American inequality

American Political Science Review (2019) 113, 4, 917–940

doi:10.1017/S0003055419000315 © American Political Science Association 2019

The Party or the Purse? Unequal Representation in the US Senate JEFFREY R. LAX Columbia University

JUSTIN H. PHILLIPS Columbia University

ADAM ZELIZER The University of Chicago

RecentworkonUSpolicymakingargues that responsiveness to public opinion is distorted bymoney, in that the preferences of the rich matter much more than those of lower-income Americans. A second distortion—partisan biases in responsiveness—has been less well studied and is often

ignored or downplayed in the literature on affluent influence. We are the first to evaluate, in tandem, these two potential distortions in representation. We do so using 49 Senate roll-call votes from 2001 to 2015. We find that affluent influence is overstated and itself contingent on partisanship—party trumps the pursewhen senators have to take sides. The poor get what theywantmore often fromDemocrats. The rich get what they wantmoreoften fromRepublicans, butonly ifRepublicanconstituents sidewith the rich.Thus,partisanship induces, shapes, and constrains affluent influence.

INTRODUCTION

The new stylized fact of American politics is that the wealthy dominate American democracy. A growing body of political science research con-

cludes that government policy is far more responsive to the preferences of the affluent than to thoseof either the middle class or poor (Bartels 2008; Ellis 2012, 2017; Gilens 2005, 2012; Gilens and Page 2014; Hayes 2013; Rigby andWright 2011; Tausanovitch 2016).Gilens, for example, concludes that “the preferences of the vast majority of Americans appear to have essentially no impact onwhich policies the government does or doesn’t adopt” (2012, 1). Such class-based distortion violates norms of equal voice and,worse still, raises the spectre of a vicious cycle in which low-income individuals are lockedoutofpower, inwhicheconomic inequalitybegets political inequality which begets still more economic inequality. This is the warning of the “economic elite domination model” (Gilens and Page 2014).

Claims of class-based inequality in responsiveness have not gone unchallenged. Some studies find evi- dence against the basic result (e.g., Bhatti and Erikson 2011; Soroka and Wlezien 2008; Ura and Ellis 2008; Wlezien and Soroka 2011). Others argue that the implications of unequal responsiveness are overstated.

For example, Enns (2015a, 2015b) notes that low- and middle-income individuals receive a great deal of co- incidental representation even when politicians respond primarily to the affluent, because preferences tend to differ little by income group. Branham, Soroka, and Wlezien(2017)findthat the ideological impactofaffluent influence is attenuated because well-to-do constituents have a mix of both liberal and conservative preferences. This important research skeptical of affluent influence has perhaps received less attention (scholarly or other- wise) than the work of Bartels or Gilens, and there re- main unresolved debates regarding the nature, pervasiveness, and substance of such influence (see generally Erikson 2015).

Independent of these debates, there is also a growing concern that elected officials care “toomuch” about the opinions of their copartisans or electoral base, creating a partisan distortion in representation (Clinton 2006; Kastellec et al. 2015; Shapiro et al. 1990; Warshaw 2012). This bias can arise from a variety of factors, most prominently the need towin primary elections (Clausen 1973; Fenno 1978; Gerber and Morton 1998). Such a bias can pull policy away from the relativelymoderate preferences of the median voter, toward the more ideologically extreme preferences of partisans. As with the economic distortion, a partisan distortion in rep- resentation violates norms of equal voice and can also become reinforced and entrenched (in this case through the manipulation of electoral rules, gerrymandering, and the like).This partisandistortion is also increasingly invokedas conventionalwisdom,but rigorous empirical investigations of partisan biases in representation re- main (at least relative to studies of affluent influence) both uncommon and somewhat limited in scope.

As work in this vein continues, scholars are finding evidence of other partisan distortions. One is that parties might not behave symmetrically with respect to constituent opinion. Some research suggests that Democratic and Republican lawmakers do not put the same weight on opinion (Clinton 2006; Krimmel, Lax, and Phillips 2016). Another is these lawmakers may discount opinion to vote the “party line” and thereby further their party’s legislative agenda (Hussey and Zaller 2011), which itself may not reflect public opinion

Jeffrey R. Lax , Professor, Department of Political Science, Co- lumbia University, [email protected].

Justin H. Phillips , Professor, Department of Political Science, Columbia University, [email protected].

Adam Zelizer , Assistant Professor, Harris School of Public Policy, The University of Chicago, [email protected].

This project was funded in part by grant from the Russell Sage Foundation (G-6789, “WhoListens toWhom?Assessing Inequalities in Representation”). We thank Andrew Guess and Michael Malecki for previous work with us on uncertainty in opinion estimation. We thank participants at Columbia University, Wake Forest University, and the American and Midwest Political Science Association Con- ferences for helpful feedback and discussion. We thank the anony- mous reviewers for their careful reading, advice, and patience. Replication files are available at the American Political Science Re- view Dataverse: https://doi.org/10.7910/DVN/MCWFCS.

Received:April 24, 2018; revised:December 4, 2018; accepted:May 3, 2019. First published online: July 12, 2019.

917

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

(Achen and Bartels 2016). As with economic dis- tortions, there remain significant unresolved debates regarding the scope and strength of partisan distortions.

Perhaps surprisingly, research on class-based and partisan distortions in representation has proceeded on two, largely separate, tracks. We argue, however, that existing debates cannot be resolved without simulta- neously considering both types of distortions. A failure todo so risks over- or underestimating theprevalenceof each and obscures interactions between them. For ex- ample, copartisan pull might constrain affluent in- fluence or vice versa. Alternatively, partisanship might be the vehicle through which affluent influence oper- ates. What might appear to be over-responsiveness to the preferences of the rich could simply be coincidental, if copartisan opinion and the rich opinion tend to agree. Moreover, potential distortions in representation may manifest differently in the behavior of Democratic and Republican lawmakers.

We juxtapose and integrate economic and partisan biases in the study of representation, thereby assessing whether these forces are indeed coincidental, comple- mentary, constraining, or conflicting. Our efforts si- multaneously serve as both the first study of economic distortions to foreground partisan opinion and the first study of partisan distortions to foreground affluent opinion.

Our analysis is the most revealing when potential influences on lawmakers conflict, and elected repre- sentatives have to take sides. What happens when party and purse pull in opposite directions, when a law- maker’s copartisan constituents and rich constituents want different things? What happens when copartisans side with rich against poor or vice versa?What happens when there is intraparty conflict between rich and poor opinion? Taking up the complications noted by Krim- mel, Lax, and Phillips (2016) and others, do the two parties respond to opinion and dealwith cross pressures the same way? Do pressures to vote the party line lead lawmakers to ignore the preferences of key constituents altogether?

We build on significant work by others, confronting many of the same challenges they did, but utilizing different solutions.These solutions requireknowing the preferences on specific issues, not only of the rich and poor but also of partisan groups and of income groups within each party. We need a sufficient number and variety of roll call votes, not only for generalizability but also so that we have sufficient instances of subcon- stituency disagreement to disentangle competing influences.

We obtain all this using a large quantity of survey data, along with the most recent advances in subgroup opinion estimation. Our dyadic analysis of represen- tation in theUS Senate uses 49 roll call votes from eight sessions of Congress (2001–15). The votes we utilize include some of the most important economic, social, and foreign policy votes cast by members of Congress during this period (e.g., health-care reform, President Obama’s stimulus bill, an extension of the Bush tax cuts on capital gains, the Federal Marriage Amendment, and a vote to withdraw American military personnel

from Iraq). We use and extend multilevel regression and poststratification (MRP) to create the necessary estimates of public opinion for partisan and class sub- groups and incorporate uncertainty around our opinion estimates.

Our baseline “taking sides” analyses reproduce the foundational findings of the economic and partisan distortion literatures, but using different and more recent evidence. We find that the affluent are more likely than the poor to get what they want, especially wheneachgroupdesires adifferent policy (not that this is common). And, we find that lawmakers more fre- quently vote in a manner that is consistent with the preferences of their home-state copartisans than with their median constituent. We also uncover, consistent with Krimmel, Lax, and Phillips (2016), broad evi- dence of asymmetric responsiveness, with Democratic senators farmore responsive toopinion in general than their Republican colleagues. However, when we consider partisan and elite opinion in tandem, our results depart from the conventional wisdom, espe- cially in regards to affluent influence. Our results show that affluent influence is largely a story of partisan politics. It both works through and is limited by partisanship.1

Republican senators are, on average, more re- sponsive to the rich than the poor, but Democratic senators are largely more responsive to the poor than rich, particularly when there is class conflict. Thus, it is Republican senators, notDemocrats, who are primarily responsible for the overall pattern of affluent influence. This does not, however, mean that Democrats are en- tirely “innocent.” In a small subset of votes in which well-to-do Democrats prefer a different policy than poor or middle-class Democrats, Democratic senators are somewhatmore likely to side with the affluent. That said, in that type of situation, Republican senators side far more with the Republican rich over Republican poor.

Howdoes this revised senseof affluent influence stand up against partisanship directly? It does not—party trumps the purse. Senators of both parties are far more responsive to copartisan opinion than rich opinion. When the two conflict, senators of both parties tend to side overwhelminglywith their copartisans over the rich.

This partisan effect not only sharply limits affluent influence but also seems to largely account for its ex- istence in the first place. Republican copartisan opinion is more likely to align with the opinions of the rich than with those of the poor (whereas Democratic copartisan opinion is more likely to align with poor). So when Republicans vote in the manner preferred by their copartisan constituents back home, these senators are also providing coincidental representation to the af- fluent more than the poor. When Democrats listen to their partisan constituents, coincidental representation favors the poor over rich.

1 Telling the story of representation in terms of the rich and not party would be like telling how Harry Potter defeated Voldemort, talking only about Ron.

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

918

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

There is, as noted above, yet another partisan con- straint on affluent influence. Pressure to vote the party line seems to trump responsiveness to opinion of any type. EvenRepublicans—who tend to side with the rich over poor, and Republican constituents over rich—will side with their fellow Republican senators against rich, against Republicans constituents, or both combined.

In sum, our analyses yield a fundamentally different understanding of the democratic deficit in legislative representation—and of affluent influence.

THINKING ABOUT REPRESENTATION

Economic Distortions of Representation

The seminal contributions on affluent influence are Bartels (2008) andGilens (2005, 2012). Bartels studied the roll call voting behavior of individual senators, using as dependent variables the overall ideological tenor of a senator’s voting record (measured using the W-Nominate scores of Poole and Rosenthal 1997) as well as individual votes on eight bills, half of which addressed abortion. He compared a senator’s roll call voting behavior with the self-reported ideology of her high-, middle-, and low-income constituents. For the abortion roll call votes, however, he departed from this strategy and employed a measure of constituent atti- tudes on abortion. Bartels found that “senators are consistently responsive to the views of affluent con- stituents but entirely unresponsive to those with low income” (275).

Gilens used a different empirical strategy, analyzing system-level outcomes. He considered the link be- tween policy change and the policy-specific prefer- ences of survey respondents from different income groups. His core data are from 1981–2002, with 1,923 survey questions (although he also considers the periods 1964–68 and 2005–06). Using the full dataset, Gilens (77) shows only small differences in the policy influence of rich, middle-income, and poor con- stituents, usually on the order of a few percentage points. Because the preferences of high-, middle-, and low-income individuals are often highly correlated, Gilens focuses the bulk of his empirical analysis on a subset of these data—those for which there is at least a 10-percentage-point gap between the preferences of the affluent and the poor. In this subset, Gilens finds that it is only the preferences of the affluent that seem to affect policy.

Other researchers, building upon the work of Bartels andGilensandusing similarmethodological approaches, have also found evidence of inequality in responsiveness (e.g., Ellis 2012, 2013; Hayes 2012, 2013; Rigby and Wright 2011; Tausanovitch 2016). Most recently, Ellis (2017), using data from the 2012 Cooperative Congres- sional Election Study, developed two measures that capture the quality of dyadic representation provided to each survey respondent by her member of the US House—the first a measure of ideological proximity and the second a measure of policy agreement. Lawmaker ideology is DW-Nominate score; respondent ideology is

self-placement. These are standardized and the differ- ence is calculated. Policy agreement is the share of comparisons inwhich the respondentandMCagreed(on five bills). For both measures, the units of analysis are individual surveyrespondents, not senators (likeBartels) or system-level policies (like Gilens). Ellis finds that wealthier citizens are more ideologically proximate to their members of Congress and receive better policy representation.

In contrast, otherworkquestionswhether class-based inequality in responsiveness (whether pervasive or not) actually leads to pervasive inequality of outcomes. Soroka and Wlezien (2008) and Wlezien and Soroka (2011), studying government spending, find that pref- erences only differ by class for welfare spending, so that it is only in this domain that differential responsiveness can matter empirically. Enns (2015a, 2015b) uses the Gilens data to demonstrate that even when respon- siveness “slopes” differ by class, there remains a great deal of coincidental representation for low- andmiddle- income constituents. Branham, Soroka, and Wlezien (2017), again utilizing the Gilens data, focus on only those policies where middle- and high-income indi- viduals prefer a different outcome (not just where there is a 10-percentage-point gap in support). They find that, over a 22-year period, the rich won only 11 more times than the middle class. They also find only a modest conservative bias among these policies, suggesting that rich influence is attenuated by the fact that the rich hold a mix of liberal and conservative preferences.

Other research challenges the very existence of unequal responsiveness. This important research skeptical of affluent influencehasperhaps received less attention, scholarly or otherwise, than the work of Bartels or Gilens. Wlezien and Soroka (2011) use a “thermostatic model” of responsiveness to study government spending across six major policy domains, finding differences in influence but not generally fa- voring the rich. Ura and Ellis (2008) reach a similar conclusion in their study of House and Senate policy liberalism and government spending. Bhatti and Erikson (2011) correct and replicate the analysis of Bartels (2008), but add data from the 2000 and 2004 Annenberg surveys. This newer survey data have much larger sample sizes and enable Bhatti and Erikson to generate more accurate preference meas- ures by income group. These measures do not reveal evidence of elite influence. Ironically, neither do some results in Gilens (2012) itself. Although often over- looked, Gilens (2012, 199) offered a positive note for contemporary politics. By 2006, the end of his study, degrees of responsiveness across income levels had converged, with the poor about to overtake the rich.

Partisan Distortions of Representation

The study of partisan distortions in representation has emerged fromtheoreticalwork that seeks tounderstand why candidates and political parties do not converge toward the preferences of the median voter in an electorate. A recurring thread is that to obtain or keep

The Party or the Purse?

919

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

elected office, politicians must first secure their party’s nomination. Nomination typically requires winning a primary election inwhich only copartisans participate, inducing particular attention to the preferences of their copartisan constituents.

Recent empiricalworkhas found that in somesettings lawmakers do indeed privilege the preferences of copartisans. Kastellec et al. (2015) demonstrate this in a study of confirmation voting onnominations to theUS Supreme Court. Using estimates of support for con- firmation by party, they show that senators vote 75% of the time with their median copartisan constituent against their median constituent when the preferences of the twoconflict.Warshaw(2012), buildingonanearly version ofKastellec et al., employs estimates of partisan opinion to examine 43 roll call votes in the House of Representatives across five session of Congress. He shows that roll call votes are most responsive to and congruent with the policy-specific opinions of a law- maker’s copartisans, even on highly salient issues.2 On theotherhand, seeWright (1989)andGerberandLewis (2004) for evidence that lawmakers do not prioritize the preferences of their copartisans.

Research in this vein also suggests that there may be partisan differences in patterns of responsiveness. For example, Warshaw notes that while both parties priv- ilege the preferences of their copartisans, Republicans are somewhat more likely to do so than their Demo- cratic colleagues. This result is not dissimilar from that of Clinton (2006), who studied roll call voting in the 106th Congress. Clinton found that while Republicans were most responsive to the self-reported ideology of their copartisan constituents, Democrats were not (indeed Clinton surprisingly finds that Democrats were also most responsive to the preferences of their Re- publican constituents). More recently, Krimmel, Lax, and Phillips (2016) in their study of Congressional bills affectingLGBT rightsfind thatDemocrats (particularly white Democrats) have been responsive to liberalizing public opinion on this issue but that Republican law- makers havenot.Collectively, these studies suggest that it need not be the case that both parties are equally responsive ingeneral orwith regard to specificgroups. If Democrats and Republicans engage with opinion dif- ferently, lumping them together can obscure distortions of various sorts.3

Less commonly, research also considers the extent to which lawmakers show fealty to their party’s legislative agenda, potentially at the expense of responsiveness to constituentpreferences. For example,HusseyandZaller (2011) study historical roll call voting in Congress and

model a lawmaker’s DW-Nominate score in a given session of Congress as a function of that lawmaker’s partisanship and the partisanship of her district (which they use to capture, albeit imperfectly, constituent preferences). Their analysis finds that a lawmaker’s partisanship is the better predictor of roll call votes, al- though constituent preferences also matter. Hussey and Zaller interpret this result as indicating that the party agenda has a large independent impact on the behavior of elected elites. This motivates our consideration of “party-line” voting in our taking sides analyses.

Combining Partisanship and Economic Distortions

Some important work on economic distortions in rep- resentation has considered the role of political parties and partisanship, although this work does not consider partisan opinion as a distinct factor as we do here. A numberof scholars have noted (aswealsofindhere) that differences in policy preferences tend to be larger be- tweenparties thanbetweeneconomic classes (cf.,Bartels 2008). Others have considered whether Democratic and Republican lawmakers differ in thedegree towhich their behavior is biased toward the preferences of the affluent. Research in this vein often finds that while both parties tend to favor the rich, Republicans do so more fre- quently. Bartels, for example, runs separate regressions for the roll call votes of Democratic and Republican senators, finding that while neither party is responsive to the preferences of low-income constituents, Republican senators are about twice as responsive to high-income constituents as are Democratic senators. Gilens (2012) compares aggregate responsiveness under periods of Democratic and Republican control of the federal government, finding that inequality in responsiveness appears to be greater not only under Republican control but also that responsiveness to all income levels is higher (180).

Ellis (2017) explores congressional district-level variation in the amount of representational in- equality. He finds that “wealthy citizens are better represented relative to the poor… in districts repre- sented by Republicans” (134). While Ellis shows that Democrats are better representing the poor, he does find that under some circumstances—in high- inequality districts and in noncompetitive dis- tricts—they too privilege the preferences of the af- fluent.4 Rigby andMaks-Solomon (2017), in a working paperusing individuals as theunit of analysis,5find that

2 In general, research suggests that the policy positions of parties tend to be congruent with those of their supporters, especially on salient issues (cf., Lefkofridi andCasado-Asensio 2013; Giger and Lefkofridi 2014), although the direction of causality has been questioned (Achen and Bartels 2016). 3 Barker and Carman (2012) show that Republican constituents are less likely toprefera“delegate”modelof responsiveness topublicwill. Broockman and Skovron (2018) show that “politicians of both parties dramatically overestimate their constituents support for conservative policies” and that “Republicans overestimate constituency conser- vatism especially.”

4 For similar findings, see Ellis (2012, 2013) and Brunner, Ross, and Washington (2013). 5 We believe the proper unit of analysis is the senator, not survey respondent (as here or Ellis 2017, inter alia). Regressing opinion of respondentson thepositionof their senators implicitly runsaweighted regression in which senators with a large number of survey respondents are given more weight. Further, this approach may consider each senator vote-constituent preference dyad as an in- dependent observation, ignoring that the senator’s vote will be the same for all constituents in her district. Thismeans standarderrorswill be incorrect if they do not deal with this inflation of the number of observations (by, say, clustering at the level of the senator’s vote).

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

920

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

while the rich are better represented overall, the issues on which they receive superior representation vary by the partisanship of lawmakers. For instance, their results suggest that Republican senators better rep- resent the rich on economic matters, while Democrats better represent the rich on moral issues. For Rigby and Maks-Solomon, parties best represent the rich on those issues where there is more intraparty disagree- ment over policy.

However, another recent study places the blame for economic biases in representation solely on Re- publican lawmakers. Rhodes and Schaffner (2017) explore the association between the ideology of in- dividual constituents measured using data from Catalist (a private political data vendor) and their representatives’Nominate scores. They also compare roll call votes with the positions of individual con- stituents using data from the 2012 CCES. Employing either approach, they find that while Republicans provide a high degree to representation to individuals in the very top income percentiles, Democrats pro- vide a level of representation that has a flat or even negative relationship to income. Democrats and Republicans are said to provide fundamentally dif- ferent types of representation (i.e., “oligarchic” versus “egalitarian”), leading to a flat relationship on average.

Not all are convinced thatDemocrats providedmore equal representation. Hayes (2012) uses DW- nominate scores and constituent ideology to study responsiveness in Congress from 2001–10, finding greater levels of responsiveness to thewealthy, but that it is Republican lawmakers as opposed to Democrats who give more weight to the preferences of middle- income constituents. He also observes a greater bias toward the rich after theDemocrats took control of the Senate.

Moving Forward

To integrate economic and partisan distortions, we make analytic choices that often differ from those noted above, especially with respect to studies of economic distortions. (There is no perfect approach to studying responsiveness, and we suggest our path in addition to, not instead of other important lines of attack.) We prioritize votes cast by elected officials (rather than system-level outcomes), use multiple metrics for rep- resentation (including both responsiveness and con- gruence with opinion majorities), and use measures of opinion specific to the choices at hand (rather than ideology or indices). These choices are informed by six related concerns that frequently arise in the empirical study of representation.

The first concern is what one could call the “False Substitutes Problem.” It is, in our view, too lenient a test to praise democratic representation for, say, making abortion policy more liberal when it is opinion on im- migration issues that gotmore liberal, or viceversa—yet indices and ideological scores do just that. To care about responsiveness as a matter of normative democratic

theory, onemust surely think that the actual contents of the policy basket matter, and not just the ideological tone of the basket.6 Pooling policies/choices together to makeanaggregate index or using ideology scores forces all the component choices of these measures to be substitutes for each other. One liberal choice becomes like any other, as though the taste is for some degree of liberalism without caring what specific choices are as- sociated with it.

The second concern, the “Non-Common Scale Problem,” was best shown graphically by Erikson, Wright, and McIver (1993, 93) (also see Achen 1978; Matsusaka 2001; Gilens 2012, 41). If the scales of opinion and of policy-making are not the same, as often happens when either the input or the output side (or both) is an ideologymeasure, the slope and intercept of a responsiveness curve do not have any direct meaning. Without knowing how the scales are connected, one cannot say what they should be for perfect represen- tation. One cannot say if there is hyper- or hypo- responsiveness (too steep or insufficiently steep a curve), or if there is liberal or conservative bias (a leftward or rightward intercept shift of the curve). A positive slope for aggregatemeasures is compatiblewith any of these.

A third concern is an odd reversal across levels of analysis generally known as Simpson’s paradox, which in this context is a “Lumping-Splitting Paradox,” demonstrated in the appendix.7 Aggregate respon- siveness is neither sufficient nor necessary for respon- siveness of specific policy choices to specific opinion. One can have real responsiveness policy-by-policy and still find aggregate anti-responsiveness; one can have perverse anti-responsiveness policy by policy and have aggregate showings of responsiveness.

The fourth concern is over responsiveness versus congruence. A responsiveness approach seeks a statis- tically significant association between opinion and policy, considering all opinion-policy dyads at once. Congruence looks at each dyad in turn to see if the majority got what it wanted. Responsiveness need not mean that opinion majorities often get what they want. Nor do high levels of congruence necessarily imply responsiveness. Responsiveness without congruence can occur due to bias or weak responsiveness (yielding large democratic deficits as in Lax and Phillips 2009b, 2012; Matsusaka 2010). Congruence without re- sponsiveness can be coincidental. This can be called “Responsiveness-Congruence Independence.”

Fifth, there is the “Delegate Paradox” (Ahler and Broockman 2018): “representatives who represent their constituencies as closely aspossibleonevery issue can appear polarized and out of step ideologically.” The intuition is that a representative who obeys mild liberal opinion majorities one by one with her votes

6 We suspect normative theories of representation would not be satisfied by these ideological correlations, but see Sabl (2015) for the can of worms this comment opens. 7 Gelman et al. (2007) shows an example where richer states vote Democratic on average, but within states, richer individuals vote Republican.

The Party or the Purse?

921

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

leads to a legislator vote score that is extremely liberal, since votes are dichotomous (yes or no) compared to “size of liberal majority” measures. Vote indices can thus drastically overstate ideological extremism and polarization.8

Thefinal concernwouldarise froma focus exclusively on system level responsiveness. Representation in the US is dyadic by construction. Political actors, not sys- tems, make choices, and we expect such actors to re- spond to their own constituencies, not national opinion. In short, there is an “Ecological Inference Problem.” Moreover, systemic policymaking has its own compli- cations that obscure responsiveness pathways. Focusing on the roll call votes of individual senators also allows us to consider whether and how responsiveness differs by legislator type (for example, are Republicans more likely than Democrats to prioritize the opinions of the wealthy). This focus also arguably better captures the link between opinion and government action.9 To be sure, the bottom line of policy does indicate the nor- mative scope of representation deficits, so considering both levels of analysis is important.

Our choices respond to all six concerns. Specifically, dyadic analysis of specific roll call votes and opinion thereon deals with “False Substitutes,” “Non-Common Scale,” “Lumping-Splitting,” and “Ecological Infer- ence.” Doing both responsiveness and congruence deals with “Responsiveness-Congruence Indepen- dence” and the “Delegate Paradox.” Moreover, by fo- cusing on votes by individual senators, we can directly compare copartisan responsiveness to class-based re- sponsiveness andassess behavioral differencesbetween Democratic and Republican senators.

In addition to the concerns above, there is a thorny empirical issue that complicates efforts to parse out the competing subgroups effects—the opinions of high-, middle-, and low-income individuals are often highly correlated, and collinearity makes it difficult to tease out influences in a simple multivariate regression ap- proach (e.g., estimates are unstable and even some- times oddly signed). The Gilens approach to this has two parts, the first being to use separate bivariate regressions on rich and on poor opinion for key results. The coefficients on opinion from each separate re- gression are then compared.10 However, running separate regressions for different independent varia- bles only “solves” the collinearity problem by creating omitted variable bias within each regression, un- dercutting simple comparisons of coefficients or their significant levels.11

Wedopresent bivariate results for comparison, along withnoisymultivariate regression results, but onlymore data can truly resolve collinearity concerns. Given that one cannot create more data, scholars need more cre- ative solutions. Gilens addresses collinearity concerns by focusing most of his inquiry on the subset of policies for which the difference between rich and poor opinion is at least 10 percentage points. Gilens defines this as disagreement.12

We instead deal with collinearity by focusing most of our inquiry on a series of taking sides analyses. In these, we define conflict as existing between groups if one is above and the other below a give opinion threshold, typically 50%.We recognize that amajoritarian thresh- old such as this can be problematic if there are many instances where subgroup opinion is clustered near 50%. This suggests the need for robustness checks (varying thresholds) and the incorporation of un- certainty in one’s estimates, as we do below. The taking sides analyses that we present focus on instances where the rich and poor disagree (as opposed to the rich and middle class). Doing so provides more observations of genuine disagreement, given that opinion differences are greatest between the poor and rich. However, our findings remain unchanged if we focus on disagreement between the rich and middle.

OPINION ESTIMATION & DATA

The survey data for estimates of constituent opinion come from the common content portion of the Co- operative Congressional Election Survey (CCES), the National Annenberg Election Survey, and a variety of other reputable polling firms such as Gallup and Pew.13

We estimate opinion by state, income group, and partisan identification using multilevel regression and poststratification (MRP). This technique, first pre- sented by Gelman and Little (1997), uses national surveys and advances in Bayesian statistics and multi- level modeling to generate opinion estimates by demographic-geographic subgroups. MRP has been shown to produce accurate estimates of public opinion by state and by congressional district (Lax and Phillips 2009a, 2013; Park,Gelman, andBafumi 2006;Warshaw and Rodden 2012), using a relatively small number of survey respondents, as few as contained in a single (moderately sized) national poll, and fairly simple demographic-geographic models of preferences (Lax and Phillips 2009a). Indeed, MRP has been called the new “gold standard for estimating constituency pref- erences fromnational surveys” (SelbandMunzert 2011, 455; cf.; Buttice and Highton 2013; Lax and Phillips 2013; Toshkov 2015).8 See also Broockman (2016) showing citizen ideology scores capture

consistency more than policy preferences. 9 With the exception of the President, no other elected official has a national electorate. Thus, it is not national opinion that should in- fluence congressional action, but rather district- or state-level opinion that should matter. 10 Gilens also runs supplemental multivariate regressions using a correlated errors approach. 11

“The Difference between ‘Significant’ and ‘Not Significant’ is not Itself [Necessarily] Statistically Significant” (Gelman and Stern 2006).

12 One concernwith theGilens approach is that a 10-percentage-point difference in the absolute levels of opinion is not necessarily indicative of a real disagreement between groups. For example, if 70% of the poor want to raise theminimumwage as do 60%of the rich, wewould be hard pressed to think of this as a policy disagreement. The data dropped by this subsetting is a potentially odd moving window. 13 See Online Appendix for more detail.

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

922

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

MRP proceeds in two stages. In the first stage, a multilevel model of individual survey response is estimated, with opinion modeled as a function of a respondent’s demographic and geographic charac- teristics. The state of the respondents is used to estimate state-level effects, which themselves are modeled using additional state-level predictors. Residents from a par- ticular state yield information on how responses within that state vary from others after accounting for de- mographics. All individuals in the survey, no matter their location, yield information about demographic patterns which can be applied to all state estimates. The second step of MRP is poststratification: the opinion estimates for eachdemographic-geographic respondent type are weighted (poststratified) by the percentages of each type in the actual population of each state. This procedure allows us to estimate the percentage of respondents within each state by income category and partisanship who have a particular issue position or policy preference.

In stage one, wemodel survey response (i.e., whether a respondent supports a given policy proposal) as a function of a respondent’s race and gender combi- nation (men and women divided into four racial cate- gories—black, Hispanic, white, and other), age (18–29, 30–39, 40–49, 50–59, 60–69, and 701 years), education (less than a high-school education, high-school gradu- ate, some college, college graduate, and postgraduate education), partisan affiliation (Democrat, Indepen- dent, orRepublican), incomecategory (number varying by survey), and state. We allow full interactions be- tween income category, state, and party so the pre- dictive effects of income can vary by states and party within states.

Income effects are modeled as follows. The CCES uses 14 to16 incomecategories, dependingon thepoll.14

Wemodel randomeffects by categorywith linear trends based on the midpoint of each category. We take the square root ofmidpoints to account for the unequal size of income categories. We allow the trend variable to vary by state and party. Trend variables are useful when modeling opinion for narrow population subgroups (Lax and Phillips 2013).

MRPsuccess dependsongoodgroup-level predictors to capture residual differences across states or the like. As a state-level predictor, we use a “demographically purged state predictor” (DPSP) (Lax and Phillips 2013).15

We face a complication that is not present in most applications ofMRP.Typically, researchers poststratify their estimates using population frequencies from the Census “five-Percent Public Use Microdata Samples” or the American Community Survey. Unfortunately, forourpurposeshere, thesedatadonot includepartisan identification (but they do include income). Thus, using standardMRP,onecanestimate the level of support for, say, President Obama’s health-care reform among middle-income college-educated black females aged 18–29 years in California, but one cannot estimate the level of support among Republican, Independent, or Democratic individuals of the same type.Kastellec et al. (2015) present a solution: “two-stage MRP.”Using the Census data as a starting point, their approach involves an additional stage of MRP to generate a new post- stratification file that includes party. We begin by col- lecting data on individual survey responses about partisan identification (i.e., whether a respondent is a Democrat, Republican, or an Independent) across multiple points in time spanning the years of our public opinion surveys. We then model partisanship as a function of demographic and geographic variables. Specifically, we treat partisanship as a response variable and apply standardMRP to estimate the distribution of partisanship across the full set of “demographic-geo- graphic types” fromabove.We then have an estimate of the proportion of Democrats, Independents, and Republicans among, say, income-category-3 (30 to 40k) college-educated black females aged 30–45 years in California.16

We construct estimates of opinion by partisan group (Democrats, Republicans, and Independents). We then construct estimates of opinion by income quintile within each state, forming five equally sized groups so that we can look at the opinion of the “rich” (top quintile), “poor” (bottom quintile), or “middle” (middle quintile). One advantage of this approach is that we compare rich and poor opinion in each state, not opinion across rich states and poor states (which would result if we used national cutoffs).17 Similarly, we examine responsiveness to rich and poor coparti- sans by taking the top and bottomquintiles of partisans within each state. From these, we derive the median position in each subgroup.

Fromthesurveys,wehave identified49questions that ask respondents their preferences on roll call votes that were actually taken bymembers of Congress (a list with details is provided in Web Appendix Table A.1). For example, in 2012, one such question asked respondents whether they would support a plan to extend Bush era tax cuts for incomes below $200,000; another asked whether the Affordable Care Act should be repealed. The surveys employed ask respondents how theywould

14 Some questions use agglomerations of other surveys. These con- structed “megapolls” may have dozens of non-overlapping income categories. Rather than estimating separate income effects by poll, we standardize incomeby assigning subjects to one of the standardCCES categories, employing weights when categories overlap, using a uni- form distribution. For example, if an individual in a megapoll has income of $8,000–$13,000, they would constitute a member of the $0–$10,000 income group with weight 0.4, and the $10,000–$20,000 group with weight 0.6. 15 DPSP is the average liberal/conservative variation in state-level public opinion that is left unexplained by a variety of demographic predictors. BecauseDPSP was estimated across a wide set of policies, it is a gooddefaultwhenusingMRPtopredictopiniononagiven issue.

16 We estimate partisanship using a five year rolling window to in- crease thenumberofobservationsandsmoothyear-to-yearchanges in partisanship. Differences are minor. 17 For example, for issues from 2001–04, the cutoff for “rich” household income varies from a low of $67k in WV, to the median of $84k inOR, toahighof $120k inNJ.Thecutoffs forpoor in these states are $16k, $24k, and $31k, respectively.

The Party or the Purse?

923

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

vote on these issues if they were amember of Congress. These questions include some of the most important economic, social, and foreign policy votes cast by members of Congress since 2000. Our sample of votes includes health-care reform, President Obama’s stim- ulus bill, an extension of the Bush tax cuts on capital gains, the Federal Marriage Amendment, and a vote to withdraw American military personnel from Iraq. For each, we focus on the share (of those with an opinion) who favor a “yes” vote. Vote data come from Con- gressional Quarterly and Congress.gov.

Where possible, given complexity and computational limits, we make use of a method sometimes called propagated uncertainty or the method of composition (Treier and Jackman 2008) to capture uncertainty around our opinion estimates, given the partisan poststratification estimates. We use empirical distri- butions to simulate uncertainty (500 simulations) from the survey response modeling stage (based on the variance–covariancematrix of a givenmultilevel model) and propagate it forward. We note where uncertainty is shown, generally for congruence scores and “taking sides” results. Where breaking down votes by senators, for clarity, or for regression results, we use point pre- dictions for opinion.18

OPINION PATTERNS

How much does state-level public opinion differ as a function of economic class and political party? We find that, on average, there are not large differences between the preferences of high- and low-income Americans. This is consistent with much existing re- search (cf. Bartels 2008; Soroka and Wlezien 2008; Gilens 2012, 2015; Enns 2015a, 2015b; Branham, Soroka, and Wlezien 2017). Across all of the roll call votes included in our empirical analysis, the average state-level difference in opinion between the top and bottom quintiles is only 10 percentage points. There are still many instances of disagreement: the top and bottom state quintiles prefer different policy choices (i.e., are on opposite sides of the 50% opinion threshold) approximately 22% of the time.

Figure 1 displays, by roll call vote, the average state- level differences in opinion between the top and bottom income quintiles, grouping the roll calls into three issue types—security, economic, and social (we discuss partisan differences later). We observe the smallest class-based differences in opinion on social issues, where the average difference between the

opinion of the top and bottom quintile is only five percentage points. On security and economic matters, class-based differences tend to be larger, around 12 points.

We often observe, however, high levels of polari- zation on issues that either largely benefit high- income earners (for example, reducing the capital gains tax) or that clearly benefit low-income indi- viduals (for example, funding the State Children’s Health Insurance Program). We also tend to observe relatively high opinion polarization on free trade issues, where the average class based difference in opinion is 15 points.

Figure 1 also showspartisandifferences.These tend to bemuch larger than class-based differences (again this is consistent with findings in the existing literature, cf., Branham, Soroka, and Wlezien 2017; Rigby and Maks- Solomon 2017). The mean state-level difference in opinion between Democrats and Republicans is ap- proximately 38 percentage points (compared to only ten for class). Thus, while the top and bottom income quintiles in a state agree on many issues, self-identified Democrats and Republicans do not. Democrats and Republicansdisagree62%of the time(comparedto22% by class). Partisan polarization is, on average, lowest for economic matters and highest on social issues. There is one issue for which class disagreement is substantially more common than partisan disagreement—support for the US-Korea Free Trade Agreement. For nearly all other issues, partisan disagreement is much more com- mon. Formany policies, all states have disagreeing party medians; for many policies, no states have disagreeing income medians.

Figure 2 brings all these together. The left-hand side shows the difference in opinion levels and the right- hand side the percentage of states in which medians disagree. Each panel compares class differences to partisan differences, with the latter presenting the starker choice for a senator seeking to please con- stituents. Coincidental representation is not available across party lines the way it is to rich and poor, who actually agree quite often.

How often do different segments of the public share the same policy preference? Table 1 displays opinion agreement rates between various subgroup medians. Democratic opinion coincides with the statewide median more than does Republican opinion; poor and rich agree with the statewide median at roughly equal rates. Importantly, Republican copartisans agree with the rich more often than with the poor, while Demo- crats agree with the poor more than the rich. The subgroup of those we study least reflective of the median voter is the rich Republicans. The subgroup most reflective would be independents, followed closely by the rich.

Party conflict is high and pervasive. Democratic and Republican medians only agree 38% of the time. If we look further, at richandpoorquantileswithin eachparty by state, poor Democrats and poor Republicans only agree42%of the time (not shown).RichDemocrats and rich Republicans only agree 30% of the time. Indeed, rich Democrats are more likely to agree with poor

18 Our estimates of opinion are quite precise. Taking the state-issue as unit, the median standard deviations across simulations are 2.3 for statewide, 3.6 for low and high income quintiles, and 2.2 and 2.6 for Democrats and Republicans respectively. Narrower categories are naturally estimated with more uncertainty: low and high income Democrats at standard deviations of 4.2 and 3.6, Republicans re- spectively at 4.0 and5.0. Some issueshavemoreprecise estimates than others: for statewide estimates, the median of the standard deviations across simulations for each issue range from 1.1 for Iraq Withdrawal (2006) to 5.8 for SCHIP (2010).

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

924

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

Republicans (41%) than with the rich ones. Rich Republicans are more likely to agree with the overall poor (44%) than with the Democratic rich. Within

parties, there is much more agreement. Rich and poor Democrats agree 89% of the time and Republicans 84%.

FIGURE 1. Opinion Polarization by Issue, Class, and Party

This figure shows the difference in opinion between the top- and bottom-income quintiles (light) and between Democratic and Republican voters (dark), averaged by issue across states. Boxes indicate the 25th and 75th percentile, with the vertical line within boxes the median opinion difference across states.

The Party or the Purse?

925

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

RESPONSIVENESS

Using these data, Figure 3 shows slopes for respon- siveness to poor, rich, statewide, and copartisan opinion first for all senators, and thenbrokendownbyparty. For all senators, the slope is steeper for rich than poor (as in Gilens), but steeper still for copartisans. Democrats seem strongly responsive to the preferences of all four normal categories of opinion (omitting the opposing partisan group). Republican voting behavior seems anti-responsive to every group, except copartisans.

Are Republicans actually responding perversely to public opinion? Sometimes, but it is not so simple. Note

that the averageDemocratic bill ismorepopular,19which can be seen in popularity distributions along the rug of Figure3ormoredirectly inFigure4, at the levelof thebill.

Figure 5 breaks down responsiveness by the parti- sanship of the bill. The party difference in voting on bills—the striking intercept shift—is important.

TABLE 1. Opinion Agreement between (Sub)Group Medians (%)

Statewide Poor Rich Democrats Republicans

Poor 87 – 78 84 51 Rich 91 78 – 70 68 Democrats 76 84 70 – 38 Republicans 62 51 68 38 –

Independents 93 84 88 74 64 Dem. poor 74 85 67 94 36 Dem. rich 74 80 69 95 37 Rep. poor 67 57 72 43 94 Rep. rich 55 44 61 31 91

Standard errors around these given uncertainty are approximately one percentage point.

FIGURE 2. Opinion Differences by Issue among Partisan and Income Groups

We plot themean difference in opinion betweenDemocrats and Republicans against the difference in opinion between rich and poor voters (left) and the percentage disagreement between Democrats and Republicans against the percentage disagreement between rich and poor voters (right). The 45° line is shown.

19 We identified the partisanship of bills by the percentage of each senate caucus voting in favor of the bills. Other codings yield similar results. This coding is endogenous to the votes themselves, but only in the aggregate, and that complicationwouldnot explain thedifferences between behavior on Democratic bills versus Republican bills.

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

926

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

Republicans vote against bills the Democrats side with, and Democratic bills are more popular, leading to their appearing anti-responsive overall (this too can be a Simpson’s paradox). On top of that, Republicans are

indeed anti-responsive within the set of Democratic bills.By contrast, theDemocrats aremore responsive to opinion on bills, both their own and those led by the other party. Republicans are mildly responsive to

FIGURE 3. Support for Each Party’s Bills among Constituent Subgroups

Weshow responsiveness for all senators (top row), Democrats (middle), andRepublicans (bottom). Outlined and shaded density curves on the x-axis show the distributions of opinion forRepublican andDemocratic bills respectively. Linear regression lines show responsiveness to subgroup opinion (thick lines for all bills, and dotted or dashed for Democratic and Republican bills respectively.

FIGURE 4. Responsiveness

The Party or the Purse?

927

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

opinion on their own bills. This result is consistent with existing work that finds partisan differences in re- sponsiveness to public opinion (Clinton 2006;Warshaw 2012; Krimmel, Lax, and Phillips 2016), a pattern that will repeat itself in our congruence analysis.

In Table 2, we build from the bivariate analyses (as shown in Figure 3) to multivariate analyses of re- sponsiveness, allowing for varying slopes and intercepts. Thefirstpanelpresents results forall senators, thesecond for Democrats, and the third for Republicans. Within each, there are three bivariate regressions of senator’s roll call vote on rich or poor or copartisan opinion, three multivariate regressions each including two of the three aforementioned, and a final regression with all three. Perhaps more than anything else, the results show how difficult it canbe,as inGilens’s data, to teaseoutdifferent effects using responsiveness regressions alone. Overall, public opinion is a robust predictor of roll call voting, but there are differences in coefficient sizes across subgroups and by senator party. In themodels that include all three preference measures, it is the coefficients on rich and copartisan opinion that are largest andmost consistently statistically significant. AmongDemocratic senators, the difference in coefficients on rich and poor opinion is smaller than that among Republican senators. (To give a sense of scale, a coefficient on opinion of 0.16 corre- sponds to up to a four-percentage-point increase in the probabilityofayesvote, fora senatoron the tippingpoint between yes and no.)

Table 3 is parallel to the previous table, but regresses vote on the dichotomous position of the median of the relevant subgroup instead of the actual opinion level. This takes a step in the direction of our taking sides analyses. Public opinion, of course, remains a robust predictor of roll call voting. In this table, however, it is the coefficient on partisan opinion, in the form of the partisan median, that consistently has the largest co- efficient (e.g., in model 7, a coefficient around 4 is a full swing in probability of a yes vote). Once again, among Democratic senators, the difference in the estimated

coefficient on rich and poor opinion is smaller than among that Republican senators.

Collectively, these regressions, although messy to interpret,20 do provide some further evidence that there are partisan differences in responsiveness to the opin- ions of poor constituents and that the opinions of copartisans have an meaningful and independent im- pact on roll call voting.

However, given problems of collinearity, we need to be careful not to place too much faith in these regressions. And, as we noted above, “Responsiveness-Congruence Independence” means that responsiveness can coexist with biased representation, without congruence, and thus with a significant democratic deficit. Thus, onemust await congruenceand“takingsides”analyses togeta fullpicture of representation.Wealsoneedtoknowhowtheoddanti- responsiveness above (among Republicans) translates into congruence, as well as how the multiple opinion influences found in the regressions translate into bottom- line representation.

CONGRUENCE

We next consider congruence, that is, whether the sub- groupmedianactuallygets thevote that it desires fromits senator. We first report congruence by subgroup and then contrast across groups. Figure 6 plots congruence rates with uncertainty, with values shown in Table 4.

Congruence averaged across all roll votes is ameager 58%. In general, the rich do a bit better than the poor, with a five-percentage-point advantage. However, this

FIGURE 5. Responsiveness by Bill

20 The results do differ a bit table to table given the different in- dependent variables (levels of opinion versus dichotomous median positions), since one can have responsiveness to opinion, or associ- ationofopinion levelsandoutcomes,withoutmajoritarianism.Table2 has the advantage of a less crude opinion predictor; Table 3 has the advantages of better capturing impact of inter-group conflict and dealing with the collinearity of opinion.

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

928

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

TABLE 3. Regressions of Vote on Subgroup Opinion Median Position

All senators Democrats Republicans

Median: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

Rich 1.65 – – 1.49 0.84 – 0.75 1.99 – – 1.74 1.63 – 1.51 1.90 – – 1.99 0.58 – 0.63 (0.24) – – (0.27) (0.46) – .(49) (0.53) – – (0.52) (0.55) – (0.57) (0.53) – – (0.55) (0.72) – (0.65)

Poor – 0.91 – 0.54 – 0.43 0.25 – 1.69 – 1.43 – 1.13 0.88 – 0.13 – 20.41 – 20.13 20.25 – (0.19) – (0.22) – (0.27) (0.26) – (0.53) – (0.50) – (0.48) (0.53) – (0.39) – (0.44) – (0.47) (0.44)

Party – – 4.58 – 4.52 4.55 4.54 – – 4.36 – 3.72 3.50 3.20 – – 2.53 – 3.42 2.97 3.60 – – (0.39) – (0.37) (0.39) (0.38) – – (0.93) – (0.99) (1.00) (1.06) – – (0.77) – (0.78) (0.82) (0.75)

AIC 5,961 6,112 4,082 5,942 4,042 4,084 4,051 1,459 1,478 1,469 1,456 1,448 1,469 1,452 1,551 1,579 1,568 1,559 1,550 1,574 1,558 PCP 66 64 80 66 80 80 80 90 89 89 90 90 90 90 88 88 88 88 88 88 88 N 4,796 2,473 2,323

Bayesian logit models, using BLGMER in R. Standard errors beneath the coefficient. Models include random intercepts for the opinion dummy variables by issue. Point predictions for opinion are used. AIC and PCP are shown.

TABLE 2. Regressions of Vote on Subgroup Support

All senators Democrats Republicans

Op.: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

Rich 0.21 – – 0.24 0.13 – 0.16 0.19 – – 0.14 0.10 – 0.10 0.26 – – 0.29 0.23 – 0.25 (0.02) – – (0.02) (0.03) – .(03) (0.03) – – (0.03) (0.03) – (0.04) (0.03) – – (0.11) (0.04) – (0.04)

Poor – 0.14 – 0.01 – 0.03 20.05 – 0.17 – 0.07 – 0.03 20.03 – 0.08 – 20.13 – 0.00 20.08 – (0.02) – (0.02) – (0.02) (0.02) – (0.02) – (0.03) – (0.04) (0.04) – (0.03) – (0.03) – (0.02) (0.03)

Party – – 0.16 – 0.16 0.16 0.16 – – 0.15 – 0.10 0.13 0.11 – – 0.17 – 0.05 0.17 0.07 – – (0.03) – (0.01) (0.01) (0.01) – – (0.02) – (0.02) (0.03) (0.03) – – (0.03) – (0.04) (0.03) (0.03)

AIC 5,147 5,775 3,087 5,118 2,934 3,025 2,930 1,324 1,379 1,362 1,325 1,311 1,354 1,321 1,420 1,551 1,480 1,546 1,424 1,481 1,420 PCP 75 70 88 75 89 88 89 91 90 90 91 91 90 91 89 89 88 89 89 89 89 N 4,796 2,473 2,323

Bayesian logit models, using BLGMER in R. Standard errors beneath the coefficient. Models include random intercepts and slopes for the opinion values by issue. Opinion (using point predictions) is centered and scaled so that a one unit change is one percentage point. Better penalized fit for a given data subset is shown by lower AIC. Percent correctly predicted (PCP) is shown.

T he

P arty

or the

P urse?

929

https://doi.org/10.1017/S0003055419000315 Published online by Cambridge University Press

advantage is notably smaller than that enjoyed by copartisan constituents who see a congruence rate 16 points higher than that of the rich. Indeed, even poor copartisans enjoy a rate of congruence that is 13 points greater than the rich (although among copartisans, the affluent do slightly better than the poor).

Onceagain,wesee thatDemocratic senatorsappear to be more responsive to public opinion than their Re- publican counterparts. Democrats more frequently cast votes consistent with the preferences of state medians, rich medians, poor medians, and copartisan medians than doRepublicans.Democrats are evenmore likely to vote in line with the rich than are Republicans, but vote more in linewith thepoor thanthe rich.Republicansvote more with the rich than the poor.

Next, Figure 7 is similar to Figure 3 butwith the y-axis capturing congruence with majority opinion instead of a yes vote. A steep U-shape would show strong ma- joritarian responsiveness, with a softer shape the more likely pattern of weak responsiveness to baremajorities and congruence at the strong extremes. Again, we look at congruence with the various subgroups. Overall, the patterns look normal, lumping all senators together… or taking just the Democrats. The Republicans look, if

anything, anti-congruent. The larger the opinion su- permajority, the more clear-cut the majoritarian posi- tion, themore likely theDemocratic senators are to side with it and Republicans against it. Again, break this down further to seewhy this is the case. The dotted lines show votes on Democratic bills and the dashed lines on Republican bills. Republicans vote for unpopular Re- publican bills (as Republican bills are typically during this time period) and against popular Democratic bills (as Democratic bills were in this time period). The Democrats meanwhile voted more often for the bills of their own party (which were popular) and against those of the other (which were not). Again, different levels of aggregation can reveal different patterns, for congru- ence aswell as responsiveness.And the degree towhich the partisan orientation of a bill drives what happens is both striking and dangerous to ignore, even for eval- uating congruence by class.

We now shift to the senator as unit of analysis, to compare degrees of congruence across subgroups. Figure 8’s top left panel plots each senator by degree of congruence with low- and high-income constituents, whether or not they agree with each other (unlike “taking sides” to come).

Democrats vote on average with public opinion of bothgroupsmore thando theRepublicans.That is, both rich and poor are more likely to see their preferences converted into actual senate votes by the Democrats. The Republicans tend to be above the 45° line and the Democrats below, showing their respective tilts toward rich and poor, even given the Democrats higher con- gruence to both rich and poor. The top right panel compares congruence rates with partisan medians, showing the expected pattern. The bottom left panel shows congruence rates with partisan medians versus statewide medians. Once again, Democrats have not only higher congruence rates with statewide medians but also (slightly) better satisfy their own partisan medians. Finally, the bottom right panel shows

TABLE 4. Congruence of Senators’ Votes (%)

Congruence with…

constituents

All Democratic Republican Senators Senators Senators

Median 58 70 46 Rich 60 67 53 Poor 55 72 35 Copartisan 76 80 71 Noncopartisan 31 35 25 Rich copartisan 79 82 75 Poor copartisan 73 79 67

FIGURE 6. Congruence of Senators’ Votes (%), with Uncertainty

A 99% confidence interval represents uncertainty around congruence estimates.

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

930

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

congruence with in-party rich and poor. Democrats here too show higher congruence on both dimensions, with both parties roughly along the 45° line.

What we have learned? What do we still need to know? Congruence thus far has not been zero- sum—rich and poor medians often agree, as do party medians and classmedians.Recall fromTable 4 that the rich saw congruence only five points higher than the poor (60 to 55). Differences in congruence rates are not large, but there are already striking partisan patterns. The Democrats are the party of higher congruence, yet tilt a bit toward the poor. The Republicans are less congruent in their votes than a random coin flip, except (barely) with the rich. Copartisan congruence is quite high, making it all themore important to let copartisans and the rich go head to head. What happens when they clash and a senator must take sides?

TAKING SIDES

Webeginwith instances where there is conflict between the opinions of a senator’s rich and poor constituents.

Senators side with the rich against the poor 63% of the 1,026 relevant votes. So far, this is similar toGilens. The parties, however, are quite different on this point: Democrats side with the rich on only 35% of 474 votes, while Republicans do so on 86% of 552 votes.21

Weexplore this type of zero-sum taking-sides conflict in a series of figures, starting with Figure 9. Each panel limits the setof votes towhereaparticular conflict exists, between one set of constituents on the left and another on the right. Each triangle plotted is a senator’s “score,” the percentage summarizing how often they voted for one side or the other, ranging from100% left (0%right) to 50–50 to 100% right (0% left). Triangle sizes are scaled to the number of votes by senator. Democrats andRepublicans are separatedbelowandabove the line, respectively. We can see how senators vary by party by

FIGURE 7. Congruence

Weshowcongruencecurves for all senators (the top row),Democrats (middle), andRepublicans (bottom). The relativedistributionof opinion is shown along the x-axis. The outlined and shaded density curves show the distributions of levels of public opinion. Locally weighted regression lines show responsiveness to poor, to rich, to statewide, to copartisan, and to opposing party opinion. The black lines cover all issues; the dotted and dashed lines cover Democratic and Republican bills, respectively.

21 If we look at collective decisions instead of individual roll call votes, we observe that amajority of richmedians (across all states) preferred a different outcome than amajority of poor medians on only 10 of the 49bills.Of these10, the richmedians received their preferredoutcome on seven bills and the poor their preferred outcomes on three.

The Party or the Purse?

931

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

FIGURE 8. Congruence of Votes with Opinion Groups by Senator: Class, Partisan, and State

Each panel shows congruence rates for Democratic (circles) and Republican (triangles) senators for two types of constituent opinion, symbols scaled to the number of votes.

FIGURE 9. Taking Sides—Class Conflict

FIGURE 10. Taking Sides—The Median Voter

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

932

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

FIGURE 11. Taking Sides—Party Conflict

FIGURE 12. Taking Sides—Party or the Purse

FIGURE 13. Taking Sides—Partisans Take Sides

The Party or the Purse?

933

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

the shadedGaussian density distributions.We also show the number of votes that fall into the categories.

Figure 9 continues exploring the results discussed above. ThemedianDemocrat tilts toward the poor with 65%. However, there is a fair amount of variation among Democrats. For example, Sen. Russ Feingold is among those on the far left, siding with the rich against the poor in only one of nine votes, while Sen. Claire McCaskill did so in four of six votes. Republicans far more strongly sidewith the rich; theirmedian senator in this subset did so 91% of the time. The overall level of pro-rich bias is being driven by Republican senators, most strongly by senators such as Sen. James Inhofe with all eleven of his votes, or Sen.Olympia Snowewith six of nine.22 These results are consistent with findings that Republicans are more responsive to the wealthy than are Democrats (cf., Bartels 2008; Ellis 2017; Rhodes and Schafner 2017).

Figure 10 brings in statewide median constituents. The median Democratic senator in the top panel sides with the statewide median more than the rich, and the median Democrat in the bottom panel sides with the poor over the statewide median (again with variation among Democrats). Republicans strongly alignwith rich over statewide andwith statewide over poor.

Figure 11 shows the expected partisan split for con- text—when the party medians disagree, senators strongly but not monolithically favor the position of copartisans.Many senators cross the aisle. A handful of “mavericks” are more likely to represent out-partisans than in-partisans: Democrats Zell Miller and Bob Torricelli; and Republicans Olympia Snowe, Susan Collins, Arlen Specter, and especially Lincoln Chafee (but not, surprisingly, John McCain, voting with his party on 22 of 30 such votes). The bottom panel shows senators match copartisan opinion more than their statewide median opinion, showing a clear and strong partisan distortion to representation.

Figure 12 combines these threads (finally!), with conflict between class and partisanship. The top panel shows that both parties’ senators mostly side with copartisan medians over the rich. The rich median may beat the poor two to one in a direct fight, but the copartisans beat the rich four to one. Both parties also side with copartisans over the poor (bottom panel). In

both panels, the Republicans tilt further to partisans than do theDemocrats. Republicanswho sidedwith the rich over the party were Snowe, Collins, Specter, and Chafee. The only Democrats who did so (and hadmore than four such votes pitting the rich against their par- tisan voter) were Nelson and Carper.

We can dig further. Figure 13 shows that what is really pivotal is where the party median stands, alongside the richmedian or instead the poor one. The poorwinwhen the copartisans are on their side.Of the 51 times a Republican senator faced a Republican median siding with the poor against the rich, the Republican senators cast 38 votes with the former (75%).Democratic senators similarly votedwithparty and poor over rich 76% of the time. While the Republicans drive thehighvictory rateof richmedians over poor medians, that in turn depends on Re- publican constituents aligningwith the rich; when they alignwith the poor, the rich advantage becomes a poor advantage. The rich do a bit better than the poor comparing both panels, but the partisan thumb on the scale is heavy.

Another way to consider partisan opinion is to look within copartisan subgroups, at conflictwithin thepartisan constituency. For example, taking the top 20% of Dem- ocrats in the state by income,what is themedian position? Out of 49 policies by 50 states (2,450 state-policy units), Democratic rich and poor quintiles disagree 11% of the time (on 263 state by issue observations). These areas of intraparty disagreement connected to 267 votes out of the 2,473 cast by Democratic senators that forced a choice between pleasing the in-state Democratic rich and Democraticpoor.Republican richandpoordisagree16% of the time (384 out of 2,450), connecting to 350 votes (out of 2,323 total) that forced a choice for Republican senators.

Here, in this limited set of votes, when there is dis- agreement between the poor and richwithin a party, we do find a pattern of affluent influence not limited to Republicans. Figure 14 shows how senators take sides between copartisan poor and copartisan rich, parallel to the top of Figure 9. Senators side with copartisan rich 72% of the time, with a moderate difference between Republicans (78%) and Democrats (63%).

All this so far has set aside yet another aspect of partisanship, the elite party line. Figure 15 explores this dimension of partisanship: how does a senator vote when themodal position of her copartisan Senate peers conflicts with that of her constituents? We do not see pure party-line voting, but the party position trumps

FIGURE 14. Taking Sides—Rich and Poor within Party

22 Limiting these “taking sides” comparisons to economic issues only, Democrats side with the rich even less often and Republicans more often.

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

934

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

that of rich and of the partisan nonelite over 80%of the time in both parties. Those Republicans who broke from their fellow partisans and aligned (coincidentally or not) with the rich were the usual suspects: Sens. Chafee, Collins, Snowe, and Specter.Whatever pull the rich have (or partisan constituents have, for that mat- ter), the party line beats it most of the time, even for Republicans. This reveals a sharp limit to affluent in- fluence and responsiveness in general. Always voting the party line, as formed by Senate partisan coalitions, would only yield 56% congruence (observed level 58%).

Taking this narrative as a whole, how well-off are the rich? Figure 16 summarizes, simplifies, and adds further information, by incorporating the uncertainty in our congruence estimates.

This figure also shows what happens when middle income constituents conflict with others: rich versus middle looks like a slightly attenuated version of rich versus poor and both are similar to middle versus poor (i.e., the same partisan pattern). Throughout, com- parisons involving the middle tell the same story as our series of taking sides graphs suggest.

FigureA.5 (in theOnlineAppendix) showsweget the same “taking sides” results if we limit votes towhere the opinion levels are divided by at least 10 percentage points, orwhere opinion levels are notwithin five points of the 50% majoritarian cutoff, or where the state Senate delegation is split with one Democrat and one Republican. These robustness checks show the 50% cutoff is not driving our findings.

The pattern in Figure 16, capturing our main inquiry, is clear.Yes, the rich getwhat theywantmoreoften than the poor… when the partisans and the rich agree. Partisanship conditions and constrains class clout. The rich largely have Republican partisanship to thank for any greater influence they have.

There are two ways the rich benefit from Repub- licans. First and foremost, Republican constituent

medians tend to side with the rich more than with the poor. Second, and secondarily, when class differences exist within the Republican constituent coalition, Re- publican senators sidewith theRepublican rich over the Republican poor. In this latter sense, so do the Dem- ocrats side with their own partisan rich over their partisan poor, albeit to a less distorting degree. Rich partisans beat out middle-class partisans and the latter beat poor partisans.

One shouldbe careful not tooverstate the substantive impact of Republican senators’ tendency toward the rich. There are only fourRepublican pro-rich votes (out of 2,323 Republican votes total) that oppose both the statewidemedian and copartisanmedian (in these same four, they also oppose the poor median and the median partisan peer).

DISCUSSION AND CONCLUSION

Our analysis of Senate roll call voting brings together, for the first time in a quantitative empirical inquiry, constituent opinion by income and party. Doing so allows us to evaluate, in tandem, class and partisan distortions to representation, to gain a more complete understanding of each, and to document the ways in which they interact. Erikson (2015, 24) described the affluent influence argument as a “consistent narrative thatpolitical representationmaybea luxuryavailable to the wealthy alone.” We instead find that the partisan distortion dominates—party beats the purse.

While the affluent dominance model is descriptively correct—in that the rich do get what they want more often than the median voter or the poor—this seems as coincidental as the oft-dismissed coincidental repre- sentation of the poor. Combining the relatively mod- erate pro-poor Democratic bias and the larger pro-rich bias of the Republicans, the result is a party system that over the last two decades favors the affluent, as a result

FIGURE 15. Taking Sides—Partisan Senate Peers

The Party or the Purse?

935

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

of the outcomes of partisan conflict. Republican par- tisanship is the key to understanding modern affluent influence. The one exception is a quite limited set of conflictswithin party: bothDemocrats andRepublicans side with their rich copartisans over poor copartisans.

If the rich are rigging the system, as some suggest, it would have to be through elections (electing Repub- licans who cater to Republican voters who more often agree with rich than poor), through convincing

Republican voters to favor the policies the rich like,23

through taking advantage of redistricting rules to ad- vantage Republicans, and/or through agenda control (sincewe study only votes that take place, not those that could).Wedonot address such pathways.At leastwhen

FIGURE 16. Taking Sides Summary

Summary percentages of voting with conflicting groups. Point predictions, based on median across simulated draws, for the Democratic percentage are the hollow circles and Republicans hollow squares. Point size reflects the number of votes. 99% confidence intervals are shown.

23 Bartels (2008) showed that theBush taxcutswerenotanexampleof the rich overpowering the poor. Many poor voters supported the tax cuts. Our results in a sense generalizes this finding.

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

936

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

it comes to one important pathway—how senators vote—we find the rigged system claim overstated.

In fact, if the rich did control how senators vote—if every senator voted in line with his or her state’s rich median, then instead of only the observed 58% con- gruence rate overall, we would observe a whopping 91%. (If majoritarian representation is the standard, affluent influence would actually help.) If the poor dominated, congruence would be 87%. If copartisan medians dominated, congruence would be 72%. Lis- tening to public opinion of even such subgroups would be much more majoritarian than the status quo.

Even this obscures party differences, in that if every senator voted theDemocratic party line (i.e., in linewith the median Democratic senator), congruence would be 67%. Itwouldonly be 36% if every senator followed the Republican party line. The poor would be better rep- resented in policymaking ifDemocratswonmore often, in that poor medians would get what they want more often… but so toowould the rich,middle, and statewide medians. If Democrats consistently controlled the levers of power (in this time period), one would likely find little (although not necessarily no) evidence of affluent influence. Affluent influence that results from partisan influence may be worrisome, but it is not the same as living in an oligarchy.

Obviously, there remain unanswered questions that are ripe for future investigation. Building upon efforts by Ellis (2017) and our results here, one could consider why some senators’ roll call voting patterns exhibit a greater bias toward copartisan or elite opinion than others (and our “taking sides” graphs clearly show that there is intriguing variation to be explained). Such an analysis might explore electoral competitiveness, the extent of state-level economic inequality, or union strength. Itmight also consider senator political ideology or wealth. While we generally focus on differences in responsiveness across senators aggregated at the party level (finding, inter alia, Democrats to be more re- sponsive), a next step might look at variation across individual legislators.

We hope that our detailed prescription for a partic- ular approach to studying these issues and our detailed implementation of this approach will influence other such work in both theoretical approach and empirical detail, building on it—or challenging it—to collectively do our best to untangle these thorny issues.

SUPPLEMENTARY MATERIAL

To view supplementary material for this article, please visit https://doi.org/10.1017/S0003055419000315.

Replication materials can be found on Dataverse at: https://doi.org/10.7910/DVN/MCWFCS.

REFERENCES

Achen,ChristopherH. 1978. “MeasuringRepresentation.”American Journal of Political Science 22 (3): 475–510.

Achen, Christopher, and Larry Bartels. 2016.Democracy for Realists: WhyElectionsDoNotProduceResponsiveGovernment. Princeton, NJ: Princeton University Press.

Ahler, Douglas J., and David E. Broockman. 2018. “The Delegate Paradox: Why Polarized Politicians Can Represent Citizens Best.” The Journal of Politics 80(4): 1117–33.

Barker, David C., and Christopher Jan Carman. 2012. Representing RedandBlue:How theCultureWarsChange theWayCitizensSpeak and Politicians Listen. New York: Oxford University Press.

Bartels, Larry M. 2008. Unequal Democracy: The Political Economy of the New Gilded Age. Princeton: Princeton University Press.

Bhatti,Yosef, andRobertS.Erikson. 2011.“HowPoorlyAre thePoor Represented in the US Senate?” In Who Gets Represented? eds. Peter K. Enns and Christopher Wlezien. New York: Russel Sage Foundation, 285–310.

Branham, J. Alexander, Stuart N. Soroka, and Christopher Wlezien. 2017. “WhenDo theRichWin?”Political ScienceQuarterly 132 (1): 43–62.

Broockman, David E. 2016. “Approaches to Studying Policy Rep- resentation.” Legislative Studies Quarterly 41 (1): 181–215.

Broockman, David E., and Christopher Skovron. 2018. “Bias in Perceptions of Public Opinion Among American Political Elites.” American Political Science Review 112 (3): 1–22.

Brunner, Eric, Stephen L. Ross, and Ebonya Washington. 2013. “Does Less Income Mean Less Representation?” American Eco- nomic Journal: Economic Policy 5 (2): 53–76.

Buttice, Matthew K., and Benjamin Highton. 2013. “How Does Multilevel Regression and Poststratification Perform with Con- ventional National Surveys?” Political Analysis 21 (4): 449–67.

Clausen, Aage R. 1973. How Congressmen Decide: A Policy Focus. New York: St. Martin’s Press.

Clinton, Joshua. 2006.“Representation inCongress:Constituents and Roll Calls in the 106th House.” The Journal of Politics 68 (2): 397–409.

Ellis, Christopher. 2012. “Understanding Economic Biases in Rep- resentation: Income, Resources, and Policy Representation in the 110th House.” Political Research Quarterly 65 (4): 938–51.

Ellis, Christopher. 2013. “Social Context and Economic Biases in Representation.” The Journal of Politics 75 (3): 773–86.

Ellis, Christopher. 2017. Putting Inequality in Context: Class, Public Opinion, and Representation in the United States. Ann Arbor, MI: University of Michigan Press.

Enns, Peter K. 2015a. “Relative Policy Support and Coincidental Representation.” Perspectives on Politics 13 (4): 1053–64.

Enns, Peter K. 2015b. “Reconsidering theMiddle: AReply toMartin Gilens.” Perspectives on Politics 13 (4): 1072.

Erikson, Robert S. 2015. “Income Inequality and Policy Re- sponsiveness.” Annual Review of Political Science 18: 11–29.

Erikson, Robert S., Gerald C. Wright, and John P. McIver. 1993. Statehouse Democracy: Public Opinion and Policy in the American States. New York: Cambridge University Press.

Fenno,RichardF. 1978.HomeStyle:Representatives inTheirDistricts. Boston: Little, Brown.

Gelman, Andrew, Boris Shor, Joseph Bafumi, and David Park. 2007. “Rich State, Poor State, Red State, Blue State: What’s the Matter with Connecticut?” Quarterly Journal of Political Science 2 (4): 345–67.

Gelman,Andrew, and ThomasC. Little. 1997. “Poststratification into Many Categories Using Hierarchical Logistic Regression.” Survey Methodology 23 (2): 127–35.

Gelman, Andrew, and Hal Stern. 2006. “The Difference Between ‘Significant’ and ‘Not Significant’ Is Not Itself Statistically Signifi- cant.” The American Statistician 60 (4): 328–31.

Gerber, Elisabeth R., and Jeffrey B. Lewis. 2004. “Beyond the Me- dian: Voter Preferences, District Heterogeneity, and Political Representation.” Journal of Political Economy 112 (6): 1364–83.

Gerber, Elisabeth R., and Rebecca B. Morton. 1998. “Primary ElectionSystemsandRepresentation.” JournalofLaw,Economics, and Organization 14 (2): 304–24.

Giger, Nathalie, and Zoe Lefkofridi. 2014. “The Role of Persoanl Issue Salience in Citizens’ Representation via Parties.” Swiss Po- litical Science Review 20 (2): 287–304.

Gilens, Martin. 2005. “Inequality and Democratic Responsiveness.” Public Opinion Quarterly 69 (5): 778–96.

The Party or the Purse?

937

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

Gilens, Martin. 2012. Affluence and Influence: Economic In- equality and Political Power in America. New York: Russell Sage Foundation.

Gilens, Martin. 2015. “The Insufficiency of ‘Democracy by Co- incidence’: A Response to Peter K. Enns.” Perspectives on Politics 13 (4): 1065–71.

Gilens, Martin, and Benjamin I. Page. 2014. “Testing Theories of American Politics: Elites, Interest Groups, and Average Citizens.” Perspectives on Politics 12 (3): 564–81.

Hayes, Thomas J. 2012. The Representational Sources of Political Inequality. Riverside: University of California.

Hayes, Thomas J. 2013. “Responsiveness in an Era of Inequality: The Case of theUSSenate.”PoliticalResearchQuarterly 66 (3): 585–99.

Hussey,Wesley, andJohnZaller. 2011.“WhoDoPartiesRepresent?” In Who Gets Represented? eds. Peter K. Enns and Christopher Wlezien. New York, NY: Russell Sage Foundation, 311–44.

Kastellec, Jonathan, Jeffrey R. Lax, Michael Malecki, and Justin H. Phillips. 2015. “Polarizing the Electoral Connection: Partisan Representation in Supreme Court Confirmation Politics.” The Journal of Politics 77 (3): 787–804.

Krimmel, Katherine, Jeffrey Lax, and Justin Phillips. 2016. “Gay Rights in Congress: Public Opinion and (Mis)Representation.” Public Opinion Quarterly 80 (4): 888–913.

Lax, Jeffrey R., and Justin H. Phillips. 2009a. “How Should We Es- timate Public Opinion in the States.”American Journal of Political Science 53 (1): 107–21.

Lax, JeffreyR., andJustinH.Phillips. 2009b.“GayRights in theStates: Public Opinion and Policy Responsiveness.” American Political Science Review 103 (3): 367–85.

Lax, Jeffrey R., and Justin H. Phillips. 2012. “The Democratic Deficit inStatePolicymaking.”AmericanJournalofPolitical Science56 (1): 148–66.

Lax, Jeffrey R., and Justin H. Phillips. 2013. “How Should We Esti- mate Sub-NationalOpinionUsingMRP?Preliminary Findings and Recommendations.” Conference paper: Midwest Political Science Association.

Lefkofridi, Zoe, and Juan Casado-Asensio. 2013. “European Vox Radicis: Representation and Policy Congruence on the Extremes.” Comparative European Politics 11 (1): 93–118.

Matsusaka, John G. 2001. “Problems with a Methodology Used to Evaluate the Voter Initiative.” The Journal of Politics 63 (4): 1250–6.

Matsusaka, John G. 2010. “Popular Control of Public Policy: A Quantitative Approach.” Quarterly Journal of Political Science 5 (2): 133–67.

Park, David K., Andrew Gelman, and Joseph Bafumi. 2006. “State- Level Opinions from National Surveys: Poststratification Using Multilevel Logistic Regression.” InPublic Opinion in State Politics, ed. J.E. Cohen. Stanford: Stanford University Press, 209–28.

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

Rhodes, Jesse H., and Brian F. Schaffner. 2017. “Testing Models of Unequal Representation: Democratic Populists and Republican Oligarchs?” Quarterly Journal of Political Science 12 (2): 185–204.

Rigby, Elizabeth, and Gerald C. Wright. 2011. “Whose Statehouse Democracy? Policy Responsiveness to Poor versus Rich Con- stituents inPoor versusRichStates.” InWhoGetsRepresented?eds. Peter K. Enns and Christopher Wlezien. New York, NY: Russell Sage Foundation, 189–22.

Rigby, Elizabeth, and Cory Maks-Solomon. 2017. “Are the Rich Always Better Represented Than the Poor? Income- and Party- StratifiedPolicyRepresentation in theUSSenate.”WorkingPaper. Paper prepared for George Washington University Political Science Department American Politics Workshop on September 19, 2017. https://corymaks.files.wordpress.com/2017/08/rigby-maks-solomon- workshop_09-09-17.pdf.

Selb, Peter, and Simon Munzert. 2011. “Estimating Constituency Preferences from Sparse Survey Data Using Auxiliary Geographic Information.” Political Analysis 19 (4): 455–70.

Sabl, Andrew. 2015. “The Two Cultures of Democratic Theory: Responsiveness, Democratic Quality, and the Empirical- Normative Divide.” Perspectives on Politics 13 (2): 345–65.

Shapiro, Catherine R., DavidW. Brady, Richard A. Brody, and John A. Ferejohn. 1990. “Linking Constituency Opinion and Senate Voting Scores: A Hybrid Explanation.” Legislative Studies Quar- terly 15 (4): 599–621.

Soroka, Stuart N., and Christopher Wleziens 2008. “On the Limits to Inequality in Representation.” PS: Political Science & Politics 41 (2): 319–27.

Tausanovitch, Chris, 2016. “Income, Ideology, and Representation.” RSF: The Russell Sage Foundation Journal of the Social Sciences, 2 (7): 33–50.

Treier, Shawn and Simon Jackman. 2008. “Democracy as Latent Variable.” American Journal of Political Science 52 (1): 201–17.

Toshkov, Dimiter. 2015. “Exploring the Performance of Multilevel Modeling and Poststratification with Eurobarometer Data.” Po- litical Analysis 23 (3): 455–60.

Ura, Joseph Daniel, and Christopher R. Ellis. 2008. “Income, Pref- erences, and theDynamics of Policy Responsiveness.”PS: Political Science & Politics 41 (4): 785–94.

Warshaw, Christopher. 2012. “Are Legislators More Responsive to PublicOpinion on Salient Issues?”WorkingPaper Prepared for the AnnualMeeting of the American Political Science Association. New Orleans, LA. August 30–September 2, 2012. https://papers. ssrn.com/sol3/papers.cfm?abstract_id=2107282.

Warshaw, Christopher, and Jonathan Rodden. 2012. “How Should We Measure District-Level Public Opinion on Individual Issues?” The Journal of Politics 74 (1): 203–19.

Wlezien, Christopher, and Stuart N. Soroka. 2011. “Inequality in Policy Responsiveness?” In Who Gets Represented? ed. Peter K. Enns and Christopher Wlezien. New York: Russel Sage Founda- tion, 285–310.

Wright, Gerald C. 1989. “Policy Voting in the US Senate: Who Is Represented?” Legislative Studies Quarterly 14 (4): 465–86.

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

938

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

APPENDIX: THE LUMPING- SPLITTING PARADOX

Consider three senators, fromconservative stateC, liberal state L, and moderate state M, defined by average opinion across issues, along with three liberal policies to vote on. Figure 17 regresses a liberal policy index (count of liberal policies sup- ported) against a liberal opinion index (average opinion) to yield see a reassuring positive responsiveness slope.

The opinion levels and votes behind these lumpy averages are shown in Table 5.

What happens when we split the lump? Figure 18 shows “splitty” responsiveness. For each policy,

responsiveness is actually perverse, with more liberal opinion “causing” a lower chance of the liberal policy. It is not that we

expect such anti-responsiveness, but rather that evenperverse split responsiveness is compatible with the appearance of lumpy responsiveness. Lumpy responsiveness is not sufficient for split responsiveness.

Next, consider a different set of three policies and opinion levels, graphing responsiveness with a splitting approach in Figure 19.

Each split responsiveness curve looksnormal: higher liberal opinion positively associates with having the liberal policy. Table 6 shows opinion andpolicy details, andFigure 20 graphs lumpy responsiveness.

Now, we find the appearance of perverse lumpy re- sponsiveness. Lumpy responsiveness is thus neither sufficient nor necessary for true responsiveness.

FIGURE 18. Splitting the Lump (from Table 5)

FIGURE 17. Lumping Three Policies (from Table 5)

TABLE 5. Splitting the Lump

State Policy

1 Policy

2 Policy

3 Opinion index

Policy index

L 99 453 453 63 2 M 413 65 65 57 1 C 55 55 55 55 0

The Party or the Purse?

939

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss

FIGURE 19. Splitting Three Policies (from Table 6)

TABLE 6. Splitting

State Policy 4 Policy 4 Policy 6 Op. avg. Pol. ind.

L 45 45 45 45 0 M 553 35 36 42 1 C 01 553 643 40 2

FIGURE 20. Lumping the Split (from Table 6)

Jeffrey R. Lax, Justin H. Phillips, and Adam Zelizer

940

ht tp

s: //

do i.o

rg /1

0. 10

17 /S

00 03

05 54

19 00

03 15

P ub

lis he

d on

lin e

by C

am br

id ge

U ni

ve rs

ity P

re ss