Viet-An Nguyen, Jordan Boyd-Graber, Philip Resnik, and Kristina Miler. Tea Party in the House: A HierarchicalIdeal Point Topic Model and Its Application to Republican Legislators in the 112th Congress. Association forComputational Linguistics, 2015, 11 pages.

@inproceedings{Nguyen:Boyd-Graber:Resnik:Miler-2015,Author = {Viet-An Nguyen and Jordan Boyd-Graber and Philip Resnik and Kristina Miler},Url = {docs/2015_acl_teaparty.pdf},Booktitle = {Association for Computational Linguistics},Location = {Beijing, China},Year = {2015},Title = {Tea Party in the House: A Hierarchical Ideal Point Topic Model and Its Application to Republican Legislators in the 112th Congress},}

Links:

• Talk [http://techtalks.tv/talks/tea-party-in-the-house-a-hierarchical-ideal-pointtopic-model-and-its-application-to-republican-legislators-in-the-112th-congress/61794/]

• Code [https://github.com/vietansegan/herbal/]• LaTeX [https://github.com/Pinafore/publications/tree/master/2015_acl_teaparty]

Downloaded from http://cs.colorado.edu/~jbg/docs/2015_acl_teaparty.pdf

1

Tea Party in the House: A Hierarchical Ideal Point Topic Modeland Its Application to Republican Legislators in the 112th Congress

Viet-An NguyenComputer Science

University of MarylandCollege Park, MDvietan@cs.umd.edu

Jordan Boyd-GraberComputer Science

University of ColoradoBoulder, CO

Jordan.Boyd.Graber

@colorado.edu

Philip ResnikLinguistics & UMIACSUniversity of Maryland

College Park, MDresnik@umd.edu

Kristina MilerGovernment and PoliticsUniversity of Maryland

College Park, MDkmiler@umd.edu

Abstract

We introduce the Hierarchical Ideal PointTopic Model, which provides a rich pictureof policy issues, framing, and voting behav-ior using a joint model of votes, bill text,and the language that legislators use whendebating bills. We use this model to look atthe relationship between Tea Party Repub-licans and “establishment” Republicans inthe U.S. House of Representatives duringthe 112th Congress.

1 Capturing Political PolarizationIdeal-point models are one of the most widely

used tools in contemporary political science re-search (Poole and Rosenthal, 2007). These modelsestimate political preferences for legislators, knownas their ideal points, from binary data such as leg-islative votes. Popular formulations analyze legis-lators’ votes and place them on a one-dimensionalscale, most often interpreted as an ideological spec-trum from liberal to conservative.

Moving beyond a single dimension is attractive,however, since people may lean differently basedon policy issues; for example, the conservativemovement in the U.S. includes fiscal conservativeswho are relatively liberal on social issues, and viceversa. In multi-dimensional ideal point models,therefore, the ideal point of each legislator is nolonger characterized by a single number, but by amulti-dimensional vector. With that move comes anew challenge, though: the additional dimensionsare often difficult to interpret. To mitigate thisproblem, recent research has introduced methodsthat estimate multi-dimensional ideal points usingboth voting data and the texts of the bills beingvoted on, e.g., using topic models and associatingeach dimension of the ideal point space with a topic.The words most strongly associated with the topiccan sometimes provide a readable description of itscorresponding dimension.

In this paper, we develop this idea further byintroducing HIPTM, the Hierarchical Ideal PointTopic Model, to estimate multi-dimensional idealpoints for legislators in the U.S. Congress. HIPTM

differs from previous models in three ways. First,HIPTM uses not only votes and associated bill text,but also the language of the legislators themselves;this allows predictions of ideal points from politi-cians’ writing alone. Second, HIPTM improvesthe interpretability of ideal-point dimensions byincorporating data from the Congressional BillsProject (Adler and Wilkerson, 2015), in which billsare labeled with major topics from the Policy Agen-das Project Topic Codebook.1 And third, HIPTM

discovers a hierarchy of topics, allowing us to ana-lyze both agenda issues and issue-specific framesthat legislators use on the congressional floor, fol-lowing Nguyen et al. (2013) in modeling framingas second-level agenda setting (McCombs, 2005).

Using this new model, we focus on Republicanlegislators during the 112th U.S. Congress, fromJanuary 2011 until January 2013. This is a par-ticularly interesting session of Congress for politi-cal scientists, because of the rise of the Tea Party,a decentralized political movement with populist,libertarian, and conservative elements. Althoughunited with “establishment” Republicans againstDemocrats in the 2010 midterm elections, lead-ing to massive Democratic defeats, the Tea Partywas—and still is—wrestling with establishmentRepublicans for control of the Republican party.

The Tea Party is a new and complex phe-nomenon for political scientists; as Carmines andD’Amico (2015) observe: “Conventional views ofideology as a single-dimensional, left-right spec-trum experience great difficulty in understandingor explaining the Tea Party.” Our model identifieslegislators who have low (or high) levels of “TeaPartiness” but are (or are not) members of the TeaParty Caucus, and providing insights into the na-

1http://www.policyagendas.org/

ture of polarization within the Republican party.HIPTM also makes it possible to investigate a num-ber of questions of interest to political scientists.For example, are there Republicans who identifythemselves as members of the Tea Party, but whosevotes and language betray a lack of enthusiasm forTea Party issues? How well can we predict fromsomeone’s language alone whether they are likelyto associate themselves with the Tea Party? Ourcomputational modeling approach to “Tea Parti-ness”, distinct from self-declared Tea Party Caucusmembership, may have particular value in under-standing Republican party politics going forwardbecause, despite the continued influence of the TeaParty, the official Tea Party Caucus in the Houseof Representatives is largely inactive and its futureuncertain (Fuller, 2015).

2 Polarization across DimensionsIdeal point models describe probabilistic rela-

tionships between observed responses (votes) ona set of items (bills) by a set of responders (legis-lators) who are characterized by continuous latenttraits (Fox, 2010). A popular formulation posits anideal point ua for each lawmaker a, a polarity xb,and popularity yb for each bill b, all being valuesin (−∞,+∞) (Martin and Quinn, 2002; Bafumiet al., 2005; Gerrish and Blei, 2011). Lawmaker avotes “Yes” on bill b with probability

p(va,b = Yes |ua, xb, yb) = Φ(uaxb + yb) (1)

where Φ(α) = exp(α)/(1 + exp(α)) is the logis-tic (or inverse-logit) function.2 Intuitively, mostlawmakers vote “Yes” on bills with high popularityyb and “No” on bills with low yb. When a bill’spopularity is lower, the outcome of the vote va,bdepends more on the interaction between the law-maker’s ideal point ua and the bill’s polarity xb.

Multi-dimensional ideal point models replacescalars ua and xb with K-dimensional vectors uaand xb (Heckman and Jr., 1997; Jackman, 2001;Clinton et al., 2004). Unfortunately, as Lauderdaleand Clark (2014) observe, the binary data used forthese models are “insufficiently informative to sup-port analyses beyond one or two dimensions”, andthe additional dimensions are difficult to interpret.To address this lack of interpretability, recent workhas proposed multi-dimensional ideal point modelsto jointly capture both binary votes and the associ-

2A probit function is also often used where Φ(α) is insteadthe cumulative distribution function of a Gaussian distribu-tion (Martin and Quinn, 2002).

ated text (Gerrish and Blei, 2012; Gu et al., 2014;Lauderdale and Clark, 2014; Sim et al., 2015).

3 Hierarchical Ideal Point Topic ModelBringing topic models (Blei and Lafferty, 2009)

into ideal-point modeling provides an interpretable,text-based foundation for political scientists to un-derstand why the models make the predictions theydo. However, both the topic—what is discussed—and the framing—how it is discussed—also revealpolitical preferences. We therefore introduce frame-specific ideal points, using a hierarchy of topics tomodel issues and their issue-specific frames. Al-though the definition of “frame” is itself a movingtarget in political science (Entman, 1993), we adoptthe theoretically motivated but pragmatic approachof Nguyen et al. (2013): just as agenda-issues mapnaturally to topics in probabilistic topic models(e.g., Grimmer (2010)), the frames as second-levelagenda-setting (McCombs, 2005) map to second-level topics in a hierarchical topic model.

Our model’s inputs are votes {va,b}, each theresponse of legislator a ∈ [1, A] to bill b ∈ [1, B].Two types of text supplement the votes: floorspeeches (documents) {wd} from legislator ad, andthe text w′b of bill b. While congressional debatesare typically about one piece of legislation, wemake no assumptions about the mapping betweenwd and w′b. In principle this allows wd to be anytext by legislator ad (e.g., not just floor speechesabout this bill, but blogs, social media, press re-leases) and—unlike Gerrish and Blei (2011)—thispermits us to make predictions about individualseven without vote data for them. Figure 1 showsthe plate notation diagram of HIPTM, which has thefollowing generative process:

1. For each issue k ∈ [1,K]

(a) Draw k’s associated topic φk ∼ Dir(β, φ?k)

(b) Draw issue-specific distribution over framesψk ∼ GEM(λ0)

(c) For each frame j ∈ [1,∞) (specific to issue k)i. Draw j’s associated topic φk,j ∼ Dir(β, φk)

ii. Draw j’s regression weight ηk,j ∼ N (0, γ)

2. For each document d ∈ [1, D] by legislator ad(a) Draw topic (i.e., issue) distribution θd ∼ Dir(α)(b) For each issue k ∈ [1,K], draw frame distribu-

tion ψd,k ∼ DP(λ, ψk)(c) For each token n ∈ [1, Nd]

i. Draw an issue zd,n ∼ Mult(θd)ii. Draw a frame td,n ∼ Mult(ψd,zd,n)

iii. Draw word wd,n ∼ Mult(φzd,n,td,n)

3. For each legislator a ∈ [1, A] on each issue k ∈ [1,K]

(a) Draw issue-specific ideal point ua,k ∼N (∑Jk

j=1 ψa,k,jηk,j , ρ) weighting ηk,j by howmuch the legislator talks about that frame

���

��

���

��

�

�

��

�����

�

��∗

���

�

����

����

�� ���

��

�

���

�

�

��

��

��

�

�

�

!

"��

#��

�

Figure 1: Plate notation diagram of HIPTM.

Agriculture: food; agriculture; loan; farm; crop; dairy;rural; conserve; commodity; eligible; farmer; margin; milk;contract; nutrition; livestock; plantHealth: drug; medicine; coverage; disease; public health;hospital; social security; health insurrance; patient; appli-cation; treatment; payment; physician; nurse; clinicLabor, Employment, and Immigration: employment;immigration; labor; paragraph; eligible; status; compen-sation; application; wage; homeland security; unemploy-ment; board; violation; file; perform; mine

Table 1: Examples of informed priors φ?k for issues.

4. For each bill b ∈ [1, B]

(a) Draw polarity xb ∼ N (0, σ)(b) Draw popularity yb ∼ N (0, σ)(c) Draw topic (i.e., issue) proportions ϑb ∼ Dir(α)(d) For each token m ∈ [1,Mb] in the text of bill b

i. Draw an issue z′b,m ∼ Mult(ϑb)ii. Draw a word type w′b,m ∼ Mult(φz′

b,m)

5. For each vote va,b of legislator a on bill b

(a) p(va,b |ua, xb, yb, ϑb) = Φ(xb∑

k ϑb,kua,k + yb)

Topic Hierarchy. With the goal of analyzingagendas and frames in mind, our topic hierarchyhas two levels: (1) issue nodes and (2) frame nodes.(Look ahead to Figure 6 for an illustration.) Morespecifically, there are K issue nodes, each with atopic φk drawn from a Dirichlet distribution withconcentration parameter β and a prior mean vectorφ?k, i.e., φk ∼ Dir(β, φ?k). In this hierarchical struc-ture, first-level nodes map to agenda issues, whichwe treat as non-polarized, and second-level nodesmap to issue-specific frames, which we assumepolarize on the issue-specific dimension.3

To improve topic interpretability, issue nodeshave an informed prior from the CongressionalBills Project {φ?k} (Table 1).4 The frame topic φk,j

3Nguyen et al. (2013) allow first-level nodes to polarizebut find first-level nodes are typically neutral.

4The Congressional Bills Project provides a large collec-tion of labeled congressional bill text. We compute {φ?

k} as

at each frame node is a Dirichlet draw centered atthe corresponding (parent) issue node. While thenumber of issues is fixed a priori, the number ofsecond-level frames is unbounded. We also asso-ciate each second-level frame node with an idealpoint ηk,j ∼ N (0, γ). This resembles how su-pervised topic models (Blei and McAuliffe, 2007;Nguyen et al., 2015) discover polarized topics’ as-sociated response variables.

Generating Text from Legislators. One of ourmodel’s goal is to study how legislators frame pol-icy agenda issues. To achieve that, we analyzecongressional speeches (documents) {wd}, eachof which is delivered by a legislator ad. To gen-erate each token wd,n of a speech d, legislator adwill (1) first choose an issue zd,n ∈ [1,K] from adocument-specific multinomial θd, then (2) choosea frame td,n from the set of infinitely many possi-ble frames of the given issue zd,n using the frameproportion ψd,k drawn from a Dirichlet process,and finally (3) choose a word type from the cho-sen frame’s topic φzd,n,td,n . In other words, ourmodel generates speeches using a mixture of KHDPs (Teh et al., 2006).5

Generating Bill Text. The bill text provides in-formation about the policy agenda issues that eachbill addresses. We use LDA to model the bill text{w′b}. Each bill b is a mixture ϑb over K is-sues, which is drawn from a symmetric Dirich-let prior, i.e., ϑb ∼ Dir(α). Each token w′b,min bill b is generated by first choosing a topicz′b,m ∼ Mult(ϑb), and then choosing a word typew′b,m ∼ Mult(φz′b,m), as in LDA.

Generating Roll Call Votes. Following recentwork on multi-dimensional ideal points (Laud-erdale and Clark, 2014; Sim et al., 2015), we definethe probability of legislator a voting “Yes” on billb as p(va,b = Yes |ua, xb, yb, ϑb) =

Φ

(xb

K∑

k=1

ϑb,kua,k + yb

)(2)

where ϑb is the empirical distribution of bill b overthe K issues and is defined as ϑb,k =

Mb,k

Mb,·. Here,

Mb,k is the number of times in which tokens in b

the empirical word distribution from all bills labeled with k.K = 19, corresponding to 19 major topic headings in thePolicy Agendas Project Topic Codebook.

5If we abandoned the labeled data from the CongressionalBills Project to obtain the prior means φ?

k, it would be rela-tively straightforward to extend to a fully nonparametric modelwith unbounded K (Ahmed et al., 2013; Paisley et al., 2014).

are assigned to issue k and Mb,· is the marginalcount, i.e., the number of tokens in bill b.

The ideal point of legislator a specifically on is-sue k is ua,k and comes from a normal distribution

N (ψTa,kηk, ρ) ≡ N

Jk∑

j=1

ψa,k,jηk,j , ρ

(3)

where Jk is the number of frames for topic k, whichis unbounded. The mean of the Normal distributionis a linear combination of the ideal points {ηk,j} ofall issue k’s frames, weighted by how much timelegislator a spends on each frame when talkingabout issue k, i.e., ψa,k,j =

Na,k,j

Na,k,·. Here, Na,k,j

is the number of tokens authored by a that areassigned to frame j of issue k, and Na,k,· is themarginal count. When Na,k,· = 0, which meansthat legislator a does not talk about issue k, weback off to an uninformed zero mean.

Equation 3 resembles how supervised topic mod-els (SLDA) link topics with a response, in that theresponse—the issue-specific ideal point ua,k—islatent. It is similar to how Gerrish and Blei (2011)use the bill text to regress on the bill’s latent polar-ity xb and popularity yb. In this paper, we only usetext from congressional speeches for regression, asthese can capture how legislators frame specific top-ics. Incorporating the bill text into the regression isan interesting direction for future work.

4 InferenceGiven observed data of (1) votes {va,b} by A

legislators on B bills, (2) speeches {wd} from leg-islators, and (3) bill text {w′b}, we estimate thelatent variables using stochastic EM. In each itera-tion, we perform the following steps: (1) samplingissue assignments {z′b,m} for bill text tokens, (2)sampling the issue assignments {zd,n} and frameassignments {td,n} for speech tokens, (3) samplingthe topics at first-level issue nodes {φk}, (4) sam-pling the distribution over frames {ψk} for all is-sues, (5) optimizing frames’ regression parameters{ηk,j} using L-BFGS (Liu and Nocedal, 1989), and(6) updating legislators’ ideal points {ua,k} andbills’ polarity {xb} and popularity {yb} using gra-dient ascent.

Sampling Issue Assignments for Bill TokensThe probability of assigning a token w′b,m in thebill text to an issue k is

p(z′b,m = k | rest) ∝M−b,mb,k + α

M−b,mb,· +Kα· φk,w′b,m (4)

where Mb,k denotes the number of tokens in billtext b assigned to issue k. The current estimatedprobability of word type v given issue k is φk,v(Equation 7). Marginal counts are denoted by · andthe superscript −b,m excludes the assignment fortoken w′b,m from the corresponding count.

Sampling Frame Assignments in Speeches Tosample the assignments for tokens in the speeches,we first sample an issue using

p(zd,n = k | rest) ∝N−d,nd,k + α

N−d,nd,· +Kα· φk,wd,n

(5)

where Nd,k similarly denotes the number of timesthat tokens in d are assigned to issue k. Given thesampled issue k, we sample the frame asp(td,n = j | zd,n = k, ad = a, rest) ∝

N (ua,k;µa,k,j , ρ) ·

(N−d,n

d,k,j

N−d,nd,k,j +λ

+λ·ψk,j

N−d,nd,k,j +λ

),

N (ua,k;µa,k,jnew , ρ) · λ

N−d,nd,k,j +λ

· ψk,jnew ,

(6)where µa,k,j = (

∑Jkj′=1 ηk,j′N

−d,nd,k,j′ + ηk,j)/Nd,k,·

for an existing frame j, and for a newlycreated frame jnew, we have µa,k,jnew =

(∑Jk

j′=1 ηk,j′N−d,nd,k,j′+ηk,jnew)/Nd,k,·, where ηk,jnew

is drawn from the Gaussian prior N (0, γ). Here,the estimated global probability of choosing aframe j of issue k is ψk,j .

Sampling Issue Topics In the generative processof HIPTM, the topic φk of issue k (1) generatestokens in the bill text and (2) provides the Dirichletpriors of the issue’s frames. Rather than collapsingmultinomials and factorizing (Hu and Boyd-Graber,2012), we follow Ahmed et al. (2013) and sample

φk ∼ Dir(mk + nk + βφ?k) (7)

wheremk ≡ (Mk,1,Mk,2, · · · ,Mk,V ) is the tokencount vector from the bill text assigned to eachissue. The vector nk ≡ (Nk,1, Nk,2, · · · , Nk,V )denotes the token counts propagated from wordsassigned to topics that are associated with frames ofissue k, approximated using minimal or maximalpath assumptions (Cowans, 2006; Wallach, 2008).

Sampling Frame Proportions Following the di-rect assignment method described in Teh et al.(2006), we sample the global frame proportion asψk ≡ (ψk,1, ψk,2, · · · , ψk,jnew)

∼ Dir(N·,k,1, N·,k,2, · · · , N·,k,Jk , λ0) (8)

where N·,k,j =∑D

d=1 Nd,k,j and Nd,k,j can besampled effectively using the Antoniak distribu-tion (Antoniak, 1974).

Optimizing Frame Regression ParametersWe update the regression parameters ηk of framesunder issue k using L-BFGS (Liu and Nocedal,1989) to optimize L(ηk)

− 1

2ρ

A∑

a=1

(ua,k − ηTk ψa,k)−1

2γ

Jk∑

j=1

η2k,j (9)

Updating Ideal Points, Polarity and PopularityWe update the multi-dimensional ideal point ua ofeach legislator a and the polarity xb and popularityyb of each bill b by optimizing the log likelihoodusing gradient ascent.

5 Data CollectionWhat makes a Tea Partier? To address that ques-

tion, we use key votes identified by Freedom Worksas the most important votes on issues of economicfreedom. Led by former House Majority LeaderDick Armey (R-TX), Freedom Works is a con-servative non-profit organization which promotes“Lower Taxes, Less Government, More Freedom”.6

Karpowitz et al. (2011) report that Freedom Worksendorsements are more effective than other TeaParty organizations at getting out votes for Repub-lican candidates in the 2010 midterms.

For the 112th Congress, Freedom Works selected60 key votes, 40 in 2011 and 20 in 2012. We areinterested in ideal points with respect to the TeaParty movement, i.e., on the anti-pro Tea Partydimension: whether a legislator agrees with Free-dom Works on a bill. More specifically, we assignva,b to be 1 if legislator a agrees with FreedomWorks on bill b, and 0 otherwise. In addition tothe votes, we obtained the bill text with labels fromthe Congressional Bills Project7 and the congres-sional speeches from GovTrack.8 In total, we have240 Republicans, 60 who self-identify with the TeaParty Caucus, and 13,856 votes.

6 Predicting Tea Party MembershipTo quantitatively evaluate the effectiveness of

HIPTM in capturing “Tea Partiness”, we predict TeaParty Caucus membership of legislators given theirvotes and text. This examines (1) how effectivethe baseline features extracted from the votes andtext are in predicting the Caucus membership, and(2) how much prediction improves using featuresextracted from HIPTM. For baselines, we consider

6http://congress.freedomworks.org/7http://congressionalbills.org/8https://www.govtrack.us/data/us/112/

TFTF−IDF

VoteHIPTM

Vote−TFVote−TF−IDFVote−HIPTM

All

0.60 0.65 0.70 0.75AUC−ROC

Figure 2: Tea Party Caucus membership prediction resultsover five folds using AUC-ROC (higher is better, random base-line achieves 0.5). The features extracted from our model areestimated using both the votes and the text.

simple feature sets where each legislator is repre-sented by their speeches as either (1) a TF-IDFvector (Salton, 1968), (2) a normalized TF-IDFvector, or (3) a binary vector containing their votes.

Our dataset for binary prediction comprises a setof 60 Republican representatives who self-identifyas Tea Party Caucus members and 180 who donot. These are divided using 5-fold cross-validationwith stratified sampling, which preserves the ratioof the two classes in both the training and test sets.We report performance using AUC-ROC (Lusted,1971) using SVMlight (Joachims, 1999).9 After pre-processing, our vocabulary contains 5,349 uniqueword types.

Membership from Votes and Text. First, giventhe votes and text of all the legislators, we runHIPTM for 1,000 iterations with a burn-in period of500 iterations. After burning in, we keep the sam-pled state of the model after every fifty iterations.The feature values are obtained by averaging overthe ten stored models as suggested in Nguyen et al.(2014). Each legislator a is represented by a vectorconcatenating:

• K-dimensional ideal point vector estimatedfrom both votes and text ua,k• K-dimensional vector, estimating the idealpoint using only text ηTk ψa,k• B probabilities estimating a’s votes on B billsΦ(xb

∑Kk=1 ϑb,kua,k + yb)

Figure 2 shows AUC-ROC results for our featuresets. VOTE-based features clearly outperform text-based features like TF and TF-IDF. CombiningVOTE with either TF or TF-IDF does not improvethe prediction performance much (i.e., VOTE-TFand VOTE-TF-IDF). Features extracted from our

9We use the default settings of SVMlight, except that weset the cost-factor equal to the ratio between the number ofnegative examples (i.e., number of non-Tea Party Caucusmembers) and the number of positive examples (i.e., numberof Tea Party Caucus members).

TF

TF−IDF

HIPTM

0.600 0.625 0.650AUC−ROC

Figure 3: Tea Party Caucus membership prediction resultsover five folds using AUC-ROC (higher is better, random base-line achieves 0.5). The features extracted from our model forunseen legislators are estimated using their text only.

model, HIPTM, also outperform TF and TF-IDFsignificantly, but only slightly better than VOTE.However, HIPTM and VOTE together significantlyoutperform VOTE alone.

Membership Prediction from Text Only. Theresults in Figure 2 require both votes and legisla-tors’ language. This is limiting, since it permitspredictions of “Tea Partiness” only for people withan established congressional voting record. A po-tentially more interesting and practical task is pre-diction based on language alone.

Thus, we first run our inference algorithm onthe training data, which includes both votes andtext. After training, using multiple models, wesample the issue and frame assignments for eachtoken of the text authored by test lawmakers.Since the votes are not available, HIPTM’s ex-tracted features here only consist of (1) the K-dimensional vector ηTk ψa,k estimating legislators’ideal point using text alone, and (2) the B proba-bilities Φ(xb

∑Kk=1 ϑb,kua,k + yb) estimating the

votes.Figure 3 compares this approach with the two

baselines capable of using text alone, TF and TF-IDF. Since HIPTM can no longer access the votesin the test data, its performance drops significantlycompared with VOTE. However, it still quitestrongly outperforms the two text-based baselines,showing that jointly modeling the voting behaviorimproves the text-based elements of the model.

7 How the Tea Party VotesIn this section, we examine legislators’ ideal

points. We first expose Tea Party-specific idealpoints by examining one-dimensional ideal pointsand then move on to the issue-specific ideal pointsthat HIPTM enables.

7.1 One-dimensional Ideal PointsFirst, as a baseline, we estimate the one-

dimensional ideal points of the legislators in our

−1 0 1Ideal Point

Tea Party Caucus Member Non−member

Figure 4: Box plots of the one-dimensional Tea Party idealpoints, estimated as a baseline in Section 7.1, for members andnon-members of the Tea Party Caucus among Republican Rep-resentatives in the 112th U.S. House. The median of members’ideal points is significantly higher than that of non-members’.

dataset.10 Figure 4 shows the box plots of esti-mated Tea Party ideal points for both members andnon-members.11 The Tea Party ideal points cor-relate with DW-NOMINATE (ρ = 0.91), and themedian ideal point of Tea Party Caucus membersis higher than non-members. This confirms thatTea Partiers are more conservative than other Re-publicans (Williamson et al., 2011; Karpowitz etal., 2011; Gervais and Morris, 2012; Gervais andMorris, 2014).

Divergences involving these ideal points helpdemonstrate the face validity of our approach. Forexample, the model gives Jeff Flake (R-AZ) the sec-ond highest ideal point; he only disagrees with Free-dom Works position on one of 60 Freedom Workskey votes, but he is not a member of the Tea PartyCaucus. Another example is Justin Amash (R-MI),who founded and is the Chairman of the LibertyCaucus. Its members are conservative and liber-tarian Republicans, and Amash has agreed withFreedom Works on every single key vote selectedby Freedom Works since 2011.

Conversely, some self-identified Tea Partiers of-ten disagree with Freedom Works and thus haverelatively low ideal points. For example, RodneyAlexander (R-LA) only agrees with Freedom Works48% of the time, and was a Democrat before 2004.Alexander and Ander Crenshaw (R-FL, 50% agree-ment) are categorized as “Green Tea” by Gervaisand Morris (2014), i.e. Republican legislators whoare “associated with the Tea Party on their owninitiative” but lack support from Tea Party organi-zations.

10We use gradient ascent to optimize the likelihood of voteswhose probabilities are defined in Equation 1. We also put aGaussian priorN (0, σ) on ua, xb, and yb.

11Estimated ideal point signs might be flipped, as uaxb =(−ua)(−xb), which makes no difference in Equation 1. Toensure that higher ideal points are “pro-Tea Party”, we firstsort the legislators according to the fraction of votes for whichthey agree with Freedom Works and initialize the ideal pointsof the top and bottom five legislators with +3σ and -3σ, whereσ is the variance of ua’s Gaussian prior.

●●

● ●

● ●

● ●

●

● ●

● ●●

● ●● ●

●● ●● ●

●

●● ●●● ●● ●●● ●

●

●

● ●● ●● ●●

●

●

● ●

●●

●

Civil Rights, Minority Issues, and Civil LibertiesAgriculture

HealthPublic Lands and Water Management

Community Development and Housing IssuesSpace, Science, Technology and Communications

Law, Crime, and Family IssuesInternational Affairs and Foreign Aid

EnergyEducation

Social WelfareEnvironment

DefenseBanking, Finance, and Domestic Commerce

Foreign TradeLabor, Employment, and Immigration

TransportationMacroeconomics

Government Operations

−6 −3 0 3Ideal PointsTea Party Caucus Member Non−member

Figure 5: Box plots of ideal points dimensions, each corresponding to a major topic in the Policy Agendas Topics Codebookestimated by our model. On most issues the ideal point distributions over the two Republican groups (member vs. non-memberof the Tea Party Caucus) overlap. The most polarized issues are Government Operations and Macroeconomics, which align wellwith the agenda of the Tea Party movement supporting small government and lower taxes.

7.2 Multi-dimensional Ideal PointsWhile it is interesting to compare holistic mea-

sures of Tea Partiness, it doesn’t reveal how leg-islators conform or deviate from what defines amainstream Tea Partier. In this section, HIPTM

reveals how issue-specific ideal points of the twogroups of Republican representatives differ.

Figure 5 shows estimated ideal points for eachpolicy agenda issue, sorted by the difference be-tween the median of the two groups’ ideal points.On most issues, the ideal point distributions of thetwo Republican groups are nigh identical.

On several issues, though, the ideal point distri-butions of the two groups of legislators diverge.In the remainder of this section, we considerthe Government Operation, Macroeconomics, andTransportation topics, and look at why HIPTM esti-mates these issues as the most polarized.

Government operations Tea Partiers differfrom their Republican colleagues on reducing gov-ernment spending on the Economic DevelopmentAdministration, the Energy Efficiency and Renew-able Energy Program and Fossil Fuel Research andDevelopment. More specifically, for example, onthe key vote to eliminate the Energy Efficiency andRenewable Energy Program, nearly 80% (41 outof 53) of Tea Partiers vote “Yea” (with FreedomWorks) but only 43% of non-Tea Partiers agree.

Macroeconomics Our model estimatesMacroeconomics policies as being the sec-ond most polarizing topic for House Republicans,which is consistent with the emphasis that the TeaParty places on issues like a balanced budget andreduced federal spending. Indeed, we see thatTea Party Republicans have distinct preferenceson these types of issues as compared to more

mainstream Republican legislators. An illustrationof this polarization can be seen in the intra-partyfight over the budget. Roll call vote 275 in 2011and roll call vote 149 in 2012 both would havereplaced Paul Ryan’s budget (the “establishment”Republican budget) with the Republican StudyCommittee’s (RSC) “Back to Basics” budget thatwould cut spending more aggressively and balancethe budget in a decade. In 2011, non-Tea PartyRepublicans were evenly split in their budgetpreferences, but three quarters of the Tea PartyCaucus supported it, which illustrates the differ-ence between the two factions of the Republicanparty. Similarly, in 2012, more than 80% of TeaPartiers voted for the the RSC budget, but fewerthan half of non-Tea Party Republicans did. Otherpolarizing votes in the Macroeconomics topicinclude votes to raise the debt ceiling and to avertthe “fiscal cliff”. In these cases, support for thesevotes was 25 percentage points higher among TeaPartiers than non-Tea Party Republicans, whichagain illustrates their distinct policy preferences.

Transportation Transportation is the third mostpolarized issue estimated by our model, with twokey votes focusing on federal spending on trans-portation that illustrate some polarization, but alsosome shared preferences among Republicans. Con-sistent with the Tea Party’s emphasis on reducinggovernment spending, Tea Party Republicans voteddifferently from their non-Tea Party colleagues onthese issues. The first key vote, roll call vote 378 in2012, caps highway spending at the amount takenin by the gas tax. More than half of Tea Party Cau-cus members (32 out of 55) voted in favor, whilenon-members voted against it by a greater than 2:1margin (122 of 172). Conversely, the second keyvote (roll call vote 451 in 2012) authorizes fed-

Macroeconomicsbalanc_budget, borrow, debt_ceil, cap, cut_spend, nation_debt, grandchildren, rais_tax, entitl, white_hous, debt_limit, prosper

Frame M1white_hous, shut,

continu_resolut, mess, hous_republican, novemb,

govern_shutdown, senat_reid, harri_reid,

vision, shutdown, liber, arriv, republican_parti, blame

Frame M2balanc_budget, debt_ceil, cap, cut_spend, debt_limit,

spend_cut, fiscal_hous, grandchildren, guarante, default, august, obama,

deficit_spend, rein, feder_budget

Frame M3borrow, nation_debt,

rais_tax, entitl, prosper, chart, grandchildren,

spend_monei, size, gdp, tax_increas, cent,

govern_spend, social_secur

-.57 -.24.56

Healthobamacar, patient, doctor, physician, afford_care, hospit, insur, replac, mandat, exchang, health_insur, coverag, medicaid, patient_protect, board

Frame H1afford_care, exchang, patient_protect, human_servic, public_health, slush_fund, ppaca, mandatori, mandatori_spend, governor, hospit, health_center, flexibl, teach_health, unlimit

Frame H2patient, doctor, physician, hospit, medicaid, board, georgia, save_medicar, nurs, tennesse, page, bureaucrat, advisori_board, medicin, independ_payment

Frame H3obamacar, replac, mandat, insur, health_insur, coverag, social_secur, premium, repeal_obamacar, entitl, govern_takeov, purchas, unconstitut, preexist_condit, employ

-.13 .04.56

Establishment Tea Party

Figure 6: Framing of Macroeconomics (top) and Health (bot-tom) among House Republicans, 2011-2012. Higher idealpoint values are associated with the Tea Party.

eral highway spending at a level that far exceedsits revenue from the gas tax, which was opposedby Freedom Works. This measure was broadlypopular with Republicans regardless of Tea Partyaffiliation and a majority of both Tea Partiers andnon-Tea Partiers opposed it.

8 How the Tea Party TalksLooking at HIPTM’s induced topic hierarchy, us-

ing labeled data to create informative priors pro-duces highly interpretable topics at the agenda-issue level; e.g., see the first-level nodes in Figure 6,which capture key issue-level debates. For exam-ple, one major event during the 112th Congress wasthe 2011 debt-ceiling crisis, which dominates dis-cussions in Macroeconomics. Similarly, Defenseis dominated by withdrawing U.S. troops from Iraq.

Turning to framing, recall that second-levelnodes of the hierarchy capture issue-specific framesof parent issues, each one associated with a frame-specific ideal point. To analyze intra-Republicanpolarization, we first compute, for each issue k, thespan of ideal points the frames associated with k,i.e., the difference between the maximum and theminimum ideal points for frames under that issue.12

We then consider several issues with a large span,i.e. whose frames are highly polarized.

Macroeconomics. The HIPTM subtree forMacroeconomics, in Figure 6 (top), foregrounds

12The frame proportions Dirichlet process ψk creates manyframes with one or two observations (Miller and Harrison,2013). We ignore those with posterior probability ψk,j < 0.1.

Republican polarization related to budget issues.The most Tea Party oriented frame node, M3,focuses on criticizing government overspending,a recurring Tea Party theme.13 In contrast,Frame M1, least oriented toward the Tea Party,focuses on the downsides of a governmentshutdown, highlighting establishment Republicanconcerns about being held responsible for thepolitical and economic consequences.

Health. Healthcare was a central issue duringthe 112th Congress, particularly the AffordableCare Act (Obamacare). Although all Republicansvoted to repeal Obamacare, Figure 6 (bottom) high-lights intra-party differences in framing the issue.Frame H1 leans strongest toward the establishmentRepublican end of the spectrum, and frames op-position in terms of the implementation of healthcare exchanges and the mandatory costs of the pro-gram. In contrast, H3 captures the more stridentTea Party framing of Obamacare as an unconstitu-tional government takeover. More neutral froman intra-party perspective, Frame H2 emphasizesMedicare, Medicaid, and the role of health careprofessionals within these systems.14

Labor, Employment and Immigration. Thediscussion of this issue illustrates how HIPTM some-times captures frames that are distinct from TeaPartiness, per se. For example, it discovered astrongly Tea Party oriented frame that focused on“union, south carolina, nlrb, boeing”. On inspec-tion, this frame reflects a controversy in which theNational Labor Relations Board accused airlinemanufacturer Boeing of violating Federal labor lawby transferring production to a non-union facilityin South Carolina “for discriminatory reasons”,15

and surfaces mainly in speeches by four legislatorsfrom South Carolina, three of whom are from theTea Party Caucus. This second-level topic illus-trates a limitation of HIPTM; it does not formallydistinguish frames from other kinds of subtopics.We observe that modeling polarization on otherkinds of sub-issues is nonetheless valuable: hereit highlights a geographic locus of conflict involv-

13E.g., Scott Garrett (R-NJ): “We will not compromise onour principles; our principles of defending the Constitutionand defending Americans and making sure that our posteritydoes not have this excessive debt on it.”

14This does not mean that discussions using this framelacked combative or partisan elements. For example, GlennThompson (R-PA) argues that “on the Democratic side, they’rejust willing to pull the plug and let [Medicare] die”.

15http://www.nlrb.gov/news-outreach/fact-sheets/fact-sheet-archives/boeing-complaint-fact-sheet

Ideal PointDistributions

Dist

ribu

tion

of

Issu

e Fr

ames

PolarizedNot

Pola

rized

Not Civil Rights, Minority

Issues, Civil LibertiesBanking and Finance;

Transportation

Health; Public Lands and Water Management

Macroeconomics; Government Operations

Table 2: Examples of agenda issues classified by polarizationof ideal points and issue frames within the Republican party.

ing South Carolina, where many representativesare Tea Party Caucus members. This may provideinsight into how geography shapes Tea Party mem-bership (Gervais and Morris, 2012).

9 Latent and Visible Disagreement

Our analyses suggest a novel framework for un-derstanding the political and policymaking impli-cations of the Tea Party in the 112th Congress, il-lustrated in Table 2. Each issue can be character-ized by two features: (1) the degree to which idealpoints among Republican legislators are polarized,and (2) the degree to which the frames used are po-larized. From these two assessments, we can orga-nize all policy issues into four categories that havemeaningful implications for congressional politicsand policy outcomes. At upper left we will findissues where HIPTM indicates low intra-party po-larization between Tea Party and non-Tea Party,and all Republicans tend to frame the issue in sim-ilar ways; e.g., Civil Rights, Minority Issues, andCivil Liberties. In such cases, we expect cooper-ation among Republicans regardless of Tea Partystatus, therefore a greater likelihood of bill pas-sage in a majority-Republican House. In stark con-trast, issues at lower right involve polarized idealpoints and polarized framing, e.g., the budget cri-sis, where many establishment Republicans balkedat a government shutdown but hard-line Tea Partylegislators did not. These issues pose the greatestchallenge to Republican party leaders.

Between these extremes are the issues in whicheither Republicans’ ideal points or their policyframes are polarized. Our model suggests that onissues at upper right, with similar framing, the TeaParty and establishment Republicans will appearto be in sync, and therefore it may seem to votersthat legislative progress is likely, but the under-lying issue polarization will make it hard to findpolicy common ground, potentially increasing pub-lic frustration. Last, at lower left are issues where

Republicans generally share similar ideal pointsand vote similarly, but frame the issue in distinctways, e.g., Obamacare. Here legislative successmay come despite the appearance that Republicanfactions are talking past each other, because thedistribution of their ideal points on the policy isactually quite similar. Put differently, Republicansshare policy goals on issues in this quadrant evenif they frame those preferences differently, and thisunderlying agreement on the ideal point may al-low Republicans to reach consensus even when thepolitical rhetoric suggests otherwise.

10 Conclusion

We introduce HIPTM, which integrates hierarchi-cal topic modeling with multi-dimensional idealpoints to jointly model voting behavior, the textcontent of bills, and the language used by legisla-tors. HIPTM is more effective than previous meth-ods on the task of predicting membership in theTea Party Caucus. This improvement is especiallyconsequential as the formal organization of the TeaParty Caucus is now defunct in the House, yet TeaParty legislators remain both numerous and influen-tial in Congress. In addition, unlike previous ideal-point methods, HIPTM makes it possible to makepredictions for members of Congress who have notyet established a voting record. More intriguingly,this also suggests the possibility of assessing the“Tea Partiness” of candidates (or, anyone else, e.g.,media outlets) based on language.

It is political conventional wisdom that the in-flux of Tea Party legislators in the 112th Congresscomplicated the task of governance and policymak-ing for Republican leaders. By looking at issue-level ideal points and issue-specific framing usingour model, we begin to address the complexityof this relationship, finding the model successfulboth in establishing face validity and in suggestingnovel insights into the dynamics of a RepublicanCongress. In future work, we plan to pursue thenew framework suggested by our analyses, inves-tigating the interaction of issue polarization andframing-based polarization. With the help of thesenew tools, we aim to both understand and predictsubstantive policy areas in which the Tea Party islikely to be most successful working with the Re-publican party, and, conversely, to flag ahead oftime policy areas in which we can expect to seelegislative gridlock and grandstanding.

AcknowledgementsThis work was supported by the NSF Grant IIS-

1211153 and IIS-1409287. Boyd-Graber is alsosupported by NSF grants IIS-1320538 and NCSE-1422492. Any opinions, findings, conclusions, orrecommendations expressed here are those of theauthors and do not necessarily reflect the view ofthe sponsor.

ReferencesE. Scott Adler and John Wilkerson. 2015. Congressional bills

project. NSF 00880066 and 00880061.

Amr Ahmed, Liangjie Hong, and Alexander Smola. 2013.Nested Chinese restaurant franchise process: Applicationsto user tracking and document modeling. In Proceedingsof the International Conference of Machine Learning.

Charles E. Antoniak. 1974. Mixtures of Dirichlet processeswith applications to Bayesian nonparametric problems.The Annals of Statistics, 2(6):1152–1174.

Joseph Bafumi, Andrew Gelman, David K Park, and NoahKaplan. 2005. Practical issues in implementing and under-standing Bayesian ideal point estimation. Political Analy-sis, 13(2):171–187.

David M. Blei and John Lafferty, 2009. Text Mining: The-ory and Applications, chapter Topic Models. Taylor andFrancis, London.

David M Blei and Jon D McAuliffe. 2007. Supervised topicmodels. In Proceedings of Advances in Neural InformationProcessing Systems, pages 121–128.

Edward G Carmines and Nicholas J D’Amico. 2015. Thenew look in political ideology research. Annual Review ofPolitical Science, 18(4).

Joshua Clinton, Simon Jackman, and Douglas Rivers. 2004.The statistical analysis of roll call data. American PoliticalScience Review, 98(02):355–370.

Philip J Cowans. 2006. Probabilistic Document Modelling.Ph.D. thesis, University of Cambridge.

Robert M Entman. 1993. Framing: Toward clarification of afractured paradigm. Journal of Communication, 43(4):51–58.

Jean-Paul Fox. 2010. Bayesian item response modeling:Theory and applications. Springer.

Matt Fuller. 2015. New Tea Party Caucus chairman:DHS fight could break the GOP. Roll Call, Febru-ary. http://blogs.rollcall.com/218/new-tea-party-caucus-chairman-dhs-fight-could-break-the-gop/.

Sean Gerrish and David M. Blei. 2011. Predicting legislativeroll calls from text. In Proceedings of the InternationalConference of Machine Learning.

Sean Gerrish and David M. Blei. 2012. How they vote: Issue-adjusted models of legislative behavior. In Proceedings ofAdvances in Neural Information Processing Systems.

Bryan T Gervais and Irwin L Morris. 2012. Reading the tealeaves: Understanding Tea Party Caucus membership inthe US House of Representatives. PS: Political Science &Politics, 45(02):245–250.

Bryan T Gervais and Irwin L Morris. 2014. Black Tea,Green Tea, White Tea, and Coffee: Understanding thevariation in attachment to the Tea Party among membersof Congress. In Annual Meeting of the American PoliticalScience Association.

Justin Grimmer. 2010. A Bayesian hierarchical topic modelfor political texts: Measuring expressed agendas in Senatepress releases. Political Analysis, 18(1):1–35.

Yupeng Gu, Yizhou Sun, Ning Jiang, Bingyu Wang, and TingChen. 2014. Topic-factorized ideal point estimation modelfor legislative voting network. In Proceedings of Inter-national Conference on Knowledge Discovery and DataMining, pages 183–192.

James J. Heckman and James M. Snyder Jr. 1997. Linearprobability models of the demand for attributes with anempirical application to estimating the preferences of legis-lators. The RAND Journal of Economics, 28:142–189.

Yuening Hu and Jordan Boyd-Graber. 2012. Efficient tree-based topic modeling. In Association for ComputationalLinguistics.

Simon Jackman. 2001. Multidimensional analysis of rollcall data via Bayesian simulation: Identification, estima-tion, inference, and model checking. Political Analysis,9(3):227–241.

Thorsten Joachims. 1999. Making large-scale SVM learningpractical. In Advances in Kernel Methods - SVM. Univer-sitat Dortmund.

Christopher F Karpowitz, J Quin Monson, Kelly D Patterson,and Jeremy C Pope. 2011. Tea time in America? Theimpact of the Tea Party movement on the 2010 midtermelections. PS: Political Science & Politics, 44(02):303–309.

Benjamin E. Lauderdale and Tom S. Clark. 2014. Scalingpolitically meaningful dimensions using texts and votes.American Journal of Political Science, 58(3):754–771.

Dong C Liu and Jorge Nocedal. 1989. On the limited memoryBFGS method for large scale optimization. Mathematicalprogramming, 45(1-3):503–528.

Lee B. Lusted. 1971. Signal Detectability and MedicalDecision-Making. Science, 171:1217–1219, March.

Andrew D Martin and Kevin M Quinn. 2002. Dynamic idealpoint estimation via Markov chain Monte Carlo for the USSupreme Court, 1953–1999. Political Analysis, 10(2):134–153.

Maxwell McCombs. 2005. A look at agenda-setting: Past,present and future. Journalism Studies, 6(4):543–557.

Jeffrey W Miller and Matthew T Harrison. 2013. A simpleexample of dirichlet process mixture inconsistency for thenumber of components. In Proceedings of Advances inNeural Information Processing Systems, pages 199–206.

Viet-An Nguyen, Jordan Boyd-Graber, and Philip Resnik.2013. Lexical and hierarchical topic regression. In Pro-ceedings of Advances in Neural Information ProcessingSystems, pages 1106–1114.

Viet-An Nguyen, Jordan Boyd-Graber, and Philip Resnik.2014. Sometimes average is best: The importance of aver-aging for prediction using MCMC inference in topic mod-eling. In Proceedings of Emperical Methods in NaturalLanguage Processing, pages 1752–1757.

Thang Nguyen, Jordan Boyd-Graber, Jeff Lund, Kevin Seppi,and Eric Ringger. 2015. Is your anchor going up or down?Fast and accurate supervised topic models. In North Amer-ican Association for Computational Linguistics.

John Paisley, Chong Wang, David M Blei, and Michael I Jor-dan. 2014. Nested hierarchical Dirichlet processes. IEEETransactions on Pattern Analysis and Machine Intelligence.

Keith T Poole and Howard L Rosenthal. 2007. Ideology andCongress. New Brunswick, NJ: Transaction Publishers.

Gerard. Salton. 1968. Automatic Information Organizationand Retrieval. McGraw Hill Text.

Yanchuan Sim, Bryan Routledge, and Noah A Smith. 2015.The utility of text: The case of Amicus briefs and theSupreme Court. In Association for the Advancement ofArtificial Intelligence.

Yee Whye Teh, Michael I Jordan, Matthew J Beal, andDavid M Blei. 2006. Hierarchical Dirichlet processes.Journal of the American Statistical Association, 101(476).

Hanna M Wallach. 2008. Structured Topic Models for Lan-guage. Ph.D. thesis, University of Cambridge.

Vanessa Williamson, Theda Skocpol, and John Coggin. 2011.The Tea Party and the remaking of Republican conser-vatism. Perspectives on Politics, 9(01):25–43.