How America Lost the Battle for the Best Scientific Guide Early in the Crisis: COVID-19 and Beyond

In any crisis, such as COVID-19, monkeypox, climate change, etc.
, it is critical
to ensure broad public access to the best scientific guidance.
We show how this battle failed on Facebook early in the COVID-19 flu pandemic and why most people in the mainstream, including many childcare communities, have moved closer
to more extreme communities by the time the vaccine arrives.
The hidden heterogeneity explains why the best scientific guidance Facebook itself promotes also seems to be missing out on key audiences
.
A simple mathematical model reproduces the system-level exposure dynamics
.
Our findings can be used to adjust guidance at scale while considering individual diversity, and to help predict tipping point behavior and system-level responses to future crisis interventions
.

Managing crises (1) such as the COVID-19 pandemic (2, iii), climate change (4) Now, misinformation about monkeypox and miscarriage requires broad public access and acceptance of guidance based on the best available science (5–12) [Guide in the Oxford Dictionary] There is a definition (12) as "advice or information intended to solve a problem or difficulty"].

However, distrust of this best scientific guidance has reached dangerous levels (7–9).
The American Physical Society, like many professional groups, calls scientific misinformation one of the most important problems of our time (8, 9).
With the greatest uncertainty and social distance ahead of COVID-19 vaccination in 2020, many people come to their online communities to guide them on how to avoid infection and propose treatment options
.
Social media has seen a huge leap (13) among users in 2020 (13.
2%), bringing the total to 4.
20 billion (53.
6% of the global population), and the number one reason for going online is seeking information (13).

Unfortunately, the guidance that many people are exposed to is not the best science (iii) and their online community may be well-intentioned, but not expert friends
.
Some people even die from drinking bleach or refusing to use masks (14).
This raises the following pressing questions that we are here to address: Who sent the guidance and to whom? Who received the guidance and from whom? When did it go wrong? What does this tell us about how, where and when to intervene in current and future crises beyond COVID-19?

This paper attempts to address these issues
through empirical and quantitative analysis of COVID-19 networks in the online community.
The study period runs from December 2019 to August 2020, months prior to the December 2020 provisional authorization for the emergency use of any COVID-19mrna vaccine in the U.
S.
or the U.
K.
(15).
We are concerned with transmitter-receiver dynamics, i.
e.
, the extent to which the population set acts as the primary source of COVID-19 guidance (i.
e.
, the
transmitter) and/or is in contact with each other for this guidance (i.
e.
, the receiver).
We supplement this with a mathematical model that reproduces system-level dynamics
.
Therefore, our research complements many excellent existing studies, but differs from many existing studies, including discussions around specific topics, sources, or target groups (16–56).

We use the data collected from Facebook because it is the dominant social media platform globally with 3 billion active users and is the top social network in 156 countries (57).
In addition, recent research confirms that people (e.
g.
, parents) tend to rely on Facebook's built-in community structure to share guidance (58–60).
Therefore, we chose our main unit of analysis to be the built-in Facebook community, especially the Facebook page
.
We refer to each Facebook page simply as a community, but we emphasize that it has nothing to do with any temporary community structure inferred from web algorithms
.
Each page gathers some common interests; It is publicly visible and its analysis does not require us to access personal information
.
Our starting point is COVID-19 (November 2019; See Materials and Methods and Supplemental Materials) before the debate around vaccine health issues on Facebook is interconnected with these community ecosystems
.
Community Link (page) My Community (page) j exists when I recommend j All page level members (i like/fan j) as opposed to the page member simply mentioning another page: the result I can t, automatically exposed to from j, i.
e.
at any time
。 , j transmits and I receive (see Section S1
).
While not all of my members must pay attention to these contents, recent work (61) proves experimentally and theoretically that the commitment of only a minority of 25% is enough to make an online community change its position
.

We collect these Facebook pages and the links between them, using the same method as our previous work (62), and then we categorize these pages in the same way as we worked before (62).
Since we reviewed this method and classification scheme (62) in Materials and Methods (and Section S1), we will only summarize
it here.
We start with a human-identified page that somehow discusses vaccines/vaccinations, and then, we use a combination of computer scripts and human cross-checking to get links
to those pages to other pages.
This process is repeated several times to produce a network
of Facebook pages (nodes) and links between them.
Our trained researchers classify each page as pro-, negative or neutral based on its most up-to-date
content.
"Pro" is a feature of a page whose content actively promotes the best scientific health guidance (Pro vaccination); "Anti" is a webpage with content that actively opposes this guide (anti-vaccination) and "neutral" is a webpage with community-level links to the former COVID-19 of the support/opposition community, but whose content focuses on other topics such as parenting (e.
g.
, child education), pets, and organic food (see Materials and Methods and Section S1).

Each researcher performs an independent manual classification of the content of each page before a consistency check
.
When there was a disagreement, they discussed and in all cases reached an agreement
.
They also further categorized neutral communities based on topics they claimed to be interested in and found 12 categories such as parenting, organic food lovers, and pet lovers (see Section S2 for full discussion and examples
).
While further subdivision is possible, this will result in too few communities in the category and blurred
boundaries.

This approach to data collection and classification provides us with 1356 interconnected communities (Facebook pages), including 86.
7 million individuals from different countries and languages, 211 pro-communities (Figure 1, blue nodes), including 13 million individuals, 501 anti-communities (Figure 1, red nodes) including 7.
5 million individuals and 644 neutral communities (Figure 1A, Golden Node) includes 66.
2 million individuals
.
We can estimate the size of each community based on the number of likes (followers) because a typical user likes only one Facebook page (13) on average: this size usually ranges from a few hundred to a few million users, but we emphasize that our analysis and conclusions do not depend on us determining the size of
the community.
The public information we collect about each page manager (63) suggests that users come from different countries (see Section S2 for
details).
The most common locations for managers are the United States, Australia, Canada, the United Kingdom, Italy, and France (Section S2
).

To complement our empirical analysis, we introduced a simple mathematical model that can simulate the collective dynamics
of these online communities.
We're not saying it represents the best mathematical model (currently unknown), but it does have some favorable features
.
First of all, it has a minimal and transparent form, it is very simple, and its output and predictions can be manually verified with standard calculus
.
Second, it can systematically aggregate or "gel" empirical facts
from first principles (see Derivative Supplemental Materials) to online communities.
These communities are then aggregated into a given type of community
.
Third, the model can be applied to different levels of aggregation: for example, some measure of the activity of all neutral particles t[i.
e.
, t[i.
e.

].
, G(t)] and the corresponding activities of all professionals [B(t)] and Antis [R( t)]] generates three coupled equations, such as equations
.
1 and 2 below, or for example, neutral subcategories can be contained separately, which will add another equation for each subcategory [G ₁(t),G].
₂(t) and so on
.
]
。 Fourth, even at the roughest approximate level (Equation 1), the model produces an output curve similar to the one observed empirically, and a more complex version can be used (for example
).
, Equation 2).
At the roughest approximate level, the rates of change of our model equations R(t), B(t), and G (t) (by · RR? ,· BB? , and.
.
.
GG? , respectively) become

·R=rR(R0? R)+rB(B? R)

·B=bB(B0? B)

·G=gG(G0? G)+gB(B? G)+gR(R? G)

(i)

As a standard for the ecosystem model, each of these three equations contains a self-acting term [g G(_{G 0}?].
G) etc] to explain the intrinsic growth or decline of each subgroup (G₀, etc.
, are constants
).
Coupling terms depend on differences and also have a simple linear form
.
Positive (or negative) coupling means positive (or negative) feedback, for example, if r B >0 (r B<0), then _B (t) The rate of change that exceeds R(t) increases (decreases) R(t) Hence increase (decrease) R(t).
The censorship of online content supports the view that (i) pro-community concerns provide the best scientific guidance to the general public, including neutrals and opponents, indicating that professionals are not materially influenced
by the activities of opponents or neutrals.
Thus, equation B(t) is not coupled to R(t) or G( t); (ii) Opponents are influenced by guidance issued by professionals because they often turn it into their own version (including misinformation) and then provide it to neutrals to draw the neutral's attention to the
best scientific guidance.
This shows that opponents are not materially influenced
by the neutral narrative.
Thus, the equation R(t) is only coupled to B(t); (iii) Neutrals are guided
by professionals and opponents.
Thus, equations G(t) are coupled to both B(t) and R( t).
These self-interacting and coupling terms are shown in Figure 2

A more complex version of Equation 1 is provided below, in which we add the completely mathematically derived feature of Equation 1, where each gel (a community or set of communities depending on the degree of aggregation) has its own time of action t_c (see derivation in the supplementary data); We also add decay terms to simulate loss of interest or suppression of moderators

·R=H(t? tc,? R)[rR(R0? R)+rB(B? R)? dRR]

·B=H(t? tc,? B)[bB(B0? B)? dBB]

·G=H(t? tc,? G)[gG(G0? G)+gB(B? G)+gR(R? G)? dGG]

(ii)

Every H(.
.
.
The Heaviside function that becomes nonzero at the respective start time t_c we use Equation 2 with Figure 3 because it provides a better fit
.
However, as shown in section S5, it is not important to include these starting and decaying terms: both are equations
.
1 and 2 produce shapes similar to empirical curves because they both have the same core structure, coupling term, and self-interaction as shown in Figure 2 Although the actual spread of information and rumors, like disease, is random, it is well known that such deterministic equations can describe the time-averaged behavior of many such random implementations
.
We have verified with stochastic simulations that our equations are equally accurate
.
In addition, parameter estimation and optimization can be difficult to perform with stochastic models (64, 65).
We also studied the impact of noise on the data, randomly removing up to 15% of COVID-19-related links from the entire network to simulate missing or non-existent links (see Additional Materials), and we found that our main results and conclusions were reliable
.

The Figure 1A results show that even with our highly simplified node and link classification scheme, exposure dynamics on any single snapshot can be very complex
even for a small subset of nodes.
The color of the node link (i.
e.
Facebook page) I go to the node (Facebook page) j is the node of me, while the arrow direction indicates the potential flow of COVID-19 guidance, so j To me
.
The color of the arrow is the color j
of the node.
If node j issues COVID-19 guideline t in time, then we place a gray border around it and around any arrows emitted from it to indicate exposure to the linked node's COVID-19 guideline j
。 If there is no link into node j, no other node will be exposed to
its COVID-19 guidance.
Each page (node) can link to other different pages, but unrelated links are filtered out, as we explained in the previous work (62) and S1 sections, resulting in a network
of several links per node.
Thus, the Venn diagram shows COVID-19 guided exposure t (golden nodes, each shape representing a separate subject)
for the neutral population.
The gray boundaries in the Venn diagram indicate that the neutral nodes on time they were exposed to came entirely from non-pro-community COVID-19 guidance, that is, the COVID-19 guidance they received came entirely from anti-satellites and/or other neutral particles
.
For example, 12 were only exposed to COVID-19 guidance from anti-node 14; Therefore, 12 is in the anti-only gray border area
.
Twelve has a link from 11 to it, so 11 is exposed to 12 of COVID-19 guidance, but this does not affect 12 itself
.

Given Figure 1A, we are not trying to assign a numerical value to the scores of each piece of scientific truth guided by COVID-19, thus complicating
this article.
In any case, such numbers are unreliable, as even the craziest anti-content can contain fragments
of reality.
For example, the false story that semiconductor chips are being injected with the COVID-19 vaccine actually has to do with a real science: a peer-reviewed publication in a top scientific journal in 2019 showed that nanoscale semiconductor structures (quantum dots) can be used as markers for injectable vaccines (66, 67).
The error, therefore, is simply that they are not used in this way, not that they are scientifically impossible
.
With that in mind, we took a simpler approach: Our team's experience analyzing all of this community content on a daily basis shows that COVID-19 content published by public research institutions promotes the best science as expected, while anti-government groups oppose it
.
The initial post of a neutral community may fall somewhere between these two extremes, but will often be further downgraded due to non-scientific comments and replies from non-expert page members, so it won't end up being the definitive best scientific guide
.
This means that we can reasonably reserve the label of "best scientific guidance" for guidance from professionals
.
While this point may be further refined in the future, we note that even though a small percentage of our classification of communities and content is wrong, our main conclusions remain unchanged because they only depend on relative numbers
.
We explicitly check the robustness
of the results by simulating errors in the classification.
We randomly picked ? Navigation links across the entire network will be removed
in 1% increments across COVID-19.
On average, removing 15% of the links will only produce a 5% difference in the amplitude of the transmitter-receiver curve, a maximum difference of 25%, and the general curve shape is preserved
.
In addition, although estimates of some parameter values may vary greatly, the general curve shape is preserved (see section S5 for
details).
Therefore, any conclusions that rely on the model are robust to noise-induced fluctuations and large variations in parameter estimates
.
We recognize that the best scientific guidance changes over time and may eventually prove wrong, but this seems to happen rarely
.

Now, we present the empirical findings of exposure dynamics that
we observed at different aggregation levels.
The analytical data was collected from December 2019 to August
2020.
Figure 1B shows how the initial conversation about COVID-19 guidelines began, mostly between anti-communities, long before the official declaration of the pandemic (March 11, 2020
).
It is constructed in the filtered version of Figure 1A: The link appears only in Figure 1B when one of the nodes (communities) to which it is connected presents COVID-19 guidance
during that time interval.
It displays the largest connected components
.
Since we use the ForceAtlas2 layout algorithm, the observed separation is self-organizing, and proximity represents stronger interlinkage, i.
e.
more linked nodes I also have j and its neighbors, closer to the visual node I Node j will be reached (see Section S4
).
It reveals how quickly the anti-community (red node) affects the system, and that neutrals (i.
e.
nodes that are not red or dark blue, for example, the nurturing community is a pale blue node) will also be selected or attached to themselves
.
The pro-community (dark blue) later enters and forms its own realm
.
This view of supporting anti-isolation suggests that when the coupling around COVID-19 originates from Figure 1C to Figure 1D in early 2020, as shown in Figure 1B.
Inch Figure 1 (C and D), only the largest components of the network are shown: it contains 91.
96% of the nodes in the system and 99.
93%.
All side figures in Figure 2C systems have 87.
24% nodes and 99.
94% edge Figures 4D

Figure 1D, prior to the introduction of the COVID-19 vaccine, provides a complete system-level view of all aspects of the flow of the COVID-19 guidelines (panels C and D Figure 1 are equivalent to Figure 1A, ignoring the yellow shadow and therefore similar to a road network that does not consider traffic, while Figure 1B is a subset of the roads that carry traffic
).
The observable changes in Figures 1C to 1D show that in the year leading up to the vaccine, anti-community subgroups not only tightened internally within the year of greatest social uncertainty, but neutrals were pulled and/or closer to anti-s, but neutral groups like parent-child communities (pale blue nodes) also tightened internally
.
The post-vaccination version visually resembles the network in Figure 1D (see Section S2
).

This has the key consequence that when the purchase of COVID-19 vaccines becomes critical (i.
e.
December 2020), many parents are responsible for their own and their children's health decisions, and older relatives are on the web with counter-spies with extreme views (including distrusting vaccines and refusing masks) and others concerned about non-vaccine and non-COVID-19 conspiracy content, 5G, fluoride, The neutral molecules of chemicals (see Pink Ring Nodes) are more closely related to genetically modified foods, and alternative healthy communities that believe in natural remedies to treat all diseases (see Section S2 for
details).
This close relationship with more extreme communities can be important for public health, as proximity in the ForceAtlas2 network layout indicates stronger interconnectedness (see section S4
).
Therefore, the closer these nodes appear in the space in the network, the more likely they are to share content, thus actually exerting influence
.
In this case, this means that the influence of extreme communities on mainstream communities, including parents, may increase
.

Facebook promotes
the best science COVID-19 guidelines from top to bottom by placing a banner at the top of some pages (i.
e.
, nodes) (see, section S3) that points to the U.
S.
Centers for Disease Control and Prevention.
However, our map Figure 1D clearly shows that these banners appear mainly in the anti-community (red node), and that the anti-government organization is mainly targeted in the gray ellipse Figure 1D (section S3).

As a result, many neutral communities are ignored, however the use of these maps may be avoidable
.

Now, we'll take a closer look at the impact that anti might have on neutral categories of parenting communities and compare
it to the impact that anti might have on a system-wide basis.
Venn Tutu Figure 4A quantifies the extent to which non-pro-communities served as the primary source (i.
e.
, transmitter) of COVID-19-directed neutral communities during periods of maximum social uncertainty prior to vaccine discovery
.
7.
19 million people received COVID-19 guidance from non-pro-communities and only 1.
28 million received COVID-19 guidance
from pro-communities.
The remaining 5.
4 million people are affected by both scenarios, which may still leave them unsure of what to think
.
The results in Figure 4B show that this imbalance is more severe for individuals in the parenting community: a total of 1.
1 million people received COVID-19 guidance from non-pro-communities, but only 503 received COVID-19 guidance
from pro-communities.

To break this down further, Figure 4C shows the long-term exposure
of neutral parenting communities in different types of communities to COVID-19 guidelines.
Since the beginning of January, anti-community guidelines have quickly produced COVID-19 guidelines, which, when combined with a large number of links from the parenting community, have led to a rapid increase in the exposure of the parenting community to the anti-community, as shown in Figure 4C.
Immediately afterward, the exposure rate from guidance from other parenting communities rose rapidly before the influenza pandemic was officially declared, There has also been a reduction
in community exposure to pre-existing non-COVID-19 diseases such as Asperger's syndrome and cancer.
Throughout the period, these high-level contacts from opposition and other parenting groups persisted
.
In stark contrast, exposure from pro-community has never shown any strong reaction and remains low
.
Section S5 shows that these plots are statistically significant
compared to the zero model of selecting a random network.
This means that we can reject assumptions
that the microstructure of the exposed network is irrelevant.
In summary, exposing the complexity of the network (Figure 1) is really the key
to understanding the dynamics of exposure over time.

The findings paint a picture of the pre-vaccination phase: In early January 2020, individuals in neutral parenting and other mainstream communities began to become aware of COVID-19 guidance from the anti-community, and then they quietly reviewed it, perhaps interacting in private groups or apps like WhatsApp, or communicating
with others offline.
By mid-February, they felt empowered to produce and share their own COVID-19 guidelines
with similar communities.
At the same time, they received only the best scientific guidance from the minimum pro-community (the dark blue curve is close to zero
).
They don't have a strong tendency to connect more with other pro-communities, probably because they have received guidance from other neutral communities that have similar interests (such as parenting) and that they think they can identify with or even trust them
more.

The findings also suggest that a missed opportunity for
intervention emerged very early in 2020.
Although in January 2020, the likelihood of providing more direct information to anti-government organizations as a guidance transmitter rises (red curve, Figure 4C) taking into account their positive opposition and possible backlash, it has been observed that in February 2020, the rise of other parenting communities as mentoring initiators (light blue curve) indicates, The best scientific COVID-19 guidelines from public research institutions may be tailored around hot topics within the parenting community at the time (which can be read from their pages), so use maps to be introduced proportionally in Figure 1B

What to do if there are also questions about other types of potential interventions
.
For example, will comprehensive interventions across all neutral categories reduce subsequent peak exposure to COVID-19 guidelines in non-pro-communities, and subsequent persistence in 2020 (Figure 4C)? The true impact of any intervention must ultimately be tested empirically
.
However, comparative discussion can benefit from an accompanying mathematical equation that reproduces exposure dynamics on a global scale and is so transparent that it is possible to gain simple insight into this "hypothetical" situation
.

Equation 2 represents such an equation at the general level of all neutrals, pro-and conspirators, as its more approximate version gives Equation 1 Figures 3 (C and D) confirm that Equation 2 has good agreement
with empirically observed exposure dynamics at this overall level.
A more similar statement is that Equation 1, too, produces acceptable consistency, but as expected, the goodness-of-fit statistic is low and therefore does not show
.
Estimate the mathematical model parameters resulting from the number of points of each curve using a full data set; Using k-Section S5 discusses the folding cross-validation and hold-and-hold validation sets
when estimating model parameters.
In contrast, the result of the zero model is that Figures 3 (A and C) show what happened in Figures 3 (B and D) if random empirical data
is used.
The observed inconsistency of this empty model is noteworthy, not only because its predictions are far from empirical data, but also because the construction of the zero model provides a rather demanding comparison: instead of randomizing (shuffling) all types of nodes [which does produce curves very different from those in the curve] Figure 3 (B and D)], we are only random within the type (shuffling), that is, we are in anti, respectively, Pro and each of the 12 neutral subcategories undergoes a random shuffle
.
Thus, this empty model contains exactly the same number of nodes in each subcategory as the empirical network: therefore, for the empty model, the network visually looks the same as the real model, because the color of the category is preserved, but the node names are shuffled
.
Repeat this operation 1000 times to obtain the frequency band shown in Figure 3 (A and C), representing the mean and 1 standard deviation
.
These bands differ significantly from empirical data (see supplementary information and mathematical and data reproduction files for
details).
This shows the importance of the actual link and the entire network in determining the online exposure dynamics, i.
e.
the observed dynamics are not a simple result
of the number of nodes of each type.
This also suggests that Equation 2 captures the true node link characteristics of an empirical network rather than simply reflecting the relative subpopulation size, and therefore, in solving the problem about online exposure, confirms the importance of
understanding the real network.

We can now use this mathematical equation to explore hypothetical interventions
.
Because we are only exploring qualitative results, we take Equation 2 (i.
e.
, .

Equation 1) because its behavior is completely solvable and understandable without the need for any computer
.
Figure 5 (A to D) shows a rough estimate of the prediction of future behavior Equation 1 uses currents R(t), B(t), and G(t) initial conditions and values under different coupling conditions (see Section S6 and Figures S13 to S15 for details
).
Based on the current hesitation about vaccines and mask wear, we assume that public research institutions have reached their maximum capacity to promote the best scientific guidance; Therefore, B(t) remains unchanged
.
Figure 5A shows what happens when the future coupling between anti, pro, and neutral is all positive (i.
e.
, positive feedback): G(t) initially peaks before reaching higher values
.
In Figure 5B, neutral and pro have negative coupling (i.
e.
negative feedback): this results in a significant upgrade
of G(t).
In Figure 5C, all couplings are negative: G(t) drops
significantly.
In Figure 5D, neutral particles and antiions have a unique positive coupling: R(t)→0, but G( t) Remain high for a long time
.
Thus, these different projections for the future provide a framework
for comparing the pros and cons of different possible interventions.

To end our analytical loop, these mathematical predictions can be made from Figure 1 using the physics of renormalization in which communities of the 12 neutral subcategories of inverse and affinity each aggregate into their own community "spheres" (see Figure 5E).
Figure 5F then shows the effect of removing anti from Figure 5E, therefore, where does the imitation of Figure 5D R(t)→0.
The comparison is only valid for a short time, Because we don't allow the entire network to be tweaked or rewired after cutting off all backlinks
.
Figure 5F shows that pro-communities still won't sit at the center of this online world because of the multifaceted interactions
with neutral sub-category communities, especially sports communities (dark green balls).
In fact, both neutrals and professionals are in Figure 5F in general with Figure 5D where G(t) is still high, and can be compared with B ( t) Despite R(t)→0

Our findings suggest that anti-communities began to dominate before the COVID-19 flu pandemic was officially announced, while neutral communities, such as parenting, subsequently moved closer to extreme communities and were therefore highly sensitive to
their content.
Back in January 2020, the parenting community received COVID-19 guidance from the anti-community for the
first time.
This continued until after the official announcement of the pandemic, the parenting community felt confident enough to start adding its own guidance
to the conversation.
Throughout, pro-community guidance remains low, consistent with the parenting community seeking other sources
.

To complement our empirical analysis, we developed a simple generative mathematical model that captures the interactions
between groups that emit and/or receive guidance at the system level.
It makes it easy to explore what-if scenarios and, therefore, can roughly predict the behavioral response
at the tipping point of different intervention strategies.
The combination of network mapping and models suggests that there are more possible ways to alter the conversation
than removing all the extreme elements from the system.
The resulting Figures 5 (D and F) show that eliminating all extremes may not even be the most appropriate solution
.
In any case, such a purge may be considered harsh; They run counter to the concept of open participation and can disrupt business models
that maximize the number of users.
Figure 5 (D and F) shows the impact
of removing counterintelligence on affinity neutral communities.
In Figure 5F, we see that the pro is not located in the center of the system (i.
e.
, the different measurements of the center of the system are not contained near or inside the pro supernode), and that the center of the system is located between
the pro and a neutral community (movement).
In Figure 5D, we see R(t)→0, G(t).
→0 also happens, meaning that the removal of anti-content may stimulate self-removal by neutral communities, suggesting that these self-removed neutral communities may have less
access to the best scientific COVID-19 guidance.
Increasing their connections with other communities in pro-communities will pull other supernodes toward professionals on Figure 5 (E and F) and move the system center closer
to the professional community.
Reducing the impact of anti-community by removing anti-community links with other communities also helps to move system centers from anti-government organizations to other communities Figure 5E
.
Because our model can be interpreted at different scales, including community communities, it can be applied to multiple platforms with built-in community capabilities and can be used to address the more pervasive online misinformation issues
beyond COVID-19 and vaccinations.

Of course, our study also has limitations, which means there is an opportunity for further research
.
We are limited to pages in languages that researchers can read; As a result, we missed out on more insights
from languages like Mandarin, Hindi, and Arabic.
In addition, Facebook may be the world's leading social network (57), and its users may not be representative of a country's population
.
It's an open question of how our results generalize to those who use the internet at low
rates.
For example, in Pakistan and Belize, Facebook is the leading social network (57); However, only 17% and 47% use the internet (68).
In addition, exposure dynamics may be influenced by small groups of so-called chaos agents used by organizations or governments (61).
However, we note that the social media community tends to self-regulate
troll-like behavior.
Another limitation is that there are many other social media platforms that can have such debates
.
Individuals may be receiving COVID-19 guidelines that they see on Facebook and discussing
them on any other social media site.
However, we believe that similar behavior will occur on any social media platform that the community can grow, and Facebook is indeed one
of the largest.

The method here follows our previous work (62).
Facebook pages consist of nodes in our data, and each link represents the case when
one page recommends another page to its members.
This avoids the need to identify personal account information, which is prohibited by Facebook's public API Terms of
Service.
This process begins with some way discussing the artificial identification of page seeds for vaccines/vaccinations, and then the connections of those pages to other pages are indexed
.
These pages were identified in 2018 and 2019 by searching Facebook pages using keywords and phrases related to vaccines
.
The findings were reviewed through human coding and computer-aided filters, and at least two different researchers then independently classified
each node.
When there was a disagreement, they discussed and in all cases reached an agreement
.
This process is repeated two more times to get the final list of candidate nodes and the links
between them.
In order to categorize the page, the post, the About section, and the categories of the self-description of the page are reviewed
.
To be categorized as pro or no, at least 2 of the last 25 posts must involve a vaccination debate or page or a self-identifying page about whether a section is for or against vaccination
.
To be classified as neutral, 5 of the last 25 posts must mention the vaccination debate, but the page does not explicitly state a position for or opposition, or the About section explicitly declares that the page is neutral in the debate, or rather, none of the last 25 pages are about vaccines, but this page claims to be an NGO, a cause, a community, or a grassroots organization
。 Therefore, our dataset only contains Facebook users
.
Our target demographics include not only those who specialize in vaccine information, but also nonprofits, public figures, government organizations, medical companies, local businesses, and more
.
Of course, we can define different nodes and links, and our dataset ends up being an imperfect sample
of some larger "right" network.
To help alleviate this, we repeated the process of manually identifying page initial torrents several times, with the goal of diversifying the torrents as much as possible by including pages published in different languages, pages focused on different geographic locations
, and pages from managers from multiple countries (Section S2).
Only those articles were written in languages that researchers could read, such as English, French, Spanish, Italian, Dutch, and Russian
.
Determining whether a post is satirical or sarcastic, fake, or villain-like (61) is a very difficult task even for subject experts; Free, fully out-of-the-box machine learning language models are now also available to generate real-world vaccine error messages (69).
However, these social media communities tend to self-regulate
bots or troll-like behavior.
The extremely difficult task of quantifying the authenticity and intent of these posts will serve as the subject of
another study.

In order to determine who is launching and who is receiving guidance from COVID-19, we must first determine which posts explicitly discuss it
.
For each post, the post message, description, image text, and link text are combined into a single string in which the situation is ignored
.
We then made a list of strings to search for these posts, for example, "corona virus", "covid", "19ncov" and other terms
.
Because we want these terms to be as flexible as possible, we use regular expressions, so terms like "corona virus" become "(c | k | [ （]））+orona（no | [[:p unct:]|\\s){,4}(virus\>| vírus)", which captures common spelling mistakes and punctuation as well as non-English languages ("vírus" for Portuguese, virus not only for English, but also for Italian, Spanish, French, etc.
).

Therefore, our method catches intentional misspellings, ignores punctuation, adds spaces between words to avoid filtering, and overrides languages
other than English.
We then filter out those posts related to COVID-19 by using filters and use that information to determine which pages issued COVID-19 guidelines at some point
.
With this information, we are able to generate the filtering structure of the network (for example
.
Figure 1A) One of the links only exists at a given time if one of the nodes to which it is connected is generating COVID-19 guidance
at the time.