In an online community, members can post their thoughts regarding a health condition or personal problems for other members to read and comment on. A support group communication episode is structured hierarchically, beginning with a person’s need for information or help (ie, locutor), to which any other member of the group may respond. Regarding the term comment, we also say post, utterance, or locution [1]. The reactions to a comment, if there are a fair number, would look like a branch of a tree split into ever more twigs. We refer to the whole branch as a bundle or a thread.

Some questions arising from this are evident: What are the influencing factors in the development of mental conditions? How do posts indicate changes in prognoses? But there are also other questions: Do posters stray? That is, do they prioritize other illnesses above or below their mental condition?

For instance, one of the key results in Cavazos-Rehg et al’s [2] study of Twitter was that depression and related factors were the subjects that drew the most attention by a large margin.

An origin or first post typically poses a question addressed to somebody’s declarative or procedural knowledge. Briefly speaking, declarative knowledge is a type of knowledge that can be verbalized and, on that basis, taught. Procedural knowledge is a person’s experience of how something is working [3]. Another typical demand in a first post would be for emotional support [4]. Griffiths et al [5] proclaimed to have captured the process rather than the end state of seeking support. Successful online support groups were also written about in scholarly journals, contributing to the impression that mental health support groups are an asset to a health care system [6,7].

In online communities for mental health symptoms (OCMHs), posts to groups and related comments have been deemed to sometimes have beneficial health consequences. Facebook-based social support, for example, was found to affect general health, mental illness, and well-being [4,8,9].

Regarding online communities, Nimrod [9] stated the following:

Employing qualitative methods, Takahashi et al [10] found both better and worse states among the members of depression help groups. The authors especially highlighted the possibility of a downward spiral of a worsening state among groups.

Deceptive or erroneous, rather than soothing or enlightening, information is also feared [11]. OCMHs may contain messages that contradict medical advice or reinforce unhealthy behaviors, such as those seen in Gavin and colleagues’ study [12], where anorexia and the misuse of diabetes medication to lower body weight were advocated on a webpage. The presence of misinformation was previously demonstrated in the context of antidepressant use on Instagram [13].

In a study of 6.5 million interactions generated by 500 posts, phatic posts constituted the strongest predictor of interactions, followed by posts with a positive emotional valence. Half of the posts were about social relations, and more than one-quarter (28%) consisted of health misinformation [14]. The article by Beck et al [15] reported on a content analysis that revealed emotional support to be the most frequently sought type of support and, at the same time, a majority of messages related to task rather than relation. Rueger and colleagues [16] reported that advice is more often followed when advice seekers perceive more similarities between themselves and advice givers (see Sillence [17]). The classical term for this is selective perception. If the perception of similarities between advice seekers and givers exerted some influence on advice acceptance, an adaptation by seekers regarding the type of internet service they chose was not observed [18].

A controversial subject is the correctness of internet content. However, recent articles posit the internet’s capacity to work against misinformation [19]. It is, however, unlikely that everyone who comes into contact with the misinformation will also be made aware of the correction. Different tensions may arise when the recognition of the increasing role of social media in health information consumption leads to the choice of traditional (ie, nonsocial) media for research [20].

The consequences of taking part in online communities can be differentiated into communicative and medical. Communicative consequences include ease of access to knowledge, enabling contact, or helping to find other people in a similar situation. These consequences are often considered the primary effects, as in a systematic review study by Moorhead et al [21]. The research methods in this area cross over into language use and mood analysis [22,23]. Medical consequences are the health effects, which are considered the secondary effects. One of the most engaging content analyses of online depression communities [9] summarized the influences of participation in terms of ameliorating the understanding of participants’ conditions and encouraging them to fight depression.

OCMHs can be moderated by health professionals; for instance, a health professional could prevent detrimental activity by establishing behavioral- and content-related rules and could moderate threads (eg, there could be restrictions on the mention of drugs and related matters). Moreover, they could censure or remove individuals from the OCMH [5]. Health professionals are defined in this study as people with a documented academic or professional background in the health field, but not restricted to mental health. These include psychiatrists or psychologists as well as nurses and general practitioners.

Very few studies have been conducted concerning the role of moderators and administrators (admins) in countering misinformation. Recent studies on moderation have shown that this type of intervention could improve the quality of online discourse (eg, Wadden et al [24]). However, these studies did not specifically investigate the impact of misinformation and its uncovering.

Italy was the first Western country to experience a nationwide lockdown and this changed health priorities, with public health authorities advising the public to limit the use of health services [25]. Psychiatric emergency department admissions and referrals, in fact, diminished [26,27]. According to recent literature, this has led to the transition of many aspects of pre–COVID-19 life to a web-based format [28].

Although many studies focused on addressing the impact of social media during the pandemic (eg, the impact on panic [29] or on educating the public on prevention measures [30]), the reverse has not yet been investigated. A few studies have examined the characteristics of informational support or public sentiment related to COVID-19 (eg, Boon-Itt and Skunkan [31]), but a broader view on the impact on mental health conversations is still lacking. To the best of our knowledge, no study has been conducted to investigate whether people with mental health disorders turned more often to OCMHs or whether they gained or lost trust in the official health care system and encountered a higher amount of erroneous information.

An interesting question that arises is whether social media subjects changed as a response to the pandemic situation and whether the management of people’s mental conditions adjusted to the challenging situation.

Only one study investigated whether social media use increased after the onset of the COVID-19 pandemic; however, the community that was investigated did not specifically include people suffering from mental health disorders [32].

In this paper, we seek to describe how posts to OCMHs and their comments can form deceptive or erroneous information in terms of false declarative or procedural knowledge, wrong judgments, and motivations for detrimental decisions.

Furthermore, an exploratory research question asks whether misleading content would be distributed unequally at the different stages of illness trajectory. It is hypothesized that advice seeking or giving related to treatment, versus diagnosis or symptoms, will have the highest prevalence of misinformation.

Our study aims to do the following:

This could be read as the foundation of a study that will test a complex causal model with demanding contemporary statistics, such as structural equation modeling. Instead, the method that will be used is a quantitative longitudinal content analysis [33]; 3 years of posts on a Facebook health community page will be analyzed, including some background and interpretive links.

Content analysis is a method to identify, record, and aggregate predefined elements of communication content [33,34]. The method assumes that the vast biological, educational, and social differences among the coders will not lead to a completely different understanding of similar text or other material. Selective perception has to be downgraded to the degree that different coders with different experience are able to understand and record communication content reliably. This is achieved by coder training and a collection of rules to understand and record the content; the rules are referred to as the codebook (Multimedia Appendix 1). The study will use utterances from Facebook OCMHs as the sampling frame. Facebook was chosen for several reasons: it is the largest social media company, it has a large presence of support groups, and it allows for easy retrieval and monitoring of member discussions, together with the number of threads present, compared to other types of OCMHs, such as blogs and forums.

The guidelines outlined in previous social media research informed this study’s procedures [35]. Permission to conduct research was requested from admins or moderators of private and public Facebook groups. A group privacy setting of “private” means that while anyone can see the group, approval by an admin or another member is required to join; the content on a group’s wall is only visible to members [36]. In public groups, any Facebook user can join the group (ie, no admin permission is required).

Furthermore, regarding data that will be reported in scientific publications, we will remove any real names or other personal information, as well as usernames, that could potentially result in a breach of anonymity and privacy that would reveal any information that could be attributed to a single individual (eg, photographs and locations) [37]. In addition, researchers abstained from any communication or interaction with the individuals in the groups. Ethics approval was obtained by the ethics committee of the Università della Svizzera italiana on April 19, 2021 (CE 2021-4).

The first draft of the codebook was developed based on the categories identified in previously published works, for example, with respect to the distinction between information seeking and giving versus emotional support (eg, Greene et al [38] and Lerman et al [39]). However, in the later stages of codebook development, an inductive approach was applied as potential themes arose from the reading of data [40].

The second step in developing these coding schemes involved applying the codebook to a sample of Facebook threads in online support groups. This was done in an iterative process. This coding often resulted in the identification of further coding criteria linked to the variables of interest.

Usually, content analysis coders do not have to have special skills. In this case, however, coders were required to have clinical experience and familiarity with mental conditions, and they were selected accordingly. Two coders were hired; both of them are psychologists with a background in health communication.

The two coders familiarized themselves with the coding procedure, the variables of interest, and the theoretical constructs underlying the codebook (eg, the concept of misinformation and the distinction between procedural and declarative knowledge). The coders were then assigned a number of threads (ie, consecutive posts) to code independently and to find shortcomings in the coding scheme. Codings were discussed among the coders and with the project management team. The iterative testing took place from November 2021 to January 2022.

When disagreement between coders and the management team had dropped to acceptable levels, the coders coded another 60 utterances for a reliability check, calculating a preliminary Krippendorff α. Coding problems and disagreements, as well as shortcomings in the codebook, were discussed, and the instrument was revised once more. In the next step, the two coders coded an additional 11 threads (61 utterances) and calculated the Krippendorff α for each variable coded (see the Results section).

The project manager downloaded the threads and inserted them into a Microsoft Word file that already matched utterances (ie, posts and comments) with the sorting variable’s number (ie, the identifier that will be used later in the coding phase). In this step, emojis were replaced with keyboard symbols; see Multimedia Appendix 2 for examples of coding. A Microsoft Excel file was then created with the columns related to the formal variables (eg, date and the number of total comments in the post) already filled out.

Data were then coded by human coders, who entered their ciphers and letters into a data file in Microsoft Excel. Coders could correct obvious errors in sorting variables. Coders had help filling out the columns through the use of the IF function in Excel (eg, “NA” [not applicable] was automatically inserted for variables of advice giving when the utterance was coded as advice seeking).

After coding was complete, the data were transferred into the data matrix of a statistics program (SPSS); data cleaning was performed to make sure that there were no undefined codes in the file and that subgroup analyses were based on the correct filters and showed a reasonable number of posts or utterances. Generally speaking, content analysis can be based on counting different units. We will distinguish the individual utterances at a detailed level, which will make up the data matrix. 

Variables were coded on the locutor level and for all utterances, meaning that they do not form their own level. Every utterance has exactly one locutor. When a new utterance is added, a new coding line is opened, and the new utterance is coded in the new line under the given thread or within the same bundle. When the data file is cleaned, different entries for the same locutor are taken together.

We first demonstrated the usability of the proposed systematic framework with 11 threads (61 utterances) coded independently by two coders. The KALPHA macro for SPSS for Windows (version 26; IBM Corp) [46] was used to calculate the Krippendorff α; the coders reached intercoder reliability (α=.89, range .73-1). Then, 1534 units of analysis (144 threads) from two Facebook groups were retrieved and analyzed by two coders. Intercoder reliability was calculated on 293 units (19.10% of the total sample; Krippendorff α=.94, range .72-1).

The α values were calculated for the following: locutor gender (α=1), locutor status (α=1), medical-scientific qualification of locutor (α=1), seeking versus giving advice (α=1), motivation for seeking advice (α=.93), action through giving locution (α=.95), type of illness (α=.95), treatment options (α=.98), treatment evaluation (α=.93), adverse effects of treatment (α=.93), treatment interruption (α=.94), health professional mentioned (α=.97), sentiment toward health professional (α=.72), doctor-patient relationship (α=.95), argument quality (α=.93), misinformation (α=.90), misinformation correction (α=.91), and illness trajectory and themes (α=.98).

In total, we have contacted 21 groups; five groups were public but were still asked for approval. Of these groups, 14 agreed to participate in the study. Among those who agreed to participate, according to the selection criteria, we chose two groups: one moderated by a health professional (n=12,058 members) and the other led by peers (n=5598). Information on the number of participants was retrieved in November 2021.

Analytic procedures will be determined in detail. In addition to descriptive statistics of the variables considered, the analysis will combine different units of analysis levels (eg, utterances in relationship to threads). The independent variables will be recoded from the locutor variables and the locutor role of seeking advice versus giving advice: put simply, the seeking of posters, their qualifications, and the illnesses by which they are afflicted. Interfering variables will be the seeking and giving dichotomy, the type of illness, the treatment mentioned, the sentiment toward the health professional, and the argument quality. Misinformation and misinformation correction will be the principal dependent variables.

The study will contribute to investigating potential detrimental effects, and possible mitigating factors, of OCMHs.

Findings will be explained by the concepts of health literacy and health empowerment. In the context of the two concepts, certain communication elements of health subjects are associated with erroneous utterances. Working with the two concepts presents the problem of a striking plurality of concepts as well as operationalization. We, therefore, will not link measures or data, but will look at the origin of erroneous health beliefs and detrimental decisions and will provide the link by way of interpretation.

Consequently, health literacy and empowerment were not coded anywhere in this content analysis. In spite of this, some understanding of the theories behind data collection might help us understand what is being measured and why. Therefore, a few more words on health literacy and empowerment are added here. Health literacy can be defined as the “capacity of individuals to obtain, process, and understand the basic information and services needed to make appropriate health decisions” [47]. Knowledge is not mentioned but is identified in the same writing as a moderator of health literacy. It is certainly a product of literacy and a facilitator of further knowledge acquisition [48]. This position of knowledge and health literacy gives us reason to consider the discussion, application, and evaluation of knowledge as major elements of erroneous consequences of health literacy.

Empowerment, in general, was defined as regaining mastery of one’s life [49]. That concise formulation can be specifically applied to health empowerment, which would be limited to matters of health, including prevention, care, disease, and treatment. Empowerment becomes harmful when thinking about and discussing competence in making medical decisions. Competence has the same meaning as in legal provisions or social norms: granting the power to make decisions. Competence in another sense refers to the ability to make decisions. That sense of the term belongs to health literacy rather than empowerment. Expressed in modal verbs, competence regarding legal and social norms that is related to empowerment is about what you may (ie, are allowed to) do; the other sense is related to health literacy and is about what you can (ie, are able to) do. Agreeing to nonprofessional decision-making (ie, supporting a patient in the belief that he does not need to seek medical consultation) belongs here. Keeping a patient away from a doctor when he or she should see one is the second element of deleterious consequences, this one resulting from an overestimation of empowerment. An example is the linkage between empowerment and using internet health services [50].

Aside from a few analyses that will compare distinct threads, individual utterances will constitute the unit of analysis in most cases [51]. More specifically, we expect misinformation to be more prevalent at specific stages of the illness trajectory, such as the treatment stage, and that professionally moderated OCMHs will present a lower number of misleading messages or negative comments toward health professionals and treatment options.

Furthermore, in addition to COVID-19 impacting the subjects of discussion, we expect the highest amount of misinformation to occur during the pandemic year 2020.

To our knowledge, this is the first piece of research that will examine OCMHs using the methodology of longitudinal quantitative content analysis. The study will permit us to overcome the limitations of automatic text analysis. However, limitations should be taken into consideration, for instance, female and male genders cannot always be reliably inferred from names, and this categorization limits the variety of gender identities.

We are not yet aware about the impact of OCMH messages on members’ intentions and behaviors in terms of professional help seeking or attitudes toward health professionals and treatment options. Future research may wish to explore this aspect.