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 Infodemiology, is properly cited. The complete bibliographic information, a link to the original publication on https://infodemiology.jmir.org/, as well as this copyright and license information must be included.
As of May 9, 2021, the United States had 32.7 million confirmed cases of COVID-19 (20.7% of confirmed cases worldwide) and 580,000 deaths (17.7% of deaths worldwide). Early on in the pandemic, widespread social, financial, and mental insecurities led to extreme and irrational coping behaviors, such as panic buying. However, despite the consistent spread of COVID-19 transmission, the public began to violate public safety measures as the pandemic got worse.
In this work, we examine the effect of fear-inducing news articles on people’s expression of anxiety on Twitter. Additionally, we investigate desensitization to fear-inducing health news over time, despite the steadily rising COVID-19 death toll.
This study examined the anxiety levels in news articles (n=1465) and corresponding user tweets containing “COVID,” “COVID-19,” “pandemic,” and “coronavirus” over 11 months, then correlated that information with the death toll of COVID-19 in the United States.
Overall, tweets that shared links to anxious articles were more likely to be anxious (odds ratio [OR] 2.65, 95% CI 1.58-4.43,
The level of anxiety in users’ tweets increased sharply in response to article anxiety early on in the COVID-19 pandemic, but as the casualty count climbed, news articles seemingly lost their ability to elicit anxiety among readers. Desensitization offers an explanation for why the increased threat is not eliciting widespread behavioral compliance with guidance from public health officials. This work investigated how individuals' emotional reactions to news of the COVID-19 pandemic manifest as the death toll increases. Findings suggest individuals became desensitized to the increased COVID-19 threat and their emotional responses were blunted over time.
The COVID-19 outbreak has spread worldwide, affecting most countries. Since the outbreak of COVID-19, the number of confirmed cases and the death toll have steadily risen. According to Johns Hopkins University, as of May 9, 2021, more than 157.9 million cases of COVID-19 and 3.2 million deaths have been reported worldwide [
Extreme safety precautions (eg, statewide lockdowns, travel bans) have impacted individuals’ physical and mental health in the United States. People experienced intense psychological frustration and anxiety regarding the virus and strict safety measures (eg, stay-at-home measures), especially during the early stages of the COVID-19 pandemic [
From such observations, two key considerations arise. First, fear-eliciting health messages have a significant effect on eliciting motivation to take action to control the threat. However, repeated exposure to these messages over long periods results in desensitization to those stimuli. In this work, we examine the effect of fear-inducing news articles on people’s expression of anxiety on Twitter. Additionally, we investigate how people are desensitized by fear-inducing news articles over time, despite the steadily rising COVID-19 death toll.
The current pandemic has fueled rapidly evolving news cycles and shaped public sentiment [
In the context of the COVID-19 pandemic, the efficacy of fear-inducing messages on behavioral compliance with public health officials is consequential. Reports of increased COVID-19 transmission and the rising death toll may elicit anxiety about the virus, consequently motivating behaviors intended to manage the problem. For instance, a national survey examining the mental health consequences of COVID-19 fear among US adults during March 2020 found that respondents generally expressed moderate to high COVID-19 fear and anxiety (7 on a scale of 10), and increased anxiety was most prevalent in areas with the highest reported COVID-19 cases [
Although fear-based health messages have been shown to motivate changes in behavior, repeated exposure to even highly arousing stimuli—such as news of the rising death toll from COVID-19—may eventually result in desensitization to those stimuli [
Numerous studies have demonstrated desensitization to media content. Research has often focused on fictional depictions of violence [
Researchers studying social media data have explored the possibility that news messages can result in desensitization. Li and colleagues [
In the context of the COVID-19 pandemic, news audiences have been repeatedly exposed to highly arousing messages related to COVID-19–related deaths—messages that inherently communicate explicit and implicit threats of serious illness and death to readers. Fundamentally, the biological response to threat communicated through text is similar to threats communicated in other ways [
The public relies heavily on news disseminated through social media for information about the spread of the virus [
The general hypothesis guiding this research is that audiences will have become desensitized to COVID-19 deaths over the course of the pandemic, decreasing the level of anxiety elicited by fearful COVID-19 health information reported in the news. To the best of our knowledge, ours is the first study to investigate whether, as the objective threat and harm of COVID-19 has increased, individuals have become desensitized to news reports of cautionary COVID-19 health information.
This study examined how anxiety levels in news articles predicted users’ tweet anxiety levels over 11 months, then correlated that information with the total death toll of COVID-19 in the United States as reported to the Centers for Disease Control and Prevention (CDC) on the day the post was shared [
The sample comprises content shared to Twitter, a popular social media platform used for sharing news [
The Python programming language was used to extract posts sharing news reports of COVID-19 health information. We collected a quota sample of 32,000 US tweets containing one of four key terms (ie, COVID, COVID-19, coronavirus, pandemic) each week from January 1 to December 2, 2020. The GetOldTweets3 Python3 library was used to scrape tweets for the months of January-July 2020 [
Human coders then filtered through the sample of 1,410,901 tweets to randomly extract a quota of 8 original tweets per key term from each week sharing a news report about COVID-19. Data collection resulted in thousands of tweets containing links per week. To facilitate the representativeness of the news articles, 32 tweets were drawn from each week from a shuffled list of tweets containing hyperlinks. Since we aimed to assess users’ reactions to the text of the article they read, without the confounding textual framing of other peoples’ commentary about an article, retweets were excluded from the analysis. If a quota of 32 tweets each week (8 per key term) was not met, additional tweets were sampled for that week. Notably, the disease and pandemic were not commonly referred to as COVID-19 in early January; accordingly, three weeks did not have 8 tweets with the terms “COVID” and “COVID-19” per week.
The news articles were collected from links shared by Twitter users in general, regardless of who posted the tweet. We only included users sharing links to news articles regarding COVID-19 in the United States; all other content was excluded (eg, news about the rock band Pandemic Fever). If all posts for that week were excluded, another sample from that week was drawn. If a tweet linked to a news article that had been taken down, a replacement post was sampled from the same week. We then extracted the text from the news articles and their corresponding tweets. The final sample was comprised of n=1465 news-sharing tweets.
Flowchart of data collection process.
Once the final sample was collected (n=1465), we analyzed articles and tweets using the Linguistic Inquiry and Word Count (LIWC) program [
Sample text from a COVID-19 news article shared to Twitter [
We paired the final sample with the CDC’s aggregate death toll on the day the tweet was posted. Contextualizing the articles and tweets allowed us to examine how fear elicited by COVID-19 health news manifests as individuals become desensitized to news of COVID-19–related deaths.
The outcome of interest was tweet anxiety. Note that the distribution of count data outcome variables (in our case, LIWC tweet anxiety) often contains excess zeros; this result is known as zero inflation. The positive values are skewed, and a considerable “clumping at zero” is trailed by a bump representing positive values [
Two distinct distributions generally characterize zero-inflation data; thus, a zero-inflated model, which separates the zero and nonzero counts, is appropriate [
We employed a zero-inflated model using a gamma distribution with a log link to examine any association between article anxiety and death toll, along with their interaction with subsequent tweet anxiety for all values of tweet anxiety greater than zero. We paired that with a model that used a binomial distribution with a logit link to determine zero anxiety versus nonzero anxiety in tweets. We recoded the death toll into categories reflecting the death count at the second quartile, the third quartile, and the fourth quartile relative to the first quartile of the total death count (see
Distribution of death toll quartiles over time.
This study only used information available in the public domain. No personally identifiable information was included in this study. Ethical review and approval was not required for this study because the institutional review board recognizes that the analysis of publicly available data does not constitute human subjects research.
Results suggest that as the death toll increased over time, the baseline level of anxiety lexicon words in articles decreased; this was evidenced by our finding that when the pandemic’s severity and threat increased, individuals shared less news coverage containing COVID-19 anxiety words (eg, “risk,” “worried,” “threatens”). When assessing the odds of a tweet having no anxiety versus anxiety, we found that the baseline odds of
The odds of a tweet containing anxiety language versus no anxiety language, as determined using a zero-inflated model with categorical deatha.
Variable | Odds ratio (95% CI) | |
Intercept | 0.11 (0.07-0.16) | <.001 |
Anxiety in article | 2.65 (1.58-4.43) | <.001 |
Second quartile (22,253-133,665 deaths) | 0.76 (0.41-1.41) | .39 |
Third quartile (133,666-193,321 deaths) | 0.41 (0.2-0.83) | .01 |
Fourth quartile (≥193,322 deaths) | 0.42 (0.21-0.85) | .02 |
Interaction anxiety in article by second quartile deaths (22,253-133,665 deaths) | 0.71 (0.34-1.48) | .36 |
Interaction anxiety in article by third quartile deaths (133,666-193,321 deaths) | 1.32 (0.54-3.24) | .55 |
Interaction anxiety in article by fourth quartile deaths (≥193,322 deaths) | 1.9 (0.75-4.83) | .18 |
aThis table reports the odds of no tweet anxiety versus tweet anxiety. Deaths were categorized based on the second, third, and fourth quartiles relative to the first quartile.
Article anxiety predicting the odds of tweet anxiety versus no tweet anxiety at the first, second, third, and fourth quartiles of the COVID-19 death toll.
We then examined the actual estimated linguistic anxiety of tweets, looking only at all of the values in a continuous distribution, excluding those values with zero anxiety (ie, the tweet did not contain any anxiety lexicon words). Although not statistically significant at
Actual anxiety expressed in tweets, as predicted by article anxiety and COVID-19 death toll: gamma regression model with categorical deatha.
Variable | Coefficient (95% CI) | |
Intercept | 3.45 (2.77-4.28) | <.001 |
Anxiety in article | 1.25 (0.99-1.59) | .07 |
Second quartile (22,253-133,665 deaths) | 1.53 (1.1-2.15) | .01 |
Third quartile (133,666-193,321 deaths) | 1.47 (0.97-2.22) | .17 |
Fourth quartile (≥193,322 deaths) | 1.21 (0.83-1.75) | .32 |
Interaction anxiety in article by second quartile deaths (22,253-133,665 deaths) | 0.78 (0.56-1.08) | .14 |
Interaction anxiety in article by third quartile deaths (133,666-193,321 deaths) | 0.61 (0.37-1.01) | .06 |
Interaction anxiety in article by fourth quartile deaths (≥193,322 deaths) | 0.78 (0.5-1.22) | .28 |
aDeaths were categorized based on the second, third, and fourth quartiles relative to the first quartile. This table reports the actual estimated anxiety in the tweet, looking only at all of the values in a continuous distribution, excluding those with zero anxiety.
Article anxiety predicting nonzero tweet anxiety at the first, second, third, and fourth quartiles of the COVID-19 death toll.
This study reports exploratory findings on the effects of fear-inducing news messages during a pandemic. Most importantly, we demonstrated a link between the anxiety expressed in news articles and the odds of anxiety being expressed by those who shared the articles to Twitter. This likely reflects the ability of pandemic-related news messages to elicit a measure of fear in their readers, consonant with public health goals. However, likely as a function of the rising COVID-19 threat over time (as indicated by LIWC news article anxiety) and a low perceived ability to prevent the rapid spread of the virus, anxiety did not increase in response to climbing death tolls over time. Instead, anxiety in tweets increased sharply in response to article anxiety early on in the pandemic, but as the death toll climbed, it flattened out, and news articles seemingly lost their ability to elicit anxiety among readers.
Such findings from this study provide several insights and directions for future research. Our findings reveal that responses to COVID-19 news as well as the rising death toll are increasingly bland. Growing desensitization in the face of threatening pandemic information impedes public health experts’ efforts to mitigate the COVID-19 crisis [
Second, since extant research shows that both statistics (eg, percentage of deaths) and cognitive dissonance can elicit desensitization [
Our findings illuminate desensitization to fear-inducing news messages during the pandemic; however, this study is not without limitations. By focusing on Twitter, we neglected to explore how anxiety manifests on other platforms for sharing news (eg, the comments section of digital news sites). As different platforms have different community norms [
This work investigates how individuals' emotional reactions to news of the COVID-19 pandemic manifest as the death toll increases. Individuals become desensitized to an increased health threat and their emotional responses are blunted over time. Our results suggest desensitized public health reactions to threatening COVID-19 news, which could affect the propensity of individuals to adopt recommended health behaviors.
Public health agencies made recommendations to slow the pandemic’s spread, including physically distancing from others when appropriate, wearing masks, engaging in frequent handwashing, and disinfecting frequently touched surfaces. The consequences of ignoring these guidelines initially incited widespread fear and anxiety around contracting the virus or having family and friends contract it and fall ill. Social scientists have tried to inform interventions aimed at promoting compliance with public health experts [
Centers for Disease Control and Prevention
Linguistic Inquiry and Word Count
odds ratio
None declared.