Tracking WhatsApp behaviors during a crisis: A longitudinal observation of messaging activities during the COVID-19 pandemic

During a crisis, the messaging platform WhatsApp allows crisis-related information to be disseminated quickly. Although case studies have documented how WhatsApp has shaped crisis outcomes in both beneficial and harmful ways, little is known about: (i) how crisis-related content is spread; (ii) characteristics of users based on usage patterns; or (iii) how usage patterns link to well-being. During the coronavirus disease 2019 (COVID-19) crisis, this study used the experience sampling method to track the daily WhatsApp usage of 151 adults throughout one week (capturing a total of 924 days of crisis-related communication). Each day, participants reported the extent to which they had received, forwarded, or discussed COVID-19- related content. During the week-long monitoring, most participants (94.7%) reported at least one COVID-19 related use of WhatsApp. Those who engaged with more COVID-19 content in personal chats were more likely to report having COVID-19 thoughts throughout the day. We further observed that around 1 in 10 individuals (14%) were chronic users who received and shared forwarded COVID-19 messages at a high volume; this group may represent everyday "super spreaders" of crisis-related content. Together, these findings provide an empirical base for policy makers to manage risk communication during large-scale crises.


Highlights
• WhatsApp is used for information-sharing during crises.
• In the COVID-19 crisis, nearly all participants used WhatsApp to receive or send crisisrelated content during one week of monitoring.
• Users who engaged with more COVID-19 content reported higher levels of COVID-19 thoughts throughout the day.
• 1 in 10 participants shared a high number of forwarded COVID-19 messages, and may represent 'super-spreaders' of crisis-related information.Tracking WhatsApp usage during COVID-19

Introduction
In April 2020, the messenger platform WhatsApp imposed new restrictions on message forwarding (Newton, 2020).Whereas users could previously forward content to multiple contacts, they could now forward only to single recipients if a message had already been shared by five users (Newton, 2020).According to WhatsApp, this was implemented to curb the spread of misinformation, following steep increases in forwarding activity during the coronavirus disease 2019 crisis (COVID-19) (Singh, 2020a).
Crises refer to situations where social, institutional, and organizational affairs are endangered, requiring rapid action from all levels of society (Rosenthal et al., 1991).Traditional examples include terrorist attacks, civil unrest, natural disasters, and epidemics (Pan & Meng, 2016).In these situations, individuals tend to seek out and share information, allowing them to learn about the crisis, resolve uncertainties, and regain personal agency over an unpredictable situation (Lin et al, 2016).
Despite this quest, a rapidly changing crisis situation makes information difficult to attain.
As such, members of the public typically rely on multiple sources, with social media increasingly turned to as the medium of choice (Malecki et al., 2020).Social media's peer-to-peer nature allows information to be exchanged quickly (Gutteling & de Vries, 2017;Lambert, 2020), and high usage has been observed during emergencies (Guetteling & de Vries, 2017).For example, when Hurricane Sandy struck in 2012, the platform Twitter recorded 12.5 posts per second (Lachlan et al., 2014;Lachlan & Spence, 2014) -a remarkable rate that highlights the sheer volume of content shared during this period.
Nonetheless, this fast-paced exchange of information can cause harm (Clark Ginsberg & Sayers, 2020).As a case in point, the 2013 Boston Marathon was targeted in a terrorist attack.When two bombs exploded close to the finish line, misinformation about the bombers' identity spread rapidly on Twitter, resulting in the inaccurate identification of suspects (Cassa et al., 2013;Starbird et al., 2014).Similarly, social media discussions generated widespread panic when a shooter entered the YouTube headquarters in 2018.Within minutes of the shooting, Tracking WhatsApp usage during COVID-19 false information was spread about the identity of the shooter, the reasons for the shooting, and the total death toll (Agence France-Presse, 2018).
Together, these accounts highlight how risk communication during crises needs to factor both the prevalence and potential harm of social media usage.In turn, this benefits from an understanding of how crisis-related content is spread from person to person.Although contentsharing has been documented extensively for social networking platforms with public posts (e.g., Twitter, Facebook) (Brandtzaeg et al., 2010;Shi et al., 2014), private messaging platforms (e.g., WhatsApp, Telegram) remain understudied.This is important to address because publicly available metrics (e.g.Facebook likes, re-tweets) largely capture passive social media usagethat is, how individuals consume crisis-related content.On the other hand, the study of private messaging platforms may allow for proactive information-sharing to be captured as well (e.g., the production and discussion of content), providing a more veridical picture of user engagement (Yu, 2016;Chen et al, 2020).Correspondingly, we report for the first time a naturalistic study tracking how WhatsApp users sent and received crisis-related content during the global COVID-19 pandemic.

Use of WhatsApp for information sharing
With 1.5 billion users across 180 countries, WhatsApp is the most commonly used messaging application worldwide (Singh, 2020b).The platform allows users to exchange encrypted texts, various multi-media, and calls with each other.Additionally, group chat features provide an avenue to reach up to 256 members at a time (WhatsApp, 2020a).Given these features, WhatsApp is used increasingly for information exchange (Bhuvana & Arul Aram, 2019).One recent global survey reported that 1 in 5 respondents turned to the platform as a source of news, with the number reaching 50% in countries with higher WhatsApp penetration (e.g.Brazil, Malaysia, and South Africa; Boczek & Koppers, 2020;Reuters, 2018).
WhatsApp facilitates information exchange through several means.First, formal organisations can send announcements through the platform's business application Tracking WhatsApp usage during COVID-19 programming interface (WhatsApp, 2020b).Introduced in 2018, the interface supports messaging capacities on a larger, automated scale, allowing organizations to reach 100,000 subscribers each day (Facebook for Developers, 2020).Correspondingly, WhatsApp has been adopted by authorities to disseminate official information (Reuter et al., 2016;Lindsay, 2011) as in the case of Israel's WhatsApp accounts to share local-and national-level updates from emergency services (Simon et al, 2016).
Apart from one-sided information flow from authorities, WhatsApp has also been used for more dynamic sharing of unofficial information amongst peers (Simon et al, 2016).Outside the crisis setting, several papers have documented how WhatsApp groups in laboratory and hospital teams facilitated the provision of instructions, quick updates, and the asking and answering of questions (Kamel Boulos et al., 2016).In one study, WhatsApp support groups also proved more successful than Facebook groups or information booklets in preventing relapse after smoking cessation; this effect was attributed to the richer discussions and stronger social support afforded by WhatsApp (Cheung et al., 2015).

WhatsApp use during crises
Returning to the crisis setting, the use of WhatsApp to disseminate information has been described in several case studies and retrospective surveys.For example, one survey of 400 residents affected by the 2015 Chennai floods found that WhatsApp had been quickly deployed to: provide updates on the weather, rescue stranded families, and allocate resources (Bhuvana & Arul Aram, 2019).Similarly, when terrorist attacks occurred in Tripoli, 43 interviewees described how they had used WhatsApp to request for help, report on the incident, or connect with family and friends (Baytiyeh, 2018).
While these examples highlight WhatsApp's key role in organizing efforts during a crisis, the ease of information-sharing means that WhatsApp has also been implicated in the spread of misinformation (Resende et al., 2019;Rossini et al., 2020).Notably, following the 2018 flood of Kerala, WhatsApp was the primary channel through which misinformation spread amongst citizens (Kumar et al., 2019).Likewise, when official information was scarce after 3 boys were kidnapped from an Israeli bus stop, the public demonstrated a clear preference for information-Tracking WhatsApp usage during COVID-19 sharing on WhatsApp.This led to the increased spread of rumours as members of the public, believing WhatsApp to be private, were less discerning of the information they received and shared (Simon et al, 2016).
Taken together, the extant literature describes various roles that WhatsApp can take on during a crisis.What remains unclear, nonetheless, is how exactly WhatsApp comes to adopt those roles.In the midst of a crisis, who sends crisis-related messages, to whom, and in what manner?This finer-grain level of understanding is missing from the current literature, but is necessary to encourage beneficial WhatsApp use amidst crisis settings.

WhatsApp use during the COVID-19 crisis
The ongoing COVID-19 crisis provides novel opportunities to understand WhatsApp usage on a fine-grained level.Just 6 months following the declaration of the COVID-19 outbreak as a pandemic (in March 2020), 27 million people have been infected worldwide with 800,000 deaths (World Health Organization, 2020).The scale of the crisis has been met with a proliferation of COVID-19 content (Depoux et al., 2020), and in line with this volume WhatsApp experienced a 40% surge in usage rates -the largest pandemic-related increase amongst social media platforms (Kantar, 2020).
The expansion of WhatsApp's user base has been accompanied by more messageforwarding activities during the pandemic (Porter, 2020).Correspondingly, WhatsApp has been linked to the spread of COVID-19 rumours (Jha, 2020), and the platform introduced messageforwarding limits to curb the spread (as described in the opening paragraph) (Singh, 2020).
Nonetheless, as in previous crises, WhatsApp user behaviours have received little research attention, with COVID-19 social media research focusing on publicly-available metrics such as Facebook likes and shares (Chen et al., 2020).

The current study
Addressing gaps in the literature, we sought in this study to document day-to-day WhatsApp usage patterns during the COVID-19 crisis.We applied the experience sampling method to track usage amidst everyday routines (Csikszentmihalyi & Larson, 2014;Hektner et al., 2007), contacting participants daily for week to capture their frequency of receiving, Tracking WhatsApp usage during COVID-19 forwarding, or discussing COVID-19-related content on WhatsApp.By tracking usage patterns in real time, this method: (i) minimized recollection biases (as compared to traditional recall surveys; Hektner et al., 2007); (ii) allowed day-to-day variations to be captured (Depoux et al, 2020); and (iii) had high ecological validity, meeting the call for research documenting the spread of COVID-19 information under naturalistic online contexts (Pennycook et al., 2020).
As a secondary aim, we sought to profile participants based on their WhatsApp usage patterns.Outside the crisis context, individual differences have been observed in how people forward messages on social media.For example, Lottridge & Bentley (2018) identified 3 categories of information-sharers: (i) those who shared information in publicly-accessible posts, private messages, and social network platforms; (ii) those who shared information in only private and social channels; and (iii) those who did not share information at all.Extending this finding to the crisis context, we likewise sought to taxonomize WhatsApp users based on their crisis-related usage behaviours.

Linking WhatsApp usage patterns to well-being
Finally, taking advantage of the experience sampling method, we also investigated whether day-to-day variations in well-being during the crisis (in terms of pandemic-related fear and intrusive thoughts) tracked day-to-day variations in WhatsApp usage.Although several surveys have linked excessive social media exposure to crisis-related fear (Gao et al, 2020;Lin et al., 2020) no study has tracked the role of specific social media activities in an intensive longitudinal design.

Study context
Our study took place in Singapore, a city state in Asia with 70% WhatsApp penetration (Steup, 2020).During the study period (March-May 2020), Singapore had a surge of daily COVID-19 cases -from 23 at the start of the survey (or 4 cases per million population; 17 March), to a peak of 1426 in mid-April (243 per million population; 20 April), before tapering off to 741 by the end of the study (126 per million population; 7 May) (World Health Organization, 2020).Throughout the second half of the study period (7 April to 7 May), participants experienced a nation-wide lockdown with movement restrictions, and business and school Tracking WhatsApp usage during COVID-19 closures (Baker, 2020).(The resident population's behavioural and emotional responses to COVID-19 during this period have been documented elsewhere; see Liu & Tong (2020); Long & Liu (2020); Saw et al. (2020)).

Participants
Between 17 March to 7 May 2020, 151 participants were recruited from the general community (104 females; mean age 36.35 ± 14.70 years; see Table 1).Participants responded to advertisements placed in Facebook and WhatsApp community groups (e.g., residential groups, workplace groups, university groups), posts on popular online forums, and paid Facebook advertisements targeting Singapore-based users.All participants fit the study inclusion criteria, and (1) were 21 years or older, (2) had lived in Singapore for at least 2 years, and (3) had a WhatsApp account.
The study protocol was approved by the Yale-NUS College Ethics Review Committee and was pre-registered at ClinicalTrials.gov(NCT04367363).

Measures
At the start of the study, participants provided informed consent through an online link.They then completed: (i) a baseline questionnaire; (ii) experience sampling responses daily for 7 days; and (iii) a final questionnaire.All study activities took place online via the survey platform Qualtrics, and participants were reimbursed SGD $5 upon study completion.

Baseline questionnaire
The baseline and final questionnaires were designed to characterise participants, allowing them to be profiled based on their WhatsApp usage patterns observed during the week.
In the baseline questionnaire, participants first reported demographic information: their age, years in Singapore, gender, religion, ethnicity, marital status, education, house type (a proxy of socioeconomic status), household size, country of birth, and citizenship.Next, participants used 4-point scales to indicate their pandemic-related concerns: (1) how confident they were that the government could control the nationwide spread of COVID-19 (1 = "Not Tracking WhatsApp usage during COVID-19 confident at all, 4 = "Very confident"); (2) how likely they judged that they (or someone in their immediate household) would be infected with COVID-19 (1 = "Not at all likely", 4 = "Very likely"); and (3) how fearful they were about the situation in the country (1 = "Not scared at all", 4 = "Very scared") (Liu & Tong, 2020;Mesch & Schwirian, 2019).

Experience sampling
The experience sampling protocol started one day after the baseline questionnaire was completed.Across 7 days, participants received a WhatsApp message each evening (2130 hrs) directing them to an online Qualtrics link.The timing was chosen to capture WhatsApp usage during the day, with a single prompt selected to reduce triggers for pandemic thoughts and fears (Fiorillo & Gorwood, 2020).
At each prompt, participants first indicated their well-being: (i) how afraid they felt ("Today, to what extent are you fearful about the COVID-19 situation", answered using a 4-point scale with 1="Not scared at all" and 4="Very scared"); and (ii) how intrusive COVID-19 thoughts had been ("Today, I thought about the COVID-19 situation all the time", answered using a 5point scale with 1="Not at all true" and 5="Very true").
Next, participants addressed the primary outcome measures on WhatsApp usage.Items were designed to capture "COVID-19 chatter" -the extent to which individuals forwarded COVID-19 content, received forwarded content, or generated novel COVID-19 messages (e.g., in a discussion).Participants responded based on the day's usage (starting from the time they woke up) and were encouraged to verify responses by referring to the application.Tracking WhatsApp usage during COVID-19 Following pilot testing for clarity, questions were organized in terms of messageforwarding, personal chat, and group chat behaviours.For forwarding behaviours, participants were first asked if they had forwarded any message related to COVID-19 (yes/no) -whether copied and pasted into a chat or shared via WhatsApp's 'forward' button.If participants responded 'yes', they were then asked: (i) how many unique COVID-19 messages they had forwarded (regardless of the number of recipients), and (ii) how many unique groups and individuals they had forwarded to (regardless of the number of messages sent).
Personal chat behaviours involved participants' one-to-one WhatsApp chats.For this category, participants first indicated whether they had received any COVID-19 messages forwarded to their personal chats (yes/no).If they responded 'yes', they were then asked: (i) how many unique messages they had received, and (ii) how many different people they had received messages from.Thereafter, participants were asked whether they had discussed COVID-19 in personal chats -defined as conversations where either they or the other person generated their own message(s) related to COVID-19 (yes/no).For participants who indicated 'yes', they were then asked how many unique chats were involved.
Finally, for group chats, participants were asked if COVID-19 had been mentioned in any of their WhatsApp groups by at least one other person (not including themselves).This could have occurred either through others forwarding messages, or through others writing their own comments (yes/no).Affirmative responses were followed with a question on how many WhatsApp groups had done so.
Participants were instructed to complete the experience sampling measures within one hour of contact.Through this protocol, we collected 924 data points across 151 participants (compliance rate: 86.2%).

Final questionnaire
A day after the experience sampling protocol ended, a final questionnaire was administered.Participants indicated which of 13 possible sources they had obtained COVID-19 news from (e.g., printed newspapers, radio, WhatsApp, YouTube), then repeated the 3 Tracking WhatsApp usage during COVID-19 pandemic-related questions and the DASS-21 (from the baseline questionnaire).The reported sources of COVID-19 news are illustrated in Figure 2.
Finally, to ensure we had captured a representative sample of WhatsApp chatter (given the cut-off time we had set for experience sampling prompts), we asked participants to indicate the primary time of day they had read and sent COVID-19 messages over the previous 7 days (mostly in the morning, afternoon, evening, late night, or throughout the day).Only 20 participants (13.2 %) responded that their WhatsApp activity had occurred mostly in the night (after 9.30pm).As our primary conclusions did not change when we removed these participants, subsequent analyses utilised the full dataset.A schematic of the study procedures is presented in Figure 1.
Figure 1.Schematic of study procedures.All participants completed a baseline questionnaire, followed by 7 days of experience sampling where participants addressed questions about wellbeing and WhatsApp usage daily.Participants completed a final questionnaire one day after the experience sampling protocol ended.Next, we sought to profile participants based on their WhatsApp usage.We first applied Latent Profile Analysis (through the R package 'mclust'; Scrucca et al, 2016) to cluster participants based on their responses to the 7 quantitative variables, with values obtained by aggregating the reported frequency of each variable over the week.Latent profile analysis is based on conditional independence, where classes are created such that within each class, indicator variables are statistically uncorrelated (Wade et al., 2006).To uncover clusters of individuals, we used Gaussian Mixture Models (GMM) and assigned cluster membership using Bayesian probabilities.The final number of clusters was determined using the Bayesian Information Criterion, Integrated Likelihood Criterion, and a Bootstrap Likelihood Test.We then ran exploratory analyses to compare the characteristics of participants in each cluster (as described below).

Statistical analyses
Finally, we examined whether day-to-day variations in well-being tracked variations in WhatsApp usage.Given prior findings that the link between WhatsApp and well-being differed as a function of individual versus group chat activities (Blabst & Diefenbach, 2017), we created two predictor variables along this dimension.For the first predictor (daily personal chats), the following variables were summed within each day and for each participant: the number of (i) individuals COVID-19 messages were forwarded to; (ii) individuals from whom forwarded messages were received; and (iii) personal conversations discussing COVID-19.For the second predictor (daily group chats), the following variables were summed: the number of (i) groups participants forwarded COVID-19 messages to; and (ii) groups where COVID-19 messages were mentioned (Appendix A.1 shows the pattern of correlations across these variables).
To facilitate interpretation of the intercept in the analyses, we grand-mean centred scores on the daily personal chats and daily group chats variable by subtracting the mean number of chats across subjects and time points (M = 2.47 and M = 1.29 respectively) from each score.In addition, because daily personal chats/daily group chats vary both between-and within-subjects, we created between-subjects and within-subjects versions of both predictors to Tracking WhatsApp usage during COVID-19 tease these effects apart (Bolger & Laurenceau, 2013).For each outcome measure (fear and thoughts about COVID-19), we ran a linear mixed-effects model with time (centred such that 0 referred to the middle of the week), daily personal chats (between-subjects), daily personal chats (within-subjects), daily group chats (between-subjects) and daily group chats (withinsubjects) entered as fixed effects.Random intercepts accounted for correlated data due to repeated measures.Across all analyses, the type 1 decision-wise error rate was controlled at α = 0.05 (where applicable).All statistical analyses were conducted in R 3.5.0(R Core Team, Vienna, Austria) and SPSS 25 (IMB Corp, Armonk, NY).
Figure 3 shows the extent to which participants engaged in each of these activities, as captured by the quantitative variables (summed across the week).On average, participants received around two times (2.27) more messages than they forwarded on and were more likely to forward to individuals (average of 5.26 messages a week) than to groups (average of 2.68 messages a week).Beyond passive engagement, participants also took part in an average of 3.84 (SD = 5.58) one-to-one conversations about COVID-19 during the week; however, these interactions occurred 2.06 times less frequently than the sending or receiving of forwarded messages in group chats.Tracking WhatsApp usage during COVID-19 Figure 3.In a week-long experience sampling procedure, participants reported the extent to which they engaged in COVID-19 WhatsApp behaviours (either by forwarding or receiving messages, or in conversations).Horizontal bars represent the total amount of each activity captured, averaged across all participants.Horizontal lines represent the 95% confidence interval for the mean.

Towards a taxonomy of WhatsApp usage (latent profile analysis)
Although most participants did receive and/or share COVID-19 content on WhatsApp over the week monitored, there were individual differences in the quantity and manner of usage (Figure 3).Correspondingly, we conducted a latent profile analysis to understand how usage patterns clustered.Applying the Bayesian Information Criterion, 4 clusters provided the optimal fit of the GMM to the data; any further increase in the number of clusters did not result in significant likelihood increases.For sensitivity analyses, we used the Integrated Likelihood Criterion and obtained similar results.Finally, we performed a Bootstrap Likelihood Ratio Test comparing model fit between the k-1 and k cluster models and found again that 4 clusters provided the best model fit.Based on our analysis, the following taxonomy emerged (Figure 4): (i) 14% of participants were "chronic users" with high counts on each of the WhatsApp usage variables; (ii) 31% were "receiving users" who were recipients of forwarded messages and who had relatively high counts of group chat discussions; (iii) 30.5% were "discursive users" with relatively high st ed pp m for ii) ly Tracking WhatsApp usage during COVID-19 engagement in group and personal chat discussions; and (iv) 24.5% were "minimal users" with low counts across all WhatsApp usage variables.

Understanding profile characteristics
As an exploratory analysis, we performed a classification tree analysis to predict participant's classification based on: their demographics, COVID-19 concerns, depression and anxiety scores (DASS-21), and the time of day for WhatsApp usage (from the baseline questionnaire; see Table 1).For this analysis, we applied recursive partitioning ("rpart"), a machine learning technique that allows multiple variables to be analysed simultaneously while permitting complex, non-linear relations between predictors (Therneau et al., 2019).To avoid overfitting, the final tree was pruned by selecting the tree size with the lowest cross-validation error (minimized with a tree size of 8 for our dataset).
As shown in Figure 5, chronic users were more likely to be married / divorced, and to send messages either throughout the day or in the afternoon.In terms of crisis-related responses, they either had: (i) extreme fears of the COVID-19 situation (low or high), or (ii) had moderate fears paired with lower confidence in the government's response (low or moderate).
On the other hand, discursive users were more likely to be single / dating, and had either: (i) extreme levels of COVID-19 fears (either high or low), or (ii) moderate fear levels alongside Christian or Taoist affiliations.A sub-group of discursive users were -like chronic users -married / divorced, with moderate levels of COVID-19 fears.However, they were distinguished from chronic users by their high confidence in government (as compared to chronic users' lower confidence).

Does COVID-19 WhatsApp usage relate to well-being?
As the final thrust of the study, we ran linear mixed effects models to examine whether WhatsApp usage related to COVID-19 fears and thoughts (see Table 2 for parameter estimates, and Figure 6 for day-to-day variations in thoughts and fears for each of the four WhatsApp user types.).As shown Figure 6, although participants' COVID-19 related concerns and DASS-21 scores were fairly stable throughout the week (range of Spearman's r = 0.41 to 0.72; see Appendix B), day-to-day COVID-19 fears and thoughts fluctuated (t(249.13)= -3.72,p = 0.00 ; t(297.02)= -2.36,p = 0.02).
For COVID-19 thoughts, there was a significant effect of WhatsApp personal chat usage at a between-subjects level; t(164.81)= 2.36, p = 0.019).That is, participants who handled higher levels of COVID-19 content in their personal chats reported more COVID-19 thoughts (relative to participants who handled lower levels of COVID-19 content).However, the corresponding effect for group chats was not significant; t(141.17)= 0.89, p = 0.37.At the level of each participant, neither day-to-day fluctuations in personal nor group chat activities significantly predicted COVID-19 thoughts; smallest p = 0.68).The parameter estimates for the multi-level model are presented in Table 2.
For COVID-19 fears, we found no significant effect of any WhatsApp usage variable (smallest p = 0.17).For sensitivity analyses, we repeated the model with group membership as a fixed factor in place of personal and group chat usage, and our primary conclusions did not change (see Appendix C).Tracking WhatsApp usage during COVID-19

Discussion
In recent crises, messaging platforms have been used increasingly by members of the public (Bucher, 2020).Whether to mobilise responses, share information, or spread rumours, public use of these platforms have been found to influence crisis outcomes (Parmer et al, 2016;Guidry et al., 2017).Despite this influence, the private nature of messages means that very little research has investigated crisis communication through these platforms.Correspondingly, we report the first observational study documenting how messages are sent and received through the platform WhatsApp.Using the experience sampling method, we tracked one week of WhatsApp usage in situ -as they occurred in the naturalistic settings of participants experiencing the global COVID-19 crisis.
As our first finding, we observed that almost every participant engaged in "COVID-19 chatter" through WhatsApp -a finding that reinforces the central place of messenger applications during a crisis.Further, echoing WhatsApp's own concerns that message forwarding can problematic (e.g., contributing to an overwhelming volume of content shared; Singh, 2020a), we observed that participants were more likely to send or receive forwarded messages than they were to engage in conversation about COVID-19.Nonetheless, the spread of forwarded messages was capped in two ways: first, participants only forwarded half the messages they received; and second, participants were more likely to forward messages to individuals than to groups.

Socio-emotional vulnerability
Moving beyond group-level analyses, we found two possible ways by which individual differences in WhatsApp usage may be associated with vulnerability during a crisis.In terms of socio-emotional vulnerability, WhatsApp users who engaged with more COVID-19 content in personal chats were more likely to think about COVID-19 throughout the day.As similar forms of rumination (involving frequent and persistent thoughts) have been linked to clinical depression (Moulds et al., 2007), this finding implicates COVID-19 chatter as a potential risk factor for poor mental health.Outside the crisis setting, a prior cross-sectional survey similarly Tracking WhatsApp usage during COVID-19 observed that individuals with greater use of WhatsApp personal chats experienced more stress (Blabst and Diefenbach, 2017); thus, our study extended this finding to the crisis context.
One reason for our observed association may be that personal chats offer intimate and private connections (Blabst & Diefenbach, 2017;Karapanos, 2020).Amidst a crisis, this feature could encourage self-disclosure about crisis-related thoughts that a person already has (Knop et al., 2016).In turn, this may promote further exchange and internalisation of thoughts between two parties in a personal chat.In other words, we tentatively suggest a bi-directional and mutually reinforcing association between WhatsApp chat activity and COVID-19 thoughts.This is currently speculative, however, and further research is needed to establish causality.
More broadly, our findings align with a large body of literature that has linked social media usage with adverse mental health (Karim et al., 2020;Schønning et al., 2020;Sharma et al., 2020).This pattern has been observed both outside and inside the crisis context: for example, even in the current COVID-19 crisis, the frequency of social media usage has been found to predict depression symptoms (Ni et al., 2020).Our study extends the existing research by demonstrating that specific messaging activities (day-to-day exchanges on WhatsApp) is linked to well-being.

Cognitive vulnerability
Aside from the socio-emotional domain, we also identified potential cognitive vulnerability owing to WhatsApp usage during a crisis.Namely, the spread of online misinformation is currently recognized as one of the most pressing risks in the 21 st century (World Economic Forum, 2018), with an estimated cost of $78 billion to the global economy each year (CHEQ, 2019).Given that specific patterns of WhatsApp usage has been linked to either the exposure or spread of misinformation (e.g., receiving and dissemination of forwarded messages, citation), we were thus able to address -albeit indirectly -the question of 'who is vulnerable'.Using latent profile analysis, we identified four categories of WhatsApp users based on their usage patterns: the 'chronic user', the 'receiving user', the 'discursive user', and the 'minimal user'.Tracking WhatsApp usage during COVID-19 First, if sheer volume of WhatsApp usage accrues vulnerability (since this could increase both exposure to and the sharing of falsehood; Arun, 2019), then approximately 1 in 10 participants (14%) were particularly vulnerable -chronic users with high levels of activity in all seven metrics assessed.Our identification of the WhatsApp 'chronic user' is reminiscent of "supersharers" and "superconsumers" on Twitter (Grinberg et al, 2019) -a small sub-set of users that were linked to 80% of the misinformation read and shared during the 2016 U.S.
presidential election (Grinberg et al, 2019).Given the potential influence that these users may have in times of crisis, it may prove prudent to identify chronic users in risk communication strategies.We thus urge further research examining characteristics that may set this group apart, and a better understanding of reasons for their high WhatsApp usage.
A second vulnerability may lie in the receipt of forwarded messages.To the extent that forwarded messages are more likely to contain misinformation than user-generated content (Melo et al., 2019;Grinberg et al, 2019), 1 in 3 of our participants were 'receiving users' who were vulnerable in this manner.This group did not spread forwarded messages, and generally presented a moderate profile -whether in terms of fear of contracting COVID-19 (moderate levels reported), or in religion (coming from a diverse religious background).Nonetheless, receiving users had high exposure to forwarded COVID-19 content.
One notable feature of receiving users is that they tended to be older.In our study, receiving participants had a mean age of 41 years -compared to the mean ages of discursive and minimal users (30 and 34, respectively).Accordingly, the tree-based model identified older age -and in particular, a cut-off of 51 years -as a predictor that WhatsApp users would fall into this category.This finding is analogous to previous research examining how false news is disseminated on Facebook during an election year (Guess et al., 2019).In this study, older Americans were found to have shared three to four times as many false articles compared to those in the youngest age bracket, with this relationship holding even when covariates such as education and overall posting activity were controlled for.Extrapolating to the current context, it seems reasonable that older adults would likewise be on the receiving end of forwarded COVID-19 messages.Correspondingly, risk communication strategies may again benefit from targeted Tracking WhatsApp usage during COVID-19 messages to this group, this time emphasising how to discern the veracity of received messages.

Understanding individual differences in crisis-related WhatsApp usage
On a theoretical level, it is unclear at this stage why individuals differ in crisis-related WhatsApp usage.Without pre-crisis data, it is difficult to know whether our observed differences apply only to crises situations, or whether they describe WhatsApp usage patterns more generally.In the former case, differences may arise from how individuals cope with stress.For example, some may 'tend and befriend' in crises (Taylor, 2012)

Limitations
At this juncture, we note the preliminary nature of our findings and highlight several study limitations.First, although the experience sampling method affords more careful documentation of WhatsApp usage than what is typical in the literature (once-off surveys asking about usage; Hektner et al., 2007), our methodology still relied on participants' self-reports.
Correspondingly, future studies may profit from objective metrics of WhatsApp usage, or from measures to verify accuracy.Second, we opted to study WhatsApp, the most widely used messenger application.However, it is unclear whether our results generalise to other messenger applications (e.g., Facebook Messenger, Telegram).Tracking WhatsApp usage during COVID-19

Policy Implications
If supported by further studies, our results have implications for crisis management strategies.In terms of risk communication, the observation that nearly all participants (%) engaged in COVID-19 chatter highlights how WhatsApp has become a major source of crisisrelated information.This lends credence to official WhatsApp accounts that disseminate crisis content (e.g., that of the World Health Organization's during the COVID-19 crisis) (Liu & Tong, 2020), providing a clear voice alongside personal and group WhatsApp chats.
For public health agencies, our findings provide a proof-of-concept for how WhatsApp usage patterns can categorize individuals and predict their well-being during crises.In the field of digital phenotyping, similar smartphone usage metrics (e.g., calls placed and received) have been used to flag out persons at risk of poor mental health (Huckvale et al., 2019).Extending these efforts to WhatsApp, it may prove profitable to personalize mental health messages based on WhatsApp usage patterns -akin to how Facebook sends prompts for suicide helplines based on Facebook post analyses (Card, 2018).

Conclusions
In conclusion, we conducted the first week-long observation of how WhatsApp is used during a crisis.Using the experience sampling method, we captured 924 days of WhatsApp usage in situ (across 151 participants), tracked daily as usage occurred in participants' natural environments.This snapshot of "COVID-19 chatter" revealed: (i) the sheer prevalence of WhatsApp usage; (ii) a typology of WhatsApp users; and (iii) the possible link between WhatsApp patterns and well-being.These findings, we hope, will spur on further research on how messenger applications steer outcomes during a crisis.Tracking WhatsApp usage during COVID-19

Acknowledgements
This research was funded by a grant awarded to JCJL from the JY Pillay Global Asia Programme (Grant number: IG20-SG002).The first author's (EYQT) involvement was funded by a centre grant awarded to the Centre for Sleep and Cognition.Tracking WhatsApp usage during COVID-19

Figure 4 .
Figure 4. Using latent profile analysis, participants were classified based on how they had used WhatsApp for COVID-19 content during a week of monitoring.The figure depicts WhatsApp usage activities for the chronic user (top left), receiving user (top right), discursive user (bottom left), and minimal user (bottom right).Horizontal lines represent the 95% confidence interval for the mean.

Finally
Figure 5. Recursive partitioning was used to predict which of 4 WhatsApp usage profiles (chronic, receiving, discursive, or minimal) participants belonged to, based on baseline questionnaire measures (demographics, COVID-19 concerns, scores on the Depression, Anxiety and Stress Scale, and time of WhatsApp usage).The final tree model is presented as a flow chart, with factors chosen at each level to categorise the maximal number of participants.Marital status, time of WhatsApp usage, and age emerged as the primary predictors (model classification accuracy: 64.2%, above the chance level of 25%).s ne n, a ts.el

Figure 6 .
Figure 6.Day-to-day variations in COVID-19 thought (top) and fear levels (bottom), as a function of WhatsApp user profiles.The shaded grey areas represent 95% confidence intervals.
, seeking to affiliate with others through WhatsApp conversations (discursive users).Others, susceptible to information overload, may distance themselves from social media in the face of overwhelming crisis information (minimal users; Austin et al. (2012); Rathore & Farooq (2020)).Alternatively, our taxonomy may simply reflect personality traits more broadly -regardless of the crisis context.In a study of 180,000 participants, Bachrach et al. (2012) found that individuals more open to experiences published and 'liked' Facebook posts more often, whereas conscientious individuals published many photos but 'liked' fewer posts.Similar dynamics may predispose individuals to the four different types of WhatsApp usage we documented.Although we did not assess factors such as personality measures or coping styles, future research could examine whether and how the WhatsApp usage patterns differ according to these variables.

Table 1 .
Participant characteristics as a function of COVID-19 WhatsApp usage patterns.

Table 2 .
Parameter estimates for the multi-level model of thoughts about COVID-19 (Model 1) and fear of COVID-19 (Model 2) as a function of participants' daily WhatsApp use (personalchat and group-chat) [Model 1 ] Outcome: Thoughts about COVID-19