AFFECTIVE POLARIZATION: CURRENT THEORY AND MEASUREMENT

Affective polarization is the tendency to evaluate in-party members positively and out-party members negatively (Iyengar and Westwood Reference Iyengar and Westwood2015). This conceptualization is rooted in social identity theory, which argues that identification with a social group (such as a political party) is sufficient to generate favoritism for that group (Huddy, Bankert, and Davies Reference Huddy, Bankert and Davies2018; Tajfel et al. Reference Tajfel, Billig, Bundy and Flament1971). According to social identity theory, people want to feel positively about the groups they belong to, which leads them to want the ingroup to do better than or compare favorably to outgroups (Tajfel and Turner Reference Tajfel, Turner, Hatch and Schultz1979). To measure affective polarization, various indices have been used, including trait ratings of the two parties (Iyengar, Sood, and Lelkes Reference Iyengar, Sood and Lelkes2012; Levendusky Reference Levendusky2018; Levendusky and Malhotra Reference Levendusky and Malhotra2016a), measures of social distance (Iyengar, Sood, and Lelkes Reference Iyengar, Sood and Lelkes2012; Klar, Krupnikov, and Ryan Reference Klar, Krupnikov and Ryan2018), the number of positive and negative things an individual says about the two parties (Levendusky and Malhotra Reference Levendusky and Malhotra2016a), and even implicit association tests (Iyengar and Westwood Reference Iyengar and Westwood2015).

Most common in the literature, however, is warmth bias. Warmth bias is assessed by taking the difference between ratings of the in-party and the out-party (assessed using thermometer ratings of the Democratic and Republican parties). One of the major advantages of using this measure is that it can be constructed using only two items, with these two items frequently present in long-running, archived surveys such as the American National Election Studies. With data spanning decades, scholars have shown that the electorate has become increasingly polarized compared to previous eras (Iyengar and Krupenkin Reference Iyengar and Krupenkin2018: Iyengar et al. Reference Iyengar, Lelkes, Levendusky, Malhotra and Westwood2019: Iyengar, Sood, and Lelkes Reference Iyengar, Sood and Lelkes2012).

Though two-party U.S. politics has been a dominant focus of the literature on affective polarization, it has also been studied in nations with multiparty systems. These studies suggest that affective polarization is not exclusive to the U.S. two-party system (Boxell, Gentzkow, and Shapiro Reference Boxell, Gentzkow and Shapiro2024; Gidron, Adams, and Horne Reference Gidron, Adams and Horne2020). Studies in this comparative literature have also relied largely on warmth-bias measures (Röllicke Reference Röllicke2023), albeit with more complex operationalizations. Warmth bias is typically measured in terms of in-party affect minus mean affect toward all out-parties (Boxell, Gentzkow, and Shapiro Reference Boxell, Gentzkow and Shapiro2024; Reiljan Reference Reiljan2020), in-party affect minus mean affect toward all parties outside the in-party’s electoral bloc (Reiljan and Ryan Reference Reiljan and Ryan2021; Wagner Reference Wagner2021), or differences in affect between parties in different ideological blocs (Bantel Reference Bantel2023).Footnote ²

Despite these advantages, there are methodological concerns with warmth bias measures that present challenges to the inferences we can make about polarization. One major methodological concern with warmth bias is identifying exactly which group of partisans respondents are thinking of when rating the political parties. Druckman and Levendusky (Reference Druckman and Levendusky2019) find that when experimentally manipulating the partisan feeling thermometer targets to be either the party itself, candidates and elected officials from the party, or voters from the party, respondents indicate greater warmth toward opposing party voters than the party itself or its politicians. Their work suggests that standard party feeling thermometers may actually be measuring feelings toward elites, not partisans in the mass public.

In addition, much of the theory surrounding affective polarization has been ad hoc. As mentioned before, the concept of affective polarization and its most common measure were first put forward in Iyengar, Sood, and Lelkes (Reference Iyengar, Sood and Lelkes2012), yet the concept was not formally defined until Iyengar and Westwood (Reference Iyengar and Westwood2015). This has led to inconsistent measurement, as a strong theory did not precede the emergence of various measurement strategies. Furthermore, theory on the political consequences of affective polarization has rarely been accompanied by empirical analyses. What empirical work has been done, however, has found little evidence for a link between warmth bias and deference for democratic norms (Broockman, Kalla, and Westwood Reference Broockman, Kalla and Westwood2023; Voelkel et al. Reference Voelkel, Chu, Stagnaro, Mernyk, Redekopp, Pink and Druckman2023, although see Kingzette et al. Reference Kingzette, Druckman, Klar, Krupnikov, Levendusky and Ryan2021).

That said, we do not think that these findings warrant an abandonment of the concept of affective polarization, efforts to measure it, or a robust interest in its correlates. Nor do we believe that existing research on the topic has been in vain. However, we do believe that the literature on affective polarization is ripe for a new approach, specifically one that more exhaustively accounts for the diversity of psychological processes that create partisan animus.

AN UPDATED AND EXPANDED MODEL OF AFFECTIVE POLARIZATION

To provide a new theoretical model of affective polarization, we must first determine what psychological processes are likely to account for inter-partisan conflict. One of the biggest limitations of most conceptualizations of affective polarization is a lack of explicit attention to the fact that it might have multiple components. Although most of the literature focuses on partisan antipathy (Abramowitz and Webster Reference Abramowitz and Webster2016; Reference Abramowitz and Webster2018; Lelkes and Westwood Reference Lelkes and Westwood2017; Westwood, Peterson, and Lelkes Reference Westwood, Peterson and Lelkes2019), strong attachment to the in-party also plays a role in partisan behavior (Huddy, Mason, and Aarøe Reference Huddy, Mason and Aarøe2015). Consequently, we argue that affective polarization is multidimensional. A multidimensional theory of affective polarization allows us to account for the possibility that different dimensions of affective polarization while correlated may predict somewhat different sets of political consequences.

But what dimensions should we attend to? In an insightful review, Finkel et al. (Reference Finkel, Bail, Cikara, Ditto, Iyengar, Klar and Mason2020) analogize polarization to religious sectarianism, and argue that three processes undergird interpartisan conflict: othering, aversion, and moralization. They define othering as the tendency to view opposing partisans as essentially different or alien to oneself, aversion as the tendency to dislike and distrust opposing partisans, and moralization as the tendency to view opposing partisans as iniquitous. Though affective polarization could conceivably consist of a wide range of perceptions, this tripartite distinction narrows things down by focusing on core psychological aspects of intergroup differentiation frequently identified by political psychologists: the tendency to exaggerate ingroup/outgroup differences (Tajfel and Turner Reference Tajfel, Turner, Hatch and Schultz1979), the tendency to like outgroups less (Iyengar et al. Reference Iyengar, Lelkes, Levendusky, Malhotra and Westwood2019; Tajfel and Turner Reference Tajfel, Turner, Jost and Sidanius2004), and the tendency to see one’s own orientations as being rooted in moral conviction (Skitka and Bauman Reference Skitka and Bauman2008; Skitka et al. Reference Skitka, Hanson, Scott Morgan and Wisneski2021). Finkel and colleagues posit that the confluence of these three ingredients which they label political sectarianism is what poses a threat to democracy.

We believe that this tripartite reconceptualization of affective polarization provides a richer depiction of the ensemble of processes that account for inter-partisan animosity, and we build on key aspects of it in the development of our own model. Following Finkel et al. (Reference Finkel, Bail, Cikara, Ditto, Iyengar, Klar and Mason2020), we believe that it is essential to understand affective polarization as consisting more than an undifferentiated tendency to feel more negatively about out-partisans than in-partisans, and we believe that the broad concepts of othering, aversion, and moralization provide a good starting point for identifying the multiple components of affective polarization. Like Finkel and colleagues, we also regard these three components as constitutive parts of a composite construct reflecting animosity rooted in partisan identity.

Nevertheless, in several respects, we depart from their understanding of othering, aversion, and moralization, how they relate to one another, and how they relate to affective polarization in general. First, whereas Finkel et al. (Reference Finkel, Bail, Cikara, Ditto, Iyengar, Klar and Mason2020) primarily conceptualize othering, aversion, and moralization as three kinds of negative attitudes toward partisan outgroups, our model defines each more broadly as a general interrelated set of beliefs about the out-party, the in-party, and the relation between the two. While out-party disdain is undoubtedly a growing concern (Abramowitz and Webster Reference Abramowitz and Webster2016; Iyengar and Krupenkin Reference Iyengar and Krupenkin2018), it is only one piece of the puzzle. In characterizing affective polarization as such, it is important to note that neither concept nor our measure of it focus on the specific content of othering, aversion, and moralization, or differences between in-partisans and out-partisans more generally. Rather, othering focuses on the subjective perception of difference, allowing partisans to “fill in the blanks.” This follows from the notion that affective polarization is rooted in perceived differences in social identity rather than substantive disagreement in terms of ideology and issues (Finkel et al. Reference Finkel, Bail, Cikara, Ditto, Iyengar, Klar and Mason2020; Iyengar, Sood, and Lelkes Reference Iyengar, Sood and Lelkes2012; Mason Reference Mason2018).

Second, whereas Finkel and colleagues assert that othering, aversion, and moralization generally co-occur and have their strongest relationship with partisan hostility in conjunction with one another, our model relaxes these assumptions. We argue that the dimensions will be positively correlated, but not redundant. In other words, we conceptualize othering, aversion, and moralization as distinguishable elements of affective polarization that should have a moderately correlated three-factor structure.Footnote ³

More importantly, we also argue that each dimension may elicit conflict on its own, even holding the others constant. That is, the confluence of othering, aversion, and moralization may not be necessary for the formation and expression of attitudes and beliefs deleterious to democracy. Rather, any one of these components may independently predict these outcomes—and not all three components may be relevant to all outcomes associated with polarization. For this reason, we believe it is important to consider the three facets independently and not just as part of a single composite. Below, we describe our own conceptualization of othering, aversion, and moralization.

OVERVIEW OF STUDIES

We empirically develop our model of affective polarization in three sets of analyses. The first focuses on the construction of a new scale measure of the three subdimensions of affective polarization, the second focuses on validation of that measure, and the third focuses on the measure’s relationship with theorized outcomes of affective polarization. These three analyses were carried out using four original surveys conducted during 2022 and 2023. The sample characteristics of all surveys as well as all survey items are available in Section B of the Supplementary Material and replication materials are available at Campos and Federico (Reference Campos and Federico2025). For initial scale-building purposes, Study 1 ( $ N=500 $ ) and Study 2 ( $ N=501 $ ) were collected through the online survey website Prolific. Prolific uses convenience samples of U.S. adults. For both Studies 1 and 2, we applied a quota limiting how many Democrats and Republicans could enroll, with the aim of having an equal number of Democrats and Republicans. Study 1 was conducted in February and March of 2022 and Study 2 was collected in March of 2022.

Study 3 ( $ N=1,346 $ ) was preregisteredFootnote ⁴ and collected via online survey website Lucid, which approximates a nationally representative sample of U.S. adults. Respondents who did not pass an initial attention check were not permitted to take the survey. Those who identified as pure Independents were also excluded from the analyses. Study 3 was conducted in September and October of 2022.

Study 4 was a three-wave panel survey ( $ {N}_{W1}=2004,{N}_{W2}=1404,{N}_{W3}=1054 $ ) conducted as part of the University of Minnesota Center for the Study of Political Psychology’s 2022–2023 Multi-Investigator Panel Study. The panel survey was collected via Bovitz/Forthright, which approximates a nationally representative sample of U.S. adults. Respondents who did not pass an initial attention check were not permitted to take the survey, and a second attention check was administered halfway through the survey, with respondents who failed the second check excluded from the analyses. Respondents who identified as pure Independents in the survey were also excluded. Wave 1 was conducted in December of 2022, Wave 2 was conducted in March of 2023, and Wave 3 was conducted in May and June of 2023.

PART I: SCALE CONSTRUCTION

The goal of Part I is to construct a scale measure of the three subdimensions of affective polarization, and test whether our tripartite conceptualization actually exists in the minds of partisans. We began this process using data from Study 1.

Further Item Reduction: Deriving the Final Nine-Item APS Scale

Although our twenty-item APS performed well, a scale of this length is not practical for most surveys. We reduced the subdimension scales for the initial twenty-item measure to three-item subscales using a combination of IRT methods and subjective item-selection criteria, including face validity and item non-overlap. To carry out the item reduction, we turned to data from Study 4 Wave 1, which had the largest sample size and the most approximately representative sample.Footnote ⁶ Scale reduction was carried out separately within each dimension to satisfy the unidimensionality assumption of IRT (Embretson and Reise Reference Embretson and Reise2000). Details of the item-reduction process are described further in Section C.4 of the Supplementary Material.

For othering, we began by fitting an IRT model for polytomous items—the graded response model (GRM; Samejima Reference Samejima1969)—to the full set of eight items. Items were then retained or discarded on the basis of the area under that item’s information curve (which plots how precisely scores on the latent trait dimension are estimated against latent trait scores; Embretson and Reise Reference Embretson and Reise2000). Using this criterion, the two lowest-information items were discarded and the two highest-information items were retained. The remaining items were similar in information. Because of this, the third item was chosen after an examination of its information curve. The information curve indicated that it provided information about differences among individuals across a wider range of the latent othering dimension than the other items, without appreciably reducing the amount of information provided by the entire three-item scale.

For aversion, a similar procedure was followed with the six items from the initial scale. On the basis of the areas under the items’ information curves, the two items with the lowest information were dropped, and the two items with the highest information were retained. The two remaining items were similarly high in information. Of these, we ultimately chose the one with lower information as the third item, as its information curve indicated greater coverage of a wider range of latent aversion scores. This item was also more face-valid than the higher-information alternative, given that the latter contained content that touched on moralizing themes (“Identifying as a [out-party] rather than a [in-party] makes someone a bad person”); the chosen third item also had the strength of being reverse-coded.

For moralization, the six items from the initial twenty-item scale were three pairs, with the only difference within each pair being whether they asked about the respondents’ “feelings about politics” or the respondents’ “identity as a [in-party],” Footnote ⁷ with the loadings being almost indistinguishable. We chose to keep the three items that prompted partisan identity on face-validity grounds, as affective polarization is more conceptually linked to partisan identity. As an additional check, a GRM like the one used above was fit to the full set of six items. Examination of the area under the items’ information curves confirmed (1) that the partisan-identity version of each item provided more information than its paired counterpart and (2) that the total information provided by a scale based on the three partisan identity items was greater than that provided a scale based on the other three. The final nine items are in Table 1.

Table 1. Final Nine-Item Affective Polarization Scale

Note: * denotes reverse coded items.

As an initial check, we estimated a three-factor CFA model in the Study 4 Wave 1 data. This model fit very well, $ {\chi}^2(24)=87.989,p<0.001,CFI=0.988,RMSEA=0.045 $ . It also fit better than an alternative single-factor model, $ p<0.001 $ . For additional confirmation, we examined the fit of the three-factor factor model in the two other datasets that were not used to generate the original scale items (Studies 2 and 3). The three-factor model fit very well in Study 2, $ {\chi}^2(24)=51.290,p<0.001,CFI=0.989,RMSEA=0.051 $ ; and in Study 3, $ {\chi}^2(24)=72.599,p<0.001,CFI=0.987,RMSEA=0.048 $ . In Studies 2 and 3, this model also fit better than the one-factor alternative ( $ ps<0.001 $ ).

In the three-factor models, the correlations between the subdimension factors were similar across Studies 2–4. Othering and aversion (Study 2: 0.55; Study 3: 0.52; Study 4W1: 0.50) and othering and moralization (Study 2: 0.53; Study 3: 0.60; Study 4W1: 0.56) were moderately correlated in all samples, whereas aversion and moralization had smaller correlations (Study 2: 0.26; Study 3: 0.25; Study 4W1: 0.24). This pattern of correlations among the latent subdimensions of the APS suggests that the facets of the construct covary, but not so highly as to make the subdimensions redundant: all correlations fall well below the cutoff of 0.80 recommended by Rönkkö and Cho (Reference Rönkkö and Cho2022) for factor discriminability in CFA.

Given its panel structure, Study 4 allowed us to examine the test-retest reliability of the APS and its subdimensions. To provide information on test-retest reliability for the three subdimensions and the full scale, we present simple between-wave test-retest correlations and more formal estimates using the intra-class correlation in Section C.5 of the Supplementary Material (Shrout and Fleiss Reference Shrout and Fleiss1979). These results suggest that the full scale and aversion show good reliability, whereas othering and moralization show moderate reliability. As one would expect, the full APS shows higher test-retest reliability, given the larger number of items. Interestingly, test–retest reliability is slightly lower for moralization (though still within acceptable range), despite the fact that the moralization scale shows similar internal consistency to the other subscales (see below). This suggests that moralization scores, while internally consistent, may vary more over time.

Using multigroup CFA, we also conducted a series of sequential measurement invariance tests comparing the properties of our three-factor solution for the final nine-item scale among Democrats and Republicans (see Section C.6 of the Supplementary Material; Brown Reference Brown2006). In both Studies 3 and 4 W1, we found that the final nine-item APS showed configural, metric, and scalar invariance when comparing factor structures for Democrats and Republicans. Given these results, cross-partisan comparisons of regression coefficients for the APS and its subdimensions in models predicting various outcome variables and cross-partisan comparison of mean scores on the APS and its subdimensions can be validly conducted (Vandenberg and Lance Reference Vandenberg and Lance2000).

Given this evidence, the nine-item reduced scale was selected as the final APS, with three items per subdimension. We examined the internal consistency of the subscales and the full scale using a variant of the $ \omega $ coefficient for multidimensional scales, which estimates reliability based on the final three-dimensional factor models above (Savalei and Reise Reference Savalei and Reise2019; see also Cho Reference Cho2016; Forbes et al. Reference Forbes, Greene, Levin-Aspenson, Watts, Hallquist, Lahey and Markon2021).Footnote ⁸ The subdimension scales were internally consistent in all studies: othering $ ({\omega}_{S1}\hskip-0.2em =0.79,{\omega}_{S2}\hskip-0.2em =\hskip-0.2em 0.78,{\omega}_{S3}=0.80,{\omega}_{S4W1}=0.80), $ aversion ( $ {\omega}_{S1}=0.90,{\omega}_{S2}=0.90,{\omega}_{S3}=0.82,{\omega}_{S4W1}=0.86), $ moralization ( $ {\omega}_{S1}=0.89,{\omega}_{S2}=0.91,{\omega}_{S3}=0.88,{\omega}_{S4W1}=0.86 $ ). To assess the adequacy of the full scale, we also examined the total $ \omega $ coefficients in each sample, which indicates the portion of variance in the full nine-item scale that is accounted for by all three APS factors (Revelle and Zinbarg Reference Revelle and Zinbarg2009). These statistics also indicated a highly-reliable full scale in each sample ( $ {\omega}_{S1}=0.92,{\omega}_{S2}=0.92,{\omega}_{S3}=0.91,{\omega}_{S4W1}=0.91 $ ). These estimates suggest that both the APS subdimension scales and full scale are sufficiently reliable for analytic use.Footnote ⁹

PART II: SCALE VALIDATION

Having constructed a new measure of affective polarization, we seek to validate the APS in Part II through testing its association with theoretically related measures. To do this, we examine the relationship between the APS and measures of partisan and ideological identity, political knowledge, and previous measures of affective polarization.

The seven-point party identification scale is possibly the most used measure in work on American political behavior, and although we are not interested in predicting which party someone identifies with, we can use this scale to test whether the APS is associated with stronger identification with a party in general.Footnote ¹⁰ We also employ a different measure of partisan identity strength that conceptualizes partisan identification as a social identity (Huddy, Mason, and Aarøe Reference Huddy, Mason and Aarøe2015; $ {\alpha}_{S1}=0.85,{\alpha}_{S2}=0.85,{\alpha}_{S3}=0.84,{\alpha}_{S4W1}=0.85 $ ).

To measure how strongly someone identifies with the liberal or conservative label, we use a seven-point symbolic ideology scale. Again, we are not interested in predicting whether someone is liberal or conservative, but rather how strongly they identify with one of these ideological labels. Another common measure is political knowledge. Knowledge represents how much a respondent knows about national politics, with those who are more knowledgeable having stronger ideological attachments as well as holding more durable and constrained issue positions (Converse Reference Converse1964; Kalmoe Reference Kalmoe2020; Kinder and Kalmoe Reference Kinder and Kalmoe2017). We used this measure to examine how much those high in affective polarization are aware of national politics ( $ {\alpha}_{S3}=0.67,{\alpha}_{S4W1}=0.48 $ ).Footnote ¹¹

Warmth bias, as noted previously, is the most common measure of affective polarization. Although we see this measure as conceptually different from our measure, we do expect the APS and warmth bias to be related.Footnote ¹² We also examine the relationship between the APS and trait ratings of partisans ( $ {\alpha}_{S2}=0.91,{\alpha}_{S3}=0.89 $ ). In Study 2, we administered an eight-item scale of trait ratings, and in Study 3, a four-item measure. These scales measured trait ratings for both Republicans and Democrats with seven response options ranging from a negative trait attribution to a positive trait attribution (e.g., lazy-hardworking). The difference between a participant’s in-party and out-party ratings was used to create a trait-rating bias measure. Again, we see this measure as conceptually different, but we still do expect it to be related to our measure.

Ordinary least squares (OLS) regressionFootnote ¹³ was used to examine the relationship between the APS and our slate of outcome variables. Because of our focus on the uniqueness of each subdimension, we estimated separate models where the subdimensions serve as separate predictors. This provides an additional test of our hypothesis that the subscales are not interchangeable, as they may predict the various outcomes in different ways. All variables were recoded to run from 0 to 1 so that coefficients represent the expected percentage change in the dependent variable associated with going from the minimum to the maximum of the respective independent variable.

The full results of our validation analyses are presented in Figure 1. We found that the full APS was positively associated with partisan identitfication (PID) extremity in all samples, with the effect of going from the minimum to the maximum value of the APS on PID extremity ranging from 73% in Study 4 W1 to 99.6% in Study 3. Looking at the subdimension models, othering was only weakly positively associated with PID extremity in Study 3 ( $ b=0.16,95\%\hskip2.22198pt CI\hskip2.22198pt [0.03,0.29] $ ), aversion was positively associated with PID extremity in all but Study 4 W1 ( $ b=0.04,95\%\hskip2.22198pt CI\hskip2.22198pt [-0.04,0.13] $ ), and moralization was strongly associated with PID extremity in all samples. The models predicting partisan social identity provide additional evidence that the APS is related to stronger partisan attachments. The full APS was positively associated with partisan social identity in all four samples, with the effect of the APS on partisan social identity ranging from 65% in Study 2 to 75% in Study 3. Looking at the subdimension models, othering was positively associated with partisan social identity across all samples (although more weakly than moralization), aversion was weakly positively associated with partisan social identity in Studies 3 and 4 W1, and moralization was associated with partisan social identity extremity across all samples.

Figure 1. APS and Subdimensions Predicting Political Identity, Knowledge, and Bias

Note: Points represent unstandardized coefficients and lines represent 95% confidence intervals. Models also included demographic controls such as age, bachelor’s degree, white, Hispanic/Latino, gender, and income. Regression tables can be found in Section D of the Supplementary Material.

Moving to the models predicting the strength of ideological identity, we found that the full APS was positively associated with ideological identity extremity in all samples, with the effect of the APS on ideological identity extremity ranging from 66% in Study 4 W1 to 75% in Study 2. Looking at the subdimension models, othering was only weakly positively associated with ideological identity extremity in Study 4 W1 ( $ b=0.10,95\%\hskip2.22198pt CI\hskip2.22198pt [0.01,0.19] $ ), aversion was positively associated with ideological identity extremity in all samples (although weaker than moralization), and moralization was associated with ideological identity extremity across all samples.

Contrary to our expectations, we found that the full APS was negatively associated with knowledge in Study 3 ( $ b=-0.15,95\%\hskip2.22198pt CI\hskip2.22198pt [-0.23,-0.07] $ ) yet positively associated with knowledge in Study 4 W1 ( $ b=0.12,95\%\hskip2.22198pt CI\hskip2.22198pt [0.05,0.20] $ ). Looking at the subdimension models, however, relationships were more consistent across samples. Othering was unrelated to knowledge in either sample, aversion was negatively associated with knowledge in both samples, and moralization was positively associated with knowledge in both samples. Though we expand more on our findings for political knowledge in the discussion, we note here that the inconsistent results for the full scale seem to be a result of cross-sample differences in the magnitude (as opposed to direction) of the subdimension relationships. Aversion and moralization have opposed relationships with political knowledge in both samples, but because the strength of these independent relationships differs across the studies, the composite effect reflected by the full scale appears inconsistent. This gives some additional credence to our argument that the subscales be treated as related but unique forms of affective polarization, as the composite measure may hide important differences between the subdimensions.

We now turn to our scale’s relationship with past measures of affective polarization. Our results suggest that the full APS was positively associated with warmth bias in all three samples, with the effect of the APS on warmth bias ranging from 44% in Study 3 to 69% in Study 2. Looking at the subdimension models, all three dimensions were positively associated with warmth bias in all samples, except for aversion in Study 3 ( $ b=0.00,95\%\hskip2.22198pt CI\hskip2.22198pt [-0.05,0.05] $ ). Models predicting trait rating bias suggest that the full APS was positively associated with trait rating bias in both samples. Looking at the subdimension models, all three subdimensions were positively associated with trait rating bias in both samples, except for aversion in Study 3 ( $ b=0.02,95\%\hskip2.22198pt CI\hskip2.22198pt [-0.03,0.07] $ ).

Taken together, these results suggest that our scale is a valid measure of affective polarization. Additionally, the differing pattern of results at the subscale level further supports our focus on the uniqueness of the subdimensions. Moralization was the most consistent predictor of political identification variables, and was the only subdimension related to political knowledge. Though less consistently related to the identity measures, othering was the most reliable predictor of previous affective polarization measures. Aversion, despite being consistently related to ideological extremity, was a less steady predictor of partisan attachment and previous forms of affective polarization, yet was negatively related to political knowledge. With our measure validated we can now examine how our theory of affective polarization relates to possible downstream consequences.

PART III: AFFECTIVE POLARIZATION AND ANTI-DEMOCRATIC ATTITUDES

Much of the recent work on outcomes of affective polarization has been interested in how increasing levels of affective polarization might lead to democratic backsliding. Counter to much of the theorizing on affective polarization, few empirical studies have found evidence for this link, leading many scholars to conclude that there is no direct relationship between affective polarization and democratic backsliding (Druckman, Green, and Iyengar Reference Druckman, Green and Iyengar2023). As stated at the outset of this paper, we argue that this conclusion is a result of current theory and measurement missing out on the complexities of affective polarization. Having constructed and validated a new multidimensional measure of affective polarization, we test this claim empirically here in Part III.

While the process of democratic backsliding involves various actors and institutions (Druckman Reference Druckman2024), we focus here on two sets of citizen attitudes that can contribute to backsliding. The first includes individual citizen’s support for anti-democratic elite action. Elites seeking to consolidate power often attack democratic institutions through the violation of norms or laws (Ahmed Reference Ahmed2023), and popular support for these actions legitimize those who seek to dismantle democracy for their own gains. The kinds of elite actions that can lead to democratic backsliding are numerous, and therefore we employ various measures of support for elite attacks on democracy.

Our first measure of support for anti-democratic action is a “rules of the game” scale (Clarke Reference Clarke2022; $ {\alpha}_{S2}=0.85,{\alpha}_{S3}=0.89 $ ). This scale measures support for breaking or circumnavigating democratic norms and laws by politicians or the ingroup. We also administered a democratic norms scale made up of various items from the ANES ( $ {\alpha}_{S3}=0.62 $ ). This scale measures individuals’ support for various norms such as freedom of the press, checks and balances, and elite accountability, therefore we would expect a negative relationship between this measure and the APS. Gidengil, Stolle, and Bergeron-Boutin (Reference Gidengil, Stolle and Bergeron-Boutin2022) find that highly partisan or ideological Americans are willing to support in-party candidates that espouse anti-democratic positions in order to make partisan or ideological gains. We measured support for an anti-democratic in-party candidate with a scale asking respondents how likely they are to vote for an in-party candidate over an out-party candidate after learning various anti-democratic positions espoused by the hypothetical in-party politician (Voelkel et al. Reference Voelkel, Chu, Stagnaro, Mernyk, Redekopp, Pink and Druckman2023; $ {\alpha}_{S3}=0.93 $ ). Anti-democratic views included suppression of the press and free speech, ignoring unfavorable court rulings, voter suppression, and ignoring unfavorable election results.

We also administered a “partisan spite” scale that measures respondents’ support of spiteful in-party elite action against the out-party (Moore-Berg et al. Reference Moore-Berg, Ankori-Karlinsky, Hameiri and Bruneau2020; $ {\alpha}_{S3}=0.87 $ ). These actions include hurting the out-party at expense of the country or economy, voter suppression, and suppression of out-party news organizations. Finally, we measured a more direct form of endorsement of anti-democratic action, support for authoritarian rule ( $ {\alpha}_{S3}=0.76 $ ). Two items measuring support for an authoritarian executive and army rule were taken from the World Values Survey.

Our second set of attitudes that contribute to democratic backsliding is support for political violence. Recently, the U.S. has seen many instances of political violence, such as an attempt to kidnap Governor Gretchen Whitmer, the January 6th insurrection at the Capitol, and even an assassination attempt against President Donald Trump (Hanna et al. Reference Hanna, Thompson, Mulvihill and Collins2024). Though support for political violence among the general population remains low (Holliday et al. Reference Holliday, Iyengar, Lelkes and Westwood2024), it is essential to identify what psychological mechanisms lead individuals to support political violence (Kalmoe and Mason Reference Kalmoe and Mason2022b). We sought to measure respondents’ support for political violence through both a scale referencing violence toward the out-party specifically (Kalmoe and Mason Reference Kalmoe and Mason2022b; $ {\alpha}_{S2}=0.81 $ ), and support for violence generally ( $ {\alpha}_{S4W1}=0.81 $ ).

Considering recent debates over how best to measure support for political violence (Kalmoe and Mason Reference Kalmoe and Mason2022a; Westwood et al. Reference Westwood, Grimmer, Tyler and Nall2022a; Reference Westwood, Grimmer, Tyler and Nall2022b), in Study 3 an experiment measuring support for specific acts of political violence was replicated from Westwood et al. (Reference Westwood, Grimmer, Tyler and Nall2022a). Respondents were randomly assigned to one of six conditions where they were told someone was recently arrested for committing a political crime. In all conditions, it is clear that this crime was against the respondents’ out-party, with the severity of the crime ranging from protesting without a permit to murder. To gauge attitudes toward the perpetrator, respondents were asked how severe of a sentence they should receive, with responses ranging continuously from community service to more than twenty years in prison. Additionally, respondents where asked whether they supported a pardon for the perpetrator. Per the recommendations of Westwood et al. (Reference Westwood, Grimmer, Tyler and Nall2022a), an attention check item was administered, with respondents who failed this additional check being excluded from the analyses. Our goal here is not to come up with an estimate of general support for political violence, but rather to examine heterogeneous treatment effects among those higher or lower in affective polarization.

We also explore different perspectives on the relationship between affective polarization and anti-democratic attitudes. Kingzette et al. (Reference Kingzette, Druckman, Klar, Krupnikov, Levendusky and Ryan2021) find that the relationship between affective polarization and anti-democratic attitudes is moderated by which party is institutionally dominant, meaning partisans support democratic norms when their party is in the minority but abandon these norms when their party is in the majority. Following this perspective, we would expect the APS and its subdimensions to have a stronger relationship with anti-democratic attitudes among Democrats than Republicans, as the surveys presented in this section were administered when the Democratic party held either a “trifecta” or both the Presidency and the Senate.

Another line of work argues that Republican elites are unique in that they are antagonistic toward the tenets of democracy, whereas Democratic elites are not (Grumbach Reference Grumbach2022; Sides, Tausanovitch, and Vavreck Reference Sides, Tausanovitch and Vavreck2022). If so, affectively polarized Republicans in the electorate should follow cues from Republican elites and hold stronger anti-democratic attitudes than Democrats. Following this perspective, we would expect the APS and its subdimensions to have a stronger relationship with anti-democratic attitudes among Republicans than Democrats. We also hypothesize that there could be no partisan differences.

To test our hypotheses, we used OLS regressionFootnote ¹⁴ to examine the relationships between the full APS and our outcome variables. Just as in Part II, we run models with the subscales as independent predictors and expect there to be different patterns of relationships between the three subdimensions and the outcome variables. We also specify separate models with and without warmth bias as an added covariate to ensure that the APS and its subscales non-redundantly predict the outcomes of interest net of the most widely used index of affective polarization in the extant literature.Footnote ¹⁵ All variables were recoded to run from 0 to 1.

Figure 2 presents the main results for anti-democratic attitudes. We found that the composite APS scale was positively related to all anti-democratic attitudes scales fielded with the exception of the “rules of the game” scale in Study 2. As expected, there was a negative relationship between the full scale and support for democratic norms, though this relationship was quite small ( $ b=-0.07,95\%\hskip2.22198pt CI\hskip2.22198pt [-0.14,-0.004] $ ). Looking at the subdimension models, othering was positively associated with all measures of anti-democratic attitudes except for democratic norms, support for authoritarian rule, and partisan violence where we found no evidence for a relationship. Aversion was positively associated with all measures of anti-democratic attitudes except for Study 2 “rules of the game,” and was negatively associated with democratic norms. In all models where both othering and aversion were related to the outcome variable in the expected direction, aversion was a substantially larger predictor than othering. Moralization was positively associated with supporting an anti-democratic candidate and partisan spite, but negatively associated with Study 2 rules of the game and positively associated with support for democratic norms. We also found no evidence for a relationship between moralization and authoritarian rule or either measure of political violence.

Figure 2. APS, Subdimensions, and Warmth Bias Predicting Anti-Democratic Attitudes

Note: Points represent unstandardized coefficients and lines represent 95% confidence intervals. Models also included demographic controls such as age, bachelor’s degree, white, Hispanic/Latino, gender, and income. Regression tables can be found in Section E of the Supplementary Material.

Again the subdimension models show that the three subscales relate to outcome variables differently. Like with political knowledge in Part II, the effect of moralization counteracts that of aversion and othering for rules of the game (Study 2) and democratic norms. Only looking at the full scale models hides the more complex relationship between the subdimensions and anti-democratic attitudes. When looking at the relationship between warmth bias itself and anti-democratic attitudes conditional on the APS, we actually found evidence for multiple negative relationships with measures of anti-democratic attitudes and a positive relationship with endorsing democratic norms. Warmth bias was also negatively associated with support for general political violence, conditional on the APS. However, warmth bias was positively associated with support for an anti-democratic candidate, conditional on the APS.

Figure 3 presents our main results but with separate models for Democrats and Republicans. There were partisan difference in the relationship between the full scale and the outcome measures for Study 2 rules of the game, support for democratic norms, and partisan violence. For these three outcome measures there was an effect in the expected direction only for Republicans, though in all other models the effect among Democrats and Republicans was substantively the same. For othering, the effects among Democrats and Republicans were similar in all models. Looking at aversion, the effects again were similar across partisanship, except for Study 2 rules of the game where there was only a positive relationship for Republicans and general political violence where the effect for Republicans was larger ( $ {b}_{Dem}=0.19,95\%\hskip2.22198pt CI\hskip2.22198pt [0.14,0.24],{b}_{Rep}=0.32,95\%\hskip2.22198pt CI\hskip2.22198pt [0.25,0.39] $ ). The relationships between moralization and anti-democratic attitudes were similar across party, except for democratic norms, where there was only a positive relationship for Democrats; and support for an anti-democratic candidate, where there was only a positive relationship for Republicans.

Figure 3. APS, Subdimensions, and Warmth Bias Predicting Anti-Democratic Attitudes, by Party

Note: Points represent unstandardized coefficients and lines represent 95% confidence intervals. Models also included demographic controls such as age, bachelor’s degree, white, Hispanic/Latino, gender, and income. Regression tables can be found in Section E of the Supplementary Material.

Given these results, we find no evidence for the institutional hypothesis and some support for the Republican hypothesis. In most cases, the effects of the full APS and the subscales were similar among Republicans and Democrats. When there were differences, Republicans high in affective polarization tended to be more inclined to endorse anti-democratic attitudes. Interestingly, there were major partisan asymmetries for the relationship between warmth bias and anti-democratic attitudes, conditional on the APS. Except for support for an anti-democratic candidate, the negative effects of warmth bias were driven by Democrats, with there being no relationship between warmth bias and these outcome measures among Republicans.

For our partisan violence experiment, the main effects followed those of Westwood et al. (Reference Westwood, Grimmer, Tyler and Nall2022a). Conviction of a more severe crime caused respondents to support harsher sentences and be less supportive of pardons compared to a non-violent crime (see Table E.15 in the Supplementary Material). Because we are interested in heterogeneous treatment effects of the APS, the marginal effects from our partisan violence experiment for the full APS and its subdimensions are shown in Figure 4. The results suggest that higher APS scores were associated with smaller treatment effects for the crimes of arson, assault with a deadly weapon, and murder on sentencing length for the convict. In no condition were different levels of the full APS related to different treatment effects on supporting a pardon for the convict. We also found no heterogenous treatment effects for othering in any condition. Aversion, however, was associated with lower treatment effects of arson, assault, and murder on sentencing length. Additionally, vandalism and assault had higher treatment effects on supporting a pardon among those higher in aversion. We found no evidence for heterogenous treatment effects at different levels of moralization on sentencing length. However, we found that the effect of arson and murder on supporting a pardon was lower for those high in moralization.

Figure 4. Marginal Effects of APS and Subdimensions

Note: Points represent unstandardized coefficients and lines represent 95% confidence intervals. Regression tables can be found in Section E of the Supplementary Material. The reference group for the condition variable is “protesting without a permit.”

In sum, our analyses suggest that affective polarization, operationalized as the APS, is associated with greater support for elite actions that dismantle democracy, less support for democratic norms, and more support for political violence. Furthermore, we found that the subscales related to anti-democratic attitudes in different ways, again supporting our notion that the subscales represent distinct, though related, forms of affective polarization. Aversion was by far the most consistent predictor of our various measures of anti-democratic attitudes, and those high in aversion were even more likely to support softer sentences for those convicted of the most violent of crimes. Othering on the other hand was less robustly associated with attitudes that support democratic backsliding. Though we found some positive relationships between othering and a few of our outcome variables, these effects were often quite small. Moralization was an even less robust predictor of anti-democratic attitudes. Though moralization was positively related to voting for an anti-democratic inparty candidates and partisan spite, we found that those high in moralization were less supportive of certain anti-democratic actions and more supportive of democratic norms. However, this relationship was limited to only a few outcome variables.