Content uploaded by Matthew A. Shapiro
Author content
All content in this area was uploaded by Matthew A. Shapiro on Oct 26, 2015
Content may be subject to copyright.
1
Title: Going ‘Bald on Record’: Relationships Among Public Officials’ Social Media Behavior
and Language Use
Authors*:
Libby Hemphill
Illinois Institute of Technology
lhemphil@iit.edu
Jahna Otterbacher
Illinois Institute of Technology
jotterba@iit.edu
Matthew A. Shapiro (corresponding author)
Illinois Institute of Technology
mshapir2@iit.edu
* Authorship is equal and listed alphabetically.
Abstract:
Public officials use polarizing language – supporting language for one’s self versus pejorative
language for others – as a means of establishing clear boundaries on certain issues. This has been
explored to some degree in terms of how such language is conveyed in the traditional media, but
minimal research has been done with regard to the role of polarizing language within social
media. This paper explores how elected U.S. officials use potentially polarizing language
(“civility,” “politeness,” and related forms) to draw in supporters. We analyze the content and
behavior of more than 30,000 tweets from the available Twitter accounts of each elected member
of Congress, particularly in terms of the nature (size and party composition) of Twitter networks
for officials who use polarizing language. Network analysis via Network Workbench and
NodeXL confirms that officials’ use Twitter for much more than broadcasting, officials’
interaction networks differ from their follower/friends networks, and polarizing language cannot
be correlated with peripheral locations in a network. These indicate that Twitter plays a more
nuanced role in political communication than previously expected.
Paper presented at Korean Association for Public Administration and American Society for Public
Administration Joint International Conference: Seoul, Korea, October 28-29
2
Introduction
Public officials use polarizing language – supporting language for one’s self versus pejorative
language for others – as a means of establishing clear boundaries on certain issues. This has been
explored to some degree in terms of how such language is conveyed in the traditional media, but
minimal research has been done with regard to the role of polarizing language within social
media. This paper explores how elected U.S. officials use potentially polarizing language to draw
in supporters. We provide preliminary analysis of content from the available Twitter accounts of
each elected member of Congress, and assume that such content is predicated on Twitter
behavior. As such, we engage in an in-depth analysis of such behavior, particularly in terms of
the nature (size and party composition) of Twitter networks for officials.
By identifying Twitter accounts for members of Congress and high-ranking public officials and
using the Twitter Database Server (Green, 2011) and Twitter-collectors (Hemphill, 2011), we
gathered 29,694 tweets posted by 411 elected members of Congress between June 14, 2011 and
August 23, 2011. We also collected tweets in which officials were explicitly mentioned by other
users who were not in our pool of public officials, and those included another 550,000+ tweets.
Twitter usage generates networks when users establish “follow,” “reply,” and “mention”
relationships with one another. Network analysis via Network Workbench (NWB Team, 2006)
and NodeXL (Smith et al., 2010) enables us to analyze those networks to reveal insights into a
group’s dynamics. For instance, we can compare the networks of elected officials to determine
whether they reach the public through Twitter or whether they establish a virtual “echo chamber”
in which they only reach themselves.
In order to identify these qualities and, as mentioned above, identify their relationship to
polarizing language, we first present a review of the relevant literature in the section immediately
following. After that, we introduce the methods of capturing the data, calculating the relevant
variables, and analyzing such data. The results are then presented, followed by our conclusions.
Literature & Hypotheses
Studies of political communication often focus on the language officials use in traditional media
(Cook et al. 1983; Edwards III and Wood 1999; Entman 2007; Graber 2000; Lee 2009), but
3
minimal research has examined language use within social media. Rarer still are studies that
examine relationships between politicians’ communication networks and political outcomes.
Our project builds on earlier research by expanding both the scale and scope of inquiry. Our
dataset (500,000+ tweets) is far larger than earlier datasets of either Congressional Twitter posts
(Golbeck, Grimes, & Rogers, 2010) or politicians’ web pages (Xenos and Foot 2005), enabling
us to address questions of the consistency and generalizability of results raised earlier. Our
dataset also includes not only elected officials but presidential appointees and the public who
tweet about both elected officials, providing us a broader scope of inquiry and allow us to
examine the interactions between officials and the public more broadly. Given the relational
nature of politics (Lazer 2011), Twitter-based networks and the interactions that occur within
them provide an interesting natural experiment.
Previous research on Twitter has focused either on the public (Golbeck & Hansen, n d; Parsing
Election Day Media - How the Midterms Message Varied by Platform, 2010) or elected officials
(Golbeck et al., 2010) but was not able to address interactions between the two. In terms of social
media, analysis of elected officials has been limited to how they address redistributive goals or
how traditional media affects them (Jennifer Golbeck, Grimes, and Rogers 2010; Xenos and Foot
2005). In sum, ours is the first examination of this scale and scope of the relationships between
social media use and political behavior among elected officials.
Our project also provides a much-needed revision to existing measures of polarization in politics.
Polarizing behavior of public officials has traditionally been identified through congressional
voting records. The obvious drawback to this measure is that it limits analyses of public officials
to only elected officials, eliminating opportunities to compare elected officials and to include
social media effects into research on public officials. On a more fundamental level, though, these
data fail to capture the underlying signals and heresthetics which now inform us of political
ideology. We will account for such signals by using Twitter content to develop a content coding
scheme and construct polarization measures which will be useful for analyzing other political
communication.
The following examples, collected on June 20, 2011 and tweeted by Speaker of the House John
Boehner and House Minority Leader Nancy Pelosi, respectively, are illustrative of officials’ use
of Twitter:
4
SpeakerBoehner “Dems admit to being nowhere on jobs http://bit.ly/jmCU52
GOP has a plan #4jobs http://jobs.gop.gov”
NancyPelosi “Today I thanked #SF @TrevorProject volunteers for helping
#LGBT youth see their future's bright!”
Both tweeters are doing more than conveying information to others. Boehner very clearly draws
a line between himself and his supporters and the Democrats. The content of the message is
essentially, “Support us. We have a plan and they don’t.” However, his audience (i.e., followers)
might not respond well to such a blatant command, and Boehner instead couches his attack in a
statement rather than in an instruction. Boehner dissociates himself from the statement by writing
in a relatively impersonal manner (e.g., avoiding first person pronouns).
Similarly, Pelosi’s message is also clear; she assures her audience that she is supportive of the
LGBT community. More generally, the statement she makes is “I support good causes that you
believe in.” By giving a specific example of thanking Trevor Project volunteers, she fosters
social relations with followers and creates a sense of common ground (i.e., the sense that “we
both believe in this cause”). Pelosi also creates an explicit, measurable connection between
herself and the Trevor Project through her use of the @reply Twitter convention.
These examples illustrate that how public officials say what they want to say on Twitter may be
as important for understanding their communication as what they say and to whom. Our study
provides a much-needed link between officials’ language behaviors within a specific medium
(i.e., Twitter), their social affiliations (i.e., networks), and their political behaviors (i.e., voting).
We predict significant relationships between the tone officials use, the networks that result from
their communication, and their voting behavior.
Political Communication and Social Network Analysis
Using social media as a means of communicating to the larger public effectively replaces
communication that was only possible through traditional media outlets (Cook et al. 1983;
Edwards III and Wood 1999; Entman 2007; Graber 2000; Lee 2009) or, more recently, websites
and blogs that reported statements and speeches of public officials (e.g., Gentzkow and Shapiro,
2010). We emphasize that Twitter allows public official to avoid the filters of traditional media
and communicate directly to their followers. This can exacerbate the negative effects of the
incomplete information held by voters, which already occurs via traditional media outlets 2010.
Thus, we expect that public officials’ frequent use of “civility,” “politeness,” and related forms
5
of polarizing language are likely to promote misconceptions about specific political or policy
issues.
Analyzing tweets allows us to test the most recent developments of political communication
theory, particularly the effects of micro-blogging efforts on party and social group formation.
Existing studies have attempted to predict political candidate success in elections (Baum and
Groeling 2008) or accurately portray public sentiment about candidates (Wang, Hanna, and
Sayre 2011). We adopt the key explanatory variables from such research – network size and
strength of ties (Tumasjan, Sprenger, Sandner, and Welpe 2010) – and apply it to the unique case
of public officials.
Much research on this topic is overly descriptive in its presentation of Twitter adoption rates by
followers of member of Congress (Siegel 2011). In one case, however, Twitter followers have
been found to aggregate into politically homogeneous or homophilous groups (e.g., Bode &
Dalrymple, 2011, Chi & Yang, 2011) .Given these findings, we predict that public officials will
create homophilous communication networks via Twitter and produce echo chambers in which
they speak primarily to one another. Acknowledging the propensity for the public to have
preconceived views about the source of information (Boutyline and Willer 2011; Himelboim,
McCreery, and Smith n.d.), we also consider whether officials’ networks grow in size and/or
strength of support with civil language.
Our approach also provides a methodological innovation: existing research relies on adoption
rates and followers, but such measures have been superseded by more appropriate measures
(e.g., “mentions,” “replies,” and TwitterRank) to measure network characteristics (Bakshy,
Hofman, Watts, & Mason, 2011, Cha, Haddadi, Benevenuto, & Gummadi, 2010, Weng, Lim,
Jiang, Search, & Information, 2010).
Linguistic Framework
To examine how public officials use polarizing language in social media and how this then
impacts their social networks, we must first specify what we mean by “polarizing” language.
Although some previous research has examined the concepts of “civility” and “politeness” in
online political discussions (e.g., McClain, 2009, Ng & Detenber, 2006, Papacharissi, 2004),
none puts forward a framework that can be readily applied to our work. A key challenge of our
work is to analyze tweets, which are extremely brief. In linguistic terms, we can describe a tweet
6
as an individual speech act (i.e., a single utterance in which the speaker asserts or requests
something) rather than a discussion. Therefore, we require a much more fine-grained framework
to analyze tweets.
We will base our framework on a well-established sociolinguistic theory of “politeness” (Brown
and Levinson 1987). Briefly, the theory proposes how a speaker (or a tweeter) goes about
performing a speech act in a way that reduces his or her risk of “losing face.” There are two
types of “face” that all adult members of society have according to the theory: one’s negative
face represents the desire to act freely, without imposition from others, while one’s positive face
represents the desire to be liked and accepted by others.
Many speech acts pose threats to the speaker’s and/or the hearer’s two faces; therefore, the
speaker may use one or more strategies to minimize threats. Table 1 provides an example of the
use of three categories of “politeness strategies” to perform a hypothetical speech act: an official
encouraging her listeners (i.e., Twitter followers) to vote. This speech act threatens the hearers’
negative face, since they would like to act freely, without being told what to do by the official.
The act also threatens the official’s positive face, since she wants followers to have a positive
opinion of her, despite the imposition.
[Table 1 here]
We leverage Brown and Levinson’s framework because it has been widely used (e.g., Jansen &
Janssen, 2010; Mboudjeke, 2010), explains language use strategies’ effects on social
relationships, and can be amended to our needs while remaining accessible to our coders and
readers. Our goal is to use the framework to develop quantitative measures of civility at both the
level of the tweet and of the tweeter. This framework will characterize the language devices that
officials use in interacting with others through their tweets, and will allow us to eventually
quantify the extent to which a given tweet (or a given individual) communicates in a civil
manner or uses polarizing language. We predict that civil language is a good predictor of
political centrism, and we will include this measure in later regressions to test the predictive
power of our civility scale.
7
Hypotheses
Based on related literature and our preliminary findings, we propose the following hypotheses
about the language officials use, the networks that result from their conversations, and the impact
of those networks on the development of social structures:
H1. Twitter is a virtual echo chamber in which officials interact mainly with themselves.
H2. Officials use Twitter primarily as a broadcast mechanism.
H3. Officials’ interaction network differs from their follower/friends network.
H4. Officials’ interaction networks exhibit higher order social structure.
H5. “Bald on record” correlates highly with peripheral location in a network.
Method
We identified Twitter accounts for members of Congress, based on listings at Congress.org.1 We
hired workers through Mechanical Turk (MTurk) to collect and recheck Twitter screen names for
all members of Congress and were able to verify 411 accounts. Mechanical Turk is a
marketplace in which requesters hire workers to complete small, self-contained tasks that require
human intelligence. We paid workers on MTurk a set fee ($0.06) for each Twitter screen name
they could find. We hired two or more workers to look up each official and then compared their
responses. In cases where the workers disagreed about the screen name, we checked the official
by hand on their websites and on Twitter. MTurk is increasingly being leveraged by researchers
to collect and evaluate the quality of social science data (e.g., (Bakshy, Hofman, Watts, and
Mason 2011; Cha, Haddadi, Benevenuto, and Gummadi 2010; Weng, Lim, Jiang, Search, et al.
2010).
Using the Twitter Database Server (Green 2011) and Twitter-collectors (Hemphill 2011), we
gathered 29,684 tweets posted by 411 elected Congressmen between June 14, 2011 and August
23, 2011. We also collected tweets in which officials were explicitly mentioned by other users
1 http://www.congress.org/congressorg/directory/congdir.tt
8
who were not in our pool of public officials, and those included another 550,000+ tweets. In this
way, we have data for the entire population of national-level public elected officials.
Twitter usage generates networks when users establish “follow,” “reply,” and “mention”
relationships with one another. Network analysis via UCINet (Borgatti, Everett, and Freeman
2002) and NodeXL (Smith et al. 2010) enables us to analyze those networks to reveal insights
into a group’s dynamics. For instance, we can compare the networks of elected officials to
determine whether they reach the public through Twitter or whether they establish a virtual
“echo chamber” in which they reach only themselves. Additionally, following someone on
Twitter allows a user to view that person’s tweets in a timeline and to be updated every time the
followed user tweets. Following enables a passive, peripheral awareness of another users’
contributions. Mentioning, on the other hand, is an active, deliberate communicative act in which
one user directs a comment to another or explicitly references another user in his own tweet. We
construct below both “follows” and “mentions” networks for the officials we studied.
To respond to virtually any of the above hypotheses and calculate a network measure, we are
faced with several methods (Borgatti, et al., 2006),2 but we are partial towards betweenness,
which is determined by first calculating the shortest path between all the pairs of vertices, and
then by summing the fraction of shortest paths between all pairs that go through the vertex in
question. We normalize this betweenness measure in order to compare between two networks.3
This is especially important as it is those nodes that fall between different clusters of individuals
that provides a unique understanding of political communication and behavior. This is because
they lie on the shortest path between the less-connected nodes of the less-connected individuals
of the established clusters and are generally members of more social groups than those with low
betweenness. In addition, they often occupy structural holes or places in the graph where if that
person were removed, the graph would no longer be connected. High betweenness individuals
often have a lot of influence because of the diversity of their connections: they have access to all
the social groups of which they are a member, and their messages/connections experience less
decay because they do not have to travel as far.
2 The most common are degree, betweenness, plainness, and eigenvector. Betweenness is often used as a measure of
influence within a network (Davis, Yoo, and Baker 2003; Newman 2005).
3 The normalization process is determined by the following equation: (n-1)/centrality*100, where n is the number of
individuals.
9
Our test of H3 is based on our collection of each Congressman’s number of friends, followers,
and tweets for the period under analysis. We predict betweenness based on these variables in
addition to a second set of variables – legislative branch membership, gender, and party
affiliation – all of which shows whether or not a Congressman is a good conduit for information,
opinion, or other content flowing over the network. If the network is relatively close, we suspect
that these conduits represent key voting positions, as they are people who can be targeted if your
intention is to have your message passed along to a group that could not be otherwise reached.
With regard to H5, our method of identifying “bald on record” – our proxy for polarizing
language among other linguistic conventions – is constantly being updated based on new insights
in what is essentially an exploratory analysis of Congressional tweets. Presented with complete
details in the Appendix, “baldness” is but a component of one of three categories which we
identify as an important component of Congressional tweets. These three categories – action of
tweet, content of tweet, and manner of tweet – have been determined through a lengthy process
of identifying specific content in tweets.4 To test H5 and see how this the speaker of such
language correlates with his/her peripheral location in the network, we first establish that the
network exhibits core-periphery qualities and then look at correlation patterns.
Results
Our first observation is that the Congress mention network does not fit into a core-periphery
structure.5 This effectively rules out any possibility of testing H5 and indicates that we must
rework our understanding of how Congressmen align themselves within the network and among
each other. Another important observation is the transitivity of the Congress mention network,
which is a measure of the triad consensus in the graph, i.e., how often two of an individual’s
connections are connected to each other.6
4 Contact corresponding author for specific details.
5 This is based on a correlation coefficient of 0.138, which represents the correlation between our matrix and the
hypothetical ideal pattern matrix were it a core-periphery matrix. It should also be noted that fifteen members of
congress who tweet are neither mentioned nor do they mention.
6 The Congress mention network has higher transitivity (0.229) than a random network of the same size and density
(0.012). Transitivity between 0.3 and 0.6 is "normal" for graphs though, so this one can still be considered low.
10
Third, the diameter of the graph – the degree of separation between members of Congress – is
six, which is surprising in such a closed network. Given these last two points, in addition to
challenges of identifying particularly important conduits in the Congress mention network in line
with H3 as well as the correlation between polarizing language (“baldness”) and an individual’s
peripheral location in the network, we might conclude that members of Congress are not using
Twitter to explicitly position themselves in terms of others, at least not through Twitter. This
provides additional evidence for existing research showing candidates more likely to provide
basic issue-related information online while avoid most other forms of issue dialogue (Xenos and
Foot 2005).
Figures 1 and 2 show respectively the networks resulting from officials mentioning and officials
following one another. In both figures, blue solid squares represent Democrats, and red hollow
circles represent Republicans. The lone Independent Senator is a yellow solid disc and is visible
only in the mentions network (Figure 1), but there is no real role of third parties, confirming
Xenos and Foot (2005). The darkness of the lines connecting nodes depends on a measure of the
strength of that relationship. In the mentions network, the darkness of a line is determined by the
number of times two individuals mention one another. Darker lines indicate more frequent
mentioning. In the follows network, gray lines indicate one-way connections (i.e., one official
follows another who does not follow him) while black lines represent reciprocal relationships
(i.e., both officials follow each other). Edge opacity here indicates the number of mentions, and
the dark self-loops on some nodes indicate that officials frequently mention themselves, partially
confirming H1.
[Insert Figure 1 here]
[Insert Figure 2 here]
A division between parties is visible in both graphs. In the mentions graph (Figure 1), we can see
that lines connecting Republicans to each other are darker, indicating Republicans mention one
another more often than Democrats do. Earlier research that explored similar mentioning
behaviors among political bloggers (using links between blogs to indicate connections) found a
similar pattern – conservative bloggers also linked to each other more often (Adamic and Glance
Transitivity is the measure used to look for in small worlds, and the Congress mention network does not seem to
have it (Faust, 2006; Newman, Strogratz, Watts, 2001).
11
2005). The division between parties in the follows graph (Figure 2) is starker. There, we see
clear clusters of Democrats and Republicans with fewer links between them. We also see isolates
(nodes that do not connect to any others) and two large, disconnected components (subgraphs
that are connected to each other but not to the rest of the graph). The smaller component on the
left of Figure 2 shows a less clear division between parties.
Differences in the strength of relationships are visible in both graphs as well. In the mentions
graph, the lines connecting Republicans are darker, indicating that they mentioning each other
more often. In the follows network, the large component has more black lines, indicating
reciprocal relationships, than does the smaller component. Taken in tandem with our earlier
evidence, a lack of transitivity suggests a lack of hierarchy in this network. With no evidence of
hierarchy and no evidence of a core-periphery structure, the network does not display any
immediately apparent structure.
The follows network is nearly ten times as dense as the mentions network (0.134 vs 0.014,
respectively), indicating that officials are passively connected to far more people than the
number of people there are actively connected to. This pattern is not surprising – it makes sense
that we can only engage with some subset of the people we know or are aware of. What is
interesting is that in both cases, the density of the graph is surprisingly low. In a network of
actors who are so similar and who, in theory, work together, we would expect to see a much
higher density. Instead, we see that, at a maximum, only thirteen percent of the possible
connections are made, indicating that even though officials clearly use Twitter, they underutilize
the ability to passively monitor one another’s behaviors.
These graphs cannot tell us why officials choose to follow or mention such a small subset of
their peers. Other research on Twitter during Congressional campaigns (Livne 2010) suggests
that the density of the network correlates to the cohesion of the network’s message meaning that
we would expect more cohesion among Republicans than among Democratics, enabling
Republicans to present a more united front. This could be a function of the campaign process,
though, which was not ongoing at the time of our data collection.
Empirical Generalizations
As was stated above, people with high betweenness are generally members of more social groups
than those with low betweenness, occupying structural holes and playing potential key roles in
12
the political process. In an attempt to further understand the impact of a member of Congress
having higher numbers of followers and friends as well as tweeting with greater frequency, we
analyze the relationship between betweenness and Congressmen characteristics. Specifically, we
are interested in the relationship between networks measured by normalized betweenness and
each Congress member’s number of followers, number of friends, and number of tweets in the
data collection period. These relationships can also be predicted by party affiliation, branch of
Congress, and gender.
Without controlling for differences in branch, gender, and party, shown in Table 2 (1), there are
no statistically significant effects from a Congressman having additional numbers of followers,
friends, or tweets. The same is true when we account for differences in branch, gender, and
party, shown in Table 2 (3), although there are indications in Table 2 (2 and 3) that branch
significantly predicts betweenness. To understand such differences further and in the context of
the number of a Congressman’s followers, friends, and tweets, interactions between each with
branch, gender, and party are introduced in stages in an attempt to eliminate confounding
variables.
In Table 2 (4), where branch is coded “1” for members of the House and zero for members of the
Senate, it is shown that a large number of followers increase betweenness for House members,
while a large number of followers decreases betweenness for Senate members. Albeit not at a
statistically significant level, there are also indications that a large number of friends decrease
betweenness for Senate members. This model of exclusivity had not previously been
documented for social media-related political communication.
There are also indications that the interaction between followers and gender and between friends
and gender significantly predicts betweenness, shown in Table 2 (5). Where gender is coded “1”
for females and zero for males, it is shown that a large number of followers increase betweenness
for females relative to males, but that a large number of friends decrease betweenness for
females relative to males. Finally, where party affiliation is coded “1” for Democrats and zero
for Republicans, it is shown in Table 2 (6) that the only significant difference between parties
occurs with regard to the number of friends: as the number of friends increase, betweenness for
13
Democrats decreases relative to Republicans. This is yet another indication of the tighter network
effects and exclusivity among Republicans.7
[Insert Table 2 here]
Qualitative/Exploratory Analysis
Our challenges in testing H5 given the lack of a formal core-periphery network structure are
compounded by our difficulty in reliably coding tweet content. Based on Congressional tweets
who are at the extremes (i.e., the ten largest and ten smallest) of friends, followers, and tweeting
frequency, we took a stratified sample of each of these groups (n = 934) and again drew a
random sample of one hundred tweets for coding along the lines of the Appendix.
Shown in Table 3 in cross-tabulation form, 44 of 100 tweets were coded as bald on record,8
meaning that Congressmen avoid hedging when making statements via Twitter. Of course, this
could reflect the abbreviated nature of tweet-based messages, but this is nowhere near a
significant amount of total tweets (in our random sample from the stratified sample). The content
of bald tweets is predominantly policy and public relations-related, although a number were also
news-based. In terms of the action of bald tweets, nearly half were efforts by the public official
to position him/herself while about another 25 percent were to provide information of some sort.
[Table 3 here]
For the time being, we reserve comment on the remaining manner of tweets coded; i.e., “other”
and “positive” (no “negative” manner tweets were coded). That is, there is nothing to say except
that manner as “other” is predominantly “other” (for content) and to provide information (for
action). Compared with bald tweets, bald tweets tend to be focused on self-promotion, policy
orientation, and/or positioning of oneself relative to other Congressmen. Without a cohesive
core-periphery structure in the Twitter network, however, little can be said as to the effects of
such potentially polarizing tweets.
7 A number of the aforementioned findings are robust to the inclusion of all interaction effects, presented in Table 1
(7).
8 These codes were assigned by a single individual without any reliability checks. Efforts to improve coding
techniques are ongoing.
14
Conclusion
Our observations above, particularly the lack of closeness in Congressional networks and the
absence of a core-periphery structure are surprising. It is also an indication that we must be
innovative in understanding and properly identifying the function of social media in political
communication. Such methods of communication are an increasingly common way for elected
officials to convey their views. It is just not (yet) common for there to be coordination among
members of Congress, although this does not preclude coordination within each party. Such
analysis has been identified only in partial form here, and we acknowledge that future study is
needed to fully address differences in party affiliation.
We also acknowledge that future study must deal with the disconnected components in both
Figures 1 and 2. Exactly who is not connecting, and why? What is preventing Democrats and
Republicans from talking? And is there a pattern to the reciprocal relationships; e.g., are
Democrats following Republicans but not getting followed in return? The answers to these
questions are expected to be answered with a much more comprehensive look at “bald on
record” tweeters. Coding for such tweets is crucial in assessing the degree to which compromise
and language are correlated. Our preliminary, exploratory analysis is a fine starting point,
showing that bald tweets have much to do with self-promotion and position making, but we must
properly compare uncompromising behavior, e.g., Congressional voting patterns, with
uncompromising language. This will also help set up what are expected to be networks of
“baldness” and polarizing behavior which have heretofore been identified only through inference
or anecdote.
15
References
Adamic, L.A., and Natalie Glance. 2005. “The political blogosphere and the 2004 US election:
divided they blog.” In Proceedings of the 3rd international workshop on Link discovery,
New York, NY, USA: ACM, p. 36–43. http://portal.acm.org/citation.cfm?id=1134277
(Accessed June 22, 2011).
Alesina, Alberto, and Guido Tabellini. 2007. “Bureaucrats or Politicians? Part I: A Single Policy
Task.” American Economic Review 97(1): 169-179.
http://pubs.aeaweb.org/doi/abs/10.1257/aer.97.1.169.
———. 2008. “Bureaucrats or politicians? Part II: Multiple policy tasks☆.” Journal of Public
Economics 92(3-4): 426-447.
http://linkinghub.elsevier.com/retrieve/pii/S0047272707000850 (Accessed August 26,
2011).
Baek, Kang Hui et al. 2011. “Love it or leave it? The relationship between polarization and
credibility of traditional and partisan media.” New York (April): 1-31.
Bakshy, Eytan, Jake M Hofman, Duncan J Watts, and Winter A Mason. 2011. “Everyone’s an
Influencer: Quantifying Influence on Twitter Categories and Subject Descriptors.” In
WSDM’11, , p. 65-74.
Baum, Matthew, and Tim Groeling. 2008. “New Media and the Polarization of American
Political Discourse.” Political Communication 25(4): 345-365.
http://www.informaworld.com/openurl?genre=article&doi=10.1080/10584600802426965&
magic=crossref||D404A21C5BB053405B1A640AFFD44AE3 (Accessed August 12, 2010).
Besley, Timothy, and Stephen Coate. 2003. “Elected Versus Appointed Regulators: Theory and
Evidence.” Journal of the European Economic Association 1(5): 1176-1206.
http://doi.wiley.com/10.1162/154247603770383424.
Bode, Leticia, and Kajsa E Dalrymple. 2011. “Profile of a Political Tweeter: Political
Candidates.”
Borgatti, Stephen P, M G Everett, and L C Freeman. 2002. “Ucinet for Windows: Software for
Social Network Analysis.” Harvard Analytic Technologies 2006.
http://www.analytictech.com/downloaduc6.htm.
Boutyline, Andrei G., and Rob Willer. 2011. “The Social Structure of Political Echo Chambers:
Ideology Leads to Asymmetries in Online Political Communication Networks.”
Brown, Penelope, and Stephen C. Levinson. 1987. 345 Politeness: some universals in language
usage. Cambridge University Press.
16
http://books.google.com/books?hl=en&lr=&id=OG7W8yA2XjcC&pgis=1 (Accessed
August 31, 2011).
Cha, Meeyoung, H. Haddadi, Fabricio Benevenuto, and K.P. Gummadi. 2010. “Measuring user
influence in twitter: The million follower fallacy.” In 4th International AAAI Conference on
Weblogs and Social Media (ICWSM), , p. 10-17.
http://www.aaai.org/ocs/index.php/ICWSM/ICWSM10/paper/download/1538/1826
(Accessed June 22, 2011).
Chi, Feng, and Nathan Yang. 2011. “Twitter Adoption in Congress.” Review of Network
Economics 10(1). http://www.bepress.com/rne/vol10/iss1/3 (Accessed April 13, 2011).
Cook, Fay Lomax et al. 1983. “Media and Agenda Setting: Effects of the Public, Interest Group
leaders, Policy Makers, and Policy.” Public Opinion Quarterly 47: 16-35.
Davis, Gerald F., Mina Yoo, and Wayne E. Baker. 2003. “The Small World of the American
Corporate Elite, 1982-2001.” Strategic Organization 1(3): 301-326.
http://soq.sagepub.com/cgi/doi/10.1177/14761270030013002 (Accessed July 26, 2011).
Edwards III, George C, and B Dan Wood. 1999. “Who Influences Whom? The President,
Congress, and the Media.” American Political Science Review 93(2): 327-344.
Entman, Robert M. 2007. “Framing Bias: Media in the Distribution of Power.” Journal of
Communication 57(1): 163-173.
Gentzkow, Matthew, and J.M. Shapiro. 2010. “Ideological segregation online and offline.”
http://www.nber.org/papers/w15916 (Accessed August 31, 2011).
Golbeck, Jen, and Derek L Hansen. “Computing Political Preference among Twitter Followers.”
Golbeck, Jennifer, Justin M. Grimes, and Anthony Rogers. 2010. “Twitter use by the U.S.
Congress.” Journal of the American Society for Information Science and Technology: n/a-
n/a. http://doi.wiley.com/10.1002/asi.21344 (Accessed May 19, 2011).
Graber, Doris A. 2000. How Members of Congress Use the Media to Influence Public Policy:
Media Power in Politics. Washington, D.C.: CQ Press.
Green, A. 2011. “Twitter Database Server.” http://140dev.com/free-twitter-api-source-code-
library/twitter-database-server/.
Hemphill, L. 2011. “Twitter-collectors.” https://github.com/casmlab/Twitter-collectors.
Himelboim, I., S. McCreery, and M. Smith. “Birds of a feather tweet together: Integrating
network and content analyses to examine cross-ideology exposure on Twitter.” Journal of
Computer-Mediated Communication.
17
Jansen, Frank, and Daniel Janssen. 2010. “Effects of positive politeness strategies in business
letters.” Journal of Pragmatics 42: 2531-2548.
Larcinese, Valentino, Riccardo Puglisi, and James M. 2009. “Partisan bias in economic news:
evidence on the agenda-setting behavior of U . S . newspapers Working paper.” Public
Choice.
Lazer, David. 2011. “Networks in Political Science: Back to the Future.” PS: Political Science &
Politics 44(01): 61-68. http://www.journals.cambridge.org/abstract_S1049096510001873
(Accessed June 16, 2011).
Lee, Han Soo. 2009. “News Media Conditions Presidential Responsiveness to the Public.” In
APSA Annual Conference, Toronto.
Mboudjeke, Jean-Guy. 2010. “Linguistic politeness in job applications in Cameroon.” Journal of
Pragmatics 42(9): 2519-2530.
http://www.sciencedirect.com/science/article/pii/S0378216610000597.
Newman, M E J. 2005. “A measure of betweenness centrality based on random walks.” Social
Networks 27(1): 1-15.
Parsing Election Day Media - How the Midterms Message Varied by Platform. 2010.
http://www.journalism.org/node/22791 (Accessed August 31, 2011).
Siegel, David a. 2011. “Social Networks in Comparative Perspective.” PS: Political Science &
Politics 44(01): 51-54. http://www.journals.cambridge.org/abstract_S104909651000185X
(Accessed June 16, 2011).
Smith, M. et al. 2010. “NodeXL: a free and open network overview, discovery and exploration
add-in for Excel 2007/2010.”
Tumasjan, a., T. O. Sprenger, P. G. Sandner, and I. M. Welpe. 2010. Social Science Computer
Review 1-37 Election Forecasts With Twitter: How 140 Characters Reflect the Political
Landscape. http://ssc.sagepub.com/cgi/doi/10.1177/0894439310386557 (Accessed July 12,
2011).
Wang, Bryan Ming, Alexander Hanna, and Ben Sayre. 2011. “Who Is Following Me?” Analysis.
Weng, Jianshu, Ee-peng Lim, Jing Jiang, H Information Search, et al. 2010. “TwitterRank:
Finding Topic-sensitive Influential Twitterers.” New York.
Xenos, Michael A., and Kirsten A. Foot. 2005. “Politics As Usual, or Politics Unusual? Position
Taking and Dialogue on Campaign Websites in the 2002 U.S. Elections.” Journal of
Communication 55(1): 169-185. http://doi.wiley.com/10.1111/j.1460-2466.2005.tb02665.x
(Accessed May 25, 2011).
18
19
Figures and Tables
Figure 1 Congressional mentions network
20
Figure 2 Congressional follows network
21
Table 1 Politeness strategies and examples
Strategy
Explanation
Example
None (i.e., “Bald on record”)
Most efficient recourse;
nothing is done to minimize
the threat
Go out and vote today!
Positive politeness
Attempts to minimize the loss
of speaker’s positive face;
promotes solidarity between
speaker and hearer
My friends, let’s all go out and
vote today!
Negative politeness
Minimizes the imposition the
act places on the hearer by the
speaker
If you can manage the time
today, please get out and vote!
22
Table 2 Predicting betweenness
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Normalized
betweenness
Normalized
betweenness
Normalized
betweenness
Normalized
betweenness
Normalized
betweenness
Normalized
betweenness
Normalized
betweenness
Followers
6.47e-09
(1.33e-08)
1.38e-09
(9.71e-09)
-1.93e-09***
(6.24e-10)
8.22e-09
(1.38e-08)
8.41e-09
(1.38e-08)
-2.72e-09**
(1.10e-09)
Friends
6.31e-07
(6.11e-07)
1.49e-08
(2.69e-07)
-1.53e-07***
(5.39e-08)
7.68e-07
(7.10e-07)
6.84e-07
(6.61e-07)
-1.85e-07***
(6.96e-08)
Tweets
0.0000156
(0.0000124)
5.91e-06
(3.90e-06)
-3.99e-06
(5.55e-06)
0.0000157
(0.000013)
0.0000183
(0.0000162)
5.12e-06
(0.0000127)
Branch
0.0027975***
(0.0009132)
0.0038813**
(0.0016898)
-0.000926
(0.0014444)
0.0041992**
(0.0020316)
0.0045672**
(0.0020705)
-0.0023189
(0.002125)
Gender
0.0004664
(0.0019)
0.0007737
(0.0019305)
-0.0003093
(0.0020894)
0.0028816
(0.0030014)
0.0003881
(0.0019567)
0.0035576
(0.0030443)
Party
-0.0019507
(0.0013123)
-0.0004532
(0.0012663)
-0.0019099
(0.0014062)
-0.0011189
(0.0012086)
0.0006461
(0.0018867)
0.0005668
(0.0016324)
Branch*
Followers
5.04e-07**
(2.36e-07)
5.97e-07***
(2.25e-07)
Branch*
Friends
-5.02e-07
(5.62e-07)
-4.48e-07
(6.89e-07)
Branch*
Tweets
0.0000131
(0.0000106)
9.80e-06
(0.0000138)
Gender*
Followers
1.56e-07*
(9.38e-08)
-2.81e-08
(1.39e-07)
Gender*
Friends
-1.97e-06**
(8.21e-07)
-1.78e-06***
(6.87e-07)
Gender*
Tweets
-0.0000223
(0.0000192)
-0.000022
(0.0000191)
Party*
Followers
9.69e-08
(9.16e-08)
-2.77e-07**
(1.15e-07)
Party*
Friends
-1.47e-06*
(8.39e-07)
-2.99e-07
(5.79e-07)
Party*
Tweets
-0.0000176
(0.0000158)
-9.83e-06
(0.0000135)
N
374
374
374
374
374
374
374
F-stat
0.64
3.16**
1.10
-
1.59
0.88
4.70***
R2
0.0392
0.0075
0.1060
0.2447
0.0600
0.0527
0.2967
23
Table 3 Cross-tabulation of tweet action, content, and manner
Content&
&
Bald&
44&
!"#$%&#'()*+,-.(/*+0&1')&'*2
32
0'4*2
52
6)-'72
82
9"%&1:2
;;2
9<2
;52
=>'*)&"(*2
;2
,7&?&.2
;2
Other&
37&
!"#$%&#'()*+,-.(/*+0&1')&'*2
82
@?'()*2
32
0'4*2
52
6)-'72
;A2
9"%&1:2
32
9<2
B2
,7&?&.2
32
Positive&
19&
!"#$%&#'()*+,-.(/*+0&1')&'*2
C2
@?'()*2
A2
6)-'72
D2
9"%&1:2
32
9<2
;2
Total&
100&
Action&
&
Bald&
44&
E&7'1)21"##>(&1.)&"(2
A2
0.77.)&?'2
32
6)-'72
32
9"*&)&"(&(F2
3;2
97"?&G&(F2&(H"7#.)&"(2
;C2
Other&
37&
E&7'1)21"##>(&1.)&"(2
82
0.77.)&?'2
52
6)-'72
B2
9"*&)&"(&(F2
A2
97"?&G&(F2&(H"7#.)&"(2
;C2
<'I>'*)&(F2.1)&"(2
;2
Positive&
19&
E&7'1)21"##>(&1.)&"(2
;2
0.77.)&?'2
82
6)-'72
82
9"*&)&"(&(F2
32
97"?&G&(F2&(H"7#.)&"(2
C2
<'I>'*)&(F2.1)&"(2
C2
Total&
100&
24
New Table
(1)
Ordered probit
(2)
Ordered probit
(3)
Ordered probit
(4)
Ordered probit
(5)
Ordered probit
(6)
Ordered probit
(7)
OLS: DV in logs
(8)
OLS: DV in logs
Followers rank
Followers rank
Friends rank
Friends rank
Tweeting rank
Tweeting rank
Polarizing votes
Polarizing votes
Narrative
-0.1450337
(0.1011615)
-0.1054102
(0.1065055)
-0.2210472**
(0.0966755)
-0.1753784*
(0.1032931)
-0.3399455***
(0.1142836)
-0.3355305***
(0.1193212)
-0.0504206
(0.0445189)
-0.0305581
(0.0377497)
Positioning
0.1859631**
(0.0858288)
0.1476951*
(0.0904867)
0.1399076*
(0.0818949)
0.0771667
(0.087537)
-0.1376972
(0.0952966)
-0.1122817
(0.1002503)
0.0870401**
(0.0383814)
0.0824784***
(0.0321193)
Providing info
0.2284105***
(0.0826242)
0.2926232***
(0.0867325)
-0.4409936***
(0.0806668)
-0.4399356***
(0.085657)
-0.3475115***
(0.0924791)
-0.336393***
(0.096873)
0.0695854*
(0.0368396)
0.0428683
(0.0308668)
Requesting
action
0.2275324
(0.2843258)
0.1476953
(0.2936186)
0.5020254*
(0.2860291)
0.4140085
(0.300259)
0.3201715
(0.3178105)
0.4061579
(0.3255249)
-0.1252984
(0.1269156)
0.0489258
(0.106614)
Thanks
0.0835565
(0.1613563)
0.2281673
(0.174364)
-0.0567282
(0.1546586)
0.0150429
(0.1693158)
-0.2547207
(0.1798114)
-0.3428282*
(0.1962796)
-0.2426859***
(0.0731434)
-0.1640155***
(0.0613789)
Other
-0.0771386
(0.2712884)
0.0149549
(0.2783337)
0.0535897
(0.2499881)
0.0881971
(0.2588879)
-0.290631
(0.2812036)
-0.3369563
(0.2927218)
0.0858603
(0.1135076
0.032759
(0.0955079)
Male
-0.7484235***
(0.1074439)
-1.531073***
(0.131873)
-0.226497**
(0.1044496)
-0.850583***
(0.127668)
1.094355***
(0.1563083)
1.156563***
(0.1744808)
0.3545207***
(0.0440087)
Republican
1.201852***
(0.0969031)
1.447268***
(0.1155909)
0.8278948***
(0.0914326)
1.201361***
(0.1142744)
0.2225375**
(0.1044148)
0.3755097***
(0.1187301)
0.3447262***
(0.0350556)
Senate
1.022279***
(0.0934136)
1.442037***
(0.1054439)
0.2382172***
(0.0883219)
0.6176648***
(0.1004663)
0.1441224
(0.1033486)
0.0881431
(0.110053)
-0.0595531*
(0.0344298)
DW Nominate
1.348098***
(0.2182876)
0.8990377***
(0.2153392)
-0.714075***
(0.2494)
N
791
711
791
711
791
711
711
711
Chi2
221.33***
346.09***
122.57***
211.27***
94.69***
93.86***
F-stat
5.16***
38.86***
Pseudo R2
0.0739
0.1202
0.0360
0.0708
0.0436
0.0485
R2
0.0421
0.3328
25
Appendix
!"#$%&'()
*+,-."#$%&'()
/$0121#1&2)
34"567$89):12#8)
;.#1&2)
<"''"#1=$)
#$7712%)")8#&'()",&+#)#:$1')>"(9)>$8.'1,12%)".#1=1#1$8)
?)'&>$)5(),1@$)#&>"()
;.#1&2)
A&81#1&212%)
81#+"#12%)&2$B8)8$70)12)'$7"#1&2)#&)"2&#:$')6&71#1.1"2)&')
6&71#1."7)188+$)
C"(&')355"2+$7)18)D+8#)
E'&2%)
;.#1&2)
A'&=1>12%)120&'5"#1&2)
%1=12%)120&)",&+#)8&5$#:12%9)$F%F9)6&12#12%)#&)")'$8&+'.$)GHI)
J181#)GHI)0&')6"'@12%)8#1.@$')
120&9)K.:$.@)#:18)&+#L)GHIK)
;.#1&2)
H$M+$8#12%)".#1&2)
$4671.1#7()#$7712%)0&77&E$'8)#&)%&)>&)8&5$#:12%)N2&#)D+8#)
=181#12%)")712@O)
:$76)E1#:)2$1%:,&':&&>).7$"2)
+6)#:18)E$$@$2>)
;.#1&2)
/1'$.#).&55+21."#1&2)
'$M+1'$8)P1>).&2=$2#1&2)
;.#1&2)
Q#:$')
>&$82B#)01#)12)"2()&#:$');.#1&2)."#$%&'()
!&2#$2#)
<$E8)
6&12#8)#&)2$E8)1#$5)$78$E:$'$)
8$$)#:18)#'1,)"'#1.7$L)GHI)
!&2#$2#)
R'1=1")
0+2)0".#89)'"2>&5)120&)",&+#)#:$)#E$$#$'9).&28#1#+$2.(9)$#.F)
!&2#$2#)
!&56715$2#8SR:"2@8S<1.$#1$8)
8"(8)21.$)#:12%8)",&+#)8&5$&2$)$78$)
!&2%'"#+7"#1&289)0'1$2>9)&2)
(&+')E$>>12%)
!&2#$2#)
3=$2#8)
6+,71.1T12%)"2)$=$2#)
U"'5$'8)C"'@$#)18)*"#+'>"()
!&2#$2#)
V+$8#1&28)
M+$'(12%)#:$)"+>1$2.$)N2&#)8"'."8#1.)&')':$#&'1."7)M+$8#1&28O)
*&4)&')!+,8W)
!&2#$2#)
A&71.()
8#"#$5$2#)",&+#)6&71.1$8)"2>)#:$1').'$"#1&2)
?8)#:18)#:$)%'$"#)Q,"5")D&,)
67"2W)
!&2#$2#)
AH)
8$70)6'&5&29),'"%%12%)
8$$)5$)&2)RJ)U'1>"()
!&2#$2#)
Q#:$')
>&$82B#)01#)12)"2()&#:$')!&2#$2#)."#$%&'()
C"22$')
X"7>)
'$"77()>1'$.#).&55$2#8Y)2&)$=1>$2.$)&0):$>%12%)&'):1>12%)
#:$)'$M+$8#)&')8#"#$5$2#)
/$58):"=$)2&)67"2)0&')D&,8Y)
ZQA):"8)")67"2Y)K8$$)5$)&2)RJ)
U'1>"(K)
C"22$')
A&81#1=$)
.&55$2#)#&)5"@$)#E$$#$')7&&@)%&&>)&')5"@$)&#:$'B8)0$$7)
:$B8)&2)#:$1')81>$Y):$)E"2#8)&#:$')6$&67$)#&)71@$):15)
I$#B8)"77)61#.:)12)*"#+'>"(Y)?B77)
,$)#:$'$)#&&Y)K67$"8$)
'$5$5,$')#&)6+#)(&+')#'"8:)
."28)"E"(K)
C"22$')
<$%"#1=$)
:$)E"2#8)#&)>&)E:"#$=$'):$)E"2#8)E1#:&+#)156$>12%)&2)
8&5$&2$9)"8@)12)")E"()#:"#)#:$)"+>1$2.$)E&2B#)0$$7)
#:'$"#$2$>Y)1F$F9),$12%)6"881=$)"%%'$881=$)
$F%F9)K85"77)0"=&'K9)KD+8#)")
512+#$K)
C"22$')
Q#:$')
>&$82B#)01#)12)#:$)&#:$')C"22$')."#$%&'()
