Abstract

What is the relationship between gender segregation in higher education and gender segregation in the labor market? Using Fossett's (2017) difference-of-means method for calculating segregation indices and data from the American Community Survey, we show that approximately 36% of occupational segregation among college-educated workers is associated with gender segregation across 173 fields of study, and roughly 64% reflects gender segregation within fields. A decomposition analysis shows that fields contribute to occupational segregation mainly through endowment effects (men's and women's uneven distribution across fields) than through the coefficient effects (gender differences in the likelihood of entering a male-dominated occupation from the same field). Endowment effects are highest in fields strongly linked to the labor market, suggesting that educational segregation among fields in which graduates tend to enter a limited set of occupations is particularly consequential for occupational segregation. Within-field occupational segregation is higher among heavily male-dominated fields than other fields, but it does not vary systematically by fields' STEM status or field–occupation linkage strength. Assuming the relationship between field segregation and occupational segregation is at least partly causal, these results imply that integrating higher education (e.g., by increasing women's representation in STEM majors) will reduce but not eliminate gender segregation in labor markets.

Introduction

The segregation of men and women into different occupations is an especially consequential form of labor market inequality because occupations are central to the work experience and are closely tied to pay, prestige, and health and mortality risks. Our analyses of data from the American Community Survey show that in the United States, 48% of women in the paid labor force would have to switch to a male-dominated occupation for all occupations to have identical shares of male and female workers. Occupational segregation among college-educated workers is less pronounced, but as many as 39% of college-educated women would need to switch to male-dominated occupations to integrate labor markets (see also England et al. 2020).

Scholars seeking to understand occupational segregation have frequently identified educational segregation as a critical precursor. This line of reasoning is especially prominent in the literature on women in STEM (science, technology, engineering, and math), which suggests that women's underrepresentation in STEM fields in college undergraduate and graduate programs contributes substantially to their underrepresentation in STEM occupations (Sassler et al. 2017; Xie and Shauman 2003). For their part, scholars of gender segregation in higher education have frequently justified their analyses by noting the consequences of educational segregation for occupational segregation, the gender gap in wages, and other forms of gender inequality in labor markets (Charles and Bradley 2009; Legewie and DiPrete 2014; Mann and DiPrete 2013; Morgan et al. 2013; Weeden et al. 2020; Xie and Shauman 2003). Although some studies have focused on men's and women's transition from particular fields and into particular types of occupations (Sassler et al. 2017; Xie and Killewald 2012), few have attempted to unpack the empirical relationship between educational and occupational segregation across the full range of fields and occupations.

Our goal is to fill this gap by assessing the relationship between educational and occupational segregation in the U.S. labor market. Our research questions are unabashedly descriptive. We ask, to what extent is occupational segregation among college graduates associated with field segregation in college, and how does this association manifest? We differentiate two proximate sources of occupational segregation among college graduates. The first is a simple compositional effect of educational segregation. If men and women obtain degrees in different fields, and if different fields lead to different occupational outcomes, it follows that educational segregation leads to occupational segregation. This argument recognizes that even in the United States, where the education system famously emphasizes generalized rather than vocational education, fields imply qualitatively different skills, job-relevant training, and credentials. To the extent that employers hire on the basis of these skills, field segregation in college will echo in occupational segregation in the labor market. The strength of this echo will depend partly on the extent to which particular fields are pathways to particular occupations (DiPrete et al. 2017).

A second proximate source of occupational segregation is within-field segregation: gender differences in occupational destinations of those who graduate with degrees in the same field (Shauman 2009). Within-field segregation may emerge through processes unrelated to the fields themselves. Alternatively, it may (partly) reflect gender-specific differences in the returns to a given field—that is, in the probability of entering a male-dominated occupation. As we discuss later, fields can either contribute to or suppress overall levels of occupational segregation, depending on the gender composition of the occupations that the fields' graduates enter.

We assess the relationship between field segregation and occupational segregation with two analyses relying on the 2019 American Community Survey five-year microdata. In the first analysis, we apply Fossett's (Crowell and Fossett 2018; Fossett 2017) difference-of-means method for calculating the index of dissimilarity, D (Duncan and Duncan 1955). This method allows us to estimate the counterfactual segregation index after adjusting for individual-level covariates, including field and (in some models) other standard sociodemographic covariates of occupation. We apply a standard regression decomposition that estimates the share of the overall segregation index that is attributable to men's and women's uneven distribution across fields and to gender differences in the occupational segregation returns to fields. We also examine how these endowment and coefficient effects differ across low-, medium-, and high-linkage fields, where linkage refers to the strength of the pathways between fields and occupations (DiPrete et al. 2017). In the second analysis, we assess the relationship between within-field occupational segregation and three field-level correlates: gender composition, STEM status, and field–occupation (FO) linkage strength.

Educational and Occupational Gender Segregation

In this section, we review potential sources of an empirical association between educational segregation and occupational segregation. The goal is to motivate the descriptive analysis that follows, recognizing that we cannot adjudicate among the mechanisms in this analysis.

Educational Segregation's Compositional Effects on Occupational Segregation

A common, instrumental view of higher education is that its main goal is to train future workers in general or specialized skills that will increase their future productivity. College degrees may signal such training or the capacity to be trained on the job (Becker 1993; Spence 1974; Stiglitz 1975). To the extent that (1) different fields are associated with particular training (or trainability) in particular types of skills required by particular occupations and (2) men and women receive degrees in different fields, gender segregation across fields will have a compositional effect on occupational segregation in labor markets.

A compositional effect could also arise through other mechanisms. For example, theories of occupational closure argue that educational credentials are institutionalized barriers that limit who can practice an occupation, especially if employers will hire only workers with specific credentials or specific credentials are required to obtain an occupational license (Bol and Weeden 2015; Weeden 2002). In this perspective, fields are relevant to occupational destinations as much for the legitimacy and legal rights they confer on workers as for the skills they signal. As with the view of education as human capital, closure predicts that field segregation will lead to occupational segregation through compositional effects, but the underlying mechanism differs.

An apparent compositional effect of educational segregation on occupational segregation could also arise through common causes that generate both field segregation and occupational segregation. Take gender discrimination as an example. In higher education, gender discrimination can manifest as chilly climates or hostile environments for women in male-dominated fields, pushing them out of these fields or pulling them into female-dominated alternatives. In the labor market, gender discrimination may similarly push women out of heavily male-dominated occupations or make female-dominated ones more attractive (Alper 1993; Moss-Racusin et al. 2012). The common mechanism of gender discrimination will generate an observed association between the two forms of segregation.

Gender discrimination is one of many potential common causes. Gender essentialism (Charles and Bradley 2009; Levanon and Grusky 2016), gendered self-evaluations of competency at particular tasks (Correll 2001, 2004), gender-differentiated household and caregiving responsibilities or anticipated responsibilities (Bianchi et al. 2000; Jacobs 1989), organizational practices that impose different constraints on men's and women's choices (Kmec et al. 2010; Reskin and Roos 2009), or deeply rooted gender-specific learning or biologically based attributes and preferences (e.g., Ceci and Williams 2011; Ceci et al. 2009; Hakim 2002) could also cause both field segregation and occupational segregation. These common causes may affect field and occupational decisions directly, or they may contribute to gender-differentiated occupational plans and aspirations that affect later educational decisions, including college entry and choice of a field (Legewie and DiPrete 2014; Mann and DiPrete 2013; Morgan et al. 2013; Weeden et al. 2020). The critical point is that these common causes can generate the observed association between field and occupational segregation, even if field segregation is not in itself a cause of occupational segregation.

Within-Field Occupational Segregation

Fields can also be related to occupational segregation if they differ in the extent to which male and female graduates enter different occupations—that is, if fields have different levels of within-field segregation. For example, suppose that 70% of male and 40% of female mechanical engineering graduates become mechanical engineers and the remaining men and women enter other occupations (as opposed to leaving the labor force). In this case, mechanical engineering will show substantial within-field occupational segregation. If, by contrast, mechanical engineering graduates sort into destination occupations at random with respect to their gender, within-field occupational segregation will be zero.

A field's contribution to overall levels of occupational segregation depends on the gender composition of the occupations its graduates enter and the gender composition of the occupations that the average college graduate enters. Returning to the previous example, if women with mechanical engineering degrees are more likely than other college-educated women to enter occupations that are male-dominated (relative to men's share of the labor force) while men with mechanical engineering degrees are equally or only slightly more likely to enter male-dominated occupations than other men, mechanical engineering will contribute negatively to overall levels of occupational segregation. In theory, this segregation-suppressing effect can occur even if men with mechanical engineering degrees enter different male-dominated occupations than women with such degrees. As we will discuss, the sum of fields' contributions to segregation through such gender differences is indicated by the coefficient effect in a standard regression decomposition. These coefficient effects of fields—the gender-differentiated segregation returns to fields—are one component of within-field segregation.

Both within-field segregation and the coefficient effects of fields emerge from processes that funnel male and female graduates of a given field into different occupational destinations. What social processes might generate this sorting? In our view, gender-differentiated occupational destinations conditional on the field are not anticipated by theories that understand educational degrees as markers of pre–labor market training or trainability. If employers are simply seeking candidates with particular training and skills for particular jobs, in the absence of other social processes (such as discrimination), one would not expect them to differentiate between male and female candidates with the relevant markers of this training. One caveat is that there may be (unobserved) subfields within an (observed) college major that signal different skills and are gender segregated. For example, women who major in mechanical engineering are more likely to focus on health-related applications (Cheryan et al. 2015), perhaps making employers more likely to hire female mechanical engineers for medical research jobs. Aside from any within-field occupational segregation driven by the aggregation of gender-differentiated subfields, the view of education as a signal of productivity does not, on its own, anticipate within-field occupational segregation.

The closure perspective is more ambiguous in its predictions about within-field segregation. Some closure theorists argue that individualist exclusionary principles, such as those based on credentials, largely replaced and delegitimized exclusion based on collectivist principles, such as gender, race, or lineage (Parkin 1979). In this view, as in the human capital perspective, gender-differentiated occupational destinations among graduates with the same credentials should be minimal. However, individualist and collectivist exclusion can coexist (Murphy 1986), and social closure around jobs can be maintained through exclusion from jobs based on gender or other sociodemographic attributes (Tomaskovic-Devey and Skaggs 2002). The latter form of the closure argument opens the door to gender-differentiated occupational destinations among workers with the same educational credentials.

Finally, the same common-cause social processes that generate field and occupational segregation through compositional effects may generate within-field segregation and aggregate-level coefficient effects. For example, gender discrimination could lead employers to prefer male applicants for certain jobs, even among a pool of applicants with the same training and credential-based signals of likely productivity (e.g., Becker 1971; Correll and Benard 2006; Quadlin 2018). Gender essentialist views may mean that male and female employees assess their own competency in particular task domains differently or that employers assess male and female applicants' competency differently, even when those applicants hold the same degree in the same field (Charles and Grusky 2004, 2011; Correll et al. 2020; Kmec et al. 2010; Levanon and Grusky 2016). Likewise, societal norms about dependent care may affect worker preferences (e.g., for part-time or flexible work) and employer hiring preferences (e.g., employers' perceptions of whether a worker will move or travel for a job) even where male and female applicants have the same level and type of degree (Bianchi et al. 2000; Cha 2013; Rivera 2017). Each mechanism can produce gender-differentiated occupational destinations, even among graduates in the same field.

Field-Level Attributes Associated With Within-Field Segregation

Although these general mechanisms may lead to gender-differentiated occupational destinations in any field, it does not follow that all fields will show identical levels of within-field segregation. We explore the relationship between three attributes of fields and within-field segregation: the gender composition of the field, the strength of its linkage to occupations, and whether the field is in STEM. We focus on these attributes not because we believe that they exhaust the potential correlates of within-field segregation but because they are especially prominent in the literatures on gender inequality in the labor market, school-to-work transitions, and women in STEM.

Consider the field's gender composition. Women who graduate in highly male-dominated fields of study may be positively selected for exceptional performance and persistence (Penner and Willer 2019) and hence may be more likely to pursue and be hired into gender-atypical occupations. Therefore, with all else equal, within-field segregation may be lower in heavily male-dominated fields than in gender-neutral or female-dominated fields. However, women in heavily male-dominated fields may be more likely to encounter employers or other gatekeepers with strong prior beliefs about women's lower competence or ability to thrive in male-dominated workplaces (Heilman 2001), and high-achieving women in male-dominated fields may be unduly penalized for lacking perceived warmth (Phelan et al. 2008; Quadlin 2018). Consistent with these latter arguments, women in male-dominated fields are less likely to enter occupations that are substantively related to their degree than male graduates from those fields (Shauman 2009). Thus, within-field occupational segregation may be especially high in heavily male-dominated fields.

What about male graduates in heavily female-dominated fields? In theory, such men may assess their own competence at female-typed tasks as lower than women with the same degree, or they may encounter employers or other gatekeepers who do. However, assumptions about women's competence at female-typed tasks are overlaid with assessments of general competence (i.e., not task-specific) that tend to favor men (Correll and Ridgeway 2006), which could offset any penalty for men in female-dominated fields. Indeed, men in female-dominated occupations (e.g., nursing, teaching, librarianship, and social work) perceive high levels of workplace support (Taylor 2010) and are more likely to encounter structural conditions that enhance their careers rather than derail them (Williams 1992, 2013). These findings imply an asymmetry in the association between a field's gender composition and its within-field occupational segregation: heavily male-dominated fields will have more within-field occupational segregation than heavily female-dominated fields.

The second attribute we examine is cross-field variations in their linkage to occupations (DiPrete et al. 2017). In fields with well-defined pathways to particular occupations, male and female workers may have very similar occupational distributions because they have very similar occupational aspirations (Shauman 2009) and because employers rely heavily on the information encapsulated in the credential itself. Thus, within-field occupational segregation may be lower in high-linkage fields relative to low-linkage fields. Conversely, fields strongly linked to occupations may also have multiple institutionalized steps for graduates in the transition from school to work (e.g., from school to internship and from internship to paid position), each of which entails a selection process in which discrimination or social exclusion can affect the gender composition of the pipeline (Tomaskovic-Devey and Skaggs 2002). This argument suggests a positive association between linkage strength and within-field segregation.

We also explore a third field-level attribute: whether the field is in STEM. Many STEM fields are heavily male-dominated, and many are comparatively strongly linked to particular occupations. As a result, a bivariate association between STEM status and within-field segregation may pick up these other field-level attributes, or vice versa. That said, the technical content of STEM fields may contribute to within-field segregation, above and beyond STEM fields' gender composition or linkage strength. Beliefs about men's greater capacity for math and scientific reasoning may affect employers' assessments of STEM field graduates applying for STEM-related jobs (see, e.g., Moss-Racusin et al. 2012) or graduates' assessments of the types of jobs for which they are most qualified (Cech et al. 2011; Correll 2001). If these beliefs operate more strongly in STEM fields than in non-STEM fields, we should find more within-field occupational segregation in STEM fields than in non-STEM fields.

The relationship between STEM status and within-field segregation is also of interest for a more pragmatic reason. A small industry of academic research has focused on the causes and consequences of gender disparities in STEM outcomes. Similarly, the federal government devotes considerable attention and resources to reducing gender and other disparities in STEM to expand the scientifically trained workforce and increase competitiveness in science. The prevalence of the distinction between STEM and non-STEM fields in these academic and policy circles makes its association with within-field segregation worth estimating, even recognizing that it may gloss over substantial differences among the detailed fields that constitute the aggregate STEM or non-STEM categorizations.

Data, Measurement, and Methods

Data and Measures

Our analyses are based on the 2019 five-year American Community Survey (ACS) public-use microdata from IPUMS (Ruggles et al. 2020). This data set contains all households and persons from the 1% ACS samples for 2015 through 2019, inclusive, which together incorporate 5% of the population. With weighting, estimates are representative of population averaged over the five-year period (Beaghen and Weidman 2008). Our analytic sample comprises 2.37 million individuals who, at the time of the survey, were employed; had obtained at least a bachelor's degree; were aged 25–65; and had no missing data on the occupation, field, education level, sex, age, race and ethnicity, and region variables.1 They are employed in 426 different occupations in the 2010 ACS scheme.

Most of our analyses use the most detailed categorization of fields available in the ACS, which includes 173 fields. For the 11% of the sample that reported more than one major, we use the information on the first major reported. We also estimate results using a 37-category scheme created by the Census Bureau and a 6-category scheme including business and law, STEM, health, social science and humanities, education, and interdisciplinary or other fields (Kim et al. 2015).

The individual-level models in the first analysis include various demographic adjustment variables likely to be associated with field and occupation, potentially in interaction with gender. Gender is coded as a binary variable. The level of education, conditional on being in the sample universe of college graduates, is indicated with a binary variable flagging college graduates with advanced degrees. (Unfortunately, the ACS does not include information on the field of the advanced degrees.) Race and ethnicity are coded as non-Hispanic White, non-Hispanic Black, Hispanic, non-Hispanic Asian, and non-Hispanic other. Region is coded as Northeast, Midwest, South, and West. Age is included to capture both potential life cycle effects, such as women's greater chance of attrition from the labor force or from STEM occupations during childbearing years (Cech and Blair-Loy 2019), and potential cohort effects. We specify age with categories corresponding to three-year intervals (Guinea-Martin et al. 2018). In models fit to data pooled across gender, all covariates are interacted with gender.

To measure FO linkage strength, we follow recent practice in the literature by relying on the local mutual information index (e.g., DiPrete et al. 2017; Elbers 2021):

(1)

In Equation (1), po|f denotes the conditional probability of working in occupation o given that one is from field of study f, and po denotes the unconditional probability of working in occupation o. Intuitively, higher FO scores indicate that a relatively large share of the field's graduates are clustered into the most common occupational destinations and that they enter relatively few occupational destinations (see DiPrete et al. 2017, especially their figure 2 and appendix table D8).2 Unlike alternative measures, such as the share of graduates in the modal occupation (or top 3 or top 5 occupations), FO scores incorporate information from across the full range of occupations of graduates of a field.3 We standardize FO scores to facilitate interpretation.

The gender composition of the field, one of the covariates in the second analysis, is indicated by the percentage of graduates who are women. To capture nonlinearities and asymmetries in the association between gender composition and within-field segregation, we rely on a three-category measure that classifies fields into male-dominated, female-dominated, and integrated. Male-dominated fields are those where the percentage female in the field falls at or below the 25th percentile among all fields, female-dominated fields are those with percentage female at or above the 75th percentile, and integrated fields are those between the 25th and 75th percentiles. Tables A1 and A2 (tables and figures designated with an “A” are available in the online appendix) show similar results using raw percentages for the cut-point values and polynomial specifications.

We categorize fields into STEM and non-STEM using the coding scheme in VanHeuvelen and Quadlin (2021). This scheme codes all fields in agriculture, environmental and natural resources, computer sciences, engineering, engineering technologies, biology and life sciences, math and statistics, physical sciences, and nuclear/industrial radiology as STEM.

Measuring Occupational Sex Segregation

This article presents results using the dissimilarity index (D) to measure occupational segregation. (Similar results using other indices are presented in Tables A3 and A4.) D is calculated with the following equation applied to tabular data:

(2)

where mo and wo are the counts of men and women in occupation o, and M and W are the total number of men and women in the employed labor force. D can be interpreted as the percentage of men who must switch to a female-typed occupation (or conversely, the percentage of women who must switch to a male-typed occupation) for all occupations to have the same sex ratio as the labor force as a whole.

Incorporating micro-level covariates into the estimation of macro-level segregation indices is a long-standing challenge for segregation research (Duncan and Duncan 1955), especially given the dimensionality problem when tabular data on gender and occupation are disaggregated by other correlates (e.g., Blau et al. 2013). To address this problem in the context of analyses of residential segregation, Fossett and colleagues (Crowell and Fossett 2018; Fossett 2017) developed a difference-of-means method for calculating D and other segregation indices. In Eqs. (3) and (4), i indexes individuals, pi represents the proportion of men in individual i's occupation, P represents the proportion of men in the employed college-educated labor force, m indexes men, and w indexes women.

  • (1)

    Assign each individual a segregation score, yi, based on a function f(pi). For D, this is defined as follows:

(3)
  • (2)

    Recover D by averaging yi over each gender, taking the difference, and multiplying by 100. Equivalently, regress yi on a binary gender indicator (e.g., male = 1); the coefficient is D divided by 100:

(4)

We are most interested in the gender coefficient, which indicates the effect of being a man (assuming the variable is coded male = 1) on the probability of being in an occupation in which the share of men is equal to or greater than the share of men in the college-educated labor force. In a model without other covariates, this coefficient (×100) will be equivalent to D calculated using Eq. (2). In a model with covariates, it is the expected D after adjusting for group differences in the distributions of those covariates of interest—in this case, field of study and other potential demographic confounders (Fossett 2017:21). Group differences in the association between these covariates and occupational segregation can be estimated as the average marginal effects (AME) of gender from models that interact gender with the other covariates (Mize 2019).4

Regression and Decomposition

Our decomposition analysis begins with linear probability models (LPMs) that estimate the relationship between the segregation score (yi) from Eq. (2), gender, and field after adjusting for gender differences in the level of education, race and ethnicity, age, and region. Results are very similar when we use logit models and fractional logit models. However, LPMs are easier to interpret, and the coefficients can be compared across model specifications.

We apply the Blinder–Oaxaca decomposition (Blinder 1973; Oaxaca 1973; see also Kitagawa 1955). This procedure begins with LPMs estimated separately for the male and female subsamples:

(5)

where Y is the segregation score vector, X is a matrix containing the predictors and a constant, β is a vector of coefficients (including the intercept), and is the error vector. The mean outcome difference is E(Ym)E(Yw); here, E(Ym) is the expected value of the segregation score for men, and E(Yw) is the expected value of the segregation score for women. The mean outcome difference is calculated as follows:

(6)

where E(βm)=βm, E(βw)=βw, E(m)=0, and E(w)=0.

The first component on the right side of Eq. (6) is the endowment effect, which indicates the contribution of gender differences in the predictors to occupational segregation. The second component is the coefficient effect, which measures the contribution of gender differences in the regression coefficients (including the difference in intercepts) to occupational segregation. The third component, the interaction term, accounts for the joint effects of endowments and coefficients. Because field-linkage scores are an attribute of fields (and hence not identified in models that also fit field effect dummy variables), we cannot fit them in the baseline LPMs to assess the relationship between gender differences in field-linkage strength and occupational segregation. Instead, we break down the field effects into shares contributed by low-, medium-, and high-linkage fields, which we define by ordering the 173 fields by linkage score and dividing them into terciles.

The estimation of the endowment, coefficient, and interaction effects is straightforward: let βm^ and βw^ be the model estimates for βm and βw, and let sample means Xm¯ and Xw¯ be estimates for E(Xm) and E(Xw). Plugging these into Eq. (4) yields Eq. (7):

(7)

We use men as the reference group.5 Results are very similar if we instead use women as the reference group (see Table A5). Because the decomposition results for categorical variables depend on the choice of the reference category, we normalize the effects so that they are expressed as deviation contrasts from the unweighted grand mean (Jann 2008:462).

Field-Level Analysis

The second analysis consists of an ordinary least-squares (OLS) regression of field-level occupational segregation on FO linkage, female representation, and whether the field is in STEM. For this analysis, we aggregate the individual-level data by field, occupation, and gender, applying survey weights.

In this analysis, we calculate D using Eq. (2) but apply it to data stratified by each of the 173 fields. Even with a sample of 2.37 million cases, some of the FO cells are quite small. In this context, occupational destinations that are random with respect to gender can generate a positive D value, a problem known as the small-unit index bias in the residential segregation literature (Cortese et al. 1976; Fossett 2017; Winship 1977). To resolve this potential source of bias, we complete three steps for each field: (1) calculate a randomized D under 50 different iterations, each randomly assigning gender to workers in the destination occupations (but fixing the total number of workers at the observed value); (2) average the 50 values of the randomized D; and (3) subtract this average from the observed D to obtain a bias-adjusted D. We log the bias-adjusted D so that it approximates a normal distribution.

Results

In the ACS analytic sample, 73.5% of men and 34.4% of women with college degrees work in male-dominated occupations (Table 1). The difference in these percentages, 39.1, is the observed level of occupational segregation as measured by D (see Eq. (4)). As expected, a smaller share of women than men earn their baccalaureate degrees in STEM fields, and a larger share of women than men earn their degrees in health, social sciences, humanities, and education fields. For ease of presentation, Table 1 presents six fields, although we use the 173-category scheme in our main analysis. Approximately half of the sample majored in a weakly linked field (the lowest tercile FO linkage score across all fields of study), whereas roughly 30% and 20% majored in moderately and strongly linked fields (middle and top terciles, respectively). Although a greater share of women majored in strongly linked fields, they are distributed in different fields of study than men. For example, whereas 44% of women in strongly linked fields majored in “medical and health sciences and services” and “education administration and teaching,” half of the men in strongly linked fields majored in a STEM field (not shown). Only 7% of women in strongly linked fields are in STEM. The distributions of men and women are largely balanced across racial and ethnic, educational attainment, age, and region categories, with two possible exceptions: (1) a higher share of female than male college graduates are non-Hispanic Black Americans, and a higher share of male than female college graduates are non-Hispanic Asian Americans; and (2) the age distribution among college-educated women skews younger than the age distribution among college-educated men. Chi-square tests show that gender differences in the distribution of workers in different fields of study and categories of the adjustment variables are statistically significant, which is to be expected given a sample of nearly 2.4 million cases.

Table 1

Descriptive statistics for U.S. college graduates, by gender (percentages)

MenWomenPooled
Y Score (individual D   
 0 (in female-dominated occupation) 26.5 65.6 46.3 
 1 (in male-dominated occupation) 73.5 34.4 53.7 
Field    
 Business and law 24.3 19.5 21.9 
 STEM 33.3 13.7 23.3 
 Health 3.0 12.7 7.9 
 Social sciences and humanities 27.7 33.9 30.8 
 Education 4.7 13.9 9.3 
 Interdisciplinary and other fields 7.0 6.4 6.7 
Field Linkage Terciles    
 Weak 48.8 50.0 49.4 
 Moderate 35.5 25.6 30.5 
 Strong 15.6 24.3 20.1 
Race and Ethnicity    
 Non-Hispanic White 71.4 69.5 70.4 
 Non-Hispanic Black 7.0 9.9 8.5 
 Hispanic 8.4 8.8 8.6 
 Non-Hispanic Asian 11.2 9.7 10.4 
 Other 2.1 2.2 2.1 
Educational Attainment    
 College degree 63.4 61.1 62.3 
 More advanced degree 36.6 38.9 37.7 
Age Group    
 25–27 7.7 9.0 8.4 
 28–30 8.6 9.3 9.0 
 31–33 8.6 8.9 8.8 
 34–36 8.6 8.8 8.7 
 37–39 8.0 8.3 8.1 
 40–42 7.8 7.9 7.8 
 43–45 7.7 7.8 7.7 
 46–48 7.8 7.8 7.8 
 49–51 7.3 7.2 7.3 
 52–54 7.1 6.7 6.9 
 55–57 6.6 6.2 6.4 
 58–60 6.1 5.6 5.8 
 61–63 5.2 4.4 4.8 
 64–65 2.8 2.1 2.5 
Region    
 Northeast 20.5 21.1 20.8 
 Midwest 20.5 20.8 20.7 
 South 34.4 35.2 34.8 
 West 24.6 22.9 23.7 
MenWomenPooled
Y Score (individual D   
 0 (in female-dominated occupation) 26.5 65.6 46.3 
 1 (in male-dominated occupation) 73.5 34.4 53.7 
Field    
 Business and law 24.3 19.5 21.9 
 STEM 33.3 13.7 23.3 
 Health 3.0 12.7 7.9 
 Social sciences and humanities 27.7 33.9 30.8 
 Education 4.7 13.9 9.3 
 Interdisciplinary and other fields 7.0 6.4 6.7 
Field Linkage Terciles    
 Weak 48.8 50.0 49.4 
 Moderate 35.5 25.6 30.5 
 Strong 15.6 24.3 20.1 
Race and Ethnicity    
 Non-Hispanic White 71.4 69.5 70.4 
 Non-Hispanic Black 7.0 9.9 8.5 
 Hispanic 8.4 8.8 8.6 
 Non-Hispanic Asian 11.2 9.7 10.4 
 Other 2.1 2.2 2.1 
Educational Attainment    
 College degree 63.4 61.1 62.3 
 More advanced degree 36.6 38.9 37.7 
Age Group    
 25–27 7.7 9.0 8.4 
 28–30 8.6 9.3 9.0 
 31–33 8.6 8.9 8.8 
 34–36 8.6 8.8 8.7 
 37–39 8.0 8.3 8.1 
 40–42 7.8 7.9 7.8 
 43–45 7.7 7.8 7.7 
 46–48 7.8 7.8 7.8 
 49–51 7.3 7.2 7.3 
 52–54 7.1 6.7 6.9 
 55–57 6.6 6.2 6.4 
 58–60 6.1 5.6 5.8 
 61–63 5.2 4.4 4.8 
 64–65 2.8 2.1 2.5 
Region    
 Northeast 20.5 21.1 20.8 
 Midwest 20.5 20.8 20.7 
 South 34.4 35.2 34.8 
 West 24.6 22.9 23.7 

Notes: Chi-square tests show statistically significant gender differences in the distribution for all variables. Y  = 1 is the percentage of men/women in a male-dominated occupation (an occupation that exceeds the overall male representation in the college-educated labor force). Field linkage terciles are based on field-level distributions. N = 2,372,463. Survey weights are applied.

Data Source: 2019 five-year American Community Survey Integrated Public Use Microdata Series (IPUMS).

Decomposing the Dissimilarity Index

Table 2 presents the results of LPMs regressing individual segregation scores on gender, fields, and the adjustment variables listed in Table 1. In all but the simplest bivariate model, all covariates are interacted with gender, and the expected D is calculated using the AME of gender.

Table 2

Expected occupational segregation (D) estimated from linear probability models with and without covariates

Model 0Model 1Model 2Model 3
Expected D 0.391*** 0.252*** 0.386*** 0.250*** 
 (0.389, 0.392) (0.251, 0.254) (0.385, 0.388) (0.249, 0.252) 
Field (173) No Yes No Yes 
Controls No No Yes Yes 
R2 .15 .25 .16 .26 
AIC 3,036,168 2,739,515 3,013,148 2,723,044 
BIC 3,036,193 2,743,902 3,013,706 2,727,963 
Model 0Model 1Model 2Model 3
Expected D 0.391*** 0.252*** 0.386*** 0.250*** 
 (0.389, 0.392) (0.251, 0.254) (0.385, 0.388) (0.249, 0.252) 
Field (173) No Yes No Yes 
Controls No No Yes Yes 
R2 .15 .25 .16 .26 
AIC 3,036,168 2,739,515 3,013,148 2,723,044 
BIC 3,036,193 2,743,902 3,013,706 2,727,963 

Notes: 95% confidence intervals are shown in parentheses. Controls include race and ethnicity (non-Hispanic White, non-Hispanic Black, Hispanic, non-Hispanic Asian, and non-Hispanic other), educational attainment, age (three-year intervals), and region (Northeast, Midwest, South, and West). N = 2,372,463. AIC = Akaike information criterion. BIC = Bayesian information criterion.

Data Source: 2019 five-year American Community Survey Integrated Public Use Microdata Series (IPUMS).

***p < .001

Model 0 is a bivariate linear probability model regressing the segregation score on the male dummy variable. The coefficient multiplied by 100 is 39.1, which is the unadjusted value of D by design. Model 1 adds field (173 categories) and yields an expected occupational segregation index of 25.2, approximately 64.7% (25.2 / 39.1  =  0.647) of the unadjusted D. Model 2, which fits the demographic covariates but not field, yields an expected value of D (38.6) that is very similar to the unadjusted D. Model 3 includes field and all demographic covariates and yields an expected D of 25.0—nearly identical to the expected D (25.2) from Model 1. In other words, approximately one third of observed occupational segregation is associated with men's and women's segregation into different fields in college, and approximately two thirds is associated with within-field segregation. After we adjust for the field of study, the residual associations between race and ethnicity, region, age, and level of education and occupational segregation are quite modest, likely because gender differences in these covariates are fairly modest (see Table 1).

Table 3 presents results from decompositions using estimates from Models 1 (field only) and 3 (field and all controls). The first panel of Table 3 shows that the two models yield very similar shares of the group differences due to endowment effects and coefficient effects including the constant term: shares due to endowment effects are 35.5% (.355  =  13.87 / 39.05) and 35.6% (.356  =  13.90 / 39.05) in Models 1 and 3, respectively; shares due to coefficient effects are 64.8% (.648  =  25.30 / 39.05) and 63.8% (.638  =  24.91 / 39.05) in Models 1 and 3, respectively. Less than 1% of macro-level occupational segregation is attributable to joint endowment and coefficient effects. We obtained 95% confidence intervals by conducting 80 bootstrap replications per model.6 The observed values of D in the top panel are identical across the two models by design, as observed rather than estimated quantities.

Table 3

Decomposition analysis of occupational segregation as measured by the dissimilarity index (D)

Model 1Model 3
Overall   
 Observed D 39.05 39.05 
  (38.91, 39.20) (38.91, 39.20) 
 Endowments 13.87 13.90 
  (13.76, 13.98) (13.78, 14.02) 
 Coefficients 25.30 24.91 
  (25.11, 25.49) (24.72, 25.10) 
 Interaction −0.12 0.24 
 (−0.31, 0.08) (0.04, 0.44) 
Endowments   
 Field of study 13.87 (13.76, 13.98) 13.66 (13.54, 13.78) 
  Weakly linked fields 1.89 1.84 
   (1.83, 1.94) (1.79, 1.89) 
  Moderately linked fields 4.34 4.22 
   (4.25, 4.44) (4.13, 4.32) 
  Strongly linked fields 7.64 7.59 
 (7.54, 7.74) (7.50, 7.69) 
 Race and ethnicity  0.22
(0.20, 0.23) 
 Educational level  −0.08
(−0.09, −0.07) 
 Age  0.10
(0.10, 0.11) 
 Region  0.00 
   (−0.01, 0.00) 
Coefficients   
 Field of study 0.88 0.60 
 (0.40, 1.37) (0.11, 1.07) 
  Weakly linked fields 2.36 2.08 
    (2.08, 2.64) (1.80, 2.36) 
  Moderately linked fields 0.20 0.19 
    (0.05, 0.34) (0.05, 0.33) 
  Strongly linked fields −1.68 −1.68 
 (−1.84, −1.52) (−1.84, −1.52) 
 Race and ethnicity  1.50 
   (1.32, 1.68) 
 Educational level  −2.15 
   (−2.26, −2.04) 
 Age  −0.29 
   (−0.34, −0.25) 
 Region  0.03 
   (0.00, 0.06) 
 Constant 24.42 25.24 
 (23.91, 24.93) (24.71, 25.78) 
Model 1Model 3
Overall   
 Observed D 39.05 39.05 
  (38.91, 39.20) (38.91, 39.20) 
 Endowments 13.87 13.90 
  (13.76, 13.98) (13.78, 14.02) 
 Coefficients 25.30 24.91 
  (25.11, 25.49) (24.72, 25.10) 
 Interaction −0.12 0.24 
 (−0.31, 0.08) (0.04, 0.44) 
Endowments   
 Field of study 13.87 (13.76, 13.98) 13.66 (13.54, 13.78) 
  Weakly linked fields 1.89 1.84 
   (1.83, 1.94) (1.79, 1.89) 
  Moderately linked fields 4.34 4.22 
   (4.25, 4.44) (4.13, 4.32) 
  Strongly linked fields 7.64 7.59 
 (7.54, 7.74) (7.50, 7.69) 
 Race and ethnicity  0.22
(0.20, 0.23) 
 Educational level  −0.08
(−0.09, −0.07) 
 Age  0.10
(0.10, 0.11) 
 Region  0.00 
   (−0.01, 0.00) 
Coefficients   
 Field of study 0.88 0.60 
 (0.40, 1.37) (0.11, 1.07) 
  Weakly linked fields 2.36 2.08 
    (2.08, 2.64) (1.80, 2.36) 
  Moderately linked fields 0.20 0.19 
    (0.05, 0.34) (0.05, 0.33) 
  Strongly linked fields −1.68 −1.68 
 (−1.84, −1.52) (−1.84, −1.52) 
 Race and ethnicity  1.50 
   (1.32, 1.68) 
 Educational level  −2.15 
   (−2.26, −2.04) 
 Age  −0.29 
   (−0.34, −0.25) 
 Region  0.03 
   (0.00, 0.06) 
 Constant 24.42 25.24 
 (23.91, 24.93) (24.71, 25.78) 

Notes: The reference group is men. Results are reported in the units of D. Bootstrapped 95% confidence intervals are shown in parentheses, with 80 replications for each model applying IPUMS ACS replication weights. The decomposition of the (very small) interaction term is omitted.

Data Source: 2019 five-year American Community Survey Integrated Public Use Microdata Series (IPUMS).

The second panel of Table 3 breaks down the endowments component, again with very similar results across Models 1 and 3. The endowment effects are driven mainly by women's and men's uneven distributions across fields in college rather than by gender differences in distributions of the adjustment variables (see Model 3). In Model 1, the gender difference in mean individual segregation scores in the fields in the top tercile ordered by linkage score contributes approximately 55.1% (.551  =  7.64 / 13.87) of the field endowment effects, whereas fields with comparatively low linkage strength contribute approximately 13.6% (.136  =  1.89 / 13.87). Model 3 shows nearly identical results.

The third panel of Table 3 identifies the major contributors to the 64% of D that is due to coefficient effects. In Models 1 and 3, the constant term dominates, contributing between 97% and 112% of the coefficient effect. (A value over 100% is possible if some coefficients suppress segregation while others exacerbate it.) Gender-differentiated returns to postbaccalaureate degrees tend to suppress occupational segregation (i.e., the coefficient effect is −9%), implying that postbaccalaureate degrees increase women's probability of entering a male-dominated occupation more than they increase men's probability of entering a male-dominated occupation. Gender-differentiated occupational returns to racial and ethnic group membership contribute roughly 6% to the total coefficient effect, and region and age contribute very little net of the other covariates.

On average, field contributes but a very small share of the total coefficient effect: approximately 3.5% (.035  =  0.88 / 25.30) of the coefficient effect in Model 1 and approximately 2.4% (.024  =  0.60 / 24.91) in Model 3. However, this cross-field average masks some variability across fields with differing degrees of linkage to occupations. In Model 3, for example, the sum of gender gaps in field coefficients across the fields with weak linkage is 2.1, or 8.4% (.084  =  2.08 / 24.91) of the total coefficient effects; that is, gender differences in the returns to weakly linked fields tend to exacerbate overall segregation. By contrast, the sum for strongly linked fields is −1.68, or −6.7% (= −1.68 / 24.91); that is, gender differences in the returns to these fields suppress occupational segregation. These results, as well as the further decomposition of field endowment effects by field linkage strength, suggest that integrating strongly linked fields will help integrate occupations, assuming that at least some of this association is causal.

These results also suggest that the small field coefficient effect (see Table 3) masks cross-field heterogeneity in differential segregation returns to fields associated with linkage strength. Figure 1 illustrates some of this heterogeneity in field-of-study coefficients by juxtaposing field coefficients with estimated coefficients from Model 3 fit to male and female subsamples separately (right panel). Rather than present these results for all 173 fields, Figure 1 focuses on fields with a gender gap statistically distinguishable from 0 (95% confidence intervals are plotted) and greater than 0.1, meaning a 10% gender difference in the probability of working in a male-dominated occupation. Because we use effect coding, field coefficients are interpreted as deviations from the (unweighted) gender-specific grand means. A field can suppress occupational segregation through a coefficient effect even if its female graduates are less likely than its male graduates to enter male-dominated occupations, as is the case with the engineering fields in Figure 1.

Fig. 1

Selected field-of-study coefficients (right panel) and gender gaps in coefficients (left panel) on the probability of entering a male-dominated occupation (i.e., in which the proportion of male workers is at or above the share of men in the college-educated employed labor force). Coefficients are based on effect coding and are interpreted as deviations from the gender-specific grand mean. Row labels are formatted according to field linkage. Row labels in bold type indicate strongly linked fields, bold and italicized row labels indicate fields in the middle tercile of linkage, and row labels in plain type indicate weakly linked fields. Data Source: 2019 five-year American Community Survey Integrated Public Use Microdata Series (IPUMS).

Fig. 1

Selected field-of-study coefficients (right panel) and gender gaps in coefficients (left panel) on the probability of entering a male-dominated occupation (i.e., in which the proportion of male workers is at or above the share of men in the college-educated employed labor force). Coefficients are based on effect coding and are interpreted as deviations from the gender-specific grand mean. Row labels are formatted according to field linkage. Row labels in bold type indicate strongly linked fields, bold and italicized row labels indicate fields in the middle tercile of linkage, and row labels in plain type indicate weakly linked fields. Data Source: 2019 five-year American Community Survey Integrated Public Use Microdata Series (IPUMS).

Close modal

Figure 1 offers additional evidence that the overall coefficient effect of fields is small (see Table 3) because of offsetting differences in individual fields' segregation returns. For example, female graduates in actuarial science are roughly 29 percentage points more likely to work in an occupation that is either at parity or male-dominated than female graduates in any field, whereas male graduates in actuarial science are approximately 11 percentage points more likely to work in a male-dominated occupation than male graduates in any field. Subtracting the women's coefficient from the men's coefficient yields a negative gap of 18 percentage points (11 – 29 = −18), meaning that actuarial science contributes negatively to (i.e., suppresses) overall levels of occupational segregation. By contrast, graduates in theology exacerbate overall occupational segregation by roughly the same amount. Female theology graduates are less likely to work in a male-dominated occupation than the average female graduate (by 11.3 percentage points), and male theology graduates are more likely to work in a male-dominated occupation than the average male graduate (6.4 percentage points), yielding a positive gender gap of approximately 18 percentage points (6.4 – −11.3 = 17.7).

Although we can adjust for level of postsecondary degree (baccalaureate or graduate), we cannot assess whether the relationship between field segregation and occupational segregation is stronger among workers with graduate degrees, given the constraints of the ACS data. In a supplemental analysis (see Table A6), we repeated the decomposition analysis based on the subset of students who received only baccalaureate degrees. We found that the overall level of D is higher in the trimmed sample (see the Discussion section), but the share of D associated with field segregation is nearly identical (at about one third) in this trimmed sample as in the full sample. A greater share of workers with postbaccalaureate degrees earned baccalaureate degrees in female-dominated or gender-neutral fields (e.g., psychology, biology, business management and administration, general education, English, political science) than in male-dominated fields (e.g., electrical engineering, economics, computer science).

How do these general patterns hold up when we use aggregated categorizations of fields, which may be the only field information available to researchers using smaller data sets? As shown in Table A7, the share of occupational segregation associated with field segregation decreases as field categories become broader. Specifically, approximately 32% of the observed occupational segregation is associated with segregation across 37 field categories, and 22% is associated with segregation across 6 broad field categories; approximately 70% and 75%, respectively, are associated with within-field segregation. Per degrees of freedom expended, the aggregate field mappings produce results that reasonably mirror the general patterns described earlier. In absolute terms, however, they underestimate the contribution of field segregation to occupational segregation.

Within-Field Segregation and Variation Across Fields

Table 4 presents descriptive statistics for our second analysis, which shifts attention to the correlates of within-field occupational segregation. Within-field segregation, as indicated by the bias-adjusted D (see the Methods section), ranges from 0.14 in nursing to 0.70 in military technologies. High-linkage fields include medical assisting services (2.80), pharmacy and pharmaceutical sciences and administration (2.35), and actuarial science (2.29). Low-linkage fields include liberal arts (0.13), multidisciplinary or general science (0.17), area ethnic and civilization studies (0.20), and intercultural and international studies (0.21). The proportion of female students in a field ranges from .06 (naval architecture and marine engineering) to .97 (early childhood education), where the average is .46. To allow for hypothesized nonlinearities in the association between field gender representation and within-field segregation, our main analyses use percentile-based categories of gender representation.7 Approximately 39% of fields are in STEM. In the regression that follows, we log D so that its distribution is closer to normal, and we standardize linkage scores (with a mean of 0).

Table 4

Descriptive statistics for fields-of-study attributes

Min.Max.MedianMeanSD
Within-Field Segregation, D 0.14 0.70 0.32 0.33 0.09 
Log Transformed D 1.99 0.35 1.14 1.14 0.26 
FO Score 0.13 2.80 0.74 0.84 0.50 
FO Score, Standardized 1.43 3.91 0.20 0.00 1.00 
Proportion Female .06 .97 .46 .47 .22 
STEM .00 1.00  .39  
Min.Max.MedianMeanSD
Within-Field Segregation, D 0.14 0.70 0.32 0.33 0.09 
Log Transformed D 1.99 0.35 1.14 1.14 0.26 
FO Score 0.13 2.80 0.74 0.84 0.50 
FO Score, Standardized 1.43 3.91 0.20 0.00 1.00 
Proportion Female .06 .97 .46 .47 .22 
STEM .00 1.00  .39  

Notes: D = dissimilarity index. FO score = field–occupation linkage score. N  = 173.

Data Source: 2019 five-year American Community Survey Integrated Public Use Microdata Series (IPUMS).

Table 5 presents estimates from three bivariate OLS models regressing within-field occupational segregation on female representation (Model 4), FO linkage (Model 5), or STEM (Model 6); a fourth model includes all three covariates (Model 7). Model 4 shows that within-field segregation is larger in heavily male-dominated fields—that is, fields in which women's representation is below the 25th percentile or, in absolute percentage terms, in which less than 28% of graduates are women. Relative to fields between the 25th and 75th percentile in gender composition, heavily male-dominated fields have approximately 15% more occupational segregation (exp[0.14] – 1  =  0.15). In the metric of D, the average within-field segregation score is roughly 31 for fields with balanced gender representation, compared with approximately 37 in heavily male-dominated fields. By contrast, fields in which women are heavily overrepresented—those above the 75th percentile of female representation (corresponding to 62% female)—do not have significantly more within-field segregation than gender-balanced fields.

Table 5

OLS regressions of within-field segregation (log D) on field female representation, FO linkage, and STEM

Model 4Model 5Model 6Model 7
Field Female Representation 
 Between 1st and 3rd quartiles (ref.) 
 <1st quartile 0.14**   0.16* 
  (0.05)   (0.06) 
 >3rd quartile 0.04   0.02 
 (0.04)   (0.05) 
FO Score, Standardized 0.03  0.02 
  (0.03)  (0.03) 
STEM   0.02 −0.05 
   (0.04) (0.05) 
Constant −1.18** −1.14** −1.15** −1.16** 
 (0.02) (0.02) (0.02) (0.03) 
Number of Observations 173 173 173 173 
R2 .05 .01 .00 .06 
AIC 18.89 23.71 25.75 21.27 
BIC 28.35 30.02 32.06 37.03 
Model 4Model 5Model 6Model 7
Field Female Representation 
 Between 1st and 3rd quartiles (ref.) 
 <1st quartile 0.14**   0.16* 
  (0.05)   (0.06) 
 >3rd quartile 0.04   0.02 
 (0.04)   (0.05) 
FO Score, Standardized 0.03  0.02 
  (0.03)  (0.03) 
STEM   0.02 −0.05 
   (0.04) (0.05) 
Constant −1.18** −1.14** −1.15** −1.16** 
 (0.02) (0.02) (0.02) (0.03) 
Number of Observations 173 173 173 173 
R2 .05 .01 .00 .06 
AIC 18.89 23.71 25.75 21.27 
BIC 28.35 30.02 32.06 37.03 

Notes: Standard errors are shown in parentheses. N = 173. The dependent variable is log dissimilarity index (D), after adjusting for small unit bias. FO score = field–occupation linkage score. AIC = Akaike information criterion. BIC = Bayesian information criterion.

Data Source:2019 five-year American Community Survey Integrated Public Use Microdata Series (IPUMS).

*p < .05; **p < .01

The field-level analyses also show a small, positive association between within-field occupational segregation and linkage score (see Model 5, Table 5). Specifically, a 1-standard-deviation increase in the linkage score is associated with a 4% increase in segregation. On average, STEM fields have 2% higher predicted within-field segregation (Model 6). However, neither the linkage nor STEM coefficients are statistically significant, and their confidence intervals include 0. The model with all three predictors (Model 7), like the individual predictor models, suggests that within-field segregation is higher in heavily male-dominated occupations but does not vary systematically with linkage score or STEM status.

Discussion

In the United States, gender segregation across baccalaureate fields is associated with roughly one third of the total occupational segregation among college graduates, and the remainder is driven by gender-specific occupational destinations of graduates in the same fields. If at least some of the association between field segregation and occupational segregation is causal, as implied by theories assuming that degrees in a given field signal a particular type of skill or a necessary credential to enter a closed occupation, this top-line result may have important implications for the trajectory of gender inequality in labor markets.

In this regard, our results have a half-empty/half-full character. On the one hand, they imply that desegregation of fields of study in higher education, which is the tacit goal of many institutionalized programs to increase women's representation in STEM, would help to integrate labor markets, even if only through a slow process of cohort replacement (see Weeden 2019). Further, reducing educational segregation would also reduce any systematic gender differences between male- and female-typed occupations in their pay, prestige, nonmonetary compensation, or other valued social goods (England et al. 2020; Petersen and Morgan 1995)—again assuming (partial) causality between field segregation and occupational segregation.

On the other hand, we also find substantially more occupational segregation within fields than between them. Even if one assumes a causal relationship between field segregation and occupational segregation, integrating fields would not entirely eliminate gender segregation across occupations simply because field segregation is not the whole story. The high within-field segregation also has implications for gender inequalities in pay, insofar as the gender-differentiated occupational destinations among graduates of the same field mean that men disproportionately enter higher-paying occupations and women disproportionately enter lower-paying occupations (see Shauman 2006). An analysis of the relationships among field segregation, occupational segregation, FO linkage, and the gender gap in pay is beyond the scope of this study, but it is a fruitful direction for further research.

Our decomposition analysis reveals that field segregation is primarily related to occupational segregation through endowment effects: the uneven distribution of men and women across fields. These endowment effects are greater in fields that have a tight linkage to particular occupational destinations. Conversely, the average effect of gender differences in the segregation returns to different fields on occupational segregation is relatively modest. However, we also find that the gender-differentiated occupational returns to fields are greater in fields that have a weak linkage to occupational destinations.

It bears reiterating that the ACS includes field information only for baccalaureate degrees, not for postbaccalaureate degrees. Excluding postbaccalaureate workers from the analysis does not appreciably change estimates of the share of occupational segregation associated with gender differences in fields of study, although the absolute level of occupational segregation is higher among undergraduate-only workers. Nevertheless, conclusions about weak-linkage and strong-linkage fields of study and gender segregation might be affected by the ACS's data limitations because some weak-linkage undergraduate fields (e.g., history) are likely feeders to strong-linkage postbaccalaureate degrees (e.g., law). It seems plausible that field segregation at the graduate level will have greater endowment effects on occupational segregation but even smaller coefficient effects than at the undergraduate level. Although we cannot evaluate this claim with the ACS data, it is a promising direction for future research.

Our second analysis shifted the focus to the field level to unpack the correlates of within-field segregation. We found that women and men with degrees in heavily male-dominated fields tend to enter more gender-differentiated occupational destinations than those with degrees in gender-neutral or female-dominated fields. However, we do not find that graduates of fields with strong linkage to particular occupational destinations or graduates of STEM fields are more (or less) segregated in the labor market than graduates of other fields. Our models, which are necessarily parsimonious given the limited number of fields in the analysis (n = 173), may gloss over interactions between STEM status, linkage, and gender composition. For example, vocationally oriented STEM fields might have less within-field occupational segregation than low-linkage STEM fields because a degree in a high-linkage field sends a stronger signal of technical skills, and employers accordingly rely less on gender and other non-task-relevant attributes. Still, it is notable that none of the field-level attributes, expressed as main effects, capture much of the variation in within-field occupational segregation. Future research could usefully attend not only to the possibility of more complex relationships among the three observed attributes and occupational segregation but also to identifying unobserved sources of variation across fields.

This article also makes two methodological contributions. First, it outlines a simple, intuitive way to adjust segregation indices for the upward bias introduced by random noise in small units—in our case, field–occupation cells. The residential segregation literature has frequently discussed this bias, but as far as we can tell, the educational and occupational segregation literatures ignore it. Second, the article applies Fossett's difference-of-means approach to estimating segregation indices. This approach can help bridge micro-level theories about the sources of gender segregation and the macro-level phenomenon of segregation and resolve the long-standing dimensionality problem in standard approaches to calculating indices from tabular data (Fossett 2017).

We see several useful extensions of this approach. Most obviously, with richer data than the ACS, scholars could assess other individual correlates of educational and occupational segregation identified in the gender (and racial) inequality literatures. One might, for example, assess how gender differences in self-assessed math ability contribute to occupational segregation by affecting the probability of graduating with a STEM degree (Correll 2001; but see Weeden et al. 2020) or the probability of entering male-dominated occupations after having earned a STEM degree. With longitudinal data, one could also examine whether field effects on aggregate levels of occupational segregation fade (or emerge) throughout the life course.

At the macro level, it would also be useful to assess whether changes in the strength and pattern of the relationship between field segregation and occupational segregation contributed to the slowdown in occupational integration observed after the 1990s (e.g., England et al. 2020; Weeden 2004), especially given the recent uptick in educational segregation among bachelor's degree recipients (England et al. 2020). Studies might also usefully explore how changes in the relationship between the two forms of segregation affected trends in labor market outcomes, such as the gender gap in wages. Also worth investigating is whether and how the strength and pattern of the covariate-adjusted relationship between field and occupational segregation differ in countries with education systems that emphasize vocational training or that are more closely coordinated with labor market institutions. The general theme of these extensions, and indeed of this article, is that the empirical literatures on gender segregation in higher education, gender segregation in labor markets, and women in STEM have for too long developed in isolation from each other.

Acknowledgments

We thank Peter Rich, Filiz Garip, Vida Maralani, Ian Lundberg, Ananda Martin-Caughey, and Wonjeong Jeong for reading and providing help on earlier versions of this article. We also gratefully acknowledge comments from participants of the Inequality Discussion Group at the Center for the Study of Inequality, Cornell University; and the Gender Segregation in Education and Occupations session at the 2022 annual meeting of the Population Association of America. A replication package is available at https://osf.io/h3t4w/?view_only=44ec8ff8e38840ec8014165d53c7e91b. All errors are our own.

Notes

1

In the ACS, 1.7% of the employed college-educated workers aged 25–65 are missing region information. Because of the ACS’s data imputation, no cases have missing information on race or ethnicity, gender, occupation, or field (approximately 3% imputed).

2

Consider two hypothetical fields. In Field 1, 60% of graduates enter Occupation 1, and 40% enter Occupation 2; in Field 2, 60% of graduates enter Occupation 1, and the remaining 40% are distributed across 10 occupations. The FO linkage score of Field 1 would be higher than that of Field 2. By contrast, a measure based on the share of field graduates who work in the modal occupation would be the same for the two fields.

3

In the ACS sample, only three fields have at least 50% of graduates in one occupation: elementary education (51% are elementary and middle school teachers); nursing (71% are registered nurses); and pharmacy, pharmaceutical sciences, and administration (53% are pharmacists).

4

AME and marginal effects at the means yield similar results.

5

Operationally, we calculate endowment effects as the estimated reduction in D after we match the women’s sample means across covariates to men’s sample means. We calculate coefficient effects as the estimated reduction in D after we match coefficients from the women’s equation to those from the men’s equation. We subtract women’s sample means or coefficients from men’s sample means or coefficients.

6

The ACS microdata include 80 replication weights. Using these replication weights in bootstrapping preserves the complex survey sampling structure.

7

Other specifications produce similar patterns: (1) alternative cut points for the percentiles, (2) a linear and quadratic term, and (3) categorical specification with cut points at 25% and 75% female. See Tables A1 and A2. For consistency with the decomposition analysis, we also test models using tercile coding for both percentage female and linkage score (see Table A8).

References

Alper, J. (
1993
).
The pipeline is leaking women all the way along
.
Science
,
260
,
409
412
.
Beaghen, M., & Weidman, L. (
2008
).
Statistical issues of interpretation of the American Community Survey's one-, three-, and five-year period estimates
(American Community Survey research memorandum series). U.S. Census Bureau. Retrieved from https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=0ad4bfb3d180bfbad0373176668fcac7ba4fdbf7
Becker, G. S. (
1971
).
The economics of discrimination
(2nd ed.).
Chicago, IL
:
University of Chicago Press
.
Becker, G. S. (
1993
).
Human capital: A theoretical and empirical analysis, with special reference to education
(3rd ed.).
Chicago, IL
:
University of Chicago Press
.
Bianchi, S. M., Milkie, M. A., Sayer, L. C., & Robinson, J. P. (
2000
).
Is anyone doing the housework? Trends in the gender division of household labor
.
Social Forces
,
79
,
191
228
.
Blau, F. D., Brummund, P., & Liu, A. Y. (
2013
).
Trends in occupational segregation by gender 1970–2009: Adjusting for the impact of changes in the occupational coding system
.
Demography
,
50
,
471
492
.
Blinder, A. S. (
1973
).
Wage discrimination: Reduced form and structural estimates
.
Journal of Human Resources
,
8
,
436
455
.
Bol, T., & Weeden, K. A. (
2015
).
Occupational closure and wage inequality in Germany and the United Kingdom
.
European Sociological Review
,
31
,
354
369
.
Cech, E., Rubineau, B., Silbey, S., & Seron, C. (
2011
).
Professional role confidence and gendered persistence in engineering
.
American Sociological Review
,
76
,
641
666
.
Cech, E. A., & Blair-Loy, M. (
2019
).
The changing career trajectories of new parents in STEM
.
Proceedings of the National Academy of Sciences
,
116
,
4182
4187
.
Ceci, S. J., & Williams, W. M. (
2011
).
Understanding current causes of women's underrepresentation in science
.
Proceedings of the National Academy of Sciences
,
108
,
3157
3162
.
Ceci, S. J., Williams, W. M., & Barnett, S. M. (
2009
).
Women's underrepresentation in science: Sociocultural and biological considerations
.
Psychological Bulletin
,
135
,
218
261
.
Cha, Y. (
2013
).
Overwork and the persistence of gender segregation in occupations
.
Gender & Society
,
27
,
158
184
.
Charles, M., & Bradley, K. (
2009
).
Indulging our gendered selves? Sex segregation by field of study in 44 countries
.
American Journal of Sociology
,
114
,
924
976
.
Charles, M., & Grusky, D. B. (
2004
).
Occupational ghettos: The worldwide segregation of women and men.
Stanford, CA
:
Stanford University Press
.
Charles, M., & Grusky, D. B. (
2011
).
Egalitarianism and gender inequality
. In Grusky, D. B. & Szelényi, S. (Eds.),
The inequality reader: Contemporary and foundational readings in race, class and gender
(2nd ed., pp.
327
342
).
Boulder, CO
:
Westview Press
.
Cheryan, S., Master, A., & Meltzoff, A. N. (
2015
).
Cultural stereotypes as gatekeepers: Increasing girls' interest in computer science and engineering by diversifying stereotypes
.
Frontiers in Psychology, 6
. https://doi.org/10.3389/fpsyg.2015.00049
Correll, S. J. (
2001
).
Gender and the career choice process: The role of biased self-assessments
.
American Journal of Sociology
,
106
,
1691
1730
.
Correll, S. J. (
2004
).
Constraints into preferences: Gender, status, and emerging career aspirations
.
American Sociological Review
,
69
,
93
113
.
Correll, S. J., & Benard, S. (
2006
).
Biased estimators? Comparing status and statistical theories of gender discrimination
. In Thye, S. R. & Lawler, E. J. (Eds.),
Advances in group processes
(Vol.
23
, pp.
89
116
).
Bingley, UK
:
Emerald Group Publishing
.
Correll, S. J., & Ridgeway, C. L. (
2006
).
Expectation states theory
. In DeLamater, J. (Ed.),
Handbooks of sociology and social research series. Handbook of social psychology
(pp.
29
51
).
New York, NY
:
Springer Science+Business Media
.
Correll, S. J., Weisshaar, K. R., Wynn, A. T., & Wehner, J. D. (
2020
).
Inside the black box of organizational life: The gendered language of performance assessment
.
American Sociological Review
,
85
,
1022
1050
.
Cortese, C. F., Falk, R. F., & Cohen, J. K. (
1976
).
Further considerations on the methodological analysis of segregation indices
.
American Sociological Review
,
41
,
630
637
.
Crowell, A. R., & Fossett, M. (
2018
).
White and Latino locational attainments: Assessing the role of race and resources in U.S. metropolitan residential segregation
.
Sociology of Race and Ethnicity
,
4
,
491
507
.
DiPrete, T. A., Bol, T., Eller, C. C., & van de Werfhorst, H. G. (
2017
).
School-to-work linkages in the United States, Germany, and France
.
American Journal of Sociology
,
122
,
1869
1938
.
Duncan, O. D., & Duncan, B. (
1955
).
A methodological analysis of segregation indexes
.
American Sociological Review
,
20
,
210
217
.
Elbers, B. (
2021
).
A method for studying differences in segregation across time and space
.
Sociological Methods & Research
,
52
,
5
42
.
England, P., Levine, A., & Mishel, E. (
2020
).
Progress toward gender equality in the United States has slowed or stalled
.
Proceedings of the National Academy of Sciences
,
117
,
6990
6997
.
Fossett, M. (
2017
).
New methods for measuring and analyzing segregation
.
Cham, Switzerland
:
Springer Nature
.
Guinea-Martin, D., Mora, R., & Ruiz-Castillo, J. (
2018
).
The evolution of gender segregation over the life course
.
American Sociological Review
,
83
,
983
1019
.
Hakim, C. (
2002
).
Lifestyle preferences as determinants of women's differentiated labor market careers
.
Work and Occupations
,
29
,
428
459
.
Heilman, M. E. (
2001
).
Description and prescription: How gender stereotypes prevent women's ascent up the organizational ladder
.
Journal of Social Issues
,
57
,
657
674
.
Jacobs, J. A. (
1989
).
Long-term trends in occupational segregation by sex
.
American Journal of Sociology
,
95
,
160
173
.
Jann, B. (
2008
).
The Blinder–Oaxaca decomposition for linear regression models
.
Stata Journal
,
8
,
453
479
.
Kim, C., Tamborini, C. R., & Sakamoto, A. (
2015
).
Field of study in college and lifetime earnings in the United States
.
Sociology of Education
,
88
,
320
339
.
Kitagawa, E. M. (
1955
).
Components of a difference between two rates
.
Journal of the American Statistical Association
,
50
,
1168
1194
.
Kmec, J. A., McDonald, S., & Trimble, L. B. (
2010
).
Making gender fit and “correcting” gender misfits: Sex segregated employment and the nonsearch process
.
Gender & Society
,
24
,
213
236
.
Legewie, J., & DiPrete, T. A. (
2014
).
Pathways to science and engineering bachelor's degrees for men and women
.
Sociological Science
,
1
,
41
48
.
Levanon, A., & Grusky, D. B. (
2016
).
The persistence of extreme gender segregation in the twenty-first century
.
American Journal of Sociology
,
122
,
573
619
.
Mann, A., & DiPrete, T. A. (
2013
).
Trends in gender segregation in the choice of science and engineering majors
.
Social Science Research
,
42
,
1519
1541
.
Mize, T. D. (
2019
).
Best practices for estimating, interpreting, and presenting nonlinear interaction effects
.
Sociological Science
,
6
,
81
117
.
Morgan, S. L., Gelbgiser, D., & Weeden, K. A. (
2013
).
Feeding the pipeline: Gender, occupational plans, and college major selection
.
Social Science Research
,
42
,
989
1005
.
Moss-Racusin, C. A., Dovidio, J. F., Brescoll, V. L., Graham, M. J., & Handelsman, J. (
2012
).
Science faculty's subtle gender biases favor male students
.
Proceedings of the National Academy of Sciences
,
109
,
16474
16479
.
Murphy, R. (
1986
).
Weberian closure theory: A contribution to the ongoing assessment
.
British Journal of Sociology
,
37
,
21
41
.
Oaxaca, R. (
1973
).
Male-female wage differentials in urban labor markets
.
International Economic Review
,
14
,
693
709
.
Parkin, F. (
1979
).
Marxism and class theory: A bourgeois critique
.
London, UK
:
Tavistock Publications
.
Penner, A. M., & Willer, R. (
2019
).
Men's overpersistence and the gender gap in science and mathematics
.
Socius, 5
. https://doi.org/10.1177/2378023118821836
Petersen, T., & Morgan, L. A. (
1995
).
Separate and unequal: Occupation-establishment sex segregation and the gender wage gap
.
American Journal of Sociology
,
101
,
329
365
.
Phelan, J. E., Moss-Racusin, C. A., & Rudman, L. A. (
2008
).
Competent yet out in the cold: Shifting criteria for hiring reflect backlash toward agentic women
.
Psychology of Women Quarterly
,
32
,
406
413
.
Quadlin, N. (
2018
).
The mark of a woman's record: Gender and academic performance in hiring
.
American Sociological Review
,
83
,
331
360
.
Reskin, B. F., & Roos, P. A. (
2009
).
Job queues, gender queues: Explaining women's inroads into male occupations
.
Philadelphia, PA
:
Temple University Press
.
Rivera, L. A. (
2017
).
When two bodies are (not) a problem: Gender and relationship status discrimination in academic hiring
.
American Sociological Review
,
82
,
1111
1138
.
Ruggles, S., Flood, S., Geoken, R., Grover, J., Meyer, E., Pacas, J., & Sobek, M. (
2020
).
IPUMS USA: Version 10.0
[Data set].
Minneapolis, MN
:
IPUMS
.
Sassler, S., Glass, J., Levitte, Y., & Michelmore, K. M. (
2017
).
The missing women in STEM? Assessing gender differentials in the factors associated with transition to first jobs
.
Social Science Research
,
63
,
192
208
.
Shauman, K. A. (
2006
).
Occupational sex segregation and the earnings of occupations: What causes the link among college-educated workers?
Social Science Research
,
35
,
577
619
.
Shauman, K. A. (
2009
).
Are there sex differences in the utilization of educational capital among college-educated workers?
Social Science Research
,
38
,
535
571
.
Spence, M. (
1974
).
Market signaling: Informational transfer in hiring and related processes.
Cambridge, MA
:
Harvard University Press
.
Stiglitz, J. E. (
1975
).
The theory of “screening,” education, and the distribution of income
.
American Economic Review
,
65
,
283
300
.
Taylor, C. J. (
2010
).
Occupational sex composition and the gendered availability of workplace support
.
Gender & Society
,
24
,
189
212
.
Tomaskovic-Devey, D., & Skaggs, S. (
2002
).
Sex segregation, labor process organization, and gender earnings inequality
.
American Journal of Sociology
,
108
,
102
128
.
VanHeuvelen, T., & Quadlin, N. (
2021
).
Gender inequality in STEM employment and earnings at career entry: Evidence from millennial birth cohorts
.
Socius
,
7
. https://doi.org/10.1177/23780231211064392
Weeden, K. A. (
2002
).
Why do some occupations pay more than others? Social closure and earnings inequality in the United States
.
American Journal of Sociology
,
108
,
55
101
.
Weeden, K. A. (
2004
).
Profiles of change: Sex segregation in the United States, 1910–2000
. In Charles, M. & Grusky, D. B. (Eds.),
Occupational ghettos: The worldwide segregation of women and men
(pp.
131
178
).
Stanford, CA
:
Stanford University Press
.
Weeden, K. A. (
2019
).
Occupational segregation
.
Pathways
,
2019
(Special issue),
33
36
.
Weeden, K. A., Gelbgiser, D., & Morgan, S. L. (
2020
).
Pipeline dreams: Occupational plans and gender differences in STEM major persistence and completion
.
Sociology of Education
,
93
,
297
314
.
Williams, C. L. (
1992
).
The glass escalator: Hidden advantages for men in the ‘female’ professions
.
Social Problems
,
39
,
253
267
.
Williams, C. L. (
2013
).
The glass escalator, revisited: Gender inequality in neoliberal times, SWS Feminist Lecturer
.
Gender & Society
,
27
,
609
629
.
Winship, C. (
1977
).
A revaluation of indexes of residential segregation
.
Social Forces
,
55
,
1058
1066
.
Xie, Y., & Killewald, A. A. (
2012
).
Is American science in decline?
Cambridge, MA
:
Harvard University Press
.
Xie, Y., & Shauman, K. A. (
2003
).
Women in science: Career processes and outcomes
.
Cambridge, MA
:
Harvard University Press
.
This is an open access article distributed under the terms of a Creative Commons license (CC BY-NC-ND 4.0).

Supplementary data