Running head: “ATTITUDE AGREEMENT IN SOCIAL NETWORKS” Real and Perceived Attitude Agreement in Social Networks

It is often asserted that friends and acquaintances have more similar beliefs and attitudes than strangers; yet empirical studies disagree over exactly how much diversity of opinion exists within local social networks, and relatedly, how much awareness individuals have of their neighbors’ views. This paper reports results from a network survey, conducted on the Facebook social networking platform, in which participants were asked about their own political attitudes, as well as their beliefs about their friends’ attitudes. Although considerable attitude similarity exists among friends, the results show that friends disagree considerably more than they think they do. In particular, friends are typically unaware of their disagreements even when they say they discuss the topic, suggesting that discussion is not the primary means by which friends infer each other’s views on particular issues. Rather, it appears that respondents infer opinions in part by relying on stereotypes of their friends, and in part by projecting their own views. The resulting gap between real and perceived agreement may have implications for the dynamics of political polarization and theories of social influence in general. Attitude Agreement in Social Networks, p. 3 The homophily principle, that “like associates with like,” is one of the more pervasive empirical regularities of the social world. With respect to a wide range of socio-demographic attributes such as education, income, gender, race, and age, numerous studies have found that couples, friends, coworkers, and members of informal organizations all tend to be more similar than randomly selected members of the same population (Kossinets & Watts, 2009; Lazarsfeld & Merton, 1954; McPherson, Smith-Lovin, & Cook, 2001). Given this plethora of findings, it is tempting to conclude that the same principle applies to psychological attributes like beliefs and attitudes as well—a conclusion that is consistent with theoretical arguments that likeminded individuals may prefer to associate with one another (Festinger, 1957), and experimental findings that processes of social influence (Cialdini & Goldstein, 2004; Deutsch & Gerard, 1955) tend to breed conformity (Asch, 1955; Bond & Smith, 1996; Sherif, 1937). In particular, a number of authors have argued recently that homophily with respect to political attitudes and beliefs is increasing in America, with individuals sorting themselves into likeminded communities that serve as echo chambers for their existing opinions (Bishop, 2008; Sunstein, 2009). In contrast with these studies, however, other empirical evidence shows that conflicting attitudes can persist within social networks (Nowak, Szamrej, & Latané, 1990; Huckfeldt, Johnson, & Sprague, 2004), and that beliefs themselves are poorly correlated across different issues (Baldassari & Gelman, 2008). Recently it has been suggested that these seemingly contrary findings may derive from a difference between real and perceived attitude agreement—specifically that people estimate that they are more similar to their friends than they really are (Baldassari & Bearman, 2007; Gelman, Park, Shor, Bafumi, & Cortina, 2008; Robbins & Krueger, 2005). If, for example, it is Attitude Agreement in Social Networks, p. 4 true that friends generally avoid discussing politics (as is frequently recommended for polite company), then in the absence of specific information to the contrary, survey respondents may systematically overestimate the probability that their friends agree with them—an example of the “false consensus” effect (Krueger & Clement, 1994; Marks & Miller, 1987; Ross, Greene, & House, 1977). Alternatively, individuals may discuss only those matters on which they expect to find agreement, or discuss only with friends with whom they expect to agree, either of which could generate a sample of opinions that is biased towards agreement. Or finally, individuals may moderate or misrepresent their true views when conversing with friends, precisely so as to avoid the appearance of disagreement, and thereby reduce the likelihood of conflict. Although different, all these mechanisms would lead people to perceive greater agreement among their friends, and also perceive greater alignment of views on different topics (for example, that if A agrees with B on topic X then he also agrees with B on topic Y) than is actually present. In principle, it is straightforward to measure the difference between real and perceived agreement, by using some variant of a “network” or “snowball” survey in which, for each pair of friends (u, v), and each topic q, one records u’s response to q, v’s response to q, and u’s perception of v’s response to q. For example, in Laumann’s (1969) study of urban men, roughly one in four respondents was asked to name at random one of three friends described in their responses to be contacted for a subsequent interview. Laumann subsequently reported on the accuracy of participants’ knowledge of friends’ demographic characteristics and also some of their attitudes expressed in the survey responses. More recently, Huckfeldt et al.’s (Huckfeldt, Sprague, & Levine, 2000) study of attitudes towards presidential candidates, conducted during Attitude Agreement in Social Networks, p. 5 the 1996 presidential election campaign, included a sub-sample of 892 interviewees who were asked to name up to five contacts and were asked to assess their level of political expertise; of those named, 1,475 were then subsequently interviewed. Another recent example is Levitan and Visser (2009), who recruited college freshmen randomly assigned to dorms for a study on attitudes, and assessed their attitudes as well as their beliefs about their friends’ attitudes. Twenty-six percent of the friends listed happened to also be in the study, so the perception of similarity and true similarity could be considered for a subset of the listed network members. Finally, related studies have been conducted in the survey methodology literature under the rubric of “proxy reporting,” in which one respondent’s response is used as a proxy for that of a close social tie, such as a spouse (Bickart, Menon, Schwarz, & Blair, 1994; Menon, Bickart, Sudman, & Blair, 1995; Sudman, Bradburn, & Schwarz, 1995). Although straightforward to design, network surveys of the kind described above are difficult to implement at a larger scale using traditional interview-based methods, in large part because the costs associated with interview-based methods typically preclude long batteries of questions about each of the named discussants; thus the scale and scope of such studies has historically been restricted either to a small number of questions (q) or a small number dyads (u,v), or both—for example, the proxy reporting studies described above examine attitude similarity only for isolated dyads, not for multiple neighbors of the same respondent. From a practical perspective, therefore, the social networking site Facebook offers some advantages for conducting large-scale network survey work. In particular, Facebook provides a ready-made “social graph” on which third party developers can build applications, including surveys, through an “Application Programming Interface” (API, see http://www.facebook.com/apps/ for Attitude Agreement in Social Networks, p. 6 examples). By using the Facebook API, a survey tool can potentially reach thousands, or even millions, of respondents at very little cost and in a short period of time. Clearly one cannot guarantee either that a user’s close friends will be named on their profile (because, for example, they are not members), or that friends who are named are necessarily the most relevant to a user’s political beliefs and attitudes. However, previous work suggests that most Facebook ties reflect real, offline relationships (Ellison, Steinfeld, & Lampe, 2007). Moreover, it is possible to restrict attention only to ties that satisfy additional criteria, such as multiple mutual friends and shared affiliations, which are more likely to be salient to political attitudes. Finally, we note that traditional respondent-driven network samples are also subject to at least two potential sources of bias. First, which friends a respondent volunteers to be recruited can be biased by their memory or motivations (Tversky & Kahneman, 1973). And second, discussants are typically identified by asking respondents to name individuals with whom they discuss “important matters” (Huckfeldt, 2001); yet questions of this sort have been shown to yield highly variable results depending on the respondents’ interpretation of the question (Bearman & Parigi, 2004). Thus although imperfect, network-oriented data collected on Facebook represents a fast, convenient and relatively reliable alternative to traditional survey methods.