1. Questions

The spread of COVID-19 posed significant health risks in the US and across the world. To reduce this spread, companies and state authorities adopted various preventive measures early in the pandemic. For instance, many US state governments implemented stay-at-home orders, and many companies allowed employees to work remotely. Regardless of the effectiveness of these preventive measures, they are likely to exert disproportionate effects on individuals with different sociodemographic characteristics. A review of the literature reveals that many studies (Askitas, Tatsiramos, and Verheyden 2020; Bonaccorsi et al. 2020; Jiao and Azimian 2021; Martin et al. 2020; Molloy et al. 2020; Ozili 2020; Saha, Barman, and Chouhan 2020) have assessed the socioeconomic effects of the pandemic over the past year. However, few ones (Kazekami 2020; Shi et al. 2020) have examined the potential relationship between socioeconomic characteristics and telework status during the pandemic. Such information is vital to navigating the current critical situation, as it can inform planners and decision makers about how to effectively support neglected groups.

2. Methods

Data were collected through the Household Pulse Survey (US Census, 2020), which was conducted by the US Census Bureau to measure the social and economic effects of the pandemic. The bureau randomly selected a limited number of addresses from across the US to represent the entire population, and these residents received a link to the online survey via email, text, or both. The questionnaire was meant to collect information on respondents’ education, demographic characteristics, and transportation use. The survey was conducted in three phases. Phase 1 ran from April 23 to July 21, 2020; phase 2 from August 19 to October 26, 2020; and phase 3 from October 28 to March 1, 2021. Each phase yielded multiple datasets, each of which contains information collected biweekly from more than 70,000 participants. In this study, we aimed to utilize the latest datasets for each phase, as they capture the latest changes in individuals’ travel behaviors for the corresponding phase. However, as the phase 1 questionnaire did not collect information on individuals’ telework status, we instead used data from the earliest phase 2 dataset (104,024 participants from August 19–31). Additionally, we used the latest datasets from phase 2 (85,350 participants from October 14–26) and phase 3 (73,858 participants February 17–March 1) in the analysis. Table 1 provides the summary statistics of the variables used Models 1, 2, and 3.

Given the categorical nature of our outcome variable (telework status) and the presence of state-level unobserved heterogeneity, we fitted three mixed logit models (with random effect terms) to each of the three datasets. Generally, developing a model with random effect terms is beneficial, as random effect terms can capture the impacts of unmeasured factors (Azimian et al. 2021; Washington, Karlaftis, and Mannering 2010). An example of such a factor in this study is the effect of state policies on travel behaviors.

3. Findings

As shown in Table 2, except for ethnicity, which was only significant in Model 1, all other variables were found to be significant in all three models. Additionally, the direction and magnitude of the relationships are consistent. The results show that the age variable, grouped as 35–50, 50–65, and ≥65, was negatively associated (significant at the .05 level), implying that adults 35 years or older were less likely to work from home than those younger than 35. Men also exhibited a lower likelihood of working from home than women. Similarly, individuals without graduate degrees were less likely to telework during the pandemic than those with such credentials. As for marital status, while being married was found to be insignificant in the models, reporting an “other marital status (divorced, separated, widowed)” was found to be significantly negatively related, suggesting that such individuals were less likely to work from home than those who had never been married. Individuals who reported some degree of financial difficulty were also less likely to telework compared with those who didn’t. In addition, the likelihood of working from home was lower for those in households with one or two people than for those in larger households.

With regard to work type, those employed by the government or the private sector were more likely than self-employed individuals to engage in telework. Individuals earning an annual income less than $100,000 were less likely than those earning more to work from home. Lastly, individuals who experienced anxiety were more inclined to work from home than those who didn’t.

The most important component for drawing implications is the random effect term, which captures the variation across states. Figures 1, 2, and 3 show the estimated random effects across the US for early phase 2, late phase 2, and late phase 3, respectively. The states highlighted in orange and red have positive values, implying that unobserved factors in those states likely incline residents to work from home. The situation is reversed in light and dark green states, which have negative values. In addition, comparing Figures 1 and 2 reveals that between early and late phase 2, the direction of the random effect values for North Carolina and Ohio changed from negative to positive, while that for Maine changed from positive to negative. Lastly, between late phase 2 and late phase 3, the random effect terms for four states (Alaska, Maine, Wisconsin, and New Mexico) changed from negative to positive, which could have resulted from state and company policy changes in those states.

Figure 1.Random Effect Estimates from Model 1 (Early Phase 2)

Figure 2.Random Effect Estimates from Model 2 (Late Phase 2)

Figure 3.Random Effect Estimates from Model 3 (Late Phase 3)