Candidate Repositioning. Michael Tomz Stanford University

Transcription

1 Candidate Repositioning Michael Tomz Stanford University Robert P. Van Houweling University of California, Berkeley Draft: November 2010 Abstract This paper examines how voters respond when candidates change positions over time. We develop a model in which repositioning affects voter behavior through two channels. First, repositioning affects perceptions about proximity by causing voters to discount the candidate s current policy pronouncements. Second, repositioning affects perceptions about valence by prompting voters to draw negative inferences about the candidate s character. The relative weights associated with these two considerations, and, therefore, the overall response to repositioning, depend on the extent to which voters regard the issues as important to them personally. We predict that voters who place higher importance on issues will be more, not less, tolerant of candidates who espouse inconsistent positions over time. We test the model by administering survey-based experiments to a representative sample of 7,495 U.S. adults. Our data strongly confirm that repositioning changes voter perceptions about both proximity and valence; that repositioning is costly on average; and that the costs of repositioning are highest among voters who do not regard the issue as personally important. We then use our data to derive the optimal strategies for candidates. Our equilibrium analysis shows how voters, by reacting negatively to repositioning, deter politicians from adjusting their positions when public opinion changes or new policy-relevant information comes to light. These findings have important implications for candidate strategy, election outcomes, and representation in democracies.

2 1. Introduction The consistency or lack thereof in candidates positions is a perennial issue in political campaigns. When candidates change positions over time, competitors often expose the inconsistency and attempt to exploit it for electoral advantage. In the 2004 U.S. presidential election, for example, George W. Bush accused Senator John Kerry of flip-flopping from supporting the war in Iraq to opposing the war. Similarly, in the 2008 Republican presidential primaries, candidates attacked John McCain for embracing the very tax cuts he had opposed in 2001 and 2003, and they criticized Mitt Romney for shifting from the pro-choice position he held as Governor of Massachusetts to an anti-abortion position that was more popular among Republican voters. How do voters respond when candidates change positions over time? The answer is fundamental to understanding candidate competition, election outcomes, and representation in democracies. If voters punish politicians for repositioning, they may unwittingly dampen governmental responsiveness to shifts in public preferences and the arrival of new policyrelevant information. Moreover, penalties for repositioning may contribute to polarization and legislative gridlock. Elected officials would feel wary of compromising with political opponents, for fear that even voters who agree with the content of the compromise would punish candidates for veering from their past commitments. Furthermore, if voters deter incumbents from changing their positions, incumbents may get pinned to non-median positions they articulated earlier in their career, thereby giving challengers the opportunity to win with relatively extreme issue stances. Voters could react to repositioning in at least three ways. First, repositioning could affect the voter s best guess about what a candidate would do if elected (Enelow and Hinich 1984, 117; Enelow and Munger 1993). Rather than trusting the candidate s new position, voters may conclude that a candidate s true preference lies between the candidate s old position and the candidate s new one. Second, repositioning could create uncertainty in the minds of voters (Alvarez 1997; Bernhardt and Ingberman 1985; Enelow and Hinich, 1984). Voters may feel less confident in their expectations about candidates who have changed positions than about candidates who have remained consistent. Third, repositioning could affect opinions about the candidate s character, or valence (Sigelman and Sigelman 1983; Tavits 2007). On the one 1

3 hand, voters might infer that position-shifting candidates are dishonest, unprincipled votegrubbers that they are weak and unfit for public office. On the other hand, citizens may interpret repositioning as evidence that a candidate is open-minded, flexible, and pragmatic. To date, we have surprisingly little micro-level evidence about these mechanisms and no theoretical framework that unites them. In this paper, we integrate these diverse perspectives into a unified theory of how voters respond to each candidate s history of positions. We model the choices of voters as a weighted function of proximity and valence, and explain how repositioning affects both components of this decision calculus. Other factors equal, each voter prefers proximate candidates: politicians whose expected positions are most like their own. Expectations, we argue, depend not only on the candidate s current proclamations but also on what the candidate previously said. Instead of accepting the candidate s current position at face value, voters form composite expectations that blend the candidate s past and the present statements. Voters also prefer candidates with good valence, or character. We predict that voters will interpret position-shifting as evidence of poor character and will penalize candidates accordingly. In our theory, the relative weights that voters place on proximity versus valence depend on issue importance, defined as the degree to which voters feel the issue is important to them personally. The higher the level of issue importance, the more the citizen will vote based on proximity, and the less the voter will allow perceptions about valence to affect their choice. Thus, the electoral consequences of repositioning should vary with the level of importance voters assign to particular issues. Of course, repositioning could also raise uncertainty in the minds of voters by making voters less confident in their expectations about what the candidate would do if elected. Whether and how this uncertainty affects voter behavior will depend on voters attitudes toward risk and the particular configuration of candidates. As discussed in greater detail below, we designed an experiment in which the risk preferences of citizens would not affect their voting behavior. Our experimental design allows us to isolate the effects of repositioning on proximity and valence, while avoiding controversial assumptions about the risk preferences of voters. Our experiments, embedded in public opinion polls, were administered to a nationally representative sample of 7,495 U.S. adults. We presented each respondent with two candidates who differed in their record of statements on either tax policy or abortion policy, and we asked 2

4 which candidate the respondent preferred. We also gathered information about respondents own policy preferences, the personal importance they assigned to the issue, their expectations about the candidates, and their evaluations of the candidates traits. Consistent with our theory, we find strong evidence that repositioning affects both proximity and valence calculations. We further find that, on average, candidates who repositioned performed worse than candidates who stood firm. This finding contrasts with related research, which shows that voters do not respond negatively to candidates who made vague policy statements (Tomz and Van Houweling 2009). Politically, then, changing positions is more costly than remaining ambiguous. Based on our model and data, we derive implications for candidate strategy and democratic representation. We show how discounting and valence considerations combine to reduce but not eliminate the incentives for candidates to respond to shifts in public opinion. Our theory predicts, and our experimental results confirm, a further counterintuitive claim: voters who place higher importance on issues will be more, not less, tolerant of candidates who espouse inconsistent positions over time. Consequently, candidates will provide better representation on issues the electorate holds dear, while remaining out of step on less important issues for which the negative effect of repositioning on valence outpaces the potentially positive effect of repositioning on proximity. By accounting for issue importance, our theory also makes sense of seemingly puzzling patterns. For example, Tavits (2007) hypothesizes that voters are more tolerant of repositioning on pragmatic issues such as economic policy than on moral issues such as abortion. We find, however, that citizens are no more likely to penalize candidates for changing positions on abortion than for changing positions on taxes. This pattern arises from the moderating effect of issue importance. Repositioning on abortion undermines perceptions about character to a greater degree than repositioning on taxes, but citizens in our sample regard abortion as more important to them personally, and therefore place less weight on perceptions of character (and relatively more weight on proximity). Our results help explain why, even on moral issues such as abortion, candidates may find it profitable to change positions if the issue is sufficiently important to constituents. In the remainder of this paper, we develop a theory of voter responses to candidate position-taking over time; describe the design and analyze the results of our survey experiment; 3

5 and discuss the implications for candidate strategy, election outcomes, and representation in democracies. 2. Theory and Hypotheses Broadly speaking, there are two schools of thought about how citizens compare candidates and decide for whom to vote. Downs (1957), and much of the rational choice literature that ensued, emphasizes the role of policy proximity. According to proximity theory, citizens prefer candidates whose policy positions are closest to their own. Stokes (1963) and much of the behavioral literature, on the other hand, stress non-positional considerations. Voters, it is argued, prefer candidates with desirable valence characteristics, including personality traits such as competence and integrity. We incorporate both perspectives into a model that explains how repositioning influences voter behavior. In our model the intertemporal consistency of a candidate s policy positions affects perceptions about both proximity and valence. Moreover, these effects often run in opposite directions: efforts to improve one s proximity could undermine one s valence, and viceversa. The net effect, we argue, depends on the importance voters attach to the issue. Highimportance voters, for whom the issue makes a great difference personally, are especially sensitive to the ways in which repositioning affects proximity. Low-importance voters, in contrast, are disproportionately responsive to the ways in which repositioning affects valence. In the remainder of this section, we present our theory of voter decision making and connect it with the previous literature. 2.1 Proximity Repositioning affects the proximity component in a voter s decision making calculus by altering expectations about what a candidate would do if elected. Compared to candidates who stand firm, candidates who reposition are likely to be seen as less credible proponents of their current position. Page (1978, 150) writes that a candidate who takes new issue stands will not be wholly convincing to the new constituencies to which he is trying to appeal. By way of 4

6 illustration, Page cites widespread doubts about whether presidential candidate George McGovern s 1972 post-primary shift towards the center was genuine. To formalize this intuition, consider a candidate who took policy positions on a single issue at two points in time. Denote the candidate s previous position as c 1, and denote the candidate s current position as c 2. When c 1 = c 2, we say the candidate has stood firm; when c 1 c 2, we regard the candidate as having repositioned. We model the voter s expectations about what the candidate would do if elected as cˆ c2 (1 ) c1. Thus, expectations are a convex combination of the candidate s policy pronouncements. The weight λ (0 λ 1), which Enelow and Munger (1993) call credibility, measures how much emphasis voters place on the candidate s current position relative to his past position. The expectation ĉ is important because it affects perceptions about proximity. Let x be the voter s policy position, or ideal point. We model the proximity component in the voter s utility function as cˆ x, such that utility declines with the absolute distance from the voter to the expected location of the candidate. The exponent ω determines the shape of the utility function. If ω=1, for example, the function is linear and the voter is risk neutral. If ω=2, on the other hand, the voter has a quadratic loss function and is, therefore, risk-averse. 2.2 Risk In addition to affecting expectations about what the candidate would do if elected, repositioning could change a voter s level of uncertainty. Voters may be less sure about the likely actions of candidates who have changed positions than about the likely actions of candidate who have stood firm. As several authors have argued, this additional uncertainty could affect voter behavior. If repositioning contributes to uncertainty and voters have risk-averse utility, two consequences ensue. First, candidates who reposition will be disadvantaged relative to otherwise identical candidates who have not changed their stances. Second, this disadvantage could produce equilibria in which the candidates diverge from each other and from the median voter (Enelow and Hinich 1984; Bernhardt and Ingberman 1985; Enelow and Munger 1993; Alvarez 1997). These theoretical results rely heavily on the assumption that voters are risk averse, however. Recent research supports an alternative view. Using data from the National Election Study, Berinsky and Lewis (2007) find that American voters are generally neutral about risk. 5

7 Tomz and Van Houweling (2009) reach a similar conclusion via experiments; they manipulated how much information voters received about the policy positions of candidates and found that voters did not prefer candidates about whom they knew more. 1 Moreover, even if voters were risk averse, their behavior would depend on the locations of the candidates. By definition, risk-neutral voters are indifferent between receiving ĉ with certainty and receiving a lottery with an expected value of ĉ. Risk-averse voters, on the other hand, would shy away from the lottery in favor of receiving ĉ with certainty. It does not follow, however, that risk-averse voters would choose high-certainty candidates (those whose position is known with confidence) at a higher rate than risk-neutral voters. To illustrate, suppose a voter with ideal point x faces a choice between two candidates, A and B. The voter feels certain that candidate A would adopt policy a and believes that candidate B would adopt a position somewhere between b 1 and b 2. The voter s beliefs about candidate B can be represented as a probability distribution p(b) over the support [b 1, b 2 ], with expected value bˆ. Under what conditions would risk aversion cause voters to choose the certain candidate over the uncertain one? To answer this question, we must identify scenarios under which a riskneutral voter would prefer B but a risk-averse voter would prefer A. A risk-neutral voter would prefer B if and only if the voter were closer to B s expected location than to A s known location. Within this subset of cases, a risk-averse voter would prefer A only if B might deliver a worse policy (a policy farther from the voter s ideal point) than A would deliver. In notation, the necessary conditions are x bˆ x a and Pr( x b x a ) 0 for some b in [b 1, b 2 ]. These conditions are necessary but not sufficient. Even if the conditions were satisfied, a risk-averse voter prefer the uncertain candidate unless the probability of bad outcomes were large enough, and the utility from the bad outcomes were low enough, to overcome B s expected proximity advantage. These facts are important for understanding voter behavior. Voters sometimes choose certain candidates over certain ones, but without additional information it would be wrong to interpret such choices as evidence of risk aversion. When the uncertain candidate poses no 1 Earlier research on this question was mixed, with some analysts concluding that voters dislike uncertainty and appear to be risk-averse (Alvarez 1997; Bartels 1986; Brady and Ansolabehere 1989) and others reaching the opposite judgment (Campbell 1983). 6

8 greater risk, in the sense that the worst possible realization in [b 1, b 2 ] is no further from the voter than a, risk aversion cannot be driving an observed preference for A over B. We designed experiments that took these theoretical points into account. With only a few exceptions, the scenarios that we administered did not satisfy the necessary conditions x bˆ x a and Pr( x b x a ) 0 for some b in [b 1, b 2 ]. Consequently, if voters in our experiments preferred certain candidates over uncertain ones, this preference typically could not be a consequence of risk aversion. Guided by theory, then, we constructed experiments that ruled out one potential causal pathway (uncertainty), in order to shine a more focused light on the other pathways proximity and valence through which repositioning might affect the preferences and behavior of voters. 2.3 Character (Valence) In addition to affecting perceptions about proximity, repositioning could trigger negative impressions about a candidate s character. Previous experiments, conducted with college students, lend plausibility to this hypothesis. In one experiment, Allegeier et al. (1979) presented students with questionnaires that had been filled out by an anonymous individual at two points in time and asked students to evaluate the individual on a variety of traits. Overall, students gave lower evaluations to individuals whose answers to the questionnaire had changed over time. Carlson and Dolan (1985) reached similar conclusions in an analogous experiment, in which the individual who filled out the surveys was described as a political candidate. 2 Admittedly, voters may see virtue as well as vice in candidates who change positions. In a novel experiment, Sigelman and Sigelman (1986) examined how people responded to fictional presidents whose policy actions were either consistent or inconsistent with their reputation as a foreign policy hawk or dove. Respondents generally perceived presidents of either persuasion who acted out-of-character as more dishonest, inconsistent, insincere, unreliable, and indecisive than presidents who stuck with their previous stands. Still, subjects also saw some merits flexibility and open-mindedness in a president who could adapt to the circumstances, a fact that 2 Other experiments, though, lead to less definitive conclusions. Hoffman and Carver (1984), for example, experimentally examined how people evaluated candidates who changed positions on gun control. Consistent candidates scored higher on only two traits: whether the candidate was decisive, and whether he was well adjusted. 7

9 raises questions about any view that casts stepping out of character in an exclusively negative light (283). On balance, we expect voters to view repositioning as evidence of bad character, while acknowledging that some voters could view repositioning as a signal of certain positive traits, such as open-mindedness. Finally, we anticipate that the valence effects of repositioning will vary by type of issue. In a pioneering nonexperimental study, Tavits (2007) examined how the electoral fates of parties in forty advanced democracies changed when they shifted platforms over time. She hypothesized that voters would be more tolerant of shifts on pragmatic issues such as fiscal policy than on moral issues such as abortion, since on pragmatic issues the optimal policy may vary with the circumstances. Her time-series cross-sectional analysis of party platform data supports these arguments. Building on Tavits, we expect that repositioning will evoke more negative assessments of character when policy issues are moral than when policy issues are pragmatic. Putting all these arguments together, we now model the effects of repositioning on both proximity and valence, and explain how voters weigh these two effects. In our framework, the utility a voter receives from a candidate is U cˆ x (1 ) v, where cˆ x and v represent the proximity and valence components in the voter s utility function. Repositioning changes this equation in two ways: by shifting ĉ away from c 2 in the direction of c 1 ; and by triggering negative inferences about character, thereby reducing the value of v. The weight α (0 α 1) measures how personally important the issue is to the voter. As α rises, voters become more sensitive to the effects of repositioning on proximity while becoming less sensitive to the effects of repositioning on valence. 3 Our model implies, somewhat counterintuitively, that voters who view an issue as important will punish candidates for repositioning on that issue less than voters who view an issue as unimportant. At one extreme, voters who do not care about the substance of an issue (α=0) will base their decisions entirely on valence, which we hypothesize to be negatively 3 Erikson and Romero (1990) propose a related model, in which voters weigh both issues and non-issue considerations. They show that as issue salience increases, citizens are increasingly likely to base their votes on issues rather than candidate characteristics (see Figures 1, 3, 4, 5 on pp ). The logic of our model is similar, but in our theoretical framework candidates take issue positions at several points in time, and those issue positions affect not only perceptions about proximity but also judgments about valence. 8

10 affected by repositioning. At the opposite extreme, voters who assign maximal importance to the issue (α=1) will focus single-mindedly on how well a candidate represents their views on the issue, giving no weight to perceptions about character. Although most voters fall between these extremes of α=0 and α=1, the general pattern should hold: the costs of repositioning should fall, not rise, with issue importance. 4 In summary, we present a model of repositioning that combines and extends insights from two scholarly traditions: proximity theory and valence theory. Our model has many behavioral implications that can be tested with micro-level data from survey experiments. If the model is correct, repositioning should affect calculations about proximity by inducing voters to discount a candidate s current policy stance. Repositioning should also affect perceptions of valence by undermining evaluations of most candidate characteristics, including core measures of competence and integrity, while enhancing scores on selected traits such as open-mindedness. Finally, the cost of repositioning should decrease, not increase, with the importance of the issue. We test these implications, and several others, in the remainder of the paper. 3. Experimental Design We designed an experiment to assess the effect of candidate repositioning on voter choice and embedded it in opinion polls that were administered over the Internet to a nationally representative sample of 7,495 U.S. adults. Our experiment involved four steps. First, we measured respondents preferences about an important public policy issue and asked how important the issue was to them. Second, we described the policy statements of two candidates who varied randomly in whether and to what degree they had changed positions on the issue, and then asked which candidate the respondent preferred. Third, we measured the expectations of respondents by asking what they thought each candidate would try to do on the issue if elected. Finally, we invited respondents to evaluate each candidate on one randomly selected trait. To see whether the effects of repositioning varied by issue, we split the sample into two groups. The first group, totaling 4,155 respondents, received a questionnaire in which the topic 4 We are not arguing that low-importance voters will draw harsher conclusions about the traits of candidates who change positions. Our model does, however, imply that low-importance voters will weigh character more (and weigh proximity less) in their assessment of the overall utility the candidate provides. 9

11 was the level of taxes on wealthy Americans. The second group, consisting of 3,340 respondents, received a questionnaire that asked about restrictions on abortion. Each respondent considered only one of these two issues. We now describe the survey instrument in greater detail. 3.1 Measuring the Respondent s Policy Preferences The survey began by asking respondents about an important policy issue: either taxes on wealthy Americans or restrictions on abortion. For those assigned to consider the issue of taxes, we noted that some people think the government should increase taxes on wealthy Americans, defined as people who make more than $250,000 per year. Other people think taxes on wealthy Americans should be kept at their current level. And still other people think the government should decrease taxes on wealthy Americans. We then asked whether the respondent thought the government should increase taxes on wealthy Americans, keep taxes on wealthy Americans at their current level, or decrease taxes on wealthy Americans (Figure 1a) [Figure 1 about here] For those assigned to consider the issue of abortion, we explained that some people think the government should increase restrictions on abortion to make it more difficult for women to get abortions. Other people think restrictions on abortion should be kept at their current level. And still other people think the government should decrease restrictions on abortion to make it easier for women to get abortions. We then asked whether the respondent thought the government should increase restrictions on abortion, keep restrictions on abortion at their current level, or decrease restrictions on abortion. The policy options were described verbally rather than numerically, though for convenience we sometimes use integers as shorthand for the three fully labeled options. In those instances, the integer one represents the most liberal position (increase taxes on wealthy Americans or decrease restrictions on abortion), two represents a moderate position (keep taxes on the wealthy or restrictions on abortion at their current levels), and three represents the most conservative position (decrease taxes on wealthy Americans or increase restrictions on abortion). 10

12 Immediately after gathering respondents opinions about the issue they were assigned, we asked, How important to you personally is the issue of [abortion/taxes on wealthy Americans]? The response options were: extremely important, very important, moderately important, slightly important, or not at all important. 3.2 Measuring Preferences about Candidates We then asked respondents to choose between two candidates who had expressed opinions about the issue. In the case of taxes, for example, we noted that Non-partisan groups often survey candidates about tax rates for wealthy Americans. We would like your views about two candidates, whose names will remain confidential. They are Candidate A and Candidate B. By denoting candidates with letters, we were able to test theories about the effects of repositioning in their purest form, without the potentially confounding effects of party or other candidate attributes. We reported what the candidates had said on two occasions: two years ago and this year. On each occasion, a candidate took one of the three positions that respondents had considered when expressing their own opinion on the issue. Let L, M, and C represent the liberal, moderate, and conservative position on each issue. Using this notation, each candidate s history of statements on the issue can be summarized as (L,L), (L,M), (L,C), (M,L), (M,M), (M,C), (C,L), (C,M), or (C,C), where the first letter represents what the candidate said two years ago and the second letter represents what the candidate said this year. After eliminating cases in which the candidates had identical histories and could, therefore, be expected to tie, we were left with thirty-six combinations of candidate histories. We randomly assigned each respondent to receive one of these thirty-six combinations and randomly labeled one of the candidates as A and the other as B. We then asked respondents On this issue, which candidate do you prefer? An example involving taxes appears in Figure 1b; the question format for abortion was identical. In only a small number of these combinations could risk aversion cause a voter to choose the consistent candidate over the inconsistent one. Specifically, risk aversion could matter if voters at the poles encountered one candidate who consistently located at the moderate position (M,M), and an opposing candidate who repositioned from one pole to the other (L,C or C,L). In 11

13 these scenarios, risk aversion could affect the voter s choice, but only if the following additional conditions are met: the voter is closer to the expected location of the inconsistent candidate than to the known location of the consistent candidate; the voter perceives a sufficiently high likelihood that the inconsistent candidate will implement a policy that is further from the voter than the known location of the consistent candidate; and the disutility arising from these bad outcomes is large enough to outweigh to the utility associated from the candidate s expected proximity advantage. By a similar logic, risk aversion could be consequential if a voter at one of the poles faces either (1) a candidate who shifts from the voter s ideal point to the center, and an opponent who moves across the entire space to the voter s ideal point; or (2) a candidate who shifts from the center to the voter s ideal point, and an opponent who flips from the voter s ideal point to the opposite pole. In either scenario, risk aversion could lead a voter to choose the candidate who moved less, but only if the candidate who moved across the entire space was the expected proximity favorite, and yet the voter regarded that candidate as being sufficiently likely to take a worse position than the candidate who moved to a smaller degree. Finally, risk is relevant when a voter in the center faces one candidate who consistently advocates the centrist position and an opposing candidate who repositions from one pole to the other. Consider the knife-edge case in which the voter expects both candidates to implement the voter s ideal policy, i.e., the expected policy of the inconsistent candidate aligns exactly with the known position of the consistent candidate and the voter s ideal point. In this special case, all voters, regardless of the power term ω in their utility function, would prefer the certain candidate, who guarantees the voter s ideal point, over the uncertain candidate, who presents some chance of delivering a less attractive policy outcome. Although possible in theory, this scenario involves a very particular set of beliefs: voters must assign exactly the same weight to the candidate s past statements as to the candidate s current statements, i.e., it requires that λ be exactly.5. The scenarios in which risk aversion could matter represent only 5% of all the scenarios in our experiment. Below, we present a pooled analysis that includes all scenarios. The results do not change noticeably when we remove the scenarios in which risk aversion could play a role. Consequently, to extent that voters in our experiment react negatively to repositioning, the negative reaction must not be due to risk aversion. 12

14 3.3 Measuring Expectations about Future Action The next phase of the experiment investigated how citizens form expectations about the likely positions of candidates. We reminded each respondent of the statements candidate A had made and asked: If you had to guess, what do you think Candidate A would try to do if elected? For respondents who were randomly assigned to the issue of taxes, the response options were increase taxes on wealthy Americans, keep taxes on wealthy Americans at their current level, or decrease taxes on wealthy Americans (see Figure 1c). For respondents who completed the questionnaire about on abortion, the question layout was identical except the phrase taxes on wealthy Americans was replaced with the phrase restrictions on abortion. We repeated this procedure for candidate B. 3.4 Evaluating Each Candidate s Traits The final phase of the experiment asked respondents to evaluate each candidate on a randomly selected trait. We included the traits honest and provides strong leadership to tap the two central dimensions of candidate evaluation, integrity and competence, which others have identified as mediating candidate support (Kinder, et. al 1980, Markus 1981). We added a third trait, open-minded, not only because voters regard it as important (Kinder et. al 1980), but also because it is a dimension on which candidates could score points by changing positions instead of standing firm (Sigelman and Sigelman 1986). We randomly assigned each respondent to evaluate the candidates on one of the three traits. To do this, we redisplayed Candidate A s history of positions on the issue and asked: In your opinion, does the phrase he [is honest / provides strong leadership / is open minded] describe Candidate A extremely well, very well, moderately well, slightly well, or not well at all? We concluded by asking respondents to judge Candidate B on the same trait (Figure 1d.). 4. Data and Findings The experiments discussed in this paper were administered by Knowledge Networks, an Internetbased polling firm. Knowledge Networks uses random digit dialing to recruit participants and 13

15 provides Internet access to households, resulting in a nationally representative sample of adults. The interviews took place between February and December of 2009, and 7,495 people (about 60% of invitees) agreed to take the survey. Table 1 presents the distribution of preferences on the two issues. [Table 1 about here] 4.1 The Average Effect of Inconsistency We begin by estimating the average effect of repositioning and later turn to a deeper evaluation of the mechanisms through which repositioning affects support for candidates. Under the null hypothesis that voters focus entirely on the current positions of candidates, thereby ignoring what candidates said in the past, support for candidates who change positions should be no higher or lower, on average, than support for candidates who stand firm. If, on the other hand, voters incorporate the past positions of candidates into their decision-making calculus, our model predicts that, on average against all possible opponents, candidates who stand pat will outperform candidates who shift. To estimate the average effect, we computed the proportion of times that respondents preferred an inconsistent candidate (a candidate with L,M; L,C; M,L; M,C; C,L; or C,M) in pairwise competition against all possible opponents, both consistent and inconsistent. Similarly, we computed the proportion of times that respondents preferred a consistent candidate (one with history L,L; MM; or C,C) in races against all conceivable opponents. The difference between these two values represents the average effect of repositioning. As noted earlier, we did not present respondents with scenarios in which the two anonymous candidates had identical histories, because respondents would have found no basis for preferring one over the other. The choice between Candidate A and Candidate B would have reduced to the flip of a coin. In actual political competition, of course, candidates sometimes take identical positions, so a complete set of opponents would include those with policy histories exactly like one s own. We reincorporated these ties into our analysis by assuming that, if candidates had competed against their clones, each would have received 50 percent of the vote. 14

16 Table 2 confirms our hypothesis that voters dislike inconsistency generally, but also holds an interesting surprise. On taxes, candidates who changed positions over time received 46.2 percent of the vote on average, whereas candidates who stood firm received 57.7 percent. The estimated cost of repositioning was, therefore, 11.5 points. The 95 percent confidence interval around this estimate ranged from 13.1 to 9.8, enabling us to reject the null hypothesis that voters are indifferent about inconsistency on taxes. [Table 2 about here] On abortion, the penalty for repositioning was slightly smaller, with the average inconsistent candidate receiving 46.7 percent of the vote and the average consistent candidate receiving 56.7 percent, for a net difference of 10 percent. The confidence interval around this difference spanned to -8.3 percent. This finding disconfirms our hypothesis that voters respond more harshly to policy shifts on the moral issue of abortion than on the pragmatic issue of taxes. In fact, our simulations suggest that the cost of repositioning on taxes exceeds the cost on abortion approximately 90% of the time. Later in the paper, we explain how our model can explain this surprising pattern. 4.2 How Past Positions Shape Expectations and Choice We now examine the mechanisms that underlie the substantial cost of repositioning. Our theory holds that voters discount (look skeptically upon) the current policy positions of candidates who espoused different policies in the past. To test this hypothesis, we measured respondents expectations about what the candidates would try to do if elected. Our evidence, presented in Figure 2, strongly supports the discounting hypothesis. For each of the nine possible histories a candidate could possess, panel (a) summarizes respondents expectations about what a candidate with that history would try to do if elected. The upper row pertains to candidates who advocated the liberal position (increasing taxes on wealthy Americans) this year. The row contains three dots: one for candidates who espoused the liberal position two years ago, one for candidates who were moderate two years ago, and one for 15

17 candidates who favored the conservative position two years ago. The horizontal coordinate of each dot indicates the mean expectation about candidates with the identified history. [Figure 2 about here] For example, the leftmost dot denotes average expectations about a candidate who had advocated increasing taxes two years ago and also advocated increasing taxes this year. When asked about such candidates, 82 percent of respondents thought the candidate would try to increase taxes. Only 15 percent felt the candidate would keep taxes at the same level, and a miniscule 3 percent guessed that the candidate would try to reduce taxes. On a scale from 1 to 3, where 1 represents an effort to increase taxes and 3 represents an effort to decrease taxes, the average expectation about candidate (L,L) was (.82 x 1) + (.15 x 2) + (.03 x 3) = 1.2. Voters were significantly more skeptical of candidates who switched from advocating higher taxes two years ago to calling for lower taxes today. The rightmost dot on the top row indicates mean expectations about these candidates. Only 56 percent of citizens who evaluated this type of candidate expected him to strive for lower taxes. Fully 32 percent guessed the candidate would maintain the status quo, and 12 percent expected the candidate to stick to his prior position and push for tax cuts. The mean placement of candidates with history (C,L) was, therefore, 1.6. Clearly, voters had less faith in the pro-tax intentions of candidates who only recently embraced that position, than of candidates who had proposed it repeatedly. The center dot in the top row refers to candidates who advocated increasing taxes this year but had previously called for retaining the status quo. A large majority of respondents, 70 percent, believed these candidates would actually try to increase taxes if elected. However, because the remaining 30 percent doubted these candidates to some degree, the mean expectation fell between expectations about candidates who had either stood firm or had flipped across the policy space. The 95% confidence intervals around these means are narrow, giving us great certainty that voters discount the present promises of candidates who have repositioned. The pattern repeats for each row in the figure, not only for taxes but also for abortion. Comparing each row to the adjacent row provides an additional insight: on average, voters place more weight on a candidate s current positions than on his past positions. Compare, for example, candidates who now advocate increasing taxes, versus candidates who call for 16

18 keeping taxes the same. The top two rows in panel (a) display expectations about these candidates conditional on their past commitments. As the figure clearly shows, voters regard candidates with histories L,M (the first dot on the second row) as more conservative than candidates who advocated C,L (the last dot on the first row). Thus, when voters form opinions about the likely behavior of candidates, even the most extreme difference in previous commitments (liberal versus conservative) cannot compensate for a more modest difference in current pronouncements (moderate vs. liberal). This pattern recurs in every comparison between adjacent rows except one. 5 Having found that expectations depend not only on current but also on previous policy statements, we next investigated how past positions affect the choices of respondents. To isolate the effect of past positions on voter choice, we defined the proximity advantage of candidate A over candidate B at time t (where t=1,2) as the degree to which the voter was closer to A s stated position than to B s stated position. Thus, A s proximity advantage two years ago would be xb 1 - x- a 1, and A s proximity advantage this year would be x- b 2 - x-a 2. Similarly, we defined candidate B s proximity advantage two years ago as x- a 1 - x-b 1, and his proximity advantage today as x- a 2 - x-b 2. In our experiment, a candidate s proximity advantage at time t could be -2, -1, 0, 1, or 2. Figure 3 displays the support candidates received, depending on their past and present proximity advantage. (The figure is a graphical representation of the 5x5 cross-tabulation of past proximity advantage by present proximity advantage.) Each dot in the figure gives the estimated level of support for a particular type of candidate. For example, the dot in the uppermost left corner of panel (a) indicates that, on the issue of taxes, candidates who held a proximity advantage of -2 two years ago and a proximity advantage of 2 this year were preferred by 97 percent of respondents over their opponents. Each line connects the set of point estimates for 5 Figure 2 also shows that people are, in general, skeptical of candidates who advocate tax cuts for wealthy Americans. This can be seen in the tendency of the expectations on taxes to be shifted to the left in the issue space. Only 69% of respondents believed that a candidate who consistently advocated lower taxes will actually attempt to reduce taxes if elected, compared to 82% of respondents who believed a candidate who repeatedly called for higher taxes on the rich. 17

19 which the proximity advantage of candidates this year is held constant while their proximity advantage two years ago varies along the horizontal axis from -2 to 2. [Figure 3 about here] Figure 3 presents the effect of past positions, controlling not only for present positions but also for valence. We achieve control over valence because each point in Figure 3 represents an equal mix of positive and negative valence scores. To see why each point is balanced in this way, consider the top left point, which summarizes support for a candidate who previously had a two-point proximity disadvantage, but who now enjoys a two-point proximity advantage. For this kind of scenario to arise, both candidates would have to switch places: one would have to move from L to C, and the other would have to move from C to L. Because both candidates repositioned to the same degree (both flipped across the entire scale), neither should have a valance advantage over the other. Next consider the adjacent point, which summarizes support for candidates who improved their proximity advantage from -1 to +2. This could happen in through two scenarios, both involving voters at the poles. In one scenario, a candidate who previously stood with the voter flips to the opposite pole, while a candidate who previously stood at the center shifts to the voter. For example, a voter at L might face candidates with histories (M,L) and (L,C). In the second scenario, a candidate who previously opposed the voter now shares the voter s position, while the opponent moved from the center to the pole opposite. This could occur if, for example, a voter at L faced one candidate with history (C,L) and another with history (M,C). Once again, the scenarios are balanced with respect to valence. In the first scenario, the candidate with the current proximity advantage shifted one position, whereas his opponent shifted two. In the second scenario, by contrast, the candidate with the current advantage shifted two positions, whereas his opponent moved only one. By design, these two scenarios counterbalance each other. Thus, on average, the candidate with the current proximity advantage has neither more nor less valence than his opponent. The same is true, either deterministically or stochastically, for every point in the figure. Consequently, the figure isolates the effect of past positions, controlling not only for current proximity but also for valence. 18

20 Figure 3 shows that the past positions of candidates have a powerful effect on voting behavior. This is most easily seen by focusing on the lines that connect the points for which the two candidates are tied in the current period, i.e. the proximity advantage this year is 0. On the issue of taxes, the leftmost point on this line indicates that candidates who suffered a two point proximity disadvantage two years ago received only 28 percent of the vote. Their competitors, identified by the rightmost point on the same line, received the other 72 percent. The analogous line for abortion, presented in panel (b), is similar if not slightly steeper. Overall, voters seem to be slightly more responsive to the past positions taken by candidates on abortion policy than on tax policy. This is most evident when considering candidates who had two-point proximity advantages two years ago. On both issues, voters almost always supported such candidates if the candidates also had a proximity advantage or were in a proximity tie this year. However, on the issue of tax policy, voters supported such candidates slightly less than 20 percent of the time if the candidates had a one-point disadvantage in the current period, and just 3 percent of the time if the candidates had a two-point disadvantage. On the issue of abortion voters offered much more support for analogous candidates, 35 percent and 10 percent, respectively. In summary, the past policy positions of candidates affect not only what citizens expect, but also how they vote. The next section considers the valence effects of repositioning. 4.3 Valence Implications of Repositioning In this section we examine whether candidates who reposition receive significantly less support than those who hold firm, after controlling for the proximity effects of repositioning. We find strong evidence that candidates pay a direct cost for changing positions, per se. We then turn our attention to evidence that bears directly on whether voters draw negative inferences about the character of candidates who change positions. To test for the valence mechanism, we first needed to control for the effects of repositioning on perceived proximity. Define the expected positions of candidates A and B as â and bˆ, respectively. For each respondent, we measured these expected positions as equal to the respondent s own perceptions about what the candidates would do if elected. With this information we then calculated an imputed proximity advantage for each candidate a voter faced. 19

21 The imputed proximity advantage for candidate A is x bˆ x aˆ, and the imputed proximity advantage for candidate B is x aˆ x bˆ, where x is a voter s ideal policy. Voters can impute candidates to occupy any point on the three point scale. Consequently, imputed proximity advantage is an integer that ranges from -2 to 2. We next defined a variable that captures the valence advantage of candidates relative to their opponents. This variable is the number of positions the opponent moved, minus the number of positions the candidate moved. In symbols, A s valence advantage is b b a, a1 while B s valence advantage is the opposite. Calculated this way, valence advantage ranges from -2 (when a candidate who moved across the entire policy space faces a candidate who stood firm) to +2 (when a candidate who stood firm faces a candidate who flipped from one pole to the other.) Figure 4 displays the share of the vote candidates received in each cell of the 5x5 crosstabulation of imputed proximity advantage by valence advantage. Each dot in the figure gives the estimated level of support for a particular type of candidate. For example, the dot in the upper left corner of panel (a) indicates that candidates who suffered a two-point valence disadvantage, but enjoyed a two-point imputed proximity advantage, were preferred by 85 percent of respondents over their opponents (on the issue of taxes). Each line connects the set of estimates for which imputed proximity advantage is held constant, while valence advantage varies along the horizontal axis from -2 to 2. [Figure 4 about here] The importance of valence considerations is most apparent when the voter perceives that the two candidates are tied in proximity. On the issue of taxes (panel a), such candidates received the support of only 35 percent of respondents when they suffered a two point valence disadvantage. Symmetrically, these candidates opponents received the support of the remaining 65 percent of respondents. Panel (b) shows very similar effects of valence when abortion policy is at stake. On both issues, the valence effect remains apparent, but is tempered to some degree, when voters perceive a one-point difference in the proximity of competing candidates. It is 20

22 tempered even further when voters perceive a two-point proximity discrepancy between their choices. Thus, candidates appear to pay a price for repositioning, per se, after controlling for the effect of repositioning on proximity. These findings must be viewed with caution, however, because they condition on post-treatment assessments about where candidates stand (see Imai, Tingley and Yamamoto (2010) and Bullock, Green, and Ha (2010) for good discussions of the stringent assumptions necessary to identify mechanisms with this inferential strategy). Furthermore, these findings leave open the question of why repositioning has negative consequences for candidates. Our experimental design rules out the possibility that risk aversion is driving the apparent distaste for inconsistent candidates. We hypothesize that repositioning is costly because it encourages voters to conclude that position-switchers have poor character. To test this hypothesis directly, we asked each respondent to assess the candidates they encountered on one of three randomly selected traits: he is a strong leader, he is honest, or he is open-minded. Higher ratings on the five-point scale correspond with more positive trait assessments. Table 3 presents the mean trait ratings of candidates who held firm over time, and also displays how these means were affected when candidates repositioned by one and two steps. [Table 3 about here] Across both issues, people were less likely to ascribe strong leadership to candidates who repositioned than to candidates who had stood firm. As expected, though, the negative effect was more evident on the issue of abortion than on tax policy. On abortion policy, the mean score for strong leader was 1.3 points lower for candidates who moved two steps in the policy space than for candidates who stood firm. The analogous effect of moving two steps on tax policy was only 0.9 points. The 95% confidence intervals around these estimates allow us to be quite certain that repositioning on abortion undermined perceptions of strong leadership to a greater degree than repositioning on tax policy. Relative to candidates who stood firm, candidates who repositioned were also more likely to be classified as dishonest. This negative effect was marginally larger on taxes than on 21

23 abortion, 1.4 and 1.3 points respectively, although the two effects were not statistically distinguishable. Respondent evaluations of candidates on the final trait, open-mindedness, reveal a small potential upside to repositioning. On both issues, a candidate who moved one position was viewed as marginally more open-minded than a candidate who stood firm. However, once a candidate moved two steps, this effect essentially disappeared, leaving us unable to reject the null hypothesis that candidates who flipped from one pole to the other were regarded as no more or less open-minded than candidates who stuck to their guns. The fact that repositioning undermines perceptions of leadership and honesty, two of the most important candidate traits scholars have identified (Kinder, et. al. 1980, Marcus 1981), helps explain why policy inconsistency has valence costs. However, these findings do not help resolve the puzzle of why, contrary to the existing literature and our expectations, voters punish candidates who reposition on the pragmatic issue of tax policy as much or more than they punish those who reposition on the moral issue of abortion policy. In the next section, we explain how our theory can solve this puzzle. 4.4 The Mediating Role of Issue Importance As issues become more personally important, we predict that voters will weigh valence less heavily, and will weigh proximity more heavily, when choosing between candidates. To test this hypothesis, we asked voters how important the issue they were assigned was to them personally. Table 4 presents the distribution of responses on this question by issue. Of special note, 23% of respondents regarded abortion policy as extremely important, while only 15% put tax policy in that high-priority category. Thus, if importance conditions how people respond to repositioning, its role may help explain why the cost of repositioning is not higher for abortion than for taxes. [Table 4 about here] To identify the effect of issue importance, we divided the sample into two groups: respondents who viewed an issue as extremely or very important and those who viewed it as less important. For each issue, we then computed the overall cost of repositioning for the high- 22

24 importance subgroup, and separately for the low-importance subgroup (Table 5). Recall from our discussion of Table 2 that this analysis is based on pairing each type of candidate against all possible opponents. The difference in the vote shares of consistent and inconsistent candidate provides an estimate of the overall cost of inconsistency. [Table 5 about here] On both taxes and abortion, respondents who see the issue as important are significantly less averse to selecting candidates who reposition than are respondents who see the issue as unimportant. The pattern is strongest on tax policy, where people who feel intensely about the issue dock candidates who reposition by only 7 points. In contrast, people who do not feel intensely about the issue punish inconsistent candidates at more than twice that rate, disfavoring them by 15 points. The analogous punishments for inconsistency on abortion are 8 and 12 points. This yields a smaller, but still statistically significant, difference of 4 percentage points between voters who regard the issue as important to them personally, and voters who do not. Moving beyond confirming our basic hypothesis about the role of issue importance, we next test our expectation that voters will put decreasing weight on valence considerations and increasing weight on issue proximity as they view an issue as more important. To test this we specified a logistic regression model in which the dependent variable is whether the respondent chose a particular candidate. The key independent variables are the valence advantage of the candidate over his opponent, the imputed proximity advantage of the candidate over his opponent, and the interactions of these two terms issue importance (dichotomized as high or low). [Table 6 about here] The results, presented in Table 6, largely conform to our expectations. On both issues, as the valence and imputed proximity advantages of candidates grow, they receive support from a significantly larger share of voters. As we hypothesized, the benefit a candidate accrues from having a valence advantage is smaller among high importance voters than low importance voters. This negative interaction effect is large and statistically significant by conservative standards. There is also more tentative evidence in favor of our hypothesis that voters who view an issue as 23

25 important place more weight on proximity than voters who do not view an issue as important. For both issues, the estimated coefficient on the interaction between a candidate s proximity advantage and importance of the issue is positive. This interaction, however, only meets the standard threshold of statistical significance on the issue of taxes. Figure 5 graphically displays the estimated interaction effects. Panel (a) presents the effects for tax policy and panel (b) for abortion policy. The top graph in each panel displays the marginal effect of valence advantage, holding proximity constant at 0. The solid line in each graph identifies the estimated effect among people who see the issue as important, while the dashed lines represent the effects for people who do not regard the issue as important. On both issues, high-importance voters choose a candidate with a two-point valence disadvantage around 35 percent of the time, and, symmetrically, choose a candidate with a two point valence advantage around 65 percent of the time. Low-importance voters, on the other hand, choose candidates with a two-point valence disadvantage only 20 percent of the time, and candidates with a two-point valence advantage 80 percent of the time. Thus, when candidates are in a proximity tie, a candidate s valence advantage can cause a massive 60 point swing in support among voters who see an issue as unimportant, and a more modest 30 point swing in the support of voters who see the issue as important. [Figure 5 about here] The lower graphs in each panel reflect the estimated consequence of proximity advantage, holding valence advantage at a tie. On both issues, these graphs conform weakly with our expectation that proximity will have a bigger impact on high-importance voters than on lowimportance voters. The substantive difference in effect of proximity on the choices of these two sets of voters is small, however. In sum, then, the importance of issues to voters conditions how they respond to candidate repositioning in general, and to the valence implications of repositioning in particular. 24

26 4.5 The Consequences for Dynamic Representation In this section we examine the implications of our findings for representation. We consider, for example, how candidates should respond when the views of voters they are courting changes which might happen, for example, because public opinion changes, because candidates run for higher office, or because candidates must triumph in a primary election before facing a general election. For each configuration of voter preferences, we analyze how candidates should behave in equilibrium this year, conditional on the position they espoused two years ago. As a first step in this analysis, we combine the abortion and tax data to consider both issues jointly. We do this because our goal in this section of the paper is to consider the general implications of repositioning, independent of the particular issue at stake or particular distribution of opinion on that issue. After combining the datasets, we divided voters into two groups by folding the issue scale. Thus, we grouped together voters who took a polar position on the issue they considered, and separated them from voters who took a moderate position. We then took the following five steps: 1) For both groups of voters we calculated the actual share of the vote each candidate received in all possible candidate strategy pairs. 2) We posited a particular distribution of opinion in the electorate. For example, that the electorate was composed of voters equally distributed across the space, with a third at each of the two poles and a third at the center. 3) We held fixed the positions candidates took last year, and calculated the vote share each candidate would receive for each possible combination of positions the two candidates could take this year. 4) We identified each candidate s best response to each possible strategy of the candidate s opponent. 5) Using this information, we determined the Nash equilibrium strategies of the two candidates, conditional on the positions they took two years ago and the distribution of opinion in the electorate. 25

27 Figure 6 displays, for all possible electorates, the equilibrium strategies we identified for candidates who started at the same position as each other either both at the same pole (top graphs) or both at the center (bottom graphs). The graphs on the left are based on the observed choices of voters who considered the issue they were assigned highly important, whereas the graphs on the right are based on the observed choices of voters who considered the issue to be of lower importance. [Figure 6 about here] Consider the graph in the upper left, which identifies the equilibrium strategies of candidates who stood at the same pole two years ago, and who now confront an electorate that sees the issue as highly important. For any point in the plotted triangle, the composition of the electorate is the percentage of voters at 1, as identified on the y-axis, the percentage of voters at 3, as identified on the x-axis, and the percentage of voters at the center, which is simply the remainder of the electorate. The graph identifies three distinct regions in which the candidates have different Nash equilibrium strategies, depending on the composition of the electorate. (Because, in this first example, the candidates took identical positions two years ago, they have identical equilibrium strategies this year. ) In the darkly shaded region at the top of the triangle, the best strategy for the two candidates is to remain at the pole they previously occupied. This equilibrium is indicated by the small pictograph in the region that has dots at the candidates starting positions. In this region, the candidates stick to their original pole even though the region clearly contains points that imply an electorate in which moderate and conservative voters are a majority, and the median voter is a moderate. For these hypothetical electorates, the candidates do not find it in their interest to converge to the median voter because they would pay a penalty for repositioning. Now consider the lightly shaded region in the middle of the triangle. If the electorate were in this region, the candidates would find it in their interest to shift to the center of the policy scale. This is indicated by the pictograph, in which the arrow runs from the candidates starting positions to their new positions. In this region, there are enough voters in the center and to the right that the spatial benefits of repositioning outweigh the valence costs. Finally, in the darkly shaded region at the right corner of the graph, both candidates find it in their interest to move all the way to the other pole to satisfy the large share of voters at 3. 26

28 It is useful to compare this graph to the adjacent, analogous graph for voters who do not regard the issue as highly important. The difference between the graphs is most easily observed by focusing on the dark region at the bottom right of each graph. This region is much smaller when voters place low importance on the issue because such voters more heavily punish candidates who reposition across the space. Figure 7 displays the remaining strategic scenarios, in which the candidates occupied different positions two year ago. As in Figure 6, there is a pure strategy equilibrium (summarized by a pictograph) for each shaded region. Unlike in Figure 6, however, Figure 7 contains unshaded regions where there are no pure strategy Nash equilibria. [Figure 7 about here] What are the implications for representation? To find out, we calculated the proportion possible electorates for which the winning candidate would occupy the position of the median voter, when the candidates are playing equilibrium strategies. Table 7 presents this summary statistic for each graph in Figures 6 and 7. [Table 7 about here] Winning candidates are more likely to represent the median voter when the electorate views the issue as important than when they regard the issue as relatively unimportant. For example, when the issue is of high importance and the candidates start in the same position, the winning candidate will, in equilibrium, take the position of the median voter for 88 percent of hypothetical electorates. In contrast, when the issue is less important, the winning candidate will serve the median voter in only 66 percent of hypothetical electorates. This pattern, which repeats for each of the four possible candidate starting positions, arises because high-importance voters place relatively less weight on valence considerations, and therefore afford candidates more slack to change positions without getting punished. 27

29 5. Conclusion To our knowledge, this paper represents the first experimental analysis of the proximity and valence effects of candidate repositioning. Our research confirms three hypotheses. First, voters discount the current policy pronouncements of candidates who have taken different positions in the past. Second, voters dislike inconsistency in itself, and therefore punish candidates for changing positions. Finally, voters who place high importance on an issue are less likely to punish candidates for flip-flopping. These findings have substantial implications for representation. By punishing politicians who adjust their issue stances, and by not accepting new candidate positions at face value, voters unwittingly dampen governmental responsiveness; they create disincentives for politicians to adjust when public preferences change or new policy-relevant information comes to light. Candidates will often prefer to stand firm, even when the median voter would prefer a different policy position. The political costs of being responsive exist in many situations, and are especially high when voters do not regard the issue as highly important to them personally. Consequently, candidates may find themselves locked into out-of-step positions on a wide range of issues of varying degrees of importance to voters. In future research, we plan to study how the attributes of candidates and the context of elections affect the consequences of repositioning. Candidates with party affiliations might evoke affective judgments from voters with partisan attachments, but such candidates might also have partisan policy reputations that could condition responses to policy. Moreover, candidates may influence the reactions of voters by attempting to excuse their own inconsistency (e.g., explaining to voters that circumstances have changed) or by criticizing the inconsistency of their opponents (e.g., claiming their opponent is not really committed to the new position). This paper offers initial answers to several fundamental questions about the politics of repositioning, while also providing an experimental template for future research about a topic at the heart of democratic politics. 28

30 References Alvarez, R. Michael Information and Elections. Ann Arbor: Univ. of Michigan Press. Allegeier, A. R., et al The Waffle Phenomenon: Negative Evaluations of Those Who Shift Attitudinally. Journal of Applied Social Psychology 9(2): Bartels, Larry M Issue Voting Under Uncertainty: An Empirical Test. American Journal of Political Science 30, 4: Berinsky, Adam J., and Jeffrey B. Lewis An Estimate of Risk Aversion in the U.S. Electorate. Quarterly Journal of Political Science 2: Bernhardt, Michael D. and Daniel E. Ingberman Candidate Reputations and the Incumbency Effect. Journal of Public Economics Brady, Henry E., and Steven Ansolabehere The Nature of Utility Functions in Mass Publics. American Political Science Review 83(1): Bullock, J., Green, D., and Ha, S Yes, But What's the Mechanism? (Don't Expect an Easy Answer). Journal of Personality and Social Psychology forthcoming. Carlson, James M., and Kathleen Dolan The Waffle Phenomenon and Candidates Image. Psychological Reports 57: Campbell, James E The Electoral Consequences of Issue Ambiguity: An Examination of the Presidential Candidates Issue Positions from 1968 to Political Behavior 5(3): Downs, Anthony An Economic Theory of Democracy. New York: Harper Collins. Erikson, Robert S., and David W. Romero Candidate Equilibrium and the Behavioral Model of the Vote. American Political Science Review 84: Enelow, James M, and Hinich, Melvin J The Spatial Theory of Voting. New York: Cambridge University Press. Enelow, James M., and Michael C. Munger The Elements of Candidate Reputation: The Effect of Record and Credibility on Optimal Spatial Location. Public Choice 77: Hoffman, Hillary S., and Charles S. Carver Political Waffling: Its Effects on the Evaluations of Observers. Journal of Applied Social Psychology 14(4): Imai, Kosuke, Dustin Tingley, and Teppei Yamamoto Experimental Identification of Causal Mechanisms. Typescript. 29

31 Kinder, Donald R., Mark D. Peters, Robert P. Abelson, and Susan T. Fiske Presidential Prototypes. Political Behavior 2(4): Markus, Gregory B Political Attitudes during an Election Year: A Report on the 1980 NES Panel Study. The American Political Science Review 76(3): Page, Benjamin I Choices and Echoes in Presidential Elections. Chicago: Univ. of Chicago Press. Sigelman, Lee, and Carol K. Sigelman Shattered Expectations: Public Responses to Outof-Character Presidential Actions. Political Behavior 8(3): Stokes, Donald E Spatial Models of Party Competition. American Political Science Review 57: Tavits, Margit Principle vs. Pragmatism: Policy Shifts and Political Competition. American Journal of Political Science 51(1): Tomz, Michael and Robert P. Van Houweling The Electoral Implications of Candidate Ambiguity. American Political Science Review 103, no. 1 (February):

32 FIGURE 1. Measuring Preferences, Expectations, and Trait Evaluations (a) Preferences about policy options (b) Preferences about candidates (c) Expectations about candidates (d) Evaluations of candidate traits