Slowly coming out of COVID-19 restrictions in Australia: Implications for working from home and commuting trips by car and public transport

4.1. Days worked and work from home

The impact of COVID-19 restrictions on the availability of work and where work is completed has been profound (see Fig. 2). Among those respondents who were working prior to the COVID-19 outbreak, after Wave 1 the number who worked 5 days per week fell from 58% to 39%, with a marginal improvement to 41% in Wave 2. Similarly, among those who worked at least one day before the pandemic, 26% found themselves without employment during the Wave 1 data collection period, though perhaps showing some form of recovery, that number reduced to 17% as of Wave 2.

While overall employment (measured in days worked) has contracted, we have seen a growth in the number of days people are working from home (see Fig. 3). Prior to COVID-19, 71% of respondents in employment, did not engage in any work from home. However, at the time of Wave 1 data collection, the number not working from home dropped to 39%, with those working 5 days at home rising from 7% to 30%. In the most recent data collected in Wave 2, however, we started to see the beginning of a return to the long term trend, with just over half the sample (54%) working no days from home, and approximately one in five (21%) working 5 days a week from home. With respect to number of days worked from home across the three time periods, prior to COVID-19 the overall average was 0.86 days per week, during Wave 1 the average rose to 2.4 days, and in Wave 2 this average fell to 1.7 days.

Fig. 4 shows the policy of the workplace with respect to work from home arrangements. Although the composition of the two samples is different (new respondents were contacted to supplement the sample of respondents who participated in Wave 1), we see a changing mix of workplace policies, with more workplaces having closed, and conversely less respondents being directed or given the choice to work from home.

4.2. Attitudes towards working from home

In Wave 2 of the survey, respondents were asked a number of attitudinal questions in order to gain insight into their experiences working from home. Fig. 5 shows the level of agreement (1 = Strongly Disagree to 7 = Strongly Agree) to five attitudinal statements. There are significant levels of agreement across all statements, with respondents finding the WFH experience to be largely positive, that they have an appropriate space from which work can be completed, and importantly, they would like to work from home more often in the future. While agreement is significant, it overall it is not extreme; however it should be noted that for many respondents, arrangements for WFH were initially haphazard as COVID-19 suddenly forced it upon many, while at the same time schools were closed. With more time to prepare for WFH and with less home-based distractions in the future, the overall experience may become more positive as we move forward, something this research intends to monitor. Respondents were also asked their state their level of agreement (1 = Strongly Disagree to 5 = Strongly Agree) with selected attitudes around COVID-19 and various responses or changes to behaviour (See Fig. 6.)

Respondents were also asked to assess their level of productivity at home relative to at work, with the result displayed in Fig. 7 . Overall, respondents rate their productivity as more or less the same while WFH as it would be when completing the same tasks in their normal work environment.

4.3. Working from home in the future

Given that working from home has been a largely positive experience, wherein the majority of respondents feel that they are at least as productive at home as they are at work, it is not surprising that overall, 71% of respondents agree with the statement that they would like to WFH more often. To gain insight into what level of WFH might persist into the future, respondents were asked how many days they would like to WFH if they could, as COVID-19 restrictions were eased. Fig. 8 displays the preferred number of days WFH. A closer analysis of the data showed older respondents (55 or more) wish to work less days from home on average, compared to younger age groups. Compared to the reported levels of WFH prior to COVID-19, it would seem that as restrictions are eased, WFH will constitute a greater proportion of working days than before.

Table 2 examines working from home as a proportion of days worked (i.e., days worked from home divided by total days worked) tabulated by quintiles. While we have seen a drift back towards the pre-COVID-19 figures with regards to the number of days worked from home, it was understood that the extremes currently seen are not likely to be sustainable. However, there might be an equilibrium that lies somewhere between the two experiences of WFH before COVID-19 and WFH during the initial height of the pandemic as measured in Wave 1. In terms of the future ideal, we can see a retraction away from the poles (0% or 100% WFH) towards a middle ground, with that middle ground being a sizeable increase in the level of WFH as a proportion of total work.

Looking at the current level of WFH as of Wave 2, overall, 52% of those currently working want to maintain the current level of WFH, but if you exclude those who currently do not work from home, then 16% of people currently WFH want to stay at the level they are currently at, 25% want to WFH more in the future than they do now, and 30% wish to WFH less in the future than they do now (9% wanting to go back to no WFH, but 21% wanting to reduce the amount of WFH, but not completely). Lastly, there is a significant and positive correlation between the proportion of days working from home currently, and the proportion of days someone would like to work in the future as COVID-19 restrictions are eased.

4.4. Overview of commuting trips

The suppression of travel activity and the increase in working from home has had a significant impact on the commuting behaviour of respondents. Fig. 9 shows the average number of one way commuting trips across the two samples, from before COVID-19 restrictions, through each of the survey waves, along with the number of commuting trips respondents are planning for the week following the Wave 2 data collection period. We can see a significant fall in commuting trips from before COVID-19 to Wave 1, but the Wave 2 results indicate that commuting trips are trending up. It should be noted, however, that the error bars on the graph display the standard deviation around the average, and indicates a very high degree of variability in behaviours, particularly in Wave 2, and even more so in the planned number of trips moving forward. This is further indication of the importance of regular data collection, analysis and modelling given the level of flux that currently exists. This diagram is a very powerful indicator of some return back to the office, but with a significant potential residual of WFH days.

4.5. Summary of descriptive analysis

We have seen a great shock to travel and work behaviours as a result of COVID-19 and associated restrictions. The aggregate analysis shows that while the impact persists, there is preliminary evidence that as restrictions are eased, behaviour will regress towards the pre-COVID-19 state, however it is clear that with respect to work from home, most respondents would like to continue to engage in this style of work at a level greater than before COVID-19. As such, developing model systems to understand the degree to which WFH is adopted, and the impact of WFH on commuting trips will be important to transport planners and authorities as it is clear that increasing working from home will need to be incorporated into strategic transport models and future transport forecasts. Section 3 proposes some key models that need to be integrated into strategic models together with a review and possible re-estimation of other models such as commuter and non-commuter mode choice and time of day models to reflect the changing travel setting.5

6.1. The ordered logit model for the incidence of working from home

The final ordered logit models6 for WFH are summarised in Table 4, with an overview of the variables in the model provided in Table 3 . In selecting and testing candidate explanatory variables, we wanted to identify influences on WFH that relate to an employee's situation where they could choose to WFH or otherwise, with the position supported or enforced by their employer, under government restrictions in the early days of the COVID-19 lockdown as well as when restrictions began to be relaxed.

In developing behaviourally rich models to represent the extreme lockdown in the latter half of March (as captured in the Wave 1 data), and the late May Wave 2 context of partial relaxation of restrictions, we recognised that the key drivers of WFH and commuting activity between these two time periods are likely to be very different. Specifically, in late March the decision to WFH and cease commuting was largely driven by mandated government directives, but with the great majority of employees and employers supporting WFH unless it is was not a feasible option. Apart from employer policies which included employees having a choice to work from home pre-COVID-19, we anticipated that employee occupation and income may have a role in determining the extent of WFH. We also expected that in late March the shock to the system was still being digested by workers with very limited knowledge of whether WFH would work out, and what strategies governments were putting in place to minimise the risk of exposure to the virus.

As time moved forward between late March and late May, questions in the survey associated with an accumulated experience in WFH and gaining an understanding of the role that government played, started to take on real meaning as people crystallised their views now that they are better informed, and indeed are reflected in the statistical significance of a number of the attitudinal questions, in contrast to late March (Wave 1) where they had no behavioural relevance.

Both Wave 1 and Wave 2 models have overall goodness of fits within the range for pseudo-R² typically obtained (i.e., 0.2 to 0.4) (Hensher et al., 2015). Both waves have identified three statistically significant employer-policy dummy variables, namely an employee having a choice to work from home pre-COVID-19, an employer directs the employee to work from home during COVID-19, and the type of work undertaken by the employee cannot be completed from home. The first two of the dummy variables have positive parameter estimates, suggesting that the probability of having more compared to less days of WFH increases when each of these policy settings are on offer; in contrast when the work cannot be completed at home, the probability of working away from home increases.

We considered all available socioeconomic characteristics (i.e., occupation, age, gender and household income), and found for Wave 1 that only one occupation class (i.e., technician and trades) based on the Australian Bureaus of Statistics 8-category Classification,7 and household income, were statistically significant indicators for establishing the extent to which WFH occurred. Technician and trade occupations tend to reduce the probability of WFH which makes good sense since such jobs typically include electricians, plumbers and builders. Household income has a positive parameter estimate, suggesting that as household income increases, we can expect a higher probability of being able to, and preferring to, WFH. The residential location of respondents was investigated, and we found that in Wave 1, people living in the three largest metropolitan areas in Australia (Sydney, Melbourne and Brisbane), tended to have a higher probability of WFH, which seems plausible given the mix of occupations compared to regional, rural and small city environments.

Wave 2 included 11 opinion and attitudinal variables together with three occupation classes that were not statistically significant in Wave 1. The three occupation classes (community and personal services, clerical and administration, and sales) were statistically significant while household income was not, suggesting that these occupation classes have a lower probability of working from home as restrictions start to relax and possibly as they are required to spend some amount of time back in the office.8 We also found that respondents residing in all capital cities had a higher probability of WFH.

The attitudinal variables are all coded as dummy variables relative to a ‘neutral’ opinion, as such the expectation on the sign is that it could be in either direction as these dummy variables now measure agreement or disagreement (rather than some level of agreement). For all attitudinal variables, it is important to emphasise that, in late May, we were still dealing with constrained choices in that people were being directed to work from home, and hence this effect as captured through the employer policy variables was having a significant and dominating influence on the extent of WFH (typically 40% of the sample were WFH 5 days a week) which in itself could prime views about working from home.

We find that productivity when WFH, associated with positive parameters, can be rationalised as follows: regardless of whether an individual believes that they are more or less productive when they WFH, they still prefer to do so, but the parameter estimate is significantly greater when productivity at home is perceived to be higher. The appropriateness of workspace at home also has positive parameters, suggesting that the less concern you have about your workspace, the more likely you are to work from home versus those that have a higher level of concern, but still a preference to WFH. On the positive experience associated with WFH, the positive parameter estimates follow a similar pattern; the more you work from home, probably the less (but still) positive you find it, given anecdotal evidence that a growing number of individuals are wanting some amount of social interaction in the workplace that is not possible online through videoconferencing and/or phone meetings. Indeed, the response to the desire to go to work from time to time, as negative parameters, reinforces the position of some likely increase in the probability of not working from home. Finally, the estimated parameter for trust in government responding to COVID-19 is negative, suggesting that as of late May, for those individuals who agreed with this statement, the greater the probability that they will be prepared to reduce the number of days working from home. This last point is powerful evidence of the importance of messaging by Government.

Although the parameter estimates are statistically significant, they are not behaviourally very interesting; instead care must be taken in interpreting the numerical magnitude of each parameter estimate since they are non-comparable in this non-linear logit form (Hensher et al., 2015). In Table 5 we present elasticities as a way of meaningfully comparing the influence of each explanatory variable on WFH. For the logit form, the elasticity of the probability is given in eq. (10) (Greene and Hensher, 2010).

where the marginal effect is given in eq. (11).

A marginal effect for continuous variables is the influence a one unit change in an explanatory variable has on the probability of selecting a particular outcome, ceteris paribus. For dummy (1,0) variables, which are the main variables in the models, the marginal effects are the derivatives of the probabilities given a change in the level of the dummy variable. The marginal effects need not have the same sign as the model parameters. Hence, the statistical significance of an estimated parameter does not imply the same significance for the marginal effect (see eq. 12). Neither the sign nor the magnitude of $\overline{\beta }$ need bear any relationship to those of β_j. β_j, equal to ∂log(P_j/P₀)/∂x, is commonly defined as an interpretation of the model parameters.

Looking first at the variables related to employer policy, the mean elasticity estimates suggest that the employer directive to WFH (associated with a government mandated restriction) has the greatest behavioural response, notably for WFH 4 and 5 days a week for both waves, and is noticeably much greater in Wave 1 (late March), as expected. There is evidence of the softening of the response in late May as we see restrictions relaxed, and possibly a greater appreciation of the risks of moving to some amount of not working from home.9 The negative elasticity estimated for WFH = 0 compared to the positive elasticity estimates for WFH 5 days a week is consistent with the view that having a choice to work from home or an employer directive to WFH, reduces the probability of going to work 5 days a week and increases the probability of WFH 5 days a week, with the latter a greater behavioural response.

A similar logic can be applied to all of the direct elasticity estimates. The household income elasticity estimates suggest that as household income increases, the probability of WFH 5 days under Wave 1 increases and decreases for WFH zero days per week. The elasticity estimates are generally greater and negative for the three employment policy variables in Wave 1 when there is WFH, and decline into the positive range for situations where there is commuting five or more days per week.

For the attitudinal variables associated with Wave 2 that have a positive parameter estimate in Table 4, we see that as we move from WFH 5 days a week to no days per week, that the behavioural responsiveness declines from a positive estimate for WFH equal to 2 to 5 days, to a negative estimate for 0 to 1 days. This is plausible and aligned with the interpretation of the Wave 2 model in Table 4 where the relaxation of restrictions is seeing a greater percentage change in the probability of WFH more days than less days over 2 to 5 days a week. The behavioural sensitivity associated with each explanatory variables is presented in more detail in a later section where we assess various scenarios.

In linking the WFH model to the modal trip frequency models, we have to calculate the probability of choosing a number of days WFH. The mean probability of each WFH level for Waves 2 and 1 (the latter in brackets) are 0 days: 0.456 (0.381), 1 day: 0.018 (0.063), 2 days: 0.068 (0.075), 3 days: 0.058 (0.089), 4 days: 0.047 (0.046), and 5 days or more: 0.29 (0.346). Given the estimated parameters obtained from the WFH ordered logit model, we can calculate the probabilities associated with each of the number of days working from home and enter them as explanatory variables into the Poisson regression model. These probabilities are obtained for each respondent using the following formulae, with an example given for four ordered alternatives.

Define U_fit = b(1) + b(2)*x1 + b(3)*x₂ as the utility expression for a constant and two explanatory variables;

f₀ = exp.(-U_fit)/(1 + exp.(-U_fit));

f₁ = exp.(μ₁- U_fit)/ (1 + exp.(μ₁- U_fit));

f₂ = exp.(μ₂- U_fit)/(1 + exp.(μ₂- U_fit));

p₀ = f₀; p1 = f₁-f₀; p2 = f₂-f₁; p3 = 1-f_{2; and}.

p_model = (y = 0)*p0 + (y = 1)*p1 + (y = 2)*p2 + (y = 3)*p3, where p = the choice probability for that level.

6.2. The Poisson regression model results for commuting activity

Turning to the Poisson regression model (in Table 7) with the number of weekly one-way modal trips defined as an integer for the Poisson count model, the overall goodness of fit (as pseudo R²) of all models are excellent for a non-linear model, varying from 0.369 to 0.681. A descriptive profile of the explanatory variables that are statistically significant is given in Table 6 . The sigma parameter that is estimated to allow for latent normal heterogeneity was statistically significant in all four models at the 1% level. Tau, the ZIP parameter is are statistically significant in Wave 2 but not in Wave 1. The Vuong statistics suggest that the estimated extended Poisson models in Table 7 for Wave 1 are favoured over an unaltered Poisson model. The Vuong statistics of 22.45 for car and 6.45 for public transport for Wave 2 also suggest that the estimated extended Poisson model is favoured over an unaltered Poisson model, but with censoring using Probit. That is, the dependent variable is over-dispersed and has an excessive number of zeros.10

Annual household income was found to be statistically significant for car and public transport trips in both Wave 1 and Wave 2. The positive sign for the car models is indicative of a greater number of weekly one-way trips for higher income households; in contrast the negative sign in the public transport models suggests fewer public transport trips for higher income households. The only other socioeconomic influences that were statistically significant are associated with the professional occupation (relative to all other occupations)11 and males for Wave 1 public transport, both positive; and professionals for Wave 2 use of both car and public transport, being also positive for both car and public transport.

Residential location was statistically significant in Wave 2, with all urban locations being statistically significant at the same level for car, but a separation of the main metropolitan locations from other capital cities for public transport produced only a marginal difference. The positive parameter estimate in the car model suggests more trips per week in all urban locations compared to regional and rural contexts. This result, may in part be attributable to generally more activity in urban areas but also because of less replacement of car travel with active transport modes. For public transport, the statistical significance was found to exist for separate variables for the three largest capital cities and other capital cities. The difference seems plausible given the greater availability of public transport in the larger cities, but also relatively good public transport in other capital cities compared to rural and regional locations. The public transport distinction compared to car, with a single parameter estimate for urban locations, is suggestive of the lack of influence of public transport on many car user preferences.

The probability of working from home no days a week was statistically significant in both waves for those using car and public transport. In both waves, it was positive for car use suggesting that as expected, all other influences remaining unchanged, as the probability of going to work 5 days a week increases, the number of one-way car commuter trips increases. For car in Wave 2, we also see a higher probability of going to work 4 days a week adding to the total switch back to the workplace. Under COVID-19 with biosecurity a concern, this makes good sense. However, for public transport, we observe a negative sign in Wave 1 which makes sense given that this was at the height of restrictions with government advice to not use public transport; but this changed to a positive sign in Wave 2 when a return, to some extent, to work resulted in an increase in public transport commuting trips, with social distancing in public transport requiring sitting only on seats with an allowable sitting sign.12 We might anticipate a continuing increase in public transport in ensuing periods as additional capacity is released and the growing evidence of PT being a relatively safe environment increases together with more positive government policy messaging. Finally, for wave 2 only, we obtained a negative sign associated with car use for the question on whether an individual sees a health risk personally to themselves under COVID-19. Although only marginally only statistically significant, this supports the position of an individual whose ranking of health risk is high, so undertaking fewer one-way commuter trips per week for car, translated into reduced commuting.

Given that the dependent variable is a count variable, and Poisson regression models the log of the expected number of weekly trips as a function of the predictor variables, we can interpret an estimated parameter as follows (Wooldridge, 2002): for a one unit change in the predictor variable, the difference in the logs of expected counts is expected to change by the respective parameter, given the other predictor variables in the model are held constant. For a binary variable such as gender, the difference in the logs of the expected number of weekly public transport trips in late March (Wave 1) is expected to be 1.231 higher for males compared to females, ceteris paribus. For a continuous explanatory variable such as the probability of WFH one day a week in late May (Wave 2), if a commuter were to increase the probability of WFH 1 day a week from say 0.1 to 0.2, the difference in the logs of expected number of weekly car trips would be expected to increase by 1.404 trips, ceteris paribus.

A more informed way of illustrating the behavioural response for one way weekly commuting trips associated with changes in the probability of WFH and other influences is to undertake a number of scenario applications, which we now present and discuss in the following section.

7.1. Scenario one: the impact of workforce directed to WFH

In this and the following four sections, we have focussed only on Wave 2 primarily because we want to speculate, through simulation, what the next period, under increasingly relaxed restrictions, might look like, and that the best data we have to make such predictions is Wave 2. The further justification is that Wave 1 is a very unusual case of immensely suppressed travel regardless of situation, attitude, work, etc. It is a baseline of “nothing” with few freedoms to vary anything. The simulation using Wave 2 produces more differentiated results compared to Wave 1, given the enhanced variability in the data.

Let us assume that all work can be performed from home instead of the sample average of 78% for Wave 2, which means all work theoretically can be completed by WFH if people were directed to do so. We simulated a scenario wherein 20% and up to 80% of the workforce were directed to work from home, to imitate the different phases of the lockdown or social distancing requirements at the workplace. In this scenario, the average number of WFH days would increase from 2.04 for 20% to 4.02 for 80%, a noticeable increase. The results are shown in Fig. 11 .

7.2. Scenario two: the impact of attitudes and experience on WFH

Using the evidence on individual's attitudes towards various aspects of the COVID-19 lockdown, and coming out of the initial height of the pandemic in late May as some restrictions are relaxed, we simulated changes in the response to the relevant scale used to measure attitudes: relative productivity; appropriateness of home working space; how positive WFH has been experienced, willingness to WFH more in the future. We varied the level of agreement for these statements from 20% to 80% to assess the impact on the average days WFH, as shown in Fig. 12 .

We see that the most influential perspective regarding WFH choice is having adequate work space at home. The more a respondent agrees to this statement, the more likely they are willing to spend more days WFH. Conversely, low levels of agreement result in the lowest uptake of WFH days (1.6 days compared to all other statements if there was only 20% of agreement). Having a positive experience with WFH on the other hand, has a relatively small role in increasing the average WFH days compared to having adequate space and being productive. This may be a result of the fact that for many, WFH is currently a necessity rather than a choice, hence personal preference and experience play a relatively more minor role. Alternatively, it might also be the case that very few people disagreed with the statement (15%), and such limited ill-will towards the experience has meant that there is insufficient variability in negative experience to estimate a sizeable impact resulting from a bad experience at this stage. If individuals expressed a desire to want to go to work from time to time to avoid social isolation, the average number of WFH days would decline from 2.3 days at the 20% agreement level to 1.8 days at the 80% agreement level.

7.3. Scenario three: the impact of WFH on the number of one-way commuting car trips

Using the estimated ordered logit model, we predicted the probabilities that people would work from home zero days per week and what this might mean for the number of predicted weekly one-way car commuting trips. We assume that the average number of WFH days would vary and that its composition may also change over time with the tightening or easing of restrictions. We simulated a scenario with 20% to 80% of people not working from home at all (i.e., WFH 0 days).There is a noticeable increase of one-way weekly commuting trips from 3 trips to 7.8 trips when the proportion of workforce not working from home increases from 20% to 80%, as shown in Fig. 13 .

The responsiveness of car travel in Wave 2 is likely to be a function of the easing of restrictions that occurred between Wave 1 and Wave 2, giving respondents more flexibility to vary their travel behaviour, within the parameters of the restrictions. Another explanation for the faster uptake in car travel for commuting is that in Wave 2, as more people start to travel, the private car becomes a preferred alternative during a pandemic for reasons of hygiene and the concern thereof related to public transport. For some individuals who used to take public transport pre-COVID-19, the choice would shift to commuting by car for this period, with an ongoing concern as to whether they will return to public transport at a later date.

7.4. Scenario four: the impact of WFH on the number of on-way commuting public transport trips

For Wave 2, the ZIP models for PT trips predicted a low level of commuting trips by public transport, at 0.89 predicted trips on average. Commuting by public transport during a pandemic lockdown period is not a preferred (or indeed desirable) choice, aided by government messaging to stay away from using public transport. A similar scenario as in Section 5.3 for car, was investigated for PT trips. The increase of PT trips is much slower than the increase for car trips when the proportion of the workforce not working from home increases from 20% to 80% (Fig. 14 ). This shows at least based on the answers taken during Wave 2 that people had no confidence to take more PT trips even if the restrictions were eased and more people were required to not work from home.

This shows that regardless of where people are and whether they work more or less from home, the commuting choice using public transport is not generally preferred for now (up to late May), despite the slow return.

7.5. Different occupations and WFH and commuting trips

We identified the influence of specific occupation categories on the probability of working for home and also on the number of one-way weekly trips by car and public transport. During Wave 2 when, on average, 2.19 days people worked from home, in the ordered logit model we found that three occupation groups went back to the workplace more often than others, and hence worked less from home. They are community and personal services workers (WFH 1.80 days), clerical and administration staff (WFH 1.86 days), and sales workers (WFH 1.97 days). Many of their duties require interacting with others and hence are more difficult to do away from the work off-site.

Occupation not only directly affects WFH choices, it also directly influences the quantum of one-way weekly commuting trips. In the Wave 2 model, for car trips, professionals were predicted to have a higher level of one-way weekly commuting trips at 5.13 trips (approximately two one-way trips for two days) compared to 4.54 trips on average. They also commute more by public transport, with predicted 1.19 PT trips compared to 0.89 PT trips on average for the overall sample.