The Economic Analysis of Labor Market Discrimination

Chapter 13
THE ECONOMIC ANALYSIS OF LABOR MARKET

DISCRIMINATION: A SURVEY
GLEN G. CAIN*
Universi(v of Wisconsin-Madison
I. Introduction
This survey of the economics of labor market discrimination is motivated by two

fundamental problems associated with income and wage differences among
groups classified by sex, race, ethnicity, and other characteristics. The first is the
inequity of long-lasting differences in economic well-being among the groups; in
particular, differences in household or family income. The second is the inequity
of long-lasting differences in the average wage rates among groups of workers
classified by these demographic traits, when the groups may be presumed to be
either equally productive or to have equal productive capacity. The second
problem also raises the question of whether a labor market that pays unequal
wages to equally productive workers is inefficient.
Economic discrimination is defined in terms of income differences among
families and wage differences among workers. In Section 2, I discuss these
definitions and present data from the United States on the income and earnings
differences of blacks, Hispanics, whites, women, and men.
Section 3 surveys theories of economic discrimination in the labor market. The
theories are classified into competitive and monopolistic neoclassical models with
(essentially) complete information, competitive neoclassical models with imper-
fect information-leading to "statistical discrimination", and institutional theo-
ries. Only neoclassical models offer generalizable theories that can be rigorously
tested, but I argue that these theories lack supporting empirical evidence.
*I am grateful to many persons, including most of the authors discussed in the chapter, for
comments, criticisms, and corrections. Even though I did not always follow their advice, the chapter
has been much improved because of their help. I am especially indebted to the following persons, who
read and commented on the entire manuscript: Francine Blau, Betty Evanson, Ross Finnie, Arthur
Goldberger, James Heckman, and Elizabeth Uhr. Research support was received from the Institute
for Research on Poverty at the University of Wisconsin-Madison and from the U.S. Department of
Health and Human Services. Any opinions expressed here are my own.
Handbook of Labor Economics, Volume L Edited by O. Ashenfelter and R. Layard

©Elsevier Science Publishers B V, 1986
694 G. G. Cain
Empirical tests of the economic theories are selectively surveyed in Section 4.

Most attention in this section is, however, given to a survey of the estimations of
wage (or earnings) functions for various groups of workers as a way of measuring
labor market discrimination, operationally defined as differences in predicted
wages (for the groups) when the prediction "holds constant" various productivity
determinants of wages.
A distinction is made between marketwide estimates of labor market dis-
crimination and estimates that apply to an individual firm. Both methods
commonly use multiple regression, but they differ primarily in the specification of
exogenous predictor variables-that is, variables that may be assumed to affect
wages but not to reflect the process of discrimination. The statistical models of
discrimination in individual firms have become widely used in recent years as
evidence in court cases or other litigation stemming from antidiscrimination laws.
Although the estimates of predicted wages in both firms and markets contain
much useful information, there are inherent weaknesses in the models in terms of
interpreting the estimates as measures of labor market discrimination.
The chapter concludes with a discussion of the policy implications of the
economic research on discrimination. Data on the changes over time in compara-
tive earnings of women and men and of black men and white men are used as a
basis for discussing the role of policies in explaining and affecting these changes.
2. The definition of economic discrimination
2.1. Concepts
Economic discrimination is a concept that defies precise definition. One difficulty

is that the intended meaning of the term differs in several contexts in which it is
used. To define economic discrimination I proceed in steps and begin with two
problems that span the economist's scope of interests and expertise, from the
practical to the theoretical.
(1) A practical problem, based on observed and quantified outcomes in the
economy and of intense concern to the public at large, is the wide disparity in
income, earnings, and wage rates among a variety of demographic groups,
classified by sex, race, ethnicity, and other characteristics. The disparities are
systematic,,~persistent, and considered by most observers to be inequitable,
although the definitions and sources of the inequities are often controversial. For
brevity, I will refer to the group experiencing lesser economic' rewards as the
" m i n o r i t y " group and the more favored group as the "majority" group. The fact
that discrimination, in the sense of disparate outcomes and inequitable treatment,
has been alleged to affect many different groups complicates its conceptual
definition and makes a review of empirical work overwhelming. In this chapter I
Ch. 13: Labor Market Discrimination 695
concentrate on discrimination in the United States against women, who are not a
numerical minority group, and blacks. References to discrimination against
certain ethnic groups, age groups, and the handicapped will sometimes be made
to elucidate certain general issues.
(2) The theoretical problem, which might be purely hypothetical except that it
has been motivated by the first problem, is: Under what conditions will essentially
identical goods have different prices in competitive markets? 1 In practice, the
question refers to goods that are, on average, the same and to a price difference
that is sustained rather than transitory. Economic discrimination refers to a
group rather than to an individual, and it is of greater concern as it persists over
time. This theoretical problem may be specified more rigorously, but let us first
consider its constituent parts to see its practical implications.
Discrimination in the labor market takes labor services as the good in question
and the wage rate as the price. Labor services are considered "essentially
identical" if they have the same productivity in the "physical" or "material"
production process; a consideration that excludes the effect of the laborer on the
psychic utility of his or her coworkers or employers. In fact, psychic disutility is
an essential part of a useful definition of economic discrimination that was
formulated by Becker (1957, rev. 1971) and which will be discussed in Section 3.
If the employers, for example, feel a disutility in hiring a minority worker solely
because of the worker's demographic characteristic, which, by itself, is irrelevant
to the worker's physical productivity, then employers may be said to be pre-
judiced. As another example, if the majority group of coworkers manifest their
feelings of psychic disutility by actions which curtail the minority worker's
physical productivity, this outcome will still be considered discriminatory, be-
cause the operative or causal variable is the majority group's prejudices, not the
minority group's productivity. Under some but not all conditions, these tastes,
which reflect the prejudices of employers and workers, will lead to discrimination,
defined by wages to the minority group being below what they would receive if
only their physical productivity were determinant. There is, therefore, a distinc-
tion between discrimination, which refers to behavioral outcomes, and prejudice,
which refers to attitudes. My point is not that tastes are the sole source of
discrimination; rather that they not be allowed to define away discrimination.
The concept of physical productivity, although it excludes the psychic compo-
nent, is intended to be broad and to include such characteristics of the workers as
their regularity in attendance at work, dependability, cooperation, expected
future productivity with the firm, and so on. A grey area occurs when there is
customer contact with the workers and when it is the customers who feel the
~Gustav Cassel, the renowned Swedisheconomist,may have been the first to state this question in
the context of labor market discriminationin his analysis in 1918 of why women doing similar work
to men received lower wages. See the citation to Cassel along with an interesting discussion of the
history of the economists' debate on labor market discrimination in Lundahl and Wadensj~5(1984,
pp. 8-80).
696 G. G. Cain
psychic disutility. Here, the distinction between physical and psychic components
of production can break down. "Service with a smile is our product", may be the
company's motto. It will be argued below that although customer prejudice c a n
lead to discriminatory outcomes, it is unlikely to be a major source of the
economywide disparities in the wages and incomes between minority and major-
ity groups.
Implicit in the foregoing two concepts of economic discrimination are two
subclassifications that are defined by the unit of the analysis; namely, (a) the
household (or family), which is generally the appropriate unit for examining the
disparity between majority and minority groups in economic well-being, usually
measured by income; and (b) the individual worker, the appropriate unit for
examining disparities in wage rates or earnings. In most of this survey the worker
is the unit of analysis, reflecting the fact that labor market discrimination,
measured by wage disparities, has been the focus of most economic-theoretical
and econometric studies. Nevertheless, attention to the family as a unit and
income as an outcome is important. The family is the principal matrix for a
worker's choices, and an understanding of labor market discrimination requires
attention to this family context. This is most clearly evident in analyzing
discrimination against women. Also, our ultimate interest in labor market dis-
crimination lies in the question of how discrimination affects the economic
well-being of people, which, as noted, is most meaningfully measured for a
household or family unit.
Each of the two units of observation, worker and household, may be analyzed
with two general types of statistical models. In Model (I), which may apply to the
short run, the outcome variable of interest-income for households or wages for
w o r k e r s - i s compared for the two groups, holding constant certain variables that
are believed (a) to affect the outcome variable (or to be relevant to the interpreta-
tion of the outcome variable), and (b) to be exogenous to the process of
discrimination under study. For example, income of households may be com-
pared, holding constant the region of residence. If region of residence is exoge-
nous and the cost of living varies across regions, then income is a better measure
of economic well-being when region is held constant in the comparison. If region
of residence is endogenous to the process of discrimination, then it is probably
not a proper control variable.
Model (I), which is distinguished by the use of contiol variables, is more
important for the second definition of discrimination-wage differences for
comparable workers. The comparability of the workers is with respect to their
productivity, which is operationally defined by measurable characteristics of the
workers that are accepted as determining productivity in the given context. Here
again we require that the productivity variables that are properly held constant
are exogenous to the process of discrimination under study.
Let us specify Model (I) in a form suitable for statistical estimation. Let
= the outcome of the process, such as the income, earnings, or wage for the ith
person; X i = a vector of productivity characteristics of the ith person that are

presumed exogenous in that they do not depend on Y nor on the particular form
of economic discrimination under study; Z i = 1 if the person is in the majority
group and 0 if in the minority group; e~ = a random error term; and let A and B
be coefficients representing the effects on Y of Z and X. Assuming a linear and
additive model for convenience and suppressing subscripts to avoid clutter, we
have
Y = X'B + AZ + e. (I)
Then, a regression in which we find A > 0 would be evidence of discrimination.
The contrary case is assumed to be A = 0, so "reverse discrimination" (A < 0) is
not being considered. In Model (I), the two groups designated by Z are assumed
to provide "essentially identical" labor services, conditional on (holding constant)
X. Equivalently, we could define market discrimination, D, as
D = (~'lX, Z = l ) - ( 9 1 X , Z=O),
where ~" is the predicted value of Y conditional on X, so in the above linear and
additive model, D = A.
Now consider Model (II), in which all X characteristics are considered
endogenous, and any difference in X across groups is attributed to the process of
discrimination under study. Model (II) may be appropriate for the long run,
although some may consider it only the limiting case in which the group averages
of all X ' s are equalized in the long run in a world without economic discrimina-
tion. The corresponding specification is
V = CZ + u, (II)
where u is a random error and C > 0 is evidence of discrimination. In this case,

we can define D = Ymaj- Y~in, now using unconditional means instead of condi-
tional means, substituting the mnemonic subscripts for the Z-values, and adopt-
ing notation suitable for describing samples instead of populations. The long-run
model deliberately ignores the common distinction between the occurrence of
discrimination "within" versus "prior to or outside" the labor market.
The practical problem of disparities in economic well-being, usually defined in
terms of differences in household incomes, is generally addressed by Model (II).
The practical-and-theoretical problem of differences in wages for equally produc-
tive workers is generally examined by Model (I). However, Model (I) may be
specified as close to Model (II) as desired by restricting the set of admissible X
characteristics.
An interesting and unusual feature of the economic analysis of discrimination
is the attention given to the roles of tastes and nonpecuniary aspects in market
698 G. G. Cain
transactions. The economist's treatment of tastes is, however, circumscribed.

Tastes are fundamentally taken as given, and explaining their sources or how
they may be changed tends to be left to the other social sciences. Instead, the
economist's main objective is to determine certain behavioral outcomes that are
the consequences of these tastes-specifically the disparities in employment,
wages, and so on. Market outcomes become indirect measures of tastes and the
focus of attention. Direct measures, such as those obtained from attitudinal
surveys, which are a staple in sociology and psychology, are seldom used in
economics. Despite these largely self-imposed limits of the economic analyses, the
goals of predicting market outcomes and predicting the effects of policies aimed
at altering these market outcomes are important and difficult.
The productivity of a given worker is also influenced by the tastes of that
worker. Adherence to Model (II) implies that minority and majority groups are
equal in both their productive capacity and their willingness to produce. Equal
productive capacity refers to a common presumption of innate equality among
racial and ethnic groups. Innate equality in "effective" capacity may also be
assumed for women, relative to men. Thus, the biological difference in physical
strength between men and women may be presumed to convey no net advantage
in earnings or productive capacity to men in today's labor market. Such differ-
ences clearly lead to differential sorting into specific occupations, just as they do
within a gender or racial group, but there is no necessary reason for this
specialization to lead to an average wage difference across groups.
Equal willingness to produce refers to equality in tastes for market work
relative to leisure when comparing racial groups and to tastes for market work
relative to the combined time allocation to housework and leisure when compar-
ing men and women. Are such tastes predetermined, or are they determined, or at
least affected, by discrimination? Prior equality in tastes between men and
women is often denied on the grounds that cultural and biological forces, which
are presumed exogenous to the economic system (or, more narrowly, to the labor
market), are the causes of a preference for market work relative to housework
among men and vice versa for women. In principle, Model (I) allows any
X-variable, including tastes, to be correlated with gender, because the gender
effect on wages is estimated net of the X's. However, as discussed below, the
choice of X ' s is often disputed.
Another conventional stance taken by economists in their study of discrimina-
tion is that the state of technology is given, which is the analogue in production
to the ass l~aption o f given tastes in consumption. The issue arises whenever a
distinctive trait of the minority group places it at some disadvantage because of
the existing state of technology. In my view, if it would be costly to change the
technology to accommodate the minority group, then there is no presumption of
discrimination. The minority group in this industry or firm would simply be
considered less productive. If the technology is not costly to change, then the
market, in the absence of discrimination, should already have provided the

accommodating change. Thus, lowering the height of shelves could equalize
the productivity of those minority groups who tend to be shorter, and new
construction offers the opportunity to build ramps instead of stairways to
accommodate people in wheelchairs. Perhaps some market impediment, such as
government regulations, might need to be eliminated to permit the accommod-
ation. These issues are interesting, but they will not be discussed in this chapter.
Technology is assumed to be exogenous, but, like physical strength, it is not
considered an important source of average productivity differences between racial
groups or between men and women.
The meaning and measure of income as an index of economic well-being and
of the "wage rate" are complicated issues in any practical or empirical examina-
tion of either Model (I) or (II). For example, measuring the wage as the price of
labor services must deal with distinctions between current and lifetime returns to
work and between pecuniary and nonpecuniary returns and, at times, with the
measure and evaluation of leisure and the rewards to housework. Some specific
examples may be helpful. Black men appear to receive fewer nonpecuniary
benefits from their market work than do white men [Lucas (1974)]. If so, the wage
advantage of white workers would be even greater if the nonpecuniary aspects of
employment were monetized and included. On the other hand, black men spend
less time at work than white men. Does this compensate them for their lower
wages and earnings? The usual answer is "yes" if the time not at work is
voluntary and perhaps considered to be leisure, but "no" if it is "involuntary
unemployment". The latter may create anxiety and distress for the unemployed
person and have a zero or negative value. Another example concerns household
work by women, for which the rewards are, let us assume, the income shared by
the family unit. Does this income compensate women for their lower market
earnings? The issue, discussed more fully below, partly depends on the degree to
which women's allocation to housework and market work is voluntary, or,
perhaps equivalently, the degree to which women's tastes for market and house-
work are exogenous.
The complexities in measuring the Y-outcomes as indicators of discrimination
should not be overemphasized. Sometimes one measure is believed to understate,
and another to overstate, discrimination, and yet both measures may give
qualitatively similar results. Evidence for this outcome is provided below. Usu-
ally, the disparity remains whether the wage or income is used, and whether the
wage is measured with or without an allowance for nonpecuniary aspects of the
job.
In summary, measures of economic discrimination in the labor market are the
positive coefficients, A and C, in Models (I) and (II), assuming the proper
measure of Y, the suitable choice of one of the two models, and, if Model I is
chosen, the suitable specification and measure of X. These qualifications and the
700 G. G. Cain
subsequent interpretation of the coefficients and their properties all require a

theoretical framework, to be discussed in Section 3.
Before presenting statistical evidence on discrimination, let us note several
strengths and weaknesses of the concepts used. Their strengths include their links
to market-based measurements of variables that are of intense concern to the
general public as well as to the technicians who study the problem. They are
robust in the face of "special cases" or individual deviations, so long as these
cases and deviations are " r a n d o m " with respect to the process that is modeled.
This last strength, however, may be viewed as a weakness from the perspective
of various ethical or legal definitions of discrimination. When Model (I) or (II)
applies to a large aggregate, such as the nationwide labor market, then a finding
of no discrimination on average could be consistent with many individual cases
of discrimination, so long as these were balanced by a sufficient number of
cases of reverse discrimination. Lawyers and philosophers need not be put out of
business by findings that A or C equal zero. [The distinctions between applying
Models (I) and (II) to marketwide versus, say, individual-firm contexts will be
discussed in Section 4.]
Now consider that the above economic measure of discrimination is silent
about segregation. Either perfect integration or complete segregation is consistent
with a finding of no discrimination. In particular, the economic definition accepts
"separate but equal (wages)" as no discrimination, even though segregation may
be considered noxiously discriminatory in legal and ethical senses.
Using wage differences rather than segregation indices to measure discrimina-
tion in the labor market is a corollary to my emphasis on wage discrimination
rather than employment discrimination in this chapter. One justification for this
emphasis, in addition to the convenience of the measurability of wages, is that
when discrimination takes the form of widespread refusals to hire or promote
minority workers, this should lower their relative wages. The rejected minorities
must bid for jobs in less favored firms, industries, occupations, and so on. This
process has been referred to as the "crowding hypothesis" [Bergmann (1974)], but
my point here is that wage outcomes will reflect this reduced demand for the
minority group. 2 A second justification is that wage discrimination can exist
irrespective of the degree of integration or segregation in the market. Thus, wage
differentials are a more fundamental measure of labor market discrimination
than are employment differentials between majority and minority groups. None
of these arguments for my use of wages as a basis for measuring and discussing
labor mark~et discrimination denies that employment and hiring statistics are
appropriate in many practical contexts, including court cases involving dis-
crimination.
2"Crowding" is an old concept. Lundahl and Wadensj8 (1984, p. 73, n. 16) trace it back to John
Stuart Mill, and they cite F. Y. Edgeworthand Millicent Fawcett as early twentieth-centuryusers of
the term regarding labor market discrimination against women.
2.2. Summary statistics for two concepts of economic discrimination
The first definition of economic discrimination, concerning differences in eco-

nomic well-being, permits a simple measure of the differences in mean household
or family income. Annual money income is assumed to be the indicator of
economic well-being, and the difference will be expressed as a ratio of the
minority group's income to that of the majority group.
Some comparisons of the incomes in 1981 among white, black, and Hispanic
households and families are shown in Table 13.1. 3 The table is detailed, and it
may be helpful to note the following highlights and interpretations.
(1) Blacks and Hispanics constitute about 17 percent of the U.S. population.
The total numbers of households and families by ethnic status are shown in
columns 6 - 8 in rows 1, 3, 5, 8, and 9. Along with other smaller minority groups,
such as American Indians and certain Asian immigrant groups, about 20 percent
of the U.S. population may be classified into ethnic minority groups that are
often believed to be victims of economic discrimination.
(2) The average income of a black household, $14900, is 63 percent of that of a
white household, which is $23 700. (See row 1, columns 1-3.) On a per-person
basis, the ratio is only 56 percent, reflecting the larger average size of black
households, as shown in row 2, columns 3, 6, and 7. As discussed below, the ratio
of black-to-white income has been fairly steady in recent years but has risen over
a longer period of time.
(3) The ratios of black-to-white and Hispanic-to-white incomes tend to be
around 0.6 or 0.7. The average income per member of a black family headed by a
woman is, however, only 32 percent of the average income per member of a white
married-couple family. 4 (Using column 2, row 6, and column 1, row 4, we obtain:
2.8/8.8 = 0.32.) This is a large difference.
(4) Poverty status for families in 1981 was officially defined to be an annual
income of $9300 or less for a family of size four and of $7300 or less for a family
of size three. Thus, a substantial proportion of black and Hispanic families
headed by women are poor, whereas only a small proportion of black and
Hispanic married-couple families are poor. For most minority-group families,
therefore, discrimination regarding family income in the United States is not so
3The term "white" will be used to refer to non-Hispanic whites. "Hispanic" refers to persons of
Spanish origin, who m a y be members of any race. Persons whose origins are Mexican, Puerto Rican,
C u b a n , or who are from other Central or South American countries constitute most of the Hispanic
group in the United States. A "household" consists of all persons who live together in a housing uhit
and includes one-person households. "Families" are defined as two or more persons related by blood,
marriage, or adoption, and residing together.
4The term "female household head" refers to a household or family in which the primary earner is
usually an adult w o m a n without a husband present. The terms "householder" and "female house-
holder," which are currently being used in the official statistics of the U.S. government, are defined in
terms of the person in the household in whose name the dwelling unit is owned or rented. Statistics
for households (or families) with a female householder are nearly the same as those that would apply
to the older designation, female-headed households (or families).
702 G. G. Cain
Table 13.1
M e a n annual incomes and income ratios of white, black, and Hispanic households and families,
United States, 1981.
N u m b e r of units (in millions);

Mean annual income ($000's) average n u m b e r of persons per
and B / W and H / W ratios a unit in parentheses
W B B~ W H H~ W W B H
Demographic unit (1) (2) (3) (4) (5) (6) (7) (8)
1. Households b $23.7 $14.9 0.63 $18.4 0.77 72.8 9.0 4.0

2. (per member) c 8.9 5,0 0.56 5.3 0.59 (2.67) (2.99) (3.49)
3. Married-couple families a 28.7 21.9 0.76 22.1 0.77 43.3 3.2 2.3
4. (per member) 8.8 5.8 0.66 5.4 0.62 (3.27) (3.79) (4.07)
5. Female-headed families e 15.3 9.8 0.61 10.8 0.70 6.6 2.6 0.7
6. (per member) 5.4 2.8 0.52 3.1 0.58 (2.84) (3.50) (3.47)
7. Female-headed families as
proportion of all families r 0.12 0.41 0.23
Families with primary earner

w o r k i n g " full time": g
8. Married-couple families h 30.5 25.9 0.85 22.3 0.73 27.8 1.9 1.5
9. Female-headed families h 18.0 13.4 0.74 15.9 0.88 2.5 0.8 0.2
Source: U.S. Bureau of the Census, Current Population Reports, Series P-60, No. 137, Money Income of
Households, Families, and Persons in the United States: 1981 (Washington, D.C.: U.S. GPO, 1983), Tables 4,
13, and 19.
a Incomes are rounded to the nearest hundred, but the ratios are based on unrounded incomes. For
example, the original mean household incomes for whites and blacks in the first row are $23 742 and $14 856.
bHouseholds consist of all persons who live together in a housing unit and include one-person households.
CMean annual income per member is money household income divided by the average size of the
household. For example, for white households: $ 2 3 7 4 2 / 2 . 6 7 = $8892, which, rounded and expressed in
thousands of dollars, is 8.9.
d The Census Bureau defines a family as two or more persons related by blood, marriage, or adoption, and
residing together. In this table, married-couple families do not include the relatively small number of families
in which the wife is listed as the owner of the housing unit. W h e n the wife is listed as the owner, the family is
classified u n d e r "female householder". The term "householder" has replaced the term "headship" in
government tables.
c Does not include the relatively small number of female-headed families with a h u s b a n d present.
f"AII families" includes the relatively small n u m b e r of female-headed families with a h u s b a n d present.
g"Full time" refers to year-round, full time, defined as working 50-52 weeks for 35 or more hours per week
in 1981.
h Median incomes are listed instead of mean incomes, which are not reported.
much a problem of poverty, at least as officially defined, as it is of inequality - their

incomes are low relative to the incomes of the white majority group.
(5) One reason black and Hispanic incomes are lower is that the fraction of
families headed by a woman is larger among these minority groups. If both
headship status and the presence of a full-time worker are held constant, the
income ratios rise to around 0.8. (See rows 7-9, columns 3 and 5.) Marital
instability and slack labor markets thus appear to be important sources of
income inequality across ethnic groups in the United States. Whether marital and
e m p l o y m e n t statuses should be held constant in assessing discrimination depends
on the particular purposes and issues in one's analysis. As noted above, one issue
is whether marital and employment statuses are affected by discrimination.
(6) This type of table is more difficult to construct for other minority groups of
interest, but consider the reported incomes for the following three groups that
faced discrimination in the United States in the past:
(a) persons of Italian a n c e s t r y - t h e largest group of immigrants to the United
States in the twentieth century;
(b) persons who state their religion as Jewish, whose ancestors had immigrated
predominantly from Eastern Europe; and
(c) persons of Japanese a n c e s t r y - t h e largest group of immigrants from Asia.
Several researchers have concluded that the average family incomes of each of
these groups was, in 1970, higher than the average in the United States for all
other white families. 5
What adjustments to the available statistics for money income are required to
measure relative economic well-being more completely? A satisfactory answer to
this question would involve the resolution of philosophical and measurement
problems that are beyond my capacity, but most of the issues that lend them-
selves to quantification or informed judgments are listed in Table 13.2. In the
table the sources of inequality and the accompanying adjustments are separated
into those pertaining to income receipts and those pertaining to expenditures. In
measuring income receipts attention is given to (a) the measures of income from
a household's assets (or wealth components); (b) the demographic unit of
analysis; (c) allowances for government taxes and subsidies; and (d) allowance
for survey biases.
There is not the space to discuss each of these adjustments, but two conjectures
may be suggested. First, the money measures in Table 13.1 probably understate
the true degree of inequality between blacks and whites, and, by extension,
between majority and minority ethnic groups generally. Seven of the 10 required
adjustments serve to widen the gap. Second, even descriptive statistics about
" i n c o m e differences" in discrimination studies are complicated.
Table 13.1 shows a static picture of income differences, and it is essential in an
analysis of discrimination to describe how these differences have changed over
time. The time-series data are, unfortunately, incomplete in several respects.
51t is more difficult to define and collect information on groups according to their ancestry and
religion than it is for gender and racial classifications,so the statements in the text are more qualified.
The problems of mixed or unknown ancestry, changes in one's religion, response refusals and errors,
and so on appear serious, and the data on income, earnings, and wage rates have not been collected
for ancestry and religion classifications as thoroughly as they have for the gender and racial groups.
The sources for the research findings in the three ethnic groups referred to in the text are Greeley
(1976, p. 52) and Sowell (1981, pp. 5, 126-127) for Italian-Americans, Chiswick (1983) for Jews, and
Sowell (1981, pp. 5, 177-178) for Japanese-Americans.
704 G. G. Cain
T a b l e 13.2
S o u r c e s o f i n e q u a l i t y in e c o n o m i c w e l l - b e i n g , i l l u s t r a t e d w i t h a c o m p a r i s o n of
black and white families in the United States.
Judgment as to whether accounting

for the source would widen or narrow
the conventional black white income
gap. (No adjustments needed, N.A.,
implies that the conventional ratio
Source already allows for the source.)
Income receipts
Asset ownership
Property (income-earning) N.A.
Property (non income- Widens gap (blacks have less wealth
earning: car, owner- in these types of durable goods)
occupied house, etc.)
Human capital (wage earnings) NA
Human capital (fringe benefits
and nonpecuniary aspects of work) Widens gap~
Defined for "' household" as unit
Adjust for family or Widens gap (unless the comparison
household size is already "per member") b
Adjust for multiple earners Narrows gap (whites have 1.65
to allow f o r " leisure" earners per family: blacks, 1.47) ~
consumption
AUowancc for government taxes,
transfers, and survey bias
Taxes Narrows gap (reflecting the moderate
degree of progressivity in the tax
system)
Money transfer payments N.A.
Nonmonetary transfer payments
to nonaged persons (primarily Narrows gap (about 25 percent of black and
Food Stamps, public housing, 8 percent of white families receive these
Medicaid) forms of noncash transfers) d
Nonmonetary transfer payments Widens gap~
to aged persons (medical care
subsidies and various tax
advantages for the aged)
Nonmonetary public benefits Widens gap r
(parks, police service, etc.)
Nonreported income ?
Lxpenditures
Discriminatory pricing housing, Widens gap I
capital markets, consumer credit,
etc.
Expenditures on "regrettables" - Widens gap eg
items that do not directly
produce utility, such as health
maintenance, transportation to
work, "waiting times"
~Fringe benefits are generally large for jobs with higher wages and salaries For evidence that
blacks have. on average, jobs with less prestige and less pleasant working conditions, see Robert E. B.
Lucas, "'The Distribution of Job Characteristics," Review of Economics and Statistics, 56 (November
1974): 530-540.
bSee Table 13.1.
Source: Table 29 in source cited in Table 13.1.
dSource: U.S, Bureau of the Census, Current Population Reports, Series P-60, No. 136, Chara¢tert~-
tt~'# of IIouseholds and Persons Receiving Selected Noncash Benefit.~, 1981 (Washington, D.C.: U.S.
GI~O, 1983). p] 3,
CMedical care subsidies are derived primarily from the Social Security system, and white persons
benefit disproportionately for two reasons: (1) eligibility and payments tend to be positively related to
earnings during preretirement years; (2) whites live longer The tax advantages of the aged are
generally greater for higher-income persons among the aged.
rA personal judgment.
gFor a definition and application of the concept of "regrettable" expenditures, see William N.
Nordhaus and James Tobin, Is Growth Obsolete? (National Bureau of Economic Research 50th
Anniversary Colloquium, Columbia University Press, New York, 1972).
Income statistics prior to 1940 are scanty. The Census Bureau's time series of
annual family income begins in 1947, and separate income statistics for blacks
begin in 1967 and for Hispanics in 1972.
The income ratios are relatively stable year by year (not shown), but the
change over decades is notable. To summarize the trends, several 10-year
averages of the annual ratios of minority-to-majority incomes for the period since
1947 are shown in Table 13.3. The ratio of nonwhite-to-white family income rose
from 0.37 in 1939, when most blacks lived in the low-income Southern region and
on farms, up to 0.6 or more in the middle 1960s, when the ratio more or less
stabilized. Since then it has been held down by the increasing proportion of black
families headed by women, and, probably, by the relatively high unemployment
levels from 1975 on. 6 Whatever the reason, progress regarding the first type of
economic discrimination, family income differences, has been painfully slow.
Table 13.4 shows the earnings of workers instead of the incomes of families. To
the extent that earnings measure the economic well-being of workers, the table
shows economic discrimination according to the definition of discrimination that
was based on disparities in well-being. According to the definition that was based
on wage rate differences among comparable workers, Table 13.4 would provide a
measure only if we considered the worker g r o u p s - t h r e e ethnic groups and two
gender g r o u p s - to be equally productive.
In Table 13.4 ratios ranging from 0.5 to 0.7 characterize most of the compari-
sons between minority men and white men and between women and men within
each ethnic group. However, minority women earn around 90 percent of the
earnings of white women. The earnings ratios of women to men and of black men
to white men are smaller for "all workers" than for "year-round, full-time
workers" (hereafter, "full-time"), because women and black men are less likely to
work full time. (The proportion of full-time workers to all workers is shown in
parentheses in the first three columns of the last two rows. More young workers
and higher unemployment among these minority groups are two sources of these
lower proportions.)
Clearly, the earnings ratios for full-time workers are closer to the ratios of
hourly wage rates, because the all-worker variation in hours worked in the
definition of e a r n i n g s - hours worked times the average wage per h o u r - i s nearly
6Family income depends importantly on the number of earners per family, and this number has
increased among white families relative to black families in the last 20 years. The main reason is that
the percentage of all families headed by women rose from 21 percent in 1960 to 42 percent in 1980
among black families and by 8 percent to 14 percent among white families [U.S. Bureau of the Census
(1983c, p. 54)]. Families headed by women tend to have fewer earners than married-couple or
male-headed families. The change in work rates amongwives, who are the largest and most important
category of secondary earners in families, did not much affect the racial difference in earners per
family. The rise in labor force participation rates of wives with husbands present was similarly rapid
for both color groups from 1960 to 1981: from 30 percent to 50 percent for white wives and from 41
percent to 60 percent for black wives [U.S. Department of Labor (1982, p. 714)].
706 G. G. Cain
Table 13.3
Median family income ratios: Black-and-other races/white; black/white;
and Hispanic/white; annual averages for five periods, 1939-1982.
Black-and-other
Year or period a races/white b Black/white c Hispanic/white d
1939 0.37 - -
1947-1956 0.54 - -
1957-1966 0.54 -
1967-1976 0.63 0.61 0.69 d
1977-1982 0.62 0.57 0.68
Sources: U.S. Bureau of the Census, Current Population Reports, Series

P-60, Nos. 43, 137, and 140, published in years 1964, 1983, and 1983
respectively. The full citation of No. 137, which gives the family income
figures for 1947-1981 is: U.S. Bureau of the Census, Current Population
Reports, Series P-60, No. 137, Money Income of Households, Families, and
Persons in the United States, 1981 (Washington, D.C.: U.S. GPO, 1983),
p. 39.
aThe years 1947-1982 are divided into four periods, and the average of
the annual ratios are reported for each period. The first year for the
continuous time series of annual incomes (see sources) is 1947.
bThe category black-and-other nonwhite races is more than 90% black
for most of the period and is the only category continuously available for
the earlier years. Except for the recent decade or so, the trends in the
ratios for nonwhites and for blacks appeared very similar, based on the
scattered evidence available. In recent years, however, the proportion of
blacks among the nonwhite races has declined. Also, the proportion of
black families headed by women has risen most sharply during the last 10
years or so, and this has tended to make the family income statistics for
blacks diverge from those of other nonwhite races.
CThe first year in which blacks are reported separately is 1967.
d Family incomes of persons of Hispanic origin were first reported in the
annual series in 1972; therefore, the period for the Hispanic/white ratio is
1972-1976.
eliminated. Among working women, minority women are more likely to be

full-time workers, so the ratios of minority women's earnings to white women's
earnings are higher for the all-worker group.
The time series of earnings ratios for full-time workers, which is shown in
Table 13.5, is useful because among the available measures it comes closest to
providing a comprehensive comparison for minority and majority workers of the
trends in the relative price (wage) of labor services. For this interpretation, one
must a s s ~ e that the full-time workers remain about the same fraction of the
total population of workers, or that deviations represent (a) voluntary shifts to
part-time work, and (b) no systematic selection regarding workers' productivity
traits, in the changing distribution of part- and full-time workers. A change in
age composition could change the distribution, and, ideally, one would want to
hold constant an exogenous trait like age when constructing the time series.
Assuming that any group differences in these types of compositional shifts are
Table 13.4
Mean earnings, earnings ratios, and numbers of all workers and of year-round, full-time workers for men
and women; whites, blacks, and Hispanics, United States, 1981.
Numbers of all workers

in millions; and
year-round, (Black/white) and
full-time workers (Hispanic/white) (Women/men)
as a ratio of all Mean annual earnings earnings ratios, earnings ratios,
workers in parenthesesa ($000's) b by gender by ethnicity
W B H W B H l~/W U/W W B H
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
All workers: 0.48 0.69 0.59
Men 58.2 5.7 3.6 $ 1 7 . 5 $11.6 $12.5 0.67 0.72
Women 45.7 5.6 2.5 8.3 8.0 7.5 0.97 0.90
Year-round,
full-time workers: 0.58 0.76 0.70
Men (0.65) (0.58) (0.61) 22.8 15.7 16.5 0.69 0.72
Women (0.44) (0.49) (0.45) 13.3 12.0 11.5 0.90 0.87
Source: Table 55 in source cited in Table 13.1.

aA year-round, full-time worker is one who works (or is paid for) 50-52 weeks and 35 or more hours per
week.
bEamings are rounded to nearest hundred, but the ratios are based on unfounded earnings. For
example, the earnings for white and black men in the first row are $17453 and $11629, respectively. The
use of median earnings, which are about 8 percent lower, would not much change the comparisons.
minor, 7 Table 13.5 shows gains over time in earnings ratios for black women
relative to black men (column 2), black men relative to white men (column 4),
and black women relative to white women (column 5). The earnings ratio of
white women to white men (column 1) has been remarkably stable at around 0.6
over this 43-year span. The ratios for Hispanics (columns 3 and 6-7) are for too
brief a period to measure a trend.
Further analysis of these trends will be presented later, but the following
points seem evident.
(1) The ratios for the most recent period, 1975-1982, generally remain so far
short of unity that "slow progress" is a fair and regrettable assessment. The
exception is the remarkable rise to near-equality for black and white women,
despite the fact that their earnings ratio in 1939 was the lowest one shown in the
table. This rise is partly explained by the huge exodus of black women from
domestic service, one of the lowest-paid occupations, and the migration of blacks
7The only check on these questions of compositional shift that is easily ascertained is that of the
age composition. A time series of five observations from the decennial censuses from 1940 to 1980 of
the percentages of the population and of the labor force that is young (age 14-24), middle-aged
(25-64), and old (65 and over) show similar trends for the race and gender groups. Thus, the age
factor is unlikely to be an important source of variation in the earnings-ratio trends in Table 13.5.
708 G. G. Cain
Table 135
Median-earnings ratios for year-round, full-time workers, gender and ethhicity
comparisons, annual averages for four periods, 1939-1982.
Women/men Black/white Hispanic/white

earnings ratio by ethnicity earnings ratio earnings ratio
Year or White Black Hispanic Men Women Men Women
period a (1) (2) (3) (4) (5) (6) (7)
1939b 0.61 0.51 - 0.45 0.38 - -

1955-1966 b 0.61 0.61 - 0.62 0.65 - -
1967-1974c 0.58 0.70 - 0.68 0.83 - -
1975-1982a 0.59 0.76 0.70 0.73 0.94 0.72 0.86
Sources: Various years for the P-60 Series of the Current Population Reports. See
Table 13.4 for full citation.
aThe years 1955-1982 are divided into three periods, and the average of the
annual ratios are reported for each period. The first year for the continuous time
series of earnings for year-round, full-time workers is 1955, but the 1940 census
provides this figure for 1939.
bRatios are for wage and salary earnings (excludes self-employedworkers) for
whites and nonwhites, who are defined as blacks and other nonwhite races in later
Census publications.
cRatios are for all earnings (includes self-employedworkers and self-employment
income) for whites and blacks. The first year for which blacks arc reported
separately is 1967. The black/white ratios are, on average, about 0.01 lower than
the nonwhite/white ratios for men, and about 0.02 lower for women. The trends in
both ratios, black/white and nonwhite/white, are virtually identical.
dSame as note c; also, 1975 is the first year in which earnings are reported
separately for Hispanic workers.
generally f r o m the low-income rural sector of the South to u r b a n places. E a r n i n g s

of d o m e s t i c servants were understated in 1939 because of the receipt of income-
i n - k i n d p a y m e n t s (meals, sometimes lodging, a n d so on).
(2) Black earnings were relatively low i n 1939, partly because of the high rate
of u n e m p l o y m e n t throughout the 1930s. Black earnings rose sharply i n World
W a r II (1941-1945). The rate of increase i n the m e n ' s black-to-white ratio has
been slow b u t steady since the mid-1950s. D u r i n g tiffs period from 1955 to 1982,
w h e n real i n c o m e s were generally rising, the modest increases in the ratio have
m a i n t a i n e d roughly the same absolute difference in real earnings b e t w e e n blacks
a n d whites.
(3) Blacks m a d e relative gains between 1940 a n d 1960 in e d u c a t i o n a l a t t a i n -
m e n t and¢~probably, in other i n v e s t m e n t s i n h u m a n capital, such as health a n d
access to b e t t e r j o b s by migration. In the 1960s and 1970s there were further
gains in relative educational a t t a i n m e n t a n d also i n legal restraints o n discrimina-
tion in e m p l o y m e n t .
(4) T h e ratio of w o m e n ' s earnings to m e n ' s a m o n g whites has b e e n stable a n d
reflects two c o u n t e r a c t i n g trends: (a) more p a r t i c i p a t i o n in the l a b o r force b y
w o m e n and, associated with this, more a c c u m u l a t e d work experience a n d ad-
vancement into higher occupations; (b) an increasing number of women who are
new entrants or reentrants into the labor force, whose average years of experience
are less than the average of the existing stock of women workers. Thus, (a) exerts
a compositional effect that raises the ratio of women's earnings to men's earnings
while (b) has the opposite effect [see Mallan (1982)].
The descriptive statistics presented in Tables 13.1-13.5 have shown two
manifestations or definitions of economic discrimination, one dealing with in-
comes and another with wage rates, for three groups affected by discrimination:
women, blacks, and Hispanics. The economic disparities are large and have
persisted over time. The fundamental theoretical challenge is the presence and
persistence of different wage rates for groups of workers for whom the assump-
tion of equal p r o d u c t i v i t y - o r equal productive c a p a c i t y - i s maintained. The next
section of the chapter surveys the economic theories that have been formulated in
response to this challenge.
3. Theories of economic discrimination in the labor market
There is no shortage of theories to rationalize the existence of different wage rates

for equally productive workers. What is scarce is a theory that is buttressed by
empirical support. As discussed in the next section, the empirical work has
seldom tested the theories. In this section I resort to informed opinion and
speculative judgment about the plausibility and robustness of the theories.
Three theories of discrimination are found in the economic literature: (1)
neoclassical, which include nonstochastic and stochastic versions, (2) institu-
tional, and (3) Marxian. Only neoclassical theories, the basis for almost all the
theoretical literature in the United States, will be examined in any detail.
Marxian theory will not be examined, although certain components of this
theory, such as exploitation, do appear in the neoclassical and institutional
theories.
The neoclassical theory of discrimination is almost entirely a demand-side
theory. The supply side of the labor market is effectively neutralized by the
assumption that minority and majority groups of workers are equally productive
(or have equal productive capacity) and have equal tastes for work. The demand
side m a y be characterized by a competitive or monopolistic structure and by
"exact" versus "stochastic" models. These characterizations define the taxonomy
used below. 8
SThe taxonomy below, in subsections 3.1 and 3.2, of seven models was initially developed by
Becker in his influential book that was published in 1957 and revised in 1971. I remind the readers of
this point because "Becker's theory of discrimination" is often incorrectly identified with only one of
his several models- that dealing with a competitivemarket and employersas agents of discrimination.
The fact that I follow Becker's taxonomy in sections 3.1 and 3.2 should not be taken to mean that he
would agree with my formulation of the models.
710 G. G. Cain
3.1. Nonstochastic competitive neoclassical models: Discrimination by consumers
Becker relabeled the abstract concept of "prejudice" into the economic concept
of "tastes", and his operational definition of "tastes for discrimination" was that
of a demand function; namely, a monetary offer for a good or service with, in this
instance, a qualitative attribute (like race) that distinguishes it from another,
otherwise identical, good or service. If the price of the labor service of the
majority worker is p, then the prejudice or tastes for discrimination of a buyer
are measured by an offer price, p - d, for the (otherwise identical) service of the
minority worker. The term d is a measure of the buyer's tastes for discrimination.
I use the small letter d to measure an individual agent's discrimination, and D
will refer to marketwide discrimination.
Several advantages of the formulation are apparent. Discrimination has the
appealing property of continuity, rather than being merely present or absent. It is
potentially measurable, and the monetary units have an intuitive meaning to
experts and laypersons a l i k e - i n contrast to various attitudinal scales ("like a
l o t " . . . " d i s l i k e a lot") that may or may not be scored numerically. There are
explicit behavioral and even policy implications in the formulation. For example,
a government subsidy to a minority-produced service could equalize the net price
to consumers.
There are some disadvantages of the measure and some properties that may be
either advantageous or disadvantageous depending on the question one is asking.
N o attention is paid to any pain or stigma felt by the victim. A lower price for
one's services appears to capture the extent of victimization and to be on the
same footing as a lower price owing to an inferior standard property of the good
being sold. However, a black insurance salesman who offered the same policy as
a white seller but sold and earned less because of customer prejudice might feel
worse than if he received less because his policy offered less coverage or smaller
settlements. Both price differentials could be the same, but only the former is
viewed as an inequity and as a social problem.
Becker (1957, rev. 1971, p. 5) used the example of physical beauty as a
qualitative attribute that leads to discrimination by demanders but is not ordin-
arily viewed as a social problem, either because beauty is considered legitimately
productive - as it is in acting and modeling- or because discriminating in favor of
this attribute is socially acceptable. On the one hand, whether discrimination in
favor of an attribute is socially approved or disapproved is a datum to economists,
just as we, usually assume that preferences are given. Economists can still be
useful if, after being informed of which attributes lead to socially disapproved
discrimination, they are able to predict behavioral consequences and, ideally,
suggest cost-effective remedies. On the other hand, inattention to the nonmone-
tary pain felt by the victim of certain types of discrimination will limit the
economist's contribution to social welfare and policy analyses (to be discussed in
the concluding section).
Economic analysts have generally concluded that consumer-based discrimina-

tion plays a minor role in the differences in average wages received by race and
sex groups. The reasoning is as follows. Assume that black workers have the same
distribution of productive skills as white workers and that consumers (who are
predominantly white) are willing to pay a price, p, for a good produced by white
workers. If, however, there is customer contact with the producers, the consumers
consider the effective price for a good produced by black workers to be p' = p + d,
where p is the cost of production and d is the monetary value of a white
consumer's distaste for contact with black producers. (For convenience, assume
temporarily that all white consumers have identical tastes.) Clearly, most goods
and services are not produced with customer contact. Thus, consumers would not
discriminate against, say, clothing or automobiles according to the color of the
workers in clothing or automobile factories. For these goods the price would
simply be p, regardless of the color of the workers.
Black workers, therefore, would specialize in the production of goods with no
customer contact and, in so doing, avoid being paid a wage lower than that of an
equally productive white worker, which would be the outcome if they competed
with whites in, say, retail selling. 9 If the concentration of black workers in
industries with no customer contact were to depress wages in these jobs, then
white workers in these jobs would move-horizontally by skill l e v e l - i n t o jobs
with customer contact until wages were equalized in the two sectors. Given that
the number of black workers is small relative to the number of jobs that have
customer contact, all black workers would be in jobs that have no customer
contact. (Realistically, some would be in the jobs with customer contact that
involve nondiscriminating customers, now recognizing that consumers have vary-
ing tastes regarding contact with black workers.) The result is some degree of job
segregation but no group difference in prices for labor services.
The assumptions that lead to this outcome are sufficiently plausible that
consumer-based discrimination has not been assigned an important role. The
market measure of discrimination, D, equals zero, even though consumers are
prejudiced and job assignments among workers are affected. Thus, Becket's
formulation provides the useful distinction between an ith individual agent's
tastes for discrimination (with potentially varying d / s ) and market discrimina-
tion, which is an aggregate that is not the sum of its parts; here, D = Pmaj - Pmin"
Discrimination, D, disappears even though ~_,d i > 0, simply because workers, in
their quest to maximize their utility, will move and bring about some degree of
segregation.
An outcome in which segregation reduces or eliminates market discrimination
occurs in several versions of Becker's model. For this reason, Welch (1975) called
Becket's theory a theory of segregation, not discrimination. Welch's point is
9Specialization, which is here associated with segregation, has been rigorously analyzed as a means
for attaining nondiscriminatory outcomes in terms of factor payments by Stiglitz (1973, 1974).
712 G. G. Cain
partly semantic, but his insight is useful and may be explained briefly as follows.
The source of market discrimination in Becker's model is on the demand
side - the willingness of an economic agent to pay to avoid contact with members
of a specific group. In a competitive model there are many employers and free
mobility among economic agents, so competition enables segregation to satisfy
this demand costlessly. The model assumes that mobility is costless (or nearly
costless), especially in the long run.
Segregation is, therefore, a means for eliminating market discrimination, but it
is not the only means. Collective action to offset the effects of discriminatory
tastes or changes in those tastes can be accomplished without seriously restricting
competition in markets. Indeed, common sense and casual observation indicate
that an integrated society is generally more competitive. It is tempting to point to
the Republic of South Africa to illustrate that segregation is not a sufficient
condition to eliminate discrimination, but this country's experience is inap-
propriate for illustrating a competitive model. Lundahl and Wadensj6 (1984, pp.
209-260) explain how a century-long pattern of private and governmental
collusive arrangements have restricted competitive forces in the South African
economy, with the undisguised purpose of concentrating wealth and power in the
hands of the white population.
3.1.1. Discrimination by workers
Assume all workers have the same skill level. If all majority workers (whites) are
prejudiced against minority workers (blacks), we may assume a white worker's
wage demand for working with other white workers is w, and his wage demand
for working with black workers is (w + d). Clearly, employers of white workers
would employ segregated work forces to pay the lower wage. Equally skilled
black workers would also receive w as a consequence of competition among
employers, mobility by workers, and the previously established sterilization of
consumers' preferences. Integrated work forces could exist among unprejudiced
white and black workers, so the worst case is when all white workers have tastes
against working with black workers. But even the worst case yields only segrega-
tion among workers, not discrimination as defined by ~maj > ~r~in-
One could postulate various impediments to competition. For example, per-
haps segregation will not permit equal wages because the black workers are too
few to allow economies of scale in production, recognizing that their numbers
must staff~,~!l skill levels. Rebuttal: Aside from examining the structure and
technology of industrial organization to determine the plausibility of this, we
should recognize the flexibility in large-scale organizations to use compartments,
work in shifts, form subgroups, provide on-the-job training, and so on to achieve
"effective" segregation of the workers. Remember, segregation is cost minimizing
when white workers are the discriminatory agents.
Another example: Assume that black workers migrate into a region populated
exclusively by prejudiced white workers. Efficient, segregated firms might take a
long time to become established. Hiring and training workers entails fixed costs
and, as Arrow has analyzed, these costs will retard any attempt by a firm to hire
an all-black work force [Arrow (1973, pp. 20-23)]. A rebuttal should not be
required because the example, although empirically relevant and interesting,
should lead to a long-run equilibrium in which the work force is segregated.
Another example: Let skills vary among workers and assume that black
workers have a legacy of low skills upon entering the labor market. White
workers with equally low skills receive w'. Assume the technology of efficient
production requires that low-skilled workers combine with complementary high-
skilled workers, all of the latter being prejudiced white workers. Black workers
must then receive W~i n < W~aj to compensate for (offset) the high labor costs they
"impose" on their complementary factor of production - the white skilled workers.
Rebuttal: Some black workers would have a particularly strong incentive to
become skilled. Those who match the skilled white workers in innate ability
would not only have the incentive to seek the normal (i.e. white workers') rate of
return on a skill investment, but they could earn extra profits by working with
low-skilled fellow black workers, because the discrimination tax, d, will not apply
to them. To see the incentives involved, we can imagine that these tax savings
could be shared among both skill levels of black workers and their employer, and
any one of these agents would have an incentive to initiate this process.
Eventually, as more black workers become skilled, the underlying source of the
(WnPaaj -- W~i n) gap withers away. Again, this scenario could take a long time, and
it may be empirically interesting. Finally, I argue in the next section that
complementary skilled white workers correspond to employers as agents of
discrimination, so the conclusions about employers also apply to complementary
skilled workers.
3.1.2. Discrimination by employers
Two versions of employer-based discrimination in competitive markets were

advanced by Becker in his analysis of racial discrimination. The first, hypotheti-
cal and pedagogic, assumed that employers all have the same prejudice against
black workers (or in favor of white workers), so a uniform lower demand for
black workers sets their market wage at ( w - D). Thus, the white workers' wages
and their monetary labor costs are higher. Competition in the product market
requires a uniform product price, but this can be achieved by the differential in
money labor costs being compensated by a differential in money profits, which,
in turn, is compensated by a differential in psychic benefits (or psychic costs - the
difference depending on whether one emphasized the employer's psychic benefit
from employing a white worker or the psychic cost from employing a black
714 G. G. Cain
worker). The psychic and money forms of profits (or employer compensation)
offset one another in equilibrium.
Other analysts suggested modifications of this model. Arrow (1972) obtained
useful insights from an assumption that the employer's discriminatory tastes were
an increasing function of the ratio of black-to-white employees, rather than being
a constant that was independent of the racial composition in the firm. Arrow
(1972, p. 89), Marshall (1974, p. 853), and Thurow (1975, p. 162) suggested that
distaste may depend on "social distance" rather than "physical distance". If true,
this would make empirical measurements complicated. For example, an em-
ployer's d might be zero for janitors but have a large negative value for
professional employees. Indeed, if the owners of capital have little or no contact
of any kind with the employees, the model would require that the discriminatory
role shifts from employers-as-capitalists to their agents, such as managers,
supervisors, foremen, or even skilled workers-all of whom are assumed to be
prejudiced white persons. These interpretations of employer discrimination add
realism to the model, but they do not negate Becker's central point, which was
the establishment of an equilibrium differential in favor of white workers.
In a second version of Becket's model of employer discrimination in a
competitive economy, tastes among employers were permitted to vary. Consider,
first, the special case of just two values of d,: low, d~, and high, d 2. Clearly,
employers with the lower value, dl, would hire all the black workers. (I will
temporarily assume that there are enough d I employers to hire all the black
workers.) The market wage differential between white and black workers under
this regime would be D 1 = d 1, a smaller differential than the average: ( N l d 1 +
N2d2)/(N1 + N2) , where N 1 and N 2 are the numbers of employers in the two
categories. Indeed, the size of d 2 is irrelevant.
Becker's insight from this model is that black workers generally benefit by a
dispersion in d~. A wider spread in the distribution of d~ could only narrow the
wage gap, assunfing some of the increased variance stretches the lower tail of the
distribution and lowers the value of the d~ of the employer with the highest d~
required to hire all black workers. Intuitively, the upper tail is irrelevant in the
setting of D because the employers with larger tastes for discrimination, d 3 > d2,
d 4 > d 3, and so on, do not bid for minority workers and they have no incentive to
pay more than the existing w for majority workers. In contrast, a widening
spread in the lower tail means that the new employers, with tastes d o < d l , would
now hire all the black workers. They increase the demand for black workers, and
the market,~differential in white and black wages becomes D O< D 1.
Two plausible extensions of the dispersion effect, as just described, will tend to
eliminate market discrimination entirely.
(1) First, the lowest value of d, call this d o, would determine the market wage
differential, even if only a small number of employers-in the limit, one per
product per market- had a value of d i as low as d 0. Clearly, this employer would
earn extra profits by hiring minority workers, benefitting monetarily from the
lower wage they receive while escaping all or some of the psychic costs that
would be experienced by employers with higher di's. Total profits could be
increased by cutting prices and hiring more black workers and expanding
production. Employers with d~ > d o would, correspondingly, lose business and
curtail production, thereby decreasing the demand for white workers. The imper-
sonal operation of the capital market would ensure an inflow of investment to the
high-profit firms. Assuming long-run constant costs, the stopping point would be
reached only when all black workers and equally paid white workers are
employed by the d o employer(s)-perhaps in newly constructed plants, each of
optimal size. White workers would lose the wage advantage they had received
from discriminating employers.
(2) Second, D O would become zero. There are several routes by which the
market should uncover one or more cost-minimizing employers (per product, per
market) with d o = 0. Some white employers might be unprejudiced. Blacks could
become employers. Capital owners, like consumers, tend to be remote from
contact with employees, so their d i's would tend to be effectively zero. (Of course,
this shifts the cost-minimizing problem to that of finding managers with low d/s.)
Indeed, consumers as well as investors would have precisely these incentives of
finding managers and other forms of complementary employees whose d o = 0.
In a phrase, competitive market forces, still assuming constant costs, tend to
drive D toward zero. Arrow, in his analysis and reformulation of Becker's model
of employer discrimination, arrived at just this conclusion: "Only the least
discriminatory firms survive. Indeed, if there were any firms which did not
discriminate at all, these would be the only ones to survive the competitive
struggle" [Arrow (1973, p. 10)]. And, "It [Becker's model of employer discrimina-
tion] predicts the absence of the phenomenon it was designed to explain" [Arrow
(1972, p. 192)].
Becker, in an article on discrimination written for the International Encyclo-
pedia of the Social Sciences and published 11 years after his book, did not reach
this conclusion.
A few of the more extreme nineteenth-century advocates of a competitive
market economy believed that eventually its extension and development would
eliminate most economic discrimination .... Unfortunately, this has not yet
taken place; discrimination exists, and at times even flourishes, in competitive
economies, the position of Negroes in the United States being a clear example
[Becker (1968, p. 210)].
Becket's disagreement with the previous scenario of the workings of competi-
tion is based on his view that the assumption of constant costs for a firm, even in
716 G. G. Cain
the long run, is a polar case and not one to be accepted generally [Becker (1957,
rev. ed. 1971, pp. 44-45)]. Entrepreneurial skill is an example that is sometimes
suggested for a factor of production that may be inelastically supplied, even in
the long run. Thus, one's judgment about the number of nondiscriminating firms
that are in or that might enter the market, about the generality of entrepreneurial
skills, and about the long-run elasticity of other factors all enter into one's
judgment about the persistence of a discriminating cost differential in the long
run under competitive conditions.
What if discrimination is redefined as nepotism and d = d b < 0 is replaced by
a term d w > 0, now adding subscripts to distinguish discrimination against blacks
from nepotism in favor of whites? This specification is examined by Goldberg
(1982), who finds that a long-run differential wage advantage in favor of whites is
sustained under competitive conditions. The result, which had been previously
advanced and then downplayed by Arrow (1972), is correct, but in my judgment
the model is not realistic. 1°
My argument begins with the observation that when a positive d w - Goldberg's
nepotism-replaces a negative d b -Becker's discrimination-the intention is to
view the tastes for whites as more than a euphemism for expressing a preference
not to be associated with blacks. This intention is clarified by a dictionary
definition of nepotism: "favoritism shown to one's nephews and other relatives;
bestowal of patronage by reason of relationship rather than merit". As defined,
nepotism is indeed real. Let us assume that only the "uncle-employers" receive
nonpecuniary utility from the employment relation. Consider two cases of wage
payment. In Case 1 the wage rate of "nephews" and all other workers is the
same, and nephews are merely sorted into jobs where their uncles are employers.
Alternatively, in Case 2 the uncles share all or some of their utility rents with
their nephews by paying them a higher than competitive wage. In Case 1 the
uncles earn extra rewards (profits plus utility), but they have no incentive to
expand production, which would (assuming constant costs) threaten other firms,
because the supply of nephews is sharply limited. In Case 2 the uncle-employers
earn lower profits, but their total utility can easily be high enough to ensure their
survival as employers.
Case 2 shows, therefore, that the dictionary definition of nepotism can coexist
with the economic definition of nepotism, according to which nephews receive a
higher wage than equally productive nonnephews (all other workers). However,
to transfer this scenario of nepotism to one in which all white workers, who
~°See Arrow (1972, pp. 91, 192). After pointing to nepotism as a source of a sustained wage
differential in favor of whites, even in the absence of any differential based on tastes against blacks,
Arrow commented: " B u t it is reasonable to postulate that any preference a firm might have for the
hiring of whites per se arises as an offset to the presence of disliked blacks. That is, for a firm that has
no black employees, d~ = 0." Furthermore, "for a firm that does not discriminate against blacks,
there will also be no reason to pay anything extra for white employees" (p. 192).
constitute 85 percent of the labor force, are the equivalent of nephews (beneficia-
ries of nepotism) seems unrealistic. Throughout this survey, therefore, discrimina-
tion against a minority group will be viewed as the operative force.
3.2. Nonstochastic monopolistic neoclassical models: Product monopoly
A monopoly has two characteristics that permit long-run discrimination: first, a

definitional uniformity in tastes, since there is only one employer; second,
above-competitive profits. The former allows a d i that will not become irrelevant
because of competition, and the latter allows the sacrifice in money profits-in
exchange for the psychic benefits from discrimination. Nevertheless, there are
several influences in the economy at large that constrain or even eliminate the
power of one or a few monopolies to sustain market discrimination.
Monopoly power in the product market does not imply monopoly power in the
labor market. If the monopoly finn cannot affect wages in the labor market, it
would not pay a higher wage than w to hire majority workers, nor could it pay a
lower wage than w to hire minority workers. In other words, the monopoly would
not be the source of discrimination, although it, like other firms with a positive
di, would employ a segregated, all-majority work force. Were the monopoly to
behave irrationally and pay higher wages to majority workers, it would create
incentives for a "takeover" by investors and managers with zero d~'s. Indeed,
Alchian and Kessel (1962) advanced the view that even where monopolists affect
wages in their labor market, they would be unlikely to sacrifice money profits
permanently by a policy of (racial) discrimination, because profit-maximizing
investors would buy them out.
But why do monopolistic enterprises discriminate.., more... ? One would ex-
pect that those who have a taste for discrimination.., would naturally gravitate
to those economic activities that, for purely pecuniary reasons, do not employ
Negroes. Free choice of economic activities implies a distribution of resources
that would minimize the costs of satisfying tastes for discrimination (p. 161).
Alchian and Kessel pointed out that a regulated monopolist or a government
monopoly, which was constrained not to maximize profits, could indulge its
tastes for discrimination at no loss in profits and, therefore, offer no incentive for
a "takeover". Such firms could, for example, engage in nepotism and consume
other nonpecuniary benefits at no cost in forgone profits, and if there were
enough such firms, they could at least contribute to a marketwide discrimination
differential.
It is useful to keep in mind two empirical characteristics of monopolies-now
using the term as shorthand for a firm that produces a "large" share of the
market. First, monopolies tend to be larger, more capital-intensive, and more
718 G. G. Cain
likely to be unionized than the average firm. Because of this, they may pay higher
wages to attract specialized skills and to ensure lower turnover. Among the
workers who apply for jobs at these monopoly firms, majority workers may be
the more skilled, as a result of previous discrimination from various channels.
The resulting combination of hiring relatively more majority workers and paying
higher wages may not be discriminatory; that is, it may be consistent with a
d i = 0 for the monopolist. In principle, a properly specified Model (I) would
permit testing whether the firm really discriminated among equally skilled
applicants, minority and majority.
Second, along with size and wealth, monopolies are often also publicly promi-
nent. They tend to be sensitive to public relations and to their "image". In the
past this sensitivity could have served to reinforce discrimination, because
government and other wielders of power in the community may have been
prejudiced and have influenced the monopoly. Today, our laws and professed
public sentiments are against discrimination or, if neutral, condone organized
pressures from minority groups on the monopolies. These forces would, if
present, tend to lower the effective d i of the monopolist below the average among
employers.
In summary, monopoly firms, particularly regulated monopolies, are in theory
capable of exerting some sustained discrimination in labor markets. There are,
however, reasons for doubting that monopoly is a major source of marketwide
discrimination.
3.2.1. Monopsony firms in labor markets
The classic case of the exploitation of labor in neoclassical economics arises

under conditions of monopsony. Workers are captive in a market where there is
only one employer, or where a group of employers collude and act as one buyer.
M o n o p s o n y represents a rare area of common ground between neoclassical and
Marxian models of the labor market, u
The model is well known: a single buyer of labor faces an upward-sloping
supply curve of labor; equates the value of labor's marginal product (VMP) and
its (rising) marginal cost; hires less labor than if the same demand for labor were
generated by many competing firms; pays labor its supply (offer) price, which is
lower than the price (wage) needed to induce the larger supply under competitive
demand conditions; and retains the positive differential between the VMP and
the wage~ as profit. Where two factors of production are supplied and demanded,
the exploitation (measured by [ V M P - w ] / w ) will be greater for the factor whose
labor supply is the more inelastic. These propositions, which were presented by
II See Lundahl and Wadensj~ (1984, pp. 49-52) for a further analysis of monopsony models of
labor market discrimination and for their critique of neo-Marxist, or radical, theories as a subset of
monopsony. See Cain (1976) for a brief discussion of radical theories of the labor market, including
the analysis of discriminationby radical theories.
Joan Robinson (1934, pp. 301-304), provide a consistent model for discrimina-
tion simply by postulating a more inelastic supply curve of labor for minority
workers. A modem application of this model is by Madden (1973).
Empirical support for the prevalence of monopsony and lower-than-competi-
tive wages is limited [see Bunting (1962)]. Labor markets that are "one-industry
towns" are increasingly uncommon, mainly because a large fraction of the
population lives in larger urban places and because the automobile has greatly
expanded the geographic boundaries of the labor market. Information about
wage rates in geographically dispersed markets is available and only those
workers "on the margin" of moving need to move to equalize wages for workers
of comparable skills. Therefore, the long-run acceptance by workers of below-
competitive wages presupposes a degree of immobility that is hard to accept. No
doubt there are some workers who are trapped by a combination of industry-
specific skills and a decline in the number of firms competing for their skills, and
who suffer long-lasting exploitation. But these are not conditions that generalize
to the entire labor market.
Because monopsony seems to have a limited application, it does not appear
worthwhile to examine more closely the requisite proposition that the supply
curve of minority workers is less elastic than the supply curve of majority
workers. However, two brief points may be useful. First, if differences in the
supply curve identify (in the econometric sense) a difference in exploitation, we
need to satisfy ourselves that the underlying sources of this difference in supply
curves are not also reasons why the workers' wages differ.
Second, regarding gender discrimination, there is a good deal of empirical
evidence and theoretical support for the finding of a greater elasticity for the
supply curve of women's labor than of men's labor. To be sure, this larger
elasticity refers to the market, not to individual firms, but as a firm (or group of
firms) becomes monopsonistic then the distinction between the supply of a factor
to the labor market and the supply of a factor to the (monopsonist) firm tends to
disappear. Thus, the larger labor supply elasticity of women in the labor market
as a whole implies a larger elasticity to a monopsonist, and this is the opposite of
the requisite condition for the exploitation of women relative to men. Again,
there may be particular circumstances when this generalization does not hold.
Nurses are sometimes used as an example of an occupation that faces a
monopsony-employer in the form of one or a few hospitals.
3.2.2. Labor unions as monopolies
In Becker's model of discrimination, white workers' prejudice against black

workers was not a sufficient condition to sustain a discriminatory wage differen-
tial. However, by forming a monopoly in the sale of labor to employers, white
workers could enforce their tastes and raise their wage above the competitive
level. Moreover, unlike monopsony, labor unions are widespread, supported by
720 G. G. Cain
laws and community approval, and have been shown in many studies to have
raised wages for their members above competitive levels.
Given that the union secures monopoly rents, some method of restricting entry
is a necessary first step in maintaining these rents. Many analysts have pointed to
the discriminatory tastes of the members as a criterion for inclusion and
exclusion. Kessel (1958) added the argument that this criterion will also be useful
in a second step in maintaining the rents; namely, in policing the existing
members to honor the union contract, even though it would often be in their
private interest to "cheat" by, say, working more for a slightly lower wage. Kessel
argued that ethnic homogeneity among the members facilitates a mutual agree-
ment to collude, making unnecessary those stronger sanctions that might be
illegal or incur community disapproval. Finally, institutional research, while
divided about the overall discriminatory impact of unions, documents many cases
of discrimination by unions [Gould (1977), Hill (1977), Marshall (1965), Ross
(1948), Northrup (1944), among many others]. Thus, the a priori case for unions
as a source for labor market discrimination appears substantial.
There are, however, a number of counterarguments. First, unions have never
organized a majority of the labor force in the United States, and before 1940
there were few periods during which more than 15 percent of the work force was
covered by collective bargaining contracts. The wage gap between blacks and
whites was larger in the pre-1940 period, although this fact by itself does not
provide direct evidence on the influence of unions on the wage gap. In 1977, only
around 25 percent of the labor force were union members or were covered by
collective bargaining contracts [U.S. Department of Labor (1979)].
Second, membership in unions is more common among blue-collar workers,
which points to a disproportionate representation among men and blacks,
although within the blue-collar ranks membership is more common among skilled
occupations, which points to a greater representation among white men. A larger
proportion of black men were members of unions in 1977 than were white men
[U.S. Department of Labor (1979)].
It is noteworthy that the few industries and occupations where unions have
grown in recent years-governments, teaching, hospitals-are disproportionately
composed of women or blacks. Ashenfelter (1972), whose study will be examined
in the next section, concluded that the white-black wage gap among men was
actually narrowed by unions as of the mid-1960s. The male-female gap was
slightly widened. His study is persuasive that labor monopoly, despite many
individuaL,.cases of discrimination by unions, is not a ,najor source for the
observed discriminatory differentials.
3.2.3. Government as a monopolist
Governments are universally monopolists in certain functions, such as providing

for national defense, police and fire-fighting, and mail services, and, most
importantly, as law-maker. With their power to tax and to punish, governments

possess more potential monopoly power than firms and unions, although the
collaboration between government and private agents may make it difficult to
isolate the source of power. Moreover, governments, unlike private monopolies,
need not be and seldom are guided by profit maximization goals. Granting that
the majority group controls the government, there is no analytical challenge to
demonstrating a theoretical case for discrimination based on government behav-
ior. Malcolm Ross, the director of the Fair Employment Practices Commission
during the 1940s, provides an exanaple of a government law that, if it did not
impose wage discrimination against blacks, at least impeded its demise. Ross's
example also illustrates one expert's skepticism about the "physical-distance"
theory of discrimination.
White and Negro workers are now [1948] and have been for decades under the
same plant roofs in the South. It is not the working associations to which the
whites object. It is the sharing of skilled wage rates .... South Carolina... re-
fuses by law to permit skilled Negro textile workers in the same plants with
whites. But that state statute (probably unconstitutional) does permit Negro
janitors and charwomen to work under the same textile plant roofs as whites.
What would you say, then, that that South Carolina law is protecting-white
workers from association with Negroes, or white jobs at the looms at white
wages? [Ross (1948, p. 307)].
The scope, history, and literature of the government's influence on labor
market discrimination are far too extensive to survey in this paper. Some
discussion about government policies is reserved for the final section.
In this chapter I generally assume that government agencies do not have
pervasive monopoly power regarding labor market discrimination, and that,
historically, their interventions in the market have had many, but more or less
offsetting, effects. In recent decades the intention of government policies has been
to reduce discrimination against minorities, but the analysis is complicated by the
claims that some actions, despite the beneficial intentions, turn out to worsen the
problem. This criticism is frequently made about minimum wage and equal pay
legislation. One fact and two theoretical-empirical points set the stage for this
criticism.
Fact. The minority group is disproportionately represented in the lower tail of
the distribution of productive skills, not because of an innate inferiority but
because of a legacy of past inequities and prelabor-market discrimination.
Theoretical Case 1. Model (I) applies with the competitive result that minority
workers receive an average pay equal to their average abilities (defined by X).
However, the minimum-standards law truncates (from below) the distribution of
X's among the work force, and relatively more members of the minority group
are disemployed from the jobs covered by the legislation. Over the full distri-
bution of the work force, minority workers are worse off, either because of their
722 G. G. Cain
excess unemployment or because those disemployed from covered jobs are

crowded into lower-paying jobs in the uncovered sector. Note that this case does
not require any tastes for discrimination, although they would exacerbate the
minorities' disadvantage (see Case 2).
Theoretical Case 2. Model (I) applies and majority workers receive a higher
wage, conditioned on X, implying A > 0 and the existence of market discrimina-
tion. A minimum wage, w, can impede the competitive forces that encourage
hiring lower-wage minority workers. Employers who might hire minority workers
at a lower wage, w - D, are prevented from doing so. Case 2 does not require
minorities to be concentrated in the lower half of the productivity distribution,
but lower-skilled minority workers face the highest risk of being without a job.
They are also prevented from competing for jobs that offer general on-the-job
training by bidding for them with lower starting wages.
To illustrate either of the two cases, consider the following historical event,
described by Ross (1948).
During the First World War the Southern [railway] carriers lost a serious
number of skilled workers to the services and munitions plants. In order to
make it attractive to Negro workers to stay on the job, Secretary McAdoo as
wartime transportation chief ruled that Negro railwaymen should receive the
same pay as whites for the same work. This 1918 move was called "a simple
act of justice," and so it was, although the far [long-run] results were anything
but just.
Forced to pay them the same wages as whites, the carriers lost interest in
Negroes as a cheap labor supply. The white workers, for their part, began to
covet the better Negro jobs. The McAdoo ruling had laid the foundation for a
coalition between the carriers and the unions against Negroes in firemen's and
other high-bracket positions (p. 119).
Blacks were driven out from these positions, but as Ross makes clear, the
government ruling was only one part of the causal chain. Also contributing to the
outcome were employer and worker prejudices, a quasi-monopolistic industry,
the antiblack environment of the South, and a labor union.
3.3. Stochastic neoclassical models and statistical discrimination
The theoretical challenge developed in the preceding discussion of neoclassical

models is to rationalize unequal pay to groups of workers who are equally
productive. The comparison between groups was intended to allow within-group
individual deviations from the equality between productivity and pay, which is
necessary if the model of pricing is to apply to the real world. However, this
stochastic feature was suppressed throughout the discussion, because the use of
average values of the wages for comparisons between groups made the models
equivalent to exact or nonstochastic models.
Attention to a stochastic model of wage determination, in which the worker's
value to the employer is not known with certainty, offers several new insights,
and more possibilities for sustained (or, at least, long-lived) group discrimination.
Whether these theories are more or less persuasive than any of the others is a
matter for judgment and empirical study.
Phelps (1972), Arrow (1972, 1973), and McCall (1972) were early authors. It is
convenient to analyze the following model of wage determination, which is due to
Phelps. Let qi be the ith worker's true productivity, which is unknown to the
employer, who must rely on some observed but imperfect indicator, Yr The
indicator may be a test score or a variable, like years of schooling, that has a
more direct connection to productivity. The notation and details of the model
below are shown in Aigner and Cain (1977), along with citations to various
authors and statistical references.
In a simple specification that brings out the main conclusions of the approach,
the relation between y and q (subscripts dropped) is
y = q + u, (1)
with E ( u ) = C(q, u) = O, E ( y ) = E ( q ) = a, V(u) = a 2, using the familiar sym-

bols for expectation, covariance, and variance. By assuming q and u are
joint-normally distributed as well as uncorrelated, we may specify a linear
regression function for the reverse regression:
q=a(1- y)+ yy+e, (2)

with e a well-behaved disturbance. Here 3' is the coefficient of determination (r 2)
between q and y; thus, 0 < 7 < 1 , and "/ measures the "reliability" of y as a
measure of q.
Assuming employers pay workers according to their expected productivity,
then
w = E ( q [ y ) = a ( 1 - 7 ) + 7Y- (3)
Equations (2) and (3) reveal the obvious point that individual discrimination,
defined as unequal pay for equally productive workers, is inevitable, given the
error component, e. In contrast, group discrimination does not follow from this
model precisely because e is considered random and has an expected value of
zero for minority and majority groups.
Letting subscripts 0 and 1 refer to minority and majority groups, eq. (3) may
be applied to each group. Assume temporarily that the minority and majority
724 G. G. Cain
groups have the same mean true productivity: a = a 0 = at, and that we compare
workers with the same y-score. If we further assume that V(q) is the same for
both groups but that Vo(u ) > Vl(U ), reasoning that the test instrument is more
unreliable for the minority group, then ~1 > ~0, and we have
w, - w0 = (y - (4)
Accordingly, for a given y-score [roughly corresponding to "holding X constant"

in Model (I)], majority workers receive a higher wage than minority workers for
y-scores above the mean, a, and lower wages for y-scores below the mean. Thus,
group discrimination, defined by E(w 1 - Wo) > 0, is not present.
Clearly, postulating a lower a for minority workers would lead to their being
paid a lower wage, but a lower wage for a given y-score, assuming y is a valid
indicator of productivity (about which, see below), would not imply economic
discrimination for the group because, on average, the minority and majority
workers continue to be paid in accordance with their average productivity. In
Figure 13.1, parts (a) and (b) show two cases for unequal a's. In part (a), where
T1 = To, the difference a~ - a 0 is evenly distributed across all y-scores. In part (b),
where T1 > To, we see that the minority workers with high y-scores who are paid
" t o o little" relative to majority workers with the same y-scores are balanced by
the low-scoring minority workers who are paid " t o o much", relative to majority
workers with the same low y-scores. As drawn in part (b), minority workers with
y-scores below y ' receive relatively higher wages than majority workers with the
same y-score.
Nevertheless, a number of economists have claimed that this model reveals,
and offers an explanation for, group discrimination. Let us examine two applica-
tions of the model. Only the second shows discrimination that is consistent with
the definition adopted in this chapter.
3.3.1. Statistical discrimination, but spurious group economic discrimination
Thurow (1975) is one of many economists who use the term "statistical dis-
crimination", when there is presumptively no economic discrimination. In the
following example, Thurow accepts the facts of (a) a higher probability of market
work by men compared with women and (b) the benefit to an employer of the
higher probability. He then says:
Any employer faced with these differences in work probabilities will practice
statistical discrimination even though there are millions of women who will be
in the full-time paid labor force for their entire lifetimes. Ex ante, he cannot
tell which women will be lifetime year-around full-time employees and which
women will leave the labor force or become part-time employees. Because the
employer provides on-the-job-training, he will want to invest in those who are
more likely to stay in the full-time labor force. If he provides training to
Ch. 13." Labor Market Discrimination 725
(a) 71 = Y0; el > a o
W= q
/
/
/
/
/ / w l = a1(1 - 7) + YYl
wo = %(1 - )1) + ')'Yo
/
/
~' Q'I
(b)[ 71 > Yo; 0'1 > °to

A
w=q I /
/
/
/
///~,,,/w~ = a1(1 - 71) + YlYl
/ . / ~ ~ w o = %(1 - 70) + "YoYo
y' % a
Figure 13.1. Predicted value of productivity (q) by indicator (y) for majority (1) and minority (0)
workers.
726 G. G. Cain
women, he is less likely to be able to recoup his investment . . . . The woman

who will participate in the paid labor force her entire lifetime is being treated
unfairly . . . . The net impact is discrimination against women as a group and as
individuals even though there is not a basic taste for discrimination against
women (p. 178).
Two points should show why this example does not imply group economic
discrimination. First, Thurow correctly indicates that the women who will
participate in the labor force their entire lifetimes are being treated unfairly and
will be underpaid. The employers cannot know an individual's future, and they
will base their wage offer partly on a 0 ( 1 - - / ) - t h a t is, partly on the known
average for all women. But this is only half the story. Women who will
participate for only the briefest period will be overpaid. As before, the employer,
not knowing these women's true low probability of working, will rely upon the
average for all women and overpay them. On average, the over- and underpay-
ments tend to cancel out. Whether the resulting average is equal to the average
for men will depend, as the next two paragraphs suggest, on whether the gender
difference (here, a commitment to full-time work) is related to productivity.
Second, suppose all the workers are the same gender, that the two groups
under study are persons with a college education and persons with less than a
college education, and that the former have a higher probability of working on
average. Thurow's entire passage could stand intact with the phrase, "persons
with less than a college education", substituted for "women". Most analysts
would agree that Thurow's case for group discrimination, even with the less-
educated group earning less on average, loses its plausibility with this substitu-
tion.
Thurow's example inadvertently raises another interesting issue. The y-indi-
cator in the stochastic model of wage determination is assumed to be unbiased on
average, even though its reliability may differ for minority and majority workers.
When, however, the y-indicator reflects discrimination, the model is no longer
appropriate for an explanation of discrimination. In Thurow's example, the
probability of working is, or could be, a reflection of discrimination. Clearly, if
women or other minorities are discriminated against by not being employed, it is
unsatisfactory to use the low probability of employment as an explanation for
discrimination in the form, say, of lower wages or some other labor market
outcome. This point will be discussed in Section 4, and here it serves to remind us
that the choice of a y-indicator is not innocent.
3.3.2. Statistical discrimination and actual group discrimination
The discussion of the stochastic model up to now has not allowed the unreliabil-
ity of the indicator to influence the average wage. Aigner and Cain (1977)
stipulated risk aversion in the employer's utility (or profit) function and ra-
tionalized a lower average wage payment to minority workers as compensation

for this undesired unreliability. A more convincing rationalization was suggested
by Rothschild and Stiglitz (1982), who specified a production function that
depended directly on matching the worker's q with a job assignment. In
particular, both undermatching and overmatching were inefficient, so the ex-
pected output, not merely its variance, depended on matching.
Either formulation may be viewed as redefining the productivity of workers to
include both the workers' physical productivity and the information workers
convey about it. Does rewarding a group for their better information constitute
economic discrimination against the group with less complete information?
Perhaps the answer depends on the fairness of the testing system and, like the
issue of the existing technology (see pp. 698- 699), on how costly it is to change,
and on whether its existing inadequacies for minorities reflect some market
failure. The important role of the government in educating, training, certifying,
and licensing workers suggests that improvements in testing minority workers
may be a public good. (In fact, improvements in testing all workers may be a
public good, but I focus here on discrimination between groups.)
If wage differentials are large merely because of differential test reliability, then
both minority workers and employers have incentives to improve the tests and
reduce this impediment to transactions. If, as is sometimes reasonable to assume,
the worker knows his or her own abilities, a low-cost private-exchange method of
minimizing this impediment is for workers to offer a trial period of employment
to demonstrate their true productivity. The cost to the worker is a low wage
during the trial period, but the benefits are higher earnings throughout the
worker's subsequent career.
A trial period of working is also a device for minimizing the private and social
costs of "signaling", as the term has come to be used following Spence (1973).
Using Spence's model, we may assume that the test or y-score (a) has no value
other than to indicate (signal) the worker's productivity, (b) is costly to obtain (as
when the signal consists of an educational degree), and (c) is more costly to
obtain for less productive (less able) workers. These assumptions imply that
workers will choose whether to invest in the signal on the basis of their
knowledge of their ability and on whether the extra pay the signal earns for them
will justify its investment costs. Employers adapt to this maximizing behavior of
workers by believing the signals and making their wage offers accordingly.
In Spence's model there is no guarantee that the equilibrium allocation of
signaling investments among workers and, correspondingly, of workers to jobs is
socially efficient, because only a "justifying" benefit/cost structure and not an
"optimizing" one is required for an equilibrium. There is a tendency for too
much investment in signaling; that is, Pareto-optimality could be achieved with
less. Without an optimizing equilibrium there is no guarantee of a unique
equilibrium. With multiple equilibria, the door is open for a benefit-cost struc-
ture that is unfavorable to a minority group compared to the majority group.
728 G. G. Cain
How robust is this discriminatory equilibrium? Even if one did not have faith
that the competitive market would facilitate efficient signaling instruments and
institutions, there remains the previously mentioned method of trial work periods
based on deals struck between individual workers and employers. The strategy
assumes that if the workers know enough about their ability to choose whether to
invest in the signal, then they can use this knowledge to offer to work for the
employer for a trial period. The strategy is better able to eliminate the Spence
type of discrimination than it is to eliminate the Spence type of social in-
efficiency. Discrimination is eliminated if the cost to minority workers of the trial
period is no higher than the cost of the majority worker's signal, even though
these costs may still be higher than the socially efficient level. 12
A recent paper by Lundberg and Startz (1983) uses certain features of both the
Phelps model of unreliable indicators and the Spence theory of signaling. They
derive a market failure in investment, although in contrast to Spence, too little
investment occurs rather than too much. Their argument may be conveyed by
reference to a commodity. Assume the commodity is produced at less quality
than would be optimal, because the information about its quality cannot be
conveyed perfectly. Specifically, the quality improvement could be produced at a
cost that is less than the benefit, if only the quality improvement were accurately
conveyed. Because the quality is imperfectly measured, however, consumers will
discount the quality signal and will pay less than the costs of the optimal amount
of the quality improvement. The situation is the same as in Model (3) above:
employers pay ~, ( < 1) for a unit more of y, instead of paying a full unit more as
they would if y were a perfect measure of q.
Lundberg and Startz apply this argument to two groups of workers, minority
and majority, and show that a less reliable signal for minority workers will lead
them to underinvest relative to majority workers. Their general conclusion of
underinvestment is opposite to that of Spence because of their contrasting
assumptions about the benefits and costs of the investment. For Spence all or
part of the benefits were merely in "signaling", whereas for Lundberg and Startz
all of the benefits are in the form of enhanced productivity. For Spence, the costs
of the investment varied inversely with the productive ability of the worker; for
Lundberg and Startz, the costs are invariant with respect to the productive ability
of the worker. Apparently, a proper mixture of the two sets of assumptions could
yield optimal investment. Both models face the criticism that the employer's
uncertainty about the productivity of workers may be inexpensively reduced by
observing,~the worker's on-the-job performance.
12An illustration of the adaption of the Spence model to an equilibrium with no discrimination is
available from the author. Also, see Riley (1975) for a critique of the robustness of Spence's
conclusions about suboptimality.
Although I do not find the empirical counterparts to the models of statistical

discrimination and signaling to be convincing in terms of the necessary empirical
magnitudes of such variables as costs of information or in terms of behavioral
patterns, what is considered convincing and realistic is a matter of judgment.
Some readers may not view trial work periods as realistic. A rigid system of
"tracking" newly hired workers, for example, could scuttle the strategy of trial
work periods. Others may believe that government and union wage floors are
pervasive and, in combination with the statistical model shown in part (b) of
Figure 13.1, block the employment of minorities on a large scale. There is need
for institutional knowledge and for judgments.
3.4. Institutional theories of discrimination
In his survey of the economics of racial discrimination, Marshall (1974) advo-

cated an institutional theory of discrimination which, although presented as an
alternative to neoclassical theories, could be viewed as a plea for more comple-
mentary attention to such factors as historical contexts, "pre-labor-market"
discrimination against minorities, group bargaining, the psychological motives of
the economic agents, monopoly elements, and a variety of societal factors
Marshall classified as environmental. Some points are well taken, and few
neoclassical economists would argue in principle against them. Others reflect
certain misunderstandings. Neoclassical theory is not, for example, synonymous
with perfect competition; monopolies, including labor unions and governments,
are not ignored in neoclassical economics. Pre-labor-market discrimination is
allowed for in Model (I), represented by group differences in X between the
minority and majority workers. Many of the societal factors Marshall mentions
(1974, p. 868), such as health, education, and business conditions, have all
received considerable attention in the neoclassical literature.
The institutional approach sometimes cuts across several disciplines. One
example is the reference to psychology and theories of adaptive behavior. Piore
(1970) argues that the initial placement of disadvantaged workers into low-wage,
low-status jobs creates attitudes and habits that perpetuate their low status.
Arrow (1973) suggested a related model in which the psychological theory of
cognitive dissonance rationalizes market exchanges that result in a suboptimal
equilibrium. In essence, expectations are formed by employers about the inferior-
ity of the group discriminated against, and the latter internalize these expecta-
tions and take actions-in particular, underinvest in human capital-which
confirm those expectations. An objection to both versions of this pattern of
self-injurious behavior is that the predicted behavior is obviously counter to the
best interests of two key actors - the group discriminated against, whose members
want to overturn the expectations, and employers, who ought to prefer to
730 G. G. Cain
augment the supply of labor by encouraging more investment in human capital

and positive attitudes towards investment and work.
Myrdal's (1944) classic work on discrimination included a similar model of
feedback effects, in which economic, attitudinal, and health variables interact
dynamically. An interventionist shock to any one of the variables sets in motion
an upward or downward spiral of all the variables. Lundahl and Wadensj/5 (1984,
pp. 16-18, 53) discuss Myrdal's model, including its similarity to Piore's, and
point to its vulnerability to the criticism of instability. A corollary objection,
similar to the one made in the preceding paragraph, is that the model's predicted
consequences from a favorable shock are so obviously beneficial to the group
discriminated against and to employers that it is difficult to see why the upward
spiral would not quickly be initiated by group intervention. These criticisms
apply, however, to the particular mathematical formulation of the model and not
to the reasonable view that economic outcomes are determined by multiple
causes, some of which are noneconomic, and that feedback relationships are part
of reality.
If institutionalism refers to historical case studies, to details of the process by
which equilibrium states (or tendencies) are reached, and to the interactions
among organized and individual agents, then the approach-while not a theory,
in my judgment- is always useful and sometimes indispensable. In the statistical
studies that are discussed in some detail in the next section, there are often
contexts in which various strata or segments of the full population are studied.
The question arises: How were the individuals selected into these strata, and does
the selection process either reflect discrimination or affect the interpretation we
give to the analysis? For example, in studies of the effect of unions on the wage
differential of black and white workers, institutional knowledge about the selec-
tion process into unions and how the process differs by race is necessary to
interpret correctly the statistical estimations. Neoclassical economists are aware
of the need for this information and, in one form or another, pay attention to the
selection process [see Ashenfelter (1972), Becker (1959), Kessel (1958), Lewis
(1959)], but they seldom have an absolute advantage in the institutional aspects
of the problem. The legal and historical studies such as those of Gould (1977),
Hill (1977), Marshall (1965), and Northrup (1944) are also useful.
Earlier, the institutional study by Ross (1948) was quoted to illustrate the harm
done to black railway workers by the interactions of government wage-fixing
(blocking the forces of wage competition), employer monopoly, Southern com-
munity prejudice, and an all-white labor union. Ross provides more institutional
detail about this episode of discrimination that is worth retelling to remind those
of us who work with austere models and simplified statistical specifications just
~,ow complex is the reality we are trying to capture. Ross recounts the advances
made by black railway workers into the higher-paying jobs of firemen on the
Southern railway carriers during World War I. Later, during two depression
periods, 1921 and again in 1931, the white workers' grasp for these jobs reached
an intensity that took on an all too typically American climax. I quote Ross:
The depression of 1921 put many Negro and white workers on the street. There
was violent competition to keep or grab places on any pay rolls. In 1921 there
began a series of shootings from ambush at Negro firemen on Southern trains.
Five were killed and eight wounded .... [In] the depression year of 1931...a
Negro fireman, Clive Sims, was wounded on duty by a shot fired out of the
dark beyond the track, the first of fourteen such attacks which stretched out
over the next twelve months. This was not a racial outbreak in hot blood. It
was a cold calculated effort to create vacancies for white firemen in the surest
way possible, death, and, by stretching out the period of uncertainty and
horror, to frighten away the others (pp. 119-120).
There are, as noted earlier, many theories or models that result in discrimina-
tory outcomes. The challenge is to determine their quantitative importance. The
instrument of terror, such as described above, no longer plays an important role
in labor market discrimination. But even when this weapon is replaced with the
milder instruments of racial and sexual harassments, we may find that the
organized, sometimes conspiratorial, activities of majority workers and employers
operate with a different set of rules than those we specify in our conventional
economic models.
4. Empirical research on labor market discrimination
Aside from descriptive statistics, empirical research on economic discrimination

may be divided into (1) tests of hypotheses suggested by the theories, such as the
proposition that wage discrimination is less in competitive industries, and (2)
estimation of the amount and determinants of discrimination; for example,
estimating the effect of race on wages (the coefficient A) in a cross-section
version of Model (I) (with productivity characteristics held constant), or estimat-
ing the change in the relative wages of minority workers over time.
4.1. Testing hypotheses suggested by theoretical models of discrimination
The hypothesis about labor market discrimination that has received the most
attention is that discrimination is greater in monopolistic industries. An early
empirical test is presented in Becker (1957, rev. ed. 1971, pp. 47-50). Many
studies have followed. 13 I do not review this hypothesis and these studies mainly
13The followingcitations refer to studies, like Becker's, in which the proportion of minority-group
employees in the firm, industry, or market is related to some measure of concentration (or degree of
competitiveness): Comanor (1973), Oster (1975), and Luksetich (1979).
732 G. G. Cain
because I am uncomfortable with two links that connect the theory and the
empirical evidence. First, I question whether product monopoly implies monop-
sony power in the labor market; the relevant labor market is usually a local area,
and we have no assurance that monopsony power is highly correlated with the
commonly used measures of monopoly, such as concentration ratios. This criti-
cism, which applies to many of the previous studies, has been recently developed
by Ashenfelter and Hannan (1986). Second, the desired theoretical measure of
discrimination is the difference in minority/majority wages for equally produc-
tive workers, but most of the studies have used minority/majority employment
differences (or ratios). While there is certainly interest in such employment ratios
and associated measures of segregation as indicators of discrimination in the
labor market, wage discrimination is not necessarily linked to segregation.
Aside from the studies of monopolies and discrimination, hypothesis testing
has been, as Masters (1975, p. 19) noted, "surprisingly limited", and this type of
study has produced few, if any, firm conclusions. In part this is because the
theories often yield ambiguous predictions. For example, discrimination may be
predicted to exist in the short run but not in the long run, but there may be no
basis for determining the time required for the transition. Also, the theories
suggest many economic influences, and the hypothesis test usually concentrates
on one influence in isolation. The disappointing yield of most hypothesis testing
may be conveyed by an examination of four studies.
(1) In his book, Reich (1981) criticized neoclassical theories of discrimination,
provided tests of neoclassical hypotheses, and developed an alternative theory of
discrimination that emphasized the role of class conflict between workers and
capitalists. I focus solely on his test of Becker's model of employer discrimination
in a competitive economy (pp. 109-163), which also appeared previously [Reich
(1971)] and was discussed by Masters (1975, pp. 19-21). Reich claimed that
Becker's model predicted a negative relation between (i) profits, which might
more accurately be identified as the employers' return on their capital and their
entrepreneurial skills, and (ii) the degree of discrimination, which is measured by
and is inversely related to the ratio of blacks' wages to whites' wages, Wb/Ww,
for equally productive black and white workers. An examination of Reich's
analysis serves to illustrate several difficulties, listed as (a)-(c) below, in testing
hypotheses.
(a) The problem of ambiguity of theoretical predictions when, as shown in Section
3, there are many plausible outcomes, eoen within the neoclassical paradigm that
Beeker employed. Reich claims that Becker's theory predicts that "white capi-
talists lose and white labor gains from racial discrimination" (1981, p. 111). This
translates into a positive relation between profits and W b / W w. To see how this
might occur, assume that white and black workers are equally productive, that
their labor is inelastically supplied, that all employers have the same tastes for
discrimination, and that employers' preferences for white workers lead to the
ratio W b / W w being less than 1. Now assume that the tastes of employers change
to a stronger preference for whites. This leads to a higher wage for white workers
and lower money profits for employers. The decline in profits is offset by a higher
psychic income to employers from their enhanced preference for white workers,
thus maintaining the total returns on their capital and entrepreneurial skills.
We here encounter a distinction, not emphasized earlier, between whether the
employers' preferences are pro-white or anti-black. Had the hypothesized exam-
ple assumed a change in preferences by employers toward greater distaste for
black workers, then W b / W w would still decline, but in this case Wb would fall
and money profits r i s e - t h e latter offsetting a decline in the psychic income of
employers. A focus on the wage ratio leaves us with an ambiguous interpretation.
There are other sources of ambiguity. The observed variables are profits and
wages, and these are predicted to change in response to an unobserved change in
employers' tastes. However, the observed variables may change for other reasons,
with a different application or interpretation of Becker's model. Assume now that
there is variation in employers' tastes for discrimination, but that the distribution
of employers' tastes does not change from one period to another. If the ratio of
black workers to white workers increases, Becket (1971, pp. 43-45 and 97)
predicts a fall in W b / W w, because the employers with stronger tastes against
blacks can only be induced to hire the increased number of blacks by a decline in
W b. In this case, money profits rise, offset again by a fall in the psychic income of
the new employers who are hiring blacks. Thus, the predicted short-run result is a
negatioe relation between profits and W b / W w --opposite of the implication Reich
draws from Becker's theory.
(b) The problem of ambiguity because the predictions depend on the length of the
time period to which they apply and because the theory offers no guidance on the
time required for certain forces to take effect. Reich's test of the relation between
profits and W b / W w is based on a 1960 cross-section of 48 standard metropolitan
statistical areas (SMSAs). Each SMSA is designated as a separate labor market.
The hypothesis Reich is testing is one that assumes that employers' tastes vary
across markets and that their tastes cause the variation in W b / W w. In 7 of his 43
reported regressions Reich (1981, pp. 135-155) controlled statistically for the
ratio of the black population to the white population in the market, so this source
of variation in W b / W w was, in principle, neutralized in these 7 regressions. In the
other 36 regressions one could argue that more black workers lower W b / W w and
increase profits and that this negative relation is consistent with Becker's model
for reasons discussed above. The simple correlation between W b / W w and the
percentage nonwhite in the SMSA is -0.71 in Reich's sample (1981, p. 149).
In the seven regressions in which the percentage nonwhite is controlled, Reich
finds a negative relation between profits and W b / W w, but whether this is
inconsistent with Becket's model depends, as we have seen, on whether one
assumes variation in pro-white or in anti-black tastes among employers. Another
734 G. G. Cain
point is that in a cross-section any nonzero relation between profits and Wb/Ww
may be viewed as a temporary disequilibrium, if the factors of production are
mobile across SMSAs. Equally productive black (or white) workers would not
remain in a market where they were underpaid relative to the wages available in
other markets. Even though pervasive tastes against blacks by employers could
lead to W b / W w < 1, the ratio should tend toward equality across markets if there
is worker mobility. Alternatively, capital flows across markets will tend to
equalize profit rates. If the profit variation is due to variation in employers'
tastes, thereby allowing for the compensating variation in psychic income among
employers, employers with the strongest tastes against blacks (or for whites)
would tend to move to markets where blacks are relatively less numerous.
Repeating the observation of Alchian and Kessel: "Free choice of economic
activities implies a distribution of resources that would minimize the cost of
satisfying tastes for discrimination" (1962, p. 161).
Neoclassical theories do not, however, tell us how long the equilibrating
process will take, so tests involving SMSA data at a point in time could be
thought of as either testing the competitive model or as testing the time of
transition to equilibrium. Alternatively, a defender of a "sluggish" competitive
model could test for the predicted equilibrating process by using SMSA data for
two or more points in time.
(c) The problem of matching the desired theoretical variables with the available
empirical variables. The hypothesis about the relation between profits and W b / W w
for equally productive workers are actually tested by Reich by a regression
between (i) a variety of measures of income inequality, such as the percentage
share of all white incomes received by the top 1 percent of white families, $1, or
the Gini coefficient of white family incomes, G, and (ii) the ratio of black to
white family income, Yb/Yw. The Gini coefficient is a commonly used measure of
overall income inequality, which includes the earnings of white workers. Becker's
theory of employer discrimination made no prediction about the effect of
W b / W w on the inequality of white workers' earnings. Nor is it obvious that S 1 is
a good measure of profits, because the incomes received by the richest 1 percent
of families will include rents, interest payments, wage and salary earnings, and
income from inherited wealth as well as current profits from businesses employ-
ing workers. 14
The theoretical variable, Wb/Ww, may diverge from Yb/Yw, and Reich
provided no control for the relative productivities of black and white workers by
such conver~onal measures as the ratios of mean educational attainments, mean
years of experience, and so on. Reich's control variables were measures of the
14It should be noted that Reich expressed interest in the relation between discrimination and
overall white inequality, so m y discussion is restricted to Reich's use of these inequality measures to
test Becker's model.
overall occupational and industrial structure, the median family income of whites
(Yw), the percentage of the SMSA population that is black (although in only one
regression were both this percentage and Yw included), and a few others.
Generally, Reich found a statistically significant negative relation between Y b / Y w
and his profit proxies, G or $1, which he interpreted as a refutation of Becker's
model of a competitive economy and discrimination based on employers' tastes.
In the light of the difficulties associated with items (a)-(c) above, I doubt that
Becker's model was or can be well tested with such data.
(2) While Reich attempted to test for a relation between W b / W w and profits,
sometimes controlling for the ratio of black workers to white workers, N J N w ,
Landes (1968) and Flanagan (1973) drew upon Becker's theories to test for a
negative relation between W b / W w and N b / N w. The justification from Becker's
theory is as follows. Assume a distribution of employers' tastes for discrimination
that is heterogeneous within a market and identical across markets. As we have
seen, a larger N b / N w leads to a smaller W b / W w because the larger is N b / N w ,
the more are employers with stronger prejudices against blacks induced to hire
black workers. The greater discrimination of these employers is manifest in a
lower W b / W w , at least during the short run.
We have noted that mobility by black workers will tend to attenuate the
negative relation between W b / W w and N b / N w, by tending to equate the ratios
across markets. Also, there are institutional reasons for doubting the assumption
of an identical distribution of tastes by employers across markets. Historically
and in 1960, discrimination against blacks was most severe in the South, the
region with the largest N b / N w. The legacy of slavery in the South was causal to
both the discrimination and the residential location of blacks.
Scholars in other disciplines have debated how prejudice is related to N b / N w
within a region. Perhaps prejudice is greater when N b / N w is greater because
whites feel threatened by a larger ratio. On the other hand, perhaps the level of
prejudice decreases as N b / N w rises because contact and familiarity erode un-
favorable stereotypes and misunderstanding. In either case the level of tastes may
change over time as experience with threats or with familiarity evolves. Thus, the
basis for testing a version of Becker's theory that depends on identical distribu-
tions of tastes across markets appears questionable, although the empirical results
of such tests are interesting on their own.
Landes (1968) found a negative correlation between W b / W w and N b / N W
across all states, but the correlation was essentially zero within both the South
and the non-South regions. However, this finding was secondary to Landes's
main interest in the effects of antidiscrimination laws on W b / W w , so I examine
the article by Flanagan (1973), whose main interest was to test the hypothesized
negative relation between an occupation-specific W b / W w and an occupation-
specific N b / N w. He used aggregated state data from the 1960 census for men in
seven, and for women in five, broadly defined (one-digit) occupations. Other
736 G. G. Cain
variables in the regressions were the black-to-white ratios of four variables - weeks
worked, educational attainment, age composition, and median family
i n c o m e - a n d two nonratio variables-a dummy variable for the South and the
percentage of the population that was foreign-born. No systematic relation
between W b / W w and N b / N w was found. This may be evidence against Becker's
theory, or it may be evidence against Flanagan's maintained assumption that the
distributions of tastes of employers are identical across states, or it may be that a
simultaneous relation between wages (prices) and the quantities of occupational
skills prevents the identification of an effect of the quantity ratios on the wage
ratios. 15
(3) A study by Chiswick (1973) is unusual for its focus on Becker's model of
workers', rather than employers', discrimination and on wage inequality among
w h i t e s - a topic not treated by Becker. Essentially, Chiswick tests the hypothesis
that a measure of the variance of white male incomes in a state is positively
related to the percentage nonwhite in the state.
Chiswick begins with Becker's definition of worker discrimination: a wage, W,
is paid to (demanded by) a white worker who works with white workers, and
( W + d ) is paid to (demanded by) a white worker who works with black workers.
As we have seen in Section 3, segregation could prevent the long-run main-
tenance of wage discrimination against blacks, but Chiswick argues that
inequality of wages is likely to persist if some white workers have skills comple-
mentary to the skills of black workers. Chiswick offers the example of "foremen"
and "laborers", presumably where whites are both foremen and laborers and
blacks are only laborers (p. 1332). 16 Chiswick apparently rules out a segregated
equilibrium in which there are some firms that hire only unskilled workers, who
would be either all white or all black, and other firms that hire workers of both
skills, who would be all white. 17
Chiswick defines a dummy variable, X, as 1 if a white worker "works with
nonwhites and ... zero if he does not" (p. 1333), and expresses the dual wage
~5Flanagan notes the potential simultaneity problem and refers in a footnote to his consideration of
it. However, not enough information is provided to determine if the simultaneity problem is
adequately handled.
16The page numbers in parentheses in the text refer to Chiswick's article. Chiswick does not discuss
the skill distribution of blacks or the possibility that blacks acquire skills. Note that if blacks acquire
complementary skills, segregation could again eliminate racial discrimination in wages. Chiswick
mentioned two other sources of integration in the work force besides complementarities in skills-
unions and fair employment laws. The operations of these sources are not explained, except to note
that they interfere with competitive market forces (p. 1332). Moreover, unions and fair employment
laws are not mentioned again and play no role m Chiswack s empirical tests.
17Firms employing all unskilled workers will pay equal wages to black and white workers. Firms
employing any white skilled workers will hire only white unskilled workers to keep their costs at a
minimum, so an equifibrium requires that all firms hiring both skills to hire only white workers. All
unskilled workers, white or black, would receive the same wage. But this scenario merely reflects the
segregation equilibrium that Chiswick has ruled out. Thus, we need to assume, as Chiswick implicitly
does, that all firms require both skills.
structure for whites as W* = W(1 + d X), where W* is the observed wage and W
is the wage paid to the white worker who works only with whites. (A skill index,
using a subscript for the j t h skill, is omitted, and my symbols differ from
Chiswick's.) The mean, X, "is the proportion of the white labor force that works
in an "integrated' situation" (p. 1333), and Chiswick represents this by the
percentage of nonwhites in the population, p = ( l O O ) [ N b / ( N b + N w ) ] (pp.
1334-1335).
The relationship between .~ and p may be justified by assuming that unskilled
workers have tastes for discrimination, so competitive forces should lead to their
segregation by race. TM There would be no wage inequality among white unskilled
workers (the laborers) within a market (or, for that matter, between
m a r k e t s - where a market is a state in Chiswick's formulation), at least as regards
the effects of workers' tastes for discrimination. White skilled workers (foremen)
would earn more if they worked in a firm with all-black unskilled workers than if
they worked in an all-white firm, and labor costs would be equalized across firms
by paying lower wages to black unskilled workers.
In this model and with the expectation that there are more firms with
segregated unskilled workers in a state with a larger proportion of blacks, the
mean wage of skilled workers should be positively correlated with p. This
correlation identifies a direct test of Chiswick's model. A second direct test is the
segregation of unskilled workers. I refer to these as direct tests because they
involve cross-state comparisons of "first-order" effects on means and proportions
rather than comparisons of "second-order" effects on within-state measures of
inequality.
As noted in Section 3, the Becker-type models in which the skilled white
workers have tastes for discrimination are similar to models with discriminating
employers. Both agents are complementary to black labor. A long-run competi-
tive equilibrium with discriminatory wage differentials paid to the skilled workers,
like the long-run equilibrium with differential profits among employers, depends
in both cases on homogeneity in the tastes of the discriminators. Or, expressed
more cautiously, the tendency for discrimination to wither away depends on the
existence of some nondiscriminating skilled workers (or employers) and on
whether they can expand production to take advantage of their cost advantage.
Chiswick's empirical work focused on the variance of the logarithm of income
for men aged 25 to 64, using midpoints of nine income classes, with an
approximation for the mean of the highest, open-ended income class. This
18If the unskilled workers did not discriminate against each other, competitiveforces would tend to
make the proportion of black and white unskilled workers equal. Otherwise, either firms with more
black workers would be at a competitive disadvantage-having to pay more to their skilled white
workers-or blacks in firms with a larger proportion of blacks would be earning less than their
counterparts in firms with a smaller proportion of blacks. See Arrow (1973, pp. 10-13) for a
discussion of this case.
738 G. G. Cain
variable was regressed on p along with controls for several market sources of
inequality in the form of variables involving the age, schooling, and weeks-worked
distributions in the state and a variable defined as the rate of return on schooling
in the state, which Chiswick had calculated in his previous research. Chiswick
assumed that tastes for discrimination and p were uncorrelated. To make this
assumption plausible, he separated the 17 Southern states from the non-Southern
states. Chiswick found that white inequality was positively related to p, within
both the South and non-South regions.
The causal inference seems shaky, but interpreting empirical tests that are
indirect is always a matter of judgment. Here, p is an indirect measure of either
the intensity of skilled workers' tastes against unskilled black workers or of the
proportion of white skilled workers who receive higher wages by working with
blacks, and the variance of income is an indirect measure of the skilled workers'
wage inequality (since there should not be inequality among the white unskilled
workers' wages). The regression for the South had only 8 degrees of freedom. In
the 31 non-Southern states, there were only 13 where blacks were more than 3
percent of the population in 1960 [U.S. Bureau of the Census (1980, p. 36)]. The
highest percentages, 8.0-10.0, were in the industrialized states: Illinois, Michigan,
Missouri, New Jersey, New York, and Ohio. The lowest percentages, 0.1-0.9,
were in relatively nonindustrialized states: Iowa, Idaho, Maine, Minnesota,
Montana, New Hampshire, North Dakota, South Dakota, Utah, Vermont, and
Wyoming. Thus, outside the South blacks were generally such a small proportion
that it is difficult to see how they could have had much effect on white income
inequality. Where they were a modest proportion, it was in states that tended to
be more industrialized and densely populated.
Aside from how one might interpret Chiswick's regressions showing a positive
relation between p and the variance of white incomes, I find them unconvincing
as a test of Becker's model in the absence of direct information on how workers'
tastes for discrimination affect (a) the segregation of workers and (b) the wages of
white skilled workers who do and do not work with black unskilled workers. On
this latter issue, Blau (1977, pp. 58-73) reports that in her study of labor market
discrimination among several white-collar occupations, men who worked in
integrated firms (with both men and women) received lower wages than men who
worked in all-male firms, and she interpreted this as evidence against the
hypothesis that workers' discriminatory tastes were causal to wage differentials) 9
A problem with these tests, however, is the necessary assumption that the
integratiort measure (say, the proportion of blacks or women in a firm) is
uncorrelated with the average skill level of the white or male workers whose wage
is the dependent variable.
19In private correspondence,Chiswick cites an unpublished study by James Ragan that also uses
data for individual firms and finds higher wages for whites who work in integrated firms. This finding
was interpreted as supporting the Becker-typemodel of worker discrimination.
(4) The final example of hypothesis testing is Ashenfelter's (1972) analysis of

the effect of unions on the white-black and male-female wage differences. The
model of discrimination under competitive conditions, which appears fragile and
difficult to test in the previous examples, is replaced here by the more robust
theory of union gains and a somewhat tentative theory of racial and gender
selection into unions. Specifically, it seems reasonable to assume that union-based
noncompetitive wage differences across racial and gender groups can be sus-
tained. The effect of union status on a worker's wage is estimated by a Model (I)
regression function, which is applied separately to the four race-gender groups.
Each of the four union effects (coefficients) is multiplied by the percentage
unionized of each race-gender group to show the difference in wages across the
groups that is attributable to unionism.
A numerical example is helpful. Assume that the effect of unions is to increase
the wages of unionized black men by 10 percent relative to nonunion black men,
while the corresponding effect for white men is 5 percent. Assume also that the
proportion unionized is 20 percent for both racial groups. A first approximation
to the union effect on W b / W w is obtained by assuming that the wages of
nonunion workers, W ", are equal to what the wages would be in the absence of
unions. Let the wage ratio for nonunion workers, black-to-white, be 70/100. This
can be compared to an estimated ratio for all workers, holding constant available
productivity characteristics. This estimated ratio is calculated as a ratio of
weighted averages of the wages of union and nonunion workers, using the
percentage union, U = 0.2, and the percentage nonunion ( = 0.8) as weights. Thus
the estimated ratio for all workers, holding constant their productivity, is
a'b ( 1 - + UI,V~ 0.8(70)+0.2(77) 71.4

-- = 0.707.
I,Vw = ( 1 - U) ^" + VWw' 0.8(100)+0.2(105) lOa.O
We see in this example that unions increase the overall wage ratio by 0.007, or by
1 percent, relative to what it would be in the absence of unions.
Clearly, the overall impact of unions on the majority-minority differential by
these calculations depends on the percentage of each group that is unionized and
the wage effect of unionism for each group. If the union effects for both racial
groups are 10 percent and the proportion unionized is 30 percent for blacks and
20 percent for whites, the same impact of unions on the black-white wage ratio
would be obtained.
Calculations like these were carried out by Ashenfelter, who first obtained
estimates for U, 1~ n, and 1~ u for the four demographic groups. He added a
refinement by computing estimates of 1~ n and I~ u for major (one-digit) occupa-
tional groups and then summing these with weights for union and nonunion
status that involve the proportion of the wage bill (total wages) received by
each union-and-occupational group. Thus, instead of weighting the W's by U,
740 G. G. Cain
Ashenfelter used U*, the proportion that union wages are of the total wage bill
earned by whites (or blacks). The U* values are larger than the U values,
especially for blacks. The low percentage unionized of both blacks and whites in
the higher-paying white-collar occuPations carries a low weight for blacks relative
to whites because relatively few blacks are in these occupations. Thus, although
23 percent of black workers in Ashenfelter's principal sample are union members,
about 34 percent of the black wage bill is from black unionized workers. The
c o m p a r a b l e figures for whites are 23 percent and 31 percent. 2°
Using U*, Ashenfelter concluded that " t h e ratio of black to white male wages
m a y have been some 3.4 percent higher in 1967 than it would have been in the
absence of all unionism" (p. 463). The ratio of female to male wages was
estimated to be 1.9 percent lower than it would have been in the absence of
unions (p. 453, n. 33). The 3.4 percent gain to blacks reflects a differential effect
of unions in favor of blacks by about 11 percentage p o i n t s - a 21 percent effect
for black men and a 10 percent effect for white men (p. 450). An illustrative
weighted average for men is
I)¢"b 0.66(70) + 0.34(84.7) 75

0.727,
6"w = 0.69(100)+0.31(110.0) 103.1
which is 3.9 percent larger than the estimated wage ratio in the absence of
unions, 0.7. Using the unrefined union weights, U = 23 percent for both blacks
and whites, the weighted ratio would be 0.717, which is a little over 2 percent
larger than 0.7.
These findings are evidence against the hypothesis that unionism in the United
States, as measured during the 1960s, is responsible for the discriminatory wage
differential in favor of whites or, with weaker evidence, in favor of men. The
data on union membership by demographic groups are not controversial, and
Ashenfelter provides alternative estimates of the effects of unions on wages,
based on his own analysis of other data sources and on the existing literature.
Overall, these checks were supportive. Ashenfelter reminds the reader that his
evidence does not say that unions are nondiscriminatory; rather that they are
shown to be no more discriminatory, or even less regarding blacks, than the
economy as a whole.
The validity of Ashenfelter's estimates of union effects depends on two key
assumptions. The first is that the estimates of union effects on union workers are
either unbiased or that they are biased equally for majority and minority groups
(hereafter, white and black men). The general issue concerning a bias is that
20These percentages are calculated using Tables 6 and 7 in Ashenfelter (1972), although I adjusted
the weights in Table 7 for whites to make them sum to 1. Apparently there is an error in Table 7 for
white workers, because the proportions sum to 1.072 instead of 1.00. I reduced each occupation's
proportion in the table by 0.92 (=1.00/1.072). In my calculation of U* I assume that the percentage
unionized for private household workers and farm workers is zero for both color groups.
union status may be correlated with unmeasured productivity variables, leading

to a misestimate of the true effect of unions. As stated, the bias could be positive
or negative, depending on whether union workers were, holding constant the
control variables in the model, less productive (owing to, say, nepotism or
perhaps because unambitious workers are more attracted to unionism) or more
productive (due, say, to the commitment of union workers to their trades or
because employers will select high-quality workers when faced with union-
imposed above-competitive wages and because high-quality workers will seek
these positions). To sharpen my argument and shorten the discussion, let me
assume that the net bias in the union effect is positive, and the coefficients of
union status on wages, 10 percent for whites and 21 percent for blacks, are both
too high. Clearly, the issue for Ashenfelter's measure of the union impact on
W b / W w is whether the bias is larger-really, much larger-for blacks than
whites.
I now argue that the bias is larger for black men and against Ashenfelter's
assumption that wages of nonunion workers represent what the wages would be
in the absence of unions. Assume that the jobs in the union sector are medium-
paying jobs in the crafts and operative occupations for both white and black
men, whereas jobs in the nonunion sector are predominantly high-paying profes-
sional, technical, managerial, and sales jobs for white men, but predominantly
low-paying laborer and unskilled service jobs for black men. Skill levels of the
jobs are assumed to be correlated with the skill abilities of workers, both innate
and acquired. Assume that these contrasting alternatives to whites and blacks
regarding nonunion jobs are entirely attributable to "pre-labor-market dis-
crimination", which is to say that they are reflections of differential family
socioeconomic backgrounds, quality and quantity of schooling, and wealth
constraints on the long-term investments required for the high-paying jobs.
Assume further that the distributions of innate ability (intelligence, "ambition",
and so on) are identical for whites and blacks. Given these assumptions, it is
reasonable to believe that if unions were nondiscriminatory then black males
would be more represented in the union jobs because black workers of above-
average ability are constrained from entering the highest-paying jobs but not (by
assumption) from crafts and operative jobs, and they will therefore gravitate
toward the better-paying crafts and operative jobs. The presence of unions
restricts numbers of both blacks and whites, but the restrictions are more binding
on blacks, since the excluded higher-ability white workers will have the highest
occupations open to them. Thus, not only should the unionized percentages be
higher for blacks in a nondiscriminatory labor market, but the effect of union
status on wages will tend to be more upward-biased for blacks. The latter bias
stems from the presumption that the omitted innate ability is, on average, higher
for black union members than for white union members. Ashenfelter's model
assumes that the occupational distribution of blacks and whites is given, and the
foregoing argument suggests that it is affected by unionism.
742 G. (7. Cain
The arguments above are admittedly speculative. An upward bias in the

estimated union effect, however, has the theoretical justification that employers
should respond to union-imposed high wages by upgrading their hiring and
retention standards. Generally, unionized employers do have control over hiring,
and they have some control over retention, at least through some probationary
period before union-imposed seniority protection commences. On the other hand,
arguments in favor of Ashenfelter's conclusion are the following. (a) The above
scenario denied any role to labor market discrimination for the dispro-
portionately low representation of black men in the white-collar occupations, and
this denial is hard to accept. (b) For a reduction in unionism to lead to relatively
more occupational upgrading among blacks than whites among the blue-collar
occupations, one must assume that the general sources of labor market dis-
crimination would not maintain the existing distribution. (c) Ashenfelter's esti-
mated union effects on wages would have to be drastically changed to reverse his
conclusion of a beneficial wage-effect for blacks among unionized workers. Recall
that his union effect for blacks (21 percent) is twice that for whites (10 percent).
Finally, Ashenfelter's rejection of the hypothesis that W b / W w would be higher
without unions is strengthened by his institutional and historical discussion about
union race policies. For historical reasons, unionism is more widespread among
the blue-collar occupations, and blacks are more likely to be competing for jobs
requiring less skill. Thus, Ashenfelter argues that because unions of lesser-skilled
workers will have more blacks in their jurisdiction, competitive forces will tend to
force the unions to include blacks. Among blue-collar occupations, therefore,
whites will be over-represented in the unionized skilled jobs and underrepre-
sented in the unionized lesser-skilled jobs, relative to blacks. An overall tendency
for equality in the incidence of union membership among whites and blacks is,
therefore, plausible.
The remaining parameter of interest, the union-effect differential, is, however,
puzzling on theoretical grounds. Ashenfelter's arguments (p. 447) about the
potential power of the skilled trades to be more restrictive in controlling
the supply of labor should be supported by larger union effects (rents) for
the more skilled groups. This result would indeed be consistent with the
a priori Marshallian arguments, found in almost every labor economics textbook,
in which skilled workers face a more inelastic demand curve and therefore have
more "bargaining power". As noted above, this result is not found by Ashen-
felter, nor by other recent analysts of union effects. [See the studies Ashenfelter
cites in his~Table 3~ p. 446; Johnson (1975), and others.] 21 Thus, the large union
2aTo be more precise, my claim is that Ashenfelter and others have found an overall negative
correlation between union effectsand skill levels, even though the constructiontrades, airline pilots,
and some other crafts have shown large union effects. Among white construction workers, inciden-
tally, the union effects of laborers exceed those of skilled workers [see Ashenfelter (1972, Table 5,
p. 450)].
effects for blacks, relative to whites, is consistent with the larger union effects for
lesser-skilled blue-collar workers, but the latter union effect remains a puzzle.
4. 2. Estimating labor market discrimination
4.2.1. Methodological points
Model (I) (in Section 2) is the basic model used to estimate labor market
discrimination. Its widespread use along with several conventions that are
customarily adopted permits a succinct summary of results, shown in Tables 13.6
and 13.7 in the next section. Unfortunately, the results are so varied that they
reveal as much about our ignorance as about our knowledge of the degree of
labor market discrimination against blacks and women. This variability is not
really surprising in light of the theoretical vagueness that underlies most of the
empirical specifications.
An inherent ambiguity, mentioned earlier (Section 2), stems from the absence
of agreement on what productivity traits-the X's in Model (I)-are ap-
propriately held constant. The criterion I suggested is that the variables held
constant in Model (I) should not be determined by the process of discrimination
under analysis. Applying the criterion requires a clear statement of the purposes
of the estimations, but this is seldom provided. Perhaps the marketwide regres-
sion studies of wage discrimination are merely intended to provide a general
social indicator of inequity in the economy, although this is ambiguous unless we
know what counterfactual regime is being compared to the current regime. This
counterfactual is usually only implicitly revealed by the set of X-variables that
have been held constant, and there is seldom discussion of whether the X's are
affected by labor market discrimination. 22 Predictions using the regression results
are not often explored, and specific remedies or policies to deal with discrimina-
tion are seldom linked to the regression results. To clarify some of these issues,
consider the following two applications of the criterion suggested above.
Case 1. Assume the analysis pertains to a given employer or firm, and that we
ask whether white workers are paid more than black workers after holding
constant the available productivity variables. Assume further that a panel of
experts provides us with the worker characteristics that determine productivity in
the firm. The productivity variables might include previous vocational training,
tests of manual dexterity, age, years of schooling, and so on. To meet the above
criterion, each variable should be exogenous to the employer; that is, the
22 Blinder (1973) is exceptional in his clear distinctions between the X s that are assumed exogenous
and those that are endogenous according to current theories of labor market behavior, specifically the
theory of human capital.
744 G. G. Cain
characteristic should not be affected by the employer's behavior. If it were, it

might reflect discrimination. Thus, a variable defined as "task-specific ability"
that is measured by "supervisor's rating" would be suspect, and perhaps not
admissible. Clearly, the presumptive identification of supervisors with manage-
ment raises suspicions about the unbiasedness of supervisors' ratings. On the
other hand, if we knew that supervisors were nondiscriminatory, their ratings
would provide direct evidence of the workers' productivity, which is usually
difficult to obtain and certainly preferable to the indirect evidence from such
variables as age and education.
Case 2. Assume the analysis pertains to the entire labor market. We ask
whether white workers are paid more than black workers after holding constant
an admissible set of productivity variables that are not affected by the process of
discrimination under analysis. Because the entire labor market is under analysis,
however, variables like "previous training" almost surely reflect previous dis-
crimination in the labor market, so they are not admissible.
There is no simple rule in marketwide studies for determining when a variable
may be appropriately held constant. Among the variables mentioned in Case 1,
age is clearly exogenous. Years of schooling are appropriately held constant if we
believe that the decision to attain schooling does not reflect discrimination in the
labor market. Perhaps the lower education among minorities reflects societal
discrimination-not labor market discrimination but "pre-labor-market dis-
crimination". Alternatively, perhaps blacks and women perceive that higher
levels of schooling yield smaller earnings for them than for white men. If this
were true, then these groups may have curtailed their schooling, in which case
educational attainment would reflect labor market discrimination. Determining
the productivity variables that are admissible is the first step in estimating Model
(I). Accurate measures of the agreed-upon variables are also needed.
Let us turn now from the conceptual issues in estimation to the mechanics of
the statistical methods. The regression specifications for Model (I) that produce
the estimates of labor market discrimination in the recent research literature
usually involve the following assumptions and procedures.
(i) Separate regression functions are estimated for majority (hereafter white, w)
and minority (hereafter black, b) groups. The equations omit the subscripts
denoting the observation. The explanatory X-variables measuring productivity
traits, and the B-coefficient of each X are collectively represented by ~BX.
I,J/w=;Y~BwXw and Ww=l~w = EBwXw, (5)
I;Vb= ~_,BbX b and Wb = l~b= EBbXb • (6)
The caret indicates predicted value, the mean of which, W, is identically equal to
the overall mean, W. The intercept term in the equation is included in ~ B X and
m a y be associated with an element in the X-vector for which X b = X w = 1 for
each observation. 23
(ii) Equations (5) and (6) are used to express eq. (7), which is a particular
decomposition of the difference in mean wages obtained by adding the term
Y'.Bw.~b to both (5) and (6) and then subtracting (6) from (5):
(7)
The firm term on the right-hand side of (7) evaluates the difference in mean
values of the X ' s at white "prices" (Bw'S), and the second term evaluates the
racial price differences at the mean value of the black X's. It turns out that, on
average, -~w > Xb and B w > Bb; more precisely, that ~ B w . ~ w > ~]Bb.~ b-
(iii) The second term on the right-hand side of (7) is a conventional measure of
labor market discrimination, with B w > B b representing a higher price received
by a white worker than by a black worker for the (assumed) same productivity
characteristic. The first term on the right-hand side of (7) involves the racial
differences in X ' s and does not have a clear interpretation. It may represent a
source of a nondiscriminatory difference in wages, because only one price is used
to evaluate different amounts of exogenous productivity characteristics. Or, it
could measure the difference in wages attributable to pre-labor-market dis-
crimination, which may explain why Xw > -~b. I n a n y case, the conventional
standard of nondiscrimination is achieved when W b / W w =1, holding the X ' s
constant.
(iv) An important reservation about the decomposition in (7) is that it is not
unique_. Each difference, B w - Bb, in the second term is evaluated as a product
with X b, but the evaluation might have used Xw or some average of Xb and Xw.
Similarly, the use of Bw as a weight fo__rthe first term, Xw - Xb, is also arbitrary.
A different decomposition of Ww - W b is obtained by adding the term ~ B b . ~ w
to both (5) and (6) and subtracting (6) from (5):
W w - Wb = E B b ( X w - -~b)+ Z X w ( B w - Bb)- (7')
The different standardizations shown by (7) and (7') reflect the familiar
index-number problem encountered whenever heterogeneous collections of goods
23Blinder (1973, pp. 438-439) separated the intercept terms from other B-coefficientsand specified
them as two components of discrimination. This procedure is not necessary or even helpful, because
the value of the intercept term will depend on the arbitrary scaring of the X-variables. Consider, for
example, the arbitrariness of defining a variable like region of residence into a set of dummy
variables, where the intercept will represent the excluded region. Which region is to be excluded is
arbitrary. See Jones (1983) for further discussion of the point.
746 G. G. Cain
( X ' s ) are s u m m e d with two sets of prices (B's). 24 In the simpl_._est case in which
all prices are the same for both racial groups, the difference Ww - 14"b is simply
equal to the first term on the right-hand side of (7), and the conceptual
experiment of assigning equal X ' s to both racial groups leaves only the difference
in intercept terms, which measures a vertical difference in 1~" between two
"parallel" linear functions. Such a difference in intercept terms is what was
previously measured by the coefficients A or C on the d u m m y variables for group
status in Models (I) and (II) in Section 2.
(v) I will rely on the following expressions for summarizing the various
estimates of eqs. (5) and (6) reported in the literature:
(i) Ur = W b / W w = the unadjusted ratio,
(ii) A r = Y ~ B b X w / ~ B w X w = ~'.BbXw/Ww,
which is an "adjusted ratio", obtained from either (7) or (7'). To arrive at (ii),
simply set (or assume) all Xb equal to Xw to eliminate the first term of the
decompositions in (7) or (7') and to reduce the__fight-hand side to its discrimina-
tion component. Then divide through by Ww and simplify to express the
following equation of ratios: Ar = WUWw, where Wff is the black mean wage
conditional on the black X ' s being set equal to the white _~'s. A~ = 1 implies no
discrimination. The amount by which the controls for X have closed the gap
between unity and U~ is the sometimes-used statistic
(iii) G = [(1- At)/(1- U,)]IO0,
called the percentage of the gap between U, and 1 that is attributable to the
difference in the X's. Thus, 1 - G is the percentage of the gap that is attributable
to labor market discrimination.
For simplicity, I will restrict my discussion of empirical results to the adjusted
and unadjusted ratios, A r and Ur. Even here, it is somewhat arbitrary to use A r
as defined b y (ii), because we could have defined
(ii') A" = E B b X b / E B w X b = Wb/Y'~BwRb,
which, like A,, holds the .~'s constant and attributes the remaining differences i_n
black and Xvhite average wages to the B's, but here the B ' s are multiplied by X b
24Blinder (1973, p. 438, n. 3) suggested that the decomposition expressed by eq. (7) is preferred to
that of (7') because he claimed that the decomposition using black prices (Bb'S) as weights for the
difference in X's leaves an interaction term as a residual, in contrast with the decomposition using
white prices (Bw'S) as weights. This is incorrect. There is no difference in the two decomposition
methods in this respect.
levels. Usually A r is presented, because the conceptual experiment of raising Xb
to the levels of -gw is more appealing and more policy-relevant than lowering
-~w to -~b levels as is done with A'r. Nevertheless, it is easy to construct examples
in which the regression results give qualitatively different measures of wage
discrimination on the basis of A r and A'. One may equal unity and the other
m a y exceed or fall short of unity. The quantity A r - Ur may be positive, showing
that the X ' s "explain" some of the gap (assuming Ur < 1), whereas A'r - Ur may
be negative, showing that the gap is made even wider after controlling for the X ' s
and using this standardization.
At the risk of belaboring the obvious, I will use the constructs of A r and A'r to
illustrate two points. One is the potential ambiguity of these ratios as measures of
discrimination, and the second is that some institutional knowledge of the
process by which discrimination occurs is necessary if the statistical measures are
to tell us anything.
A s s u m e eqs. (8) and (9) refer to males (subscript m) and females (subscript f)
and that the only explanatory variable, X, in the wage function is the number of
young children present in the household of the worker. The wage functions,
evaluated at means, are
Wm = B0m -I- Blm.~m = 10+ 1( Sm = 2) = 12, (8)

<= B0r + B , f X e = 9 - 1 ( X f = I ) =8. (9)
In eq. (8), I assume that all men are working, that they are in families with an
average of two young children, and that the presence of young children has a
positive effect on the wages earned by the men. (Perhaps additional dependents
lead them to work harder.) Eq. (9) is assumed to express the wage equation for
employed women. I assume that half the women are employed (and thus have a
market wage), that women with fewer young children are more likely to be in the
labor force, and that the presence of children is negatively related to the wages of
women, which is discussed below.
Clearly, A r = ~ - ~ B f ~ ' m / W m = 7 / 1 2 = 0.58, and A r - U~ = 0 . 5 8 - 0 . 6 7 = - 0 . 0 9 .
Thus, the unadjusted ratio is higher than the adjusted ratio. The women's wage
would be less than their current wage if they had the same values of X as men, so
we may conclude that discrimination is even more severe than shown by the
unadjusted wages.
On the other hand, A~ = ~/Y'~BmX f = 8/11 = 0.73, and A'r - Ur = 0.73 - 0 . 6 7
= 0.06. This shows that discrimination against women would be less if men had
the same values of X as women. Since the regression method shows that
discrimination is both worse and better than the unadjusted wage comparisons,
what should we conclude? Or, consider the following specification:
Wr~ = 6 + l(.Xm = 2) = 8 (10)

748 G. G. Cain
and
~ =9-1(Xf=l) =8. (11)
A r ( = 0.875) shows discrimination against women, whereas A'r ( = 1.14) shows

discrimination against men.
The statistical procedures cannot tell us the correct answer. Let us consider two
hypothetical processes by which employers pay wages to examine how we wright
determine whether there is discrimination.
Case 1. Assume men and women are equally productive, but that employers
discriminate against women with children, as shown in (9) or (11). Let us assume
that children have no real effect on productivity of either men or women but that
employers have a uniform preference for paying men with children more and
w o m e n with children less. Both sets of equations, (8)-(9) and (10)-(11), show this
discriminatory behavior, and constructing ratios and making decompositions do
not add to our knowledge.
Equations (10)-(11) do not show a gender difference in average wages, but one
could argue that the discrimination against women with children expressed in
(11) is a cause of the low labor force participation of women. Assume that all
w o m e n enter the labor market and that demand conditions remain the same.
Then (11) becomes Wf = 9 - 1 ( X ~ = 2) = 7, and Ur becomes 7 / 8 = 0.875. The
discrimination, expressed in (11) and revealed by a "potential" b~ < 1 for the full
population, is a cause of the fact that only half the women are working. This is a
reminder that the values of the X ' s may reflect discrimination and are not always
exogenous to the process under study.
Case 2. Assume that men and women have the same productive capacity, but
women with children are less committed to market work than women with no
children, and men with no children are less committed to their jobs than men
with children. If these commitments reflect people's preferences about how they
want to live, and if the presence of children is an accurate signal of this
commitment, then there is no presumption of discrimination. The two sets of
equations show the relation between productivity (commitment) and children and
also show how the relation differs for men and women. Again, the ratios A~ and
A r do not add anything useful to our knowledge.
The message is that the original data and statistical functions are mere
description. Knowledge of the process by which wages are set and, perhaps, by
which workers are selected into the market are necessary.
4. 2.2. A survey of selected estimated wage functions
In contrast to the research that tests hypotheses, the studies presenting empirical
estimates of labor market discrimination are numerous. Only about 20 of these
studies are selected for mention in this section, and they will be summarized in
two tables. Methodological issues are emphasized to aid in understanding the
strengths and weaknesses of the research and its theoretical and policy content.
In addition, these empirical studies contain useful descriptive statistics.
Labor market discrimination, or wage discrimination, has been defined in this
paper by using Model (I) to isolate the net effect of minority-group status on
wages, holding constant the productivity characteristics of the workers. Two
crucial questions invariably arise: (a) Do the variables measuring productivity
reflect discrimination? (b) Do the variables measure productivity comprehen-
sively, aside from factors that can be assumed to be random with respect to
group status? If the answer to (a) is yes, we may presume that the estimate
understates discrimination. If the answer to (b) is no, the estimate of discrimina-
tion may be biased up or down. To avoid prejudging these answers, I will use the
term "wage gap" rather than "wage discrimination". The wage gap will be
measured by the unadjusted ratio, Ur, and by the adjusted ratio, A r (or, rarely,
At).
The estimated wage gaps are based on cross-section or time-series studies,
which were the classifications used in the descriptive statistics presented in
Section 2. Cross-section studies generally are interpreted as representing normal
or equilibrium conditions. Trends over time may be inferred from successive
cross-sections, allowing for changing compositional effects (like the age distribu-
tion) or specific period effects (such as the business cycle). Trends may be directly
measured in a time series by introducing time as an independent variable and
determining if its effect differs for the different groups, but time-series studies are
hampered by the fewness of observations. Almost all the empirical work in the
published literature uses cross-sectional data, and this section will be devoted to
these, leaving the few time-series studies for the final section on policy analysis.
Another classification of the studies is by the type of minority and majority
groups being compared, and I will continue to focus on black-white and
w o m e n - m e n comparisons. Some estimates of wage gaps according to national
origins and religions will be briefly mentioned.
I also concentrate on studies that intend to estimate the overall wage gap,
rather than on studies that focus on the differential effects on wages of particular
variables, like education, years of work experience, union status, or participation
in some government program. Finally, I concentrate on studies that measure the
wage gap for the entire labor force, or at least for large groups in the labor force.
Only limited attention is given to the many studies of the wage gap within
individual firms or within occupations.
4.2.2.1. Comparisons of the earnings gap between women and men. A summary
of studies of the wage gap between women and men is shown in Table 13.6. The
wage is, as discussed below, the most appropriate simple measure for examining
750 G. G. Cain
Table 13.6
S u m m a r y of studies of ratios of women's earnings to men's earnings, unadjusted and adjusted for
various characteristics of workers and jobs.
Statistical Women's earnings as a

Author and Data source Measure method and ratio of men's
year of and population of explanatory Observed e Adjusted f
publication a studied b earnings~ variables a = Ur = Ar
Gwartney and Census, age 25 + T, R: 1,2

Stroup (1973) with positive ~, 1959 (grouped 0.33 0.39
incomes ~, 1969 data) 0.32 0.40
Featherman and OCG, married y, 1961 R, S: 1,(2), 0.38 0.48*
Hauser (1976) workers 1972 7, 23
Blinder (1973) PSID, white w, 1969 R, S: 2, (3), 0.54 0.54
working house- 9,12-14,
hold heads and 32, 34
spouses, age 25 +
Sawhill (1973) CPS, wage and y, 1966 R: 1,3,10 0.46 0.56
salary workers, 13
age 14+
Gwartney and Census, age 25 +, .Vr, 1959 T, R: 1,2, 0.56 0.58
Stroup (1973) full-time, year- (10)
round workers
Suter and NLS, CPS wage- y, R,S: 1,(2), 0.39 0.62*
Miller (1973) and-salary 1966 6,10, 23
workers, age
30-44
Roos (1981) GSS, white y, R,S.: 1,2, 0.46 0.63*
workers, age 1974-1977 10,22,23,
25-64 26,29-31
Fuchs (1971) Census (1/1000 w, 1959 R: 1,2,3, 0.60 0.66
sample), nonfarm 8,12, 25, 33
workers
Treiman and NLS: married y, R,S: 1,(2),
Terrell (1975) workers, age 1966 6, 7,10,17, 22
30-44: whitc 0.42 0.67
nonwhite 0.54 0.68
Cohen (1971) Survey of vr, 1969 R, S. : 1,2,10, 0.55 0.69
working 11,16, 24, 27,
conditions, 28
full-time
nonprofessional
wage-and-salary
workers, age
22-64
Blinder (1973) PSID: white, w, 1969 R,S.: 1,2,(3), 0.54 0.70*
working house- 5,9,11 14,21,
hold heads and 27, 32, 34
spouses, age 25 +
Oaxaca (1973~ SEO, urban w, 1967 R,S.: 1(2),
workers, age 16 + 3,7-10,12,13
white 0.65 0.72
nonwhite 0.67 0.69
Sanborn (1964) Census, wage and ~, T: 1,2,3,10, 0.58 0.76*
salary workers 1949 18, 20
Oaxaca (1973) SEO, urban w, 1967 R,S.: 1,(2),
workers, age 16 + 3,7-10,12,13,
white 21,25- 27 0.65 0.78"
nonwhite 0.67 0.80 *
Ch. )3: Labor Market Discrimination 751
Table 13.6 continued
Statistical Women's earnings as a

Author and Data source Measure method and ratio of men's
year of and population of explanatory Observed e Adjusted r
publication~ studied b earningsc variablesd = U r = Ar
Kohen and NLS, full-time w r, R, S: 1,3,4,

Roderick wage and salary 1968-1969 7-9,13-15
(1975) workers, age
18-25
white 0.76 0.78
black 0.82 0.81
Mincer and NLS, SEO, white w, 1967 R,S: 1,(2),(3)
Polachek (1974) wage and salary 6,11
workers, age
30-44
married 0.66 O.80
single 0.86 0.87
Corcoran and PSID, working w, 1975 R,S: 1,(2), 0.74 0.85
Duncan (1979) household heads, age (3), 5, 6, 9,
18-64, white 11-13,16,17
Sanborn (1964) Census, wage ~, T: 1-3,6,10, 0.58 0.88*
and salary 1949 16,18-20, 24
workers
Malkiel and Professional annual salary R,S: 1,6,8,16, 0.66* 0.77*
Malkiel (1973) full-time 1966, publications,
employees in 1969-1971 Ph.D., field
one company
above + job
level g 0.66 * 0.86 *
Astin and Survey of annual salary R: rank, 0.78* 0.87*
Bayer (1972) college faculty degree, field,
research output,
type of college
Johnson and Survey of 9-month R: years since 0.85 * 0.93*
Stafford Ph.D.s in salary degree, field,
(1974) college sector,
faculties experienced
( = 10 years)
Sources: D. J. Treiman and H. I. Hartmann, eds., Women, Work, and Wages: Equal Pay for Jobs of
Equal Value (Washington, D.C.: National Academy Press, 1981), pp. 20-37, along with additional
material and changes in the tables presented in Treiman a n d Hartmann.
aFull citations are given in the references. The same study m a y appear more than once in the table.
bSources for the individual studies use the following shorthand terms:
Census = the decennial Census of the United States.
CPS = the Current Population Survey of the U.S. Bureau of the Census
O C G = the CPS survey of Occupational Change in a Generation, 1962 and 1972
GSS = General Social Survey, 1975-1978
SEO = Survey of Economic Opportunity, 1966-1967
N L S = National Longitudinal Survey, 1967 and subsequent years (Ohio State University)
PSID = Panel Study of Income Dynamics, 1968 and subsequent years (University of Michigan)
Terms such as "aged 25 + " refer to "workers aged 25 or older", and so on.
Cy = annual earnings (or income)
) = a group's m e a n (or median) annual earnings (or income)
w = wage per hour
subscript f = refers to full-time workers
752 G. G. Cain
dT = tabular standardization
R = regression analysis
S = regression analysis using separate equations for men and women
Explanatory variables are listed by number at the end of these notes. The use of parentheses
around a number indicates that this variable is implicitly held constant, either because of the sample
selection or because another variable effectively controls for the variable in question. For example, if
only whites are sampled, then race is being held constant.
e Ur = the ratio of mean female earnings (or income or wage) to mean male earnings.
fA r = adjusted mean-earnings ratio, which is the ratio of the conditional mean earnings of women
to the mean earnings of men. The conditional mean earnings of women is the earnings predicted for
women if they had the same values of the explanatory variables as do men.
SThe term "above + " means that the explanatory variables used are the same as those in the
preceding list, plus whatever new variables are listed.
* The explanatory variables include a control for the occupation of the worker. Controlling for
occupation is especially likely to raise the ratio of women's to men's earnings, for reasons discussed in
the text.
Explanatory variables:
1. Education
2. Age
3. Race
4. Mental ability (intelligence)
5. Formal training
6. Actual labor market experience
7. Proxy for labor market experience
8. Marital status
9. Health
10. Hours of work (annual, weekly, full-time/part-time)
11. Tenure (length of service with current employer)
12. Size of city of residence
13. Region of residence
14. SES background (parental education, occupation, income, number of siblings, migration history,
ethnicity, etc.)
15. Quality of schooling
16. Absenteeism record
17. Dual burden (number of children, limits on hours of location, plans to stop work for reasons
other than training, etc.)
18. U r b a n / r u r a l
19. Turnover
20. Occupation (census three-digit)
21. Occupation (census one-digit)
22. Occupational prestige
23. Occupational SEI (Duncan scale of a socioeconomic index)
24. Other occupational classification or scale
25. Class of worker (self-employed, government, or private wage and salary)
26. Industry
27. Union membership
28. Type of employer (government/private, sex segregated/integrated, size of work force)
29. Supervi~0ry status
30. Percentage female in work group
31. Median income of male incumbents
32. Local labor market conditions
33. Length of trip of work
34. Veteran status
35. Migration status
labor market discrimination between men and women, but most of the studies
use earnings or incomes. For brevity, I will refer to the earnings gap. The style
and much of the content of the table are taken from the compilation of studies in
Treiman and Hartmann (1981). The columns denote the authors of the studies,
the data sources, the measure of the dependent variable, the statistical method
(usually regression analysis) and the control variables used, the unadjusted ratio,
Ur, and the adjusted ratio, At, which is the ratio of the average predicted
earnings of women to the average earnings of men. For Ar the earnings of women
are usually predicted by a regression equation. When separate regressions for
men and women are used, the earnings of women are predicted by assigning
men's mean values for the predictor variables along with the regression coeffi-
cients from the women's equation.
The studies are listed in rough order of the size of A~, which, although
cautiously referred to as the wage gap in the sample, given the particular control
variables used, is sometimes referred to as a measure of labor market discrimina-
tion. An asterisk next to the ratio indicates that some measure of occupational
status was held constant. Among all the commonly used control variables,
occupation is perhaps the one most "suspect" or "tainted" as being a reflection
of labor market discrimination. It is an inappropriate control variable by the
criterion I have proposed, although as noted above almost any variable that is
subject to some choice by the individual worker and to some influence by the
market may be suspect according to this criterion. Of course, occupation may
have been advisedly included because the investigator wanted to measure the
wage gap conditional on being in a given occupation. Thus, the asterisk is not an
indicator of a defective study but rather of a study that does not measure
marketwide discrimination as I have chosen to define it.
Reading down the rows of Table 13.6, we see that Ur and A~ have a similar
ranking, and both range from 0.3 or 0.4 to around 0.8 or 0.9. The high figures
usually refer to restricted samples. Much of the variation in these estimates may
be explained in common-sense terms according to the following characteristics of
the studies.
(1) The use of earnings (or in rare cases income) for persons in the labor force
tends to give a lower ratio than the use of wage rates. The latter holds constant
the unit of time for which earnings are measured. I prefer the hourly wage for
gender comparisons because the amount of time spent at work will partly reflect
voluntary choice. In contrast, an earnings comparison may be more useful for
comparing white and black men, because working less than full-time by black
men often reflects discrimination rather than voluntary choices.
(2) Samples that represent the full population generally show a smaller ratio.
There is no necessary reason for this pattern; rather, the restricted samples
happen to be for groups where the gap is narrower, such as for young age groups,
for single women, or for certain occupations or industries. The wage gap is
754 G. G. Cain
narrow for young people, and it widens with age. This could mean that there is
little discrimination by gender for young people and that the widened gap among
older workers merely reflects the voluntary choices of women and men to
specialize later on in housework or market work, respectively. Or it could mean
that discrimination takes the form of providing women fewer chances for
promotion or for on-the-job training. If it is the latter, then some part of the
lesser market work (and more housework) by women may reflect market dis-
crimination.
(3) Black women tend to have a higher Ur and A r than white women. Again,
the research challenge is to determine the extent to which this is attributable to
differential discrimination on the demand side compared with the difference in
supply-side characteristics between white and black women. Keep in mind that
certain supply-side factors, such as the century-long commitment of black women
to market work, their lower probability of marriage, and their higher probability
of marital dissolution are all plausible reflections of the labor market discrimina-
tion faced by black men. 25 Thus, the low earnings of black men are in part a
cause of the high work rates of black wives and, perhaps of the lower proportion
of black adults who are married.
(4) Adding more control oariables usually raises A r, and there is a noteworthy
pattern to this. Various "pre-labor-market" controls, such as education, age,
family background, and residential location, are all very similar for men and
women, unlike the case of white men compared to black men. Standardizing the
women's predicted earnings with men's mean values for these control variables
can hardly close the gap by much. Thus, Blinder shows no difference between Ur
and A r (both equal 0.54) when he holds constant age, health, residence, and
family background. Nor would education have made much difference, because
the means for men and women in his sample are about the same. These
pre-labor-market variables often differ substantially between blacks and whites,
however, so controlling for them does raise A r relative to U~, as we shall see in
Table 13.7.
The variables that reflect work experience, such as the number of years spent in
the labor force and the worker's tenure with a firm have, on the other hand,
substantially different mean values for men and women. When Blinder added
tenure, union status, and a one-digit occupational classification, tile A r rose to
0.70 (from a U~ = 0.54). A strong point of Blinder's study is his distinction
between variables that are reasonably viewed as being exogenous to the process
of labor market discrimination from the variables that are likely to reflect labor
market discrimination.
2SThe percentage of black women who had ever been married tends to be slightly lower than this
percentage a m o n g white women, holding age constant. The percentage of black women who were
divorced or separated at the time of the surveys is two to three times as large as this percentage
a m o n g white women, holding age constant. These statistics refer to the years 1970 and 1982. See U.S.
Bureau of the Census (1983c, pp. 33, 44-45).
(5) Using a wage rate as the dependent variable and controlling for years of
experience usually raises A~, as is illustrated by Corcoran and Duncan and by
Mincer and Polachek (M&P). Both studies measure a relatively narrow definition
of discrimination in which the years of experience of the workers are carefully
controlled. In these studies the A~'s rise to 0.80 and 0.85 for married persons and
to 0.87 for single persons. These ratios are almost as close to unity as those for
which occupation is controlled (Sanborn, Malkiel and Malkiel, Astin and Bayer,
and Johnson and Stafford). However, if tenure is a reflection of
discrimination-"last hired, first fired"-and if years of experience are less for
women because of the lower wage offered to them, then tenure and experience are
in the same category as occupation; that is, invalid control variables because they
reflect discrimination.
M&P deal with the endogeneity of experience in one of their models by
substituting the predicted value of experience in place of actual experience in the
wage equation for women. [This technique is also used by Zabalza and Arrufat
(1983), who estimate the wage difference between women and men in Great
Britain.] The validity of this technique, however, depends crucially on two
assumptions. (1) There is at least one variable in the equation predicting
experience that is excluded from the equation predicting wages. (2) The excluded
variable, which serves to identify the "experience effect" in the wage equation,
does not reflect labor market discrimination. The key predictor variable that is
excluded from the wage equation and included in th,: experience equation is the
woman's number of children. Are the above two assumptions satisfied? The
question is debatable, but I believe the presence and number of children shifts
the issue of discrimination onto another dimension of what are simultaneously
determined behavioral outcomes: time in market work, time in housework,
numbers of children, occupational choices and career plans, and so on.
Polachek (1979) does treat experience and occupation similarly, because he
views both as simultaneously chosen by women in view of their greater commit-
ment to housework and their lesser commitment to market work compared to
men. Polachek argues that women will choose occupations that facilitate their
intended short and intermittent stays in the labor market; specifically, occupa-
tions with relatively flat age-earnings profiles that do not offer the large rewards
to experience as do occupations that provide relatively steeply rising age-
earnings profiles and which tend to be male-dominated. The theory of Polachek
and M&P of the time allocations to work over the life cycle offers an explanation
for why market experience is less for women and also for why women's wage
returns to experience are less-that is, why their age-earnings profile is flatter.
Figure 13.2 clarifies these ideas. Consider the three age-earnings paths, DF,
EG, and DH, drawn linearly to simplify the exposition. Equally productive
workers, who start at age A 0 and who retire no later than A~, may be assumed to
be indifferent between occupations with the age-earnings profile DH or EG
because, let us assume, the present values of the two streams of earnings are the
756 G. G. Cain
Wages
(or earnings)
U
xAi
D
Ao A× An Age = A
A o = Age at which worker enters labor force
A n = Retirement age
EG, DH = Two age-earnings profiles chosen by equally

productive workers
DF = Age-earnings profile chosen by a worker who has

chosen to invest less in earnings capacity (or
human capital)
Figure 13.2. Three hypothetical age-earnings profiles. A0 = age at which worker enters labor force;
A,, = retirement age; EG, DH = two age-earnings profiles chosen by equally productive workers;
DF = age-earnings profile chosen by a worker who has chosen to invest less in earnings capacity (or
human capital).
same. The cross-hatched area, D E X , is drawn to be smaller than X G H to allow

for the discounting of future receipts.
If the equally productive workers are m e n and women, the M & P theory is that
w o m e n wil,t~choose E G instead of D H because they are less likely to want to stay
in their j o b continually (from ages A 0 to A , ) , and they will prefer the higher
wages in E G up to age A x instead of choosing the D H path. This part of the
M & P theory would not, by itself, be supported empirically: we do not observe
w o m e n earning higher wages than men between ages A 0 to A x. A second part of
the theory offers a supply and d e m a n d explanation for why w o m e n ' s profiles will
be like D F instead of EG. On the demand side the claim is that employers
generally prefer workers who are willing to work continually and who are willing
to accept the D H profiles (and the on-the-job investment that D H implies). This
decreases the demand for the EG workers and lowers their earnings profile. On
the supply side, the choice by women to work intermittently implies that they will
not invest as much in marketable human capital, because they will have fewer
years to receive returns on their investment. This lowers the earnings path still
further, say to DF.
Thus, the supply-side argument for lower earnings paths for women is that
they invest less. However, when the investment is measured by years of schooling,
we do not observe important differences between men and women. The argument
sometimes shifts to an emphasis on less observable variables, such as the intensity
of investment in schooling or in on-the-job training. See the comments by June
O'Neill, below.
In principle, these sorts of assumptions about gender behavior can rationalize
an A r of 0.85 and explain away the remaining gap of 0.15 by references to
"measurement error" and other sources of omitted productivity variables that, if
corrected, would show men to be more productive. In summary, the argument is
that the lesser investment, lesser experience, greater time in housework, and lower
occupational attainments of women (a) are voluntary choices made by women
and (b) are choices that causally precede the gender difference in the demand
structure. But because our economic theories and statistical techniques cannot
tell us what is or is not voluntary, I doubt that the computations of Ar's will
measure labor market discrimination in any fundamental sense.
Perhaps the most emphatic argument in support of the "voluntary view" is
presented in a series of papers by O'Neill (1983a, 1983b, 1984). Her discussion of
the lower occupational status of women is replete with references to choices
(from 1983b):
The investment component of schooling ... varies by subject matter. Women
have traditionally chosen majors such as education, arts, and humanities,
which have lower pecuniary returns than subjects such as business or science
(p. 19).
Since many women continue to be responsible for a disproportionate share of
household maintenance and child care even after they enter employment, they
are likely to evaluate certain job characteristics differently than men...[so]
predominantly female occupations were much more likely to offer part-time
work and less likely to require very long work weeks (p. 19).
... there is a strong element of personal choice in the occupations held by
women.., the dominant variables explaining whether a woman is in a typically
female occupation were those describing plans and expectations held five years
758 G. G. Cain
earlier. W o m e n who said they planned to be a homemaker at age 35, who had
children, were married, and who did not attend college, were more likely to be
in stereotypically female occupations (p. 22).
Women who five years earlier said they planned to work at age 35 and desired
to be working in a male-typed occupation, who attended college, and majored
in a scientific subject, were in fact highly likely to be in male-dominated
occupations. These findings would appear to contradict the presumption that
barriers to entry are the primary reason why women are poorly represented in
many occupations (p. 22).
Notice that none of these arguments, which emphasize choice rather than
discrimination, is persuasive if discrimination is believed to be causal to the
choice of majors, to the time devoted to housework, to the employers' offers of
part- versus full-time jobs, and to the "plans and expectations" women had at
ages 19-29 regarding their career at age 35. In other words, if labor market
discrimination does restrict the quality of jobs and wages available to women, it
is reasonable to believe that this affects their plans and expectations regarding
school majors, fertility, and their time allocation to home and market sectors. The
last quote by O'Neill seems to say that, for example, (a) if a woman who received
a degree in electrical engineering is working as an electrical engineer, then (b) the
poor representation of women in electrical engineering is not to be considered
evidence for discrimination against women ("barriers to entry") in that occupa-
tion. It seems to me that (b) does not follow from (a).
(6) Restricting the sample to unmarried women and men usually shows higher
values of Us and Ar, as is illustrated by the ratios 0.86-0.87 reported by M&P.
Studies that compare single women to either single men or all men might be
viewed as providing a purer measure of gender discrimination by avoiding the
troublesome issue of the dual career that is associated with married women.
Unfortunately, the issue remains. The never-married single women tend to be
young, under 25 or so, and the Us for young people is relatively high (see the
Kohen and Roderick entry in Table 13.6). However, a ratio that is less than 1
may reflect the employers' expectations that the women are likely to marry and to
be less committed to their jobs than men of the same age. If the sample were
restricted to never-married women in their 40s or older, for whom a strong
commitment to market work may be presumed, the sample would be relatively
small and probably selective of women who were either unusually dedicated to a
career or ~ausually adverse to marriage. Arguments could be made that these
women would be likely to earn more, or less, than men who are comparable in
the conventional characteristics used in earnings functions. Indeed, single men
tend to earn less than married men, holding constant conventional variables. A
full understanding of these selective traits determining marital status involves
more than just economics.
(7) Restricting the comparison to a narrowly selected group of jobs tends to
produce higher Ur and Ar ratios, as is illustrated by Malkiel and Malkiel (1973).
This study is the only one in Table 13.6 for a single company, and I will have
more to say about this type of sampling restriction later: Also, it is not only a
sample of a relatively narrowly defined occupational g r o u p - all college-educated
professionals who work for a particular research f i r m - b u t it offers a control over
"job level", which further narrowly defines the tasks, duties, and responsibilities
of the employees. By controlling for job level, the adjusted ratio rises from 0.77 to
0.86. This is evidence for the claim made earlier that with a sufficiently narrowly
defined job almost all ratios would be unity. Indeed, if not, companies would risk
violating the law. Finally, the study offers a rather striking example of the
importance of the method of standa___rdization. The conventional A~ is equal to
W f / W m, and an alternative is A~ = W J W m. The alternative adjusted ratios in the
Malkiel and Malkiel study are 0.85 (instead of 0.77) and 0.99 (instead of 0.86).
The 0.99 ratio was the one emphasized by Malkiel and Malkiel and used by
O'Neill (1984, pp. 79-82).
4.2.2.2. Black-white earnings gap for men. The empirical measurements of

wage discrimination between blacks and whites in the United States involve the
same procedures as those just described for men and women, but there are
differences in results and interpretations. I focus on men in the racial comparison
in order to separate this from the gender factor. A difference in the normative
interpretation is that the wage ratios for women-to-men, Ur and At, that are less
than one may be rationalized by claiming that women choose to specialize in
home production. No such alternative employment is credible for black men.
Furthermore, even if women suffered lower market wages because of discrimina-
tion, they might recover all or part of these losses by marrying men, who are the
favored group. Consideration of total household income for comparisons be-
tween men and women will be discussed in the last section.
The unadjusted wage ratios for black and white men shown in Table 13.7 are
similar to those for women-to-men in Table 13.6, if we restrict the comparisons to
large populations and exclude the comparisons for young people and for selected
occupations. The average Ur is 0.58 in the black-to-white male ratios and 0.55 for
women-to-men. The increase in the adjusted ratios, Ar - Ur, is, however, gener-
ally larger for the race ratios in Table 13.7 than those for the gender ratios in
Table 13.6, if we exclude the comparisons with asterisks where occupation is held
constant. The average increase in A~ relative to Ur is about 0.16 (0.74-0.58) for
men, black-to-white, and about 0.09 (0.65-0.56) for women-to-men. Despite the
crudity of these comparisons-they are not confined to comparisons with similar
variables held constant, for example-we may conclude that holding constant the
usual available productivity variables has a larger effect in reducing the wage gap
between black and white men. The reason, mentioned earlier, is that the usual
760 G. G. Cain
Table 13.7
Summary of studies of ratios of black men's earnings to white men's earnings, unadjusted and adjusted
for various characteristics of workers and jobs.
Statistical
method and Blacks' earnings as a
Author and year Data source and Measure of explanatory ratio of whites'
of publication a population studied b earnings c variables d Observed e Adjusted f
Masters (1975) Census (1/1000 sample), y, 1959 R, S: 1,2,12, 0.50 0.59

blacks, non-Hispanic 13, 25
whites, age 17-64,
worked in 1959, above + 10 g 0.50 0.80
civilian, nonstudent
Same as above but y, 1959 R,S: 1,2,12 0.64 0.61
for non-South 13, 25
above + 10 g 0.64 0.74
Blinder (1973) PSID, white and black w, 1969 R,S: 2,9,12-14, 0.49 0.64
working household heads 32, 34
above +1,5,11, 0.49 0.80*
21,27 g
Masters (1975) SEO, blacks, non-Hispanic y, 1966 R,S: 1,2,12, 0.55 0.66
whites, age 17-64, worked 13,25
in 1966, civilian, above + l0 g 0.55 0.72
nonstudent
Same as above but R,S: 1,2,12, 0.68 0.75
for non-South 13, 25
Duncan (1968) OOC, native, nonfarm, y, 1961 R,S: 1,4,8 0.46 0.68
25-64
above +4g 0.46 0.75
Corcoran and PSID, household heads, w, 1975 R,S: 1,(2), 0.77 0.89
Duncan (1979) worked 500 hours or more, 5,6,9,11-13,
age 18-64 16,17
Flanagan (1974) NLS, men 15-25 or 46-60 w, 1967 R,S: 1,(2),5,
6,8,9,11,13,
26, 35
Age 46-60 0.58 0.84
Age 15-25 0.72 0.94
Sources: Full citations of the original studies are in the bibliography. In some cases the adjusted
ratios that are shown in this table do not appear in the original studies or appear in a different form.
Note: All footnotes are identical to the footnotes in Table 13.6.
c o n t r o l v a r i a b l e s g e n e r a l l y r e p r e s e n t e i t h e r e x o g e n o u s characteristics, like age, o r

p r e - l a b o r - m a r k e t characteristics, like y e a r s o f s c h o o l i n g , a n d t h e s e t y p e s of
v a r i a b l e s t e n d to b e m o r e s i m i l a r f o r m e n a n d w o m e n t h a n t h e y a r e for w h i t e s
a n d b l a c k s . ~,
B l i n d e r ' s (1973) c o m p a r i s o n s a r e a g a i n i n s t r u c t i v e . H i s c o n t r o l f o r e x o g e n o u s
v a r i a b l e s m a d e n o difference in t h e w o m e n / m e n ratios; b o t h U r a n d A r e q u a l
0.54, b u t t h e s a m e c o n t r o l v a r i a b l e s raised t h e b l a c k / w h i t e ratios f r o m 0.49 to
0.64.
T h e 1979 s t u d y b y C o r c o r a n a n d D u n c a n ( C & D ) is also i n t e r e s t i n g , a n d s o m e
c o m m e n t s a b o u t it serve to raise several g e n e r a l p o i n t s . T h e u n a d j u s t e d ratio,
Ur = 0.77, is the highest in Table 13.7. C&D use a wage rate rather than earnings
as the dependent variable, and they exclude workers who worked less than 500
hours during the survey year. Both restrictions raise the black/white ratio,
because black men are likely to suffer more unemployment, including unemploy-
ment for 10 months or longer. Recall that restricting w o m e n / m e n comparisons
to full-time workers (or controlling for hours worked) was primarily justified on
grounds that the frequency of part-time jobs among women was often voluntary.
This is seldom true among prime-age black men.
C&D also restrict their sample to men who are are household heads, and there
is likely to be some selectivity bias here that raises the wage ratio above what it
would be for the full population of black and white male workers. A smaller
proportion of black men aged 18-64 are household heads compared to white men
of these ages; the wages of those who are not household heads in both races are
lower than the wages of heads of households; the wage ratio of blacks-to-whites
among men who are not household heads is slightly less than 0.77; and household
headship may reflect labor market discrimination. 26 Therefore, the black-to-white
wage ratio for all men would be less than 0.77.
The adjusted ratio in the C&D study is 0.89, also high relative to other Ar'S.
The control variables include years of experience which, given a control for years
of schooling, is representing age and therefore almost purely exogenous. Let us
assume, for the sake of argument, that years of schooling, city size, region of
residence, and health, are exogenous. The remaining control variables, formal
training and tenure with one's current employer are, however, likely to reflect
labor market discrimination.
The ratios by Blinder and by C&D are virtually the lower and upper bounds in
Table 13.7. The other studies suggest several additional methodological issues,
but I will be brief. Masters' studies (1975) show the importance of the
South/non-South differential. More recent data in the C&D study show that this
differential is still important, although smaller. Masters clearly brings out the
effect of controlling for time worked, because each of the A r'S for the "above + 10"
comparisons (see the fourth and sixth columns) allows only the additional control
for weeks worked, and these Ar's are much larger. Finally, although I do not
show them, there are some striking differences between the A r and A'r ratios in
Masters' study.
26Among men aged 18-64, which is the popuiation frame for Corcoran and Duncan, 79 percent of
white men and 58 percent of black men were household heads at the time of the Census Bureau's
survey in 1981 [U.S. Bureau of the Census (1983a, P-20, No. 372, Table 2)]. The incomes of
year-round full-time workers who are male heads of households is about 25 percent higher than for
similar workers who are not heads of households. The ratio of black-to-white income for men who
were not household heads in 1981 is around 0.70 for all workers and 0.76 for full-time workers [U.S.
Bureau of the Census (1983b, P-60, No. 137, Tables 44 and 55)]. Income figures are used instead of
earnings because earnings are not reported for persons classified by their relationship to the
household head.
762 G. G. Cain
The study by the sociologist Otis Dudley Duncan (1968) may have been the
first to use separate regressions and to construct the "decompositions" of wage
(or income) differences. Economists usually cite later studies by economists for
these procedures.
Flanagan's studies show the frequently observed result that black/white wage
ratios are relatively high for young workers, which was also true for the gender
ratios. However, the smaller ratio among older men surely does not reflect a
voluntary choice by black men to work less in market employment, as might be
claimed for women. On the other hand, so-called "vintage effects" may be
revealed in the different wage ratios for younger and older black and white
groups, whereby the current period's larger ratio for young workers may reflect a
true long-run improvement in the relative earnings capacities of black m e n -
perhaps reflecting, in turn, recent improvements in the quality and quantity of
education. Welch (1973) and Smith and Welch (1977, 1978) have stressed this
source of a vintage effect. Others have emphasized the civil rights movement in
the last 20 years, a reduction in discrimination in society, and the increase in
antidiscriminatory legislation, all of which may be having a larger positive effect
on young blacks than on older blacks [see Freeman (1981)].
Clearly, current wages of young blacks and whites cannot conclusively reveal a
lifetime comparison. Cohort analyses of previous generations show that only part
of the improvement in wage ratios among previous generations of young people
is sustained. See Freeman (1973a), Chiswick (1974, pp. 116-118), Smith and
Welch (1978), and Hoffman (1979).
4.2.2.3. The earnings gap for other ethnic groups. The nationality group in the
United States that has received the most attention in discrimination studies in
recent years is Hispanics, which consists predominantly of persons with Mexican,
Puerto Rican, and Cuban ancestry, in that order. There is not the space to review
the empirical estimates of Model (I) for these groups and to display a correspond-
ing table. Nevertheless, some new and interesting methodological issues may be
mentioned briefly in connection with the general finding that relatively small
unadjusted wage ratios (Ur), Hispanic-to-white, coexist with relatively large
adjusted ratios (At). ("White" refers to non-Hispanic white.) For example,
Reimers (1983) finds that Ur for Mexican Americans (hereafter Mexicans)
relative to whites is about 0.70, whereas the adjusted ratio after fitting separate
Model (I)-type regressions is about 0.94. See also Abowd and Killingsworth
(1982) and Grenier (1984) for similar results. Adjusted ratios of 0.9 or higher
imply a minor role for labor market discrimination.
There are four main sources for the increases from U~ to A~ in these studies: (1)
Age differences: Hispanics tend to be younger than whites, so part of the wage
difference is explained by this exogenous variable. (2) Education differences:
Hispanics, particularly Mexicans and Puerto Ricans, have substantially lower
average educational levels. (3) Immigration status, as measured by years in the

United States. (4) Language, as measured, for example, by a categorical variable
defined as the ability to speak and understand English. All these controls seem
reasonable, but let us critically examine the last three and raise again the point
that definitive empirical measures of wage discrimination require knowledge
about the underlying processes and institutions.
Education. Assume that education contributes to one's earnings capacity in
skilled jobs but not in unskilled jobs. If this were true, then the fact that, say,
Mexicans have a low average years-of-schooling-completed ( E D ) should not
place them at an earnings disadvantage relative to whites in the unskilled jobs,
even though whites in these jobs may have a larger mean ED, perhaps because of
school attendance laws in the U.S. Nevertheless, the usual regression procedure,
with E D entered linearly in the earnings regression, will tend to assign a lower
predicted wage to Mexicans relative to whites in these unskilled jobs. Actual
discrimination in these jobs may be masked. The following simple hypothetical
example brings out this point. Assume that only ED, among the observable
variables, has a systematic effect on wages, and that the distribution of wages
(W) and E D is as shown in Panel A of Table 13.8. There are two examples each
of distributions for whites and Hispanics, which permit four possible calculations
of Ur and Ar. To obtain A r with the data of example 1, we obtain the average
predicted wages for Hispanics and whites:
1~w = -0.46+0.357ED, so W = -0.46+0.357(E-D =11.93) = 3.80 (12)
and
l~h= --0.60+0.375ED, SO W = -0.60+0.375(ED = 8) = 2.40. (13)
The corresponding regression equations using example 2 data are
1~"w = 0.025+0.322ED, so W = 0.025+0.322(ff?D =11.73) = 3.80 (14)
and
I~h = -0.300+0.313ED, so W = -0.300+0.313(/TD = 8) = 2.20. (15)
The A r and A'r formulas, using (12)-(15), give us the A r entries in Panel B of
Table 13.8.
The data have been constructed to reveal wage discrimination against HisPanics
that is assumed among unskilled workers, who are all classified with low levels of
E D - an average of 4 for Hispanics and an average of either 4 or 5 for whites. In
example 1 of Hispanic data there is no discrimination among the higher ED
764 G. G. Cain
Table 13.8
Hypothetical distribution of years of schooling ( E D ) and wages ( W ) among whites and Hispanics
and resulting comparisons of unadjusted and adjusted wage ratios
Panel A. Distribution
Whites Hispanics
Years of
schooling Example 1 Example 2 Example 1 Example 2
(ED) Number. Wage Number Wage Number Wage Number Wage
4 1 $2 3 $2 3 $1 3 $I
5 1 2
6 1 2
12 7 3 7 3 1 3 1 3
16 5 6 5 6 1 6 1 5
Number 15 15 5
Means E D = 11.93 E D = 11.73 ED = 8 ED = 8
= $3.80 W = $3.80 W = $2.40 W = $2.20
Panel B: Comparisons of unadjusted (Ur) and adjusted ( Fh, At, A'r) wage ratios*
Hispanic distribution
White distribution Example 1 Example 2
/Jr Fh Ar A'r Ur Fh Ar A~
Example I 0.63 1.02 1.00 0.58 1.02 0,94
Example 2 0.63 0.70 0.99 0.93 0.58 0.67 0.88 0.85
Source: Hypothetical numbers.

Definitions and explanations of ratios in Panel B:
u~=wh/Ww.
mr= ~Vh/Ww= E b h X w / W w .
F h for example 2 of whites with example 1 of Hispanics:

Fh = ( 3 / 5 ) ( 1 / 2 ) + (1/5)(1) + (1/5)(1) = 0.70.
F h for example 2 of whites with example 2 of Hispanics:
Fh = ( 3 / 5 ) ( 1 / 2 ) + (1/5)(1) + (1/5)(5/6) = 0.67.
categories, and in example 2 there is discrimination for the college (ED = 1 6 )

c a t e g o r y - a wage ratio of 5 / 6 = 0.83.
The unadjusted wage ratios with the data reveal typical values: 0.63 and 0.58.
An exact measure of wage discrimination, given the assumptions used in con-
structing the data, is available with example 2 of the white data and both sets of
the Hisp~hic data. This measure, expressed as a ratio, will be labeled F h to
indicate that the ratio is based on a "free functional form" (using categorical
variables for each level of ED) and that it uses Hispanic numbers of workers as
weights. It is calculated as follows: With example 2 of the white data, we see that
3 of the 5 Hispanics have the same ED as whites and earn only half as much.
With example 1 of the Hispanic data, the other two Hispanic workers with higher
ED earn the same as whites with the same ED values. F h is the Hispanic/white
Ch. 13: labor Market Discrimination 765
wage ratio, adjusted for ED values in each of the three ED categories.

F h = ( 3 / 5 ) ( 1 / 2 ) + ( 1 / 5 ) ( 3 / 3 ) + ( 1 / 5 ) ( 6 / 6 ) = 0.70.
N o functional form has been imposed on the E D / W relation, and the wage
comparisons at each ED level are weighted by the relative frequencies in the
Hispanic distribution. 27
The main conclusions from Table 13.8 are that the A r and A'r ratios show
either minimal or no discrimination, despite its "constructed" presence, and that
the sources of the discrepancy are the incorrect linear relation between ED and
W and the disparity in numbers of observations in the categories- few Hispanics
in the higher ED categories and relatively few whites in the low ED category. The
example is oversimplified, of course, but the problem it reveals with the typical
econometric estimation for Hispanics, particularly Mexicans, is, I believe, worth
considering. The general issue is that a trait, in this case ED, may vary in its
validity as an indicator of productivity over certain ranges of the variable and in
certain work situations.
Immigration. Years since arrival in the United States may be an indicator of
productivity, because the variable may represent English-language skills, labor
market information (including investments in job mobility), and various aspects
of cultural assimilation that may enhance workers' productivity in their contacts
with supervisors, co-workers, customers, and so on. If the lack of cultural
assimilation has nothing to do with one's productivity and everything to do with
discrimination, the variable loses its appeal as a control variable on this account,
although it retains its appeal on the other two accounts.
Language skills. As a control variable, language skills may be partially con-
taminated by simultaneity if the skills depend on the type of job available to the
worker, on the relationship to one's co-workers, and on other outcomes of the
labor market. In one study the variable was highly correlated with "place of
birth" and " n u m b e r of years spent in the U.S." and the latter variables were
omitted from the regression. The author [Grenier (1984, p. 42)] commented:
" O n e consequence of these omissions, however, is that estimated coefficients of
the language variables may include some effects of these other variables."
In addition to the study and measurement of discrimination regarding
Hispanics, there have been a few econometric and historical-institutional studies
27Using white relative frequencies, we have Fw = (3/15)(1/2) + (7/15)(3/3) + (5/15)(6/6) = 0.80,

but even this moderate measure of discrimination is not appropriate to describe the situation of
Hispanics in this hypothetical example, where 60 percent (3 of 5) are experiencing a large measure of
discrimination. A more appropriate use of the F-ratio for whites is to define
F~' = ( 3 / 1 5 ) ( 2 / 1 ) + ( 7 / 1 5 ) ( 3 / 3 ) + ( 5 / 1 5 ) ( 6 / 6 ) = 1.20,
which shows that the large (two-to-one) wage advantage of the unskilled whites raises their overall
wages relative to Hispanics by 20 percent, holding education constant. Finally, if we assume, with
example 1 data for whites, that the ED levels of 4, 5, and 6 provide no productivity differences among
these workers, then F h and the other F-ratios would be defined exactly as they are for the example 2
data for whites.
766 G.G. Cain
of several European and Asian nationality groups and various religious groups.
One important result of these studies is that for many ethnic groups, the ratio of
their earnings to those for a more broadly defined white group (sometimes as
narrow as those with an English ancestry) is larger than one. This was found, for
example, for Catholic Irish-Americans [Greeley (1976, p. 52), and (1981, pp.
110-120)], Japanese-Americans [Petersen (1978), Sowell (1981)], and Jews
[Chiswick (1983), Sowell (1981)]. The current advantaged status of these groups
has been explained by particular historical and institutional developments, rather
than as revealing "reverse discrimination", and these explanations persuade me
of the value of this method of analysis.
In a study of the relation between the larger religious groups and earnings,
T o m e s (1984) found small and statistically insignificant effects among Catholic,
Protestant, and " N o n e / O t h e r " categories. 28 He also found no statistically sig-
nificant difference among various Protestant denominations. Tomes provides a
useful distinction between estimates with purely exogenous variables held con-
stant - family background, age (and age squared), and location of residence - and
estimates in which potential " o u t c o m e " variables, like education and self-
employment, were additional control variables. Another interesting feature of this
study was the distinction between one's current religious affiliation and that of
one's upbringing. Current affiliation is, to some extent, endogenous, and there
were some interesting, although not startling, differences in the estimates when
the two definitions of religious affiliation were used.
4.2.3. Empirical studies of wage discrimination in indioidual firms

and "reoerse regression"
A brief examination of the econometric analysis of discrimination in individual

firms is useful for three reasons. First, discrimination by firms, identified singly,
has come under close scrutiny by various groups in society, mainly as a result of
antidiscrimination laws and executive orders. Regression analyses of the Model
(I) type are frequently offered as evidence in cou,~ cases and other litigation
proceedings stemming from these laws and regulations. 29 Discrimination in firms
may consist of the differential treatment of majority and minority groups in
hiring, placement, wages, promotion, layoffs, and in other ways. I will refer only
to wage discrimination in this discussion.
A second reason for interest in these studies is that they have the advantages,
relative ~o marketwide studies, of explicit and well-defined objectives and
2~Tomes found relativelylarge positive effects for Jews, but this group was numericallysmall in his
sample and the differences were sometimes not statistically significant.
29See Baldus and Cole (1980), Finkelstein (1980), and Fisher (1980) for a discussion of statistical
analyses in court cases of discrimination and for extensive citations. An example of an econometric
study of wage discrimination in a single firm, listed in Table 13.6, is that by Malkiel and Malkiel
(1973), although this study was not used in litigation proceedings.
straightforward procedures. The analysts' objectives are usually to support their

clients, and the objective of the court is to use these studies to determine whether
an employer is guilty or innocent. The procedures involve a Model (I)-type
regression in which the employer's criteria for wage payments may be specified in
detail, and those criteria that are correlated with race or gender can be explicitly
included in the estimation model. Moreover, many characteristics associated with
the workers' productivity may be clearly exogenous to the employer, even though
they are not exogenous to the market as a whole. Regression analyses with
nationwide samples usually suffer from ambiguities and vagueness about both
their objectives and model specifications.
A third reason for our attention is the innovative use of "reverse regression",
which raises some interesting methodological points even if, as I will argue, it
does not offer a preferred model for estimating wage discrimination. Indeed, this
section will examine only this aspect of the studies of individual firms. The actual
analyses are often buried in trial proceedings, and there is no space to present
them here. More important, I believe that studies of discrimination by single
firms do not provide useful measures of marketwide discrimination, which is my
main interest. The reason is that the samples are based on company records, and
the selection rules for inclusion in the sample are seldom known. The companies
studied in court cases are not a random sample of all companies, and their
recruitment policies do not pretend to yield random samples.
Company records generally apply to a single industry and a few occupations,
and the question of how market discrimination affects the distribution of
minority workers in the industry and occupation is not examined. More gener-
ally, no valid conclusions about discrimination can be reached without attention
to the company's recruitment or selection procedures. For example, perhaps
the company has a reputation for discrimination against women that restricts
the pool of female applicants. Maybe only a small number of newcomers to the
community constitute this pool. The statistical analyst usually deals with the
employees on board or, at best, with persons who have applied to the company.
A famous study by Conway and Roberts (1983) of alleged salary discrimination
against women in a large bank, for example, involved a sample of 274 employees,
of whom only 37 are women. Under these circumstances, generalizing about
discrimination in the market as a whole on the basis of studies of one or several
companies is not valid.
Studies by Roberts (1980) and by Conway and Roberts have given prominence
to the statistical model known as "reverse regression" (hereafter RR). 3° The term
refers to "reversing" the Model (I) regression in which the wage, IV, is regressed
3°A large number of articles on reverse regression in discrimination analyses appeared around the
time and soon after Roberts's article (1980), and a symposium on the issue appears in Journal of
Business and Economic Statistics, vol. 1, January 1983. My understanding of the issues owes much to
Arthur S. Goldberger, and my discussion is based on Goldberger (1984), but I am solely responsible
for any errors in this section.
768 G. G. Cain
on a vector of productivity variables, X, and a minorty/majofity-status variable,

Z. In R R , X is in effect regressed on W and Z. Assume Z refers to gender. One
motivation for R R may lie in the question it addresses: "Holding wages constant,
are men less qualified than women?" This reverses the customary question in
direct regression: "Holding qualifications (the X's) constant, are men paid more
than women?"
Another motivation for R R is the classic problem, in regression analysis, of
errors in the independent variables. Given the regression, W = a + bX, in which
X contains random measurement error but wherein the regression is otherwise
correctly specified, R R - h e r e , ) ( = c + d W - p e r m i t s the coefficients b and 1 / d
to give boundaries on the true linear relation between W and the correctly
measured X. When X is a collection of variables, the dependent variable in R R
may either be (a) each individual X-component regressed against W and Z in a
system of equations, or (b) the fitted part of the direct regression for the case in
which Z = 0, specifically, W = X ' B . Thus, X ' B is regressed on W and Z in R R .
I will sometimes simply refer to X as the "dependent variable" in R R .
The initial appeal of R R is the reasonable proposition that the econometrician's
usual set of productivity variables, X, is not a perfect measure of a worker's
productivity. Assuming that W is a function of "true productivity", X*, the
following model may be specified:
W=X*+AZ+v, where Z = 1 for men and Z = 0 for women; (16)
X * = G Z + u, where G > 0; (17)
X = X* + e. (18)
By eq. (16) the wage is a function of true productivity, X*, which is unobserved
by the econometrician. Holding X* constant, A = 0 implies no discrimination
against women, and A > 0 implies discrimination against women. Equation (17)
says that men are more productive than women; this assumption will be
maintained throughout this discussion. Equation (18) says that X is a fallible
measure of X*. The usual assumptions about the error terms are that v, u, and e
are independent of each other; that v and e are independent of X* and Z; and
that u is independent of Z.
Assfffiae that the econometrician estimates the direct regression:
W = B X + C Z + v'. (19)
It follows from (16)-(18) that C is an upwardly biased estimate of the true

relation, A. Assuming A = 0, C will be greater than zero, implying discrimina-

tion against women when none exists. [The algebra and further discussion of
these qualitative results are found in Goldberger (1984).] In eq. (19), X* is an
omitted variable, as revealed by the assumed correct wage eq. (16). By (17), X* is
systematically related to gender. Thus, positive values of Z in (19) partly
represent " m o r e " X* in addition to representing "maleness" (since Z = I for
men). Therefore, the coefficient C will be positive owing to the relation of X* to
Z even if the true male effect, A, is zero. Clearly, C is more upwardly biased as G
is larger in (17) and as the variance of e is larger (implying that X is a less
accurate measure of X*) in (18).
At first glance, eqs. (16)-(18) may seem so plausible that the systematic
upward bias of C in (19) seems incontrovertible. The first glance is misleading.
There is no basis for assuming that the employer pays according to "true
productivity", X*. On the contrary, we should expect that the systematic basis
on which employers pay their workers is a basis of observable variables. A
random error term should also be included to allow for miscellaneous factors that
may be assumed unrelated to gender. This latter point reintroduces the earlier
argument that in a discrimination case, any systematic productivity trait that is
correlated with minority-group status should be included in the employer's list of
the X-variables. These assumptions effectively restore to legitimacy either eq.
(19), for the case of a single X-variable, or eq. (20), as written below, with
multiple X-variables:
W = X ' B + C Z + v'. (20)
We may further assume that there is generally a positive relation between X

and Z in the sense that the men's mean of the X-vector, weighted by their
" p r i c e s " - t h e B-vector-is larger than that for women, that is
F.( x - ) s > E( X'w)S. (21)
This inequality is the analogue of (17), which expressed the assumption that men
are more productive than women. Given the assumption that the X's are
positively correlated with X*, (21) implies G > 0 in (17), and conversely. How-
ever, (17) and (21) do not imply that the expected value of X*, holding constant
the observable X's, is greater for men than for women, and without this
assumption there is no basis for assuming that C is upwardly biased in (20).
Despite the assumed correctness of (17) and (21), the direct regression of (20)
gives an unbiased estimate of the gender (male) coefficient. Equation (20) has the
virtue of focusing attention on the explicit measure of alleged discrimination, W,
and of leading all the interested parties-the econometrician, the defendant
employer, the plaintiff, and the adjudicator- to address the same questions raised
770 G. G. Cain
earlier in this chapter about Model (I):

(a) Are the variables in X proper control variables; that is, are they exogenous
with respect to the employer's behavior?
(b) Are there omitted productivity variables that are systematically related to
gender?
It should be clear that eq. (20), which expresses an observable relation and is to
be analyzed in conjunction with an inquiry into questions (a) and (b), allows for
the model expressed as eq. (16) to be a special case. As mentioned earlier, X*
could be assumed to be an omitted variable and, given (17)-(18), the upward bias
in C follows. But there is no reason to believe that any such concept as "true
productivity" will be operational. Furthermore, if X* is redefined to be the
employer's "assessed productivity" [see Roberts (1980), Conway and Roberts
(1983)], then we return to the reasonable expectation (or requirement) that the
employers simply point out which criteria in their assessment have a net correla-
tion with gender-that is, a correlation that persists after controlling for the
observable X's. Again, a straightforward analysis of "omitted variable" bias in
eq. (20) could provide a qualitative answer about the direction of bias in C. The
direct regression associated with (20) leads us to focus attention on specific
sources of any bias. This has more appeal to me than a model in which the wage
is presumed to be determined by "true" (or "assessed") productivity combined
with the presumption that male superiority regarding that productivity is
maintained after holding constant observable measures of productivity. These
presumptions prejudge the very issue that is to be investigated, and these
presumptions are precisely what eqs. (16)-(18) embody.
The model in eqs. (16)-(18) has two multivariate analogues that have been
analyzed by Goldberger (1984). In one, each X-variable in the vector of observed
indicators, X, is assumed to be a fallible measure of the corresponding element in
a vector of unobserved true productivity determinants, X*. In this case R R is
not necessarily superior to direct regression even though one may assume,
analogous to (17), that
X* = HZ + u, (17a)
with all elements of H > 0.

In the other nmltivariate model the X-vector is assumed to provide multiple
indicators of a single X*, so (18) becomes
X = OX* + e. (18a)
Let a representative indicator be Xj. The RR for the j-indicator is
L=c +dy+ z, (18b)

and the estimator, A*, of the gender coefficient in the true wage equation, (16), is
A T = - ~ / d j.
This multiple-indicator model is said to be the one "stochastic specification
under which reverse regression provides a valid estimator" of the gender coeffi-
cient [Goldberger (1984, p. 314)]. Because each indicator of X is assumed to
provide a consistent estimator of the true gender coefficient, these implied
restrictions on the different estimators can be tested. Few such tests have been
tried, and those few are hardly supportive of the model [see Goldberger (1984),
Green and Ferber (1984)].
One of the questionable assumptions of the multiple-indicator model as it
applies to gender discrimination is that each indicator, Xj in X, is independent
of gender, holding constant X*. This assumption is associated with eq. (18a),
where e is assumed to be independent of gender, or with eq. (18) in the classical
errors-in-variable model with one independent variable. To illustrate why this
assumption and the implied restrictions on the coefficients in (18b) are not likely
to hold, consider an X that has the following four elements, which are positively
correlated with true (or assessed) productivity, X*, but which differ in their
correlation with gender: X 1 = verbal test score; X 2 = mathematics test score;
X 3 = manual dexterity; X 4 = physical strength.
The gender correlations will reflect, let us assume, an advantage of women over
men in verbal ability and manual dexterity and an advantage of men over women
in mathematical ability and physical strength, which are gender-linked relation-
ships that have been generally found [see Anastasi (1969, p. 421)]. To keep the
statistical relations simple, assume that each element of D is the same. Then if a
nondiscriminating employer hires men and women at a specified wage, we should
observe that women's low scores on X 2 and X 4 will tend to be compensated by
higher scores on X 1 and X 3, and conversely for men. The dependence of e with
gender ( = Z ) in (18a) will induce different signs of the fj in (18b). In particular,
with Z = 1 for men, f l and f3 will tend to be negative and f2 and f4 will tend to
be positive. With dg positive, the estimators of the true gender effect, A~, will not
all have the same sign.
The reasons why the multiple-indicator model is invalid in this example are, I
believe, realistic and prevalent; specifically, the presence of some gender-linked
productivity traits that favor women, the fact that employee productivity typi-
cally involves multiple skills, and the fact that employers recognize that skills
(and traits) are typically compensating (or substitutable). 3~
31Ahypothetical example of substitutable gender-linkedtraits that brings out these statistical points
is available from the author. It should be noted that a model in which wages are determined by
gender-linked traits, X, along with an assumption that other traits, holding constant X, are
uncorrelated with gender simply illustrates the model in eq. (20), which is the same as the "multiple
cause model" that Goldberger (1984) discusses.
772 G, G. Cain
5. Policy implications and conclusions
5.1. Explaining and judging discrimination: The diversity of cases
This chapter began with the normative issue of equity in outcomes measuring
economic well-being among racial, ethnic, and gender groups. Inequities appear
to be widespread, and our economic theories of why they persist are only
moderately helpful.
At one extreme, the outcomes experienced by earlier immigrants to the United
States suggest an optimistic view of both the ethical and the scientific judgments
about the workings of the economy generally and of labor markets more
particularly. 32 Although discrimination against early immigrant groups was not
analyzed in this chapter, the references to the achievements of immigrants who
were Irish Catholics, Italians (mainly from southern Italy), Japanese, and Jews
(mainly from eastern Europe) seem to show a pattern in which groups who were
initially " h a v e nots" in the United States and who faced discrimination gradually
attained an equal economic status to whites whose ancestry was Anglo-Saxon
Protestant and who were the "haves". Such an evolution is consistent with a
neoclassical view of the workings of competitive markets, assuming that the
productive capacities of the different ethnic groups are equal and that the
economy is sufficiently competitive.
A more specific application of economic principles to an analysis of dis-
crimination involves Hispanic Americans, who are mainly recent immigrants.
Their lower relative earnings may be rationalized by a theory of the determinants
of earnings that assigns important roles to information about the labor market, to
facility in the English language, and to education, measured by years of school-
ing. Such theories are qualitatively supported by empirical evidence. Whether the
evidence shows that the quantitative gap in earnings between Hispanic and
non-Hispanic whites is explained by these theories is not clear to me.
The difference in market earnings between men and women can be rationalized
by economic theories of the gains from specialization and investment in human
capital, combined with an assumption of voluntary choices by women to special-
ize in the home sector. This earnings gap, particularly between white men and
women, is one of the largest and most time-persistent of the comparisons
discussed in this paper (see Table 13.5).
In ano@er paper [Cain (1985)] I have suggested that the theory of voluntary
choice regarding labor market activities should lead to equality in total incomes
32Hispanic immigrants are considered to be recent immigrants, and most blacks who came to the
United States were not voluntary immigrants, so these two groups are not included in the group
referred to as the earlier immigrants.
received by men and women, if not to equality in labor market wages. I assume
equality in women's and men's productive capacity, in the nonpecuniary aspects
of their work, and in their leisure consumption. I then test for the equality in
income received by assuming that husbands and wives share their household
income equally while married. Even with this assumption women were found to
have a substantially lower present value of lifetime income than men: the ratios
were between 0.7 and 0.9 (depending on various assumptions). These are,
however, closer to unity than the usual measure of women-to-men ratios of
wages, as reported in Table 13.6.
I also examined the total time spent in housework and market work combined
for men and women. The data are weak, but the available evidence suggests near
equality among husbands and wives [Cain (1984b)]. It is not clear how the
inclusion of men and women without spouses present would affect this compari-
son. More women than men are likely to head single-parent families, and many
of these women have the double burden of market work combined with a heavy
workload at home, especially child care. On the other hand, those female
heads-of-household who are recipients of public welfare tend not to work much
in the market; indeed, the conditions of their welfare receipt discourage market
work. Men who are single-parent heads-of-households are not likely to be on
welfare. Another important unknown factor in my attempted comprehensive
measure of economic well-being is the nonpecuniary utility (or disutility) that
men and women obtain from their work.
The wages, earnings, and incomes of black workers and black households are
substantially less than those of whites, and the conventional human capital
variables, such as education, training, and health care leave much of the differ-
ence unexplained. Even if they explained more, the question would then be: Why
is the market for such human capital investments functioning so poorly that
blacks continue to be shortchanged? If whites find it profitable-in terms of
higher earnings and better j o b s - t o make these investments, why are blacks'
opportunities for these investments so curtailed? If the answer is not labor
market discrimination, is it discrimination in the capital markets that supply
funds or other sources of human capital investments? It is not scientifically
satisfactory for economists to argue that labor market discrimination is minimal
if they then have no explanation for how discrimination in capital markets
creates and sustains the inequities we measure in the labor market.
The case of blacks in the United States appears to offer the strongest evidence
for the reality of labor market discrimination and, given existing economic
theories, for flaws in the competitive functioning of the market. In these respects,
the case of blacks is at the other end of the spectrum from that of non-Hispanic
early-immigrant groups. Economic discrimination, whether measured by average
family incomes or by comparing wages when exogenous productivity factors have
been held constant, is substantial for blacks and is nonexistent or insubstantial
774 G. G. Cain
for various former-immigrant white (and some Oriental) groups. For those
groups, but not for blacks, the market has virtually eliminated the differences in
economic attainment that were present decades ago.
5.2. The effects of discrimination on total output
The foregoing remarks refer to descriptive statistics, including regression analyses

of earnings functions, and to the normative issue of equity. Let us turn to positive
economics and the tasks of analysis, prediction, and explanation. One issue that
has not been much studied is' the implications of discrimination for economic
efficiency as measured by total social income. I have elsewhere [Cain (1985)]
addressed this question regarding discrimination against women, and my conclu-
sions were embarrassingly thin. The neoclassical economist's convention, and
perhaps it is an obligation, to take tastes-individual preferences-as given
prevents the automatic translation of "different prices (wages) for the same good
(labor)" into a loss in total social income (or total utility). 33 In a competitive
economy in which tastes are the fundamental cause of discrimination there is no
presumptive case for inefficiency.
Perhaps the underlying atomistic competitive model with only private (inter-
nalized) benefits and costs is too narrow. Alexis (1973, p. 297) distinguishes
between discrimination, in which the discriminator is indifferent " t o the welfare
of the avoided [black] person", and racism, where "the decision maker is not
indifferent to the relative economic status of nonpreferred persons". Alexis's
point may be extended to posit racist discrimination as a pubhc "good" among
the majority group, and this could explain the persistence of labor market
discrimination. However, it is difficult to reconcile the idea of racism as a public
good with the proliferation during the last 30 years of antidiscriminatory legis-
lation and court decisions, which ought to reflect the public's externalities
regarding discrimination. On the other hand, there have also been numerous
references in recent American politics to the "silent majority" among whites, who
oppose the pro-civil-rights legislation. The purely political aspects of this topic
are well beyond the scope of this chapter. The issue of externalities is not, but I
do not pursue it.
Discrimination that results from private and government monopolies deserves
further study, particularly if the term monopoly is extended to include collusive
action B a t deprives minorities of access to various opportunities, some of which,
like housing, may be only indirectly related to labor markets. Akerlof (1976)
33Ambiguities about total welfare when tastes for discrimination are part of a person's utility
function are discussed by Thurow (1969, pp. 116-138), Toikka (1976), and Lundahl and Wadensj6
(1984, pp. 81-108).
analyzes several models illustrating collusion, including conspiratorial acts and

intimidation, that are sources of the oppression of minorities. Even if I am
correct in my judgment that monopoly is not the predominant source of
discriminatory wage differences, this does not imply that the benefits from
attacking this source are less than the associated costs.
Aside from monopoly, the standard cases of market failure that point to
inefficiencies that might be overcome through government intervention do not
emerge from economic theories of discrimination. Perhaps an exception is the
externalities of information concerning the productive capacities of minorities.
Tests, licenses, certifications, and other such signals are used extensively in labor
markets, and the private costs of obtaining accurate information about workers'
productivity may be high relative to the private benefits, which are not neces-
sarily appropriable as a private good. Clearly, those who believe in the equality of
productive capacities across the groups under study will be more likely to believe
in the benefits of more scientific information about productive capacities. The
history of the stereotypes of inferiority among ethnic minorities in the United
States is too familiar to cite, and the demise of these stereotypes regarding the
earlier immigrant groups is part of the latter's success story.
In the main, however, I was not able to extract efficiency losses from the
economic theories of discrimination [Cain (19~5)], which reflects my agnostic
view of these theories expressed in Section 3. Something is amiss. Discrimination
in its many forms, not only economic, is widely believed to suppress the
achievements of the minority group with no fully offsetting gains to the majority
group. The economists that I know agree on this, yet conventional economic
theories do not, to my knowledge, explain or analyze this widely shared belief.
Economists have prescribed limits for themselves in many policy spheres.
Economics does not distinguish among the ethical merits of different tastes;
between, say preferences for physical attractiveness or for race. As economists we
have nothing to say about the justness of laws that prohibit an employer from
refusing to hire someone on the basis of color but that permit hiring on the basis
of physical attractiveness. As citizens we may, of course, have strong opinions
about such matters.
5.3. Measuring the impact of policies
A more promising role for economic analysis lies in the measurements and
methods that permit prediction. Empirical regularities, such as time trends, may
be established and be useful even in the absence of fully developed theories. At a
minimum, the measurements provide valuable data for monitoring progress or
regress regarding discrimination.
776 G.G. Cain
A more ambitious form of prediction is that concerning policy instruments.

The policies available to government may be classified into three categories: (a)
macroeconomic (mainly monetary and fiscal policies affecting aggregate demand),
(b) income transfers, and (c) microeconomic structural labor market policies. I
discuss only the third. Supply-side structural policies typically comprise educa-
tion and training programs. They have tended to be directed to low-income
workers, with no special targeting to workers of a particular race or gender. In
this regard, the supply-side policies tend to differ from the demand-side policies.
Microeconomic demand-side policies might also be general, such as public
employment programs or wage subsidies for unemployed and low-income workers,
but the demand policies that have received most attention are those that directly
forbid discrimination or promote preferential treatment of minority workers in
hiring, placement, pay, or employment security. Preferential treatment is also
called affirmative action.
Research aimed at evaluating these policies is abundant and controversial. See,
for example, the proceedings of a conference on the labor market effects of
antidiscrimination legislation that appeared in the Industrial and Labor Relations
R e v i e w (vol. 29, July 1976). The essential difficulty in evaluating these programs
is the classic problem of trying to make inferences from an uncontrolled
experiment. We observe an outcome for a group of workers, some of whom
participated in the program or, alternatively, had the program imposed on them.
To establish causality between program status and the outcome, the factors that
selected the workers into the program must be either (a) known and controlled
for in the evaluation, or (b) known to be unrelated to the outcome.
It is difficult to know enough about the selection process and about all the
causes of the outcome to justify either (a) or (b). Random assignment would
satisfy condition (b), but this selection procedure is rare. Legislators and courts,
therefore, seldom rely on the research of economists to determine the fate of
government programs.
Decades of empirical research on labor markets, much of it like the research
reported in Section 4, can be helpful in estimating the effects of a variety of
policy variables on the earnings of workers, even if the research does not provide
conclusive answers. Customarily we use cross-sectional data for empirical work,
although the policy question is invariably one of predicting a change over time.
The implications for policy from cross-sectional research arose in the previous
section. Some variables may be only minimally affected by policy, like "Fears
since imm~igration" among the existing stock of immigrants; some partly affected,
like "years of schooling"; and some almost wholly affected, like "participation in
a government training program". Policies related to the cross-sectional findings
will not be discussed, but some of the issues about how the above-mentioned
variables affect outcomes in the labor market appear in time-series analyses of
discrimination, which is the final topic of this chapter.
5.4. Results o f time-series analyses
For my purposes the essential facts from time-series data that pertain to
economic discrimination have been revealed by Table 13.5 in Section 2. This
table shows two major challenges. One is the near-constant ratio of women's-to-
m e n ' s wages over a 40-year period, using the data on earnings of year-round,
full-time workers. The second is the slow increase in the ratio of black-to-white
wages a m o n g men, which in 1982 was only 0.72. These trends in wages could be
usefully supplemented with an analysis of trends in other measures of attainment
in the labor market, including occupational attainment, labor force participation,
and e m p l o y m e n t / u n e m p l o y m e n t rates, but space limitations preclude more than
brief remarks.
The sharp increase in labor force participation rates (LFPRs) by women and
the moderate decrease in LFPRs of men during the last 40 years have brought
men and women into closer equality with respect to the quantity of time spent in
market work, although men still spend about twice as many hours of their adult
life in market work as do women [Cain (1984a)].
The increase in L F P R s for women has been the result of two trends which, as
noted in Section 3, have contrasting effects on the trend in average wages of
women: (a) women who work are working more continuously and for more years
of their adulthood; (b) a larger fraction of women are entering and reentering the
work force. Trend (a) should increase the average wage, because the wage should
increase with experience and seniority. Trend (b) probably decreases the average
wage because the composition of workers is altered by the influx of women with
less-than-average experience-referred to as "adverse selection". The "adverse
selection" hypothesis is strongly advocated b y Smith and Ward (1984), but see
Fuchs (1984) for counter arguments and Mallan (1982) for counter evidence. 34
The L F P R s of men between the ages of 50 and 65 have declined during the last
20 years or so. Are these early retirements and disability-related retirements
concentrated among low-wage workers? The substitution effect of wages on labor
supply suggests a yes answer, but the income effect suggests otherwise. Retire-
ment may be considered a "luxury good" that is selected by workers with
above-average incomes. The net result of these contrasting effects on the trends
a m o n g men needs to be studied.
The changing composition of the male labor force has also been examined in
analyzing trends in wages of men, black and white. Butler and Heckman (1977)
34Further counter evidence to Smith and Ward is Maloney's finding that the wages of husbands
and wives who work in every year covered by the Michigan Panel Study of Income Dynamics show
declining women-to-men ratios from 1975 through 1980 [Maloney (1983, pp. 135-139)]. These data
may, of course, simply reflect the contrasting age/earnings profiles shown in Figure 13.2. Whether the
declining ratios imply labor market discriminationdepends on whether they reflect voluntary choices
of these wives to commit less effort than their husbands to the market sector.
778 (7. G. Cain
Table 13.9
Median money incomes and income ratios for black and white male workers, 1948-1982,
in constant 1982 dollars
White Black Ratio White Black Ratio

Year medianS medianS B/W Year medianS medianS B/W
1948 $10064 $5465 0.54 1965 $16185 $8710 0.54

1949 10006 4844 0.48 1966 16631 9212 0.55
1950 10862 5899 0.54 1967 16901 9653 0.57
1951 11524 6 346 0.55 1968 17 388 10 551 0.61
1952 1J 837 6487 0.55 1969 17812 10508 0.59
1953 12237 6760 0.55 1970 17428 10490 0.60
1954 12080 6011 0.50 1971 17248 10351 0.60
1955 12776 6724 0.53 1972 18029 11100 0.62
1956 13 558 7113 0.52 1973 18 360 11551 0.63
1957 13 402 7096 0.53 1974 17 330 11135 0.64
1958 13275 6614 0.50 1975 16679 10511 0.63
1959 13937 6561 0.47 1976 16849 10540 0.63
1960 14003 7 367 0.53 1977 16 889 10 326 0.61
1961 14290 7385 0.52 1978 16945 10796 0.64
1962 14859 7318 0.49 1979 16363 10604 0.65
1963 15151 7 874 0.52 1980 15 612 9 786 0.63
1964 15361 8708 0.57 1981 15172 9624 0.63
1982 14748 9493 0.64
Source: U.S. Bureau of the Census, Current Population Reports, Series P-60, No. 142,
Money Income of Households, Families, and Persons in the Unites States: 1982 (Washing-
ton, D.C.: U.S. GPO, 1984), Table 40.
suggested that part of the rise in the ratio of black-to-white wages through the
1960s and into the early 1970s could be attributable to a selection of higher-earn-
ing workers among blacks, relative to whites. Their argument is as follows. Black
men's LFPRs have declined more than have the LFPRs of white men. Assume
that the men who drop out of the labor force tend to be low earners. The black
male labor force would then have relatively fewer of the low earners remaining,
and the average earnings of blacks -which is measured only for those in the labor
force-will rise relative to that of whites.
The issue is not resolved, and it illustrates themes that I wish to stress in this
survey: the need for closer attention to descriptive economic statistics about the
labor market statuses of majority and minority groups, and the need to specify
the purposes of one's analysis. A problem in charting trends is that the use of
broad pol~ulation groups may introduce exogenous compositional changes (like
the age distribution) that should be held constant, while narrowing the compari-
son to various subgroups may reflect selection according to an endogenous
characteristic (like full-time work status) that should not be held constant.
Table 13.9 shows another version of the black-to-white comparison among
men, but this time for all men who worked, rather than for year-round, full-time
workers, as in Table 13.5. Table 13.9 shows the time series for money incomes in
constant i982 dollars for each year, 1948-1982. (The trend in money income is
very similar to the trend in money earnings.) The dollar figures show the striking
reversal from 1973 on, of the long-term growth in real incomes for both groups
that prevailed from 1948 to 1973.
The overall picture regarding the black-to-white (B/W) ratios in Table 13.9 is
similar to that summarized in Table 13.5, but the year-by-year statistics bring out
more clearly the three periods of stability and change in the B / W ratios:
1948-1965, stability; 1966-1974, growth; 1975-1982, stability. Presumably a
theory or an empirical evaluation of specific hypotheses about labor market
discrimination against blacks should be able to explain these stylized facts.
Freeman (1981) discusses three main contending explanations, in addition to
the Butler-Heckman hypothesis about selectivity in the composition of the
populations. One is that the B / W ratio is pro-cyclical-rising during periods of
prosperity, when unemployment is low, and falling during recessions. This is
consistent with its growth in 1966-1974, when labor markets were relatively tight,
and with the ratios that are low relative to surrounding years in the recession
troughs of 1949, 1954, and 1958. However, the hypothesis is not supported by the
stability of the ratios throughout the period of 1948-1965, when unemployment
rates were relatively low, or by the behavior of the ratios in the later recession
troughs, 1961, 1971, 1975, and 1982.
A second hypothesis is that the B / W ratios were affected by the surge of legal
measures that may be said to have begun with an Executive Order in 1961 (No.
10925) that reinforced a somewhat dormant ban on discrimination by firms doing
business with the federal government, followed by the Civil Rights Act of 1964
and subsequent legislation. Measuring the impact of these various forms of
government intervention is difficult, however. How does one quantify the re-
sources devoted to the intervention? How do we separate the effects of the
legislation from the effects of the political and social climates that fostered the
legislation? It is no surprise that the attempted evaluations of the legislation have
not been conclusive.
A third hypothesis to explain the time series of black-to-white earnings ratios
focuses on education, where this may be interpreted narrowly as years of
schooling completed or broadly as a general indicator of human capital, includ-
ing qualitative aspects of schooling as well as the training, information, and
mobility that are affected by schooling. In either case, the emphasis is on the
supply side of the labor market and the relative increases in the human capital of
black men.
The role of education in this stream of research has had a curious history.
Early quantitative studies based on the 1940, 1950, and 1960 censuses con-
sistently showed two rather pessimistic regularities about male workers: (a)
holding age constant, the B / W earnings ratio generally declined as years-of-
schooling increased; (b) holding years of schooling constant, the B / W ratio
generally declined as age increased. See Zeman (1955) as quoted in Becker, (1971,
780 G. G. Cain
Table 13.10
Median years of schoofingcompleted and schooling ratios,
white and black men, various populations, 1940-1980.
Median years of
Population, by educational attainment
age and labor Ratio
Year force status White Black B/W
1940 All males 25 + 8.6 5.7 0.66

Males, 25-59 10.3 7.0 0.68
1950 All males 25 + 9.3 6.8 0.73
Males 25-29 12.0 8.6 0.72
1952 Males, 18+ in
civilian labor
force 10.8 7.2 0.67
1959 Males, 18+ in
civilianlabor
force 11.9 8.3 0.70
1970 All males 25 + 12.2 9.9 0.81
Males, 25-29 12.6 12.2 0.97
Males, 18+ in
civilianlabor
force 12.1 11.1 0.89
1980 All males 25 + 12.5 12.0 0.96
Males, 25-29 12.9 12.6 0.98
Males, 18 + in
civilianlabor
force 12.7 12.2 0.96
Sources: "All males" and "Males, 25-29": U.S. Bureau of

the Census, Statistical Abstract of the United States: 1982 - 83,
103rd edition (Washington, D.C.: U.S. GPO, 1983) p. 143;
"Males, 18+ in civilian labor force", U.S. Department of
Labor, Handbook of Labor Statistics 1978, Bulletin 2000
(Washington, D.C.: U.S. GPO, 1979) p. 68; "Males, 16+ in
civilian labor force", U.S. Department of Labor, Bureau of
Labor Statistics, Handbook of Labor Statistics, Bulletin 2175
(Washington, D.C.: U.S. GPO, 1981), pp. 152-153.
p. 111) for 1940, A n d e r s o n (1955) and F r e e m a n (1973b, p. 85) for 1950, a n d

H a n o c h (1967) for 1960.
A t face value, (a) implies that a n equal growth in educational a t t a i n m e n t over
time w o u l d widen the B / W difference in earnings. However, the increase in
e d u c a t i o n a l a t t a i n m e n t by black m e n has exceeded that of white men, particu-
larly b e t w e e n 1960 a n d 1980 (see Table 13.10). This period includes, b u t does n o t
c o i n c i d e with, the period of the rise in the earnings ratios, 1967-1974. By 1980
the B / W differences in m e d i a n e d u c a t i o n a l a t t a i n m e n t s had been virtually
e l i m i n a t e d , although there may well r e m a i n differences in the quality of school-
ing.
In addition to the relative increase in the quantity of schooling obtained by

blacks, a new development in research findings in the 1970s re-emphasized the
i m p o r t a n c e of schooling. Recent studies show that the wage returns to schooling
were b e c o m i n g m o r e equalized between blacks and whites, although this was
mainly true for the y o u n g age groups and those with some college [see Smith and
Welch (1977), F r e e m a n (1977)]. Like the vintage hypothesis regarding age effects,
these new and higher education effects for black men relative to white men
remain to be tested in the years to come.
This brief survey of proposed explanations for the trend in black-to-white male
earnings illustrates the tentativeness of empirical regularities regarding labor
market discrimination and the consequent difficulty in drawing policy implica-
tions. Estimated relations from cross-sections at different times show widely
varying results, and the time series, with relatively few observations and m a n y
c o m p e t i n g hypotheses, do not yield firm empirical regularities either.
5.5. Final word
T h e economics of discrimination is a particularly complex subject. My j u d g m e n t

is that the theories of discrimination have been useful for providing definitions
and for suggesting measurements of discrimination but not for providing con-
vincing explanations of the p h e n o m e n o n nor of its patterns. The econometric
work has also been useful, but to my eyes more so for its descriptive content than
for testing hypotheses or for providing estimates of causal relations.
References
Abowd, John M. and Mark Killingsworth (1982) "Employment, wages, and earnings of Hispanics in
the federal and non-federal sectors: methodological issues and their empirical consequences", in:
G. Borjas and M. Tienda, eds., Hispanic labor conference. Washington, D.C.: National Commission
on Employment Policy. (Subsequently published in G. J. Borjas and M. Tienda, eds., Hispanics in
the U.S. Economy. Orlando, Fla.: Academic Press, (1985, 77-125.)
Aigner, Dennis J. and Glen G. Cain (1977) "Statistical theories of discrimination in the labor
market", Industrial and Labor Relations Review, 30:175-187.
Akerlof, George A. (1976) "The economics of caste and of the rat race and other woeful tales",
Quarterly Journal of Economics, 94:599-617.
Alchian, Armen A. and Reuben Kessel (1962) "Competition, monopoly and the pursuit of pecuniary
gain", in: Aspects of labor economics, Conference Volume of the Universities-National Bureau
Committee for Economic Research. Princeton, N.J.: Princeton University Press, 156-175.
Alexis, Marcus (1973) "A theory of labor market discrimination with interdependent utilities",
American Economic Review, 63:296-302
Anastasi, Anne (1969) "Differential psychology", in: Encyclopedia Britanniea. Chicago: Encyclopedia
Britannica, Inc., William Benton, vol. 7.
Anderson, C. Arnold (1955) "Regional and racial differences in the relation between income and
education", The school review, 63:38-45.
Arrow, Kenneth (1972) "Models of job discrimination", and "Some mathematical models of race in
the labor market", in: A. H. Pascal, ed., Racial discrimination in economic life. Lexington, Mass.:
Lexington Books, 83-102 and 187-204.
782 G. G. Cain
Arrow, Kenneth (1973) "The theory of discrimination", in: O. A. Ashenfelter and A. Rees, eds.,
Discrimination in labor markets. Princeton, N.J.: Princeton University Press, 3-33.
Ashenfelter, Orley (1972) "Racial discrimination and trade unions", Journal of Political Economy,
80:435-464.
Ashenfelter, Orley and Timothy Hannan (1986) "Sex discrimination and product market compe-
tition: the case of the banking industry", Quarterly Journal of Economics, 101: 149-173.
Astin, Helen S. and Alan E. Bayer (1972) "Sex discrimination in academe", Educational Record,
53:101-118.
Baldus, D. C. and J. W. L. Cole (1980) Statistical proof of discrimination. New York: McGraw Hill.
Becker, Gary S. (1971) The economics of discrimination. Chicago: The University of Chicago Press.
(Original edition, 1957.)
Becker, Gary S. (1959) "Union restrictions on entry", in: P. D. Bradley, ed., Thepublic stake in union
power. Charlottesville, Va.: University of Virginia Press, 209-224.
Becker, Gary S. (1968)"Discrimination, economic", in: D. L. Sills, ed., International encyclopedia of
the social sciences. New York: Macmillan and Free Press, vol. 4, 208-210. (Reprint edition, 1972.)
Bergmann, Barbara R. (1974) "Occupational segregation, wages and profits when employers dis-
criminate by race or sex", Eastern Economic Journal, 1:103-110.
Blau, Francine D. (1977) Equalpay in the office. Lexington, Mass.: D.C. Heath and Co.
Blinder, Alan S. (1973) "Wage discrimination: reduced form and structural variables", Journal of
Human Resources, 8:436-455.
Bunting, Robert L. (1962) Employer concentration in local labor markets. Chapel Hilh University of
North Carolina Press.
Butler, Richard and James J. Heckman (1977) "Government's impact on the labor market status of
black Americans: a critical review", in: Equal rights and industrial relations. Madison, Wis.:
University of Wisconsin, Industrial Relations Research Association, 235-281.
Cain, Glen G. (1976) "The challenge of segmented labor market theories to orthodox theory: a
survey", Journal of Economic Literature, 14:1215-1257.
Cain, Glen G. (1984a) "Lifetime measures of labor supply of men and women", Discussion Paper
No. 749-84, Institute for Research on Poverty, University of Wisconsin-Madison.
Cain, Glen G. (1984b) "Women and work: trends in time spent in housework", Discussion Paper No.
745-84, Institute for Research on Poverty, University of Wisconsin-Madison.
Cain, Glen G. (1985) "Welfare economics of policies toward women", Journal of Labor Economics, 3,
part II:375-396.
Chiswick, Barry R. (1973) "Racial discrimination in the labor market: a test of alternative hypothe-
ses", Journal of Political Economy, 81:1330-1352.
Chiswick, Barry R. (1974) Income inequality. New York: National Bureau of Economic Research,
distributed by Columbia University Press.
Chiswick, Barry R. (1983) "The earnings and human capital of American Jews", Journal of Human
Resources, 18:313-336.
Cohen, Malcolm S. (1971) "Sex differences in compensation", Journal of Human Resources, 6:434-447.
Comanor, William S. (1973) "Racial discrimination in American industry", Economica, 40:363-378.
Conway, Delores A. and Harry V. Roberts (1983) "Reverse regression, fairness and employment
discrimination", Journal of Business and Economic Statistics, 1:75-85.
Corcoran, Mary and Greg J. Duncan (1979) "Work history, labor force attachment, and earnings
differences between the races and sexes", Journal of Human Resources, 14:497-520.
Duncan, Otis Dudley (1968) "Inheritance of poverty or inheritance of race?", in: D. P. Moynihan,
ed., On understanding poverty. New York: Basic Books, 85-110.
Featherman David L. and Robert M. Hauser (1976) "Sexual inequalities and socioeconomic
achievement m the U.S., 1962-1973 ', Amertcan Sociological Review, 41:462-483.
Finkelstein, Michael O. (1980) "The judicial reception of multiple regression studies in race and sex
discrimination cases", Columbia Law Review, 80:737-754.
Fisher, Franklin M. (1980) "Multiple regression in legal proceedings", Columbia Law Review,
80:702-736.
Flanagan, Robert J. (1973) "Racial wage discrimination and employment segregation", Journal of
Human Resources, 8:465-471.
Flanagam Robert J. (1974) "Labor force experience, job turnover, and racial wage differentials",
Review of Economics and Statistics, 56:521-529.
Freeman, Richard B. (1973a) "Decline of labor market discrimination and economic analysis",
American Economic Review, 63:280-286.
Freeman, Richard B. (1973b) "Comment", in: O. Ashenfelter and A. Rees, eds., Discrimination in
labor markets. Princeton, N.J.: Princeton University Press, 82-87.
Freeman, Richard B. (1977) Black elite: the new market for highly qualified black Americans. New
York: McGraw-Hill.
Freeman, Richard B. (1981) "Black progress after 1964: who has gained and why?", in: S. Rosen, ed.,
Studies in labor markets. Chicago: University of Chicago Press, 247-294.
Fuchs, Victor R. (1971) "Differences in hourly earnings between men and women", Monthly Labor
Review, 94:9-15.
Fuchs, Victor R. (1984) "His and hers: gender differences in work and income, 1959-1979", Working
Paper No. 1501, National Bureau of Economic Research, Cambridge, Mass. Forthcoming in
Journal of Labor Economics.
Goldberg, Matthew S. (1982) "Discrimination, nepotism, and long-run wage differentials", Quarterly
Journal of Economics, 96: 307- 319.
Goldberger, Arthur S. (1984) "Reverse regression and salary discrimination", Journal of Human
Gould, William B. (1977) Black workers in white unions:job discrimination in the United States. Ithaca,
N.Y.: Cornell University Press.
Greeley, Andrew M. (1976) Ethnicity, denomination and inequality. Beverly Hills, Calif.: Sage
Publications.
Greeley, Andrew M. (1981) The Irish Americans. New York: Harper and Row.
Green, Carol A. and Marianne A. Ferber (1984) "Employment discrimination: an empirical test of
forward versus reverse regression", Journal of ttuman Resources, 19:557-569.
Grenier, Gilles (1984) "The effect of language characteristics on the wage of Hispanic-American
males", Journal of Human Resources, 19:35-52.
Gwartney, James D. and Richard Stroup (1973) "Measurement of employment discrimination
according to sex", Southern Economic Journal, 39:575-587.
Hanoch, Gloria (1967) "An economic analysis of earnings and schooling", Journal of Human
Hill, Herbert (1977) Black labor and the American legal system. Washington, D.C.: The Bureau of
National Affairs.
Hoffman, Saul (1979)"Black-wtfite life cycle earnings differences and the vintage hypothesis: a
longitudinal analysis", American Economic Review, 69:855-867.
Johnson, George E. (1975) "Economic analysis of trade unionism", American Economic Review,
Papers and Proceedings, 65:23-28.
Johnson, George E. and Frank P. Stafford (1974) "The earnings and promotion of women faculty",
American Economic Review, 64:888-903.
Jones, F. L. (1983) "On decomposing the wage gap: a critical comment on Blinder's method",
Journal of Human Resources, 18:126-130.
Kessel, Reuben A. (1958) "Price discrimination in medicine", Journal of Law and Economics,
1:20-53.
Kohen, Andrew and Roger Roderick (1975) "The effect of race and sex discrimination on early career
earnings", unpublished paper, Center for Human Resource Research, Ohio State University,
Columbus, Ohio.
Landes, William (1968) "The economics of fair employment laws", Journal of Political Economy,
76:507-552.
Lewis, H. Gregg (1959) "Competitive and monopoly unionism", in: P. D. Bradley, ed., The public
stake in union power. Charlottesville, Va.: University of Virginia Press, 181-208.
Lucas, Robert E. B. (1974) "The distribution of job characteristics", Review of Economics and
Statistics, 56:530-540.
Luksetich, W. A. (1979) "Market power and sex discrimination in white-collar employment", Review
of Social Economy, 37:211-224.
Lundahl, Mats and Eskil Wadensj5 (1984) Unequal treatment: a study in the neo-classical theory of
discrimination. London: Croom Helm.
Lundberg, Shelly J. and Richard Startz (1983) "Private discrimination and social intervention in
competitive labor markets", American Economic Review, 73:340-347.
Madden, Janice F. (1973) The economics of sex discrimination. Lexington, Mass.: D.C. Heath and Co,
784 G. G. Cain
Malkiel, Burton G. and Judith A. Malkiel (1973) "Male-female pay differentials in professional
employment", American Economic Review, 63:693-705.
Mallan, Lucy B. (1982) "Labor force participation, work experience, and the pay gap between men
and women", Journal of Human Resources, 17:437-448.
Maloney, Timothy J. (1983) "The cyclical labor supply response of married women and disequi-
librium unemployment", unpublished Ph.D. dissertation, Department of Economics, University of
Wisconsin-Madison.
Marshall, F. Ray (1965) The Negro and organized labor. New York: John Wiley.
Marshall, F. Ray (1974) "The economics of racial discrimination: a survey", Journal of Economic
Literature, 12:849-871.
Masters, Stanley H. (1975) Black-white income differentials. New York: Academic Press.
McCall, John J. (1972) "The simple mathematics of information, job search, and prejudice", in: A.
Pascal, ed., Racial discrimination in economic life. Lexington, Mass.: Lexington Books, 205 224.
Mincer, Jacob and Solomon Polachek (1974) "Family investments in human capital earnings of
women", Journal of Political Economy, 82, part 2:$76-S108.
Myrdal, Gunnar (1944) An American dilemma. New York: Harpers.
Northrup, Herbert R. (1944) Organized labor and the Negro. New York: Harpers.
Oaxaca, Ronald L. (1973) "Male-female wage differentials in urban labor markets", International
Economic Review, 14:693-709.
O'Neill, June (1983a) "The determinants and wage effects of occupational segregation", working
paper, Urban Institute, Washington, D.C.
O'Neill, June (1983b) "The trend in sex differentials in wages", paper presented at the Conference on
Trends in Women's Work, Education, and Family Building, Urban Institute, Washington, D.C.
O'Neill, June (1984) "Earnings differentials: empirical evidence and causes", in: G. Schmid and R.
Weitzel, eds., Sex discrimination and equal opportunity. London: Gower Publishing Co., 69-91.
Oster, Sylvia M. (1975) "Industry differences in discrimination against women", Quarterly Journal of
Economics, 89:215-229.
Petcrsen, William (1978) "Chinese Americans and Japanese Americans", in: T. Sowell, ed., Essays
and data on American ethnic groups. Washington, D.C.: The Urban Institute, 65-106.
Phclps, Edmund S. (1972) "The statistical theory of racism and sexism", American Economic Review,
62:659-661.
Piore, Michael J. (1970) "Jobs and training", in: S. H. Beer, ed., The State and the Poor. Cambridge,
Mass.: Winthrop Press, 53-83.
Polachek, Solomon W. (1979) "Occupational segregation among women: theory, evidence, and
prognosis", in: C. B. Lloyd, E. S. Andrews and C. L. Gilroy, eds., Women in the Labor Market.
New York: Columbia University Press, 137-157.
Reich, Michael (1971) "The economics of racism", in: David M. Gordon, ed., Problems in economy:
an urban perspective. Lexington, Mass.: D.C. Heath and Co., 107-113.
Reich, Michael (1981) Racial inequality. Princeton, N.J.: Princeton University Press.
Reimers, Cordelia W. (1983) "Labor market discrimination against hispanic and black men", Review
of Economics and Statistics, 65:570-579.
Riley, John G. (1975) "Competitive signaling", Journal of Economic' Theory, 10:174-186.
Roberts, Harry V. (1980) "Statistical biases in the measurement of employment discrimination", in:
E. R. Livernash, eds., Comparable worth: issues and alternatives, Washington, D.C.: Equal Employ-
ment Advisory Council, 173-195.
Robinson, Joan (1934) The economics of imperfect competition. London: Macmillan.
Roos, Patricia A. (1981) "Sex stratification in the workplace: male-female differences in economic
returns to occupations", Social Science Researeh, 10:195-224.
Ross, Malcol~a (1948) All manner of men. New York: Reynal and Hitchcock.
Rothschild, MiChael and Joseph E. Stiglitz (1982) "A model of employment outcomes illustrating the
effect of the structure of information on the level and the distribution of income", Economics
Letters, 10:231-236.
Sanborn, Henry (1964) "Pay differences between men and women", Industrial and Labor Relations
Review, 17:534-550.
Sawhill, Isabel (1973) "The economics of discrimination against women: some new findings", Journal
of Human Resources, 8:383-396.
Smith, James P. and Michael P. Ward (1984) "Women's wages and work in the twentieth century",
Report R-3119-NICHD, RAND, Santa Monica, Calif.
Smith, James P. and Finis Welch (1977) "Black-white male wage ratios: 1960-1970", American
Smith, James P. and Finis Welch (1978) "Race differences in earnings: a survey and new evidence",
unpublished ms., R-2295-NSF, RAND, Santa Monica, Calif.
Sowell, Thomas (1981) Ethnic Americans. New York: Basic Books.
Spence, A. Michael (1973) "Job market signaling", Quarterly Journal of Economics, 87:355-374.
Stiglitz, Joseph E. (1973) "Approaches to the economics of discrimination", American Economic
Review, Papers and Proceedings, 63:287-295.
Stiglitz, Joseph E. (1974) "Theories of discrimination and economic policy", in: G. M. von
Furstenberg, A. R. Horowitz, and B. Harrison, eds., Patterns of racial discrimination. Volume B:
employment and income. Lexington, Mass.: D.C. Heath and Co., 5-26.
Suter, Larry E. and Herman P. Miller (1973) "Income differences between men and career women",
American Journal of Sociology, 78:962-974.
Thurow, Lester C. (1969) Poverty and discrimination. Washington, D.C.: The Brookings Institution.
Thurow, Lester C. (1975) Generating inequality. New York: Basic Books.
Toikka, Richard S. (1976) "The welfare implications of Becker's discrimination coefficient", Journal
of Economic Theory, 13:472-477.
Tomes, Nigel (1984) "The effects of religion and denomination on earnings and the returns to human
capital", Journal of Human Resources, 19:472-488.
Treiman, Donald J. and Heidi I. Hartmann, eds. (1981) Women, work, and wages: equal pay for jobs
of equal value. Washington, D.C.: National Academy Press.
Treiman, Donald J. and Kermit Terrell (1975) "Sex and the process of status attainment: a
comparison of working women and men", American Sociological Review, 40:174-200.
U.S. Bureau of the Census (1980) Statistical abstract of the United States, 1980. Washington, D.C.:
U.S. GPO, 101st ed.
U.S. Bureau of the Census (1983a) Current population reports, Series P-20, No. 372, Marital status
and living arrangements, March 1981. Washington, D.C.: U.S. GPO.
U.S. Bureau of the Census (1983b) "Current population reports", Series P-60, No. 137, Money income
of households, families, and persons in the United States: 1981. Washington, D.C.: U.S. GPO.
U.S. Bureau of the Census (1983c) Statistical abstract of the United States, 1984. Washington, D.C.:
U.S. GPO, 104th ed.
U.S. Bureau of the Census (1984) "Current population reports", Series P-60, No. 142, Money income
of households, families, and persons in the United States: 1982. Washington, D.C.: U.S. GPO.
U.S. Department of Labor, Bureau of Labor Statistics (1979) Earnings and other characteristics of
organized labor, May 1977, Report 556. Washington, D.C.: U.S. GPO.
U.S. Department of Labor, Bureau of Labor Statistics (1982) Labor force statistics derived from the
current population survey: a databook, Volume I, Bulletin 2096. Washington, D.C.: U.S. GPO.
Welch, Finis (1973) "'Education and racial discrimination", in: O. A. Ashenfelter and A. Rees, eds.,
Discrimination in labor markets. Princeton, N.J.: Princeton University Press, 43-81.
Welch, Finis (1975) "Human capital theory: education, discrimination, and life cycles", American
Zabalza, Anthony and J. L. Arrufat (1983) "Wage differences between married men and married
women in Great Britain: the depreciation effect of nonparticipation", unpublished Discussion
Paper No. 151, Centre for Labour Economics, London School of Economics.
Zeman, Mordechai (1955) "A quantitative analysis of white-nonwhhe income differentials in the
United States in 1939", unpublished Ph.D. dissertation, University of Chicago.

The Economic Analysis of Labor Market Discrimination

Uploaded by

Copyright:

Available Formats

The Economic Analysis of Labor Market Discrimination

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Economic Analysis of Labor Market Discrimination

Uploaded by

Copyright:

Available Formats

Chapter 13

THE ECONOMIC ANALYSIS OF LABOR MARKET

This survey of the economics of labor market discrimination is motivated by two

Handbook of Labor Economics, Volume L Edited by O. Ashenfelter and R. Layard

Empirical tests of the economic theories are selectively surveyed in Section 4.

2. The definition of economic discrimination

Economic discrimination is a concept that defies precise definition. One difficulty

person; X i = a vector of productivity characteristics of the ith person that are

where u is a random error and C > 0 is evidence of discrimination. In this case,

transactions. The economist's treatment of tastes is, however, circumscribed.

market, in the absence of discrimination, should already have provided the

subsequent interpretation of the coefficients and their properties all require a

2.2. Summary statistics for two concepts of economic discrimination

The first definition of economic discrimination, concerning differences in eco-

N u m b e r of units (in millions);

1. Households b $23.7 $14.9 0.63 $18.4 0.77 72.8 9.0 4.0

Families with primary earner

much a problem of poverty, at least as officially defined, as it is of inequality - their

Judgment as to whether accounting

Sources: U.S. Bureau of the Census, Current Population Reports, Series

eliminated. Among working women, minority women are more likely to be

Numbers of all workers

Source: Table 55 in source cited in Table 13.1.

Women/men Black/white Hispanic/white

1939b 0.61 0.51 - 0.45 0.38 - -

generally f r o m the low-income rural sector of the South to u r b a n places. E a r n i n g s

3. Theories of economic discrimination in the labor market

There is no shortage of theories to rationalize the existence of different wage rates

3.1. Nonstochastic competitive neoclassical models: Discrimination by consumers

Economic analysts have generally concluded that consumer-based discrimina-

3.1.1. Discrimination by workers

3.1.2. Discrimination by employers

Two versions of employer-based discrimination in competitive markets were

3.2. Nonstochastic monopolistic neoclassical models: Product monopoly

A monopoly has two characteristics that permit long-run discrimination: first, a

3.2.1. Monopsony firms in labor markets

The classic case of the exploitation of labor in neoclassical economics arises

3.2.2. Labor unions as monopolies

In Becker's model of discrimination, white workers' prejudice against black

3.2.3. Government as a monopolist

Governments are universally monopolists in certain functions, such as providing

importantly, as law-maker. With their power to tax and to punish, governments

excess unemployment or because those disemployed from covered jobs are

3.3. Stochastic neoclassical models and statistical discrimination

The theoretical challenge developed in the preceding discussion of neoclassical

with E ( u ) = C(q, u) = O, E ( y ) = E ( q ) = a, V(u) = a 2, using the familiar sym-

q=a(1- y)+ yy+e, (2)

Accordingly, for a given y-score [roughly corresponding to "holding X constant"

3.3.1. Statistical discrimination, but spurious group economic discrimination

(a) 71 = Y0; el > a o

wo = %(1 - )1) + ')'Yo

(b)[ 71 > Yo; 0'1 > °to

/ . / ~ ~ w o = %(1 - 70) + "YoYo

women, he is less likely to be able to recoup his investment . . . . The woman

3.3.2. Statistical discrimination and actual group discrimination

tionalized a lower average wage payment to minority workers as compensation

Although I do not find the empirical counterparts to the models of statistical

3.4. Institutional theories of discrimination

In his survey of the economics of racial discrimination, Marshall (1974) advo-

augment the supply of labor by encouraging more investment in human capital