Is Working From Home The New Norm? An Observational Study Based On A Large Geo-tagged COVID-19 Twitter Dataset

Figure 1 shows the daily distribution of geo-tagged tweets within the top 10 states with the highest tweet volumes.¹¹1As mentioned in Appendix A.1, we lost around one-third detailed tweets between Mar. 18 and Apr. 4 due to the corrupted data. But we recorded the daily tweet counts during this period (see the dashed lines in Figure 13). Our crawlers shut down for 8 hours and 9 hours On Mar. 27 and Apr. 23 respectively, which caused the data gaps in the two days. We can see that daily tweet volumes generated by different states show similar trends. In fact, we tested statistical relationships regarding the daily volumes over time for all pairs of two arbitrary states, and found strong linear correlations existed among 93.2% state pairs with a Pearson’s $r>0.8$ and $p<0.001$.

Based on key dates, we split the entire observation period into the following three phases.

• •

  Phase 1 (from Jan. 25 to Feb. 24, 31 days): people mentioned little about COVID-19 except for a small peak at the end of January.

• •

  Phase 2 (from Feb. 25 to Mar. 14, 19 days): the number of COVID-19 related tweets began to increase quickly. On Feb. 25 U.S. health officials warned the COVID-19 community spread in America was coming [1]. On March 13, the U.S. declared the national emergency due to COVID-19 [2].

• •

  Phase 3 (from Mar. 15 to May 10, 57 days): people began to adjust to the new normal caused by COVID-19, such as working from home and city lockdowns.

For each tweet, we converted the UTC time zone to its local time zone²²2For states spanning multiple time zones, we took the time zone covering most areas inside the state. For example, we used Eastern Standard Time (EST) when processing tweets from Michigan because EST is adopted by most of the state. Except for Arizona and Hawaii, we switched to Daylight Saving Time (DST) for all states after Mar. 8, 2020 according to the state where it was posted. The aggregated hourly tweeting distributions in different phases are shown in Figure 2. The tweeting behaviors on workdays and weekends were studied separately because we wanted to figure out how the working status impacted on tweeting patterns. We colored the tweeting frequency gaps during business hours (8:00-16:59) as green if people tweeted more frequently on weekends than workdays. Otherwise, the hourly gap is colored as red.

In Phase 1, there existed a tweeting gap from 8:00 to 16:59 between workdays and weekends. The tweeting peak occurred at 12:00-12:59 on weekends but at 17:00-17:59 on workdays. We think it may be explained by the fact that people engage at work during regular working hours and have little time to post tweets on workdays. But they become free to express concerns on COVID-19 on Twitter after work.

The hourly distribution patterns changed in Phase 2 when confirmed COVID-19 cases increased quickly in the United States. People posted COVID-19 tweets more frequently during business hours than at the same time slots on weekends, indicating COVID-19 had drawn great attention of workers when they were working.

It is interesting to note that a green tweeting gap from 8:00 to 16:59 reappeared in Phase 3 when most people had worked from home. These findings motivated us to take advantage of the tweeting frequencies on workdays and weekends to estimate work engagement in the COVID-19 crisis (see Section 4).

We extracted the state information from both general and exact tweet locations and calculated tweet volume percentages for each state, as shown in Figure 3(a). The most populated states, i.e., California, Texas, New York, and Florida, contributed the most tweets. In contrast, less populated states such as Wyoming, Montana, North Dakota, and South Dakota created the least tweets. We measured the relationship between tweet volumes and populations for all states, and found a strong linear correlation existed (Pearson’s $r=0.977$ and $p<0.001$).

Then we normalized tweet volumes using state residential populations. Figure 3(b) illustrates Washington D.C. posted the highest volume of tweets by every 1000 residents, followed by Nevada, New York, California, and Maryland. The rest states demonstrate similar patterns. We think the top ranking of Washington D.C. might be caused by its functionality serving as a political center, where COVID-19 news and policies were spawned.

Unlike state populations, we did not find strong correlations between tweet counts and cumulative confirmed COVID-19 cases (Pearson’s $r=0.544$ and $p<0.001$) or deaths (Pearson’s $r=0.450$ and $p<0.001$). We further normalized tweet volumes based on COVID-19 cumulative number of cases and deaths in each state. Figure 3(c) and Figure 3(d) shows the average number of tweets generated by each COVID-19 case and each death respectively. Note that Hawaii and Alaska (not plotted in Figure 3(c) and Figure 3(d)) ranked as the first and second in both scenarios. Residents in states like Oregon, Montana, Texas, and California reacted sensitively to both confirmed cases and deaths, as these states dominated in Figure 3(c) and Figure 3(d).

We utilized GPS locations to profile the geographic distribution of COVID-19 tweets at the county level because general tweet locations might not contain county information. In our collected geo-tagged tweets, 3.95% of them contained GPS locations. We resorted to Nominatim [3], a search engine for OpenStreetMap data, to identify the counties where each tweet GPS coordinate lay. Figure 4(a) and Figure 4(b) visualize the raw GPS coordinate and corresponding county distributions. Large cities in each state demonstrated a higher tweeting density than small ones. In fact, we found a strong correlations between GPS-tagged tweet counts and county populations (Pearson’s $r=0.871$ and $p<0.001$). But such correlations did not hold true for cumulative confirmed COVID-19 cases (Pearson’s $r=0.590$ and $p<0.001$) or deaths (Pearson’s $r=0.497$ and $p<0.001$).

We assumed that (1) people would tweet less frequently during working hours if they engaged more on their working tasks; (2) if people spent no time on work tasks during business hours on workdays, their tweeting behaviors kept the same as that on weekends, especially when people were confined at home. We think the two assumptions are intuitive and reasonable as both Phase 1 and Phase 3 showed the meaningful working-hour tweeting gaps in Figure 2(b) and Figure 2(d).

We took the tweeting gap size as an indicator of work engagement. More specifically, let $h_{i}^{j}$ denote the tweet volume at the $i$-th hour on the $j$-th day in a week. For example, $h_{8}^{2}$ meant the number of tweets posted from 8:00 to 8:59 on Tuesdays (we took Monday as the first day in one week). Accordingly, the total number of tweets on the $j$-th day in a week was represented by $T_{j}=\sum_{i=1}^{24}h_{i}^{j}$. The total tweet volumes on workdays and weekends can be expressed as $T_{workday}=\sum_{j=1}^{5}T_{j}$ and $T_{weekend}=\sum_{j=6}^{7}T_{j}$.

Note that we estimated both hourly and daily work engagement by considering at least seven days, because the data sparsity would lead to unreliable results if fewer days were involved. The work engagement at $i$-th hour $H(i)$ could be defined as the ratio of the normalized tweeting frequency at $i$-th hour on weekends over that on workdays and minus one, as expressed in Equation 1.

where $i\in\{8,10,11,12,13,14,15,16\}$. A larger positive $H(i)$ indicates higher work engagement. When $H(i)$ equals 0, it means there exists no difference on work engagement at $i$-th hour between workdays and weekends. A positive value of $H(i)$ implies people are more engaged at work on workdays than weekends. Although it is rare, a negative $H(i)$ means people fail to focus more on their work on workdays than weekends.

We also defined the daily (from Monday to Friday only) work engagement by aggregating tweeting frequencies in regular working hours (from 8:00 to 16:59). The daily work engagement on $j$-th day was expressed as:

where $j\in\{1,2,3,4,5\}$ and $T_{j}$ was the total tweet count on the $j$-th day. Similar to hourly engagement, a larger positive $D(j)$ means higher work engagement.

Hashtags are widely used on social networks to categorize topics and increase engagement. According to Twitter, hashtagged words that become very popular are often trending topics. We found #hashtags were extensively used in geo-tagged COVID-19 tweets — each tweet contained 0.68 #hashtags on average. Our dataset covered more than 86,000 unique #hashtags, and 95.2% of them appeared less than 10 times. #COVID-19 and its variations (e.g., “Covid_19”, and “Coronavid19”) were the most popular ones, accounting for over 25% of all #hashtags. To make the visualization of #hashtags more readable, we did not plot #COVID-19 and its variations in Figure 7. In other words, Figure 7 displays top #hashtags starting from the second most popular one. All #hashtags were grouped into five categories, namely COVID-19, Healthcare, Place, Politics, and Others.

People use @mentions to get someone’s attention on social networks. We found most of the frequent mentions were about politicians and news media, as illustrated in Figure 8. The mention of @realDonaldTrump accounted for 4.5% of all mentions and was the most popular one. To make Figure 8 more readable, the mention of @realDonaldTrump was not plotted. Other national (e.g., @VP, and @JeoBiden) and regional (e.g., @NYGovCuomo, and @GavinNewsom) politicians were mentioned many times. As news channels played a crucial role in broadcasting the updated COVID-19 news and policies to the public, it is not surprising to observe news media such as @CNN, @FoxNews, @nytimes, and @YouTube are prevalent in Figure 8. In addition, the World Health Organization @WHO, the beer brand @corona, and Elon Musk @elonmusk were among the top 40 mentions.

To further explore what people tweeted, we adopted latent Dirichlet allocation (LDA) [6] to infer coherent topics from plain-text tweets. We created a tweet corpus by treating each unique tweet as one document. Commonly used text preprocessing techniques, such as tokenization, lemmatization, and removing stop words, were then applied on each document to improve modeling performance. Next, we performed the term frequency-inverse document frequency (TF-IDF) on the whole tweet corpus to assign higher weights to most import words. Finally, the LDA model was applied on the TF-IDF corpus to extract latent topics.

We determined the optimal number of topics in LDA using $C_{v}$ metric, which was reported as the best coherence measure by combining normalized pointwise mutual information (NPMI) and the cosine similarity [7]. For each topic number, we trained 500-pass LDA models for ten times. We found the average $C_{v}$ scores demonstrated an increasing trend as the topic number became larger. But the increasing speed became relatively slow if more than ten topics were considered. Therefore, we chose ten as the most suitable topic number in our study.

The ten topics and words in each topic are illustrated in Table 5. We can see that Topic 1 is mostly related to statistics of COVID-19, such as deaths, cases, tests, and rates. In the topic of treatment, healthcare related words, e.g., “mask”, “patient”, “hospital”, “nurse”, “medical”, and “PPE”, are clustered together. Topic 3 is about politics, as top keywords include “Trump”, “president”, “vote”, and “democratic”. The emotion topic mainly consists of informal language expressing emotions. Topic 5 is related to impact on work, businesses, and schools. We believe Americans who are bilingual in Spanish and English contributed to the topic of Spanish. Topic 7 is calling for unity in the community. In the topic of places, many states (e.g., Florida and California) and cities (e.g., New York and San Francisco) are mentioned. The last two topics are about praying and home activities when people followed stay-at-home orders. These topics are very informative and well summarize the overall conversations on social media.

Emotionally polarized words or sentences express either positive or negative emotions. We leveraged TextBlob [8] to estimate the sentimental polarity of words and tweets. For each word, TextBlob offers a subjectivity score within the range [0.0, 1.0] where 0.0 is most objective and 1.0 is most subjective, and a polarity score within the range [-1.0, 1.0] where -0.1 is the most negative and 1.0 is the most positive. We used the subjectivity threshold to filter out objective tweets, and used the polarity threshold to determine the sentiment. For example, a subjectivity threshold of 0.5 would only select the tweets with a subjectivity score greater than 0.5 as the candidates for polarity checking. A polarity threshold of 0.7 treated tweets with a polarity score greater than 0.7 as positive and those with a polarity less than -0.7 as negative.

Figure 9 illustrates the ratio of the number of positive tweets over negative ones with different combinations of subjectivity and polarity thresholds. Positive and negative emotions evenly matched with each other when the ratio equaled one. We can see that emotion patterns changes along with threshold settings. Specifically, positive emotions dominated on Twitter with small polarity and subjectivity thresholds. However, negative emotions became to overshadow the positive ones under large polarity and subjectivity thresholds. Figure 10 shows three examples of polarized word clouds where the ratio was greater than 1 (subjectivity=0.2, polarity=0.7), equal to 1 (subjectivity=0.8, polarity=0.2), and less than 1 (subjectivity=0.8, polarity=0.7).

Besides polarized-text based sentiment analysis, we took advantage of facial emojis to further study the public emotions. Facial emojis are suitable to measure tweet sentiments because they are ubiquitous on social media, conveying diverse positive, neutral, and negative feelings. We grouped the sub-categories of facial emojis suggested by the Unicode Consortium into positive, neutral, and negative categories. Specifically, all face-smiling, face-affection, face-tongue, face-hat emojis, and     were regarded as positive; all face-neutral-skeptical, face-glasses emojis, and     were grouped as neutral; and all face-sleepy, face-unwell, face-concerned, face-negative emojis were treated as negative. A full list of our emoji emotion categories are available at http://covid19research.site/emoji-category/.

We detected 4,739 (35.2%) positive emojis, 2,438 (18.1%) neutral emojis, and 6,271 (46.6%) negative emojis in our dataset. Negative emojis accounted for almost half of all emoji usages. Table 6 illustrates top emojis by sentiment categories with their usage frequencies. The most frequent emojis in the three categories were very representative. As expected,  still was the most popular emojis in all categories, which kept consistent with many other recent research findings[9, 10]. The thinking face emoji  was the most widely used neutral facial emoji, indicting people were puzzled on COVID-19. Surprisingly, the face with medical mask emoji  ranked higher than any other negative emojis. The skull emoji  appeared more frequently than any other positive and neutral emojis except    . We think the sneezing face  and the hot face emoji  are very likely to be relative to suspected symptoms of COVID-19.

We used facial emojis to track the different types of public sentiment during the COVID-19 pandemic. Figure 11(a) shows the daily overall emotions aggregated by all states. In Phase 1 (from Jan. 25 to Feb. 24), the publish emotions changed in large ranges due to the data sparsity. In Phase 2 (from Feb. 25 to Mar. 14), positive and negative emotions overshadowed each other dynamically but demonstrated stable trends. In Phase 3 (from Mar. 15 to May 10), negative sentiment dominated both positive and neutral emotions, expressing the public’s concerns on COVID-19.

We also investigated the daily positive, neutral, and negative sentiment of different states as presented in Figure 11(b), Figure 11(c), and Figure 11(d) respectively. The top five states with the highest tweet volumes, i.e., CA, TX, NY, FL, and PA, were taken as examples. Similar to Phase 1 patterns in Figure 11(a), the expression of emotion by people in different states varied greatly. In Phase 2 and Phase 3, the five states demonstrated similar positive, neutral, and negative patterns at most dates, as their sentiment percentages were cluttered together and even overlapped. However, there existed state-specific emotion outliers in Phase 2 and Phase 3. For example, the positive sentiment went up to 70% in PA when the Allegheny County Health Department (ACHD) announced there were no confirmed cases of COVID-19 in Pennsylvania on Mar. 5. People in New York state expressed more than 75% neutral sentiments on Feb. 27 when the New York City Health Department announced that it was investigating a possible COVID-19 case in the city. On Mar. 17 and Mar. 18, residents in PA demonstrated almost 100% negative sentiments when the statewide COVID-19 confirmed cases climbed to 100.

We studied the event-specific sentiment by aggregating tweets posted from different states when the same critical COVID-19 events occurred. We focused on the following eight events:

• •

  The first, the $100^{th}$, and the $1000^{th}$ confirmed COVID-19 cases,

• •

  The first, the $100^{th}$, and the $1000^{th}$ confirmed COVID-19 deaths,

• •

  Lockdown

• •

  Reopen

For the first seven events, we aggregated the tweets in CA, TX, FL, NY, GA, PA, IL, MD, VA, and AZ, which were also studied in Subsection 4.1.2. For the last event, we investigated the nine states of TX, GA, TN, CO, AL, MS, ID, AK, and MT, which kept consistent with Subsection 4.1.2. To our surprise, the average percentages of each sentiment type in the eight events demonstrated similar patterns, as shown in Figure 12. We carried out the one-way multivariate analysis of variance (MANOVA) and found the p-value was nearly 1.0, indicating there was no significant difference among these eight event-specific sentiments. When the first case, $100^{th}$ cases, and first death were confirmed, sentiment standard deviations were much larger than the rest events, suggesting people in different states expressed varying and diverse sentiments at the beginning of COVID-19 outbreak. The negative emotion reached the highest level among all events when $1000^{th}$ deaths were reported. The positive emotion achieved the highest level among all events when states began to reopen.