Wage Inequality and the Rise in Returns to Skill

Author(s): Chinhui Juhn, Kevin M. Murphy, Brooks Pierce

Source: The Journal of Political Economy, Vol. 101, No. 3 (Jun., 1993), pp. 410-442
Published by: The University of Chicago Press
Stable URL: http://www.jstor.org/stable/2138770
Accessed: 06/10/2010 11:49

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=ucpress.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

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

The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access to The
Journal of Political Economy.

http://www.jstor.org
Wage Inequality and the Rise
in Returns to Skill

ChinhuiJuhn
Universityof Houston

KevinM. Murphy
Universityof Chicago

Brooks Pierce
TexasA&M University

Using data from the March Current Population Survey, we docu-

ment an increase over the past 30 years in wage inequality for males.
Between 1963 and 1989, real average weekly wages for the least
skilled workers (as measured by the tenth percentile of the wage
distribution) declined by about 5 percent, whereas wages for the
most skilled workers (as measured by the ninetieth percentile of the
wage distribution) rose by about 40 percent. We find that the trend
toward increased wage inequality is apparent within narrowly de-
fined education and labor market experience groups. Our interpre-
tation is that much of the increase in wage inequality for males over
the last 20 years is due to increased returns to the components of
skill other than years of schooling and years of labor market experi-
ence. Our primary explanation for the general rise in returns to skill
is that the demand for skill rose in the United States over this period.

I. Introduction
Between 1963 and 1989, the average weekly wage of working men
increased by about 20 percent. However, as we show in this paper,

This research was supported by grants from the National Science Foundation and
the Sloan Foundation.
[Journal of Political Economy, 1993, vol. 101, no. 3]
) 1993 by The University of Chicago. All rights reserved. 0022-3808/93/0103-0001$01.50

410
WAGE INEQUALITY 411
these real wage gains were not spread equally across workers. Wages
for the least skilled, as measured by the tenth percentile of the wage
distribution, fell by about 5 percent, and wages for the most skilled,
as measured by the ninetieth percentile of the wage distribution, in-
creased by about 40 percent. Comparisons of 1989 with 1970 reveal
similar differences. While real wages for the median worker were 5
percent lower in 1989 than in 1970, wages for workers at the ninetieth
percentile were more than 15 percent higher in 1989 than in 1970.
This 15 percent increase in real wages contrasts even more sharply
with a 25 percent decline in real wages for workers at the tenth per-
centile of the wage distribution. Looking within education and experi-
ence categories reveals even more striking changes. Real wages for
tenth percentile high school graduates with 1-10 years of experience
are about 15 percent lower today than wages for the same group in
1963. In fact almost 40 percent of today's workers with 1-10 years
of experience earn less than workers at the corresponding percentile
in 1963.
The net result of this divergence in earnings between the most
skilled and the least skilled has been an enormous increase in wage
inequality. According to our calculations, the variance of log weekly
wages increased by about 72 percent from 1963 through 1989. These
findings are broadly consistent with the findings of others who have
recently looked at inequality (see, e.g., Bluestone and Harrison
[1988], Bluestone [1989], Levy [1989], or many of the studies cited
in Sawhill [1988]). While the basic facts support the general trends
identified in previous work, we feel that there are many unique com-
ponents of our analysis as well. First, much of the inequality literature
has focused on earnings or income as a measure of welfare rather
than on wages, which are more closely related to market prices for
human capital components. We believe that the emphasis on wages
as prices, instead of on incomes, allows us to make several important
contributions to this literature. In addition, we push our analysis in
a somewhat different direction and describe more clearly the timing
of the increase in overall inequality, decompose the increase in in-
equality into components accounted for by observable differences
across workers (e.g., age and education) and inequality within these
classifications, and finally evaluate some simple alternative explana-
tions. Each of these points adds significantly to our understanding of
the recent history of wage inequality.
Throughout the paper we interpret the dispersion in wages, after
controlling for observable skill determinants, as a distribution of un-
observable ability in the population in conjunction with a current
market value of this unobservable ability. In particular, we view the
trend toward increased inequality not in terms of increased disper-
412 JOURNAL OF POLITICAL ECONOMY

sion in unobserved ability (or increased measurement error etc.) but

rather as an increasing market return to skill. This characterization
seems the most complete and compact available for the empirical
facts we present. In addition, this characterization points to changes
in skill demand as a likely explanation of the facts, an explanation
that has some support in our data.
Our basic empirical findings are that wage inequality remained very
stable or even declined slightly from 1960 to 1968 or 1969 and then
increased relatively steadily through the end of the data. When we
decompose this increase in inequality into a rise in inequality across
observable dimensions of skill (i.e., experience and education) and a
rise in inequality within schooling and experience groups, we find a
substantial difference in the timing of the increase. For example,
wage differentials by education actually increased from 1960 through
1970 (while overall inequality remained relatively stable) and then
fell significantly over the 1970s (while overall inequality was rising).
In contrast, the recent sharp rise in the returns to education (identi-
fied by Murphy and Welch [1992] and others) has been much greater
than the recent rise in overall inequality. Our conclusion is that the
general rise in inequality and the rise in education premia are actually
distinct economic phenomena. In contrast to the education story, in-
equality within education and experience categories remained stable
or fell over the 1960s and then rose steadily through the end of our
data. The differences in the timing of the increases in wage inequality
within and between groups represent one of the most important
findings of this paper and point to a rise in the demand for skill that
predates the recent rise in returns to education by about a decade.
In spite of these differences in timing, we find that, on average, wage
differentials both between and within groups have increased by be-
tween 30 and 50 percent since the late 1960s.
We consider two basic types of explanations for the general rise in
returns to skill. The first explanation links the rise in returns to skill
with a shift in labor demand both across and within industries that
favors the most skilled. Using observed changes in the industry and
occupation mix of the labor force as measures of the changing de-
mand for skill, we find that employment has shifted toward industries
and occupations that demand more skilled workers even in the face
of rising skill premia. Overall, we find that the "demand" for the most
skilled increased by about 50 percent relative to the demand for the
least skilled. We do not as yet know what underlying forces (such as
international competition) have led to these dramatic shifts in labor
demand, but whatever the source, these changes would seem to be a
major factor accounting for the rise in skill premia.
We also attempt to determine the extent to which changes in the
WAGE INEQUALITY 413
composition of the work force (i.e., changes in age and education
composition) and changes in the patterns of employment across occu-
pations and industries have affected the level of wage inequality. On
both margins we find that these composition effects are relatively
unimportant. In particular, the shift of the economy to the service
sector and changes in the occupation distribution have had only very
minor influences on wage inequality (accounting for less than 15 per-
cent of the increase).
It is important to stress the relevance of this conclusion to current
research topics in labor economics. In particular, studies that make
wage comparisons over time for groups that are relatively concen-
trated in the upper or lower tail of the wage distribution must con-
sider the importance of rising skill prices. For example, the well-
documented convergence of black and white wage rates over the
1940-80 period (Smith and Welch 1986) has slowed in the 1980s
(Juhn, Murphy, and Pierce 1991). A substantial part of this slowdown
in convergence is explainable by rising skill prices in combination
with the fact that blacks are relatively concentrated in the lower half
of the skill distribution (because, say, premarket discrimination has
been manifested in lower quantity and quality of schooling for
blacks). Similarly, some small part of the findings that entering co-
horts of immigrants have become less skilled may be due to the fact
that skill prices have risen (as opposed to skill quantities falling) over
time (see, e.g., Borjas 1985; Chiswick 1986; LaLonde and Topel
1991).
The paper is organized as follows. Section II describes the data we
use in our analysis. Section III presents our results on wage inequality
and the rise in returns to skill. Section IV decomposes the rise in
inequality into components associated with observed differences (ed-
ucation and experience) and the residual or unobserved component
and identifies the time patterns of returns to alternative measures of
skill. Section V attempts to explain the overall rise in the returns to
skill in terms of demand changes and composition effects. Section VI
relates our findings on wage inequality to changes in earnings in-
equality and points out several key differences between the series.
Section VII presents concluding remarks.

II. The Data

The results in this paper are based on wage data for men drawn from
27 years of the March Current Population Survey (CPS), survey years
1964-90, and the 1960 decennial census. The data from the CPS
come from the March annual demographic supplement and refer to
earnings and weeks worked in the calendar year preceding the March
414 JOURNAL OF POLITICAL ECONOMY

survey. The data from the 1960 census refer to earnings and weeks
worked in 1959. As a result, our sample measures wages for 1959
and 1963-89. Throughout the paper we focus on log weekly wages
or log hourly wages for full-time workers (defined as those that usu-
ally work 35 hours or more per week). We deflate annual earnings
by the personal consumption expenditure deflator from the National
Income and Product Accounts and define the log average weekly
wage as the natural logarithm of deflated annual wage and salary
earnings divided by weeks worked and the log hourly wage as the
natural logarithm of deflated annual wage and salary earnings di-
vided by the product of weeks worked and usual weekly hours. We
present initial analyses for both weekly and hourly wages and focus
on weekly wages for the remainder of the results.
For purposes of analysis we selected a sample that we felt would
be representative of workers with a reasonably strong labor force
attachment. Our sample inclusion criteria were that the workers be
aged 18-65, work full time, not be self-employed or working without
pay, not live in group quarters, work at least 14 weeks, have a positive
number of years of potential labor market experience, not work part
of a year because of retirement or school, and earn a minimum of
$67 per week in 1982 dollars (equal to one-half of the 1982 real
minimum based on a 40-hour week). We imputed weekly earnings
for workers top-coded at the census maximum as 1.33 times the top-
coded value. For the most part, the qualitative results in the paper
are not sensitive to the exclusion and imputation procedures we used.
The one major exception is the exclusion of workers earning less
than half of the real 1982 minimum wage. When these workers are
included, wage inequality actually declines somewhat more over the
1963-69 period. For example, when these workers are included, the
log weekly wage differential between the ninetieth and tenth percen-
tiles falls by three points from 1963 to 1969, whereas in our reported
results it actually rises slightly. This distinction, however, does arise
for other time periods. The question of whether one believes that
wage differentials fell or remained stable over the late 1960s turns
on the credibility one lends to these relatively extreme observations.
In previous versions of this paper we noted that because of changes
in the survey structure between the 1975 and 1976 surveys (wage
data for 1974 and 1975), measured inequality dropped significantly.
We have subsequently been able to overcome most of this problem
by improving our predictions of weeks and hours worked last year,
deleting all individuals with imputed wage and salary information,
and excluding self-employed workers on the basis of the existence of
any significant self-employment income (i.e., those with nega-
tive self-employment income or more than $100 [1982] of self-
WAGE INEQUALITY 415

employment income). While it appears that some spurious fall in

inequality still occurs between 1974 and 1975, particularly in the low-
est and highest deciles, we do not use any correction factors in this
version. In fact, the results reported here are very close to our previ-
ous corrected numbers and not the previous uncorrected numbers,
which gives us considerable confidence that our prior corrections
were in fact moving our answers in the right direction.

III. The Rise in Inequality

Figure 1 graphs the median, tenth percentile, and ninetieth percen-
tile of the real weekly wage distribution of men for 1963-89. For
ease of comparison, wages for the three groups are indexed to an
average of 100 in 1963 and 1964 for all three series. The median
wage series tells a well-known story. Real wages increase relatively
steadily from 1963 through about 1973, decline sharply from 1973
through 1975, rise slightly from 1975 through 1977, decline from
1977 through 1982, and then recover somewhat from 1983 through
1985, only to fall back again from 1985 to 1989. The basic story is
that real wages were about 25 percent higher in 1973 than in 1963
and 1964 and about 5 percent lower in 1989 than in 1973. As is clear
from the figure, the story is significantly different for the tenth and
ninetieth percentiles. For the least skilled (proxied here by the tenth

iOth Percentile (o) Median( ) 90th Percentile (+)

140.00 -

g130.00-
3U

.00
M 1200
V
0)
D 110.00
'-4

100.00

[ I III 1II _I IIIIII

65 70 75 80 85
YEAR
FIG. 1.-Indexed real weekly wages by percentile, 1963-89
416 JOURNAL OF POLITICAL ECONOMY

percentile), wages rose by about 20 percent from 1963 through 1970

and then declined by 25 percent from 1970 through 1989. In con-
trast, real wages for workers at the ninetieth percentile rose steadily
from 1963 through 1973 (by about 31 percent), declined by about 10
percent over the next 2 years, and then rose about 20 percent from
1975 through 1989. After about two and one-half decades, workers
in the top 10 percent of the wage distribution have gained almost 40
percent, whereas workers in the bottom 10 percent have lost over 5
percent in real terms.
The data before and after 1970 provide a stark contrast. Between
1963 and 1970, wages for both the ninetieth and tenth percentiles
increased by about 20 percent. Since 1970, workers at the tenth per-
centile have lost about 25 percent in real terms, whereas workers at
the ninetieth percentile have gained about 15 percent. Figure 2 tells
a similar story for hourly wages. The only significant differences are
that real hourly wages grow more for both low-wage (tenth percen-
tile) and high-wage (ninetieth percentile) workers than for the me-
dian worker and that median hourly wages fall more rapidly than
median weekly wages over the late 1980s. Figure 3 shows that this
rapid divergence in wages is not limited to comparisons of the most
and least skilled. The figure gives the percentage change in real
weekly wages by percentile over the 1963-89 period. As the figure
illustrates, the change in log real wages from 1964 to 1988 is basically

1Oth Percentile (0) Median ( ) 90th Percentile (+)

140.00

130.00 -

o 120.00-
cc

v110.00

100.00

10.0 I f I I I I l I Ii I I I I I I I I I I I I I I I
65 70 75 80 85
YEAR
FIG.2.-Indexed real hourly wages by percentile, 1963-89
WAGE INEQUALITY 417

0.40

0
0I~~~~~~~~~~~~~~~~~~~~~~~~~
to
X
W
0.20

.' 0.10-

Co

D 0.00

-0.10

0 10 20 30 40 50 60 70 80 90 100
PERCENTILE
FIG. 3.-Log real wage changes by percentile, 1964-88

a linear function of the percentile, with workers at the upper end

gaining about 40 percent and workers in the lowest percentiles losing
about 5 percent. As the figure shows, the divergence in wages across
percentiles is pervasive and is not limited to a specific part of the
wage distribution. Skill differentials have increased at all points in
the skill distribution.
The four panels in figure 4 decompose the change in weekly wage
inequality from 1963 through 1989 into four subperiods. In order to
minimize the effects of measurement error, we look at changes be-
tween 3-year averages. Panel A compares wages by percentile for
the period 1969-71 with wages for 1963-65. The wage changes are
normalized by comparing the change at each percentile with the
mean change in log wages over that time interval. As panel A shows,
the increase in inequality between 1963-65 and 1969-71 was quite
modest overall. Over the period, workers at the lowest percentiles
(below the tenth) actually gain slightly on the mean and median
worker (thus reducing wage inequality), and workers at the upper
percentiles (above the eightieth) also gain on the median worker (thus
raising inequality). The basic message of panel A is that inequality
between the most skilled and the average worker increased slightly
during the late 1960s but inequality between the median worker and
the lowest percentiles may have actually declined slightly. As we men-
tioned in Section II, this convergence may be even somewhat larger
depending on how one treats low-wage observations.
 0
coI

00

to~~~~~~~~~~~~~~~~~~~~~~~~b

I~~

U, ~~~ , 0 0 0
W
'4~~~~~~~~~-
ww , , '4
w
0 w 0 '

a o In o f o o~~~~~n
o - o In l~~~~~~~~~~~~~
I

Ia a in 0 0 i0 0 0I
0

0 0 0 00 0 V
1 a, V

861M OT4916 501 U ONRU 609M AT41M 001 UT~5UlV

WAGE INEQUALITY 419
The next three panels show much clearer moves toward greater
wage inequality. In the six years from 1969-71 to 1975-77, workers
at or below the tenth percentile of the wage distribution lost about 7
percent relative to the average (9 percent relative to the median), and
workers in the upper quartile gained about 3-4 percent on the aver-
age worker. The changes from 1975-77 to 1981-83 are slightly
larger, particularly at the extreme upper percentiles, but the overall
nature of change remains the same. Workers at the upper percentiles
gained significantly (about 7 percent) relative to the average, and
workers at the lowest percentiles lost about 6-7 percent. The change
over the most recent period (from 1981-83 to 1987-89) is about the
same. Workers at the lowest percentiles lose about 7 percent relative
to the mean, and workers at the highest percentiles gain about 7
percent. Again, as in the change for 1963-89 as a whole, the expan-
sion in wage differentials by percentiles is pervasive. The percentage
increase in wages is roughly a linear function of the percentile, with
wage increases being 1.4 percent higher for each 10 percentile points
up in the wage distribution.
Table 1 quantifies these changes by giving inequality measures for
1959 (i.e., from the 1960 census) and the same 3-year intervals used
to summarize changes in figure 4. Panel A presents results for weekly
wages, and panel B presents results for hourly wages. From 1959 to
1988 (the mean year of the final interval), the standard deviation of
weekly wages increases from .44 to .59, an increase of about 33 per-
cent. Similarly, the log wage differential between the ninetieth and
tenth percentiles increases from 1.05 to 1.46, an increase of 42 per-
cent. Over the full period the increase in inequality has been slightly
larger below the mean than above the mean. The differential between
the ninetieth percentile and the median increased by .17, and the
difference between the median and the tenth percentile increased by
.23. Similarly, the seventy-fifth percentile-median differential in-
creased by .10, and the median-twenty-fifth percentile differential
increased by .15. In spite of the differences, the basic message is
that inequality has increased substantially in all parts of the wage
distribution. The general pattern of accelerating change shown in
figure 4 comes through in the comparisons in table 1 as well.
The results presented so far refer only to changes in the overall
wage distribution and do not tell us how these changes break down
into changes within groups (defined by education and experience)
and changes between groups. They also do not tell us whether the
changes have been greater for some subgroups than for others. Fig-
ure 5 addresses both of these issues by looking at log wage changes
by percentile separately for workers with 1-10 years of experience
and workers with 21-30 years of experience. The percentile numbers
420 JOURNAL OF POLITICAL ECONOMY

TABLE 1
INEQUALITY MEASURES FOR LOG WAGES OF MEN, 1959-88

1959 1964 1970 1976 1982 1988

A. Weekly Wages

Standard deviation .44 .45 .46 .49 .54 .59

Percentile differential:
90-10 1.05 1.09 1.12 1.21 1.34 1.46
(.0029) (.0066) (.0050) (.0054) (.0053) (.0057)
75-25 .50 .53 .54 .61 .69 .75
(.0015) (.0033) (.0029) (.0031) (.0033) (.0036)
90-50 .49 .51 .54 .55 .61 .66
(.0021) (.0042) (.0035) (.0036) (.0039) (.0042)
50-10 .57 .57 .58 .66 .74 .80
(.0021) (.0053) (.0037) (.0042) (.0041) (.0045)
75-50 .24 .26 .26 .29 .31 .34
(.0010) (.0026) (.0021) (.0022) (.0024) (.0027)
50-25 .26 .26 .28 .32 .37 .41
(.0013) (.0025) (.0021) (.0025) (.0028) (.0031)
Observations 212,127 42,780 54,369 54,760 59,922 66,669
B. Hourly Wages

Standard deviation .45 .46 .46 .49 .53 .57

Percentile differential:
90-10 1.05 1.11 1.10 1.21 1.32 1.43
(.0027) (.0063) (.0050) (.0052) (.0051) (.0052)
75-25 .51 .53 .53 .62 .70 .75
(.0015) (.0036) (.0029) (.0032) (.0034) (.0033)
90-50 .49 .51 .53 .54 .60 .66
(.0018) (.0039) (.0034) (.0035) (.0036) (.0038)
50-10 .56 .60 .57 .67 .72 .77
(.0020) (.0050) (.0038) (.0041) (.0040) (.0041)
75-50 .25 .26 .26 .29 .32 .35
(.0010) (.0024) (.0020) (.0022) (.0024) (.0024)
50-25 .26 .28 .27 .33 .38 .40
(.0012) (.0028) (.0022) (.0026) (.0028) (.0028)
Observations 212,127 42,780 54,369 54,760 59,922 66,669

NOTE.-Standard errors are in parentheses. Data for 1964-88 are 3-year averages of surrounding years from
1964-90 March Current Population Surveys. Data for 1959 are taken from the 1960 Public Use Micro Census
Tapes.

on the bottom refer to percentiles of the individual groups' wage

distribution. As the figure illustrates, wage differentials across experi-
ence groups have increased at all percentiles (with a slightly larger
divergence at the mean than at the extremes). Overall, the wage dif-
ferential between the groups increased by about .20. Given the ex-
isting positive wage differential between those with 21-30 years of
experience and those with 1-10 years of experience, this change has
served to increase overall wage inequality. Inequality has also gone
up enormously within each group. For the youngest group we find
WAGE INEQUALITY 42 1

Experience i-10 (a) Experience 21-30 (+

0.40 -

0.30 -

,?0.20 -

a 0.10

0.00

-0.10
10 20 30 40 50 60 70 80 90
PERCENTILE
FIG. 5.-Estimated wage change by percentile, 1964-88

that workers at the tenth percentile have lost about 7 percent in real
terms, whereas workers at the ninetieth percentile have gained almost
25 percent. Among the older group, workers at the tenth percentile
gain about 3 percent (about the same as the median worker in the
younger group) and workers at the ninetieth percentile gain about
39 percent. On the basis of these calculations, it seems that the in-
crease in inequality is about the same for both experience groups (in
fact it is also quite similar for those with 11-20 years of experience
and only slightly smaller for those with more than 30 years of expe-
rience).
One message from figure 5 that will continue throughout the paper
is that the increase in wage inequality shows up both between and
within groups. As a result, while we can say that older workers have
gained relative to younger workers as a whole, it is still true that the
highest-paid among the youngest workers (say the ninetieth percen-
tile) have gained relative to the lowest-paid (or even the twenty-fifth
percentile) older worker.
One of the most striking things about figure 5 is that for the lowest
40 percent of younger workers, real wages are lower in 1988 than
for the corresponding group in 1964! Hence for two-fifths of all
younger workers there has been no increase in economic opportunity
as measured by weekly wage rates in about two and one-half decades.
Figure 6 breaks things down still further by looking at real wage
422 JOURNAL OF POLITICAL ECONOMY

Experience 1-10: High School Graduates (o) College Graduates C+

0.30 -

0.20 -

a)
C"

c 0.10
a)

3
0.00
0 _
-J

-0.10

-0.20
10 20 30 40 50 60 70 80 90
PERCENTILE
FIG. 6.-Estimated wage change by percentile, 1964-88

changes for high school and college graduates for workers with 1-10
years of experience. As with the experience contrasts shown in figure
5, the between-group differential moved in the direction of greater
inequality, with college graduates gaining on high school graduates
at all percentile levels. The gains for college graduates relative to high
school graduates are very similar at all percentiles above the twenti-
eth, with college graduates gaining about 20 percent relative to high
school graduates.
The increases in inequality within a group are equally striking.
High school graduates at the ninetieth percentile gained about 9 per-
cent in real terms from 1964 to 1988, whereas high school graduates
at the tenth percentile lost about 15 percent in real terms. The data
in figure 6 imply that the bottom 40 percent of high school graduates
with 1-10 years of experience earn 10- 17 percent less than the corre-
sponding workers in 1964. In fact only the top 30 percent of young
high school graduates have gained in real terms since 1964. The
relative wage changes for college graduates show a similar increase
in inequality, with the ninetieth percentile college graduates gaining
about 25 percent and the tenth percentile college graduates losing
slightly less than 5 percent. As in the experience calculation described
above, this large increase in within-group inequality implies that while
college graduates gained on high school graduates as a whole, the
WAGE INEQUALITY 423
best high school graduates (say the ninetieth percentile) gained sig-
nificantly on the low-end college graduates (say the tenth percentile).
The basic message of figures 5 and 6 is that wage inequality has
increased significantly within groups defined by experience and edu-
cation. Table 2 takes this analysis one step further by looking at the
distribution of regression residuals from a regression of log weekly
wages on a very flexible specification of education and experience
effects.' Looking at regression residuals allows us to look within very
narrowly defined education and experience categories. A striking fea-
ture of the table is the similarity of the inequality measures for 1959
and 1970. Apparently there was very little change in within-group
inequality over the 11 years from 1959 to 1970. In contrast, the pe-
riod from 1970 to 1988 is characterized by an enormous increase in
inequality, with workers at the ninetieth percentile of the residual
distribution gaining about 26 percent relative to workers at the tenth
percentile.
To understand the magnitude of the changes we describe, consider
the following frame of reference. In 1964 the standard deviation of
log weekly wages was .45, from table 1. In 1988 the standard devia-
tion based on regression residuals was approximately .49, from table
2. This means that wage inequality as measured by standard devia-
tions rose by an amount greater than the predictive power of the
observables in our wage equation. The magnitude of the inequality
increase is greater than the wage variation explainable by experience
and education combined.2
As stated in the Introduction, we view this increase in within-group
wage inequality as a trend toward higher skill prices. However, the
argument could be made that it is the result of increased dispersion
in unobserved ability within recent entry cohorts due to, say, increas-
ingly unequal educational opportunities. To evaluate this potential
objection, table 3 documents wage inequality growth over time within
synthetic cohort groups. Panel A gives the ninetieth-tenth percentile
differential for log weekly wages of various 6-year entry cohorts. One
follows a cohort over time by moving horizontally across columns
within the same row. One follows the same experience group over
time by moving upward along a diagonal (those who enter in 1959-64
have the same experience level in 1964 as entrants in 1965-70 do in
1970). Within cohorts, inequality changes over time are attributable

1 We estimated a wage equation with education dummies for less than high school,
high school, some college, and college graduates and with linear terms in education
within these groups. The regressions also include a quartic in experience fully inter-
acted with the education variables and regional dummies.
2 Typically, education and experience observables can explain about a quarter to a
third of the observed log weekly wage variation in a cross-sectional regression.
424 JOURNAL OF POLITICAL ECONOMY

TABLE 2
INEQUALITY MEASURES BASED ON REGRESSION RESIDUALS FOR MEN, 1959-88

1959 1964 1970 1976 1982 1988

Standard deviation .38 .39 .39 .41 .45 .49

Percentile differential:
90-10 .89 .94 .92 .99 1.10 1.18
(.0022) (.0053) (.0042) (.0044) (.0046) (.0048)
75-25 .43 .46 .46 .50 .57 .60
(.0012) (.0030) (.0024) (.0026) (.0028) (.0028)
90-50 .42 .44 .44 .46 .51 .54
(.0014) (.0035) (.0029) (.0029) (.0031) (.0033)
50-10 .47 .50 .48 .53 .59 .64
(.0017) (.0042) (.0033) (.0036) (.0037) (.0038)
75-50 .22 .22 .22 .24 .28 .28
(.0008) (.0021) (.0017) (.0018) (.0020) (.0020)
50-25 .22 .24 .23 .26 .30 .32
(.0009) (.0023) (.0019) (.0020) (.0022) (.0023)
Observations 212,127 42,780 54,369 54,760 59,922 66,669

NOTE.-Data for 1964-88 are 3-year averages of surrounding years from 1964-90 March Current Population
Surveys. Data for 1959 are taken from the 1960 Public Use Micro Census Tapes.

to age or time effects. In contrast, changes in inequality within an

experience group are due to cohort or time effects. As in tables 1
and 2, data are 3-year averages for surrounding years.
Obviously changes within recent cohorts cannot adequately explain
the trends toward greater inequality because past cohorts have also
experienced this trend in recent years. But the striking feature of
table 3 is that changes over time within cohorts (along rows) and
within experience groups (along diagonals) show remarkably similar
patterns. For example, the ninetieth-tenth percentile wage differen-
tial within the 1959-64 entry cohort increased from 1.13 in 1964 to
1.40 in 1988, and the differential for the 1-6 years of experience
group increased from 1.13 to 1.38. Further, the timing of the in-
creases is quite similar across the two series.
We summarize the growth of inequality within experience and co-
hort groups by averaging 6-year changes across the six cohort groups
for which comparisons can be made. The average change within a
cohort has the same magnitude and time pattern as that within an
experience group. This basically means that the older cohorts leaving
the data and the new cohorts entering the data across a 6-year span
have similar levels of inequality. We obtain the same results using
wage residuals in panel B of table 3. Given the fact that inequality
increases are age-neutral in the cross section (fig. 5), we surmise that
the similarity of entering and exiting cohorts implies that trends to-
ward inequality are due mainly to increasing skill prices over time,
and not to increasing dispersion of quality in more recent cohorts.
WAGE INEQUALITY 425

TABLE 3

CHANGES IN INEQUALITY BY COHORT, 1963-89

A. 90-10 DIFFERENTIALS FOR LOG WEEKLY WAGES

Year of Market Entry 1964 1970 1976 1982 1988

1983-88 1.38
1977-82 1.27 1.38
1971-76 1.13 1.24 1.38
1965-70 1.08 1.12 1.29 1.42
1959-64 1.13 1.01 1.13 1.30 1.40
1953-58 1.02 1.07 1.16 1.32 1.43
1947-52 1.02 1.11 1.15 1.30
1941-46 1.02 1.07 1.16
1935-40 1.06 1.09
1929-34 1.09
Average Changes within Cohorts and Experience Levels

Average Change 1964-70 1970-76 1976-82 1982-88

Within cohorts .018 .073 .153 .115

Within experience levels .015 .069 .145 .110

B. 90-10 DIFFERENTIALS FOR LOG WAGE RESIDUALS

1964 1970 1976 1982 1988

1983-88 1.09
1977-82 1.06 1.16
1971-76 .96 1.09 1.18
1965-70 .86 .96 1.12 1.23
1959-64 .92 .86 .98 1.12 1.21
1953-58 .88 .91 .99 1.15 1.26
1947-52 .89 .94 .99 1.14
1941-46 .94 .94 1.05
1935-40 .95 .98
1929-34 .99
Average Changes within Cohorts and Experience Levels

Average Change 1964-70 1970-76 1976-82 1982-88

Within cohorts .007 .096 .145 .101

Within experience levels - .016 .076 .123 .076

Of course, it is possible that cohort and age effects are equal and
have such a magnitude as to appear to be time effects. This follows
from the usual identification problem arising when one tries to sepa-
rate cohort, age, and time effects. While we cannot calculate growth
in inequality over time separately from inequality growth across co-
horts and ages, we can identify the change in inequality growth over
time. We can difference inequality measures within a cohort (elimi-
nating the cohort effect) and compare this difference across adjacent
426 JOURNAL OF POLITICAL ECONOMY

cohorts (eliminating the age effect and leaving only a change in in-
equality growth over time). From table 3 we can see that this is identi-
cal to comparing the average changes within cohorts over time. Since
the aver-age change within a cohort from 1964 to 1970 is much
smaller than the changes across later 6-year periods, we know that
the data indicate an accelerating increase in inequality with time that
cannot be explained by any combinations of cohort and age effects.
Given these data, we interpret increasing inequality over time as
arising mainly from increasing skill prices, and not from changes in
the distributions of unobservable skill quantities. In the following
section we present a framework that attempts to quantify the contri-
bution of changing observable quantities, observable prices, and
prices of unobservables to the overall increase in inequality.

IV. Components of Change in Wage Inequality

Panel A in figure 7 plots the time series of overall wage inequality as
measured by the log wage differential between the ninetieth and
tenth percentiles of the wage distribution. As the figure illustrates,
overall inequality was quite stable from 1963 to 1968. After 1968,
wage inequality increased relatively steadily through the end of our
data. Our basic interpretation of this change is that the wage pre-
mium for skill was stable during the 1960s and the skill premium has
increased by about 35 percent since the late 1960s. As we have shown,
this rise in skill premia applies to both observable dimensions of skill
(i.e., education, experience, and occupation) and unobservable di-
mensions of skill (the residual). A useful framework for isolating
these effects is to write a simple wage equation such as
Yit=Xitrt + t (1)
where Yit is the log weekly wage for individual i in year t, Xit is a
vector of individual characteristics (including experience and educa-
tion effects), and uit is the component of wages accounted for by the
unobservables. For our purposes it will be useful to think of this
residual as two components: an individual's percentile in the residual
distribution, 0jt' and the distribution function of the wage equation
residuals, Ft( ). By definition of the cumulative distribution function,
we have
Uit = Ft7'(OtIXjt), (2)
where F7-' QX-t) is the inverse cumulative residual distribution for
workers with characteristics Xit in year t.
In this framework changes in inequality come from three sources:
changes in the distribution of individual characteristics (i.e., changes
 0 ~~~~~~~~~~0

Cqc~~~~~~~~~~~~C
=~~~~~~~~~~~

.1

* S~~~~~~

q 0 .4
a 0a 0 0 0 a 0 0aa 0 0 0
I I I I I I I I

rT;iuojo.ina o.ki DBa, CI0 a; uoa;;lqj;.Juo3 TvluWmdu;TO *Bmg bol 0;-O* D; ua;~nqj;.iuaj

bC

4V-

I.

0~~~~~~~~

.4
* ~ ~ ~ ~ ~4~~~
.4
0 ~~0
.4 .4 .4
0
0
.4
#a
.4
0
cm
.
0
"a
0
0
I
a0
0
00
q
00
"a
00 0
Pd
.4
U
.

4 0~ I

TomaisJUI 'Ben oi5o ;-oG TSIWUJOMlOclii Dol o;-Oe 9a; uoy;no;.j;uo3

428 JOURNAL OF POLITICAL ECONOMY

in the distribution of the X's), changes in the prices of observable

skills (i.e., changes in the it's), and changes in the distribution of the
residuals. If we define S to be the average prices for observables over
the whole period and FQ(IXit) to be the average cumulative distribu-
tion, we can decompose the level of inequality into corresponding
components as

t = XiJ3 + Xit(t-A) + F -(Oi X)it)

? [F71(0it|Xit) F-1(0it|Xit)].
The first term captures the effect of a changing education and experi-
ence distribution at fixed prices. The second term captures the effects
of changing skill prices for observables at fixed X's, and the final term
captures the effects of changes in the distribution of wage residuals.
Armed with this simple framework, we can reconstruct what the wage
distribution would look like with any subset of components held fixed.
For example, with fixed observable prices and a fixed residual distri-
bution, wages would be determined as

it= XitJ + F(O0itiXit). (4)

In practice, we can estimate how this distribution would have changed

through time by predicting wages for all workers in the sample in
year t using the average coefficients, A, and computing a residual for
each worker based on his actual percentile in that year's residual
distribution and the average cumulative distribution over the full
sample. The major advantage of this over the more standard variance
accounting framework is that it allows us to look at how composition
changes have affected the entire wage distribution and not just the
variance. We can determine how changes in the distribution of ob-
servables have affected other inequality measures such as the inter-
quartile range or the ninetieth-tenth percentile differential or how
the effects have been different for inequality above and below the
mean.
If we want to allow both observable prices and observable quantities
to vary through time, then we can generate wages by

y2 = XitAt + F'l(0it1X_). (5)

In this case we predict wages for each worker in year t given his
observable characteristics and the wage equation estimated for year t
and again assign him a residual based on the cumulative distribution
for all years. Finally, if we allow observable prices and quantities and
the distribution of residuals to change through time, we obtain

it= XIitt + F7-(0it1Xit) = XI-t3 + uit = Yit' (6)

WAGE INEQUALITY 429

which replicates the actual wage distribution since ujt = F7-1(OIXit)

by definition of the cumulative wage distribution.
Our basic technique will be to calculate the distribution of Y,'i,Y.,
and Yi for each year and attribute the change through time in in-
equality in the Y t distribution to changes in observable quantities. We
then attribute any additional change in inequality in Y? to changes in
observable prices, and finally we attribute any additional changes in
inequality for Y' beyond those found for Y? to changes in the distri-
bution of unobservables (i.e., changes in unmeasured prices and
quantities). The same analysis can be done in other orders and would
simply rearrange the assignment of interaction terms.
The remaining three panels in figure 7 give the part of the nine-
tieth-tenth percentile log wage differential accounted for by each of
these three components (each is measured as a deviation from its
overall mean). Panel B gives the effects of changes in the distribution
of observables. As is clear from the figure, changes in observable
characteristics have had only a very modest impact on overall inequal-
ity. This implies that the changes in the age and education composi-
tion of the work force have not had a direct effect on the level of
inequality (with skill prices held fixed). Panel C looks at the compo-
nent of changes in inequality due to changes in observable prices (i.e.,
changes in the returns to education and experience). As the figure
illustrates, changes in observable prices had a very modest effect on
inequality until about 1980. Since 1980, however, the rapid increases
in education differentials and returns to experience (for the less edu-
cated groups) have increased the ninetieth-tenth percentile log wage
differential by about 12 percentage points (more than half of the
total increase over the 1980s).
Panel D examines the component due to changes in unmeasured
prices and quantities (i.e., the residual). As the figure shows, this
component is by far the most important for the overall increase in
inequality (accounting for about two-thirds of the increase). More-
over, unlike the increase in observable skill prices, the increase in
inequality based on unobservables has been operating since the late
1960s. While it is fair to say that skill premia based on both observed
skill differences and unobserved skill differences have increased since
1970, it is important to note that the timing of these changes is very
different. The rise in within-group inequality (measured by the resid-
ual component) preceded the increase in returns to observables by
over a decade. On the basis of this difference in timing, it seems clear
to us that there are at least two unique dimensions of skill (education
and skill differences within an education group) that receive unique
prices in the labor market.
Table 4 quantifies the contribution of observed quantities and
430 JOURNAL OF POLITICAL ECONOMY

TABLE 4
OBSERVABLE AND UNOBSERVABLE COMPONENTS OF CHANGES IN INEQUALITY

Unobserved
Total Observed Observed Prices and
Change Quantities Prices Quantities
Differential (1) (2) (3) (4)
A. 1964-88

90-10 .373 .035 .128 .208

90-50 .146 .020 .068 .061
50-10 .227 .015 .060 .147
B. 1964-79

90-10 .165 .029 .014 .119

90-50 .059 .011 .010 .036
50-10 .106 .018 .004 .083
C. 1979-88

90-10 .208 .006 .114 .089

90-50 .087 .009 .058 .026
50-10 .121 - .003 .056 .064

NOTE.-The years refer to the middle point of the 3-year interval. Col. I gives the change in the indicated
statistics over the years shown. Components in cols. 2-4 are calculated on the basis of the full distribution accounting
scheme outlined in the text.

prices and the unobservables to the increase in the standard deviation

as well as the increase in the ninetieth-tenth percentile differential.
The major information contained in the table that was not apparent
in figure 7 is the difference in explanatory power for inequality above
and below the mean. This ability to estimate how different parts of
the wage distribution have been affected by the various components
is the major advantage of the full distribution accounting scheme
proposed here over the more conventional variance accounting
framework.
Panel A in table 4 refers to the change over the period 1964-88
(the years refer to the middle year of the 3-year interval). As the
table shows, changes in observed quantities account for only about 7
percent of the increase in wage inequality between the fiftieth and
tenth percentiles (primarily because of the decline in the number of
men with very low education levels) but have accounted for about 14
percent (i.e., .020/.146) of the increase in inequality above the mean.
Similarly, the rise in observable prices (primarily the increase in col-
lege returns) has accounted for about 47 percent of the increase in
the ninetieth-fiftieth percentile differential but only about 26 percent
of the growth in the fiftieth-tenth percentile differential. As a result,
the unobserved component accounts for 65 percent of the increase
WAGE INEQUALITY 431
in inequality for those below the median but less than half of the
increase in inequality for those above the median. Apparently, the
increasing wage gap between the highest wage earners and the aver-
age is much more understandable in terms of observables than the
even larger increase in the wage gap between the average and the
low end.
Panels B and C perform the same decomposition for the periods
before and after 1979 (the time in which education returns began to
rise). For the 1964-79 period, the observables combine for a slight
increase in overall inequality (with the increase in inequality gener-
ated by changing composition exceeding any effect on inequality gen-
erated by the fall in returns to education). Unobservables account for
the vast majority of the increase in inequality, accounting for over 60
percent of the rise in the ninetieth-fiftieth percentile differential and
78 percent of the increase in inequality below the median. Panel C
looks at the most recent period. For this period the changes in ob-
served prices account for the dominant portion of the increase in
inequality. However, even for this period, unobservables remain im-
portant, accounting for about 53 percent of the increase in inequality
below the median.
The results in table 4 reinforce the finding of figure 7 that the
timing of the increase in inequality has been very different for the
observables and unobservables. This suggests that the notion of skill
or a rise in returns to skill may be too broad and that it may be useful
to distinguish alternative types and hence alternative "prices" of skill.
Under this interpretation, the 1970s were characterized by a rapid
rise in the returns to skills within education and experience categories
but no significant rise in returns to skill across these categories. In
contrast, the 1980s were characterized by a rise in returns to both
observed and unobserved skill. Figure 8 addresses this issue by plot-
ting three skill "prices" for education, experience, and within-group
skills. The price series were derived from yearly regressions of log
weekly wages on education and experience effects, as follows. Within-
group skill price is the ninetieth-tenth percentile log wage differential
from the regression residuals. The education skill price is an un-
weighted average of the college-high school log wage differential
across experience levels. Similarly, skill prices for experience are con-
structed from the average log wage differential within education lev-
els between the 26-35 and 1-10 experience groups. Each of these
differentials is indexed to its 1963-64 average.
The pattern for the within-group differential tells a familiar story.
Skill differentials within a group were steady until the late 1960s and
then increased greatly from 1970 to 1989. Within-group differentials
end up being about 30 percent larger in 1989 than in 1963. The
432 JOURNAL OF POLITICAL ECONOMY

Unobserved Skill (o) Experience ( ) Education (+)

170 -

160 -

150-
U

m. 140-
H
130-
V
D 1
1204
V

110

100

65 70 75 80 85
YEAR
FIG. 8.-Skill price indexes for men, 1963-89 (1963/64 = 100)

price of education shows a very different pattern. Education returns

increased over the 1960s (within-group differentials were stable) and
decreased significantly over the next decade. By 1979 the education
premium had fallen roughly 10 percent below its value in 1963. Since
1979 the education premium has skyrocketed, so that by 1989 it stood
more than 25 percent above its value in 1963. The experience returns
show a still different pattern. Experience returns fall slightly over the
late 1960s and then rise rapidly (primarily for college graduates) with
the arrival of the baby boom cohorts in the early 1970s. Experience
returns then rise slowly over the next 7-8 years before rising rapidly
in the early 1980s with the collapse of wages for less skilled younger
workers. The basic message we take away from figure 8 is that all
skill premia have increased greatly since 1963 but that these increases
have shown quite different timing (particularly for education and
within-group skills).

V. Accounting for the Rise in Returns to Skill

The skill prices shown in figure 8 show that returns to a wide variety
of skills have increased over the past two and one-half decades. Given
the rapid rise in average education levels over this period (from 10.9
years of schooling in 1963 to 12.3 years of schooling in 1989), such
a large increase in skill premia must have resulted from a significant
WAGE INEQUALITY 433
demand shift toward the most skilled. Such a shift in labor demand
can be due to either a shift across industries toward industries that
demand more skilled workers or a technological shift within indus-
tries toward production methods that favor the most skilled. To mea-
sure these shifts empirically, we divided the economy into the 12
industries and 11 occupation categories given in table 5. The columns
of the table give the fractions of workers in the bottom 10 percent, the
middle 10 percent, and the top 10 percent of the wage distribution
employed in each occupation and industry. As the table shows, skill
composition varies significantly across industry and occupation cate-
gories. As a result, shifts in industrial and occupational composition
should generate significant changes in relative demand.
To construct a demand index, we write the output of an occupation
by industry cell (this can be thought of as an intermediate good) as
Yij= Fij(X ij,...* , X loij)q
g (7)

TABLE 5
INDUSTRY AND OCCUPATION DISTRIBUTION BY PERCENTILES, 1959-89

PERCENTILES

0-10 45-55 90-100

Industry:
Agriculture/mining 1.20 2.33 2.63
Construction 13.34 8.66 8.93
Manufacturing:
Low-tech 9.03 2.75 1.87
Basic 11.85 17.26 10.18
High-tech 6.54 15.52 17.77
Commercial transportation
and utilities 5.66 10.84 9.49
Wholesale 4.86 5.55 6.75
Retail 21.66 9.60 5.96
Professional services, finance,
insurance, and real estate 9.91 8.58 21.90
Education and welfare 5.26 6.13 5.16
Public administration 2.56 10.02 7.75
Other service 8.14 2.75 1.62
Occupation:
Professional/technical 5.19 15.27 31.11
Managers 5.09 12.71 38.64
Sales 4.99 5.29 8.88
Clerical 6.44 8.71 2.17
Craft 20.14 27.20 12.50
Operatives 20.37 14.18 2.36
Transportation operatives 8.94 6.69 2.35
Laborer 12.76 4.74 .69
Farm private household .57 .09 .00
Service 15.52 5.13 1.30
434 JOURNAL OF POLITICAL ECONOMY

where Y. is the output produced by industry/occupation cell ij, and

Fij is the corresponding production function giving output as a func-
tion of the number of workers from each percentile (XIj, ... , X1000.

With constant returns to scale the dual of this problem is then

Xii = Yi x Dj(W,... , W100), (8)

where Xii is the 100 X 1 vector of employment by percentile in indus-
try/occupation cell ij, Dij is the vector of unit demand functions for
each percentile, and W1, . . . , W100are the wages at the different
percentiles. The change in labor demand associated with a given
change in the industrial and occupational structure is then
dYi
dXd = Z i ]Xi,. (9)
ij V

Empirically, we measure the change in output of an industry by

occupation cell by the change in factor inputs at fixed reference prices
(note that this also allows for factor-neutral technological change
within industries and occupations) and measure Xi by the employ-
ment distribution across industries and occupations by percentile over
the entire sample. Demand growth for any group of workers is mea-
sured as a weighted average of the growth in factor inputs in indus-
try/occupation cells; the weights are industry by occupation shares
for that group. Therefore, groups employed largely in expanding
sectors will experience rising demand, and groups employed largely
in contracting sectors will have reduced demand. Changes in the cod-
ing of industries and occupations limit our ability to make these calcu-
lations for years prior to 1967, so we limit this analysis to the 1967-89
period and 1959.
Figure 9 graphs the percentage change in relative demand at each
percentile of the overall wage distribution accounted for by shifts in
employment across our industry and occupation categories for the
whole period and three subperiods.3 Panel A gives the change in
demand over the full period, 1959-89. As the figure shows, demand
fell by roughly 10 percent for workers below the median and in-
creased by between 5 and 40 percent for workers in the top quartile.
Since the demand shifts that can be proxied by our crude demand
measures are likely to be a small part of the true change in demand,
such large relative movements suggest a significant shift in favor of
the most skilled. The remaining three panels compute demand

3 Given the rapid rise in skill premia over the period, these "measured" demand
shifts must understate the "true" changes in demand that would have occurred with
fixed skill prices.
 00

o 0

* 0

a 0

la
so 1~~~~~~~~~~~~~~m~C
di~~~~~~~~~~~~~~~nI

-0~~~~~~~~~~

W4 I IR W!~ ~~~~~~~~~~~~~"4 b

"4
0 0 0
.4 .4 "
0 0
0 "
0

-0 CZ-

b0
z

-.M

00

pug" T ougo oipueeaUTfougoDo

436 JOURNAL OF POLITICAL ECONOMY

changes for three subperiods: 1959-69 (when skill premia showed

little change) and 1969-79 and 1979-89 (when skill premia increased
rapidly).
As can be seen in the figure, rising demand for skill has character-
ized the past three decades. The demand for skill increased over the
1960s (panel B), whereas skill premia did not change. In fact, the
only real contrast in the later periods appears to occur at the highest
percentiles, where demand growth accelerated from the 1960s to the
1970s and again from the 1970s to the 1980s. We interpret the four
panels of figure 9 to show that growing demand for skill is an impor-
tant factor leading to the growth in wage inequality over the past two
decades (causing the demand for workers in the top decile to grow
about 30-40 percent relative to the demand for workers in the lowest
deciles). However, figure 9 also shows that growth in the demand for
skill was not limited to these two recent decades. Unless wage inequal-
ity is particularly sensitive to demand shifts for the highest-decile
workers, changes in demand growth as we have measured it here do
not explain the contrast between the events of the 1960s and the two
most recent decades.
The change in occupation and industry composition has been hy-
pothesized to have had an effect on the wage structure quite separate
from that described here. Some authors (Bluestone and Harrison
1988) and the popular press have emphasized the shift in industrial
composition toward services and away from manufacturing as a shift
toward low-wage jobs and a shift toward industries in which high-
skilled or highly educated workers do well but less educated and less
skilled workers do poorly. This alternative theory suggests that in fact
wage inequality has risen as a result of a shift in employment toward
low-wage jobs or a shift in employment toward both high- and low-
wage jobs. The demand index numbers clearly reject the hypothesis
that employment has shifted toward low-wage jobs but support the
view of a shift toward high-wage jobs.
A direct way to evaluate these types of explanations is to look at
a decomposition of the variance into within- and between-industry
components (the same can be done by occupation or by both industry
and occupation). The basic decomposition is that
2
2ZPitu? + PtWit~Wtt) (10)
I ~~~~~~~~~i
where a 2 is the variance of weekly wages in year t, Pit is the fraction
of workers in industry i in year t, a. is the variance of wages in
industry i in year t, and witis the average log weekly wage in industry
i in year t. The change in variance through time can then be decom-
posed into a shift in industry composition (i.e., a shift in the Pit)
WAGE INEQUALITY 437
and shifts in within- and between-industry wage differences. The
composition effects will raise the variance to the extent that employ-
ment shifts either toward industries with a high within-industry vari-
ance or toward industries with average wages very different from the
mean. The skill price effects show up as either increases in wage
inequality within an industry or increases in the wage differentials
between industries.
Table 6 quantifies the total change in variance into four compo-
nents for the period as a whole and for subperiods before and after
1979. For the full period, wage changes account for .142 out of the
.158 change in variance. Of the remainder almost all of the change
is accounted for by a shift to industries with a higher within-industry
variance. The subperiods show a similar decomposition, with relative
wage changes accounting for .048 of the total change of .054 for the
1967-79 period and .093 out of .104 for the 1979-89 period. These
numbers would seem to imply that the shift in industrial composition
has not had an important composition effect on the wage distribution.
If the shift in industrial composition is important, it must have af-
fected the returns to skill as emphasized by our demand shift theory.
Our basic interpretation of these results and those in figure 9 is that
the fall in wages for the least skilled is symptomatic of a fall in de-
mand for low-wage workers and not a rise in the number of low-
skilled jobs. There are simply too few low-wage jobs, to use the com-
mon jargon, and not too many, as previous authors have contended.
While the story of a rapid increase in the demand for skill offers
some explanation for why skill premia are roughly 30-50 percent
higher today than in 1963, it does not explain why the time series of
skill prices are so different for education, experience, and within-
group skill. One potential explanation for the returns to education
story is that supply growth as well as demand growth has been an
important factor in determining education returns. The first thing to
note along these lines is that the supply of educated workers has
increased enormously over the past two decades (increasing by

TABLE 6
EFFECTS OF INDUSTRIAL COMPOSITION ON WAGE INEQUALITY, 1967-89

WITHIN-INDUSTRY BETWEEN-INDUSTRY

TOTAL Variance Composition Wage Composition

YEARS CHANGE Changes Effect Changes Effect

1967-89 .158 .119 .015 .023 .001

1967-79 .054 .039 .006 .009 .000
1979-89 .104 .079 .009 .014 .001
438 JOURNAL OF POLITICAL ECONOMY

roughly 100 percent). Given that wage premia have increased in spite
of this enormous growth in supply, it seems clear that there must
have been significant growth in the demand for education. One view
for the fall in education returns over the 1970s and the subsequent
rapid rise in the 1980s (suggested by Murphy and Welch [1989]) is
that supply grew faster than demand over the 1970s and then slower
than demand during the 1980s. Similarly, perhaps much of the in-
crease in the returns to experience can be attributed to the arrival of
the baby boom cohorts and the associated youthening of the labor
force. However, as Katz and Murphy (1992) point out, this does not
seem to be a sufficient explanation for the rapid rise in experience
returns in the 1980s as the baby boom cohorts moved up in the age
distribution.
A complete explanation of the economic phenomena behind the
data we have described would appear to require at least two addi-
tional components. First, it seems clear from the data that the demand
for skill has risen. Further research must identify the sources of this
demand shift; likely but untested candidates are biased rates of tech-
nological progress and changes in the world economy. Also, an accu-
rate description of the determinants of the timing in observable skill
prices requires explicit consideration of supply and demand forces
for the skill in question. Continued success in understanding the
forces bringing about greater wage inequality hinges on the progress
of future work in these areas.

VI. Wage Inequality versus Income Inequality

Our discussion to this point has focused on inequality in weekly and
hourly wages. While this focus is probably correct for purposes of
measuring the returns to skill, income rather than wage inequality has
historically been more widely described and analyzed. Our purpose in
this section is not to choose between these two concepts as measures
of welfare or anything else but simply to show that empirically the
two concepts are quite distinct. For our purpose here we focus on
the contrast between inequality in annual earnings and inequality in
weekly wages. Since annual earnings and weekly wages for an individ-
ual are much more closely linked than total family income (a common
income measure) and weekly wages, any contrast we find here for
weekly wages and annual earnings is likely to greatly understate the
true income/wage distinction.
Panel A of figure 10 graphs the variance of log annual earnings
for 1963-89 using the same CPS data used for our wage inequality
calculations, except that we no longer exclude those working 1-13
weeks. As is clear from the figure, earnings inequality declined some-
 r;u

3~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~1

001~~~~~~~~~
h 0

s r ra rw o o o o o. o.o.Ii

0 0 0 0 0 0 0 0 0 0 00

C
u/>m1 A14u" SD01 40 *uUTJUA on ;uuOdUo3 *3UUIJBoA
C"

bio

CZ

CZ

I~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
CZ
I~~~~~~~~~~ c

ri~~~~~~~~~~~~~~~~~~~~~~~~~~r

I IJ? IO U I0 I0 IMT? PMI I IO 0 IO D IW IA

440 JOURNAL OF POLITICAL ECONOMY

what from the mid-1960s through about 1968 and has increased sig-
nificantly since. Business cycle swings are also clearly important, as
evidenced by the large increases in inequality during the recessions
of 1971, 1975, and 1982. Overall from 1968 through 1989 the vari-
ance of log earnings increased by about 80 percent (from .25 to .45).
The remaining panels of the figure decompose this increase into the
weeks worked variance, the weekly wage variance, and the covariance
of weekly wages and weeks worked. Since annual earnings are simply
the product of weeks worked and the weekly wage, we can write the
log of annual earnings as
y =+ w, (11)
where y is the log of annual earnings, 1 is the log of weeks worked,
and w is the log weekly wage. Using this notation we can write the
variance of log annual earnings, r2, as
cr2 = as-2 + (3-2 + 2cwx(12)
where a 2 is the variance of log weekly wages, ar is the variance of
log weeks worked, and au1 is the covariance of log earnings and log
weeks. The variance of weekly wages is shown in panel B of figure
10 and follows a pattern quite similar to the ninetieth-tenth percentile
log wage differential shown in figure 7. The variance of weekly wages
is very steady from 1963 through 1968 and then increases relatively
smoothly from 1968 until 1986. In contrast, the variance of log weeks
(panel C) shows a distinct cyclic pattern, with sharp rises in 1970-71,
1975, and 1982. This cyclic pattern results from the fact that reduc-
tions in annual hours are very unevenly distributed across workers
(i.e., relatively few workers work many fewer weeks).
The covariance term in panel D reflects the growing positive associ-
ation between wages and time worked. As two of us have found in
our other work (Juhn, Murphy, and Topel 1991; Juhn 1992), this
strengthening of the cross-sectional labor supply relationship occurs
mostly prior to 1975, after which the relationship is relatively stable.
Over the period as a whole, the increase in the weekly wage variance
accounts for about .14 of the overall .18 increase in the annual earn-
ings variance. The remainder is attributable to the increased variance
of weeks worked and a small rise in the covariance of weekly wages
and weeks worked.
We stress the need to distinguish between the earnings and wage
inequality concepts. For example, because of the highly cyclic pattern
of weeks worked, the variance of log annual earnings is actually lower
in 1989 than in 1982, whereas the variance of weekly wages is actually
about 20 percent higher in 1989 than in 1982. As measured by the
variance of log annual earnings, inequality is lower now than in 1982,
WAGE INEQUALITY 441
but as measured by weekly wages, it is significantly higher. This strik-
ing example serves as a warning to keep in mind which inequality
concept is of interest for a particular problem.

VII. Conclusion
In this paper we have identified the enormous increase in wage in-
equality among male workers over the past two decades. The trend
toward greater wage inequality is attributable primarily to increases
in the premia on both unobserved and observed (such as education)
dimensions of skill, with the majority of the increase over the period
due to the unobserved component. We also show that the timing of
the increased premium on the unobserved components of skill differs
from the timing of changes in the skill premia on education and labor
market experience. In particular, returns to unobservable skills have
shown a steady increase since 1970.
Our basic rationale for this increase is the rapid growth in demand
for skilled workers. While it seems clear that skill premia have risen
and hence that the demand for skill has risen as well, the exact source
of this demand increase is as yet unknown; likely candidates are bi-
ased rates of technological progress and changes in the world econ-
omy. We feel that further progress in comprehending the increases
in wage inequality documented here will require a greater under-
standing of these fundamental underlying forces.

References
Bluestone, Barry. "The Changing Nature of Employment and Earnings in
the U.S. Economy: 1963-1987." Paper prepared for the Conference on
Job Creation in America, Univ. North Carolina, Chapel Hill, April 1989.
Bluestone, Barry, and Harrison, Bennett. "The Growth of Low-Wage Em-
ployment: 1963-86." A.E.R. Papers and Proc. 78 (May 1988): 124-28.
Borjas, George J. "Assimilation, Changes in Cohort Quality, and the Earnings
of Immigrants."J. Labor Econ. 3 (October 1985): 463-89.
Chiswick, Barry R. "Is the New Immigration Less Skilled than the Old?"
J. Labor Econ. 4 (April 1986): 168-92.
Juhn, Chinhui. "Decline of Male Labor Market Participation: The Role of
Declining Market Opportunities." QJ.E. 107 (February 1992): 79-121.
Juhn, Chinhui; Murphy, Kevin M.; and Pierce, Brooks. "Accounting for the
Slowdown in Black-White Wage Convergence." In Workersand Their Wages:
Changing Patterns in the United States, edited by Marvin H. Kosters. Washing-
ton: American Enterprise Inst. Press, 1991.
Juhn, Chinhui; Murphy, Kevin M.; and Topel, Robert H. "Why Has the
Natural Rate of Unemployment Increased over Time?" Brookings Papers
Econ. Activity, no. 2 (1991), pp. 75-142.
Katz, Lawrence, and Murphy, Kevin M. "Changes in the Wage Structure
442 JOURNAL OF POLITICAL ECONOMY

1963-87: Supply and Demand Factors." QJ.E. 107 (February 1992):

35-78.
LaLonde, Robert J., and Topel, Robert H. "Labor Market Adjustments to
Increased Immigration." In Immigration, Trade, and theLaborMarket, edited
byJohn M. Abowd and Richard B. Freeman. Chicago: Univ. Chicago Press
(for NBER), 1991.
Levy, Frank. "Recent Trends in U.S. Earnings and Family Incomes." In
NBER MacroeconomicsAnnual, vol. 4, edited by Olivier J. Blanchard and
Stanley Fischer. Cambridge, Mass.: MIT Press, 1989.
Murphy, Kevin M., and Welch, Finis. "Wage Premiums for College Gradu-
ates: Recent Growth and Possible Explanations." Educational Researcher 18
(May 1989): 17-26.
."The Structure of Wages." QJ.E. 107 (February 1992): 215-326.
Sawhill, Isabel V. "Poverty in the U.S.: Why Is It So Persistent?" J. Econ.
Literature26 (September 1988): 1073-1119.
Smith, James P., and Welch, Finis. Closing the Gap: Forty Years of Economic
Progressfor Blacks. Santa Monica, Calif.: Rand Corp., 1986.