Data collection and preprocessing

During the 2022 Brazilian presidential election (from October to November 2022), a national survey was conducted by NetQuest, an international survey company. The data were gathered using NetQuest’s exclusive online panel, using stratified quota sampling aligned with national distributions of age, gender, and geographic area to ensure national demographic representativeness. Statistical tests indicate that the survey participants constitute a representative sample of the national electorate (see Fig B in S1 Text). Participants were enrolled through a double opt-in process and completed the survey online in return for incentives. Approximately 40 questions and 189 variables regarding socio-demographics, news consumption behaviors, and political predispositions were asked in the survey (see the questionnaire in section 2 in S1 Text). We also requested respondents’ consent to provide their Twitter/X handles. Of the 1,018 respondents, 403 consented, and 271 were verified as existing accounts on Twitter/X, accounting for 26.62% of the whole panel.

Based on the 271 survey respondents, we collected the Twitter/X accounts followed by them (57,645 accounts and 73,755 following pairs were collected) using the Twitter/X API. We use “following” relations to represent individuals’ preference to be exposed to information posted by the accounts they follow. We then identify political influencers from the 57,645 accounts. We define political influencers as a composition of both ordinary citizens and celebrities (e.g., politicians, parties, media outlets, journalists, and individual opinion leaders) who satisfy two conditions: 1) with at least more 1,000 followers and 2) indicate political related information in their profiles or are known political figures [47,48] (see section 3 in S1 Text for more details about the identification of political influencers). After filtering out the survey respondents who did not follow any political influencers, we obtained 204 survey respondents (also referred to as political consumers in the text), 2,307 political influencers (also referred to as political suppliers in the text), and 4,107 following pairs between the two groups. Pearson’s chi-squared tests (for categorical variables) and Mann–Whitney U tests (for discrete variables) comparing the demographic distributions of the 204 subsample with the 1,018 survey respondents do not reject the null hypothesis that the 204 samples are drawn from the same distribution as the survey respondents at the 5% significance level (see Fig C in S1 Text). Therefore, the survey respondents who provided their handles are very likely to form a representative sample of the Brazilian electorate.

This study was reviewed and approved by the Ethics Committee of the Faculty of Arts and Social Sciences, University of Zurich. Informed electronic consent was obtained prior to participation via the NetQuest survey platform. Participants were informed of the study purpose, data processing, voluntary participation, and their right to withdraw at any time. Twitter/X data were collected in accordance with the platform’s Terms of Service and API policies and were limited to publicly available content.

Multi-scale community detection

To analyze the relationships between political consumers and political influencers, we construct a bipartite network , where is the set of ordinary users who primarily consume political information online, and is the set of political influencers who produce such content. The set contains edges that indicate which users follow which influencers. We then project this bipartite network onto the influencer side, resulting in a new graph , where edges in indicate that the connected influencers are followed by at least one common user. The weight w_ij of an edge between influencers s_i and s_j in the projected graph is equal to the number of shared consumers following both influencers. Similarly, we create a consumer network where edges between two consumer nodes represent the number of common influencers they follow.

We then apply a multi-scale community detection approach to the projected influencer graph . Community detection is the process of identifying groups of nodes that are more densely connected to each other than to the rest of the network. It is a common strategy to study phenomena like selective exposure, polarization, and echo chambers [5,27,59,60]. However, most studies detect communities using single-scale approach, such as modularity optimization [45]. But real-world social networks are usually multi-scale and hierarchical, for example, smaller class groups are embedded within the broader structure of a school [36,37]. Multi-scale community detection approaches allow for the discovery of communities at different resolution scales. This flexibility also enables selecting the scales closest to the real situation among all the experimental scales, depending on the research goals [38].

A variety of techniques have been designed for multi-scale community detection, such as methods with adjustable resolution parameters (e.g., [35,39,40]) and methods which automatically detect relevant scales (e.g., [41–43]). Here, we use Markov Stability [35,44] with automatic scale selection [43] and optimization using the Leiden algorithm [46]. We note that our approach is applicable to the results of any multiscale community detection approach. The rationale of the Markov Stability method is to identify communities as groups of nodes where a random walk remains confined over varying time scales, thus revealing community structures at different resolutions. We give a detailed mathematical explanation of the Markov Stability method in section 5 in S1 Text.

Multidimensional annotation of identities

We then annotate four identity attributes of political influencer nodes in the network: political ideology, campaign support, social identity, and account type. The data are manually annotated based on the self-presentation of political and social identities in users’ Twitter/X profiles, combined with a formal list of politicians and parties participating in the election. It is shown that users who disclose their political identities publicly are generally more active than those who do not [63]. Almost all political influencers have profile descriptions.

Political ideology is the most traditional and straightforward indicator of individuals’ political opinions and comprises categories {Left, Right, Center}. During the 2022 Brazilian Presidential Election, Lula and Bolsonaro are the two candidates who represent leftist and rightist leaders, respectively. Nevertheless, it is not necessarily the case that all left-wing individuals support Lula and all right-wing individuals support Bolsonaro, due to increasing cleavages on issues and socio-cultural factors, even though there is still significant overlap. To differentiate between these situations, we also annotate them as “Pro-Lula” or “Pro-Bolsonaro” if they support certain candidates during the election campaign.

Many users also indicate their social identities, as cultural issues have become increasingly related to politics. We identify four social identities most frequently revealed in the profiles: Women (in the context of feminism or supporting women’s rights), Religious, Black, and LGBTQ. For religious, Black, and LGBTQ identities, annotation is based on either users’ self-identification with the group or their expression of support for its rights. Note that, unlike the fixed options of political ideology, social identity encompasses more categories; accordingly, we include only those observed in our samples. The same situation applies to account type, which indicates what the accounts represent. We consider media (including media outlets and journalists), politicians, and individual opinion leaders, as they are the most common accounts that produce political information.

As not all political influencers reveal their identities at all dimensions, we assign those unrevealed profiles as “Unlabeled”. The annotation results on political ideology, campaign support, social identity, and account types for influencers in communities of all scale levels are included in section 4 in S1 Text.

Measurements of selective exposure indices

We introduce five indices – Community Overlap, Identity Diversity, Information Diversity, Structural Integration, and Connectivity Inequality to measure individuals’ selective exposure patterns at different levels. We consider a followship bipartite network between political influencers and survey respondents, and project the network onto both the survey participants’ side and the political influencers’ side. From the projection on the survey participants, we capture the number of influencer communities in which they participate. From the projection on the political influencers, we detect communities and identify the characteristics of influencer communities followed by each survey participant. In this context, we introduce the following measures:

Community overlap.

This index is derived from the network’s projection on the survey participants and reflects the number of influencer communities they engage with. It captures how online users are divided into fragmented communities in which individuals primarily interact with others who share similar interests, and how they could act as bridges between multiple communities. This index is based on the phenomenon of cyberbalkanization introduced by Alstyne and Brynjolfsson in their 2005 study [7] which showed that some internet users are more inclined to have diverse interests and engage with multiple communities, while others prefer to remain within a single community. Survey users who connect with multiple communities have a higher potential to act as bridgers—individuals who consume information from a variety of sources [49]. To quantify survey respondents’ capacity to bridge multiple communities, we introduce the Community Overlap Index.

Community Overlap is measured for each consumer node and at each level and corresponds to the number of influencer communities a consumer belongs to. Each consumer node can follow influencers either in a single community or from multiple communities.

At scale s, for a consumer node , let be an indicator function that is 1 if node v belongs to the hyperedge - a hyperedge represents at least one connection to influencer community i, with K communities in total, and 0 otherwise. The number of hyperedges node v belongs to, i.e., the number of influencer communities that the user follows, is expressed as

(1)

Identity diversity.

This index is measured from the network’s projection on the political influencers and indicates the diversity of identities among the influencers engaged by survey participants. According to social identity theory and social categorization theory, people may act based on inter-group behaviors arising from a sense of belonging to certain social groups [50,51]. It can be applied to the political domain, where political identities advance social identities, such as national identities, partisan affiliations, and political ideologies [52]. The diversity of influencers’ political identity in a community is measured by the Identity Diversity Index.

We annotate identities for influencers on four dimensions: political ideology, campaign support, social identity, and account type, as indicated in their profile descriptions. Influencers who strongly exhibit certain attributes may lead individuals to filter into specific community structures. However, not all influencers disclose identities in all dimensions. To address unlabeled accounts, we apply probability assignment, as described below (also see the percentages of unlabeled identities in Figs E–H in S1 Text). We only measure the diversity of labels on the political ideology dimension as it is the most direct indicator of political identity.

We measure this index using the Gini-Simpson Diversity Index, used in fields such as ecology, sociology, and psychology to measure the diversity of types in a dataset [55]. The Gini-Simpson Diversity Index gives the probability that two entities taken at random from the dataset of interest represent different labels. It ranges from 0 to 1, where 0 represents a totally homogeneous distribution of ideology labels, and 1 represents totally diversified labels. For an influencer community at scale s, the Gini-Simpson Diversity is computed as

(2)

Where L is the number of ideological labels (here: “Left”, “Right” and “Center”), is the number of influencers of label in community , and N is the total number of influencers in community . As not all the influencers are labeled, we apply two approaches to deal with the missing values: in one approach, we assign probabilities of labels to the unlabeled influencer accounts based on the proportions of labels present in each local community; in the other approach, we only consider the influencer accounts that are labeled. The two approaches yield similar results (see section 6 in S1 Text).

Structural integration.

This is a structural measure of the influencer communities engaged by survey participants. Each influencer community occupies a distinctive structural position within the network. Some communities are relatively isolated from others, while some are more integrated into the overall structure. In the co-following influencer network, a more isolated community indicates that it caters to a more targeted consumer group, whereas greater integration suggests exposure to a broader audience.

To quantify this, we introduce the Structural Integration Index. This index is derived from the Normalized Cut, a widely used metric in image segmentation and graph partitioning that evaluates the connectivity between two disjoint subsets of nodes in a network [56]. We employ the Normalized Cut to assess the degree of integration of a given community by comparing its internal and external connectivity. For a community at scale s, the Structural Integration is defined as:

(3)

where is the size of the cut, i.e., the sum of the weights of the edges crossing the boundary of the community , is the sum of the weights of the edges inside of the community, i.e. , and m is the sum total of the edge weights in the influencer network, . The Structural Integration Index reaches a minimum value of 0 and a maximum value of 2 for two degenerate cuts: the empty cut, and the cut containing all edges. In this work, all the values of the index were between 0 and 1. Lower values indicate stronger isolation with the rest of the network, and larger values reflect greater integration, meaning the community has relatively more or stronger connections to other parts of the network. This measure enables us to evaluate whether survey participants are primarily exposed to influencers from structurally peripheral communities or from those that are more centrally embedded in the network.

Connectivity inequality.

This index is also based on the influencer communities but focuses on their internal structures. Some communities exhibit an unequal distribution of (weighted) degrees, with certain nodes having a larger number of connections and potentially playing a more central role, while others display a more even distribution.

We quantify the inequality in the distribution of influencer degree weights—defined as the number of shared survey consumers each influencer has with others within a community—using the Gini Index [57]. It captures the skewness of the degree distribution. A highly skewed distribution may arise due to the presence of “super-influencers”, such as prominent politicians, who share survey consumers with many other influencers. The metric is computed as

(4)

where is the number of nodes in the community , is the weighted internal degree of node u and is the average weighted internal degree of the nodes in community .

Finally, all influencer community-level indices—Identity Diversity, Information Diversity, Structural Integration, and Connectivity Inequality—are aggregated at the user level for the regression analysis by calculating the weighted average across all influencer communities with which the individual is connected.

Specifically, let denote the set of influencer communities followed by survey participant u, and let represent the number of influencers in community followed by user u. For each index , the aggregated individual-level score for user u is computed as a weighted average of the community-level values:

(5)

This step ensures that communities with more connections to a given user contribute proportionally more to their overall exposure, reflecting the strength of affiliation and potential information influence. In contrast, the Community Overlap Index () is inherently computed at the individual level based on the number of influencer communities followed, and thus does not require further aggregation.

Variable dimension reduction and regression analysis

Having measured the patterns of consumers’ selective exposure, we continue to examine the association of ten groups of individual attributes obtained from the survey: Demographics, News Consumption, Political Communication, Political Identification, Political Engagement, Perceptions of Incivility, Perceptions of Disinformation, Authority Trust, Populism, and Attitudes toward Democracy, with the five selective exposure indices (see section 7 in S1 Text for detailed description of the ten groups of individual attributes).

We perform a regression analysis to explore factors related to the five indices of selective exposure patterns. The independent variables are consumers’ individual attributes. To examine a wide range of consumers’ attributes, we construct 10 groups of 189 variables obtained from the National Survey in Brazil, including Demographics, News Consumption, Political Communication, Political Identification, Political Engagement, Perceptions of Incivility, Perceptions of Disinformation, Authority Trust, Populism, and Attitudes to Democracy. These groups encompass individuals’ social, political, and media attributes. We then reduce the dimensions of variables in each group by applying PCA, retaining only one variable for each significant dimension, resulting in 25 variables in total. (See section 7 in S1 Text).

The dependent variables are measurements of five indices of selective exposure patterns. As mentioned before, except for Community Overlap, the other four indices—Identity Diversity, Information Diversity, Structural Integration, and Connectivity Inequality—are initially measured based on influencer communities. We transform these four indices onto individual levels by calculating the average values of influencer communities each consumer interacts with, weighted by the number of links between consumers and influencers in each community.

Finally, we conduct a regression analysis between the reduced attribute variables and the transformed measurements of five indices. We employ a forward selection approach with BIC to select the appropriate regression models for all the selective exposure indices; variables are added iteratively until the BIC does not improve anymore. We use a Zero-truncated Negative Binomial regression (implemented with vglm R package, family = posnegbinomial) for Community Overlap, since it accepts values in positive integers, and Beta regression with a logit link function for the other indices, which all range between zero and one (implemented with betareg R package [58]).

For a more intuitive understanding of our methodological design, please refer to Fig A in S1 Text.

Multidimensional profiles of political influencers and their interconnections

We build a bipartite network between 204 political consumers and 2,307 political influencers by combining datasets of surveys and Twitter/X. Political consumers are sampled from respondents of a national survey conducted during the 2022 Brazilian Presidential Election. Political influencers are identified based on the Twitter/X accounts followed by these survey respondents and can be politicians, media, and individual opinion leaders. Finally, 4,107 following pairs are found in the followship network.

Based on our annotation of political influencers on four dimensions - political ideology, campaign support, social identity, and account type, We find that influencers’ attributes across different dimensions are correlated to some degree (see Fig 1). A Fisher’s exact test confirms that accounts with left-wing ideologies are associated with the support of Lula - the leftist leader and current president of Brazil and advocate for women’s rights (Fisher’s Exact Test p-value <0.01). Conversely, right-wing accounts are more likely to support Bolsonaro - the rightist leader, former president of Brazil, and be associated with religious groups (p < 0.01). Most accounts that express support for electoral candidates are from individual opinion leaders, showing only weak dependency between dimensions of Campaign Support and Account Type (p = 1.00). However, left-wing accounts demonstrate higher proportions of politician accounts compared to others, and religious accounts show higher correlations with individual opinion leaders compared to other identities (p < 0.01).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Contingency tables of multidimensional annotations of political influencer Twitter/X accounts.

Influencer accounts are annotated based on the dimensions of Political Ideology (N = 886), Campaign Support (N = 356), Social Identity (N = 438), and Account Type (N = 2,129). Correlations among categories are examined for each dimension pair (a–f). A multivariate Fisher’s exact test is applied to each contingency table to assess dependencies between dimensions, with significance levels indicated. In Social Identity, Women denotes feminism or support for women’s rights; Religious, Black, and LGBTQ refer to self-identification or expressed support for the rights of these groups.

https://doi.org/10.1371/journal.pone.0332663.g001

These correlation patterns can suggest a dependency relation or potentially hierarchical relation between dimensions. To understand whether there is a more complex structure of individuals’ selective exposure to political influencers, we investigate how the political influencers are organized in the network with a multi-level analysis.

Hierarchical structure of online political selective exposure

We then project the bipartite network onto the influencer side and conduct multi-scale community detection on the influencer-projected network to measure the complexity of individuals’ selective exposure structures (see the schematic diagram in Fig 2(a) and 2(b) and the Materials and methods section for more details).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Schematic diagram of network construction and community detection.

(a) The procedure used to construct the bipartite network between political consumer nodes (orange triangles) and political influencers nodes (blue circles). We project the bipartite network onto political influencers. The directed arrows between consumers and influencers represent the following relationships on Twitter/X. The undirected lines between influencers indicate their shared consumers, weighted by the number of shared consumers. (b) Illustration of multi-scale community detection in the projected influencer network. At the lower level, there are more communities, and the size of each community is smaller (shown in grey). At the higher level, the smaller communities merge, resulting in fewer, larger communities (shown in orange). (c) Histograms of community sizes at the five detected levels.

https://doi.org/10.1371/journal.pone.0332663.g002

To capture the multi-scale and hierarchical organization of real-world social networks [36,37], we conduct multi-scale community detection [35,43] on the influencer network (see details in the Materials and methods section and section 5 in S1 Text). We detect five significant levels (or scales). The number of influencer communities detected at five levels is shown in Fig 2(c), ranging from 46 communities at the lowest level and 2 communities at the highest level. In this setting, a community represents a set of political influencers more likely to be co-followed by the same set of consumers.

Fig 3 shows the results of the multi-scale community detection of the influencer network as an alluvial diagram, revealing its hierarchical structure. As we can see, at the most granular level (Level 1), diverse communities are formed based not only on ideologies but also on political support, social group affiliation, and account types. At the second scale (Level 2), communities with mostly the same categories merge, resulting in more unified left-wing and right-wing communities, as well as distinct Lula and Bolsonaro camps. From level 3 onward, the left-wing and right-wing communities, along with Lula and Bolsonaro supporters, merge together. At the same time, many categories based on social identity and account types become less prominent. By levels 4 and 5, we observe a large mixed community and predominantly right-wing communities and Bolsonaro supporters on the side, who are more isolated and mainly come from religious groups and individual opinion leaders.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Alluvial diagram showing the hierarchical organization of political influencer communities across five levels.

The different panels display the labeling of political influencer communities according to the following dimensions: Political Ideology (a), Campaign Support (b), Social Identity (c), and Account Type (d). The partition of the influencers in communities from a granular level to coarser levels is shown by the different grouping and arrangement of the bundles. We assign a color to a community if more than half of the influencer accounts within the community belong to one category. The transparency of the colors is proportional to the homogeneity of the labels; the higher the proportion of the majority category, the more salient the color. We only color communities when more than 30% of influencer accounts are annotated. In Social Identity, Women denotes feminism or support for women’s rights; Religious, Black, and LGBTQ refer to self-identification or expressed support for the rights of these groups.

https://doi.org/10.1371/journal.pone.0332663.g003

The results show that a single scale in community detection may not always detect the polarized selective exposure pattern between left-wing and right-wing groups. Lower-level scales can uncover more fragmented communities nested within ideological groups, while higher-level scales may capture more integrated ideological communities. Our findings also suggest that the higher-level ideological segmentation in Brazil is asymmetric, with some right-leaning communities being more isolated while others are integrated into larger, more mixed communities. This aligns with findings on the asymmetric political news consumption in the United States [6].

Measurements of online political selective exposure patterns

We also project the consumer-influencer bipartite network onto the consumer side. We group consumers together if they follow political influencers in the same influencer community. The consumers can, therefore, be involved in multiple or just a single influencer community (see Fig 4(a)). We observe that as the level increases, the number of individuals involved in multiple influencer communities decreases. In contrast, the number of individuals who belong to only a single influencer community increases.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Projection of the consumer-influencer bipartite network onto both the consumer and influencer sides.

(a) Influencer communities projected on the consumer network. Each node represents a consumer who can follow single or multiple influencer communities. Consumers following the same communities are encircled. Consumer nodes that follow a single community are annotated in green, while those that follow more than one community are annotated in purple. All five levels are displayed. (b) Projected pie-chart graph of the influencer community network. The nodes are influencer communities, indicated by a pie chart of labels on four dimensions: Political Ideology, Campaign Support, Social Identity, and Account Type. Proportions of categories in each dimension are shown inside the pie chart. Edges between two nodes denote the consumers who follow both influencer communities. In Social Identity, Women denotes feminism or support for women’s rights; Religious, Black, and LGBTQ refer to self-identification or expressed support for the rights of these groups.

https://doi.org/10.1371/journal.pone.0332663.g004

Consumers who are involved in multiple influencer communities create closer connections between these influencer communities. We visualize the influencer community network connected by shared consumers in Fig 4(b), where each pie chart represents an influencer community node and shows the proportions of different identity categories across four dimensions. The edges represent consumers who follow both communities. At Level 1 and 2, we observe that influencer communities closely connected by shared consumers exhibit high proportions of left-wings and media categories. Consumers following these categories of influencer accounts are more versatile and fragmented than others, and this pattern can be mainly observed at the lower level.

We also show the values of the four community-level indices in Fig 5. The granular levels (Level 1 and Level 2) demonstrate diverse values across the four indices. As the levels increase, there is a large merged influencer community where the political ideologies are more diverse and the (weighted) degree distribution is more unequal compared to other communities. Information Diversity and Structural Integration show little distinction among communities at higher levels.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Visualization of measurements of four community-level indices in the influencer community network.

(a) Identity Diversity, (b) Information Diversity, (c) Structural Integration, and (d) Connectivity Inequality. The nodes in each network represent detected influencer communities, and the edges represent the shared consumers between two influencer communities. The area of each node is proportional to the number of political influencers in the community. The color spectrum ranges from black to blue, with a darker color indicating more homogeneous ideological identities and domain information, more integration from other communities, and a more equal distribution of degree weights.

https://doi.org/10.1371/journal.pone.0332663.g005

Individual attributes related to multi-level selective exposure patterns

We continue to investigate how certain consumer attributes are related to selective exposure patterns at various levels with a regression analysis. Individual attributes are conceptualized into 10 groups, encompassing 189 variables, including Demographics, News Consumption, Political Communication, Political Identification, Political Engagement, Perceptions of Incivility, Perceptions of Disinformation, Authority Trust, Populism, and Attitudes toward Democracy. We conduct Principal Component Analysis (PCA) within each variable group to address the issue of multicollinearity. This results in 25 variables after PCA dimension reduction, which are then regressed against the five indices. It is worth noting that we convert the four community-level indices—Identity Diversity, Information Diversity, Structural Integration, and Connectivity Inequality—into individual-level indices by calculating the average values of the influencer communities each consumer interacts with, weighted by the number of influencers each consumer follows (see mathematical definitions in the Materials and methods section).

We identify significant explanatory variables among the 25 initial variables for each index at each level using a forward selection approach based on the Bayesian Information Criterion (BIC). This model selection procedure aims to find the statistical model with the fewest explanatory variables by iteratively adding them until the statistical significance of the model, as measured by the BIC, no longer improves (see section 7 in S1 Text for more details).

We fit the minimal explanatory variables into the regression model for each index and each level. After conducting distribution checks for the values of dependent variables, the zero-truncated Negative Binomial Regression model is chosen as the basic model for the Community Overlap Index. For the indices of Identity Diversity, Information Diversity, Structural Integration, and Connectivity Inequality, the Beta Regression model is employed. For more details, refer to the Materials and methods section and to section 7 in S1 Text. The regression results for the first four levels are shown in Fig 6.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Regression plot of significant individual attributes related to selective exposure patterns.

Regression analyses of consumers’ attributes on Demographics, News Consumption, Political Communication, Political Identification, Political Engagement, Perceptions of Incivility, Perceptions of Disinformation, Authority Trust, Populism, and Attitudes toward Democracy are conducted across five indices of selective exposure patterns: (a) Community Overlap, (b) Identity Diversity, (c) Information Diversity, (d) Structural Integration, and (e) Connectivity Inequality at four network community levels. The variables displayed are significant factors resulting from a forward selection based on BIC. The coefficient values are displayed in blue, and the confidence intervals at the 95% confidence level are shown with error bars. The red lines indicate the 0 value.

https://doi.org/10.1371/journal.pone.0332663.g006

For the Community Overlap Index, we can observe that the ideological position (measured on a scale from 0 to 10, where 0 means “Very left-wing” and 10 means “Very right-wing”) of individuals plays a significant role in explaining their involvement in multiple communities at all levels. However, it is not the only factor; at lower levels of communities, more factors contribute to the versatile engagement of consumers in influencer communities (see Fig 6(a)). The impact of ideological position corresponds to the fragmentation pattern displayed in Fig 4(b), suggesting that left-wing influencer communities are more likely to be shared by consumers with high bridging capacity. Another two significant factors that remain at the first two levels are the frequency of news consumption on Twitter/X and age. At Level 3, the sign of the ideological position coefficient changes, demonstrating a turning point from which right-wing consumers are more inclined to be involved in multiple influencer communities. This is because as the left-wing influencers merge, most of the remaining isolated communities are right-wings. The ideological position plays an increasingly important role at higher levels, especially at Level 4, where only the ideological position remains significant.

The other community-level indices show correlations with different attributes, but similarly to the Community Overlap, they demonstrate various associations across levels. For instance, military trust, with collinearity among variables such as trust in the police and trust in the government, and the frequency of participating political activities primarily explain Identity Diversity at Level 1 and Level 2 (see Fig 6(b)). However, demographic variables play a more important role at Levels 3 and 4, where the religious level is consistent at both levels.

The ideological position of individuals plays a more significant role in Information Diversity compared to other indices. As shown in Fig 6(c), the ideological position is the only significant factor at Levels 1, 3, and 4, and a sub-dimension of populist ideology—people centralism, which is the demand for prioritizing the rights of the majority, is a significant factor at Level 2.

For Structural Integration, there are many explanatory factors at the finer levels, while the degree of integration of communities is mainly related to consumers’ support for democracy at Levels 3 and 4 (see Fig 6(d)). Regarding Connectivity Inequality, similar to Community Overlap, at lower scales, ideological position, as well as a variety of other factors such as age, gender, religious level, and the frequency of news consumption on Twitter/X, have explanatory power. Only the ideological position remains significant at Level 4 (see Fig 6(e)).