Keywords

Introduction

To join a work organization is to be plunged into a world of similar and dissimilar others. Of the many people one encounters at work, only some become our personal friends. The question of why people become friends with some people and not others is important because we know that workplace friendship ties influence key outcomes, such as job satisfaction, turnover, job performance, and promotions (for a review, see Brass, 2022). Perhaps the most influential response to this question draws on Lazarsfeld and Merton’s classic work on “homophily,” defined as “a tendency for friendships to form between those who are alike in some designated respect” (1954, p. 23). Similarity, from the homophily perspective, breeds friendship.

But if similarity is to be used to explain patterns of friendship choice, on what basis is similarity to be assessed? In their study of friendships in “Hilltown” (a bi-racial, low-rent housing project in Pennsylvania) and “Crafttown” (a housing project consisting of mostly white families in New Jersey), Lazarsfeld and Merton distinguished between two different bases of homophily. The first—which they labeled “status homophily”—focused on sociodemographic dimensions that stratify society, such as race and gender. The other—“value homophily”—included a wide variety of internal psychological states and attitudes, such as opinions about whether or not “colored and white people should live together in housing projects” (Lazarsfeld & Merton, 1954, p. 26). Similarly, the conceptually related work on “The Attraction Paradigm” (Byrne, 1971) examined both demographic variables and psychological attitudes as precursors of interpersonal attraction. This line of work found that similarity in terms of demography and psychological beliefs were both related to interpersonal attraction, but the effect sizes for the demographic variables were anemic compared to those for attitude similarity: Individuals tended to be attracted to people who shared similar attitudes, even if these individuals happened to be of a different gender or race (Byrne, 1971, pp. 127–163).

Classic work on homophily recognized the importance of assessing perceptions of similarity (and dissimilarity). But most of the work on homophily in the workplace has tended to focus on demographic characteristics to the exclusion of cognitive ones (e.g., Brass, 1985; Ibarra, 1992, 1995; Lawrence & Shah, 2020; Lincoln & Miller, 1979, for a review). The structural take on homophily focuses on readily observable demographic variables (such as sex and race) rather than on an examination of underlying psychological states. There are good reasons for the popularity of the structural perspective on homophily in investigations of workplace friendship. For one thing, it possesses the virtues of parsimony and expediency: Individuals' cognitions don't have to be assessed; they can be inferred by examining easily observed demographic proxies (Pfeffer, 1983). Demographic similarity is treated as a proxy for underlying but unobserved cognitive processes, such as shared cultural beliefs and attitudes (McPherson et al., 2001, p. 435). On the other hand, the strategy of bypassing individuals' psychological beliefs and attitudes “obscures clear thinking” (Lawrence, 1997, p. 3; Lawrence & Shah, 2020). This is because the structural approach assumes that demographic variables also capture the variation in underlying, but unmeasured, psychological concepts. Although this “congruence assumption” (Lawrence, 1997, p. 3) is regularly invoked in explanations of why demographic similarity is related to friendship choice, it is seldom examined in the empirical research in work organizations (Harrison et al., 2002; cf. Kilduff et al., 2000). This raises a troubling possibility: We know that demographic similarity is related to friendship choice in work organizations (Brass, 1985; Ibarra, 1992; e.g., Lincoln & Miller, 1979), but is it possible the rationale underlying this observed relationship is mis-specified?Footnote 1

The primary goal of our field-based study is to submit to empirical test the explanatory theory that is implicit in structural accounts of workplace homophily. Specifically, we examine whether interpersonal cognitions mediate the relationship between demographic characteristics (we examined rank and sex) and workplace friendship choice at the dyadic level of analysis. Looking inside the black box of homophily, we directly examine the perceptions of similarity that presumably explain why demographically similar people are more likely to be friends in the workplace. We also examine individuals’ perceptions of dissimilarity. As Simmel (1950, p. 30) pointed out long ago: “… for the actions of the individual, his difference from others is of far greater interest than his similarity with them.” The tendency to ignore dissimilarity as an independent force in friendship formation is perhaps not surprising given the relative neglect of negative ties in organizational research (Labianca & Brass, 2006). In allowing for the possibility that perceptions of dissimilarity play a unique role in the choice of friends, we hope to provide a “useful antidote” to the “emphasis on similarity as the pervasive dynamic in groups” (Weick, 1969, p. 14).

A second and related goal of this research is to expand the current focus of organizational homophily research from visible demographic characteristics to include the less visible but psychologically relevant characteristics that make up individual personality. Although personality characteristics were examined in classic work on the similarity attraction paradigm (see the summary in Byrne, 1971, pp. 164–187), they have received relatively little attention in subsequent research on homophily (see the review in Ertug et al., 2022). There is, however, a reason to suspect that personality plays an important role in friendship formation (e.g., Klohnen & Luo, 2003; Sasovova et al., 2010). Of the many possible personality theories that can be used to differentiate individuals, we focused on one that has been shown to be of importance to the structure of workplace friendship relations: self-monitoring (e.g., Fang et al., 2015; Mehra et al., 2001). Self-monitoring theory is concerned with systematic “differences in the extent to which people value, create, cultivate, and project social images and public appearances” (Gangestad & Snyder, 2000, p. 531; for an organizationally oriented review of self-monitoring theory and evidence, see Day & Schleicher, 2006; Kudret et al., 2019). Some people (“high self-monitors”), out of a concern for situational appropriateness, monitor and regulate their self-presentation for the sake of creating desired public appearances. The behaviors of others (“low self-monitors”), by contrast, are largely a reflection of their inner attitudes and dispositions. Whereas high self-monitors are akin to social pragmatists who are willing and able to project images intended to impress others, low self-monitors seem unwilling and unable to carry off appearances (Gangestad & Snyder, 2000, p. 531).

Given these sharply contrasting approaches to managing their social worlds, it seems likely that people of the same self-monitoring orientation would see each other as similar and those of opposite self-monitoring orientation would see each other as dissimilar. The homophily principle can be used to anticipate that, in work organizations, individuals will tend to befriend others of the same self-monitoring orientation as themselves because they will be more likely to perceive these others as similar to themselves and those of the opposite self-monitoring orientation as dissimilar to themselves (Snyder & Smith, 1986, pp. 71–73). Whether this is in fact the case is an empirical question, and it is one this research attempts to answer.

Methods

Site

We collected survey-based data from a small high-tech organization that researched and produced sophisticated chromatographic equipment. The organization had been kept deliberately flat to enhance speed and responsiveness, which were critical to success in a competitive environment that pitted the firm against bigger rivals. The small size meant that people at the company knew each other on a first name basis and regularly ran into each other in the spacious atrium, that included a cafeteria, surrounded by plants and small trees, where employees from all parts of the organization ran into each other. The firm had won industry awards for its innovative products, customer service, and its inclusive culture.

Data

Data on social networks and self-monitoring were collected through a questionnaire sent to all 116 employees (68 men, 48 women). 102 people provided data on their friendship ties. We used the roster method to collect this “whole-network” data. People were free to nominate as many individuals as they liked as friends. Missing data on self-monitoring reduced the total usable sample in this study to 93.

Measures

Friendship

The raw data on friendship relations were arranged in a 93 × 93 binary matrix. Each cell Xij in this matrix initially corresponded to one individual i's relation to another individual j as reported by i. For example, if i reported j as a friend, then cell Xij was coded as 1. Because we were interested in this study in reciprocal friendship choice, we symmetrized this matrix using the rule that Xij = 1 if and only if Xij = Xij = 1. That is, both i and j had to list each other as a friend for the pair to be considered friends.

Perceived Similarity

We used the same sociometric approach to capture data on perceived interpersonal similarity as we did for friendship (for a previous application of this measure, see Mehra et al., 1998, p. 443). Individuals were asked, on the questionnaire, to identify those individuals they considered “especially similar” to themselves. We noted, on the questionnaire, that we were interested in the respondents’ perceptions and the basis of judging similarity was entirely up to them. As with the friendship data, we initially arranged these responses in a 93 × 93 binary matrix, which we then symmetrized using the rule that cell Xij was coded as 1 if and only if both i and j reported the other as someone who was especially similar to themselves.

Perceived Dissimilarity

We used the same approach to code and symmetrize this 93 × 93 binary matrix as we did the perceived similarity matrix. The question we used to gather data asked individuals to identify individuals at the firm who they thought of as “especially dissimilar” to themselves. Cell Xij in the final matrix was coded as 1 if and only if both i and j reported the other as someone who they considered “especially dissimilar.”

Self-monitoring

Self-monitoring was measured with an 18-item true–false questionnaire (Snyder & Gangestad, 1986). A sample item is “In different situations with different people, I often act like very different persons.” This measure correlates highly (r = 0.93) with the original 25-item measure (Snyder, 1974) and has been demonstrated to be both more reliable and factorially pure (Snyder & Gangestad, 1986). In the present study, the reliability for this scale as measured by Cronbach's alpha was 0.80.

Personality Difference

This dyadic measure was computed, for each pair of individuals in the sample, as the absolute difference in the self-monitoring scores between the two individuals. This variable was arranged as a 93 × 93 valued matrix where cell Xij was coded as the absolute difference between self-monitoring orientations of i and j. The observed values range from 0 to 17. To ease interpretation of the MRQAP regression coefficient associated with this variable, we rescaled this variable by dividing it by 10.

Gender Similarity

Data on gender was obtained from company records. We converted these data into a 93 × 93 binary matrix where cell Xij was 1 if and only if both i and j were of the same gender.

Rank Similarity

Data on rank came from company records. We simplified data on rank into two categories: supervisors (i.e., those who had one or more persons formally reporting to them) and non-supervisors (all others). These data were converted into a 93 × 93 binary matrix such that cell Xij was coded as 1 if and only if both i and j were of the same rank.

Control Variables

Workflow

Because friendships are more likely to form between people who come into regular contact with each other (Festinger et al., 1950), we included as a control the 93 × 93 binary “workflow” matrix. Data on workflow relations were obtained from the sociometric survey by asking people to identify their workflow contacts (i.e., the set of people from whom one regularly exchanged workflow inputs and/or outputs—see Mehra et al., 2001 for additional details). Cell Xij in the symmetrized workflow matrix used for analysis in this paper was coded as 1 if and only if both i and j identified the other as a workflow partner.

Analysis

We used the Matrix Regression Quadratic Assignment Procedure (MRQAP) in UCINET 6 (Borgatti et al., 2002) to test our ideas. This analytic approach is a better choice than OLS regression because MRQAP is a procedure designed specifically to account for the lack of independence in network data (Krackhardt, 1988; see the discussion in the methodological appendix of Chen et al., 2021). We used the standard procedure to examine mediation (Baron & Kenny, 1986).

Results

Table 1 reports descriptive statistics and zero-order correlations for each of the matrixes used in the MRQAP analyses.

Table 1 Means, standard deviations, and QAP correlations among matrices
Full size table

The first question we sought to investigate was: Is the relationship between demographic similarity and friendship choice mediated by psychological perceptions of similarity and dissimilarity? To test for mediation, we examined, first, if there was a significant relationship between demographic similarity in terms of rank and gender and friendship choice. Second, we checked to see if this relationship was eliminated or significantly reduced once perceptions of similarity and dissimilarity were added to the regression model. The results presented in model 1 of Table 2 show that, controlling for the significant effects of being workflow partners on friendship (b = 0.04, p < 0.001), gender similarity (b = 0.02, p < 0.001) predicted dyadic friendship choice. Similarity in rank, by contrast, was not a significant predictor of dyadic friendship (b = 0.01, p = n.s). Model 2 in Table 2 shows that perceived similarity (b = 0.43, p < 0.001)—but not perceived dissimilarity (b =  − 0.02, p = n.s.)—predicted dyadic friendship choice. However, the relationship between gender similarity and dyadic friendship choice was not significantly diminished once perceived similarity and dissimilarity were also included in the regression (see model 3 in Table 2). Moreover, as shown in Table 3, the demographic similarity was only inconsistently related to underlying perceptions of similarity and dissimilarity: gender similarity predicted perceived similarity (b = 0.01, p < 0.01) but did not predict perceived dissimilarity. And rank similarity did not significantly predict perceived similarity or perceived dissimilarity. Overall, these results indicate that gender similarity and perceived similarity had significant but independent effects on dyadic friendship choice. Thus, there was no support for the argument that the effects of demographic similarity on friendship are mediated by underlying perceptions of similarity and dissimilarity.

Table 2 MRQAP analyses predicting dyadic friendship choice
Full size table
Table 3 MRQAP analyses predicting perceived similarity and dissimilarity
Full size table

A second question we sought to investigate was: Does self-monitoring personality predict friendship choice? If so, is the effect mediated by perceptions of similarity and dissimilarity? The results in Table 2, model 1 show that the greater the difference in the self-monitoring score of two individuals, the more likely they were to be friends (b = 0.00, p < 0.10). There was, however, no evidence that differences in self-monitoring personality were significantly related to perceptions of similarity and dissimilarity (see Table 3); and, even with the inclusion of these perceptual variables in the model, the greater the difference in the self-monitoring scores of two individuals, the more likely they were to be friends (b = 0.00, p < 0.05; Table 2, model 3).

We found further evidence of this complementarity in analysis that, first, coded a person as a high self-monitor if that person scored above 11 on the 18-point self-monitoring scale and as a low self-monitor otherwise.Footnote 2 Next, we examined the mean score for “preference-based homophily,” computed using a formula that adjusted for the relative numbers of high and low self-monitors in our sample (for the formula, see Ibarra, 1992; Krackhardt, 1990). We found that the mean for both high self-monitors (mean =  − 0.01, SD = 0.08) and low self-monitors (mean =  − 0.01, SD = 0.12) was negative, an indication that high self-monitors were more likely to prefer friends who were low self-monitors, and vice versa. These mean scores were not significantly different (t =  − 0.18, p = n.s.), so high self-monitors and low self-monitors did not differ to the extent to which they were heterophilous. Complementarity, it appears, rather than similarity is the principle that governs the relationship between self-monitoring personality and friendship choice.

Discussion

This paper reported an empirical test of the idea, implicit in structural research on homophily, that demographic indicators, such as gender and rank, accurately capture underlying psychological perceptions of similarity and dissimilarity, and these, in turn, drive friendship choice. The results of our investigation suggest that this mediation model is mis-specified. Looking inside the black box of homophily, we examined whether perceptions of similarity and dissimilarity contributed to the observed tendency for people to be friends with those of the same gender and rank as themselves. We found that similarity in gender (but not rank) predicted dyadic friendship, but this effect was not mediated by underlying perceptions of similarity or dissimilarity. Instead, both gender similarity and perceived similarity independently predicted dyadic friendship.

Although several theorists (e.g., Simmel, 1950; Weick, 1969) have argued for greater attention to dissimilarity as an independent force in organizational life, these negative cognitive ties—like negative social ties more generally—have received relatively little attention in the organizational literature. We found no evidence that perceptions of dissimilarity were associated with differences in gender, rank, or personality. Moreover, perceptions of dissimilarity did not predict dyadic friendship. It may be that perceptions of dissimilarity only infrequently play a role in friendship choice, and this rarity of dissimilarity ties may also make it difficult to detect this otherwise plausible effect. Nonetheless, even if infrequent, perceptions of dissimilarity may play a decisive role in friendship dynamics. This is a topic that deserves more empirical attention.

A goal of our study was to examine whether the homophily principle applied to self-monitoring personality (Snyder, 1974). We focused on self-monitoring personality because it has been shown to have clear implications for friendship dynamics in work settings (e.g., Sasovova et al., 2010). Moreover, past work suggests that individuals can accurately identify the self-monitoring orientation of people with whom they interact (Snyder, 1974). From the homophily perspective, individuals of the same self-monitoring orientation should be more likely to choose one another as friends than individuals of opposite self-monitoring orientations. What our study found, however, was that the greater the difference in the self-monitoring score of two individuals, the more likely they were to be friends. Heterophily,Footnote 3 not homophily, seems to be the principle that governs the relationship between self-monitoring personality and dyadic friendship choices in organizations. Further, as with gender, we found no evidence that perceptions of similarity and dissimilarity mediated the effects of self-monitoring on dyadic friendship.

Implications for Theory and Future Research

Our results have three implications for theory. First, our results suggest that although the structural perspective on homophily in work organizations is parsimonious and accounts for much of the variance in friendship relations in work organizations, the rationale for why it does so deserves further scrutiny. We found that similarity in gender was related to friendship choice, but this relationship was not mediated by underlying psychological perceptions of similarity and dissimilarity. Alternative theoretical accounts may be needed to explain why similarity in gender is related to friendship choice in the workplace even in the absence of corresponding psychological perceptions of similarity and dissimilarity (see Montoya et al., 2008). An intriguing possibility is that workplace friendship means different things to women and men. There is some evidence, for example, that women tend to see their friendships as a vehicle for the receiving and giving of social support whereas men think of work friends as someone to do things with, such as socialize outside work (Argyle & Henderson, 1985, pp. 75–81). These gendered differences in what friendship means may provide fresh clues about the mechanisms that explain friendship choices at work (for a plea for greater attention to mechanisms underlying the relationship between gender and social networks, see, e.g., Brands et al., 2022; Woehler et al., 2021).

A second and related implication of our study is that the homophily perspective in organizational research may need to re-inject the earlier emphasis on underlying perceptions of similarity as an independent basis for homophily in friendship choice. We found that people were more likely to befriend those they perceived to be especially similar to themselves, irrespective of whether those people were demographically similar to them. By ignoring the psychological bases of friendship formation in work organizations, structural research on homophily offers an incomplete account of the antecedents of friendship choice in organizations. Perceptions of similarity matter for friendship, above and beyond demographic similarity. We are not driven in our choice of friends at work by structural forces alone; our inner beliefs and cognitive processes also play an important if neglected role.

The third implication of our work is that the homophily principle may explain why similarity in demographic characteristics is related to friendship, but heterophily may be the more appropriate principle when it comes to how personality influences friendship choice. In contrast to the pervasive emphasis on how “birds of a feather flock together,” it appears that sometimes it is birds of different feathers that flock together. Past work has argued that similar individuals are preferred as friends because similar others provide consensual validation of one's own views and beliefs (e.g., Byrne & Clore, 1967; Snyder & Smith, 1986). However, individuals may sometimes prefer different yet complementary others. Plato, in his famous dialogue, “Lysis,” argued that those who resemble us excite in us feelings of envy and competitiveness whereas dissimilar others can inspire and attract. Despite the intuitive appeal of Plato’s argument, the idea that similarity is the basis for interpersonal attraction has come to be accepted as axiomatic. Organizational research and practice could profit from a richer understanding of the conditions under which opposites attract.

The questions we are in pursuit of beg for longitudinal data. The analytical approach we used for analyzing our cross-sectional data is a type of regression in which the cases are dyads, the dependent variable is the state of the dyad, and the independent variables are dyadic properties, such as differences in gender or self-monitoring. The p-values in MRQAP, however, are calculated via a permutation-based method that avoids assuming a mathematical distribution and instead generates its own distribution of beta coefficients. With cross-sectional data, the MRQAP can be considered as modeling tie formation (see the appendix in Chen et al., 2021). Nonetheless, it would be fruitful to explicitly observe and model changes in workplace friendships over time (Nestler et al., 2015). With longitudinal data, coevolution modeling can be used to distinguish whether people select those they perceive as similar or whether, over time, people come to influence each other, thereby coming to see each other as similar (e.g., de Klepper et al., 2010). It might also be worth manipulating perceptions of similarity and dissimilarity to see if changes in similarity/dissimilarity shape network dynamics. It is possible that whereas perceived dissimilarity did not emerge in our analysis as a significant predictor of friendship choice, it could be a potent predictor of friendship dissolution. It is also possible that whereas dissimilarity does not predict friendship choice, it may predict other relationships, such as advice, or role modeling, that we did not examine in this paper.

Our study is limited in its reliance on just two items to measure the underlying perceptions that presumably mediate the relationship between visible, demographic markers and friendship choice. The benefit of this approach is that respondents were free to make a determination based on individually-salient criteria unbiased by researcher-imposed categories. However, an alternative strategy would be to ask people about the extent to which others are similar/dissimilar to them regarding their views on specific topics and/or values. It may be that our approach to measuring the psychological beliefs that mediate between demographic similarity and friendship was simply too coarse-grained.

Managerial Implications

Work organizations in the post-(Covid) pandemic world have been scrambling to find ways to enhance employees’ sense of connectedness to their colleagues at work. A firm called Imperative is using an interesting approach to addressing this challenge, one that dovetails nicely with our findings (Hurst, 2022). In a nutshell, the firm uses software to have each employee identify their “purpose drivers,” the things that they are intrinsically motivated by at work, and the outcomes they most care about. Drawing on this psychological data on attitudes and beliefs, the firm creates meaningful connections between people who were otherwise unlikely to connect. This approach to directly collecting data on relevant attitudes could be one way that firms can grow workplace friendships. As our research shows, irrespective of whether people were demographically similar, they were more likely to be friends if they perceived one another as similar. Managers cannot control the gender or personality of the people they would like to turn into workplace friends, but what they can do is shape people’s sense of interpersonal similarity by helping them see what they have in common despite their demographic differences.

Conclusion

Why do people befriend certain others at work? An influential answer relies on the “homophily” principle, which posits that similarity breeds friendship. Although early formulations of this principle accounted for similarity in terms of both observable demographic markers (such as gender and race) and underlying perceptions of similarity and dissimilarity, the study of friendship choice in work settings has since come to focus almost exclusively on readily observable surface characteristics. Looking inside the black box of homophily, our research suggests that the gains in methodological expediency that result from the adoption of a structural perspective on homophily need to be carefully weighed against the possibility that the underlying psychological assumptions it relies upon are mis-specified. It can be pragmatic for a theory to focus on predictive success at the expense of explanatory accuracy. The trouble arises when the model starts to be confused with reality.Footnote 4