Strategic identity signaling in heterogeneous networks

Individuals often signal identity information to facilitate assortment with partners who are likely to share norms, values, and goals. However, individuals may also be incentivized to encrypt their identity signals to avoid detection by dissimilar receivers, particularly when such detection is costly. Using mathematical modeling, this idea has previously been formalized into a theory of covert signaling. In this paper, we provide an empirical test of the theory of covert signaling in the context of political identity signaling surrounding the 2020 US presidential elections. To identify likely covert and overt signals on Twitter, we use methods relying on differences in detection between ingroup and outgroup receivers. We strengthen our experimental predictions with additional mathematical modeling and examine the usage of selected covert and overt tweets in a behavioral experiment. We find that participants strategically adjust their signaling behavior in response to the political constitution of their audiences. These results support our predictions and point to opportunities for further theoretical development. Our findings have implications for our understanding of political communication, social identity, pragmatics, hate speech, and the maintenance of cooperation in diverse populations.

Individuals constantly emit signals of their identity, consciously and unconsciously, informing others about the sort of person they are. Identity signals are any components of communication that inform receivers of the signaler’s membership (or nonmembership) in a subset of individuals (1–3). Such subsets can reflect strong social boundaries, such as “Republican” or “Democrat” in the United States, or reflect subtler intragroup variations, such as differences among Democrats regarding government regulations. In large, multicultural nations like the United States, identities such as Republican or Democrat can serve to organize like-minded communities or coalitions (4–7). Although the specific style of communication may vary with cultural context (8), identity signaling serves a key social function by enabling individuals to rapidly characterize others as similar or dissimilar (2, 3, 9). Finding similar others has many proximate psychological benefits, such as better mental health (10) and the security that results from a stronger sense of group identity (11, 12). Our emphasis here is on the role of identity signaling to facilitate social assortment: preferentially interacting with similar individuals and reaping the benefits of coordinating on norms, goals, and values (2, 3, 9, 13–15).Identity signaling is especially important in vast and diverse social communities, in which little can be assumed about strangers in the absence of identity information. This type of scenario is made all the more common in the digital age (16). In recent years, online social media has both expanded the pool of potential partners and enabled easier formation of communities across traditional, social, and geographic boundaries. This presents new challenges and opportunities for signalers to successfully find niche communities (17, 18). On the one hand, large online communities have arisen dedicated to worldviews that are otherwise rare in most local communities. An individual expressing a viewpoint that is rare in their locality can nevertheless become part of a flourishing, geographically dilute collective. On the other hand, online signaling also carries new risks that come from expanding one’s audience far beyond one’s local social network, sometimes without the signaler’s knowledge (e.g., ref. 19).Given the social importance of political identity in the United States and other countries (5, 20), we expect much identity signaling to be about political views and related coalitional affiliations. Political views are often expressed on social media using obvious signals like slogans, partisan memes, and other declarations of partisanship. However, the United States is also highly polarized (5, 6), and obvious political signals are not always advisable. Partisans often hold deeply negative feelings toward members of groups perceived as opposed or even simply different to their own (21–24). Signaling one’s political affiliation to outgroup members can therefore be costly, with costs ranging from the loss of social standing or relationship opportunities to the loss of an employment (25) or even becoming the victim of violence (26). For example, Van Duyn (27) documents a group of anti-Trump women in rural Texas who met in secret to discuss politics because they feared negative consequences for their business or marriages if their views became known. Exactly who is considered a member of one’s outgroup also varies over time and context. In the context of political identities, debates during US presidential primaries tend to be between members of the same political party, and so, a perceived outgroup may be copartisans that support different candidates or policy goals. During national presidential elections, cross-partisan differences become more salient. In both cases, the assortative benefits of overt identity signaling must be weighed against the potential costs of being identified by outgroup individuals in situations in which identification has consequences.Overt, unambiguous signals of identity are useful when individuals can sufficiently benefit from their role in supporting positive assortment—preferentially interacting with similar others. A wide literature on social tags and ethnic markers has documented and modeled the utility and likely emergence of such signals for this purpose (9, 13, 28–31). The benefits of overt identity signaling, however, must sufficiently outweigh any risks that come from alienating others or revealing oneself to be misaligned with their interests. If those risks and their associated costs are high enough, we should expect cultural or psychological processes (such as cultural evolution or strategic decision-making) to favor subtler signaling strategies that encode information in such a way that it is detectable only by those who share relevant worldviews.We refer to identity signals that are accurately received by their intended audience but obscured when received by others as covert signals (3, 32, 33). Covert signals allow individuals to reap at least some benefit from being identified by similar others, when possible, while simultaneously avoiding the costs associated with detection by dissimilar others. Covert signals work because communication often contains multiple, simultaneous layers of meaning, which are not all equally apparent to all receivers. A receiver’s background knowledge and perception of context affects whether and how those layers will be revealed (34–37). Individuals should, therefore, adjust their patterns of communication, based not only on their intended audience but also on the likely third parties that will perceive those communications (37–39). Loury (39) captures the essence of the idea: “If the significance of some words as signals of belief is known only to ‘insiders,’ their use in public allows the speaker to convey a reassuring message to some listener—‘I share your values’—without alarming the others.” A well-documented example is the routine remarks made by former US president George W. Bush concerning his opposition to the 1857 Dred Scott Supreme Court decision. While it might seem unnoteworthy to oppose a judicial decision that upheld slavery, the mention was seen by many evangelical conservatives as morally analogous to the 1973 Roe v. Wade case that upheld the right to abortion and so subtly communicated to these audiences the president’s commitment to overturning that decision (40).Recently, Smaldino and colleagues have developed a theory of covert signaling, using formal mathematical and agent-based models of cultural evolution to examine the circumstances under which overt or covert identity-signaling strategies should be favored (32, 33). The theory of covert signaling provides a formalism for identity signaling in the context of third-party receivers and describes how signalers should communicate based on their likely audiences and the consequences for both successful and failed communication. The models derived from this theory make general predictions about strategies for identity signaling related to both the ability of individuals to preferentially assort with similar others and the costs of failing to assort accordingly. Covert signaling can achieve higher payoffs than overt signaling when individuals are likely to have interactions with dissimilar individuals and when those interactions incur high costs once the dissimilarity is revealed. The theory of covert signaling is consistent with a number of common signaling domains, including the use of humor as an encrypted signal of similarity (41, 42), the use of fashion to subtly signal insider status (43), political dog whistles (44, 45), and signals used by LGBTQ+ individuals (46, 47) or political dissidents (48–50) to assort without detection. It is also consistent with the fact that signals of political identity need not be obviously political in nature, as reliable associations with certain products and activities may be used as heuristics to differentiate partisans (51–53).It is likely that a great deal of online speech is covert, especially on social media platforms on which users can be personally identified, such as Twitter. Although other social media sites have more users than Twitter (54), Twitter is a particularly important forum for public discourse on current events and as such is valuable for studying covert identity signals that are likely to be both relevant and visible to diverse audiences. For example, the strategy of “subtweeting” is well documented and refers to online communications that are interpretable only to individuals who have relevant information that is not provided in the communication itself (55). As another example, a search for tweets containing the phrase “remember that scene” sent on November 9, 2016 (the day after Donald Trump was elected as US president) returned a number of candidate covert tweets concerning feelings about the election results, many from users unhappy with the outcome (Fig. 1). Each of these requires background knowledge about the cultural artifact (i.e., film) being referenced as well as an understanding of recent political events, as the relevant contextual backdrop for interpreting the analogy implied by those references.Open in a separate windowFig. 1.Potentially covert tweets related to the 2016 US presidential election. To understand each tweet, the reader must be familiar with both the ingroup conversations about the relevant political events as well as the cultural artifacts being referenced.Our data were collected during an especially salient period for political identity signaling on Twitter: the 2020 COVID-19 pandemic, in which millions of people were restricted from gathering and communicating in person, making online engagement one of the principal ways to interact with other humans. Moreover, our tweets were collected and assessed in the wake of many high-profile sociopolitical events in the United States, including the Black Lives Matter protests following the murder of George Floyd, the confirmation of Supreme Court Justice Amy Coney Barrett, and the 2020 presidential election. Opinions on these events were often polarized, making expressions of those opinions into signals of political identity.While they are likely to be common, covert signals are also inherently challenging to study empirically because, by definition, they require insider knowledge to be detected. In this paper, we introduce a theoretically motivated measurement of covertness, focusing on identity signals in the context of political speech online. Essentially, covertness was measured in terms of how people from ingroup and outgroup political groups perceived different tweets. On Twitter, cross-partisan followers are rare (56), and thus, we focused on differences between copartisan radicals and moderates. We downloaded tweets from politically engaged Twitter users with heterogeneous follower networks, thus increasing our chances to collect tweets with some covert political identity signaling according to the theory (Fig. 2A). Then, we asked ingroup and outgroup members to guess the political identity of the tweet author and to report their affective responses to the tweet. Tweets were considered to be more likely to serve as covert identity signals if there was a large difference in responses of ingroup and outgroup raters (Fig. 2B).Open in a separate windowFig. 2.Empirical process to test the theory of covert signaling: (A) selecting Twitter users who might be more likely to use covert or overt political speech and downloading their tweets, (B) rating of tweets on two dimensions and using these ratings to select a subset of covert or overt tweets, and (C) conducting a behavioral experiment with the selected covert and over tweets to test how people use them to communicate their political belief.Our paper constitutes a direct empirical test of the theory of covert signaling. Based on this theory, we predict that covert signaling will be more prevalent among 1) individuals in more heterogeneous communities or individuals with minority status and 2) individuals who face higher costs from being recognized as dissimilar. We derive a simple mathematical model of our experimental design, which yields more precise predictions concerning the relationships between covert signaling, the frequency of outgroup members in the audience, and the cost of being disliked. We test these predictions in a behavioral experiment in which participants select from a set of tweets that contain either overt or covert political signals to communicate with and be evaluated by an audience of varying partisanship (Fig. 2C). We compare signaling strategies when the outgroup audience consists of copartisan members (more or less radical) and cross-partisan members (left or right). Although previous studies have also considered the use of covert or encrypted signals (41, 43, 47, 57), our study tests predictions derived from a formal model with relatively unambiguous predictions and a clear scope of applicability (sensu ref. 58). By doing so, we can show where the existing theory fits the real world and where we need to direct our future efforts to refine the theory. Our empirical pipeline (Fig. 2) is described in more detail in the Materials and Methods.