A Language-Based Method for Assessing Symbolic Boundary Maintenance between Social Groups

Symbolic Boundaries and Boundary Maintenance

A large body of sociological research on boundaries has studied how individuals and groups maintain social boundaries—the objectified ways through which people of different groups gain differential access to resources—through the construction and perpetuation of symbolic boundaries—conceptual distinctions, interpretive strategies, and cultural traditions that people use to differentiate themselves from others. Early work in this vein assumed that symbolic boundaries are, for the most part, consistent across social group members and stable over time (Durkheim et al. 1912; Grillo 2003). In this tradition, symbolic boundaries serve to maintain social boundaries across long timescales. Thus, the maintenance of symbolic boundaries results from enduring cultural markers. For example, Bourdieu’s influential studies of cultural reproduction argued that it is precisely because cultural distinctions of taste are widely accepted and legitimized that they can serve as vehicles for the intergenerational reproduction of social class (Bourdieu and Passeron 1970; Bourdieu 1984). The reproduction of existing cultural symbols is especially important in the face of inevitable turnover in group membership. For example, religious groups, military branches, and college fraternities and sororities all have elaborate and enduring markers—initiation rituals, shared symbols, and specialized jargon—to demarcate and preserve group boundaries (Dacin, Munir and Tracey 2010). The legitimacy and authenticity of these cultural traditions derives in large part from their durability (Berger and Luckmann 1966; Carroll 2015). Moreover, symbols endure because people tend to exhibit cultural inertia (Vaisey and Lizardo 2016; Kiley and Vaisey 2020). Bourdieu’s notion of habitus, for example, is rooted in an understanding of cultural attitudes and behaviors as deeply ingrained and difficult to shift.

In contrast to this view of symbolic boundaries as mostly uniform and static, a separate stream of work has shown that symbolic boundaries vary considerably across individuals and that new cultural markers emerge over time. For instance, Lamont (1992) points out the diverse understandings of high status culture across national contexts, Bail (2008) illuminates multiple configurations of symbolic boundaries that define European nationalism and attitudes towards immigration, and Goldberg (2011) identifies three competing logics of cultural distinction in American musical tastes. In even earlier work, anthropologist Barth (1969) recognized that attempts to classify individuals’ culture and ethnicity according to the same unchanging categories was futile because the relationship of cultural behaviors to ethnic groups and individuals’ ethnic group membership were in constant flux.

Further support for the view that boundary maintenance involves symbolic innovation—that is, the ongoing creation of new cultural markers—comes from studies that show how boundaries evolve over time. For example, elite consumers of the New York Philharmonic responded to the opening of subscriptions to middle class audiences by erecting new symbolic distinctions of exclusivity through their patterns of attendance (Accominotti et al. 2018). The analysis of baby name trends by Lieberson and Bell (1992) points out that fashion cycles are a form of symbolic innovation that reinforce symbolic boundaries of social class.

In sum, the extant literature makes two implicit and disconnected assumptions about how people maintain symbolic boundaries—through reinforcing existing symbols and innovating new ones. Moving beyond conceptions of symbolic boundary maintenance as involving cultural markers that are either static or constantly evolving, we propose a theoretical integration of these two perspectives. Recent research lends support for this dual approach to boundary maintenance. For instance, Khan (2012) finds that the cultural capital inculcated in modern elite boarding schools blends longstanding highbrow education (e.g., “Beowulf”) with more recent elite symbols of cosmopolitan and worldly consumption (e.g., “Jaws”). In a parallel vein, Goldberg, Hannan and Kovàcs (2016a) demonstrate how high status tastes in cuisine and film are characterized by both retention of classical genres and more recent boundary refinement according to genre purity. Finally, recent methodological advances in natural language processing have illuminated how the semantic structure underlying gender and class boundaries have simultaneously evolved while key symbolic features remain stable (Garg et al. 2018; Kozlowski, Taddy and Evans 2019).

We label the two forms symbolic boundary maintenance as boundary retention—the extent to which individuals sustain and entrench in pre-existing symbolic differences between social groups—and boundary reformation—the degree to which individuals alter the substance of symbolic boundaries by creating new cultural markers of distinction between social groups. Conceptualizing these two forms of boundary maintenance allows us to theorize about how different types of individuals will engage in each activity and how their behavior might vary across social contexts. Although we focus on the maintenance of symbolic boundaries between particular social groups that exist in organizations—the legacy organizations people belonged to before their organizations merged—our theory can be readily extended to other social groups both within and outside formal organizations.

Boundary Retention and Reformation: The Roles of Social Position and Context

We first consider how an individual’s social position—specifically the relative power of her group—might influence how she differentially employs the two boundary maintenance strategies. We expect that individuals from high-power groups—in our setting, employees from the acquiring firm—will be more invested in preserving the status quo and less inclined to adapt to a new or changing context. This could be because their group dominance is based on existing cultural markers or have previously invested considerable effort to enculturate to existing styles. We therefore anticipate that members of high-power groups will exhibit more boundary retention of existing symbolic markers. On the other hand, individuals from lower power groups—that is, employees from acquired firms—will tend to engage more in boundary reformation given that they face greater pressure (and control by dominant group members) to adapt to the pre-existing culture of the dominant group. Innovating new symbolic markers of distinction might provide lower-power group members with an alternative source of agency to respond to the identity threat that arises through exposure to dominant individuals.

Next we consider how an individual’s social context influences the form of boundary maintenance she engages in. Individuals who interact more frequently with outgroup members may develop more towards them—in large part because intergroup contact promotes cultural openness and exchange (Suzuki 1997; Brannon and Walton 2013). In other words, individuals are more inclined to blur symbolic boundaries when they interact more frequently with members of their outgroup. The converse is that infrequent exposure to the outgroup might result in less cultural exchange and greater perpetuation of existing symbolic differences (Coleman 1988). Comparing boundary retention and reformation, we therefore anticipate that individuals surrounded by a greater proportion ingroup, rather than outgroup, members have more exposure to existing symbolic markers of distinction. As a consequence, they will be more likely to reenact these pre-existing markers—that is, engage in boundary retention—than to culturally adapt and innovate new ones—that is, to engage in boundary reformation.

Finally, we conduct exploratory analyses related to the question of how the two forms of boundary maintenance might relate to an individual’s performance in a post-merger organizational setting. Prior work has shown that individuals who do not adapt to the prevailing organizational culture are likely to suffer performance penalties (Srivastava et al. 2018). Building on this insight, we anticipate that boundary retention will result in more negative evaluations for individuals from lower status groups—employees from acquired firms—because retaining pre-existing cultural differences signals the failure to successfully enculturate to the dominant culture. In contrast, when higher status groups—employees from the acquiring firm—exhibit boundary retention, they are still conforming to the dominant culture and will therefore avoid receiving harsh evaluations. As for boundary reformation, it is theoretically unclear whether it will be a boost or detriment to an individual’s performance. In some social contexts, symbolic innovation may be perceived positively, while in others, it may be construed as disruptive. Thus, we expect that the performance implications of boundary reformation are likely to be context specific.

Empirical Setting

We focus on the context of organizational mergers and acquisitions (M&A). Most mergers and acquisitions involve some level of integration between the two organizations (Bodner and Capron 2018). ¹ Thus, most mergers result in a clear erosion of existing social boundaries between two groups. Moreover, many mergers are implemented with an eye to creating a unified combined culture (Nahavandi and Malekzadeh 1988; Weber and Camerer 2003; Van den Steen 2010). In some cases, this uniform culture is mostly reflective of the one that prevailed in the acquiring organization, while in other instances the emergent culture is a blend of the pre-existing cultures or a novel recombination of them. Regardless, the erosion of social boundaries between organizations in a post-merger context requires people to engage in various forms of boundary work as they make sense of the new social order and seek to find their place within it (Amiot, Terry and McKimmie 2012; Drori, Wrzesniewski and Ellis 2013).

Our research setting is a mid-sized regional bank in the U.S. ( n 1 = 306 ) that acquired and integrated two other regional banks ( n 2 = 247 , n 3 = 51 ) over the course of two years. The second acquisition took place exactly 12 months after the first. Conversations with both employees and senior management, conducted nine months after the second merger and prior to archival data analysis, revealed several key features of the setting that informed our analytical approach.

First, both mergers constituted a sudden blurring of the institutionalized boundaries between organizations. Neither merger event was known to or anticipated by rank-and-file employees until announced. In fact, the vast majority of employees did not experience significant change or threat to their routines or communication patterns until the week the merger went into effect, with all architectural integration occurring on what one of our interviewees referred to as a single “flipping weekend…when all the signage and systems are flipped.” This quasi-exogenous discontinuity to individuals’ experience inspired key decisions around data segmentation and variable construction discussed below in our measurement approach.

Second, employees became acutely aware of cultural differences following the mergers, making this empirical setting especially appropriate for the study of symbolic boundary work. Given that the rationale for the mergers focused on geographic complementarities between the branch networks of the different banks rather than on their anticipated cultural alignment, interview accounts suggested that employees experienced varying degrees of “culture clash.” For example, while one employee perceived an “easy transition as far as the culture…because it was the same,” another felt that “the merger was so hard because you thought you were getting a bank with the same culture [but] it was the opposite.”

Third, because there was both employee hiring and attrition after each merger, our sample selection for analyses includes those individuals who were present before and after a given merger. Interviews mitigated concerns of any systematic selection of these individuals based on their cultural behavior. Inspecting the archival data confirms that employees who departed at the point of merger or in the following month did so mostly because they were in job roles that were deemed redundant rather than for reasons of cultural fit (i.e., they did not exhibit meaningfully different interactional language).

Finally, significant cultural variation was apparent both before and after the mergers. Because the bank branches were geographically dispersed and staff relied on local knowledge and customer relationships, management acknowledged and allowed for some level of cultural variation within the firm. Variation also existed at the department level, since different occupations and specializations lent themselves to different norms. For example, it was widely acknowledged among employees that aggression and competitiveness were the norm in commercial banking, while these tendencies were frowned upon in retail banking. This cultural variation has two important implications for our analysis: first, we exploit precisely this variation to examine differences between individuals in their symbolic boundary maintenance. Second, it necessitates that we model between-group differences in a manner that accommodates this within-organization heterogeneity. We elaborate on this point in discussing our analytical strategy.

Beyond access to a subset of employees whom we interviewed, the firm provided us with the complete corpus of internal email communication data (totaling roughly 1.5 million messages), including both metadata and hashed content, spanning 23 months—from three months before the first merger through eight months after the second merger. In addition to communication data, we had access to pre- and post-merger personnel records for both the acquiring firm and the two target firms. For ethical and security purposes, all data were deidentified by replacing names, addresses, and other identifying information with codes. To safeguard the content of email messages, we extracted information about the linguistic styles used in the text of emails, which were represented as numerical vectors, and then hashed the message content such that the original text was unrecoverable.

Language as a Measure of Boundary Maintenance

Studies of symbolic boundaries have drawn on various types of data, from in-depth interviews and ethnographic accounts to large-sample surveys (Lamont 1992; Bail 2008). Although these methods have yielded valuable insights about how people maintain and manage cultural distinctions, they also have important limitations in uncovering fine-grained variation across individuals and over time. Survey-based approaches, though designed to assess individual-level variation, capture symbolic boundaries at a single point in time and are subject to various forms of self-report bias (Greenwald and Banaji 1995). Moreover, survey instruments are based on coarse-grained cultural categories that are defined by the researcher and that may not be relevant for some individuals and in particular contexts. Ethnographic approaches are well-suited to inductively detecting subtle symbolic distinctions; however, they are difficult to scale to large groups and therefore require extrapolating from a subset of observations over a focused period of time.

As an alternative, we propose that the natural language people use in communicating with members of their own group and other groups can provide a granular window into how they maintain and manage cultural distinctions over time. Indeed, scholars have long recognized the primary role of language in demarcating boundaries (Gumperz 1977; Gieryn 1983). Because language is laden with cultural meaning (Fishman 2012), it is especially well-suited to inductively identifying symbolic distinctions. Moreover, language-based analyses do not require the researcher to impose a pre-defined set of cultural attributes and are thus a good match for our focal task of disentangling the use of novel versus pre-existing symbolic distinctions. Finally, the tools of computational linguistics and machine learning make it easier than ever to extract individuals’ cultural signals from increasingly widespread digital trace data—for example, emails exchanged among employees (Goldberg et al. 2016b), messages posted to online platforms such as Slack (Lix et al. 2022), and worker ratings of their employers on platforms such as Glassdoor (Corritore, Goldberg and Srivastava 2020).

To tease apart symbolic distinctions from the functional aspects of language, we specifically measure the linguistic style of communication as reflected in emotional, cognitive, and cultural categories, rather than context-dependent topics such as particular people or events that different social groups might reference. Our approach is related to and builds on the interactional language use model (Goldberg et al. 2016b; Srivastava et al. 2018), which uses the well-established Linguistic Inquiry and Word Count (LIWC) lexicon to measure linguistic style (Pennebaker et al. 2007). The LIWC lexicon consists of a dictionary of words corresponding to a different dimension of linguistic style, such as Achievement, Anger, or Assent (see Appendix A for a table of these categories). Prior studies have used the interactional language use model to derive a time-varying measure of individual-level cultural fit in an organization (Goldberg et al. 2016b; Doyle et al. 2017; Srivastava et al. 2018). In these studies, one’s cultural fit increases with the degree of convergence between one’s own linguistic style and that expressed by one’s interaction partners at a given point in time. By abstracting away from specific words to higher-order categories, this technique allows one to assess cultural alignment independent of specific issues that are being discussed in a group. In contrast to most other language-based analyses such as topic models that measure what content is communicated, a LIWC-based approach probes how this content is communicated.

Building on this insight, here we use the LIWC lexicon to assess the extent to which a given communication matches the “group style” of a given social group at a particular point in time (Eliasoph and Lichterman 2003). We can then aggregate these artifact-based measures at the person-level to derive time-varying measures of how individuals use cultural markers to maintain symbolic group boundaries. Before describing our language-based measures in detail, we begin with a brief description of the empirical setting in which we sought to demonstrate the utility of this approach.

Analytical Strategy

Our approach to measuring symbolic boundary maintenance is based on the extent to which a given employee’s linguistic style is aligned with that of her legacy organization rather than that of the new organization to which she is exposed following a merger. One way to identify communication as stylistically representative of one organization versus the other is to infer what constitutes the general linguistic style of each one—for instance, based on how the “typical” employee uses specific LIWC categories. Yet this approach has a critical drawback: it imposes a monolithic or prototypical group style that prevails in the organization. As discussed earlier, there is ample evidence that cultural boundaries are instead characterized by significant heterogeneity (Bail 2008). Indeed, as suggested in our interviews and illustrated in Appendix B, there is more within-bank heterogeneity in linguistic style than there is between-bank variation. Traditional regression and clustering models suffer from this drawback because they require a mathematical continuous mapping between the multidimensional space of LIWC features and group style. A second challenge involves accounting for the complexity of language—specifically, the many interactions between LIWC features that characterize group style. For instance, the symbolic implications of language employing future tense might vary based on the presence of positive versus negative emotion. Moreover, these interactions are largely unknown ex ante, handicapping any parametric modeling approach. These two issues point to the need for a model of group style that is neither parametric nor a continuous mapping.

To accommodate these requirements, we draw on the idea that the quantified output of a flexible machine learning classifier can itself be a measure of divergence between group styles (cf. Gentzkow, Shapiro and Taddy 2019). We distinguish between group styles by training a classifier to predict the legacy organizational affiliation of the sender of a focal email message based on the distribution of LIWC counts in the message. To obtain a precise, fine-grained measure of boundary work, we choose to model linguistic style at the level of analysis of a single email communication, rather than at the level of a multi-email thread (which involves multiple senders) or bundling a single sender’s emails (which might distort our LIWC counts by averaging across distinct styles). ²

Among classifiers, random forest models are appropriately suited to the complexity of this task: being completely nonparametric and allowing for discontinuities, they are built to deal with multiple heterogeneous patterns within each classification, and they permit nonlinear and interdependent relationships between linguistic features. ³ An ancillary benefit of a random forest model is that it eliminates the need for normalizing LIWC counts of emails by message length. While the relationship between the inputted features (in this case, LIWC counts) and the classification of bank origin is not easily interpretable, the output is a model that inductively maps patterns and interactions of linguistic features to each organization. This mapping is sufficient for our purposes as we are not concerned with the content of each organization’s linguistic signature.

Measurement of Symbolic Boundary Maintenance

To derive our main measures of symbolic boundary maintenance—boundary retention and reformation—we took advantage of organizational merger events as quasi-exogenous breachings of social boundaries and made use of our longitudinal email corpus by segmenting our classifier training data into pre- and post-merger periods. We trained two sets of classification models: one of historical differentiation and another of contemporaneous differentiation (illustrated in Figure 1). Our model of historical differentiation was constructed from a single random forest classifier trained on messages sent in the three months prior to each merger. Intuitively, this model corresponds to the symbolic boundaries that existed when the focal merger was impending but prior to actual integration and significant intergroup contact. In contrast, our model of contemporaneous differentiation was based on multiple monthly classifiers trained on messages sent in the same post-merger month as the test data. Each of the trained models represents the linguistic markers that characterized symbolic boundaries in a given month. We applied these models to paired test datasets from contemporaneous months to measure the extent to which messages could be classified according to contemporaneous symbolic distinctions.

OPEN IN VIEWER

Figure 1.

Illustration of the Modeling Approach for Historical and Contemporaneous Differentiation.

Our measure of boundary retention follows straightforwardly from the historical differentiation model. Theoretically, boundary retention reflects the extent to which pre-existing symbols of distinction are perpetuated. We obtained this empirically by applying the historical differentiation model to test datasets from subsequent months to measure the extent to which post-merger messages reflected the maintenance of pre-merger distinctions. Although the performance of the historical classifier trained on pre-merger data is inevitably worse when applied to post-merger time periods, the objective is to examine the extent to which the historical model retains explanatory power for a given individual in future time periods. We calculated boundary retention as the proportion of an individual employee’s ( i ) outgoing messages ( m → i , n ) in the six months post-merger (in the test dataset) that were classified correctly according to the historical differentiation (HD) model—that is, as more aligned with the pre-merger linguistic style of the focal employee’s legacy organization ( org i ) than of the new organization to which the person was exposed (see Equation 1).

Boundary Retention i = 1 N i ∑ n = 1 N i 1 ( or ∧ g H D ( m → i , n ) = org i ) )

(1)

Boundary Reformation i = 1 N i ∑ n = 1 N i 1 ( or ∧ g C D ( m → i , n ) = org i ) ) * 1 ( or ∧ g H D ( m → i , n ) ≠ org i ) )

(2)

Data Sampling and Model Training

We undertook several critical steps to construct the random forest models of historical and contemporaneous differentiation, on which the measures of boundary retention and reformation are based. First, for each of the two mergers, we constructed separate panel datasets of messages sent both before and after the merger. With these two panel datasets, we filtered the communication data along a number of dimensions. We excluded messages from individuals who did not experience the focal merger (i.e., who departed prior or joined following) or who sent fewer than 20 total messages in the six months following the focal merger. We also filtered out idiosyncratic messages with over 500 terms. For the second merger, we filtered out messages sent by individuals who had been acquired in the first merger, since their social group identity was already in flux and potentially ambiguous.

Second, we carefully constructed multiple training and test datasets according to the needs of each theoretical construct. We divided the two panel datasets of messages (one for each merger) into separate monthly datasets and split each dataset of monthly messages into equal-sized training and test sets. ⁵ The monthly training datasets were downsampled to a maximum of 50 messages per individual so that no single individual’s linguistic style dominated the model. We made this choice to avoid overfitting to a handful of prolific individuals. Using training data from the appropriate months (as outlined above and illustrated in Figure 1), we trained multiple random forest classifiers. Models were trained in R using the caret and ranger packages, with 5-repeated 10-fold cross-validation to tune parameters and conventional downsampling to balance classes. Class balancing helped ensure that, even in the absence of linguistic information, messages were equally likely to be classified as belonging to either of the two organizations—even if employees in one organization tended to send more messages than their counterparts in the other. See Figure 2 for an illustration of the procedure for sampling training data.

OPEN IN VIEWER

Figure 2.

Procedure for Sampling Training Data for Each Month.

Finally, to validate our models of historical and contemporaneous differentiation, we calculated an aggregate month-level indicator of model fit based on a well-established machine learning metric. Specifically, for both the historical and contemporaneous models, we measured the area under the receiver operating characteristic curve (AUC) for each monthly test dataset (including messages from all individuals). The monthly AUC is an explanatory measure (akin to an R 2 metric), representing how well a model can distinguish organizational affiliation of messages in a given month according to their linguistic style. We interpret this ability to distinguish as an aggregate measure of the strength of the historical or contemporaneous symbolic boundary between the merged organizations. Note that this aggregate measure of boundary strength is not the main dependent variable in our analyses; it is separate from and in addition to the individual-level measures of boundary retention and reformation described earlier.

Figure 3 summarizes the entire data processing pipeline.

OPEN IN VIEWER

Figure 3.

Flowchart of the Entire Data Processing Pipeline.

Additional Variables

In addition to these cultural measures, we assembled a number of additional variables for the regression analyses described below. First, because the nature of individuals’ boundary work typically differs based on the relative power of their social group, we defined an indicator, acquired, which was set to 1 for employees from the acquired (i.e., less dominant) firm and to 0 for employees from the acquiring (i.e., dominant) firm. Given that power also derives from the structural positions one occupies within an organization, we controlled for individuals’ hierarchical rank based on their pre-merger job titles. Specifically, we identified whether individuals were senior leaders, middle managers, or individual contributors. Hierarchical position is coded ordinally, with 2 indicating senior leaders, 1 middle managers, and 0 individual contributors. Because employees’ cultural behavior typically evolves over their tenure in an organization, we also included employee tenure as measured in years at the date of the merger (Srivastava et al. 2018). In our models of individual performance, we also included gender (to account for gender bias in evaluations) using an indicator set to 1 for male employees.

To understand how an individual’s local context shaped their tendency to engage in boundary maintenance through retention versus reformation, we developed a measure of local ingroup density, which is defined as the proportion of employees in the same bank branch as the focal individual who share the same pre-merger affiliation (averaged over the six months post-merger). Finally, for exploratory analyses of how boundary maintenance might relate to perceived performance, we used individuals’ post-merger performance ratings, which were assessed on a 1 to 5 scale in the six months following each merger. ⁶

Summary statistics and pairwise correlations of measures are reported in Appendix C.

Estimation

We estimated ordinary least squares regression models to explain variation in the extent to which individuals engage in boundary reformation and retention. Our dependent variables were boundary reformation and retention, and we included as regressors individuals’ local ingroup density, tenure at date of merger, and an indicator of acquired status. We also included controls for individuals’ hierarchical position and the interaction between acquired status and local ingroup density. ⁷

In our exploratory analyses of the association between boundary maintenance and individuals’ post-merger performance ratings, we also estimated ordinary least squares regression models that included boundary reformation and retention as regressors, as well as controls for acquired status, hierarchical position, gender, and tenure. We also included interaction terms between acquired status and the two measures of boundary maintenance.