Existence of Differential Belief Proles of COVID-19 Narratives: The Role of Trust in Science

The global spread of coronavirus disease (COVID-19) diffusion of misinformation and conspiracy theories about its origins (such as 5G cellular networks) and the motivations of preventive measures like vaccination, social distancing, and face masks (for example, as a political ploy). These beliefs have resulted in substantive, negative real-world outcomes but remain largely unstudied.


Abstract Background
The global spread of coronavirus disease 2019 (COVID -19) has been mirrored by diffusion of misinformation and conspiracy theories about its origins (such as 5G cellular networks) and the motivations of preventive measures like vaccination, social distancing, and face masks (for example, as a political ploy). These beliefs have resulted in substantive, negative real-world outcomes but remain largely unstudied.

Methods
This was a cross-sectional, online survey (n=660). Participants were asked about the believability of ve selected COVID-19 narratives, their political orientation, their religious commitment, and their trust in science (a 21-item scale), along with sociodemographic items. Data were assessed descriptively, then latent pro le analysis was used to identify subgroups with similar believability pro les. Bivariate (ANOVA) analyses were run, then multivariable, multivariate logistic regression was used to identify factors associated with membership in speci c COVID-19 narrative believability pro les.

Results
For the full sample, believability of the narratives varied, from a low of 1.94 (SD=1.72) for the 5G narrative to a high of 4.56 (SD=1.64) for the zoonotic (scienti c consensus) narrative. Four distinct belief pro les emerged, with the preponderance (70%) of the sample falling into Pro le 1, which believed the scienti cally accepted narrative (zoonotic origin) but not the misinformed or conspiratorial narratives.
Other pro les did not disbelieve the zoonotic explanation, but rather believed additional misinformation to varying degrees. Controlling for sociodemographics, political orientation and religious commitment were marginally, and typically non-signi cantly, associated with COVID-19 belief pro le membership. However, trust in science was a strong, signi cant predictor of pro le membership, with lower trust being substantively associated with belonging to Pro les 2 through 4.

Conclusions
Belief in misinformation or conspiratorial narratives may not be mutually exclusive from belief in the narrative re ecting scienti c consensus; that is, pro les were distinguished not by belief in the zoonotic narrative, but rather by concomitant belief or disbelief in additional narratives. Additional, renewed dissemination of scienti cally accepted narratives may not attenuate belief in misinformation. However, prophylaxis of COVID-19 misinformation might be achieved by taking concrete steps to improve trust in science and scientists, such as building understanding of the scienti c process and supporting open science initiatives. Background Page 3/25 As coronavirus disease 2019  has spread around the globe, the scienti c community has responded by conducting and providing unprecedented access to research studies related to  Early in the course of the pandemic, researchers noticed the spread of misinformation, conspiracy theories (causal attribution to "machinations of powerful people who attempt to conceal their role"), 2 and unveri ed information about COVID-19, 3,4 which has taken the form of false/fabricated content and true information presented in misleading ways. 5 This deluge of information has introduced confusion among the public in terms which sources of information are trustworthy, 6 despite the open conduct of epidemiological research and other scienti c work on COVID-19.
Although one might expect that improved access and visibility of research would result in increased trust being placed in scientists and the scienti c enterprise, a preliminary study failed to nd such a change between December 2019 and March 2020 in the United States (US). 7 Peer reviewed studies exist alongside misinformation about medical topics, which is easily accessible in the US and is associated with differential health behaviors (e.g., who gets a vaccine, or who takes herbal supplements). 8 As we describe and demonstrate subsequently, belief in misleading narratives about COVID-19 can have substantive, real-world consequences that makes this both an important theoretical and practical area of study. At the same time, evidence suggests that belief in misinformation is not pathological, but rather that it merits treatment as a serious area of scienti c inquiry. 9

Misinformation and Conspiracy Theories
Research on misinformation and conspiratorial thinking has burgeoned in recent years. Because this work has focused both on misinformation and conspiratorial thinking, we use these terms consistently with the speci c studies cited, but somewhat interchangeably.
Consistent with the proliferation of misinformation about COVID-19, it has been proposed that conspiratorial thinking is more likely to emerge during times of societal crisis 10 and may stem from heuristic reasoning (e.g., "a major event must have a major cause"). 11 At the same time, endorsement of misinformation or conspiracy seems to be common, with evidence from nationally representative research indicating that approximately half of US residents endorsed at least one conspiracy in surveys from 2006 to 2011, even when only offered a short list of possibilities. 12 A recent study of COVID-19 conspiracy theories similarly found that nearly 85% of a representative US sample believed that at least one COVID-19 conspiracy theory was "probably" or "de nitely" true. 13 The widespread nature of this phenomenon logically suggests that endorsing misinformation is unlikely to be caused by delusions or discreet pathology.

Factors Associated with Beliefs
Previous research on factors associated with belief in misinformation or conspiracy theories has produced varying, and sometimes inconsistent, ndings. The endorsement of misinformation has been found to vary across sociodemographic groups. For example, studies have identi ed that both low 14 and high 15 education levels are positively associated with belief in certain conspiratorial ideas. In addition, individuals who are contextually low-status (e.g., part of a minority group within a culture) may be more likely to endorse conspiracy theories, especially about high-status groups, but social dynamics likely affect this substantively. 16 Political orientation is generally believed to be associated with conspiratorial endorsement or belief in misinformation; some studies have reported that conservatism predicts such beliefs, such as a nding that sharing fake news on Facebook during the 2016 US presidential election was associated with political conservatism and being age 65 or older, but researchers acknowledged potential omitted variable bias and pointed to the potential confounding (unmeasured) role of digital media literacy. 17 However, other researchers have suggested that strong political ideology on either side (left or right) is more explanatory, 18 and that associations vary depending on the political orientation of the conspiracy or misinformation itself. 19 Consistent with the latter explanations, a working paper by Pennycook et al. examined data from the US, Canada, and the UK and found that cognitive sophistication (e.g., analytic thinking, basic science knowledge) was a superior predictor of endorsing misinformation about COVID-19 than political ideology, though none of the included variables predicted behavior change intentions. 20 This mirrored his prior nding that lower levels of analytic thinking were associated with inability to differentiate between real and fake news. 21 Though less well studied, religiosity, too, may impact endorsement of misinformation, but the relationship is likely complex and mediated by trust in political institutions. 22 Researchers have posited positive, indirect relationships between religion, in general, and endorsement of conspiracy theories, such as the conceptual similarity between an all-powerful being and a hidden power orchestrating events, which is a core feature of conspiratorial thinking. 15 The Importance of Misinformation about COVID-19 Misinformation about COVID-19 is an important area of study not just theoretically, but also because of the potential for these beliefs to lead to real-world consequences. The present study examined four core misperceptions about COVID-19 that contributed to short-term adverse consequences (situated alongside a fth narrative that re ects scienti c consensus). The misperceptions were drawn from Cornell University's Alliance for Science, which prepared a list of current COVID-19 conspiracy theories in April, 2020. 23 33 and there has been a series of incidents where preventive measures like mask-wearing in public have become brief, violet ashpoints, resulting in outcomes up to and including murder. 34 We cannot be certain, yet, about the long-term effects of beliefs about COVID-19 on the landscape of US politics, treatment of vulnerable populations, and other longer-term outcomes, though lessons from prior viral epidemics such as human immunode ciency virus (HIV) suggest that misinformation like AIDS denialism, when embedded, can result in avoidable morbidity and mortality. 35 Addressing Misinformation Misinformation can be di cult to address in the public sphere because it requires the source of information be trusted, 36 while the very nature of misinformation often hypothesizes that experts or authorities are working to conceal the truth. Krause and colleagues note that it is important for scholars to be honest and transparent about the limits of knowledge (e.g., uncertainty), and that simply asserting one's trustworthiness or accuracy is likely an insu cient step to take. 36 Further, one cannot assume that "fact checkers" are trusted by the public to be objective, or that objective presentation of data will simply overturn misinformation, especially when it is value-laden. 36 Timing of information provision may also matter; studies have suggested that people may be less inclined to share or endorse misinformation or conspiracy theories if they are presented with reasoned, factual explanations prior to their exposure to misinformation, 37 but this was not found to be true after exposure; stated differently, factual information may be capable of prevention, but not treatment. This is consistent with theories about fact-based inoculation conferring resistance to argumentative persuasion. 38 Adding additional complication, in the same way that misinformation tends to proliferate within a social echo chamber where few individuals interact with content "debunking" misinformation, scienti c information tends to be shared within its own echo chamber, rarely being interacted with by those who do not already agree with the content. 39 So even if a scienti c source of information is trusted, and "gets out ahead" of misinformation, there is a risk it will never reach its intended audience. The summed total of this information led us to conclude that: (a) it is both practically and theoretically important to understand the factors underlying endorsement of misinformation about COVID-19, (b) certain indicators might be, but are not de nitively, associated with endorsement of misinformation, including political orientation, religious commitment, and education level, and (c) if scientists and "fact checkers" are not trusted by some individuals (whether rightly or wrongly), the degree of trustworthiness assigned to scientists may be an underlying mechanism that can explain belief in conspiratorial theories about COVID-19.
To investigate this question, we adopted a person-centered approach to identify pro les of beliefs about COVID-19 narratives. Importantly, these pro les incorporated perceived believability not only of misinformation, but also of a scienti cally-accepted statement about the zoonotic source of COVID-19.
To identify belief pro les, we used Latent Pro le Analysis (LPA), a speci c case of a nite mixture model that enables identi cation of subgroups of people according to patterns of relationships among selected continuous variables (i.e., "indicators," in mixture modelling terminology). 40 The goal of LPA is to identify the fewest number of latent classes (i.e., homogenous groups of individuals) that adequately explains the unobserved heterogeneity of the relationships between indicators within a population.
We hypothesized that 1) there are distinct pro les of individuals' beliefs in different narratives related to COVID-19; 2) trust in science and scientists, as conceptualized in prior research on this topic, 7,41 is lower among subgroups that endorse misinformation or conspiracy theories about COVID-19, even after controlling individuals' sociodemographic characteristics, political orientation and religious commitment (see Figure 1 as a conceptual model).

Data Collection
Data were obtained on May 22, 2020, from a sample of 660 US-based Amazon Mechanical Turk (mTurk) users ages 18 and older. A relatively new data collection platform, mTurk allows for rapid, inexpensive data collection mirroring quality that has been observed through traditional data collection methods, 42,43 including generally high reliability and validity. 44 Though not a mechanism for probability sampling, 44 mTurk samples appear to mirror the US population in terms of intellectual ability 45 and most, but not all, sociodemographic characteristics. 46 To ensure data quality, minimum quali cations were speci ed to initiate the survey (task approval rating >97%, successful completion of more than 100, but fewer than 10,000 tasks, US-based IP address). 46 Additional checks were embedded within the survey to screen out potential use of virtual private networks (VPNs) to mimic US-based IP addresses, eliminate bots, and manage careless responses. 47 Failing at these checkpoints resulted in immediate termination of the task and exclusion from the study. Four narrative statements were drawn and synthesized from Cornell University's Alliance for Science. 23 An additional statement was based on the zoonotic explanation. 28,29 The statements were prefaced with a single prompt, reading: "There is a lot of information available right now about the origins of the COVID-19 virus. We are interested in learning how believable you nd the following explanations of COVID-19." The statements were as below and used to form the pro les of believability of COVID-19 narratives: 1. "The recent rollout of 5G cellphone networks caused the spread of COVID-19." 2. "The COVID-19 virus originated in animals (like bats) and spread to humans." 3. "Bill Gates caused (or helped cause) the spread of COVID-19 in order to expand his vaccination programs." 4. "COVID-19 was developed as a military weapon (by China, the United States, or some other country)." 5. "COVID-19 is no more dangerous than the u, but the risks have been exaggerated as a way to restrict liberties in the United States."

Trust in Science and Scientists
Participants were asked to complete the Trust in Science and Scientist Inventory consisting of 21 questions with 5-point Likert-type response scales ranging from 1 (Strongly disagree) to 5 (Strongly agree). After adjusting for reverse-coded items, the mean value of the summed scores of 21 questions was used to indicate a level of trust ranging from 1 (Low Trust) to 5 (High Trust). 41 The scale demonstrated excellent reliability for this sample (α = .931).

Religious Commitment
Participants were asked to describe their "level of religious commitment" on a scale from 1 (Low) to 10 (High).

Political Orientation
Participants were asked to describe their "political orientation" on a scale from 1 (Liberal) to 10 (Conservative).

Statistical Analysis
Four stages of analyses were conducted. First, descriptive statistics were computed and reported for believability of COVID-19 narratives, religious commitment, political orientation, trust in science, and sociodemographic characteristics (e.g., race/ethnicity, sex, sexual orientation, education level). Means and standard deviations (SD) were used to describe continuous variables (e.g., believability of COVID-19 narratives, age Angeles, CA) to delineate subgroups of belief patterns related to COVID-19 among participants ( Figure   1). 40 We used maximum likelihood and a robust estimator (Huber-White, MLR estimator in Mplus) to handle the non-normal distribution of the indicators (absolute value of skew ranged from 0.30 to 1.67, and of kurtosis ranged from 1.70 to 4.39). LPA is an unsupervised machine learning technique to identify unobserved groups or patterns from the observed data. 40,49 Compared to traditional cluster analysis, LPA adapts a person-centered approach to identify the classes of participants who may follow different patterns of beliefs in COVID-19 narratives with unique estimates of variances and covariate in uences.
Since no other study has investigated this question or these variables, we followed an exploratory approach to identifying the number of classes by testing increasingly more classes until the value of the log likelihood began to level off (1-5 latent classes).
To determine the nal number of classes, we systematically considered conceptual meaning, 50  separation. 56 Models that included class sizes with less than 1% of the sample or that did not converge were not considered due to the risk of poor generalizability. 57 The Vuong-Lo-Mendel-Rubin Likelihood Ratio Test (LMR) 58 was further used to test whether models with k classes improved the model t versus models with k-1 classes (a signi cant p-value <.05 suggested such improvement). Full information maximum likelihood (FIML) estimation was used to handle missing data. [59][60][61] Third, bivariate analyses were conducted between the study variables and the classi ed groups using analysis of variance (ANOVA). A Bonferroni correction for multiple comparisons was applied. Finally, multivariate multinomial logistic regressions were used to examine the utility of trust in science in identifying COVID-19 narrative groups, adjusting for all sociodemographic variables, political orientation, and religious commitment. Signi cance testing was 2-sided and carried out at the 5% signi cance level.

Descriptive Statistics
Of the 660 participants (see Table 1 Note. GED = General Education Development or General Education Diploma. SD=standard deviation.
a Educational levels were treated as continuous variables in later analyses.

Pro les of Beliefs in COVID-19 Narratives
Based on model t statistics (see Table 2), we selected a 4-class model. The LMR test was nonsigni cant when comparing the 5-class to the 4-class model, the model t indices of the 4-class model were the smallest among 1-to 4-class models, and the entropy was over 0.60 and the highest of all the estimated models (entropy = 0.994). The smallest class of the 4-class model was also larger than 5% of the total sample (8.18%). Pro le 3 (n= 77, 11.67%) reported low-to-moderate believability for all of the narrative statements. In most cases, this class had the second-lowest belief scores for narrative statements, but also, notably, the lowest score for the zoonotic narrative (mean = 4.59, SD=1.71).
Pro le 4 (n= 66, 10.00%) reported fairly high believability for most narratives (similarly to Pro le 2). However, this group diverged from Pro le 2 in indicating lower plausibility of the 5G narrative (mean = 4.55, SD=0.50), though it was still a higher level of belief than for Pro les 1 and 3. Table 3, pro les differed signi cantly across racial/ethnic groups, education levels, political orientation, religious commitment, and trust in science. These ndings are provided for transparency and context, but the primary associative ndings are those in the next subsection (e.g., the multivariate models). Note. n = 660; SD=standard deviation.

Multivariate Models Predicting COVID-19 Belief Pro les
The multivariate regression models contrasted Pro les 2 through 4 with Pro le 1 (which was the pro le expressing belief in the zoonotic narrative but the lowest belief in the other narratives). Controlling for race/ethnicity, gender, age, and education level, individuals with greater trust in science were less likely to be in Pro le 2 (AOR=0.07, 95%CI=0.03-0.16), Pro le 3 (AOR=0.20, 95%CI=0.12-0.33), and Pro le 4 (AOR=0.07, 95%CI=0.03-0.15) than Pro le 1. In addition, study participants with greater religious commitment were more likely to be in Pro le 3 (AOR=1.12, 95% CI = 1.02-1.22) than Pro le 1. No other signi cant differences related to religious commitment were observed, though it appears that with a larger sample size a similar religious effect may have been signi cant for Pro les 2 and 4. Political orientation was not associated with belief pro les in the multivariate models. Trust in science 0.07 (0.03-0.16)*** 0.20 (0.12-0.33)*** 0.07 (0.03-0.15)*** Note. * P<0.05 ** P<0.01 *** P<0.00. AOR = Adjusted Odds Ratio; 95% CI = 95% Con dence Interval; Controlled for race/ethnicity, gender, age, and educational levels.

Discussion
This study tested two preliminary hypotheses about beliefs in narratives about COVID-19. We had hypothesized that individuals would be separable into distinct latent classes based on belief in various narratives about COVID-19, and the LPA analysis identi ed four statistically and conceptually different subgroups. Further, we speculated that trust in science was lower among that groups that high believability for misinformation about COVID-19, which was partially supported by our results. These results should be interpreted as supporting the plausibility of these explanations, but as always, should be replicated and further investigated before de nitive conclusions are made. We speci cally encourage further replication and extensions of this work and support open dialogue about the ndings and their implications.

Pro les of COVID-19 Belief Subgroups
Prior research on conspiracy theories has suggested that many people in the US believe in at least one conspiracy theory, 12 and that those who do may believe in multiple conspiracy theories. 13 Our LPA analysis, which included believability not only of conspiracy theories/misinformation, but also of the scienti cally-accepted zoonotic explanation for COVID-19, a rmed this nding and added considerable detail.
Pro le 1 reported the lowest believability for each conspiratorial narrative and reported high believability of the zoonotic narrative. This may suggest that people who are skeptical of misinformation tend to believe the scienti cally accepted narrative. Interestingly, however, the converse was not true. In fact, the highest believability in the zoonotic explanation was observed for Pro le 2, which reported the highest believability for all explanations. Further, Pro le 4 was fairly similar to Pro le 2, except for lower endorsement of the 5G theory, which we subjectively note is the least plausible theory on its face, given a complete lack of scienti c evidence that wireless technology can transmit a virus. Finally, Pro le 3 reported low to moderate believability for all narrative statements but reported the lowest endorsement for the zoonotic explanation. This is also important to note, as it rea rms that a generally neutral position on the believability of misinformed narratives does not necessarily translate to endorsement of a scienti cally-accepted narrative.
Our data support the existence of multiple and distinct belief pro les for COVID-19 misinformation. Based on these ndings, we speculate that one reason providing factual information has not always reduced endorsement of misinformation 37 is that latent groups of people exist for whom belief in a scienti callyaccepted explanation is not a mutually exclusive alternative to belief in misinformation (e.g., Pro les 2 and 4). For people belonging to these subgroups, convincing them of the validity of the scienti callyaccepted explanation may simply increase their belief in that explanation, without concomitant reductions in belief in alternative narratives. In addition, it is important to note that even Pro le 1, which was the most skeptical of misinformation and which expressed high believability for the zoonotic explanation, reported a mean believability value >2 for two alternative narratives (laboratory development and liberty restriction). Though such narratives are not strongly supported by currently-available evidence, neither are they scienti cally impossible (such as the 5G theory). In some ways, failure to reject such alternative narratives with complete certainty better re ects true scienti c work better than would absolute rejection of all alternative narratives. 36 Predictors of COVID-19 Belief Subgroups In our multivariate, multinomial logistic regression models, controlling for race/ethnicity, gender, age, and education level (as well as the other predictor variables), political orientation was not signi cantly associated with belonging to any particular COVID-19 belief subgroup. This is consistent with some prior hypotheses, 12 but it is important to note, given the tenor of current political discussion in the US, that endorsement of misinformation about COVID-19 does not appear to be a function of politics.
Although religious commitment was signi cantly associated with being part of Pro le 3 versus Pro le 1, the magnitude of this association was not particularly large in distinction to the ndings related to trust in science. In addition, examining the con dence intervals independently of signi cance levels, one might reasonably speculate that belonging to any of Pro les 2 through 4 might be potentially associated with increased religious commitment. It may be the case that the trust in science variable captures some of the complexity that has been observed in associating religion and belief in misinformation. 22 Finally, low trust in science was substantially and signi cantly predictive of belonging to Pro les 2, 3, and 4, relative to Pro le 1. However, those pro les were distinguished from Pro le 1 not by their failure to believe in the zoonotic explanation, but by their endorsement of alternate explanations. In other words, trusting science and scientists appears to be associated with lower likelihood of expressing a belief pattern that endorses narratives that are de nitively, or likely to be, misinformed. In this sense, trust in science was conceptually less related to what narrative to believe, and more related to what narrative(s) are more appropriate to disbelieve.
It is important, on a surface level, to understand the potential importance that trust in science has in understanding how people perceive competing narrative explanations about a major event like the COVID-19 pandemic. However, the solution is not likely to be as simply as "just building more trust in science." First, consider the con ict described earlier in this manuscript, where there is an inherent tension between conspiratorial thinking and trusting expert opinion. If it were true, for example, that 5G networks were being used to spread COVID-19, then the authorities doing so, and desiring to hide it, would have an interest in debunking the 5G narrative. If "science" and "authority" or "government bodies" become con ated, then lower trust in science may result from distrust of authority, thereby affecting believability of explanations. 62 Thus, one important consideration might be the importance of working to ensure that science remains non-partisan, including careful vigilance for white hat bias (distortion of ndings to support the "right" outcome). 63 Second, although as researchers we believe in the power of the scienti c approach to uncover knowledge, there have been well-documented cases of scienti c misconduct, such as the 1998 Wake eld et al. paper linking vaccines and autism, 64 as well as other concerns about adherence to high-integrity research procedures. 65 Anomalies or other issues related to research partnerships can occur as well; while this paper was being revised for submission, a major COVID-19 study on hydroxychloroquine was retracted due to issues with data access for replication. 66 Thus, to some degree, scienti c opinion on the use of hydroxychloroquine may have shifted in the past 24 hours. At the same time, as researchers, we understand that a single study does not constitute consensus. Science, as a eld, scrutinizes itself and tends to be self-correcting -though not always as rapidly as one might wish, and systems regularly have been recon gured to ensure integrity. 67 However, to a person not embedded within the scienti c research infrastructure, it is not necessarily irrational to report a low level of trust in science on the basis of the idea that certain scienti c theories have been wrong, or ndings fraudulently obtained.
Given that trust in science and scientists was the most meaningful factor predicting pro le membership, accounting for a wide variety of potential covariates, systematically building trust in science and scientists might be an effective way to inoculate populations against misinformation related to COVID-19, and potentially other misinformation. Based on this study's ndings, this would speci cally not take the form of repeatedly articulating factual explanations (especially within a scienti c echo chamber 39 ), as this might potentially increase believability of accurate narratives, but only as one among other equally believable narratives. Rather, to improve trust in science, we might consider demonstratinghonestly and openly -how science works, and then articulating why it can be trusted. 36 Parallel processes such as implementing recommendations to facilitate open science 68 may also have the secondary effect of improving overall public trust in science. Individuals who both understand 20,21 and trust science 7,41 appear to be most likely to reject explanations with less supporting evidence while accepting narratives with more supporting evidence.

Limitations
This study has several limitations. First, to conduct rapid research amid a pandemic, we used the mTurk survey platform. As noted in our Methods, this is a widely accepted research platform across multiple disciplines, but it does not produce nationally representative data. Thus, the ndings should not be generalized to any speci c population without further study. Second, due to the newness of COVID-19, no questionnaire had been validated to articulate or measure speci c explanations of COVID-19. However, we suggest some face validity to our questions because the response scale was established in prior research. Third, as with all inferential models, this study is subject to omitted variable bias, 69 though the magnitude of the association between the latent pro les and the trust in science variable somewhat attenuates this concern. Fourth, since this was a cross-sectional study, we cannot assert any causality or directionality.

Conclusions
We propose several next steps after the current work. First, a larger, nationally representative sample of individuals in the US should complete these items, potentially also including common misinformation or conspiracy theories about other topics likely to affect health behaviors, like vaccination.70 Second, it will be important for future studies to determine whether our study's ndings can be replicated, are highly generalizable, and whether additional nuances to the ndings can be identi ed by the broader scienti c community. Further, longitudinal studies could be structured to enable causal inferences from the pro les.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.