ASRXXX10.1177/0003122420922989American Sociological ReviewDellaPosta research-article9229892020

American Sociological Review 2020, Vol. 85(3) 507­–536 Pluralistic Collapse: The “Oil © American Sociological Association 2020 https://doi.org/10.1177/0003122420922989DOI: 10.1177/0003122420922989 Spill” Model of Mass Opinion journals.sagepub.com/home/asr Polarization

Daniel DellaPostaa

Abstract Despite widespread feeling that public opinion in the United States has become dramatically polarized along political lines, empirical support for such a pattern is surprisingly elusive. Reporting little evidence of mass polarization, previous studies assume polarization is evidenced via the amplification of existing political alignments. This article considers a different pathway: polarization occurring via social, cultural, and political alignments coming to encompass an increasingly diverse array of opinions and attitudes. The study uses 44 years of data from the General Social Survey representing opinions and attitudes across a wide array of domains as elements in an evolving belief network. Analyses of this network produce evidence that mass polarization has increased via a process of belief consolidation, entailing the collapse of previously cross-cutting alignments, thus creating increasingly broad and encompassing clusters organized around cohesive packages of beliefs. Further, the increasing salience of political ideology and partisanship only partly explains this trend. The structure of U.S. opinion has shifted in ways suggesting troubling implications for proponents of political and social pluralism.

Keywords political polarization, public opinion, belief networks, political pluralism

Social and political theorists have long argued prevent disagreement in one arena from esca- that differentiated societies are best able to lating into all-out (metaphorical or literal) integrate diverse interests when relying on a warfare. As long as opinion cleavages remain foundation of cross-cutting social and political cross-cutting rather than all-encompassing, cleavages (Blau and Schwartz 1984; Dahl pluralistic disagreement should channel social 1961; Durkheim 1933; Lipset 1963; Simmel conflict toward mutual tolerance (Stouffer 1955; Truman 1951). Unlike stultifying mono- 1955), political forbearance (Levitsky and cultures built on public consensus (Arendt Ziblatt 2018), and other liberal ends. 1968), cross-cutting lines of conflict and dis- agreement found in pluralistic societies embed individuals in overlapping groups (Baldassarri aPennsylvania State University and Diani 2007; Centola 2015; Mutz 2002). Even if people fight over differing views on Corresponding Author: taxation, they may still agree on issues of for- Daniel DellaPosta, 503 Oswald Tower, Pennsylvania State University, University Park, eign policy or overlap in religious views; PA 16802 these areas of agreement are thought to Email: [email protected] 508 American Sociological Review 85(3)

Against this backdrop, recent decades have they have? Why, in other words, does polari- brought mounting concern among social sci- zation appear as such a salient social fact entists, journalists, policymakers, and the gen- despite its apparent absence in the U.S. eral public that the pluralistic structure of U.S. population? politics has turned to all-encompassing con- Previous studies suggest the gap between flict between increasingly polarized camps. In popular and scholarly consensus could be due their provocatively titled book How Democra- to the tendency for media reports to pay selec- cies Die, Levitsky and Ziblatt (2018) frame tive attention to polarized “hot-button” issues polarization as a fundamental problem for like abortion (Baldassarri and Bearman 2007; U.S. democracy—not merely a nuisance but a DiMaggio et al. 1996; Evans 2003); outsized potentially existential threat. Such alarmism is focus on political elites who have become not uncommon in mainstream discourse. Yet increasingly partisan and polarized them- despite the common belief that the United selves (Hetherington 2001); social sorting States has become dramatically more politi- leading to a sense of political homogeneity in cally polarized, social-scientific support for our personal networks (Baldassarri and Bear- such a pattern has been surprisingly elusive. man 2007; Cowan and Baldassarri 2018; Del- In the rarefied air of the U.S. Congress, laPosta, Shi, and Macy 2015); or increasing there is clear evidence that the Democratic geographic polarization in which regions and and Republican parties have moved far apart locales increasingly feature one-party domi- in recent decades (Andris et al. 2015; Lee nance even when elections split evenly at the 2009; Moody and Mucha 2013; Poole and national level (Bishop 2008; but see Abrams Rosenthal 1984). In the broader public, adher- and Fiorina 2012). ents of the two major political parties have become increasingly distinct, as Democrats Measuring Polarization have become more consistently liberal and Republicans more reliably conservative (Bal- Nevertheless, although many scholars have dassarri and Gelman 2008). However, critiqued the lack of a clear index of polariza- although people have become better sorted by tion in previous work (e.g., Fiorina and party and political ideology, public attitudes Abrams 2008), less attention has been paid to on most political issues seem to have remained the ubiquity of a particular (and implicit) unpolarized to a remarkable degree (DiMag- model of how polarization would appear in gio, Evans, and Bryson 1996; Evans 2003; population-level survey data. The common Fiorina 2004; Hill and Tausanovitch 2015). empirical strategy in prior research is to select Although knowing someone’s political and analyze a limited set of survey questions party or self-described ideology allows us to on political issues, focusing on questions the predict their attitudes with increased accu- researcher identifies as most politically rele- racy, these attitudes themselves have not vant (e.g., opinions on abortion, climate become much more strongly aligned with change, taxation, or foreign policy). By select- other attitudes, as we would expect in a world ing on an issue’s already-established political of mass polarization (Baldassarri and Gelman relevance, however, the researcher implicitly 2008; Fiorina and Abrams 2008; Fischer and assumes that polarization occurs via height- Mattson 2009; Park 2018).1 Unlike the politi- ened alignment along existing lines of politi- cal elite, the broader public is composed of cal debate, akin to a fence the researcher can large numbers of “ideological innocents” observe getting taller over time. (Converse 1964; Kinder and Kalmoe 2017) But what if polarization is less like a fence who espouse cross-cutting and often incon- getting taller over time and more like an oil sistent beliefs across political issues. This spill that spreads from its source to gradually raises the question: If opinions have not taint more and more previously “apolitical” become markedly more polarized, why is it so attitudes, opinions, and preferences? DellaPosta widely believed—and so easy to intuit—that and colleagues (2015) show that even many DellaPosta 509 initially apolitical lifestyle characteristics, from to analyze how these cultural elements cohere musical taste to belief in astrology, can become into a broader structure of meaning (DiMag- politicized as signals for deeper beliefs and gio 2011:287; see also Lizardo 2014; Mohr preferences—a tendency most saliently cap- and White 2008; Pachucki and Breiger 2010; tured in the popular image of the “latte liberal” Vilhena et al. 2014). (see also Hetherington and Weiler 2018; Mutz This conceptual approach dovetails with and Rao 2018; Shi et al. 2017). classic work on beliefs and opinions (e.g., This suggests a different answer to the Converse 1964) in which the meaning and puzzle posed at the outset: rather than height- importance of a particular belief derives from ened alignment across already-politicized its relationships to other beliefs (for some opinion dimensions, the crux of contempo- recent examples, see Baldassarri and Gold- rary polarization might lie in the increased berg 2014; Boutyline and Vaisey 2017; Fried- breadth of opinions and preferences that have kin et al. 2016). Whereas recent work typically come to be associated with political identities conceives of polarization as increasing and beliefs. This broadening of opinion align- extremism on issues (e.g., DiMaggio et al. ment to encompass areas typically thought of 1996) or heightened alignment across pairs of as nonpolitical would not be picked up in issues (e.g., Baldassarri and Gelman 2008), studies that only consider polarization along the belief network approach recasts polariza- existing lines of political debate. However, tion as a particular structure of opinion. the widespread impression that polarization Scholars of public opinion have long rec- has dramatically increased may reflect real ognized the importance of belief structure, politicization of an increasingly diverse span meaning an ordered arrangement of beliefs of beliefs and preferences that once cut across such that holding one belief implies holding (or lay apart from) ideological divides. other beliefs (Kinder and Kalmoe 2017:13; see also Converse 1964). However, recent research on polarization conceptualizes struc- Belief Network Analysis ture in terms of pairwise relationships among Rather than focusing on a selected subset of beliefs, ignoring the larger webs of associa- political opinion items, I broaden the scope of tion and mutual implication in which these analysis to include any opinion question ever dyadic relationships are embedded (Baldas- presented in the General Social Survey (GSS) sarri and Goldberg 2014). between 1972 and 2016. Building on To understand what polarization looks like Boutyline and Vaisey (2017), I represent cor- in a structure of associated beliefs, we can relations between opinions in a representative compare it to its opposite: cross-cutting opin- sample of the U.S. population as relational ion pluralism. In a pluralistic belief network, ties in a network of interconnected beliefs. attitudes may be correlated with one another, Rather than a traditional social network anal- but these correlations are cross-cutting. For ysis where nodes are human or organizational example, perhaps attitudes toward gun con- actors and the ties or edges between nodes trol are correlated with attitudes toward abor- represent social connections (Wasserman and tion. But attitudes toward gun control may Faust 1994), the belief network conceives of also be correlated with attitudes toward racial each attitude, opinion, or belief as a node; the justice, which may themselves not be corre- strength of the edges or ties between beliefs lated with abortion attitudes; abortion atti- reflects the extent to which two beliefs are tudes, in turn, may be correlated with attitudes correlated in survey data. The analysis of toward LGBT rights, despite the latter remain- belief networks represents one strand in the ing uncorrelated with gun control attitudes. growing field of cultural network analysis, As a whole, this hypothetical belief network where “networks of co-occurrence relations consists of relationships that cut across and among such cultural elements as words, offset each other without cohering into large tropes, attitudes, symbols, or tastes” are used cohesive clusters of densely connected attitudes. 510 American Sociological Review 85(3)

Substantively, this lack of coherence means belief formation and change. Yet, by demon- two people who disagree on one or even most strating in what way mass beliefs have things will still likely be able to find some become more polarized, and by using net- issue on which they agree, creating an oppor- work analytic tools to describe the structure tunity to bridge the ideological distance of this polarization, this article provides three between them. linked contributions. In this conceptualization, opinion plural- First, the article fleshes out a novel con- ism collapses into polarization via the con- ceptualization of polarization—the earlier solidation of previously cross-cutting referenced “oil spill” model—rooted in the alignments into increasingly broad and breadth of opinion alignment, and places this encompassing ones. Drawing on theories of conceptualization in dialogue with previous belief formation, I argue that beliefs will tend ones. Second, I use the formal tools of net- to cohere into modular “packages” that can be work analysis to develop previous research distinguished based on clustered items in the on opinion alignment in a way that more fully belief network. Beliefs falling within the accounts for the structure of associations same package tend to be more strongly con- among beliefs. Finally, the article builds on nected to one another, and items in different and extends Boutyline and Vaisey’s (2017) modules tend to be more weakly connected. cross-sectional analysis of belief networks to Simply put, a belief network becomes more show how this approach can be used to track consolidated as the distinct modules of beliefs changes over time in mass beliefs. become larger in size but fewer in number. I find that—in line with claims for increased polarization—the structure of U.S. opinion The Illusion of has shifted in recent decades away from one Polarization? made up of narrower but cross-cutting mod- Previous work describes the process of mass ules of beliefs and toward fewer modules, opinion polarization in one of two ways: (1) indicating broadly-encompassing alignments. as increasing extremism (i.e., bimodality) of Representing mass beliefs as networks, as attitudes on political issues or (2) as increas- in the present study, does not directly tell us ing alignment (i.e., correlation) of attitudes about the underlying mechanisms giving rise across political issues. In the first tradition, to belief polarization. Correlations between DiMaggio and colleagues (1996) used data beliefs in the population aggregate need not from the General Social Survey (GSS) and imply cognitive associations between these the American National Election Studies beliefs at the individual level (Martin 2000), (ANES) to show that U.S. attitudes and opin- although aggregate belief structures can shape ions had largely not become more extreme individual belief formation by determining over time, with the notable exceptions of which attitudes and behaviors are most likely attitudes toward abortion and opinion differ- (or unlikely) to co-occur in the population, ences between partisan identifiers (see also which in turn affects how individuals will Evans 2003; Fiorina 2004; Fiorina and perceive the associations among different Abrams 2008). The alignment tradition, attitudes and behaviors (Goldberg and Stein which provides the starting point for the pres- 2018). Individuals, social groups, and the ent study, can be traced to the work of Con- contentions and conflicts among them will verse (1964), who proposed that belief effectively “fall out” of the analysis presented structures should be described in terms of here, leaving a descriptive picture—albeit a their degree of constraint (see also Baldas- rich one—of larger-scale trends in population sarri and Gelman 2008; Friedkin et al. 2016; belief structure. Experimental and other Martin 2002). approaches (e.g., Bail et al. 2018; Hunzaker A highly constrained belief structure is one and Valentino 2019) remain necessary for dis- in which holding some beliefs strongly implies covering the underlying root causes driving holding other beliefs, producing correlations DellaPosta 511 across a wide range of issues. If knowing Gelman (2008:441) summarize, any apparent someone’s attitude on issue X allows us to polarization in the population should be seen predict attitudes on issues Y and Z, then X, Y, as an “illusory” artifact of the reshuffling of and Z make up a constrained set of beliefs. partisan labels. Applying this logic to the polarization debate, Despite key differences in how scholars researchers examine correlations or align- conceptualize attitude polarization and the ments between issues rather than distributions mechanisms producing it, previous studies of attitudes within issues. When strong corre- have one thing in common: they approach lations across multiple dimensions of opinion polarization by focusing on a selective set of indicate high levels of constraint—for exam- political issues curated by the researcher. ple, when knowing someone’s attitude on Whether conceiving of polarization as extrem- abortion helps predict their attitude on gun ism on single issues or as alignment across control—the population can be said to be pairs of issues, studies of mass polarization more polarized. In contrast, if there are null or seem to invariably begin with the researcher weak alignments across pairs of issues, the selecting a set of political opinion questions belief structure as a whole would be made up from a national survey. In their exemplary of intersecting, cross-cutting dimensions of study, Baldassarri and Gelman (2008) analyze opinion that resist polarization. pairwise alignments across 47 issues sorted Tracking alignment across a broad set of into four topical domains (economic, civil political opinions, Baldassarri and Gelman rights, moral, and security and foreign policy). (2008) report that, on one hand, attitudes on This approach is eminently sensible because it particular issues have become increasingly focuses attention on the issues we would pre- aligned in recent decades with both self- sume to be most relevant to understanding reported political ideology and party identifi- polarization and reduces the complexity that cation. This finding fits with other studies would emerge from taking a more inclusive reporting that the Democratic and Republican approach absent such presumptions. However, parties (and liberal and conservative identifi- this strategy also imposes an implicit con- ers) have become increasingly ideologically straint on how polarization can appear in the coherent (e.g., Abramowitz and Saunders data. Namely, the only residue of polarization 1998, 2008; Fiorina and Abrams 2008; Lay- that can be uncovered using this approach man and Carsey 2002; Levendusky 2009; would be heightened alignment across issues McCarty, Poole, and Rosenthal 2006; but see that the researcher has defined a priori as Kinder and Kalmoe 2017). On the other hand, being most relevant to the study of polariza- what Baldassarri and Gelman (2008) call “issue tion, typically issues that can be widely cate- alignment”—or the actual extent of correla- gorized as “political” in nature. tion across different attitude dimensions—has More recently, DellaPosta and colleagues remained relatively stable. Self-identification (2015) replicated Baldassarri and Gelman’s as a liberal now increasingly predicts one’s (2008) approach using an eclectic set of items attitudes on specific issues, yet attitudes on from the General Social Survey (GSS), issue X evidently do not predict one’s attitudes including a mix of political “hot-button” on issue Y with significantly greater accuracy issues and seemingly less political lifestyle than in past decades thought to be less polar- questions, such as musical taste, attitudes ized. With these patterns at loggerheads— toward art, parenting beliefs, and even belief partisan alignment evidently on the rise yet lit- in astrology. Their cross-sectional focus was tle evidence of issue alignment—Baldassarri not designed to identify over-time shifts in and Gelman (2008) conclude that the U.S. alignment across these dimensions, yet their population has followed the lead of party elites analysis is notable for the surprisingly large in becoming increasingly consistent partisans number of such seemingly nonpolitical items but that polarization at the level of actual atti- that nevertheless displayed strong and statisti- tudes has not increased. As Baldassarri and cally significant correlations with both 512 American Sociological Review 85(3) self-described political ideology and more and Napier (2009) define ideology as a network explicitly political attitudes. of interconnected beliefs that, taken together, This motivates the questions addressed in form a coherent whole or “world view.” the present study: Is mass polarization an illu- In a study of the emergence of cultural sion caused by the reshuffling of party labels, variation, Goldberg and Stein (2018:899) as much previous work suggests, or have we define culture in terms of “social conventions not taken a broad enough view of the beliefs that associate practices with meanings,” that that might be involved in producing polariza- is, cognitive frames via which people come to tion? Put differently, is the evident lack of understand distinct beliefs and ideas as being belief polarization found in previous work a functionally dependent on one another (see true signal of stably cross-cutting opinions in also DiMaggio 1997; Ghaziani and Baldas- the U.S. population, or a result of under- sarri 2011; Lizardo 2017; Mohr 1998; Patter- specifying the full scope of beliefs relevant to son 2014). What these different perspectives understanding the structure of polarization? have in common is the shift from studying To address these questions, this study rep- distributions of particular beliefs or even cor- licates key features of the correlational or relations among pairs of beliefs to represent- alignment-based approach offered by Baldas- ing the overall structure of beliefs in relation sarri and Gelman (2008) while going beyond to one another. Rather than describing partic- this previous work in two respects. First, I re- ular beliefs, the key question concerns how assess the claim that partisan alignment has various beliefs cohere into a larger network increased but mass polarization has not, and I and the structural properties of that network. do this for a broader cross-section of issues Previous work often uses network meta- drawn from the totality of opinion questions phors to describe structures of interconnected ever presented in the General Social Survey beliefs, yet most empirical studies of belief (GSS) in place of a more narrowly selected structure and polarization nonetheless limit subset of explicitly political items. Second, I themselves to methods that do not fully use the pairwise alignments between issues as account for patterns of relations among a baseline for further analysis rather than as an beliefs. Summarizing this critique, Baldas- end point. Specifically, the correlations between sarri and Goldberg (2014:54) write, “Most pairs of beliefs are used to induce a holistic scholars, following Converse, measure ‘con- network depicting the structure of U.S. mass straint’ using bivariate relationships (e.g., cor- opinion. The structural properties of this belief relation coefficients) or, alternatively, summary network provide a novel grammar for measur- indices (e.g., factor scores). Such approaches, ing mass polarization. however, either presuppose or overlook the overall pattern of political attitudes that char- acterize a belief structure.” As previously Beliefs as Networks noted, for example, Baldassarri and Gelman Sociologists, political scientists, psychologists, (2008) track patterns of belief constraint in and cognitive scientists alike have long pro- the U.S. population by estimating weighted posed that beliefs are not held in isolation; averages of bivariate correlations among rather, they are held in relation to other beliefs. beliefs. Earlier work often conceived of For example, Converse (1964:207) defines a beliefs as being collapsible into one or several belief system as a “configuration of ideas and underlying linear dimensions. For example, attitudes in which the elements are bound Fleishman (1988) applies factor-analytic together by some form of constraint or func- methods to survey data to show that attitudes tional interdependence” (see also Baldassarri are generally organized along two broad and Goldberg 2014; Bonikowski and DiMag- dimensions: one concerning attitudes toward gio 2016; Friedkin et al. 2016; Kinder and individual liberty and the other pertaining to Kalmoe 2017; Martin 2002). Jost, Federico, economic welfare policy.2 DellaPosta 513

More recently, scholars have begun to use that liberal–conservative ideology plays in network-analytic techniques to examine the organizing the political beliefs of the U.S. pub- formal properties of belief structures (Baldas- lic. They contrast their findings with Lakoff’s sarri and Goldberg 2014; Boutyline and Vai- (1996) famous claim—which finds little sup- sey 2017; Friedkin et al. 2016; for an earlier port in the aggregate data—that Americans forerunner, see Martin 2002). The several derive their beliefs on political issues from formal approaches that have emerged are metaphorical models of the family (liberals based on the duality between persons and the favoring a “nurturant parent” model versus beliefs they hold, practices they exhibit, or conservatives favoring a “strict father” model). objects with which they associate (e.g., Lee Previous studies in both of these molds and Martin 2018; Lizardo 2014; Pachucki and have been cross-sectional in focus, rather than Breiger 2010; Puetz 2017). Goldberg’s (2011) studying change over time in mass opinion. relational class analysis (RCA) approach RCA and related approaches seek to identify exploits the duality of persons and beliefs to heterogeneous cognitive mappings of issues inductively uncover groups of individuals in within a population. Any such over-time anal- survey data who organize their attitudes in ysis would likely require a constant set of similar ways (see also Baldassarri and Gold- survey items appearing in every cross-section berg 2014; DiMaggio and Goldberg 2018; of the data. In comparison, aggregate belief- DiMaggio et al. 2018). For example, two centric approaches, such as Boutyline and people who share the same opinions across a Vaisey’s (2017), provide a less granular focus range of questions organize their beliefs simi- on population heterogeneity within cross- larly, but so do two people who hold directly sections while providing the potential for opposed positions on those same questions. greater comparison across cross-sections. RCA provides the counterintuitive insight This is because the aggregate belief network that two people who disagree with regard to for a representative sample of the population every issue still share the same fundamental allows individuals to “drop out” of the analy- cognitive model of the way those issues are sis.4 As with previous studies of pairwise connected to one another.3 opinion alignment (Baldassarri and Gelman Rather than seeking to uncover groups of 2008; DellaPosta et al. 2015), the researcher individuals who “construe” the world in simi- can then statistically model population time lar ways (DiMaggio and Goldberg 2018), the trends in the correlations between beliefs, belief-networks approach proposed by Boutyline using estimates from these models to interpo- and Vaisey (2017) exploits the duality of per- late missing observations for years in which a sons and beliefs to study the patterns of con- particular survey item did not appear. This nections among those beliefs at the population interpolation makes focusing directly on rela- level. In their approach, responses to attitudi- tionships among beliefs themselves especially nal survey questions generate networks of useful for tracking changes over long periods interconnected beliefs linked by the magnitude of time in the structure of mass beliefs. of their bivariate associations with one another. Individuals within a population may vary In the belief network, each opinion or attitude in their conformity to an aggregate belief is a node and its correlation with another opin- structure (Baldassarri and Goldberg 2014; ion or attitude produces a network tie or edge Martin 2002), and such heterogeneity would weighted according to the strength of the cor- not be captured by studying the population- relation. The goal of this approach is to repre- level belief network.5 Yet, as recently illus- sent the central elements organizing beliefs in trated in a theoretical model by Goldberg and the population aggregate. To this end, Boutyline Stein (2018), aggregate belief structures can and Vaisey (2017) represent aggregate correla- shape individual belief formation by provid- tions among political beliefs as a network ing ready instances of association between configuration to show the central role beliefs and practices that are likelier than not 514 American Sociological Review 85(3) to co-occur in the population. For example, if Baldassarri and Gelman 2008) is akin to every time I go to a football game, I see park- observing the height of a fence. The researcher ing lots full of cars with Trump bumper- takes observations at multiple points in time stickers, I will tend to see football fandom as to see whether the alignment between these being associated with Trump support. If I issues has increased—in other words, whether already like football but do not yet support the fence has gotten higher. This fence model Trump, I might conclude from this that I of polarization is illustrated in Panel A of should naturally support Trump due to my Figure 1, where beliefs are represented as other preferences. In other words, to the nodes in a network and the lines connecting extent that individuals (a) associate com- these beliefs represent the absolute correla- monly co-occurring beliefs and practices and tion between those beliefs in a hypothetical (b) update their own beliefs and practices to population. At Time 1, beliefs about abortion, maintain consistency with these mental maps climate policy, and taxation are all weakly and avoid cognitive strain, aggregate belief aligned with one another; by Time 2, each of structures can be expected—at least to a modest these pairwise alignments has become stron- extent—to shape individual beliefs (Festinger ger. Two people who disagreed about abor- 1957; Goldberg and Stein 2018; Heider 1946; tion at Time 1 may still have agreed on Hunzaker 2016; Kunda 1990). climate policy or taxation, given that these The present study builds on the methodo- beliefs were only weakly correlated. By Time logical framework introduced by Boutyline 2, however, the same two people would be and Vaisey (2017). Whereas Boutyline and likelier to either agree or disagree on all Vaisey analyzed a selected cross-section of issues, as the three beliefs have all become beliefs from the 2000 edition of the American more strongly correlated with one another. National Election Studies, the present study This is the model of polarization as the applies the belief-network approach to map heightening of existing opinion alignments the structure of a large number of attitudes across pairs of issues. and opinions over more than 40 years of the Cross-cutting opinion pluralism can col- General Social Survey. By comparing the lapse not just through heightening alignment same belief network over long periods of time across pairs of issues but also through broad- among representative aggregates of the popu- ening alignment across a wider range of lation, the present study brings the belief- issues than were previously aligned. Broad- network approach to bear on the question of ening alignment is less like a fence and more whether and how mass opinion has shifted in like an oil spill. In Panel B of Figure 1, initial the United States. Previous public opinion alignments on abortion, climate policy, and research in the tradition of Converse (1964) taxation at Time 1 expand to implicate LGBT frequently invokes the importance of struc- rights and religious beliefs at Time 2. It is not ture, but the belief-network approach allows that the particular alignments between the for a more fully network-based view of mass original set of issues have become stronger, polarization. necessarily, but rather that the introduction of new issue dimensions has expanded the range of these previous alignments.6 At Time 1, The Network Structure people may tend to either agree or disagree on of Belief Polarization a set of core political issues; at Time 2, as new Fence and Oil Spill Models (political, lifestyle, and other) issues are of Polarization introduced into existing alignments, the range of agreement and disagreement broadens. To return to an earlier metaphor, measuring Tracing the expanding boundaries of the “oil the average alignment across pairs of quintes- spill” as it taints an increasingly broad set of sentially political issue dimensions (e.g., issues represents a distinct challenge for DellaPosta 515

Figure 1. Two Models of Polarization Note: Each node represents an opinion or attitude and each solid line indicates an absolute correlation or alignment observed between those opinions or attitudes in the population. Thicker lines indicate higher-magnitude correlations than thinner lines.

empirical research on polarization. Unless the clearly defined belief modules that are rela- researcher knows ahead of time which issues tively large in size and few in number. In will (and will not) be drawn into range of the contrast, an unconsolidated belief network spill, they cannot define the relevant set of could simply lack any clear modular struc- issues a priori but instead must allow the ture, such that the belief network cannot be broadest possible range of data to speak to divided into distinct “packages” of correlated this question. beliefs. Or, in another scenario, a relatively unconsolidated belief network might feature a larger number of modules where no single Belief Modules and Consolidation module or handful of modules dominates by Network structures commonly feature groups containing all or most beliefs. By analogy, we or clusters of nodes that associate more could compare a highly consolidated belief strongly with one another than with others. network to a monopolistic or oligopolistic The modularity or “clusteredness” of a net- market or industry dominated by one or a few work refers to the extent to which the network firms holding giant market shares. In contrast, is divided into neatly compartmentalized a less consolidated belief network would modules with more edges occurring within— more resemble a competitive marketplace as opposed to across—modules (Newman with many smaller firms competing for mar- 2006). In a modular belief network, beliefs ket share. group into cohesive “packages” featuring Figure 2 illustrates the difference between strong associations among the beliefs con- modular belief networks with differing levels tained in a given package (i.e., believing of consolidation using simple “toy” networks. something about one issue in the package As in Figure 1, each node in the network rep- strongly implies believing something about resents a belief on an issue or topic as meas- another issue in the same package). ured by a survey question. A line or edge Increasing mass polarization in an aggre- between two issues indicates that opinions on gate belief network should bear a distinct those issues tend to be strongly correlated in empirical signature, which I call consolida- the population. These correlations can be posi- tion. Simply put, belief networks with high tive or negative—it is the magnitude, not the consolidation are composed of cohesive and direction, that matters. For example, if support 516 American Sociological Review 85(3)

Figure 2. Illustrative Examples of Belief Networks Note: Each node represents an opinion or attitude and each solid line indicates a strong absolute correlation. Patterned fills indicate different belief modules. for legal abortion is strongly negatively cor- lifestyle and culture (Hunter 1991), or under- related with support for teaching creationism lying sociodemographic and material causes in public schools, beliefs on these two issues (Brooks and Manza 1997; Lipset and Rokkan should be likelier ceteris paribus to be placed 1967; Marx and Engels 1978). As I will dis- into the same belief module. Finally, the pat- cuss, however, such formations could also terned fills on the nodes indicate distinct mod- plausibly emerge even in the absence of any ules, each containing an internally cohesive coherent organizing principle (DellaPosta et (correlated or aligned) “package” of beliefs al. 2015).7 (setting aside for now the question of how one As discussed earlier, previous work on would formally identify these modules beyond cross-issue alignment in the tradition of Con- mere visual inspection). verse (1964) generally focuses on the average Panel A in Figure 2 displays a belief net- correlation strength across pairs of beliefs work with modular groupings but little con- (e.g., Baldassarri and Gelman 2008; Del- solidation. The network is modular because laPosta et al. 2015). Yet, even two networks the only strong correlations occur within that are identical with regard to the average clearly differentiated packages of attitudes or strength of the correlations between beliefs opinions; despite the strong clustering evident could evince different structural arrangements in the network, however, modules are similar depending on how those individual beliefs fit in size and none dominate the network. This to produce the larger gestalt of the belief net- unconsolidated arrangement suggests cross- work.8 To be clear, it is not my argument that cutting packages of opinions that do not the overall strength of correlation between cohere into a larger whole. In the more con- pairs of beliefs is irrelevant to polarization. solidated belief network in Panel B, by con- Rather, my argument is that structural differ- trast, most beliefs adhere to a single central ences that can be observed independently module. In other words, the belief network in from the aggregate strength of alignments Panel B is dominated by a central cluster of between issues are also empirically and theo- beliefs that correlate together closely. In this retically significant but have been compara- consolidated arrangement, one might hypoth- tively neglected in previous work. Thus, esize the existence of a central organizing analysis of structure provides a key avenue principle from which all the beliefs in this through which we can apply a new lens to the central module originate—for example, lib- question of polarization. eral or conservative political identity (Boutyline Importantly, even a highly modular belief and Vaisey 2017; Jost et al. 2009), underlying network could be seen as supporting pluralis- moral frames (Haidt 2012), deep-rooted tic, cross-cutting alignments rather than cognitive associations (Lakoff 1996), shared polarization. Panel A of Figure 2 illustrates DellaPosta 517 such a dynamic, where opinions cluster in range of opinions and attitudes encompassing several clearly differentiated modules lacking not just straightforwardly political issues but a broader alignment with one another. Theo- also social, moral, and cultural values. To ries of political pluralism suggest that such examine structural alignments across the wid- cross-cutting alignments, far from represent- est available array of attitudes over the entire ing a social ill, are in fact central to organic span of years covered by the GSS, I sought to solidarity in modern, complex societies com- include as many of these items as possible. posed of many heterogeneous subpopulations Following Boutyline and Vaisey (2017), I (Baldassarri and Diani 2007; Baldassarri and adopt Alwin’s (2007) distinction between Gelman 2008; Mutz 2002). The real concern “factual” questions—for which the answers for polarization occurs as these networks can be objectively verified—and “non- cease to consist of such cross-cutting align- factual” questions—namely attitudes, opinions, ments. This happens when previously distinct beliefs, and values—that reflect respondents’ belief modules are subsumed into increas- subjective viewpoints. I sought to include ingly encompassing alignments—as reflected every item ever asked in the GSS that fell into in the high-consolidation example of Panel B. the latter category of presenting respondents’ As a belief network becomes more consoli- subjective viewpoints. This agnostic approach dated, the breadth of polarization increases— to item inclusion broadens the scope of previ- two people who disagree on one issue, for ous analyses by considering all available example, become increasingly likely to disa- opinions and attitudes rather than a smaller gree on an expanded range of other issues.9 curated subset of charged political questions, To summarize the argument: the structural but we are still constrained by what the pur- properties of belief networks can shed light veyors of the GSS chose—or did not choose— on whether the U.S. public has become more to ask about in the first place.10 Because the polarized, or whether cross-cutting pluralistic focus of the analysis is on the correlation alignments remain. Drawing on theories of between items, I only included items with belief formation, networks of interconnected binary, interval-scaled, or continuous beliefs can be approached by identifying the responses, or items that could be recoded to different modules or “packages” of correlated reflect such a scale. In other cases, multiple beliefs that comprise the broader network. versions of the same question had to be rec- Then, drawing on theories of political plural- onciled by recoding and then consolidated. ism, polarization should manifest in increas- This initial selection process produced a ing consolidation as previously cross-cutting list of 1,357 GSS items. Adapting a procedure clusters of beliefs collapse to form increas- developed by Baldassarri and Gelman (2008), ingly broad and encompassing groupings, I next obtained the zero-order Pearson correla- thus shrinking the space available for agree- tion between each unique pair of items for ment across the divides created by opinion every year in which those two items appeared cleavages. together in the GSS.11 Because we are inter- ested in the magnitude of the correlation—or the strength of the alignment between any two Data and Methods items—rather than the direction, I use the Survey Data absolute value of the correlation coefficients throughout the analysis (Boutyline and Vaisey The General Social Survey (GSS) is a nation- 2017; DellaPosta et al. 2015). Over the 31 edi- ally representative survey of U.S. attitudes, tions of the GSS issued between 1972 and behaviors, and demographics fielded by the 2016, this procedure produced 474,199 unique National Opinion Research Center (NORC) year-specific correlations between pairs of at the University of Chicago since 1972. The opinions, attitudes, and beliefs. However, GSS allows for consideration of a broad these year-specific instances of correlations 518 American Sociological Review 85(3) between items only reflect a total of 216,704 desire to see in their children to attitudes unique pairs of items—in other words, the toward abortion, foreign policy, and taxation average pair of items appeared together in the and economic regulation. same edition of the GSS just between two and three times. Moreover, the modal pair would Constructing the Belief Network have appeared together just once. This reflects the structure of the GSS, in which a relatively In the belief network, each GSS item becomes small number of items appears reliably across a node and the correlations between items successive editions of the survey, and other become edges linking these nodes into a questions are asked just once or a handful of broader structure of connection (Boutyline times, typically as part of a larger set of items and Vaisey 2017). Edges are weighted by the rotated into the survey to address an area of magnitude of the correlation between the two timely interest to researchers. connected nodes. The goal is to create a The sparsity of data for most unique pairs graphical representation of this belief network of GSS items presents a double-edged prob- for each of the 31 years in which the GSS was lem. First, it makes the process of inferring an conducted from its 1972 inception to the 2016 over-time trend of increasing, decreasing, or edition. To make apples-to-apples compari- stable alignment between any such pair of sons across the belief networks for different items arguably tenuous, due to the lack of years, I hold constant the set of items captured data points with which to observe such a in the network and focus only on changes in trend. One possible solution would be to the relationships among these items. restrict analysis to only the observed correla- Building on previous work (Baldassarri tions without any reliance on statistical mod- and Gelman 2008; DellaPosta et al. 2015), I eling of over-time trends. However, this use a multilevel mixed-effects model (Gel- solution would only induce a much deeper man and Hill 2007) to estimate the linear selection problem: the list of items analyzed trend of increasing or decreasing magnitude in any given year would differ radically from of correlation over time for each pair of GSS other years. Because the goal of the analysis items (see the Appendix for more detail). For is to identify changes over time in the belief years in which two items co-appeared in the network owing to corresponding changes in GSS, the weight of the edge between those the alignment of different attitudes, it is two items is the absolute value of their essential that the population of GSS items observed correlation for that year. Then, for represented in the belief network remain con- years in which those two items did not co- stant. Otherwise, observed changes in the appear, I retain them in the network and esti- structure of the belief network could easily be mate the weight of the edge between them explained by the somewhat arbitrary mix of using predicted values from the mixed-effects questions that happened to be asked from one model, incorporating intercepts and time edition to the next. trends that vary for each pair of items. In seeking a solution that balances both In effect, missing observations stemming edges of this data sparsity problem, I restricted from the absence of some items in certain analysis to the 14,910 pairs of items that years of the GSS are filled in based on the appeared together in at least five years of the statistical trends observed for those items in GSS. Although this subset of reliably observed the years for which they did appear. This correlations includes only approximately 7 approach requires a heavy reliance on statisti- percent of all possible pairs of items, it still cal estimates to fill gaps in the record of retains roughly 43 percent of all year-specific empirical observation, but avoids the more correlations observed in the original set (a vexing dilemma of accounting for the count- total of 202,928 such correlations).12 The less idiosyncratic changes in the number and retained questions cover a wide array of top- composition of items appearing in the GSS for ics, ranging from the traits people would any given year. Yet, there are consequent DellaPosta 519 limitations in how the results of the analyses To further distinguish time trends owing to should be interpreted. Namely, because the partisan sorting from those reflecting broader network tie between any given pair of items in shifts in the belief structure, I replicate the any given year will often be based on a statisti- analysis while removing from the belief net- cal estimate, we should be cautious in inter- work all correlations featuring either political preting short-run year-by-year changes as ideology or party identification as one of the substantively meaningful. Rather, our focus two focal variables. I call this second condition should remain on the longer-term trends the the ideology-removed network; comparing the mixed-effects model is designed to capture.13 baseline and ideology-removed networks will The belief network for each year contains help answer the question of whether any 219 nodes, each representing a GSS item. To observed changes over time persist beyond the inductively uncover the distinct belief direct influence of partisan sorting (in which modules—internally cohesive clusters of highly particular opinions become more strongly cor- correlated opinions—present in the network related with partisan and ideological identity for a given year, I rely on widely-used com- without the relationships among the opinions munity detection algorithms designed for use themselves necessarily shifting). in complex networks. In particular, I imple- I then go a step further by also adjusting ment the “walktrap” method, which discovers for self-reported ideology and party identifi- communities in a network based on the logic cation in all the remaining correlations that a random walk through the network will between belief-pairs. This third condition, tend to become trapped within the dense clus- which I call the ideology-controlled network, ters of nodes belonging to the same commu- includes the partial correlations between nity (Pons and Latapy 2006). In a simulation belief items after controlling for ideology and study, Gates and colleagues (2016) found that party and removing all correlations that the walktrap method outperformed other directly involve either of these variables from community detection algorithms in recover- the network. Clearly, this third condition pre- ing community structures for dense weighted sents the most stringent test of whether any networks; the belief network, derived from observed shifts in the belief network persist correlation matrices among pairs of survey apart from sorting, because it adjusts for both items, is just such a structure (for other appli- direct (i.e., partisan identity’s correlation with cations of the walktrap approach in correla- particular beliefs) and indirect (i.e., the extent tion networks, see Dalege et al. 2017; Golino to which partisan identity moderates the cor- and Epskamp 2017). relation between other beliefs) pathways through which partisan and ideological iden- tity may influence the belief network struc- Accounting for Partisan and ture. Figure 3 summarizes the key steps in the Ideological Sorting analytic procedures used to generate the Although self-reported political ideology (on a belief networks.14 scale from “Extremely Liberal” to “Extremely Conservative”) and party identification (on a scale from “Strong Democrat” to “Strong Analysis Republican”) could be seen as sociodemo- Describing the Belief Network graphic traits rather than beliefs, I include both in the belief network to account for the Figure 4 presents a visual depiction of the role of partisan and ideological sorting. I call belief network for 1972, the first year in the this the baseline network because it simply observation period.15 Belief modules are dis- contains the full set of zero-order correlations tinguished by color (see the online version of among all belief items (including ideology this article for color figures); note, however, and party identification). that the belief modules are induced separately 520 American Sociological Review 85(3)

Figure 3. Analytic Procedures Used to Generate Belief Networks

for each year, so the colors are not necessarily three visible central modules, the network consistent across years. The largest module in contains a number of peripheral modules that the 1972 network contains 30 percent of are less clearly connected to the central core. beliefs (purple-colored in the right region of When looking at the network snapshot for the graph). This module is organized around any single year, the relative placement of beliefs about race, gender, and civil liberties— nodes is partly arbitrary, because an edge is it features questions (labeled in the graph by only present between two beliefs if those their GSS mnemonic) such as whether African items co-appeared in enough editions of the Americans have “worse jobs, income, and GSS for the correlation between them to be housing than white people” due to having “less included in the analysis (as discussed earlier, I in-born ability to learn” (racdif2), whether it is use five co-appearances as the threshold for “much better for everyone involved if the man inclusion). However, these snapshots become is the achiever outside the home and the meaningful when compared to earlier or later woman takes care of the home and family” snapshots. The set of nodes (beliefs) and edges (fefam), and whether books written by com- (correlations) is the same in every snapshot of munists should be permitted in libraries (lib- the network, so any change over time in the com). The second largest module (green), structure of the belief network stems from containing 24 percent of beliefs, includes changes in the relative weights of the edges (among other things) questions gauging confi- that are present. As the alignments between dence in core social and political institutions pairs of beliefs change over time, the structure like big business (conbus), the military of the belief network shifts accordingly. (conarmy), and the clergy (conclerg). A third Figure 5 shows the 2016 belief network. central module (orange), which cuts across the Each of the two largest belief modules have first two and contains 10 percent of beliefs, now increased in size relative to the others. includes law-and-order-tinged beliefs about The largest module (purple) now contains 37 capital punishment (cappun), crime (nat- percent of all items, and the second largest crime), and guns (gunlaw). In addition to these (green) now contains another 32 percent. This DellaPosta 521

Figure 4. Belief Network for 1972

trend suggests that previously peripheral into fewer and more encompassing modules. beliefs belonging to their own smaller mod- To provide a more rigorous test of these ules have been pulled closer over time to the claims, Figure 6 shows changes over time in dense core of political identity reflected at the formal properties of the belief network. Each center of the network. The set of beliefs that panel displays local polynomial regression moved from smaller modules into the two fits of over-time network properties in the largest ones over the observation period reflect baseline condition containing all zero-order a wide-ranging set of concerns, including correlations (solid lines), the ideology- race-based affirmative action (affrmact), the removed condition where self-reported ideol- appropriateness of older people sharing a ogy and party identification are excluded home with their children (aged), whether most from the network (dashed lines), and the people can be trusted (trust), the scientific ideology-controlled condition featuring par- credibility of astrology (astrosci), and the tial correlations between beliefs after adjust- moral infallibility of the Pope (popespks). ing for political ideology and party identification (dotted lines). To assess the significance of the observed trends relative to Trends in Belief Network Structure those that could plausibly emerge through Visual inspection of these belief networks sampling variability, I generated 5,000 boot- provides initial evidence that beliefs have strapped replications of the belief network for become increasingly consolidated, meaning each year of the GSS. Adapting the procedure previously cross-cutting (or relatively uncor- used by Boutyline and Vaisey (2017), each related) modules of opinion have collapsed nonparametric bootstrap replication involved 522 American Sociological Review 85(3)

Figure 5. Belief Network for 2016 resampling the GSS respondents for each with density because this metric provides an year with replacement, re-computing the analogue to measures of average alignment observed correlations in each bootstrapped across pairs of issues in previous studies of sample, re-generating the statistical models of opinion alignment. The overall trend for the these observed correlations, and using the ideology-removed network is similar to that results of the model to produce yearly belief of the baseline network. However, it is nota- networks (thus repeating the procedures ble that the ideology-controlled network fea- described in Figure 3 for each bootstrap rep- turing partial correlations (adjusting for lication). The trends depicted in Figure 6 are political ideology and party identification) based on the yearly means across all 5,000 between beliefs saw a slight decrease in den- bootstrap replications. sity over the same span. These trends support Panel A of Figure 6 finds that the overall previous work (e.g., Baldassarri and Gelman network density (i.e., the average weight of 2008) by showing that the increase in average the edges across all pairs of beliefs) increased alignment or correlation between pairs of modestly across the 1972 to 2016 observation beliefs owes much to the mediating effect of period in the baseline belief network. I begin partisan and ideological labels. Yet, the fact DellaPosta 523

Figure 6. Time Trends in the Belief Network Note: Displayed lines are based on local polynomial regression fits from 5,000 bootstrap replications across all years in the observation window. Solid lines reflect the baseline condition featuring unadjusted zero-order correlations; dashed lines reflect the ideology-removed condition with correlations directly involving political ideology and party identification removed from the network; and dotted lines reflect the ideology-controlled condition featuring partial correlations adjusting for the influence of political ideology and party identification.

that this pattern is confirmed here puts even distinct) in construction, and large in size. greater importance on where the belief net- Corresponding to these three structural signa- work analysis diverges from previous studies, tures, the main analyses presented here rely which is in examining the broader structure of on simple network metrics capturing the ways aggregate beliefs rather than the average pair- beliefs are arranged across different belief wise alignment. modules in the network. Panel B in Figure 6 As described earlier, belief consolidation shows the number of belief modules uncov- will be most evident when belief modules are ered by the walktrap community detection few in number, modular (i.e., cohesive and algorithm across the observation period. 524 American Sociological Review 85(3)

Consistent with increasing consolidation, the percentage linear change between 1972 and number of modules decreased over time in all 2016 based on these regression fits, both in three network conditions. terms of the mean change and the variability Panel C tracks the modularity of the belief across bootstrap replications. Clearly, each network. As mentioned earlier, modularity bootstrap replication can produce substan- captures the “clusteredness” of the network, tially variable results due to the several ana- conceived as the amount of information a lytic steps involved in each such replication given clustering solution gives us about the and the path-dependence among these steps actual structure of relationships in the network (as summarized in Figure 3). This high degree (Newman 2006). More formally, modularity of variability is reflected in the wide confi- represents the proportion of edges (or weight dence intervals (90 and 95 percent intervals attached to those edges, as in the present case) are shown) around the mean estimates in occurring within clusters or modules compared Figure 7. Still, some notable trends emerged to the proportion one would expect by chance reliably across a high percentage of bootstrap in a network with the same degree distribution replications. as the observed one. Importantly, this means First, Panel A in Figure 7 shows there are modularity captures the extent of clustering statistically significant changes in density within dense communities while adjusting for over the observation period: we see positively effects of the overall density of ties in the net- increasing density in the baseline and ideology- work, such that time trends in the modularity removed conditions and decreasing density in of a given network can be interpreted inde- the ideology-controlled condition. With regard pendent of changes in the network’s underly- to the number of belief modules, Panel B ing degree distribution. As shown in Panel C, shows that the observed decrease is most reli- modularity increased similarly over time ably observed in the baseline network and less across all three network conditions. consistently seen in the other two network Panel D of Figure 6 shows the percentage conditions. In these latter two conditions, the of beliefs placed into the largest module by trend still appears to be in the same direction the community detection algorithm; in other as in the baseline network, but the trend is words, the size of the largest single grouping clearly weaker and less distinguishable from a of cohesively tied beliefs in the network. In null pattern. As Panel C shows, the 12.3 per- the visual inspections, we saw that growth in cent average increase for modularity in the the second largest module also seemed to baseline network and the 11 percent increase indicate consolidation, so Panel E tracks the in the ideology-removed network are both percentage of nodes in the two largest mod- observed consistently enough across bootstrap ules combined. The percentage of nodes replications that the 95 percent confidence placed in the single largest module remained intervals do not include zero. Strikingly, even relatively consistent over time. However, the the ideology-controlled network—in which combined size of the top two modules saw a political ideology and party identification are more visible increase, a pattern consistent not just removed from the network but further with increasing consolidation. Panel F shows adjusted for in computing the partial correla- an overall summary measure of the concen- tions between all other pairs of beliefs—sees a tration of beliefs within modules using the marginally significant 10 percent average Rosenbluth market concentration index (Hall increase in modularity (p < .10). and Tideman 1967).16 As with the previous Panel D confirms the suspicion that the measures, this summary metric increased size of the single largest belief module did not over the observation period. noticeably change over the observation To test the significance of these over-time period. Panel E shows that although an aver- differences, I estimated a linear regression fit age increase in the combined size of the two of the time trends for each of the 5,000 boot- largest modules is observed across all net- strap replications. Figure 7 plots the work conditions, the point estimates vary DellaPosta 525

Figure 7. Bootstrapped Confidence Intervals for Network Trends Note: Results are based on linear regression fits of the time trends across 5,000 sets of bootstrapped belief networks. For each network condition, panels show the mean percent linear change across all bootstrapped replications and the 90 and 95 percent confidence intervals. widely enough that the confidence intervals In short, analyses of the aggregate belief still intersect zero (although barely) in all networks suggest that, on the whole, the three cases. A similar pattern appears in Panel structure of beliefs has moved in meaningful F for the summary measure of module con- ways toward increasing consolidation— centration, which sees a marginally signifi- beliefs increasingly concentrated within fewer cant (p < .10) increase in the baseline and more encompassing modules rather than condition and slightly less consistent increases a greater number of more cross-cutting mod- in the other two conditions. ules. As in most such studies that attempt to 526 American Sociological Review 85(3) encompass a wide and agnostically-selected polarization; however, important questions range of opinions, the overall time trends are concerning underlying causal mechanisms substantively modest; yet, they are suffi- remain for future work. Because survey data ciently consistent with one another that the based on probability sampling isolate indi- overall shift in the structure of the belief net- viduals from the networks and social contexts work is unlikely to be a product of chance in which they are embedded (McPherson statistical variation. Even still, the observed 2004), these data do not allow us to rigorously trends encompass substantial uncertainty and disentangle the various within-individual (e.g., merit cautious interpretation. effects of sociodemographic background) and Like previous work, I find clearer evi- between-individual (e.g., effects of social dence of a shift with regard to opinion align- influence) mechanisms that may give rise to ments that include salient markers of partisan belief structures (DellaPosta et al. 2015). and ideological identity. However, even after Moreover, because the repeated cross- removing these partisan identity markers sectional data do not follow the same indi- from the belief network, I still observe a sta- viduals over time, little can be gleaned about tistically significant increase in network den- individual-level processes of belief formation sity (i.e., average correlation strength) and and change (but see Kiley and Vaisey 2020). modularity (i.e., the tendency for beliefs to By focusing on aggregate belief structure in a cluster into coherent and internally cohesive representative sample of the U.S. population, packages). The trends observed for even the the analysis also leaves open the question of most stringent condition—where partisan and group-level opinion cleavages and how belief ideological identification are both removed heterogeneity within the population has from the network and controlled while com- evolved over time. puting the correlations between other Correlated sets of beliefs may give rise to beliefs—are less precise but still suggestive. social identities organized around those In particular, the fact that this network fea- beliefs (e.g., people who oppose legal abor- tured a significant decrease in the average tion and gay rights organizing as the “Reli- correlation while still likely moving toward gious Right”). Just as importantly, however, consolidation in other ways (e.g., increasing the existence of certain social identities pro- modularity) that appear similar to the baseline vides signals for how individuals should form network suggests the network approach cap- new beliefs in connection with others, revers- tures shifts in the structure of beliefs that ing the causal sequence (Achen and Bartels would not be obvious using traditional 2016; Converse 1964; Lazarsfeld, Berelson, techniques—in particular, approaches that and Gaudet 1944). Political ideology and measure belief structure by averaging the party identification are commonly proposed dyadic alignments among pairs of beliefs as the connective tissues linking otherwise without placing these relationships in broader unconnected beliefs; for a signal of how one relational context. should form beliefs on a given issue, partisans look to what their co-partisans—including public figures—are saying about the issue. Discussion and Through such influence processes, the origi- Conclusions nal or orienting belief (e.g., belief in liberal- Limitations and Suggestions ism) acts as a constraint on the formation of for Future Research subsequent subsidiary beliefs (e.g., support for abortion rights). The analysis presented here remains limited In still other cases, cohesive belief modules in several important ways. The aim of the may be epiphenomenal to other underlying present study was to conceptualize and mea- social forces only loosely related (or perhaps sure aggregate trends reflecting mass belief unrelated) to social identity. For example, a DellaPosta 527 variety of attitudes toward civil liberties, non- Perlstein’s (2001) narrative history of the U.S. conformity, and the value of societal tolerance conservative movement describes how con- might cluster because they all stem from a greater servative intellectuals and activists fused the “open-mindedness” rooted in sociodem- previously unaligned (or weakly-aligned) ographic variables like education, urbanism, identities of anti-communist Cold Warriors, and occupation (Stouffer 1955). Or, beliefs socially conservative Christians, and eco- may become clustered in sociodemographic nomic libertarians to craft a combined politi- space due to “bottom-up” processes of homo- cal identity powerful enough to take over the phily and social influence whereby people Republican Party. with similar lifestyles or cultural values have These ideological bridges can also emerge greater chances to interact and thereby further from a process that Goldberg and Stein (2018) influence one another’s opinions (DellaPosta call associative diffusion. In their model, peo- et al. 2015; McPherson 2004; McPherson, ple are influenced by observing what behav- Smith-Lovin, and Cook 2001; Vaisey and Liz- iors tend to co-occur among their peers (who ardo 2010). Via a gradual process of social need not be closely tied to the actors, and in influence channeled through these homophil- fact might be strangers). Observing the co- ous ties, DellaPosta and colleagues (2015) occurrence of two behaviors creates a cogni- show that initially small alignments between tive association between those behaviors, substantively unconnected issues can tip the leading to people seeing them as naturally population into highly polarized alignments related to each other. Like beliefs, practices, (see also Baldassarri and Bearman 2007; Del- and behaviors themselves, Goldberg and laPosta and Macy 2015; Flache and Macy Stein (2018) convincingly argue, cultural 2011; Macy et al. 2003). associations among such objects can also dif- The present study mostly abstracts away fuse through repeated observations of the from the historically specific processes social world. If two otherwise unrelated through which mass opinion in the United beliefs or opinions (i.e., liberalism and latte- States became increasingly consolidated, but drinking or football fandom and Trump sup- the belief-network approach can also be used port) come to be seen as associated through to home in on such questions. Structurally, enough instances of co-occurrence, the result- two previously distinct belief modules would ing associative bridge might gradually col- become likelier to collapse into a single mod- lapse the distance between them. Once I come ule after the emergence of new “bridging” around to drinking lattes through this prac- associations linking beliefs located in the pre- tice’s association with liberalism, I might also viously separate modules. These bridges play proceed to adopt a series of other beliefs and a role analogous to the “weak ties” described practices associated with latte-drinking— by Granovetter (1973), bridging ties spanning such as driving a hybrid electric car or listen- structural holes (Burt 1992; for a related dis- ing to indie rock. cussion of cultural holes, see Pachucki and Yet, there is nothing inherently irreversible Breiger 2010), and “rewired” ties that link about these dynamics. It would also be pos- otherwise disconnected segments of small- sible via the reverse process for a previously world networks (Watts and Strogatz 1998). In cohesive module of beliefs to split into two other words, these are associations or align- through the emergence of a new fault line— ments between beliefs that one would not that is, when one or more associations hold- otherwise expect to be associated, given their ing together beliefs in the same module different locations in the overall network. The weaken, causing the module to disintegrate. addition of such distance-spanning bridges Particular beliefs can move over time from collapses the network into fewer modules by one module to another when the package of reducing the distance between otherwise- other beliefs with which they are associated disconnected sets of beliefs. For example, become “rewired.” For example, it is 528 American Sociological Review 85(3) reasonable to suspect that in the wake of the between polarized political and social identi- Republican Party’s tacit acceptance of Rus- ties can be seen as contributing factors to a sian interference in the 2016 presidential politics steeped in racial resentment (Bon- election, the set of beliefs associated with ikowski 2017; McVeigh and Estep 2019; Americans’ liking for Russia (a question fea- Sides, Tesler, and Vavreck 2017), cultural tured in the General Social Survey) may backlash (Hochschild 2016; Skocpol and Wil- become quite different than during the height liamson 2012), and demagoguery encouraged of Cold War hawkishness. by a lack of binding legitimate institutions (Hahl, Kim, and Zuckerman Sivan 2018). Bail and colleagues (2018) even cast doubt Implications for the Pluralism Debate on political pluralists’ favored argument that Conceived as a structure of interrelated atti- more cross-cutting interactions would lead to tudes, opinions, and beliefs, the U.S. belief greater mutual understanding and less polari- network as captured in General Social Survey zation. The researchers randomly assigned a (GSS) data experienced increasing consolida- sample of Twitter users to follow a “bot” tion over time. Yet, even knowing these account that would share content from the trends, it is difficult to know what baseline to opposing side of the political divide. Rather compare them against. Is the United States than decreasing polarization, they found that today polarized or pluralistic? One’s answer exposure to opposing viewpoints pushed peo- likely depends on how one perceives the pres- ple even further toward their own side. One ent sociopolitical moment. For those inclined might ask, however, whether this polarizing to feel U.S. society and the political system is effect of outgroup exposure would material- currently paralyzed by partisan conflict ize if the political and social identities associ- between polarized groups, the evidence pre- ated with the two sides had not already sented here supports the hypothesis that atti- become increasingly totalizing over recent tudes have indeed shifted in a way likely to decades, as Mason (2018) argues. intensify such conflict. For others inclined to Perhaps interactions with members of the respond that Americans retain many cross- political outgroup would work as pluralists cutting and pluralistic attitudes, the evidence expect if the population featured enough presented here may suggest that, although this cross-cutting cleavages, but these effects dis- pluralism is apparently on the decline, a num- sipate once those cross-cutting cleavages ber of cross-cutting dimensions remain. have collapsed to form more encompassing History may not furnish us with enough partisan identities with little common ground examples to validate the harshest warnings between them. Whereas overlapping identi- one could offer based on these findings. Yet, ties generate common ground and the possi- neither would this history make us particularly bility of mutual understanding, the existence optimistic. The starting point of analysis here of polarized “super-identities” feeds affective follows the widespread conflict and violence polarization by leading people to simplify the of the late 1960s and early 1970s, often cited outgroup (e.g., as an evil force unworthy of as the closest comparison to the current civil engagement) and attach negative stereo- moment. The toll in human carnage of this types (Brewer and Pierce 2005). Theories previous era of polarization—assassinations, from social psychology (e.g., Roccas and pitched street battles between police and pro- Brewer 2002) suggest such a collapse of pre- testers, and corruption driven by partisan viously cross-cutting identities will likely malice—is well-documented (e.g., Perlstein reduce outgroup tolerance and might thereby 2008), and the fact that polarization proceeded account for the increasingly harsh tenor of apace from this starting point is suggestive of political disagreement perceived by scholars a potential for further deepening conflict. As of affective polarization (Iyengar, Sood, and recent scholarship highlights, the deep fissures Lelkes 2012; Mason 2018). DellaPosta 529

At the same time, the present study quali- instance, the pairing of premarsx, a question fies previous work on identity-based polariza- eliciting respondents’ views on the morality tion by suggesting that polarization has also of premarital sex, and gunlaw, which asks occurred at the level of beliefs and issue posi- whether respondents would favor a law tions, once we conceive more fully of the requiring a police permit before buying a gun, structure of polarization among beliefs. This would constitute a pair indexed by j. Then, is most clearly evident in the fact that some the 23 (in this case) measured correlations markers of belief consolidation persist in the between responses to these two items between over-time trends even after adjusting for par- 1972 and 2016 would represent observations tisan and ideological identity. This observa- nested within the pairing. The resulting tion should complement ample existing mixed-effects model across all such pairs j evidence for polarization in social and politi- takes the form cal identity, which can be a cause of issue polarization (i.e., when people form opinions rt=+αβ+ s jt, jjjt, (1) based on signals from their co-partisans) but also a consequence of issue polarization (i.e., where |r|j,t is the magnitude of the observed when consistent identities are easier to form correlation between items in pairing j in year and maintain due to their being rooted in t; αj is an intercept that varies by pair j; βj is coherent sets of issue positions). a time trend that also varies by pair j; and sj,t According to the political pluralism narra- captures residual variation across different tive, a high-dimensional public space in years within the same pair of items. which agreement on some dimensions bal- The purpose of this model is to produce ances against disagreement on others is self- year-specific estimates of the correlations reinforcing as long as divisions remain between items while adjusting for the fact cross-cutting and overlapping rather than that pairs of items appear in different years absolute and all-encompassing. From this and with varying frequency (Baldassarri and perspective, the analysis presented here car- Gelman 2008; DellaPosta et al. 2015). The ries troubling implications because it suggests main strength of the mixed-effects strategy is the number and plurality of cross-cutting that it allows for estimation of separate inter- alignments in U.S. public opinion has declined. cepts and time trends for each pair of items, Such collapses in pluralistic structure lead to avoiding the assumption of homogeneity in disagreements becoming increasingly all- the tendency for any two items in a pair to encompassing because there are fewer miti- become more or less aligned over time. The gating domains in which two opponents may assumption of linearity in the trend toward find agreement. Conceived in terms of struc- increasing or decreasing alignment for any tures of attitudes, the trend toward an increas- given pair of items remains somewhat restric- ingly polarized and consolidated opinion tive. However, estimation of the linear trend space—and thus one that amplifies rather is still sufficient for the purpose of assessing than tamping down conflict—may be clearer long-term trends in the structure of the and less elusive than previous studies would correlations. suggest. Table A1 gives the results of the mixed- effects model when all zero-order correla- tions between beliefs are included as well as Appendix: Statistical when the model uses partial correlations to Model adjust for the influence of political ideology In the multilevel framework, year-specific and party identification on each bivariate bivariate correlations between any given pair association. In both cases, the typical pair of of items x and y are treated as observations beliefs would have seen only very modest nested within item-pairs indexed by j. For changes over time. In the baseline condition, 530 American Sociological Review 85(3)

Figure A1. Yearly Numbers of Observed Correlations Before and After Sample Restriction Note: Each bar chart shows the number of correlations observed for each year of the GSS. Panel A shows all observed correlations; Panel B shows correlations retained for analysis after dropping those for items that co-appeared fewer than five times. the typical correlation increased in magni- reversed but the coefficient remains substan- tude by less than .01 with each decade, mean- tively small in magnitude. In both cases, the ing a typical pair of items would become variance across pairs of items for this time slightly more strongly correlated between trend far outweighs the magnitude of the 1972 and 2016. For the ideology-controlled average or typical time trend across all such condition, the sign of the time coefficient is pairs. DellaPosta 531

Table A1. Multilevel Mixed-Effects Model Predicting Magnitude of Correlations between Beliefs

Zero-Order Correlations Partial Correlations

b s.e. b s.e.

Fixed-Effects Intercept .09 <.01*** .08 <.01*** Time (decades) <.01 <.01*** >–.01 <.01***

Residual SD Intercepts .08 .08 Time trends .01 .01 Within-pairs .03 .03

Note: N = 202,928 unique correlations nested within 14,910 unique pairs of items for the model using all zero-order bivariate correlations; N = 191,890 unique correlations nested within 14,476 unique pairs for the partial-correlation model that excludes and controls for political ideology and party identification. Estimated coefficients and standard errors are shown for fixed effects, and residual standard deviations are shown for random effects. Time is grand-mean centered. ***p < .001 (two-tailed tests).

Acknowledgments 3. Boutyline (2017) argues for a similar approach (called correlational class analysis [CCA]) to the I benefited greatly from helpful comments on earlier ver- same problem of inductively uncovering groups sions of this paper from Gary Adler, Dave Baker, Marjan organizing their beliefs in similar ways (see also Davoodi, Fedor Dokshin, Liam Essig, Diane Felmlee, Daenekindt, de Koster, and van der Waal 2017), Roger Finke, Sang Won Han, Melissa Hardy, Minjae finding that this approach out-performs RCA across Kim, Michael Macy, John McCarthy, Victor Nee, Eric a wide range of starting assumptions. Neither RCA Plutzer, Charles Seguin, and participants in the Culture nor CCA is used in the present study, however. and Politics Workshop as well as the Innovative Methods 4. More specifically, the main difference is that belief Working Group at Penn State. Marjan Davoodi and Kelli networks represent correlations between beliefs in a Knipe provided helpful research assistance in the early population aggregate, rather than directly clustering stages of this project. I also gratefully acknowledge sup- individual survey respondents as RCA and related port during the writing of this article from the Department methods do. To maintain a constant set of beliefs of Sociology and Criminology at Penn State. Previous in the network, person-clustering approaches like versions of this paper were presented at the 2018 meeting RCA would seem to require modeling and imput- of the International Network of Analytical Sociologists, ing missing “cells” in the respondent-belief matrix the 2019 meeting of the American Sociological Associa- separately for each individual survey respondent tion, and the Economic Sociology Colloquium at Cornell every time a particular survey item did not appear University in 2019. The usual disclaimers apply. for a given cross-section of the survey. Because the assumptions required for such an exercise would be ORCID iD difficult to justify, one would likely have to restrict analysis to the small and selective set of belief items Daniel DellaPosta https://orcid.org/0000-0002 that appear in every cross-section of the survey. -5307-413X 5. Notably, however, it is not clear that the U.S. population features much of this heterogeneity. Boutyline and Vaisey (2017) examined 44 subpopu- Notes lations and found remarkable consistency in belief- 1. This is true, at least, of the population in aggregate. network structure across these groups; some groups However, there is evidence that polarization of atti- exhibited greater or less amounts of overall orga- tudes has increased among some groups, such as nization and cohesion in their belief networks, but strong partisans and socioeconomic elites (Baldas- the beliefs that held relatively central or peripheral sarri and Gelman 2008; Evans 2003). positions in the networks were virtually the same. 2. Dimensional methods have also been influential in 6. The “oil spill” model is in some ways anticipated studies of congressional polarization, most nota- by Converse’s (1964) original formulation of con- bly with the widespread use of the DW-Nominate straint, although it departs from the measurement of scores of political ideology used by Poole and constraint offered by him and subsequent scholars. Rosenthal (1984). Converse (1964:208) points out that “belief systems 532 American Sociological Review 85(3)

may . . . be compared in a rough way with respect ever, Salganik and Levy (2015) proposed a new to the range of objects that are referents for the set of survey techniques, which they term “Wiki ideas and attitudes in the system.” An increase in Surveys,” that would allow for a greater balance the range of beliefs that come to be incorporated in between researcher-generated and respondent- a structure of opinion alignment might be usefully generated content in survey construction. If such an thought of as the breadth of constraint or alignment. approach were scaled to the level of repeated and 7. The two examples in Figure 2 do not exhaust nationally representative surveys of attitudes, this the possibilities for how belief networks may be might allow us to more fully overcome the limita- arranged. As mentioned previously, a belief network tions imposed by researcher selectivity in survey could lack a modular structure, meaning attempts construction. to partition the network into modules would pro- 11. I only included correlations based on more than 100 duce results that do not much describe the actual responses, however, to ensure sufficient precision arrangement of cross-issue alignments. Such a in the observations used for subsequent analysis. “non-modular” structure may be cross-cutting at the 12. Appendix Figure A1 shows the number of observed level of individual issues, rather than the example correlations from each edition of the GSS before in Panel A of Figure 2 where links between beliefs and after this sample restriction process. One con- cut across modules. Furthermore, a belief network sequence of dropping pairs of items that seldom could be organized around a single belief rather appear together is that the resulting analytic sample than a module containing multiple beliefs. For retains a relatively more even number of correla- example, Boutyline and Vaisey (2017) found politi- tions from year to year than it would otherwise. cal ideology played a key centralizing role among 13. Similarly, because the focus will be on changes over the smaller set of issues they analyzed. Yet, the a long time in the structure of relationships among a centralizing role of any single belief in the network fixed set of GSS items, I did not engage in any pro- is notably distinct from what I call consolidation; a cedure to combine substantively linked items into more consolidated belief network does not neces- common scales or dimensions. Even if, for exam- sarily become more centralized in graph-theoretic ple, two different questions about abortion were terms (see Freeman 1978:227ff), because the more so strongly correlated as to suggest they belong to beliefs getting pulled into large central modules can one underlying scale, we might still be interested in offset the ability of any one or a few central beliefs whether the correlation between the two items grew to dominate connectivity in the network. stronger or weaker over time. As with the decision 8. In fact, the simplest way to illustrate such a ten- to include any GSS item that could be considered dency empirically would be to generate a consoli- an opinion, belief, or attitude, the goal is to analyze dated network such as the one in Panel B and then the data in a way that loses as little information as randomly “rewire” all or most edges in that net- possible, allowing patterns in the data to emerge work (Maslov and Sneppen 2002). The result is a naturally with few externally-imposed constraints structure with the same average strength of correla- from the researcher. tion but in which the network does not divide into 14. Replication data and code for all analyses are avail- modular groupings. As the patterning of ties comes able on the author’s “dataverse” at https://dataverse. closer to a pattern of “random association,” the harvard.edu/dataverse/dellaposta. belief network will increasingly lack any coherent 15. Figure 4 and all other network visualizations in this structure linking the beliefs as a whole. Formally article were generated using the Fruchterman-Rein- speaking, modularity is measured by comparing the gold spring-embedding algorithm (Fruchterman extent of within-module ties in an observed network and Reingold 1991).

to the extent of within-module ties that would be 16. Formally, the index is computed as C = 1/(2ΣiPi) expected in a randomly permuted version of the – 1 where i indicates the size rank of a given belief

same network, one that features the same nodes and module, and Pi is the proportion of nodes belonging degree distribution but in which the arrangement of to the module with rank i. ties or edges is random (Newman 2006). 9. In a limiting case, consolidation could also increase due to the emergence of a consensual “monoculture” References in which everyone agrees with everyone else (Arendt Abramowitz, Alan I., and Kyle L. Saunders. 1998. “Ideo- 1968). In the U.S. context, this may seem to be a logical Realignment in the U.S. Electorate.” Journal relatively remote possibility. In robustness checks, I of Politics 60(3):634–52. found that relatively few beliefs in the GSS became Abramowitz, Alan I., and Kyle L. Saunders. 2008. “Is more consensual over time, and these beliefs did Polarization a Myth?” Journal of Politics 70(2):542– not account for the time trends reported here. These 55. analyses are available upon request from the author. Abrams, Samuel J., and Morris P. Fiorina. 2012. “The Big 10. I am not aware of existing survey data sources that Sort That Wasn’t: A Skeptical Reexamination.” PS: could overcome this limitation. Recently, how- Political Science and Politics 45(2):203–10. DellaPosta 533

Achen, Christopher H., and Larry M. Bartels. 2016. Brooks, Clem, and Jeff Manza. 1997. “Social Cleavages Democracy for Realists: Why Elections Do Not Pro- and Political Alignments: U.S. Presidential Elec- duce Responsive Government. Princeton, NJ: Princ- tions, 1960 to 1992.” American Sociological Review eton University Press. 62(6):937–46. Alwin, Duane F. 2007. Margins of Error: A Study of Reli- Burt, Ronald S. 1992. Structural Holes: The Social Struc- ability in Survey Measurement. Hoboken, NJ: Wiley- ture of Competition. Cambridge, MA: Harvard Uni- Interscience. versity Press. Andris, Clio, David Lee, Marcus J. Hamilton, Mauro Centola, Damon. 2015. “The Social Origins of Networks Martino, Christian E. Gunning, and John Armistead and Diffusion.” American Journal of Sociology Selden. 2015. “The Rise of Partisanship and Super- 120(5):1295–338. Cooperators in the U.S. House of Representatives.” Converse, Philip E. 1964. “The Nature of Belief Systems in PLoS ONE 10(4):e0123507. Mass Publics.” Pp. 206–61 in Ideology and Discontent, Arendt, Hannah. 1968. “Truth and Politics.” Pp. 227–64 Vol. 5, International Yearbook of Political Behavior in Between Past and Future: Eight Exercises in Polit- Research, edited by D. Apter. New York: Free Press. ical Thought. New York: Viking Press. Cowan, Sarah K., and Delia Baldassarri. 2018. “‘It Could Bail, Christopher A., Lisa P. Argyle, Taylor W. Brown, Turn Ugly’: Selective Disclosure of Attitudes in Polit- John P. Bumpus, Haohan Chen, M. B. Fallin Hunza- ical Discussion Networks.” Social Networks 52:1–17. ker, Jaemin Lee, Marcus Mann, Friedolin Merhout, Daenekindt, Stijn, Willem de Koster, and Jeroen van der and Alexander Volfovsky. 2018. “Exposure to Oppos- Waal. 2017. “How People Organise Cultural Atti- ing Views on Social Media Can Increase Political tudes: Cultural Belief Systems and the Populist Radi- Polarization.” Proceedings of the National Academy cal Right.” West European Politics 40(4):791–811. of Sciences 115(37):9216–21. Dahl, Robert. 1961. Who Governs? Democracy and Baldassarri, Delia, and Peter Bearman. 2007. “Dynam- Power in an American City. New Haven, CT: Yale ics of Political Polarization.” American Sociological University Press. Review 72(5):784–811. Dalege, Jonas, Denny Borsboom, Frenk van Harreveld, Baldassarri, Delia, and Mario Diani. 2007. “The Integra- and Han L. J. van der Maas. 2017. “Network Analysis tive Power of Civic Networks.” American Journal of on Attitudes: A Brief Tutorial.” Social Psychological Sociology 113(3):735–80. and Personality Science 8(5):528–37. Baldassarri, Delia, and Andrew Gelman. 2008. “Partisans DellaPosta, Daniel, and Michael Macy. 2015. “The Cen- without Constraint: Political Polarization and Trends ter Cannot Hold: Networks, Echo Chambers, and in American Public Opinion.” American Journal of Polarization.” Pp. 86–104 in Order on the Edge of Sociology 114(2):408–46. Chaos: Social Psychology and the Problem of Social Baldassarri, Delia, and Amir Goldberg. 2014. “Neither Order, edited by E. J. Lawler, S. R. Thye, and J. Yoon. Ideologues nor Agnostics: Alternative Voters’ Belief Cambridge, UK: Cambridge University Press. System in an Age of Partisan Politics.” American DellaPosta, Daniel, Yongren Shi, and Michael Macy. Journal of Sociology 120(1):45–95. 2015. “Why Do Liberals Drink Lattes?” American Bishop, Bill. 2008. The Big Sort: Why the Clustering of Journal of Sociology 120(5):1473–511. Like-Minded America Is Tearing Us Apart. Boston: DiMaggio, Paul. 1997. “Culture and Cognition.” Annual Houghton Mifflin. Review of Sociology 23(1):263–87. Blau, Peter M., and Joseph E. Schwartz. 1984. Crosscut- DiMaggio, Paul. 2011. “Cultural Networks.” Pp. 286– ting Social Circles: Testing a Macrostructural Theory 300 in The Sage Handbook of Social Network Analy- of Intergroup Relations. Orlando, FL: Free Press. sis. Thousand Oaks, CA: Sage. Bonikowski, Bart. 2017. “Ethno-Nationalist Populism DiMaggio, Paul, John Evans, and Bethany Bryson. and the Mobilization of Collective Resentment.” Brit- 1996. “Have Americans’ Social Attitudes Become ish Journal of Sociology 68(S1):S181–S213. More Polarized?” American Journal of Sociology Bonikowski, Bart, and Paul DiMaggio. 2016. “Varieties 102(3):690–755. of American Popular Nationalism.” American Socio- DiMaggio, Paul, and Amir Goldberg. 2018. “Searching logical Review 81(5):949–80. for Homo Economicus: Variation in Americans’ Con- Boutyline, Andrei. 2017. “Improving the Measurement struals of and Attitudes toward Markets.” European of Shared Cultural Schemas with Correlational Class Journal of Sociology 59(2):151–89. Analysis: Theory and Method.” Sociological Science DiMaggio, Paul, Ramina Sotoudeh, Amir Goldberg, and 4(15):353–93. Hana Shepherd. 2018. “Culture Out of Attitudes: Boutyline, Andrei, and Stephen Vaisey. 2017. “Belief Relationality, Population Heterogeneity and Attitudes Network Analysis: A Relational Approach to Under- toward Science and Religion in the U.S.” Poetics standing the Structure of Attitudes.” American Jour- 68:31–51. nal of Sociology 122(5):1371–447. Durkheim, Emile. 1933. The Division of Labor in Soci- Brewer, Marilynn B., and Kathleen P. Pierce. 2005. ety. New York: Free Press. “Social Identity Complexity and Out-group Toler- Evans, John H. 2003. “Have Americans’ Attitudes ance.” Personality and Social Psychology Bulletin Become More Polarized? An Update.” Social Science 31(3):428–347. Quarterly 84(1):71–90. 534 American Sociological Review 85(3)

Festinger, Leon. 1957. A Theory of Cognitive Disso- Hall, Marshall, and Nicolaus Tideman. 1967. “Measures nance. Stanford, CA: Stanford University Press. of Concentration.” Journal of the American Statisti- Fiorina, Morris P. 2004. Culture War? The Myth of a cal Association 62(317):162–68. Polarized America. New York: Pearson Longman. Heider, Fritz. 1946. “Attitudes and Cognitive Organiza- Fiorina, Morris P., and Samuel J. Abrams. 2008. “Politi- tion.” Journal of Psychology: Interdisciplinary and cal Polarization in the American Public.” Annual Applied 21(1):107–12. Review of Political Science 11(1):563–88. Hetherington, Marc J. 2001. “Resurgent Mass Partisan- Fischer, Claude S., and Greggor Mattson. 2009. “Is ship: The Role of Elite Polarization.” American Polit- America Fragmenting?” Annual Review of Sociology ical Science Review 95(3):619–31. 35(1):435–55. Hetherington, Marc J., and Jonathan Weiler. 2018. Flache, Andreas, and Michael W. Macy. 2011. “Small Prius or Pickup? How the Answers to Four Simple Worlds and Cultural Polarization.” Journal of Math- Questions Explain America’s Great Divide. Boston: ematical Sociology 35(1–3):146–76. Houghton Mifflin Harcourt. Fleishman, John A. 1988. “Attitude Organization in the Hill, Seth J., and Chris Tausanovitch. 2015. “A Discon- General Public: Evidence for a Bidimensional Struc- nect in Representation? Comparison of Trends in ture.” Social Forces 67(1):159–84. Congressional and Public Polarization.” Journal of Freeman, Linton C. 1978. “Centrality in Social Net- Politics 77(4):1058–75. works Conceptual Clarification.” Social Networks Hochschild, Arlie Russell. 2016. Strangers in Their Own 1(3):215–39. Land. New York: The New Press. Friedkin, Noah E., Anton V. Proskurnikov, Roberto Hunter, James Davison. 1991. Culture Wars: The Strug- Tempo, and Sergey E. Parsegov. 2016. “Network Sci- gle to Define America. New York: Basic Books. ence on Belief System Dynamics under Logic Con- Hunzaker, M. B. Fallin. 2016. “Cultural Sentiments and straints.” Science 354(6310):321–26. Schema-Consistency Bias in Information Transmis- Fruchterman, Thomas M. J., and Edward M. Rein- sion.” American Sociological Review 81(6):1223–50. gold. 1991. “Graph Drawing by Force-Directed Hunzaker, M. B. Fallin, and Lauren Valentino. 2019. Placement.” Software: Practice and Experience “Mapping Cultural Schemas: From Theory to 21(11):1129–64. Method.” American Sociological Review 84(5):950– Gates, Kathleen M., Teague Henry, Doug Steinley, and 81. Damien A. Fair. 2016. “A Monte Carlo Evaluation Iyengar, Shanto, Gaurav Sood, and Yphtach Lelkes. of Weighted Community Detection Algorithms.” 2012. “Affect, Not Ideology: A Social Identity Per- Frontiers in Neuroinformatics 10:45 (https://doi.org/ spective on Polarization.” Public Opinion Quarterly 10.3389/fninf.2016.00045). 76(3):405–31. Gelman, Andrew, and Jennifer Hill. 2007. Data Analysis Jost, John T., Christopher M. Federico, and Jaime L. Using Regression and Multi-level/Hierarchical Mod- Napier. 2009. “Political Ideology: Its Structure, Func- els. Cambridge, UK: Cambridge University Press. tions, and Elective Affinities.” Annual Review of Psy- Ghaziani, Amin, and Delia Baldassarri. 2011. “Cultural chology 60(1):307–37. Anchors and the Organization of Differences.” Amer- Kiley, Kevin, and Stephen Vaisey. 2020. “Measuring Sta- ican Sociological Review 76(2):179–206. bility and Change in Personal Culture Using Panel Goldberg, Amir. 2011. “Mapping Shared Understand- Data.” American Sociological Review 85(3):477– ings Using Relational Class Analysis: The Case of the 506. Cultural Omnivore Reexamined.” American Journal Kinder, Donald R., and Nathan P. Kalmoe. 2017. Neither of Sociology 116(5):1397–436. Liberal nor Conservative: Ideological Innocence in Goldberg, Amir, and Sarah K. Stein. 2018. “Beyond the American Public. Chicago: University of Chicago Social Contagion: Associative Diffusion and the Press. Emergence of Cultural Variation.” American Socio- Kunda, Ziva. 1990. “The Case for Motivated Reasoning.” logical Review 83(5):897–932. Psychological Bulletin 108(3):480–98. Golino, Hudson F., and Sacha Epskamp. 2017. “Explor- Lakoff, George. 1996. Moral Politics: How Liberals and atory Graph Analysis: A New Approach for Estimat- Conservatives Think. Chicago: University of Chicago ing the Number of Dimensions in Psychological Press. Research.” PLoS ONE 12(6):e0174035. Layman, Geoffrey C., and Thomas M. Carsey. 2002. Granovetter, Mark S. 1973. “The Strength of Weak Ties.” “Party Polarization and ‘Conflict Extension’ in the American Journal of Sociology 78(6):1360–80. American Electorate.” American Journal of Political Hahl, Oliver, Minjae Kim, and Ezra W. Zuckerman Science 46(4):786–802. Sivan. 2018. “The Authentic Appeal of the Lying Lazarsfeld, Paul F., Bernard Berelson, and Hazel Gaudet. Demagogue: Proclaiming the Deeper Truth about 1944. The People’s Choice: How the Voter Makes Up Political Illegitimacy.” American Sociological His Mind in a Presidential Campaign. New York: Review 83(1):1–33. Duell, Sloan, and Pearce. Haidt, Jonathan. 2012. The Righteous Mind: Why Good Lee, Frances E. 2009. Beyond Ideology: Politics, Prin- People Are Divided by Politics and Religion. New ciples, and Partisanship in the U.S. Senate. Chicago: York: Pantheon. University of Chicago Press. DellaPosta 535

Lee, Monica, and John Levi Martin. 2018. “Doorway to Mutz, Diana C. 2002. “Cross-Cutting Social Networks: the Dharma of Duality.” Poetics 68:18–30. Testing Democratic Theory in Practice.” American Levendusky, Matthew. 2009. The Partisan Sort: How Political Science Review 96(1):111–26. Liberals Became Democrats and Conservatives Mutz, Diana C., and Jahnavi S. Rao. 2018. “The Real Became Republicans. Chicago: The University of Reason Liberals Drink Lattes.” PS: Political Science Chicago Press. and Politics 51(4):762–67. Levitsky, Steven, and Daniel Ziblatt. 2018. How Democ- Newman, M. E. J. 2006. “Modularity and Community racies Die. New York: Crown. Structure in Networks.” Proceedings of the National Lipset, Seymour Martin. 1963. Political Man: The Social Academy of Sciences 103(23):8577–82. Bases of Politics. New York: Anchor Books. Pachucki, Mark A., and Ronald Breiger. 2010. “Cul- Lipset, Seymour Martin, and Stein Rokkan. 1967. Party tural Holes: Beyond Relationality in Social Net- Systems and Voter Alignments: Cross-National Per- works and Culture.” Annual Review of Sociology spectives. New York: Free Press. 36(1):205–24. Lizardo, Omar. 2014. “Omnivorousness as the Bridging Park, Barum. 2018. “How Are We Apart? Continuity and of Cultural Holes: A Measurement Strategy.” Theory Change in the Structure of Ideological Disagreement and Society 43(3):395–419. in the American Public, 1980–2012.” Social Forces Lizardo, Omar. 2017. “Improving Cultural Analysis: 96(4):1757–84. Considering Personal Culture in its Declarative and Patterson, Orlando. 2014. “Making Sense of Culture.” Nondeclarative Modes.” American Sociological Annual Review of Sociology 40(1):1–30. Review 82(1):88–115. Perlstein, Rick. 2001. Before the Storm: Barry Goldwater Macy, Michael Walton, James A. Kitts, Andreas Flache, and the Unmaking of the American Consensus. New and Stephen Benard. 2003. “Polarization in Dynamic York: Nation Books. Networks: A Hopfield Model of Emergent Structure.” Perlstein, Rick. 2008. Nixonland: The Rise of a President Pp. 162–73 in Dynamic Social Network Modeling and the Fracturing of America. New York: Scribner. and Analysis: Workshop Summary and Papers, edited Pons, Pascal, and Matthieu Latapy. 2006. “Computing by R. Breiger, K. Carley, and P. Pattison. Washington, Communities in Large Networks Using Random DC: National Academies Press. Walks.” Journal of Graph Algorithms and Applica- Martin, John Levi. 2000. “The Relation of Aggregate Sta- tions 10(2):191–218. tistics on Beliefs to Culture and Cognition.” Poetics Poole, Keith T., and Howard Rosenthal. 1984. “The 28(1):5–20. Polarization of American Politics.” Journal of Poli- Martin, John Levi. 2002. “Power, Authority, and the Con- tics 46(4):1061–79. straint of Belief Systems.” American Journal of Soci- Puetz, Kyle. 2017. “Fields of Mutual Alignment: A Dual- ology 107(4):861–904. Order Approach to the Study of Cultural Holes.” Marx, Karl, and Friedrich Engels. 1978. The Marx- Sociological Theory 35(3):228–60. Engels Reader. New York: Norton. Roccas, Sonia, and Marilynn B. Brewer. 2002. “Social Maslov, Sergei, and Kim Sneppen. 2002. “Specificity and Identity Complexity.” Personality and Social Psy- Stability in Topology of Protein Networks.” Science chology Review 6(2):88–106. 296(5569):910–13. Salganik, Matthew J., and Karen E. C. Levy. 2015. “Wiki Mason, Lilliana. 2018. Uncivil Agreement: How Politics Surveys: Open and Quantifiable Social Data Collec- Became Our Identity. Chicago: The University of tion.” PLoS ONE 10(5):e0123483. Chicago Press. Sides, John, Michael Tesler, and Lynn Vavreck. 2017. McCarty, Nolan, Keith T. Poole, and Howard Rosenthal. “The 2016 U.S. Election: How Trump Lost and 2006. Polarized America: The Dance of Ideology and Won.” Journal of Democracy 28(2):34–44. Unequal Riches. Cambridge, MA: MIT Press. Simmel, Georg. 1955. Conflict and the Web of Group- McPherson, Miller. 2004. “A Blau Space Primer: Prole- Affiliations. New York: Free Press. gomenon to an Ecology of Affiliation.” Industrial and Skocpol, Theda, and Vanessa Williamson. 2012. The Tea Corporate Change 13(1):263–80. Party and the Remaking of Republican Conservatism. McPherson, Miller, Lynn Smith-Lovin, and James M. Oxford, UK: Oxford University Press. Cook. 2001. “Birds of a Feather: Homophily in Social Stouffer, Samuel A. 1955. Communism, Conformity, and Networks.” Annual Review of Sociology 27(1):415–44. Civil Liberties: A Cross-Section of the Nation Speaks McVeigh, Rory, and Kevin Estep. 2019. The Politics of Its Mind. Garden City, NJ: Doubleday. Losing: Trump, the Klan, and the Mainstreaming of Truman, David. 1951. The Governmental Process: Politi- Resentment. New York: Columbia University Press. cal Interests and Public Opinion. New York: Alfred Mohr, John W. 1998. “Measuring Meaning Structures.” A. Knopf. Annual Review of Sociology 24(1):345–70. Vaisey, Stephen, and Omar Lizardo. 2010. “Can Cul- Mohr, John W., and Harrison C. White. 2008. “How to Model tural Worldviews Influence Network Composition?” an Institution.” Theory and Society 37(5):485–512. Social Forces 88(4):1595–618. Moody, James, and Peter J. Mucha. 2013. “Portrait Vilhena, Daril A., Jacob G. Foster, Martin Rosvall, Jevin of Political Party Polarization.” Network Science D. West, James Evans, and Carl T. Bergstroma. 2014. 1(2):119–21. “Finding Cultural Holes: How Structure and Culture 536 American Sociological Review 85(3)

Diverge in Networks of Scholarly Communication.” Daniel DellaPosta is an Assistant Professor of Sociology Sociological Science 1:221–38. and Social Data Analytics at the Pennsylvania State Uni- Wasserman, Stanley, and Katherine Faust. 1994. Social versity. His research applies network analysis across Network Analysis: Methods and Applications. Cam- diverse research contexts, with a core focus on the mecha- bridge, UK: Cambridge University Press. nisms giving rise to intergroup cooperation and conflict, Watts, Duncan J., and Steven H. Strogatz. 1998. “Collec- political and attitudinal alignments, and social and eco- tive Dynamics of ‘Small-World’ Networks.” Nature nomic organization. His previous work has been published 393(6684):440–42. in the American Journal of Sociology, Social Forces, Shi, Yongren, Kai Mast, Ingmar Weber, Agrippa Kellum, Social Networks, and several other outlets. His research and Michael Macy. 2017. “Cultural Fault Lines and has also received awards from the American Sociological Political Polarization.” Pp. 213–17 in WebSci ’17: Association’s sections on Mathematical Sociology, Eco- Proceedings of the 2017 ACM on Web Science Con- nomic Sociology, and Rationality and Society. ference (https://doi.org/10.1145/3091478.3091520).