Computing Political Preference among Twitter Followers
Jennifer Golbeck
Human-Computer Interaction Lab
College of Information Studies
University of Maryland
College Park, MD 20742
jgolbeck@umd.edu
Derek L. Hansen
CASCI, Human-Computer Interaction Lab
College of Information Studies
University of Maryland
College Park, MD 20742
dlhansen@umd.edu

ABSTRACT
There is great interest in understanding media bias and
political information seeking preferences. As many media
outlets create online personas, we seek to automatically
estimate the political preferences of their audience, rather
than of the outlet itself. In this paper, we present a novel
method for computing preference among an organization’s
Twitter followers. We present an application of this
technique to estimate political preference of the audiences
of U.S. media outlets. We also discuss how these results
may be used and extended.
Author Keywords
Twitter, politics, liberal, conservative, news, journalism
ACM Classification Keywords
H5.m. Information interfaces and presentation (e.g., HCI):
Miscellaneous.
General Terms
Human Factors
INTRODUCTION
As major media outlets establish online presences in social
media, understanding the characteristics of those audiences
is an important task. It has implications for how information
is presented in this environment where personalization is
expected. Furthermore, it can provide valuable information
to marketers and social media analysts.
As a first step toward understanding audiences, we present
a technique for estimating audience preferences in a given
domain on the microblogging service Twitter. We use U.S.
politics as our motivating example by estimating the
political preferences of media outlets’ audiences.
BACKGROUND
While we are not studying media bias, but rather the
political preferences of audiences, it is worth briefly
discussing the extensive research on analyzing media bias.
A subset of this work uses automated methods to infer
liberal/conservative bias of news stories and outlets. These
automated methods do not depend on subjective
measurements of bias, although the specific techniques used
to infer bias can be problematic and are highly contested.
One approach is to compute a media bias score based on
citations in the news story – news outlets that cite “think
tanks” that are also cited by Congressperson’s with a
known liberal bias are assumed to be more liberal [8].
Another approach is to compare keywords and phrases used
by Congresspeople of known political persuasions with
those used in news articles – news outlets that use terms
like “death tax” and “illegal immigration” are more likely
to be conservative [7]. A final approach assigns a
liberal/conservative score to web documents based on the
number of times they are co-cited with other web
documents that have a known political bias [3].
In contrast to these approaches, we estimate the political
preferences of news outlet audiences, not the news outlet
content itself. Our strategy is similar to [8] in that we use
Congresspeople’s American for Democratic Action’s
(ADA) scores as a starting point for our scoring; however,
we use Twitter Follow relationships rather than article
citations. Using Follow relationships avoids the concern
with [8] that results rely too much on the citation practices
of journalists and Congresspeople. Our approach does not
require coding of data (as in [8]) or access to large corpuses
of news stories and congressional speeches; it relies instead
on freely available and open access data from Twitter.
METHOD AND SAMPLING
Unlike [8], we are not interested in predicting media bias.
Instead, we are interested in predicting the political
preference of the audience of different media outlets and
organizations by using sites like Twitter that embed social
ties. Our examples and applications are in the political
domain, but the technique is generalizable when the right
background information is available. Our approach includes
the following steps:
Step 1: Apply known scores to a seed group, in this case
Congresspeople using Twitter. The base data of
liberal/conservative scores are obtained from Americans for
Democratic Action (ADA), who puts out an annual report
that considers the voting record of members of Congress
[1]. ADA defines a key set of votes that indicate liberal and

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise,
or republish, to post on servers or to redistribute to lists, requires prior
specific permission and/or a fee.
CHI 2011, May 7–12, 2011, Vancouver, BC, Canada.
Copyright 2011 ACM 978-1-4503-0267-8/11/05....$10.00.

CHI 2011 • Session: Microblogging Behavior May 7–12, 2011 • Vancouver, BC, Canada
1105

conservative positions, and uses the Congressperson’s
voting record to assign each a score. The most liberal score
is a 1.0, and the most conservative is 0.0. This is a widely
accepted measure of political position. We apply the 2009
ADA ratings to the 111th Congress members.
Step 2: Map the scores of the seed group onto their
followers. We collect the list of followers for each member
of Congress on Twitter (i.e., Congress Followers). An
inferred political preference score (P Score) for each
Congress Follower is computed as the simple average of the
ADA scores for all Congresspeople he or she follows. Our
approach relies on the assumption that people’s political
preferences will, on average, reflect those of the
Congresspeople they follow. Prior literature on “selective
exposure” to political information suggests this assumption
is reasonable since people seek after information from those
with similar political views [4, 5]. We tested this
assumption in the Twitter context by surveying a
convenience sample (recruited via Twitter and email by the
authors and friends) of 47 subjects who follow politicians
on Twitter. Of those, 66% stated that the politicians they
follow “mostly share [their] political views”, while only 4%
follow politicians who “mostly hold political views that
oppose [their] own.” The remaining 30% reported
following a mix of both, with an average of 56% of the
politicians they follow sharing their own views. Thus,
overall, users tend to follow politicians with similar views;
even when there is a mix of political views, those that
match the user's views tend to dominate. These preliminary
results suggest that our assignment of a P Score to
congressional followers is a reasonable approximation of
real political preference. An adjusted P Score could be
computed based on data from more comprehensive surveys
that compare self-reported liberal/conservative scores with
inferred P Scores.
Step 3: Map the inferred scores of the seed group followers
(i.e., Congress Followers) onto the target of the
investigation, in this case the Twitter accounts of media
outlets. A simple approach is to assign the average of the
liberal/conservative scores of all Congress Followers who
also follow the target media outlet. However, this approach
raises a problem: Twitter users may not represent the
population well.
Indeed, on Twitter Republican members of Congress
significantly outnumber members who are Democrats (127
to 103), though Democrats outnumber Republicans in the
111th Congress. Furthermore, Republican Twitter users tend
to have disproportionately more followers than Democrats.
Even excluding John McCain, who has over 1.7 million
followers as a result of the 2008 Presidential election (more
than 30 times the next most followed member of Congress
and nearly double the total number of followers of all other
Congresspeople), the sum of Republican followers is
581,997, compared to 291,050 for Democrats. As a result,
the pool of Congress Followers is significantly biased
toward conservatives.
Without any adjustment, the audiences of news outlets will
incorrectly appear more conservative. Our solution is to
selectively sample Congress Followers so they more closely
match the roughly equal ratio of Republicans and
Democrats in the general population, minimizing the effect
of the initial selection bias. John McCain was excluded as
an outlier because his Presidential run makes him a
particularly abnormal and overly influential data point.
Congresspeople were broken into groups by the number of
followers they had: over 10,000, 5,000, 1,000, 500, 100, 50,
and 10. Within each group, we randomly selected equal
numbers of Republican and Democratic representatives
until we had the maximum number for the least represented
group. We chose equal numbers from each party since in
recent elections there are roughly equal numbers in
Congress. For each selected Congressperson, we randomly
selected the number of followers equal to the group in
which he was a member (e.g. for a Congressperson from the
“over 1,000” group, we selected 1,000 random followers).
To compute the final media audience P Scores, we used a
10-fold validation; we drew 10 samples using this
technique, computed the P Score for each media audience
using each sample, and averaged the scores across samples.
CONGRESS FOLLOWER SCORES
Figure 1 shows a distribution of scores for all Congress
Followers, as well as the distribution of ADA scores for all
Congresspeople (not just those on Twitter).

Figure 1. The distribution of P Scores among all Congress
Followers and ADA scores for Congresspeople. A score of 0.5
is moderate, 0.0 is very conservative and 1.0 is very liberal.
The distribution of scores among the Congress Followers
and the sampled follower population is strongly bimodal,
since it is based upon the congressional ADA scores which
are bimodal. This is in contrast to other evidence that
suggests the distribution of political ideologies among the
general public is a more normal distribution, with most of
the population as moderate [6]. This discrepancy may arise
from two sources. First, followers of Congresspeople may
not share the views of those they follow, although our
survey and prior “selective exposure” literature makes this
unlikely.
CHI 2011 • Session: Microblogging Behavior May 7–12, 2011 • Vancouver, BC, Canada
1106

Media Outlet Twitter User ID Total
Followers
Avg. % of Total
Followers sampled
Avg. P Score
Fox News foxnews 266,121 7.33% 0.26
The Drudge Report Drudge_Report 102,981 20.52% 0.27
Washington Times washtimes 13,545 33.75% 0.29
Wall Street Journal WSJ 392,332 5.75% 0.41
US News & World Report usnews 7,836 28.60% 0.49
The L.A. Times latimes 72,296 9.91% 0.51
USA Today USATODAY 63,714 9.90% 0.51
Good Morning America gma 1,698,875 0.72% 0.51
The News Hour NewsHour 54,787 12.79% 0.52
CBS News CBSNews 1,578,599 0.94% 0.53
Newsweek Newsweek 1,256,536 1.25% 0.54
Washington Post washingtonpost 154,400 9.51% 0.54
The Today Show todayshow 629,088 1.89% 0.55
The Early Show theearlyshow 12,628 11.16% 0.55
Time Magazine TIME 2,134,411 0.99% 0.57
ABC World News abcworldnews 12,973 14.93% 0.57
CNN Breaking News Cnnbrk 3,314,716 1.40% 0.58
NBC Nightly News nbcnightlynews 27,137 16.85% 0.58
The New York Times nytimes 2,553,291 1.40% 0.60
Morning Edition (NPR) MorningEdition 6,146 17.22% 0.66
Table 1. The average audience P Scores for the twenty popular media outlets studied in [8]. Averages are computed over 10
randomly drawn samples using the sampling method described above. Results are sorted from most conservative to most liberal.
Second, those who follow Congresspeople on Twitter may
have more polarized political tendencies than the overall
US population. This is likely the case, as those who actively
decide to follow a Congressperson are likely more
politically aware and active, characteristics that are often
accompanied by more extreme political tendencies. If this is
true, our estimated audience political preference scores for
media outlets will rely on the Twitter follow practices of
“politically savvy users” rather than the general population.
This is not necessarily problematic, although it does impact
the meaning of the P Scores. In our political example with
the bi-modal distribution, the Avg. P Score for a media
outlet is roughly equivalent to the ratio of “political savvy”
liberals and conservatives that follow that news outlet.
MEDIA AUDIENCES’ POLITICAL PREFERENCE
Our first application of this method was to estimate the
audiences' political preferences, through their P Scores, of
the same popular media outlets evaluated in [8]. Table 1
shows the audience bias scores along with the percentage of
each outlet’s followers who were considered when
computing the value.
For traditionally conservative outlets [9], such as Fox News
[2], we found audiences with correspondingly conservative
P Scores: Fox News (0.256), the Drudge Report (0.265),
and the Washington Times (0.290). There are no outlets
with audiences that have P Scores that are liberal to the
same extent that these are conservative, but some liberal
preference is visible in the audiences of outlets like the New
York Times (0.604) (considered to be liberal leaning [9,
11]) and NPR’s Morning Edition (0.659).
The vast majority of these media outlets’ audiences – 15 out
of 20 – fall between the moderate scores of 0.4 and 0.6.
Half are even closer to the midpoint, falling within 0.05 of
the perfect moderate 0.5 value. This suggests that most
media outlets audiences have roughly equal numbers of
liberal and conservative Congress Followers.
Our findings of audience preference are similar to the
estimates of political orientation of media outlets found in
[3], which used co-citation of hyperlinks to infer political
orientation of web documents and their associated news
outlets. Although, the scales differ, like [3] we found that
the Wall Street Journal and The New York Times equally
deviated from the middle in opposite directions
(conservative and liberal respectively). This similarity may
come from the fact that [3] relies partially on news outlet
audiences by using website linking behavior. Our method
provides a more direct estimate of audience political
preferences by focusing on follower relationships.
Note that these scores do not imply that the outlets
themselves present news in a way that reflects their
audience's political preferences.
CHI 2011 • Session: Microblogging Behavior May 7–12, 2011 • Vancouver, BC, Canada
1107

DISCUSSION
There are a number of implications and areas for future
work that follow from these results. Many users expect
personalized web-based content. Some personalization
comes from the structure of the services themselves (e.g.
who users choose to follow on Twitter affects which tweets
they see). However, there are many opportunities to further
personalize and enhance the way information is presented.
Understanding the political preference of an audience can
be important for presenting tailored information (including
or excluding information according to the user's tastes) and
personalizing the user’s experience (e.g. through
recommender systems). For example, an audience’s
political preference can be used as input into recommender
systems. In collaborative filtering systems, items are
recommended by finding people with tastes similar to the
user and recommending things those people like. In this
context, if we know a user’s political preferences, we can
find media outlets that have audiences with a similar
preference, mimicking the basic idea behind collaborative
filtering. Tweets (or information provided on other social
media sites) can be highlighted, filtered out, or sorted based
on the similarity of their audiences’ political preferences to
those of the user. Alternatively, our method could be used
to help recommend tweets commonly read by people on
both sides of the political spectrum reducing homophily
[10]. Finally, marketers and analysts can use our method to
measure their Twitter reach within different political
markets to see if they are reaching their intended audience.
Outside of personalization, this technique may have
applications for studying media bias in social media. While
we have set out to estimate audience political preference,
not media bias, previous work has shown that news
consumers have a significant preference for like-minded
media outlets [4, 5] and use new social media tools to
actively seek out those with similar views. This implies that
people may choose a media outlet because its presentation
of the news reflects their own political beliefs, and thus the
preference of an audience may generally reflect the bias of
the outlet. While we do not have evidence to support this
connection, predicting media bias based on audience
preferences is an area for future research.
Finally, while we have used political preference and media
outlets as our example case, we believe this technique is
applicable in other domains. We have begun to use this
technique to estimage the political preferences of Twitter
audiences of government agencies, think tanks, political
organizations, and individuals. It could also be used for
non-political analysis. For example, a similar analysis could
be done using the Green Scores that rate how
environmentally responsible companies are. These could be
used to create an environmental score for Twitter followers
of organizations, politicians, and other Twitter accounts.
CONCLUSION
In this paper, we have presented a technique for estimating
the political preferences of Twitter account followers.
Using the political domain as our motivating example, we
present a method that uses the follower connections in the
Twitter network to propagate liberal/conservative scores
from members of Congress to Congress Followers and then
to audiences of media outlets. Our results show that the
estimated political preferences of media outlets' audiences
reflect the liberal/conservative leanings of the media outlets
as presented in prior literature.
The results have potential applications for motivating new
interface personalization techniques, understanding media
bias, and for detecting other types of audience preferences
in different domains. There is much future work to be done
in this space and we hope this initial work serves as
motivation to pursue those issues.
ACKNOWLEDGMENTS
Thanks to members of the University of Maryland's Human
Computer Interaction Lab (HCIL) for their helpful
comments on an earlier draft.
REFERENCES
1.Americans for Democratic Action. Annual Voting
Records, 2009. Available at http://www.adaction.org/
2.DellaVigna, S. and Kaplan, E. The Fox News Effect:
Media Bias and Voting. Quarterly Journal of Economics,
122, 3 (2007), 1187-1234.
3.Efron, M. 2004. The liberal media and right-wing
conspiracies: using cocitation information to estimate
political orientation in web documents. In Proceedings of
CIKM’04. (Washington, D.C., November 2004), ACM
Press, 390-398.
4.Frey, D. Recent research on selective exposure to
information. Advances in experimental social psychology,
19, 1986.
5.Garret, K. Politically Motivated Reinforcement Seeking:
Reframing the Selective Exposure Debate. Journal of
Communications, 59, 4 (December 2009), 676-699.
6.Gelman, A. Red state, blue state, rich state, poor state:
why Americans vote the way they do. Princeton Univ
Press, Princeton NJ, 2008.
7.Gentzkow, M. and Shapiro, J. What drives media slant?
Evidence from US daily newspapers. Econometrica, 78, 1
(2010), 35–71.
8.Groseclose, T. and Milyo, J. A Measure of Media Bias.
The Quarterly Journal of Economics, 120, 4 (2005),
1191–1237.
9.Mondo Times: The worldwide news media directory.
Available at http://www.mondotimes.com
10.Munson, S. A. and Resnick, P. 2010. Presenting diverse
political opinions: how and how much. In Proceedings of
CHI’10. (Atlanta, GA, April 2010). ACM Press, 1457-
1466.
11.Puglisi, R. Being the New York Times: The Political
Behaviour of a Newspaper. Available at SSRN:
http://ssrn.com/abstract=573801
CHI 2011 • Session: Microblogging Behavior May 7–12, 2011 • Vancouver, BC, Canada
1108