Exploring Echo Chambers in Twitter during Two Spanish Regional Elections: An Analysis of Community Interactions

The integration of digital technology in modern society has led to an increased importance of the study and analysis of the digital environment in political campaigns and elections. The use of information operations, such as disinformation, propaganda, astroturfing, and other techniques, can negatively impact public debate, shape public opinion, and polarize societies, which can lead to disastrous consequences. This issue has been extensively studied in the literature, with one notable report being “The Global Disinformation Order” (Bradshaw & Howard, 2019), which provides a global overview of the rise of propaganda on social media worldwide.

Previous research has shown that social network conversations tend to occur within echo chambers, commonly referred to as digital bubbles, where users engage primarily with like-minded others (Terren & Borge, 2021). This is often associated with homophily or confirmation bias, as well as the impact of social media algorithms (Thorson et al., 2019). This phenomenon creates a suboptimal deliberative environment that is characterized by a lack of cross-community conversation and debate, which is crucial for the maintenance of healthy democratic societies (Sunstein, 2017). Several studies have explored the impact of homophily and echo chambers on digital platforms in the United States and other Western democracies. However, there is limited literature available (Aragón et al., 2013; Balcells & Padró-Solanet, 2020; Esteve Del Valle & Borge Bravo, 2018; Guerrero-Solé, 2017) that examines the existence and impact of echo chambers on the political digital debate in the Spanish context.

This study explores Twitter interactions around two regional Spanish elections: the #4M 2021 Madrid and #19J 2022 Andalusia. These elections had a special significance due to the pandemic situation and the highly polarized political climate in Spain. The data extraction process was focused on Twitter, the most prominent social network for political discussions, with approximately 30% of the Spanish population using Twitter and 18% using it for news (Newman et al., 2023). Additionally, Twitter offered numerous tools for analyzing public data, and there is already a significant body of research on the platform.

The concept of echo chambers in social networks continues to be a complex and evolving academic topic of debate. Although there is literature indicating the existence of echo chambers and their capacity to strengthen pre-existing beliefs and polarization, the full scope of their real existence and influence remain topics of continual investigation and discussion. In this study, our goal is to analyze interactions among political communities and the presence of echo chambers in two regional Spanish elections. This research aims to contribute to the global analysis of this topic by providing relevant data and reinforcing aspects studied by previous researchers. The investigation contributes to the confirmation that echo chambers exist within each political community, and these echo chambers become even stronger when communities are grouped by political affiliation. Additionally, the text analysis conducted in this research supports the community-interaction analysis and reveals possible evidence of polarization.

This paper is structured in three main parts. The first part consists in the presentation of the study framework with an overview of the election results (Section “Election Results Overview”), a review of literature and theoretical concepts (Section “Literature Review”), and the research questions that englobe the investigation (Section “Research Questions”). The second part presents the methodology used to collect and analyze the data (Section “Methodology”), as well as the results obtained from the analysis (Section “Results”). Finally, the third part discusses and compares the results to other studies (Section “Discussion”) and presents the main conclusions of the research (Section “Conclusions”).

As mentioned earlier, the two regional elections under study held particular importance because of the pandemic context and the deeply polarized political atmosphere in Spain. Madrid and Andalusia are among the most populated regions in the country, and their results could set the trend for future electoral contests. Both elections were advanced for tactical reasons of the main party in the coalition government (Partido Popular [PP]) at a time where the coalition partner (Ciudadanos) was aiming for a sinking in their national expectations. Another important factor was the emergence of the far-right VOX party, with a clear upward trend in the Spanish scenario. Additionally, both elections saw a divided left electoral space, with two different parties to the left of Partido Socialista Obrero Español (PSOE) (center-left).

Tables 1 and 2 show the results of these two elections, focusing on the parties that obtained parliamentary representation (adding also the Ciudadanos party, which was present in the previous Government coalitions and lost its representation). In both cases, the previous ruling government coalition leader (PP) emerged victorious in the elections, securing a mandate to govern independently in spite of the collapse of their previous coalition partner, Ciudadanos. In Madrid, the results of PP came to a short distance of a majority, meaning they will likely only require occasional support (or mere abstentions) from VOX. Meanwhile, the results of the Andalusia parliament elections have given the PP a comfortable absolute majority.

Lastly, it is crucial to highlight the significance of these examined cases as they can potentially serve as an initial model for future investigations into whether similar scenarios occur in different Spanish regions or in national elections. This could lead to the identification of common patterns in how political communities engage with social media in Spain.

The concept of echo chambers and their influence on social networks has received significant attention in recent years within academic investigations, particularly in the fields of digital communication and social network analysis. In this section, we navigate through the body of literature on echo chambers using a logical progression that moves from the abstract to the specific. We begin by exploring the theoretical concepts and pivotal literature reviews around this topic. Later, prominent studies conducted in the political domain are reviewed. Finally, we explore particular investigations related to electoral campaigns and the relationship between echo chambers and polarization.

Before entering into the specific contributions of individual studies, it is necessary to highlight the importance of two literature reviews that have been used as essential references in this investigation: Terren and Borge (2021) and Ross Arguedas et al. (2022). Terren and Borge (2021) provide an overview of the literature debate on this topic, focusing on the differences obtained between studies depending on the data collected or the platform (or platforms) under study. Meanwhile, Ross Arguedas et al. concentrate on defining the concepts and their relationship with media usage effects.

Echo chambers are often described as the situation in which users mostly communicate with and consume content from like-minded others (Terren & Borge, 2021). The phenomenon of echo chambers is closely tied to several key concepts, especially homophily, the human tendency to interact and associate with similar others (McPherson et al., 2001). This natural inclination toward like-minded connections provides the foundation upon which echo chambers are built, as users tend to engage primarily with those who reinforce their existing views (Barberá et al., 2015; McPherson et al., 2001).

Selective exposure (Garrett, 2009; Stroud, 2010) or confirmation bias (Nickerson, 1998) are other critical elements that are linked to the formation of echo chambers. There are also other similar and related concepts, such as the filter bubble, which place more emphasis on the algorithmic selection of content (Ross Arguedas et al., 2022; Thorson et al., 2019). In order to simplify the concepts discussed in this article, the focus will primarily be on echo chambers and homophily, wherein some authors emphasize on the interactions between users and communities (Barberá et al., 2015).

Research on echo chambers continues to be a topic of rich and ongoing debate, with their impact and even their existence remaining subjects of discussion. Terren and Borge’s (2021) literature review points out the influence of conceptual and methodological choices on research outcomes. The review highlights two primary approaches: one focused on communication and interactions between users, while the other is centered on content exposure. Regarding content exposure, certain studies indicate that algorithmically personalized content, as observed in the works of Bakshy et al. (2015) and Flaxman et al. (2016), tends to reduce exposure to and interaction with diverse perspectives. Additionally, Guess et al. (2019) found that users do encounter diverse viewpoints but may not necessarily engage with them. Concurrently, a methodological division arises based on the type of data employed, distinguishing between digital trace data and self-reported data. Studies reliant on self-reported data generally tend to identify less evidence supporting the existence of echo chambers (Vaccari et al., 2016).

This study is centered on the analysis of users and community interactions using digital trace data, and these are the investigations that merit particular attention in our literature review and discussion. The idea in the studies under this approach is that segmented, like-minded, and homogenous online audiences within social networks are indicative of echo chambers and polarization (Batorski & Grzywinska, 2018).

Barberá et al. (2015), one of the highly cited researchers in this field, reported diverse outcomes concerning echo chambers in social media. They observed that strong echo chambers were evident when discussions revolved around political subjects, with users primarily exchanging content with those who shared similar ideological preferences. However, when the conversation topics were more general, there were instances of cross-ideological interactions, and the echo chamber effect was less pronounced. Additionally, Barberá et al. (2015) examined communication between polarized groups, such as liberals and conservatives, and found that liberals were more inclined than conservatives to engage in interactions across ideological lines. Two separate social network analyses focusing on climate change (Williams et al., 2015) and gun control (Merry, 2015) revealed that a majority of Twitter users mainly interacted with individuals who shared their views, avoiding direct confrontation with their ideological opponents. Furthermore, Tsai et al. (2020) determined that the echo chamber effect was particularly noticeable in the retweet and mention network, while in the reply network, the intergroup communication was more likely to occur.

Additionally, within the Spanish context, several specific studies have investigated into this topic. Guerrero-Solé (2017) and Esteve Del Valle and Borge Bravo (2018) have both highlighted the presence of highly polarized interactions occurring within echo chambers. Esteve Del Valle and Borge Bravo (2018) specifically pointed out that the follower and retweet networks tend to exhibit greater political homophily. Furthermore, Balcells and Padró-Solanet (2020) conducted an interesting qualitative analysis focused on cross-conversations between polarized communities, with a particular emphasis on examining the replies to specific tweets. The results of Balcells and Padró-Solanet (2020) showed that users interact with people holding opposing views on the debate on Catalonia Independence.

Another crucial factor that merits special attention is the specific context of our research, which focuses on an electoral campaign. International studies (Hayat & Samuel-Azran, 2017; Park et al., 2016) found a significant prevalence of retweets and a limited number of replies and mentions. This pattern suggests that, rather than engaging in active debates with one another, users tend to amplify content created by other users, further supporting the notion of Twitter as echo chambers where prevailing opinions are strengthened at the cost of diversity.

Furthermore, Aragón et al. (2013) conducted an examination of Twitter usage by the major Spanish political parties and their respective communities during the national election in 2011. Their findings supported the notion of a “balkanization” of the online political landscape in Spain, which was determined through an analysis of diffusion and conversational patterns, as well as the use of Twitter as a one-way broadcasting medium by political parties and candidates. Similar to other studies, this research observed a limited occurrence of retweets between members of different political communities, but it identified more interactions between communities when analyzing replies. Notably, the increase in intercommunities interactions was most pronounced among parties with similarities in terms of size and/or geographical focus, although the majority of interactions still took place within each respective party community.

Finally, research on polarization presents a multifaceted view and offers a complex picture. While some studies have suggested a potential connection between echo chambers and polarization, this relationship remains limited in scope and is subject to ongoing debate (Quattrociocchi et al., 2016; Törnberg, 2018). Conversely, a recent study by Nyhan et al. (2023) contends that exposure to content from like-minded sources is widespread, but reducing its prevalence does not necessarily lead to a corresponding decrease in polarization. In terms of interactions, Buder et al. (2021) demonstrated that the sentiment of the message can serve as a precursor to attitude polarization. Aragón et al. (2013) have highlighted variations in emotional content depending on political community, noting that the most popular party tends to have a more positive tone in its messages. Additionally, the work of Casañ et al. (2022), which includes community and text analysis of the Catalonia 2021 regional elections, indicates how Twitter can contribute to increased polarization on specific topic debates. As mentioned previously, the discussion surrounding polarization on social media remains highly active, and this research intends to make a contribution to this ongoing debate.

Through this study, our aim is to analyze interactions within political communities and scrutinize the existence of echo chambers during these two regional Spanish elections. We have addressed these objectives through a discussion based on the existing literature, with the following research questions presented in this section.

Firstly, a fundamental point in the literature revolves around the presence of echo chamber effects and their specific attributes. Numerous prior investigations have studied this phenomenon on Twitter, both internationally (Barberá et al., 2015) and in the context of Spain (Balcells & Padró-Solanet, 2020; Esteve Del Valle & Borge Bravo, 2018; Guerrero-Solé, 2017). Some of these studies have even examined echo chambers within the framework of electoral campaigns (Aragón et al., 2013). In our study, we intend to analyze the interactions within and between political communities during our specific case, allowing us to draw comparisons with these previous investigations.

Furthermore, continuing the preceding question, the academic debate on echo chamber effects is closely interconnected with the dynamics of polarized political groups (Barberá et al., 2015; Merry, 2015; Williams et al., 2015). Our intention is to replicate these analyses and identify parallels in our own research. Additionally, in line with Barberá et al.’s findings regarding communication between polarized groups, which suggest that liberals are more likely to engage in intergroup interactions than conservatives, this pattern is likely to be analyzed in our investigation.

Finally, as highlighted in the preceding literature review, the association between echo chambers and polarization remains a subject in its preliminary stages, characterized by ongoing discussions. This study introduces innovative analyses, examining the vocabulary employed within each echo-chambered community (Casañ et al., 2022). Additionally, it initiates an investigation into sentiment analysis of tweets (Aragón et al., 2013; Buder et al., 2021), seeking supplementary insights that may serve as indicators of polarization within the electoral context.

This section outlines the methodology used to analyze the data for the study, with a focus on specific decisions made. Additionally, introductory findings are presented to justify these decisions. A significant number of results were obtained from the investigation, and while the most interesting ones are included in the main text of the paper, some secondary results (which are still considered noteworthy but to a lesser extent) are included in Appendices A and B.

To collect data, we utilized the Twitter Streaming API with the assistance of the “Tweepy” Python library (Roesslein, 2020). The collection period spanned during the campaign period in each election. The tweets collected contained specific hashtags or keywords, and a comprehensive list of these can be found in Appendix C.

The list of target hashtags and keywords used in data collection was manually updated throughout the entire collection period, following the approach used by Abilov et al. (2021). Special attention was given to the hashtags used by the political parties running in the election. These hashtags were selected from a list of parties with the potential to obtain parliamentary representation based on: -

Madrid elections: Last “Barómetro del CIS” before the elections (Centro de Investigaciones Sociológicas, April 2021), which included PP, PSOE, Más Madrid (MM), VOX, Podemos, and Ciudadanos.

Each party had a single hashtag or group of hashtags related to their campaign slogan. In addition to the party-specific hashtags, other hashtags directly related to the election period were also collected, such as #4M, 4M, and #elecciones4M for the Madrid case, and #19J, 19J, and #elecciones19J for the Andalusia case (among others, for the complete list of hashtags and keywords, see Appendix C). Hashtags related to significant events such as TV debates were also included.

In both datasets, the distribution of tweets by type indicates that most of the tweets (around 85%) were retweets, followed by original and quote tweets (around 5%), while replies made up a small percentage of the total tweets collected (concrete information can be consulted in Table A1). The percentages varied slightly across the different datasets collected but were not significant. Additionally, Figures 1 and 2 depict the temporal analysis of the collection period, grouping the number of tweets collected each day. Both graphs show a similar distribution pattern of tweets, with noticeable spikes on significant dates, especially on TV debates. However, there was a data collection issue during the voting day (June 19th) in the Andalusia #19J dataset, which resulted in a lower number of tweets on that day.

Figure 1:

Figure 2:

It is important to consider that social media public data often includes personal information that could be used to identify individuals. Therefore, we followed Twitter’s responsible data and platform-usage guidelines (Twitter Developer Policy and Terms) and received approval from our University Ethics Committee for all data-processing procedures.

To ensure anonymity, we randomized the user ID and tweet ID for each tweet captured using an algorithm that generated a new globally unique identifier (GUID). Additionally, the usernames were anonymized, but only after the community detection analysis (the username of public figures and political parties is used for labeling on the community detection process). Furthermore, whenever possible, we used the data in an aggregated manner.

The code used in this investigation can be accessed through this link: https://github.com/raulbroto/Twitter4M_19JSpanishElections. The Github repository URL will be provided in the camera-ready version of the paper.

The tweet dataset was analyzed using various tools to identify communities, starting with transforming it into a directed graph using the “NetworkX” Python library (Hagberg et al., 2008). The graph included all the relationships between users involved in the dataset, and retweets, quotes, and replies were considered in identifying community affinities. Additionally, as detailed in Sections “Analysis separating tweets by type (RT–Quote–Reply)” and “Analysis separating tweets by type of keyword/hashtag,” supplementary experiments were conducted to assess the significance of hashtag choice and tweet types. The graph was divided and filtered for each specific case under examination in these experiments. Regarding mentions, we did not include them in our analysis as they were not the primary focus of our data collection. Mentions could be a valuable aspect to investigate directly, but we chose not to analyze them because our dataset could be incomplete.

We have defined G = (V, E) as a directed weighted graph, where V is the set of Twitter users and E is the set of directed edges that represent interactions between those users through tweets (retweets, quotes, and replies). The edges in E have a weight that represents the number of times the interaction has occurred. This type of graph is commonly used in the literature (Pastor-Galindo et al., 2020; Stella et al., 2019). Table 5 summarizes the general metrics of this graph, where “average degree” represents the average number of edges that go in/out of graph nodes, considering the weight of edges. It is worth noting that in both cases, the number of nodes is slightly superior to the number of users in Table 3. This is because a tweet captured can be a reply or quote of a tweet not captured (does not have the keywords/hashtags), and thus the user of the tweet not captured appears in the graph as a passive user.

The community detection process was carried out using the Gephi software, where a k-core filter was first applied to remove isolated nodes and make operations easier. The Louvain community partition algorithm was then executed, which splits groups of users with strong connections between them. These partitions are not directly related to a specific political community a priori, so the method applied to rename these communities is a manual inspection of the top in-degree users on each of them, considering that the users with more incoming relations inside their community are the users who flag it. As discussed in the results section, the accounts with higher in-degree typically correspond to the official party account, the party candidate, or other important party figures, simplifying the categorization of each community.

To visualize the graph, the ForceAtlas2 algorithm (Jacomy et al., 2014) was used. This algorithm is designed to create an optimal layout of the graph where the nodes are placed according to their connection strengths, with closer nodes indicating stronger connections. Each community was assigned a specific color that matched the party iconography.

To complement the community analysis, we conducted a text analysis to identify similarities and differences in the words used by each community, providing additional insights into the results obtained in the community interaction study. Furthermore, through the sentiment analysis, we evaluated the words used in each tweet to determine whether they have associated a negative, positive, or neutral emotion. So the text analysis performed can be separated into two main methodological lines: word analysis and sentiment analysis.

To perform the word analysis, two units were considered: user bio (where users write a short description of themselves) and tweet text. Both units had a similar methodology, starting with dividing the text string by each word and filtering out nonalphabetical characters and web links. The stopwords for the English and Spanish languages were filtered using the “NLTK” Python library (Bird et al., 2009). An overview of the most common words in the Madrid and Andalusia datasets can be consulted in Tables A2 and A3, which will be further analyzed by communities in the Results section.

The sentiment analysis was conducted using the Polyglot toolkit (Chen & Skiena, 2014), which is capable of analyzing sentiment in Spanish. To ensure the accuracy of the analysis, the datasets were filtered based on the language detected by Polyglot and the language specified in the tweet metadata (“lang” field). The Polyglot toolkit assigns a value of +1 for positive sentiment words, -1 for negative sentiment words, and 0 for neutral sentiment words. After classifying each word, the average of all words in the tweet was calculated and used to determine the tweet’s sentiment as either NEGATIVE (minor than zero), POSITIVE (greater than zero), or NEUTRAL (equals to zero).

It is essential to recognize the limitations of the methodology used in this study, particularly in terms of the tools and methods utilized. Firstly, the data collected was limited to a single social network and a specific set of hashtags/keywords, which may not reflect the entirety of the interactions and debates on Twitter. Secondly, some methodological decisions made could have an effect on the outcomes obtained, which forces us to carry out complementary experiments to reinforce our results.

Finally, the analysis of text sentiment was constrained by the limited availability of sentiment-analysis tools in Spanish, and there is a possibility of biases or errors within the Polyglot model, which only utilizes a simple model that may not capture the nuances of language.

This section presents the results obtained from the methodology and operations described in the previous section. The results are divided into two parts: community analysis and text analysis. The community analysis focuses on the political communities detected and the interactions within and between them. Finally, the text analysis looks at the words used by each community, as well as the sentiment associated with those words, with the objective of complementing the community interaction study.

Total Network Communities Graph

The graph in Figure 3 effectively illustrates each of the six political party communities analyzed in the #4M Madrid election, allowing for observation of the polarization and proximity of the communities based on their left and right political classifications.

Figure 3:

A similar pattern is shown in Figure 4, with the six political party communities analyzed in the #19J Andalusia election, again with a clear grouping of the communities around the left-right political classification and with a noticeable polarization.

Figure 4:

Table 6 shows the number of users detected in each political community after identifying the top three-most significant users in each community (Table A4), which allowed us to tag them without contradictions, for the Madrid elections. Similarly, Table 7 presents the same for the Andalusia elections (with Table A5 with the top in-degree users). The tables show that each community can be associated with its corresponding main accounts, including official party accounts, candidates, and other notable political figures.

Table 6.

Number of users in each political community (Madrid).

Community	# Users
Podemos	78.464
VOX	71.994
PSOE	12.638
PP	9.818
MM	9.513
Cs	5.247

Cs, Ciudadanos; MM, Más Madrid; PP, Partido Popular; PSOE, Partido Socialista Obrero Español.

Table 7.

Number of users in each political community (Andalusia).

Community	# Users
VOX	37.846
PA	37.378
PSOE	10.184
PP	8.152
AA	6.289
Cs	3.487

AA, Adelante Andalucía; Cs, Ciudadanos; PA, Por Andalucía; PP, Partido Popular; PSOE, Partido Socialista Obrero Español.

However, in Madrid, there is a notable exception with the political figure of Isabel Diaz Ayuso (highlighted in Table A4), the PP candidate and winner of the elections, who is positioned within the VOX community instead of the PP community. One possible explanation for this is her high popularity among the VOX community, combined with the considerable difference in size between the VOX and PP communities, which allowed her to have stronger connections with the VOX community than her own party.

Community Interactions Analysis

The community interactions are divided into two categories: active interactions, where a user retweets, replies, or quotes another user, and passive interactions, where a user is retweeted, replied to, or quoted by another user. A general overview of the interactions of users in each political community detected for the Madrid and Andalusia elections can be found in Tables A6 and A7.

Moreover, the active and passive interactions can also be grouped by communities. The active and passive interactions between all the communities are illustrated in Figure 5 (Madrid) and Figure 6 (Andalusia). The detailed 15 most frequent relations between an active user community and their respective passive user community, including the count and percentage of these interactions in the active interactions of the community, can be examined in Table A8 (Madrid) and Table A9 (Andalusia). As expected, the majority of these interactions occur inside each community.

Figure 5:

Figure 6:

Analysis Separating Tweets by Type (RT–Quote–Reply)

The high frequency of retweets (approximately 85% of all tweets in both datasets) might influence the analysis and outcomes in favor of a particular perspective, potentially obscuring other forms of interactions that occurred on Twitter among the communities. To address this issue, we have conducted a segmented analysis in this section, distinguishing between different types of tweets. This approach aims to uncover potential differences that may arise when interactions depend on tweet types.

We start by examining the retweet network, a type of network that has traditionally been viewed in prior research (Aragón et al., 2013; Esteve Del Valle & Borge Bravo, 2018) as diffusion networks commonly employed for disseminating messages during election campaigns. As depicted in Figure B1, the network graph that exclusively includes RT-type tweets closely resembles the previous Figures 3 and 4. In both networks, we are able to discern each of the six communities and the division between the left and right supra-communities.

When examining the interactions among these communities (Figure B2), the outcomes once more reveal a highly comparable pattern to what was previously illustrated in Figures 5 and 6. There is a slight increase in all intracommunity interactions (the primary diagonal) and a decrease in intercommunity interactions. A parallel scenario occurs when communities are grouped into supra-communities (Table A11), mirroring the findings presented in Table 8. In this context, supra-communities exhibit even greater robustness, while intercommunity interactions remain almost negligible.

Next, we examine the network graph and interactions associated with quote-type tweets, constituting approximately 5% of the total tweets in both datasets. Unlike retweets, a quoted tweet can be employed for both positive and negative commentary on the quoted content. Consequently, it does not inherently function as a diffusion network in the same manner as retweets.

Figure 7 displays the resulting graph in the case of quote-type tweets, revealing a higher level of intercommunity interactions compared to the retweet graph. Nonetheless, it remains possible to distinguish each community and the division between the left and right groups. The heatmaps in Figure 8 confirm this trend by focusing on the decrease of the values on the main diagonal. The findings indicate a decrease in intracommunity interactions, with the specific decrease varying within each community but averaging around 10%–20% in each case. When replicating the analysis of supra-communities, as shown in Table 11, we observe that while the supra-communities remain relatively robust (with over 80% of intercommunity interactions in each case), there is a significant uptick in interactions between the left and right communities. This could be attributed to the use of quoting as a means to criticize and attack political opponents. Particularly noteworthy is the greater volume of interactions from the right to the left block compared to the reverse direction, which was observed the opposite in previous studies (Barberá et al., 2015).

Figure 7:

Figure 8:

Table 11.

Number of interactions between community blocks by quote-type tweet.

Active user community	Passive user community	Madrid Count	Percentage in active interactions (%)	Andalusia Count	Percentage in active interactions (%)
Left	Left	53.757	86.08	34.289	89.27
Right	Right	43.914	80.01	30.180	84.82
Right	Left	7.711	14.05	4.570	12.84
Left	Right	3.865	6.19	3.019	7.86

Finally, we analyze the reply network graph, comprising approximately 2% of tweets in each dataset. When a user replies to a tweet, various motivations may underlie this action. It could be an expression of agreement or support for the original tweet, a critique of the statement presented, or a genuine engagement in a debate with other users. This study does not enter into these specific motivations, instead focuses on how reply interactions occur between users based on their respective communities. It is reasonable to expect that replies would show a higher degree of cross-community interaction, but we now quantify how this happened in our captured datasets.

Figure 9 displays the network graph for reply tweets, which presents distinct results compared to the total or retweet network graphs but shares similarities with the quote graph. While we can identify the larger communities (in terms of number of users), the smaller ones do not have a significant presence because their users are integrated within the larger ones, primarily Podemos-PSOE and VOX. Additionally, there is a higher level of connection between communities, similar to the quote scenario. An interesting observation from these graphs is the centrality of the winning candidates in the election (Isabel Diaz Ayuso and Juanma Moreno), indicating that they received replies from users across various communities.

Figure 9:

Despite these graph differences, an examination of Figure 10 and Table 12 reveals results similar to those observed in the quoted tweets study. Intracommunity interactions are lower than in the retweet scenario, ranging from approximately 73% to 90%, although interestingly slightly higher than those observed in the quote heatmaps. A similar trend is evident in Table 12, where an analysis of supra-communities demonstrates significant robustness, with cross-community values generally higher from the right community to the left than vice versa.

Figure 10:

Table 12.

Number of interactions between community blocks by reply-type tweet.

Active user community	Passive user community	Madrid Count	Percentage in active interactions (%)	Andalusia Count	Percentage in active interactions (%)
Left	Left	18.406	85.79	10.101	89.88
Right	Right	15.422	77.87	15.103	89.67
Right	Left	3.049	15.39	1.216	7.22
Left	Right	1.627	7.58	704	6.26

Analysis Separating Tweets by Type of Keyword/Hashtag

In this section, we investigate the extent to which the earlier findings are influenced by the specific choice of keywords and hashtags used in data collection. To address this, we conduct an analysis that categorizes tweets based on the type of keyword or hashtag that triggered their collection. We classify tweets into three types: partisan (when the hashtag aligns with the ones proposed by the political parties), general (hashtags or keywords related to the elections, such as #4M-#19J, #Elecciones4M-#Elecciones19J, TV debates, etc.), or a combination of both.

The results of this categorization can be found in Table 13, while Tables A12 and A13 present the results broken down by community.

Table 13.

Number of tweets separating by type of keyword/hashtag.

Type of keyword/hashtag	Madrid	Andalusia
Partisan	664.205	827.771
General	1,979.082	393.090
Both	206.421	303.402

The results indicate notable differences between the two election cases, which can be attributed to the Madrid election having a more pronounced national and general focus, while the Andalusia case has a more regional and partisan focus. Additionally, as we will further analyze, these differences align with the respective party strategies employed in each case.

Next, we investigate the results of the community analysis within this context. For simplicity, we present the “partisan” and “both” types together, as the presence of a partisan hashtag has a more significant influence on the tweets, and the results are almost identical when presented together or separately.

Beginning with the analysis of tweets categorized as partisan–both types, the results align with the expectations, revealing highly clustered communities shown in Figures B3 and B4. These clusters are similar to those observed in the retweet network, which primarily serves as a diffusion network.

On the other hand, when examining general-type tweets, some noteworthy insights emerge. Firstly, when observing the network graph in Figure 11, the Andalusia case resembles most of the previous cases. However, in the Madrid graph, a significant difference becomes apparent: there is no PP community by itself.

Figure 11:

Further analyzing community interactions, it becomes evident from the heatmaps in Figure 12 that the percentages of intercommunity interactions (main diagonal) are significantly lower in certain cases, which matches with what was expected. But again, in the Madrid case, the PP community is an outlier, with over 58% of interactions directed toward the VOX community. As mentioned in Section “Total network communities graph,” the placement of the PP candidate, Isabel Diaz Ayuso, within the VOX community could be an important indicator of this result and will be explored at the end of this section, along with additional information that connects this fact to the candidate’s strategy.

Figure 12:

Finally, despite the presence of intercommunity interactions as observed in Figure 12, when grouping into left-right supra-communities, the results presented in Table 14 consistently demonstrate again the presence of robust echo chambers in each case.

Table 14.

Number of interactions between community blocks by general-type tweet.

Active user community	Passive user community	Madrid Count	Percentage of active interactions (%)	Andalusia Count	Percentage of active interactions (%)
Left	Left	694.935	94.94	166.326	94.37
Right	Right	483.508	95.31	133.027	93.88
Right	Left	10.620	2.09	3.638	2.57
Left	Right	5.904	0.80	2.938	1.66

Before concluding this section, it is important to highlight a noteworthy observation regarding the types of hashtags used by Isabel Diaz Ayuso, the PP candidate in the Madrid elections. Her official account posted 27 original tweets during the electoral period, and it is remarkable that she only used the partisan slogan hashtag (#VotaLibertad) on two occasions. The rest of her tweets included only general hashtags (#4M, #DebateTelemadrid, and #Elecciones4M). This behavior could be related to her strategy of reaching and engaging with voters from other parties, particularly on the right wing, an aspect this study demonstrates that she effectively accomplished.

In this section, we present and discuss the results obtained in this study, which aim to address the research questions and objectives outlined in Section “Research Questions”.

Based on the analysis conducted in this study, it can be expressed that homophily does exist in the interactions around the Madrid and Andalusia elections on Twitter. Each political party community tends to interact primarily with others who share similar political views, resulting in the creation of clear echo chambers. This situation was previously noticed and studied by Barberá et al. (2015), Aragón et al. (2013), Esteve Del Valle and Borge Bravo (2018), among others.

The analysis of the community interactions, revealed in Figures 5 and 6, showed that the communities were highly clustered, with strong ties within the same political party community and weaker ties between different communities. All communities detected have interactions inside their own community over 75% in the Madrid study case and over 87% in Andalusia. When dividing the dataset by the type of tweet, the retweet network exhibits an even more pronounced isolation, reaffirming its role as a primary diffusion network, consistent with prior research (Esteve Del Valle & Borge Bravo, 2018; Guerrero-Solé, 2017). Conversely, when examining the network of replies and quotes, the outcomes generally indicate a greater degree of cross-interaction. However, it is noteworthy that intracommunity interactions continue to dominate, consistent with the findings of Aragón et al. (2013).

The only exception to the echo-chambered detected situation is the previously mentioned categorization of Isabel Diaz Ayuso (the PP candidate in Madrid) in the VOX community, which can be attributed to her popularity and the difference in the size of the political communities. When analyzing separately the networks depending on the hashtag/keyword collected, we have noticed that Isabel Diaz Ayuso barely used the partisan hashtag and instead used the general hashtags to reach out to the public in other communities on the right. This strategy allowed her to overcome its party echo chamber and interact with other parties on the right spectrum.

Furthermore, the analysis of the most frequently used words in tweets by each community ( Tables 17 and 18) showed that the words used by different political party communities were distinct, with clear references to the party name, candidate names, and other ideological concepts. Even though we have detected some words that possibly show intercommunity attacks, these are used as an internal product to attack the rivals, and not to engage in a real conversation with the opposite communities. This represents a novel and innovative analysis that, as far as we are aware, has not been previously presented in the literature.

Also, Tables 9 and 10 indicate that the PP, the winner of both elections, did not have the most interactions or the largest community. Moreover, the largest communities, which also had the most interactions, were not the ones that received the most votes by far. One possible explanation could be the homophily and echo chambers that exist in social networks, particularly Twitter, which lead to interactions primarily among individuals with more extreme political positions. However, previous research has suggested an alternative explanation, indicating that these new parties are stronger and well adapted to social networks due to their limited access to traditional mass media channels, unlike the traditional and larger political parties (Sampedro, 2021).

In conclusion, there is certain evidence of homophily in the interactions surrounding the Madrid and Andalusia elections on Twitter. These results contribute to the topic knowledge, confirming what was found in Barberá et al. (2015), Aragón et al. (2013), Guerrero-Solé (2017), and Esteve Del Valle and Borge Bravo (2018), contributing additional insights into the existence of echo chamber effects. Even in scenarios where cross-community interactions were expected to be more prevalent (such as in quote-reply interactions or when using general hashtags exclusively), the quantity and strength of intracommunity interactions remained significantly higher than intercommunity interactions, consistent with the results reported in the study by Aragón et al. (2013), which also is framed in an electoral campaign on the Spanish scenario. Thus, it can be concluded that the majority of user interactions occur with individuals who share similar ideological views.

Certainly, the results from Section “Supra-Community analysis” suggest that when we group the political communities into left-right polarized blocks, the echo chambers become even more isolated. The data reveals an astonishing result of more than 97% of interactions occurring within each left-right block, which essentially means that there is practically no interactions outside of these bubbles. The interblock community interactions are only about 1% in all cases.

These findings align with studies on echo chambers involving interactions between polarized political groups in social media, as demonstrated by Barberá et al. (2015), Williams et al. (2015), and Merry (2015). They also corroborate the existence of echo chambers in the Spanish context, as evidenced by Guerrero-Solé (2017) and Esteve Del Valle and Borge Bravo (2018). However, these findings are in contrast to the research conducted by Balcells and Padró-Solanet (2020), in which pro and anti-independence blocks interacted with each other. We attribute the variation in results between our study and that of Balcells and Padró-Solanet (2020) to differences in methodology. Specifically, their study focused on a qualitative analysis of the replies network, whereas our approach involved capturing replies associated with specific hashtags or keywords, which may not encompass the entire spectrum of replies involved.

Furthermore, these results gain even greater significance when considering the construction of supra-communities based on the reply/quote network or the use of general hashtags/keywords. In these cases, a high degree of clustering is observed in tools designed to encourage cross-community interactions between users. Additionally, the higher interaction observed from the right block to the left block, as compared to the reverse, is a noteworthy finding and stands in contrast to some previous studies conducted in other countries, such as Barberá et al. (2015), where liberals were more inclined than conservatives to engage in interactions across ideological lines.

This analysis is reinforced by the tweet text analysis in Section “Community interactions analysis” ( Tables 17 and 18), which highlights the previously mentioned opposition attack strategy, where community criticism has been detected toward a party (or candidate) from the opposite left-right block.

The text analysis conducted in this study supports the community-interaction analysis by revealing the text patterns and ideological concepts used by each political party community. As shown in Tables 17 and 18, the most frequently used words in tweets by each community reflected clear references to the party name, candidate names, and other ideological concepts. This indicates that users in each community tend to interact primarily with others who share similar political views, resulting in signs of echo chambers effects (Aragón et al., 2013; Casañ et al., 2022).

This finding is consistent with the homophily patterns observed in Q1 and Q2, where each political party community tends to interact primarily with others who share similar political views. The highly clustered community interactions revealed in Figures 5 and 6 indicate that there is limited interaction between communities, a fact that is reinforced by the text analysis performed.

The sentiment analysis conducted in this study could evidence a possible polarization in the tweets analyzed, as evidenced by the predominance of negative sentiment (Buder et al., 2021). Specifically, negative sentiment tweets were almost three times higher than positive sentiment tweets. While most interactions occur within communities, this negative sentiment is likely related to the prevalence of cross-community attacks detected in the tweet text analysis presented in Tables 17 and 18.

Overall, the text analysis provides additional support for the homophily patterns observed in the community-interaction analysis and reveals possible evidence of polarization among political party communities on Twitter.

The research results indicate that within each political party community, interactions predominantly occur among individuals who hold similar political views, leading to the creation of echo chambers. These echo chambers exhibit significant robustness, with an even higher degree of isolation (over 97% of interactions) when political party communities are combined into left-right groups. Despite taking into account the methodological limitations mentioned earlier, this study conducted additional experiments to validate the presented results. It is worth highlighting that intracommunity interactions remain dominant, even in scenarios such as the reply and quote network or when exclusively examining the general hashtags network. This outcome is relevant, considering that these elements are designed to promote cross-community interactions between users.

Furthermore, the higher interaction observed from the right block to the left block, in comparison to the opposite direction, is a significant finding. This result contradicts prior studies conducted in other nations, like Barberá et al. (2015), which found that liberals were more inclined than conservatives to engage in interactions across ideological lines. Additionally, the text analysis performed in this study provides additional support for the homophily patterns identified in the community-interaction analysis and could indicate possible presence of polarization among political party communities on Twitter.

It is also worth mentioning that the findings and deductions of this study are remarkably similar in both election campaigns, even though they took place more than a year apart, and despite the differences in the participating parties and the concrete characteristics of each region. This could suggest that political parties may follow a certain pattern in their digital strategies, particularly on Twitter. The most important differences include the dominance of general hashtags in the Madrid election (in contrast to the prevalence of partisan hashtags in the Andalusia elections), and the specific strategy employed by Isabel Diaz Ayuso (PP Madrid candidate) to surpass its political echo chamber in her objective for supremacy within the right-wing election block.

We are experiencing a rise in polarization and propaganda around the world, and the omnipresence of social media could be behind these trends. The existence of echo chambers leads to poor cross-community conversation and debate, which would be the desirable environment in a healthy democracy, in parallel with the debate in the real world. The concept of echo chambers and their influence on social networks has received significant attention in recent years within academic investigations, particularly in the fields of digital communication and social network analysis, and the debate about their existence and effects is still ongoing.

In conclusion, this study added results and analysis to the ongoing academic debate regarding the digital political ecosystem during political campaigns, specifically in two very specific cases during the Madrid and Andalusia regional elections.