Southern Illinois University Carbondale
OpenSIUC
Working Papers Political Networks Paper Archive
2011
Mapping the Political Twierverse: Finding
Connections Between Political Elites
Leticia Bode
University of Wisconsin - Madison, lbode@wisc.edu
Alexander Hanna
University of Wisconsin - Madison, [email protected]
Ben Sayre
University of Wisconsin - Madison, [email protected]
JungHwan Yang
University of Wisconsin - Madison, jyang66@wisc.edu
Dhavan V. Shah
University of Wisconsin - Madison, dshah@wisc.edu
Follow this and additional works at: hp://opensiuc.lib.siu.edu/pn_wp
is Article is brought to you for free and open access by the Political Networks Paper Archive at OpenSIUC. It has been accepted for inclusion in
Working Papers by an authorized administrator of OpenSIUC. For more information, please contact [email protected].edu.
Recommended Citation
Bode, Leticia; Hanna, Alexander; Sayre, Ben; Yang, JungHwan; and Shah, Dhavan V., "Mapping the Political Twierverse: Finding
Connections Between Political Elites" (2011). Working Papers. Paper 59.
hp://opensiuc.lib.siu.edu/pn_wp/59
Mapping the Political Twitterverse:
Finding Connections Between Political Elites
∗
Leticia Bode
Alexander Hanna, Ben Sayre, JungHwan Yang, Dhavan Shah
University of Wisconsin-Madison
June 17, 2011
∗
This paper was prepared for the 2011 Political Networks Conference, Ann Arbor, MI.
1
Abstract
Twitter provides a new and important tool for politi-
cal actors, and is increasingly being used as such. In the
2010 midterm elections, the vast majority of candidates for
the U.S. House of Representatives and virtually all can-
didates for U.S. Senate and governorships used Twitter to
reach out to potential supporters, direct them to particular
pieces of information, request campaign contributions, and
mobilize their political action. Despite the level of activity,
we have little understanding of what the political Twitter-
verse looks like in terms of communication and discourse.
This project seeks to remedy that lack of understanding
by mapping candidates for federal office in 2010 and their
followers, according to their use of the 4016 most used hash-
tags (keywords). Our data set is uniquely constructed from
tweets of most of the candidates running for the U.S. House
of Representatives in 2010, all the candidates for the Senate
and governorships, and a random sample of their followers.
From this we utilize multidimensional scaling to construct
a visual map based on hashtag usage. We find that our
data have both local and global interpretations that reflect
not only political leaning but also strategies of communi-
cation. This study provides insight into innovation in new
media usage in political behavior, as well as a snapshot of
the political twitterverse in 2010.
2
1 Introduction
Twitter as a medium of communication has rapidly come of age since its
creation in 2006. As usership of Twitter has grown, so too has its adoption
in new arenas, including that occupied by politicians in the United States.
As political officials adopt and use Twitter in larger numbers, it becomes
important for us to understand how strategic politicians and political elites
are making use of this new media platform.
This study takes a focused and unique sample from the political Twit-
terverse – candidates in the U.S. Congress in 2010 and their followers – to
create a map of the space occupied by political elites on Twitter. Based on
elements of Twitter speech (namely hashtags) we are able to gain a more
nuanced understanding of what these universes of discourse look like, and
how political users are connecting with one another in this new medium.
With the dramatic growth in p opularity of Twitter, political actors have
increasingly begun to use tweets as one of many campaign tools. The vast
majority of candidates for the U.S. Congress in 2010, for instance, employed
Twitter at least marginally in their campaign strategy. Political use of Twit-
ter by candidates included efforts to reach out to potential supporters, direct
them to particular pieces of information, request campaign contributions, and
mobilize political action. As a result, a large amount of valuable information
regarding political behavior is embedded in the political Twitterverse. While
this is only one of many campaign tools and strategies, it is useful in that it
3
gives us an easily measurable proxy for candidate outreach efforts, as well as
offering an understanding of organic connections amongst political elites on
Twitter.
1.1 Political Use of Twitter
The vast majority of research to date on the political use of Twitter has
focused on members of Congress. Scholars have considered both what en-
courages members of Congress to adopt use of Twitter, and what helps them
to be “successful” in such use. Lassen & Brown (2010) found members are
more likely to adopt Twitter if their party leaders urge them to, if they are
young, or if they serve in the Senate, whereas Gulati & Williams (2010) de-
termined that party (Republicans adopt more) and campaign resources were
the most important predictors of adoption. Chi & Yang (2010a) suggest
that adoption is driven by a desire for constituency outreach, rather than a
transparency motivation. Adoption may be accelerated by evidence of past
users’ success with the medium (Chi & Yang 2010b), and factors includ-
ing vote share, funding, usage and influence may help to explain why some
congressional users have more followers than their colleagues.
A single study to date has examined Twitter use within the electoral
context, in an attempt to predict election outcomes. Tumasjan, Sprenger,
Sandner & Welpe (2010) searched for mentions of political candidates and
political parties in tweets. Simple word count analysis of this sample of
explicitly political tweets revealed that the more frequently a candidate or
4
party was mentioned, the more likely electoral victory for that entity.
Our study hopes to further this literature by incorporating a global un-
derstanding of the political Twitterverse and using elements from this under-
standing as a way to gain insight into political outcomes. This is by design
an exploratory study, not aiming at explanation of any explicit outcome, but
attempting to describe the shape of a communicative space with regard to a
particular subject matter.
1.2 Multi-dimensional Issue Spaces
For decades, scholars have been concerned that issues are too often described
merely in terms of “right” and “left.” A single dimension of almost any
issue space is likely to be flawed, as there are multiple conflicting factors in
competition for any given issue. Whether to drill in the Alaskan wilderness,
for instance, may pit environmental concerns against fiscal concerns, and
also includes issues like security and federalism. A fiscal conservative who is
socially or environmentally liberal may be truly conflicted on where she fits
on a uni-dimensional spectrum of liberal to conservative. This is apparent
empirically, in that some issues simply do not conform to a left-right spectrum
(see Norton (1999) for a discussion of the impact of gender, and Anderson
(2007) for a discussion of various issues, including agriculture), as well as
theoretically, in modeling multi-dimensional issue voting (see for example
Downs (1957), Calvert (1985)).
The problem with multi-dimensional issue spaces is the complexity of
5
understanding them. This is the true benefit provided by our analysis of
the political twitterverse. In our mapping, we are able to identify users
who are more similar to each other by virtue of what they actually say –
the elements of sp eech that they share. At the same time, we are able to
identify which articles of political speech are more alike since they share
the same users. Thus an understanding of connections between political
elites emerges organically, by virtue of their own behavior, rather than by
researchers imposing a known spectrum of understanding (most frequently
political ideology) upon their actions. In this way, we are able to achieve
greater understanding of how active users in the political twitterverse connect
to one another and self-identify.
We have two main expectations of what the analysis of the map will
produce. First, we expect major clusters to emerge representing the tradi-
tional ideological extremes of the left-right spectrum. Much like Adamic and
Glance’s famous work mapping the political blogosphere (2005), we expect
to find a sharply divided Twitterverse along party lines. Those who are on
the left will tend to say similar things and use similar hashtags, clustering
together and apart from those on the right who will similarly connect with
co-hashtagging behavior. Second, we expect to pick up not only a global di-
vision based on partisanship, but also to identify local clusters of users who
engage in types of political behavior within or between the classic under-
standings of ideological right and left. Sub-partisan clusters may represent
different understanding of ideology, or strategic attempts at communication
6
by political elites. As Twitter is a nascent communication medium, people
likely attempt to exploit it for their own purposes. Applying this perspective
to the political space, strategy may be to disseminate information as widely
and effectively as possible. Therefore, we expect to detect in our data local
clusters of users who engage in similar diffusion-maximizing behavior.
2 Data
Data for this project were gathered in two waves. The initial wave began
on Labor Day 2010 and was based on a list of 404 candidates in 103 races
for seats in the House of Representatives. At the time when this sample was
started, the total number of Twitter accounts we could follow was too low
for us to be able to follow all candidates for House races along with samples
from their follower lists, so we had to select a subset of races to focus on.
The strategy employed was to include all candidates in races that were either
tossups or leaning to one side or the other (as judged by the New York Times
in the last week of August), along with a handful of noncompetitive races
chosen at random. 16 of these races had not held their primaries by Labor
Day, and for these races all candidates running in the primary were included
in the sample. Of a total of 404 candidates, 253 were from one of the two
major parties and the rest were independents or third-party candidates. Out
of this list of candidates, 233 were found to have Twitter accounts, with 201
of those being major-party candidates.
7
A random sample of followers was taken for each candidate such that
it proportionally decreased as the sample approached the maximum sample
size of 50. The size of the sample per candidate was calculated by
n
c
=
50
1 + (50 − 1)/F
c
(1)
in which F
c
is the total number of followers for candidate C at the beginning
of the measurement p eriod.
The second wave of data collection was started at the beginning of Octo-
ber and included all gubernatorial candidates and all US Senate candidates as
well as a replication of the first wave which resampled the House candidates
using the same sampling formula. Among the races for governor, only three
included a third-party or independent candidate in the race, and only the two
candidates in the Nebraska race had not identifiable Twitter account. The
Senate races were very similar, with only occasional races with third-party
candidates and few candidates without identifiable Twitter accounts.
A new random sample of Twitter followers of all candidates (N = 409)
was added to the existing sample of users being followed, resulting in a total
sample of 23,466 followers. Collection of tweets continued for one month
after the election on Novemb er 2, 2010.
Over this time period (88 days total), nearly 9 million tweets were gath-
ered, either directly tweeted by users in the sample or distributed (retweeted)
by those users. The data were collected by using Twitter’s Streaming API.
8
A feature of the collection with the Streaming API is that the data structure
returned by the API has a number of different elements included along with
the actual tweet. This includes information such as all public user informa-
tion and geolocation, and most relevant for the current project, what the
API labels as “entities”, parts of the text which can be identified as either
a URL (e.g. http://www.website.com), a hashtag (e.g. #politics), or user
mention (e.g. @johndoe). The current project takes advantage of the distinct
enumeration of these “entities” which allow us to parse important informa-
tion from the data structure itself. For this particular study, we focus on
the use of hashtags by users, allowing the data to inform which hashtags are
important within this sample, and then further how groups of users coalesce
based on similar use of hashtags.
3 Methodology
In order to identify our variable of interest – the clustering of Twitter users
together – we performed a multidimensional scaling (MDS) analysis (Kruskal
& Wish 1981) for use of hashtags by every unique user in our dataset. First
we constructed a two-mode matrix of users by hashtags used. Entries in the
matrix were the number of times the user used the hashtag U
ei
normalized
by the user’s total usage of hashtags U
e•
, then weighted by the population’s
total usage of that hashtag P
•i
. Normalizing hashtag use by user helps to
distinguish a user who tweets once and uses a particular hashtag from a
9
user who tweets a thousand times, hashtagging each time, but using that
same hashtag only once. We believe this is an important distinction when
attempting to classify between types of political users of Twitter. We fur-
ther weight by the population’s use of a particular hashtag so that not all
hashtags are assumed equal. Because some hashtags are used with much
greater frequency than others, it is important to give those hashtags greater
importance in our classification scheme as well. Equation 2 expresses this in
mathematical notation.
M = [
U
ei
U
e•
P
•i
] (2)
The MDS analysis was performed using non-metric MDS. Non-metric
MDS attempts to retain rank order of entries as ordered by distance while
at the same time attempting to minimize the badness-of-fit (stress) itera-
tively (Kruskal & Wish 1981). We used the Kruskal’s Non-metric Multidi-
mensional Scaling function included in the R MASS package (Venables &
Ripley 2002). Input to the MDS was a dissimilarity matrix calculated from
Euclidean distance between rows of matrix M in equation 2. We allowed
for two dimensions in order to most easily interpret the results graphically
and substantively (two dimensions map nicely on X and Y axes). Using
additional dimensions did not dramatically decrease the stress.
We gain two main insights from this analysis. First we can see how actors
cluster together in a two-dimensional space by virtue of what they say. This
10
means we can discern distinct groupings of individuals based on their shared
Twitter behavior. Second, we distinguish which entities are substantively
closer to each other. Presumably the concordance of entities like hashtags
uttered similarly by multiple users suggest some other shared unobservable
or latent variable. That is, we imagine that the shared use of language on
Twitter is a proxy for other similarities amongst users within clusters. In the
context of our study, the most likely latent variable is political sentiment.
The clearest understanding of political sentiment is represented by the left-
right spectrum of political ideology. Thus we expect clusters to reflect a clear
left-right division, with the potential for one or many middle categories as
well. Empirically, we expect to see a clustering of hashtags such as #tcot
(top conservatives on Twitter) and #teaparty, and on the other end #tlot
(Top Liberals on Twitter) and #p2 (Progressives 2.0).
The local interpretation of the analysis relies on the ability to observe
clustering in the MDS output. This lends itself to more nuanced interpreta-
tions that do not accord to what may be considered only political leaning.
We can also pick up on the variations in the types of political behavior in
which users engage. Visually, we can discern clusters and attempt to assign
meaning to them based on our knowledge of the cases. We also can identify
clusters systematically using hierarchical cluster analysis (HCA). We used
the output of the MDS analysis to generate a fitted distance matrix based on
Euclidean distance between rows and used this as the input to HCA using
Ward’s method. We use this method to minimize the loss of information we
11
get from the clustering process. This results in compact, spherical clusters
of actors.
4 Making and Interpreting the Map
To generate the MDS, we used the 4979 users who used the most hashtags
and the 4016 top hashtags. We found that using more users and hashtags did
not change the analysis dramatically. Because the same essential dynamics
underly Twitter use as do those behind political blogs, we expect similar
polarization in the political Twitterverse (based on the left/right or Demo-
crat/Republican dichotomy) to that of the political blogosphere (Adamic &
Glance 2005). We can achieve an understanding of this potential polarization
both visually and computationally.
First, we need to be able to understand the map in a meaningful way
and say more about its variance. There are two ways we can interpret the
mapping, the global and local interpretation. The global interpretation lends
itself to interpretation upon the axes. To assess the significance of any global
interpretation, we created a variable based on the number of candidates the
user followed, separated into five categories: Democrat, Republican, Indepen-
dent, Third Party Left, and Third Party Right. The first three categories are
self-explanatory, while last two were generated by categorizing various third-
party groups according to their political leaning (i.e. Green for 3rd Party
Left, and Libertarian and Tea Party for 3rd Party Right). We regressed the
12
coordinates from the MDS using generalized linear modeling (GLM) on this
variable. Essentially, we are interested in seeing how the slopes of the vari-
ous lines generated by this regression vary - the further the distance between
lines, the more distinct the follow patterns of the users. In addition, we cre-
ated separate variables from the two most popular hashtags in the political
Twitterverse: #tcot and #p2. This allows us to see how hashtagging be-
havior falls in terms of the political divide generated by candidate following
behavior, and allows us to answer a number of interesting questions. Do
users employing conservative hashtags follow conservative candidates? Do
third party followers use traditionally party-specific hashtags? Again, we
used GLM to assess the direction to which elements in the map lean.
Figure 1 displays the results for the the map based upon hashtags. We
see the #p2 and #tcot curves approaching the point of being orthogonal
to each other, which is expected given the way that the map is constructed
(remember we imposed two dimensions upon the graph, so it is not surprising
that a conservative hashtag occupies one dimension while a liberal hashtag
occupies another). It seems as though the #p2 line accords with the users
on the left of the map (convenient since it is the hashtag that is supposed to
represent the left of the political spectrum) and the #tcot with those on the
bottom and the right of the map. Again, this is not terribly surprising. We
also see that the lines of the Democrats and the Third Party Left are nearly
identical, and that they are the closest to #p2.
We would expect that the lines for the Republicans and Third Party
13
Figure 1: Multidimensional scaling with trend lines. Stress: 2.55
14
Right approach the #tcot line. Oddly enough, however, their slopes actually
have the opposite sign of their respective hashtag. This could be an artifact
of the estimation procedure, or could reflect unexpected behavior amongst
these Twitter users. Our cluster analysis will speak to this possibility to
some extent, but future research should also consider more explicitly the
relationship between party and hashtag use.
Interestingly, the line for Independents seems to be far removed from
all the other plots. This may reflect a number of possibilities. First, there
were not a large number of independent candidates in our sample. Thus
this particular slope suffers from a relatively small N, and as a result we are
less sure of its slope than we are for some other lines (particularly those for
the major two parties). Additionally, it is quite possible that independents
choose not to engage in use of the two major hashtags we plot, as they are
commonly associated with the two ends of the ideological spectrum, whereas
independents by definition choose a third path. As such their slope in relation
to the two lines reflecting hashtag use holds little true meaning.
The global analysis is revealing in two ways. It confirms that there is po-
larization in the political Twitterverse, in that conservatives and liberals tend
to align separately, and also describes how disparate users are with regard
to hashtag usage. Secondly, and perhaps more importantly to understanding
political behavior, the global analysis indicates that a traditional partisan
dichotomy is not sufficient to explain political behavior (at least in terms of
hashtag use) on Twitter. There seem to be other important mechanisms at
15
Figure 2: MDS with clusters
work with regard to this behavior. Thus we turn to cluster analysis to reveal
other potentially non-partisan breakdowns among political tweeters.
The results of the hierarchical cluster analysis are reported in Figure
2. We chose to separate the map into six clusters, although using five or
seven clusters would not have changed the analysis dramatically. Figure 2
demonstrates these distinct clusters, which change consistently with the X
axis, but change much less on the Y dimension. Again, these clusters are
based on hashtagging behavior on Twitter, suggesting that different types of
such behavior occur within each of the six clusters.
In order to better understand the details of this behavior, we conducted
16
Cluster 1 2 3 4 5 6
N 223 708 1011 1603 616 817
Dem -0.32 -0.25 -0.30 0.08 -0.05 0.06
Rep 0.01 0.04 0.03 -0.06 0.03 -0.11
Indep. 0.16 (ns)* -0.29 -0.05 0.25 (ns) 0.42 0.01 (ns)
3rd R
†
0.30 0.05 (ns) 0.05 (ns) 0.14 (ns) 0.20 (ns) -0.12 (ns)
tcot 0.01 0.01 0.01 -0.10 -0.01 -0.06
p2 -0.01 -0.01 0.01 -0.15 -0.01 0.07
(ns) denotes non-significance at the p ≤ 0.05 level
Logistic analysis performed for inclusion in each cluster individually. † All coefficients for third party left
were insignificant and so are not shown.
Table 1: Local interpretation of clusters
a series of regression analyses, to determine what behaviors predicted a user
falling in any given cluster. Because there are six clusters, six separate re-
gressions were estimated.
The regression analysis is shown in Table 1. Clusters 1 and 6 are the most
straightforward to interpret. Cluster 1 is positively related to all conserva-
tive concepts – following Republicans or members of a right-leaning third
party, and using the hashtag #tcot – and negatively correlated with each
liberal concepts – following Democratic politicians and using the hashtag
#p2 (note that the left-leaning third party following is not shown, though it
was included in each model, as no coefficients representing this behavior were
significant). Cluster 6 represents essentially the opposite set of behaviors –
following Democrats and using #p2, but less likely to follow Republicans or
use the hashtag #tcot. These clusters fit clearly within the left-right di-
chotomy into which American politics is most frequently divided. However,
there are four other clusters of political Twitter users that fall somewhere
outside of that clear spectrum. This is where a local interpretation is of most
use to our understanding of the politics on Twitter.
17
We will address each remaining cluster in turn, and then offer some over-
arching observations regarding the clustering in its entirety. Clusters 2 and 3
are nearly identical – users in both groups are less likely to follow Democratic
and Independent candidates, more likely to follow Republican candidates,
and more likely to use the hashtag #tcot. The notable difference between
the two clusters is their use or avoidance of the hashtag #p2, with cluster
2 avoiding use and cluster 3 positively associated with use. This suggests
that users in cluster 2 resemble cluster 1 users (classic conservatives, avoid-
ing liberal entities), whereas users in cluster 3 offer a new type of political
behavior on Twitter. Cluster 3 users seem to employ both of the two top
political hashtags, even though in terms of following behavior, they seem to
lean right. This may suggest strategic hashtagging, in that Twitter users
following a hashtag either on the left or the right (rather than another indi-
vidual Twitter user) would still be likely to see a tweet, thus disseminating
the information farther than it would otherwise travel, and to a different
audience.
Similarly, clusters 4 and 5 somewhat resemble one another. Users in both
clusters are less likely to use the top political hashtags from either ideological
persuasion. This could be strategic, in that independent-minded users might
choose to avoid classifying their tweets into one camp or the other. These
users might also be employing other, more specialized political hashtags.
Future work within this framework should further determine which hashtags
are most common within each cluster, but particularly in those less likely to
18
use #tcot and #p2. Alternatively, this could represent a cluster of users that
is less sophisticated in Twitter use, and thus less likely to employ hashtags at
all. The main difference between clusters 4 and 5 is their opposite affiliations
in terms of following Democrats and Republicans for office. Cluster 4 users
are more likely to follow Democrats and less likely to follow Republicans,
whereas cluster 5 users are more likely to follow Republicans and less likely
to follow Democrats. This may suggest a slight leaning toward or away
from each of the main parties. Interestingly, cluster 5 users also have strong
likelihood of following Independent candidates (the largest coefficient seen
in any of our models). Thus, in terms of a classic understanding of political
ideology, cluster 5 seems to best represent the true Independents.
How are we to characterize these clusters, then? The major insight pro-
vided by our local analysis is that we cannot assume all political activity
on Twitter falls neatly into the left-right dichotomy to which political scien-
tists are accustomed. Unlike the blogosphere, which has very little political
middle ground (Adamic & Glance 2005), many of the users on Twitter have
mixed following patterns and hashtagging behavior, suggesting greater nu-
ance in the political behaviors and discussion occurring within Twitter. At
the very least, this may serve as a call for greater research into the burgeoning
political Twitterverse.
19
5 Conclusion
Through this project we have developed a method of creating a map of the
political Twitterverse, using the built-in functionality of Twitter. We found
that solely left/right distinctions, while useful in some ways, inadequately
describe political behavior on the platform. Rather, we find it much more
fruitful to discuss how users employ Twitter for political purposes in more
nuanced ways, including how they interact with one another and how they
self-affiliate and self-identify using the tools available to them within Twit-
ter. Most notably, we think it fruitful to consider the strategic nature of
political action and conversation on Twitter, particularly in terms of strate-
gic hashtagging, such as encroaching on others’ keywords. Moreover, the
construction of this map may ultimately be useful in attempting to explain
political outcomes such as elections, referenda, protests, and the like.
This analysis only chose to look at the use of hashtags in mapping the
political Twitterverse. We hope that this method can b e generally extrapo-
lated to mapping any bounded space of discourses in the social media sphere,
and attempting to explain outcomes in that space by virtue of elements of
the map. By the same token, we could have used other entities used within
the realm of Twitter, such as URLs and user mentions. Future research
should consider whether similar rules apply in alternative areas of Twitter,
or whether the more directed user mention is employed differently. Addition-
ally, with computer-aided content analysis software we can create an entirely
20
new set of attributes from which to categorize tweets, using political tweeters
own language to inform us.
While the potential for research within this subfield of study is enormous,
our project represents an important step in understanding the political in-
teractions and connections happening every day on Twitter.
21
Works Cited
Adamic, Lada A. and Natalie Glance. 2005. “The Political Blogosphere
and the 2004 U.S. Election: Divided They Blog.” WWW-2005 Work-
shop on the Weblogging Ecosystem, May 10-14, 2005, Chiba, Japan.
Anderson, Sarah E. 2007. “Are We Missing Something? Assessing
Measures of Congressional Ideology.” Unpublished manuscript.
Calvert, Randall L. 1985. “Robustness of the Multidimensional Vot-
ing Model: Candidate Motivations, Uncertainty, and Convergence.”
American Journal of Political Science Vol 29:1. 69-95.
Cha, Meeyoung, Hamed Haddadi, Fabricio Benevenuto, and Krishna
P. Gummadi. 2010. “Measuring User Influence in Twitter: The Mil-
lion Follower Fallacy.” Proceedings of the Fourth International AAAI
Conference on Weblogs and Social Media.
Chi, Feng, and Nathan Yang. 2010a. “Twitter in Congress: Outreach
vs. Transparency.” Available at SSRN: http://ssrn.com/abstract=1630943.
Chi, Feng, and Nathan Yang. 2010b. “Twitter Adoption in Congress.”
Review of Network Economics.
Downs, Anthony. 1957. An Economic Theory of Democracy. Harper
and Row.
22
Gulati, Girish J. and Christine B. Williams. 2010. “Communicating
with Constituents in 140 Characters or Less: Twitter and the Diffusion
of Technology Innovation in the United States Congress.” Presented at
the annual convention of the Midwest Political Science Association.
Kruskal, Joseph B. and Myron Wish. 1981. Multidimensional Scaling.
Sage Publications.
Lassen, David S. and Adam R. Brown. 2010. “Twitter: The Electoral
Connection?” Social Science Computer Review.
Lerman, Kristina and Rumi Ghosh. 2010. “Information Contagion:
An Empirical Study of the Spread of News on Digg and Twitter Social
Networks.” Proceedings of the Fourth International AAAI Conference
on Weblogs and Social Media
Norton, Noelle H. 1999. “Uncovering the Dimensionality of Gender
Voting in Congress.” Legislative Studies Quarterly Vol 24:1. 65-86.
O’Connor, Brendan, Ramnath Balasubramanyan, Bryan R. Routledge,
and Noah A. Smith. 2010. “From Tweets to Polls: Linking Text
Sentiment to Public Opinion Time Series.” Proceedings of the Fourth
International AAAI Conference on Weblogs and Social Media.
Tumasjan, Andranik, Timm O. Sprenger, Phillipp G. Sandner, and
Isabell M. Welpe. 2010. “Predicting Elections with Twitter: What
23
140 Characters Reveal about Political Sentiment.” Proceedings of the
Fourth International AAAI Conference on Weblogs and Social Media.
Venables, W. N., and B.D. Ripley. 2002. Modern Applied Statistics
with S, Fourth Edition. New York: Springer.
Yang, Jiang, and Scott Counts. 2010a. “Comparing Information Diffu-
sion Structure in Weblogs and Microblogs.” Proceedings of the Fourth
International AAAI Conference on Weblogs and Social Media.
Yang, Jiang, and Scott Counts. 2010b. “Predicting the Speed, Scale,
and Range of Information Diffusion in Twitter.” Proceedings of the
Fourth International AAAI Conference on Weblogs and Social Media.
24
