NORC at the University of Chicago
Interfirm Segregation and the Black/White Wage Gap
Author(s): William J. Carrington and Kenneth R. Troske
Source: Journal of Labor Economics, Vol. 16, No. 2 (April 1998), pp. 231-260
Published by: The University of Chicago Press on behalf of the Society of Labor
Economists and the NORC at the University of Chicago
Stable URL: http://www.jstor.org/stable/10.1086/209888
Accessed: 23-03-2018 13:17 UTC
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
http://about.jstor.org/terms
NORC at the University of Chicago, Society of Labor Economists, The University of
Chicago Press are collaborating with JSTOR to digitize, preserve and extend access to Journal
of Labor Economics
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
Interﬁrm Segregation and the
Black/White Wage Gap
William J. Carrington, Charles River Associates
Kenneth R. Troske, University of Missouri—Columbia
This article studies interﬁrm racial segregation in two newly developed
ﬁrm-level databases. We ﬁnd that the interﬁrm distribution of
black and white workers is close to what would be implied by random
assignment. We also ﬁnd that black workers are clustered in
employers where managers, owners, and customers are also black.
These ﬁndings may be reconciled by the facts that (a) there are not
enough black employers to generate much segregation and that (b)
other forces may systematically integrate black and white workers.
Finally, we ﬁnd that the black/white wage gap is primarily a withinﬁrm
phenomenon.
I. Introduction
Interracial contact is an important index of social health. Yet while
there are innumerable studies of residential racial segregation, we know
very little about the extent to which blacks and whites are integrated at
work. This is unfortunate because employed adults spend a large fraction
We thank Bill Evans and seminar participants at the Center for Economic
Studies, the University of Maryland, and the National Bureau of Economic Development
for helpful comments. The opinions expressed here are solely ours and
do not reﬂect the opinions of the U.S. Census Bureau. The data used in this
article were collected under the provisions of Title 13 U.S. Code and are available
only at the Center for Economic Studies, U.S. Census Bureau. To obtain access
to these data, contact The Chief, Center for Economic Studies, U.S. Census
Bureau, RM 211/WPII, Washington, DC 20233.
[ Journal of Labor Economics, 1998, vol. 16, no. 2]
᭧ 1998 by The University of Chicago. All rights reserved.
0734-306X/98/1602-0003$02.50
231
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
232 Carrington/Troske
of their time at work and because perhaps an even larger fraction of
nonfamily social interaction takes place at work. Thus, an understanding
of racial integration in society as a whole requires some understanding
of racial segregation at work.
This article advances our understanding of interﬁrm racial segregation
in three ways. First, there are very few studies of interﬁrm racial segregation,
largely, we believe, because the appropriate data have not been
available. Thus, our work simply lengthens a short literature on this topic.
Second, such studies as exist use data from the mid-1970s (Becker 1980),
the late 1960s (Flanagan 1973), or from the early part of the twentieth
century (Higgs 1977). In contrast, we use data from the late 1980s. Finally,
commonly used measures of segregation conﬂate systematic segregation,
which occurs in some models of labor market discrimination, for
example, with segregation due to the random allocation of workers to
ﬁrms. In contrast, we use new methods that better sort out the systematic
and random components of segregation.
Our ﬁrst results simply measure interﬁrm segregation. We ﬁnd that
black and white workers are quite segregated by conventional measures.
However, when we look within metropolitan statistical areas (MSAs),
we ﬁnd that the distribution of black and white workers is surprisingly
close to what would be implied by random allocation of workers to ﬁrms.
We ﬁnd virtually no systematic segregation in our ﬁrst sample of large
manufacturing establishments, and while we ﬁnd some systematic segregation
within our second sample of smaller ﬁrms drawn from a range of
industries, segregation is still far from complete. This is particularly true
among workers stratiﬁed by industry, occupation, or educational attainment.
Indeed, in some cases we ﬁnd that black and white workers of
similar skill are actually more integrated than random allocation would
predict. In sum, we ﬁnd that the distribution of black and white workers,
particularly when workers are grouped by skill, can be modeled reasonably
well by a random hiring model.
Our second set of results examines the matching of workers to ﬁrms.
Our main ﬁnding is that black workers are disproportionately sorted
into ﬁrms in which owners, managers, and customers are also black.
These ﬁndings are at ﬁrst contradictory, because how can the distribution
of workers look nearly random, while at the same time black
workers are sorted into ‘‘black’’ employers (i.e., those with black owners,
managers, and customers)? There are two complementary answers
to this question. First, while the relationship between the race of employers
and the race of workers is moderately strong, there are not
many black employers, so the relationship does not generate much
segregation. Second, it is possible that Title VII, afﬁrmative action,
and other difﬁcult-to-identify forces systematically integrate black and
white workers. On balance, these weak opposing forces induce an
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
233The Black/White Wage Gap
interﬁrm distribution of workers that is roughly consistent with the
random allocation of workers to employers.
Our ﬁnal results decompose the black/white wage gap in the manufacturing
industry into within- and between-plant components. The central
ﬁnding is that most of the black/white wage gap among men is accounted
for by within-plant differences in pay. In fact, we ﬁnd that both black
men and black women are disproportionately sorted into plants with
above-average wages. In addition, we ﬁnd that the wages for all workers
tend to increase with the fraction of their coworkers that are of the
opposite race. Thus, whites earn the most in plants with many blacks,
and blacks earn the most in plants that are nearly all white.
This article proceeds as follows. Section II describes our data drawn
from the Worker-Establishment Characteristics Database and the Characteristics
of Business Owners survey. Section III discusses our approach
to measuring the systematic component of segregation, and Section IV
uses this approach to measure and interpret interﬁrm racial segregation.
Section V then analyzes the determinants of interﬁrm segregation, and
Section VI assesses the role of segregation in accounting for the black/
white wage gap. Although this article is primarily descriptive, Section
VII discusses some implications of the results.
II. Data Sources
This section describes the two data sets used in this article, the ﬁrst of
which is the Worker-Establishment Characteristics Database (WECD).1
In the WECD workers are identiﬁed through household responses to the
1990 census long form, which contains standard demographic information
as well as the location and industry code for each respondent’s place of
work.2
All workers located in an industry-location cell with a unique
establishment are matched to that establishment, whereas workers in cells
without a unique establishment are not included in the sample. Establishments
themselves are identiﬁed as being unique or not based on our
analysis of the Census Bureau’s list of manufacturing establishments.
This matching process results in a data set consisting of 199,558 workers
matched to 16,144 plants. Since the WECD matches workers only to
employers that are unique in an industry-location cell, it tends to eliminate
small plants that are less likely to meet this criterion. Thus, our WECD
results describe segregation in large manufacturing plants, and the results
may not generalize to the economy at large. However, Bound and Free-
1
See Troske (in press) for a more complete description of the WECD.
2
Industry information is recorded at the three-digit level. For establishments
in urban areas (primarily MSAs), a plant’s location is coded at the block level.
For establishments in rural areas, a plant’s location is coded at the place level.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
234 Carrington/Troske
man (1992) argue that the post-1975 decline in the relative wages of
young black men is partially attributable to the employment declines in
manufacturing, so manufacturing is a particularly interesting sector to
study from the perspective of racial segregation and wage differences.
The large plant focus of the WECD leads us to also study segregation
among the smaller ﬁrms surveyed in the Characteristics of Business Owners
(CBO) database. The CBO is the result of a 1987 Census Bureau
survey of 125,000 small business owners.3
The CBO is essentially a survey
of ﬁrm owners, but one can merge owner responses to obtain a ﬁrm-level
data set (Carrington and Troske 1995). The resulting database includes
information on the ﬁrm itself (location, industry, receipts, capital), the
ﬁrm’s owners (income, race, sex, age, education), the ﬁrm’s customers
(% minority), and the ﬁrm’s workforce (number of employees, % minority,
% female, payroll).4
These latter data allow calculations of interﬁrm
racial segregation. Compared with the WECD, the CBO is limited
in that it has no information on workers’ characteristics other than race
and sex and reveals nothing about the within-ﬁrm distribution of pay.
However, it complements the WECD with demographic information on
owners and customers and is not restricted to the manufacturing industry.
It bears emphasis that in the CBO we analyze segregation across small
ﬁrms, while in the WECD we analyze segregation across large plants. In
3
Owners were included in the sample frame if they ﬁled their tax return with
one of the following IRS forms: 1040 (Schedule C), 1065, or 1120S. The 1040
(Schedule C) returns correspond to individual proprietorships, or unincorporated
businesses owned by an individual. The 1065 returns include unincorporated
businesses owned by two or more persons. Finally, the 1120S returns correspond
to subchapter S corporations, which are legally incorporated businesses with 35 or
fewer shareholders who elect to be taxed as individuals rather than corporations.
Corporations ﬁling a regular 1120 tax return were excluded from the sample.
The CBO comprises ﬁve equal-sized panels of business owners drawn from ﬁve
demographic groups: Hispanics, blacks, other minorities, women, and nonminority
men. The equal size of the panels obviously required that the CBO oversample
certain ownership groups, but sample weights can be used to recover the attributes
of a random sample.
4
Linking owner characteristics to the ﬁrm is trivial for ﬁrms owned by one
person, but multiowner ﬁrms are slightly tricky because not all owners are alike.
Following the work of previous CBO users (Bates 1988; Carrington and Troske
1995), we use the cross-owner mean for continuous variables (such as education)
and the cross-owner mode for discrete variables (such as sex or race). In cases
of ties for the discrete variables, we use the mode containing the owner that
reports spending the most hours per week at the business. For example, if a ﬁrm
has two white and one black owners, then we describe the ﬁrm as being ‘‘whiteowned.’’
If a ﬁrm has one white and one black owner, then we describe the ﬁrm
as ‘‘white-owned’’ if the white reports working more weekly hours at the ﬁrm,
and as ‘‘black-owned’’ if not. Single-owner ﬁrms account for 57% of the ﬁrms
and 39% of the employment in our sample.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
235The Black/White Wage Gap
statements referring to both data sets, we generally speak of ‘‘interﬁrm’’
segregation, though this applies directly only to the CBO.5
One complication with the CBO is that employees are only identiﬁed
as ‘‘minority’’ or not.6
Since only about 50% of all minorities (i.e., those
reported as black, Asian, or Hispanic) are black, we imposed two restrictions
to minimize the number of nonblack minorities in our CBO sample
of employees. First, we excluded CBO ﬁrms owned by Asians and Hispanics
because we suspect that minority employees of such ﬁrms are
much less likely to be black. Second, we restricted the CBO sample to
states and MSAs where blacks account for over 75% of all minorities.7
This amounts to restricting the sample to southern states such as Mississippi,
Georgia, and North Carolina, and to certain northern cities such
as Detroit.8
In addition, this restriction increases the black share of em-
5
Because the CBO oversamples black and minority business owners, we use
sample weights in our CBO analysis. The sample weights in the CBO are inversely
proportional to the estimated probability that each ﬁrm entered the sample. We
scaled the weights so that they summed to the sample size.
6
More particularly, employers reported the fraction of ‘‘white non-Hispanics’’
in their workforce.
7
We treated the non-MSA portions of each state as a single independent MSA
(e.g., rural Texas).
8
These states and MSAs were identiﬁed from our analysis of the outgoing
rotation groups of the Current Population Survey (CPS) for the months of
January 1987 to December 1991. We restricted the sample to those employed at
the time of the survey and to those aged 16 or older. In addition, we restricted
the sample to those in their ﬁrst year of CPS interviews (out of a maximum of
two outgoing CPS interviews) to avoid double-counting individuals. The resulting
sample included roughly 570,000 individuals, of whom approximately 150,000
were minorities. Blacks account for a small fraction of all minorities in almost all
western cities, 21% in Los Angeles, for example. In contrast, blacks account for
a very high share of minorities in the South, with the exception of the Miami MSA
and the MSAs in Texas. For example, blacks account for 99% of the minorities in
Birmingham, Alabama. While there is a sharp distinction between the South and
West along these lines, MSAs within the Northeast and Midwest are more varied.
Minorities are relatively heterogeneous in some northern cities such as New York
and Chicago, but blacks account for a very large share of all minorities in other
northern cities. For example, blacks account for 90% of all minorities in the
Detroit metropolitan area. Thus, restricting the sample to MSAs where blacks
account for a large share of all minorities is equivalent to restricting the sample
to all southern MSAs save for Miami and those in Texas and to a smattering of
MSAs in the North. It is worth noting that black workers are disproportionately
located in states and MSAs where blacks account for the bulk of all minorities.
For example, 75% of all nonrural blacks (i.e., those who live in an MSA) live in
MSAs where blacks account for at least 54% of all minorities. Similarly, 50% of
blacks live in MSAs where blacks account for more than 83% of all minorities,
and 25% of blacks live in MSAs where blacks account for more than 92% of all
minorities.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
236 Carrington/Troske
ployment in our CBO sample to roughly 30%. While this does not
completely avoid the confusion of blacks with other minorities, these
restrictions provide some legitimacy to the interpretation of our CBO
results as reﬂecting black/white differences.9
Since neither the WECD nor the CBO contain a random sample of
workers, it is important to ask whether patterns of segregation found in
these data are likely to parallel the rest of the economy. Of course, a
precise answer to this question would require a random sample of workers
that contained employer identiﬁers, and it is the absence of such data that
led us to the WECD and the CBO in the ﬁrst place. Nevertheless, table
1 considers the measurable similarities between workers in the WECD,
the CBO, and two relatively representative data sets. Columns 1 and 2
present summary statistics for manufacturing workers in the 1990 Decennial
Census, for blacks and whites separately. Columns 3 and 4 present
analogous statistics for the WECD, columns 5 and 6 present results for
the May 1988 Current Population Survey, and columns 7 and 8 present
results for the CBO.10
The comparisons between the Decennial Census
data and the WECD suggest several ways in which WECD workers are
different from the manufacturing industry at large. First, WECD workers
are more likely to live in the Northeast and Midwest and less likely to
live in the South and West. Second, average hourly wages and annual
earnings for workers are higher in the WECD than in the Decennial
Census. Finally, the black/white wage and earnings gap is somewhat
lower in the WECD than in the Decennial Census. These facts suggest
that, while the WECD is not wildly different from the rest of the manufacturing
industry, there is some legitimate concern about the representativeness
of the WECD.
Table 1 also illustrates some special features of our CBO sample. First,
the black population share is much larger than that in the WECD or
in the general U.S. population. This arises because we have lumped all
‘‘minorities’’ into the ‘‘black’’ group and because we have restricted the
9
Our CBO analysis is also complicated by the fact that while most multiunit
CBO ﬁrms operate in a single MSA, a small fraction (280 out of more than 7,000
ﬁrms in the CBO) have establishments in more than one MSA. This is potentially
problematic when we later assign each ﬁrm as belonging to a particular MSA,
particularly because these multi-MSA ﬁrms are somewhat bigger than average
and because 20% of their employment occurs outside of their main MSA. Unfortunately,
the available data do not allow us to break out each ﬁrm’s employment
by MSA and race simultaneously. Thus, we assigned all employment to the ﬁrm’s
main MSA even though this is not strictly accurate. We examined the sensitivity
of our results to the inclusion of these multi-MSA ﬁrms, and our results are not
substantively dependent on whether they are included in the sample.
10
We chose the May 1988 CPS for comparison because it included questions
about the number of employees in establishments where respondents work.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
237The Black/White Wage Gap
Table 1
Sample Characteristics of Workers
Sample
1990 Decennial
Census
Manufacturing WECD May 1988 CPS CBO
White Black White Black White Black White Black
Characteristic (1) (2) (3) (4) (5) (6) (7) (8)
Population share 92.2 7.8 93.4 6.6 92.0 8.0 72.0 28.0
Age 39.2 38.6 40.3 40.0 36.5 35.6 . . . . . .
Education:
Less than high school 18.6 28.3 17.0 24.6 14.6 25.8 . . . . . .
High school diploma 39.0 41.9 42.9 42.9 44.0 45.6 . . . . . .
Some college 26.1 23.3 25.7 25.7 21.7 18.5 . . . . . .
College 12.3 5.22 11.1 5.6 12.8 7.5 . . . . . .
Advanced degree 4.0 1.37 3.3 1.3 6.8 2.5 . . . . . .
Experience 21.0 21.1 22.2 22.4 17.7 17.6 . . . . . .
% Women 30.7 41.5 27.6 37.9 45.6 51.8 . . . . . .
Region:
Northeast 21.5 11.7 28.6 13.7 23.2 14.0 11.3 8.5
South 28.7 64.0 19.7 50.4 30.3 62.7 66.7 73.1
Midwest 35.4 18.9 45.8 33.8 29.0 17.8 22.0 18.4
West 14.4 5.5 5.9 2.1 17.4 5.4 .0 .0
Occupational shares:
Professionals/
technicans/
managers 19.6 6.7 17.7 6.1 22.3 10.7 . . . . . .
Sales and service 7.6 3.8 5.8 3.3 16.2 10.9 . . . . . .
Clerical 13.7 12.9 13.4 11.1 28.6 36.2 . . . . . .
Craftsmen 20.2 16.2 21.2 17.1 13.9 10.2 . . . . . .
Operatives 33.9 51.9 35.9 53.8 14.1 23.6 . . . . . .
Laborers 5.1 8.5 6.0 8.7 4.8 8.4 . . . . . .
Establishment size:
1–9 . . . . . . 2.2 1.4 22.5 15.5 13.3 12.7
10–24 . . . . . . 3.8 2.2 15.0 11.4 18.3 15.9
25–49 . . . . . . 4.9 3.3 12.8 11.8 16.4 16.1
50–99 . . . . . . 9.1 7.3 10.9 10.4 15.1 17.5
100–249 . . . . . . 17.8 14.5 13.1 14.3 18.2 20.7
250/ . . . . . . 62.2 71.3 25.7 36.6 18.7 17.2
Hourly wages 12.86 10.24 13.56 11.96 9.93 7.56 . . . . . .
Log(hourly wages) 2.40 2.19 2.48 2.35 2.08 1.82 . . . . . .
Annual earnings 28,275 21,396 29,863 25,124 19,897 14,121 14,283 12,927
Log(annual earnings) 10.07 9.82 10.16 9.99 9.53 9.18 9.29 9.21
NOTE.—Results for cols. 1–2 are based on a sample of manufacturing workers drawn from the Sample
Detail File. Results for cols. 3–4 are based on all workers in the Worker-Establishment Characteristics
Database (WECD), who are by construction employed in the manufacturing industry. Results for cols.
5–6 are based on workers in the May 1988 Current Population Survey (CPS), with no restriction on
industry. For each of these samples, workers were included only if they had a reasonable degree of labor
force attachment. Finally, results for cols. 7–8 are based on our sample from the Characteristics of
Business Owners (CBO) database. Since we have very crude information on workers in the CBO, we
could not apply any restrictions associated with labor force attachment.
sample to locales with large black populations. Second, these geographic
restrictions have led to a sample overrepresented in the South and in
certain cities in the Northeast and Midwest. Third, CBO ﬁrms are small,
as less than 20% of CBO workers are employed at ﬁrms with more than
250 employees. Finally, CBO earnings are lower than in the rest of the
economy, partly because of the focus on the South but also because wages
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
238 Carrington/Troske
and earnings are lower in the small ﬁrms overrepresented in the CBO.
The lack of information on workers’ age and education also illustrates an
important weakness of the CBO; all we really know about the workers
is their race and sex.
Table 2 presents summary statistics on the ﬁrms in our two samples.
The left side of the table presents data on our WECD sample of manufacturing
plants. The average plant in our sample employs roughly 155
workers and ships about $130,000 worth of goods per worker per year.
Note that while the average establishment represented in the WECD is
large, the number of workers per establishment in our sample is only
about 12. The incomplete sampling within establishments occurs because
workers had to both answer the long form of the 1990 Decennial Census
and give accurate information about their employer’s address. The right
side of table 2 presents information on the ﬁrms in our CBO sample.
The table reports ﬁgures for all ﬁrms and reports them separately for
black- and white-owned businesses. Rows 1 and 2 show that the CBO
ﬁrms are obviously much smaller than our WECD ﬁrms. The rows also
show that the white-owned businesses are larger than those owned by
Table 2
Characteristics of Firms and Owners
WECD CBO
All Black- WhiteCharacteristic
M Characteristic Firms Owned Owned
1. Total workers in plant 155.3 1. Workers in ﬁrm 6.4 4.8 6.4
2. Sample workers in plant 12.0 2. Receipts/worker 77,196 57,776 77,790
3. Shipments/worker 131,606 3. Average earnings ($) 11,271 8,826 11,346
4. Annual earnings/worker ($) 24,542 4. Education of owner
5. % multiunit 46 (in %):
6. % in MSA 88 Elementary school 4.6 10.3 4.4
7. Region (in %): Some high school 10.7 15.5 10.6
Northeast 32.9 High school/General
South 22.9 Equivalency Degree 29.3 25.9 29.4
Midwest 36.2 Some college 19.9 17.4 20.0
West 8.1 College graduate 14.6 11.5 14.7
Graduate school 20.8 19.4 20.9
5. Age of owner (in %):
Under 25 .5 .8 .4
25–34 11.7 9.0 11.8
35–44 32.2 26.1 32.4
45–54 26.1 28.5 26.0
55–64 21.7 24.0 21.6
65/ 7.8 11.6 7.7
6. Region (in %): 7.1
Northeast 12.8 13.0
South 60.8 75.0 60.4
Midwest 26.4 17.9 26.7
West .0 .0 .0
7. % in MSA 70.4 72.9 70.3
8. Number of owners 1.1 1.1 1.1
NOTE.—The unit of observation in this table is an establishment (in the WECD) or a ﬁrm (in the
CBO). MSA Å metropolitan statistical area.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
239The Black/White Wage Gap
blacks. The table also reveals that we know the education, age, race, and
sex of CBO owners. Since CBO owners presumably make hiring decisions,
this information is useful in assessing the role of employer discrimination
in generating interﬁrm segregation.
III. On Measuring Segregation
This section considers a methodological question that we need to address
before we can adequately analyze the data discussed in Section II.
Economists typically summarize segregation patterns with a segregation
index, which is simply a statistic that summarizes the extent to which
two groups interact in a sample. Although there are many such indices,
our work focusses on the Gini coefﬁcient of segregation because of its
desirable properties (Hutchens 1991). This index measures the extent to
which the distribution of blacks and whites across ﬁrms deviates from
an even distribution in which each group is proportionately represented
in each ﬁrm. To present this more formally, let there be T ﬁrms, let sbi
and swi be ﬁrm i’s share of the black and white sample populations,
respectively, and let the ﬁrms be sorted in ascending order of sbi /swi .
Then the Gini coefﬁcient is
G Å 1 0 ∑
T
iÅ1
(sbi )ͩswi / 2 ∑
T
jÅi/1
swjͪ. (1)
The Gini coefﬁcient varies between zero and one, with zero corresponding
to complete evenness and one corresponding to complete unevenness,
which occurs when blacks and whites never work for the same em-
ployer.11
An often overlooked feature of the Gini coefﬁcient of segregation (and
other popular indices) is that it can be positive when workers are allocated
randomly across ﬁrms.12
This occurs for two reasons. First, there is a
simple integer constraint in that each worker must be uniquely allocated
to one ﬁrm. In a sample with 10 black workers and 20 ﬁrms, for example,
evenness is unobtainable because it is impossible for each ﬁrm to get half
a black worker. Second, and more important, the random allocation of
11
If every ﬁrm has exactly the same share of black workers, say 10%, then the
sample is completely even and G Å 0. In contrast, if every ﬁrm is either all white
or all black, then the sample is completely uneven, and G Å 1. The Gini coefﬁcient
of segregation has the same geometric interpretation as the Gini coefﬁcient of
income inequality, as the indices are analytically equivalent.
12
This discussion is drawn from Carrington and Troske (1997).
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
240 Carrington/Troske
workers to ﬁrms typically generates some deviation from evenness.13
It
is easy to show that such random unevenness leads to positive and sometimes
large Gini coefﬁcients, particularly in samples with small ﬁrms and
small minority shares. This is a problem in our study because the average
ﬁrm has 6.4 workers in the CBO and 12.0 workers in the WECD and
because blacks are a small minority, particularly in the WECD. Thus,
random allocation would lead to Gini coefﬁcients on the order of .65 for
the WECD and .40 for the CBO.
This is unfortunate because we would like to use the term ‘‘segregated’’
only in instances where the distribution of workers across ﬁrms
is more uneven than would be implied by random allocation.14
Carrington
and Troske (1997) propose the following modiﬁcations of the Gini
coefﬁcient as a means of distinguishing between systematic and random
segregation. Let the Gini coefﬁcient of random segregation G* be the
Gini coefﬁcient that would occur if a very large number of workers
were allocated randomly to employers, taking the black population
share and the size distribution of plants as determined by the sample.15
Put slightly differently, G* is the average Gini coefﬁcient obtained if
sample workers are assigned randomly to sample plants. Thus, if a
sample contains mostly large plants and roughly equal numbers of
blacks and whites, then G* will be close to zero. In contrast, if the
sample contains mostly small plants and is either mostly black or
mostly white, then G* will be closer to one.
We use the Gini coefﬁcient of random segregation to adjust the standard
Gini coefﬁcient so that it accounts for the role of random assignment
13
As a stark example, consider a large sample of two-person ﬁrms that, in
aggregate, employ a 50/50 mix of blacks and whites. Random allocation of workers
to ﬁrms will result in 25% of the ﬁrms employing two blacks, 50% of the
ﬁrms employing one black and one white, and 25% of the ﬁrms employing two
whites. The Gini coefﬁcient of segregation would be .75 in this instance.
14
Our interest in separating systematic from random effects, as well as our ﬁx
to the problem, are similar in spirit to Ellison and Glaeser’s (1997) work on
indices of geographical concentration.
15
See Carrington and Troske (1997) for a complete discussion of this issue.
The basic idea behind the computation of G* is as follows. The ﬁrst step is to
calculate the empirical number of ﬁrms in each size class s, gˆ(s). Within any size
class s, random allocation implies that the binomial function B(m; s, p) is the
fraction of ﬁrms that will have m blacks if p is the black population share. Thus,
random allocation implies that the number of sample units with size s and m
blacks should be N(m, s; p) Å B(m; s, p)gˆ(s). Thus, the support of N(m, s; p)
is m Å 0 to s for every s in the support of gˆ(s). This artiﬁcial distribution
corresponds to what would be expected if workers were allocated randomly,
given p and gˆ(s). The Gini coefﬁcient of segregation computed from this artiﬁcial
distribution is what we call the Gini coefﬁcient of random segregation.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
241The Black/White Wage Gap
in causing unevenness. In particular, we deﬁne the Gini coefﬁcient of
systematic segregation as
GO Å
G 0 G*
1 0 G*
if G 0 G* ¢ 0
and
GO Å
G 0 G*
G*
if G 0 G* õ 0.
If the sample is more uneven than random allocation would imply—that
is, G ú G*—then Gˆ ú 0 is simply excess unevenness (G 0 G*) expressed
as a fraction of the maximum amount of such excess segregation that
could possibly occur (1 0 G*). When Gˆ Å 1, it is analogous to complete
unevenness, as with the standard gini coefﬁcient, but Gˆ Å 0 implies that
the sample is equivalent to random allocation. If, in contrast, there is
excess evenness—that is, G õ G*—then Gˆ is negative and represents
excess evenness in the sample (G 0 G*), expressed as a fraction of
the maximum amount of such excess evenness that could possibly
occur (G*).16
To summarize, our Gˆ is different from the standard Gini coefﬁcient
in two ways. First, we have set the baseline of zero to correspond to
random allocation rather than complete evenness. Second, we have remapped
values of G that are greater than G* into the [0, 1] interval and
remapped values of G that are less than G* into the [01, 0] interval. We
think that this modiﬁed index provides more useful information than the
traditional Gini coefﬁcient. However, we recognize that some readers
will be more comfortable with the traditional index. Thus, in the work
that follows we report traditional indices of segregation, indices of random
segregation, and the indices of systematic segregation that we developed
here. Together, these indices provide a useful summary of segregation
patterns.
16
The example from n. 13 above illustrates the difference between G and Gˆ .
First, consider a large sample of two-person plants with equal aggregate numbers
of men and women and a traditional Gini coefﬁcient of .75. In this case, Gˆ Å 0
because .75 is exactly what random assignment would imply. Second, consider a
large sample of 1,000-worker plants with equal aggregate numbers of men and
women and a traditional Gini coefﬁcient of .30. In this case, Gˆ would be very
close to .30 as well, as random assignment implies very little unevenness. As these
examples illustrate, the extent to which Gˆ differs from G is entirely dependent
on the likely role of random allocation in the particular sample at hand.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
242 Carrington/Troske
IV. Interﬁrm Racial Segregation
Table 3 begins this section’s analysis of interﬁrm racial segregation
with results from both the WECD (rows 1–13) and the CBO (rows 14–
20). Rows 1 and 14 report Gini coefﬁcients for the entire WECD and
CBO samples, respectively, while the other rows present results for subsets
of the data.17
For example, rows 2 and 3 analyze the WECD data when
the manufacturing industry is broken into durables and nondurables,
and rows 4–9 analyze the WECD data when workers are stratiﬁed by
occupation. Note that while both the CBO and the WECD allow separate
analyses by industry, only the WECD admits to separate analyses by
workers’ occupation. This is a result of the relatively crude demographic
information available in the CBO.
While the rows of table 3 vary by data set and by samples deﬁned by
characteristics of workers or ﬁrms, the columns of table 3 vary on two
separate dimensions. First, columns 1–3 differ from columns 4–6 in the
geographic sampling scheme under consideration. (Table 4 considers two
additional ways of geographically organizing the data.) Second, within
any geographic scheme, we report estimates of the traditional Gini coefﬁcient,
the Gini coefﬁcient of random segregation, and the Gini coefﬁcient
of systematic segregation. In addition to the index values, bootstrap standard
errors are reported in parentheses to the right of each index value.18
Thus, if we look at row 1 of column 1, we see that the Gini coefﬁcient
for all workers in the WECD is .78 with a bootstrap standard error
of .01.19
There is a range of views on the causal relationship between residential
segregation and workplace segregation. At one extreme, one might view
all residential decisions as the outcome of employment opportunities, in
which case residential segregation is entirely due to employment segregation.
At the opposite extreme, one might view residential decisions as
completely exogenous to the labor market, in which case the causality
runs entirely from residential segregation to employment segregation. We
17
We have conducted an analogous set of calculations using the dissimilarity
index (Duncan and Duncan 1955). In spite of its problems (Hutchens 1991),
the dissimilarity index remains popular because of its intuitive interpretation. In
particular, the dissimilarity index reports the share of black (or white) workers
that would have to switch ﬁrms in order for the sample to be completely even.
The results of our analysis using the dissimilarity index are broadly similar to
those reported here, but full results are available from us on request.
18
Boisso et al. (1994) have shown that bootstrapping provides a useful measure
of sampling variation in this context.
19
We also tested the hypothesis of random allocation for each of our samples
with the x2
test proposed by Blau (1977). Most entries in tables 3 and 4 represent
a statistically signiﬁcant departure from random allocation, although we will argue
that the differences are not always economically signiﬁcant.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
243The Black/White Wage Gap
Table 3
Gini Coefﬁcients of Interﬁrm Racial Segregation
Geographic Scheme
National Within MSAs
Standard Random Systematic Standard Random Systematic
Gini Gini Gini Gini Gini Gini
(1) (2) (3) (4) (5) (6)
A. WECD:
1. All workers .78 (.01) .48 (.01) .57 (.01) .60 (.02) .56 (.02) .09 (.03)
Within industry:
2. Nondurable
manufacturing .81 (.01) .51 (.01) .61 (.02) .63 (.02) .62 (.01) .06 (.02)
3. Durable manufacturing .74 (.01) .46 (.02) .52 (.02) .53 (.02) .51 (.02) .02 (.03)
Within occupation:
4. Professionals/
technicans/
managers .86 (.03) .86 (.02) 0.03 (.06) .73 (.05) .80 (.05) 0.13 (.07)
5. Sales and service .93 (.01) .92 (.01) .06 (.06) .82 (.03) .89 (.02) 0.05 (.05)
6. Clerical .89 (.01) .86 (.01) .20 (.03) .79 (.02) .85 (.01) 0.06 (.02)
7. Craftsmen .85 (.01) .73 (.01) .45 (.02) .71 (.02) .74 (.02) 0.00 (.03)
8. Operatives .84 (.01) .58 (.01) .62 (.02) .68 (.02) .66 (.01) .11 (.03)
9. Laborers .90 (.01) .75 (.01) .60 (.03) .76 (.01) .78 (.01) .08 (.03)
Within plant size group:
10. 15 or fewer employees .97 (.00) .97 (.00) .23 (.07) .95 (.01) .97 (.00) .15 (.04)
11. Between 16 and 50
employees .95 (.01) .91 (.01) .41 (.05) .90 (.01) .92 (.01) .13 (.04)
12. Between 51 and 125
employees .90 (.01) .76 (.01) .56 (.02) .82 (.01) .82 (.01) .15 (.02)
13. More than 125
employees .72 (.02) .31 (.01) .59 (.02) .47 (.02) .41 (.01) .06 (.04)
B. CBO:
14. All workers .87 (.01) .25 (.01) .83 (.01) .81 (.02) .28 (.01) .74 (.02)
Within industry:
15. Nondurable
manufacturing .76 (.04) .13 (.02) .72 (.04) .48 (.07) .28 (.06) .16 (.13)
16. Durable manufacturing .86 (.02) .21 (.01) .82 (.02) .54 (.05) .30 (.04) .30 (.09)
17. Construction .88 (.02) .26 (.01) .84 (.02) .70 (.05) .33 (.04) .54 (.07)
18. Wholesale trade .81 (.03) .29 (.02) .74 (.04) .63 (.04) .34 (.02) .42 (.12)
19. Retail trade .86 (.02) .28 (.01) .81 (.03) .68 (.04) .35 (.03) .47 (.07)
20. Services .90 (.01) .23 (.01) .87 (.01) .76 (.04) .39 (.04) .62 (.05)
NOTE.—The geographic schemes vary as follows. The national measures of segregation presented in
cols. 1–3 are computed for the entire United States at a single step. The within-MSA measures of cols.
4–6 are the employment-weighted average of the indices for each MSA. Note that (3) Å col. 3 Å (col.
1 0 col. 2)/(1 0 col. 2) by deﬁnition. However, col. 6 Å/ (col. 4 0 col. 5)/(1 0 col. 5) because each of
the three indices is the average across MSAs, and the relationship has to hold only within MSAs, and
not for the averages. Numbers in parentheses are bootstrap standard errors. See text for description of
the Gini coefﬁcient of segregation, the expected Gini coefﬁcient of segregation, and the systematic Gini
coefﬁcient of segregation. Also see the text for sample descriptions. In rows 10–13, plants are classiﬁed
by their total employment, not by the sample of their employees that are in our data.
will examine both of these extremes, as well as some intermediate views.
As a ﬁrst step, columns 1–3 of table 3 adopt the ‘‘employment segregation
causes residential segregation’’ view by studying interﬁrm segregation at
a national level. The traditional Gini coefﬁcients of column 1 suggest that
there is substantial interﬁrm segregation of black and white workers in
the United States, as none of the Gini coefﬁcients are less than .72. For
example, the Gini coefﬁcients for all workers are .78 in the WECD and
.87 in the CBO, and both are precisely measured. One explanation for
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
244 Carrington/Troske
Table 4
Gini Coefﬁcients of Interﬁrm Racial Segregation
Geographic Scheme
Within Relatively Integrated
MSAs Within Small MSAs
Expected Systematic Expected Systematic
Gini Gini Gini Gini Gini Gini
(1) (2) (3) (4) (5) (6)
A. WECD:
1. All workers .60 (.05) .63 (.05) 0.07 (.06) .59 (.03) .62 (.03) 0.03 (.04)
Within industry:
2. Nondurable
manufacturing .64 (.03) .68 (.03) 0.08 (.03) .58 (.04) .63 (.03) 0.13 (.05)
3. Durable manufacturing .48 (.06) .58 (.06) 0.26 (.09) .52 (.04) .60 (.03) 0.17 (.05)
Within occupation:
4. Professionals/
technicians/managers .63 (.16) .68 (.16) 0.26 (.20) .83 (.03) .91 (.02) 0.22 (.08)
5. Sales and service .74 (.08) .83 (.06) 0.27 (.13) .89 (.03) .95 (.01) 0.25 (.09)
6. Clerical .77 (.07) .83 (.07) 0.20 (.09) .82 (.03) .89 (.02) 0.09 (.06)
7. Craftsmen .80 (.04) .84 (.03) .01 (.09) .73 (.04) .79 (.03) 0.10 (.08)
8. Operatives .74 (.02) .76 (.02) .05 (.05) .69 (.02) .71 (.02) .03 (.05)
9. Laborers .85 (.02) .88 (.02) .11 (.12) .75 (.04) .81 (.03) 0.05 (.08)
Within plant size group:
10. 15 or fewer employees .96 (.01) .98 (.01) .14 (.12) .96 (.01) .97 (.01) .03 (.18)
11. Between 16 and 50
employees .91 (.02) .93 (.04) 0.01 (.08) .90 (.03) .93 (.02) 0.02 (.11)
12. Between 51 and 125
employees .81 (.02) .30 (.02) 0.10 (.02) .81 (.02) .86 (.01) 0.12 (.05)
13. More than 125
employees .49 (.05) .52 (.05) 0.15 (.08) .49 (.03) .52 (.03) 0.08 (.05)
B. CBO:
14. All workers .70 (.07) .28 (.03) .58 (.12) .67 (.06) .38 (.06) .51 (.09)
Within industry:
15. Nondurable
manufacturing .22 (.14) .44 (.09) 0.79 (.16) .24 (.13) .42 (.10) 0.91 (.07)
16. Durable manufacturing .37 (.15) .47 (.08) 0.27 (.32) .24 (.18) .22 (.18) .00 (.34)
17. Construction .73 (.14) .31 (.08) .49 (.28) .26 (.13) .57 (.08) 0.68 (.20)
18. Wholesale trade .42 (.09) .31 (.05) .01 (.23) .40 (.07) .35 (.06) 0.10 (.22)
19. Retail trade .38 (.08) .32 (.07) 0.04 (.16) .52 (.10) .52 (.10) 0.09 (.17)
20. Services .80 (.05) .46 (.07) .60 (.13) .68 (.12) .64 (.09) .19 (.25)
NOTE.—To be included in col. 1, MSAs must be below the median of the residential Gini coefﬁcient
distribution or have a residential Gini coefﬁcient of .712 or less. To be included in col. 2, MSAs must
be below the median of the MSA population distribution, or have a population of 255,301 or less.
Numbers in parentheses are bootstrap standard errors. See text for description of the Gini coefﬁcient
of segregation, the expected Gini coefﬁcient of segregation, and the systematic Gini coefﬁcient of segregation.
Also see the text for sample descriptions.
this pattern is that blacks and whites have different skills and that workers
of all races are segregated by skill (Kremer and Maskin 1994). However,
the other rows of column 1 show that these indices are not much reduced
when we look at relatively homogeneous groups of workers. For example,
row 9 shows that the national Gini coefﬁcient among laborers is .90 in
the WECD, even higher than the Gini coefﬁcient for all workers. Similarly,
the Gini coefﬁcient among construction-industry workers is .88 in
the CBO, which is again even higher than that observed among all workers.
Thus, by this traditional measure, the U.S. workforce is quite seg-
regated.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
245The Black/White Wage Gap
Are these patterns evidence of systematic sorting of blacks and whites
to different employers? Column 2 of table 3 addresses this question by
reporting the Gini coefﬁcient of random segregation, which again is the
Gini coefﬁcient that would arise if workers were randomly allocated to
employers. Inspection shows that these random Gini coefﬁcients are often
quite high. For example, the random Gini coefﬁcient for the entire sample
is .48 for the WECD and .25 for the CBO, and generally even higher in
the subsamples.20
This suggests that much of the unevenness measured
by the Gini coefﬁcient is potentially attributable to random allocation
rather than to systematic forces such as employer discrimination. However,
the systematic Gini coefﬁcients of column 3 show that there is an
important systematic component to segregation in the United States. For
example, the systematic Gini coefﬁcient of .57 in row 1 implies that actual
excess unevenness (G 0 G*) is 57% of the maximum that could possibly
be observed (1 0 G*). Similarly, the systematic Gini coefﬁcient is substantial
within each industry of the WECD and the CBO and for most
of the other subgroups. Yet there are some subgroups in which the systematic
Gini coefﬁcient is not substantially different from zero. For example,
the systematic Gini coefﬁcient in the WECD is 0.03 for professionals,
technicians, and managers and .06 for sales and service occupations. Thus,
for these subsamples, the traditional Gini coefﬁcient is quite close to what
would be implied by random allocation. These results are consistent with
the notion that the market for such skilled labor is relatively nationalized
and, therefore, less likely to reﬂect historical racial differences in residential
patterns.
Much of the segregation measured in columns 1–3 of table 3 is due to
the dissimilar distributions of blacks and whites across states and MSAs.
For example, blacks are relatively likely to live in the South and in the
central cities of the North and relatively unlikely to live in states like
Colorado, Utah, and Montana. To the extent that such patterns reﬂect
historical and cultural factors not directly related to the contemporary
labor market, columns 1–3 do not tell us much about current labor
20
The role of random allocation varies across samples for two reasons. First,
random allocation generates more unevenness when the black sample share is
small. This leads CBO indices to be smaller than those in the WECD. Second,
the role of random allocation varies with the number of sample employees per
ﬁrm, which varies considerably across samples. In particular, note that per-ﬁrm
sample sizes are much reduced when we restrict attention to the WECD subsamples
stratiﬁed by worker characteristics. This causes the Gini coefﬁcients of random
segregation for these samples to be higher than those observed in the entire
data set. It may seem odd that the expected Gini coefﬁcient has a standard error,
but this too is stochastic because the expected Gini coefﬁcient takes the empirical
ﬁrm size distribution as ﬁxed, and this distribution varies across bootstrap replica-
tions.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
246 Carrington/Troske
market segregation. Therefore, columns 4–6 of table 3 take a more reﬁned
geographic approach by measuring segregation within MSAs. In particular,
we computed our three Gini coefﬁcients separately for each MSA
and then reported the mean of these indices across MSAs.21
While this
measure may still reﬂect within-MSA residential segregation, it is not
affected by the uneven distribution of blacks and whites across MSAs.
The traditional Gini coefﬁcients of column 4 of table 3 show that
there is substantial unevenness within the typical MSA. For example, the
average within-MSA Gini coefﬁcient among all workers is .60 in the
WECD and .81 in the CBO. However, it is not always clear that this
within-MSA unevenness represents a substantial departure from random
allocation. For example, column 5 of row 1 shows that the within-MSA
Gini coefﬁcient of random segregation is .56 in the WECD as a whole,
which implies that observed unevenness within MSAs is only marginally
greater than that implied by random allocation. In fact, in none of the
rows of panel A is workplace segregation substantially greater than that
predicted by random allocation. Thus, there is little evidence in the
WECD of any important systematic component to within-MSA segregation.
In contrast, there is substantially more unevenness than random
allocation would imply in the CBO. Row 14 shows that in these data the
traditional Gini coefﬁcient is .81 within MSAs, while random allocation
implies an index of .28.
The other rows of columns 4–6 analyze racial segregation within relatively
homogenous groups of workers. These specialized results are analogous
to the general results of rows 1 and 14: there is systematic segregation
within-MSAs in the CBO but not in the WECD. In particular, there are
several groups for which there is signiﬁcantly less within-MSA segregation
in the WECD than random allocation would predict. For example, the
‘‘sales and service’’ occupation within manufacturing has a Gini coefﬁcient
of .82, whereas random allocation predicts a Gini coefﬁcient of .89.
In contrast, the within-industry CBO estimates of rows 15–20 show that
there is signiﬁcantly more unevenness than random allocation would
imply in these data, particularly in the service and trade industries.
Why do we ﬁnd more systematic segregation in the CBO than in the
WECD? The difference is partly driven by the CBO’s broader industrial
coverage, as the systematic Gini coefﬁcients for CBO manufacturing industries
are substantially lower than those of other CBO industries. The
difference is also driven by the greater size of the WECD plants. Rows
10–13 of table 3 show that there is more systematic segregation among
smaller plants in the WECD, and these plants are closer in size to those
21
Note that the move to an MSA-based sample excludes the roughly 28% of
whites and 17% of blacks who do not live in an identiﬁable MSA.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
247The Black/White Wage Gap
typical of the CBO. Thus, the comparisons both across data sets and
within the WECD suggest that large ﬁrms are less systematically segregated.
Finally, the difference may also be driven by the greater frequency
of family-owned ﬁrms in the CBO, as owners of such businesses are
probably more likely to hire members of the same ethnic group. In contrast,
the difference between the CBO and the WECD is apparently not
driven by the CBO sample’s focus on the South. We base this conclusion
on within-MSA indices calculated separately by region in the WECD.22
These ﬁgures show that segregation is, if anything, less systematic in the
South than in the rest of the country.23
The WECD results of table 3 show that, once we restrict attention to
within-MSA patterns, black and white workers are not systematically
segregated. Nevertheless, the CBO results suggest that perhaps there is
some systematic component to these patterns. To what extent is this
systematic interﬁrm segregation the outcome of ‘‘spatial mismatch,’’
which occurs when (a) blacks and whites are residentially segregated
within MSAs and (b) all workers tend to ﬁnd jobs near their home
(Wilson 1987)? Table 4 addresses this question by examining segregation
within MSAs where spatial mismatch is likely to be a minor factor. In
particular, for spatial mismatch of black workers and jobs to explain
workplace racial segregation, it must be that blacks and whites are residentially
segregated. Thus, looking at MSAs where there is little residential
segregation is one way to reduce the effect of within-MSA spatial
mismatch.
In this spirit, columns 1–3 of table 4 repeat the within-MSA analysis
of table 3 for a sample of MSAs that are relatively residentially integrated.
In particular, we used the residential Gini coefﬁcients of Harrison and
Weinberg (1992) to restrict the sample to MSAs below the median of the
distribution of MSA (residential) Gini coefﬁcients. The results for the
WECD are similar to those of table 4, as there is little evidence that
within-MSA segregation is more pervasive than random allocation would
imply. If anything, the WECD results suggest that there is less segregation
than random allocation would predict. In contrast, the CBO results differ
slightly from table 3. It is still true that the traditional Gini coefﬁcient
for all workers in the CBO (.70) is higher than that implied by random
allocation (.28). However, when we restrict attention to industries where
22
The average within-MSA systematic Gini coefﬁcients were .13 for the Northeast,
.13 for the Midwest, .06 for the South, and 0.29 for the West.
23
We are aware of one factor that works in the opposite direction. Since the
WECD is a sample of establishments while the CBO is a sample of ﬁrms, establishment
level segregation within ﬁrms will be missed in the CBO but not in the
WECD. This factor appears to be overwhelmed by the other forces creating more
segregation in the CBO.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
248 Carrington/Troske
workers are likely to be relatively homogenous, there is no longer much
evidence of systematic segregation. In fact, there is some evidence that
the distribution of workers is less uneven than randomness would imply.
For example, among workers in nondurable manufacturing, the average
Gini coefﬁcient within MSAs is .22, while random allocation implies a
Gini coefﬁcient of .44. Only in the construction and service industries
are black and white workers systematically segregated in an important
way.24
Of course, even ‘‘relatively residentially integrated MSAs’’ are rather
segregated. For example, Bloomington, Indiana, is a relatively integrated
city, but its residential Gini coefﬁcient is still .342. Therefore, an alternate
approach to minimizing the role of spatial mismatch is to measure segregation
in relatively small MSAs. While living on one side of the Los Angeles
MSA may preclude holding a job on the other side, there are smaller
MSAs where even a cross-town commute takes only a few minutes. This
reasoning suggests that spatial mismatch is less likely to explain workplace
segregation in small MSAs. Therefore, columns 4–6 of table 4 repeat the
within-MSA analysis for MSAs below the median of the MSA population
distribution. The results are quite similar to those of columns 1–3 of the
same table. Namely, there is some evidence of systematic segregation
among all workers in the CBO. For all other samples from both data
sets, however, there is no evidence that black and white workers are
distributed across employers more unevenly than would be suggested by
random allocation.25
The results of this section suggest the following view. The national
distribution of black and white employees across employers is far from
even, as some employers have predominantly white workforces while
others are predominantly black. However, systematic national segregation
is largely due to black/white differences in MSA residence. When we
look within MSAs, there is still substantial interﬁrm segregation by conventional
measures, but this is mostly explained by racial differences in
occupations and industry and by simple random allocation. Of the modest
amount of within-MSA segregation that remains unexplained by random
24
We should note that the samples of both data sets become smaller as we
restrict the sample to particular MSAs or to particular groups of workers. Thus,
our power to distinguish between alternative hypotheses is reduced in these subsamples,
as reﬂected in the increased standard errors.
25
We also made an attempt to explain the cross-MSA variation in interﬁrm
segregation. In particular, we regressed both the standard and the systematic Gini
coefﬁcients on log MSA population and on the residential Gini coefﬁcient. We
did this for both the CBO and the WECD. Few consistent patterns emerged,
although MSAs with large residiential Gini coefﬁcients tended to have larger
systematic Gini coefﬁcients of interﬁrm segregation. This provides some support
for the spatial mismatch hypothesis.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
249The Black/White Wage Gap
allocation and group skill differences, much appears to be attributable to
spatial mismatch.26
However, it is worth noting that while the data are
consistent with random allocation in many ways, it is entirely possible
that there are strong systematic forces (e.g., discrimination) at work. All
we can really say is that the data do not radically reject the hypothesis
of random allocation.
These results bear comparison with the earlier work of Flanagan
(1973), Higgs (1977), and Becker (1980). Higgs (1977) found substantial
racial segregation in a sample of Virginia ﬁrms that were surveyed in 1900
and 1909. However, data constraints prevented Higgs from assessing the
importance of intrastate residential segregation, and Higgs’s analysis did
not reduce patterns of segregation into a single summary index. These
facts make it difﬁcult to compare his results directly with those presented
here. Flanagan (1973) studied segregation within a sample of Chicago
ﬁrms surveyed in 1967 as part of the Equal Employment Opportunity
Commission’s (EEOC) monitoring process. As with the Higgs study,
methodological differences preclude a numerical comparison with our
study. However, Flanagan notes that ‘‘observed segregation of blacks is
closer to the random (allocation model) than to the general equilibrium
prediction of the utility analysis’’ (p. 468); that is, a random allocation
model did a reasonable job of explaining the data. Finally, Becker (1980)
analyzed segregation in a nationwide sample of employers drawn from
the 1975 EEOC data. He too ﬁnds substantial segregation, but the results
of his paper are particularly difﬁcult to compare with those presented
here. This is because the data are not disaggregated by MSA, because the
paper used a somewhat exotic index, and because the role of random
allocation was not discussed. In sum, previous studies have found substantial
interﬁrm racial segregation in years past, at least by traditional measures,
but the evidence is mixed as to whether the data were inconsistent
with random allocation. Unfortunately, methodological differences preclude
a more direct quantitative comparison between them and the present
study.
26
Since neither the CBO nor the WECD is a random sample of all employers
or employees, we wondered whether these ﬁgures were representative of the
broader population. The National Survey of Black Americans (NSBA) is of some
assistance in this regard (Jackson and Gurin 1987). Administered in 1979 and
1980, the NSBA asked a national sample of blacks about the fraction of blacks
in their ‘‘workplace,’’ which was probably interpreted as the worker’s establishment
but could also have been interpreted as their ﬁrm. In any case, their answers
were tabulated as follows: all black (17.7%), mostly black (26.3%), about half
black (20.6%), mostly white (23.6%), all white except you (11.7%). While the
qualitative nature of the NSBA question and the lack of ﬁrm size information
preclude a precise comparison, the ﬁgures appear to be closer to those of the
CBO than the WECD.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
250 Carrington/Troske
V. Are Black and White Workers Sorted into Different Types
of Employers?
Section IV showed that the distribution of black and white workers
across ﬁrms is not too far from what would be implied by random allocation.
Yet it is important to recognize that the observed patterns could be
systematic even though they look random. In an effort to look for systematic
patterns, we present in table 5 estimates of a WECD establishmentlevel
regression where the dependent variable is the black share of nonsupervisory
employment in the establishment.27
Column 1 includes as
regressors the share of black supervisors in the establishment, the black
sample share within each MSA, the log of establishment employment,
the average age and education of nonsupervisory employers, and dummy
variables for industry and region. Column 2 adds a set of MSA dummy
variables and takes out the black share of MSA population and the region
dummies. The regressions show that there is a statistically strong relationship
between the race of supervisors and nonsupervisors; black workers
tend to be supervised by black managers. This relationship is nowhere
near one-for-one, however, as managers of one race often supervise workers
of another.
The correlation between the race of managers and the race of nonmanagers
may be explained by residential segregation, so columns 3 and 4
present two crude ways of trying to minimize its role. In particular,
column 3 interacts percent black supervisors with the residential Gini
coefﬁcient in the MSA, and column 4 interacts percent black supervisors
with the log of MSA population.28
Our earlier arguments suggest that if
residential segregation is responsible for the correlation between the race
of supervisors and nonsupervisors, then the coefﬁcient on percent black
supervisors should be bigger in large or segregated MSAs. The interaction
terms in columns 3 and 4 are small and, in the case of column 3, statistically
insigniﬁcant. While obviously not conclusive, these results suggest
that residential segregation is not an attractive explanation for the interplant
connection between black supervisors and black nonsupervisors
in the WECD.
We pursue a similar analysis in the CBO. The CBO records the fraction
of nonminority employees within six brackets (0%–25%, 26%–50%,
27
To be in this sample, an establishment had to have at least one supervisor
and one nonsupervisor in the WECD. This led to a 50% reduction in the number
of plants in our sample, as the smallest establishments seldom met this restriction.
Regressions are weighted by total employment in the plant.
28
Note that this speciﬁcation results in a smaller sample because we throw out
non-MSA establishments. The slight change in the coefﬁcient on percent black
supervisors between cols. 1 and 2 and cols. 3 and 4 is driven by the change in
speciﬁcation rather than the change in sample.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
251The Black/White Wage Gap
Table 5
Plant-Level OLS Models of Employee Racial Composition: WECD
Dependent Variable Å Black Share of Nonsupervisory Employment
Independent Variable (1) (2) (3) (4)
1. Percent black supervisors: .261 .301 .178 .179
(.013) (.014) (.013) (.013)
1 residential Gini coefﬁcient in MSA . . . . . . .001 . . .
(.001)
1 log MSA population . . . . . . . . . .001
(.000)
2. Percent black population in MSA sample .873 . . . 1.056 1.054
(.024) (.024) (.024)
3. Log of establishment employment .003 .005 .003 .003
(.001) (.001) (.001) (.001)
4. 2-digit industry dummies yes yes yes yes
5. Region dummies yes no yes yes
6. MSA dummies no yes no no
7. Number of plants in sample 7,813 7,813 6,562 6,562
8. R2
.386 .426 .404 .405
NOTE.—All data are drawn from the WECD. Standard errors are in parentheses. All regressions
included controls for the average age and education of nonsupervisory employees, industry, and region.
To be in the sample for this table, establishments had to both have at least one supervisor and one
nonsupervisor in the WECD. Regressions are weighted by total employment in the plant.
51%–75%, 76%–90%, 91%–99%, and 100%), making it natural to
apply the ordered probit model.29
Table 6 presents estimates of a series
of such models for the CBO where the dependent variable is the fraction
of black employees in the ﬁrm.30
The speciﬁcation in column 1 includes
the race and sex of the owner, the log of ﬁrm employment, the black
population share of the MSA, and a set of controls for the minority share
of a ﬁrm’s customers. The speciﬁcation indicates that the race of the
owner has a fairly strong inﬂuence on the minority share of the ﬁrm’s
workforce. Given the estimated cut points, the black owner coefﬁcient
is sufﬁcient to move the median predicted response of a ﬁrm from a
10%–24% black workforce to one that is 50%–74% black. The relationship
between race of owner and race of workers is estimated to be particularly
strong for male owners. Row 6’s coefﬁcients for percent minority
customers are relative to the omitted group of ﬁrms that did not report
this information. The results indicate a strong relationship between the
race of a ﬁrm’s customers and the race of its employees, even after we
control for the race of its owner.
The rest of table 6 presents alternative speciﬁcations of the model.
Column 2 adds MSA dummies and takes out the percent black population
29
Regressions are weighted by the product of each observations sampling
weight and its total employment.
30
These results are similar to Bates’s (1988) analysis of the 1982 CBO.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
252 Carrington/Troske
Table 6
Firm-Level Ordered Probit Models of Employee Racial Composition: CBO
Independent Variable (1) (2) (3) (4)
1. Black owner: 1.293 1.320 1.353 1.347
(.137) (.139) (.160) (.160)
1 MSA Gini coefﬁcient . . . . . . .056 . . .
(1.267)
1 log MSA population . . . . . . . . . .098
(.123)
1 female owner 0.508 0.450 0.497 0.488
(.237) (.240) (.271) (.270)
2. Female owner 0.072 0.081 0.200 0.198
(.099) (.103) (.116) (.116)
3. Log(ﬁrm employment) .132 .148 .172 .172
(.012) (.013) (.014) (.014)
4. Black share of MSA sample population 2.992 . . . 1.605 1.646
(.938) (1.015) (1.016)
5. MSA dummies no yes no no
6. Minority customer share:
75%–100% 1.195 1.473 1.152 1.152
(.059) (.063) (.071) (.071)
50%–74% 1.051 1.221 1.063 1.061
(.086) (.089) (.110) (.110)
25%–49% .584 .681 .374 .375
(.058) (.061) (.068) (.068)
10%–24% .240 .308 .126 .126
(.049) (.053) (.057) (.057)
1%–9% 0.207 0.134 0.264 0.264
(.047) (.050) (.055) (.055)
0% 0.332 0.275 0.530 0.530
(.057) (.061) (.069) (.069)
7. Cut Points:
0% black r 1%–9% black 0 0 0 0
1%–9% black r 10%–24% black .321 .341 .342 .342
10%–24% black r 25%–49% black .668 .712 .713 .713
25%–49% black r 50%–74% black 1.109 1.182 1.253 1.253
50%–74% black r 75%–100% black 1.535 1.634 1.701 1.701
8. Number of ﬁrms in sample 6,043 6,040 4,149 4,149
NOTE.—All data are drawn from the 1987 CBO. Standard errors are in parentheses. All regressions
also included controls for log of establishment employment, 2-digit industry, region (except col. 2), and
the age and educational attainment of the business owner. In row 6, the left-out group is those ﬁrms
for whom minority customer share was missing. Regression weights are the product of each observation’s
sample weight and its employment. Regression weights were then scaled to sum to sample size.
in the MSA, which increases the coefﬁcient on black ownership and
slightly strengthens the relationship between the race of customers and
the race of employees. Thus, the racial nexus between owners, employees,
and customers is not solely due to the distribution of blacks and whites
across MSAs. As before, spatial mismatch arguments suggest that the
relationship between the race of owners and customers and the race of
employees is stronger in large, segregated MSAs. Thus, columns 3 and 4
interact the MSA residential Gini coefﬁcient and the log MSA population,
respectively, with the black owner variable. As in table 5, there is little
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
253The Black/White Wage Gap
difference between those MSAs with high and low residential Gini coef-
ﬁcients.
The results of this section suggest that there is a systematic component
to the sorting of black and white workers across establishments in our
two data sets. In the WECD, black nonsupervisors are more likely than
whites to work for black supervisors, and vice versa. In the CBO, black
workers are much more likely to work for ﬁrms with black owners and
customers. If one accepts the view that black supervisors, owners, and
customers are less likely than whites to discriminate against black employees,
then these results are consistent with the hypothesis that tastes for
discrimination cause some systematic workplace segregation. Of course,
it is also possible that these correlations are driven by group differences
in skill or by residential segregation. The available data do not support
these hypotheses, but the analysis is sufﬁciently crude that we can not
rule out their importance.
Section IV found that the distribution of workers to ﬁrms in the WECD
was roughly consistent with random allocation, while this section documents
a modest systematic component to the matching of black and white
workers to particular ﬁrms. There are two ways to reconcile these results.
The ﬁrst is that, although black employers are more likely to hire black
workers, there simply are not enough black employers for this to generate
much systematic segregation. This is particularly true in the WECD,
where the black share of supervisors has an interplant mean of .026 and
an interplant variance of .013. Applying these supervisor shares and the
parameters of table 5 to an initially random distribution of black and
white workers leads to a trivial amount of segregation. In contrast, the
supervisor/nonsupervisor relationship can generate nontrivial segregation
in our CBO sample, as there are a relatively large number of ‘‘black’’
ﬁrms (i.e., those with black owners) in that data set. This difference
between the two databases may be another reason why we ﬁnd the CBO
to be relatively segregated.
The results may also be explained by recognizing that, while employer
tastes may induce systematic segregation, other forces might induce systematic
integration. For example, if black and white labor are complementary
inputs, then ﬁrms might systematically integrate their workforces
even in the presence of discriminatory attitudes. Title VII and afﬁrmative
action provide similar incentives to integrate.31
Thus, the apparent randomness
of the distribution of black and white workers may represent a
31
We explored this idea in the following way. For a subsample of plants in the
WECD and for most ﬁrms in the CBO, we know the fraction of sales made to
the federal government. In these samples, we reestimated the regressions of tables
5 and 6 including the federal government’s share of sales as a regressor. In both
databases, this added coefﬁcient was signiﬁcantly positive but not large.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
254 Carrington/Troske
balance between several systematic forces, each of which is rather weak.32
This brings up an important point, namely that Section IV only demonstrated
that the data were roughly consistent with random allocation.
We reiterate that this may reﬂect a balance between several not-wellunderstood
systematic forces that tend to cancel each other.
VI. Interﬁrm Segregation and the Black/White Wage Gap
This section investigates two aspects of the relationship between interﬁrm
segregation and the black/white wage gap. Our analysis in this
section is based solely on the WECD because it records wages and personal
characteristics for each worker individually. In contrast, the CBO
only records a ﬁrm’s total payroll without providing information on
how wages are distributed to individual workers, making it relatively
unsuitable for the study of black/white wage differences.33
It bears repeating
that the gap between black and white wages is smaller in the
WECD than in the economy at large. For example, the difference in mean
log hourly wages between blacks and whites is .21 among manufacturing
workers in the 1990 Decennial Census, while the analogous gap in the
WECD is only .13. Even bigger differences arise if we compare the
WECD to the nonmanufacturing component of the Decennial Census.34
Thus, the ensuing analysis considers a segment of the economy where
32
The following analogy may help explain our results. Suppose that you observe
black and red checkers strewn across a checkerboard and that you are interested
in whether there is any systematic pattern to their distribution. Initially suppose
that you observe the color of the checkers but not the color of the checkerboard
squares on which they lie and that the distribution of checkers looks roughly
random. With a relatively small number of checkers per square, this is consistent
with some squares having only red checkers and some squares having only black
checkers. Then suppose that you observe the color of the checkerboard squares
as well as the color of the checkers, and that you ﬁnd that the squares with only
red checkers tend to be red themselves, and that the squares with black checkers
tend to be black. This would now suggest some sort of systematic phenomenon.
This situation is analogous to what we ﬁnd in the distribution of workers to
ﬁrms. When we ignore the ‘‘race’’ of the ﬁrms (i.e., the color of the checkerboard
squares), we ﬁnd nothing systematic in the distribution of workers (i.e., the
checkers). But when we look at the matching of workers to ﬁrms (i.e., the
matching of checkers to squares), then systematic patterns emerge.
33
Carrington and Troske (1995) combines information from the CPS and the
CBO to estimate the within- and between-ﬁrm components of the wage gap
between two groups. However, since this procedure requires difﬁcult-to-verify
assumptions on the compatibility of the CBO and the CPS, our analysis here
concentrates on the WECD.
34
The smaller black/white wage gap in the WECD reﬂects the fact that (a)
the WECD contains mostly large manufacturing plants and that (b) the black/
white wage gap is partially accounted for by whites’ greater representation in
such plants.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
255The Black/White Wage Gap
black and white earnings are relatively similar, and our results may not
apply directly to the broader economy. Nevertheless, there is a substantial
black/white wage gap in the WECD, and we believe that a better understanding
of it has implications for the rest of the economy.35
Our ﬁrst exercise decomposes the black/white wage gap into a between-
and within-plant component. In particular, we regress wages on
a set of plant ﬁxed effects before, after, or at the same time that we control
for workers’ personal characteristics in order to see how much of the
black/white wage gap can be explained by the location of black and white
workers in different plants. We do this for all workers at one time, but
we also perform separate analyses for men and women because the size
of the black/white wage gap varies so much by sex. Let Y Å log hourly
wages; X Å a set of personal characteristics including terms in experience,
education, sex, marital status, and dummies for occupation, industry, and
region; and Z Å a set of plant ﬁxed effects. Columns 1–3 of table 7 then
report results from a two-step procedure in which we ﬁrst estimated Y
Å X ؅b and then regressed the residuals of this ﬁrst regression on the
plant ﬁxed effects Z. Column 1 reports results for all workers, while
columns 2 and 3 report results for men and women separately. Row 1
shows that the unadjusted difference in log hourly wages between blacks
and whites is 12.7% for all workers, 12.5% for men, and 0.5% for
women.36
(While we include full results for women, the fact that there is
virtually no black/white wage gap among women leads us to focus our
remaining discussion of table 8 on the results for men.)37
Row 2 of table 7 reports the residual wage gap that is left after we
control for the effect of personal characteristics. This is approximately
2% for all workers and 5% for men, which are somewhat smaller residual
wage gaps than one typically sees in these sort of data. Row 3 reports
the residual black/white wage gap after controlling for the ﬁxed effects
in the second step, without controlling for personal characteristics, and
row 4 reports the residual wage gap after we control for both the personal
35
It bears emphasis, however, that patterns of wage variation in the WECD
are generally not pathological. An appendix table (available from us on request)
presents selected coefﬁcients from a log wage regression estimated on our WECD
sample. The results are generally quite similar to those obtained from a similar
regression estimated on all manufacturing workers in the 1990 Decennial Census.
36
We restricted the sample for this analysis to ﬁrms where we had at least three
workers matched to the establishment. The fact that we imposed this requirement
at each stage of the analysis (e.g., to be in the men’s sample a ﬁrm had to have
three men in our sample) means that the total sample is not the union of the
male and female samples. Thus, there is no requirement that the total wage gap
of col. 1 lie between the men’s and women’s wage gaps of cols. 2 and 3.
37
The rough equality of the wages of black and white women is not unique
to the WECD. Similar results obtain from the Decennial Census or the CPS.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
256 Carrington/Troske
Table 7
Decomposing the Black/White Wage Gap into Within- and Between-Plant
Components
Order of the Decomposition
Step 1: Y Å X؅b / u1 Step 1: Y Å Z؅g / u1
Step 2: Y 0 X؅bO Step 2: Y 0 Z؅gP Step 1: Y Å X؅b
Å Z؅g / u2 Å X؅b / u2 / Z؅g / u2
Total Men Women Total Men Women Total Men Women
(1) (2) (3) (4) (5) (6) (7) (8) (9)
1. Unadjusted black/white
wage gap .127 .125 .005 .127 .125 .005 .127 .125 .005
YU w 0 YU b
2. Black/white wage gap
adjusted for personal
characteristics .019 .049 0.036 .052 .074 0.015 .037 .053 0.021
(YU w 0 XU ؅wbO ) 0 (YU b 0 XU ؅bbO )
3. Black/white wage gap
adjusted for plant ﬁxed
effects .137 .134 .043 .118 .123 .031 .137 .131 .055
(YU w 0 ZU ؅wgP ) 0 (YU b 0 ZU ؅bgP )
4. Black/white wage gap
adjusted for personal
characteristics and plant
ﬁxed effects .029 .058 .002 .043 .063 .011 .045 .059 .027
(YU w 0 XU ؅wbO 0 ZU ؅wgP )
0 (YU b 0 XU ؅bbO 0 ZU ؅bgP )
NOTE.—Y Å hourly wages. X Å worker characteristics including ﬂexible terms in experience and
education, sex (in cols. 1, 4, and 7), marital status, sex 1 marital status, 10 occupation dummies, 4-digit
industry dummies, MSA, region, MSA 1 region. Z Å a set of plant ﬁxed effects. In addition to our
previous restrictions, we required that each individual be in a plant with at least three people in our
sample. This reduced the sample size by approximately 10%. The remaining samples were (a) 172,056
for the total columns, of whom 160,598 were white and 11,458 were black, (b) 121,056 for the male
sample, 114,076 of whom were white and 6,980 of whom were black, and 44,997 for the female sample,
of whom 40,866 were white and 4,131 were black. Observations in the male and female samples do not
equal those of the total sample because we imposed the ‘‘must have three people in the sample’’ restriction
for each sample. Thus, some plants had three individuals without having either three women or three
men in our sample.
characteristics and the plant ﬁxed effects. Rows 3 and 4 suggest that, if
anything, black men tend to work in plants that pay slightly aboveaverage
wages once we control for personal characteristics.38
Columns 4–6 reverse the order of the decomposition by regressing
wages on the plant ﬁxed effects in the ﬁrst step and then regressing the
residuals on the personal characteristics in the second step. In effect, this
gives the plant effects ﬁrst crack at explaining the black/white wage gap.
38
It is worth noting that the plant effects alter black/white wage comparisons
for women much more than they do for men. Black and white women have
similar average wages in the WECD, but this is a combination of two effects:
black women earn somewhat less than white women within any given plant, but
this is compensated for by the fact that black women tend to work in plants with
higher average salaries.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
257The Black/White Wage Gap
Row 3 shows that the black/white wage gap that remains after controlling
for plant is .118 for all workers, or about 93% of what it was without
controlling for these effects. For men, column 5 shows that plant ﬁxed
effects explain virtually none of the black/white wage gap. Columns 7–
9 regress Y on X and Z simultaneously, so that personal characteristics
and plant ﬁxed effects are given equal opportunity to explain the black/
white wage gap. The results are similar to those of the previous columns,
as black men (and women) are sorted into what are, if anything, highpaying
plants.
These results suggest that the black/white wage gap is primarily a
within-plant phenomenon.39
In contrast, very little of the black/white
wage gap in our data is accounted for by the allocation of black and white
workers to ﬁrms that pay systematically different salaries. Most of the
within-plant pay gap is accounted for by observable characteristics, but
a signiﬁcant component remains unaccounted for. At least in the sample
of manufacturing workers studied here, the problem is not that blacks
are not getting jobs at the ‘‘good’’ plants but, rather, that they receive
lower wages on average within any given plant.40
Our second exercise relates the wages of black and white workers to
the racial makeup of their coworkers.41
Table 8 presents estimates of
individual-level hourly wage regressions with personal and plant characteristics
on the right hand side.42
The personal characteristics include
terms in experience and education, sex, marital status, occupation, and
race, and the plant characteristics include total employment and the black
share of establishment employment. Columns 1 and 2 present results
when we combine men and women into one regression. The regressions
differ only in that column 1 includes a plant-level control for labor productivity
(deﬁned as value of shipments/employment), while column 2
does not. The coefﬁcients on most regressors are consistent with similar
data sets, so we will not discuss them. The unique results are in rows 4
39
This result is consistent with the ﬁndings of Hellerstein, Neumark, and
Troske (1996).
40
To the extent that we could, we conducted a similar analysis in the CBO.
There are two problems with the CBO. First, we only know the annual earnings
distributed to the ﬁrm’s employees and not their hourly wages, and second, we
only know aggregate earnings and not how they were distributed among the
employees. Nevertheless, if we assign ﬁrm average earnings to each worker, then
white earnings exceed black earnings by roughly 10% in the CBO. It is impossible
to know whether this cross-ﬁrm racial difference in earnings is due to analogous
differences in wages, and to what extent it is due to black/white differences in
hours worked.
41
See Ragan and Tremblay (1988) for a similar analysis.
42
The standard errors in table 8 have been corrected for heteroscedasticity and
for the clustered sampling design.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
258 Carrington/Troske
Table 8
Individual-Level Models of Wage Determination: WECD
Dependent Variable Å Log Hourly Wage
Sample
All Workers Men Women
Independent Variable (1) (2) (3) (4) (5) (6)
1. Black 0.008 .001 0.006 .003 .051 .056
(.009) (.010) (.010) (.011) (.011) (.011)
2. Female 0.145 0.145 . . . . . . . . . . . .
(.006) (.006)
3. Black 1 female .080 .076 . . . . . . . . . . . .
(.009) (.009)
4. Black share of establishment
employment .158 .197 .177 .223 .116 .136
(.027) (.032) (.035) (.040) (.033) (.035)
5. Black 1 black share of
establishment employment 0.349 0.387 0.370 0.412 0.295 0.319
(.035) (.038) (.046) (.051) (.047) (.048)
6. Log(establishment employment) .062 .062 .058 .058 .071 .072
(.003) (.004) (.004) (.004) (.004) (.004)
7. Labor productivity (1 1,000) .200 . . . .188 . . . .236 . . .
(.027) (.029) (.038)
8. R2
.40 .40 .37 .37 .24 .23
9. Number of observations 168,125 168,125 120,574 120,574 47,551 47,551
NOTE.—Each regression also included controls for 4-digit industry, 1-digit occupation, MSA, region,
MSA 1 region, marital status, marital status 1 sex, ﬁve education categories, a quartic in experience,
and the interaction of the education and experience terms. Standard errors are in parentheses. All data
are drawn from the WECD. Labor productivity is deﬁned as the dollar value of shipments divided by
establishment employment.
and 5, which relate wages to the black share of sample employment within
each worker’s establishment. Since this share is interacted with race, the
direct coefﬁcient in row 4 indicates that wages of white workers are
increasing in the black share of their coworkers. For example, the results
in column 1 imply that a white worker in a plant that is 50% black earns
roughly 8% higher wages than an observationally equivalent worker in
an all-white plant. Columns 3–6 present analogous regressions for men
and women separately. The results are quite similar to those of column
1, although of course the black/white wage gap among women is, if
anything, positive. The wages of white workers increase with the black
share of establishment employment, while the wages of black workers
decrease along the same dimension. How does one reconcile these results
with those of table 7, which demonstrated that the black/white wage gap
is primarily a within-plant phenomenon? The answer is that, while blackmajority
plants pay their black employees less than plants where blacks
are a minority, these same plants tend to pay their white workers more.
On balance, this implies that plants with substantial black minorities do
not pay particularly low wages on average but that the black/white wage
gap is biggest within such plants.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
259The Black/White Wage Gap
VII. Conclusions
This article has shown that interﬁrm segregation within the average
MSA is close to what would be expected by the random allocation of
workers to ﬁrms; that blacks are more likely to work at ﬁrms with black
owners, black supervisors, and black customers; and that the black/white
wage gap is predominantly a within-plant phenomenon. We would like
to brieﬂy speculate about the implications of our results for theories of
black/white wage differentials. Becker’s (1957) theory of discrimination
and Wilson’s (1987) spatial mismatch hypothesis are two of the most
inﬂuential theories of black economic disadvantage, and both imply that
blacks and whites will be systematically segregated in the workplace. The
fact that we do not ﬁnd much systematic segregation in our data provides
a modest challenge to the ability of these theories to explain the black/
white wage gap. The results are particularly damaging to the spatial mismatch
hypothesis, as the problem does not appear to be that blacks are
spatially separated from high-paying employers but, rather, that they
receive low pay even when they work with whites. We also view the
results as damaging to the formulation of Becker’s theory where whites
have a distaste for physical proximity to blacks. In contrast, the results
may be fully consistent with a version of Becker’s theory in which whites
have a distaste for certain types of social proximity to blacks, such as
having a black boss or coworker of equal rank. However, we have relatively
little to say about such within-ﬁrm segregation.
There are some caveats to these conclusions associated with our data.
Both of the databases analyzed here are special in some way, and thus
the results found here may not apply to the broader population of ﬁrms,
establishments, and workers. We hope that a more representative database
suitable for the study of interﬁrm segregation will be developed in the
near future. Until that date, however, our results will have to stand as is.
References
Bates, Timothy. ‘‘Do Black-Owned Businesses Employ Minority Workers?
New Evidence.’’ Review of Black Political Economy 17 (1988):
51–60.
Becker, Gary S. The Economics of Discrimination. Chicago: University
of Chicago Press, 1957.
Becker, Henry Jay. ‘‘Racial Segregation among Places of Employment.’’
Social Forces 58 (1980): 761–75.
Bound, John, and Freeman, Richard B. ‘‘What Went Wrong? The Erosion
of Relative Earnings and Employment among Young Black Men in the
1980s.’’ Quarterly Journal of Economics 107 (February 1992): 201–32.
Blau, Francine. Equal Pay in the Ofﬁce. Lexington, MA: D.C. Heath,
1977.
Boisso, Dale; Hayes, Kathy; Hirschberg, Joseph; and Silber, Jacques.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms
260 Carrington/Troske
‘‘Occupational Segregation in the Multidimensional Case.’’ Journal of
Econometrics 61 (1994): 161–71.
Carrington, William J., and Troske, Kenneth R. ‘‘Gender Segregation in
Small Firms.’’ Journal of Human Resources 30 (Summer 1995): 503–
33.
. ‘‘On Measuring Segregation in Samples with Small Units.’’ Journal
of Business and Economic Statistics 15 (October 1997): 402–9.
Duncan, Otis, and Duncan, Beverly. ‘‘A Methodological Analysis of Segregation
Indices.’’ American Sociological Review 20 (1955): 210–17.
Ellison, Glenn, and Glaeser, Edward L. ‘‘Geographic Concentration in
U.S. Manufacturing Industries: A Dartboard Approach.’’ Journal of
Political Economy 105 (October 1997): 889–927.
Flanagan, Robert J. ‘‘Racial Wage Discrimination and Employment Segregation.’’
Journal of Human Resources 8 (1973): 456–71.
Harrison, Roderick J., and Weinberg, Daniel H. ‘‘Racial and Ethnic Residential
Segregation in 1990.’’ Paper presented at Population Association
of America meetings, May 1992.
Hellerstein, Judith; Neumark, David; and Troske, Kenneth. ‘‘Wages, Productivity,
and Worker Characteristics.’’ Working Paper no. 5626. National
Bureau of Economic Research, Cambridge, MA: 1996.
Higgs, Robert. ‘‘Firm-Speciﬁc Evidence on Racial Wage Differentials and
Workforce Segregation.’’ American Economic Review 67 (1977): 236–
45.
Hutchens, Robert M. ‘‘Segregation Curves, Lorenz Curves, and Inequality
in the Distribution of People across Occupations.’’ Mathematical
Social Sciences 21 (1991): 31–51.
Jackson, James S., and Gurin, Gerald. National Survey of Black Americans,
1979–1980 (machine-readable data ﬁle). 1st ICPSR ed. Ann
Arbor, MI: Inter-university Consortium for Political and Social Research
(ICPSR), 1987.
Kremer, Michael, and Maskin, Eric. ‘‘Segregation by Skill and the Rise
in Inequality.’’ Working paper. Cambridge, MA: Harvard University,
1994.
Ragan, James F., Jr., and Tremblay, Carol Horton. ‘‘Testing for Employee
Discrimination by Race and Sex.’’ Journal of Human Resources 23
(Spring 1988): 123–37.
Troske, Kenneth R. ‘‘The Worker-Establishment Characteristics Database.’’
In Labor Statistics Measurement Issues, edited by John Haltiwanger,
Marilyn Manser, and Robert Topel. Chicago: National Bureau
of Economic Research, in press.
Wilson, William Julius. The Truly Disadvantaged. Chicago: University
of Chicago Press, 1987.
/ 9e10$$ap01 02-10-98 21:31:38 laeca UC: Labor Econ
This content downloaded from 147.251.55.81 on Fri, 23 Mar 2018 13:17:37 UTC
All use subject to http://about.jstor.org/terms