The Electoral Costs of Party Loyalty in Congress

Jamie L. Carson The University of GeorgiaGregory Koger University of MiamiMatthew J. Lebo Stony Brook UniversityEverett Young Stony Brook University

To what extent is party loyalty a liability for incumbent legislators? Past research on legislative voting and elections suggeststhat voters punish members who are ideologically “out of step” with their districts. In seeking to move beyond the emphasis inthe literature on the effects of ideological extremity on legislative vote share, we examine how partisan loyalty can adverselyaffect legislators’ electoral fortunes. Specifically, we estimate the effects of each legislator’s party unity—the tendency of amember to vote with his or her party on salient issues that divide the two major parties—on vote margin when running forreelection. Our results suggest that party loyalty on divisive votes can indeed be a liability for incumbent House members.In fact, we find that voters are not punishing elected representatives for being too ideological; they are punishing them forbeing too partisan.

In 2004, U.S. Representative Rob Simmons (R) wasreelected to his eastern Connecticut congressionaldistrict with 54% of the vote even though voters in

the district favored John Kerry over George W. Bush by a10% margin. Simmons won his Democratic-leaning dis-trict with a combination of personal appeal (he had servedas an Army intelligence officer) and a moderate votingrecord. Two years later, however, Simmons’s opponentand the Democratic Congressional Campaign Commit-tee criticized his voting record as being too loyally Re-publican for his district. Although Simmons’s 2005 partyunity score of 74% was relatively low when comparedwith most House Republicans, by the end of the cam-paign even a Republican who voted with the party 74%of the time was too much for the voters of eastern Con-necticut; Simmons lost by 83 votes.1 In the end, the only

Jamie L. Carson is Associate Professor of Political Science, The University of Georgia, 104 Baldwin Hall, Athens, GA 30602-1615(carson@uga.edu). Gregory Koger is Associate Professor of Political Science, University of Miami, 314 Jenkins Building, Coral Gables, FL33124-6534 (gregory.koger@gmail.com). Matthew J. Lebo is Associate Professor of Political Science, Stony Brook University, SBS S-749,Stony Brook, NY 11794-4392 (matthew.lebo@sunysb.edu). Everett Young is a 2009 PhD in political science from Stony Brook University,Stony Brook, NY 11794-4392 (vertie@earthlink.net).

This is a substantially revised version of a paper presented at the 2006 annual meeting of the Midwest Political Science Association. Theauthors wish to thank Ben Bishin, Justin Buchler, Stanley Feldman, Jeff Grynaviski, Gary Jacobson, Phil Paolino, Doug Roscoe, the politicalscience colloquium at the University of Miami, and three anonymous referees for comments; Ellen Key, David Perkins, and Andrew Sidmanfor research assistance; and the Dirksen Center for its financial support of this project.

1This account of Simmons’s defeat is drawn from “The Northeast: Some Prime Takeover Targets.” CQ Weekly Online (April 24, 2006):1088–89. Accessed November 7, 2007, at http://iiiprxy.library.miami.edu: 2346/cqweekly/weeklyreport109–000002161511; and KadyII, Martin. “Party Unity: Learning to Stick Together.” CQ Weekly Online (January 9, 2006): 92–95. Accessed November 7, 2007, athttp://iiiprxy.library.miami.edu: 2346/cqweekly/weeklyreport109–000002027054. As of December 23, 2009, a DCCC ad against Sim-mons can be seen at http://www.youtube.com/watch?v=DMFOptPBB7k and a Joe Courtney ad against Simmons can be seen athttp://www.youtube.com/watch?v=toyD3fnwG4I.

Republican reelected from Connecticut was Chris Shays,who eked out a 51%–48% win with a party unity score of67%.

Simmons’s defeat illustrates how party loyalty canhave severe consequences for an incumbent seeking re-election. While it is possible that a single legislative votewill have such detrimental electoral effects, legislators alsoworry that a pattern of controversial roll-call votes mayresult in defeat during the subsequent election (Arnold1990; Bovitz and Carson 2006). At the same time, legis-lators care about policy outcomes and may accept someelectoral risk to pass a bill for the sake of personal prefer-ences or their party’s overall goals (Cox and McCubbins2005). Thus, when studying voting behavior in Congress,it is important to recognize the effects of competing influ-ences on votes. To date, these effects have been examined

American Journal of Political Science, Vol. 54, No. 3, July 2010, Pp. 598–616

C!2010, Midwest Political Science Association ISSN 0092-5853

598

ELECTORAL COSTS 599

in a limited setting or in the context of a single electioncycle, but we lack a systematic examination of the elec-toral consequences of loyalty to one’s party on contestedissues.

To what extent do voters punish members of Congressfor a record of voting against constituency interests? Arelegislators penalized for voting too often with their par-ties or are they more likely to be punished for ideologicalextremity? In this article, we seek to address these impor-tant questions by examining the electoral consequencesof party unity on divisive roll-call votes in the U.S. Houseof Representatives. Previous work has emphasized the im-portance of ideological extremity in affecting vote sharefor incumbents and suggested this relationship is constantacross all districts (see, e.g., Canes-Wrone et al. 2002;Erikson 1971). We contend, however, that ideologicalpreferences are not directly related to electoral outcomes.Instead, revealed ideology (measured with NOMINATEscores) is antecedent to the more important factor of loy-alty to one’s party as reflected by party unity (voting witha majority of members of one’s party on issues dividingthe political parties).2 In our account, ideology is one ofseveral factors influencing party unity in voting, and in-cumbents’ vote share tends to decrease as their party unityincreases. Instead of this effect being constant across dis-tricts, we expect the partisan makeup of constituencies toplay a part in determining these effects.

Further, party loyalty is not exogenous to legislators’reelection prospects. Party leaders on both sides of theaisle expend considerable effort to increase the unity oftheir party but simultaneously attempt to limit the elec-toral costs of doing so. Thus, while they may twist mem-bers’ arms to win a key vote, they carefully choose whicharms to twist. Party leaders will be more likely to ask leg-islators in “safe” seats to take risks on behalf of the teamsince safe members can afford to lose a modest amount ofsupport from their constituents. In this way, leaders cananticipate future elections as they apply pressure withintheir caucuses. This implies that we should observe re-ciprocal causality, with voters responding to a legislator’sparty unity when casting their ballots and party leadersanticipating voters’ sanctions when they ask members tosupport the partisan cause. We model this strategic inter-action in our empirical analysis of the electoral effects ofparty unity.

2We recognize there are limitations of using party unity scores inanalyses of legislative behavior, especially in terms of measuringhow much influence parties exert over member actions indepen-dent of preferences (see, e.g., Krehbiel 1993). Since we are inter-ested in understanding the electoral consequences of party unityon a subset of divisive roll-call votes and not trying to measureparty influence on these roll calls, however, we do not view thislimitation as adversely affecting the scope of our analysis.

Next, we review prior research examining the rela-tionship between roll-call behavior and electoral effectsto illustrate how our view of electoral accountability hasevolved. From there, we turn our attention to the linkagesbetween these factors and argue why higher party unitymay have a conditional effect on legislators’ reelection.We then present our empirical results, which consist ofan experiment highlighting the individual-level mecha-nisms at work and analyses of the determinants of incum-bent vote share in the House from 1956 to 2004. Lastly,we review the implications of our findings for electoralaccountability and representation in Congress.

Evidence from the Literature

In a democratic system of government, few issues aremore important than accountability and representation.In his classic work examining the electoral connection inCongress, Mayhew (1974) maintains that incumbents areextremely sensitive to the potential electoral implicationsof their votes, and as a result, behave strategically whenannouncing a position on a roll-call vote. When requiredto reveal a position on a controversial or highly contestedvote, Mayhew argues that those members who are the“safest” will be more likely to vote with the leadership,while those who are not will take the politically expedi-ent position, all else equal. If legislators are strategic andgood at anticipating voters’ responses, they might nevercast a vote that costs them electoral support. However,“mistakes” on roll calls may occur, or legislators may re-spond to peer pressure and requests from party leadersby voting with their party more often than they may like.In such cases we should observe that legislators who areviewed as being “out of touch” by their constituents willbe punished in an upcoming election.

Many studies of congressional behavior examine theeffects of ideological extremity on electoral outcomes. Us-ing electoral and survey data from 1952 to 1968, Erikson(1971) found that conservatism among Republican leg-islators had a pronounced, negative effect on their votemargins. In his study of voting behavior in Congress,Kingdon (1989) concludes that members act strategicallywhen casting roll-call votes and are careful to avoid ex-pressing positions that may be viewed as too “extreme”by constituents. This is true, Fenno (1978, 142) suggests,even though most votes are not visible; legislators do notknow ex ante which votes will be important to their con-stituents or attacked by challengers, so they must act asif every vote is potential campaign fodder. More recently,Ansolabehere et al. (2001) and Erikson and Wright (2001)

600 JAMIE L. CARSON, GREGORY KOGER, MATTHEW J. LEBO, AND EVERETT YOUNG

found that candidates’ vote share is inversely related toideologically extreme voting.

In one of the most comprehensive and related anal-yses to date, Canes-Wrone, Brady, and Cogan (2002) ex-amine the relationship between members’ electoral mar-gins and their overall ideological support as reflected byADA scores and demonstrate that legislators are indeedheld accountable for their roll-call behavior. Using datafrom the 1956–96 elections, they find that incumbentlegislators tend to receive smaller electoral margins astheir ADA scores become more extreme. Moreover, theyillustrate that this effect is comparable to other determi-nants of electoral margins, including challenger qualityand campaign spending. Finally, they contend that thedistinction between “safe” and “marginal” representa-tives is tenuous at best, since this phenomenon affectsall members equally, regardless of their previous electoralperformance.3

The implicit mechanism in research on electoral ef-fects of ideological extremism is that voters disapproveof incumbents whose policy views are inconsistent withtheir own (see, e.g., Miller and Stokes 1963). In the longrun, representatives who vote like staunch liberals or con-servatives will find it difficult to be reelected unless mostof their constituents share their strong views, or unlessthey moderate their voting behavior. Given perfect infor-mation about their constituents, it is not clear why anyambitious legislator would lose for being too extreme.Ideologically extreme members could restrict their “ex-tremism” to issues their constituents care little about ordo not monitor closely. We need some explanation, there-fore, for why legislators (many of whom are risk-aversewhen it comes to voting against their constituents) wouldcast a pattern of votes that is “out of step” with voters.

One explanation for legislators’ casting electorallyrisky votes is that they are members of parties whichinduce them to cast risky votes for the collective interestsof the party. More recent work examines the relationshipbetween election outcomes and party loyalty to determinewhether partisan cooperation is a liability for legislators.Carson (2005), for instance, examines the voting patternsof legislators on key votes since the early 1970s and findsthat experienced candidates are more likely to run as anincumbent’s party unity score increases on these highlyvisible roll calls. This suggests that legislators’ high levelsof partisan loyalty on salient votes often diverge fromconstituent interests, especially when those interests differfrom the policy goals of the party leaders. As such, strong

3See Bovitz and Carson (2006) for related evidence of electoralaccountability in conjunction with legislative voting on individualroll calls in the U.S. House and Bonney, Canes-Wrone, and Minozzi(2007) for similar evidence with respect to crime policy in Congress.

challengers are more likely to run against incumbentswho have high party unity scores on key votes since suchbehavior can be framed as being out of touch with theirconstituents—especially if the legislators represent moremoderate districts.

In a recent article, Lebo, McGlynn, and Koger (2007)propose a model of “strategic party government” and in-vestigate the relationship between aggregate party behav-ior in Congress and electoral outcomes over time. Theytest their model on macrolevel data from 1789 to 2000 andfind that an increase in party unity on voting has adverseelectoral costs for both parties during this time period.Although these findings do not speak to the presence ofmicrolevel effects, they do raise a number of questionsfor students of representation and electoral accountabil-ity. Is there a connection between party loyalty on divisivevotes and how well incumbents perform at the polls? Towhat extent do constituents hold legislators accountablefor party unity above and beyond ideological extremity?If the effects of party unity at the macrolevel are indeedpresent, when and under what conditions might we ex-pect these same effects to show up at the microlevel? Itis to these questions that we turn our attention in theremainder of the article.

Linking Electoral Accountabilityand Party Loyalty Theoretically

The literature on congressional elections suggests thatmembers seeking reelection often have several distinctadvantages. They typically can raise money and outspendtheir opponents, they have better name recognition, andthey can run on their accomplishments as legislators(Jacobson 2009). Legislators also use party affiliation as a“brand name,” which offers benefits or liabilities as theyseek reelection (Cox and McCubbins 2007). More specif-ically, the benefits incumbents provide for constituentsstem from serving as a member of one of the two parties,and advantages accruing to majority party members inparticular are a function of the agenda-setting opportu-nities and powers that majority status entails (Cox andMcCubbins 2005; Finocchiaro and Rohde 2008; Monroeand Robinson 2008; Rohde 1991).

Of course, electoral considerations represent only oneof several possible determinants of congressional vot-ing (Kingdon 1989). Legislators may have personal policyviews that conflict with their constituents’ preferences. AsBianco, Spence, and Wilkerson note,

. . .to say that an electoral connection exists doesnot imply that a legislator will invariably comply

ELECTORAL COSTS 601

with constituent demands. A legislator who hasintense policy concerns may decide to risk re-election by voting against constituents in orderto promote a preferred policy. (1996, 151)

When a member’s policy-driven votes reflect a generalapproach to policy issues, we can label that ideology andevaluate whether a member is too liberal or conservativefor her district. On the other hand, party leaders mayask legislators to support the party’s position on contro-versial and tightly contested issues that might be diffi-cult to explain in their home districts. This is most likelywhen the party position is inconsistent with, or unrelatedto, the representative’s ideology. If a legislator is pulledin multiple directions by competing electoral and policyconsiderations, she must often make a tough choice ona roll call, especially since it remains to be seen whetherthat position will prove consequential in the upcomingelection.

What happens when legislators face contradictorypressures from their constituents and party leaders? Ifelected from a district that largely overlaps with the un-derlying preferences of the party, legislators have littleneed to worry that supporting the party on highly visibleor closely contested votes will adversely affect their reelec-tion chances. When the preferences of their constituencydo not overlap with those of their political party, how-ever, then their electoral status becomes more uncertain.These cross-pressured members may be called upon tomake tough choices on important votes in Congress. Ifthey vote with the party on controversial or highly salientissues, they risk alienating their political base in the nextelection. But, if they repeatedly vote in line with theirdistrict and against the party, then they may lose favorwith the party leadership and risk sanctions (Cox andMcCubbins 2005). Regardless of their choice, they in-crease the risk of isolating one of their two core bases ofsupport.

The preceding discussion raises an important ques-tion for students of legislative politics: What exactly isthe type of behavior most likely to upset voters? To besure, an overall pattern of legislative voting that is viewedas too dissimilar from constituent interests can result ina premature departure from Congress. As noted above,many prior studies have sought to explain the effects ofroll-call voting on election results in terms of the under-lying ideology of votes cast by a legislator. Yet, what weknow about the individual psychology of voters tells usthat voters neither understand ideological concepts likeliberalism and conservatism very well nor do they have aneasy time placing events in theoretical terms like ideology(see, e.g., Converse 1964; Sniderman, Brody, and Tetlock1991, 95–96, 188).

Perhaps more central to voters in congressional elec-tions are concepts of partisanship. Voters may be muchmore comfortable identifying themselves as Democratsor Republicans than they are as liberals or conserva-tives (Bartels 2000; Campbell et al. 1960; Converse 1964;Kinder 1998). If so, they would have a much stronger re-action to political actions taken by a legislator classifiedin terms of the legislator’s partisan opposition than his orher ideological opposition. Since “members of Congressare far less familiar than the president to most voters,it is easy to accept the proposition that party identifica-tion is a far more important determinant of congressionalthan of presidential voting” (Mann and Wolfinger 1980,619). In the current era of highly polarized politics, in-group/outgroup dynamics may be more prevalent withrespect to Democrats versus Republicans than to liberalsversus conservatives. If partisan considerations motivatecitizens’ voting decisions, legislators may pay steep elec-toral costs for a pattern of extreme party loyalty.

If voters place greater weight on partisanship than onideology when evaluating behavior in Congress, then pastresearch has largely mischaracterized the connection be-tween roll-call voting and citizen evaluation. Many studieshave used roll calls to understand congressional structureand behavior but have mostly framed these votes in ideo-logical terms. While ideology and partisanship are closelyrelated—that is, ideology can predict partisanship—it isperhaps the latter concept, and not the former, that is themore direct cause of electoral outcomes.4 If so, the par-tisan nature of roll calls should help us understand howvoters hold legislators accountable.

How do voters become aware of excess partisan-ship by their elected representatives? While most rollcalls remain outside of the electoral arena, studies oflegislative behavior suggest that incumbents are risk-averse and worry about their votes because they suspectsome roll calls may become electorally salient (Arnold1990; Bianco, Spence, and Wilkerson 1996; Fiorina 1974;Mayhew 1974), and they can never be absolutely certainwhich roll calls will figure prominently in the next elec-tion. In fact, the more divisive or salient the roll call, themore likely it is to generate attention from outside inter-ests, especially if the overall pattern of legislative voting isperceived as being too far out of touch with district senti-ment (Carson 2005; Jacobson 2009; Mann and Wolfinger1980).

Arnold (1990, 46) argues that legislators must be care-ful when casting roll calls because citizens may use any

4In one of the few studies that have systematically examined theelectoral effects of both ideological and partisan extremity, Erikson(1971) found that the effects of both factors are very similar interms of electoral accountability.

602 JAMIE L. CARSON, GREGORY KOGER, MATTHEW J. LEBO, AND EVERETT YOUNG

number of the incumbents’ actions as cues when engag-ing in retrospective voting. Incumbents must also estab-lish voting records that seek to dissuade challengers fromexposing inconsistencies. “The fear is not simply that cit-izens will notice on their own when a legislator errs, butthat challengers will investigate fully a legislator’s votingrecord and then share their interpretations of how he orshe has gone wrong” (Arnold 1990, 272–73). Indeed, “aprominent position on the wrong side of a major issue[can]. . .galvanize potential opponents” (Jacobson 1987,139). Wright (1978, 446) maintains that it is not necessaryfor issues to account for much variance in election out-comes since electorally insecure legislators only requiretoken incentives to take note of the policy interests oftheir constituents. He also suggests that “since the candi-date’s issue stance is one of the few factors relevant to hisreelection that is also within his control, the representa-tive is well advised to bring his issue positions into linewith those of his constituency. Not to do so could be thedetermining factor in electoral defeat” (459).

Media coverage surrounding salient votes inCongress suggests that legislators have sufficient reasonto worry about their roll calls being politicized. Mediacoverage of showdown votes in Congress can expose “at-tentive publics” (Arnold 1990, 64–65) to an incumbent’sroll-call choices. As a result, those individuals most ca-pable of exploiting legislators’ voting records for politi-cal advantage—prospective candidates, political activists,and social elites—have ample opportunities to becomeaware of roll-call votes and transform them into elec-torally salient political issues. This, in turn, can be suffi-cient for the roll call to have an impact at the polls.5 It iseven more likely that party unity votes will attract consid-erable media attention given that a majority of membersin both parties are taking opposing sides on the issuesunder consideration.

Whenever possible, the party leadership attempts tomake roll-call voting decisions as easy as it can for itsmembers on most legislation that comes before Congress.While the majority party wants to win, it frequently doesnot need its entire membership to fall in line to do so.Moreover, the party leadership is preoccupied with main-taining its majority status, and doing so occasionallyrequires placating representatives who represent cross-pressured districts (Arnold 1990). In order to maximize

5As Fiorina asserts, “an informed, issue conscious citizenry (in thebest traditions of democratic theory) may not be crucially impor-tant for representative government. The entire district need not bewatching, just some part of it—a potential challenger, newspapereditor, interest group, or lone, informed citizen. Nor need they bewatching at the time of the vote; just so they dig up the dirt beforethe election” (1974, 123).

the likelihood the party will win while simultaneouslyminimizing its loss of seats, the party leadership mustbe strategic in choosing when to pressure members oncontroversial legislation. Thus, when the party leadershipfinds it has more votes than necessary to pass a bill (orrealizes it does not have enough votes to win), it will re-lease extraneous, cross-pressured members to vote withtheir constituencies.6 These legislators might otherwisefind it difficult to support the party position on contro-versial legislation (Cox and McCubbins 2007; Desposatoand Petrocik 2003; King and Zeckhauser 2003; Mayhew1974).

At other times, i.e., when an upcoming vote becomescritical to the collective reputation of a party, the politicalstakes may be too high to allow cross-pressured membersto defect on important legislative issues. Party leaders maybe more reluctant to allow a member to defect if it is on aprocedural issue, as this is where we should expect to seethe majority party attempting to structure the legislativeagenda (Cox and McCubbins 2005; Cox and Poole 2002;Jenkins, Crespin, and Carson 2005; Theriault 2008). Partyloyalty in voting can thus become a campaign issue in twodistinct ways: a challenger can criticize an incumbent’soverall pattern of voting (“Smith votes in lockstep with herparty”) or specific votes on which an incumbent joinedwith his or her party (“Smith sided with her party bosseson the minimum wage”).

This article explores the extent to which elected rep-resentatives are held accountable for loyalty to their polit-ical party. When the stakes are high, and policy outcomesare at risk, the majority party leadership has a lot ridingon the final outcome.7 In such cases, party leaders mayplace greater pressure on rank-and-file members to sup-port the party’s position. In other words, there is reasonto suspect theoretically that the degree of party pressureplaced on legislators may be conditional upon a variety ofcircumstances as well as the specific electoral prospects ofthe legislators asked to support the political party. Draw-ing upon legislative voting behavior since the 1950s, wemove beyond prior work that has focused exclusively onideological extremity and examine whether party unity

6Griffin (2006) finds that elected officials who represent competi-tive House districts tend to be more responsive to their constituents’preferences.

7This is not meant to suggest, however, that minority party leadersfail to recognize the importance of these types of votes. Indeed,minority party leaders often try to maintain high levels of partycohesion on the off chance they can pick off a few members ofthe majority and “roll” them in the process (Cox and McCubbins2005). Failing that, they may seek to force as many majority mem-bers as possible to cast votes that can be used against them in thenext election.

ELECTORAL COSTS 603

directly affects incumbent electoral success in the U.S.House given the theoretical linkages outlined above.

One thing that is missing from studies of congres-sional elections is an understanding of how voters rewardor punish legislative behavior. Many studies (e.g., Canes-Wrone et al. 2002) implicitly assume that voters punishlegislators when their records are seen as ideologicallyextreme. But, since voters tend to think more easily inpartisan terms than in ideological terms (Kinder 1998;Sniderman et al. 1991), this is a point where our under-standing of congressional elections is at odds with ex-tant research on voting behavior. Moreover, even if vot-ers identified themselves as strongly within ideologicalgroups as they do within partisan groups, they may bemore forgiving of ideological extremism than they are ofexcessive partisanship. For example, voters may view anideologically extreme legislator as principled, while theymay consider a legislator who consistently votes with herparty as a “party hack.”

In light of this disconnect between studies of legisla-tive and voting behavior, our article offers us an oppor-tunity to test four distinct claims about the relationshipbetween voting in Congress and election outcomes. First,an ideologically extreme voting pattern on roll calls maycause incumbents’ reelection margins to decrease as vot-ers punish these politicians for their divergent behavior.Second, voters punish legislators who are overly parti-san in their voting on divisive issues on the House floor.Third, both partisanship and ideological extremity maybe important factors in predicting how well incumbentsdo at the polls. Finally, we may simply observe no dis-cernible relationship between congressional voting andcongressional elections, which would cast doubt on somerecent scholarly evidence on this issue. In the next section,we offer some experimental evidence to illustrate moreclearly the linkage between party loyalty and individualvoting behavior in congressional races. Following that,we present analyses of 7,939 congressional races over the1956–2004 period.

Experimental Evidence

We designed an experiment8 to determine whether indi-vidual voters punish legislators more severely for locksteppartisanship than for ideological extremism. Each partic-ipant was first provided with bland information about a

8For a more detailed discussion of our sample, experi-mental procedures, and results, see our online appendix athttp://ms.cc.sunysb.edu/"mlebo/details.htm.

fictional member of the U.S. House—name, marital sta-tus, number of children, profession prior to election topublic office, and charitable works. We gave our subjectsjust enough information so that we could measure theirinitial propensity to support the incumbent for reelec-tion, “despite having very little to go on.” This initial levelof support was recorded on a 7-point scale ranging fromstrong support to strong opposition.

Next, participants were exposed to a list of 10 pol-icy positions taken by the incumbents. It was explainedto participants that roll-call votes could be character-ized either in partisan terms where the parties were spliton the issue—voting “with the Republicans and againstthe Democrats”—or in ideological terms—as rated byAmericans for Democratic Action (ADA). Each partici-pant was exposed to one “partisan-characterized” and one“ideology-characterized” incumbent. That is, one incum-bent’s 10 roll-call votes were characterized as votes withone party and against the other. The other incumbent’s 10roll-call votes were characterized by their reported ADArating as being on either the liberal or the conservativeside of the issue. The 10 issues were chosen to balancefiscal, moral, and military and foreign policy issues.

A participant would see an incumbent who had votedwith one particular party either 6, 7, 8, 9, or 10 out of 10times, and an incumbent who had voted with one partic-ular ideology 6, 7, 8, 9, or 10 out of 10 times. This createdvariance in how partisan or how ideologically extremethe incumbent might be. After viewing each incumbent’s10 positions, the participant answered between three andfive distracter questions (for the purposes of allowing theparticipant to forget exactly how many times the politi-cian had voted with one party or ideology and to retainonly a vague impression of the politician’s partisanship orideological extremity). Finally, the participant indicatedhis post-exposure propensity to support the incumbentand answered a number of questions about his impres-sions of the candidate.

The main dependent variable for this analysis, then,is the difference between the pre- and the post-exposurelevel of electoral support. The main explanatory variableis the amount of lockstep partisanship or ideological ex-tremity of the legislator. This variable was dichotomizedso that a participant’s exposure to a legislator voting withone side 6 or 7 out of 10 times was coded 0 and ex-posure to a legislator voting with one side 8, 9, or 10out of 10 times was coded 1. We believe this dichoto-mous coding reflects the likelihood that the differencesbetween, say, a 9-out-of-10 and a 10-out-of-10 politicianmay not seem so large to a voter. Instead, voters more gen-erally may see politicians as being predictably one-sided orless so.

604 JAMIE L. CARSON, GREGORY KOGER, MATTHEW J. LEBO, AND EVERETT YOUNG

We include in our analysis a single control variable:the extent to which the voter’s party identification or ide-ological self-identification (measured on a 7-point scale)opposes the incumbent’s party. In other words, if the in-cumbent votes more often with the Republicans and theparticipant is a strong Republican (7 on the 7-point scale),the participant is scored 1 on this variable, indicating thatthe voter and the legislator match. If the incumbent votesmore often with the Democrats, this same voter would bescored a 7, indicating a strong mismatch. This variable wasrecoded to range from 0 to 1, with 0 a voter-incumbentmatch and 1 a voter-incumbent mismatch. This covariatecontrols for the tendency of voters to evaluate incum-bents’ partisanship or ideology from their own perspec-tive. We also want to minimize the possibility that ourfindings are the result of the characteristics of our partic-ipants rather than the fictional incumbents we presentedto them.

The primary means, however, of minimizing thispossibility was thorough randomization. Randomizedaspects of the experimental design were (1) whetherthe participant saw the partisanship-characterized or theideology-characterized legislator first; (2) the order inwhich the 10 issues were presented; (3) whether the legis-lator leaned Republican or Democratic, liberal or conser-vative; (4) how many times out of 10 the incumbent votedwith his preferred party or ideology; and (5) on which is-sues the incumbent voted with his preferred as opposedto his less-often-preferred party or ideology. Because ofthe randomization, it is difficult to argue that our resultsare biased by certain issues appearing systematically ear-lier in the experiment, or participants making the votedecisions based on specific “hot-button” issues.

We first administered the experiment to a sample ofstudents (N = 143) at a northeastern university. The re-sults are shown in the first two columns of Table 1, wherewe report results from seemingly unrelated regressions.The SUR model allows the creation of a single covari-ance matrix of coefficients for the purposes of comparingcoefficient sizes between the partisan-incumbent and theideological-incumbent equations. When the dependentvariable is the change in support after seeing the incum-bent’s positions characterized as partisan, partisan lock-step voting has a significant negative effect on supportfor the legislator. Extreme legislators suffer a decline of0.64 scale-points (on the 7-point scale) for being lock-step or near-lockstep roll-call voters over and above theexpected penalty (2.67 points) exacted when the voter’sparty identification changes from maximally for to max-imally against the legislator’s preferred party.

On the other hand, while there is still a penalty (1.20points) for a legislator’s preferred ideological position

being opposite the participant’s ideological self-ID, theeffect of ideological extremity is nonsignificant and pos-itive. A post-SUR Wald test shows that the coefficientsfor partisan and ideological extremity are significantlydifferent from each other (p = .003): partisan lockstepvoting yields electoral sanctions (and is, indeed, the onlykind of extremity which brings punishment at all) whileideological extremity does not.

One potential concern with this finding is that it couldbe unique to the northeast. In other regions of the country,such as the more conservative south, the word liberaland, perhaps even the word conservative, might be moretoxic than in the northeast. To answer this challenge, wegathered an additional nonstudent “adult” sample nearlybalanced between 50 adults from a southern capital cityand 54 subjects from an area within approximately 30miles of the northeastern university for our experimentalanalyses. The results are shown in the next two columns ofTable 1. The inclusion of southerners as half of the adultsample leaves the results unchanged. As before, there isa significant negative effect of partisan lockstep votingon electoral support, there is a nonsignificant positiveeffect of ideological extremity, and, again, the differencebetween the two coefficients is significant.9

That an incumbent’s partisan lockstep voting dam-ages electoral support at the individual level, while ide-ological extremity does not, appears to be a robust find-ing regardless of how the sample is divided. Combiningthe student and adult samples and then partitioning byvarious demographic and other variables, the coefficientfor partisan lockstep voting was between #0.5 and #1.0and significant at p < .05 (one-tailed test) whether welook only at liberals or conservatives, only Republicans orDemocrats, only higher- or lower-educated, only male oronly female, only strongly religious or only less so, onlyin the upper or only in the lower half of incomes,10 oronly the more or only the less politically knowledgeable.Also, for every one of these subsamples, the coefficientfor ideological extremity was positive.

9The southern-only results are not shown, but they mainly fol-low the results of Table 1 while offering no support whatsoever tothe counterhypothesis that southerners might punish ideologicalextremity more than they punish partisan lockstepping; the coeffi-cient for partisan lockstep voting is #0.52, negative as in the caseof the full adult sample and the student sample, but with a p-valueof only 0.12 in a one-tailed test. The coefficient for ideological ex-tremity is, as in the full adult and the student samples, positive butcuriously larger in contradiction to the southern counterhypothe-sis: ! = 0.87, p = .057 two-tailed. Again, the Wald test reveals thetwo coefficients are significantly different from each other.

10One exception here: the coefficient for partisan lockstep voting isabsolutely larger at#1.22 for the higher incomes in the adult sample(p = 0.04), the only sample for which income was asked.

ELECTORAL COSTS 605

TABLE 1 Seemingly Unrelated Regressions (SUR) of Change in Support for Legislator$

Northern Student Sample Adult Sample

Dependent variable: change in support before/ Coefficient p-value Coefficient p-valueafter seeing legislator’s positions in partisan terms (std. err.) (2-tailed) (std. err.) (2-tailed)

Independent variablesLegislator was very partisan #0.64$ 0.007 #0.83$ 0.016

(0.24) (0.34)Voter’s party ID opposite of legislator #2.67$ 0.000 #1.94$ 0.000

(0.33) (0.44)Constant 1.18$ 0.000 0.65 0.051

(0.26) (0.33)Dep. var.: change in support before/after seeinglegislator’s positions in ideological terms

Independent variablesLegislator was ideologically extreme 0.39 0.123 0.22 0.516

(0.25) (0.34)Voter’s ideological ID opposite of legislator #1.20$ 0.002 #2.95$ 0.000

(0.42) (0.46)Constant 0.05 0.862 0.96 0.009

(0.29) (0.37)

Wald test of H0: effect of hyper-partisanship = " 2 = 8.8, p = .003 " 2 = 4.61, p = 0.032effect of ideological extremityN 137 104

$After seeing 10 policy positions characterized in either partisan or ideological terms.

In sum, the results are unambiguous: when an incum-bent is known to have voted with one particular party agreat majority of the time, voters of various demographicstripes withdraw electoral support. When an incumbent’sroll calls are known to have supported one particular ide-ology all or almost all of the time, voters do not appearto exact electoral punishment for this fact alone, and caneven reward such behavior, although they do not do soreliably. In all cases, members of Congress (MCs) with arecord of lockstep partisanship incur a significant loss ofelectoral support.

Party Unity in House Elections,1956–2004

Next, we seek to explain electoral success for membersof the House running for reelection from 1956 to 2004.Consistent with Jacobson (1996), the dependent variableis the percent of incumbent i’s two-party vote share; fol-lowing Canes-Wrone et al. (2002) this varies by year t .Since there are many races in which incumbents ran un-opposed or faced only token opposition, we examine onlythose races in which a major party challenger received atleast 1,000 votes. The mean vote share for incumbents is64.8% with a standard deviation of 9.8.

The independent variable we are most interested infor our analysis is the level of party unity for the in-cumbent legislator. Party Unity scores for individual rep-resentatives were collected from Congressional QuarterlyAlmanac from 1955 to 2004. CQ first identifies every“party” vote that pits a majority of one party’s membersagainst a majority of the opposing party, then calculatesfor each legislator the percentage of these party votes onwhich the member voted with his or her party. For someyears, CQ counted absences against one’s unity score,while for other years it did not. We recode the data sothat absences do not count for or against one’s score. Asthe proportion of times a member votes with her party,unity is scored from 0 to 1, with an actual minimum of0.039 for Larry McDonald’s (D-GA) voting in the 95thCongress.11

11What kind of representative votes with his party less than 4% ofthe time? Serving as a Democrat from Georgia’s 7th district from1975 to 1983, McDonald once said, “We have four boxes with whichto defend our freedom: the soap box, the ballot box, the jury box,and the cartridge box.” He also became the president of the JohnBirch Society while serving in the House. McDonald was aboardKorean Air Flight KAL-007, shot down by the Soviets in 1983, thusbecoming the only U.S. member of Congress killed by the SovietUnion during the Cold War. Or, “knowingly” killed by the SovietUnion, as the John Birch Society might report.

606 JAMIE L. CARSON, GREGORY KOGER, MATTHEW J. LEBO, AND EVERETT YOUNG

FIGURE 1 Average Range in House Members’ Party Unity Score byTerms of Service

We note that this variable provides three forms ofvariation: individual legislators change their unity overtime, the unity of legislators varies within a Congress,and there is variation as a district changes its representa-tive over time. Even if every member of Congress had aconstant rate of party unity throughout his or her career,this would still give us a great deal of leverage to makeinferences about different levels of party unity. But, ofcourse, there is much more variation than this over thecourse of a legislator’s career. We summarize the differ-ence between a legislator’s highest and lowest party unityscore across his or her career and present these ranges,aggregated by the number of terms served, in Figure 1.The longer a member’s career, the greater the average dif-ference between high and low party unity scores, withlegislators serving 18+ years varying (on average) 20% ormore. Even at the low end, though, legislators who servetwo terms change (on average) 5%. So even if we focus onwithin-legislator variation, there is a great deal to explain.

Next, we view ideological extremism as a determinantof party unity and measure it using first-dimension DW-NOMINATE scores (Poole and Rosenthal 2007), whichassess legislators’ voting tendencies on the traditionalliberal-conservative economic dimension. These scoresreflect measures of revealed ideology relative to other leg-islators in Congress. More specifically, they are derivedusing an item-response model to estimate the number ofsubstantive dimensions in legislative voting and each leg-

islator’s ideal point on these dimensions. A single dimen-sion typically explains most of the variance in congres-sional voting, so we use first-dimension DW-NOMINATEscores. We calculated the absolute value of these scores sothat high values of the variable Ideological Extremism in-dicate that a member has an “extreme” voting record,while low values suggest that a member’s preferences areat odds with the mainstream of her party.

To isolate the effects of representatives’ roll-call be-havior on their electoral fortunes, we control for a va-riety of factors that have previously been shown toaffect incumbents’ electoral performance, including priorelectoral success, the partisanship of the district, chal-lenger quality, incumbent and challenger spending (inyears where available), freshman status, presidential ap-proval, the change in real disposable personal income,and in-party versus out-party status. We explain ouroperationalization of each below.

Incumbent Vote, our measure of electoral security, isthe incumbent’s percentage share of the vote in the previ-ous election. Since we are only interested in incumbentsrunning for reelection, there is no loss of cases by includ-ing a lagged variable.12

District partisanship is measured as the share of thetwo-party vote that the presidential candidate of the

12Given concerns about the consequences to inference of includinga lagged endogenous variable (Achen 2000), this variable does notappear in all of our models.

ELECTORAL COSTS 607

incumbent’s party received in the congressional districtin the most recent presidential election. Presidential voteat the district level is typically employed as a measure ofdistrict preferences and can be a useful proxy for evaluat-ing the degree to which legislators who support the partyon controversial votes are cross-pressured by constituents(Jacobson 2009).

Challenger quality is coded 1 if the candidate previ-ously held elected office, 0 otherwise. This coding alsofollows Jacobson’s (1980) classic study that views havingrun a successful elective campaign as a proxy for candi-date quality.13 Spending Gap is included to control for theeffects of incumbent and challenger spending, followingJacobson’s work on money in elections. It is calculated asthe natural logarithm of challenger expenditures minusthe natural logarithm of incumbent expenditures.14 Sincespending data are not available prior to 1978, using themrequires leaving out elections from 1956 to 1976. Thus, wetest our hypotheses using both our full historical periodof data and, separately, our full complement of variablesfor 1978 to 2004.

Freshman is a dummy variable controlling for fresh-man status among incumbent legislators. In Party is adummy variable that measures whether legislators aremembers of the president’s party. Midterm accounts forthe effects of midterm elections and is scored “1” formidterm elections with a president of the legislator’sparty, “#1” with a president of the opposite party, and“0” in presidential election years. Presidential Approvaland Change in Personal Income are also coded for in-partystatus so that a popular president and growth in incomelevels are expected to help members of the president’sparty and hurt those in the out-party.15

To examine our expectations systematically, we usea generalized two-stage least-squares panel data es-timator with fixed effects and instrumental variables

13See Gary C. Jacobson, Money in Congressional Elections (1980),106–7. For an alternative view on how to measure challenger quality,see Green and Krasno (1988).

14As Jacobson (1980) argues, the advantage of using the naturallogarithm of campaign expenditures for analyses of elections isthat doing so avoids the assumption of a linear relationship betweenmoney and votes, thus accounting for diminishing marginal returnsfrom campaign spending. Data on presidential vote share, qualitychallengers, and campaign spending were generously provided byGary Jacobson.

15Presidential approval is scored as the raw approval number minus50% and multiplied by #1 for members not of the president’s party.Change in real disposable income is taken from the third quarterof the election year and is multiplied by #1 for members of theout-party.

(Baltagi 2005).16 The use of single-stage pooled regres-sion models is common in the congressional elections lit-erature (see, e.g., Canes-Wrone, Brady, and Cogan 2002;Jacobson 1993), yet it would be preferable to incorporatethe efforts of legislators to anticipate the effects of theirlegislative behavior on electoral outcomes. Instead, a two-stage modeling strategy has two distinct advantages oversingle-stage estimation techniques. First, party unity is it-self an endogenous variable that is affected by, and affects,a legislator’s electoral vote margin. Party leaders use theirexpectations about upcoming elections to decide whichlegislators’ arms will be twisted in terms of higher levelsof unity. Since we also expect constituents to react to lev-els of party unity, this results in reciprocal causality andmakes a two-stage approach appropriate for estimationpurposes.17

Endogeneity is also a problem due to the effects otherindependent variables in the model have on Unity. Thus,a second advantage of our two-stage approach is that itaccounts for the plausible effects on Unity that Freshmanstatus, District Partisanship, and Approval likely have.

A third advantage of our estimation strategy is thatit gives us the opportunity to evaluate whether revealedlegislator preferences are an antecedent variable in the re-lationships between members’ actions and their electoralfortunes. That is, we expect that revealed preferences area reasonably good predictor of unity but are not by them-selves a predictor of vote share—voters are more likelyto punish legislators for being too partisan rather thansimply being too ideological. When we use party unityas an endogenous variable and revealed preferences asan instrument in a first-stage equation, our instrumenttests can tell us whether preferences predict unity, butnot incumbent vote share. If preferences are not a goodinstrument, they belong in the second-stage equation asa direct predictor of incumbent vote share. This is one ofthe major differences between our analyses and those ofCanes-Wrone et al. (2002), who omit what we view as theproximate predictor from their equations—party unity—and use revealed preferences instead. Just to be thorough,

16We also estimated the models of Tables 2–6 using random effectsmodels and random coefficient models. In every case, our findingswere nearly identical. We present the fixed effects models due totheir ease of interpretation—the coefficients are the within-yeareffects on the dependent variable due to a one-unit increase in theindependent variable of interest. Note that low values of # in thetables indicate that the vast majority of unexplained variance isgoing on within years rather than between years, thus making theresults of the various approaches quite similar.

17Our study is the first to apply a two-stage approach to electoraloutcomes, but Canes-Wrone, Rabinovich, and Volden (2007) usea two-stage model to explain ideological extremity in legislativevoting.

608 JAMIE L. CARSON, GREGORY KOGER, MATTHEW J. LEBO, AND EVERETT YOUNG

however, we rerun our models and present abbreviatedresults with voting extremity moved to the second stage,thus allowing us to evaluate its impact relative to unity.

In sum, then, we have the following basic model ofrepresentatives’ electoral vote margins estimated as thesecond stage of our simultaneous equations model:

Incumbent Voteit = $2t + %1 !Unityit + !0 + !1District i t

+ !2Challengeri t + !3SpendGapit + !4Freshmanit

+ !5InPartyit + !6Midtermit + !7Approvali t+ !8!Incomeit + !9Incumbent Voteit#1 + &it

where $2t is the year-specific effect, &i t is the overall er-ror component, !0 is the estimated constant, !1 # !9 areregression coefficients, and %1 estimates the effect of pre-dicted values of the endogenous variable "Unityit , whichis estimated in a first-stage equation:

"Unityit = $1t + '1District i t

+ '2Challengerit + '3SpendGapit + '4Freshmanit

+ '5InPartyit + '6Midtermit + '7Approvalit+ '8!Incomeit + '9IncumbentVoteit#1

+ (1Extremismit + (2Unityit#1

In this equation, (1 and (2 estimate the effects of instru-mental variables excluded from the second stage aboveand U nityit#1 is the lagged value of party unity for leg-islator i, which serves as a useful instrument.18 Also, $1t

is the year-specific effect for the first-stage equation. Toinvestigate our questions fully, we estimate several mod-els that vary slightly from the above equations, but allmaintain the same basic structure.19

Results

We begin by testing our central claims that incumbentsexperience a decrease in their share of the two-party voteas their party unity in Congress increases, and that leg-islators’ party loyalty decreases when they are electorallyvulnerable. We then test variations on our central argu-ment by determining whether the relationship is strongerfor vulnerable or secure legislators, we control for cam-paign finance spending, we explore interactive effects, and

18However, this costs us cases, including all of 1956 when lookingat the complete time period.

19To use data beginning in 1956, SpendGap is left out of Models 1and 2 of Table 2 and Model 1 of Table 5. To compare our modelswhen not accounting for electoral security, Incumbent Voteit#1 is leftout of both equations in Models 1 and 3 of Table 2. Throughout,we aid identification problems by having at least two instrumentalvariables that are not correlated with the errors of the second stage(see Wooldridge 2002, 90).

we look at a model that excludes southern representatives.Finally, we calculate our model using change in vote shareas the dependent variable, thereby controlling for vari-ables that are constant for each legislator (e.g., ideology).

Vote Share Declines with Party Unity

Table 2 estimates four slightly different versions of ourbasic model. Models 1 and 2 use our entire sample periodbut exclude the spending gap variable, and Models 1 and3 exclude our control variable tapping electoral safety. Allour independent variables perform as expected.

To begin, the character of the district is very im-portant in predicting incumbent vote share. As a districtbecomes more supportive of presidential candidates fromthe legislator’s party, the incumbent does better. Aspectsof the individual race matter as well. Facing a qualitychallenger makes reelection more difficult, costing a leg-islator just over 5% of the vote share in Model 1 but less inthe shorter period of data. This difference could well bedue to the introduction of a spending gap variable in thelater period (see Jacobson 2009). Indeed, the differencein the natural logs of incumbent and challenger spendingis a powerful predictor of incumbent vote. Also, gettingreelected as a freshman appears to have a significant im-pact on vote share, but the sign depends upon whetherelectoral security is included in the model.

National-level conditions are also important deter-minants of congressional races. As others have found,higher levels of presidential approval help legislators ofthe president’s party and hurt members of the oppositeparty. Incumbents also benefit from sharing the presi-dent’s party affiliation as personal income is growing.Consistent with regularized patterns of surge-and-declinein House races, midterm elections reduce vote share formembers of the president’s party, all else equal (Jacobson2009).

The key result from the four models is that party unityhas a significant negative impact on an incumbent’s voteshare. Over the 1956–2004 period, and holding all othervariables constant, voters consistently punish legislatorsfor voting too often with their party. In Model 1, a 50-point increase in a member’s party unity score will cost anincumbent nearly 5% of the vote share in the subsequentelection—a loss equal to that of a quality challenger enteringthe race.20 When we also factor in the effect of electoralsecurity in Model 2, the size of the party unity effect isdiminished but it is still statistically significant. Over the

20Also note that this is a “typical” effect, so actual effects may varywith factors like seat marginality, campaign intensity, and the depthof party loyalty in a district’s votes.

ELECTORAL COSTS 609

TAB

LE2

Expl

aini

ngH

ouse

Incu

mbe

nts’

Shar

eof

the

Two-

Part

yVo

te,1

956–

2004

†

1956

–200

419

78–2

004

Mod

el1

Mod

el2

Mod

el3

Mod

el4

Coe

f.(s

.e.)

zC

oef.

(s.e

.)z

Coe

f.(s

.e.)

zC

oef.

(s.e

.)z

Con

stan

t50

.23

(0.6

8)74

.36$$

$24

.90

(0.8

3)30

.12$$

$47

.87

(0.8

2)58

.18$$

$30

.01

(0.9

9)30

.43$$

$

Vot

esha

ret#

10.

48(0

.01)

44.8

4$$$

0.36

(0.0

1)29

.39$$

$

Dis

tric

tPar

tisa

nshi

p43

.90

(0.9

8)44

.79$$

$23

.17

(0.9

9)23

.33$$

$37

.60

(1.0

9)34

.42$$

$23

.21

(1.1

3)20

.59$$

$

Qua

lity

Cha

lleng

er#

5.20

(0.2

3)#

22.2

0$$$

#3.

25(0

.21)

#15

.54$$

$#

3.07

(0.2

8)#

11.0

8$$$

#2.

07(0

.26)

#8.

05$$

$

Spen

dgap

#2.

83(0

.08)

#37

.02$$

$#

2.25

(0.0

7)#

30.0

4$$$

Fres

hman

#2.

74(0

.33)

#8.

29$$

$1.

72(0

.31)

5.59

$$$

#0.

94(0

.35)

#2.

67$$

0.93

(0.3

3)2.

78$$

Pres

iden

tial

App

rova

l(co

ded

byin

-par

ty)

0.07

(0.0

1)6.

29$$

$0.

15(0

.01)

14.5

9$$$

0.06

(0.0

1)4.

83$$

$0.

12(0

.01)

9.44

$$$

Mid

term

Elec

tion

(cod

edby

in-p

arty

)#

4.05

(0.2

0)#

20.2

0$$$

#5.

00(0

.18)

#28

.02$$

$#

2.54

(0.2

8)#

9.21

$$$

#3.

34(0

.26)

#12

.88$$

$

!Pe

rson

alIn

com

e(c

oded

byin

-par

ty)

#0.

43(0

.04)

#10

.53$$

$#

0.41

(0.0

4)#

11.4

8$$$

0.03

(0.0

6)0.

44#

0.10

(0.0

6)#

1.71

$

In-p

arty

0.61

(0.2

1)2.

95$$

2.00

(0.1

9)10

.78$$

$#

0.26

(0.3

1)#

0.86

1.03

(0.2

9)3.

50$$

$

Part

yU

nity

††#

9.55

(0.8

2)#

11.6

6$$$

#4.

15(0

.75)

#5.

50$$

$#

9.19

(0.9

8)#

9.33

$$$

#5.

85(0

.92)

#6.

34$$

$

R2

wit

hin/

betw

een/

all

0.36

/0.1

0/0

.35

0.52

/0.6

0/0

.53

0.55

/0.1

4/0

.52

0.64

/0.2

4/0

.62

#.0

60.0

37.0

94.0

86O

bser

vati

ons

6669

6031

3967

3519

Gro

ups

2424

1414

Obs

/Gro

upM

in/A

vg/M

ax12

0/2

77.9

/316

104

/251

.3/2

9324

9/2

83.4

/310

228

/251

.4/2

72In

stru

men

tsex

clud

edfr

omse

cond

stag

eEx

trem

ity,

Part

yU

nity

t#1

Extr

emit

y,Pa

rty

Uni

tyt#

1Ex

trem

ity,

Part

yU

nity

t#1

Extr

emit

y,Pa

rty

Uni

tyt#

1

F-te

stof

Excl

uded

IVs(

p)77

68(0

.00)

6800

(0.0

0)55

96(0

.00)

4849

(0.0

0)Sa

rgan

"2

stat

isti

c(p

)0.

19(0

.66)

0.18

(0.6

7)0.

93(0

.36)

0.28

(0.6

0)

Effe

ctof

Extr

emit

yon

z=

41.0

2$$

$z

=41

.22$$

$z

=34

.61$$

$z

=33

.17$$

$

Uni

ty–

1stSt

age

Rer

unni

ngA

bove

butw

ith

Mod

el1b

†††

Mod

el2b

†††

Mod

el3b

†††

Mod

el4b

†††

Extr

emit

yin

Seco

ndSt

age

Coe

f.(s

.e.)

zC

oef.

(s.e

.)z

Coe

f.(s

.e.)

zC

oef.

(s.e

.)z

Part

yU

nity

††#

9.76

(1.3

7)#

7.13

$$$

#3.

97(1

.30)

#3.

06$$

$#

8.73

(1.7

6)#

4.97

$$$

#8.

27(1

.68)

#4.

92$$

$

Extr

emit

y0.

21(1

.10)

0.19

0.17

(1.0

1)0.

17#

0.42

(1.3

2)#

0.32

#0.

68(1

.28)

#0.

53

†Fi

xed

effe

cts

pane

l-da

tam

odel

sw

ith

inst

rum

enta

lva

riab

les

and

2-SL

S.T

here

sult

sof

both

rand

omef

fect

sm

odel

san

dm

ulti

leve

lra

ndom

coef

fici

ent

mod

els

are

near

lyid

enti

cal.

††In

stru

men

ted

vari

able

.$ p<

.05,

$$p

<.0

1,$$

$ p<

.001

,one

-tai

led

test

s.†††

Mod

els

are

othe

rwis

eth

esa

me

asab

ove,

butw

esh

owon

lyth

eke

yre

sults

.

610 JAMIE L. CARSON, GREGORY KOGER, MATTHEW J. LEBO, AND EVERETT YOUNG

shorter period from 1978 to 2004, the variable is againsignificant and negative with or without the control forelectoral security. On the whole, the models do quite well,with Model 2 (1956–2004) explaining 53% of the varianceand Model 4 (1978–2004) explaining 62% of the variancein incumbent vote share.21

The finer points of our modeling strategy provide uswith some additional insights into incumbent electoralperformance. First, ideological extremity does not be-long in a model predicting incumbent vote share. It does,however, prove to be a useful instrumental variable as apredictor of party unity, with z statistics of 41.02, 41.22,34.61, and 33.17 in the first stage of the four respectivemodels. Hansen-Sargan " 2 tests indicate that extremityis properly excluded from the second-stage equation withp-values not dropping below .36. Additionally, the lasttwo rows of Table 2 present abbreviated results when wemove extremity from the first stage to the second stage.22

Our instrument tests suggest that Models 1b-4b are mis-specified, but they do allow extremity to predict vote shareside-by-side with predicted values of party unity. Thusour empirical conclusions need not rest on one particularmodeling strategy. Since extremity fails to approach sta-tistical significance in any of these revised models, we gainconfidence that it is an indirect predictor of incumbentvote share—voters appear to penalize a voting patternof partisan loyalty by representatives, not the underlyingideology that may predict their voting record.23

Electoral Risk Predicts Party Unity

Additionally, our analysis finds reciprocal causality be-tween incumbent vote and party unity. Like Lebo,McGlynn, and Koger (2007), we expect that parties at-tempt to win legislative contests while minimizing elec-toral costs. To do this, party leaders will be careful aboutwhom they pressure to vote with the party, choosing legis-lators who are expected to do well in upcoming elections.In this way, predictions about future election outcomescondition levels of party unity. In Table 3 we show thatthis is indeed the case by switching the causal ordering of

21Again, random effects models give very similar results to thesefixed effects models and Hausman tests establish that there are nodistinguishable differences in our coefficient estimates.

22For the complete results of these models, see our online appendix.

23This is also evident when we forgo our two-stage models and sim-ply estimate the effects of ideological extremity and party unity to-gether in a one-stage pooled regression model. In such cases, partyunity is the superior predictor, with extremity failing to achievestatistical significance on its own. Note, however, that these modelsmay suffer from the endogeneity bias that partially motivates ourtwo-stage approach.

TABLE 3 Explaining Change in Party UnityUsing Future Incumbent Share,1978–2004†

Coef. (s.e.) Z

Extremity 9.34 (0.89) 10.55$$$

Votesharet†† 0.09 (0.02) 3.99$$$

District Partisanship #15.71 (1.61) #9.73$$$

Midterm (by in-party) #1.41 (0.27) #5.23$$$

! In Party #0.26 (0.12) #2.12$

Constant 0.55 (1.23) 0.45R2 0.05Observations 3519Instruments excluded Quality Challenger,

from second stage Spendgap, PresidentialApproval, ! PersonalIncome, Votesharet #1,

FreshmanF-test of Excluded IVs (p) 475 (0.00)Sargan " 2 statistic (p) 4.01 (0.55)

† This is a pooled cross-sectional time-series model with instru-mental variables and two-stage least squares. †† Instrumented vari-able.$p < .05, $$p < .01, $$$p < .001, one-tailed tests.

our earlier models. Here, we predict changes in the levelof party unity from one Congress to the next. Predictedvalues of incumbent vote share in the election to followare generated from a first-stage regression and are a strongpositive predictor (z = 3.99) of levels of party unity in thesecond stage.24 In addition, we can see some of the causalfactors predicting changes in party unity. One interestingpoint is that legislators who are more extreme have a ten-dency to increase their level of unity from one Congressto the next once electoral safety is accounted for. That is,not only does extremity predict levels of party unity, but italso forecasts the inclination to increase the level of partyunity controlling for an MC’s chances for reelection.

Close vs. Safe Districts

Returning to our predictions regarding vote share for in-cumbent legislators, we do not expect these relationshipsto stay the same regardless of the partisan landscape in-cumbents represent. The punishment constituents doleout for partisan behavior should be greatly reduced tothe extent that the district leans toward the same party.

24Challenger quality, spending gap, lagged values of incumbent voteshare, freshman status, change in income growth, and presidentialapproval are used as instruments excluded from the second stageof the estimation.

ELECTORAL COSTS 611

TAB

LE4

Incu

mbe

ntSh

are

byC

lose

ness

ofD

istr

ict†

Pre

s.Vo

te<

40%

1978

–200

4P

res.

Vote

<50

%19

78–2

004

Pre

s.Vo

te<

65%

1978

–200

4P

res.

Vote

>65

%19

78–2

004

Coe

f.(s

.e.)

zC

oef.

(s.e

.)z

Coe

f.(s

.e.)

zC

oef.

(s.e

.)z

Con

stan

t38

.34

(4.6

2)8.

30$$

$39

.36

(2.1

4)18

.37$$

$37

.68

(1.2

4)30

.31$$

$9.

39(3

.19)

2.94

$$

Vot

esha

ret#

10.

29(0

.05)

5.92

$$$

0.28

(0.0

2)11

.77$$

$0.

32(0

.02)

22.8

2$$$

0.35

(0.0

3)13

.11$$

$

Dis

tric

tPar

tisa

nshi

p23

.00

(9.7

4)2.

36$

23.1

2(3

.78)

6.12

$$$

13.4

8(1

.52)

8.89

$$$

57.2

1(3

.98)

14.3

7$$$

Qua

lity

Cha

lleng

er#

1.36

(0.7

6)#

1.76

$#

1.53

(0.4

0)#

3.82

$$$

#1.

99(0

.27)

#7.

50$$

$#

3.04

(0.8

0)#

3.82

$$$

Spen

dgap

#3.

00(.

300)

#9.

99$$

$#

2.59

(0.1

3)#

19.2

2$$$

#2.

40(0

.08)

#29

.60$$

$#

1.86

(0.1

6)#

11.2

9$$$

Fres

hman

2.13

(1.1

9)1.

78$

0.82

(0.5

6)1.

470.

99(0

.36)

2.75

$$0.

05(0

.74)

0.07

Pres

iden

tial

App

rova

l(co

ded

byin

-par

ty)

#0.

21(0

.10)

#2.

150.

13(0

.03)

4.40

$$$

0.13

(0.0

1)9.

47$$

$0.

14(0

.03)

4.57

$$$

Mid

term

Elec

tion

(cod

edby

in-p

arty

)2.

84(1

.98)

1.43

#2.

26(0

.57)

#3.

92$$

$#

3.48

(0.2

8)#

12.5

2$$$

#3.

83(0

.73)

#5.

23$$

$

!Pe

rson

alIn

com

e(c

oded

byin

-par

ty)

0.87

(0.4

2)2.

05#

0.19

(0.1

4)#

1.30

#0.

15(0

.07)

#2.

27$

#0.

51(0

.16)

#3.

26$

In-p

arty

#3.

54(2

.16)

#1.

641.

51(0

.67)

2.26

$1.

51(0

.32)

4.69

$$$

2.41

(0.7

6)3.

19$$

$

Part

yU

nity

††#

10.6

7(2

.99)

#3.

57$$

$#

10.0

6(1

.57)

#6.

40$$

$#

6.25

(0.9

4)#

6.67

$$$

#7.

27(3

.42)

#2.

12$$

$

R2

wit

hin/

betw

een/

all

0.55

/0.1

7/0

.42

0.52

/0.2

0/0

.49

0.53

/0.1

4/0

.51

0.70

/0.3

9/0

.68

#.3

67.1

62.0

96.1

86O

bser

vati

ons

268

1074

2857

662

Gro

ups

1414

1414

Obs

/Gro

upM

in/A

vg/M

ax5

/19.

1/5

537

/76.

7/1

1518

7/2

04.1

/230

23/4

7.3

/77

Inst

rum

ents

excl

uded

from

seco

ndst

age

Extr

emit

y,Pa

rty

Uni

tyt#

1Ex

trem

ity,

Part

yU

nity

t#1

Extr

emit

y,Pa

rty

Uni

tyt#

1Ex

trem

ity,

Part

yU

nity

t#1

F-te

stof

Excl

uded

IVs

(p)

305

(0.0

0)11

85(0

.00)

2845

(0.0

0)57

9(0

.00)

Sarg

an"

2st

atis

tic

(p)

1.72

(0.1

9)0.

03(0

.86)

0.02

(0.9

7)0.

27(0

.60)

Effe

ctof

Extr

emit

yon

z=

9.60

$$$

z=

18.2

9$$$

z=

29.5

5$$$

z=

15.4

4$$$

Uni

ty–

1stSt

age

Rer

unni

ngA

bove

butw

ith

Mod

el1b

†††

Mod

el2b

†††

Mod

el3b

†††

Mod

el4b

†††

Extr

emit

yin

Seco

ndSt

age

Coe

f.(s

.e.)

zC

oef.

(s.e

.)z

Coe

f.(s

.e.)

zC

oef.

(s.e

.)z

Part

yU

nity

††#

21.1

6(8

.80)

#2.

41$$

#10

.19

(3.0

7)#

3.32

$$$

#4.

71(1

.73)

#2.

72$$

#6.

33(6

.17)

#1.

03Ex

trem

ity

10.4

4(8

.21)

1.27

0.12

(2.5

4)0.

05#

1.45

(1.3

8)#

1.05

#0.

51(2

.81)

#0.

18

†Fi

xed

effe

cts

pane

l-da

tam

odel

sw

ith

inst

rum

enta

lva

riab

les

and

2-SL

S.T

here

sult

sof

both

rand

omef

fect

sm

odel

san

dm

ulti

leve

lra

ndom

coef

fici

ent

mod

els

are

near

lyid

enti

cal.

††In

stru

men

ted

vari

able

.$ p<

.05,

$$p

<.0

1,$$

$ p<

.001

,one

-tai

led

test

s.†††

Mod

els

are

othe

rwis

eth

esa

me

asab

ove,

butw

esh

owon

lyth

eke

yre

sults

.

612 JAMIE L. CARSON, GREGORY KOGER, MATTHEW J. LEBO, AND EVERETT YOUNG

TAB

LE5

Incu

mbe

ntSh

are—

Inte

ract

ive

Rel

atio

nshi

psan

dN

on-S

outh

ern

Stat

es†

Mod

el1

Mod

el2

Mod

el3

Mod

el4

Inte

ract

ing

Part

isan

ship

&U

nity

,195

6–20

04In

tera

ctin

gPa

rtis

ansh

ip&

Uni

ty,1

978–

2004

+O

ther

Inte

ract

ions

,19

56–2

004

All

Non

-Sou

th,1

978–

2004

Coe

f.(s

.e.)

zC

oef.

(s.e

.)z

Coe

f.(s

.e.)

zC

oef.

(s.e

.)z

Con

stan

t49

.36

(6.1

0)8.

09$$

$77

.75

(9.2

4)8.

42$$

$79

.64

(9.8

4)8.

10$$

$29

.32

(1.1

2)26

.07$$

$

Vot

esha

ret#

10.

49(0

.01)

37.9

8$$$

0.32

(0.0

1)22

.44$$

$0.

33(0

.01)

22.3

7$$$

0.36

(0.0

1)26

.69$$

$

Dis

tric

tPar

tisa

nshi

p#

22.5

8(1

0.58

)#

2.13

$#

64.2

0(1

6.04

)#

4.00

$$$

#69

.41

(17.

33)

#4.

01$$

$23

.62

(1.2

4)19

.03$$

$

Qua

lity

Cha

lleng

er#

3.19

(0.2

1)#

15.2

8$$$

#2.

09(0

.26)

#8.

13$$

$#

2.07

(0.2

6)#

8.08

$$$

#2.

07(0

.27)

#7.

30$$

$

Spen

dgap

#2.

29(0

.08)

#30

.48$$

$#

2.28

(0.0

8)#

30.3

7$$$

#2.

18(0

.08)

#26

.30$$

$

Fres

hman

1.58

(0.3

1)5.

13$$

$0.

73(0

.34)

2.17

$0.

78(0

.33)

2.34

$$1.

07(0

.38)

2.85

$$

Pres

iden

tial

App

rova

l(co

ded

byin

-par

ty)

0.15

(0.0

1)14

.89$$

$0.

12(0

.01)

9.96

$$$

0.18

(0.0

8)2.

39$$

0.12

(0.0

1)8.

70$$

$

Mid

term

Elec

tion

(cod

edby

in-p

arty

)#

4.84

(0.1

8)#

26.7

7$$$

#3.

34(0

.26)

#12

.93$$

$#

3.46

(0.2

6)#

13.2

9$$$

#3.

12(0

.29)

#10

.87$$

$

!Pe

rson

alIn

com

e(c

oded

byin

-par

ty)

#0.

39(0

.04)

#10

.84$$

$#

0.16

(0.0

6)#

2.53

$$0.

63(0

.21)

3.02

$$#

0.10

(0.0

7)#

1.43

In-p

arty

1.92

(0.1

9)10

.34$$

$1.

26(0

.29)

4.30

$$$

1.40

(0.3

0)4.

70$$

$0.

98(0

.33)

2.98

$$

Part

yU

nity

††#

32.0

2(7

.05)

#4.

54$$

$#

59.4

4(1

0.47

)#

5.68

$$$

#61

.72

(11.

15)

#5.

53$$

$#

5.24

(1.0

4)#

5.02

$$$

Dis

tric

tPar

tisa

nshi

p$

54.6

1(1

2.87

)4.

24$$

$10

2.31

(18.

99)

5.39

$$$

108.

20(2

0.46

)5.

29$$

$

Part

yU

nity

App

rova

l$U

nity

#0.

07(0

.09)

#0.

73!

P.In

com

e$

Uni

ty#

0.95

(0.2

5)#

3.87

$$$

R2

wit

hin/

betw

een/

all

0.53

/0.6

0/0

.53

0.64

/0.2

0/0

.61

0.64

/0.2

2/0

.61

0.64

/0.3

0/0

.62

#.0

38.1

04.0

96.0

87O

bser

vati

ons

/Gro

ups

6031

/24

3519

/14

3519

/14

2911

/14

Obs

/Gro

upM

in/A

vg/M

ax10

4/2

51.3

/293

228

/251

.4/2

7222

8/2

51.4

/272

179

/207

.9/2

25In

stru

men

tsex

clud

edfr

omse

cond

stag

eEx

trem

ity,

Part

yU

nity

t#1

Extr

emit

y,Pa

rty

Uni

tyt#

1Ex

trem

ity,

Part

yU

nity

t#1

Extr

emit

y,Pa

rty

Uni

tyt#

1

F-te

stof

Excl

uded

IVs

(p)

391

(0.0

0)21

2(0

.00)

192

(0.0

0)40

09(0

.00)

Sarg

an"

2st

atis

tic

(p)

1.95

(0.1

6)1.

46(0

.23)

0.02

(0.9

7)0.

10(0

.75)

Effe

ctof

Extr

emit

yon

Uni

ty–

1stSt

age

z=

3.60

$$$

z=

5.33

$$$

z=

4.04

$$$

z=

33.7

2$$$

†Fi

xed

effe

cts

pane

l-da

tam

odel

sw

ith

inst

rum

enta

lva

riab

les

and

2-SL

S.T

here

sult

sof

both

rand

omef

fect

sm

odel

san

dm

ulti

leve

lra

ndom

coef

fici

ent

mod

els

are

near

lyid

enti

cal.

††In

stru

men

ted

vari

able

.$ p<

.05,

$$p

<.0

1,$$

$ p<

.001

,one

-tai

led

test

s.

ELECTORAL COSTS 613

For example, we do not expect two districts—one heav-ily Democratic and one heavily Republican—to react inthe same way to a member of Congress with a recordof supporting the Democrats on a large percentage ofvotes. Table 4 shows the results when we look at foursubsets of our 1978–2004 sample—cases where the mostrecent presidential candidate of the incumbent’s party re-ceived less than 40% of the vote (n = 268), cases wherehe won less than 50% of the vote (n = 1,074), caseswhere he won less than 65% of the vote (n = 2,857),and cases where he won more than 65% of the vote (n= 662).25 We see in Models 1 and 2 a strong punish-ment for unity when only a minority of a district votedfor the incumbent’s presidential candidate. As expected,the electoral effect of unity does vary significantly acrossdistricts. When cases between 50% and 65% are added tothe sample, the coefficient for Unity drops significantly to#6.25. Thus, in more competitive districts, party unity isof far more importance to voters looking at the recordsof their incumbents.26 To again ensure the soundness ofour conclusions that unity directly affects Voteshare andthat extremity is antecedent, we show abbreviated resultsfor Models 1b–4b that move extremity to the first stage ofour 2-SLS approach. Unity again matters and extremityis not statistically significant.27

District Partisanship and CampaignFinance

Table 5 adds some additional tests. The first model ex-amines the difference between close and safe districtsover our entire data period using an interaction term.This variable, District Partisanship $ Party Unity, is pos-itive and significant, indicating that the negative effectsof unity are attenuated by higher partisan support in alegislator’s district. In districts with extremely high levelsof partisan support, higher levels of unity can actuallyhelp, since constituents expect higher degrees of parti-san loyalty. This finding is consistent with our theoreticalexpectations that a legislator’s support of the party ondivisive votes will be conditional upon the underlyingpartisanship of the district. The second model of Table 5examines the interactive nature of this relationship usingthe spending gap variable and the shortened time period.

25This follows the most liberal definition of safe versus contesteddistricts, but the results hold when we set the dividing line at 55%and 60%.

26When an incumbent represents a safer district, party unity stillmatters, but it matters much less compared to legislators represent-ing moderate and competitive districts.

27Consult our online appendix for complete results.

Here the interactive relationship gets even stronger. Look-ing at the extreme cases, going from the lowest level ofunity to the highest in a district with no partisans wouldcost a legislator 64.20% of the vote; however, doing thesame in a district with 100% partisans would increase alegislator’s vote share by 38.11% (102.31–64.20).

To this conditional effect, we add two more inModel 3 to account for the varying effects of unity accord-ing to national-level conditions. The coefficient on PartyUnity $ Change in Personal Income tests the extent towhich partisanship is harmful contingent upon the stateof the economy. This interaction does not achieve statis-tical significance, but interacting unity with presidentialapproval does yield a significant result. This in turn tells usthat higher levels of unity are more harmful as either thepresident of one’s own party declines in popularity (e.g.,Nancy Johnson, R-CT, in 2006) or as the president ofthe opposition party increases in popularity. While theseeffects are of modest interest, they do not change the es-sential results reported in the earlier models. Model 4of Table 5 checks the possibility that our findings aredriven by noncompetitive southern districts. While all ofour models exclude cases where no challenger receivedmore than 1,000 votes, here we exclude all cases fromthe 11 Confederate states. Once again the punishment forparty voting is evident.

Explaining Change in Vote Share

Lastly, we use a dynamic-panel approach to test our expec-tations and present the results in Table 6. In this model, allvariables are differenced so that we are explaining changesin vote share using changes in the independent variables.That is, we explain how much better/worse an incumbentdid in each election as compared with how she did inthe previous election. Independent variables also measuredifferences between elections. For example, ! Presiden-tial Approval measures how much more (or less) popularthe president is in one election year as compared with theprevious election year.

This strategy allows us to hold a great many thingsconstant—such as immeasurable personality or districtcharacteristics. Most importantly, this method ensures thatideology, while omitted, is perfectly accounted for in themodel. Since each legislator’s ideology is likely unchang-ing between successive Congresses, it should not have aneffect on the change in vote share for successive elec-tions.28 The Party Unity coefficient tells us how much

28The ideal points of legislators can change slightly between Con-gresses as a result of additional votes being added to the roll-callmatrix (Poole and Rosenthal 2007). We thank an anonymous re-viewer for mentioning this.

614 JAMIE L. CARSON, GREGORY KOGER, MATTHEW J. LEBO, AND EVERETT YOUNG

TABLE 6 Dynamic Model of House IncumbentShare, 1980–2004†

Coefficient (s.e.) Z

Constant #2.15 (1.65) #12.98$$$

! District Partisanship 23.10 (2.61) 8.86$$$

! Quality Challenger #1.81 (0.33) #5.46$$$

! Spending Gap #2.86 (0.10) #27.75$$$

! Presidential Approval(coded by in-party)

0.18 (0.01) 12.75$$$

! Midterm Election(coded by in-party)

#3.75 (0.30) #12.38$$$

! Personal Income (codedby in-party)

#0.34 (0.09) #3.76$$$

! In Party 2.19(0.40) 5.43$$$

! Party Unity†† #9.39 (3.68) #2.55$$

R2 within/between/all 0.28 / 0.08 / 0.27P .069Observations 3518Groups 13Obs. per Group

min/avg/max244 / 270.6 / 306

Instruments excludedfrom second stage

! Extremity, ! Freshman

Anderson LR test 1330(p) (0.00)Sargan " 2 statistic 0.86(p) (0.35)

†These are fixed effects panel-data models with instrumental vari-ables and two-stage least squares. The results of both a randomeffects model and a multilevel random coefficient model are nearlyidentical. †† Instrumented variable.$p < .05, $$p < .01, $$$p < .001, one-tailed tests.

better (or worse) a House member will do compared tohis or her last reelection based on how much more (orless) partisan he or she was in the present Congress com-pared to the previous Congress. The coefficient of #9.39indicates a considerable loss when so many other effectsare held constant. Although this model estimates effectson legislators who have been elected at least twice before,voters can still dole out a hefty punishment for increasesin the level of partisan voting regardless of how “safe”legislators perceive themselves to be.

Conclusion

This article began with a simple, yet importantpremise—constituents evaluate the voting records of theirmembers of Congress, including their tendency to vote

with their parties. While past research has focused al-most exclusively on the relationship between legislativeideology and electoral outcomes, we argue that suchan approach misses an important linkage in legislative-constituency dynamics. Voters may prefer an ideologi-cally “extreme” legislator as their elected representative inmany districts to mirror their own preferences. They mayeven accept members who are ideological “mavericks” byvoting their own preferences as a matter of conscience.At the same time, representatives may be more likely tobe punished if they are perceived as being too partisanin their voting behavior in Congress, especially in an erawhen extreme partisanship has a strong negative conno-tation (Jacobson 2000; McCarty, Poole, and Rosenthal2006).

By focusing almost exclusively on the connection be-tween ideological extremity and electoral accountability,we believe that past research has overlooked the role ofparty loyalty in congressional elections. As the literatureon individual voting behavior shows, partisanship is amuch more prevalent force in the minds of voters thanis ideology when choosing between candidates (see, e.g.,Kinder 1998). Ideological considerations may be impor-tant when choosing between different candidates runningfor president. However, this is less often the case with rep-resentatives since far less information is available to eval-uate legislators on this dimension. Indeed, as Mann andWolfinger (1980) have shown, voters are more likely torespond to partisan than ideological considerations whenconsidering who to support for Congress at the polls.

Given the disconnect between studies of legislativeand individual voting, we investigate the relationship be-tween partisanship and electoral outcomes through ex-perimental and empirical evidence. The results of theexperiment are striking: when an incumbent is known tohave voted with one party a great majority of the time,voters are less likely to support him or her. In contrast, pat-terns of extreme ideological voting among legislators donot yield significant negative effects in terms of electoralsupport. Regardless of the sample we surveyed, legisla-tors with a record of lockstep, partisan voting were mostlikely to suffer electoral losses in the subsequent election.These results help explain our main finding: legislatorspay an electoral price for party loyalty in legislative vot-ing, but party unity is partially a function of legislators’strategic calculations. We show this through a dynamicanalysis that allows for reciprocal causation and incorpo-rates the indirect effect of legislators’ policy preferenceson incumbent vote share.

By focusing explicitly on the effects of party unity inCongress, we find that incumbent House members’ voteshare declines the more they vote with their own party on

ELECTORAL COSTS 615

issues that divide the two major parties. While ideologicalextremity is correlated with party unity, we find that ithas little direct effect on vote share. Moreover, party unityis almost twice as costly for members from moderatedistricts compared to lopsided districts, as defined bypresidential vote share. Taken together, our work suggeststhat many legislators face electoral penalties for votingtoo often with their parties on divisive issues that attractthe attention of challengers and attentive constituents.Legislators nonetheless support their party positions (andpay the penalties) out of some combination of personalconviction and support for the collective interests of theirparties.

We hope that our findings stimulate additional re-search on congressional elections and legislative behav-ior. For example, it is unclear whether legislators are morelikely to be pilloried for a single “wrong” vote (on behalfof a party goal) or for a pattern of party unity. In practice,we observe both types of claims in campaign ads. It wouldbe useful to know more about the mechanism by whichlegislative partisanship becomes a campaign issue, or howpartisanship is incorporated into campaign themes (e.g.,tying a local MC to an unpopular party leader, or asproof that the local MC is “out of touch” or has “goneWashington”). Finally, more evidence is needed to eval-uate which voters are likely to be swayed by patterns ofpartisan behavior.

Furthermore, our findings may stimulate additionalresearch on the strategies that parties use to mitigate thecosts of loyalty. While it is well established that congres-sional parties strive to minimize their costs by manipulat-ing the rules and agenda of their chamber, further researchis needed to test whether “procedural partisanship” is lesscostly in electoral terms than patterns of party loyalty onpolicy issues. Even more important, however, is the needfor further refinement that explains why legislators incurthe costs of voting with their parties in the first place.

References

Achen, Christopher H. 2000. “Why Lagged Dependent VariablesCan Suppress the Explanatory Power of Other IndependentVariables.” Princeton University. Typescript.

Ansolabehere, Stephen, James M. Snyder, Jr., and CharlesStewart, III. 2001. “Candidate Positioning in U.S. HouseElections.” American Journal of Political Science 45: 136–59.

Arnold, R. Douglas. 1990. The Logic of Congressional Action.New Haven, CT: Yale University Press.

Baltagi, Badi H. 2005. Econometric Analysis of Panel Data.3rd ed. West Sussex, UK: John Wiley.

Bartels, Larry M. 2000. “Partisanship and Voting Behavior,1952–1996.” American Journal of Political Science 44: 35–50.

Bianco, William T., David B. Spence, and John D. Wilkerson.1996. “The Electoral Connection in the Early Congress: TheCase of the Compensation Act of 1816.” American Journalof Political Science 40(February): 145–71.

Bonney, Jessica, Brandice Canes-Wrone, and William Minozzi.2007. “Issue Accountability and the Mass Public: The Elec-toral Consequences of Legislative Voting on Crime Policy.”Working Paper. Princeton University.

Bovitz, Gregory L., and Jamie L. Carson. 2006. “Position Tak-ing and Electoral Accountability in the U.S. House of Rep-resentatives.” Political Research Quarterly 59(June): 297–312.

Campbell, Angus, Philip Converse, Warren Miller, and DonaldStokes. 1960. The American Voter. Chicago: University ofChicago Press.

Canes-Wrone, Brandice, David W. Brady, and John F. Cogan.2002. “Out of Step, Out of Office: Electoral Accountabilityand House Members’ Voting.” American Political ScienceReview 96(March): 127–40.

Canes-Wrone, Brandice, Julia Rabinovich, and Craig Volden.2007. “Who Parties? Floor Voting, District Ideology, andElectoral Margins.” In Party, Process, and Political Changein Congress, Volume 2: Further New Perspectives on the His-tory of Congress, ed. David Brady and Mathew McCubbins.Stanford, CA: Stanford University Press, 113–25.

Carson, Jamie L. 2005. “Strategy, Selection, and CandidateCompetition in U.S. House and Senate Elections.” The Jour-nal of Politics 67(February): 1–28.

Converse, Philip. 1964. “The Nature of Belief Systems in MassPublics.” In Ideology and Discontent , ed. D. Apter. Glencoe,IL: The Free Press, 206–61.

Cox, Gary W., and Mathew D. McCubbins. 2005. Setting theAgenda: Responsible Party Government in the U.S. House.New York: Cambridge University Press.

Cox, Gary W., and Mathew D. McCubbins. 2007. LegislativeLeviathan: Party Government in the House. 2nd ed. NewYork: Cambridge University Press.

Cox, Gary W., and Keith T. Poole. 2002. “On Measuring Par-tisanship in Roll Call Voting: The U.S. House, 1877–1999.”American Journal of Political Science 46(3): 477–89.

Desposato, Scott, and John Petrocik. 2003. “The Variable In-cumbency Advantage: New Voters, Redistricting, and thePersonal Vote.” American Journal of Political Science 47:18–32.

Erikson, Robert S. 1971. “The Electoral Impact of CongressionalRoll Call Voting.” American Political Science Review 65(July):1018–32.

Erikson, Robert S., and Gerald C. Wright. 2001. “Voters,Candidates, and Issues in Congressional Elections.” InCongress Reconsidered, 7th ed., ed. Lawrence C. Dodd andBruce I. Oppenheimer. Washington, DC: CQ Press, 67–96.

Fenno, Richard F. Jr. 1978. Home Style. Boston: Little, Brown.

Finocchario, Charles, and David Rohde. 2008. “War for theFloor: Partisan Theory and Agenda Control in the U.S.House of Representatives.” Legislative Studies Quarterly 33:35–61.

616 JAMIE L. CARSON, GREGORY KOGER, MATTHEW J. LEBO, AND EVERETT YOUNG

Fiorina, Morris P. 1974. Representatives, Roll Calls, and Con-stituencies. Lexington, KY: Lexington Books.

Griffin, John D. 2006. “Electoral Competition and DemocraticResponsiveness: A Defense of the Marginality Hypothesis.”The Journal of Politics 68(November): 911–21.

Jacobson, Gary. 1980. Money in Congressional Elections. NewHaven, CT: Yale University Press.

Jacobson, Gary C. 1987. “The Marginals Never Vanished: In-cumbency and Competition in Elections to the U.S. Houseof Representatives, 1952–1982.” American Journal of PoliticalScience 31: 126–41.

Jacobson, Gary C. 1996. “The 1994 House Elections in Perspec-tive.” Political Science Quarterly 111: 202–23.

Jacobson, Gary C. 2000. “Party Polarization in National Politics:The Electoral Connection.” In Polarized Politics: Congressand the President in a Partisan Era, ed. Jon R. Bond andRichard Fleisher. Washington, DC: CQ Press, 9–30.

Jacobson, Gary C. 2009. The Politics of Congressional Elections.7th ed. New York: Pearson.

Jenkins, Jeffery A., Michael A. Crespin, and Jamie L. Carson.2005. “Parties as Procedural Coalitions in Congress: An Ex-amination of Differing Career Tracks.” Legislative StudiesQuarterly 30(August): 365–89.

Kinder, Donald. 1998. “Opinion and Action in the Realm ofPolitics.” In The Handbook of Social Psychology, 4th ed., ed.Daniel T. Gilbert, Susan T. Fiske, and Gardner Lindzey. NewYork: McGraw-Hill, 778–867.

King, David C., and Richard Zeckhauser. 2003. “CongressionalVote Options.” Legislative Studies Quarterly 28(3): 387–411.

Kingdon, John W. 1989. Congressmen’s Voting Decisions. AnnArbor: University of Michigan Press.

Krehbiel, Keith. 1993. “Where’s the Party?” British Journal ofPolitical Science 23: 235–66.

Lebo, Matthew, Adam J. McGlynn, and Greg Koger. 2007.“Strategic Party Government: Party Influence in Congress,1789–2000.” American Journal of Political Science 51(3):464–81.

Mann, Thomas E., and Raymond E. Wolfinger. 1980. “Can-didates and Parties in Congressional Elections.” AmericanPolitical Science Review 74(September): 617–32.

Mayhew, David. 1974. Congress: The Electoral Connection. NewHaven, CT: Yale University Press.

McCarty, Nolan, Keith T. Poole, and Howard Rosenthal. 2006.Polarized America: The Dance of Ideology and Unequal Riches.Cambridge, MA: MIT Press.

Miller, Warren E., and Donald E. Stokes. 1963. “ConstituencyInfluence in Congress.” American Political Science Review57(March): 45–56.

Monroe, Nathan, and Greg Robinson. 2008. “Do RestrictiveRules Yield Nonmedian Outcomes? A Theory with Evi-dence from the 101st–108th Congresses.” Journal of Politics70: 217–31.

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

Rohde, David W. 1991. Parties and Leaders in the PostreformHouse. Chicago: University of Chicago Press.

Sniderman, Paul M., Richard A. Brody, and Philip E. Tetlock.1991. Reasoning and Choice: Explorations in Political Psychol-ogy. New York: Cambridge University Press.

Theriault, Sean. 2008. Party Polarization in Congress. New York:Cambridge University Press.

Wooldridge, Jeffrey M. 2002. Econometric Analysis of Cross Sec-tion and Panel Data. Cambridge, MA: MIT Press.

Wright, Gerald C., Jr. 1978. “Candidates’ Policy Positions andVoting in U.S. Congressional Elections.” Legislative StudiesQuarterly 3: 445–64.