You are on page 1of 13

Occupational Attainment and Segregation by Sex

Author(s): Randall S. Brown, Marilyn Moon and Barbara S. Zoloth


Source: ILR Review , Jul., 1980, Vol. 33, No. 4 (Jul., 1980), pp. 506-517
Published by: Sage Publications, Inc.

Stable URL: https://www.jstor.org/stable/2522696

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Sage Publications, Inc. is collaborating with JSTOR to digitize, preserve and extend access to
ILR Review

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
OCCUPATIONAL ATTAINMENT AND
SEGREGATION BY SEX

RANDALL S. BROWN, MARILYN MOON, and BARBARA S. ZOLOTH*

The authors use multinomial logit and multiple discriminant analyses to


predict the probabilities that an individual will attain each of several oc-
cupational categories based on that individual's characteristics and
qualifications. By estimating the parameters of this model from a sample of
men and then applying them to a sample of women, the authors simulate the
occupational distribution that these women would have attained had they
been treated as if they were men. Even after making adjustments for taste
differences between men and women, the authors find that their hypothetical
results vary substantially from women's actual occupational distribution.
They conclude that a significant portion of occupational segregation by sex
can be attributed to discrimination.

THE two most important manifestations earnings across occupations.- We choose


of sex discrimination in the labor mar- instead to focus directly on the process of
ket are wage differentials within occupa- occupational attainment.2 We develop a
tions and differences in the occupational model that uses microdata to predict the
distributions of men and women. To under- probabilities that an individual will end up
stand fully the nature and impact of this in each of several occupational categories.
discrimination, we must understand both
underlying processes: that of wage deter- 'This commonly takes the form either of including
occupational variables as regressors or of looking only
mination and that of occupational attain-
at intra-occupational wage differentials. See, for ex-
ment. Many researchers have used wage ample, Alan S. Blinder, "Wage Discrimination: Re-
determination models to measure the extent duced Form and Structural Estimates," Journal of
of sex discrimination, frequently including Human Resources, Vol. 8, No. 4 (Fall 1973), pp. 436-

some attempt to account for the variance in 55, and Michael J. Boskin, "Unions and Relative
Wages," American Economic Review, Vol. 62, No. 3
(June 1972), pp. 466-72. For an interesting way of
directly comparing earnings distributions both within
*The authors are an economist at Mathematica and between occupations, see Edward N. Wolff, "Oc-
Policy Research in Princeton, assistant professor at the cupational Earnings Behavior and the Inequality of
University of Wisconsin-Milwaukee, and assistant Earnings by Sex and Race in the United States," Re-
professor of agricultural economics at the University of view of Income and Wealth, Series 22, No. 2 (June
California-Davis, respectively. This study was sup- 1976), pp. 151 -66.
ported by funds granted to the Institute for Research on 2We use the term "occupational attainment" to
Poverty at the University of Wisconsin-Madison by the mean only the net effect of all the forces and processes
Department of Health, Education, and Welfare pur- that determine the occupation of a particular in-
suant to the Economic Opportunity Act of 1964. The dividual. This should not be confused with the socio-
authors wish to thank Nancy Williamson for her pro- logical use of the term nor with the sociological mod-
gramming assistance. els of status attainment. Readers interested in a model

Industrial and Labor Relations Review, Vol. 33, No. 4 (July 1980). ? 1980 by Cornell University.
0019-7939/80/3304-0506$00.75

506

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
OCCUPATIONAL ATTAINMENT & SEGREGATION BY SEX 507

By estimating the coefficients of this model allows us to simulate more accurately a


from a sample of men and then using these "nondiscriminatory" occupational distri-
estimates to simulate the occupational bution for women. As a result, we are better
distribution of a sample of women, we can able to measure the extent to which women
identify the impact of sex discrimination on are excluded from various occupations.
the occupational distribution of women. In
order to make the comparison less subject to A Model of Occupational Attainment
the criticism that women's "inferior" oc-
Heretofore, although some empirical
cupational distribution is attributable to a
work has been done on models of occupa-
lower career commitment, we apply our
tional choice, there has been little done on
procedures to a subsample of women whom
the subject of occupational attainment. Our
we identify as having a relatively strong
concern is in explaining the specific occu-
attachment to the labor force.
pation in which an individual is actually
Our primary interest is in comparing the
employed. Since the model is therefore in
actual occupational distribution of women
reduced form, reflecting both supply and
with the distribution that would prevail in
demand considerations, it should include
the absence of discrimination. Although
variables relevant to both sides of the mar-
other studies have asked similar questions,
ket. The individual seeking to maximize
very few have taken this particular ap-
his or her utility function chooses to work in
proach. Most of them have chosen instead
a certain occupation. Variables that may
to focus on the impact that occupational
affect this choice include socioeconomic
segregation and other forms of past dis-
background, marital-family status, aware-
crimination have on earnings differentials.'
ness of available alternatives, personal char-
One notable exception is the study by
acteristics, and individual tastes. Employ-
Schmidt and Strauss, which predicted in-
ers, on the other hand, hire individuals who
dividual probabilities of occupational at-
have qualifications fitting their job speci-
tainment. However, their model considered
fications. Such qualifications include edu-
far fewer occupational categories and used
cation, previous work experience, and voca-
less precise measures of education, experi-
tional training.
ence, and other personal characteristics than
From this underlying structure, it should
are available to US.4 Our approach thus
be theoretically possible to derive a condi-
tional probability function for the proba-
of an individual's occupational choice are referred to bility that an individual becomes employed
Michael J. Boskin, "A Conditional Logit Model of
in any one occupation. Thus, we have:
Occupational Choice," Journal of Political Economy,
Vol. 82, No. 2 (March 1974), pp. 389-98.
3For example, see Marianne A. Ferber and Helen M. (1) Pij|Xi = f(X,1. Xik)
Lowry, "The Sex Differential in Earnings: A Reap-
praisal," Industrial and Labor Relations Review, Vol. where Pij = the probability that indivi
29, No. 1 (October 1975), pp. 377 -87. They estimate ends up in occupation j; XIk = the value of
median occupational earnings using percent male in the kth personal characteristics for indi-
the occupation as an independent variable and find vidual i; and Xi = a K x 1 vector of personal
that differences in occupational distributions account characteristics for individual i for
for part of the male-female earnings gap. We have
elsewhere developed a model that allows the effects of i=l, ... N;
occupational distribution and wage discrimination to
be empirically disaggregated. See Randall S. Brown,
Marilyn Moon, and Barbara S. Zoloth, "Incorporat- k=l,... K.
ing Occupational Attainment in Studies of Male-
Female Earnings Differentials," Journal of Human
Resources, Vol. 15, No. 1 (Winter 1980), pp. 1 -28. that an individual will be in any one of five occupa-
4Peter Schmidt and Robert P. Strauss, "The Pre- tional categories ranging from professional to menial.
diction of Occupation Using Multiple Logit Mod- They then infer from the sign and significance of the
els," International Economic Review, Vol. 16, No. 2, race and sex dummies that race and sex discrimination
(June 1975), pp. 471 -85. They use education, po- do exist in access to occupations. Note that in order to
tential experience (age minus education minus 5), and isolate the effects of sex on occupational attainment we
dummies for race and sex to predict the probability restrict our analysis to whites only.

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
508 INDUSTRIAL AND LABOR RELATIONS REVIEW

The personal characteristics, not all of The Methodology


which would be exogenous to the indi- We assume that the unknown functional
vidual's occupational choice, are predeter- form for our conditional probability func-
mined for the purposes of predicting occu- tion can be adequately approximated by a
pational attainment. multinomial logistic model as follows:
We assume that, were it not for discrim-
ination and differences in tastes, the struc-
(2) e_ _ _
ture of occupational attainment would be = e -i--=-)
the same for men and women with compar-
L e X f
able personal characteristics. Consequently,
h =1
the ability to predict occupations for men
accurately on the basis of their backgrounds
where = a K x 1 vector of coefficients for
and qualifications allows us to construct a
occupation j and the other variables are as
hypothetical "discrimination-free" dis-
defined above. This specification yields a
tribution of occupations for women, but
separate coefficient vector for each possible
only if we assume that women's tastes for
occupation. The multinomial logit tech-
different occupations are the same as men's.
nique, when applied to our problem, al-
If we include in this new simulation of oc-
lows us to predict the probability that an
cupations only those women whose tastes
individual with a particular set of personal
about work are likely to be most similar to
characteristics is in each possible occupa-
men's, then we can begin to identify the
tion.6 The estimation requires the use of an
effects of discrimination on occupational
iterative maximum likelihood procedure.
attainment. The portion of the occupa-
To test the robustness of the logit model
tional distribution that we attribute to dis-
and to accommodate a larger number of
crimination is a "residual," since it repre-
occupational categories, we also use mul-
sents those differences in occupation be-
tiple discriminant analysis (DA). DA, while
tween the original distribution and the dis-
closely related to logit, differs in the dis-
tribution simulated for career-oriented
tributional assumptions.7 Discriminant
women that are not explained by their
analysis uses differences in group means to
socioeconomic characteristics. This ap-
construct classification schemes that are
proach is analogous to that used in studies
based on maximizing the within-group
that attribute to discrimination the unex-
relative to the between-group variances. It
plained wage differential that remains after
requires the assumptions that, for individ-
accounting for productivity differences.5
uals in occupation i, the independent vari-
ables have a multivariate normal distribu-
50ur attempt to account for productivity qualifi- tion and that the variance-covariance ma-
cations and for tastes raises the question of how to trix is constant across occupations.
define discrimination. One could certainly argue that Both techniques provide us with a set of
sex differences in personal preferences for alternative
occupations reflect past discrimination. Women and
men are raised to view their roles in society differ- 6A good explanation of the multinomial logit model
entially; thus, existing differences in occupational can be found in Marc Nerlove and S. James Press,
tastes reflect differential socialization. Some of the Univariate and Multivariate Log-Linear and Logistic
differences in personal characteristics that are used to Models, Manuscript R-1306-EDA/NIH (Santa Mon-
justify male/female wage differentials and acquisi- ica, CA: Rand Corporation, December 1973). For
tion of different occupations would instead be at- further details of this approach as applied here readers
tributed by this argument to sex role stereotyping are referred to Schmidt and Strauss, "The Prediction
and hence discrimination. If one chooses to define oc- of Occupation Using Multiple Logit Models," and to
cupational discrimination in this broad manner, then Brown, Moon, and Zoloth, "Incorporating Occupa-
very few personal characteristics should be controlled tional Attainment in Studies of Male-Female Earn-
for in our model. At the other extreme, if one were ings Differentials."
only interested in actual labor market discrimination, 7For a detailed comparison of logit and discrimi-
then all relevant pre-occupational-attainment factors nant analyses, see Daniel McFadden, "A Comment on
should be included in the specification. In this study, Discriminant Analysis 'Versus' Logit Analysis,"
to the extent permitted by the data, we employ the Annals of Economic and Social Measurement, Vol. 5,
latter approach. No. 4 (Fall 1976), pp. 511 -23.

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
OCCUPATIONAL ATTAINMENT & SEGREGATION BY SEX 509

J predicted probabilities for any individual information not commonly found in other
with characteristics Xi. These probabilities data sources.
for individuals are then used to obtain a After excluding those observations with
predicted occupational distribution for the missing data and those individuals who
sample. Two different classification tech- never participated in the labor force, we
niques are possible. One is to place each have a sample of 2,277 white males aged 45
individual in the occupation for which he through 59 (based on 1966 NLS data). In
or she has the highest probability of mem- 1967, the first survey year for women, the
bership. Although this is the technique sample of 1,968 women ranged in age from
generally applied in discriminant analysis, 30 to 44. In an effort to make the men and
it unnecessarily restricts the amount of women more comparable in age, we use 1971
information used from the results of both data for the women, aging that sample by
the logit and DA methods. For example, the four years.8 Because we use 1966 data for
probability that a particular individual is in men and 1971 data for women, we are im-
one occupation may be only slightly higher plicitly assuming that the structure of oc-
than for one or more of the remaining occu- cupational attainment for men did not
pations, or it may clearly be higher than change between 1966 and 1971. Although
that of all others. This information would older men may well have different occupa-
be lost if occupational assignment were tions than younger ones, they also have
determined solely by the maximum prob- different attributes. The difference in ages
ability. A more appealing technique (par- between men and women thus poses a prob-
ticularly when multiple groups are in- lem only if the structure of determination
volved) is to take the predicted probabilities (established through the coefficients)
of belonging to a specific occupational changes with age. Although some bias prob-
group and sum them over all observations. ably does arise from the age difference be-
This calculation yields the expected number tween the samples, it should understate
of sample members who would belong to the effects of discrimination on the occupa-
the specified occupation if the model were tional assignment of women since age tends
an accurate description of reality. By using to be positively related to occupational ad-
the sum-of-probabilities technique, we vancement.
therefore obtain predictions of the number Tastes. Because of the relatively compre-
of sample individuals who would be in each hensive work histories for the women in our
occupation. sample, we are able to make some judgments
Data. The data for this study were taken about their commitment to the labor force.
from the National Longitudinal Survey We assert that women with strong labor-
(NLS). The NLS has the most compre- force attachment, either by choice or neces-
hensive information on work history for a sity, are likely to be as career-oriented as
large sample of adult workers, a fact that men. Consequently, we identify a sub-
allows us to calculate a better experience sample of career-oriented women and simu-
variable for women than would other data late a new occupational distribution for
sets. Since women are more likely than men them alone, leaving the rest of the women
to have extended periods out of the labor in their actual occupations.
force, it is important to calculate this vari- Our segmented group of career-oriented
able more directly for them, rather than women includes all females who are heads
simply to use age minus years of schooling. of households or who earn more than their
Our experience variable for women is thus
more reliable and allows us to overcome one 8An alternative to aging the sample of women would
of the serious shortcomings of studies that have been to include age directly as an explanatory
use other data sources. The NLS also allows variable. However, since age is more likely to affect
the entire structure of the equation than it is to affect
us to separate current from previous job
the probability of a person attaining a particular
experience-another improvement in our occupation, we decided that making the ages of the
work history information. Finally, the study sample as close as possible was the preferable pro-
contains attitudinal and family history cedure.

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
510 INDUSTRIAL AND LABOR RELATIONS REVIEW

spouses. In addition, we include women ous experience.' For the men, previous ex-
who have both been active in the labor force perience is defined as the age at which their
at some time during the last five years and current job began minus YRSSCH minus
have spent one-half or more of their "po- VOCTRG/12 minus 6. Since actual labor
tential" working years in the labor force. We history data are available for women on the
define "potential" experience as age minus NLS, we can observe their levels of PRE VEX
years of all schooling minus six. The ratio directly. Number of children is included to
of actual to potential participation we use capture the pressure for a stable career often
to identify a woman as career-oriented rises attributed to those with family responsi-
as a linear function of age to a maximum of bilities.10 Both OCFATH and RES15 at-
.7 for the oldest women because several tempt to measure early influences on career
years absence from the labor force to have a choice, with the former a proxy for socio-
child will command a larger proportion of economic background and the latter indicat-
potential experience for a young woman ing exposure to available alternatives. The
than for an older woman. This segmented Rotter scale variable is an indicator of the
group of women totals 829, or 41 percent of individual's sense of control, providing a
the full sample, and, as one would expect, a qualitative scale ranging from 11 to 44: the
greater proportion of them are originally in lower the score, the greater the degree of
the higher status, male-dominated pro- internality or initiative."
fessional, managerial, and craft occupa- Using both logit and discriminant analy-
tions. ses, we first estimate occupational assign-
ment using the eight highly aggregated
The empirical specification. For the
categories listed in the Appendix. Since
occupational simulations, we use the fol-
much segregation exists within each of these
lowing independent variables constructed
eight categories, the use of so few categories
from information on the NLS files for ma-
of occupations obscures part of the occupa-
ture men and women:
KIDS = number of children; 9We do not use total experience since tenure on cur-
OCFA TH = Duncan index of father's rent job should not relate to current job attainment.
occupation when individ- '0Although marital status is frequently used for this
ual was age 15; purpose, there is good reason to believe this variable
would have a different effect on occupational grouping
YRSSCH = years of schooling;
for men than for women. Number of children is less
VOCTRG = number of months of vo- likely to do so, although access to some occupations
cational training; may be biased against women with children because of
RES15 = dummy variable for resi- competing demands on their time. The simulations
dence at age 15: urban = 1, were run without the KIDS variable, and, although the
numbers changed, the conclusions remained essen-
rural = 0;
tially the same. See Brown, Moon, and Zoloth, "In-
IESUM = sum of 11 Rotter scale corporating Occupational Attainment in Studies of
variables pertaining to Male-Female Earnings Differentials."
the individual's sense of "1The Rotter scale variables are responses by in-
dividuals to questions on attitude and motivation.
internal and external con-
Responses range from one through four, lower re-
trol; and sponses implying a greater sense of internal control,
PREVEX = labor market experience which has been interpreted to mean a greater sense of
prior to current job. initiative. To some extent, of course, aspirations are
endogenous: women with better jobs may score higher
on the Rotter scale and hence be qualified for even
PREVEX, YRSSCH, and VOCTRG re-
better jobs. This is similar to the impact of past dis-
flect different kinds of preparation for vari- crimination and socialization in that it will cause us to
ous careers. Experience before current or last understate the effects of discrimination on occupa-
job should be important for those occupa- tional attainment. For a thorough explanation of the
tions, such as foremen, that require experi- Rotter scale, see Paul J. Andrisani and Gilbert Nestel,
"Internal-External Control and Labor Market Ex-
ence in other jobs. Professional occupa-
perience," in U.S. Department of Labor, The Pre-
tions, on the other hand, are likely to re- Retirement Years, Vol. 4 (Washington, D.C.: G.P.O.,
quire substantial education but little previ- 1975), pp. 197 - 235.

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
OCCUPATIONAL ATTAINMENT & SEGREGATION BY SEX 511

tional differentiation between the sexes. of rows (2) and (3) reveals very little differ-
Consequently, we further subdivide these ence in the results from multiple discrimi-
eight aggregated groups into a total of 37 nant and multinomial logit analyses. This
categories on the basis of the type of work is indicative of the similarity of the two
they involve.12 Our purpose is to analyze methods.
occupational segregation better by both The only available measure of "success"
identifying traditionally female occupa- for the predicted occupational distributions
tions and grouping together occupations is how accurately the discriminant analysis
with similar requirements regardless of sex technique predicts for the men.15 Unfor-
roles. Information from the Dictionary of tunately, this criterion is not helpful for
Occupational Titles was used to combine logit since the sum of the predicted prob-
several of the three-digit occupations with abilities is constrained to equal the sample
similar skill levels and job qualifications frequency, thus guaranteeing an exact pre-
into each category. We use only the multiple diction. The x 2 statistic, which tests the
discriminant technique to analyze these 37 hypothesis that all the coefficients (except
categories since such a large number of the constant terms) of the model are zero,
categories is not tractable with logit analy- was computed and found to be significant
sis.13 at the .005 level for each of the logit models
estimated. This result implies that the struc-
tural model specified does a much better
Results job of predicting occupation (in terms of
maximizing the likelihood function) than
The occupational distributions. The does the use of sample proportions. Esti-
actual sample distributions of women and mated coefficients on education, vocational
men across the eight major occupational training, and number of children were gen-
groups are shown in rows 1 and 4 of Table 1. erally found to be highly significant. One
The second and third rows give the women's can see from the last two rows of Table 1
simulated occupational distributions based that the DA distribution for the men very
on redistributing only career-oriented closely approximates the actual eight-
women (row 2) and based on redistributing category occupational breakdown. Only for
all women (row 3), using both multiple crafts and laborers do the predicted pro-
discriminant analysis (a) and multinomial portions differ substantially from the actual
logit analysis (b).14 Comparing the simu- distribution.
lated distributions between parts (a) and (b) Of greatest interest are the differences
that appear between the actual and simu-
lated occupational distributions of women.
12Although it would have been preferable to use each In general, the simulated occupational
separate three-digit occupational category, we needed distribution of women resembles the actual
to ensure an adequate number of observations of men distribution of men much more closely than
in each category in order to perform the estimations.
it does the women's actual distribution. In
In "Occupational Earnings Behavior...", Wolff used
particular, a much larger proportion of
interval analysis to establish 32 categories of occupa-
tions with similar earnings. We did not use this type of women "become" managers and crafts
technique because we needed to consider several char- workers, while their representation in cleri-
acteristics at the same time. cal, sales, and service occupations drops
13By imbedding occupational attainment models
dramatically. In fact, when the full sample
within each of the eight broad occupational categories,
this limitation could be overcome. Since the results
from the logit and DA models are so similar for the
smaller number of categories, however, this has not 15Although the discriminant analysis program pro-
been pursued. vides measures such as the percent of sample success-
"4Length considerations preclude the reporting of fully classified, they are available only for the miaxi-
detailed coefficients. The complete logit estimation mum probability method. For details see Robeit A.
results are presented in Brown, Moon, and Zoloth, Eisenbeis and Robert B. Avery, Discrirninant Analysis
"Incorporating Occupational Attainment in Studies and Classification Procedures (Lexington: D.C. Heath
of Male-Female Earnings Differentials." and Co., 1972).

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
512 INDUSTRIAL AND LABOR RELATIONS REVIEW

00 0~0
- ~ l or) c' or)C' al 00

00 CZ 00

C,)~~~~~~~~~/

O ~~00 00
-bt -7 00 00? - - ie t1 *

t ~ ~ ~ ~ ~ ~ ~~~o a; o6 o
< C Q t- G00 Cl a) 0

Cq O C

.- .-

V) = .: N M m o. s X c c N UO<) O

00 00c t)
,:S.:C~
.- b-0 z 0s
0 Cb

cl D C 7 o6 o6 C7; C C
tu ~ ~ ~ ~ ~ -

k
~ ~ >Ct G
CZ~~~~~~~~~~~

-~~~~~~~7$T
0 0

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
OCCUPATIONAL ATTAINMENT & SEGREGATION BY SEX 513

of women is reclassified, the new percentage out of other occupations is also striking.
of women managers exceeds the actual pro- Within the sales category, women "leave"
portion of men in that category. the lower status sales jobs and "enter" the
Recalling that the women are classified higher status sales group. Women increase
by subjecting their own characteristics to in the crafts categories across the board,
classification functions based on the struc- while in some of the operative groups they
ture of the men's occupational attainment rise (drivers and heavy labor) and in others
helps us explain these patterns. While some they fall (graders, sorters, and wrappers,
of the "new" women managers come from and light labor).
highly educated, professional and technical Our occupational simulations for the
categories, most of the increase results from women are clearly hypothetical. We do not
the assignment of high probabilities of be- deal with the general equilibrium issues of
coming a manager to women who were what would happen to men (and women) if
originally in the less prestigious clerical the structure of occupational attainment
and service categories. Thus, although these were to change for women. It seems un-
women have the same characteristics as likely, for example, that one-fourth of all
male managers do in terms of training and men and women could be managers. More-
experience (especially the clerical workers), over, the assumptions that women not only
they have not attained managerial rank. gain employment in specific occupations in
Similarly, the large increase in female craft the same way as men but also have the same
workers appears to come from service occu- a priori probability of being in each occupa-
pations, indicating that many of these tion as men "ensures" that some women will
women have backgrounds similar to those of leave the traditionally "female" occupa-
men in crafts. In all cases the redistributions tions and enter those with greater propor-
resulting from reassigning all women ex- tions of men. To test the effect of this last
hibits the same pattern as those based on assumption about probabilities, we re-
only the "career-oriented" group but are classified women by DA, this time basing
much more extreme. our a priori probabilities on the total oc-
As shown in Table 2, the results for the cupational distribution of both men and
37-category distributions yield similar pat- women. The results were quite similar to
terns. Multiple discriminant analysis again those reported above. The women moved
accurately predicts occupations for the men out of the sales and clerical categories and
(compare columns 4 and 5). Only categories into the managerial group in similar mag-
4, 20, 29, and 31 display substantial devia- nitudes. Moreover, the resulting proportion
tion from the actual distribution. Moreover, of women in the managerial category again
we are now able to trace more specifically substantially exceeded the actual propor-
which occupations change for the women. tion of men. For the eight-category dis-
For example, the decline in women in the tributions, proportionately more men were
professional category occurs almost ex- predicted to be operatives and sales workers
clusively in groups 5 and 7, which include than in the original distribution for women;
teachers and nurses, respectively. Other yet the percentage of women in these occu-
groups within the professional category pations increased only slightly with multi-
gain, rather than decrease, in size. For ex- ple discriminant analysis and actually fell
ample, the proportion of doctors and law- by logit analysis. Similarly, for both the
yers rises from .1 to .4 percent and engineers 37-category and the professional simula-
rise from .4 to 2.6 percent (categories 4 and tions, the percentage of men in categories 5
3, respectively).16 The movement into and (teachers) and 7 (nurses) exceeded the simu-

results are consistent with those of both the eight- and


I6We also applied both the logit and DA techniques 37-category findings: the women leave categories 5
to the subsample consisting of the eight categories of and 7 (teachers and nurses) and are reclassified to
professionals identified in the Appendix. There are categories 2, 3, and 6 (social workers, clergy, engi-
208 men and 242 women in these occupations, in- neers, and technicians). The results, which are similar
cluding a total of 137 career-oriented women. The for both logit and DA, are available from the authors.

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
514 INDUSTRIAL AND LABOR RELATIONS REVIEW

Table 2. Actual and Predicted Distributions of Men and Women,


Using Multiple Discriminant Analysis on 37 Occupational Categories.
(in percentages)

Women Men

Predicted Predicted Predicted


Occupational Actual (intermediate)a (all women) (all men) Actual
Category (1) (2) (3) (4) (5)

(1) Professional
1 0.8 .7 0.7 0.8 0.6
2 0.4 .7 0.8 1.0 0.8
3 0.4 1.7 2.6 2.8 3.3
4 0.1 .4 0.4 1.5 0.7
5 6.3 3.1 0.6 1.5 1.8
6 0.7 1.4 3.4 1.7 2.0
7 3.8 1.9 0.05 0.4 0.4
8 0.1 .2 0.3 0.4 0.5

(2) Managers
9 0.4 .6 0.7 1.1 1.3
10 0.1 .7 1.9 0.8 0.7
11 4.4 10.4 20.7 15.0 16.2
12 0.4 .7 0.9 1.2 1.2

(3) Clerical
13 2.8 1.7 0.1 0.4 0.4
14 10.0 7.5 2.3 1.8 1.4
15 17.9 10.6 1.9 0.6 0.7
16 0.2 1.1 3.8 0.8 0.6
17 10.3 7.6 2.1 2.8 2.6

(4) Sales
18 0.4 .7 1.5 1.6 1.5
19 7.3 6.2 2.2 3.4 3.8

(5) Crafts
20 0.1 1.9 4.1 5.4 7.1
21 0.1 1.2 2.6 3.9 4.9
22 0.0 .9 2.4 2.1 2.7
23 0.5 2.7 6.2 4.3 5.3
24 0.1 2.3 5.3 4.7 4.2
25 0.4 .4 0.6 0.7 0.9
26 0.2 1.0 1.8 3.1 3.4

(6) Operatives
27 4.4 3.5 1.5 2.8 1.7
28 0.7 2.6 6.6 5.3 5.6
29 0.1 1.6 4.5 5.9 1.9
30 5.2 4.4 4.6 3.9 4.2
31 6.5 7.1 8.5 5.5 7.2

(7) Service
32 6.3 4.7 0.7 3.4 2.5
33 0.0 .5 0.9 1.6 1.9
34 8.9 5.8 0.3 1.7 1.1

(8) Laborers
35 0.1 .2 0.2 1.1 0.5
36 0.0 .4 0.6 2.0 0.9
37 0.3 1.0 2.0 3.2 3.6

a The percentages include all women, butonly the "career-oriented" group has been redistributed according to DA
classification.

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
OCCUPATIONAL ATTAINMENT 8& SEGREGATION BY SEX 515

Table 3. Segregation Indexes for Actual and Simulated Occupational Distributions,


Based on All NLS Men and Women.

Reassigning Reassigning Only


All Women "Labor-Force" Women

Sample Index a Index a

D S D S

For the 8 one-digit occupations


(2,277 men; 829 LF women; 1,169 other
women)
Actual distribution .5050 .3478 .5050 .3478
Discriminant analysis simulation .1500 .0311 .2955 .1398
Logit simulation .1507 .0338 .2965 .1378

For the 37 occupational categories


(2,050 6men; 807 LF women; 1,141 other
women)
Actual distribution .6261 .3457 .6261 .3457
Discriminant analysis simulation .2094 .0316 .3617 .1810

a The indexes are defined as follows:

D = Dissimilarity index = - -M| where Wi and M are the number of women and men res
2 W M
in the th occupation.
T~ (p -p)2
S = Segregation index where P is the proportion of women and Ti is the total number
of people in the ith occupation.

lated percentage of women. Women also to another in order to obtain equal propor-
exceeded men in percentage terms in two of tions across all occupations. The segrega-
the four managerial subcategories and three tion index (S) is a measure of association
of the seven craft categories after reclassi- between a person's occupation and sex, with
fication. In short, it is clear that labor market a higher degree of association indicating
discrimination contributes to a significant more segregation by sex across occupa-
waste of human resources with women tions.'7
being relegated too frequently to jobs for The calculated index values reveal sub-
which they would be considered over- stantial differences in distributional pat-
qualified if they were men. terns. For example, the actual value of D of
.505 in the eight-category case implies that
Segregation indexes. To compare the ef-
more than half of either the men or the wo-
fects of our redistributions on occupational
men would need to be transferred between
segregation, we calculated two different
segregation indexes for each of the occupa-
occupations in order for their proportions
to be exactly equal. However, our first simu-
tional distributions (see Table 3). Both
lation, based on reassigning all women,
measure the degree of segregation and both
produces a value of .150. This large drop
range from 0 to 1. A zero value indicates
reflects the strong combined effects of dis-
equal proportions of men and women in
each occupation, while a value of one re-
flects total segregation of the sexes. The
"7These indexes are more fully described and com-
dissimilarity index (D) represents the pro- pared in Barbara S. Zoloth, "Alternative Measures
portion of either men or women who would of School Segregation," Land Economics, Vol. 52,
have to be transferred from one occupation No. 3 (August 1976), pp. 278-98.

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
516 INDUSTRIAL AND LABOR RELATIONS REVIEW

crimination in the labor market and of dif- large differences between our actual and
ferences in tastes."8 However, even when we simulated distributions suggest that ap-
adjust for taste differences by reassigning propriate antidiscrimination policies could
only the "career-oriented" women, the re- be extremely fruitful in equalizing occupa-
sulting index value of .296 is still substan- tional opportunities between men and
tially below the actual value of .505. Identi- women. Repeating this type of analysis
cal conclusions are reached in both the periodically, coupled with the calculation
eight-category and the 37-category cases as of segregation indexes, could provide a use-
well as for both indexes."9 It is clear, in short, ful way of tracking the progress of these
that occupational segregation between policies. Finally the ability to predict occu-
men and women would, in each case, be pations for men and to simulate a new dis-
substantially reduced by assigning women tribution for women brings us closer to a
to occupations according to the men's comprehensive analysis of labor market
model of occupational attainment. discrimination.

Future Directions
Appendix
The methodology outlined in this paper
Representative 3-Digit Occupations Included in
can be used to differentiate between "justi- Each of the 37 Occupational Groupings
fied" (explained) and "unjustified" (unex-
plained) differences in occupational dis-
(1) Professional and technical
tribution in much the same way that wage 1. Authors, editors, librarians, personnel workers,
differentials are often separated into those public relations workers
two categories. We do need to be concerned 2. Clergy, recreation and group workers, social
and welfare workers
that the variables used to classify occupa-
3. Architects, engineers, draftsmen, designers
tions capture those personal characteristics 4. Lawyers, judges, physicians
that are relevant to career attainment. More- 5. Teachers, actors, artists, musicians, dieticians
over, taste for occupation, including the 6. Technicians (except medical and dental)
7. Physical and natural scientists, health practi-
working conditions associated with the tioners and technicians
occupation, is likely to be particularly im- 8. College professors, social scientists
portant. Nevertheless, our results suggest
that personal characteristics and tastes ex- (2) Managerial
plain only part of the differences in the 9. Buyers, purchasing agents, store department
heads
distribution of occupations between men 10. Floor managers, credit managers, inspectors,
and women and that a substantial portion conductors, ship officers
of the segregation in occupations can be 11. Building managers and superintendents, man-
agers nec.'
attributed to discrimination. As we argued
12. Officials and administrators (public admin.)
above, analyses of wage differentials be-
tween men and women, particularly those
(3) Clerical
that attempt to measure the effects of sex 13. Agents, cashiers
discrimination, need to deal explicitly with 14. Attendants, clerks, office machine operators,
occupational segregation as well as with un- receptionists, typists
15. Tellers, secretaries, bookkeepers, stenog-
equal pay for equal work. Furthermore, the
raphers
16. Messengers
17. Clerical workers nec.
'8Since the difference is a residual it may also con-
tain justifiable differences to the extent that we have
(4) Sales
omitted or incorrectly specified personal character-
18. Real estate and insurance agents or brokers,
istics that affect occupational attainment.
advertising agents
'9Note also that the similarity of the index values for 19. Auctioneers, demonstrators, peddlers, news-
the discriminant analysis and logit simulations in boys, sales nec.
Table 3 further supports the claims made above that
these two procedures appear to perform equally well
in predicting distributions across the eight one-digit
occupations. Inec. = not elsewhere classified.

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms
OCCUPATIONAL ATTAINMENT 8c SEGREGATION BY SEX 517

(5) Crafts 30. Light labor (examiners and checkers, textile


20. Mechanics and repair operatives)
21. Cabinetmakers, carpenters, electricians, 31. Operatives nec.
plumbers
22. Other construction workers
(7) Service
23. Foremen, inspectors
24. Big machine and heavy metal workers 32. Cleaning and laundry services (housekeepers,
janitorial services)
25. Higher level crafts (engravers, jewelers, op-
33. Protective service (police, fire, guards)
ticians, upholsterers)
34. Personal service (barbers, bartenders, cooks,
26. Lower level crafts (bookbinders, typesetters,
practical nurses)
furriers, millers)

(6) Operatives (8) Laborers


27. Apprentices, attendants, graders and sorters 35. Heavy laborers (longshoremen, lumbermen,
and wrappers teamsters)
28. Drivers (bus, truck, taxicab), other motormen 36. Other laborers (fishing, garage laborers, gar-
29. Heavy labor (blasters, smelters, mine opera- deners)
tives, brakemen) 37. Laborers nec.

This content downloaded from


202.41.10.93 on Fri, 08 Sep 2023 05:01:21 +00:00
All use subject to https://about.jstor.org/terms

You might also like