Distorting the Electoral Connection? Partisan Representation in Supreme Court Confirmation Politics

Transcription

1 Distorting the Electoral Connection? Partisan Representation in Supreme Court Confirmation Politics Jonathan P. Kastellec Dept. of Politics, Princeton University Je rey R. Lax Dept. of Political Science, Columbia University Michael Malecki Applied Statistics Center, Columbia University Justin H. Phillips Dept. of Political Science, Columbia University April 25, 2013 Abstract Do senators respond to the preferences of their state s median voter or only to the preferences of their co-partisans? We develop an approach for distinguishing between general and partisan responsiveness, and we develop a method for estimating statelevel public opinion broken down by partisanship. We use these estimates to study responsiveness in the context of Senate confirmation votes on Supreme Court nominees. We find that senators more heavily weight their partisan base when casting such roll call votes. Indeed, when their state median voter and party median voter disagree, senators strongly favor the latter. This has significant implications for the study of legislative responsiveness, the role of public opinion in shaping the personnel of the nation s highest court, and the degree to which we should expect the Supreme Court to be counter-majoritarian. The methodological approach developed in the paper greatly expands the scope of sub-state opinion estimation, and incorporates full uncertainty surrounding our opinion estimates. It can be applied elsewhere to estimate opinion by state and partisan group, or by many other typologies, so as to study other important questions of democratic responsiveness and performance.

2 1 Introduction Whom do legislators represent? Most scholars of public opinion and legislative behavior agree that constituents preferences shape the behavior of their representatives in Congress (Mayhew 1974, Arnold 1990, Erikson, Mackuen and Stimson 2002). There is, however, less consensus about whose opinion matters. Are some constituents better represented than others? Are lawmakers most responsive to the median voter or to subconstituencies, particularly their own partisans? The answers to these questions are important for understanding the quality of American democracy if members of Congress are primarily (or only) responsive to their same-party constituents, it raises normative concerns of democratic performance and has implications for the study of legislatures, elections, and other features of American politics. That the electoral connection is the dominant motivator for members of Congress can be reassuring on normative grounds, even if self-interest is the key mover if seeking reelection leads to representative legislative choice. But if representation is distorted by paying more attention to one s fellow partisans than the typical voter back home, then the electoral connection has its limitations as well. The possibility that lawmakers are most responsive to their co-partisans has long been recognized (Fenno 1978). Still, there is little systematic evidence that co-partisans actually have such influence relative to average voters. The co-partisan hypothesis is di cult to test due to the absence of accurate measures of preferences across subconstituencies. Researchers have often compensated by using demographic and economic proxies or by using di use survey-based measures such as average preferences across a range of policies or self-identified ideology. Such measures, while often rigorously constructed, can be problematic in two ways. First, they do not directly measure constituent preferences on the specific roll call votes being studied. Second, they do not share a common metric with roll call votes, seriously limiting the inferences that can reasonably be drawn. The di culty of measuring subconstituency opinion means that the question of whose opinion matters is far from settled. Indeed, the 1

3 existing literature has produced a set of conflicting findings. We overcome the existing methodological limitations by generating estimates of opinion on specific votes broken down by subconstituency in particular, partisan subconstituencies within a given state. To do so, we build on recent advances in opinion estimation specifically, the technique of multilevel regression and poststratification [MRP]. We develop an extension of this method that allows more fine-grained estimates of public opinion by subgroup in our case, a senator s in-party, opposite-party, and independent constituents. We use this innovation to conduct a fine-grained study of responsiveness and representation, focusing on how senators cast votes on Supreme Court nominees. Whereas most prior work uses aggregated indices of voting behavior across across a wide range of issues, we connect senatorial roll call votes to roll-call specific subconstituency preferences. Since our opinion estimates and roll call votes are on a common scale, we can not only estimate the strength of the relationship between opinion and senatorial vote choice by subconstituency, but also how often a senator s vote is congruent with the preferences of same-party, oppositeparty, and independent voters. This linkage generates more nuanced assessments of responsiveness in nomination politics than was previously possible. The technology we develop in this paper creating sub-state estimates when Census data necessary for basic MRP is not available will be useful for many further applications of MRP and for studying a wide range of substantive questions. In particular, it will make possible the generalization of our substantive research on nomination voting to other types of votes. From a substantive perspective, the question of who gets represented is most important when evaluating key votes cast by legislators these votes are likely to have a lasting impact on their constituents. Few decisions are as consequential for and visible to the public as their votes to confirm or reject a nominee to the United States Supreme Court. While the outcomes of many votes are ambiguous or obscured in procedural detail, the result of a vote on a Supreme Court nomination is stark: either the nominee is confirmed, allowing her to serve on the nation s highest court, or she is rejected, forcing the president to name another 2

4 candidate. Once a justice is confirmed, she serves, with life tenure, on the world s most powerful Court, whose policy reach extends to everything from the enforcement of contracts to whether a convicted murderer lives or dies. Thus, every nominee has the potential to shift the direction of the Court s jurisprudence (Cameron and Park 2009, Krehbiel 2007). From a research design perspective, votes on Supreme Court nominees are valuable because constituents from across the political spectrum care deeply about their outcomes (Gimpel and Wolpert 1996, Gibson and Caldeira 2009), several national public opinion polls are usually conducted for each nominee, and public opinion can vary widely across states and nominees and has been shown to influence senatorial voting on nominees (Kastellec, Lax and Phillips 2010). Our previous work has demonstrated that senators respond to state opinion on Supreme Court nominees. However, it did not (and could not, given existing methodological approaches) evaluate whether senators respond di erently to di erent constituencies. In this paper, we document that opinion on Supreme Court nominees frequently varies across partisan constituencies. Given this divergence, whether senators follow the preferences of the median voter or their co-partisans can mean the di erence between a vote to confirm and a vote to reject. Using these opinion estimates, we show clearly and robustly that senators heavily weight the opinion of their fellow partisans. After controlling for ideology and party, we find that Democrats still listen more to Democrats and Republicans more to Republicans. Simply changing the composition of a nominee s supporters (as opposed to her overall level of support) can have striking e ects on the likelihood that a senator votes to confirm the nominee. Indeed, we find that when the preference of the median voter and the party median di er, senators side with their co-partisans 88% of the time. Thus, our paper makes both amethodologicalandsubstantivecontribution: themethodwedeveloptoanalyzeroll-call specific opinion leads us to the conclusion that the extra weight given to partisan subconstituencies filters responsiveness to the public will in ways that are troubling to normative democratic theory. 3

5 2 Subconstituencies and legislators: theory and measurement The usual starting point for assessing the linkages between voters and legislators the Median Voter Theorem predicts that if candidates and representatives are motivated solely by o ce-seeking, they will locate themselves at the ideal point of the median voter of the lawmaker s constituency (Downs 1957). The Downsian model thus predicts that representation should follow the average voter and not any other constituency that diverges from the preference of the median. However, as discussed in Clinton (2006), much empirical evidence suggests that lawmakers often do not converge to the median voter. For example, House candidates from the same district often adopt divergent ideological positions (Ansolabehere, Snyder and Stewart 2001), and same-state senators frequently disagree (Bullock and Brady 1983, Krehbiel 1993). The theoretical literature on representation provides several reasons why the empirical predictions of the Downsian model generally are not supported. First, if candidates and politicians care not just about winning elections but also about implementing their preferred policies, they will have incentives to diverge from the median voter (Wittman 1983). Second, the desire to please party activists and interest groups who are more ideologically extreme than the average voter may induce divergence (Aldrich 1983). Third, the fact that o ce holders are usually members of a political party may induce both representatives and parties to adopt more extreme positions to advance the party s brand (Aldrich 1995, Snyder and Ting 2002). Fourth, the fact that challengers and incumbents often must first win a primary election before running in a general election may lead an o ce holder to favor her primary constituency over the median voter (Owen and Grofman 2006, Hirano, Snyder Jr and Ting 2009) especially if she serves in a jurisdiction with a closed primary in which only selfidentified partisans may vote (Gerber and Morton 1998). The empirical literature on this question (with respect to representation, rather than elections) has largely flowed from Fenno s (1978) canonical work on how members of Congress 4

6 respond to di erent subconstituencies. Whereas the median voter can be thought to represent what Fenno calls the geographical constituency (i.e. the entire district or state), members of Congress will also focus on both the reelection constituency and the primary constituency. The former comprises people in a district or state that a member thinks will vote support her, while the latter comprises a subset of these voters those that are the member s strongest supporters. These supporters, of course, are most likely to be members of the legislator s party. There are several reasons to expect that members of Congress will, on balance, favor the primary constituency over the geographic one. As noted above, members face the task of winning a primary before running in the general election; voters who participate in primaries are more likely be ideological extreme relative to those who participate in general elections, which may pull representation towards the primary (and thus partisan) constituency (Wright 1989, 468). Relatedly, if there exists a high degree of preference hetereogenenity across a state or district, it may be di cult to accurately represent the median vote. In contrast, partisans are more homogenous, probably more communicative, and hence easier to represent than than the full constituency (Wright 1989, 469). Finally, candidates may seek to vote more extremely than the median voter would want in order to mobilize core supporters or activists, and to win (or keep) the approval of key interest groups (Adams and Merrill 2003, Miller and Schofield 2003). Yet, despite conventional wisdom and these strong reasons to believe that non-median representation exists, there remains is no clear empirical answer to the question of whom legislator represents. This literature is too vast to summarize in detail, but it is worth providing a brief sense of the conflicting findings on this question. Examining the responsiveness of senators to di erent constituencies, Shapiro et al. (1990) find that senator s votes are strongly related to the preferences of their in-party constituents, while (Wright 1989) finds that same-party preferences have no direct e ect on representation. More recently, Clinton (2006) finds that House Republicans in the 106th Congress were strongly responsiveness to 5

7 the preferences of Republicans in their districts. However, he finds not only that Democrats do not follow the preferences of their partisan constituency, but actually seem to be more responsive to Republicans. Finally, in an innovative study of representation among three legislative bodies that represent California, Gerber and Lewis (2004) finds responsiveness to the median voter (especially when preferences within a district are more homogenous), but no e ect of in-party preferences on members voting behavior. Methodological challenges To be sure, all these studies use di erent datasets and ask di erent questions. But one possible reason for the lack of clarity in the empirical literature is that testing expectations of di erential representation raises several methodological concerns. Foremost among these is the di culty of accurately measuring the preferences of various subconstituencies. This challenge arises from a harsh constraint the frequent lack of comparable public opinion polls across states or congressional districts. To compensate for this, scholars have pursued several alternatives, each with its own limitations. Early empirical research often used demographic and economic data as proxies for policy preferences (see e.g. Peltzman 1984). Recognizing the limitations of this approach, recent analyses have transitioned to survey-based measures of preferences. These measures are typically created by disaggregating respondents from national polls so that opinion percentages can be calculated for each state or district. To generate adequate sub-sample sizes, either many national surveys must be pooled (sometimes over numerous years) or very large surveys must be found. This severely restricts the type of preference measures that can be constructed (often in ways that make it di cult to accurately gauge the relative influence of di erent constituencies). Studies that have examined the relationship between legislators and constituency opinion have therefore relied on general measures of preference aggregated across hundreds or even thousands of votes covering various types and issues. There are some limitations to this particular approach. First, responses are not directly matched with relevant roll call votes. Instead, an assumption is made that voters who hold 6

8 liberal, moderate, or conservative opinions on one set of policies will do so on the set of roll call votes being analyzed. However, other research has shown that survey respondents often hold ideologically inconsistent preferences across policy areas (Converse 1964). Furthermore, without accurate measures as to how voters want specific roll calls to be cast, there is not a common metric for opinion and votes, limiting the inferences that we can draw. A high correlation between roll call votes and the policy liberalness of a senator s same-party constituency reveals a strong relationship between the two, but it does not allow us to conclude whether same-party constituents are actually getting their senator to vote the way they want more often than the median voter or opposite-party constituents. As Clinton (2006, 407) notes, the inability to measure subconstituency preferences and voting behavior on a common scale prevents a definitive answer we simply cannot see which constituency is closer to the legislator s revealed preferences. Finally, most papers in this literature aggregate many di erent types of votes. To be sure, pooling hundreds of types of votes also has its advantages, in that idiosyncrasies or details of any one policy area are averaged out. However, examining the general relationship between constituent opinion and roll call voting means that the two cannot be directly compared, complicating any analysis of representation (Bishin and Dennis 2002). Evaluating Supreme Court nominations The politics of Supreme Court nominations illustrate both the importance of adjudicating between median and non-median theories of representation and the methodological di culties in carrying out such adjudication. Kastellec, Lax and Phillips (2010) show that senators respond to state-level public opinion when casting roll call votes on Supreme Court nominees. This finding seemingly ties the Supreme Court, a potentially counter-majoritarian institution, back to majority will. However, this study does not explore to whom senators respond. If senators respond with special attention to particular subconstituencies, this would undercut the majoritarian linkage once again. Which subconstituencies in Supreme Court confirmation politics are likely to influence 7

9 senators? One possibility is racial or ethnic groups. For example, public opinion among African-Americans and Hispanics loomed large in the politics surrounding the respective nominations of Justice Thomas in 1992 and Justice Sotomayor in 2009 (Overby et al. 1992, Bishin 2009). In general, however, given the overall importance of partisanship in the Senate confirmation process (Epstein et al. 2006, Shipan 2008), and for the theoretical reasons discussed above, we would expect the views of partisan subconstituencies to play an important role in senators voting decisions. Perhaps most importantly, primary elections allow such challengers to attack incumbents who do not heed their partisan constituents opinion. Indeed, Senate lore contains ominous warnings on this front. Despite being virtually unknown, Carol Moseley Braun defeated incumbent Senator Alan Dixon in the Illinois Democratic primary, principally campaigning against his vote to confirm Clarence Thomas (McGrory 1992). Testing whether senators respond more to the median voter or their in-party median requires us to generate nominee-specific estimates of public support, broken down by partisan constituencies. In doing so we must overcome the methodological limitations outlined above. Specifically, we need to have measures of subconstituency policy preferences that relate directly to roll call votes on Supreme Court nominees and that are on the same scale. 3 Data and Methods Estimating state-level and constituency-level opinion To evaluate the role of subsconstituency opinion on roll call voting on Supreme Court nominees, we estimate opinion by party for 11 recent nominees: Rehnquist (for Chief Justice in 1986), Bork (1987), Souter (1990), Thomas (1991), Ginsburg (1993), Breyer (1994), Roberts (2005), Miers (2005), Alito (2005), Sotomayor (2009), and Kagan (2010). 1 Each of 1 These nominees are the only ones for whom su cient polling data is available. For nominees who featured in only a handful of polls, we gathered every poll containing su cient demographic and geographic information on individual respondents. For nominees with a large number of such polls, we only used the polls closest to their confirmation vote. For Thomas, we only retained polls taken after the Anita Hill allegations surfaced. This procedure helped ensure as much as possible that our estimates would tap state opinion as it stood at the time a senator cast his vote. 8

10 these nominees was eventually confirmed except Bork, who was defeated in a floor vote, and Miers, whose nomination was withdrawn before a vote could take place. To generate the required measures of public opinion, we employ Multilevel Regression and Poststratification, or MRP, a technique developed and assessed by Gelman and Little (1997), Park, Gelman and Bafumi (2006), Lax and Phillips (2009a), and Lax and Phillips (2013). It combines detailed national survey data and Census data with multilevel modeling and poststratification to estimate public opinion at the subnational level. Importantly, it can generate accurate estimates of state or district-level opinion using a relatively small number of survey respondents as few data as contained in a single national poll so that we do not need to limit ourselves only to those questions that are asked across many large national polls (as the existing literature is forced to do). There are two stages to MRP. In the first, individual survey response is modeled as a function of demographic and geographic predictors included in the survey data, with individual responses nested within states, which are in turn nested within regions. The state of the respondents is used to estimate state-level e ects, which themselves are modeled using additional state-level predictors such as region or state-level aggregate demographics. Those residents from a particular state or region yield information on how responses within that state or region vary from others after controlling for demographics. All individuals in the survey, no matter their location, yield information about demographic patterns which can be applied to all state estimates. The second stage is poststratification: the estimates for each demographic-geographic respondent type are weighted (poststratified) by the percentages of each type in actual state populations, so that we can estimate the percentage of respondents within each state who have a particular issue position. As previous evaluations have demonstrated, MRP performs very well in generating accurate state-level estimates of public opinion (Gelman and Little 1997, Park, Gelman and Bafumi 2006, Lax and Phillips 2009a;b, Pacheco2009). Itconsistentlyoutperformsraw state breakdowns, even for large samples, and it yields results similar to actual state polls. 9

11 Asinglenationalpollandsimpledemographic-geographicmodels(indeed,simplerthanwe use herein) su ce for MRP to produce highly accurate and reliable estimates. MRP compensates for small within-state samples by using demographic and geographic correlations. To put this another way, there is much information within surveys that is typically thrown away, in that the demographic and geographic correlations capture patterns which MRP relies on. If the response model captures variation in survey response, if predictive accuracy is high enough and enough variation is captured by it, then the MRP process can produce accurate estimates, particularly given the detailed weighting information from Census data. The response model need not be substantively useful; it must simply allow for su ciently accurate predictions of response. Furthermore, since we will, as we discuss below, incorporate uncertainty from our response models in our estimates of opinion and throughout the analysis, we can show that our results do not depend on assuming we have perfect models of response. In general, we follow the suggestions for MRP laid out in Lax and Phillips (2013). Aproblemarises,however,inusingMRPtodevelopestimatesofopinionbypartisanship. A standard use of MRP is su cient to generate state-level estimates of opinion. It cannot, however, estimate support across partisan constituencies. This is because the second stage of MRP involves poststratifying the estimates based on the Census 5-Percent Public Use Microdata Sample s population frequencies for every demographic-geographic type (e.g. college-educated Hispanic males aged in New Jersey). Unfortunately, this Census data does not include partisan identification. Thus, using basic MRP, as in Kastellec, Lax and Phillips (2010), one can estimate the level of support for, say, Samuel Alito among college-educated Hispanic males aged in New Jersey, but one cannot estimate the level of support among Republican, Independent or Democratic college-educated Hispanic males aged in New Jersey. This means that using standard MRP to generate fine-grained estimates by variables not gathered by the Census (such as party or religion) is not possible. We have, however, devised a method for doing so. Full details of the procedure are given in the Appendix, where we explain how all estimates are produced. Here we give a 10

12 brief overview. Our approach involves using an additional level of MRP to generate a more nuanced poststratification file than one can obtain from the Census alone. We begin with a large dataset of individual survey responses about partisan identification (i.e. whether a respondent is a Democrat, Republican, or an Independent) and then model partisanship as a function of demographic and geographic variables. In other words, we treat partisanship as a response variable and apply standard MRP to estimate the distribution of partisanship across the full set of demographic-geographic types (4,890 in all). This allows us to ultimately create a much more expansive poststratification file consisting of 14,688 partisandemographic-geographic types. In short, the extra level of MRP provides us with an estimate of the information that would be readily estimated via standard MRP if the Census data included partisan identification. With these estimates of demographic-geographic-partisanship types, we can thus proceed with MRP on survey responses on Supreme Court nominees, generating simple estimates of state-level opinion but using a more finely grained weighting scheme to estimate the opinion of all three partisan subconstituencies in each state. Beyond the substantive payo we can get in this paper, importantly, this process extends earlier work on generating model-based estimates of opinion from national polls by tripling the number of subgroups available for analysis, and by allowing for partisanship to enter MRP estimates, which is crucial given the role of partisan identification in opinion formation on many issues. Moreover, this methodology could be applied to any feature of individuals for which Census data is not available but which is included in polling data, so long as there are su cient data for the purpose. At the same time, because we are estimating the distribution of partisanship across the Census types using a model, this procedure comes with inherent uncertainty. We incorporate this uncertainty into our analyses, not through analytical means but through empirical means, by generating empirical distributions of results throughout and carrying forward the accumulated uncertainty from all stages in turn. That is, we use the method of propagated uncertainty, simulating uncertainty from each stage and pushing it through the rest of the 11

13 analysis. We take our initial models of response, generating not just a set of point predictions but rather generating many sets of coe cients and opinion estimates, with the distribution of them capturing uncertainty on the basis of variances and covariances, each time combining the uncertainty of previous stages of analysis with the added uncertainty from subsequent stages, until we wind up with a distribution of results for our ultimate substantive analysis that reflects all the underlying uncertainty. Visualizing subconstituency opinion We begin our exploration of the opinion estimates in Figure 1, which depicts kernel density plots of our estimates of support among opinion holders, broken down by Democrats, Independents, and Republicans, across states. Republican and then Democratic nominees are ordered by increasing state-level mean support. That is, the unit of analysis is states, broken down by each type of opinion (so each density plot summarizes 50 estimates of opinion). The dots under each distribution depict the mean of that respective distribution. The vertical dashed lines depict median state-level support. Not surprisingly, the graph reveals that support for nominees is always higher on average, and indeed very high in absolute terms, among constituents from the president s party. Figure 1 also reveals that polarization defined as the di erence between median Democratic and Republican opinion varies significantly across nominees. Recent nominees Miers, Alito, Sotomayor, and Kagan generated large divisions of opinion, as did Bork. At the other end of the scale, the nominations of Souter, Ginsburg, and Breyer genereated little polarization and substantial overlap across constituencies. We observe the widest di erences within party for the nomination of Rehnquist to become Chief Justice. Figure 2 shows how opinion varies across both states and constituencies, as well as the degree of uncertainty in our opinion estimates. In the interest of space, we focus on two nominees: Alito and Kagan. For each nominee, the top panels depicts our estimates of statelevel opinion (among opinion holders) in each state; the bottom panels are broken down by Democratic, Independent and Republican opinion. The vertical lines connect the median 12

14 estimate for each state (for the respective constituency). We also depict the uncertainty in the estimates: for each constituency and state, we plot the 95% confidence interval for each set of estimates in the form of an empirical distribution of opinion estimates: to depict each distribution, we plot the inner 95% of estimates as translucent dots such that the darker regions depict the center of the distribution and lighter region depicting the tails. For example, Republican support for Alito is more precisely estimated than the other subgroups for Alito and even than Democratic support for Kagan. On the other hand, we more precisely estimate overall support for Kagan than for Alito. For each nominee, the states are ordered from lowest levels of state support to highest. (Note that ordering by overall opinion by state is not the same as ordering by opinion for independents; this is because subgroups are weighted by their size to form the overall state distributions.) There is substantial variance in opinion within the same constituency but also across states; that is, moving from the more conservative to liberal states. However, the variance across parties is even larger, with Democrats and Republicans far apart from each other in every state (for these nominees). Taken together, Figures 1 and 2 illustrate if senators respond di erently to partisan constituencies the e ects on roll call voting would potentially be quite consequential. Modeling roll-call votes We now move to a multivariate analysis of roll call voting on Supreme Court nominees by individual senators, so that we can control for other influences. Excluding Miers, and after abstentions, a total of 1,090 confirmation votes were cast on our remaining ten nominees, 73% of which were to confir the nominee. 2 Our key tests evaluate how the probability of a confirmation vote changes as subconstituency opinion increases or decreases. Doing so requires careful accounting of not just nominee support by a particular group, but also potentially the size of that group. To illustrate our opinion measures, consider public opinion in Ohio on the confirmation of Justice Sotomayor. We limit the denominator to those with an opinion, which is 82.5% of Ohioans 2 Roll call and other data for all nominees except Alito, Sotomayor, and Kagan come from Epstein et al. (2006); we collected data on the latter nominees. 13

15 (this is from our median across simulated samples). Of those with an opinion, 33.3% are Democrats, 83.8% of whom say confirm; 32.2% are Republicans, 23.6% of whom say confirm, and 34.6% are Independents, 50.6% of whom say confirm. Of all Ohio opinion holders, 53.0% support confirmation. Our main set of models measures supporters as the share of state opinion holders who support the nominee. A one-unit shift means that 1% of state opinion holders who fall into a particular category, such as constituents in a senator s party, switch from non-support to support. This shift is relative to the size of the state s opinion-holding population; what share of the party population this is depends on party size. That is, this unit shift flips afixedshareofthestatepopulation,butanunfixedshareofthepartypopulation. (One cannot scale to both at the same time.) Consider Senator Voinovich, the Republican senator from Ohio in 2009, for example. A unit shift in support consisting only of in-party opinion holders means that 1% of the total number of opinion holders in Ohio switch from no to yes, where the switchers consist only of Republicans. Support goes from 53.0% to 54.0% overall in Ohio, but only Republicans change, so that this shift means that 3.1% (= ) of Republicans move from no to yes, increasing support among Republicans from 23.6% to 26.7%. Next, consider Senator Sherrod Brown, the Democratic senator from Ohio in Now, a unit shift in opinion holder support consisting only of Democrats still moves total support in Ohio from 53.0% to 54.0%, but this means that 3.0% (= )ofdemocrats shifted from no to yes (83.8% becomes 86.8%). The unit shift in opinion holders correlates to a di erent size share within party because party sizes di er. 3 3 We have checked our results by operationalizing public opinion and partisan composition in two ways, which are similar but require a di erent interpretation of key variables. We stick to one of these methods but have verified results using the second. The alternate models (not shown) invokes the percentage of the relevant group who support the nominee; for instance, the percent support among opinion holders in the senator s party. Here, the scale is to the in-party size directly, but not to the state s opinion holding population as a whole. A one-unit shift in in-party percentage support means a shift of 1% of such constituents but the actual share of opinion holders overall captured by such a shift depends on party size. So, to be clear, this unit shift flips an unfixed share of the state (opinion-holding) population, but a fixed share of the party population. Now, for Voinovich, if in-party support increases by one unit, this means that 24.6% of Republicans say yes instead of 23.6%. For Brown, a unit shift in-party means 84.8% of Democrats say yes instead of 83.8%. But this 1% of Republicans correlates to 3.1% of Ohioans, while the 1% of Democrats correlates to 3.0% of Ohioans. It is always possible to translate changes in one type of unit to the other, but the 14

16 Our main predictors are defined as follows: Supporters out of all opinion holders: thepercentageofopinionholdersthatsupportthe nominee. Supporters in senator s party: the percentage of opinion holders that share the party a l- iation with the senator in question and support the nominee (SSP < PSP, by definition). In some models, we include the following predictors to separate independent supporters from opposite-party supporters: Independent supporters: the percentage of opinion holders who are Independents and support the nominee. We also account for the size of each partisan group, as follows: Percentage of opinion holders in the senator s party: the percentage of opinion holders that share the senator s party a liation. Percentage of opinion holders in the opposite party: thepercentageofopinionholdersthat do not share the senator s party a liation. Based on the existing literature on Supreme Court confirmation voting, we also include several additional predictors as control variables. These include: nominee quality, ideological distance between a senator and a nominee, partisanship, and presidential strength (Epstein et al. 2006). These studies show that senators are likely to support high quality nominees and less likely to support ideologically distant nominees. Senators are also more likely to support nominees appointed by presidents of the same party, and by presidents with greater popular support. We add an additional control for state ideology. Details on each predictor are as follows: Quality: The degree to which a nominee is judged to be qualified to join the Court (according to an ideologically balanced set of newspaper editorials (Cameron, Cover and Segal 1990)). It ranges from 0 (least qualified) to 1 (most). Ideological distance or location: Ideological distance between the senator and the nomconversion depends on party sizes in the state. The partisan constituency di erences we find are robust to which measure is used. 15

17 inee, or the location of one or both. For senators, we use their dw-nominate scores. For nominees, we employ the scores developed and used in Cameron and Park (2009) and Cameron, Kastellec and Park (2013). The authors use the past experience of each nominee (e.g. whether they served in Congress) to develop Nominate-scaled perception scores, which e ectively place nominees on the same scale as senators (see Cameron and Park (2009, ) for further details.) For models that make use of location rather than distance various specifications, we flip the location measure around its mean so that higher values mean greater distance from the ideology of the president. Senator in president s party: Coded 1 if the senator is a co-partisan of the president. Presidential capital: Weusetwomeasurestocapturepresidentialcapital.Thefirst,Strong president, is coded 1 if the president was not in his 4th year of o ce and his party controlled the Senate at the time (Cameron, Cover and Segal 1990). We also use the more direct measure of public approval of the president, Presidential approval, basedonthemostrecent Gallup poll taken before a nominee s confirmation vote. State voter ideology: Wecontrolforthepossibilitythatsenatorsrespondtodi usestatelevel ideology (rather than nominee-specific opinion) by including updated scores created by Erikson, Wright and McIver (1993). We recode this variable to match whether nominees are liberal or conservative (i.e. nominated by Democratic or Republican presidents, respectively), such that higher values indicate greater ideological support for the nominee. If the appointing president is a Democrat, higher numbers mean a greater percentage of liberal voters in a state; if the appointing president is a Republican, higher numbers mean a greater percentage of conservative voters. Thus, higher values of our recoded measure should always increase the probability that the senator votes to confirm a nominee. 4 Models and Results To evaluate the systematic e ect of subconstituency opinion on roll call voting, we estimate a series of logit models. The models vary in several ways. First, in some models we split opinion 16

18 into two components: in-party opinion versus not-in-opinion party. In other models, we break opinion down into three components: in-party, independent, and out-party. Second, we vary the way in which we estimate the e ect of senator and nominee ideology. In some models, we look only at the location of the senator, while in others we employ the distance between the senator and the nominee. Third, in some models we employ random e ects to estimate varying intercepts for each nominee this allows us to include nominee-level predictors such as quality and presidential popularity. In other models, we employ nominee fixed e ects, which ensures that our results are not a ected by unobserved heterogeneity across nominations. Finally, in some models we take our point estimates of nominee support as fixed; in other we incorporate the full uncertainty of our estimates into the model estimation. This allows us to gain a sense of whether the uncertainty in the opinion estimates influences our results. All in all, we estimate 24 models. In the interests of space, Table 1 present the model estimates from eight of these models (full results are available on request). We return to a discussion of the control predictors below. Of more substantive interest is the e ect size of changing each type of opinion, as defined above. Table 2 presents these e ects across each of the models we estimate. The table caption explains how each model varies. 90% confidence intervals for each estimate appear in parentheses. Column (1) shows the coe cient estimates on all opinion: that is, without breaking down by constituency. This simply replicates our prior findings that senators respond to state-level opinion. Column (2) depicts, for models that separate in-party, our estimates of in-party opinion e ects. That is, the e ect size here captures the e ect of raising in-party support (and therefore total support). The results across the models produce our main finding: there is indeed a large partisan constituency e ect. In-party nominee support has a strong and significant e ect on roll call voting behavior, holding constant support in the other party, support among Independents, support among both of these categories combined, or overall support. This finding remains no matter which set of control variables we include, holding 17

19 constant the party composition of the state, and varying how we operationalize opinion. Most importantly, the partisan constituency e ect remains even incorporating uncertainty in our poststratification weights that define our populations, uncertainty from our model of party ID, uncertainty from our models of nominee support, and, of course, the uncertainty from the final-stage model of roll call voting. Given the measurement error incorporated under these multiple levels of uncertainty, the substantive point predictions for e ects of opinion are smaller, but still substantively meaningful. Column (3) shows the estimates, for three-way models that separate independent and inparty and out-party opinion, of the e ect of independent opinion, which is never statistically di erent from zero, once we control for in-party opinion. In addition, in most models, the magnitude of the e ect of independents is also much smaller compared to the magnitude of the estimates on in-party opinion. Column (4) shows the e ect of either out-party opinion or not-in-party, which are both always statistically insignificant. Columns (5)-(7) depict our estimates of the di erence between the various combinations of subconstituency opinion. Most importantly, the di erence between in-party and independent opinion is always positive, and in some models the confidence interval does not include zero. However, in several models, the di erence between the two is not statistically distinguishable at conventional levels of significance (across the models, the probability that in-party opinion is greater than independent opinion is about 80-85%). Note, however, that this result is driven in large part by the fact we have the greatest uncertainty around our estimates of support by independents, which makes it more di cult to conclude statistically that there is a di erence between the two types of opinion. And, again, the di erence in magnitude in the coe cients on in-party and independent opinion is sizable in most models. To give a sense of the substantive implications of these results, consider a senator who is on the fence. Based on the results from model 5.2, suppose we flipped 1% of state opinion holders consisting of in-party constituents from opposition to support, while at the same time decreasing not-in-party support such that total support remained the same. In this scenario, 18

20 the probability of a yes vote increases by up to 4% (with that large an e ect occurring for a senator otherwise on the fence of a yes or not vote, for whom the e ect of opinion will always be largest given the logit curve). Shifting that 1% and not holding total support constant would increase the probability by only a bit more than that, given that the e ect of not-in support is substantively small and statistically insignificant. Across all the certainty models, the smallest upper bound on the e ect of an added one point of in support (again, for asenatoronthefence,andtakingthesmallestcoe cienton in support)isabout3.5 percentage points. Across the full uncertainty models, the smallest substantive e ect is over 2percentagepoints. To give a further sense of the substantive magnitude of these di erences, Figure 3 plots the e ects (using point-predictions) of adding support of di erent types (calculated based on Model 5.3). The y-axis depicts the probability of a senator voting to confirm, while the x-axis depicts support relative to the average support among in-party, out-party, and independent constituents. For each curve, as the slope becomes more steep, the e ect becomes larger. It is clear that in-party support has the largest substantive e ect moving from 10% below average to 10% above increases the probability of a yes vote from about.3 to about.9. We now return to evaluating the e ect of the control predictors. as presented in Table 1. The usual suspects among predictors of confirmation roll-call voting perform as expected throughout, with the exception of the party identification of the senator himself. Once we control for in-party opinion, senators in the president s party are not more likely to approve a nominee than senators of the opposite party, ceteris parabus. We return to this point in the discussion. Congruence between opinion and votes What is the bottom line for democratic representation given the partisan constituency e ect that we find? To answer this, we turn to a congruence analysis, measuring how often a senator s vote on a nominee matches what the median voter among opinion holders in his state wants, and how often these votes match the median voter within either the senator s 19

21 own party or opposition party. We present this information in the top part of Figure 4 (with 95% confidence intervals depicted by the horizontal line around each estimate). We find congruence with the median voter of the entire state 73% of the time. This statistic, however, obscures a big di erence in terms of partisan representation: majorities among opinion holders in the senator s own party will see their senator vote they want 86% of the time, whereas those in the opposing party will only see the vote they want 42% of the time. What happens when a senator s constituencies are in conflict? How does a senator weigh his or her competing constituencies? The bottom part of the graph depicts the percent of yes votes for all the nominees, according to which constituencies favor confirmation. The percentages under the labels in this part of the graph depict the proportion of observations that fall into each category, with the numbers of parentheses depicting 95% confidence intervals. Both the state median and party median favor confirmation around 68% of the time. When this happens, a senator votes yes 89% of the time. The party median favors confirmation and the state median does not 7% of the time when this happens, the percentage of confirmation votes (88%) is essentially unchanged from the case where both constituencies agree. That is, flipping the median voter in the state but keeping the in-party median voter as is barely reduces the chances of getting a yes vote. Conversely, the state median favors confirmation in opposition to the party median 20% of the time, and then a yes vote occurs only 27% of the time. Finally, in five percent of cases neither median favors confirmation, and there are only 5% of yes votes when it happens. In short, a nominee seeking a senator s vote would much rather have the median voter in the senator s party on her side than the median voter in the state. 4 To put this in context, consider the confirmation of Justice Sotomayor, in Thirtyfour of the senators (one-third of the Senate) who cast votes on her nomination faced conflicting constituencies. The five conflicted Democrats all voted with their party against the state median. And of the 29 conflicted Republicans, all but nine sided with their party median 4 Note that a model with only the party median correctly predicts about 10% more votes than a model with only the state median, and adding in the state median adds only negligible predictive success. 20

22 against their state median by voting against confirmation. Thus, this nomination illustrates how the partisan constituency e ect can strongly shape voting behavior on Supreme Court nominations. 5 Discussion and Conclusion Confirmation politics is responsive to the public will but not as broadly as we previously thought. Our fine-grained study of representation, focusing on votes by senators for or against the confirmation of Supreme Court nominees, reveals a distorted electoral connection. We find that senators weigh the opinion of their fellow partisans more heavily, and that this partisan constituency e ect has significant substantive e ects on voting behavior. One advantage of our approach to studying representation is that we used a subset of Senate votes that are already well understood, allowing for the inclusion of context and controls to better assess partisan e ects. While we find that senators are strongly responsive to public opinion when voting on Supreme Court nominees, their partisan loyalties filter this responsiveness in ways perhaps troubling to normative democratic theory. When voting on Supreme Court nominees, senators engage in a tradeo in representing their median constituent and median party constituent. When that choice is pulled away from the median voter, a distortion to median representation occurs on top of any distorting e ects due to the co-partisan electorate itself. That is, if senatorial candidates are chosen by somewhat more extreme electorates, that alone can mean that senators will not be perfectly representative of the state median. Thus, on top of their own ideological score, they still give more attention to their in-party constituents can mean they are pulled further away still in their voting behavior. Majority control over policy becomes more di cult when the two parties do not converge towards the median, but instead represent influences of those to either side thereof. Our results suggest how elites can become polarized by electoral incentives. Even in a relatively smooth distribution of opinion, partisan groupings that have disproportionate influence can lead to polarized voting behavior. 21