transport but not so waiting for public transport; and that an overall regular and reliable service (as proxied by shorter queuing times) creates a positive public transport experience even in the context of COVID-19. This result is also important as it suggests that maintaining rather than reducing services may be key to the recovery of public transport.

Another key result from the models, in particular the elasticities, is the ability for the respondent to have flexibility in deciding if they work from home increases their positive attitude towards WFH and therefore, the probability of doing so. Through the elasticities and simulations, we believe that respondents seem to prefer a balance in WFH, rather than being directed to do so, and that the WFH policies and workspace facilities linked to productivity are one of the most influential factors in the probability to WFH. These are structural issues that are difficult for transport policy makers to directly influence. It seems like a hybrid work model, as opposed to working entirely from the office or from home, will make people more productive and satisfied, obtaining the best of both worlds. To that end, the mechanism for transport policy makers is not only to think of the provision of transport, but how can investment in good technology be seen as a transport investment (i.e., in a new WFH future ICT like the NBN is now also a transport investment).

Overall, these findings have a great deal of plausibility, but what is especially relevant is that we have a series of behaviourally informative model outputs that can be used to provide useful advice into the policy debate on what WFH and resistance to using PT means in respect of changing levels of commuting activity by specific modes.

CRediT authorship contribution statement

Camila Balbontin: Conceptualization, Data curation. David A. Hensher: . Matthew J. Beck: Data curation.

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.