Echo chamber amplification and disagreement effects in the political activity of Twitter users

Echo chamber amplification and disagreement effects in the political activity of Twitter users

Abstract

Online social networks have emerged as a significant platform for political discourse. In this paper we investigate what affects the level of participation of users in the political discussion. Specifically, are users more likely to be active when they are surrounded by like-minded individuals, or, alternatively, when their environment is heterogeneous, and so their messages might be carried to people with differing views.

To answer this question, we analyzed the activity of approximately twitter users who expressed explicit support for one of the presidential candidates of the US presidential election. We quantified the level of political activity (PA) of users by the fraction of political tweets in their posts, and analyzed the relationship between PA and measures of the users’ political environment. These measures were designed to assess the likemindedness, e.g., the fraction of users with similar political views, of their virtual environment and, for a subset of approximately twitter users, of their geographic environment.

Our results showed that the highest levels of PA are usually obtained by users in politically balanced virtual environment. This is in line with the disagreement theory of political science that states that a user’s political activity is invigorated by the disagreement with their peers. Our results also show that users surrounded by politically like-minded virtual peers tend to have a low level of PA. This observation contradicts the echo chamber amplification theory that states that a person tends to be more politically active when surrounded by like-minded people. Finally, we observe that the likemindedness of the geographical environment does not have a significant effect on the level of PA of users.

We thus conclude that the level of political activity of the Twitter users is independent of the likemindedness of their geographical environment and is correlated with likemindedness of their virtual environment. The exact form of correlation manifests the phenomenon of disagreement and, in a majority of settings, contradicts the echo chamber amplification theory.

Twitter, echo chamber, disagreement effect, politics
\numberofauthors

2

\category

H.2.8Data MiningInformation Retrieval

\terms

Twitter, politics

Introduction

In the recent years, online social networks have emerged as a significant platform for discussion and dissemination of political information. For example, Pew surveys [15] found that 22% of adult Internet users participated in political campaigns through at least one of the major social media platforms (Twitter, Facebook, Myspace) during the 2010 US elections. Similarly, it was found in [13] that, in year 2012, 34% of social network users posted their own thoughts on political or social issues, and 38% of users “liked” and reposted political posts of others.

This increasing importance of social media and the relative convenience of its analysis attracted attention from academic researchers. Among the questions that have been investigated are: can future election results be predicted (e.g., [5]), is political information on Twitter credible [2, 6, 14], who are users whose opinion on a certain subject is influential [1, 18], and how to leverage anonymized web search queries to analyze and visualize political issues [17].

In this paper we analyze social factors associated with the level of participation of users in the political discussion. Two complementing theories were suggested in the scientific literature to explain the interaction between likemindedness of ones’ social environment and of the level of political activity:

  • Echo chamber and echo chamber amplification. In [16], the author shows that people tend to look for cognitive comfort by discussing their opinions with like-minded people. Their opinions are thus echoed and reinforced by their social peers; creating an echo chamber effect. In the context of the web, the echo chamber effect is achieved when people follow blogs and news sources that do not challenge their political opinions. This theory predicts that people in comforting environments such as echo chambers will exhibit an amplified level of political activity.

  • Disagreement effect. A large body of political-science literature [7, 8, 9, 11] explores the effect that disagreement, i.e., having political opinions different from your (non-virtual) social peers, has on your political activity. Nir [9] shows that disagreement has a dual effect. A politically isolated person, in the sense that all of their peers disagree with their opinion, tends to exhibit a lower than average level of political activity. However, a person with politically heterogeneous social peers tends to exhibit a higher level of political activity (see also [11]), even when compared to people completely surrounded with like-minded peers (echo chamber). This theory predicts that the level of political activities should be attenuated for isolated users and have a dominant peak for heterogeneous social environments.

In this paper we test the presence and the relative importance of these two phenomena in political discussions in Twitter. We also test a conjecture that likemindedness of both the virtual (web) environment and the physical (geographical) environment have effect on a user’s level of political activity.

Methods

In this paper we analyze data from Twitter - a micro-blogging service that allows users to post short “tweets” and to receive tweets made by other users by “following” their Twitter feeds, thus creating a social network of Twitter accounts. Our data extraction technique largely follows methods from our previous work in [4].

In our analysis we extract and analyze four overlapping sets of Twitter users. We begin by extracting Raw-DS, which is a large set of users with known political affiliations and known levels of political publishing activity before 2012 US presidential election. We then extract a Retweet-network graph that is used as an approximation of a social network connecting Twitters users in Raw-DS. Using connections defined by Retweet-network, we extract two subsets of Raw-DS: Follower-DS and Followed-DS. These sets of users are used to analyze the relation between users’ level of political activity and the likemindedness of their neighbors in the Retweet-network. Finally, we extract a small subset Geographical-DS of users of Raw-DS that have enough available geo-spatial information to identify counties they reside in. The Geographical-DS data set is then used to analyze the relation between users’ level of political activity and the likemindedness of their physical environment. We proceed formally define each one of these sets.

Raw-DS. We begin by extracting a large set of users with known political affiliations and known levels of political publishing activity before 2012 US presidential election. We denote this set of users by Raw-DS. In what follows, we alternatively refer to these users as pro-Obama and pro-Romney, respectively.

To this end, we employed the method described in [4]. Namely, we looked for specific highly-partisan hashtags (a single word preceded by “#” sign, listed in Table 1) among the tweets made in the 10 days following Election Day. We picked this method over other existing methods (e.g., [3, 10, 12]) because of ease of its implementation and accuracy (above ) that is higher than in other solutions. The simplicity and the higher accuracy of our method come at the expense of a smaller recall than that of other existing methods. However, the obtained user population was large enough for meaningful analysis.

As in [4], we found a total of pro-Obama users and pro-Romney users.

Affiliation Used hashtags
Pro-Obama #voteobama, #obama2012, #goobama,
#obamabiden, #guardthechange,
#4moreyears,
#forward, #forwardwithobama,
#obamaforpresident,#igoobama
Pro-Romney #romenyryan2012, #voteromneyryan,
#voteromney, #benghazi, #nobama,
#imwithmitt, #americascomebackteam,
#fireobama, #teamprolife, #gogop
Table 1: List of hashtags used for identification of the political affiliation of users.

We then extracted all tweets published by users in Raw-DS during the three and a half months period between Aug. 1st and Nov. 15th . This interval includes tweets published, roughly, three months before the Election Day (Nov. 6th) and ten days after it.

Overall, there were tweets. As in [4], we used hashtags such as ”#election2012” (see Table 2 for the complete list) to extract a subset of tweets on political issues.

List of hashtags:
All hashtags from Table 1, #tcot , #election2012, #gop,
#romney,#obama, #elections, #president
Table 2: List of hashtags used to identify political tweets.

Given the number of political and non-political tweets made by each user in Raw-DS, we were able to calculate their political activity (PA), which we quantified as the fraction of political tweets among their posts.

Retweet-network. We next inferred the edges of the social graph that connects users in Raw-DS. There are several commonly used proxies for the social connections between Twitter users. For instance, one approach is to assume that there exits a directed edge from user A to user B if user A follows user B’s Twitter feed. Another approach is to define an edge from user A to user B if user A “retweeted” one of user B’s posts. We choose the latter approach as it indicates a stronger connection between users. Namely, user A is more than just skimming through user B’s Twitter feed, user A also actively engages with its content. We refer to corresponding social network over users in Raw-DS as the Retweet-network. We keep the terms “follower” and “followed” to describe the relationship between users in the Retweet-network.

The Retweet-network contains edges for nodes, which implies that the network is highly sparse. Table 3 shows the number of edges between users for the four possible pairs of political affiliations.

User pair Number of edges
(in thousands)
(pO,pO)
(pO,pR)
(pR,pO)
(pR,pR)
Table 3: Connectivity statistics in Retweet-network. Acronym pO stands for pro-Obama user and acronym pR stands for pro-Romney user. The first user in the pair follows the second user in the pair, e.g., in a pair (pO,pR) a pro-Obama user follows a pro-Romney user.

We note a large number of retweets that cross party lines, i.e., a tweet made by a pro-Obama user is retweeted by a pro-Romney user or vice versa. In [4], we show that some of these retweets are part of a political debate. In particular, when a link to a political article published by a pro-Obama user is retweeted by a pro-Romney user, the text accompanying the retweeted link is likely to be modified, usually to interpret the link according to the users’ own point of view [4].

Follower-DS. Using Retweet-network, we extracted a subset
Follower-DS of users in Raw-DS that have at least one follower. This subset contains pro-Romney users and pro-Obama users. For users in Follower-DS, we refer to the likemindedness of their followers as follower-LM and quantify it as the fraction of users’ followers that share their choice of candidate.

Followed-DS. Similarly to Follower-DS, we extracted a subset
Followed-DS of users in Raw-DS that follow at least one user. This subset contains pro-Romney users and pro-Obama users. For users in Followed-DS, we refer to the likemindedness of Twitter feeds they follow as followed-LM and quantify it as the fraction of Twitter feeds followed by these users that share their choice of candidate. We note that this separation between Follower-DS and Followed-DS is meaningful since only % of edges in Retweet-network are reciprocated.

Geographical-DS. Finally, we identified a subset Geographical-DS of users in Raw-DS with enough geo-spatial information to identify counties they reside in. To this end, we began by extracting a larger subset of users that provided their geographical location (in terms of GPS coordinates) in at least two of their tweets. For each such user, we calculated their average location by taking the mean value of GPS coordinates. In order for this average location to be representative, we discarded all users with the maximal distance between user’s locations greater than kilometers. We further discarded all users with the average location outside of the United States. The remaining subset Geographical-DS contains a total of pro-Romney users and pro-Obama users. For each user in Geographical-DS we use their average location to identify the county this user resides in and obtain the official voting record for this county. Given this information we are able to calculate the likemindedness of user’s geographical environment (geographical-LM), which is quantified as voting share of user’s candidate in their county.

Results

We begin by using data from Follower-DS to analyze the dependence of the median level of political activity on the likemindedness of users that read posts of the considered user. I.e., the dependence of median PA on follower-LM. To this end, we divide the range of possible values of follower-LM to equally-sized bins. For each bin, we calculate the median PA of all users with the value of follower-LM that falls in this bin. Figure 6 and Figure 6 depict these dependencies for pro-Obama and pro-Romney users, respectively.

Both for pro-Obama and pro-Romney users we observe a unimodal dependency of median PA on follower-LM. The single peak corresponds to the disagreement effect and is obtained for medium values of follower-LM (between and for pro-Obama users and between and for pro-Romney users). The level of political activity of politically isolated users (follower-LM smaller than for pro-Obama users and smaller than for pro-Romney users) is negligible. This observation is also in line with the disagreement effect. The level of political activity of users in the echo chamber environment (large values of follower-LM) is also very low, which is in contradiction to the echo chamber amplification theory.

We now use data from Followed-DS to analyze the dependence of the median level of political activity on the likemindedness of the Twitter feeds read by the considered user, i.e. the dependence of median PA on followed-LM. Again, we divide the range of possible values of followed-LM to equally-sized bins. For each bin, we calculate the median PA of all users with the value of followed-LM that falls in this bin. Figure 6 and Figure 6 depict these dependencies for pro-Obama and pro-Romney users, respectively. For both pro-Obama and pro-Romney users, small values of followed-LM (below ) and high values of followed-LM (above ) correspond to very low levels of political activity. The first observation is in line with the disagreement effect and the second observation contradicts the echo chamber amplification theory. For both pro-Obama and pro-Romney users, the dominant peak corresponds to the disagreement effect and is obtained for values of followed-LM close to . However, in contrast to the dependence of PA on follower-LM, there are also secondary peaks. For Obama supporters, there is a secondary peak for the values of followed-LM between and . We hypothesize that this secondary peak is also a manifestation of the disagreement effect. For Romney supporters, there is a secondary peak for the values of followed-LM between and . This secondary peak may be a manifestation of both the disagreement effect and the echo chamber amplification.

Finally, we analyzed the relationship between the median PA and geographical-LM. Similarly to figures above, we divide the range of possible values of geographical-LM to equally-sized bins. For each bin, we calculate the median PA of all users with the value of geographical-LM that falls in this bin. The results are depicted on Figure 6 for pro-Obama users and on Figure 6 for pro-Romney users. In contrast to the dependence of median PA on likemindedness of user’s virtual environment (follower-LM and followed-LM), the dependence of PA on geographical-LM does not differ for pro-Obama and pro-Romney users. In fact, it seems that the level of political activity of users is independent of geographical-LM, hence does not exhibit neither the echo chamber amplification nor the disagreement effect.

Figure 1: Median PA (fraction of political tweets) vs. follower-LM (fraction of users’ followers that support their candidate) for pro-Obama users. Median PA is depicted as a dashed line. Lower confidence interval is given by the 20th percentile and the upper confidence interval is given by the 80th percentile. Each bin contains at least points. One bin with less than points was omitted.
Figure 2: Median PA (fraction of political tweets) vs. follower-LM (fraction of users’ followers that support their candidate) for pro-Romney users. Median PA is depicted as a dashed line. Lower confidence interval is given by the 20th percentile and the upper confidence interval is given by the 80th percentile. Each bin contains at least points.
Figure 3: Median PA (fraction of political tweets) vs. followed-LM (fraction of people followed by the user that support user’s candidate) for pro-Obama users. Median PA is depicted as a dashed line. Lower confidence interval is given by the 20th percentile and the upper confidence interval is given by the 80th percentile. Each bin contains at least points.
Figure 4: Median PA (fraction of political tweets) vs. followed-LM (fraction of people followed by the user that support user’s candidate) for pro-Romney users. Median PA is depicted as a dashed line. Lower confidence interval is given by the 20th percentile and the upper confidence interval is given by the 80th percentile. Each bin contains at least points.
Figure 5: Median PA (fraction of political tweets) vs. geographical-LM (voting share users’ candidate in their county). Median PA is depicted as a dashed line. Lower confidence interval is given by the 20th percentile and the upper confidence interval is given by the 80th percentile. Each bin has at least 10 points. Three bins with less than points were omitted.
Figure 6: Median PA (fraction of political tweets) vs. geographical-LM (voting share users’ candidate in their county). Median PA is depicted as a dashed line. Lower confidence interval is given by the 20th percentile and the upper confidence interval is given by the 80th percentile. Each bin has at least 10 points. Three bins with less than points were omitted.

Discussion

In this paper we analyzed the connection between user’s level of political activity on Twitter (PA) and the likemindedness of its virtual (follower-LM, followed-LM) and geographical environments (geographical-LM). Specifically, we focused on the presence of the echo chamber amplification and the disagreement effect.

We showed that user’s PA as a function of follower-LM has a similar form for both pro-Obama and pro-Romney users. In both cases, high level of political activity of a user correlates with a politically diverse set of followers, i.e., readers of user’s Twitter feed. The dependence of PA on follower-LM exhibits strong disagreement effect, but does not manifest the echo chamber amplification.

The dependence of user’s PA on followed-LM is different for users supporting different candidates. Pro-Romney users exhibit high level of political activity when Twitter feeds they follow are predominantly, but not purely, pro-Romney. Namely, pro-Romney users are most active in the mostly likeminded environments. This behavior is a manifestation of the combination of the disagreement effect and the echo chamber amplification. In contrast, pro-Obama users tend to have a high level of political activity in the politically adverse environments. Namely, when Twitter feeds they follow are predominantly pro-Romney but not exclusively pro-Romney. This behavior is in line with the disagreement effect and contradicts the echo chamber amplification theory.

We also show that the level of users’ political activity is independent of the likemindedness of their geographical environment, for both pro-Obama and pro-Romney users. In particular, this implies that the dependence of PA on geographical-LM does not manifest neither the disagreement effect nor the echo chamber amplification.

We thus conclude that the level of political activity of the Twitter users correlated with likemindedness of their virtual environment and is independent of the likemindedness of their geographical environment. The exact form of correlation between the PA and the likemindedness of the virtual environment is in line with the disagreement effect and, with the exception of PA as a function of followed-LM for Romney supporters, contradicts the echo chamber amplification theory.

The main limitation of our approach is in the selection of users for our analysis. Specifically, we ignored users that are politically active but did not express explicit support for neither of the presidential candidates of the US presidential election. This obviously introduced a bias to our measurements of user’s follower-LM and followed-LM.

References

  1. N. Barbieri, F. Bonchi, and G. Manco. Topic-aware social influence propagation models. In ICDM’12, pages 81–90, 2012.
  2. C. Castillo, M. Mendoza, and B. Poblete. Information credibility on twitter. In WWW’11, pages 675–684, 2011.
  3. M. D. Conover, B. Gonçalves, J. Ratkiewicz, A. Flammini, and F. Menczer. Predicting the political alignment of twitter users. In SocialCom’11, pages 192–199, 2011.
  4. K. Dyagilev and E. Yom-Tov. Linguistic factors associated with propagation of political opinions in twitter. Social Science Computer Review, 2013.
  5. D. Gayo-Avello, P. T. Metaxas, and E. Mustafaraj. Limits of electoral predictions using twitter. In ICWSM’11, 2011.
  6. M. R. Morris, S. Counts, A. Roseway, A. Hoff, and J. Schwarz. Tweeting is believing?: understanding microblog credibility perceptions. In CSCW’12, pages 441–450, 2012.
  7. P. Moy and J. Gastil. Predicting deliberative conversation: The impact of discussion networks, media use, and political cognitions. Political Communication, 23(4):443–460, 2006.
  8. D. C. Mutz. Hearing the other side: Deliberative versus participatory democracy. Cambridge University Press, 2006.
  9. L. Nir. Disagreement and opposition in social networks: Does disagreement discourage turnout? Political Studies, 59(3):674–692, 2011.
  10. S. O’Banion and L. Birnbaum. Using explicit linguistic expressions of preference in social media to predict voting behavior. In ICASNAM’13, pages 207–214, 2013.
  11. C. J. Pattie and R. J. Johnston. Conversation, disagreement and political participation. Political Behavior, 31(2):261–285, 2009.
  12. M. Pennacchiotti and A.-M. Popescu. Democrats, republicans and starbucks afficionados: user classification in twitter. In KDD’11, pages 430–438. ACM, 2011.
  13. L. Rainie, A. Smith, K. L. Schlozman, H. Brady, and S. Verba. Social media and political engagement. Pew Internet & American Life Project, 2012.
  14. J. Ratkiewicz, M. Conover, M. Meiss, B. Gonçalves, A. Flammini, and F. Menczer. Detecting and tracking political abuse in social media. In ICWSM’11, 2011.
  15. A. Smith. Twitter and social networking in the 2010 midterm elections. Pew Research, 2011. http://bit.ly/heGpQX.
  16. A. L. Stinchcombe. Going to extremes: How like minds unite and divide. Contemporary Sociology: A Journal of Reviews, 39(2):205–206, 2010.
  17. I. Weber, V. R. K. Garimella, and E. Borra. Mining web query logs to analyze political issues. In WebSci’12, pages 330–334, 2012.
  18. J. Weng, E.-P. Lim, J. Jiang, and Q. He. Twitterrank: finding topic-sensitive influential twitterers. In WSDM’10, pages 261–270, 2010.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
""
The feedback must be of minumum 40 characters
Add comment
Cancel
Loading ...
116833
This is a comment super asjknd jkasnjk adsnkj
Upvote
Downvote
""
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters
Submit
Cancel

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test
Test description