The causes and consequences of COVID-19 misperceptions: Understanding the role of news and social media

---

Implications

The COVID-19 pandemic has been accompanied by a so-called “infodemic”—a global spread of misinformation that poses a serious problem for public health. Infodemics are concerning because the spread of false or misleading information has the capacity to change transmission patterns (Kim et al., 2019) and consequently the scale and lethality of a pandemic. This information can be shared by any media, but there is reason to be particularly concerned about the role that social media, such as Facebook and Twitter, play in incidentally boosting misperceptions. These platforms are increasingly relied upon as primary sources of news (Mitchell et al., 2016) and misinformation has been heavily documented on them (Garrett, 2019; Vicario et al., 2016). Scholars have found medical and health misinformation on the platforms, including that related to vaccines (Radzikowski et al., 2016) and other virus epidemics such as Ebola (Fung et al., 2016) and Zika (Sharma et al., 2017). 

However, misinformation content typically makes up a low percentage of overall discussion of a topic (e.g. Fung et al., 2016) and mere exposure to misinformation does not guarantee belief in that misinformation. More research is thus needed to understand the extent and consequences of misinformation surrounding COVID-19 on social media. During the COVID-19 pandemic, Twitter, Facebook and other platforms have engaged in efforts to combat misinformation but they have continued to receive widespread criticism that misinformation is still appearing on prominent pages and groups (Kouzy et al., 2020; NewsGuard, 2020). The extent to which misinformation continues to circulate on these platforms and influence people’s attitudes and behaviours is still very much an open question.

Here, we draw on three data sets and a sequential mixed method approach to better understand the consequences of online misinformation for important behaviours and attitudes. First, we collected nearly 2.5 million tweets explicitly referring to COVID-19 in the Canadian context. Second, we collected just over 9 thousand articles from nineteen Canadian English-language news sites from the same time period. We coded both of these media sets for misinformation and public health recommendations. Third, we conducted a nationally representative survey that included questions related to media consumption habits, COVID-19 perceptions and misperceptions, and social distancing compliance. As our outcome variables are continuous, we use Ordinary Least Squares (OLS) regression to identify relationships between news and social media exposure, misperceptions, compliance with social distancing measures, and risk perceptions. We use these data to illustrate: 1) the relative prevalence of misinformation on Twitter; and 2) a powerful association between social media usage and misperceptions, on the one hand, and social distancing non-compliance on the other.

Misinformation and compliance with social distancing

We first compare the presence of misinformation on Twitter with that on news media and find, consistent with the other country cases (Chadwick & Vaccari, 2019; Vicario et al., 2016), comparatively higher levels of misinformation circulating on the social media platform. We also found that recommendations for safe practices during the pandemic (e.g. washing hands, social distancing) appeared much more frequently in the Canadian news media. These findings are in line with literature examining fake news which finds a large difference in information quality across media (Al-Rawi, 2019; Guess & Nyhan, 2018).

Spending time in a media environment that contains misinformation is likely to change attitudes and behaviours. Even if users are not nested in networks that propagate misinformation, they are likely to be incidentally exposed to information from a variety of perspectives (Feezell, 2018; Fletcher & Nielsen, 2018; Weeks et al., 2017). Even a highly curated social media feed is thus still likely to contain misinformation. As cumulative exposure to misinformation increases, users are likely to experience a reinforcement effect whereby familiarity leads to stronger belief (Dechêne et al., 2010).

To evaluate this empirically, we conducted a national survey that included questions on information consumption habits and a battery of COVID-19 misperceptions that could be the result of exposure to misinformation. We find that those who self-report exposure to the misinformation-rich social media environment do tend to have more misperceptions regarding COVID-19. These findings are consistent with others that link exposure to misinformation and misperceptions (Garrett et al., 2016; Jamieson & Albarracín, 2020). Social media users also self-report less compliance with social distancing.

Misperceptions are most meaningful when they impact behaviors in dangerous ways. During a pandemic, misperceptions can be fatal. In this case, we find that misperceptions are associated with reduced COVID-19 risk perceptions and with lower compliance with social distancing measures. We continue to find strong effects after controlling for socio-economic characteristics as well as scientific literacy. After accounting for the effect of misperceptions on social distancing non-compliance, social media usage no longer has a significant association with non-compliance, providing evidence that social media may lead to less social distancing compliance through its effect on COVID-19 misperceptions.

While some social media companies have made efforts to suppress misinformation on their platforms, there continues to be a high level of misinformation relative to news media. Highly polarized political environments and media ecosystems can lead to the spread of misinformation, such as in the United States during the COVID-19 pandemic (Allcott et al., 2020; Motta et al., 2020). But even in healthy media ecosystems with less partisan news (Owen et al., 2020), social media can continue to facilitate the spread of misinformation. There is a real danger that without concerted efforts to reduce the amount of misinformation shared on social media, the large-scale social efforts required to combat COVID-19 will be undermined. 

We contribute to a growing base of evidence that misinformation circulating on social media poses public health risks and join others in calling for social media companies to put greater focus on flattening the curve of misinformation (Donovan, 2020). These findings also provide governments with stronger evidence that the misinformation circulating on social media can be directly linked to misperceptions and public health risks. Such evidence is essential for them to chart an effective policy course. Finally, the methods and approach developed in this paper can be fruitfully applied to study other waves of misinformation and the research community can build upon the link clearly drawn between misinformation exposure, misperceptions, and downstream attitudes and behaviours.

Next steps

We found use of social media platforms broadly contributes to misperceptions but were unable to precise the overall level of misinformation circulating on non-Twitter social media. Data access for researchers to platforms such as Facebook, YouTube, and Instagram is limited and virtually non-existent for SnapChat, WhatsApp, and WeChat. Cross-platform content comparisons are an important ingredient for a rich understand of the social media environment and these social media companies must better open their platforms to research in the public interest. 

Findings

Finding 1: Misinformation about COVID-19 is circulated more on Twitter as compared to traditional media.

We find large differences between the quality of information shared about COVID-19 on traditional news and Twitter. Figure 1 shows the percentage of COVID-19 related content that contains information linked to a particular theme. The plot reports the prevalence of information on both social and news media for: 1) three specific pieces of misinformation; 2) a general set of content that describes the pandemic itself as a conspiracy or a hoax; and 3) advice about hygiene and social distancing during the pandemic. We differentiate content that shared misinformation (red in the plot) from content that debunked misinformation (green in the plot). 

There are large differences between the levels of misinformation on Twitter and news media. Misinformation was comparatively more common on Twitter across all four categories, while debunking was relatively more common in traditional news. Meanwhile, advice on hygiene and social distancing appeared much more frequently in news media. Note that higher percentages are to be expected for longer format news articles since we rely on keyword searches for identification. This makes the misinformation findings even starker – despite much higher average word counts, far fewer news articles propagate misinformation.

Finding 2: There is a strong association between social media exposure and misperceptions about COVID-19. The inverse is true for exposure to traditional news.

Among our survey respondents we find a corresponding strong association between social media exposure and misperceptions about COVID-19. These results are plotted in Figure 2, with controls included for both socioeconomic characteristics and demographics. Moving from no social media exposure to its maximum is expected to increase one’s misperceptions of COVID-19 by 0.22 on the 0-1 scale and decreased self-reported social distancing compliance by 0.12 on that same scale.

This result stands in stark contrast with the observed relationship between traditional news exposure and our outcome measures. Traditional news exposure is positively associated with correct perceptions regarding COVID-19. Moving from no news exposure to its highest level is expected to reduce misperceptions by 0.12 on the 0-1 scale and to increase social distancing compliance by 0.28 on that same scale. The effects are plotted in Figure 2. Social media usage appears to be correlated with COVID-19 misperceptions, suggesting these misperceptions are partially a result of misinformation on social media. The same cannot be said of traditional news exposure.

Finding 3: Misperceptions about the pandemic are associated with lower levels of risk perceptions and social distancing compliance.

COVID-19 misperceptions are also powerfully associated with lower levels of social distancing compliance. Moving from the lowest level of COVID-19 misperceptions to its maximum is associated with a reduction of one’s social distancing by 0.39 on the 0-1 scale. The previously observed relationship between social media exposure and misperceptions disappears, suggestive of a mediated relationship. That is, social media exposure increases misperceptions, which in turn reduces social distancing compliance. Misperceptions is also weakly associated with lower COVID-19 risk perceptions. Estimates from our models using COVID-19 concern as the outcome can be found in the left panel of Figure 3, while social distancing can be found in the right panel.

Finally, we also see that the relationship between misinformation and both social distancing compliance and COVID-19 concern hold when including controls for science literacy and a number of fundamental predispositions that are likely associated with both misperceptions and following the advice of scientific experts, such as anti-intellectualism, pseudoscientific beliefs, and left-right ideology. These estimates can similarly be found in Figure 3.

Methods

Canadian Twitter and news data were collected from March 26^th to April 6^th, 2020. We collected all English-language tweets from a set of 620,000 users that have been determined to be likely Canadians. For inclusion, a given user must self-identify as Canadian-based, follow a large number of Canadian political elite accounts, or frequently use Canadian-specific hashtags. News media was collected from nineteen prominent Canadian news sites with active RSS feeds. These tweets and news articles were searched for “covid” or “coronavirus”, leaving a sample of 2.25 million tweets and 8,857 news articles.

Of the COVID-19 related content, we searched for terms associated with four instances of misinformation that circulated during the COVID-19 pandemic: that COVID-19 was no more serious than the flu, that vitamin C or other supplements will prevent contraction of the virus, that the initial animal-to-human transfer of the virus was the direct result of eating bats, or that COVID-19 was a hoax or conspiracy. Given that we used keyword searches to identify content, we manually reviewed a random sample of 500 tweets from each instance of misinformation. Each tweet was coded as one of four categories: propagating misinformation, combatting misinformation, content with the relevant keywords but unrelated to misinformation, or content that refers to the misinformation but does not offer comment. 

We then calculated the overall level of misinformation for that instance on Twitter by multiplying the overall volume of tweets by the proportion of hand-coded content where misinformation was identified. Each news article that included relevant keywords was similarly coded. The volume of the news mentioning these terms was sufficiently low that all news articles were hand coded. To identify health recommendations, we used a similar keyword search for terms associated with particular recommendations: 1) social distancing including staying at home, staying at least 6 feet or 2 meters away and avoiding gatherings; and 2) washing hands and not touching any part of your face. ¹Further details on the media collection strategy and hand-coding schema are available in the supporting materials.

For survey data, we used a sample of nearly 2,500 Canadian citizens 18 years or older drawn from a probability-based online national panel fielded from April 2-6, 2020. Quotas we set on age, gender, region, and language to ensure sample representativeness, and data was further weighted within region by gender and age based on the 2016 Canadian census.

We measure levels of COVID-19 misperceptions by asking respondents to rate the truthfulness of a series of nine false claims, such as the coronavirus being no worse than the seasonal flu or that it can be warded off with Vitamin C. Each was asked on a scale from definitely false (0) to definitely true (5). We use Cronbach’s Alpha as an indicator of scale reliability. Cronbach’s Alpha ranges from 0-1, with scores above 0.8 indicating the reliability is “good.” These items score 0.88, so we can safely construct a 0-1 scale of misperceptions from them. 

We evaluate COVID-19 risk perceptions with a pair of questions asking respondents how serious of a threat they believe the pandemic to be for themselves and for Canadians, respectively. Each question was asked on a scale from not at all (0) to very (4). We construct a continuous index with these items.

We quantify social distancing by asking respondents to indicate which of a series of behaviours they had undertaken in response to the pandemic, such as working from home or avoiding in-person contact with friends, family, and acquaintances. We use principal component analysis (PCA) to reduce the number of dimensions in these data while minimizing information loss. The analysis revealed 2 distinct dimensions in our questions. One dimension includes factors strongly determined by occupation, such as working from home and switching to online meetings. The other dimension contains more inclusive behaviours such as avoiding contact, travel, and crowded places. We generate predictions from the PCA for this latter dimension to use in our analyses. The factor loadings can be found in Table A1 of the supporting materials.

 We gauge news and social media consumption by asking respondents to identify news outlets and social media platforms they have used over the past week for political news. The list of news outlets included 17 organizations such as mainstream sources like CBC and Global, and partisan outlets like Rebel Media and National Observer. The list of social media platforms included 10 options such as Facebook, Twitter, YouTube, and Instagram. We sum the total number of outlets/platforms respondents report using and take the log to adjust for extreme values. We measure offline political discussion with an index based on questions asking how often respondents have discussed politics with family, friends, and acquaintances over the past week. Descriptions of our primary variables can be found in Table A2 of the supporting materials. 

We evaluate our hypotheses using a standard design that evaluates the association between our explanatory and outcome variables controlling for other observable factors we measured. In practice, randomly assigning social media exposure is impractical, while randomly assigning misinformation is unethical. This approach allows us to describe these relationships, though we cannot make definite claims to causality.

We hypothesize that social media exposure is associated with misinformation on COVID-19. Figure 2 presents the coefficients of models predicting the effects of news exposure, social media exposure, and political discussion on COVID-19 misinformation, risk perceptions, and social distancing. Socio-economic and demographic control estimates are not displayed. Full estimation results can be found in the Table A3 of the supporting materials. 

We further hypothesize that COVID-19 misinformation is associated with lower COVID-19 risk perceptions and less social distancing compliance. Figure 3 presents the coefficients for models predicting the effects of misinformation, news exposure, and social media exposure on severity perceptions and social distancing. We show models with and without controls for science literacy and other predispositions. Full estimation results can be found in the Table A4 of the supporting materials.

Limitations and robustness

A study such as this comes with clear limitations. First, we have evaluated information coming from only a section of the overall media ecosystem and during a specific time-period. The level of misinformation differs across platforms and online news sites and a more granular investigation into these dynamics would be valuable. Our analysis suggests that similar dynamics exist across social media platforms, however. In the supplementary materials we show that associations between misperceptions and social media usage are even higher for other social media platforms, suggesting that our analysis of Twitter content may underrepresent the prevalence of misinformation on social media writ large. As noted above, existing limitations on data access make such cross-platform research difficult.

Second, our data is drawn from a single country and language case study and other countries may have different media environments and levels of misinformation circulating on social media. We anticipate the underlying dynamics found in this paper to hold across these contexts, however. Those who consume information from platforms where misinformation is more prevalent will have greater misperceptions and that these misperceptions will be linked to lower compliance with social distancing and lower risk perceptions. Third, an ecological problem is present wherein we do not link survey respondents directly to their social media consumption (and evaluation of the misinformation they are exposed to) and lack the ability to randomly assign social media exposure to make a strong causal argument. We cannot and do not make a causal argument here but argue instead that there is strong evidence for a misinformation to misperceptions to lower social distancing compliance link.