Started From the Bottom,Now We’re (Working) Here: Workplace Segregation by Social Class Origin

Abstract

Segregation is a pervasive social phenomenon that separates individuals across social settings. Workplaces, as key hubs of the labor market, are not only central to economic outcomes but also potential sites for cross-cutting social ties. Yet workplace segregation can exacerbate inequality and weaken social cohesion. Existing research has largely examined highly visible categories such as race and ethnicity, while much less is known about segregation by subtler statuses. Social class origin is one such case: it is harder for individuals to observe in interaction and more difficult to measure at the population level. In this paper, we analyze workplace segregation by social class origin using Swedish register data from 2012. We find substantial segregation by social origin, well above what would be expected if employees were randomly allocated across workplaces. Segregation also persists beyond what can be explained by individual and workplace characteristics such as education, occupation, or region. Even after adjusting for our richest set of covariates, employees remain somewhat overexposed to coworkers who share their class origin. Notably, excess segregation is smallest among workers from the lowest origins and largest among those from the highest.

Being the hub of the labor market, the workplace is central to accessing labor market rewards by providing employment, job ladders, and wages. Research demonstrates that workplace segregation structures substantial portions of earnings inequality. But the workplace is also a potential facilitator of social interactions, providing a setting where individuals from diverse social backgrounds encounter one another. It may enforce social interaction, influence social cohesion, the formation of political attitudes, and therefore shape broader patterns of inequality and social integration.

Nevertheless, segregation is a pervasive social phenomenon, separating individuals of various categories or groups across schools, neighborhoods, and workplaces. Workers sort into workplaces based on their preferences, the employers’ preferences, the job structure of the workplace, but also on their labor specialization, such as occupation and educational qualifications, and other merits. This might give rise to workplace segregation and create systematic differences in access to rewards, the opportunity structure for social contact, and exposure to various subpopulations in society.

Despite that persistence in income or social class over generations of parents and children is pervasive (Solon, 1999), our knowledge of workplace segregation by social background is virtually non-existent. Although workplace segregation is well documented for salient social categories like gender (Reskin, 1993) and race/ethnicity/immigration status (Hellerstein & Neumark, 2008; Åslund & Skans, 2010), meaning that females and males often work together with coworkers of their own gender and minorities tend to be concentrated in certain workplaces and overexposed to coworkers of their own kind, little is known about segregation by a fundamental stratification dimension, namely class origins. Social class origin is special in that it is not salient but only weakly discoverable by indirect cues such as class-linked behavioral and interactional signs (Kohn, 1969; Groves, 2005), language use, speech, and communicative style (Bernstein, 1971), or via cultural practices and signals (Chan & Goldthorpe 2007).

In this paper, we provide (most likely) the first documentation of class origin segregation between workplaces. With social class, we mean large clusters of occupations sharing the same working conditions, skills, and employer-employee relations (Erikson & Goldthorpe, 1992; Goldthorpe, 2000; Tåhlin, 2007). Since segregation requires measurement of the whole workforce, this requires both population-level data and measurement of parents’ social class. Since the workforce at a given point in time is very heterogeneous in age (from approximately 20-year-olds to 70-year-olds), a very long observation window back in time is required to capture parents’ social class. Swedish register data offers a unique opportunity to study this phenomenon.

Since workplace segregation can arise both because of random and systematic sorting, we estimate segregation above both a baseline counterfactual random distribution of workers across workplaces and a counterfactual random distribution conditional on multiple observed characteristics. The key research question is: “To what extent are workers segregated across workplaces by social class origin, and how much of that segregation is accounted for by randomness and sorting into workplaces based on educational attainment, occupational specialization, sector, and geography?” We use register data on the workplaces of the entire Swedish workforce in 2012. We find substantial segregation by social background in the baseline, i.e., over and above what one would expect under a purely random assignment of working individuals to workplaces. When comparing to counterfactual segregation levels under random allocation conditional on covariates, in particular education and occupation, this ‘’excess’’ sorting by social class origin is much smaller. That is, our results indicate moderate amounts of workplace segregation by class origin after accounting for, e.g., occupational sorting and sorting according to human capital. One possible conclusion from these results would be that the observed segregation levels can largely be explained by the factors conditioned on in the analysis. However, another possible interpretation is that despite conditioning on rich sets of covariates, employees of a certain class origin still have (slightly) more coworkers sharing a similar origin compared to what would be expected under a conditionally random distribution.

Our study of class-origin segregation in workplaces contributes to a broader literature documenting how organizations reproduce inequality. Research shows that boundaries of gender (Fuller & Kim, 2024; Hart, 2025), sexuality and gender identity (Hutchinson et al., 2025), and migration background (Sprengholz & Hamjediers, 2024) structure workplace opportunities. Moreover, organizational logics—such as hiring for ‘fit’ (Nichols, Pedulla and Sheng 2025) or divergent diversity trajectories (Han & Tomaskovic-Devey, 2025)—play a central role in sustaining or altering inequality regimes. At the same time, macro-level labor market transformations, including union decline and debt (Rhodes & Dwyer, 2025), gig work (Auguste et al., 2024), and automation (Damelang & Otto, 2024), reshape the terrain of risk and opportunity. We extend this scholarship by showing how class origin itself functions as a principle of workplace segregation.

Background

Workplace segregation is important because the workplace is the nexus of the labor market and where individual careers are determined. We first discuss the implications of potential workplace segregation and then address why it may matter for social relations.

The Role of Workplaces for Inequality

The sorting of individuals to workplaces is an important determinant of individuals’ labor market and career outcomes. Lazear and Shaw (2009a) documented that workplaces structured a great deal of earnings inequality over and above individual characteristics. For Sweden, Skans et al. (2009: 244) estimated that workplaces alone explain some 30 percent of the variance in labor earnings in the early 2000s. This is on a par with the contributions of an individual's education, age, gender, and migration status (for the same year). Recently, comparative research has suggested a universal increase in wage inequality across workplaces, highlighting how workplace segregation and inequality are tightly linked (Tomaskovic-Devey et al., 2020).

Abowd et al. (1999) outlined a model to analyze which more productive workers select (or match) into high-paying workplaces. Although the literature has produced disparate findings, ranging from negative to positive estimated correlations between worker and firm pay levels, some of the recent literature suggests that such selection may be increasing over time (Card et al., 2013). To the extent that sorting depends on productivity, this could be interpreted as some form of economic efficiency. However, to the extent that sorting is also contingent on non-productive factors, workplace segregation might exacerbate social inequality of opportunity by, e.g., structuring access to rewards. Hällsten (2010; 2013) found substantial wage gaps by class origin when adjusting for individual characteristics. When adjusting for detailed education information, the wage difference between children originating in the upper-service class and the unskilled manual class was some 4–5 percent. Hällsten (2010) showed that adjusting for workplace segregation led to a non-negligible further reduction in the estimated wage gap, suggesting a small but possible role of segregation in social persistence across generations in line with the views of many scholars, who have theorized that workplaces are central to social inequality because they are arenas where actors struggle over limited resources (Baron & Bielby, 1980; Tilly, 1999; Tomaskovic-Devey & Avent-Holt, 2019). Hence, workplaces may play an important role in intergenerational inequality, as the better rewards of some workplaces are only available for insiders who get access. At the same time, persons who are not employed will be outsiders in a double sense.

Workplaces and Social Integration

Workplaces play a double role in social integration. First, they are a foci (Feld, 1981) where people meet and interact. The sorting of individuals into the workplace foci is what generates segregation. Second, social class origin may continue to structure interaction patterns within the workplace. Workplaces may mirror outside hierarchies through homophily and organizational culture, privileging dominant-class practices and interactional styles (Lamont & Lareau, 1988; McPherson et al., 2001; Rivera, 2012; Nichols, Pedulla & Sheng, 2025). However, because firms coordinate diverse workers toward common goals rather than only status domination, they can foster cross-boundary contact (Feld, 1982; Marsden, 1990; Zetka, 1992; Erickson, 1996; Mutz & Mondak, 2006), sometimes yielding more heterogeneous networks than neighborhoods (De Souza Briggs, 2007). Crucially, a workplace's integrative potential depends on who gets in. When a work unit pools similar people, it reproduces similarity; when it pools dissimilar people, it creates opportunities for cross-boundary ties (Feld, 1982). In practice, work settings often generate more cross-cutting interaction than other everyday contexts: using five surveys, Mutz and Mondak (2006) show the workplace is the setting most likely to expose people to discussants with dissimilar political views—more so than family, neighborhoods, or voluntary associations—largely because conversations at work involve more discussants and greater viewpoint diversity. This helps explain why work ties can be more heterogeneous than neighborhood or kin-centered networks: Marsden's classic study of Americans’ core discussion networks shows they are small and heavily kin-based, which tends to limit heterogeneity—a contrast with the larger, mixed-origin networks people encounter at work (Marsden, 1987). Inside firms, what makes coordination effective is not “high-status culture” but practical, shared repertoires that cut across status lines—Erickson (1996) argues that elite culture is largely “profitless irrelevancies” for private-sector coordination, whereas common talk (e.g., sports) facilitates collaboration—though such cultures can still exclude women and minorities. Evidence shows that diverse workplaces increase cross-boundary contacts.

Research on interracial friendship finds that socializing with coworkers (and community involvement) is strongly associated with having friends of another race or natives having immigrant backgrounds, underscoring workplaces’ capacity to bridge social divides beyond what residential contexts alone provide (De Souza Briggs, 2007; Kokkonen et al., 2015). Taken together, these findings suggest that workplaces can foster cross-boundary contact and, at times, produce more heterogeneous networks than other foci such as neighborhoods. Yet such integrative effects materialize only when selection, discrimination, and homophily do not first segregate the workforce. Where mixed-origin teams form, workplaces can bridge class boundaries; where segregation prevails, the same mechanisms undermine that potential.

In conclusion, whether workplaces integrate or segregate is an open question, but beyond the scope of this paper. In this paper, our focus is on the set of mechanisms that sort people into firms.

Mechanisms of Workplace Sorting

We organize the mechanisms that structure workplace segregation as self-selection, discrimination, and homophily.

Self-Selection

Several mechanisms may explain how people sort into workplaces based on preferences or incentives that are hard to observe. These will contribute to baseline workplace segregation that we cannot further disentangle. One such form of self-selection is the inheritance of employers—parents and children working in the same firm (Corak & Piraino, 2011). In Canada and Denmark, 30–40 percent of young adults have worked at a parent's employer (often in their teens), and 4–6 percent hold their main adult job there (Bingley et al., 2011). This mixes family-business succession with occupational inheritance (Jonsson et al., 2009), as when two generations of doctors work in the same large hospital.

Firm wage premia also attract unequal origins. High-origin workers disproportionately match high-paying firms, while lower-origin workers are channeled to weaker wage structures (Lazear & Shaw, 2009b; Tomaskovic-Devey et al., 2020). Using Swedish administrative data, Engzell and Wilmers (2025) estimate that about a quarter of the midlife intergenerational earnings correlation reflects sorting across unequally paying firms. Only a small share is direct employer inheritance; most operate via selection on education and occupation into high-paying firms.

On top of these hard-to-observe factors, structural factors will also determine workplace segregation. Since social background has a pervasive effect on educational attainment both internationally (Shavit & Blossfeld, 1993; Breen & Jonsson, 2005) and in Sweden (Erikson & Jonsson, 1996), this means that education channels workers into different occupations—higher education into white-collar roles, vocational tracks into blue-collar work—so, because much intergenerational reproduction runs through education, it mediates class-origin segregation. Further, occupations are transmitted across generations (Jonsson et al., 2009), so a workplace's occupational mix strongly predicts its class-origin composition. One important channel is the horizontal stratification in education (Hällsten & Thaning, 2018), which helps translate social origin into occupational segregation. Moreover, status orders of occupations vary little across firms and tend to reflect a national hierarchy (Avent-Holt et al., 2020). This implies that hospitals with many physicians will also employ many workers whose parents were physicians, and analogous patterns hold elsewhere.

Geography further structures selection as urban areas host more high-skilled workplaces. However, there is both an origin (where they grew up) and a destination (where career options are) effect involved here. Individuals of lower origins are often concentrated in local areas, more often rural, with poorer access to skilled occupations. For them, upward mobility often requires internal migration. Adjusting for geography will capture both the effect of geographic origin and destination effects.

Discrimination

A second mechanism is discrimination against applicants or employees from less advantaged origins. Again, this is not something we can observe, but it will contribute to baseline segregation. Employers may devalue candidates based on presumed competence, ambition, or “fit,” often via subtle cues in speech, accents, manners, clothing, and interactional styles (Kohn, 1969; Bernstein, 1971; Groves, 2005; Chan & Goldthorpe, 2007). Hiring often favors the employer's own class background (Rivera, 2012; Rivera & Tilcsik, 2016). In UK elite law firms, Ashley and Empson (2013) show class-based screening justified by dual imperatives: attracting talent while reducing risk and protecting image. Because gatekeeping roles (owners, senior partners, hiring agents) are themselves high-class positions and reproduce at a micro-class level (Jonsson et al., 2009), access advantages accrue to higher-origin candidates.

That said, workplace specialization constrains discrimination's scope: a carpentry firm mostly needs carpenters; a law firm, lawyers. Individuals of lower origin may escape discrimination if they are socially immobile. Especially in smaller firms, employers may not be of higher-class origin, which can temper access asymmetries. Still, where discriminatory judgments prevail, they convert pre-existing segregation into current segregation by filtering who enters which firms and roles.

Homophily

The third mechanism is homophily—ingroup favoritism without necessarily negative intent. In networks, people prefer similar others (McPherson et al., 2001), and this extends to workplaces (Williams & O’Reilly, 1998; van Knippenberg & Schippers, 2007). Employers and employees may gravitate toward, mentor, and evaluate peers of similar class origin; job seekers may avoid settings where they anticipate marginalization or a poor “fit.” In class-origin terms, homophily often rests on cultural comfort rather than visibly marked boundaries, and judgments are frequently framed as neutral assessments of “fit” (Nichols et al., 2025). While homogeneity can sometimes aid coordination (Alesina & La Ferrara, 2005), class-origin favoritism relies on shared subcultural repertoires—reinforcing everyday practices that align with dominant-class norms and potentially misrecognizing working-class styles as deviant. This sustains divisions in daily interaction and can inhibit trust and collaboration unless countered by organizational routines that reward cross-class coordination.

Even if entries are the main focus, workplace segregation can also be sustained by differential exits: if workers who are “out of place” in a workplace composition are more likely to leave, segregation can persist (or re-emerge) even when hiring becomes more integrated. A plausible mechanism is that workers from a minority origin in a given work unit may experience weaker social integration, stereotyping, or a sense of misfit—processes that your framework already treats as part of discrimination and homophily (“fit” judgments; anticipating marginalization). In gender research, exits from settings where one is numerically underrepresented have been shown to contribute to the persistence of segregation (Hamjediers & Peters, 2024). More broadly, organizational demography and workgroup diversity research finds higher turnover risks among isolated or dissimilar individuals (e.g., relational demography and isolation arguments), which is consistent with the idea that “being different” in a unit can reduce attachment and increase exit (O'Reilly et al., 1989; Williams & O’Reilly, 1998) Classic tokenism theory likewise emphasizes pressures on numerical minorities that can motivate withdrawal (Kanter, 1977). Once again, this is not something we can observe, but it will nonetheless potentially contribute to workplace segregation.

Summary

Our research question is helped by five hypotheses. The first is that there is an observable, unconditional workplace segregation by social class origin.

The second hypothesis is that this is not entirely driven by randomness

H1: There is an unconditional workplace class origin segregation, net of randomness

The second hypothesis

H2: Educational attainment explains part of workplace class origin segregation

is guided by the unanimous finding in previous literature of strong social origin effects on education. The third hypothesis similarly

H3: Occupational specialization (occupation, sector) explains part of the workplace class origin segregation

is motivated by various literatures showing that social origin effects do not end with educational attainment, but generate occupational segregation or specialization. The fourth hypothesis is

H4: Geography explains part of the workplace class origin segregation, based on the insight that labour markets are geographically segmented, with a concentration of skilled occupations in urban areas.

Our final hypothesis is

H5: Net of structural factors, there will be residual workplace class origin segregation

which captures all those mechanisms discussed above, for both entry and exit sorting, that we cannot measure and condition on in our analyses.

Data

To study segregation levels at a particular moment in time, we use Swedish population-level register data for the year 2012. This is linked employer-employee data for the full population. We collect information on workplaces from tax registers and link this to data on individuals from other registers. Workplaces are defined as the part of a firm that exists in a unique physical location, delimited by a postal address (coded by the Swedish CFAR number, which we use in a pseudonymized form). Our sample for the cross-sectional segregation analysis includes all employed individuals in workplaces with 2 or more employees during the single year 2012. We remove individuals with missing information on the workplace, gender, and birth country

Most importantly, our data allows us to measure class origin for a large majority of workers. The data demands are quite high since we both need to link workers to their parents, and observe their parents’ occupation in a reasonable age. Our class origin is established through the use of the multigenerational register (Statistics Sweden, 2010), which links parents and children based on birth records. We then construct measures of social class based on census data and the occupation register. The censuses go back to 1960, then they are repeated every five years to 1990. The occupation register, which exists annually from 2001 and onwards (Statistics Sweden, 2004). We code social class using the scheme SEI (in Swedish: socioekonomisk indelning, Statistics Sweden, 1982), which is a Swedish variant of EGP (the Erikson-Goldthorpe-Portocarero class schema; Erikson & Goldthorpe, 1992). This measure groups occupations into broader aggregate classes using working conditions, skills, and employer-employee relations (Erikson & Goldthorpe, 1992; Goldthorpe, 2000; Tåhlin, 2007). Overall, the measure has high criterion and construct validity (Smallenbroek et al., 2022). For the censuses, we lack a consistent native measurement of SEI (which exists only in the 1985 and 1990 censuses) and therefore use the information on NYK (in Swedish: Nordisk yrkeklassificering, based on ISCO-58) that has been linked to a Swedish class scheme SEI via the modal SEI for each NYK code in the 1990 census (Erikson & Jonsson, 1993). For the occupation register, the occupational information is three-digit ISCO-88(com) in its Swedish version, SSYK (in Swedish: Standard för svensk yrkesklassificering). The SEI classes have been constructed in similar ways using the 1985 and 1990 census by taking the modal SEI category for each SSYK category, also using industry information (four categories). We then take the highest (dominant, see Erikson, 1984) observed class position over the career for each parent. Class origin is, in turn, constructed as the dominating class (Erikson, 1984) across both parents and across time. ¹ Due to the high accuracy of the multigenerational links and our very long time span of data, we can measure class origin for nearly 90 percent of the Swedish workforce in 2012. A large part of those with missing information are immigrants born outside Sweden, for whom we have no information on parents. For the Swedish-born population, we have information on 99 percent of the population. We will include a category of missing in order to keep our population sample intact, but one must keep in mind that the analysis we present here represents the native population well, but not the immigrant population.

We will use a version of SEI with eight categories (see Table 1) to analyze segregation, but to simplify our analysis, we also focus on the two extremes of the class scheme and estimate separate models for manual-class origin (both skilled and unskilled) and upper service-class origin, both coded 0/1. ² Our multigroup segregation analyses incorporate all the origin classes: unskilled manual, skilled manual, routine non-manual, lower service, upper service, entrepreneurs, farmers, and a missing category (including e.g., immigrants lacking data on class origin).

Table 1.

Descriptive Statistics for Static Analysis (Year 2012).

	Mean	St. Dev.	Min	Max
Manual-class origin (0/1)	0.396	0.489	0	1
Upper service-class origin (0/1)	0.089	0.285	0	1
Class origin categories
Unskilled manual	0.199	0.399	0	1
Skilled manual	0.197	0.398	0	1
Routine non-manual	0.113	0.316	0	1
Lower service	0.184	0.387	0	1
Upper service	0.089	0.285	0	1
Entrepreneurs	0.063	0.242	0	1
Farmers	0.039	0.193	0	1
Missing	0.117	0.321	0	1
Year	2012	0	2012	2012
Female	0.487	0.5	0	1
Age < 30	0.152	0.359	0	1
Age 30 to 49	0.533	0.499	0	1
Age 50 or above	0.315	0.464	0	1
Non-western immigrant	0.106	0.308	0	1
Own education
Basic education	0.139	0.346	0	1
Non-acad US	0.347	0.476	0	1
Acad US	0.112	0.315	0	1
Post-sec	0.197	0.398	0	1
Tertiary	0.183	0.387	0	1
Post-grad	0.012	0.11	0	1
Missing	0.009	0.097	0	1
Sector
Private Sector	0.725	0.446	0	1
Public sector, State Sector	0.054	0.225	0	1
Public sector, county & municipality	0.221	0.415	0	1
Nr. of observations	4,573,909

Notes: All individuals in workplaces with at least two employees during the year 2012.

In addition to the measures of class origin, we include several covariates for our analyses of conditional segregation. These include non-Western immigrant status (i.e, where the social class origin is missing to a large degree), a gender dummy, dummies for age (age < 30; age 30 to 49; age 50 or above), occupation, education, sector, and county.

Moreover, we included covariates to capture the self-selection mechanisms discussed above for hypotheses H2, H3, and H4. For H2, education is taken from the education register, which in turn is based on graduation records, other administrative data, and self-reports in the censuses. Self-reports make up to 30 percent of all information for the eldest workers, but this declines by birth cohort (see Table B4 in Hällsten & Kolk, 2022). Information on education for immigrants stems mainly from a dedicated survey, but also from other administrative contacts. We have coded this as a categorical variable with six levels. This measure separates both major levels (elementary, secondary, tertiary) and minor levels within each level (e.g., academic vs. vocational secondary education; undergraduate vs. post-graduate tertiary).

For H3, individuals’ current occupation is collected from the above-mentioned occupation register. As said, this is coded to three-digit ISCO, and we further code this into 69 categories similar to micro classes (Weeden & Grusky, 2005; Jonsson et al., 2009) to achieve conceptually meaningful categories while simultaneously reducing the skew in size (see Table A1 for coding and descriptives). Sector separates between private sector, the public state sector, and the public county/municipality. For H4, Counties (in Swedish Län) are 21 different geographical entities included as dummies.

Method

We aim to analyze segregation in the cross-section of 2012. One important feature of workplaces is that many of them are small in size. This means that randomness will play a large role in determining the workplace composition of workers. We thus want to explicitly take randomness into account. Segregation also arises because of the selection on characteristics of workers, for example, educational merits. Our strategy is therefore to analyze segregation above what would be expected under a random allocation of employees to workplaces, conditional on the distribution of such characteristics. Conditioning means keeping specific workers fixed in the workplaces, but reassigning their social origin based on randomness or observed covariates. In this way, we create counterfactual segregation measures that we can compare to the real observed measures.

We study workplace segregation using indices of (systematic) segregation (Carrington & Troske, 1997; Bygren, 2013) and (over)exposure, conditional on the distribution of covariates using a method developed by Åslund and Skans (2005; 2009). We use the classic D segregation index (Duncan & Duncan, 1955) when we analyze dichotomous segregation by either manual-class origin and upper service-class origin, and the H information theory index for the multigroup case of all origin classes (Reardon & Firebaugh, 2002). Both are staple measures of segregation in the literature, and we refer to the mentioned sources for their formulas. The D has an intuitive interpretation (the share of all individuals of origin s who would have to change the workplace to achieve an even distribution), whereas the H index is more obscure since its interpretation is “a normalized likelihood-ratio measure of association between two nominal variables indexing group and unit, respectively” (Reardon & Firebaugh, 2002, p. 42).

For exposure, we borrow from the discussion of residential segregation in Massey and Denton (1988) and think of exposure as the extent to which individuals of different social origins meet by sharing the same workplace. Accordingly, we define exposure E to individuals of some origin of interest s as the fraction of an individual's coworkers of this particular origin s, in line with Åslund and Skans (2010) in their study of immigrant exposure. Similar to the proposed segregation measure of Echenique and Fryer Jr (2007), this directs one's attention to the relationship between segregation and social interactions. ³ In keeping with the reasoning of Åslund and Skans (2005), we argue that a direct focus on the number of (potential) contacts with individuals of some origin of interest is particularly useful when studying the non-ephemeral, non-market interactions between coworkers in a workplace. ⁴

We use simulations to produce counterfactual ‘’expected’’ segregation and exposure indices. As illustrated by Carrington and Troske (1997), and more recently Bygren (2013), a completely random allocation of people over units is expected to produce an uneven distribution when units or minority shares are small. Thus, one should compare any measure of actually observed workplace segregation to a random rather than a perfectly even allocation. To illustrate, assume that one has a large sample of two-employee workplaces and that individuals are of either white-collar or working-class origin. If 50 percent of employed individuals are of working-class origin (in reality it is around 40 percent, but we keep the example simple), random allocation implies 25 percent of workplaces employing two individuals of worker-origin, 50 percent of workplaces employing one individual from each group, and 25 percent employing only individuals of white-collar origin. Thus, we define workplace segregation as occurring when the actual distribution differs from the result of a random allocation. Furthermore, we also account for the distribution of covariates, thus testing for random allocation conditional on some observed characteristics.

The Segregation Outcome

Using a method suggested by Åslund and Nordström Skans (2009), we will compute counterfactual segregation conditional on covariates as follows. Assume that we have a vector X of some discrete characteristics. For each combination of these characteristics, we have a unique type of employee, and we calculate the fraction of individuals of origin s (which takes eight values). In the total population of the employed, let $N^s(X=x)$ be the total number of individuals of origin s and type x, and $N(X=x)$ the total number of individuals of type x. Then, for each combination of characteristics x, we obtain a propensity for being of class origin s equal to the fraction of all x-type employees that are of this particular origin s:

$$p^s(x)={\displaystyle \frac{N^s(X=x)}{N(X=x)}}$$

(1)

In the total population of employed individuals, p^s(x) would be interpreted as the probability that a person with characteristics x is of origin s. To illustrate, assume that X includes immigrant status, gender, age categories, education level, and county. A particular type of employee x could be, say, a non-immigrant female between 29 and 50 years of age, who has completed only a basic level of education and is currently working in the county of Stockholm. Then, p^s(x) would be the probability that this particular type x is of manual class origin. ⁵

We first calculate the observed E, D, and H indices on the population using our real data. Then, we obtain a counterfactual distribution through random assignment of class origins s to employees within each cell defined by a particular combination of characteristics x, where the x:s for each workplace are fixed, using p^s(x) in equation (1) as the probability of being of origin s. ⁶ We calculate counterfactual E*, D*, or H* on the simulated data, providing measures of the expected segregation under random assignment of employees to workplaces conditional on the characteristics in X (if any). The intuition behind this is that we allow sorting according to any characteristics in X, but not according to social origin s (in the binary case). For each index and covariate specification, we use the mean from 100 repeated simulations to average out random variations (since our sample is very large, not many iterations are needed).

To help clarify, we provide an example. Suppose we have two workplaces with ten employees each. The first has 20 percent of workers of manual origin, and the second 60 percent. We now calculate the observed segregation, e.g., D, using standard formulas. To account for randomness, we simulate data without covariates. We thus replace the share of workers of manual origin with draws from the national distribution of workers of manual origin, with a mean of 40 percent. The simulated numbers are 30 and 50 percent, respectively. We can now calculate counterfactual measures, e.g., D*, on these data. Furthermore, we can adjust for covariates. Say that the first workplace is 75 percent young workers, whereas the second workplace is 25 percent. And the national means for manual origin are 30 percent among young workers and 50 percent among older workers. To simulate, we split the workplaces by age into two groups and assign random values drawn from the age-specific national distributions. The simulated numbers for workers of manual origins are now 35 and 45 percent, respectively. To complete the example, we do the simulated parts 100 times each and then average the counterfactual indices.

To analyze the degree of systematic segregation, above randomness, we follow Carrington and Troske (1997), Åslund and Nordström Skans (2005, p. 12–14), and Bygren (2013), and define a systematic segregation index, using D as the example, as

$$D_{systematic}=\{\begin{array}{c}{\displaystyle \frac{(D-D*)}{(1-D*)}\in [0,1]ifD\ge D*}\\ {\displaystyle \frac{(D-D*)}{D*}\in [-1,0)ifD<D*}\end{array}$$

(2)

where D is the observed Duncan and Duncan dissimilarity index, and D* is the simulated counterfactual Duncan and Duncan index under random allocation conditional on covariates (if any). ⁷ H_systematic, is calculated using H and H* and incorporates multiple origin classes, but is otherwise analogously defined.

Finally, for employees of a given class origin s, we define overexposure OEXP as the average actual exposure to coworkers sharing this origin s, divided by the counterfactual average expected exposure to individuals of origin s under random allocation, conditional on any covariates: ⁸

$$\mathrm{OEXP}={\displaystyle \frac{E}{E*}}$$

(3)

The average person of a given origin s then has OEXP times as many coworkers of origin s as would be expected if employees were randomly assigned to workplaces conditional on the observed characteristics in X (if any). To summarize, positive values of D_systematic, H_systematic or OEXP imply that there is systematic segregation over and above random sorting of individuals across workplaces, conditional on any included covariates. ⁹

As noted by Åslund and Skans (2005), there are endogeneity issues associated with this type of conditioning on intermediary covariates. ¹⁰ For example, geographic sorting of workers by county may not be exogenous to the process of workplace segregation by class origin if class origin drives geographic mobility to pursue educational or occupational careers. However, social origin can be geographically clustered and thereby relate to workplace segregation. Likewise, an individual's choice of education and occupation is certainly not exogenous to class origin. Åslund and Skans (2009; 2010) conclude that tightly conditioning on a rich set of covariates might lead to understating true segregation, and argue that the analyses would (hopefully) provide conservative measures of segregation. While our analyses should not be given a causal interpretation, they might nevertheless provide some important descriptive insights into workplace sorting by class origin and other characteristics.

Results

Before discussing our main results, an initial glance at Table 1 shows that our cross-sectional sample, used for the analysis of segregation in 2012, consists of roughly 4.58 million employees. For our measures of segregation in the extremes, measured via D and OEXP, 39.6 percent of these working individuals are of what we define as manual-class origin (skilled and unskilled), while 8.9 percent are of upper service-class origin. We also display the full class origin scheme that we use to calculate overall multigroup segregation with the H index. Moreover, some 10 percent of the individuals are non-Western immigrants, and about 49 percent are female. The workforce is also dominated by middle-aged individuals; around 53 percent are between 30 and 49 years of age.

In Table 2, we present calculations for D and E for both manual class origin and upper service class origin, as well as the H index for all classes. We discuss these three forms of segregation in turn. The first row presents the observed measures, and the later rows present estimated expected or benchmark values obtained through simulations. After the random segregation baseline in the second row, we start conditioning on covariates in our simulations from the third row onwards to account for sorting on various characteristics besides randomness. The last row presents results when conditioning only on the occupation.

Table 2.

Class Origin Workplace Segregation in 2012.

	Manual Origin				Upper Service Origin				All Origin Classes
	Index of dissimilarity		Exposure		Index of dissimilarity		Exposure		Information theory index
	D	Dsystematic	E	OEXP	D	Dsystematic	E	OEXP	H	Hsystematic
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
Measured segregation	0.29154		0.44324		0.42698		0.14359		0.17813
Random segregation baseline	0.15650	0.16009	0.39565	1.12029	0.26866	0.21648	0.08885	1.61599	0.10911	0.07757
	(0.00030)		(0.00025)		(0.00050)		(0.00018)		(0.00008)
+ Age, gender, immigration	0.17305	0.14328	0.40206	1.10244	0.28331	0.20046	0.09237	1.55446	0.13129	0.05391
	(0.00032)		(0.00021)		(0.00053)		(0.00022)		(0.00009)
+ Education	0.21736	0.09478	0.41857	1.05893	0.33085	0.14366	0.11118	1.29145	0.13914	0.04518
	(0.00036)		(0.00025)		(0.00057)		(0.00023)		(0.00010)
+ Occupation	0.23548	0.07332	0.42468	1.04371	0.35561	0.11075	0.11810	1.21580	0.14736	0.03613
	(0.00039)		(0.00028)		(0.00057)		(0.00023)		(0.00009)
+ Sector	0.23696	0.07152	0.42557	1.04152	0.35836	0.10695	0.11919	1.20467	0.14832	0.03515
	(0.00035)		(0.00022)		(0.00057)		(0.00024)		(0.00011)
+ County	0.26080	0.04158	0.43191	1.02622	0.38393	0.06988	0.12702	1.13040	0.15564	0.02661
	(0.00032)		(0.00023)		(0.00059)		(0.00022)		(0.00009)
Occupation as sole covariate	0.20490	0.10896	0.41079	1.07900	0.33283	0.14112	0.10775	1.33255	0.12209	0.06384
	(0.00036)		(0.00029)		(0.00053)		(0.00021)		(0.00009)
Nr. of observations	4,573,909		4,573,909		4,573,909		4,573,909		4,573,909

Note: D_systematic measures systematic (de)segregation in excess of (or below the level produced by) randomness and any conditioning variables. D_systematic = (D-D*)/(1-D*) $\in$ [0, 1] if D ≥ D*, and D_systematic = (D-D*)/D* $\in$ [-1, 0) if D < D*. D is Duncan and Duncan's index of dissimilarity. D* is simulated segregation that adjusts for randomness, with and without covariates. E measures the average actual exposure to coworkers of a certain origin among individuals sharing this same origin. E* is the simulated average expected exposure that adjusts for randomness, with and without covariates. OEXP measures overexposure and is calculated as E divided by E*. The average person of a particular origin then has OEXP times as many coworkers of this same origin as would be expected if individuals were randomly assigned to workplaces conditional on the observed characteristics in X (if any). 100 repeated simulations are used for each estimation. Standard deviation (error) of the simulation distribution within parentheses. H is the information theory index, H* and H_systematic are computed similarly to D* and D_systematic.

As can be seen in the first two rows of the first two coefficient columns in Table 2, more than half of the observed segregation across workplaces for manual-class origin would be expected also if the distribution of employees across workplaces was entirely random. We observe a D of .29, meaning that 29 percent of the workforce would have to change occupations for the occupational distribution to be even in class origin. Under random sorting, however, we would expect D to be .156, and D_systematic takes the value 0.16. A lot of segregation is just randomness. Remember that in this covariate-free case, D_systematic measures the observed level of systematic (non-random) segregation as a share of the maximum possible attainable value of such systematic (non-random) segregation. Differently stated, D_systematic, in this case, tells us how far, percentage-wise, relative to the expected random baseline ( $\mathrm{D}*$ ), non-random processes have moved the observed segregation (D) closer to perfect segregation (or desegregation). Hence, we cannot reject our first hypothesis, H1, of workplace sorting net of randomness.

In the third row, we start conditioning on covariates in our simulations to adjust for sorting across workplaces based on observable characteristics of individuals and their workplaces. As we account for sorting on additional covariates, the index of systematic dissimilarity steadily shrinks. First, we condition on basic demographics (age, gender, immigration), and this explains almost .02 out of the D_systematic of .16. For hypothesis H2, we condition on education, and find that it explains almost .05, a quarter, of D_systematic. For hypothesis H3, we condition on occupation and sector, where occupation explains another .02 but sectors explain virtually nothing. For hypothesis H4, we condition on geography, which explains a further .03). Hence, we cannot reject our hypotheses H1 to H3. Educational level explains a quarter of the variance, and so do occupation and geography in combination.

Once we have accounted for all our included covariates, the value on D_systematic has come much closer to, but not reaching, zero. A D_systematic equal to zero would imply no segregation level under random allocation conditional on the included covariates. In the fully controlled simulation specification for manual-class origin, the level of expected conditional segregation is now almost 90 percent of the level of actually observed segregation (compare D of .26 to the observed .29), and we get an index value for D_systematic of 0.042, meaning that net of everything so far, the observed segregation has moved 4.2 percent closer to perfect segregation. We can also tease out the remaining D as .29-.26 = .03 as the proportion that needs to change their workplace to achieve an even distribution (conditional on covariates and randomness). We cannot reject hypothesis H5 of the remaining residual workplace sorting net of both covariates and randomness. The main conclusion here is that a very large part of the systematic sorting for manual-class origin may be accounted for by workplace sorting according to the observable characteristics included in our list of covariates. Interestingly, as can be seen in the final row of Table 2, conditioning solely on occupation produces levels of expected segregation that are often not very far from the levels of actually observed segregation. This indicates moderate amounts of workplace segregation by class origin after accounting for, e.g., occupational sorting of workers across workplaces.

Columns 3 and 4 of Table 2 present the corresponding results using average exposure (E) as an alternative measure of segregation, focusing on the number of (potential) contacts with individuals of the same class origin. This segregation index measures the proportion of coworkers who are of manual-class origin, averaged over all individuals who are themselves of manual-class origin. Thus, the E used in our analysis is the average ‘’own-group’’ exposure among individuals of manual-class origin. We also include measures of overexposure (OEXP) to coworkers of manual-class origin, i.e., the average individual of manual-class origin has OEXP times as many coworkers of manual-class origin as one would expect under random sorting of individuals across workplaces, conditional on the included covariates (if any). The overall pattern is quite similar to the results for D, so all conclusions regarding our hypotheses H1 to H5 also apply here; we cannot reject any of them. The degree of overexposure decreases steadily as we condition on additional covariates in the simulation process. Equally similar to the case of the dissimilarity index, when including the full list of controls, we observe some overexposure to coworkers of the manual-class origin among individuals who are themselves of such manual-class origin. The average individual of manual-class origin has 1.026 as many coworkers (or 2.6 percent more) of manual-class origin than as expected under random sorting, conditional on age, gender, immigration status, education, occupation, sector, and county.

In the coefficient columns 5 to 8, we study segregation by upper-service origin. One may initially note that, throughout all tables studying segregation levels for class origin, the systematic segregation levels (as measured by D_systematic and OEXP) are always larger for upper service-class origin compared to the manual-class origin. Furthermore, in the analysis of the upper service-class origin, conditioning on (the full list of) covariates never brings us as close to zero systematic dissimilarity, or towards equally low levels of overexposure, as was the case for manual-class origin.

As can be seen in the first two rows of column five, a completely random allocation of individuals across workplaces would imply a value of D of above 60 percent of the value observed. As we account for sorting on covariates, D (found in column six) steadily shrinks, and once we have accounted for all our included covariates, the value is much closer to zero than was initially the case. However, as was also the case for manual-class origin, we still obtain some systematic segregation left to explain. However, as mentioned above, for upper service-class origin, this remaining systematic segregation is larger than the remaining segregation in the analysis of the manual-class origin of columns one and two.

Columns seven and eight present the corresponding results for average exposure (E) and overexposure (OEXP) to coworkers of upper service-class origin among employees sharing the same origin. The overall pattern is quite similar to the results for D. The degree of overexposure decreases steadily as covariates are added to the simulation process, but also when including the full list of controls, the results suggest some overexposure to coworkers sharing one's own upper service-class origin. The average individual of upper service-class origin still has about 1.13 as many coworkers of upper service-class origin as one would expect under a random allocation conditional on age, gender, immigration status, education, occupation, sector, and county. As was also the case for the dissimilarity measures, the remaining overexposure is larger than the corresponding number in the analysis of the manual-class origin of column four.

In columns nine and ten, we use the full scheme and analyze multigroup segregation. Again, we find that random sorting blows up the segregation measures. We observe an H index of .17, out of which .11 is to be expected under a random allocation of employees across workplaces. Accordingly, the H_systematic in the second row is around .08. As we add covariates, this declines gradually. Also in the full model, we observe segregation above what would be expected under a conditionally random allocation, as the H is .026 in the full model. Interestingly enough, the differences between segregation above what we expect under random sorting, and segregation above what we expect under random sorting conditional on occupation, are rather small.

Discussion

We present one of the first analyses of workplace segregation by class origin. We document non-negligible segregation by class origin. The average individual of manual-class origin has about 12 percent more coworkers sharing this manual-class origin than one would expect under a completely random allocation. The average individual of upper service-class origin has 62 percent more coworkers of the same upper service-class origin than expected under purely random sorting. We also find that the random sorting of employees across workplaces can explain a lot of the observed workplace segregation by class origin. Even in a counterfactual scenario where working individuals were completely randomly sorted over workplaces, we would still obtain segregation levels that are about 54 percent of the observed level of segregation (as measured by the dissimilarity index) in 2012 for manual-class origin, and 63 percent of the corresponding segregation for upper service-class origin. In the extremes of the class origin distribution, we observe non-negligible segregation levels also above a counterfactually random allocation conditional on a rich set of covariates, including e.g., occupation, education, age, and gender, especially for upper-service class origin. In fact, throughout all our analyses, the systematic segregation levels are always larger for the upper service-class origin when compared to the manual-class origin. The age gradient in segregation also appears weak, so segregation by class origin is not changing much over time. A note of caution on the possible endogeneity issues in our analyses is warranted. For instance, some of the variables we control for may be caused by class origin, so controlling for them could remove part of the class-origin sorting we want to capture. Furthermore, it is important to remember that measured segregation indices in the first row of Table 2 capture the extent to which employees of a specific social class origin actually are concentrated in certain workplaces, regardless of their explanation. Our measures of segregation to be expected under a random, or conditionally random, allocation should then be regarded as a descriptive device to obtain insight into the segregation processes. Similar to Åslund and Skans (2010), we might note the risk of overcontrolling in our richest specification. For example, part of the geographical composition might be a function of those forces that also drive segregation between workplaces within counties. Nevertheless, also under such tight conditioning, we still obtain some excess levels of segregation by class origin. We should also emphasize that our study only offers a cross-sectional account of segregation. It does not show whether coworkers engage in meaningful interaction across class origins, nor does it provide longitudinal evidence about how this segregation emerges. Rather, it presents what is likely the first documentation of segregation by class origin across workplaces.

Conclusion

We find clear evidence of unconditional workplace segregation by social class origin that cannot be attributed to random allocation alone (H1 not rejected). Consistent with prior work on social origin effects in schooling, educational attainment explains a meaningful share of this segregation, but does not eliminate it (H2 not rejected). We also show that occupational and sectoral specialization contributes additional explanatory power, indicating that class-origin sorting continues beyond education into the structure of jobs (H3 not rejected). Further, geography accounts for part of the observed segregation, in line with geographically segmented labour markets and the concentration of skilled work in urban areas (H4 not rejected). Finally, even after conditioning on education, occupation/sector, and geography, a residual level of workplace class-origin segregation remains, pointing to additional unobserved mechanisms shaping both entry and exit sorting across workplaces (H5 not rejected).

The empirical patterns uncovered here invite further theoretical and empirical inquiry into the mechanisms through which class origin shapes workplace sorting. Recent work on labor market inequality points to the enduring role of elite educational trajectories and organizational gatekeeping in reproducing advantage across generations (Rivera, 2012; Ashley & Empson, 2013; Friedman & Laurison, 2020). Such processes highlight how cultural signals, hiring heuristics, and organizational preferences systematically favor individuals from privileged backgrounds, thereby reinforcing workplace segregation at the upper end of the class distribution. The asymmetric segregation patterns we observe between manual-class and upper service-class origins, therefore, underscore the need to study not only barriers to mobility among disadvantaged groups but also the organizational mechanisms through which privilege is preserved. Future research should also investigate sectoral and institutional variation in these patterns, and explore the long-run consequences for career trajectories, wage inequality, and intergenerational persistence of class advantage.

In line with our finding that segregation is most pronounced at the top of the social-origin distribution, it is useful to note that this pattern mirrors recent comparative evidence on the workplace concentration of economic elites. Using linked employer–employee data from a broad set of advanced capitalist economies, Godechot et al. (2024) document substantial “top earner segregation,” showing that high earners are disproportionately clustered within particular workplaces and work units rather than being diffusely spread across firms. While class origin and earnings are not the same dimension of stratification, the parallel is suggestive: the organizational processes that concentrate those at the top—through selective recruitment channels, internal labor markets, and elite career tracks—may also contribute to the stronger systematic segregation we observe among individuals of upper service-class origin. Connecting these literatures underlines a broader conclusion of our study: workplace sorting appears to be especially consequential for understanding how advantage is maintained at the upper end of the distribution, not only how disadvantage is experienced at the lower end.

An important implication of segregation, given that class shapes political preferences (Evans, 2000), is that enforced contact at work can affect political attitudes (Mutz & Mondak, 2006) and broader social trust via intergenerational transmission (Achen, 2002; Hooghe & Boonen, 2015). The workplace segregation we observe can weaken this tendency. Although our analysis has not centered on policy evaluation, the findings nonetheless carry implications for equality and diversity frameworks. Our results suggest that workplace segregation by class origin constitutes an overlooked but consequential axis of inequality. Addressing this blind spot could entail measures such as monitoring recruitment pipelines, increasing transparency in hiring practices, and broadening access to elite career tracks. Moreover, organizational interventions that explicitly acknowledge class-based disadvantage—such as outreach to less advantaged schools, mentorship schemes, or formal anonymized recruitment procedures—could mitigate the clustering of employees by origin and contribute to more inclusive workplaces. More broadly, incorporating class origin into policy and organizational practice would align with a growing recognition that socioeconomic background is a critical but under-researched dimension of labor market stratification.