We excluded retweets and addresses to focus solely on personal opinions or statements. First, the collected tweets were preprocessed by removing stop words as well as keywords with IDs and hashtags. Second, tweets were represented using unigrams, bigrams, and trigrams, such as “results,” “lab results,” and “check test results.” Word-level n-grams features were selected to represent tweets instead of the bag-of-words (ie, single words) feature because the latter has two major drawbacks: (1) they lose the ordering of the words and (2) they ignore semantics of the words [33,34].

To discover the abstract “topics” that occur in the collected posts, we ran a topic mining model, specifically the LDA algorithm, with 50 topics. Given a set of documents, D = {d₁, d₂, … , d_n}; a number of topics, T = {t₁, t₂, … , t_m}; and a number of words in each topic, W = {w₁, w₂, … , w_k}; the LDA algorithm generates the following:

The corresponding reproductive process is shown below [35,36]:

βt is the word distribution for topic t, and θd is the topic distribution for document d. The notations η and α are model parameters.

Topic models are statistical based models for uncovering the main themes (ie, set of topics) that depict a large and unstructured collection of documents. Topic models make it possible to summarize textual data at a scale that cannot possibly be tackled by human annotation. In this study, we chose the LDA algorithm [25] owing to its conceptual advantage over other latent topic models [35-38].

The 50 topics from the LDA were labeled by first author and validated by second author. The identified topics were further analyzed and grouped into 10 representative categories. The grouping was done based on semantic similarities between the topics identified. For example, the topics “build herd immunity,” “herd Immunity,” and “build immune system” could be grouped into in one main topic, namely, “herd immunity and dependency on the immune system.” Overall, we discovered and collected 10 different categories.

Brandwatch employs ReadMe, a supervised algorithm developed by Hopkins and King [26]. The algorithm is particularly suited when the objective is to know the proportion of all posts that fit in specific categories. Rather than calculating these proportions based on the categorization of individual posts, ReadMe gives approximately unbiased estimates of category proportions even when the optimal classifier performs poorly [26].

The ReadMe algorithm requires the researcher to hand-code a “training set” of documents into a set of predefined categories. In this study, the tweets represent the set of documents and the predefined categories are obtained using the LDA algorithms. The authors hand-coded 20 tweets into each predefined category obtained from the LDA and then ran the ReadMe algorithm iteratively on the remaining posts, ensuring that the examples clearly outline each category. Then, based on the training phase, the algorithm builds a model that can automatically assign the remaining tweets into categories and obtain the total number of tweets in each category. Brandwatch automatically generates the trends of tweet volumes over time.

To analyze the relationship between the volume of tweets against mask-wearing and the daily volume of new COVID-19 cases, we plotted two time-series over the time span from January to October 2020 and calculated the Pearson correlation coefficient, which measures how two continuous waves co-vary over time and indicate the linear relationship as a number ranging from –1 (negatively correlated) to 0 (not correlated) to 1 (perfectly correlated) [39]. The correlation is a snapshot measure of global synchrony. Although the Pearson correlation coefficient provides a very simple way to compute both global and local synchrony, it does not provide insights into signal dynamics such as which signal occurs first or which can be measured via cross-correlations. A TLCC can identify directionality between two signals such as a leader-follower relationship. We can get a sense of which signal occurs first by looking at cross-correlations. A TLCC is measured by incrementally shifting one time-series vector and repeatedly calculating the correlation between two signals. If the peak correlation is at the center (offset=0), this indicates that the two time-series are perfectly synchronized at that time. However, the peak correlation may have a different offset if one signal leads another [40]. To analyze the relationship between the two time-series, the volume of tweets against mask-wearing, and the daily volume of new COVID-19 cases, we calculated the Pearson correlation coefficient and TLCC in Python using the SciPy package.

Figure 6 depicts the volume of tweets against wearing masks and the number of newly reported COVID-19 cases over the study period. The two time-series exhibit a high positive Pearson correlation (r=0.77). Since information about directionality between the two waves—leading and following—cannot be interpreted solely from this data, we further studied the relationship between both waves (Figure 7). Overall, the results show a 9-day lead for tweet volume over the number of new COVID-19 cases. This 9-day lag is considered comparable to the number of days after which people can develop COVID-19 symptoms. According to a previous study, approximately, 97% of people infected with COVID-19 developed symptoms within 12 days after exposure [43].

This study analyzed the negative stance regarding masks on social media, the specific themes within this discourse, and how this discourse could be associated with the prevalence of new COVID-19 cases. The study reported Twitter users’ concerns related to constitutional rights and freedom of choice, conspiracy theory, misinformation, health issues, fearmongering, and other concerns related to the use of face masks during the COVID-19 pandemic. Furthermore, the time-series analysis demonstrated a strong correlation between the number of tweets posted against mask-wearing and the actual number of COVID-19 new cases, with the volume of negative tweets leading the number of newly reported COVID-19 cases by 9 days.

The study findings emphasize the potential relationship between social media behavior and its manifestation in the physical world. Such findings highlight the importance of listening to social media and proactively reacting to public perception in fighting COVID-19. Lyu and Wehby [44] showed that mask mandates in a number of states were associated with lowering infection rates by 0.9%-2% after wearing masks for 1-21 days. However, when the government mandates mask-wearing in public, many people feel their constitutional rights and freedom of choice are being violated [45]. As a result, there is a need to increase awareness about the fact that wearing masks can protect others from contracting COVID-19 even though they do not fully protect the person wearing the mask from the infection [46]. The government should also address the challenges faced by implementing a balanced mask-wearing mandate that considers protecting people's lives while also protecting their freedom of choice [47].

Social media platforms have been used to spread fake news, lies, and conspiracy theories, all of which have a strong impact on people and society [48]. As a result of such an impact, the public is less likely view actions like wearing masks as a necessity to mitigate the spread of the virus during a pandemic [49]. Therefore, it is crucial that, as we seek to control the spread of COVID-19 and future viruses, we develop policies to fight against misleading and damaging conspiracy rhetoric. Similarly, there should be policies in place to combat fake news, lies, and misinformation, especially on social media, which could negatively affect the public’s trust in science [49].

Health care professionals should actively engage in the conversation with the public in order to discuss scientific evidence supporting the importance of wearing a mask and debunk rumors on social media that promotes discussions related to masks causing low oxygen levels or lung infections. They should also discuss evidence and guidelines such as “wearing a mask does not raise the carbon dioxide (CO₂) level in the air you breathe” [50] and “people aged 2 and older should wear masks in public settings and when around people who don’t live in their household” [50] to increase awareness regarding the effectiveness of masks in protecting the wearer from inhaling and spreading airborne particles.

Children of specific age groups should be encouraged to wear masks to protect them from COVID-19. However, protecting these age groups only by using a mask could prove very difficult [51]. To overcome these challenges, there is a need to advocate for parental involvement and support for the initiatives aimed at increasing mask-wearing among children [51]. Children should be encouraged to “take off their masks to breathe in fresh air after wearing masks for a certain amount of time,” and they should not wear masks in certain cases, such as while exercising [51]. In the case of noncompliance, it would be a better option for children to not wear masks and follow other measures to reduce infection risk and remain at home [51].

Following an empathetic approach to motivate people to wear masks and adhere to physical distancing could be an effective alternative [52] for fearmongering that focuses only on presenting a negative picture of the COVID-19 pandemic [53]. In addition, policy makers could use positive messaging to curb the spread of fear while still maintaining a transparent and accurate depiction of the situation [53].

With physical, mental, social, and economic burdens imposed by the pandemic, many populations may experience increased suicide risk [54]. Furthermore, the prevalence of anxiety, depression, posttraumatic stress disorder, and stress was reported to have increased in a number of countries during the COVID-19 pandemic [55]. Data analysis and event surveillance conducted during the first 6 months of the pandemic have shown impacts on suicide risk [54]. As a result, knowing the facts about masks and containing the spread of rumors can reduce stress and the adverse impacts on people’s mental health [56]. Finally, since many people believe that herd immunity is the best solution to this public health crisis and to strengthen their immune systems, a scientific and fact-driven view should be shared with the public explaining why herd immunity is not an ideal solution as has been reported by many researchers [57].

By carefully analyzing social media posts, policy and decision makers are in a better position to tailor public health awareness campaigns to respond to specific themes and thereby improve their effectiveness in a crisis situation such as the COVID-19 pandemic. Thus, exploring the categories of tweets surrounding the topic of mask-wearing during the COVID-19 pandemic may help reveal a number of insights that could help better design and implement awareness campaigns.

This study has some limitations that could be addressed in future research. First, although we identified a very strong correlation between the increase in the volume of tweets against wearing masks and the rise in the number of COVID-19 cases, we cannot claim causality, as the rise in COVID-19 cases could be attributed to population density, government-enforced lockdown restrictions, and other factors that are beyond the scope of this study. Second, the study focused on analyzing English tweets in the United States. Future studies need to address and compare the public discourse on masks across different social media platforms and in different countries. Third, given the number of tweets collected and the focus on Twitter as a data source, the public discourse might not reflect the actual public opinion against masks. According to Wojcik and Hughes [58], Twitter has been found to have much younger audiences, with the most prolific 10% of users creating 80% of all tweets published. Finally, we did not separately analyze the opinions of Twitter users against masks in the early and later stages of the pandemic. Such analysis could unmask other important trends that are not discussed in this paper.

In this study, we analyzed tweets against wearing masks on social media to understand topics, insights, and information about user-reported issues. We used data analytics to identify trending themes and topics of concern by the public about wearing face masks. The most discussed issues were related to the constitutional rights and the freedom of choice, conspiracy theory, misinformation, health issues, fearmongering, and the ineffectiveness of masks, followed by issues related to mental health, herd immunity, child abuse, and virus-related statistics. Another key finding of this study is that it highlights the strong correlation between the increase in the volume of tweets against wearing masks and new COVID-19 cases and the lead of negative tweets published in comparison with the rise in new COVID-19 cases in the time-series analysis. In effect, these findings demonstrated the impact of social media not only on people’s opinion or perceptions about public topics but also the potential impact on real events such as changing the course of the pandemic. The significance and implication of this research transcends the COVID-19 pandemic, as it demonstrates the importance of social media mining and its potential to support public health–related policies and decisions. Government officials and decision makers could tailor and fine-tune public awareness campaigns and prioritize policy interventions toward the most discussed topics. In case of a future massive-scale health crisis such as the COVID-19 pandemic, government officials and policy makers could leverage social media analytics and surveillance as important tools in proactively responding to the impending crisis. Policy makers need to proactively address public perception and work on shaping this perception through raising awareness, debunking negative sentiments, and adopting early policy intervention to steer the wheel towards public acceptance of more precautionary measures and thereby containing the situation.