Ticking all the boxes? A comparative study of social sorting and affective polarization

In the United States, the growing hostility between supporters of the Democrats and Republicans is wellestablished. Research shows that such ‘affective polarization’ exists in other countries too, but to varying degrees. What makes compatriots with different political preferences opponents in some contexts and enemies in others? This manuscript shows, using two complementary studies, that a contributing factor is the alignment of political and non-political identities. First, I employ CSES data to predict the level of affective polarization by such ‘social sorting’ at 119 elections in 40 countries, showing that greater overlap of partisan divisions with non-political divisions in a society (along lines of income, education, religion and region) is associated with more dislike of political outgroups. Second, using Dutch panel data I show that individuals who fit the socio-demographic ‘profile’ of their party better tend to be more affectively polarized. This has important implications for our understanding of affective polarization.

Mutual dislike across political camps is not just a 'sign of the times', but rather something that thrives under certain conditions but not others. This study aims to contribute to our understanding of these conditions by investigating a factor that might explain (macro-and microlevel) variation in affective polarization: the alignment of political and non-political divisions in society ('social sorting').
It is a classic insight that cross-cutting cleavages decrease social tensions (Lipset 1960). Mason (2015;2016; demonstrates that Americans with aligned religious, racial and partisan identities -for instance, Christian and non-White Republicans -are more antagonistic towards partisan outgroups, regardless of the extremity of their views. The contribution of the present study is to test whether this phenomenon travels to a broader set of cases. The hypothesis is that if politically like-minded individuals in any society also share other non-political identities -i.e., if they are socially sorted along political lines -they will be less tolerant towards those with opposing views. As will be discussed below, this mechanism builds on insights from Social Identity Theory and the sociological literature on cleavages that are not likely to be restricted to the United States.
Different societies experience different non-political divisions: between religious groups, between city dwellers and rural inhabitants, between the poorest and the middle classes, etcetera.
Regardless of the nature of the divisions involved, if they align with political divisions, antagonism between political camps can be expected to be amplified. Political 'ingroups' and 'outgroups' will appear even more distinct and homogeneous, fostering negative outgroup affect and behavior (Brewer and Roccas 2002). I investigate the relation between social sorting and affective polarization comparatively using two complementary studies, one on the aggregate and one on the individual level. This should confirm whether the relation found in the United States indeed travels to other contexts, and hence shed light on the causes (and prospects) of affective polarization around the globe. First, I analyze CSES data on 40 countries between 1996 and 2018 to investigate the relation between the alignment of political and non-political identities (on the one hand) and affective polarization (on the other). A within-country analysis shows that in periods in which political divisions correlate more strongly with non-political ones, affective polarization tends to be higher too. To further gauge the likelihood that such macro patterns are brought about by the hypothesized relations on the individual level, I supplement this with a study of individual-level panel data from the Netherlands (2007-2018; N=8205) that allows me to track the extent to which a respondent fits the socio-demographic profile of their party. This shows that individuals who become more strongly 'sorted' in this way are also more likely to become more affectively polarized. The Conclusion section discusses the implications for our understanding of affective polarization.

Theory Politics as a (divisive) social identity
Our political stances -whether it is a partisan identity such as being a Social Democratic supporter, an ideological identity such as being conservative (Mason 2018), or even an issue identity such as being a 'Remainer' (Hobolt et al. 2019) -can constitute a social identity. A social identity is "that part of an individual's self-concept which derives from his knowledge of his membership of a group […] together with the value and emotional significance attached to the membership" (Tajfel 1979). Depending on intergroup dynamics, this can also involve varying degrees of negative bias towards the out-group (Brewer 1999). The resulting 'affect gap' towards the political outgroup is now commonly called affective polarization (Iyengar et al. 2012). 1 Research in the United States has a longstanding tradition of conceiving political identities predominantly along partisanship lines (Campbell et al. 1960). Iyengar & Westwood (2015) define affective polarization in the American context as "the tendency of people identifying as Republicans or Democrats to view opposing partisans negatively and copartisans positively". They demonstrate that this 'affect gap' has grown steadily. 'Feeling thermometers' testify that Americans feel increasingly colder towards adherents of the opposite party. They are also increasingly unwilling to have social interactions with outpartisans -most vividly illustrated by the steep increase in the percentage of Americans that would dislike their son or daughter marrying an outpartisan, which increased from 4-5% in 1960 to 33% in 2010 (Iyengar et al. 2012) and to 45% in 2019 (Jones & Naijle 2019). These patterns are more than mere survey artefacts: implicit association tests (IATs) show that unconscious prejudice based on partisanship is (now) stronger in the US than negative affect based on race (Iyengar & Westwood 2015).
At the same time, affective polarization is far from restricted to the United States (Reiljan 2020;Wagner 2020). The configuration of political antipathy does depend on the features of a political system. In the US, affective polarization takes place between two clearly defined partisan camps, which in turn increasingly align with ideological divisions (Baldassarri & Gelman 2008;Webster and Abramowitz 2017;Banda and Cluverius 2018). In multiparty systems, selfidentification with any single party is generally weaker (though not absent; Huddy et al. 2018;Bankert et al. 2017), and any partisan ingroup is not opposed to every possible partisan outgroup to the same extent (Wagner 2020). 'Political identities' in the broad sense of the word also reflect ideological identities (Devine 2015;Malka & Lelkes 2010) 2 and even specific issue identities (Hobolt et al. 2019). A supposedly shared ideological identification as 'Left-wing' might lead a Social Democratic supporter to feel more warmly towards supporters of Green parties than towards supporters of a mainstream, let alone radical, right-wing party. As Reiljan (2020) andWagner (2020) show, this provides not only an ingroup but also an outgroup, which tend to disliked.
Excessive affective polarization is likely to have a negative impact on the functioning of democracy. Next to competitive elections and strong institutions, a well-functioning democracy requires norms that allow for deliberation, compromise, and forbearance between citizens and elites of opposing political camps (Ziblatt & Levitsky 2018;Strickler 2017). If political identities become strongly entrenched, and those voting for the 'wrong party' become loathed, such acceptance (or at least pragmatism) might dwindle. It might even legitimate violence against opponents (Kalmoe and Mason 2018). 3 It is crucial to recognize that affective polarization is not the same as ideological polarization -the extent to which citizens disagree on matters of ideology and issue positions. This relation is endogenous and not straightforward (Iyengar et al. 2018;Ward and Tavits 2019). While affective polarization has grown according to most accounts, there is less evidence for surging ideological polarization: regarding most topics, Americans' and Europeans' actual views have become less, rather than more, divided (Baldassarri & Gelman 2008;Adams et al. 2012;Nuesser et al. 2014). Indeed, Reiljan (2019) and Gidron et al. (2019a) find a weak or inconsistent relation between ideological polarization on affective polarization. Individual-level evidence is mixed too (Rogowski & Sutherland 2016;Bougher 2017). All in all, it is clear that affective polarization depends partly -perhaps mostly -on other factors than the strength of ideological disagreement between camps.

The role of social sorting along political lines
This paper builds on the work by Mason (2015;2016; by studying the role of social sorting along political lines ('social sorting' in short 4 ), or the alignment of political identities with nonpolitical identities. The general expectation is that such alignment fosters antipathy towards political opponents, and that this relation holds everywhere -not only in the US where it has been documented so far. The reason is that the underlying mechanisms, which involve intergroup relations, are plausibly universal. They reiterate the insight that cross-cutting cleavages work in the interest of social harmony (Lipset 1960). Most concretely, social sorting 'infuses' political divisions with the tensions characterizing other divisions in society. A political outgroup (say, Green party supporters) will be consciously or subconsciously associated with non-political outgroups (say, the highly educated or city-dwellers). To the extent that prejudice or bias exists towards the latter (based on allegedly competing values, lifestyle or interests), this amplifies any purely political antipathy (say, over environmental policy). While this amplified dislike might not be purely political, it does bear political consequences.
In addition, from a Social Identity Theory perspective, increasing alignment can easily distort citizens' perceptions of the world and reduce 'social identity complexity' (Roccas and Brewer 2002). Imagine a situation in which a person associates a political outgroup (say, again, Green party supporters) to typically be an outgroup in many other respects as well (say, university graduates, atheists and city dwellers). In that case their 'social identity complexity' can be said to be low. Reversely, if the same person would perceive that the political outgroup (Green party supporters) is often on the ingroup side in other respects ('us lower educated' or 'us rural folk'), this would signal a higher 'social identity complexity'. A more complex social identity generally enhances tolerances for outgroups (Roccas and Brewer 2002). It decreases intracategory assimilation and intercategory accentuation -in other words, the tendency to overestimate the internal homogeneity of, as well as differences between, the ingroup and the outgroup. 5 As a result, groups such as 'Green voters' and 'Conservative voters' will appear less like a threatening homogeneous and distinct block. Indeed, stressing (in experimental settings) that outgroup members are ingroup members on another dimension decreases bias (Gaertner et al. 1993;Levendusky 2018). It follows from this that having cross-cutting political and non-political identities should reduce negative affect towards the political outgroup.
In the US context, Mason (2015;2016; demonstrates that social sorting is indeed associated with stronger affective polarization. She shows that Republicans who are also white and Christian are more affectively polarized than Republicans with a cross-cutting identity -regardless of ideological extremity. The same mechanism is observable among Democrats who identify as secular and non-white, as opposed to Democrats with cross-cutting identities. Mason and Wronski (2018) therefore stress the need to study the "cumulative relationship between social identities and partisan identities". The fact that affective polarization depends on the configuration of a broader set of identities is also apparent from the work of Levendusky (2018). He finds affective polarization to be lower among respondents who were cued to be aware of their overarching American identity and among respondents interviewed on the national holiday July 4 th , which signals that the political outgroup is a at the same time a national ingroup (although see Brandt and Turner-Zwinkels 2020 for a non-replication of this finding).
Of course, the fact that non-political attributes tend to generally go together with a particular political position does not mean that individuals will universally perceive it as such. Roccas and Brewer (2002) and Mason (2018) therefore also study the cognitive representation of such alignment. Still, research in US shows partisans are generally aware of, and often even grossly overestimate, the features associated with Democratic or Republic support, such as religion, union membership or income (Ahler and Sood 2018). In short, the alignment of social groups with political camps (the empirical focus of this study) can be expected to feature systematically, if imperfectly, in citizens' perceptions. Through the mechanisms discussed above, this is expected to foster affective polarization.

Hypothesis: higher levels of social sorting along political lines are associated with higher levels of affective polarization
Of course, political differences have always been connected to social divisions. Political parties mobilize voters by appealing to one side of a salient current or historical cleavage (Lipset and Rokkan 1967). For instance, studies on Western Europe have shown patterns of both dealignment and realignment (Kitschelt and Rehm 2015), including a decline in the association of vote choice with class and somewhat of an increase in the predictive power of education (e.g. Rekker and Van der Brug 2020). Historical cleavages such as a religious, territorial and urban-rural still play a role to varying extent (e.g. Knutsen 2010). It is therefore important to stress that neither the concept or the phenomenon of social sorting is new. Rather, the extent to which political groups are socially sorted, and the lines along which they are, varies between contexts and over time. For instance, spatial segregation can create neighborhoods or entire regions in which citizens share not only a worldview but also other social identities such as education or ethnicity (Bishop 2008; but see Abrams and Fiorina 2012). Cross-cutting institutions that socially integrate socio-demographic groups and a plethora of political views -for instance churches or sport clubs -can decline (but also grow) in importance or become organized along more homogeneous (or heterogeneous) lines.
Affective polarization might in turn encourage people to 'sort', i.e. withdraw in homogeneous groups, which would make the relation endogenous. Establishing the causal direction with certainty is beyond the observational data presented in this paper. While the US studies discussed above provided strong evidence for the hypothesized mechanism, any observed correlation between affective polarization on social sorting might partly reflect a reciprocal effect.
Nevertheless, it remains important to study whether the relation between social sorting and affective polarization travels beyond the US context.

Design
In the remainder of this paper, I investigate the relation between social sorting and affective polarization empirically in two complementary studies. The first study does so on the aggregate level using the Comparative Study of Electoral Systems (CSES) data, covering 119 elections in 40 countries between 1996 and 2018. The goal is to assess whether this relation indeed holds across the world. I calculate a Social Sorting Score inspired by Selway (2011) that captures the extent to which political divisions align with non-political cleavages in a society. I then use the Social Sorting Score to predict affective polarization, which I measure following the operationalization of Reiljan (2019). The time span provides sufficient variation within countries, which isolates the correlations as far as possible from (relatively) time-invariant confounding factors such as political system or culture. The repeated cross-sectional nature of the CSES data makes it less suitable to make strong inferences about mechanisms on the individual level. I therefore complement my design with the second study, which employs panel data to study the mechanism on the individual level. I use the population-representative Dutch LISS data that allows to track the social sorting of individuals over a period of up to ten years (2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018). I develop a measure of the extent to which individuals fit the socio-demographic profile of a party, and use a hybrid model to study if affective polarization moves in tandem with an individual's social sorting.

Data and cases
The Comparative Study of Electoral Systems 6 data consists of harmonized election studies in countries around the world between 1996 and 2018. This provides a unique collection of equivalent survey items across all continents over 20 years. 40 countries have valid observations of the key independent and dependent variables at least once (see Appendix A in the Supporting Information). 7 In total this yields a maximum of 119 (country-election) observations. While the main interest of this study is to study the relation between social sorting and affective polarization on the aggregate level (allowing for a longitudinal element), I replicate the main correlation on the individual level in the robustness section.
Like the measures proposed by Wagner (2019) and Gidron et al. (2019), this is based on sympathy scores towards out-parties. Although related, this is not the same as affect towards partisans.
However, because of the correlation between the two (r = 0.69 in the US in Iyengar et al. 2012; see also Druckman et al. 2019) it can be used -with some caveats -to study affective polarization. The measure also likely captures ideological polarization (Wagner 2020), so I control for ideological polarization in each model (see below). The Affective Polarization Index can be thought of as 'weighted sympathy' towards outparties. The formula below (taken from Reiljan 2020: 381) summarizes the procedure. First, for each partisan group n (i.e. respondents who say they identify with a particular party 8 ) the average evaluation towards all other parties m is subtracted from the respondent's evaluation of their inparty. Each difference is weighted by the relative size of a party (measured as vote share) and summed over all outparties. This is repeated for each of the partisan groups, and all these 'relative AP' scores are weighted by vote shared and summed up as well.

Affective Polarization Indexc =
(1) Note: formula from Reiljan (2020: 381) Social sorting. The operationalization of the main independent variable is inspired by the work on cross-cuttingness by Selway (2011). Cross-cuttingness is the reverse of sorting, and can be defined 7 I opted not to restrict the analysis to a particular subset, because a priori the mechanisms should apply in all contexts. 8 Based on the question "Do you feel closer to any particular party?".
as the extent to which "group i on cleavage x is identically distributed among groups on cleavage y with all other groups on cleavage x" (Selway 2011: 51). Selway (2011) measures the crosscuttingness of two divisions by calculating Cramér's V, a measure of association for categorical variables, and subtracting it from 1. Because I am interested in sorting, I simply use Cramér's V.
Another application of Cramer's V is the work of Knutsen (e.g. 2004;2010), who measured the strength of cleavages in a country through Cramer's V associations with party choice.
For each of the elections covered in the dataset, I calculated four Cramér's V associations between on the one hand party identification 9 and on the other hand each of the following nonpolitical variables: income (5 quintiles), education (5 harmonized categories), region (number of categories differs per country), and religion (operationalized as 'majority religion' vs the rest). 10 The selection of non-political divisions was restricted to variables consistently available across most election studies, 11 but it does capture arguably most of the major social fault lines that tend to be expressed politically: class (approximated by income), the 'new class' or 'winners and losers of globalization' division (education), as well as center-periphery (region) and religious oppositions.

/ 4 (2)
Because it is based on a correlation coefficient for categorical variables, the Social Sorting Score is a priori agnostic about the shape of the relation between party preference and each of the other identities. For instance, in some countries, the lower educated are relative often supporters of the Social Democratic party, in other countries of the Conservative party; in again other countries, they might usually vote for far left or far right alternatives, rather than centrist ones. All of these patterns are captured by the cross-cuttingness measure, as it reflects how well categories of education (or income, region, or religion) predict vote choice. The better it does, the higher the Social Sorting Score.
The measure comes with several caveats. First, taking the average Cramer's V means that countries that are (moderately) sorted along multiple lines get higher scores than countries that are (perhaps strongly) sorted along a single dimension. While it is plausible that the former cumulative understanding matters more than the latter (Mason and Wronski 2018), the analysis will be replicated with a measure taking the maximum Cramer's V. Second, the Social Sorting Score relies on ascribed identities that can be 'objectively' measured (education, income, region, and religion), rather than directly assessing citizens' subjective identities. However, it seems plausible that the factors included in the measure have some systematic bearing on many citizens' of the share of the manifesto's of all parties in a given election that is devoted to cultural issues or economic issues, respectively. 14 Because the main interest is to explain changes over time within countries (see below), the models do not include slow-moving economic indicators such as GDP or inequality (Gidron et al. 2019). I do include the control variable Number of parties. A larger set of parties changes the logic of competition and cooperation, and possibly reflects a more consensual political culture. Arguably, this indicator too varies more substantially between countries than over time. Still, I include it because I expect it to correlate with the outcome variable in a mechanical way, too: the more parties are available in the questionnaire, the more negative affect towards one of them gets averaged out by others. This indicator should absorb much of that correlation.

Design
The independent and dependent variables are available for 119 waves across 40 countries. Affective polarization and social sorting are measured at the level of country-election, providing repeated observations at the country level. The analysis is performed at the aggregated level, using both a fixed effects (FE) and between-effects (BE) estimation. The FE specification restricts the estimation to variation within countries, between elections. In other words: at elections in which a country is more sorted than on average, is affective polarization relatively high too? This absorbs potential time-invariant confounders at the country level, and also reduces the problem that the values on Social Sorting are likely to be influenced to a substantive extent by whether the chosen nonpolitical divisions matter in a particular country. The between-effects (BE) specification is based on variation between countries. Do those countries that are (on average) more sorted also experience more affective polarization? Of the 40 countries under study, 10 appeared in the dataset only once; these are only included in the between-effects analyses. 15 14 Coded as cultural issues were: environmentalism, culture, equality, national way of live, law and order, multiculturalism. Coded as economic issues were: free market economy; incentives; market regulation; economic planning; corporatism; protectionism; economic goals; demand management; economic growth; controlled economy; nationalization; economic orthodoxy; Marxist analysis. 15 The average number of observations for a country is 2975.

Results
Descriptives. Appendix B reports the trends in affective polarization, social sorting (including the individual components), and ideological polarization on the mass level. It testifies that affective polarization has different levels and trends across the world. Replicating Reiljan (2019), Southern Europe and East-Central Europe generally show very high scores, followed by the US.
Furthermore, there is ample variation in the trend of social sorting (see Boxell et al. 2019). It also shows that affective polarization and ideological polarization sometimes, but not always, move in tandem. The same is true for affective polarization and social sorting. However, the large number of cases defies visual identification of correlations. For that we turn to a regression analysis.
Regression. Figure 1 shows regression coefficients of the within-country (FE) and between-country (BE) models. To allow for some comparability, both the Affective Polarization Index and the independent variable were standardized across the sample as a whole. Each model contains all control variables as well as dummies for the CSES wave to capture possible trends in the dependent variable. Table 1 in Appendix C shows the full regression tables.
The open red dots (which represents the coefficients of the fixed-effects model) shows that the Social Sorting Score has a positive effect (p = 0.013) on the Affective Polarization Index. The effect size is substantial: an increase of 1 standard deviation in social sorting is associated with a 0.37 standard deviation increase in affective polarization. The between-country coefficient (closed blue dots), by contrast, provides no evidence for an effect of the Social Sorting Score. In other words, only variation in the social sorting measure over time predicts affective polarization. This might be due to lower comparability of the social sorting measure across countries. It could also mean, plausibly, that many unobserved factors determine the level of affective polarization in a country that add noise to the correlation. At any rate, the existence of a within-country effect supports the hypothesis that social sorting goes together with higher affective polarization outside the US, too.  Tables 2 in Appendix C). Using the maximum (rather than the average) Cramer's V score yields a somewhat weaker correlation that is not significant at p = 0.11. This suggests that having multiple cumulative alignments is more polarizing than a single one. The last four coefficients split up the Social Sorting Score by its individual components. Theoretically, given that cumulative overlap matters most, so it is a priori unlikely that individual overlap scores are as strong predictors of affective polarization as the aggregate scores. Still, splitting them allows to explore whether some dimensions are more polarizing than others. It shows that political alignment along all four nonpolitical dimensions is (descriptively) associated with more affective polarization, but only significantly so (at the 10% and 5% level respectively) in the case of overlap with regional and religious divides. The latter two arguably often constitute stronger identities than income and education, which makes their preeminence plausible. At the same time, even they will not constitute the most relevant cleavage in every country.

FIGURE 2. REGRESSION MODELS WITH DIFFERENT OPERATIONALIZATIONS OF SOCIAL SORTING
Note: Different operationalizations of social sorting, each of them estimated in a separate model. All models controls for the same variables as Figure 1, including round dummies. Fixed effects for countries.
Having explored the effects of sorting, we can return to the control variables of Figure 1, which perform largely as expected, even if not all of them are robust predictors. The number of parties is not significantly associated with affective polarization, despite the expectation that fragmentation would dampen polarization. Ideological polarization, both on the level of voters and parties, is associated with more affective polarization. Ideological polarization among voters is somewhat associated (but not significantly so) with more affective polarization when comparing both between and within countries; ideological polarization among parties is a solid predictor of variation in affective polarization within countries. It is notable that the Social Sorting Score has an effect size that is comparable to that of ideological polarization. This confirms that affective polarization is rooted in much more than substantive disagreement alone (Iyengar et al. 2018). The salience of either cultural or economic issues does not play a role, in contrast to Gidron et al. (2019). The dummies for CSES waves (not included in the figure) suggest a small upward trend in affective polarization exists, net of the included predictors.
Robustness. I conducted a set of additional analyses to further gauge the robustness of the above findings. First, I replicated the analysis on the individual level using the individual-level measures for affective polarization (Wagner's [2020] weighted affective polarization score) and social sorting (the respondent's 'fit' to a party in socio-demographic terms 16 , or reverse residual score), both of which are further explained in Study 2 below. Table 3 in Appendix C shows the results of a multilevel model in which (standardized) affective polarization was predicted by (standardized) social sorting as well as a set of control variables at the individual and country-wave level. 17 This confirms that social sorting, also when measured on the individual level, has a significant positive effect (p < 0.01; b = 0.144).
Second, I performed two robustness checks on the aggregate models of Figure 1 above, which are reported in Tables 4 and 5 of Appendix C. I estimated a first-differences model (i.e. testing whether the change in API is predicted by the change in the independent variables), which, too, isolates the effect from time-invariant confounders but is more robust in case of high serial correlation. This yielded a significant effect of similar magnitude (b = 0.35; p = 0.020). I also replicated the fixed-effects analysis using a 'jackknife' strategy (removing one country at a time) to test the sensitivity of the findings for the patterns in individual countries. Under this specification the p-value of the effect of social sorting increases from 0.025 to 0.069. Given the more conservative specification and directional hypothesis this presents evidence for the hypothesis.

Study 2: a Dutch panel study on social sorting and affective polarization Data and case
In Study 1, we learned that when societies become more sorted along political lines, affective polarization is generally higher. In this study, I zoom in on one case, the Netherlands, and track the 16 The same factors were used as in the aggregate analysis: income, education, region and religion. 'social sortedness' and affective polarization among a total of 8205 unique individuals over periods up to 10 years (from 2008 to 2018). If the pattern of Study 1 is brought about by the theoretical mechanism discussed earlier, we would expect to see that individuals who fit their ingroup party better in socio-demographic terms should generally be more affectively polarized.
The Netherlands experiences only weak affective polarization according to Reiljan (2019) and is not characterized by the unique linguistic, ethnic or religious cleavages studied by Westwood et al. (2015). This thus makes for a less-likely case to explain any variation in affective polarization by social sorting. To the extent that the two do correlate, it is plausible that effects are at least as pronounced in contexts that are both more affectively polarized and sorted.

Affective Polariationi = (3)
The element likei is the weighted sympathy towards all parties: Social sorting. In the two-party context of the United States, Mason (2015)  political interest (three categories), two variables that respectively capture ideological polarization 19 A Christian democratic party might be sorted above all by religion, a social democratic party by class, a green party by education, etcetera. 20 In contrast to Study 1, this variable includes urbanity rather than region, because urban-rural differences in political support are generally larger than regional differences in the small and dense country of the Netherlands. Religion is not available in all waves and therefore excluded. Age is used to capture the age differential in support for parties that was visible in recent elections (Van der Meer et al. 2019) 21 This residual is calculated based on an OLS rather than logistic model to obtain continuous residuals. The result is a roughly normally distributed absolute residual. and the salience of political identities, and which might well change over time. I also control for education, age and gender. I formulate no expectations for the latter variables, but for their role in the US see Iyengar et al. (2012: 425).

Design
As in Study 1, the goal of the analysis is to establish the relationship between social sorting and affective polarization. This focus of the main analysis on explaining both variation between and within respondents, of which the latter makes most use of panel data by isolating effects from confounding by time-invariant individual characteristics (that is, given that for various reasons some citizens are both more sorted and polarized). I estimate a hybrid model which simultaneously models both sources of variation, providing both a within-and between-estimator for each variable (Mundlak 1978;Bell and Jones 2015).
Of course, the focus on within-respondent variation raises the question whether individuals change in their level of sorting, as the features on which sorting is based often do not change (although they sometimes do). However, respondents can switch parties and hence end up in a more (or less) sorted party, and the party they belong to can change in composition to become more or less similar like the respondent. Even if sorting would vary more between individuals than within individuals, a focus on the latter makes for the more robust control strategy.

Results
Regression. Figure 3 shows the unstandardized coefficients 22 of the hybrid model explaining Weighted Affective Polarization score by Social sorting and other variables. Table 1 in Appendix E provides the full regression table. Each variable (except gender 23 ) has a within-and betweenrespondents coefficient. The first coefficient confirms that in waves in which individuals score relatively high on Social sorting, this person also tends to score higher on Weighted Affective Polarization. Or, in other words, the better an individual's support for a party can be predicted by their socio-demographics, the more disliking this individual generally is of political outgroups. The between-respondent coefficient of Social sorting is in the same direction and somewhat smaller in size. This means that those respondents who are most sorted (keeping a range of socio- Figure 2 is therefore in line with the hypothesis. All of this is in line with the hypothesis and further strengthens the likelihood that the aggregated pattern found in study 1 are based on the same mechanisms as observed in the work of Mason (2018 As an alternative specification, I again ran a first difference model, regressing change on change (see Table 2 in Appendix E). This results in a significant effect of a comparable size (b = 0.57, p < 0.001).

Conclusions
Cross-cutting cleavages have long been argued to decrease social tensions (Lipset 1960). Building on the research of Mason (2015;2016;, this study investigated whether -in a broader set of cases than hitherto studied -citizens' dislike towards political opponents is stronger if political identities align with non-political identities. Comparing 119 elections in 40 countries over two decades, I found that social sorting along political lines is indeed associated with affective polarization across a range of specifications. The effect size is substantive and similar to that of ideological polarization -the extent to which citizens disagree substantially on issues. I furthermore show, using Dutch panel data that allows to make stronger claims on the individual level, that when citizens better fit the socio-demographic profile of their party they also tend to be more affective polarized. It is telling that this mechanism is visible even in a country with very weak affective polarization (Reiljan 2020). Both studies strongly suggest that the hostility fostered by political identities depends on its interplay with non-political identities.
In line with previous work, this study thus confirms that (growing) affective polarization is about more than just (growing) issue disagreement (Iyengar et al. 2018;Reiljan 2019;Lelkes 2018). It raises a question with important implications: how purely political is affective polarization? To the extent that affective polarization is partly a reproduction -or amplification -of tensions smoldering outside the realm of politics, there are limits to the extent it can be abated by reforming a political culture or adjusting political institutions. Efforts to prevent affective polarization from becoming 'toxic' (or even lethal; Kalmoe and Mason 2018) would need to grapple with other dimensions of social divisions and segregation too. Empirically, it is difficult to isolate the relative role of the political and the non-political using observational data, precisely because citizens have conscious and subconscious associations between political and non-political groups.
Experimental or conjoint designs, such as employed by Helbling and Jungkunz (2020), are important steps forward in this respect. Does flagging a non-sorted identity (i.e., portraying somebody as a Green supporter and lower educated) as opposed to a sorted identity (Green supporter and higher educated) reduce affective polarization towards that person?
The present study took a relatively coarse approach by assuming that respondents identify with the 'objective' categories presented to them in the questionnaires and are aware of how these are aligned with the political camps in the population at large. As Ahler and Sood (2018) show, real associations between political and nonpolitical groups are reflected in people's perceptions, although in a heavily distorted way. Still, the fact remains some citizens will identify more or less strongly with -say -their class, religious, or regional group than others, and this matters for the extent to which such identities influence their views of political opponents. It is therefore plausible that the correlation between 'objective' sorting and affective polarization found in the present study is an underestimation of the complete role played by social sorting. Following Mason (2015), future comparative studies on social sorting could benefit from including measures of subjective identification. Although this requires more elaborate data, it does provide for a more direct evidence of the mechanism at hand.
The social sorting measure presented in this paper suggests that sorting is (somewhat) on the rise across the globe, but with substantial variation in levels and trends between countries.
This calls for furthering our theoretical and empirical understanding of social sorting and its causes. It also raises the question who tends to be sorted most. Even if only a subset of society, such as the most politically most engaged, finds itself increasingly sorted, then the associated higher levels of affective polarization might spill over to the political debate and the public at large.
Next to continuing the comparative agenda, an important next step is therefore to further explore its interaction with the individual level: who is most 'at risk' of becoming affectively polarized?