Birds of a Feather Tweet Together: PartisanStructure in Online Social Networks

David Sparks∗

September 7, 2010

Abstract

Online social networking sites give observers a unique perspective on thepurposive communication and positioning of political elites as they interactwith each other and with non-elites. A novel dataset culled from the webservice Twitter of the connections and communications between political actorsgives us insight into the purposively expressed preferences and priorities ofpoliticians and the public at large.

Social network analysis and statistical scaling techniques show that a net-work of political elites has a strong partisan structure, and further that thereis an ideological dimension to the choices made by members of the public as towhich political elites to attend to. The aggregation of many individual atten-tion allocation decisions is shown to reveal a sophisticated level of ideologicaldiscernment.

Dating from the work of Mayhew (1974) and Fenno (1978), Political Scientistshave recognized that much of politicians’ time and energy is spent communicatingwith constituents. While politicians act as though they believe constituents are pay-ing attention (Kingdon, 1967), there is at best only mixed support for this notion(see Sniderman, 1993). It has become easier, however, for interested non-elites toselectively attend to the behavior and communications of politicians, governmentagencies, and media, just as the venues for selectively targeted messaging have ex-panded.

Two of the more notable recent innovations in communication are the onlinesocial network and the related, though separate, set of “microblogging” services.

∗Graduate in Political Science, Duke University, Durham, NC; d.sparks@duke.edu; (919) 724-4443

1

Prominent features found among many instances of both of these types of onlineapplications are the ability for users to post “status updates,” indicating what theuser is doing, how they are feeling, etc., and some functionality that allows usersmake connections with others, as co-workers, friends, followers, or fans.

Such services are a goldmine for social scientists, as they offer a massive sourceof data about users’ interests, perceptions, and behavior, as conveyed voluntarilyby users themselves in the messages they send and the updates they post. Addi-tionally, there is substantial information encoded in the patterns of behavior andconnections between users. This paper focuses on the latter data source, as foundin the microblogging service Twitter1. Even without studying the content of onlinepolitical communication, there are interesting implications of the structure of suchcommunication.

Literature

Twitter

Still a relatively new phenomenon, microblogging, and specifically Twitter, have notyet received much academic attention, especially from a Political Science perspective.Java et al. (2009) offer a very thorough overview of the structure of the overall Twitternetwork, as well as the nature of sub-community formation, and a classification ofuser intentions and types (see also Krishnamurthy, Gill and Arlitt, 2008, for an earlieranalysis). Huberman, Romero and Wu (2008) find that, compared to the number ofaccounts followed by a user, the number of other users with whom any individualhas significant interaction (they label these as “friends”) is very small.

As the use of Twitter has become more widespread, increasing attention is beingpaid to the way the service is used by politicians. Some media outlets have examinedthe relative influence of congressional Twitterers, using an online service called Kloutthat purports to measure the “size and strength of a person’s sphere of influence onTwitter”2. One media study (Sifry, 2009) identified Congressman Joe Wilson (R-SC) as the most influential Congressman, shortly after his outburst during a speechgiven by President Barack Obama in September, 2009. In the wake of his nearlyinstantaneous rise to Internet fame, Wilson actually hired a “new-media strategist”to handle his online communications, including the posting of status updates toTwitter (Zimmerman and Wilkie, 2009). Despite its seeming popularity, not even

1http://twitter.com/2http://klout.com/kscore/

2

half of U.S. Representatives and Senators have Twitter accounts; many claim thatit isn’t the best way to communicate with constituents, or they are unsure of itslong-term viability as a medium (Beam, 2009).

Golbeck, Grimes and Rogers (2009) content-analyze 6,000 Tweets from the ac-counts of members of Congress, and find that over 80% of all messages either linkto news articles or press releases, or present basic information about the Member’swhereabouts, activities, or schedule. A Congressional Research Service study (Glass-man, Straus and Shogan, 2009) of lawmakers’ use of Twitter during a two-week periodin 2009 finds that Republican Congressional Twitterers outnumbered their Demo-cratic counterparts by roughly 2 : 1, and that members, as a whole, sent 85 Tweetsper day. This study also replicates the finding that the plurality of updates offerlinks to news or press releases, while only 1.4% of all Tweets could be characterizedas replies to messages posted by others. Interestingly Glassman, Straus and Shogan,find that the content of Member Tweets varies by whether Congress is in session oron recess: unsurprisingly, members refer more frequently to their districts when onrecess.

Politics in the Structure of Online Connections

Twitter is certainly not the only outlet for online communications used by officials,candidates, and the politically interested, and there is a growing literature devotedto other such outlets, like campaign websites and political blogs.

There has been some study of the content of candidate campaign websites (Con-ners, 2005; Trammell et al., 2006) and e-mail communication (Williams and Tram-mell, 2005; Giammo, 2009). It has also been found that increased involvement andhigher incidence of linkages in online social networks like Facebook correlates withelection competitiveness, electoral support, and turnout (Williams and Gulati, 2007;Martin and Schmeisser, 2009; Williams, 2009)

There are also some interesting findings about the nature of the network of po-litical blogs, written by non-elites, or at least, non-politicians. It has been notedthat popular political bloggers are more likely to link to others who share a similarideological viewpoint (Adamic and Glance, 2005; Hargittai, Gallo and Kane, 2008,in a special issue devoted entirely to “Blogs, Politics and Power”), and that blogsare able to exert political influence by framing issues and being read by mainstreamjournalists and other political elites (Farrell and Drezner, 2008, in the same issue).There is even some recent evidence of partisan filtering of content from web sources,suggesting that some “new media” outlets prioritize politics over newsworthiness(Baum and Groeling, 2008).

3

Political Network Connections

The literature on social network analysis of the U.S. Congress is varied and growing.Fowler (2005, 2006b,a) uses co-sponsorship data to construct a network of ties amonglawmakers, identifying the most central and most “connected” legislators. He findsthat this measure of connectedness is predictive of success in passing amendments,which is a common proxy for legislative influence.

Zhang et al. (2008) uses the same co-sponsorship data to identify communitystructure in the House and Senate. Finding the maximum modularity derived by theleading-eigenvector method, it is shown that this measure of modularity tracks withestimates of partisan polarization, a finding supported using roll-call voting networksby Waugh et al. (2009).

There have also been studies using committee (Porter et al., 2005) and caucusmembership (Victor and Ringe, 2009), finding a hierarchical and partisan structurein the network, that supports conventionally held understandings of the social andpolitical organization of the Congress.

Even when looking outside of the institution itself, networks peripheral to Congressevince a high degree of partisan structure. Grossmann and Dominguez (2009) findthat interest groups tend to cluster along partisan and issue lines during the campaign(although less so with respect to behavior during the Congressional session). Kogerand Victor (2009), studying the campaign contributions of lobbyists, and Koger,Masket and Noel (2009), looking at the network created by the transfer of mail-ing list data, identify largely distinct ideological/partisan groupings, even amongstgroups outside of the directive of a formal party organization.

Jakulin et al. (2009) use patterns of similarity in roll-call voting among Senatorsto identify sub-partisan blocs, each of which exhibit more uniform behavior than doparties, as well as more influence.

Ideological Scaling

Ideological scaling is of interest to many Political Scientists, and has been thoroughlyexplored within Congress (among many others, Poole and Rosenthal, 1997; Clinton,Jackman and Rivers, 2004), and the Supreme Court (Martin and Quinn, 2002).

Some attention has recently been paid to ideology within state legislatures (Shor,2008) and among the electorate (Bafumi and Herron, 2007; Shor, 2009), with theintent of scaling all actors on a single ideological continuum. These studies havegenerally found a high degree of bimodality/polarization among elites, with somemoderation amongst the public at large.

4

This paper seeks to extend our knowledge of networks amongst elites, as well asto generate insight into the ideological distribution of the public at large.

Theory and Implications

Twitter users are sending several related messages when they choose to Follow an-other user. Not only are they are suggesting that the Followed individual is worthpaying attention to, but also, that they are the sort of person who would pay atten-tion to such an individual.

People have limited time and attention, so attention is costly. Thus, peoplemust prioritize those sources to which they attend, presumably in a latent utility-maximizing fashion. The specific reasons such choices are made are less importantthan the fact that they are made, and that they vary by individual.

One likely reason to attend to an information source is that one agrees withthe position of that source, and is interested in sharing ideas or new information,or hearing what a more authoritative figure has to say about that shared interest.If the Followed entity is most saliently a part of the political sphere, we can expectthat many of their linkages are based on shared ideology, policy positions, or politicalopinions. Users have a limited amount of attention to spend on their Twitter streamconsumption, and to the extent the utility of such consumption is a function ofconsonance and agreement, they will choose to follow other users who have similarpoints of view, which in the political realm is generally synonymous with ideology.3

Further, Following is a signal that the Followed is worthy of attention. Thus,being listed as a Follower is a vote of confidence in another user, or a show ofsupport. To the extent that one user is opposed to the goals and ideals of another,such a signifier of merit should not be granted. For this reason, we would expectpolitically opposed entities to avoid Following each other, in order to avoid lendingcredibility to an antagonistic cause.

As such, we might see evidence of ideological preferences even even amongsta large population of potentially less-politically-sophisticated Twitter users. Whenaggregated across an entire network, individual inconsistencies can be smoothed over,and patterns of Following can reveal sophisticated behavior on the part of the groupas a whole.

3Other motivations certainly exist: a user may be “spying” on those with whom he or shedisagrees, perhaps in order to take better rhetorical positions in opposition. In such a case, Followerideology would be negatively associated with the ideology of the Followed.

5

Hypotheses

In the political realm, one possible criterion for the allocation of attention is polit-ical ideology. If it is the case that useful information about ideology is encoded inthe structure of network ties, we should be able to find evidence of behavior thatcorrelates with ideology/partisanship, using no information other than that which isto be found in the network structure.

Hypothesis 1 The online network of political elites is distinctly partisan.

If this hypothesis is true, we should see evidence of significantly greater propensityamong political elites to attend to in-party colleagues versus cross-party. I offertwo tests of this hypothesis. The first uses known information about Lawmakers’partisanship and compares the incidence of in-party Following ties to cross-partyties; in-party ties should be substantially more likely. A second test uses a clusteringalgorithm on node linkages to identify two groups within the network of Membersof Congress on Twitter. Assignment one cluster rather than the other should be, ifthis theory is correct, a good predictor of member partisanship.

Hypothesis 2 The allocation of attention by non-elites reflects a partisan dimen-sion.

If this is the case, we would expect user network connections to show a tendencytoward ties to elites exclusively of one party or the other, with relatively few indi-viduals attending equally to elites of both parties. Thus, we would expect a negativerelationship between the counts of Democrats and Republicans being followed byeach individual.

Hypothesis 3 Attention allocation preferences reflect political ideology.

If this hypothesis is true, it should be possible to identify an ideological dimen-sion to observations of non-elites’ Following choices. Evidence for this could befound in the substantive nature of communication from opposite ends of such anempirically-derived dimension; but more crucially, in the process of attempting toexplain variation in Following behavior, the political elites being Followed will them-selves be scaled. Their arrangement on analogous dimensions should correlate withcommonly-accepted measures of ideology.

To verify this hypothesis, I employ three tests. The first serves as a non-statisticalcheck of the substantive differences between individuals located at the extremes of anideological dimension. If the content of their status updates and account biographies

6

does not seem to differ in ideologically meaningful ways, it implies that the dimensionalong which they are arrayed does not actually capture ideology.

A second test uses a cutline which optimally classifies a subset of Democratic andRepublican lawmakers in a two-dimensional space created by the scaling technique.The test is whether this classification criterion correctly predicts partisanship at abetter-than-random rate.

A third, and possibly most exacting, test of whether or not an ideological dimen-sion is recovered is to compare the scaling of Members with other commonly-acceptedcontinuous measures of political ideology, based on observed roll-call behavior. Itmay not be surprising, given a partisan clustering of the network of political elites,that it is possible to identify partisanship. A more difficult challenge is to show thatthe Following behavior of non-elites is structured in such a way as to reflect a con-tinuous liberal-moderate-conservative array of politicians; that is, civilians allocatetheir attention according not just to partisan allegiances, but more specifically, to anideologically-coherent subset of legislators.

Analysis

With a list of Congressional Twitter account users compiled by TweetCongress4,I used the Twitter application programming interface (API) to identify which ofthe Congressional accounts were being Followed by other Congresspersons. Thisgenerated a 133-member adjacency matrix, which can be seen as a directed graph inFigure 1. As the graph indicates, Republicans outnumber Democrats by about 2:1within the Twitter network.

Two major clusters are also evident. The first is a very distinct dense cluster ofRepublican Twitter users, and the second is a less tightly-knit (due to fewer in-groupties) grouping of Democrats. Our interest in ascertaining whether the CongressionalTwitter network is partisan leads us to consider the incidence of in-group and cross-group ties among partisans (Table 1).

The vast majority of Republican ties are to other Republicans, and Democrats areroughly twice as likely to Follow (or be Followed by) other Democrats. The Freeman’s(1972) Segregation Index of this configuration is 0.59, and the log odds ratio of thisobserved pattern of ties is 3.14, which is statistically significant at p < 0.001.

It is possible to refine this analysis by using fuzzy analysis clustering (Kaufmanand Rosseeuw, 1990) on the distance matrix computed from the adjacencies. Thisalgorithm returns probabilities that each observation fits into one of n clusters, where

4http://tweetcongress.org/, all data collected in October, 2009.

7

AnderCrenshaw

AnhJosephCao

ArturDavis

askgeorge

Barbara_Boxer

boblatta

BuckMcKeon

CandiceMillercathymcmorris

cbrangel

chelliepingree

chrislee_ny26ChuckGrassley

clairecmc

CongJoeWilson

CongressmanGT

CynthiaLummis

DanaRohrabacher

danburton

DarrellIssadavereichert

DavidVitter

DeanHeller

Dennis_Kucinich

DennyRehberg

donmanzullo

duncanhunter

EricCantor

Erik_Paulsen

GreggHarper

greshambarrett

IRL

jaredpolis

JeffFlake

JeffFortenberry

Jim_Jordan

JimDeMint

JimGerlach

jiminhofe

JimOberstar

JoeSestakJohnBoozman

JohnCornyn

johnculberson

JohnEnsign

JohnKerry

johnthune

JudyBiggert

keithellison

kendrickbmeek

kevinomccarthy

LamarSmithTX21

LeonardBoswell

MarkeyMemo

markkirk

MarkUdall

MarkWarner

MaryBonoMack

MaryjoKilroy

MelMartinez

michaelcburgess

MicheleBachmann

mlfudge

NancyPelosi

neilabercrombie

OrrinHatch

PatrickMcHenry

PaulHodes

petehoekstra

PeteOlson

PeterRoskam

Randy_Forbes

repaaronschock

repbenraylujan

RepBillShuster

repblumenauer

RepCarnahan

RepDaveCamp

RepGoodlatte

repgregwalden

RepGusBilirakis

RepGwenMoore

RepJoeBarton

repjohnbarrow

RepJohnFleming

RepKenMarchant

RepKevinBrady

RepLynnJenkins

RepMaryFallin

repmikecastle

RepMikeCoffman

RepMikeHonda

RepMikeMichaud

RepMikePence

RepMikeRoss

RepPaulBrounMD

reppaulryan

RepPerlmutter

RepRadanovich

repsandylevin

repschrader

RepShimkus

RepSteveIsraelRepTomPrice

RepTrentFranks

RepWalterJones

RobWittman

rogerwicker

RoyBlunt

russfeingold

SenArlenSpecter

senatorcollins

SenatorMenendez

SenatorReid

SenBillNelson

SenChrisDodd

SenJeffMerkley

SenJohnMcCain

SenSamBrownback

SenSherrodBrown

ShaysCongress

SteveAustria

ThadMcCotter

timryanTomCoburn

TomLatham

tomperriello

tomudall

USRepMikeDoyle

USRepSullivan

VernBuchanan

virginiafoxx

zachwamp

Figure 1: Graph of the Congressional Twitter Following network, color-coded by party. Nodesare scaled according to eigenvector centrality.

8

Dem RepDem 64 28Rep 34 344

Table 1: Incidence of in-group and cross-group ties within the Congressional Twitter Followingnetwork.

in this case, n = 2. These probabilities are plotted in Figure 2, by actual partymembership. The clustering works well for Democratic members, but there is greaterspread in the probabilities for Republicans fitting into a Republican cluster. Thegraph in Figure 3 depicts the Congressional Twitter network, color-scaled accordingto probability of fitting into the Democratic or Republican cluster. The colorationclosely resembles that seen in Figure 1, but many peripheral Republicans are seenas unlikely to be part of their party’s dense central cluster. If we partition nodesaccording to their most-likely cluster, we correctly classify 74.4% of members, whichis a 9% improvement over guessing the modal group for all nodes.

Cluster Assignment Probabilities, by Party

Party

Pro

babi

lity

0.2

0.4

0.6

0.8 —

—

—

—

—————

—

———————————

—

—

—

—

————

—

—

———

—

—

—

——

—

—

—

——

—

——

—

—

—

—

—

——

—

—

—

—

——

—

————

—

—

———

———

—

—

—

———

—

—

—

—

—

——

—

—

——

—

—

—

—

—

—

—

——

—

—

—

—

—

—

—

—————

—

—

—

—

—

—

—

Democrat Republican

Figure 2: Probability of assignment into partisan clusters.

Again with the Twitter API, I collected the IDs of all users who were listed asFollowing any of the Congressional Twitterers. This list contains 117, 837 uniqueusers, each following an average of 2.2 Members of congress. Figure 4 depicts the

9

AnderCrenshaw

AnhJosephCao

ArturDavis

askgeorge

Barbara_Boxer

boblatta

BuckMcKeon

CandiceMiller

cathymcmorris

cbrangel

chelliepingree

chrislee_ny26

ChuckGrassley

clairecmc

CongJoeWilson

CongressmanGT

CynthiaLummis

DanaRohrabacher

danburton

DarrellIssa

davereichert

DavidVitter

DeanHeller

Dennis_Kucinich

DennyRehberg

donmanzullo

duncanhunter

EricCantor

Erik_Paulsen

GreggHarper greshambarrett

IRL

jaredpolis

JeffFlake

JeffFortenberry

Jim_Jordan

JimDeMint

JimGerlach

jiminhofeJimOberstar

JoeSestak

JohnBoozman

JohnCornyn

johnculberson

JohnEnsignJohnKerry

johnthune

JudyBiggert

keithellison

kendrickbmeek

kevinomccarthy

LamarSmithTX21

LeonardBoswell

MarkeyMemo

markkirk

MarkUdall

MarkWarner

MaryBonoMack

MaryjoKilroy

MelMartinez

michaelcburgess

MicheleBachmann

mlfudge

NancyPelosi

neilabercrombie

OrrinHatch

PatrickMcHenry

PaulHodes

petehoekstra

PeteOlson

PeterRoskam

Randy_Forbes

repaaronschock

repbenraylujan

RepBillShusterrepblumenauer

RepCarnahan

RepDaveCamp

RepGoodlatte

repgregwaldenRepGusBilirakis

RepGwenMoore

RepJoeBarton

repjohnbarrow

RepJohnFlemingRepKenMarchant

RepKevinBrady

RepLynnJenkins

RepMaryFallinrepmikecastle

RepMikeCoffman

RepMikeHonda

RepMikeMichaud

RepMikePence

RepMikeRoss

RepPaulBrounMD

reppaulryan

RepPerlmutter

RepRadanovichrepsandylevin

repschraderRepShimkus

RepSteveIsrael

RepTomPrice

RepTrentFranks

RepWalterJonesRobWittman

rogerwicker

RoyBlunt

russfeingold

SenArlenSpecter

senatorcollins

SenatorMenendez

SenatorReid

SenBillNelson

SenChrisDodd

SenJeffMerkley

SenJohnMcCain

SenSamBrownback

SenSherrodBrown

ShaysCongress

SteveAustria

ThadMcCotter

timryan

TomCoburn

TomLatham

tomperriello

tomudall

USRepMikeDoyle

USRepSullivan

VernBuchanan

virginiafoxx

zachwamp

Figure 3: Graph of the Congressional Twitter Following network, color-scaled by probability ofassignment to one of two clusters. Nodes are scaled according to eigenvector centrality.

10

distribution of observed Follow ties from users at large and to Members of Congress.5

As can be seen in the graphic, most users are following just one or a small numberof politicians’ accounts, while the mean number of Followers per legislator is 3, 308.

Log of Followings and Followers

Den

sity

0

2

4

6

8

10

Member Accounts Followed

1 10 100

0.0

0.2

0.4

0.6

0.8

1.0

Member Followers

1000 10000

Figure 4: Distribution of logged number of lawmakers Followed, and logged number of Followersper Member.

It is of interest to compare the number of Republicans and Democrats followedby a given user. If there is a partisan dimension to the decision to Follow or not(rather than merely newsworthiness, name-recognition, or general interest in poli-tics), we would expect a negative relationship between the count of Republicans andDemocrats followed by each user. Figure 5 plots the number of Democratic andRepublican member accounts followed by each user in the sample.

There are several interesting patterns here. First, there is a sizable subset ofFollowers who converge on Following both parties in proportion commensurate withthe number of members from each party who have accounts. These individuals arelikely members of the media, or other political completists, with a high degree ofinterest in the Congress as a whole.

The second pattern in evidence is seen along the axes of the graph, where thereare distinct groups of partisan Followers, who selectively attend only to elites of oneparty or the other. Some of these are the mass majority of users following only asingle member, though others are choosing to follow more than a single account, butonly one party’s perspective.

Thus, it makes sense to account for these different types of users when tryingto understand the partisanship of the Twitter network at large. Table 2 shows the

5Note that the mass at the upper end of the Followers count graph is due to a ceiling on thenumber of usernames per account accessible through the Twitter API.

11

Figure 5: Number of Democratic and Republican member accounts followed, by user.

estimated relationship between increasing Followship of Democratic and RepublicanMembers, controlling for whether or not a user is a “Major Follower,” here defined asthose following more than 10 Congressional accounts, the 90th percentile of users andabove. The finding here is that at relatively low levels of interest in politics in general,users are likely to exhibit a partisan bias in their Following choices, whereas morepolitically-interested users are likely to follow both Republicans and Democrats.

Effect on Number of Republicans FollowedEstimate Std. Error t value Pr(>|t|)

(Intercept) 1.2249 0.0071 173.07 0.0000Dems Followed -0.6745 0.0082 -82.02 0.0000

Major Follower × Dems Followed 0.7805 0.0081 96.22 0.0000

Table 2: Predicting the number of Republican Members Followed as a function of Democratsfollowed, and user type.

To test for evidence of ideology, as distinct from simple partisanship, in thestructure of the Twitter network, I assess the person-to-group network formed bythe adjacency list of all non-politician users Following the set of Congressional Twit-terers. To identify the underlying dimensionality of this network, I employ principalcomponent analysis (PCA)6, treating users as observations and Members of Congress

6Buntine and Jakulin (2004, 2005) develop a discrete multinomial version of probabilistic PCA,but the present work uses classical PCA, per Jolliffe (2004, pg. 68).

12

as binary measurements.For the purposes of this analysis, the components of greatest interest are the first

two, the loadings on which can be seen in Figure 6. The graphic depicts the seconddimension arrayed along the X-axis, as it is the most clearly ideological component.

Using the scores estimated by the PCA for users who Follow politicians, we cando a basic check of verisimilitude by reference to the accounts of some of the moreextreme users. At the lower extreme of the partisan dimension, we see one userwhose account bio reads, “Conservative, Sports, Coach, Politics, Teacher, Blogger,Republican, Catholic, Religion, Baseball,” and another who proclaims “Liberals andRhinos must go! Save the USA; vote Conservative! Our countrys future is at stake!”At the opposite extreme, we encounter a “Mom, wife, blogger, democrat, liberal,progressive,” and “An American, a Democrat, a husband, a dog owner, a NY Yankeesand NY Giants fan, a Catholic and a proud graduate of SUNY Cortland.”7 Thesesubstantive details certainly reinforce the notion that this component correlates withideology.

To further substantiate this claim, I perform linear discriminant analysis, usingthe loadings on the first two components to predict the known partisanship of Mem-bers. The cutline, displayed in Figure 6, correctly identifies partisanship in 94.9% ofcases. Clearly, there is substantial ideological information encoded in these first twodimensions.

As a final test, I am interested in ascertaining whether or not the second principalcomponent correlates well with a widely-accepted measure of ideology in Congress,Lewis and Poole’s Optimal Classification scalings (Lewis and Poole, 2000). Usingin-Chamber rank data for the 111th House and Senate8 as estimates of ideology, this“true” measure correlates with Component 2 at 0.762. As shown in Table 3, infor-mation about Members’ Party affiliation and second component loadings explains81% of the variation in ideology as derived from roll-call voting.

As Table 3 and Figure 7 show, ideology is better explained for Republicans thanDemocrats by the pattern of Twitter Followship, but significant information is en-coded in the network, telling us more about ideology than does partisanship alone.

7http://twitter.com/MattSaliga, http://twitter.com/Goolsby2012,http://twitter.com/angellivingston, and http://twitter.com/tkbva, respectively.

8From http://voteview.org/

13

First Two Principal Components of the Followers Network

Second Component Loading

Firs

t Com

pone

nt L

oadi

ng

−0.15

−0.10

−0.05

0.00

●

●

●

●

●

● ●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●●

●

●

● ●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

AnderCrenshaw

AnhJosephCao

askgeorge

Barbara_Boxer

boblatta

cathymcmorris

cbrangelchelliepingree

chrislee_ny26

ChuckGrassley

clairecmc

CongJoeWilson

danburton

davereichert

donmanzullo

duncanhunter

EricCantor

greshambarrett

IRL

jaredpolis

JeffFlake

JimGerlach

jiminhofe

joebiden

JohnBoozman

JohnCornyn

johnculberson

JohnKerry

johnthune

keithellison

kendrickbmeek

kevinomccarthy

markkirk

MarkUdall

michaelcburgess

MicheleBachmann

mlfudge

PatrickMcHenry

PaulHodes

petehoekstra

Randy_Forbes

RandyKuhlrepaaronschock

repbenraylujan

repblumenauer

RepCarnahan

RepDaveCamp

RepGoodlatte

repgregwalden

RepGwenMoore

RepJohnFleming

RepKenMarchantrepmikecastle

RepMikeCoffman

RepMikeRoss

reppaulryan

repsandylevin

repschrader

RepShimkus

RepWalterJones

rogerwicker RoyBlunt

russfeingold

SenArlenSpecter

senatorcollins

SenatorReid

SenChrisDodd

SenJeffMerkley

SenSamBrownback

SenSherrodBrown

ShaysCongress

ThadMcCottertimryan

TomLatham

tomperriello

tomudall

USRepMikeDoyle

virginiafoxx

zachwamp

−0.2 −0.1 0.0 0.1

Figure 6: Loadings on the first two principal components derived from the person-to-group Fol-lowing network, scaled by number of Followers.

14

Second PCA Component versus Optimal Classification Ranking

Second Component Loading

Opt

imal

Cla

ssifi

catio

n C

ham

ber

Per

cent

ile

0.2

0.4

0.6

0.8

AnderCrenshaw

AnhJosephCao

askgeorge

Barbara_Boxer

boblatta

cathymcmorris

cbrangel

chelliepingree

chrislee_ny26

ChuckGrassley

clairecmc

CongJoeWilsondanburton

davereichert

donmanzullo

duncanhunter

EricCantorgreshambarrett

IRL

jaredpolis

JeffFlake

JimGerlach

jiminhofe

JohnBoozman

JohnCornyn

johnculberson

JohnKerry

johnthune

keithellison

kendrickbmeek

kevinomccarthymarkkirk

MarkUdall

michaelcburgess

MicheleBachmann

mlfudge

PatrickMcHenry

PaulHodes

petehoekstra

Randy_Forbes

repaaronschock

repbenraylujan

repblumenauer

RepCarnahan

RepDaveCamp

RepGoodlatte

repgregwalden

RepGwenMoore

RepJohnFleming

RepKenMarchant

repmikecastle

RepMikeCoffman

RepMikeRoss

reppaulryan

repsandylevin

repschrader

RepShimkus

RepWalterJones

rogerwicker

RoyBlunt

russfeingold

SenArlenSpecter

senatorcollins

SenatorReid

SenChrisDodd

SenJeffMerkley

SenSamBrownback

SenSherrodBrown

ThadMcCotter

timryan

TomLatham

tomperriello

tomudall

USRepMikeDoyle

virginiafoxx

zachwamp

−0.2 −0.1 0.0 0.1

Figure 7: Comparison of adjacency-matrix-derived scaling and optimal classification rankings.Ideology is better explained by the Twitter Following network for Republicans than for Democrats.

15

Effect on In-Chamber Optimal Classification PercentileEstimate Std. Error t value Pr(>|t|)

(Intercept) 0.7640 0.0251 30.48 0.0000Party -0.5299 0.0610 -8.69 0.0000

Component 2 0.6534 0.2952 2.21 0.0300Party × Component 2 -0.8026 0.4738 -1.69 0.0946

R2 = 0.8056

Table 3: Predicting optimal classification rankings with ideology as derived from the TwitterCongressional Following network.

Conclusions

I have shown that the network of Congresspersons on Twitter is distinctly parti-san, as are the choices made by the vast majority of Twitter users, those who wewould not characterize as highly-politically interested. Furthermore, I have shownthat using only the structure of the person-to-group network formed by individualsFollowing political figures, it is possible to identify an ideological dimension latentin the Following decision. In fact, information about the individuals following thestatus updates of a given legislator offers better predictions about that legislators’“true” ideology than does party alone. The implications of these findings are many,as are the opportunities for extensions of this analysis.

The first implication, not explored here, but a direct extension of the methodsused above, is that any entity on Twitter can be arrayed along the same ideologicaldimension. This may work in one of two ways: Additional entities, such as membersof the media or the executive branch, are included in the principal componentsanalysis, and used along with legislators in the scaling of civilians. Another methodmay be to use the estimates of civilian placements on the ideological dimensionderived above to identify, for each entity of interest, a mean and variance of Followerideological placement.

The most interesting potential use for such an approach might be to place chal-lengers for legislative office on the same scale as incumbent elected officials, or toarray, on the same scale, multiple entities from disparate bodies, such as across statelegislatures. Applications like these would take advantage of the fact that the Twitternetwork connects a wide variety of potentially interesting actors. Currently, to scalelegislators from different chambers in a common space, “bridge actors” who haveserved in more than one body are used–but this generally assumes that ideology isfixed over time, which may not be a valid assumption, given changing constituencies.

16

Additionally, some potentially interesting political elites, such as electoral challengerswho have not previously held office, cannot be scaled using techniques that rely onroll-call votes. Networks such as the Twitter Following network would permit acomparison of ideological extremism of these heretofore unquantifiable politicians,allowing us to answer questions about, for example, whether polarization may becaused by the success of ideologically extreme candidates in party primaries.

Just as easily, the Following patterns of those identified as liberal or conservative,Democratic- or Republican-leaning, users could be contrasted–do these politically-different groups attend to different media sources or Follow the status of differentcelebrities? Are certain corporations more closely identifiable on Twitter with oneparty than the other? Though we would not likely expect many Members of Congressto publicly follow certain entities, with whose reputation an electorally-accountablepolitician may not want to be associated, their proximate presence within the overallnetwork may allow us to make connections that they themselves would not explicitlysignal.

Another theoretically interesting use for this Twitter network data takes advan-tage of the expressed pattern of diffusion of content within the network. ManyTwitter users “re-tweet” some of the more personally compelling or interesting sta-tus messages of those they follow, and each such message is time-stamped. Tracingthe paths of these retweets would allow us a look at the diffusion of ideas and agen-das amongst political elites, and from elites to their civilian followers. A directednetwork of influence could be constructed, giving insight into a hierarchy of agendaentrepreneurs and opinion leaders, and an interesting look at whether such com-munication is top-down, bottom-up, or bi-directional between elected officials andinterested constituents.

These more dynamic questions do not even speak to the litany of more descrip-tive aspects of the Twitter network, such as, What types of communication do wesee between politicians and constituents? What are the correlates of high levels ofMember-constituent interaction online? Does the content of communication changemeaningfully during the campaign? Why do Republican legislators so greatly out-number their Democratic counterparts on Twitter?

One final conclusion suggested by this analysis is that while Twitter and sim-ilar online social networks, may seem ephemeral and superficial, they are not un-intelligent. Though online attention may be distributed in a multitude of idiosyn-cratic ways, the collective decisions of a large number of even moderately-politically-interested individuals reveal a sophisticated ideological discernment.

17

References

Adamic, Lada A. and Natalie Glance. 2005. The political blogosphere and the 2004U.S. election: divided they blog. In LinkKDD ’05: Proceedings of the 3rd interna-tional workshop on Link discovery. New York, NY, USA: ACM pp. 36–43.

Bafumi, Joseph and Michael C. Herron. 2007. “Preference Aggregation, Representa-tion, and Elected American Political Institutions.” [Under Review].

Baum, Matthew A. and Tim Groeling. 2008. “New Media and the Polarization ofAmerican Political Discourse.” Political Communication 25:345–365.

Beam, Christopher. 2009. “The Twitter Holdouts: Why is more than half of Congressstill not on Twitter?”.URL: http://www.slate.com/id/2237243/pagenum/all/

Buntine, Wray and Aleks Jakulin. 2004. Applying discrete PCA in data analysis.In UAI ’04: Proceedings of the 20th conference on Uncertainty in artificial intel-ligence. Arlington, Virginia, United States: AUAI Press pp. 59–66.

Buntine, Wray and Aleks Jakulin. 2005. Discrete Principal Component Analysis.Technical report Helsinki Institute for Information Technology Helsinki, Finland:.

Clinton, Joshua, Simon Jackman and Doug Rivers. 2004. “The Statistical Analysis ofRoll Call Data: A Unified Approach.” American Political Science Review 98:355–370.

Conners, Joan. 2005. Meetup, Blogs, and Online Involvement: U.S. Senate CampaignWebsites of 2004. Annual Meeting of the American Political Science Association.

Farrell, Henry and Daniel W. Drezner. 2008. “The power and politics of blogs.”Public Choice 134:15–30.

Fenno, Richard F. 1978. Home Style: House Members in Their Districts. Boston:Little, Brown.

Fowler, James. 2005. Who is the Best Connected Legislator? A Study of Cospon-sorship Networks. American Political Science Association.

Fowler, James H. 2006a. “Connecting the Congress: A Study of CosponsorshipNetworks.” Political Analysis 14(4):456–487.

18

Fowler, James H. 2006b. “Legislative cosponsorship networks in the US House andSenate.” Social Networks 28(4):454 – 465.

Freeman, Linton C. 1972. “Segregation in Social Networks.” Sociological Methods &Research 6:411–30.

Giammo, Joseph. 2009. When Candidates Attack: Inter- and Intra-Party Criticismsin Campaign E-mails in the 2008 Presidential Nominating Contests. MidwestPolitical Science Association 67th Annual National Conference.

Glassman, Matthew Eric, Jacob R. Straus and Colleen J. Shogan. 2009. “SocialNetworking and Constituent Communication: Member Use of Twitter During aTwo-Week Period in the 111th Congress.” Congressional Research Service .

Golbeck, Jennifer, Justin Grimes and Anthony Rogers. 2009. “Twitter Use by theU.S. Congress.” working paper.

Grossmann, Matt and Casey B. K. Dominguez. 2009. “Party Coalitions and InterestGroup Networks.” American Politics Research 37(5):767–800.

Hargittai, Eszter, Jason Gallo and Matthew Kane. 2008. “Cross-ideological discus-sions among conservative and liberal bloggers.” Public Choice 134:67–86.

Huberman, Bernardo A., Daniel M. Romero and Fang Wu. 2008. “Social Networksthat Matter: Twitter Under the Microscope.” NBER Working Paper .

Jakulin, Aleks, Wray Buntine, Timothy M. La Pira and Holly Brasher. 2009. “Ana-lyzing the U.S. Senate in 2003: Similarities, Clusters, and Blocs.” Political Analysis17(3):291–310.

Java, Ashkay, Xiaodan Song, Tim Finin and Belle Tseng. 2009. Why We Twitter:An Analysis of a Microblogging Community. In Advances in Web Mining and WebUsage Analysis. Berlin: Springer pp. 118–138.

Jolliffe, Ian T. 2004. Principal Component Analysis. Springer Series in Statisticssecond ed. New York: Springer.

Kaufman, Leonard and Peter J. Rosseeuw. 1990. Finding Groups in Data: AnIntroduction to Cluster Analysis. New York: Wiley.

Kingdon, John W. 1967. “Politicians’ Beliefs About Voters.” The American PoliticalScience Review 61(1):137–145.

19

Koger, Gregory and Jennifer Nicoll Victor. 2009. “Polarized Agents: CampaignContributions by Lobbyists.” PS: Political Science & Politics 42(03):485–488.

Koger, Gregory, Seth Masket and Hans Noel. 2009. “Partisan Webs: InformationExchange and Party Networks.” British Journal of Political Science 39(03):633–653.

Krishnamurthy, Balachander, Phillipa Gill and Martin Arlitt. 2008. A few chirpsabout twitter. In WOSP ’08: Proceedings of the first workshop on Online socialnetworks. New York, NY: ACM pp. 19–24.

Lewis, Jeff and Keith Poole. 2000. “Non-Parametric Unfolding of Binary ChoiceData.” Political Analysis 8:211–237.

Martin, Andrew D. and Kevin M. Quinn. 2002. “Dynamic Ideal Point Estimationvia Markov Chain Monte Carlo for the U.S. Supreme Court, 1953-1999.” PoliticalAnalysis 10(2):134–153.

Martin, Kimberly. and Hans Schmeisser. 2009. Emerging Trends in Youth VotingBehavior: Social Networking Websites and the 2008 General Election. SouthernPolitical Science Association.

Mayhew, David R. 1974. Congress: The Electoral Connection. New Haven: YaleUniversity Press.

Poole, Keith T. and Howard Rosenthal. 1997. Congress: A Political-Economic His-tory of Roll Call Voting. New York: Oxford University Press.

Porter, Mason A., Peter J. Mucha, M. E. J. Newman and Casey M. Warmbrand.2005. “A network analysis of committees in the U.S. House of Representatives.”Proceedings of the National Academy of Sciences of the United States of America102(20):7057–7062.

Shor, Boris. 2008. Stronger Bridges: Putting Congress and State Legislatures inCommon Ideological Space. Working Papers 0814 Harris School of Public PolicyStudies, University of Chicago.

Shor, Boris. 2009. All Together Now: Putting Congress, State Legislatures, andIndividuals on a Common Ideological Space. Midwest Political Science AssociationChicago, IL: .

20

Sifry, Micah L. 2009. “Who Has the Most Twitter Klout in Congress? (The AnswerWill Surprise You).” techPresident .URL: http://techpresident.com/blog-entry/who-has-most-twitter-klout-congress-answer-will-surprise-you

Sniderman, Paul M. 1993. The New Look in Public Opinion Research. In The Stateof The Discipline II, ed. Ada Finifter. Washington, D.C.: The American PoliticalScience Association.

Trammell, Kaye D., Andrew Paul Williams, Monica Postelnicu and Kristen D. Lan-dreville. 2006. “Evolution of Online Campaigning: Increasing Interactivity inCandidate Web Sites and Blogs Through Text and Technical Features.” MassCommunication & Society 9(1):21–44.

Victor, Jennifer Nicoll and Nils Ringe. 2009. “The Social Utility of Informal In-stitutions: Caucuses as Networks in the 110th U.S. House of Representatives.”American Politics Research 37(5):742–766.URL: http://apr.sagepub.com/cgi/content/abstract/37/5/742

Waugh, Andrew S., Liuyi Pei, James H. Fowler, Peter J. Mucha and Mason A.Porter. 2009. “Party Polarization in Congress: A Social Networks Approach.”SSRN eLibrary .

Williams, Andrew Paul and Kaye D. Trammell. 2005. “Candidate Campaign E-Mail Messages in the Presidential Election 2004.” American Behavioral Scientist49(4):560–574.URL: http://abs.sagepub.com/cgi/content/abstract/49/4/560

Williams, Christine. 2009. What Is a Social Network Worth? Facebook and VoteShare in the 2008 Presidential Primaries. American Political Science AssociationBoston, MA: .

Williams, Christine and Girish Gulati. 2007. Social Networks in Political Campaigns:Facebook and the 2006 Midterm Elections. American Political Science AssociationChicago, IL: .

Zhang, Yan, A.J. Friend, Amanda L. Traud, Mason A. Porter, James H. Fowlerand Peter J. Mucha. 2008. “Community structure in Congressional cosponsorshipnetworks.” Physica A: Statistical Mechanics and its Applications 387(7):1705 –1712.

21

Zimmerman, Eric and Christina Wilkie. 2009. “Wilson hires professional Tweeter.”The Hill .URL: http://thehill.com/homenews/house/58351-wilson-hires-a-pro-to-tweet

22