This is an open access article distributed under the terms of the Creative Commons
Attribution License (Americans increasingly like their own political party while disliking the opposing party (Iyengar & Krupenkin, 2018; Már, 2020). Much work has examined the growth of such so-called “affective polarization” (e.g., Iyengar & Krupenkin, 2018; Iyengar & Westwood, 2015; Iyengar et al., 2012; Már, 2020). This includes its causes, such as partisan sorting and changes in media consumption (e.g., Garrett et al., 2014; Mason, 2016; Rogowski & Sutherland, 2016), and interventions for reducing it, such as correcting partisan stereotypes (e.g., Ahler & Sood, 2018; Levendusky, 2018).
However, two elements of this literature remain less well-explored. The first relates to a lack of clear consensus on how to measure and conceptualize American partisan affect. Contemporary measures of partisan affect vary considerably. For example, items range from emotional responses toward partisans (e.g., do you feel angry toward Democrats, yes or no; e.g., Mason 2015, 2016, 2018) and candidates (e.g., Garrett et al., 2014), to social distance measures (e.g., how would you feel about having a Republican neighbor, measured on a three-point scale; e.g., Iyengar et al., 2012), and 100-point feeling thermometers (e.g., Iyengar & Krupenkin, 2018; see online Appendix A for a complete list of measures).
One way to get around this problem is to use theoretical or empirical approaches that permit aggregation across these different measures in different ways. In terms of theoretical approaches, some works have examined inparty and outparty feelings separately (e.g., Iyengar et al., 2012), whereas others—such as those in the negative partisanship tradition—have focused primarily on outparty animosity (e.g., Abramowitz & Webster, 2016; Levendusky & Malhotra, 2016). Others, still—such as more classical approaches to political polarization— have collapsed across the two components, investigating partisan affect as the difference between inparty and outparty affect (e.g., Iyengar & Westwood, 2015; Lelkes et al., 2017; Mason, 2015). These theories have different implications for the conceptualization of partisan affect, and what forms of behavior it may drive.
Moreover, although there are sophisticated empirical methods for comparing the suitability of these different theoretical models, the recent proliferation of the different types of partisan affect measures has presented a technical challenge. More classical approaches to these empirical methods (such as conventional forms of factor analyses, like those used by Green, 1988) may be less appropriate for modeling responses across a broad set of partisan affect measures, as they are better suited to consistent, non-categorical item formats (Wirth & Edwards, 2007). Accordingly, the current study aims to compare the fit of these different theoretical models using a more appropriate method in the form of categorical factor analyses.
The second under-explored element of the partisan affect literature involves the political consequences of rising affective polarization (as noted explicitly by Iyengar et al., 2019), particularly in terms of political behavior. While some recent work has suggested that affective polarization has little influence on outcomes such as support for democratic norms, support for bipartisanship, or adoption of inparty policy positions (e.g., Broockman et al., 2022), this work has not addressed political participation directly. Such a gap is relevant given the importance of political participation for the healthy functioning of democracies (Powell, 2000), and the association between affective polarization and other behaviors, such as dating (Huber & Malhotra, 2017) and hiring discrimination (Gift & Gift, 2015).
Investigating what elements of partisan affect predict different aspects of political behavior also permits us to compare the predictive validity of different theoretical models of partisan affect. Specifically, the prior literature suggests three competing hypotheses. First, theoretical approaches that focus primarily on negative affect toward the outparty (e.g., Abramowitz & Webster, 2016) imply the importance of outparty animosity in driving political behavior, suggesting that (1) outparty animosity alone should predict most forms of political behavior. Alternatively, some approaches that distinguish between inparty affinity and outparty animosity (e.g., Iyengar & Krupenkin, 2018) may leave more room for differential effects of inparty affinity and outparty animosity (e.g., Huddy et al., 2015, 2018). For example, in line with prior work on emotions suggesting the motivational power of negative emotions such as anger, in particular (e.g., Harmon-Jones, 2019; Valentino et al., 2011), it may be that (2) outparty animosity will motivate costlier behaviors. In contrast, inparty affinity should motivate less costly behaviors, requiring comparatively little resources or effort. Finally, classical approaches to affective polarization (e.g., Iyengar & Westwood, 2015), might identify the combination of inparty affinity and outparty animosity as a primary driver of behavior. This implies that (3) the interaction between inparty affinity and outparty animosity should predict most forms of political participation. Insofar as one of these predictions is more consistent with the data, this can be taken as empirical support for that theoretical approach to partisan affect over others.
In the present research, we review and build upon existing studies of partisan affect and participation before addressing these two gaps—comparing different conceptualizations of partisan affect and their ramifications for political participation outcomes, in ways that help adjudicate between different theories. We do so using data from the Pew American Trends Panel (Pew Research Center, 2016)—an extensive, nationally-representative survey—which covers a comprehensive set of partisan affect measures from the existing literature and various measures of political participation.
We address the first gap in Study 1A using contemporary categorical approaches to factor analyses (Wirth & Edwards, 2007) to empirically derive and compare the appropriateness of different theoretical models of partisan affect across a particularly extensive set of partisan affect measures. This permitted us to investigate whether partisan affect is better conceptualized as inparty affinity and outparty affinity treated as two separate constructs or as a combined construct.
We address the second gap in Study 1B, in which we examine the associations between partisan affect and different political behaviors, using our partisan affect measures alongside standard covariates in logistic regressions to predict different forms of political participation. This allowed us to investigate the three competing hypotheses detailed above. Moving beyond differentiating between the appropriateness of one- and two-dimensional models of partisan affect, this allowed us to investigate whether models that emphasize inparty affinity or outparty animosity provide better predictive validity when it comes to political participation. In so doing, we also contribute to our understanding of partisan affect’s political consequences.
Our results across Study 1A and 1B indicate three critical conclusions. First, we find that, across a robust set of measures, two separate dimensions—inparty affinity and outparty animosity—best characterize contemporary American partisan affect, consistent with theoretical approaches that divide between these two aspects of partisan affect (e.g., Iyengar et al., 2012). This suggests that scholars may be best served by including two distinct measures of partisan affect rather than a single, cumulative scale, which may help bring more consensus to the measurement of partisan affect.
Second, regarding the predictive validity of different aspects of partisan affect in terms of political behavior, we found the most robust support for our second hypothesis, that outparty animosity and inparty affinity predict different forms of political participation. This is also consistent with approaches that divide between inparty and outparty affect (e.g., Iyengar et al., 2012). However, outparty animosity predicted more, and rarer, forms of political participation than did inparty affinity, which predicted fewer and more common forms of political participation, suggesting some support for approaches that emphasize negative affect (e.g., Abramowitz & Webster, 2016). Surprisingly, and inconsistent with classical approaches to affective polarization, we found that the interaction between inparty affinity and outparty animosity was not a significant predictor in our analyses. Collectively, these results add nuance to our understanding of the appropriateness of different models of American partisan affect and their ramifications for political behavior.
Theoretical and Empirical Approaches to Partisan Affect
Americans’ feelings about U.S. parties and partisans have shifted rapidly in recent years, with partisans reporting notably more negative feelings toward their outparty (Finkel et al., 2020; Iyengar & Krupenkin, 2018; Már, 2020). However, although many conceptualizations and measures of partisan affect have been used in recent years (see online Appendix A), the literature still lacks a unified conceptualization of the construct and its political consequences.
Theoretical Approaches to Partisan Affect
Much of the work on partisan affect is, at least implicitly, rooted in social identity theory. Social identity theory posits that people first categorize themselves and others into different social groups. They then identify with and strengthen their emotional commitment to their social ingroup and seek to feel better at the expense of outgroups (Tajfel, 1982; Tajfel & Turner, 1986). In the political sphere, people form partisan attachments. They have positive and negative feelings about their respective political ingroups and outgroups, and this affective gap tends to be larger when a person has a stronger party identification (Huddy et al., 2018).
Despite this common thread of social identity theory, different theoretical approaches in the partisan affect literature vary in how they conceptualize the links between inparty and outparty feelings. Figure 1 illustrates some of this conceptual space. For instance, whereas some theoretical approaches consider inparty affinity and outparty animosity as separate constructs (as depicted in the x and y axis in Figure 1), others collapse across the two dimensions (as depicted in the dotted diagonal line in Figure 1).
Figure 1
Conceptualizations of Partisan Affect
Importantly, differences in these conceptualizations of partisan affect also have ramifications for the behaviors these models might predict, given work in the emotions literature suggesting that different affective states can motivate different behaviors (Harmon-Jones, 2019; Roseman et al., 1994). For instance, some work on emotions and political behavior suggests that although positive and negative emotions may both lead to greater political participation, positive emotions such as enthusiasm tend to reinforce people’s practiced behavioral tendencies. In contrast, some negative emotions such as anxiety tend to disrupt people’s practiced behavioral tendencies (Marcus et al., 2000). Accordingly, one way to compare different models of partisan affect is to compare how well different components of those models predict different behaviors. We consider each of these models of partisan affect and their implied patterns of behavioral prediction in turn.
One approach that suggests a distinction between ingroup positivity and outgroup negativity, as depicted in the x and y-axis in Figure 1, comes from the intergroup relations tradition (Allport, 1954; Brewer, 1999). For example, in the context of African-American group identity, Herring et al. (1999) find that “pro-Black doesn’t mean anti-White”; whereas African-American group identity is associated with a sense of shared fate (positive ingroup sentiment), it is not necessarily associated with negative attitudes toward Whites (negative outgroup sentiment). This suggests that ingroup and outgroup affect may be distinct, an approach sometimes taken in the partisan affect literature.
Negative Partisanship
However, not all work acknowledging two separate dimensions of partisan affect equally emphasizes inparty affinity and outparty animosity. Given recent declines in inparty affinity (e.g., Groenendyk et al., 2020) and the finding that changes in affective polarization have been primarily driven by outparty animosity (Iyengar et al., 2012), some work has redefined the term “affective polarization” to focus primarily on negative affect, as seen on the x-axis in Figure 1. For example, “This phenomenon of animosity between the parties is known as affective polarization.” (Iyengar et al., 2019, p. 130). This outparty-focused conceptualization of affective polarization overlaps with another body of literature, “negative partisanship.” Negative partisanship argues that as socio-economic and cultural divisions align, one’s outparty is seen as increasingly alien, resulting in stronger negative feelings toward that outparty (Abramowitz & McCoy, 2019; Abramowitz & Webster, 2018; Bankert, 2021).
The importance of outparty animosity has also been justified using the predictive validity of various outcomes. For instance, Nelson (2022) finds that greater affective polarization—in terms of a greater number of negative emotions experienced toward the outparty—is associated with worse self-reported health, though these negative health consequences may be somewhat buffered by the increased political engagement associated with negative partisan affect. Work in the negative partisanship tradition has also noted that as outparty animosity has increased, so too has party loyalty in vote choice, as Americans are increasingly unwilling to vote for the opposing party despite intra-party disagreement (Abramowitz & McCoy, 2019; Abramowitz & Webster, 2018). Ultimately, these approaches have emphasized the role and predictive power of negative outparty affect in particular, while deemphasizing the role of inparty affect.
Models That Emphasize Inparty Affinity
However, other approaches have also considered the contributions of inparty affect. Expressive partisanship, for instance, focuses on partisanship as a social identity along with the emotional attachments one has to their party (Huddy et al., 2015), thus more explicitly addressing the role of (positive) inparty feelings (e.g., Huddy et al., 2018).1
Like outparty animosity, inparty affinity predicts outcomes such as vote choice and turnout. Positive affect toward a candidate, in the form of hope, is a key motivator for choosing that candidate over an opponent (Just et al., 2007). Furthermore, there is a long-standing connection between liking one’s party, partisanship, and voting. In their seminal work on voting, Campbell et al. (1980) discussed the effects of affect toward parties, candidates, and political events on voter turnout, noting that “Orientations to these objects, seen by the voter as positive or negative, comprise a system of partisan attitudes that is of primary importance for the voting act” (p. 66). Some aggressive forms of inparty affinity, such as collective narcissism—an emotional commitment to the exceptionalism of the ingroup and the idealization of its qualities (Cichocka, 2016; Golec de Zavala et al., 2009, 2013)—have also been associated with behavior. Although also linked with outparty animosity, collective narcissism has been associated with support for populist candidates who promote ingroup exceptionalism, such as Donald Trump (Federico & Golec de Zavala, 2018). This suggests the utility of considering inparty affinity alongside outparty animosity when predicting political behavior.
Affective Polarization
Finally, one recent theoretical approach—“affective polarization” – is a more uni-dimensional model of partisan affect, akin the dotted diagonal line in Figure 1. This approach focuses on those who “view opposing partisans negatively and copartisans positively” (Iyengar & Westwood, 2015, p. 691), and is often measured using the difference between inparty and outparty feeling thermometers. This combination, which we refer to as “classical” affective polarization, attempts to intuitively capture and describe the growing affective distance between partisans in the United States (e.g., Iyengar et al., 2012). Research on affective polarization in this form has primarily focused on its causes and remedies (Iyengar et al., 2019). For example, a high-choice media environment has been associated with greater affective polarization (Lelkes et al., 2017), and correcting partisan stereotypes has been found to reduce it (Ahler & Sood, 2018). However, the theoretical premise of classical affective polarization implies that the combination of high inparty affinity and high outparty animosity may be part of what drives political behavior, as well.
Empirical Approaches to Partisan Affect
Many of the above approaches assume a-priori that partisan affect consists of either two dimensions, in the case of the intergroup relations approaches, or one dimension, in the case of the classical model of affective polarization. Empirical methods—such as examining the correlations between inparty affect and outparty affect, or factor analyses, which provide comparative fit statistics—can test which of these two broad conceptual models best fits the partisan affect data.
However, the results of such analyses have been somewhat inconsistent. For example, work factor-analyzing ten items from the 1972 and 1976 National Election Studies suggested the presence of one dimension underlying partisan affect (Green, 1988). This single dimension reflected the co-occurrence of positive feelings toward the inparty and negative feelings toward the outparty, consistent with classical approaches to affective polarization. By contrast, other work showed a relatively low correlation between feeling thermometer ratings toward Democrats and Republicans (Weisberg, 1980) and conservatives and liberals (Conover & Feldman, 1981), suggesting the presence of two different dimensions of partisan affect. Trends in inparty affinity and outparty animosity are also asynchronous (Finkel et al., 2020; Iyengar et al., 2012; Iyengar & Krupenkin, 2018), suggesting the two dimensions change separtely from one another, further implying distinctions between them.
Although studies from the 1980s suggested a unidimensional model of partisan affect, it is unclear whether this is still the case. Whereas outparty feelings have become cooler in recent years, so have inparty feelings (Groenendyk, 2018), meaning that partisan affect may no longer be unidimensional in today’s era of greater polarization. Accordingly, it is unclear whether the results from Green (1988) still hold today, including with more contemporary factor-analytic methods (e.g., Wirth & Edwards, 2007)—which allow for the inclusion of multiple operationalizations of partisan affect, including across partisan targets such as the mass public and partisan elites (e.g., Druckman & Levendusky, 2019). Furthermore, even among models that acknowledge two dimensions of partisan affect, it is unclear whether models that emphasize only outparty animosity, or inparty affinity and outparty animosity alike (either in combination, or separately), do a better job predicting political behavior than others.
The Present Study
The present study empirically tests which conceptualization of partisan affect best characterizes partisan feelings in two ways. First, it uses contemporary categorical factor analyses (Wirth & Edwards, 2007) to appropriately examine whether a one or two-dimensional conceptual model better characterizes how American partisans respond to contemporary set of partisan measures (Study 1A), using two waves of the Pew Research Center’s 2016 American Trends Panel (hereafter, ATP; see online Appendix B for details). This provides a methodological and temporal update to previous work (e.g., Green, 1988). Second, it examines which theoretical model provides the best predictive validity for different forms of political behavior (Study 1B). This is particularly important given a recent call for more research investigating the political consequences of partisan affect (Iyengar et al., 2019) and the fact that different theories imply that different aspects of partisan affect— either outparty animosity (e.g., Abramowitz & Webster, 2016), inparty affinity and outparty animosity separately (e.g., Huddy et al., 2018; Valentino et al., 2011), or the interaction between inparty affinity and outparty animosity (e.g., Iyengar & Westwood, 2015)— should matter most for predicting different types of political behavior.
Specifically, we had five hypotheses across Study 1A and 1B. We had two competing hypotheses for Study 1A, concerning the structure of partisan affect. Hypothesis 1A-1 predicted that partisan affect would be one-dimensional, consistent with the classical affective polarization approach. By contrast, Hypothesis 1A-2 predicted that partisan affect would be two-dimensional, consistent with intergroup relations approaches. We had three competing hypotheses for Study 1B, concerning the predictive validity of partisan affect. Hypothesis 1B-1 predicted that outparty animosity (but not inparty affinity) would predict most forms of political behavior, consistent with negative partisanship approaches to partisan affect. Hypothesis 1B-2 predicted that outparty animosity would predict higher-cost forms of political behavior, and inparty affinity would predict lower-cost forms of political behavior, consistent with approaches highlighting inparty affinity and with other work on emotions in politics. Hypothesis 1B-3 predicted that the combination/interaction of outparty animosity and inparty affinity would predict most forms of political behavior, consistent with classical affective polarization approaches to partisan affect.
Study 1A: Empirically Investigating the Dimensionality of Partisan Affect Across Multiple Measures
Study 1A investigated how many dimensions characterize partisan affect, with special interest in whether partisan affect has (Hypothesis 1A-1) one dimension (e.g., Green, 1988; Iyengar & Westwood, 2015) or (Hypothesis 1A-2) two dimensions, dividing between inparty and outparty feelings (e.g., Allport, 1954; Brewer, 1999; Herring et al., 1999). Factor analyses are an ideal method for this investigation because they statistically examine the latent structure—in terms of dimensions or factors— underlying the responses to a particular set of items (Wirth & Edwards, 2007). The quantity and content of those factors are used to empirically determine the number and nature of the dimensions underlying a given construct. Whereas exploratory factor analyses (EFAs) help determine how many models should be considered, confirmatory factor analyses (CFAs) provide formal fit statistics to compare models.
Given the various response types in the present data (including binary items), categorical factor analyses were used. This method was chosen because it can handle binary and ordinal (e.g., 5-point scales) data in the same model without risking the attenuation that can occur when incorrectly estimating ordinal responses as continuous (Byrne, 2005). This provides an advantage to conventional forms of factor analysis, which are best suited to continuous items (e.g., Wirth & Edwards, 2007). To the authors' knowledge, the present work represents the first application of categorical factor analyses to partisan affect data.
Analyses proceeded in two parts. First, exploratory analyses were used to identify which models were plausible candidates for the dimensions underlying partisan affect in a data-driven way. Then, confirmatory factor analyses were used to examine which candidate models best fit the data. Consistent with best practices in examining and confirming factor structures, the total sample was split in half, with one half used for exploratory analyses (for space reasons, reported in online Appendix C) and the other half for confirmatory analyses (reported in the main text). Splitting helped ensure models were not merely overfitted to the same data.
Method
Participants
A total of 846 participants had complete responses on all items of interest, and were split into exploratory and confirmatory datasets for a robustness check. The analyses reported here use the confirmatory dataset (n = 423), which is similar to the exploratory dataset used in online Appendix C. Of the respondents in the confirmatory dataset, 51.1% reported being female and ranged in age (reported in categories) from 18-29 to 65+, with a modal age category of 50-64 years old. Approximately 77.8% of respondents reported being White; 7.2% African-American; 8.6% Hispanic; and 6.4% Other. Respondents ranged in income (reported in categories) from <$10,000 to $150,000+, with a modal income category of $50,000 to under $75,000. Regarding political affiliation, 48.2% of respondents identified as Democrats/Democrat-leaners, and 51.8% identified as Republicans/Republican-leaners.
Measures
A pooled cross-sectional dataset from Waves 14 and 23 of the ATP (collected during April and November of 2016, respectively) was constructed, containing 19 measures of political affect. These measures covered item types commonly used in the literature (see online Appendix A for a detailed list of extant measures). They reflected positive and negative emotions, inparty and outparty targets, and elite and civilian targets (see Table 1 for details), including feeling thermometers about candidates and parties (e.g., Iyengar et al., 2012; Mason, 2018), reactions to ingroup and outgroup partisans moving next door (e.g., Garrett et al., 2014; Levendusky & Malhotra, 2016; Luttig, 2018), and emotions (i.e., anger, fear, hope, frustration, enthusiasm, and pride) felt toward the parties (e.g., Klar et al., 2018; Mason, 2016). More information about item wording and inclusion decisions is available in online Appendix B.
Table 1
American Trends Panel Partisan Affect Questions
| Question Types | Question Description(s) | Response Formats | Time of Measurement |
|---|---|---|---|
| Social Distance | Preference/Disdain for Inter-Party Neighbors | 5-point scales (Harder to live near in/outparty – Easier) | April 2016 |
| Feeling Thermometers | Party Feeling Thermometers | 0-100 scales (Cooler toward in/outparty – Warmer) | November 2016 |
| Candidate Feeling Thermometers (Clinton, Trump, Cruz, and Sanders) |
April 2016 | ||
| Affective Response | Angry, Afraid, Hopeful, Proud, and Enthusiastic, Frustrated about Parties | Binary responses (Yes/No) | April 2016 |
To ensure adequate response variability for analyses across the diverse item types, binary emotion items with less than a 10% endorsement rate were excluded from analyses. Thus, “enthusiastic,” “proud,” and “hopeful” toward the outparty and “afraid” of the inparty were excluded (but see online Appendix D for a replication of the reported results with all affect items included).
Because the selected items (see Table 2) refer to either Democrats or Republicans, each item was also re-coded to reflect respondents’ reactions to either their inparty or outparty. Items about Democrats were coded as “inparty” measures for Democrats/Democrat-leaners and as “outparty” measures for Republicans/Republican-leaners. In contrast, items about Republicans were coded as “inparty” measures for Republicans/Republican-leaners and as “outparty” measures for Democrats/Democrat-leaners.
For the candidate feeling thermometers, comparisons between multiple candidates (across Senator Ted Cruz, Donald Trump, Senator Bernie Sanders, and Secretary Hillary Clinton) were used due to the April 2016 timing of data collection. At that point, the 2016 presidential nominee was not yet decided. Thus, to ensure the broadest possible assortment of elites among items, each available comparison of Republican and Democrat candidates was included.2 Moreover, for each comparison, feelings toward one of the candidates were coded to reflect participants’ evaluation of their inparty candidate; feelings toward the other were coded to reflect participants’ evaluation of their outparty candidate.
Analysis
Loading patterns for the one and two-factor solutions were entered as specified in the EFAs (see online Appendix C) to compare their fit. To identify the models and standardize solutions (thus aiding with interpretability), the variance of each latent factor was set to one, and the first item on each factor could load freely. Categorical CFAs were conducted using WLSMV estimation in MPlus version 8 (Muthén & Muthén, 2017), permitting the specification of continuous (e.g., feeling thermometer) variables in the same model.3
Results
Consistent with the results of the EFAs in online Appendix C, the one-factor solution reflected more ease of getting along with inparty neighbors and difficulty getting along with outparty neighbors, positive inparty candidate evaluations, negative outparty candidate evaluations, inparty hope and enthusiasm, and more outparty anger and fear.
The two-factor solution reflected one dimension which accounted for more ease of getting along with inparty neighbors, difficulty getting along with outparty neighbors, unfavorable evaluations of outparty candidates, and more outparty anger and fear. It also included a second dimension which accounted primarily for more favorable evaluations of inparty candidates, as well as inparty hope and enthusiasm. Finally, the correlation between the two factors was only moderate—r(421) = .448, p < .001 (see Figure 2), consistent with the low correlations in the EFAs (see online Appendix C).
Figure 2
Two-Factor CFA Path Diagram
Note. Factor one generally reflects outparty animosity, whereas factor two reflects inparty affinity. High factor loadings are not unexpected for items with more response options, given the variability in item response formats.
In terms of fit, as reported in Table 2, the one-factor solution had an RMSEA of 0.106 (90% confidence interval: 0.099, 0.113), suggesting that the one-factor model did not do a good job accounting for the data, as the cutoff for adequate RMSEA is < 0.10 (Browne & Cudeck, 1993).
Table 2
Fit Indices for Competing CFA Models in Confirmatory Dataset
| Model | RMSEA | RMSEA 90% CI | CFI | TLI | WRMR |
|---|---|---|---|---|---|
| One-Factor Model | 0.106 | [0.099, 0.113] | 0.599 | 0.549 | 2.182 |
| Two-Factor Model | 0.084 | [0.077, 0.091] | 0.749 | 0.716 | 1.738 |
Note. RMSEA = root square error of approximation (lower values are better); 90% CI = 90% confidence interval; CFI = comparative fit index (lower values are better); TLI = Tucker Lewis Index (higher values are better); WRMR = weighted root mean square residual, an approximation of standardized root mean square residual (lower values are better).
The two-factor model, by contrast, met this criterion, with an RMSEA of 0.084 (90% confidence interval: 0.077, 0.091), suggesting the two-factor model fit adequately, particularly given the diversity of item response types. Notably, these two solutions had non-overlapping 90% RMSEA confidence intervals, suggesting the two-factor solution fit significantly better than the one-factor solution. In addition, the WRMR—a badness of fit index—was lower, and the TLI and CFI—which are goodness of fit indexes (West et al., 2012)—were higher in the two-factor solution than the one-factor solution.4
Discussion
Study 1A suggested that a two-factor model, reflecting a distinction between outparty animosity and inparty affinity, fit the confirmatory sample better than a one-factor model collapsing across the two dimensions. The correlation between the two factors was relatively low, suggesting that only about 20% of the variance in outparty animosity and inparty affinity was shared in this split-half of the data, corroborating the results of the EFAs reported in online Appendix C.
This suggests support for Hypothesis 1A-2: a two-factor model of American partisan affect was present across two samples and many contemporary item types, helping to bring cohesion across the literature. This model divides between outparty animosity on the one hand and inparty affinity on the other. It is most consistent with prior work on intergroup relations suggesting a distinction between ingroup love and outgroup hate (e.g., Allport, 1954; Brewer, 1999; Herring et al., 1999). Notably, the data included positive and negative evaluations about the ingroup and the outgroup, as well as ratings of elites and ordinary citizens. Thus, it was well-positioned to adjudicate between many theoretical possibilities, such as those that imply that (Hypothesis 1A-1) partisan affect is collapsible across one dimension (e.g., Green, 1988, Iyengar & Westwood, 2015), or those that divide even further by target type (e.g., elites versus the mass public; Druckman & Levendusky, 2019). However, those other models were not as interpretable or appropriate for the data relative to the two-factor division between outparty animosity and inparty affinity (see online Appendix C).
Study 1B: Partisan Affect and Participation
Study 1A provided robust empirical support for conceptualizing partisan affect using two dimensions—inparty affinity and outparty animosity. However, even among approaches that separate between inparty affinity and outparty animosity, not all equally consider both in terms of predictive validity.
For example, work in the negative partisanship tradition suggests that inparty vote choice is primarily driven by the desire to prevent the outparty from gaining power (Abramowitz & Webster, 2016). Thus, a negative partisanship approach might expect that (Hypothesis 1B-1) outparty animosity, but not inparty affinity, will significantly predict most forms of political participation since such participation might aid in outparty losses.
By contrast, more inparty-affinity focused perspectives, such as collective narcissism, imply that inparty affinity could also motivate a desire to participate (and win ingroup power) in order to demonstrate ingroup exceptionalism (Federico & Golec de Zavala, 2018), suggesting that both outparty animosity and inparty affinity may predict political behavior. But, in what ways? One possibility has to do with the costliness of behavior.
Prior work has shown that certain negative emotions, such as anger, are associated with high approach motivation, whereas certain positive emotions, such as happiness, are associated with lower approach motivation (Harmon-Jones, 2019). This suggests that emotions like anger might lead people to seek out behaviors more strongly than emotions like happiness. Some work has also investigated the relationship between emotions and political participation specifically (e.g., Valentino et al., 2009), finding that while anger, fear, and enthusiasm all seem to motivate less costly behaviors, such as wearing a button; only enthusiasm, and most powerfully, anger, seem to motivate costlier behaviors, such as working for a campaign or attending a rally (Valentino et al., 2011). This may imply that (Hypothesis 1B-2) outparty animosity predicts costlier forms of participation in terms of time, money, and effort. On the other hand, inparty affinity might be associated with less costly forms of participation, as many positive emotions are associated with weaker approach motivations (e.g., Harmon-Jones, 2019). Accordingly, different types of political participation, which varied in costliness, such as attending a rally and wearing a button, needed to be investigated separately in the present analyses.
Finally, unidimensional approaches to partisan affect, such as classical approaches to affective polarization, which focus on the difference between inparty and outparty affect, were not found to be empirically preferable in Study 1A. However, they may still imply that (Hypothesis 1B-3) the combination of high inparty affinity and low outparty animosity guides political participation. This approach might predict that the interaction between inparty affinity and outparty animosity drives most forms of political participation.
Method
Participants
Study 1B again used the Pew ATP data to test the three competing hypotheses listed above. However, to maximize sample size, the sample was no longer split in two. The sample was also restricted to registered voters to ensure all respondents were equally able to participate in all behaviors (n = 761). For space reasons, see demographic information in online Appendix F.
Measures
Study 1B used the same affect measures as Study 1A, collected in April 2016, except for party feeling thermometers, which were collected starting in November 2016, after the election. However, Study 1B added participation measures from September 2016, collected after most of the affect measures. These included all of the relevant political participation measures available in the ATP data, ordered here from most to least frequent (see online Appendix G for frequencies): displaying campaign material such as a bumper sticker, posting political material online, contacting an elected official, attending a rally, contributing money to a campaign, being a member of a lobbying organization, volunteering for a campaign, and attending a protest.
Contrary to Iyengar and Krupenkin (2018), who operationalize participation using a single index measure,5 we analyzed each form of participation separately to investigate qualitative differences in the types of behavior predicted by each affective dimension. We did this to examine potential differences in the costliness of behavior, as has been examined elsewhere (Valentino et al., 2011). To approximate the costliness of the eight forms of participation listed above, we ranked them from most common (displaying campaign material) to least common (attending a protest).6
Analysis
Eight separate logistic regressions examined whether and how inparty affinity and outparty animosity predicted distinct forms of political participation. These regressions were arranged from least costly (displaying campaign material) to most costly (attending a protest) political behaviors. The outcome variable in each logistic regression indicated whether respondents had performed the named political behavior in the last year (see exact wording in online Appendix B).
For the focal independent variables of inparty affinity and outparty animosity, two z-scored means were computed—one z-scored mean reflected the items that went into the outparty animosity factor (Factor 1), and another z-scored mean reflected the items that went into the inparty affinity factor (Factor 2). Z-scores were used because the items that comprised outparty animosity and inparty affinity varied in response types (binary, 5-point scales, and feeling thermometers), making a mean score inappropriate. Including these z-scores allowed for comparing the relative impact of outparty animosity and inparty on each of the eight political behaviors.
An interaction between the inparty affinity and outparty animosity z-scores was also included in each regression to capture whether the influence of one of these variables was contingent on the level of the other.7 Such a conditional relationship between inparty affinity and outparty animosity would be consistent with Hypothesis 1B-3, which contends that combining inparty affinity and outparty animosity should predict most forms of political behavior. It would also be consistent with “classical” approaches to affective polarization, which often use a difference score (e.g., Hutchens et al., 2019; Lelkes et al., 2017) to model that relationship. Here, we instead used an interaction for two reasons. First, because z-scores do not reflect mean values, the subtraction of two z-scored variables is not as easily interpretable as a mean difference score. Second, including an interaction term in a model that includes inparty affinity and outparty animosity permits follow-up analyses examining whether particular levels of inparty affinity and outparty animosity drive any significant effects. Unlike a difference score—in which a feeling thermometer score of 10 could be equally obtained by a partisan who scores their inparty 10 and their outparty 0, or by a partisan who scores their inparty 100 and their outparty as 90—a significant interaction would allow for follow-up tests investigating whether a conditional relationship between inparty affinity and outparty animosity matters in one of these cases, but not the other.
Relevant demographic control variables were also included, measuring participants’ age, education level,8 sex, and race. Because we did not include independents in analyses, we also included a 4-point measure of political partisanship (1 = Democrat, 2 = Democrat-Leaner, 3 = Republican-Leaner, and 4 = Republican)9, with “Democrat” serving as the reference category.
Results
Table 3 displays the results of eight logistic regressions for the self-reported forms of political participation. To account for inflated familywise error rates due to multiple simultaneous tests, target p values were adjusted using Bonferroni corrections. These corrections were made because, without them, there was a 33.7% chance of committing a Type I error given eight regressions (Cabin & Mitchell, 2000). Consistent with standard practices, Table 3 also indicates the results without the corrections.
Table 3
Logistic Regressions Predicting Political Participation From Inparty Affinity and Outparty Animosity
| Variable | (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) |
|---|---|---|---|---|---|---|---|---|
| Displayed Campaign Material | Posted Material Online | Contacted an Official | Attended Rally | Contributed Money | Lobbying Org. Member | Campaign Volunteer | Attended Protest | |
| Outparty Animosity | 1.211 | 1.472 | 1.306 | 1.397 | 1.619 | 1.145 | 0.998 | 1.233 |
| (0.050) | (0.000) | (0.004) | (0.004) | (0.000) | (0.248) | (0.989) | (0.251) | |
| Inparty Affinity | 1.468 | 1.836 | 1.072 | 1.081 | 1.272 | 1.154 | 1.216 | 1.118 |
| (0.001) | (0.000) | (0.494) | (0.533) | (0.031) | (0.274) | (0.218) | (0.574) | |
| Outparty Animosity x Inparty Affinity | 0.961 | 1.073 | 0.969 | 1.020 | 1.165 | 1.076 | 1.008 | 0.983 |
| (0.675) | (0.446) | (0.709) | (0.851) | (0.116) | (0.482) | (0.950) | (0.914) | |
| Partisanship | ||||||||
| Dem-Leaner | 0.822 | 1.827 | 1.088 | 1.436 | 1.024 | 1.086 | 0.491 | 2.046 |
| (0.501) | (0.021) | (0.746) | (0.214) | (0.932) | (0.798) | (0.097) | (0.076) | |
| Rep-Leaner | 1.138 | 2.313 | 1.641 | 0.906 | 0.772 | 0.750 | 0.404 | 0.749 |
| (0.666) | (0.003) | (0.069) | (0.773) | (0.378) | (0.435) | (0.054) | (0.593) | |
| Republican | 0.978 | 1.391 | 0.644 | 0.661 | 0.554 | 0.821 | 0.572 | 0.317 |
| (0.923) | (0.138) | (0.053) | (0.134) | (0.012) | (0.478) | (0.092) | (0.033) | |
| Age | 1.302 | 0.949 | 1.257 | 1.056 | 1.729 | 1.174 | 1.484 | 1.364 |
| (0.004) | (0.525) | (0.006) | (0.583) | (0.000) | (0.131) | (0.003) | (0.058) | |
| Education | 0.978 | 0.985 | 0.879 | 0.805 | 1.093 | 0.791 | 0.877 | 1.060 |
| (0.801) | (0.857) | (0.112) | (0.027) | (0.299) | (0.024) | (0.287) | (0.706) | |
| Sex (Male) | 1.042 | 0.896 | 1.053 | 0.969 | 1.191 | 1.039 | 0.863 | 1.071 |
| (0.640) | (0.182) | (0.526) | (0.750) | (0.044) | (0.713) | (0.238) | (0.654) | |
| Race (White) | 1.027 | 0.993 | 1.206 | 0.894 | 1.185 | 1.151 | 0.882 | 0.846 |
| (0.777) | (0.932) | (0.038) | (0.247) | (0.076) | (0.230) | (0.300) | (0.240) | |
| Constant | 0.316 | 0.317 | 0.443 | 0.214 | 0.461 | 0.185 | 0.151 | 0.0702 |
| (0.000) | (0.000) | (0.000) | (0.000) | (0.000) | (0.000) | (0.000) | (0.000) | |
| Observations | 761 | 761 | 761 | 761 | 761 | 761 | 761 | 761 |
| Pseudo R2 | 0.046 | 0.075 | 0.046 | 0.034 | 0.129 | 0.028 | 0.057 | 0.063 |
Note. Odds ratio coefficients; p-values in parentheses. Item wording is available in online Appendix B. Significant effects (p < .05) that survive the Bonferroni correction (p < .00625) are in bold, significant effects that do not are in italics.
Most of the eight forms of political participation displayed in Table 3 were predicted either by inparty affinity, outparty animosity, or both. Outparty animosity was a significant and unique predictor of increased odds of later rally attendance (OR = 1.397, p = .004), contacting an elected official (OR = 1.306, p = .004), and contributing money (OR = 1.619, p < .001).10 By contrast, inparty affinity was a significant and unique predictor of increased odds of displaying campaign material (OR = 1.468, p = .001).
Neither inparty affinity nor outparty animosity significantly predicted protest attendance, campaign volunteering, or membership in a lobbying organization. Only one form of political participation, posting about politics on social media, was significantly predicted by both outparty animosity (OR = 1.472, p < .001) and inparty affinity (OR = 1.836, p < .001). However, a Wald test comparing these coefficients showed that they were not significantly similar (χ2 = 2.090, p = 0.148); inparty affinity was a significantly stronger predictor of posting political material on social media.
Finally, the interaction term for outparty animosity and inparty affinity did not reach significance in any regression models. Thus, the effects of inparty affinity and outparty animosity were not significantly conditional upon one other when predicting different forms of political participation.
Discussion
Study 1B examined whether and how inparty affinity and outparty animosity relate to political participation to help adjudicate between different theoretical models of partisan affect. Consistent with Hypothesis 1B-2, outparty animosity uniquely predicted forms of political participation that were costlier in terms of time/effort (i.e., less common—contacting an elected official, attending a political rally, and contributing money to a campaign). Inparty affinity was uniquely related to a less costly behavior (displaying campaign material). Posting political material online was predicted by both inparty affinity and outparty animosity. However, inparty affinity was a stronger predictor of this relatively lower-cost behavior. Moreover, additional analyses in online Appendix L suggested that people with high inparty affinity were likelier to have more politically like-minded social media networks, whereas people with high outparty animosity were not. Thus, social media posting might also be lower-cost for people with high inparty affinity, as their posts may be more ingroup-oriented—and thus potentially less likely to receive backlash—relative to posts from people with high outparty animosity.
Moreover, outparty animosity uniquely predicted more forms of participation than inparty affinity, suggesting that outparty animosity may be more politically motivating. This is partially consistent with Hypothesis 1B-1 and recent work highlighting the primacy of outparty animosity (e.g., Abramowitz & Webster, 2016; Finkel et al., 2020) and anger’s strength in motivating costly forms of political participation (Valentino et al., 2011). However, the rarest behaviors in our sample (lobbying membership, campaign volunteering, and protest attendance) were not predicted by either dimension of partisan affect. Affect may not be sufficient to overcome the cost of those behaviors.
Interestingly, Hypothesis 1B-3, which anticipated that the interaction between inparty affinity and outparty animosity would predict most forms of participation—was not supported for any of the eight participation outcomes. Despite a thriving body of work on classical approaches to affective polarization, the present work suggests the effects of inparty affinity and outparty animosity on political participation may be additive rather than conditional upon one another.
In short, the distinct patterns of results for inparty affinity and outparty animosity suggest that each is separately important for predicting qualitatively different behaviors—providing further support for intergroup relations models that divide between ingroup affinity and outgroup animosity, and for inparty-focused and outparty-focused accounts of partisan affect. Separating between inparty affinity and outparty animosity thus adds value to our understanding of partisan affect’s political consequences.
General Discussion
This paper addressed two gaps in the literature on partisan affect. The first relates to measurement and conceptualization, and the latter, to the predictive validity of different aspects of partisan affect for political participation. Both of these helped adjudicate between different theories of partisan affect. In Study 1A, consistent with intergroup relations approaches (e.g., Allport, 1954; Brewer, 1999; Herring et al., 1999) and Hypothesis 1A-2, we found that across two samples and multiple measures of partisan affect, two separate dimensions—inparty affinity and outparty animosity—better characterized contemporary American partisan affect than a single dimension which collapsed across the two (Hypothesis 1A-1). Moreover, these two dimensions were not strongly correlated, suggesting that people who dislike their outparty do not always like their inparty, and vice-versa. The distinction between these two dimensions also bore out regarding their consequences for political participation in Study 1B.
From a theoretical perspective, this suggests that American partisan affect is no longer unidimensional, as was suggested by work in the 1980s (Green, 1988). It is unclear whether this change is due to updates in analytical approaches (e.g., the use of categorical analyses across a more extensive set of measures), changes in how people feel about the political parties, or both. There is reason to believe that the structure of partisan affect itself may have changed (e.g., Groenendyk, 2018; Iyengar & Krupenkin, 2018). Some work suggests that outparty feelings have declined alongside inparty feelings (Groenendyk, 2018), suggesting that outparty animosity does not always pair with inparty affinity. Regardless of the reason, however, the present work suggests that at least now and in the United States, outparty animosity and inparty affinity are relatively distinct.
From a practical measurement perspective, this implies that cumulative scales— such as difference scores between inparty affinity and outparty—may fail to fully capture American partisan affect. Different people may have high inparty affinity versus high outparty animosity, suggesting the literature would benefit from including separate measures of the two.
Regarding the second gap, and answering a call for work examining the political consequences of partisan affect (Iyengar et al., 2019), the two dimensions predicted qualitatively different forms of political participation. This also has ramifications for theories of partisan affect. Consistent with Hypothesis 1B-1—and negative partisanship approaches to partisan affect— outparty animosity uniquely predicted more medium-cost forms of participation. However, more consistent with Hypothesis 1B-2 and work highlighting the role of certain positive emotions in predicting lower-cost behaviors (e.g., Harmon-Jones, 2019; Valentino et al., 2011), inparty affinity also predicted fewer, lower-cost forms of political participation. Another interpretation of these results is that inparty affinity was associated with the behaviors participants felt more efficacious about, given that the behaviors were more common. Future work should aim to tease these explanations further, though in either case, the present work suggests the utility of separating the two dimensions of partisan affect.
Interestingly, and in contrast to predictions from the large literature on classical affective polarization and Hypothesis 1B-3, in no case was the interaction between outparty animosity and inparty affinity a significant predictor of participation; a result that was also consistent with the results of Study 1A. This further suggests that the contributions of inparty affinity and outparty animosity should be considered separately rather than in combination.
Strengths and Limitations
One strength of the present work is the use of multiple operationalizations of partisan affect. To our knowledge, this investigation is the first to use categorical factor analyses to examine the structure of partisan affect across such a diverse set of contemporary measures. Another strength is the use of various political participation measures, which allowed us to address a call for further research on the political consequences of partisan affect (Iyengar et al., 2019), and examine whether different aspects of partisan affect were associated with distinct political behaviors.
However, the present work also has some limitations. For example, the diversity of measures may have made the data noisier than if a more consistent item format had been used. It is notable that, despite this, a coherent two-dimensional structure of partisan affect still emerged. Moreover, like much of partisan affect literature, this paper does not address the affect of independents, who do not have a readily identifiable inparty and outparty (e.g., Klar & Krupnikov, 2015). Finally, the data used here focuses on self-report data from one dataset at one time point, potentially during a particularly contentious election cycle (2016). But, prior research suggests that the relationship between inparty and outparty feelings may change over time (Groenendyk, 2018; Iyengar & Krupenkin, 2018). Thus, future research may benefit from developing a more standardized measure of inparty affinity and outparty animosity and investigating whether partisan affect’s structure and behavioral consequences look the same over time, including in terms of more direct behavioral measures.
Conclusion
Partisan affect plays a vital role in current events. Thus, we should be concerned with accurately understanding the concept and its political consequences. Using a variety of measures in a nationally presentative sample, we found that two dimensions best characterize American partisan affect: inparty affinity and outparty animosity, each of which predicts distinct aspects of political participation, in ways consistent with both negative partisanship and more inparty-affinity-centric theoretical approaches to partisan affect. We believe this work suggests the promise of continuing to examine partisan affect in a nuanced manner.