Exploring the Impact of COVID-19 Lockdown on Social Roles and Emotions while Working from Home

There has been several investigations into social role theory and it’s impact on the way that we interact with computers. People combine their work and private lives in different ways. Recently, Anderson et al. contributed to this by looking into how social roles can be inferred from ICT usage (Anderson et al., 2018). This is moving towards building intelligent attention management systems, controlling when people can be interrupted for tasks that relate to different social roles (Anderson et al., 2019).

The literature on COVID-19’s impact on valence and individual lives is still developing. Kleinberg et al has recently introduced a new dataset¹¹1https://osf.io/awy7r/ Accessed 3rd of July 2020 collecting emotions of individuals expressed through texts during COVID-19 (Kleinberg et al., 2020). Furthermore Zhang et al have investigated the impact of COVID-19 in Wuhan, and call to pay attention to the mental and physical impacts of those not affected by the virus epidemiological (Zhang et al., 2020).

Another source of COVID-19’s impact on individual happiness comes from the Hedonometer. The Hedonometer is a way of attempting to measure the general happiness of the population by observing the word usage of tweets (Dodds et al., 2011). This project found that although the COVID-19 caused a large dip in happiness, over time happiness has rebounded to be close to levels before COVID-19, asides from the dip in happiness caused by the murder of George Floyd and the consequent protests against police brutality. ²²2https://hedonometer.org/timeseries/en_all/ Accessed on the 1st of July 2020. This would hypothesise that although a dip in initial happiness may be observed, over time, the impact on happiness would be negligible.

Many of the existing literature regarding app usage behaviour analysis focused on how people use the apps in their smartphones (Ferreira et al., 2014; Verkasalo, 2009). These research suggested that mobile usage pattern of a person contain many micro-usage e.g, locking or unlocking phones. Some researchers worked towards predicting the next app which is more likely to be opened by the user (Huang et al., 2012; Liao et al., 2013). Jones et al. tried to determine which smartphone apps encourage mobile phone usage behaviour. The have analysed the revisiting patterns of different apps for 165 users. They found that the app usage behaviour on macro-level is technology independent (Jones et al., 2015).

The effects of working from home and work-related mobile phone usage in the after-hours have been investigated in the field of organizational psychology and work management. On the one hand, positive effects on individuals’ well-being have been found when, for example, work-related usage of mobile phones is in line with their motivation and individual preferences(Ohly and Latour, 2014). Also, the possibility to meet demands from both domains – work and private – have been shown to have positive effects on well-being, especially for individuals who like to integrate work and private-related matters(Derks et al., 2016). On the other hand, adverse effects related to conflicts between work and private domains(Delanoeije et al., 2019) as well as on individuals’ well-being have been reported – originating from the blurring of work and non-work domains(Mellner, 2016).

Due to the pandemic, a large number of people are currently working from home which has impact on daily app usage patterns. The social role and temporal patterns in app use are different than previous times. Being attentive to the feelings plays vital role in staying healthy during this tough time. This study shed light on all of these by analyzing the data during COVID-19.

This application used Experience Sampling Method ESM and gave a survey for the user to complete every 90 minutes that the user was active on their devices. This survey asked:

• •

  How happy were you feeling in the last hour? (Valence: Scale from 1 to 5)

• •

  How excited were you feeling in the last hour? (Arousal: Scale from 1 to 5)

• •

  What role you were fulfilling in the past 15 minutes? (Work, Private, or Both)

These surveys were optional and were collected before the outbreak of COVID-19 and during the outbreak. A total of 488 and 436 surveys were gathered before and during COVID-19, respectively.

For the comparison in happiness and roles, only the surveys of those who participated during and before the pandemic were used. The number of survey responses that were before and during COVID-19 from each participant is noted in Table 1. The histogram of surveys collected during COVID-19 is illustrated in Figure 1. The left group is the surveys collected before COVID-19 and the right the surveys collected during.

The application also recorded the application usage of the user. Recording the application that is currently active and the time it was opened for both smartphone and desktop.

This study filtered out windows utilities, (Desktop, start menu etc), Android home screen, and applications that were only used very little to make a dataset of over 174,000 records of usage of different applications over the time measurement period. The sample size used for the analysis is noted in Table 2.

These applications were then categorised into 12 categories. The categories and examples of applications are below:

• •

  Browser (Chrome, Microsoft Edge)

• •

  Communication (Discord, Messenger, WeChat, Zoom)

• •

  Developer Tools (Command line, Python, Sublime Text)

• •

  Entertainment (YouTube, Google Play Music, Spotify)

• •

  Game (Steam and other Games)

• •

  Media Viewing (VLC, Windows Media Player, Android Photos)

• •

  Reading (PDF viewers, Audible, Wuxia World)

• •

  Settings (Device maintenance, Android Settings)

• •

  Tools (Google Drive, ING Banking, tramTRACKER)

• •

  Work Communication (Email tools, Teams)

• •

  Utilities (Snipping Tool, Camera, Contacts, Phone)

• •

  Writing (TeX Studio, Notepad, One Note, Microsoft PowerPoint)

This study first compares the roles assumed by the participants. The aim of this is two fold. Firstly, to determine whether the pandemic introduced a change in social roles. Secondly, whether the introduction of COVID-19 caused a caused the proportion of work roles chosen each day to change.

We used Fisher’s exact test for the first comparison. Fisher’s exact test is a statistical method of determining whether two variables are independent or dependent of each other. It is particularly appropriate when the number of samples in the distribution is small(Campbell, 2007), which is true in our case and particularly for participant 2. The null hypothesis of the test is that the two distributions are independent.

The Fisher’s exact test used to compare the dependence of two variables, the amount a person chose a role, and whether or not the participant was taking the surveys during the lockdown. Below is a sample table shown for participant 1 for which Fisher’s exact test was run.

To determine whether the amount of work done in a day changed, we found the proportion that a user would choose a work role (that is, Work or Work and Private). Then did a t-test to determine whether these proportions changed before and during COVID-19. Box plots were also used to show the distribution of data.

To determine if the mean of the valence distribution changed before and during COVID-19, a t-test, which is a test to determine whether the mean of two distributions are different, was deployed. This test is to determine whether the average valence of each participant before and after COVID-19 changed.

One of our research question was to determine whether before COVID-19, the app usage activity, happiness, excitability and social role would be more consistent than the earlier days of COVID-19 and whether the early days of COVID-19 would be more varied than the later days of COVID-19.

This study takes inspiration from Sadri et al(Sadri et al., 2018) who tried to predict the afternoon movements of individuals based off what they did in the morning. They did this by determining how similar each day was to all other days within the observed window of time. A similar approach was taken in looking at how similar each day was to each other day in different time periods.

This study investigates the impact of the COVID-19 lockdown on these four variables. The first three were gathered in the ESM surveys presented to the users:

First, valence. Valence was gathered by asking the happiness of the individual. The valence is measured between 1-5.

Secondly, arousal. Arousal was gathered by asking participants how excited they feel, and is also measured between 1-5.

Thirdly, social role. The social role was a categorical value of ”work” ”private” or ”work and private”

Finally, app usage. The app usage was correlated by segmenting the data with IGTS (described below)

To determine how the valence, arousal and social role of a participant changed over the days before and during COVID-19, a four dimensional vector was created for each day, where each element was the average value recorded in the surveys between hours $[9-12)$, $[12-15)$, $[15-18)$, $[21-24)$ respectively, four three-hour segments. We chose these time intervals as these were the times that our participants were working.

The similarity matrix between each day for each of the ESM survey vectors was measured using Jaccard index, which is a popular method of similarity measurement between two finite sets specifically for sets categorical variables (Niwattanakul et al., 2013). It measures the similarity between two sets by dividing the common number of features with total number of available features. Jaccard distance is the inverse of jaccard similarity which can be obtained by using the equation 1.

As social role is a categorical variable, we compute Jaccard distance to find the non-similar days of a user during different periods of interest (before, 2 weeks after and 6 weeks after the pandemic declaration).

Information Gain Temporal Segmentation (Sadri et al., 2017) (IGTS) was applied to segment the app usage time series into different activities that the participant did throughout their day. IGTS is a method of segmenting multivariate times series by attempting to reduce the amount of entropy in each individual segment. The method creates segments that in total have the lowest entropy. IGTS was applied to each day to discern the activities that were completed throughout the day. To use IGTS, an appropriate value for the parameter $k$ must be chosen. $k$ is the amount of splits to be found in each day (the amount of segments is therefore $k+1$). Including more segments increases information gain. Information gain is the amount of entropy lost by segmenting the time series. However, too many segments decreases the size and interpretability of the segments found. So choosing a $k$ that finds a good balance between the interpretability and size of clusters is important. The original IGTS paper offers a method of finding the knee point in the information gain vs $k$ curve. This method was used over all the segments for all users for all days and the median $k$ was chosen for this analysis. The median $k$ was found to be 27. An example segmentation for a day is in figure 4.

IGTS was then applied over all the time series, giving a list of segments for each user.

We compare the similarities between days in terms of app usage using cosine similarity. Cosine similarity is measured based on the dot product of two vectors. Thus, the angle between two vectors is considered in measuring the similarity. Given two $N$ dimension vectors $\vec{A}$ and $\vec{B}$, the cosine similarity is computed using the equation 2.

In our experiment, the vector of each day contains the information gain of each type of app usage in a segment. We considered maximum activity of an app category in a segment. Considering total 12 app categories and 28 segments, each vector contains 336 elements. We applied the pairwise cosine similarity on the vectors corresponding to each day to identify the similar days of each user based on app usage activity during different periods.

The venn diagram in Figure 5 shows the proportions for which participants found themselves in different roles. Participant 2 and 3 showed a large increase in the proportion of work roles assumed during the lockdown. It also shows the heavily skewed work role being assumed during the lockdown period in comparison to private only and both work and private roles.

To determine whether the results shown in the Figure 5 were significant, we used fisher’s exact test twice. The null hypothesis for the first test is that the roles that an individual chose were independent of whether COVID-19 is currently occurring. The $p$ values for this test are reported in Table 4. The null hypothesis for the second test was that choosing a combined role over a separate role was independent of the pandemic. The $p$ values are reported in Table 5.

As shown in Table 4, the pandemic had a significant impact on the social roles specifically for participants 2 and 3. In Table 5, we observe that only participant 3 had combined social roles significantly.

To investigate whether COVID-19 caused a change in the size of proportion of work related roles assumed in a day, we charted the proportions of work assumed by each participant in a box plot shown in Figure 6. The significance of these results were determined using t-test and are found in Table 6.

This shows that 2 out of 3 participants showed a significant change in work, increasing the amount of work in a day during COVID-19, and all participants increased the proportion of work in their days. This confirms our hypothesis that the lockdown has an influence on longer hours dedicated to work daily.

Next, We analysed whether the valence distribution significantly changed during COVID-19 compare to before COVID-19 or not. A t-test of independence was performed to compare the distribution before and during COVID-19. The null hypothesis of the test was that the distributions of scores given (before and during COVID-19) have the same mean.

In Table 7, $\mu$ is the sample mean before COVID. $\sigma$ is the sample standard deviation before COVID-19. $\mu^{\prime}$ is the sample mean during COVID-19 and $\sigma^{\prime}$ is the standard deviation during COVID-19.

It was found that only one participant (again, Participant 3) had a significant difference in reported happiness. This also shows that for the majority of participants, there was not a significant change in valence. This does not provide strong evidence that valence was impacted by the lockdown and working from home. This, however, only shows an aggregate analysis. Later, in section 5.4 we present an analysis of the fine-grained changes in the distribution of valence during lockdown.

Figure 7 shows the distributions of reported happiness for each participant. Participant 1 and 3 are similar. Participant 2 shows a difference in distribution, however, it is not significant because Participant 2 did not have a large sample size of surveys during the pandemic.

The next research question is how COVID-19 impacted the stability of valence, arousal, social role and app usage. We are reporting the behaviour distribution similarity for participant 1 (with his associated desktop app usage) and participant 4 (with his associated mobile app usage).

Figure 8 shows the similarity of valence, arousal, social roles and desktop app usage day to day for Participant 1 including periods of before, 2 weeks into and 6 weeks into the pandemic. The grey cells represent days that we do not have data.

Figure 8 shows that for valence, this user reports much more stable valence before the pandemic. Two weeks after the pandemic declaration, the stability of the participant’s valence was affected, causing it to change every two to three days. After four weeks, the valence of the participant has seems to have stabilised, showing large portions of time reporting similar distributions of valance each day.

Figure 8 also suggests that the participant’s arousal distribution is highly varied during the weekend and early part of the week before the pandemic. During lockdown, the participant’s arousal distribution changed across different days, including weekdays and weekends. In the third phase (third row shown in Figure 8), signifying sixth and seventh week into the lockdown, the arousal seems to stabilise, however, around 11th May, the arousal distribution was highly dissimilar. We found that this was the time when restriction was loosened in Victoria, and up to five visitors were allowed first time indoors and up to 10 outdoors. This potentially explains why the valence on 10th May also shows a different distribution given that the news was released this day.

Interestingly, the social role of the participant shows a large amount of instability at 6 weeks after the pandemic declaration for this participant. This could be caused by changing between work and private roles as they are becoming accommodated to working at home. This is especially interesting considering the charts do not show a large change in stability before the pandemic and 2 weeks in. We can also observe similar behaviour in social role for an another participants in Figure 9.

Figure 9 shows similarity matrices for participant 4’s survey responses and the mobile app usage distribution. The figure suggests that the valence of the individual varied less in the later weeks of the pandemic, along with the arousal. At the start of the pandemic, the valence distribution was found to be different from day to day. Six weeks into the pandemic, the behaviours had stabilised and only changed on the Tuesday and Wednesday and Thursday of each week.

During the pandemic, participant 4 had more varied work and private roles later in the pandemic than earlier in the pandemic.