Polarization and the Decline of the American Floating Voter

Abstract

The observed rate of Americans voting for a different party across successive presidential elections has never been lower. This trend is largely explained by the clarity of party differences reducing indecision and ambivalence and increasing reliability in presidential voting. American National Election Studies (ANES) Times Series study data show that recent independent, less engaged voters perceive candidate differences as clearly as partisan, engaged voters of past elections and with declining rates of ambivalence, being undecided, and floating. Analysis of ANES inter‐election panel studies shows the decline in switching is present among nonvoters too, as pure independents are as reliable in their party support as strong partisans of prior eras. These findings show parties benefit from the behavioral response of all Americans to polarization. By providing an ideological anchor to candidate evaluations, polarization produces a reliable base of party support that is less responsive to short‐term forces.

Replication Materials

The data, code, and any additional materials required to replicate all analyses in this article are available on the American Journal of Political Science Dataverse within the Harvard Dataverse Network, at: https://doi.org/10.7910/DVN/4FCVUR.

How is it that American voters have allowed elite party polarization to flourish? Whether their opinions are polarized or not, the primary way that Americans reward or punish polarization is through voting. Yet current evidence provides an incomplete picture of how voting behavior responds to polarization. We know that elite polarization invigorates partisan attitudes (Hetherington 2001) and makes the issue opinions of engaged partisans more consistent with their party's stance (Druckman, Peterson, and Slothuus 2013; Layman and Carsey 2002; Levendusky 2010), such that partisanship's association with vote choice has strengthened (Bartels 2000). However, associations within elections cannot tell us whether voters change their opinions or partisanship across elections to reward or punish a party. An even greater concern is that these patterns fail to comment on the role of independent voters, who provide the strongest incentive for party moderation and are as large a component of the electorate (and arguably more decisive) as they were in the 1950s or 1960s.11 Despite declining turnout rates (Abramowitz 2010), the average percentage of presidential voters who are pure independents in the American National Election Studies (ANES) is slightly greater across recent presidential elections (7.3% from 2000 to 2012) than what it was from 1952 to 1968 (7.1%) since more Americans are independent.

To better understand the electoral consequences of polarization, we need to consider how its effects extend beyond engaged partisans. It is along these lines that I propose elite polarization partially sustains itself by turning independent and detached voters into loyal party supporters. In short, there are fewer floating voters. Scholars emphasize that floating voters provide valuable elements of flexibility and pragmatism within the electorate (Berelson, Lazarsfeld, and McPhee 1954; Zaller 2004), and that they waver in their party support because they are either unaware, indifferent, or conflicted (Converse 1962; Hillygus and Shields 2008; Kelley 1983; Lavine 2001; Lavine, Johnston, and Steenbergen 2012; Mayer 2008; Zaller 2004). Greater clarity of party differences reduces all of these attributes. It makes it easier for Americans to recognize the meaning and consequences of candidate differences, such that it reduces ambivalence and indecision in voting. This makes Americans less open to a change in their behavior and ultimately more reliable in which party they support across time. Moreover, these dynamics are especially consequential for those voters who otherwise lack partisan attachments and are most likely to switch their support, specifically independents and the less aware.

I proceed by first presenting evidence of the unprecedented decline of floating voters. I then discuss why this is likely in response to elite polarization and verify this link across multiple examinations. This analysis shows that, despite declining rates of partisanship, polarization has enabled independents and the politically inattentive to become like loyal partisans in their attitudes and behavior. The ANES Time Series studies estimate that recognition of party and candidate differences has become so pervasive that independent and less aware Americans of recent elections recognize party and candidate issue differences as clearly as Americans who were strong partisans or highly attentive in prior eras. Further analysis finds this growing recognition of differences predominately explains the shrinking percentage of Americans who are ambivalent or undecided in their choice for president. This combination of trends representing greater clarity of choice predominately explains the observed decline of floating voters as well. Finally, an examination of ANES panel studies validates that these patterns come from greater loyalty among all Americans, not the withdrawal from voting of those who typically switch their support, as it finds that pure independents were as stable in their party support across the 2000–04 presidential elections as were strong partisans across the 1956–60 or 1972–76 elections.

These findings have important implications for our understanding of elite polarization, campaigns, and voting behavior. They illustrate that parties likely benefit from polarization by gaining reliable supporters among even nonpartisans. Importantly, these gains need not be a function of polarization in voters' attitudes but simply their recognition of party differences. Furthermore, they show that elite polarization has promoted major changes among what we traditionally consider to be unattached or inattentive voters. These voters now appear more similar to past party stalwarts. They are as aware of candidate differences on the issues and as loyal in their presidential support in recent elections as the strong partisans or highly attentive Americans of prior eras. More broadly, the findings suggest greater voter loyalty has consequences for candidates' electoral incentives that can then lead to further polarization. Polarization limits the number of floating voters and the electoral payoff of appealing to their moderate or pragmatic concerns, thereby strengthening candidate incentives to court and mobilize only their likely supporters.

Floating Voters in the Face of Polarization

Popular control in a democratic system depends on the capacity of the electorate to remove a party from power. If stable voters for each party are relatively equal in number, then floating voters, those persons most likely to support different parties in successive elections, are largely responsible for bringing about democratic accountability. When sufficient in number, the fear of their movements “powerfully disciplines the actions of governments” (Key 1966, 10).

Scholars have emphasized various related markers of voters who are relatively pliable in their party support across circumstances by characterizing them as being either independent, less informed (Zaller 2004), cross‐pressured (Hillygus and Shields 2008), indifferent (Kelley 1983; Mayer 2008), or ambivalent (Lavine 2001; Lavine, Johnston, and Steenbergen 2012).22 Ticket splitting also relates to stability in party support and being cross‐pressured and ambivalent (Mulligan 2011). However, rates of ticket splitting are also contingent on incumbency, challenger quality, the southern realignment, and campaign competition (Burden and Kim ball 2002), factors that have less relevance to presidential elections.
These studies share a portrayal of floating voters as failing to perceive a clear home within existing partisan divisions, either because they are sporadically engaged in politics or because their values and attitudes are incongruent with major‐party alignments. They also commonly describe floating voters as being more pragmatic in comparison to partisan stalwarts. Floating voters are more likely to reward or punish the incumbent party for the country's economic performance or its successes and failures in foreign affairs (Key 1966; Lavine, Johnston, and Steenbergen 2012; Zaller 2004), and they are more responsive to the ideological positioning of candidates (Zaller 2004).

Prominent claims that politicians are fearful of and attentive to pragmatic swing voters stand in sharp contrast to conventional descriptions of today's polarized party elites catering to and mobilizing their existing supporters. But why is this? Although some scholars suggest floating voters have withdrawn from participating in politics (Prior 2007), an emerging body of public opinion research points to a more likely explanation. Party elites can ignore the moderating specter of floating voters because polarization has changed many of them into loyal supporters. Specifically, I propose that the clarity of elite polarization makes voters more attentive to the nature and consequences of party differences, which, in turn, reduces the indifference, uncertainty, or ambivalence in candidate evaluations that would otherwise make these voters less stable in their party support.

Recent studies document that these consequences exist for partisan opinions but have yet to consider their effects on stability in presidential voting. ANES surveys and experimental evidence find greater partisan constraint in mass public opinion when party cues are clearer and more salient (Layman and Carsey 2002; Levendusky 2010; Nicholson 2012). In a set of experiments, Druckman, Peterson, and Slothuus (2013) find that clearer differences between the two parties make partisans more confident in their party‐directed opinions and weaken the role that relevant substantive information might play in promoting countervailing consideration, which enables partisans to be less ambivalent and more stable in supporting their party's position.

Of even greater consequence, however, is that independent and detached voters likely react just like partisans. The findings among partisans are consistent with the more general claim that clearer cues enable voters to accept a relatively homogeneous pool of considerations that are consistent with their values when forming their candidate evaluations (Zaller 1992). Independents may fail to show greater ideological consistency in response to elite polarization (Layman and Carsey 2002), but their greater recognition of issue differences still enables them to identify the policy consequences of their vote more clearly. This information can stabilize party support by either reducing a voter's uncertainty over which party or candidate is the lesser of two perceived evils or strengthening a voter's resolve to equivocate between considerations that others perceive as conflicting (i.e., Alvarez and Brehm 2002).

Polarization can also convert voters who lack clear party attachments into stable supporters by clarifying each party's association and agreement with other social identities and group attachments. Polarization has strengthened cues from socioeconomic groups by making opinions within those groups more homogeneous (Garner and Palmer 2011). In addition, these affective cues from social groups can reduce instability among members who might otherwise be ambivalent (Rudolph 2005) or provoke changes in social behaviors without changing political beliefs (Mason 2015).

The strong association between elite polarization and greater loyalty in presidential voting is clear from looking at trends in survey measures of repeat voter behavior. ANES estimates indicate recent presidential elections exhibit the lowest levels of floating voters ever recorded. The ANES usually asks voters to recall whether they voted in the previous presidential election and which nominee they supported. Following Key (1966), I code respondents as being either standpatters (repeat voters with no change in major‐party support), floating voters (repeat voters with a change in major‐party support), surge‐and‐decliners (voted in only one election), or repeat nonvoters.33 Third‐party voters are excluded from the measure since variations in third‐party support mostly reflect the entry of third parties instead of tendencies to vote for a different party. Supporters of Perot in 1992 or Wallace in 1968 were previously unable to support their party. The results in Tables 4–6 address this limitation by imputing an estimate of major‐party switching among nonvoters and third‐party voters as well. They suggest most Perot voters would not have switched their votes if Perot had not run. Among the 1992 fresh cross‐section, the estimated floating voter rate is 20.4% when including Perot voters, 10.6% when excluding Perot voters, and 13.1% when imputing whether a Perot voter would have floated if Perot had not run.

Figure 1 plots the estimated percentage of Americans who are in each category. Polarization of the parties in Congress started its rapid growth during the 1980s (McCarty, Poole, and Rosenthal 2006), and there is a clear declining trend in floating voters since the 1980 election. Before then, the lowest percentage of vote switching was 9.1% in 1956, which reflected the uncommon circumstance of both parties nominating their previous nominee. In stark contrast, the percentage of floating voters has never exceeded 9% since 1996. Only 8.1% of Americans reported voting for a different party's nominee in 2008 than in 2004, despite that year's economic recession and reversal in party presidential fortunes. The year 2012 represented a new low in Americans' tendency to switch their party support, with only 5.2% of Americans voting for a different major‐party nominee. The average rate of vote switching over the past four elections (6.2%) is approximately half of the average rate across 1952 to 1980 (12.0%). Moreover, the decline in floating voters is most closely associated with an upward trend in standpatters, who reach their highest observed levels in 2012.

Figure 1

The Decline of Floating Voters

Note: ANES estimates using poststratification weights from the post‐election interview, when available. The 1984 ANES time‐series study did not ask for a respondent's prior presidential vote in 1980. Percentages calculated with third‐party candidate supporters excluded from the measure.

Considering the important role floating voters play in deciding elections and enhancing democratic accountability, it is curious that little attention has been paid to their rapid decline and its association with elite polarization. More Americans today may not identify with a party, but their behavior indicates we have never observed as many loyal supporters. ANES estimates of standpatters have never been higher, and estimates of floating voters have never been lower.44 The ANES provides the most comparable survey evidence over time, but changes in the design and implementation of the ANES through the years admittedly complicate any inferences. These complications include missing years (1984) and changes in the time and mode of interview, response rates, and weighting. Most notably, response rates have declined from 80% to 60%. Election‐to‐election changes in response rates, however, have a weak correlation with changes in floating voter rates (). I use multiple imputation in the analysis of the ANES panel studies to account for differences in panel attrition.

The observed timing of these contrasting trends in standpatters and floating voters points to elite polarization's likely role, but many other political or social factors may also be responsible for this pattern. In response, the following analyses validate polarization's role by documenting how America's growing recognition of party and candidate differences has ultimately erased many of the attributes that make voters, especially independents and the less aware, more open to a change in their party support.

The Now‐Aware Floating Voter

Scholars often describe floating voters as detached from politics because of either indifference (Mayer 2008) or inattentiveness (Zaller 2004). But the clarity of differences between the two polarized parties makes it easy for even Americans who ignore politics to recognize that there are meaningful differences between them. Indeed, this growing recognition of differences has been so extreme and universal that the inattentive and independent voters of today perceive party and candidate differences to be as meaningful and as prominent as the informed and partisan voters of the past saw them.

Figure 2 compares the percentage of respondents who claim there are important differences between the two parties by their strength of partisanship or their level of political awareness (as rated by interviewers). Americans in the lower half of political awareness have been more likely to claim there are important party differences during recent elections than were those in the upper half prior to 1980. A greater percentage of independents perceive important differences during recent elections than the percentage of weak or leaning partisans did prior to 1980.

Figure 2

Comparing Perceptions of Important Party Differences across Time by Attachment to Politics, 1960–2012 ANES

Note: Percentage among each category saying there are important differences between the two parties, calculated from the ANES time‐series studies. Respondents are considered low in awareness if interviewers rate them as having “average” or worse political knowledge, where high awareness represents “fairly high” and above.

The clarity of party differences that comes from polarization extends to voter perceptions of presidential candidate differences across multiple issues. Elite polarization has a strong positive association with the issue awareness of Americans and their capacity to recognize differences between the two parties on political issues (Layman and Carsey 2002). It is less recognized, however, that independents and the less aware of recent years have been as aware of candidate differences on the issues as the strong partisans and highly aware of earlier elections.

To demonstrate these changes, I generated an issue awareness score for each ANES respondent that measures the percent of issues on which he or she met the basic awareness criteria for issue voting: holding an issue opinion and seeing differences in the stances of the two party nominees on that issue. As shown in Section A of the supporting information (SI), awareness of issue differences rose from an average rate of 31.4% of issues in 1956 to 76.5% of issues in 2012, with a sharp upward trend following 1980.55 Scores prior to 1972 rely on perceptions of party differences since candidate perceptions were not measured. Analyses fail to indicate substantial differences in presidential nominee and party placements in those instances when the ANES asks respondents to place both. Since ANES surveys differ in the number and type of issues measured across the years, I restrict the score to the five most salient issues of each year, where the most salient issues were determined by the lowest percent of “don't knows” or no opinions. I include scores based on four issues in 1992, which only included four candidate issue placement items, or if a respondent refused to answer one issue item. Issue awareness trends within the consistently asked items, such as those concerning jobs and standard of living, also show positive swings. SI Section A plots national averages by year, and SI Section C provides a list of the variables used to create each year's score.

Further tests indicate this trend is strongly associated with polarization, even after accounting for the effects of a more educated public and other demographic changes. Figure 3 compares the estimated growth in issue awareness associated with greater elite polarization by strength of partisanship and political awareness. The specifications for these linear regression estimates follow Hetherington (2001) and account for a respondent's age, gender, race, and education. Additional details and coefficient estimates are presented in SI Section A.

Figure 3

Comparing Awareness of Issue Differences across Time by Attachment to Politics (ANES 1956–2012, Controlling for Other Demographic Changes)

Note: Issue awareness represents percent of issues for which the respondent holds an opinion and sees a difference between the two presidential candidates (or parties prior to 1972). Estimates are generated from bias‐corrected linear regression estimates from a bootstrap sample of election studies for a 40‐year‐old white male with a high school degree and, in the case of awareness, average partisan strength. Shading indicates 90% confidence intervals from the bootstrap sample of coefficients.

As the largely uniform trends in Figure 3 demonstrate, interaction tests fail to indicate polarization's effects are weaker among independents or the less aware.66 As in Figure 2, I code weak partisans and leaning independents to be equally strong in their identifications. Preliminary tests failed to show substantial differences between weak partisans and leaning independents for all models (with leaning independents slightly more partisan than weak partisans) and supported the collapsed measure's interval coding.
But what is most telling is that polarization's association is so strong that the independents and the less aware of recent elections appear to be as capable of voting on issue differences as the strong partisans and highly aware before polarization's rapid growth. Even when holding demographic factors constant, an independent in 2012 is estimated to recognize candidate differences across more issues than a strong partisan prior to 1984. Likewise, the estimates find that someone whom ANES interviewers rate as having “fairly low” knowledge of politics in 2012 perceives candidate differences across a relatively equal percentage of issues as someone ANES interviewers rate as having had “fairly high” knowledge prior to 1980.

Even if independents and those less interested in politics have greater opportunity to ignore politics in modern society, the growth of elite polarization has more than compensated. Consequently, independent and inattentive voters exhibit an awareness of candidate differences across more political issues than strong partisan or politically attentive Americans prior to 1980.77 Further examination verifies the changes in Figure 3 reflect Americans being more likely to recognize differences on issues on which they have opinions, and not changes in their ability to hold opinions on these issues (see also Abramson, Aldrich, and Rohde 2010, 158).
The people we consider to be the detached voters of recent elections recognize party and candidate policy differences across as many issues and perceive these differences to be as meaningful as did the attached voters of past elections.

How Clarity in Party Differences Enhances Clarity of Choice

Greater perception and salience of party and candidate differences are consequential because they enhance the capacity of Americans to recognize and understand the policy consequences of a party holding power (Hetherington 2001; Layman and Carsey 2002). As discussed, this clarity likely reduces the tendency of voters to be ambivalent or undecided in their candidate support. By seeing the consequences of their vote choice more clearly, voters are better able to hold a set of candidate evaluations that are consistent and homogeneous with their values and identifications, making them less ambivalent or indifferent and less likely to waver in their support.

Evidence from the ANES time‐series studies confirms these expectations. There has been a steady decline in both the rate of Americans who are ambivalent in their candidate evaluations and the rate of Americans who are undecided in their candidate preference since 1980 (see Figure B.1 in the SI). The last two ANES surveys that measure ambivalence from number of likes and dislikes, 2000 and 2004, had the smallest percentage of ambivalent Americans ever observed in the series (18.4 and 20.6, respectively).88 Following Basinger and Lavine (2005) and Thompson, Zanna, and Griffn (1995), let D indicate the average number of likes the respondent provided for the Democratic presidential nominee and dislikes of the Republican nominee, with R indicating the same for the Republican nominee. A respondent's ambivalence is then measured as . As discussed by Rudolph (2005) and Thornton (2011), I treat ambivalence as a nominal trait and identify a respondent as ambivalent if he or she has a score greater than 0. The ANES has yet to release counts and codes of party and candidate comments for 2008 or 2012.
Only once prior to 1984 was the percentage of Americans with ambivalent candidate evaluations observed below 24%. Since 1984, the national average of ambivalence has exceeded that rate only once, in 1992. Each study prior to 1996 had more than 10% of Americans claiming they were undecided, with a high of 17.1% in 1976. The percentages since 1996, however, have been in the single digits, with an average rate of 7.0%.

To evaluate whether greater clarity in differences is responsible for this trend, I combine a respondent's perception of important party differences and issue awareness score into a summated scale that represents his or her recognition of party and candidate differences. Importantly, this scale measures both facets that counter indifference or detachment from party politics: the degree to which a respondent is aware of differences and considers them to be meaningful.99 The combination of party and candidate perceptions is necessary since respondents are not asked weather presidential candidate differences are important. But the similarity of the party and candidate placement measures suggests respondents view parties and their nominees as similar entities during presidential elections. Tests that include the awareness and importance components separately justify the additive scale. The 1956 ANES did not ask respondents whether they saw important differences between the two parties. I use a respondent's issue awareness level and other attributes to randomly impute a value based on a logit model estimate of the relationship from other years.
I then test the relationship between a respondent's recognition of party and candidate differences and the probability of being ambivalent or undecided in his or her candidate evaluations. The model also accounts for changes associated with education, strength of partisanship, and other demographic factors. Since early studies do not include interviewer ratings of a respondent's political awareness, I follow Lavine (2001) and use the respondent's total number of likes and dislikes made in discussing the parties and candidates as an indicator of awareness and cognitive ability.1010 Interviewer ratings of awareness have a .55 correlation with total number of comments; it only has a .39 correlation with education. I limit the highest value of total comments to 20, since such values are only possible if one does not have univalent candidate and party evaluations.
In addition to the individual‐level variables, I account for the potential election‐level effects of economic conditions, with the percentage point change in America's per capita real disposable personal income from the previous year, and reelection campaigns.

I utilize a bivariate probit model to test these relationships. Bivariate probit estimates are more efficient and especially useful in this case because being undecided and being ambivalent are positively correlated with each other, but many characteristics that are associated with being ambivalent differ from those associated with being undecided. The bivariate probit model estimates the extent to which any unmodeled dependencies drive voters to be both ambivalent and undecided. These estimates, and all others from the combined time‐series studies, account for the dependency of observations within each election by drawing 2,000 bootstrap samples of the election studies to generate bias‐corrected coefficient and standard error estimates.1111 The ANES Time Series studies are multilevel data in that they are a sample of individuals within 12–15 elections. With so few years, however, asymptotic assumptions of maximum likelihood estimates are biased, primarily from standard error estimates that are too small (Stegmueller 2013). As discussed by Hox (2010, 236), it is best to either use a stronger set of Bayesian priors on the variance components to inform these estimates or, as I do, use a bootstrap sample of the elections to provide a nonparametric estimate of how much these associations vary depending on the sample of elections included. An additional advantage in using bootstrap estimates is that they provide bias‐corrected estimates that account for outlying elections having possible undue influence. SI Section B provides further discussion and demonstrates that in most instances the bootstrap estimates match the alternative asymptotic likelihood coefficient estimates but with slightly larger standard errors.

Table 1 presents the bivariate probit estimates of the relationship between recognition of party and candidate differences and the probability of being ambivalent or being undecided in one's candidate evaluations. The lack of candidate ambivalence measures in 2008 and 2012 restricts the analysis from 1956 to 2004. The left column under each dependent variable presents the model's parameter estimates and significance tests. The right column under each dependent variable presents the model estimate of each variable's across‐era effect, which I operationalize as the change in the sample average predicted probability when changing each variable from its average observed during the prepolarized era (pre‐1980) to the average recently observed during the polarized era (2000 and beyond). Each calculation is made while keeping all other variables at their observed value (Hanmer and Kalkan 2013). By combining the model parameter estimates with the observed changes in each variable across time, the estimates of the across‐era effects identify what changes are most associated with the observed declines in ambivalence and being undecided.

Table 1. Explaining the Presence of Ambivalence and Indecision about Presidential Candidates, 1956–2004

	Ambivalent		Undecided
Variable	Parameter	Across‐Era Effect	Parameter	Across‐Era Effect
Recognition of Differences Scale	−0.277*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−1.54*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−0.436*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−1.42*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.
	(.049)	(.26)	(.045)	(.15)
Number of Likes/Dislikes	0.078*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	0.04*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−0.033*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−0.01*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.
	(.004)	(.00)	(.005)	(.00)
Education	0.102*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	1.93*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	0.030	0.33
	(.015)	(.29)	(.022)	(.25)
Strength of Partisanship	−0.296*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−0.05*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−0.529*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−0.05*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.
	(.028)	(.00)	(.032)	(.00)
Age	−0.005*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−0.13*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	0.005*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	0.08*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.
	(.001)	(.02)	(.001)	(.02)
Female	−0.042*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	0.02*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	0.081*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−0.02*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.
	(.022)	(.01)	(.025)	(.01)
African American	−0.349*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−0.63*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	0.046	0.05
	(.057)	(.11)	(.069)	(.07)
Change in per Capita Income	−0.045*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	0.99*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.	−0.026	0.34
	(.023)	(.51)	(.024)	(.30)
Reelection Campaign	−0.007	−0.01	−0.139	−0.11
	(.077)	(.10)	(.093)	(.07)
Constant	−0.579*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.		−0.433*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.
	(.099)		(.134)
aROC	0.718		0.734
	0.327*Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.
Log likelihood	−15942.54
Elections	13
n	19,639

• Note. Bias‐corrected bivariate probit estimates from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Across‐era effects represent average percentage point change in respondents' predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's pre‐election interview weight (if available). , two‐tailed test.

The estimates presented in the top row of Table 1 show that greater recognition of differences is significantly associated with respondents who are less likely to be ambivalent and undecided. The consistency of the negative estimate across dependent variables is important because it shows greater recognition of party and candidates differences has effects that differ from what we would expect if it were simply a measure of political awareness or education. Americans with more education and who are politically aware are significantly more likely to be ambivalent.1212 This is a common result when accounting for strength of partisanship; see Keele and Wolak (2008) for further discussion and evidence of a positive relationship between awareness, education, and presidential candidate ambivalence.
Consequently, the across‐era comparison estimates that the growth of formal education among Americans would have produced a rise in ambivalence rates of 1.93 percentage points, when holding other factors constant.

We have not seen a rise in ambivalence, however, because Americans' concurrent greater recognition of differences has had the opposite effect. Indeed, the consistent negative association between a person's recognition of differences and being ambivalent or undecided is most similar to that of strength of partisanship. But strength of partisanship shows only slight gains over recent decades and is only slightly responsible for the across‐era decline. Americans' growing recognition of differences, however, has effects that are very similar to what we would expect if Americans had much stronger party identities. This massive shift respectively explains 1.5 and 1.4 percentage point declines in being ambivalent and undecided across eras, estimates that are over twice as large as any other factor with significant negative across‐era effects.1313 Figure B.2 in the SI demonstrates the strong negative aggregate association between rates of ambivalence and being undecided and the average recognition of differences score.
Americans are also less likely to be ambivalent and undecided when America's personal income is growing at a high rate and during reelection campaigns, but only the relationship between ambivalence and change in per capita income is significant at the .05 level, where the relatively low rates of income growth in recent elections suggests there should have been higher rates of ambivalence. In summary, these estimates illustrate that a growing recognition of candidate and party differences predominately explains the decline in national rates of candidate ambivalence and being undecided.

Further examination of these results also suggests that these changes have had larger consequences for nonleaning independents. When independents were completely ignorant of party differences, as they were prior to polarization, they were much more likely to be undecided or ambivalent. For instance, 64% of independents in 1956 failed to see a difference between the two parties on at least one of the five most salient issues, compared to 21% of strong partisans. Now most independents recognize some issue differences, as a similar calculation in the 2012 ANES estimates that only 14% of independents (and 3% of strong partisans) failed to recognize a difference between Obama and Romney on any issue.

Table 2 compares the estimated consequence of the observed growth in recognition among independents and strong partisans, by setting each profile to its observed average for income, age, change in per capita income, and number of comments and the most common category of gender, race, and reelection since the 2000 ANES. When an independent changes her rate of recognizing differences from the 1956 average for independents to the 2012 average, she is 6.7 percentage points more likely to be neither ambivalent nor undecided and almost 4 percentage points less likely to be both undecided and ambivalent.1414 My use of the average score in 2012 is not an out‐of‐sample prediction since there are plenty of independents and strong partisans with recognition of differences scores of either 0 or 100 from 1956 to 2004.
In contrast, her estimates as a strong partisan are significantly smaller because her existing level of recognition already made her unlikely to be either undecided or ambivalent. Comparatively larger effects exist for those less aware as well. Observed rates of being undecided reached their peak in 1976 and then bottomed in recent elections. When holding the economy and reelection campaigns constant, the average predicted probability of being undecided drops by 3.3 percentage points (16.8 to 13.5) among those interviews rated in the bottom half of the political awareness scale, but only by 1.4 percentage points for those rated in the top half (9.5 to 8.1).

Table 2. The Effects of Greater Clarity of Differences by Strength of Partisanship

	Change in Predicted
	Probability of Being
	Neither Undecided	Both Undecided
1956 to 2012 Change in Rec. of Differences Scale	nor Ambivalent	and Ambivalent
Independents	6.68*Note. Percentage point change in predicted probability as recognition of candidate and party differences moves from its 1956 average to its 2012 average for that profile. Standard errors in parentheses. Both are calculated using the bootstrap sample of coefficients. Individual profiles of the remaining variables are set to the mean or modal value observed since 2000 within each profile. , two‐tailed test.	−3.78*Note. Percentage point change in predicted probability as recognition of candidate and party differences moves from its 1956 average to its 2012 average for that profile. Standard errors in parentheses. Both are calculated using the bootstrap sample of coefficients. Individual profiles of the remaining variables are set to the mean or modal value observed since 2000 within each profile. , two‐tailed test.
	(0.73)	(0.53)
Strong Partisans	3.37*Note. Percentage point change in predicted probability as recognition of candidate and party differences moves from its 1956 average to its 2012 average for that profile. Standard errors in parentheses. Both are calculated using the bootstrap sample of coefficients. Individual profiles of the remaining variables are set to the mean or modal value observed since 2000 within each profile. , two‐tailed test.	−0.69*Note. Percentage point change in predicted probability as recognition of candidate and party differences moves from its 1956 average to its 2012 average for that profile. Standard errors in parentheses. Both are calculated using the bootstrap sample of coefficients. Individual profiles of the remaining variables are set to the mean or modal value observed since 2000 within each profile. , two‐tailed test.
	(0.49)	(0.16)
Difference	3.30*Note. Percentage point change in predicted probability as recognition of candidate and party differences moves from its 1956 average to its 2012 average for that profile. Standard errors in parentheses. Both are calculated using the bootstrap sample of coefficients. Individual profiles of the remaining variables are set to the mean or modal value observed since 2000 within each profile. , two‐tailed test.	−3.09*Note. Percentage point change in predicted probability as recognition of candidate and party differences moves from its 1956 average to its 2012 average for that profile. Standard errors in parentheses. Both are calculated using the bootstrap sample of coefficients. Individual profiles of the remaining variables are set to the mean or modal value observed since 2000 within each profile. , two‐tailed test.
	(0.40)	(0.42)

• Note. Percentage point change in predicted probability as recognition of candidate and party differences moves from its 1956 average to its 2012 average for that profile. Standard errors in parentheses. Both are calculated using the bootstrap sample of coefficients. Individual profiles of the remaining variables are set to the mean or modal value observed since 2000 within each profile. , two‐tailed test.

The Consequences of Clarity for Stability in Party Support

The results in Tables 1 and 2 are consistent with elite polarization making party and candidate differences clear to Americans and providing them with greater clarity of mind in their vote choice. Americans are less ambivalent and more decisive in their presidential support, and these changes are relatively more prevalent among independents and the less aware. These dynamics among detached voters are especially important because, as Table 1 also shows, the consequences of greater recognition of party and candidate differences for being ambivalent and undecided are much more similar to the effects of having stronger partisan identities as opposed to the effects of simply more education or being more attentive. In short, it enables detached voters to evaluate candidates like partisan voters even if they do not hold those identities.

If this is the case, then the consequences of clarity of choice should importantly extend to the stability of behavior across elections as well. With fewer ambivalent or undecided voters, there should be declining rates of their more open or pragmatic behavior since these voters are typically much more responsive to the effects of short‐term forces like economic performance evaluations or candidate positioning (Key 1966; Lavine, Johnston, and Steen‐ bergen 2012; Zaller 2004). Along these lines, polarization has made Americans less open to a change in their party preference and more reliable in which party they support across elections by enabling them to approach recent elections with much greater clarity of choice.

Table 3 presents multinomial logit estimates that test which factors are most responsible for the observed decline in floating voters, using as the dependent variable the same four‐part categorization of voters as in Figure 1.1515 Mixed logit estimates of the same model with correlated errors of different scales across choices failed to indicate the multinomial logit specification provided a significant loss of fit, supporting its IIA assumption.
I test for the effects that clarity of choice has on the inter‐election voting behavior of Americans by including the three associated variables: the index of a respondent's recognition of party and candidate differences and two dummy variables indicating whether someone was undecided or ambivalent in the pre‐election interview. The other independent variables follow from the models in Table 1, but I also include each person's level of external efficacy since its general decline might be associated with the abstention of those who might otherwise be likely floating voters.

Table 3. Explaining the Decline of Floating Voters, 1956–2004

		Across‐Era Effect
Variable	MNL Estimate Standpatter— Floating	Single	Combined
Clarity of Choice
Recognition of Differences Scale	−0.440*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.	−0.45*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.
	(.150)	(.19)
Ambivalent	0.516*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.	−0.26*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.	−0.95*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.
	(.095)	(.04)	(.20)
Undecided	0.583*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.	−0.22*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.
	(.117)	(.04)
General Political Orientation
External Efficacy	−0.054	−0.42*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.
	(.101)	(.16)
Number of Likes/Dislikes	−0.005	0.00*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.	−0.41*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.
	(.008)	(0.00)	(.17)
Strength of Partisanship	−0.737*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.	−0.03*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.
	(.048)	(.00)
Demographic Profile
Age	−0.008*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.	0.08*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.
	(.002)	(.02)
Female	0.077	−0.01
	(.049)	(.01)	0.14
African American	−0.372	−0.27*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.	(.24)
	(.225)	(.11)
Education	−0.163*Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.	0.24
	(.042)	(.16)
Election‐Level Factors
Change in per Capita Income	−0.014	0.14
	(.053)	(.32)	0.08
Reelection Campaign	−0.135	−0.06	(.31)
	(.183)	(.06)
Constant	0.118
	(.264)
Log‐likelihood	−14738.00
Elections	12
n	14,424

• Note. Bias‐corrected maximum likelihood parameter estimates of a multinomial logit from a bootstrap sample of election studies. Bootstrapped standard errors in parentheses. Standpatters are baseline comparison group; coefficient estimates for repeat nonvoters and surge and decliners are presented in SI Section D. Across‐era effects represent average percentage point change in the respondent's predicted probability when changing the variable's value from its pre‐1980 mean to its observed mean since 2000 and holding all other variables at their observed value. Both estimates weighted by each year's post‐election interview weight (if available). , two‐tailed test.

The first column of results in Table 3 presents the coefficient estimates for the floating voter category (with standpatters as the comparison group). SI Section D presents coefficient estimates and predicted probability estimates for all categories. The last two columns of Table 3 present the across‐era effect for changes in each independent variable and the combined effect associated with each broader class of changes. Once again, the across‐era effect is the average change in the predicted probability when changing each variable or set of variables from the prepolarized era average (pre‐1980) to the polarized era average (2000 and beyond), holding all other variables at their observed values.

The top three variables account for polarization's effects on voter loyalty through greater clarity of choice. Each variable has a significant and relatively strong association with the probability of being a floating voter. Individuals are significantly less likely to float to another party when they have greater recognition of candidate differences and are neither ambivalent nor undecided during the current election. Across the independent variables, greater recognition of party and candidate differences has the largest estimated single effect on the decline of floating voters. Moreover, the trends in these variables combine to produce the largest joint effect estimate.

Among other factors, only changes in the political orientations of Americans come close to making a substantial contribution. Some of the decline in floating voters is associated with Americans becoming less efficacious and having slightly stronger partisan identifications. Only efficacy, however, has made a substantial contribution to the decline, as it has lowered turnout among potential floaters.1616 The recent growth of strength of partisanship also fails to show stronger effects when restricting the analysis to post‐1980 studies, as clarity of choice factors continue to have the largest effects.
Many demographic changes have produced significant changes in floating across eras. Americans who are African American are significantly less likely to be floating voters, but being older and more educated increases the probability of being a floating voter since it has a very strong effect on reducing rates of abstention. Consequently, demographic changes fail to show a significant effect in combination since the turnout effects of an older and more educated electorate negate the effects of a diverse electorate. Finally, when holding these individual‐level attitudes constant, none of the election‐level covariates have significant associations with rates of floating voters.1717 Poor income growth can indirectly lead to more floating through its effects on ambivalence (Table 1). Measures of extremity of economic conditions (very good or very poor income growth) failed to show significant associations with rates of floating.

It is not surprising that Americans now perceive that differences between the parties are more meaningful and exist across more issues in response to polarization, but this natural response has pervasive consequences for the behavior of American voters across elections. It enables Americans to be less ambivalent and undecided in their candidate evaluations, and this clarity of choice has made Americans less open to changing their behavior in response to short‐term forces as they approach presidential elections.1818 For example, SI Section D shows that poor economic growth is no longer positively associated with rates of voters switching to punish incumbents.
Consequently, we find that Americans are much less likely to vote for a different party's nominee across successive presidential elections.

Evidence from the ANES Panel Studies

There are some limitations to the preceding analysis of floating voters. First, the identification of floating voters relies on respondents' recollections of their behavior four years prior. But, as Converse (1962) notes, recalling one's vote from four years prior varies in accuracy by factors associated with switching, such as political sophistication, thus confounding comparisons of floating and nonfloating voters. More generally, this measure of floating combines both voter participation and preference switching, as ANES Time Series studies fail to ascertain the candidate preference of those who did not vote in the previous elections. Since I have not accounted for the stability of major‐party support among habitual nonvoters, it is possible that Americans who typically switch their party support across elections have simply withdrawn from voting (Prior 2007) instead of polarization having reduced the propensity for major‐party switching among all Americans.

I address these concerns by comparing estimates from the four inter‐election ANES panel studies (i.e., 1956–60, 1972–76, 1992–96, and 2000–04). These surveys usually interview respondents before and after two successive presidential elections, and they typically measure the preferred candidate for nonvoters as well. This allows one to estimate the percentage of party switchers, individuals who support a different party's nominee across successive elections, who become floating voters by turning out to vote across both elections.

These four studies clarify the contribution of repeat turnout and preference switching to floating's decline but potentially overstate trends. The first two studies start with a landslide election that might have promoted higher levels of switching. The last two studies have a repeat nominee for one party, which also limits comparability with the first two studies. However, there is no evidence to suggest a significant reelection bias. Tables 1 and 3 fail to show reelections having a significant or strong influence on either clarity of candidate evaluations or the probability of floating. Compared to other elections, there is an observed tendency for floating voter rates to be 1.5–2 percentage points higher in elections without a candidate who ran previously, but this difference is insignificant.1919 About 8.65% of Americans are floating voters when presidents run a second time, compared to 10.69% otherwise, and 10.07% if 1968 is excluded, when Johnson opted against running out of fear of losing.
Many reelections also fail to exhibit greater stability, as both 1992 and 2004 were reelection campaigns with higher rates of floating voters compared to the prior election.

Not all respondents interviewed in the first wave of the ANES panel studies participated in the next election's wave, especially within the recent two studies.2020 Observation rates for both voting turnout measures are (in chronological order) 91%, 92%, 54%, and 46%.
Panel attrition lessens the representativeness of the sample, which can be problematic if respondents who fail to complete the survey are those less likely to vote (Bartels 1999) or more likely to switch. To remedy attrition bias, I use multiple imputation to simulate a distribution of possible values for missing observations (King et al. 2001). Along with the independent variables, I include a set of auxiliary variables (retrospective evaluations, thermometer scores, survey compliance, and other demographic measures) to impute possible values of missing observations. By using multiple imputation, I can also estimate rates of major‐party switching among third‐party supporters from their assessments of the major‐party candidates across the two studies. I present estimates summarizing across 10 imputed data sets for the first two panels and 20 imputed data sets for the recent two panels with higher attrition rates.2121 I exclude respondents not interviewed in the original wave, since these interviews were added to correct for attrition biases within the original sample. I use the bootstrapping‐based EM algorithm within R's version of Amelia II (Honaker, King, and Blackwell 2008) to impute values of each variable, except candidate preference. I rely on an additional set of transformed (i.e., passive) variables to sharpen the imputation of candidate preference within a wave‐specific logit model. SI Section E provides a list of the variables and specifications used in estimation.

As shown in the top entry of Table 4, the estimated proportion of Americans who switch their party support is much smaller across recent elections. The proportion of switchers across the 1956–60 and 1972–76 elections is larger than the rate across 1992–96 and approximately twice the rate across 2000–04. The next two rows show how the decline in party switching is much sharper than the decline in repeat turnout among each group. Switchers are less likely to be voters than nonswitchers, and they did show a decline in repeat turnout rates across 1992–96. But their repeat turnout levels rose across 2000–04, to the point it is not significantly different from the turnout rates of the first two studies. There is also no evidence of party switchers lowering their turnout rate relative to nonswitchers, as their rate is 15.0 percentage points lower for 1956–60 but only 9.1 percentage points lower for 2000–04.

Table 4. Estimates of Switching, Repeat Turnout, and Floating by ANES Panel Study

	Study
Percentage	1956–60	1972–76	1992–96	2000–04
Switching Parties	26.7	29.7	22.5	13.5
	(1.3)	(1.4)	(2.1)	(1.1)
Repeat Turnout
Switchers	60.8	61.5	49.6	56.9
	(2.8)	(2.6)	(4.9)	(4.7)
Nonswitchers	75.8	72.4	65.2	66.0
	(1.5)	(1.5)	(2.3)	(1.6)
Floating Voters	16.2	18.3	11.2	7.7
	(1.1)	(1.1)	(1.3)	(0.8)
n	1,239	1,320	1,005	1,807

• Note. Sample average rate of switches in preferred major‐party presidential candidate, repeat turnout, and being a floating voter with standard errors in parentheses (1992–96 and 2000–04 estimates are weighted) using multiple imputation to correct for panel attrition.

The bottom entry presents the percentage of Americans who were floating voters. These are switchers who turned out to vote in both elections. The proportion of Americans who were floating voters from 1992 to 1996 was 5 to 7 percentage points lower relative to the previous two observations. This is a product of the decline in switchers and the decline in repeat turnout among switchers. Across 2000–04, however, repeat turnout rates rose but rates of switching continued to decline, such that only 7.7% of Americans voted for a different party's nominee. Compared to the first two studies, this is a significant decline in floating voter rates of approximately 10 percentage points.

Moreover, further examination indicates the decline of floating voters is largely a product of independents and the less aware being much more stable in their party preferences across presidential elections. Table 5 compares rates of switching, repeat turnout, and being floating voters by strength of partisanship. The most dramatic change is the large decline in average rates of switching, not repeat turnout. Independents remain more likely to switch than partisans, but the differences by strength of partisanship become much smaller. From 1956–60 to 2000–04, the decline in switching rates among leaning/weak partisans was about 13 percentage points and only 10 percentage points for strong partisans, but pure independents show a much larger drop of 25 percentage points. After having gone through a recession and a war, pure independents were more stable in their party support across 2000–04 than strong partisans were across 1972–76 and about as stable as strong partisans across 1956–60. In combination, the percentage of independents or leaning independents and weak partisans who were floating voters across 2000–04 is lower than the percentage of strong partisans who were floating voters across either 1956–60 or 1972–76.

Table 5. Sample Rates of Switching, Repeat Turnout, and Floating by Strength of Partisanship

		Respondent's Average Strength of Partisanship
Percentage	Study	Independent	Leaning/Weak Partisan	Strong Partisan
Switching Parties	1956–60	45.2	28.7	19.4
	1972–76	34.9	32.8	23.3
	1992–96	35.9	25.1	13.6
	2000–04	19.7	16.1	9.9
Repeat Turnout	1956–60	55.4	73.6	75.1
	1972–76	50.3	70.3	77.0
	1992–96	42.6	61.3	71.5
	2000–04	44.3	64.4	70.8
Floating Voters	1956–60	24.0	18.9	11.6
	1972–76	16.6	20.6	16.3
	1992–96	14.3	13.3	8.0
	2000–04	9.4	9.1	6.1

• Note. Average rates of major‐party switching, repeat turnout, and floating (1992–96 and 2000–04 estimates are weighted) using multiple imputation to correct for panel attrition.

Similarly, Table 6 uses multiple measures to show the largest declines in switching rates are among those least attentive to politics. The top panel compares Americans by the interviewers' rating of their knowledge of public affairs during the first wave. Respondents interviewers rated as “very low” or “low” in their level of political knowledge exhibit the greatest decline in rates of switching. Their average rate of switching declines by 19.6 percentage points from 1956–60 to 2000–04, compared to only 8.8 percentage points among those interviewers rated “very high.” However, the rate of repeat turnout across 2000–04 for those with low ratings is not significantly lower than the rate across 1972–76.

Table 6. Sample Rates of Switching, Repeat Turnout, and Floating by Measures of Political Awareness

		Interviewer's Rating of Respondent's Political Knowledge
Percentage	Study	Very Low–Low	Average	Fairly High	Very High
Switching Parties	1972–76	36.0	30.1	28.5	19.0
	1992–96	31.5	23.6	17.7	15.6
	2000–04	16.4	14.5	10.8	10.2
Repeat Turnout	1972–76	43.9	69.1	81.2	85.9
	1992–96	37.6	60.4	74.0	78.4
	2000–04	39.0	66.3	80.3	85.9
Floating Voters	1972–76	15.1	18.2	21.6	15.0
	1992–96	10.0	12.8	11.4	9.0
	2000–04	5.6	8.7	7.9	8.8
		Average Total Mentions (t)
Percentage	Study
Switching Parties	1956–60	36.1	32.0	24.3	13.8
	1972–76	36.1	31.1	28.7	23.5
	1992–96	31.5	26.2	19.3	14.0
	2000–04	16.7	14.5	10.7	9.5
Repeat Turnout	1956–60	48.1	72.8	80.3	87.3
	1972–76	51.6	64.0	74.8	85.4
	1992–96	40.4	55.6	68.7	80.7
	2000–04	47.5	69.6	72.6	83.6
Floating Voters	1956–60	14.5	22.1	16.4	11.9
	1972–76	18.7	17.0	18.3	19.1
	1992–96	10.1	13.5	12.4	9.7
	2000–04	7.5	9.6	7.6	6.8

• Note. Average rates of major‐party switching, repeat turnout, and floating (1992–96 and 2000–04 estimates are weighted) using multiple imputation to correct for panel attrition.

The bottom panel compares respondents across categories marking quartiles of number of mentions.2222 In comparing these four studies, the bottom quartile of average total mentions was consistently five or below, with the top quartile showing mention rates above 11.
When comparing rates of switching between 1956–60 and 2000–04, the top quartile shows only a small decline of 6.8 percentage points, whereas declines of 23.0, 14.1, and 12.8 percentage points are respectively estimated for the lowest, second, and third quartiles. However, outside of 1992–96, rates of repeat turnout are very much consistent for those in the lower two categories. Both of these comparisons suggest floating voter rates have largely declined because rates of switching have declined, with repeat turnout rates remaining steady.

In combination, a consistent picture of the decline of the American floating voter emerges. All Americans, not just partisans or the politically engaged, have become less likely to switch their vote across presidential elections. This is primarily a function of greater stability in party preferences as opposed to switchers exiting politics. The panel evidence also documents that there have been dramatic changes among Americans whom scholars typically consider to be unattached to parties or inattentive to politics. Pure independents and the less aware of 2000–04 were more likely to switch than the strong partisans or the highly aware of those elections, but they were also as loyal as were strong partisans and the highly aware prior to polarization. Moreover, when factoring in turnout, independents and weak/leaning partisans are essentially equal in their propensity to be floating voters, and, according to interviewer ratings of political awareness, the most aware Americans were most likely to vote for different party nominees across the 2000 and 2004 elections.

Discussion and Conclusion

Berelson, Lazarsfeld, and McPhee (1954) suggest that our electoral system requires behaviors that are so incompatible within each individual that we can only expect to attain them through a heterogeneous electorate. From this perspective, even the “least admirable” voters (316) enrich our democratic political system by providing it with the flexibility and indifference that it needs. The pragmatic tendency of these voters enables and motivates party compromises, and their willingness to change their party support enhances accountability within democratic elections.

This analysis finds, however, that this diversifying and flexible group of voters has been on the decline, as there have been major changes in the prominence and characteristics of floating voters over the past 30 years. Despite a declining percentage of Americans claiming a party identification, Americans now exhibit the highest observed rates of party allegiance when voting across successive presidential elections. This decline in floating voter rates is primarily a result of all Americans being more stable in their party support, not switchers abstaining from voting. By making it easy for Americans to recognize party differences, polarization has reduced ambivalence and indecisiveness and provided a strong and consistent ideological anchor to Americans' presidential preferences across time, even for independents and the less aware. Although independents remain slightly less stable in their party support across recent presidential elections, it is difficult to consider them similar to the independents of the past since they are as aware of party differences and as loyal in their party support as strong partisans were prior to polarization. Polarization has not strengthened their sense of partisan loyalty, but the clarity of polarization has effectively allowed independents and the politically inattentive to act as loyal partisans in their behavior.

More generally, these findings document that the clarity of elite polarization changes American voting behavior and election outcomes. Americans' party support in presidential voting has strengthened in its durability and pervasiveness, as stable party divisions have extended to independents and the politically inattentive. This growth in loyalty lessens voter responsiveness to short‐term forces and means that the party elites of today have much less reason to fear losing their supporters to the other party or need to appeal to the pragmatic concerns of the remaining few floating voters, as long as they are the only two parties. These findings only examine polarization's effects on choices, not opinions, but they remain informative nonetheless since they indicate that a polarization in choices is sufficient for a growth in loyalty. Whether or not they are polarized in their issue attitudes, Americans show little conflict or concern in picking a side and sticking with that choice. Consequently, even small divisions in American public opinion can have pervasive consequences for voting behavior if they clearly align with elite party differences.

Since these changes largely follow trends of elite polarization, where the mass public's issue awareness is especially higher on moral issues where party divisions used to be nonexistent (Fiorina, Abrams, and Pope 2006), the decline in switching appears to be in response to polarized parties. But to claim it is simply caused by elites is an oversimplification. The mass public's response to party positioning is not immediate but takes time. In terms of abortion, voters of 1980 were largely unaware and unresponsive to the differences of Reagan and Carter (Granberg and Burlison 1983). Recognition of party differences only emerged in 1984, once issue activists gained footing within each party (Carmines and Woods 2002), and its influence on partisan opinions took much longer (Fiorina, Abrams, and Pope 2008).

A clearer conclusion is that the response of voters and elites to each other's behavior has helped to facilitate this trend. By being less conflicted and more stable in their support of party nominees, the response of Americans to elite polarization changes the electoral incentives of party candidates. When party preferences are more predictable and less responsive to short‐term forces, candidates have little need to appeal to the moderate or pragmatic concerns of the few remaining floating voters. For those states where one party has an advantage, greater stability enhances the dominant party's certainty of winning and reduces its need to appeal to moderates or floating voters. But even if stalwart supporters are evenly matched, and floating voters are potentially decisive, the predictability of support among likely nonvoters may make a mobilization strategy more appealing. Candidate efforts to persuade and win over a small group of floating voters are costly and provide uncertain payoffs. In contrast, and as was exemplified by the winning campaigns of 2004 and 2012, presidential candidates often see larger payoffs in trying to mobilize those who previously did not vote but are highly reliable in their support.

Indeed, beyond studies of campaign fundraising in Congress, we currently have little understanding of elite polarization's relationship with campaign strategy and how elections are won. Polarization has occurred as institutional barriers to voting have eased and as campaign organizations have made major advancements in their ability to contact likely supporters and mobilize them to vote. Game‐theoretic models find that the participation choices of voters modify a candidates desire to appeal to the median voter (e.g., Callander and Wilson 2007). Along these lines, it is possible that polarization has been facilitated by changes in how candidates view and experience campaigning and what type of voters they view as easiest or most desirable to win.

The candidate‐centric nature of modern campaigns may also play a part in answering why we have varying evidence about the strength of partisanship among contemporary Americans. Americans are less willing to claim a party identification, and they have shown slightly higher levels of ambivalence in their evaluations of the political parties more broadly since 1980 (Thornton 2013). This latter trend should theoretically lessen the direct effects of partisanship by promoting informed and ideological voting (Basinger and Lavine 2005; Lavine, Johnston, and Steenbergen 2012). But perhaps there has not been a decline in reliable party support because clear candidate positioning can stand in as a substitute for party attachments among independents and conflicted partisans by clarifying the policy consequences of their support.2323 It is along these lines that Thornton (2014) fails to find party ambivalence promoting defection in presidential voting when the candidates are ideologically more extreme.
Hence, Americans can distance themselves from the parties in their political identities but fail to show similar conflict in their more proximal candidate evaluations. There is a clarity of mind among Americans when they vote, but perhaps not of purpose.

1 Despite declining turnout rates (Abramowitz 2010), the average percentage of presidential voters who are pure independents in the American National Election Studies (ANES) is slightly greater across recent presidential elections (7.3% from 2000 to 2012) than what it was from 1952 to 1968 (7.1%) since more Americans are independent.

2 Ticket splitting also relates to stability in party support and being cross‐pressured and ambivalent (Mulligan 2011). However, rates of ticket splitting are also contingent on incumbency, challenger quality, the southern realignment, and campaign competition (Burden and Kim ball 2002), factors that have less relevance to presidential elections.

3 Third‐party voters are excluded from the measure since variations in third‐party support mostly reflect the entry of third parties instead of tendencies to vote for a different party. Supporters of Perot in 1992 or Wallace in 1968 were previously unable to support their party. The results in Tables 4–6 address this limitation by imputing an estimate of major‐party switching among nonvoters and third‐party voters as well. They suggest most Perot voters would not have switched their votes if Perot had not run. Among the 1992 fresh cross‐section, the estimated floating voter rate is 20.4% when including Perot voters, 10.6% when excluding Perot voters, and 13.1% when imputing whether a Perot voter would have floated if Perot had not run.

4 The ANES provides the most comparable survey evidence over time, but changes in the design and implementation of the ANES through the years admittedly complicate any inferences. These complications include missing years (1984) and changes in the time and mode of interview, response rates, and weighting. Most notably, response rates have declined from 80% to 60%. Election‐to‐election changes in response rates, however, have a weak correlation with changes in floating voter rates (). I use multiple imputation in the analysis of the ANES panel studies to account for differences in panel attrition.

5 Scores prior to 1972 rely on perceptions of party differences since candidate perceptions were not measured. Analyses fail to indicate substantial differences in presidential nominee and party placements in those instances when the ANES asks respondents to place both. Since ANES surveys differ in the number and type of issues measured across the years, I restrict the score to the five most salient issues of each year, where the most salient issues were determined by the lowest percent of “don't knows” or no opinions. I include scores based on four issues in 1992, which only included four candidate issue placement items, or if a respondent refused to answer one issue item. Issue awareness trends within the consistently asked items, such as those concerning jobs and standard of living, also show positive swings. SI Section A plots national averages by year, and SI Section C provides a list of the variables used to create each year's score.

6 As in Figure 2, I code weak partisans and leaning independents to be equally strong in their identifications. Preliminary tests failed to show substantial differences between weak partisans and leaning independents for all models (with leaning independents slightly more partisan than weak partisans) and supported the collapsed measure's interval coding.

7 Further examination verifies the changes in Figure 3 reflect Americans being more likely to recognize differences on issues on which they have opinions, and not changes in their ability to hold opinions on these issues (see also Abramson, Aldrich, and Rohde 2010, 158).

8 Following Basinger and Lavine (2005) and Thompson, Zanna, and Griffn (1995), let D indicate the average number of likes the respondent provided for the Democratic presidential nominee and dislikes of the Republican nominee, with R indicating the same for the Republican nominee. A respondent's ambivalence is then measured as . As discussed by Rudolph (2005) and Thornton (2011), I treat ambivalence as a nominal trait and identify a respondent as ambivalent if he or she has a score greater than 0. The ANES has yet to release counts and codes of party and candidate comments for 2008 or 2012.

9 The combination of party and candidate perceptions is necessary since respondents are not asked weather presidential candidate differences are important. But the similarity of the party and candidate placement measures suggests respondents view parties and their nominees as similar entities during presidential elections. Tests that include the awareness and importance components separately justify the additive scale. The 1956 ANES did not ask respondents whether they saw important differences between the two parties. I use a respondent's issue awareness level and other attributes to randomly impute a value based on a logit model estimate of the relationship from other years.

10 Interviewer ratings of awareness have a .55 correlation with total number of comments; it only has a .39 correlation with education. I limit the highest value of total comments to 20, since such values are only possible if one does not have univalent candidate and party evaluations.

11 The ANES Time Series studies are multilevel data in that they are a sample of individuals within 12–15 elections. With so few years, however, asymptotic assumptions of maximum likelihood estimates are biased, primarily from standard error estimates that are too small (Stegmueller 2013). As discussed by Hox (2010, 236), it is best to either use a stronger set of Bayesian priors on the variance components to inform these estimates or, as I do, use a bootstrap sample of the elections to provide a nonparametric estimate of how much these associations vary depending on the sample of elections included. An additional advantage in using bootstrap estimates is that they provide bias‐corrected estimates that account for outlying elections having possible undue influence. SI Section B provides further discussion and demonstrates that in most instances the bootstrap estimates match the alternative asymptotic likelihood coefficient estimates but with slightly larger standard errors.

12 This is a common result when accounting for strength of partisanship; see Keele and Wolak (2008) for further discussion and evidence of a positive relationship between awareness, education, and presidential candidate ambivalence.

13 Figure B.2 in the SI demonstrates the strong negative aggregate association between rates of ambivalence and being undecided and the average recognition of differences score.

14 My use of the average score in 2012 is not an out‐of‐sample prediction since there are plenty of independents and strong partisans with recognition of differences scores of either 0 or 100 from 1956 to 2004.

15 Mixed logit estimates of the same model with correlated errors of different scales across choices failed to indicate the multinomial logit specification provided a significant loss of fit, supporting its IIA assumption.

16 The recent growth of strength of partisanship also fails to show stronger effects when restricting the analysis to post‐1980 studies, as clarity of choice factors continue to have the largest effects.

17 Poor income growth can indirectly lead to more floating through its effects on ambivalence (Table 1). Measures of extremity of economic conditions (very good or very poor income growth) failed to show significant associations with rates of floating.

18 For example, SI Section D shows that poor economic growth is no longer positively associated with rates of voters switching to punish incumbents.

19 About 8.65% of Americans are floating voters when presidents run a second time, compared to 10.69% otherwise, and 10.07% if 1968 is excluded, when Johnson opted against running out of fear of losing.

20 Observation rates for both voting turnout measures are (in chronological order) 91%, 92%, 54%, and 46%.

21 I exclude respondents not interviewed in the original wave, since these interviews were added to correct for attrition biases within the original sample. I use the bootstrapping‐based EM algorithm within R's version of Amelia II (Honaker, King, and Blackwell 2008) to impute values of each variable, except candidate preference. I rely on an additional set of transformed (i.e., passive) variables to sharpen the imputation of candidate preference within a wave‐specific logit model. SI Section E provides a list of the variables and specifications used in estimation.

22 In comparing these four studies, the bottom quartile of average total mentions was consistently five or below, with the top quartile showing mention rates above 11.

23 It is along these lines that Thornton (2014) fails to find party ambivalence promoting defection in presidential voting when the candidates are ideologically more extreme.

Biography

• Corwin D. Smidt is Assistant Professor, Department of Political Science, Michigan State University, South Kedzie Hall, 368 Farm Lane, S303, East Lansing, MI 48824 ([e-mail]).