Causes and Effects of Gender Occupational Segregation: Overlooked Obstacles to Gender Equality

Abstract

In this chapter, an extremely uneven distribution of professional occupations that women enter is revealed and the problems associated with these biases are elucidated. Professions are classified into two types: Type 1 professions , which include three representative human service professions of high socioeconomic status—namely, medical doctors, dentists, college professors—and all professions that are not human service professions; and Type 2 professions , which include human-service professions other than medical doctors, dentists, and college professors. Both the United States and Japan have a large number of women employed in Type 2 professions and clerical jobs; however, Japan has a markedly lower proportion of women in Type 1 professions and managerial positions than the United States. Moreover, an examination of the gender wage gap by occupation shows that the differences between men and women are relatively small within Type 1 professions and managerial positions, and women’s average wages in Type 2 professions and clerical occupations are much lower than those of men within the same occupations and are significantly lower than men’s average wages in blue-collar occupations. Thus, women are subject to a two-fold wage disadvantage. On the one hand, the proportion of women is miniscule in occupations with relatively high wage and smaller gender wage gaps (Type 1 professions and managerial positions). On the other hand, the proportion of women is large in white-collar occupations exhibiting the largest wage gaps by gender (Type 2 professions and clerical occupations). Further, in this chapter, whether gender occupational segregation can be explained by gender disparities in human capital is analyzed. The results, although paradoxical, indicate that the gender equalization of human capital intensifies occupational gender segregation between men and women. This segregation occurs because the increases of women in female-dominated Type 2 professions and the decreases of women in non-service manual occupations in which women are already underrepresented—as a result of more human capital —outpace the increases of women in underrepresented Type 1 professions and managerial positions. In this chapter, theories on gender occupational segregation are also reviewed and their consistency with the empirical results is examined.

Appendices

Appendix 1: Matching Model

Regarding the DiNardo-Fortin-Lemieux (DFL) method, please refer to Appendix 1 of Chap. 2. Of the assumptions made by the inverse-probability of treatment (IPT) weighting method of Rubin’s causal model (Rubin 1985), on which the DFL method is based, the matching model developed by Yamaguchi (2017) eliminates only the “Stable Unit Treatment Value (SUTVA) Assumption” and replaces it with the assumption that the outcome’s marginal distribution is independent of the covariate distribution. SUTVA is the assumption that the assignment of other people to the treatment or control group has no effect on the outcome when each individual i is assigned to the treatment group, Y_1i, or on the outcome when the individual is assigned to the control group, Y_0i. In reality, each individual is assigned to either the treatment or the control group. Hence, one of the two potential outcomes is the observed outcome, and the other is the unobserved counterfactual outcome for each individual. Therefore, the SUTVA assumption implies that the values of Y_1i and Y_0i remain unchanged, given the attributes of individual i, even if the relationship between treatment variable X and the outcome-affecting covariates V change in the population. However, the distribution of outcomes changes when the distributions of individual attributes change due to change in the relationship between X and V. When X represents the gender dummy variable—as opposed to a regular treatment variable—and covariates V consist of mediating variables, the DFL method decomposes the gender difference in the outcomes into the component explained as the result of gender difference in the mediating variables V, and the remaining unexplained component.

The matching model omits the SUTVA assumption from the DFL method and, replaces it an alternative assumption that the probability of each individual obtaining job outcome j when the relationship between X and V changes within the population is determined by retaining the parameters that represent the effects of X and V on outcome Y implicitly assumed by the DFL model, and by social forces that keep the marginal distribution of outcome Y unchanged given the fixed demand for job j. We now assume that the observed result is expressed by the following “saturated model” and is represented by Eq. (3.2):

$$ P_{ij} = \alpha ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{0j} ) + \beta ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{1j} )X_{i} ,\quad {\text{for }}j = 1, \ldots ,J - 1 $$

(3.2)

where $P_{ij}$ is the probability of individual i with attribute ${\mathbf{v}}_{i}$ to have job j, and $\alpha ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{0j} )$ is the “main effects” for attributes v where ${\varvec{\uptheta}}_{0j}$ is the parameter dependent on outcome j. Variable $x_{i}$ is the gender dummy variable, whereby women are assumed to have the value 0 and men are assumed to have the value 1, $\beta ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{1j} )$ is the effect of gender on the outcome, and ${\varvec{\uptheta}}_{1j}$ is the parameter for the dependence of β on ${\mathbf{v}}_{i}$ arising from the assumed interaction effects of attribute ${\mathbf{v}}_{i}$ and gender $x_{i}$ on the outcome. Attributes V are all categorical variables, and Eq. (3.2) assumes the presence of all high-order interaction effects of gender X and attributes V. The “saturated model” thus produces probability $P_{ij}$ for each combination of gender and attribute set, which is equal to the sample estimate of the proportion for the combination.

The gender difference in the mean of Eq. (3.2) is then given as:

$$ \begin{aligned} \overline{P}_{j}^{M} - \overline{P}_{j}^{W} & = \left\{ {E(\alpha ({\mathbf{v}}|{\varvec{\uptheta}}_{0j} )|X = 1) - E(\alpha ({\mathbf{v}}|{\varvec{\uptheta}}_{0j} )|X = 0)} \right\} \\ & \quad + E(\beta ({\mathbf{v}}|{\varvec{\uptheta}}_{1j} )|X = 1). \\ \end{aligned} $$

If we apply the inverse probability of treatment (IPT) weights $\omega_{DFL} ({\mathbf{v}},x)$ that attain statistical independence between X and V, then the weighted gender difference is given as

$$ E_{{\omega_{DFL} }} (\overline{P}_{j}^{M} ) - E_{{\omega_{DFL} }} (\overline{P}_{j}^{W} ) = E_{{\omega_{DFL} }} (\beta ({\mathbf{v}}|{\varvec{\uptheta}}_{1j} )|X = 1). $$

Here $E_{{\omega_{DFL} }}$ denotes the weighted mean by using weights $\omega_{DFL} ({\mathbf{v}},x)$. This weighted gender difference in the mean of the DFL estimate for the gender difference in the proportion of attaining occupation j unexplained by gender differences in the attributes.

In contrast, the matching model is assumed to be represented by Eq. (3.3):

$$ P_{ij} = \phi_{j} ({\mathbf{V}},{\mathbf{x}})\left\{ {\alpha ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{0j} ) + \beta ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{1j} )x_{i} } \right\},\quad {\text{for }}j = 1, \ldots ,J - 1. $$

(3.3)

In Eq. (3.3), factors $\alpha ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{0j} )$ and $\beta ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{1j} )$ retain the same values as in Eq. (3.2), but Eq. (3.3) adds the new parameter set, {$\,\phi_{j}$}. Those parameters take a value of 1 for the observed joint distribution of (X, V), whereas for a counterfactual joint distribution of (X, V), they are assumed to be determined to satisfy the condition that the marginal distribution of occupations equals the observed distribution. $\phi_{j} ({\mathbf{V}},{\mathbf{x}})$ depends only on the macro joint distribution of V and X, and not on the individual values of $x_{i}$ and ${\mathbf{v}}_{i}$.

Generally, each counterfactual situation assumed in the analysis in the matching model , Eq. (3.3) is assumed to satisfy the following five conditions.

• (a) With respect to each value of v_i, parameters $\alpha ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{0j} )$ and $\beta ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{1j} )$ have the same values as in Eq. (3.2).

• (b) The marginal distribution for gender X is equal to the observed distribution.

• (c) Under the counterfactual situation of gender equality in the distribution of V, the counterfactual female distribution of the attributes become identical to the observed male distribution of the attributes.

• (d) For each counterfactual situation , gender X and attributes V are statistically independent.

• (e) Parameter $\,\phi_{j}$ is adjusted such that the marginal distribution for occupations agrees with the observed marginal distribution in the counterfactual situation .

When condition (d) is satisfied, the gender differences in the probability of attaining job j that cannot be explained by the gender differences in attributes are given, by using the matching method weight $\omega_{M} ({\mathbf{v}},x)$ which is subsequently discussed, by:

$$ \overline{P}_{j}^{M} - \overline{P}_{j}^{W} = \phi_{j} E_{{\omega_{M} }} (\beta ({\mathbf{v}}|{\varvec{\uptheta}}_{1j} )) $$

(3.4)

Here, the methodological problem is to estimate the set of $\{ \phi_{j} \}$ that satisfies conditions (a)–(e). We now denote by Z the category variable representing occupations, by $f(x,z,{\mathbf{v}})$ the observed joint frequency distribution of X, Z and V, in the sample, and by $F_{1} (x,z,{\mathbf{v}})$ is the estimate for the joint distribution from the DFL method that satisfies the counterfactual statistical independence between X and V. An appropriate propensity score estimation makes the joint distribution $F_{1} (x,z,{\mathbf{v}})$ satisfy conditions (a)–(d); however, condition (e) is not satisfied. Moreover, $\,\phi_{j} = 1$ for $F_{1} (x,z,{\mathbf{v}})$. Here, if the joint distribution $F(x,z,{\mathbf{v}})$ satisfies the following two conditions, conditions (b)–(e) hold. Conditions (b)–(d) hold as a result of condition (A), and condition (e) holds as a result of condition (B).

• (A) $F(x, + ,{\mathbf{v}}) = F_{1} (x, + ,{\mathbf{v}})$

• (B) $F( + ,z, + ) = f( + ,z, + )$

Here, the + symbol indicates the sum across categories. For the remaining condition (a), it is satisfied because the following iterative estimation of parameters keeps the values of the parameters $\alpha ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{0j} )$ and $\beta ({\mathbf{v}}_{i} |{\varvec{\uptheta}}_{1j} )$ unchanged.

Taking the estimated values of the joint distribution under the DFL method as the initial value and using the following Deming–Stephan iterative proportional fitting, the converged value $F^{*} (x,z,{\mathbf{v}})$ obtained from performing the iterative fittings with the following equations satisfies both conditions (A) and (B).

First, assuming an estimated value of the joint distributions from the DFL method as $F_{1} (x,z,{\mathbf{v}})$, begin with t = 1 and increase the value of t until it converges, repeating the following four steps.

• (S1) Calculate: $F_{2t - 1} ( + ,z, + ) = \sum\nolimits_{x} {\sum\nolimits_{v} {F_{2t - 1} (x,z,{\mathbf{v}})} }$

• (S2) Calculate: $F_{2t} (x,z,{\mathbf{v}}) = F_{2t - 1} (x,z,{\mathbf{v}})\frac{f( + ,z, + )}{{F_{2t - 1} ( + ,z, + )}}$

• (S3) Calculate: $F_{2t} (x, + ,{\mathbf{v}}) = \sum\nolimits_{z} {F_{2t} (x,z,{\mathbf{v}})}$

• (S4) Only if $F_{1} (x, + ,{\mathbf{v}}) > 0$ is satisfied, then

  $$ {\text{Calculate:}}\;F_{2t + 1} (x,z,{\mathbf{v}}) = F_{2t} (x,z,{\mathbf{v}})\frac{{F_{1} (x, + ,{\mathbf{v}})}}{{F_{2t} (x, + ,{\mathbf{v}})}} $$

The converged value is expressed by $F^{*} (x,z,{\mathbf{v}})$. The product of the iterative fitting value z = j from step (S2) is the estimated value $\widehat{\phi }_{j}$ of $\phi_{j}$. Further, the product of the iterative fitting value for step (S4) is represented by $\varphi ({\mathbf{v}},x)$, and if the weights for the DFL and matching methods are expressed by $\omega_{DFL} ({\mathbf{v}},x)$ and $\omega_{M} ({\mathbf{v}} , { }x),$ respectively, the following two equations are satisfied.

$$ \omega_{M} ({\mathbf{v}},x) = \omega_{DFL} ({\mathbf{v,}}x)\varphi ({\mathbf{v}},x) $$

(3.5)

$$ \phi_{j} \varphi ({\mathbf{v}},x) = F^{*} ({\mathbf{v}},j,x)/F_{1} ({\mathbf{v}},j,x) $$

(3.6)

For a more detailed explanation of the matching method , please refer to Yamaguchi (2017).

Appendix 2: Application of the Eight Occupational Classifications Used in the Tables to the Sub-classifications of Japanese and U.S. Data

• 1. Occupational classification codes for sub-major groups used in the Population Census of Japan, and application of eight categories

501–509:

Type 1 professions

510–516:

Type 2 professions

517–519:

Type 1 professions

520–523:

Type 2 professions

524:

Type 1 professions

525–526:

Type 2 professions

527–536:

Type 1 professions

537–539:

Type 2 professions

540–544:

Type 1 professions

545–553:

Managerial occupations

554–565:

Clerical occupations

566–577:

Sales occupations

578–592:

Service work

593–622:

Other

623–684:

Non-service manual jobs

685–686:

Other

687:

Service work

688–689:

Other

701:

Clerical occupations

702:

Non-service manual jobs

703:

Type 2 professions

704:

Non-service manual jobs

706:

Other

999:

Other

• 2. Occupational classification codes for sub-major groups used in the U.S. Census, and application of the eight categories

0010–0430:

Managerial occupations

0500–1965:

Type 1 professions

2000–2025:

Type 2 professions

2040–2060:

Other (clergy, religious workers)

2100–2200:

Type 1 professions

2300–2340:

Type 2 professions

2400–2440:

Type 1 professions

2540–2550:

Type 2 professions

2600–3100:

Type 1 professions

3030:

Type 2 professions

3040:

Type 1 professions

3050:

Type 2 professions

3060–3140:

Type 1 professions

3150–3245:

Type 2 professions

3250:

Type 1 professions

3255–3655:

Type 2 professions

3700–3955:

Other

4000–4650:

Service work

4700–4955:

Sales occupations

5000–5940:

Clerical jobs

6005–6130:

Other

6200–8965:

Non service manual jobs

9030–9750:

Other