This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Formative Research, is properly cited. The complete bibliographic information, a link to the original publication on https://formative.jmir.org, as well as this copyright and license information must be included.
A cross-sectional study (Miyara et al, 2020) conducted by French researchers showed that the rate of current daily smoking was significantly lower in patients with COVID-19 than in the French general population, implying a potentially protective effect of smoking.
We aimed to examine the dissemination of the Miyara et al study among Twitter users and whether a shift in their attitudes toward smoking occurred after its publication as preprint on April 21, 2020.
Twitter posts were crawled between April 14 and May 4, 2020, by the Tweepy stream application programming interface, using a COVID-19–related keyword query. After filtering, the final 1929 tweets were classified into three groups: (1) tweets that were not related to the Miyara et al study before it was published, (2) tweets that were not related to Miyara et al study after it was published, and (3) tweets that were related to Miyara et al study after it was published. The attitudes toward smoking, as expressed in the tweets, were compared among the above three groups using multinomial logistic regression models in the statistical analysis software R (The R Foundation).
Temporal analysis showed a peak in the number of tweets discussing the results from the Miyara et al study right after its publication. Multinomial logistic regression models on sentiment scores showed that the proportion of negative attitudes toward smoking in tweets related to the Miyara et al study after it was published (17.07%) was significantly lower than the proportion in tweets that were not related to the Miyara et al study, either before (44/126, 34.9%;
The public’s attitude toward smoking shifted in a positive direction after the Miyara et al study found a lower incidence of COVID-19 cases among daily smokers.
COVID-19 is caused by SARS-CoV-2 [
Researchers in France—Miyara et al—conducted a cross-sectional study on patients infected with COVID-19 in a large French university hospital, which was placed on a preprint server on April 21, 2020 [
Twitter, a
A previous Twitter study on COVID-19 and smoking—the only other Twitter analysis on this topic, to our knowledge—showed that preprints suggesting the benefits of smoking might increase reactions to tweets on tobacco products and the virus [
The related tweets (ie, Twitter posts) posted from April 14 to May 4, 2020, were crawled by the Tweepy stream application programming interface using keyword queries with COVID-19–related keywords, including “CORONA,” “corona,” “COVID19,” “covid19,” “covid,” “coronavirus,” “Coronavirus,” “CoronaVirus,” and “NCOV.” The analysis period was chosen due to the nature of our study objective. Because we were interested in the change in sentiment before and after publication of the paper, we analyzed tweets that were posted immediately before and after the day it was published. Next, retweets without comments were deleted, since simple retweets typically do not explicitly reflect personal opinions; the behavior of retweeting can mean supportive, oppositional, or neutral attitudes toward the original tweet. Repetitive tweets were also removed from the collected data set, as the majority were copied news headlines without personal sentiments. Afterward, research- and tobacco-related tweets were filtered out in sequence using keyword matching: we first filtered research-related tweets using “study” and “research,” then tobacco-related tweets were filtered using “smok,” “cigarette,” “tobacco,” “nicotine,” and “ace2.” Tweets discussing studies without clear findings were removed as those having no impact in shifting people’s opinions. Finally, 1929 tobacco- and research-related tweets remained and discussed, in some way, the effects of smoking on COVID-19 infections and symptom development.
Data set preparation procedures.
Three categories were used to categorize each tweet: (1) whether it was about the Miyara et al study (yes or no), (2) the article’s attitude toward smoking (positive, negative, or neutral), and (3) the user’s attitude toward smoking as expressed in the tweet (positive, negative, or neutral). For the first category, whether or not each tweet was related to the published Miyara et al study was manually coded for all the selected tweets. For the second category, the article’s attitude toward smoking as discussed in the tweet was manually coded as positive, neutral, or negative. For example, if the tweet discussed an article that found smokers were at more risk for COVID-19, it was considered negative. For the third category, each tweet was manually coded as positive, neutral, or negative based on the user’s attitude toward smoking. For example, if the user said, “smoking is bad for you,” the tweet was categorized as negative. The term
Two coders from the study team coded each tweet separately and disagreements were discussed among team members to achieve final agreements. High agreements were achieved between the two independent coders on coding the 1931 selected tweets into different categories. The Cohen κ value was 0.92 (95% CI 0.90-0.94) for categorizing whether the tweet discussed the French study. Regarding the article’s attitude toward smoking, the Cohen κ was 0.93 (95% CI 0.92-0.95) for categorizing the attitude into positive, neutral, and negative groups, indicating very high agreement. The Cohen κ was also very high when categorizing the tweet’s attitude toward smoking into positive, neutral, and negative groups, with almost perfect agreement (κ=0.86, 95% CI 0.85-0.88).
According to whether the tweet was related to the Miyara et al study and whether the tweet was posted before or after the publication date of the Miyara et al study (ie, April 21, 2020), all selected tweets were classified into three groups: (1) tweets not related to the Miyara et al study before publication, (2) tweets not related to the Miyara et al study after publication, and (3) tweets related to the Miyara et al study after publication. The tweets’ attitudes toward smoking and the articles’ attitudes toward smoking were compared among the above three groups using multinomial logistic regression models through the
In order to capture certain themes that were prevalent within the tweets, the two members of the study team that hand-coded the tweets also created topics based on the content they read. The tweets were separated into two basic categories to allow for a more efficient comparison of themes: tweets that were not about the Miyara et al study and tweets that were about the Miyara et al study. Different focuses were adopted during the theme-capturing process, considering the fact that people react differently within these two groups. Specifically, themes from tweets about the Miyara et al study were mainly people’s attitudes toward the research itself or speculations about unseen driving forces. Contrary to a comparably narrow but concentrated scope, tweets that were not related to the Miyara et al study discussed diverse aspects of the field. For example, users showed sentiments beyond attitudes toward smoking, such as generally distrusting research, stating reasons to support their stances, and requesting information for truth finding. Such diversity has also been observed through various research directions that analyzed smokers’ risks of COVID-19 infections, which include but are not limited to the analysis of existing health conditions and harmful life habits. To obtain a comprehensive understanding of these themes, which are not necessarily correlated with each other, multiple groups are, thus, needed for illustrations. As shown in the Results section below (Tweets Related to the Miyara et al Study subsection), groups defined as
The temporal analysis of the Miyara et al study was done by compiling the hand-coding results after the sentiment analysis. A clear comparison between numbers of tweets that were about the Miyara et al study versus those that were not about the Miyara et al study can be drawn from
Temporal analysis of the Miyara et al (2020) study’s influence.
Final tweet classification into different categories.
Proportion of negative, neutral, and positive tweets regarding their attitudes toward smoking in three different tweet groups: (1) tweets not related to the Miyara et al study (Non-Miyara et al) before April 21, 2020; (2) tweets not related to the Miyara et al study (Non-Miyara et al) after April 21, 2020; and (3) tweets related to the Miyara et al study.
For the 324 tweets that were not related to the Miyara et al study, 126 (38.9%) tweets were posted before April 21, 2020, and 198 tweets (61.1%) were posted on April 21 or after April 21, 2020. The 324 tweets were categorized into three groups of perspectives to analyze users’ opinions, various sentiments, and the research focus of the studies mentioned:
Topic categories for tweets not related to the Miyara et al study.
Group and topics | Tweets before April 21, 2020 (n=126), n (%) | Tweets after April 21, 2020 (n=198), n (%) | Example tweetsa | ||||
|
|||||||
|
Stating the finding | 58 (46.0) | 97 (49.0) | “shocking result smokers are far less likely to be hospitalized with coronavirus than non smokers” |
|||
|
Discourage tobacco | 39 (31.0) | 67 (33.8) | “this study is simple and others published since have made it pretty clear smokers die more from covid than nonsmokers” |
|||
|
Encourage tobacco | 13 (10.3) | 23 (11.6) | “bibber indepth if you split current and former smokers current smokers will end up with a lower risk for hospitalization than or 0 7 and former smokers with a higher or there is no misleading this is a well done study of 4 103 corona patients in a nyc health system” |
|||
|
|||||||
|
Advocate quitting smoking | 13 (10.3) | 7 (3.5) | “a new study shows that cigarettes can help the coronavirus enter lung cells meaning it’s time to stop smoking” |
|||
|
More info requested | 12 (9.5) | 5 (2.5) | “are there any official recording of those who smoke are more likely to suffer from covid19” |
|||
|
Debate with others | 4 (3.2) | 5 (2.5) | “kwaza really please reference one scientific study that proves that smoking decreases your chance of surviving a covid19 infection specifically a study by the cdc found that just over 1 of those who died from a covid19 infection were smokers and just over 2 were previous smokers” |
|||
|
General distrust in research | 3 (2.4) | 5 (2.5) | “such a musical chair this research on covid19 has become smokers will be at higher risk then smokers have a better chance of surviving this is when you shut it all off fda says smokers may have higher risk of catching covid 19” |
|||
|
|||||||
|
Smoking itself | 97 (77.0) | 185 (90.9) | “smoking does not increase covid 19 susceptibility” |
|||
|
Health conditions | 18 (14.3) | 3 (1.5) | “only half of urban and rural pakistanis are aware of the fact that diabetics smokers and asthmatics are at a higher risk of falling severely ill from covid 19 study by aga khan university aku coronavirus” |
|||
|
ACE2b related | 11 (8.7) | 12 (5.6) | “icilondres one theory is that nicotin blocks ace2 receptors the backdoor to the lungs for covid19 however research needs to be done before we all reach out to our patches” |
aDuring preprocessing and before hand-coding, punctuation and capitalization were removed from the tweets to facilitate content analysis.
bACE2: angiotensin-converting enzyme 2.
While manually coding the 1605 tweets about the Miyara et al study, certain common themes arose: encouraging tobacco use, expressing feelings of surprise about the findings of the Miyara et al study, stating the main finding of the Miyara et al study, discussing the follow-up nicotine patch study, highlighting the negative consequences of tobacco use, and questioning whether or not the Miyara et al study was funded by the tobacco industry.
Topic categories for tweets related to the Miyara et al study.
Group and topics | Tweets (n=1605), n (%) | Example tweetsa | |
|
|||
|
Encouraging tobacco use | 51 (3.18) | “that's hilarious so nicotine is actually good for something lol smoke em if you got em lol study finds smoking reduces chance of getting coronavirus symptoms” |
Stating the finding | 643 (40.06) | “a cross sectional study strongly suggests that those who smoke every day are much less likely to develop a symptomatic or severe infection with covid 19 compared with the general population” |
|
|
|||
|
Shocked or surprised | 28 (1.74) | “shocking study supports smoking as preventive measure against covid19” |
Tobacco consequences highlighted | 24 (1.50) | “french research suggests nicotine could protect against covid 19 but smoking remains biggest killer in france 75 000 people die every year from smoking related complications” |
|
Tobacco industry funded | 34 (2.12) | “this is insane to even suggest using substance such nicotine less likely to catch coronavirus is it tobacco companies is paying for this study” |
|
Research focus in the future: |
497 (30.97) | “french researchers to test nicotine patches on coronavirus patients” |
|
Unsorted: miscellaneous | 328 (20.44) | “treat this research with caution it was my solid understanding that this virus affects smokers far worse than a non smoker as the lungs are already compromised from damage by smoking” |
aDuring preprocessing and before hand-coding, punctuation and capitalization were removed from the tweets to facilitate content analysis.
By topic comparison, both similarities and differences were seen. Regarding consistency, topics found within tweets related to the Miyara et al study and those found in tweets not related to the study showed that the majority simply stated the finding of the study; 643 out of 1605 (40.06%) tweets related to the Miyara et al study belonged to this category, compared to 155 out of 324 (47.8%) tweets not related to the study. However, while both groups contained tweets that held negative attitudes toward smoking, tweets related to the Miyara et al study discouraged tobacco use in a more comprehensive way. For example, within the 1605 tweets related to the Miyara et al study, 24 (1.50%) highlighted the risks of tobacco use and 34 (2.12%) speculated whether the study was funded by the tobacco industry. Within the 198 tweets that were not related to the study, 67 (33.8%) discouraged tobacco use, while 23 (11.6%) encouraged tobacco use. Furthermore, there were tweets that demonstrated doubts regarding the potential benefits of smoking, but many did not absolutely reject such possibilities. This suggests that users may have been open to exploring whether there was a positive effect of smoking on COVID-19 but that they proceeded with caution. Before the publication of the Miyara et al study, such rejections were frequently observed, which indicates that most users shared a neutral or more negative opinion of smoking’s influence on COVID-19 and suggests that users may have begun to think more critically about smoking’s impact on the novel coronavirus after the Miyara et al study was published. The distributions of the number of followers of Twitter users who posted either tweets related to the Miyara et al study or tweets not related to the study were both highly skewed to the left. The median number of followers of Twitter users who posted tweets related to the Miyara et al study was 585 (IQR 3407). The median number of followers of Twitter users who posted tweets that were not related to the study was 630 (IQR 2681). The number of followers of Twitter users indicated the bandwidth of outreach of those posted tweets.
This report presents novel findings observing a shift in attitudes toward smoking among Twitter users after publication of a Miyara et al study that reported lower rates of daily smoking among COVID-19 cases. The relatively large number of median followers of Twitter users who posted tweets related to the Miyara et al study indicated that the results of the Miyara et al study were widely disseminated. Overall, the findings suggest that this study was successfully disseminated and appears to have led to more positive attitudes toward smoking among our population. Every post that stated the main finding of the Miyara et al study, which was considered a neutral sentiment, can be looked at as an instance of spread of information from one user to other users, leading to even greater spread of the study. When comparing tweet sentiments before and after April 21, 2020, there was a significantly more positive attitude toward smoking among all tweets.
The largest percentage of tweets, from before and after April 21, 2020, and that were about the Miyara et al study, had a neutral sentiment toward smoking, including mentions of wanting more information and a sense of uncertainty regarding the study’s findings. The most prevalent tweets stated the main finding of the study and/or mentioned the nicotine patch follow-up study. The main findings of the articles tweeted about in this sample—those that were not about the Miyara et al study—reflected both the benefits and risks of smoking on COVID-19 [
Nevertheless, differences in people’s opinions and study focuses were observed in some ways. Even among posts not directly discussing the Miyara et al study, there was still a shift to a more positive attitude toward smoking after its publication. Not necessarily shifting all others’ attitudes toward smoking from negative to positive, the release of Miyara et al study did at least waver the stances of those who opposed smoking, confirming its positive impact. This speculation was supported by a 6.82% weight decrease of tweets that were not related to Miyara et al study that advocated quitting smoking. To conclude, a different reaction pattern was seen between tweets within French and non-French groups. Beyond the study’s influence of changing people’s attitudes toward smoking, it also informs a critical thinking mindset behind how people observe the effects of smoking. For example, while people’s stances varied regarding the effect of smoking on COVID-19, more tweets (10.3%) were observed that called on people to quit smoking before April 21, 2020, compared to after (3.5%). This decrease was perhaps due to the impact of the Miyara et al study, which highlighted a potential benefit of smoking.
Significant changes were also seen regarding how people analyzed the impacts of smoking on COVID-19. Before April 21, 2020, many tweets (14.3%) discussed the increased risks of infection if smokers had existing health conditions, such as diabetes and asthma [
Twitter is a valuable tool used in health research and can be used to analyze up-to-date data about a specific topic while it is at its peak discussion point [
Although the results met our original expectations of the impact of the Miyara et al study on people’s opinion changes, several limitations of our analysis can be found. Firstly, some posts might be missing, since keywords were used to filter out tweets before hand-coding. The potential problems of this are whether or not all posts about our topic contained the chosen keywords. For example, the keywords “study” and “research” were used to filter out research-related posts; however, users might use words like “result” and “report,” among other words, to reflect a research finding. In our case, “result” and “report” would bring in a lot of off-topic tweets; thus, those were not included in the keyword list. Nevertheless, an alternative analysis could consider adding those tweets and starting the filtering process afterward. Similar problems might be relevant if some users comment on a research result without referring to the subject, in which case users’ sentiments would still be related to our analysis but would be ignored. Secondly, we analyzed tweets written only in English and missed tweets in other languages, which might bias the study results. However, analyzing tweets written only in English could avoid misinterpreting translations. Thirdly, we did not collect and analyze the numbers of likes and retweets of the original posts, which could also help to analyze the impact of those tweets. Furthermore, although intercoder variances were small enough to make the results valid, hand-coding is a subjective method, implying potential cognitive differences in coding. Tweets were posted globally, which suggests that tones from distinct regions might imply different meanings (eg, sarcasm). Though hard to eliminate, such limitations could be alleviated by incorporating more coders with diverse backgrounds. Lastly, it is possible that smokers tended to tweet more about smoking-related studies. They may even have had positive attitudes toward smoking before seeing any studies demonstrating smoking’s beneficial effects. To address that, studies quantifying such likelihood may be needed and should be taken into consideration during analysis to more precisely observe the impact of the Miyara et al study.
Note that this paper is not a criticism of the Miyara et al study’s authors or their research—our focus here is on using that paper and its findings as a jumping-off point for exploring how a particular study is disseminated on Twitter, and how that information may influence the sentiment of tweets moving forward. Information-based communication strategies can be used to modify people’s attitudes by providing evidence for or against an idea. Previous literature has described the impact of research and research dissemination as affecting knowledge, attitudes, and behavior with respect to health risks [
angiotensin-converting enzyme 2
US Food and Drug Administration
National Institutes of Health
Research reported in this publication was supported by the National Cancer Institute of the National Institutes of Health (NIH) and the US Food and Drug Administration (FDA) Center for Tobacco Products under Award Number U54CA228110. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH or the FDA.
ZX, DL, and ROC conceived and designed the study. ZX, CT, and LC extracted and preprocessed the data. CT, DD, and DL analyzed the data. CT, DD, ZX, LC, DL, and ROC wrote the manuscript, assisted with interpretation of the analyses and results, and edited the manuscript.
None declared.