COVID-19 was first reported in China and then spread worldwide and has caused 22.3 million confirmed cases and more than 373,000 deaths in the United States as on January 11, 2021 [1]. To help prevent the virus from spreading and to salvage the economy, companies and schools nationwide have started work and study from home, respectively. According to a Gartner survey of 880 global human resources executives on March 17, 2020, almost 88% organizations have encouraged or required employees to work from home (WFH) [2]. Barrero et al [3] found that WFH might persist even after the pandemic. Concerns may arise regarding productivity [4], willingness [5], and future trends [3,6] regarding work and study from home.

WFH has been a controversial issue that merits closer scrutiny. Palumbo [5] reported that WFH might incur side-effects such as a negative impact on work-life balance. This would lead to negative opinions on WFH when people tweet about it. Other studies have focused on specific categories. A survey of employees in Lithuania [7] reported that female employees appreciate WFH more than male employees because female employees can enjoy a healthier lifestyle, while male employees worry about career constraints. However, another survey conducted in the United States [4] shows “a gender gap in perceived work productivity”: before implementing WFH, female and male employees reported the same level of self-rated work productivity. After transitioning to WFH, male employees performed with higher productivity than female employees [4]. Regarding age, people in their 40s have more negative opinions on WFH because of their unfamiliarity with teleworking. People aged 30-39 years have the most positive opinions because they can enjoy time with their families and they are already accustomed with new technologies for teleworking [7]. Previous studies [3,7,8] also show that opinions concerning WFH vary across different socioeconomic groups. A similar social media study of public sentiments on WFH has been conducted in the United Kingdom [9] and reported that more than 70% of tweets concerning WFH expressed a positive sentiment, with the main topics including “traffic,” “drink,” and “e-commerce.”

Similar approaches have been implemented by researchers who mined Twitter posts on public attitudes toward face masks through natural language processing [10] using the Valence Aware Dictionary and Sentiment Reasoner (VADER) model [11] to perform sentiment analysis. Moreover, Twitter data have been used to study many different aspects of COVID-19, such as mining of overall public perception of COVID-19 [12], college students’ attitudes toward the pandemic [13], people’s attitude toward potential COVID-19 vaccines [14,15], sentiment analysis among pregnant women during quarantine [16], and monitoring of depression trends on Twitter during the COVID-19 pandemic [17]. These studies have used VADER for sentiment analysis, and most of them also include a time-series analysis. In addition, we followed the practice of using latent Dirichlet allocation (LDA) [18] to identify topics among a large text corpus. The M3-inference model [10] was used to portray different demographic groups.

In this study, we intend to understand public opinions on WFH, using large-scale social media data. Twitter has been a popular social media platform for people, especially in the United States, to express their opinions on what is happening around them. In contrast, the Boston Consulting Group used survey data to study employees’ opinions regarding WFH owing to the COVID-19 pandemic [19]. However, social media data allow an opportunity for conducting a timelier study of many population-level issues on a larger scale [20]. We acquired data with an authorized Twitter developer account using Tweepy. This ensures reliability by acquiring first-hand and sufficient data for the study.

In this study, we also inferred user demographic information using Twitter user information. This is important, since we can carry out an in-depth assessment of the characteristics of those who are more pro-WFH. For example, when we consider gender, we understand that historically mothers have been mostly responsible for caring for children [21]. Therefore, we need information regarding the users’ gender to determine whether there is any difference in sentiment toward WFH between women and men, as WFH would allow female employees to allocate more time to spend with their children.

Our goal is to understand the public opinions on WFH in the United States during the COVID-19 pandemic. In particular, we focused on the following research questions:

To summarize, in a large-scale data set of publicly available Twitter posts concerning WFH during April 10-22, 2020, we found that women and older people are more likely to tweet about WFH. On performing ordinary least squares regression analysis, we confirm that sentiment of WFH varies across user characteristics. In particular, women tend to be more positive about WFH than men. People in their 40s are more positive toward WFH than those in other age groups. People from high-income areas are more likely to have positive opinions about WFH.

These nuanced differences are supported by a more fine-grained topic analysis. At a higher level, we found that the most negative sentiment about WFH roughly corresponds to discussions on government policy. However, people express more positive sentiment when discussing topics on “remote work and study” and “encouragement.” Furthermore, topic distributions vary across different user groups.

In this section, we summarized the data collection process and the methods we used in the analyses. To address research questions 1 and 2, we discuss how we inferred user characteristics and the sentiment in the “Feature Inference” subsection. To address research question 3, we describe how we extracted the topics of tweets in the “Topic Modeling” subsection.

We collected relevant English-language tweets through the Tweepy stream application programming interface (API) using keywords and hashtag-filtering. The filter keywords and hashtags are “WFH,” “workfromhome,” “work from home,” “#wfh,” and “#workingfromhome.” In total, 553,166 unique tweets with 23 attributes posted by 405,455 unique Twitter users during April 5-26, 2020, were sampled. We attempted to infer the gender, age, and ethnicity of these Twitter users, extract the population density of the area they resided in, and estimate the sentiment of the tweets. There are 405,455 unique users in our data set, 313,815 (77.3%) of whom only tweeted once. After excluding duplicates and users with incomplete features, 28,579 unique Twitter users with all features were included in the data set.

We used LDA [18] to extract topics from the tweets. In our study, we used the stop words package of the Natural Language Toolkit library, extended with topic-related words (eg, “work” and “home”). To extract the most relevant topics, we only collected nouns, verbs, adjectives, and adverb lemmas. We use the spaCy package to screen all the words of the tweets and only includes those words whose postag is “NOUN,” “ADJ,” “VERB,” or “ADV.” We tuned the hyperparameters with nested looping topic numbers α and β. Finally, we chose num_topics=9, α=0.91, β=0.31, and a coherence score C_v of 0.379.

As indicated above, Twitter users’ opinions of WFH were slightly positive. We attempted to investigate the relationship between user characteristics and the sentiment of discussions on WFH. We performed ordinary least squares regression analysis on our data set (n=28,579). Descriptive statistics and bivariate correlations are shown in Table 1. Table 2 summarizes the results of ordinary least squares regression analysis.

Further, we attempted to investigate what Twitter users mainly discuss with reference to WFH. In particular, we investigated how the contents of the tweets correlate with the sentiment of WFH. Table 3 shows the 9 topics extracted by the LDA model. We assigned each topic a title on the basis of the top 10 keywords.

Figure 1 shows the proportions of the topics. Topic 1 (family activities) contained the keywords “dog,” “wife,” and “watch” and accounts for 17.7% of the total tweets. Topic 2 (remote work or study) accounts for 16.7% of the total tweets, where people mostly tweeted about remote work and study. Topic 3 (quarantine) contained the keywords “pandemic,” “force,” and “stay.” Topic 4 (dressing) contained the keywords “dress” and “wear,” and most of the tweets based on this topic discuss what people wear when they WFH. Topic 5 (government and policy) accounts for 6.8% of the total tweets and contained the keywords “money,” “job,” and “force”; in this topic, many of the tweets mention the names of governors and express their concerns about WFH-related policies. An example of such a tweet is as follows:

Topic 6 (COVID-19 side effects) contained the keywords “still” and “job” and accounted for 8.5% of the total tweets. In this topic, people mostly complained about the influence of COVID-19, such as “job changing” and “staying home for so long.”An example of such a tweet is as follows：

Topic 7 (encouragement) accounted for 7.5% of the total tweets, where people express their support and inspire people to overcome difficulties together. One example of such a tweet is as follows:

Topic 8 (back to office) contained the keywords “back” and “office” and accounted for 13.2% of the total tweets. Under this topic, people mostly tweet about their opinions toward the office, including those inquiring if the return to “office” will finally materialize or when people will be able to go back to office. Topic 9 (leasing) contains keywords “year,” “lease,” “renew,” and “office,” where people argue that some companies might not renew their office lease for the next year because of how well-suited WFH is for these companies.

Figure 2 shows the average sentiment score of each topic. Topic 7 (encouragement) had the highest average sentiment score (0.460) and was considered the most positive topic; in contrast, topic 5 (government and policy) was the least positive topic (average sentiment score=0.129). As indicated above (“Feature Inference” subsection), the average sentiment score of all the users in our data set was 0.242. Among these 9 topics, all expressed a positive sentiment toward WFH. Moreover, topics 2 (remote work or study), 7 (encouragement), and 8 (back to office) had a sentiment score above the average, while topics 1 (family activities), 3 (quarantine), 4 (dressing), 5 (government and policy), 6 (COVID-19 side effects), and 9 (leasing) had a sentiment score below the average.

This study represents a large-scale quantitative analysis of public opinions on WFH in the United States during the COVID-19 pandemic. Through the lens of social media, we found that gender and age are the most influential features to public opinions about WFH. On performing ordinary least squares regression analysis, we found that the sentiment toward WFH varies across user characteristics. In particular, women are more positive about WFH, which could be related to the change in working styles [30] and reduced work hours compared to those of men [31]. People aged ≥40 years tend to be the most pro-WFH than other age groups. This could be owing to the fact that people of those ages are the most vulnerable to COVID-19, while also being the most difficult people to be re-employed upon losing their jobs. These people also need to work to mitigate the shrinkage of retirement savings that were invested in the rather inert stock market. People from high-income areas are more likely to have positive opinions about WFH, which echoes the findings of Barrero et al [3].

These nuanced differences are supported by a more fine-grained topic analysis. At a higher level, we found that all the topics expressed a positive sentiment about WFH. However, people expressed a more negative sentiment toward family activity and the government. Under the topic of family activity, we noticed that women pay more attention to family than men, and we identified superwomen in WFH. When people talk about government and policy, money and jobs are 2 major concerns. Furthermore, based on our analysis by age groups, we noticed that the family-work boundary is another issue that varies among different age groups. As age increases, more people prefer to discuss work rather than family, which implies that family boundaries are becoming more permeable than work boundaries.

Barrero et al [3] reported that WFH would persist even after the pandemic. It is critical to understand public opinions on WFH to help improve their experience and to design a more suitable and flexible work policy. Our study suggests that there are nuanced differences across user characteristics. Government and company policymakers could design a more customized work policy to not only increase work productivity but also improve work satisfaction among their employees. It is also important to address the WFH-related disparities that have been reported among different racial and socioeconomic groups [36,37].

Our study is focused on the relationship between user characteristics and the sentiment about WFH. However, user occupation can be included in future analyses. Since the ability to WFH varies among different jobs [28], 1 potential hypothesis could be that people of different occupations have different opinions about WFH; thus, occupations would have an impact on the sentiment of WFH. In addition, there are some limitations of only using the Ethnicolr API to infer ethnicity. The Ethnicolr API is trained on voter registration data from Florida. First, using data from a single state (albeit a representative state) may not be ideal, since the pattern of names can be different among states. Another limitation is that our training data set was imbalanced (8,757,268 non-Hispanic White people, 1,853,690 non-Hispanic Black people, 2,179,106 Hispanic people, and 253,808 Asian people). Although these numbers are consistent with the population distribution in the United States, when training an inference model, we believe that the use of a more balanced data set could provide a better outcome.

This is a large-scale social media–based study on people who are more likely to tweet about WFH. On performing ordinary least squares regression analysis, our study shows how the sentiment of WFH varies across user characteristics. On conducting a content-based analysis, we carried out an in-depth analysis to determine what Twitter users mainly discuss and how the content of their tweets correlates with the sentiment of WFH. This paper contributes to a better understanding of public opinions on WFH in the United States during the COVID-19 pandemic and contributes to making policies both at national and institution or company levels to improve the overall population’s experience of WFH.