Click, Share & Conspire—The Way Out of Misinformation Maze

Table of Links

Abstract

Transnational Network Dynamics of Problematic Information Diffusion

Theoretical Framework

Methodology

Statistical Analysis and Results

Discussion and Limitations

Conclusions and Policy Implications, and References

Methodology

Data, sampling and measures

The units of analysis are public Facebook groups that shared native or externally hosted audiovisual content that included mentions to the key terms Médicos por la Verdad or Natalia Prego (one of the group’s founders and most vocal member) for the MPV case in Spanish language, and Didier Raoult or hydroxychloroquine, for the DR case in French language. The data were queried using CrowdTangle, a tool developed by Facebook that allows researchers to obtain and analyze public data shared on the social media platform (Bleakley, 2021). Our case studies center therefore on the Facebook network of groups that shared videos featuring problematic information across the Spanish and French speaking postcolonial world. Combined, the Facebook groups amassed audiences exceeding several orders of magnitude.

Case 1: Médicos por la Verdad and the false pandemic

The first case looks at conspiracy theories in Spanish that were promoted by the organization Médicos por la Verdad (MPV), based in Spain and accruing significant reception across Latin American countries throughout the pandemic (Maldita.es, 2021). Mentions of MPV reached close to 16,000 across public Facebook groups from Spain and Latin America during the pandemic, and in 2020 alone local chapters of the organization began emerging in more than 10 countries (Gardel, 2020). Due to their viral audiovisual diffusion on social media (Knuutila et al., 2020), MPV conspiracies have posed a significant increase in public health risks in Latin America and other Hispanic diasporas across the continent by promoting the use of unauthorized medicine, challenging the efficacy of face masks, doubting the accuracy of testing kits, all while arguing that the pandemic is a “false pandemic” or a “plandemic”.

The study analyzes Facebook data of non-geolocated posts made by public groups or pages containing mentions of a prominent COVID-19 conspiracy organization from Spain and of one of its most vocal members. The Facebook data is non-geolocated to allow for posts that lack this metadatum to also be included in the query results. The complete dataset contains posts from February 1, 2020, until October 1, 2021, with video content (native, native live, and externally hosted videos) for a total of around 15,336 posts. These data are used to build a bipartite network where nodes are Facebook groups and video URLs, and a directed link is formed between them when a group or page shares a URL. The full network comprises 5781 unique nodes and 2243 unique video URLs. Based on the number of times a node shares a video, a cutoff threshold is set to retain only those groups that participated in video co-sharing behavior at least twice during the sampling period to ensure minimal sharing behavior in the subsequent analysis. This operation yields a directed bipartite network composed of 2168 nodes and 1968 unique video URLs.

This bipartite network is then projected onto a one-mode co-share undirected network where Facebook groups are nodes and ties are formed between two nodes if they have co-shared the same video URL. To focus only on the super-spreader core of the network, we trimmed it based on edge weight (i.e., the times that two groups co-shared a video), retaining only the edges whose weight belongs to the upper percentile of the edge weight distribution. The resulting undirected one-mode network contains 518 nodes and 2390 edges.

The resulting super-spreader network was then manually labeled to classify each node according to their geographical location, the language of their posts and their category as given by the overarching theme of the group’s activity and content. Due to the limited availability of metadata about several of the Facebook groups, a series of indicators were used to assign the labels. For the language label, indicators in the group’s name, description or recent activity were used to determine the most common (if not unique) language used and the corresponding language ISO codes were used as labels. The country or domain label was assigned by looking at indicators like the mention of a geographical location (country, city, town, region, etc.) in the group’s name or description, or, alternatively, by the nationality of the group's administrators, depending on availability, after which the country ISO codes were employed as labels. Finally, the group category label was assigned by reviewing a mixture of indicators whenever they were found in the group’s name, the group’s description and/or most recent group’s posts and activity, depending on their availability for each group. Then, each group was assigned to five possible categories: Conspiracy Theories, Media, Politics, Religion & Spirituality, and Community/Other. After removing those groups whose country was not included in the SCI (such as Russia, Cuba or Venezuela) and those that could not be reasonably inferred, we retained 467 nodes and 2170 edges. Figures 1, 2 and 3 showcase the distribution of nodal attributes in the MPV network data.

Case 2: Didier Raoult and the alleged effectiveness of hydroxychloroquine

The second case looks at problematic information in French that was ignited by early declarations made public by the controversial French microbiologist Didier Raoult, who promoted the use of hydroxychloroquine to effectively treat patients infected with the coronavirus (Sayare, 2020). The data analyzed for this case comprises Facebook data of non-geolocated posts made by public groups containing mentions of “Didier Raoult” and “Hydroxychloroquine” from February 1, 2020, until October 1, 2021, with audiovisual content (native, native live, and externally hosted videos) for a total of 11,122 posts.

Following the exact same procedure as with the MPV case data, we built a bipartite network where nodes were Facebook groups and video URLs, and a directed link is formed between them when a group or page shares a URL. The full network comprised 1170 unique nodes and 1195 unique video URLs. This bipartite network was also projected onto a one-mode co-share undirected network where groups and pages are nodes and ties are formed between two nodes if they have co-shared the same video URL. The largest component of the resulting undirected one-mode network contained 1156 nodes and 120274 edges. As in the MPV case, edges were trimmed based on their weight to retain only the upper percentile of heaviest edges, that is, those that account for most of the co-sharing behavior in the network. After removing nodes whose country could not reasonably be inferred, the resulting subset of the network comprised 454 nodes and 1191 edges.

This subset was also manually labeled to classify each node according to their geographical location, the language of their posts and their category as given by the theme of their activity. Due to the same limitations in available information in several of the Facebook groups, a series of heuristics were used to assign the labels. For the language label, cues in the group’s name, description or recent activity were used to determine the most common (if not unique) language used and the corresponding language ISO codes were used as label. The country or domain label was assigned by looking at cues like the mention of a geographical location (country, city, town, region, etc.) in the group’s name or description, or, alternatively, by the nationality of the group's administrators, depending on availability, after which the country ISO codes were employed as labels. Finally, the group category label was assigned by reviewing a mixture of cues available in the group’s name, description or most recent activity and assigned to five possible categories: Conspiracy Theories, Media, Politics, and Community/Other (unlike the MPV data, there were no groups with overarching Religion & Spirituality themes in the DR data). Figures 4, 5 and 6 showcase the distribution of nodal attributes in the DR network data.