Polarization on Twitter: An Analysis of Media
Selection in the American Political Landscape
Tom Verburg 10769633
September 13, 2018
Abstract
Social media such as Twitter are a very important source of information for many people in the world. The ease and speed of access to new information and the building of your own network online is something which we are all familiar with in this modern era. Yet, the information which pops up on your Twitter feed will probably differ from that of other people. This is often due to factors like interest, different kinds of friends or your political views. These will result in you following different influencers and media on this same online medium. In this research, we investigate how the media selection differs between people who follow different political influencers and whether there are differences in polarization levels in media selection on Twitter.
1
Introduction
Since the recent presidential elections in the United States of America, the infamous term of "fake news" has become a very hot topic. The American people were bom-barded by information through online social media, some of which was not validated or even downright false. However, at the time people did not take this into account and research suggests that this had significant influence on the results of the 2016 elections (Allcott & Gentzkow, 2017). Seeing as most Americans initially believe the fake news when they see it (Silverman & Singer-Vine, 2016), one can imagine how votes were swayed and how people started to be more critical of their information intake on social media. One of the more popular social media the American people use to read the news is Twitter.
1.1
Twitter is a social medium which is unique in its own right. Launched in October 2006, Twitter is one of the most popular micro blogging tools on the web. Micro blogging means that it will: "allow users to exchange small elements of content such as short sentences, individual images, or video links" (Micro-blogging, n.d.). In the case of Twitter, these micro blogs are called tweets and can only be 280 characters long. Each user can be followed and can request to follow other users. Tweets can have a certain set of traits, such as the number of reposts (also known as re-tweets) and favorites. Furthermore, users who have access to a tweet are also able to reply to it and thus makes it possible to create an entire chain of tweets.
Twitter has more than 330 million active users as of the 4th quarter of 2017 showing no signs of declining (Statista, 2017). At the time of the elections in 2016, almost a quarter of the American adults were using Twitter (Statista, 2016). Yet, research has shown that the average Twitter user in the US is not representative of the average American when looking at the geography, gender and ethnicity (Mislove, Lehmann, Ahn, Onnela, & Rosenquist, 2011). Seeing as the blogs are so short, peoples tweets are often about very current events or the current state of mind of the user.
Furthermore, it is such a real-time social medium that it can be used to analyze real-time events such as the occurrence of earthquakes (Sakaki, Okazaki, &
Mat-1.2 Fake News 1 INTRODUCTION
suo, 2010) or the current sentiment towards presidential candidates (Wang, Can, Kazemzadeh, Bar, & Narayanan, 2012).
1.2
Fake News
The boundaries of what is fake news and what is not are ill defined and it has proven to be quite complicated to get an exact definition. A satirical article may be labeled as fake news, yet some see satire as an important element when taking into consideration freedom of speech. In the case of satire, it is important to keep an eye on the context in which an article is presented (Allcott & Gentzkow, 2017). This is only a single example of many that illustrates how complicated it can be to label something as fake news. In this research we define fake news as news articles/posts on social media with the only purpose to attract the attention of its users, either through creating controversy or if the article contains incorrect information and/or can not substantiated by facts or evidence.
Fake news also seems to have a special appeal to adolescents, seeing as they gather most of their information from the web (social media) and find the controversy and authentic feel of hard truths as presented by biased media more interesting than the ones spread by objective hard truth media. This is because the objective media are bound to certain rules (Marchi, 2012). Seeing as adolescents are a group of the population who receive most of their information from social media (Marchi, 2012), it would suggest that they are a susceptible audience to the exposure of fake and/or controversial news.
1.3
Trump and Social Media
Through the use of free media (such as Twitter) president Trump was able to do catch the attention of these adolescents and all other social media users who follow him. Seeing as he wasn’t bound to any rules, he repeatedly created controversy online which was picked up by news agencies and resulted in free coverage. This was probably done deliberately and is a strategy which has existed long before the era of social media (Francia, 2017). In this current age, the strategy only has more benefits compared to back then. Trump is able to completely put traditional media out of play by addressing the American people directly through his social media accounts, without making any costs and discrediting opposing news media by calling them fake. The only thing Trump needed was a following, something which he already had and kept growing due to his controversial remarks he kept making. By catching everyone’s attention online he was able to take on a persona who is provocative, has a bare knuckle authenticity, speaks half truths and has a very politically incorrect speaking style: all of which greatly appealed to the disaffected voters and adolescents who are interested in these hard ’truths’ (Gurevitch, Coleman, & Blumler, 2009; Francia, 2017)(Marchi, 2012).
1.4
Relevance
On Twitter, president Trump is currently still posting tweets about controversial rants concerning fake news and bias against him. He repeatedly criticizes news agencies and tries to discredit them by selecting his own sources/other news agencies who agree with him and present information favoring him.
In his tweets, Trump calls out CNN, NBC news and the NYT for their allegedly poor reporting whilst praising Fox news for reporting the news correctly. Seeing as people are very aware of the circulation of fake news, it results in them selecting their sources of information very carefully. This and the crude manner of shaming and praising by Trump may have a polarizing effect on what is already a very divided country. Furthermore, Trump’s Twitter profile is still one of the fastest growing profiles when it comes to the gathering followers with having over 60 million followers at the moment of writing this paper. However, he is not the only one who addresses his "followers" directly through the use of social media. There are a variety of influencers active on Twitter who use it as a medium to inform and connect to their followers.
1.5 Research Question 1 INTRODUCTION
The influencers whose networks were analyzed for this research can be seen in Table 2 and are mainly American politicians.
1.5
Research Question
With the issues described above, this research attempts to explore whether networks of followers (such as Trumps) differ between politicians and groups of politicians. The research will focus on the polarization of information selection by the users who follow various politicians and analyze the differences between these networks. This resulted in the following research question:
To what extent does the political media polarization of the politically active network on Twitter differ between American political influencers?
The research question is posed in this manner because there are a lot of people who follow various politicians on Twitter, but do not have a strong political opinion, as well as accounts that follow politicians but are either bots or influencer accounts which only exist to build as large a network as possible. Therefore both these type of followers should not be taken into account when analyzing the polarization between different networks. It is important to note that all use of the phrase "polarization" is meant in the American political context. Furthermore, we only look at American politicians because of how the American political landscape is built. Bluntly stated, there are only two sides: the Republicans and the Democrats. Because of this fact polarization in the American landscape is one dimensional and can be defined fairly simple. Lastly, the network is the selection of users who chose to follow certain influencers and media on Twitter. In other words, each influencer is a central node to which ’normal’ Twitter users connect to. The interconnectivity between ’normal’ users is not analyzed, only their connectivity to media outlets as well as other influencers.
2 PREVIOUS RESEARCH
2
Previous Research
In this section the existing research concerning the analysis of tweets and polarization is presented. Firstly, Twitter usage and the medium itself are discussed. Secondly, there is the manner of what relationships have been explored between the polarization of users as well as the use of social media. Thirdly, there are the tweets of president Trump and the existing research concerning these tweets. Fourthly, research concern-ing the analysis of tweets in various contexts is discussed as well as how polarization will be measured.
2.1
Trust in Media
The American people have long lost their trust in the media, especially the conser-vative republicans. A large amount of the republicans who follow politics, believe that the media has a liberal bias (Jones, 2004), despite there being evidence to the contrary (Lee, 2010). However, there has been evidence that there is a link between the claims of the news media having a liberal bias, and the trust of the public in these same news media (Domke, Watts, Shah, & Fan, 1999). This evidence shows how powerful the Twitter activity of politicians can be in relation to the public view and trust in the media, especially if they are not dependent on traditional media to spread their thoughts on the liberal bias/fake news but simply use social media to reach out to the public.
2.2
Polarization on Twitter
Social media has opened up many different possibilities for people to personalize the manner in which they take in information. First of all, a user is able to connect to a vast network of other users and exchange information. Secondly, a user is able to select their own sources of information from other parties. This can result in the user only selecting sources which share stories that align with the users’ view of the world and become unfamiliar with news which contradicts the users’ belief. Furthermore, the source of the news becomes less important in the online landscape, seeing as people tend to believe something if it is in alignment with their own beliefs. Research has shown that this can cause fragmentation in media environment online and polarizes individual-level attitudes (Stroud, 2010). Social media (and the rest of the Internet) tend to show users what is most interesting to them based on a variety of algorithms with the end goal to create the largest revenue and/or to catch the users interest. Due to this, a user can end up in a filter bubble in which all that is shown to the user is what aligns with his/her interests and concurs with their view of the world (Pariser, 2011). This is one of the greater threats in the battle against polarization online (Messing & Westwood, 2014).
As mentioned earlier, personal selection can lead to fragmentation between users, which has proven to lead to polarization (Stroud, 2010). Yet, because of the exposure to the opinions of other people reacting upon these partisan sources online, users still come in contact with information which is counter attitudinal compared to their offline interpersonal discussions and traditional news media(Messing & Westwood, 2014). Even though users are not "connected" to other users by by a direct link (friends on Facebook or following each other on Twitter), their opposing attitudes still clash because social media facilitates a platform which allows for discussion. However, recent study (Garimella & Weber, 2017) suggests that political polarization has increased between 10% and 20% over the last 8 years on Twitter. Moreover, this same research goes even further by suggesting that polarization will likely increase even more due to politicians like Trump who generate a lot of controversy. This research, among many others, draw its foundations from other research about online blogging, the front runner of the micro blogging service Twitter. There are various different methods to measuring the polarity of a text which have been used in the past for analyzing either blogs and/or micro blogs.
2.3 Polarization in Tweets 2 PREVIOUS RESEARCH
2.3
Polarization in Tweets
One of the ways polarization can be measured is by applying natural language pro-cessing (NLP) techniques on the public tweets of people who are actively sharing their thoughts and opinions on Twitter. However, it has proven to be quite a chal-lenge to use NLP techniques to accurately classify the subjectivity of texts. In the next part, NLP and a few other techniques will be addressed and this will illustrate the difficulties associated with the measuring of polarization on Twitter.
2.3.1 Manual Labeling of Tweets
There has been a lot of research concerning the content of blogs/tweets that rely upon the manual labeling of texts (or in this case tweets).
Most of the time there are three different classes in which the texts can be placed such as positive, neutral and negative. These can have different labels such as left, neu-tral and right when talking about politics (Conover et al., 2011; Gilbert, Bergstrom, & Karahalios, 2009). The advantage of this method is that the researchers do not need to build a specific classifier for the specific set of data in a specific domain, but the main disadvantage is that there is an enormous bias based on that the classification. This is because the classification is prone to the subjectivity of the person who clas-sifies the data. However, this can be overcome if the data is classified a second time by another individual using the same classes and then statistically compared to the original to see whether there are significant differences between the two individuals.
One of the researches which applied manual labeling is the research by Yardi and Boyd in which group polarization on Twitter was analyzed concerning the shooting of Tiller, a late term abortion doctor. Seeing as this topic has quite a political load (especially in the US), it is very relevant to this research. Their results concerning polarization suggest that the discussion on Twitter could have positive effects on group polarization, seeing as people are exposed to a variety of opinions. Yet, a real discussion or debate is hard to facilitate between opposing sides.
In this case, the decision was made to label all the tweets manually because the author put the argument forward that it was unclear how NLP techniques would be applicable on texts such as tweets (Yardi & Boyd, 2010).
2.3.2 Algorithmic Classifier and NLP
Another method which has also has been used to measure polarization in tweets is the use of a algorithmic classifier. By extracting features from the text using a variety of NLP techniques and combining this with machine learning, a classifier can be trained to detect polarization in text.
One of the more interesting and modern NLP techniques is sentiment analysis. Sentiment analysis is a machine learning method which computationally deals with the opinion, sentiment and subjectivity of a text (Pang, Lee, et al., 2008). Therefore, it is applicable when looking to analyze the political opinions/polarization of individuals on Twitter. However, sentiment analysis can run into trouble when sarcasm is present, which is the case for most social media including Twitter (Maynard & Greenwood, 2014). This machine learning technique has been used before for the analysis of social media texts (Ceron, Curini, Iacus, & Porro, 2014), algorithmically measuring the agreements of comments with corresponding blogs (Gilbert et al., 2009) and predicting elections (Tumasjan, Sprenger, Sandner, & Welpe, 2010).
However, when looking at the research done concerning polarization on social media, sentiment analysis is often only taken as a feature to be taken into account when creating an algorithmic classifier (Gilbert et al., 2009). Furthermore, such a classifier also needs training data which would have to be manually labeled. Still, research also suggests that sentiment analysis could illuminate some existing questions concerning polarization on Twitter (Conover et al., 2011).
2.4 Polarization in Twitter Network 2 PREVIOUS RESEARCH
2.4
Polarization in Twitter Network
The research of Garimella and Weber performed resulted in a long term analysis of polarization of Twitter of which the focus lay on the retweets, follows and hashtag use of politically active Twitter users. Their dataset was collected by analyzing the presidential and vice-presidential candidates of the previous 8 years and their parties, as well as the Twitter accounts of popular media outlets. With all these accounts in mind (each with their own left or right label), all of their public followers and online behavior since 2009 were collected. The behavior entails their followers, retweets and hashtags in political context and all of this was analyzed over time.
Therefore, the focus of this particular research was not on the content of the tweets (with the exception of the politically loaded hashtags) but on the interaction and re-lations between users and politicians/politically biased news outlets. The research of Garimella and Weber laid bare the network of politically active users and show its polarization on 3 different fronts based on the interaction between users and the politically biased Twitter accounts (either politicians or news outlets) instead of ana-lyzing the text in the tweets themselves. The conclusion of the research of Garimella and Weber research was already mentioned before in this paper, which was that there was an increase between 10% and 20% in polarization over the last 8 years on Twitter over all 3 different types of polarization: hashtags, followers and retweets.
Another research which also only looks as the network aspects of Twitter when researching polarization was the research of Conover et al.. In this particular paper, the focus lies on the retweets and the mentions of politically active Twitter users. In this case, political conversations were found through their use of politically biased hashtags. They then used cluster analysis to analyze both the mention and retweet networks, as well as analyze the content of the tweets by performing a cosine similarity on the hashtag use. The result was that there was strong polarization in the retweet network, but not in the mention network on Twitter (Conover et al., 2011).
2.5
The Polarization Hypotheses
In previous research, polarization online could often be split into two different cate-gories. Firstly, there is polarization which can be perceived in the content of what people were posting (content), the second is polarization in how different users and influencers are connected (network).
Seeing as NLP techniques are very difficult to implement onto 280 character tweets (which include sarcasm, emojis and sentiment), previous work concerning content often focused on the political biased hashtags and not the text itself.
For the network polarization we make the distinction between the mention network and the friend/follower network. The mention network consists of what kind of other accounts with political bias users mention in their tweets. The friend/follower network consists of the network of other users which has been built around a user.
As stated before, research has shown that the calling out of political bias has an influence on the view of the media. The same applies when influencers directly discredit certain media outlets (Domke et al., 1999). In this research however, the focus lies on how different groups of political followers inform themselves on social media. This is because people are able to follow a variety of different media outlets on Twitter that give them an up to date news feed. However, this news feed can be its own filter bubble due to the polarization in the selection of sources by the user (seeing as there are many biased and partisan news outlets active on Twitter). Therefore, the type of polarization this research focuses on is the polarization of the friend/follower network to see how people (who are part of different followings and political affiliation) wish to inform themselves on Twitter.
Based on these ideas and concepts the following hypotheses are presented: H0: There is no difference in polarization between networks of American political influencers on Twitter.
H1: There is a difference in polarization between networks of American influencers on Twitter.
2.5 The Polarization Hypotheses 2 PREVIOUS RESEARCH
As discussed before, polarization is American political polarization which is mea-sured by the selection of partisan/biased media outlets. The networks are the users who follow these influencers, and it is their polarization which is measured.
3 METHOD
3
Method
In this section the measuring method, as well as the exact manner in which the data was collected from Twitter are discussed.
3.1
Method of Polarization
One of the first issues which needs to be addressed is how polarization is measured for this research.
This research will not focus on any natural language processing techniques seeing as it would result in this research becoming a classification problem, as well as that it would have a heavy bias (seeing as the training data would be labeled by a single person). Instead, the polarization is measured based on the relative amounts of left to right leaning media outlets a user follows. This research lends its foundation from a previous research which revolves around the connectivity between left and right leaning Twitter media outlets, influencers and the connectivity of politically active users to these accounts.
So, to define the measurement of polarization we first explore how this research differentiates between different kinds of media outlets and influencers.
Table 1: Media Outlets from Garimella and Weber(2017)
3.1.1 Media Outlets and Influencers
In the research of Garimella and Weber (2017), there was a distinction made between left and right leaning Twitter accounts to measure polarization. This distinction was based off the Pew Research Center (Gottfried, Kiley, Matsa, & Mitchell, 2014). For this research, the same distinction and selection concerning media outlets and influencers will be made. These can be seen in Table 1 and 2.
3.1.2 User Dataset
The user dataset consists of politically active Twitter users. This means that users will only be selected if they have tweeted, retweeted, or replied to a tweet which mentions one or more of the influencers mentioned in Table 2. More on how this was done is explained in section 3.2.
This research makes the distinction between different types of followers based on which influencers they follow. The reason for this is that the goal of this research is to compare the polarization between different networks of people around certain influencers on Twitter.
3.2 Data Collection 3 METHOD
3.1.3 Measuring Individual Polarization
The actual measuring of the individual user polarization is based on the research of Garimella and Weber (2017) as well. In their research the focus lies on several different kinds of polarization, but we will only adapt the method used for measuring the follower polarization. A Bayesian methodology is used in which results in each user having a prior probability to follow certain media outlet accounts from either side. It is a beta distribution in which the prior probability to a user leaning to either side is the same:
α = β = 1 (1)
Following this, every follow of a media outlet increases the α or the β by 1, like a repeated coin toss which then changes the beta distribution according to whichever side the observed user is leaning towards. This leaning factor is defined as:
γ = α/(α + β) (2)
Based on the leaning value of γ the polarization is defined as follows:
ρ = 2× | 0.5 − γ | (3)
This results in a polarization value (ρ) between 0 and 1.0 and is the deviation from a balanced leaning. All three of the above formulas and definitions are completely identical to the ones used for measuring polarization over time in previous research by Garimella and Weber (2017).
3.2
Data Collection
The data was collected through the official Twitter API as well as an unofficial Twitter scraper. For this research, the official Twitter API was avoided as much as possible seeing as the rate limits pose a problem for the gathering of large amounts of data. However, some data could not be mined through the use of unofficial scrapers which meant that the official Twitter API still needed to be used. The initial data which was gathered was scraped by an unofficial scraper and resulted in a large amount of tweets which were directed to one or more of the influencer accounts which are described in Table 2. The ids of the authors/users of these tweets were then collected. With the use of the official Twitter API the friend lists of these users were collected which included which media and influencers they followed.
Follower lists can only be accessed through the official Twitter API and it only allows for a maximum of 5000 followers to be returned after a single call. This means that it would take over a week with the current rate limits to gather the entire follower list of Trump (seeing as he has over 60 million followers), but this list is useless without gathering the followers of these followers (depth = 1) and see what kind of partisan media outlets they follow. Furthermore, the manner in which the Twitter API returns the follower list is in a chronological order. This means you only get the most recent followers, which contains a large amount of bot accounts (seeing as there are more than 48 million bots on Twitter (Varol, Ferrara, Davis, Menczer, & Flammini, 2017) and they are not yet ’caught’ by the system and banned, because they have only recently been created) as well as non American people who do not necessarily have any political opinion regarding American politics but only follow high profile influencers because they are celebrities.
To overcome the rate limit problem, as well as the chronological bias, a large set of tweets that contain, reply or retweet a mention of one or more of the influencer accounts (see Table 2) were collected using an unofficial Twitter scraper. These tweets could then be traced back to a unique account.
The users who were collected based on these tweets are users who have expressed their political opinion on Twitter seeing as they had actively shared, commented or replied about one or more of the influencers. With this in mind, their Twitter account would probably reflect their political sympathies, as well as show how they wish to inform themselves by their selection of media outlets.
3.3 Processing the Raw Data 3 METHOD
The friends (the other accounts they follow) of these politically engaged Twitter users were collected through the official Twitter API and resulted in a dataset which contained the Twitter ID’s of these users and their corresponding friend lists (the other users they chose to follow). However, because we used this earlier scraper to search for tweets based on a certain query, we have to take into account that there may be a small bias to the set of tweets that were collected. This can however not be determined accurately, seeing as the exact manner in which the scraper interacts with Twitter is unknown.
All of the programming was done in Python using iPython notebook and a variety of the Anaconda package libraries were used, as well as some external libraries for the gathering and processing of the Twitter data which was collected.
3.3
Processing the Raw Data
After the process of gathering data using the official API, over 10 000 Twitter users and data about their friend lists had been gathered. Due to the privacy reasons of Twitter, no demographic data of these users could be accessed through this API and therefore there is very little known about individual users except for the fact that they tweeted or replied to something related to an influencer, and which other Twitter accounts they follow on Twitter. For each individual account the friend list was cross referenced with the accounts of interest for this research, both the list of media accounts and the list of influencers (Table 1 and 2). Those who did not follow either a media account or an influencer were automatically removed seeing as they were not exposed to influencer or media alike. This resulted in the dataset shrinking to 9148 individual users with their own selection of influencers and media outlets. 3.3.1 Grouping Variables
The next part consisted of the grouping of the users based on their exposure to differ-ent influencers. Seeing as following one influencer does not exclude the possibility of the following of other influencers, initially all combinations were taken into account. This means that people who only followed Obama were assigned to a different group than people who followed Obama and Trump. With this dataset, taking into account all the possible variations, this resulted in 887 different variations of influencers being followed, with some combinations only recurring a single time. Taking into account the overlaps these variations would have (2 different groups who completely overlap in terms of influencers, except for a single one should perhaps not be treated as com-pletely different groups) it would be a very time consuming and an overly complicated task to find anything substantially relevant.
For the statistical analyses between these groups, small samples such as these would decrease the statistical power significantly (Burns & Burns, 2008). Even when only taking into account variations which had an occurrence of n>30, the dataset still returned 44 different groups which meant adopting this method would result in more than 900 pairwise comparisons for a basic means comparison. The baseline for n>30 was initially held seeing as this is where the distinction is made between small and large samples (Triola, 2008).
However, seeing as this still resulted in 44 groups, several other grouping methods are proposed for making a distinction between different groups of politicians and their networks.
The first method is splitting the dataset of users based on the political affiliation of the influencers they were following, instead of looking at the individual influencers. This grouping variable is labeled "influencer polarization values". Based on the po-litical affiliation of the influencers, users were given a influencer score from 1 till 6, with 1 for users who only follow left influencers, and 5 for users who only follow right influencers. Group 6 is for users who do not follow any influencers at all and are classed as "No Influencer", whereas users who follow the same amount left and right influencers, are categorized as "Neutral", also known as group 3. All of these can be seen in Table 3.
3.3 Processing the Raw Data 3 METHOD
The second method is an approach to see what kind of influence individual in-fluencers may have on polarization. This grouping variable is labeled "individual influencers", which repeatedly compares the polarization value of users who follow a certain individual influencer to those who do not. The main motivation for this method was to compare a network of people who choose to follow a specific person to those who do not. Certainly with the high profile Twitter influencers, not deciding to follow a specific influencer can be a strong statement on its own. Therefore the individual groups of users following each of the influencers were selected, as well as the groups who did not follow each specific influencer. Every group who followed a single influencer was then pairwise compared to the group who did not follow that particular influencer, and this was repeated for all 12 influencers noted in Table 2.
The third grouping method focuses on the total amount of influencers a user follows. Based on this amount, users are divided into one of 4 categories. In the rest of the research, this grouping variable is labeled as "amount of influencers". The first category are users who follow 1-4 influencers, the second 5-8 influencers, the third 9-12 influencers and the fourth group are the users who do not follow any influencers at all. This grouping method was used to see whether users who follow different amounts of influencers have different polarization levels and whether those who follow more influencers become less polarized due to the fact they are exposed to both left and right influencers.
4 RESULTS
4
Results
In this section all of the results of the statistical analyses performed are presented.
4.1
Descriptives
After the raw data was processed, the dataset contained 9148 different users and what accounts they followed. For the first grouping variable, we looked at what influencers the users followed. This resulted in the users being categorized in 6 different groups based on the political affiliation of the influencers in their follower list. The grouping of these values was under the variable name "influencer polarization values", seeing as each user had a polarization value based on which influencers they were following.
Table 3: Descriptives Influencer Polarization Groups
The descriptives containing the names and amount of groups can be seen in Table 3, containing the polarization mean, standard deviation and median values per group for this first grouping variable.
As can be seen in the Table 3, the groups are not similar in size but are all of significant size. Furthermore, the mean polarization values of these different groups can also be observed in this Table.
The descriptives for the grouping variable based on the amount of influencers followed against Polarization can be observed in Table 4 along with the mean, median and standard deviation values of polarization for all the different groups respectively. The descriptives for the grouping method individual influencers can be found in the appendix.
Table 4: Descriptives Amount of Influencers followed across Polarization
4.2
Exploring the Data
A Kolmogorov-Smirnov test as well as the Shapiro-Wilk test were repeatedly used to test for normality on the dependent variable Polarization and 3 independent grouping variables. These include the independent variable influencer polarization groups (left, slightly left, neutral etc.), the variable based on whether a user follows a individual
4.3 Non-Parametric Kruskal-Wallis Test 4 RESULTS
influencer and the final grouping variable which categorizes based on the amount of influencers a user follows on Twitter.
For all these different grouping variables a large part of the corresponding groups showed a significant result for both tests (p<0.05). Even though both the Kolmogorov-Smirnov and the Shapiro Wilkens are very strict for large samples, visual inspection of both the boxplots and the histograms confirm the significant results and suggest that the data of each sample per group is not distributed normally, but this is also due to the measuring method of polarization that is being used for this research, seeing as it is a normalized measure of left and right leaning Twitter users instead of left versus right. Yet, seeing as this is what will be compared between different networks around influencers, the assumption can be made that the groups do not have a normal distribution.
Upon further exploration of the data, a non-parametric variant of the Levenes test was performed for all the groups of the 3 grouping variables separately, to see whether there is homogeneity of variances. This method is based on the research of Nordstokke and Zumbo (2010) and performed with the use of ranked means. The Levenes tests indicated unequal variances (p<0.05) for all pairs for 2 out of the 3 grouping variables, suggesting heterogeneity of variances. However, the remaining grouping variable based on individual influencers indicated unequal variances for most pairs (p<0.05) with the exception of 2. These can be viewed in the appendix. Seeing as most pairs still indicated heterogeneity of variances, it was decided to repeatedly perform the Mann-Whitney U for all the pairs.
4.3
Non-Parametric Kruskal-Wallis Test
Seeing as there are 2 violations of assumption for the parametric one-way ANOVA for all three grouping variables, the asymptotic non-parametric Kruskal-Wallis test was performed for two of the grouping variables (those with the number of groups>2) to compare the ranked means of the dependent variable polarization against the different groups. The ranked means of the Kruskal-Wallis test for both grouping variables can be seen in Table 5 and 6
Table 5: Mean Ranks Influencer Po-larization Values
Table 6: Mean Ranks Amount of Influ-encers
The first Kruskal-Wallis test compares the independent grouping variable influ-encer polarization values against the rank means of the dependent variable polar-ization. This resulted in a statistically significant difference (H=684.814, p<0.001). Upon analyzing the post-hoc analysis of the different pairwise comparisons between the 6 groups the results indicated that 9 out of the 15 pairs showed a significant difference between samples (p<0.05). The exact results can be seen in Table 7.
4.3 Non-Parametric Kruskal-Wallis Test 4 RESULTS
Table 7: Kruskal-Wallis: Pairwise Comparisons Influencer Polarization The second Kruskal-Wallis test compares the rank means of polarization between users who follow different amounts of influencers. The results indicated a significant difference (H = 811.098, p<0.001), which suggests that there are differences in po-larization between people who follow different amount of influencers. The post-hoc pairwise analysis can be observed in Table 8.
4.4 Non-Parametric Mann-Whitney U test 4 RESULTS
4.4
Non-Parametric Mann-Whitney U test
Based on the results of the of the Kruskal-Wallis test, a variety of Mann-Whitney U tests were performed to see how large the differences were between the different pairs for different grouping variables by calculating the effect size based on the Z value of the Mann-Whitney U. Furthermore, a repeated Mann-Whitney U test was performed for the second grouping variable: individual influencers.
Table 9: Mann-Whitney U: Influencer Polarization Groups across Polarization For the grouping variable influencer polarization values, this meant performing 15 pair analyses.Te results are shown in Table 9, with the p value adjusted using the Bonferroni correction to minimize the type I error.
The next Mann-Whitney U test was performed repeatedly with the grouping vari-able individual influencers against the dependent varivari-able polarization. In the ex-ploration of the data, two assumptions for the T-test (homogeneity of variance and normal distribution) were violated for this grouping variable, which means the Mann-Whitney U test had to be performed. More about this can be read in the Exploring the Data section and the appendix.
4.4 Non-Parametric Mann-Whitney U test 4 RESULTS
Table 10: Mann-Whitney U: Including/Excluding Individual Influencers across Po-larization
All of the 12 tests resulted in a significant difference (p<0.001) for all pairs, as can be seen in Table 10. Yet, this can also be written off to the size of the samples. Therefore, the effect sizes are more interesting to compare which are also displayed in Table 10.
Table 11: Mann-Whitney U: Amount of Influencers across Polarization The third and final repeated Mann-Whitney U can be observed in Table 11. This was for the grouping variable amount of influencers against the dependent variable polarization. Seeing as a Kruskal-Wallis was already performed on this grouping variable as well, this was done as a post-hoc method to calculate the effect size and confirm the results of the previously performed Kruskal-Wallis test. All of the pairs showed a significant difference (p<0.05). Furthermore, the p value was adjusted using the Bonferroni correction to minimize the type I error.
There are many significant results for all the Mann-Whitney U tests for all 3 grouping variables, yet it is important to note that the sample was quite large for all of these which quickly results in significant results. The effect size therefore becomes an important factor for whether the inferred difference is either worth mentioning, or insignificantly small but bloated due to the large sample size. For this research the effect size was calculated as eta squared and calculated according to the formula of Fritz, Morris, and Richler (2012):
η2= z 2
4.4 Non-Parametric Mann-Whitney U test 4 RESULTS
Here η2 is Eta squared, z is the z statistic and N is the sample size. The effect sizes based on this equation 4 can be seen in Table 9, 10 and 11. The rule of thumb for interpreting η2 are that a large effect is larger than 0.14, a medium effect larger than 0.06, and a small effect is larger than 0.01 (Cohen, 1988).
4.4.1 Absolute and Real values for Polarization
Lastly, the polarization measuring method was analyzed by calculating the real values. This was done by removing the modulus from equation 3. It resulted in a new variable which differentiates between left and right leaning polarization, left being < 0 and right being > 0. For the two grouping variables where the core focus lay on the difference of the political affiliation between the influencers followed (Influencer polarization values and the Individual influencers), the values for the absolute and real polarization variables were compared. The values for the grouping variable influencer polarization values can be observed in Table 12. The results for grouping variable individual influencers can be consulted in the appendix.
5 DISCUSSION
5
Discussion
The question was posed whether there are differences in the levels of polarization between the networks surrounding different political influencers on Twitter. Based on previous research, a set of influencers, biased media outlets and a measure of polarization were adopted to answer this question. By scraping a large amount of tweets and retracing the users who wrote these tweets, a large set of users and their friend network on Twitter were analyzed, with a focus on whether the polarity of the media outlet selection differed between different sets of users based on what kind of influencers they followed. The results of this paper suggest that there are differences in polarization between different networks of individual influencers and groups of influencers and how this was deduced is explained in this section.
5.1
Sample, Data and Measuring Method
The sample consisted of 10 000 random Twitter users whose only prerequisite is to have tweeted or replied to a tweet which contained a mention to one or more of the influencers. This condition was an attempt to exclude bots as well as include users who actively have and express their political opinion on Twitter. The reasoning behind this is that these Twitter profiles and their corresponding network of friends would probably be a more accurate reflection of their own political views and attitude towards news media, compared to users who do not express their political opinion online.
For the selection of media outlets, measure of polarization and the set of influ-encers, the article of Garimella and Weber (2017) was consulted and several definitions were assumed unaltered and make up a large part of the academic substantiation for the interpretation of the results for answering the research question.
When looking at the descriptives in Table 3 of the first test performed between the groups of different influencers, only a small group of users from the sample did not follow any influencer at all. This only confirms what the idea behind the sampling method was, namely selecting users who have and express their political opinion on Twitter.
5.1.1 Difference between Absolute and Real Polarization values: Analysing the Measuring Method
For this research the measuring method of polarization as well as the media selec-tion were adapted unaltered from the research of Garimella and Weber (2017). It is of importance to note that although both this research as well as the research of Garimella and Weber (2017) measure polarization, the distinction between left and right polarization was not made up till now. After removing the modulus from equa-tion 3 the equaequa-tion yields a value of polarizaequa-tion which leans either left or right. After analyzing the results for the grouping variables which (assumingly) yield either left or right leaning groups of users, it was clear that the distribution of polarization is not divided right through the middle. The deviation from 0 (which is the value for users who are not polarized at all) differs greatly between users who follow left and right influencers. For the left leaning users the distance is almost the mean and me-dian polarization values are almost the same compared to the absolute polarization values. However, the users leaning right have very different values as can be observed in Table 12 and the appendix for the grouping variables influencer polarization values and individual influencers. When observing the latter, it important to focus on the group of people who chose to follow a certain influencer seeing as this is always a conscious choice by the user. It can be observed in the appendix that the users who follow the left influencers have very similar mean deviations from 0 (eg: 0.57 and -0.56 for Joe Biden) between the real and non-absolute values for polarization, whilst right influencers have very different deviations between the two measuring methods (eg: 0.46 and -0.10 for Paul Ryan). By making the differentiation between left and right, it suggests that there is a bias to the left/negative side of the measuring scale and results in users who follow right influencers having right to slightly left polarization
5.2 Interpreting Results 5 DISCUSSION
values, in comparison to users who follow left influencers who have very strong left leaning polarization values.
Also, when looking at the boxplot for the distribution of the polarization values for all the users, it can be observed that the median not 0, but slightly negative when using this non-absolute measure of polarization. This only supports the previous observations about users favoring left media. The boxplot can be observed in the appendix.
As mentioned earlier, the selection of media outlets selection (Table 1) used for this research was not balanced equally when looking at both the amount of right and left partisan media. This could be an explanation to why the mean non-absolute/real polarization value is on the negative/left side of the distribution and why all users have a tendency to be polarized in the left direction.
5.2
Interpreting Results
In the following subsection, the interpretation and underlying statistics of the results for each of the three grouping methods are discussed.
5.2.1 Influencer Polarization Values: Discussion
The Kruskal-Wallis test measuring polarization between groups based on their influ-encer polarization values, showed that the polarization between these groups differed significantly, which allows us to reject the H0 hypothesis and accept the alternative H1 hypothesis.
Table 7 shows the pairwise analysis between the different rank of means between all the groups, and shows that there is a significant difference (p<0.05) between 9 out of 15 pairs. The pairs that jump out are those whose mean ranks do not differ significantly: Right-Neutral, Right-Slightly Right, Slightly Right-Neutral, Neutral-No Influencer, Slightly Right-No influencer and Left-No influencer. Still, of these pairs, only 3 showed a non significant result (even after the Bonferroni correction) for the Mann-Whitney as can be seen in Table 9. This could be due to the Mann-Whitney being too powerful with these large samples and resulting in the Type I error being too large. Therefore, the pairwise post-hoc analysis of the Kruskal-Wallis provided by SPSS is used for looking at the significance, whilst the Mann-Whitney is used for calculating the effect size.
When consulting the descriptives in Table 3 and the mean ranks in Table 5, it can be observed that the group ’Right’ has the lowest mean followed by the group ’Neutral’. Furthermore, users who only follow right partisan influencers don’t only have the lowest mean value for polarization, the means are lower than users who do not follow any influencers at all. On the other hand, people who follow mostly left influencers have a higher mean and mean rank than those who only follow left influencers.
The pairwise analysis of the Kruskal-Wallis shows us that the pair Right-Neutral does not differ significantly and this is supported by the significance displayed in Table 9 (which was the result of pairwise Mann-Whitney U test to calculate the effect sizes). On the other hand, there was a significant difference (p<0.05) between the ’No Influencer’ group and the ’Right’ group for both the post-hoc pairwise comparison of the Kruskal-Wallis (Table 7), as well as the Mann-Whitney U test (Table 9) which suggest that there is a significant difference in polarization between these two groups. Significant differences alone are not conclusive with samples as large as these. Therefore, it is of importance to analyze the effect sizes as well. As can be seen in Table 9, 3 pairs had a medium effect size and 5 show a small effect size .The remaining pairs either showed no significant difference (p<0.05) or had an effect size too small to have any significant meaning when upholding the rules of thumb set by Cohen (1988). However, the other 8 pairs did show effect sizes of significant size, which suggests that there are significant and relevant differences of polarization between groups, based on the grouping variable influencer polarization values.
5.2 Interpreting Results 5 DISCUSSION
5.2.2 Influencer Polarization Values: Further Observations
One can imagine that those who do not follow any influencer or follow a variety of left and right influencers have a similar low polarization value. Yet the question which comes to mind is: how does a group which solely follows right influencers have a lower polarization mean rank and differs significantly from a group of users who follows no influencers at all? Furthermore, how is it that the ’Slightly Left’ group has the highest mean and mean rank value for polarization?
When looking at Table 1, it can be observed that the total amount of left and right media outlets which were used to calculate the polarization are not equal. There are 18 media outlet accounts leaning left, compared to 7 media accounts leaning right. This was the case because both the list of influencers and the media outlets were adopted from the research of Garimella and Weber (2017) and no adaptations were to these lists. Seeing as there are fewer right compared to left media outlets, it could be argued that a user who randomly follows a selection of media outlets, will have a large chance of leaning left and have a polarization value. A user who is right will probably follow a couple of (extreme) right partisan media outlets, as well as some moderate left partisan mainstream media. On the other hand, seeing as the extremity of the bias of the various media outlets is not taken into account, the right user will probably have a lower polarization value compared to a user who randomly follows a couple of media outlets on Twitter according to this method of measuring polarization.
Another possible explanation could be that this is due to the fact that around half of the people who only follow right influencers, only follow Trump. These are then labeled as absolute right leaning users, while they might only follow him because he is a controversial figure on Twitter and following him can be seen as the most direct link to the current president. When analyzing whether this is the same for the users who only follow left media, only about a quarter of the left group follow only Obama or only Clinton.
As was mentioned earlier in this paper, right (republican) people have a history of mistrust in news media. Along with the fake news on social media online, as well as Trump calling out fake media, republicans will probably follow fewer media online compared to democrats. This could be another possible explanation for lower polarization values for the groups who follow Right influencers, seeing as these users are more particular in which media they choose to follow.
A final observation is how the group who follow no influencer at all have a higher mean ranking than a lot of other groups. This could be due to the fact that not following a influencer does not strip you of a political opinion, but may even show you have a distaste for the current influencers and politicians. However, the results suggest that this is a very mixed group because the pair No Influencer-Neutral did not show a significant difference for the Kruskal-Wallis post-hoc. So, it is more likely that this is due to the fact that the measuring method used is a little skewed towards the left, and this in turn is probably because of the uneven ratio left to right media outlets.
5.2.3 Individual Influencers: Discussion
A repeated Mann-Whitney U test was performed between users who followed specific influencers and those who did not for the second grouping variable. For all the cases there was a significant (p<0.001) difference between the network that followed a specific user, and those who chose not to. Yet, this difference could be written off to the sample size. Therefore, the effect size was calculated in the same manner as with the influencer polarization values (see equation 4) and the results can be seen in Table 10. As can be observed, all the rows in the table (which is ordered from small to large on the η2 column) can be split in two based on the political affiliation. The upper half contains the lower effect sizes and are the right influencers, whilst the lower half has the larger effect sizes and are all the left influencers.
Beside Trump, there were 3 influencers who had a small effect size and 4 who had a medium effect size when upholding the rules of thumb by Cohen (1988).
5.3 Limitations and Future Research 5 DISCUSSION
This suggests that for most influencers, it can be inferred that there is a significant and relevant difference in polarization between people who follow a specific influencer, and those who do not. This substantiates the earlier rejection of the H0 hypothesis. Still, there are some other interesting observations worth noting.
5.2.4 Individual Influencers: Further Observations
Yet again, when a user chooses to follow a left influencer it has a bigger impact on the polarization value than a user who chooses to follow a right influencer. As mentioned before, this could be due to the fact that a left user is easier polarized with this current measuring method than a right user, mainly because the amount of left and right media which are taken into account are not equal as can be observed in Table 1. The fact that right users are more mistrusting of media outlets and wary of fake news could also have an influence. Especially because of the large selection of media outlets users can choose from and when a user follows influencers who repeatedly call out various media for being fake.
5.2.5 Amount of Influencers followed: Discussion
For the final grouping variable, a final Kruskal-Wallis was performed which returned a significant result (p<0.01) which suggests that there are differences in polarization between different groups based on the amount of influencers one follows. This in turn again substantiates the rejection of the H0 hypothesis for the third and final grouping variable, if the effect sizes are of significant size.
The post-hoc as well as the Mann-Whitney U test which were performed show that there are significant differences between all pairs with the exception of the pair Small Amount-None. However, for this grouping variable the effect size for all comparisons was calculated as well. These show 2 insignificant, 3 small and 1 medium effect size. Again, for this grouping variable it can be inferred that there are significant and relevant differences in polarization between different groups based on the influencer selection.
5.2.6 Amount of Influencers followed: Further Observations
The only pair which almost showed a significant difference in polarization are the pair Small-None as can be seen in Table 11. This probably is because the users in these 2 groups may only differ slightly, seeing as following one or two influencers doesn’t differ a lot compared to following none at all. Especially if the influencers are only followed because of their fame factor.
Also, it must be noted that the group who follow no influencer at all have a larger mean and mean ranking for polarization compared to those who follow a small amount of influencers as can be seen in Table 6 and 4. This could be because choosing not to follow an influencer could be because of a strong political opinion, and does not necessarily mean you have no political opinion whatsoever. Furthermore, people who only follow the major influencers may only do so because they are famous.
Another interesting observation is that when looking at the descriptives in Table 4 is that both the mean and median polarization is the largest for the medium group. This makes sense because one would expect the polarization to become less when being exposed to a variety of left and right influencers, seeing as a user can only fall in the Large category if the user follows both left and right influencers.
5.3
Limitations and Future Research
Twitter is a interesting medium used for many different goals. For this research, users were selected who had shown that they were active on Twitter and replied or retweeted something which mentioned one or more influencers. This method of sampling was done to exclude bots and users who do not have any political opinion but are pure spectators, but it is important to question whether this method was appropriate and effective enough to sample a representative population of users who have a political
5.4 Conclusion 5 DISCUSSION
opinion on Twitter. One of the ideas to change this up for future research would be to look at tweets and analyze whether these contain a political hashtags or a retweet of one or more political influencers.
Another point is that after users expressed their political opinion on Twitter, it was concluded that the polarization and political profile could be inferred by analyz-ing that user’s follower network. Yet, the results have shown that some influencers seem to rise above their political ideology in popularity and have almost become su-perstars with tens of millions of followers in the process, resulting in an attainment of followers who do not follow them for their political views, but simply because they are symbols of success, fame and (in some cases) controversy. Perhaps political figures are more like celebrities than politicians on Twitter, especially seeing as the connection is almost between ’ordinary’ users and influencers. For future research, it would be an interesting idea to add a fame factor to the equation, a factor which could could take into account why users would choose to follow a specific influencer based on their fame instead of their political views.
Another element which follows up on this is the measure of polarization. For this research the entire focus lies on the amount and political affiliation of media outlets for determining the polarization level. However, the results indicate that there is a bias to the left side of the scale, probably due to the difference in amount of left and right media. Furthermore, the severity of the bias of various media are not taken into account, only the direction. For future research one could also take a look at combining different kinds of polarization (hashtag, other networks of users, etc.) and combine these to have a different and perhaps more general view of polarization on Twitter. Also, after comparing the polarization method to a non-absolute/real measuring method, it was revealed that users have a tendency to lean left when it comes to media selection. Therefore one could argue that this measuring method of polarization is not completely neutral, seeing as the media selection which was used for this research was unbalanced. A suggestion to correct this would be to take more right media into consideration as to compensate for the left bias and see whether this makes a difference. Adding to this it would probably be interesting to differentiate between different levels of bias for various media outlets, seeing as these are not all equal to each other.
Lastly one could try and adapt the method of measuring polarization proposed in this paper measuring political polarization for users from a different country. However, it would probably be difficult to find such a country where the politicians are as active on Twitter as they are in the US, not to mention who have an equally large and active following. To add to this, the political system would have to be similar to that of the US: it would have to consist of two large opposing parties instead of many small ones with minor differences.
5.4
Conclusion
This research has analyzed the difference in polarization between users based on which media outlets they followed on Twitter and which American influencers they chose to follow. After a variety of analyses, both looking at individual influencers, groups of influencers based on their political standing and the total amount of influencers users follow, it can be concluded that there are significant differences in polarization between different networks surrounding different American influencers on Twitter.
However, there were a couple of interesting results concerning the right/republican influencers having relatively low levels of polarization compared to the left/democrat influencers. There is a possibility that this could be written of to the popularity of Trump on Twitter, who attracts followers who do not follow him as an influencer for their political views but possibly because he generates a lot of attention as a result of creating controversy. Also the general mistrust and fear of fake news is a factor which may contribute to the more careful attitude of right leaning users. Another possible reason which was put forward was the unequal amounts of left and right media which were considered for this research, as well as not taking into account the difference in severity of bias between different media. This unequal distribution and levels of bias
5.4 Conclusion 5 DISCUSSION
could have skewed the polarization values and resulted in left leaning users appearing more polarized compared to right leaning users.
References References
References
Allcott, H., & Gentzkow, M. (2017). Social media and fake news in the 2016 election. Journal of Economic Perspectives, 31 (2), 211–36.
Burns, R. P., & Burns, R. (2008). Business research methods and statistics using spss. Sage.
Ceron, A., Curini, L., Iacus, S. M., & Porro, G. (2014). Every tweet counts? how sentiment analysis of social media can improve our knowledge of citizens’ polit-ical preferences with an application to italy and france. New Media & Society, 16 (2), 340–358.
Cohen, J. (1988). Statistical power analysis for the behavioral sciences 2nd edn. Erlbaum Associates, Hillsdale.
Conover, M., Ratkiewicz, J., Francisco, M. R., Gonçalves, B., Menczer, F., & Flam-mini, A. (2011). Political polarization on twitter. ICWSM , 133 , 89–96. Domke, D., Watts, M. D., Shah, D. V., & Fan, D. P. (1999). The politics of
conserva-tive elites and the ‘liberal media’argument. Journal of Communication, 49 (4), 35–58.
Francia, P. L. (2017). Free media and twitter in the 2016 presidential election: The unconventional campaign of donald trump. Social Science Computer Review , 0894439317730302.
Fritz, C. O., Morris, P. E., & Richler, J. J. (2012). Effect size estimates: current use, calculations, and interpretation. Journal of experimental psychology: General , 141 (1), 2.
Garimella, V. R. K., & Weber, I. (2017). A long-term analysis of polarization on twitter. arXiv preprint arXiv:1703.02769 .
Gilbert, E., Bergstrom, T., & Karahalios, K. (2009). Blogs are echo chambers: Blogs are echo chambers. In System sciences, 2009. hicss’09. 42nd hawaii international conference on (pp. 1–10).
Gottfried, J., Kiley, J., Matsa, K. E., & Mitchell, A. (2014). Political po-larization & media habits. Retrieved on the 23rd of March, 2018, from http://www.journalism.org/2014/10/21/political-polarization-media-habits/. Gurevitch, M., Coleman, S., & Blumler, J. G. (2009). Political communication—old
and new media relationships. The ANNALS of the American Academy of Po-litical and Social Science, 625 (1), 164–181.
Jones, D. A. (2004). Why americans don’t trust the media: A preliminary analysis. Harvard International Journal of Press/Politics, 9 (2), 60–75.
Lee, T.-T. (2010). Why they don’t trust the media: An examination of factors predicting trust. American Behavioral Scientist , 54 (1), 8–21.
Marchi, R. (2012). With facebook, blogs, and fake news, teens reject journalistic “objectivity”. Journal of Communication Inquiry, 36 (3), 246–262.
Maynard, D., & Greenwood, M. A. (2014). Who cares about sarcastic tweets? inves-tigating the impact of sarcasm on sentiment analysis. In Lrec (pp. 4238–4243). Messing, S., & Westwood, S. J. (2014). Selective exposure in the age of social
media: Endorsements trump partisan source affiliation when selecting news online. Communication Research, 41 (8), 1042–1063.
Micro-blogging. (n.d.). Retrieved on 7th of March, 2018, from http://en.wikipedia.org/wiki/Micro-blogging.
Mislove, A., Lehmann, S., Ahn, Y.-Y., Onnela, J.-P., & Rosenquist, J. N. (2011). Understanding the demographics of twitter users. ICWSM , 11 (5th), 25. Nordstokke, D. W., & Zumbo, B. D. (2010). A new nonparametric levene test for
equal variances. Psicológica, 31 (2).
Pang, B., Lee, L., et al. (2008). Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval , 2 (1–2), 1–135.R
Pariser, E. (2011). The filter bubble: What the internet is hiding from you. Penguin UK.
Sakaki, T., Okazaki, M., & Matsuo, Y. (2010). Earthquake shakes twitter users: real-time event detection by social sensors. In Proceedings of the 19th international conference on world wide web (pp. 851–860).
References References
Silverman, C., & Singer-Vine, J. (2016). Most americans who see fake news believe it, new survey says. BuzzFeed News.
Statista. (2016). Statista statistics: Share of online adults in the united states using twitter from 2009 to 2016. Retrieved on the 7th of March, 2018, from https://www.statista.com/statistics/186675/share-of-us-adults-using-twitter-since-2009/.
Statista. (2017). Statista statistics: Number of monthly active twitter users worldwide from 1st quarter 2010 to 4th quarter 2017 (in millions). Retrieved on the 7th of March, 2018, from https://www.statista.com/statistics/282087/number-of-monthly-active-twitter-users/.
Stroud, N. J. (2010). Polarization and partisan selective exposure. Journal of com-munication, 60 (3), 556–576.
Triola, M. F. (2008). Essentials of statistics. Pearson Addison Wesley Boston, MA, USA:.
Tumasjan, A., Sprenger, T. O., Sandner, P. G., & Welpe, I. M. (2010). Predicting elections with twitter: What 140 characters reveal about political sentiment. Icwsm, 10 (1), 178–185.
Varol, O., Ferrara, E., Davis, C. A., Menczer, F., & Flammini, A. (2017). Online human-bot interactions: Detection, estimation, and characterization. arXiv preprint arXiv:1703.03107 .
Wang, H., Can, D., Kazemzadeh, A., Bar, F., & Narayanan, S. (2012). A system for real-time twitter sentiment analysis of 2012 us presidential election cycle. In Proceedings of the acl 2012 system demonstrations (pp. 115–120).
Yardi, S., & Boyd, D. (2010). Dynamic debates: An analysis of group polarization over time on twitter. Bulletin of Science, Technology & Society , 30 (5), 316–327.
Polarization * Tim Kaine Polarization Polarization Polarization Tim Kaine Polarization
Mean N Std. Deviation Median
0 1 Total .3808178045 7633 .2717943289 .3333333333 .6190944646 1515 .2338708372 .6923076923 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization
Polarization * The Democrats Polarization
Polarization Polarization
The Democrats Mean N Std. Deviation Median 0 1 Total .3708468521 7194 .2698128528 .3333333333 .6022713211 1954 .2399340879 .6666666667 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization
Polarization * Barack Obama Polarization
Polarization Polarization
Barack Obama Mean N Std. Deviation Median 0 1 Total .3614248957 3758 .2537799771 .3333333333 .4613127379 5390 .2903508974 .5000000000 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization Polarization * GOP Polarization Polarization Polarization GOP Polarization
Mean N Std. Deviation Median
0 1 Total .4114856814 7497 .2868801847 .5000000000 .4602073059 1651 .2440035581 .5000000000 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization
Polarization * John McCain Polarization
Polarization Polarization
John McCain Mean N Std. Deviation Median 0 1 Total .4054197681 7732 .2816002383 .5000000000 .5014157967 1416 .2581320560 .5555555556 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization Page 1
A STATISTICS: DESCRIPTIVES INDIVIDUAL INFLUENCERS
Polarization * Mitt Romney Polarization
Polarization Polarization
Mitt Romney Mean N Std. Deviation Median 0 1 Total .4134781011 8167 .2828542720 .5000000000 .4768957839 981 .2505422586 .5000000000 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization
Polarization * Paul Ryan Polarization
Polarization Polarization
Paul Ryan
Polarization
Mean N Std. Deviation Median 0 1 Total .4069012709 7108 .2875045908 .5000000000 .4668902851 2040 .2478062266 .5000000000 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization
Polarization * Joe Biden Polarization
Polarization Polarization
Joe Biden
Polarization
Mean N Std. Deviation Median 0 1 Total .3645311800 6758 .2687803006 .3333333333 .5779115903 2390 .2504389707 .6666666667 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization
Polarization * Hillary Clinton Polarization
Polarization Polarization
Hillary Clinton Mean N Std. Deviation Median 0 1 Total .3480703868 4930 .2655179389 .3333333333 .5046760096 4218 .2733645227 .6000000000 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization
Polarization * Mike Pence Polarization
Polarization Polarization
Mike Pence
Polarization
Mean N Std. Deviation Median
0 1 Total .4108702762 6582 .2942699956 .5000000000 .4444124151 2566 .2389172737 .5000000000 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization Page 2
Polarization * Sarah Palin Polarization Polarization Polarization Sarah Palin Polarization
Mean N Std. Deviation Median
0 1 Total .4154278183 7896 .2862259476 .5000000000 .4508724937 1252 .2369096513 .5000000000 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization
Polarization * Donald Trump Polarization
Polarization Polarization
Donald Trump Mean N Std. Deviation Median 0 1 Total .4690021060 3465 .2936467923 .5000000000 .3905715499 5683 .2674515165 .4285714286 .4202787949 9148 .2802433176 .5000000000 Polarization Polarization Page 3
Tests of Normality
Infl_pol_value
Kolmogorov-Smirnova Shapiro-Wilk Statistic df Sig. Statistic df Sig.
Polarization Left Slightly Left Neutral Slightly Right Right No Influencer .193 2215 .000 .866 2215 .000 .157 1734 .000 .881 1734 .000 .194 907 .000 .895 907 .000 .097 1102 .000 .953 1102 .000 .177 2861 .000 .887 2861 .000 .279 329 .000 .880 329 .000
Lilliefors Significance Correction a.
Created Variablesa
Source Variable Function New Variable Label Polarizationb Rank RPolariz Rank of Polarization
Mean rank of tied values is used for ties. a.
Ranks are in ascending order. b. ANOVA polarization_dif_rank polarization_dif_rank polarization_dif_rank polarization_dif_rank Sum of
Squares df Mean Square F Sig.
Between Groups Within Groups Total 718041702.7 5 143608340.5 94.469 .000 1.390E+10 9142 1520167.666 1.462E+10 9147 polarization_dif_rank polarization_dif_rank Page 1
B STATISTICS: ASSUMPTIONS POLARIZATION VALUES
