Telecommuting and Travel Behaviour: A Survey of White-Collar Employees in Adelaide, Australia
The forthcoming sections are dedicated to the presentation and discussion of the results. During the survey, respondents were prompted with multiple questions designed to produce a comparative analysis of their behaviours pre-COVID restrictions and as of the survey timeframe, which was August 2022. This latter period is referenced in the figures as “post-COVID” for consistency and because all COVID-19 related restrictions were removed.
3.1. Telecommuting Behaviour
The survey’s design incorporated a dual-question filter to gauge the potential for remote work among respondents. The initial question probed whether their current job included tasks that can be performed from home. A follow-up query was presented to those who responded negatively, asking if they could work from home under government or health mandates. This design aims to minimize underestimation or overestimation of remote work capabilities by prompting respondents to consider their telecommuting feasibility under extraordinary circumstances, akin to a lockdown.
Figure 2 shows that there is a consistent pattern among respondents regarding their telecommuting capability. Those who initially answered “No” to the ability to work from home but indicated “Yes” when faced with a government mandate scenario, largely did not report any frequency of working from home, aligning with their stated restrictions—either from their employer or the nature of their job, as shown in Figure 3. Figure 4 confirms that there was no indication of overestimation from those who answered “Yes”, as none of them reported a complete lack of telecommuting options or a “Never” frequency. Interestingly, casual telecommuting, defined as less than once a month, was noted among those with employer restrictions or a personal preference against working from home, indicating occasional remote work in special circumstances. These findings help delineate the boundaries of remote work practices within the surveyed population.
Additionally, it is important to point out that approximately 17% of individuals who indicated their jobs could be conducted from home were nonetheless restricted from telecommuting by their employers. This underscores the pivotal role that employer and managerial attitudes will have in shaping telecommuting practices in the post-pandemic world.
These statements in Figure 4 were further investigated in the regression analysis. Figure 5 illustrates the survey design and the four dependent variables that will be used in the analysis: contingency telecommuting, non-telecommuter by choice, non-telecommuter by policy, and regular telecommuter. The dependent variables were established as follow:
Capability: Can the job be done remotely?
Opportunity: Is the employee allowed by their employer to work remotely?
Adoption: Has the employee chosen to work remotely given the capability and opportunity?
As part of data pre-processing, the one-hot encoding technique was implemented on some of the categorical variables. This method transforms categorical variables into binary vectors, creating new columns for each category, with values of 0 or 1. Meanwhile, certain variables—specifically household size, number of children, and number of cars—were considered as discrete ordinal. These were represented under individual columns with values constituting whole numbers such as 1, 2, or 3, each representing the respective counts in each category.
A variance inflation factor (VIF) assessment was conducted and it highlighted issues of multicollinearity among several variables, prompting us to reevaluate their inclusion. Specifically, we discovered a high degree of multicollinearity between the “distance to work” variable and “travel time” and “distance to CBD”. This close relationship suggested that these variables contributed similar information regarding commuting patterns, which could distort the analysis. To maintain the integrity of our statistical model and ensure accurate interpretation of the results, we decided to exclude the “distance to work” variable from further analysis. Additionally, the variable “Ind_Administrative_and_Support_Services” was identified as having multicollinearity with “Ocu_Clerical_administrative” and was therefore removed. The analysis also determined that household size exhibited multicollinearity with the number of children in the household, leading to the exclusion of the household size variable from the analysis. The multinomial regression analysis uses a reference category to make result interpretation easier. The reference category serves as a baseline against which the effects of other categories are evaluated. The reference category, often the one with the lowest occurrence in the dataset, was deliberately omitted from the analysis. This exclusion is a standard statistical procedure that aids in averting multicollinearity, a situation where two or more variables in the model are highly correlated.
The final independent variables were selected using the random forest-recursive feature elimination algorithm (RF-RFE) [55] to avoid overfitting and to follow the common rule of thumb of 10–20 observations per independent variable [56]. RF-RFE is a variable selection technique in machine learning that enhances model performance by systematically removing less important features. It involves using random forest, training a model on all available features, and then evaluating the importance of each feature [57]. Features are ranked based on their importance, and the least important ones are pruned from the dataset. This process is repeated iteratively, where in each step one or more of the least significant features are removed and the model is retrained on the remaining features. The cycle continues until a predetermined number of features is reached.
The modelling results for each telecommuting choice are presented in Table 2, where the “B” column represents the estimated logistic regression coefficients for each feature and the “S.E” column shows the standard error of the coefficients (a smaller standard error would suggest that the estimate of the parameter is more precise and reliable). In Table 2, the “p-value” column represents the statistical significance of each coefficient (if the p-value is small (typically ≤0.05), it indicates strong evidence against the null hypothesis). Exp(B) is the odds ratio which indicates by how much the odds of the outcome (in this case, being able to work from home) increase for each unit increase in the predictor (independent variable), while holding other independent variables constant.
Several significant factors were identified for non-telecommuters by choice, based on a p-value of less than 0.05. This group represents individuals who have the option to telecommute but opt not to. For this category, living within 0–5 km of the central business district (CBD) emerged as a highly significant factor (B = 3.49, p = 0.01), with an odds ratio of 32.86 when compared to the reference category of contingency telecommuters. This suggests that individuals living very close to the CBD are significantly more likely not to choose telecommuting compared to those who telecommute under exceptional circumstances. Those who mainly travel by bus are more inclined to be non-telecommuters by choice (B = 1.73, p = 0.08, odds ratio = 5.65) in comparison to individuals who telecommute solely under exceptional circumstances. This trend indicates a possible preference or necessity for these individuals to be present at their place of work. Moreover, the tendency to choose bus travel might suggest that the bus system in Adelaide is seen as a comfortable and dependable mode of transportation. This perception could make commuting by bus a more attractive and feasible daily option, especially when compared to the reference category of tram users, for individuals in this group. Regarding occupational roles, individuals in managerial positions were found to be significantly more inclined to choose not to telecommute (B = 2.12, p = 0.01, odds ratio = 8.33). This suggests that managers may have job responsibilities or personal preferences that predispose them to working on-site. Additionally, the significant odds ratio implies that managers might possess more autonomy in their decision-making regarding their work location, facing fewer restrictions in choosing whether to work from home or the office. Income level appeared as a non-significant factor for non-telecommuters by choice, with higher-income individuals (>125k) slightly more likely to be non-telecommuters by choice (B = −0.48, p = 0.35, odds ratio = 0.62), but again this was not statistically significant.
In the same multivariate nominal logistic (MNL) model, distinctive factors were identified for individuals who are non-telecommuters due to employer restrictions. This category includes those who could potentially telecommute but are not permitted to do so by their employer. When looking at occupational categories, academic staff or students are particularly affected by employer restrictions on telecommuting. They are significantly more likely to be non-telecommuters due to employer policies (B = 2.41, p = 0.01, odds ratio = 11.18) compared to the contingency telecommuters. This could be due to the nature of academic work, which often requires on-site presence for activities such as teaching, laboratory work, or other in-person responsibilities. Moreover, managerial positions also show a significant effect (B = 0.65, p = 0.41, odds ratio = 1.91), suggesting that while managers are more likely to have the autonomy to work from home, they might also face company policies that limit this option. This is a lower odds ratio than seen in the non-telecommuter by choice category, implying that while managers have a higher likelihood of being non-telecommuters, the effect of employer policy is less pronounced in this occupational group.
Within the multivariate nominal logistic (MNL) framework, analysing the factors influencing regular telecommuters—those who telecommute at least once a week—yields intriguing insights. Regular telecommuters are distinguished from those who telecommute under exceptional circumstances only. One notable finding is the significant impact of living more than 20 km from the central business district (CBD). Individuals in this category are significantly more likely to be regular telecommuters (B = 1.13, p = 0.03, odds ratio = 3.08) compared to the reference group. This suggests that the greater the distance from the CBD, the higher the propensity to telecommute regularly, possibly due to longer commute times and the convenience that telecommuting offers. Occupational roles also play a pivotal role in regular telecommuting. Notably, those in professional scientific and technical services are significantly more likely to telecommute regularly (B = 1.65, p < 0.01, odds ratio = 5.19), underscoring the flexibility and possible work-from-home compatibility of such jobs. Additionally, managerial positions also show a high likelihood of regular telecommuting (B = 1.20, p = 0.05, odds ratio = 3.33), suggesting that managerial duties can often be fulfilled remotely, aligning with the trend towards more flexible working arrangements in such roles. Regarding income, individuals earning more than 125k are more likely to be regular telecommuters (B = 0.71, p = 0.06, odds ratio = 2.04), hinting at a correlation between higher income levels and the likelihood of telecommuting. This may reflect the nature of higher-paying jobs that often offer greater flexibility or the ability for these individuals to create a conducive environment for telecommuting at home.
In the context of contingency telecommuting, a significant finding emerges indicating a higher propensity for individuals in clerical/administrative roles to opt for contingency telecommuting over regular telecommuting (B = −1.21, p = 0.03, odds ratio = 0.29). The negative odds ratio of 0.29 suggests that those in clerical/administrative positions are approximately 3.45 times more likely to engage in contingency telecommuting compared to regular telecommuters. This interpretation stems from the negative sign of the odds ratio, indicating a reduced likelihood, leading to the calculation (1/0.29) resulting in a multiplier of approximately 3.45.
Figure 6 lists the top reasons for working from home for those who indicated a positive frequency of working from home. The reasons included personal choices such as needing to be home to do other chores, followed by having a pre-agreed schedule with the employer or manager. Adverse weather conditions were also a factor for some people in their daily choices. Interestingly, a significant percentage of respondents indicated that they work from home when they must do drop-off and pick-up trips. This suggested that flexible working arrangements, such as working from home, can help individuals balance work and personal responsibilities.
Figure 7 shows the change in frequency of telecommuting. About 30% of telecommuters had no experience of telecommuting before the pandemic. This suggests the pandemic may had introduced 30% of the sample to the working from home phenomenon for the first time. In addition, the percentage of workers who telecommuted 3–4 days each week increased from 6% before the pandemic to 26% after the pandemic, indicating that more workers shifted to full-time or near-full-time remote work after the pandemic. The most common telecommuting frequency before the pandemic was 1–2 days each week (21% of respondents), which could suggest that this option was popular among workers who wanted some flexibility and balance between home and office work. Interestingly, the most preferred telecommuting frequency after the pandemic was also 1–2 days each week (but a much larger percentage, around 41% of respondents), which could indicate this option remained attractive for workers who could not or did not want to telecommute daily.
Figure 8 shows that post-COVID telecommuting was more frequent among higher-income groups than lower-income groups. For example, 66% of people who earned more than AUD 100k telecommuted 3 to 4 days each week, while only 2% of people who earned AUD 40k or less telecommuted. Higher-income workers may have more autonomy, motivation, and resources to work from home effectively, while lower-income workers may face more challenges, such as lack of space, equipment, or support. Figure 9 also shows that post-COVID telecommuting is less frequent among the lowest and highest income groups than the middle-income groups. For example, only 22% of people who earned more than AUD 100k telecommuted less than once a month, while 40% of people who earned AUD 60k–80k telecommuted. This may be attributed to the fact that the lowest and highest-income workers have more extreme preferences or constraints regarding telecommuting, while middle-income workers have more balanced or mixed views. Income by telecommuting frequency.
Figure 9 displays the relationship between income levels and telecommuting frequency. Notably, there is a distinct pattern for individuals working from home five days a week: their proportion increases with increasing income. Conversely, for those who telecommute less than once a month, their proportion decreases as income goes up. For other telecommuting frequencies, no consistent trend is evident.
