Women's Gains or Men's Losses? A Closer Look at the Shrinking Gender Gap in Earnings
by
Annette Bernhardt, Martina Morris, Mark S. Handcock
Citation
Title:
Women's Gains or Men's Losses? A Closer Look at the Shrinking Gender Gap in Earnings
Author:
Annette Bernhardt, Martina Morris, Mark S. Handcock
Year:
1995
Publication:
The American Journal of Sociology
Volume:
101
Issue:
2
Start Page:
302
End Page:
328
Publisher:
Language:
English
URL:
Select license:
Select License
DOI:
PMID:
ISSN:
Updated: January 30th, 2013
Abstract:
Women's Gains or Men's Losses? A Closer Look at the Shrinking Gender Gap in ~arnin~s'
Annette Bernhardt
Institute on Education and the Economy
Martina Morris
Columbia University
Mark S. Handcock
New York University
The recent closing of the gender wage gap is often attributed to increases in women's human capital. This explanation neglects the effect of growing inequality in men's earnings. The authors develop a decomposition that allows them to test how distributional changes in men's and women's earnings combine to yield changes in women's economic status. Using Current Population Survey data from 1967 to 1987, the authors find that the striking polarization in white men's earnings has played a critical role in generating women's relative economic gains, though more for white women than for black women. For both groups, the results predict a future slowing of women's relative progress.
THE GENDER WAGE GAP
For American women who work, the 1980s brought good news. After years of stagnation, their economic standing in the labor market finally began to improve. The most visible and publicized gain took place in women's earning power relative to men's. The gender wage gap had in
'We would like to thank John Myles, Kevin Murphy, James Coleman, and Aage Serensen for their input at various stages of this research, as well as several AJS reviewers for their careful reading and helpful comments. We also thank Robert D. Mare and Christopher Winship for the use of the uniform series of the March Current Population Survey files. Please direct all correspondence to Annette Bernhardt, Insti tute on Education and the Economy, Teachers College, Box 174, Columbia Univer sity, New York, New York 10027.
O 1995 by The University of Chicago. All rights reserved. 00029602/96/101020002$01.50
302 AJS Volume 101 Number 2 (September 1995): 30228
recent history been an obstinate statistic, remaining unchanged for close
to half a century despite progress on social, political, and legal fronts.
The ratio of female to male wages hovered around .60 throughout much
of the postWorld War I1 period. Only during the 1980s did the wage
gap begin to close. Current estimates range from a low of .65 to a high
of .72, with analysts often citing a ballpark figure of .70 (for a detailed
review, see Marini [1989]).' When black women have been studied, the
results suggest even stronger economic gains (Blau and Beller 1992).
Thus, many have interpreted the current trends as a significant turning
point, an indication that full gender equality may finally be within sight
(e.g., Heidi Hartmann quoted in Nasar 1992).
Explanations for the decline in the gender gap have generally focused on increases in women's qualifications, such as experience, education, and related skills. While work in this area is still young, some basic conclusions can be made. Women's human capital and monetary returns to that capital have grown relative to men'sthis is especially true of work experience and job tenure. Estimates suggest that these factors are responsible for 30%50% of the reduction in the wage gap (O'Neill and Polachek 1993; Wellington 1993). To a significant degree, these findings are driven by a cohort effect, with new generations of women bringing higher skill and experience levels to the labor market at the same time that older generations are leaving it.
At this cursory level, then, it seems that the future of women's economic welfare is a positive one. This optimism is nicely illustrated by a recent New YOY~Times article, "Women's Progress Stalled? Just Not So" (Nasar 1992). The article emphasizes that women are entering the workforce in greater numbers with more education and work experience, fewer drop out of the labor market and for shorter periods of time, and more are entering formerly maledominated occupations. With new cohorts of women offering ever improving qualifications, the trend to ward equalization of wages seems assured, and one may therefore con clude that "today's gains in women's pay reflect solid, enduring trends"
(P 10).
THE ROLE OF CHANGES IN MEN'S EARNINGS
While such an optimistic assessment is certainly welcome, it may be somewhat premature. Recent work suggests that we need to look further in explaining the shrinking gender gap. Specifically, researchers are be ginning to examine the impact of industrial restructuringAmerica's
The variation in estimates stems from the use of different data sets, summary mea sures, and sample definitions.
shift to a postindustrial, servicebased economy. The processes involved are often quite complex and there is a plethora of competing arguments about the net effects of these changes on the gender gap. Depending on the author, the shift to services is seen as leading to either greater or less economic equality, for either men or women or both (see below). This variation stems from the time period studied, whether industrial or occu pational shifts are analyzed, whether aggregate or micro data are used, and whether white workers are the sole focus of the analysis. Given these variations, it is not surprising that the resulting estimates differ.
One clear theme that has emerged, however, is the possibility that women gained economic ground on men partly because male workers suffered disproportionately from deindustrialization. That is, men's earn ings were more vulnerable to steep declines in manufacturing jobs than were women's earnings, and the growing service sector did not provide sufficient equivalents to the wellpaid and relatively lowskill jobs that were lost. Researchers have tested this idea in a variety of ways, for example, by linking declines in unionization to differential earnings gains for men and women over time or by documenting that the effect of industry indicators on average wages has changed between two time points (see, e.g., Lorence 1991; Even and Macpherson 1993; Montgomery and Wascher 1987). But the validity of the thesis has yet to be estab lished. It is rarely tested directly, by modeling how structural changes have affected changes in the gender gap (Lorence 1991). Part of the problem is that rigorous tests are difficult to construct given data con straints. The effects of industrial and occupational shifts on the gender gap are especially hard to disentangle, because they have worked in a variety of conflicting directions (Tienda, Smith, and Ortiz 1987).
Somewhat more success has been had by focusing exclusively on earn ings. This approach does not allow one to identify underlying structural causes, but it does speak much more directly to the question of whether changes in men's earnings have contributed to women's recent economic gains. The intuition behind this approach comes from the striking in crease in earnings inequality that American workers experienced during the 1970s and 1980s. That the gap between the rich and poor has grown is by now a wellpublicized fact (Harrison and Bluestone 1988; Kuttner 1983; Danziger and Gottschalk 1993). The magnitude of this change is astonishing. For example, between 1979 and 1987 the income of the top tenth of American families grew by 10.1%, while that for the bottom tenth actually fell by 8.7% (Karoly 1993). In general, there is agreement that earnings inequality rose by 20%30% since the early 1970s after years of stability.
The important point is that the timing and magnitude of the increase in inequality differed between men and women (Morris, Bernhardt, and Handcock 1994; Karoly 1993). White men experienced the strongest in crease. It was driven by a pronounced polarization in their earnings stemming from the beginning of the 1970s' with growth at both the top and bottom of the distribution and a decline in the middle. By contrast, white women's earnings first improved significantly across the board and only began to polarize at the start of the 1980s. And for black women, the 1980s brought a pronounced rise in lowwage earners, after a decade in which the middle of the earnings distribution had swelled consider ably. However, in 1987 the degree of inequality in white men's earnings was still higher than that for either of these two groups of women.
In short, white men experienced a marked and consistent rise in earn ings inequality to an extent not matched by women. This difference in distributional changes suggests that white and black women's standing in the labor market may have improved partly because some men's stand ing has worsened. A number of recent studies have taken up this theme. Simple decompositions of trends in the gender wage gap suggest that declines in men's average earnings have been as important as increases in women's average earnings (Institute for Women's Policy Research 1992). More complex links between wage inequality and group inequality have been suggested, both generally (Juhn, Murphy, and Pierce 1991; Blau and Kahn 1992) and specifically in terms of the recent closing of the gender gap (Blau and Kahn 1994).
But we do not yet have a full account of how the evolving structure of Americans' earnings has given rise to women's economic gains. In this article, we attempt to provide such an account. Our analysis moves beyond the typical focus on average or median earnings differences, to ward a focus on how the full distribution of women's earnings relative to men's has changed over time. We try to identify exactly where that change has come fromwhether the convergence in median earnings forms the whole picture or whether changes in distributional inequality for each sex contributed as well. As we will see, these analyses recast our understanding of what women's economic progress really means and dampen our expectations for a continuing rapid decline in the gender wage gap. The findings also underscore that race remains a critical inde pendent dimension in understanding earnings inequality. Finally, the methods developed for this analysis constitute an important contribution in and of themselves.
DATA
The data used here are the March uniform series of the Current Popula tion Survey for the years 196787. We focus on yearly earnings of indi vidual workers and follow the convention of restricting the sample to
workers 1665 years old, not in school, the military, or farming. The
sample is further restricted to fulltime, yearround workers. This set of
restrictions is the common method of limiting the variation in hours
worked, so that differences in earnings are not confounded by differences
in amount of work performed. The analysis focuses on nonHispanic
white men and women and nonHispanic black women. With all of the
above restrictions in place, the size of the resulting samples is roughly
20,000 per year for white males, 10,000 per year for white females, and
1,500 per year for black females.
The Current Population Survey, like many employment data sets, suf fers from problems in the data on yearly earnings. First, there is the presence of reported earnings that fall below the minimum amount a worker could legally earn. An initial response might be to simply delete or recode any cases falling below the legal minimum. Our decision, an accepted one, was not to do so: the percentage of low reports declines over time, in part due to increased Census Bureau cleaning. The result is that both keeping and deleting the cases introduces bias across years. Also, low reports tend to bunch close to the legal minimum, suggesting the presence of valid earnings that one would not want to delete. There is, however, no way to disentangle real from erroneous low earnings.
Second is the wellknown "heaping" problem in reported wages: re spondents tend to round their yearly earnings to whole figures. For this analysis, the heaping has been removed by an imputation scheme that reallocates earnings reported in round numbers within a range of values. For example, a report of $10,000 is assumed to be a rounded version of yearly earnings between $7,500 and $12,500, so the new imputed value is chosen randomly from this range. The resulting distribution has the basic statistical characteristics of the original reported distribution (for additional information see Handcock [1995]). Imputation was also used to eliminate "topcoding." The Census Bureau uses a single code (or topcode) for earnings above a certain threshold (e.g., in 1987 all cases above $99,999 were coded as $99,999). The imputation scheme imputes this upper tail using a Pareto distribution. The resulting mean for the topcoded values is very close to 1.45 times the topcode, which is consis tent with the value commonly used to recode such observation^.^
METHODS
The Relative Distribution
Much of the argument in this article rests on moving the methodological focus of the debate from simple averages to full earnings distributions.
The reader may skip the following, more technical section without loss of continuity.
The concept of relative distributions (Morris et al. 1994) is central to this move. The relative distribution can be thought of as a ratio. It compares the density of two groups of earners at each earnings level, where the groups can be defined by sex, race, year, or any category of interest. One group forms the reference earnings distribution, and another forms the comparison distribution. The relative distribution is then the ratio of comparison group earners to reference group earners at each level of the earnings distribution. It takes the value "1" when the numerator and denominator are equal, which occurs when the fraction of earners at that level of the earnings distribution is the same for both groups. When the fraction of comparison earners is higher or lower than that of the refer ence group, the value takenby the relative distribution is greater or less than one, respectively. The value itself can be interpreted as the number of comparison group earners per reference group earner at each earnings level if the groups were of equal size.
For notation, let yo and y, equal reference group and comparison group earnings, respectively, and fo(.) and fl(.) equal the density function for each group's earnings distribution. The relative distribution of group 1 (yl) to group 0 (yo) earnings may then be expressed as4
It is important to understand that the numerator uses the density function of the comparison group, f,, but gives it the argument yo, the earnings level of the reference group. The effect is to compare the earnings density of the two groups at the same point on the scale, defined here by yo. For example, if yo represents men's earnings and y, women's earnings, then the relative distribution compares the density of women earners to male earners at each level of the earnings distribution. We refer to this as a betweengroup comparison.
Relative distributions can also be used for analyzing changes in a single group's earnings over time. In this case, one would use real earnings in a reference year to define the reference distribution and real earnings in subsequent years to define the comparison distribution. The result is a time series of relative distributions. We refer to this as a withingroup comparison over time. As we will see, the extension to a betweengroup analysis over time is straightforward.
Technically, the relative distribution is a proper probability density function and takes as its arguments the densities of each group at each percentile rank of the reference earnings level. It therefore represents the density of comparison group earn ers at each percentile of the reference group earnings scale. The precise technical definition of the relative distribution is given in the appendix.
These approaches readily lend themselves to graphical methods. Spe cifically, the reference earnings distribution is partitioned into deciles (so the denominator of the relative distribution always takes the value .lo), and the fraction of the comparison earners falling into each of these deciles is used to form the numerator. When the resulting ratio is multi plied by 10, the value represents the percentage of comparison earners falling into each reference decile. This decilebased version is adopted for all of the figures below, although the methods in this section continue to be developed in continuous form.
Multiplicative Decomposition
The critical point of this article is that changes over time in a relative distribution can be decomposed into two parts: changes in the median and changes in shape. The intuition behind the decomposition is straight forward and easy to motivate if we first focus on a single group's earnings over time: A median shift would occur if every earnings level were multi plied by the same factor, for example, if every earner received the same costofliving adjustment. The entire earnings distribution would then be moved up (or down) on the dollar scale, but the underlying shape of the distribution would remain constant. A shape shift, by contrast, would occur if earners were redistributed along the earnings scale. The "declin ing middle class" scenario provides one example of such a redistribution, with earners moving from the middle of the distribution into the upper and lower tails. But other shape changes are also possible, with growth occurring only in the upper tail, the lower tail, or the middle of the distribution.
One or both of these changes may be operating over time. Because the relative distribution is a ratio, the natural decomposition is multiplica tive. To formalize the definitions of the two components, let the distribu tion of earnings at time t relative to time 0 be
This relative distribution can be decomposed into the two components:
wheref,d(~)is the density of year t earnings deflated by the median ratio, median(yo)l median(yt). The first term in equation (3) can be interpreted as the median shift effect. The numerator and denominator differ only in the median of the density functions, because the ratio holds the distri butional shape constant over time. A median shift in this case simply represents the effect of inflation and is sometimes of little substantive interest. The second term can be interpreted as the effect of the shape shift and is usually of more interest. Here, deflation equalizes the medi ans of the reference and comparison distributions. As a result, the numer ator and denominator differ only in the shape of the earnings density, which is allowed to change over time (the basis of fig. 3 below).
This simple decomposition can easily be extended to the more compli cated betweengroup analyses that form the backbone of the article. We begin by defining the relative distribution of women's to men's earnings over time. Let
mt, w, = men's and women's earnings in year t, respectively;
m,d, wp = for both groups, earnings deflated by the men's
median ratio, median(m,)lmedian(m,);
g,(.) = the density function for men's earnings in year t;
h,(.) = the density function for women's earnings in year t;
g,d(.),h,d(.) = density function for each group based on its earnings
deflated by the men's median ratio;
hyd(.) = density function for women's earnings deflated by the
women's median ratio, median (wt)/median (zoo).
Then, with men as the reference group and women as the comparison group, the distribution of women's earnings relative to men's at time t is given by
ht(mt)
Rd(w,, m,) =
gt(mt)
This equation gives the relative distribution that, in its decile version, is graphed in figure 2. The last equality follows because the relative density of the two earnings distributions is unchanged when both (and their common scale, m,) are multiplied by the same constant. Here, both were deflated by the men's median ratio.' This step allows us to compare real earnings, net of the general effect of inflation. Note, however, that we do preserve changes over time in the ratio of women's to men's median earnings.
Using the same logic as the withingroup decomposition, the relative
In general, the relative distribution is invariant to any monotonic transformation, while the components of the decomposition are invariant to multiplicative transforma tions.
distribution in equation (4) can be decomposed into a series of terms isolating median and shape shift effects, with a few additional terms needed for rescaling.
The first term in the expression, h,(m,) l g,(m,), is the constant term: the initial relative distribution of women's to men's earnings. The remaining terms are components of change, multiplicatively acting on the initial distribution to yield the distribution at time t.
The second term, the median ratio effect, isolates the effects of the relative gains women made in median earnings. It takes the women's density deflated by the men's median ratio, hf, and compares it to the women's density deflated by their own median ratio, hYd, thus holding the basic distributional shape, h,, constant. The third and fourth terms represent the effect of the female and male shape shifts, respectively. They compare withingroup deflated densities at time t to those at time 0 and therefore isolate changes in shape for each group (analogous to the withingroup shape shift effects identified in eq. [3] above).
The two terms on the second line of the decomposition are rescaling effects. Mathematically, they rescale the baseyear earnings (m,) used in the constant term into the currentyear mediandeflated dollars (mf) that are used in the reference distribution. In our application, however, these ratios also serve to isolate another part of the male shape shift effect (note that the density functions and earnings medians are held constant within each term). For the purposes of this article, we combine these two rescaling terms with the male shape shift term to represent the total male shape shift effect.
For notational simplicity, we denote the multiplicative coefficients at earnings level m, at time t by
the median ratio shift coefficient; the female shape shift coefficient; and
the male shape shift coefficient. Then the decomposition from equation
(5) can be reexpressed as
Rd(wt, m,) = Rd(w,, m,) x RM(m,) x FS(mt) x MS(m,). (9)
In words, the relative distribution of women's to men's real earnings at time t can be expressed as the product of the initial relative distribution at time 0, changes in the ratio of women's to men's median real earnings, changes in the shape of the women's earnings distribution, and changes in the shape of the men's earnings distribution.
The last three components, therefore, jointly generate the change from the initial time 0 relative distribution to the time t relative distrib~tion.~ The coefficients themselves have a simple interpretation. They act as a multiplier, either increasing or decreasing the fraction of women at each level of the men's earning scale over time, depending on whether the value of the coefficient is above or below one. In terms of value, the coefficients can be interpreted as the percentage change in the fraction of women at each earnings level.
There is one other decomposition approach in the literature that is worth briefly mentioning here for comparison. Like the approach pre sented above, it is aimed at separating the effects of changes in average wages from changes in the shape of the wage distribution. The method was developed by Juhn et al. (1991) in the context of earnings differences between blacks and whites and has been applied recently to the gender earnings gap by Blau and Kahn (1994). This approach uses linear regres sion to partial out a series of terms representing changes in mean human capital, mean returns to that capital, changes in the mean residual earn ings gap between men and women, and changes in the standard deviation of the men's residual earnings variation (interested readers are referred to the articles cited above for a more detailed explication).
While taking account of the human capital component is an important feature of this approach, there are two problems with this decomposition from our perspective. First, it works only with average differences. Even though the residual wage distribution and the men's standard deviation
Dividing eq. (9) on both sides by the first component on the right side makes this clear.
are included, these distributions are summarized into the average residual wage gap and a single number summary of the dispersion in the men's residual wage distribution, respectively. This approach makes it impossi ble to examine how changes in the distributions affect men and women at different levels of the earnings scale, and, as will be shown below, being able to view the effects at different levels is a critical issue. Second, this decomposition does not separately identify and estimate the effects of the male and female shape shifts. Instead, the two are summarized and combined in the third term, which simply reflects the changes in the mean residual wage gap multiplied by the men's standard deviation. This approach confounds the two shape shifts, again removing the level of detail needed to answer the question of interest here: How have sex differences in earnings polarization, upgrading, and downgrading combined to produce the net changes in women's relative economic status?
Additive Decomposition While parsimonious, the multiplicative decomposition in equation (9) is often not as useful as one might wish. The problem is that a coefficient of 1.6 clearly has a stronger absolute effect on an earnings level with a high percentage of women than on one with a small percentage. Taken alone, the multiplicative coefficients do not indicate the absolute amount of change generated. An additive decomposition is therefore necessary. Its derivation is presented in the appendix, and we simply present its final form here:
Rd(w,, m,) Rd(wo, m,) = RM(m,) + FS(m,) (10)
+ MS(m,) + ZNT(m,).
Note that there is now an additional term, a residual interaction effect. The interaction exists because each coefficient multiplies the effects of the others in the multiplicative version of the decomposition (see eq. [9]), and it is usually quite small in magnitude. In the analyses below, all results will be given in their additive form.
A CLOSER LOOK AT THE GENDER GAP
We begin with the type of trend that most research on the gender gap has dealt with to date. Figure 1 shows the ratio of women's to men's median yearly earnings between 1967 and 1987, for white and black
Year
+White Women Black Women
FIG.1.Ratio of white and black women's median earnings to white men's median earnings, 196787.
women separately.7 The unprecedented closing of the gender gap is clearly evident for both black and white women. The race difference, however, is equally striking. Because of the prevailing focus on white women, most analysts pinpoint the 1980s as the watershed for women's recent economic gains. Yet it is clear that for black women the strongest gains actually occurred during the 1970s. And the net increase in the median ratio for black women was much stronger than for white women, a rise of 17 percentage points versus six percentage points.
While suggestive, the median trends perhaps raise more questions than they answer. For example, this plot gives us no idea of where a wellpaid woman worker would rank on the men's scale or what proportion of women earn less than the average white male. And while both of the median ratios increased, we do not know whether women at all earnings levels benefited to the same degree from this increase. Such questions are critical and can only be answered by shifting the focus from median or average earnings to the full distribution of earnings.
' Median, rather than average, earnings will be used throughout this article. We have chosen 1967 to be the reference year for all analyses. Using 1967 as the reference year is common when using the Current Population Survey series, because that year is the first that permits exact tracking of the earnings measures over time.
Figure 2 provides the starting point for such an approach. Shown is
the relative distribution of women's to men's earnings from 1967 to 1987,
in panel A for white women and panel B for black women. The relative
distribution is formed as follows: White men's earnings constitute the
reference point and, in each year, are divided into even deciles, each
containing 10% of male earners. Decile 1 is the lowest earnings decile,
decile 10 the highest. The cutoff points defining these deciles are re
corded, and then the percentage of women earners falling into each decile
is tracked over time.
The immediate and striking impressions conveyed by the graphs are that women earn dramatically less than white men and that this inequity has changed little over the past two decades. The great majority of the women are concentrated in the lower tail of the men's distribution. This pattern is especially pronounced for black women. In 1967 48% of white women and a staggering 71% of black women fell in the bottom decile of the men's distribution. For both races, more than 90% of the women earned less than the median male worker (the cumulative sum of all those in levels 15). By 1987 this pattern had changed somewhat, but more than 80% of women still earned less than the median male worker and over onequarter remained in the bottom decile of the men's distribution. The absence of women in the upper tail of the men's earnings distribution is just as striking. The number of women in the top decile does not reach 2%.
Moreover, while the median ratios suggested that women made real progress over the past two decades, the relative distributions make clear that this progress was limited to women at the bottom of the earnings distribution. In both panels, more than threequarters of the change over time occurred below the male median, half of it in the lowest decile alone. This finding runs counter to the usual intuition about how women's economic progress works. When presented with evidence of a closing earnings gap, most observers think of women making strong in roads into wellpaid jobs that men had dominated. This distributional approach, however, reveals that white and black women's gains occurred largely at the bottom of the relative earnings scale.
The median ratios in figure 1 are thus to some extent misleading: they suggest an acrosstheboard gain for women at all earnings levels, which we have just seen is not the case. Certainly, the convergence in median earnings is an important part of the story, but it is not the only one. Because we have shifted our focus to the level of distributions, two other changes come into play in generating the trends in figure 2. Not only may the average or median level of an earnings distribution change over time, but so may its shapethat is, the degree and nature of earnings inequality. We must therefore consider whether the distributions of men's
A. White Women B. Black Women
Men's Earnings Deciles
Fig. 2.Relative distribution of white and black women's earnings to white men's earnings, 196787
and women's earnings, viewed separately, have undergone any changes in shape.
At the outset, we cited research that documented that the inequality in men's earnings increased over the past two decades. Panel A of figure 3 reflects these findings. For this graph, the 1967 earnings of white men were divided into deciles, each containing 10% of earners. Again, decile 1 is the lowest earnings level, decile 10 the highest. Men's earnings in each of the following years were deflated,' then the percentage of men falling into each 1967 decile was tracked over time.9 Clearly, both the proportion of men with high earnings and the proportion with low earn ings increased over the 21year period, with growth at the bottom of the distribution especially pronounced. Conversely, the proportion in the middle of the distribution decreased. It is this strong and consistent polar ization in earnings that has generated the widespread public concern over a declining middle class.
Panel B of figure 3 shows that the earnings of white women underwent a different set of changes during the same time span. Their earnings at first actually increased, with growth in the middle and top of the distribu tion. A distinct polarization then set in and continued unabated. This finding may surprise readers who have seen evidence only of the closing gender gap in average earnings. The growing inequality in earnings among white women has largely been overlooked but would seem an important dimension on which to measure whether women are in fact doing better (see Smith [I9911 for one of the few studies to address the issue). By 1987 the polarization in white women's earnings was marked but still not as strong as that evidenced by white men.
Black women show a remarkably different pattern (panel C of fig. 3). They initially experienced a convergence toward the middle of the earn ings distribution. The dominant trend, however, is a pronounced bifurcation in earnings during the 1980s. The lowest deciles in particular (some of which are obscured in the graph) grew strongly during this decade. Conversely, growth at the top of the distribution was negligible. This pattern stands in clear contrast to that of white women, whose earnings polarized in a balanced way between 1979 and 1987, with the bottom fifth growing by 3.5 % and the top fifth by 4.5 %. During the same
Earnings were deflated with the ratio of the 1967 median to the current year's median. This is an internally consistent deflation method that sidesteps the pitfalls of the census deflators (e.g., the Consumer Price Index and its variants). It is not influ enced by changes in average earnings and is technically congruent with the decomposi tion technique to be used later in the article.
Choosing an alternative reference year would change the view provided by these decile graphs, but it would not change the central trends or yeartoyear comparisons that one would infer from them.
TABLE 1
CHANGEIN THE DISTRIBUTIONOF WHITE WOMEN'S TO WHITE MEN'S EARNINGS, 196787
1 (low) ......................................
9 .............................................
10 (high) .....................................
time span, however, black women saw a 6.7% rise in the bottom fifth and a decline of 1.7% in the top fifth. Thus we are again faced with a worrisome trend. Figure 1 suggests that black women have rapidly gained on white men, which is encouraging. Yet the growing number of lowwage earners among black women cannot but cause concern.
DECOMPOSING THE CHANGE IN WOMEN'S ECONOMIC STATUS
In sum, the converging gender gap has been accompanied by a much more complex set of shifts at the distributional level. The remainder of the article is dedicated to tracing the relationship between these shifts. Substantively, we return to the question raised at the outset: Did women gain economic ground on men largely because they closed the gap in median earnings and moved into highincome jobs, as is commonly held? Or did the strong growth in earnings inequality among men play a role as well? Technically, the problem is one of linking distributional and median shifts and evaluating the impact of each on women's relative economic gains. Specifically, we want to use the method outlined above to decompose the observed changes in the relative distributions shown in figure 2.
White Women
While the graphs in figure 2 show an entire 21year series, the simplest way to proceed is to focus on the overall net change between 1967 and 1987. A numerical summary of this change is shown in table 1. The
decomposition focuses on explaining the gains or losses in each of the 10
deciles in the last column of the table. While the methods section above
presented the mathematical decomposition, we describe it here at a con
ceptual level. The change in the relative distribution of women's to men's
earnings can be broken down into three sources:
Change in the median ratio over time.This component measures
how much ground women would have gained if only the convergence in
women's and men's median earnings had taken place.
Change in the shape of the men's distribution over time.This component measures how much ground women would have gained if only the polarization in men's earnings had taken place.
Change in the shape of the women's distribution over time.This component measures how much ground women would have gained if only the degree of inequality in women's earnings had changed.
In the decomposition's additive form, these three components, plus a small interaction term, sum to produce the changes in the last column of table 1. We are interested in which component has contributed the most to the shifts in each of the 10 deciles. If only the convergence in median earnings has a strong effect, then we have little of substance to add to current research. But if any of the changes in earnings inequalityeither men's or women's or bothmake themselves felt, then the story of women's recent economic gains becomes more complicated.
Figure 4 summarizes the results of the additive decomposition for white women. The solid bars merely show the changes that we are trying to explain, from the last column of table 1. Each of the lines then represents one of the three components in the decomp~sition.'~
As a guide to inter pretation, we begin with the first and lowest decile. We know that the percentage of white women with earnings in this decile fell by roughly 2 1 percentage points between 1967 and 1987. The polarization in men's earnings (the men's shape shift) was clearly the strongest factor in this drop, singlehandedly contributing 15 points to the decline. The conver gence in median earnings (the median ratio shift) added a further decline of 9 points. The shape shift of white women's earnings yields a prediction that points in the wrong direction: it would have resulted in an increase in this bottom decile if no other changes had taken place.
Nearer the middle of the distribution (i.e., deciles 2, 3, and 5), the men's shape effect is clearly the only substantial one that points in the accurate direction and thus was the main contributing factor. Note that the magnitude of the effects here are smaller than they were in the first
lo Plus the interaction terms for deciles 110, which are, respectively, 1.88, .69, 1.13, .20, .61, .64, .48, .11, .04, and .lo.
............................
................................................
.............................................................
*
a" .

...............................................................
2
'b
4
20
!?. +Changein medianratio
t .........................................
White women's shape effect
White men's shape effect
30 .........................................
t
FIG. 4.Decomposition of change in the relative distribution of white women's to white men's earnings, 196787.
decile, because the empirical changes being decomposed are themselves smaller.
In the upper half of the distribution, we see a different pattern. Changes in men's earnings no longer dominate. Rather, the convergence in median earnings and the polarization in white women's earnings now become the dominant factors, especially in generating the gains in deciles 7, 9, and 10.
In sum, the strong polarization in men's earnings played an important roleit helped to push white women out of the lower earnings levels toward the middle of the distribution. The convergence in median wages also added to the movement of white women out of the lowest earnings decile but otherwise made itself felt in the middle of the distribution. Because we know that the gender gap has narrowed in the past decade, this finding should come as no surprise. Last, the growing inequality in white women's earnings during the 1980s was beneficial to their economic standing relative to men. If it had occurred alone, this polarization would have been detrimental and actually increased the number of women who were earning at the bottom of the men's distribution. But because the polarization took place in the context of the other two changes, only its
TABLE 2
% OF BLACKWOMENIN MEN'S EARNINGS DECILES
CHANGEFROM  

MEN'S EARNINGS DECILE  1967  1987  1967 TO 1987 
1 (low) ................................. 
9 .............................................
10 (high) ..................................
positive effects manifested themselves, boosting the number of women who were earning at the highest levels.
It is important, however, to stress that the striking polarization in men's earnings had the strongest effect overall because it was concen trated in the lowest deciles, where women saw the greatest movement. Between 1967 and 1987 close to half (48%) of the total change that oc curred in the relative distribution was generated by the rising inequality in men's earnings, with the convergence in median earnings adding 26% and the rising inequality in women's earnings 18%.
Black Women
We now turn to an analysis of black women's relative gains. The experi ence of white and black women has generally been similar, in that both made economic progress over the past 20 years. But we have also noted a number of specific and important differences. Overall, black women's gains have been more dramatic. Recall that their earnings gap closed by 17 percentage points, which is nearly three times the sixpoint reduction in the white women's gap. The shift in their relative distribution was also more marked. Contrast the steep drop of 37% in the bottom men's decile with that of 2 1% for white women (tables 2 and 1, respectively). And last, changes among black women also looked different, with a more pronounced growth in lowwage earners that was not offset by growth in highwage earners.
...........................................................
............................................................
..................................
Black women's shape effwt
.....................................
White men's shape effect
...............................................................
...............................................................
FIG. 5.Decomposition of change in the relative distribution of black women's to white men's earnings, 196787.
It is therefore an open question whether these race differences will translate into a different configuration of effects in the decomposition. Figure 5 presents the additive decomposition results for the relative distri bution of black women's to white men's earnings." Again, the focus is on explaining the changes in the last column of table 2, represented in the graph by the solid bars. The massive movement of black women out of the bottom decile is clearly the largest change over the 2 1year period. The convergence in median earnings has by far the strongest effect in this decile. It contributes a drop of 32 percentage points all by itself, which should come as no surprise considering the magnitude of that convergence. The polarization in men's earnings, so important for white women, does play a role here but in the end is dwarfed by the median effect.
For the remaining deciles, the clear impression is that the convergence in median earnings has driven most of the observed shifts, almost exactly mirroring them. With the exception of the second decile, median conver gence served to pull black women out of the lowest earnings levels and
l1 The interaction terms for deciles 110 are, respectively, 4.47, 1.42, 1.75, 1.02, .79, .13, 4.38, 2.24, .04, and .20.
push them toward the middle levels. The growing inequality in men's
earnings also contributed a good amount of the change in deciles 25 but
again simply does not compare in strength.
The effects of changes in black women's earnings make an interesting and important point. We know that the number of lowwage earners grew strongly among black women during the 1980s. This downgrading had a noticeable impact in the second decile, where it roughly doubled the presence of black women and effectively put a brake on the massive exodus out of the first decile. Unlike white women, then, black women have not been insulated from some of the more negative earnings trends of the 1980s.
All of these effects are well summarized by a breakdown of the overall contributions to the observed changes. Between 1967 and 1987 close to twothirds (63%) of the change in the relative distribution was generated by the marked convergence in median earnings, with the men's shape shift adding 23% and the black women's shape shift only 8%.
In sum, white and black women both experienced important economic gains over the past two decades, but there was a marked difference in how those gains were generated. In particular, the growing inequality in white men's earnings played the strongest role for white women. But for black women, it was the substantial convergence in median earnings that proved decisive.
DISCUSSION
There are clearly several stories to be told here. When one looks at the full distribution of women's earnings compared to men's earnings over time, the overriding impression is one of substantial and persistent in equality. More than 80% of women still earned less than the median male worker in 1987, and the gains that women did make over the 20 years from 1967 to 1987 largely consisted of movement out of very low earnings levels, not into high earnings levels. These realities shed a sober light on the current pronouncements that gender equality is imminent.
We therefore reiterate the point made at the outset. An exclusive focus on median or average earnings levels can be seriously misleading when one analyzes inequality between two groups. Our analysis indicates that the convergence in median earnings is only one element in the changing economic welfare of women. It is not always the most important.12
l2 At this point the reader should recall that the analyses are based on samples of fulltime, yearround workers. Expanding the analysis to parttime and partyear workers would likely change some of the results, given that our focus is on gender differences and that lowwage, contingent work has grown disproportionately among women.
Just as critical is the fact that men's earnings worsened and became more unequal during the 1970s and 1980s. In general, the striking polar ization of men's earnings served to push significant numbers of women out of the lowest earnings levels. Because the lowest levels are precisely where most of women's earnings were concentrated, we conclude that trends in men's earnings contributed significantly to women's relative progress. In the case of white women, this effect clearly dominated and singlehandedly accounted for 48% of the gains made. For black women, the effect was important as well, although to a lesser degree.
Surprisingly, the inroads that women made at the higher earnings levels were small and played a weaker role. To many, however, these are the important gains, because they indicate entrance into the most prestigious and best rewarded occupations. Thus it is of interest that for white women the distributional change in their earnings, more than the increase in the median level, generated these gains. In other words, it is largely because white women's earnings polarized and became more un equal during the 1980s that some of them were able to move into the upper rungs of men's jobs. The women's distributional shift, then, is very much a twoedged sword. As Shelley Smith (199 1, p. 117) argues, we must recognize that "the kinds of changes that have enhanced the economic position of employed women have also produced conditions that increase earnings inequality among women."
These findings also force us to recast optimistic predictions about the future of gender inequality. Contrary to prevailing opinion, our decompo sition analysis would predict a slowing of the convergence of women's on men's wages.13 There are several reasons for this prediction.
For white women, the growing inequality in men's earnings over the past two decades was highly beneficial. But this growth is not likely to continue at the same rate in the future, because one of its causes was deindustrialization, which is by now mostly played out. We also know that white women's earnings began to polarize in the 1980s, and this trend, by contrast, does not show any sign of lessening. In fact, it appears to be escalating, and we can expect a significant growth in lowwage women workers as a result. Thus the irony: the fact that men's earnings worsened so strongly in the recent past may well have insulated white women from feeling the detrimental impact of growing inequality in their own earnings. In coming years, however, this insulation will lessen and the negative consequences will probably set in.
For black women also, the factors responsible for their economic prog ress are unlikely to continue. The most important of these was the conver
l3 In fact, the most recent data indicate that the gender wage gap has stagnated since 1990 (Nomani 1994).
gence in median earnings (in contrast to white women), but it was concen
trated in the 1970s and caused mainly by a dramatic movement of black
women out of domestic work. Fully 36% of black women were employed
as private household workers in 1960, but by 1987 only 5% were (Blau
and Beller 1992). Clearly, this occupational shift has now stabilized,
opening the door far the recent growth in lowwage black women workers
to make itself felt.
The argument, then, is that changes in earnings among women workers
will determine the extent of their future economic gains. These changes
have so far been detrimental: again, white women's earnings are polariz
ing, black women's are being downgraded. Such trends could well put
a brake on women's progress if they continue, which is likely given
several longterm transformations in the service sector.
Most important is the possibility that labormarket segmentation is increasing and that it is doing so along race and gender lines (e.g., Har rison 1994; Pfeffer and Baron 1988). Likely causes include the ascent of lowwage, highturnover business strategies and in particular the increas ing reliance on involuntary parttime labor; the formation of a protected, specialized, and wellpaid stratum of knowledge workers; and the flat tening and "rationalization" of organizations in order to cut costs and improve productivity. The upshot is that mobility between core and pe riphery starts to break down, and thus entrylevel service workers may find themselves increasingly cut off from traditional career ladders. More over, history suggests that gender often plays an important role in this type of transformation.14 At a minimum, we know that women are still concentrated in the lower rungs of the service sector. Any widening of the gap between good jobs and bad jobs will therefore disproportionately affect them and slow their relative economic progress.
For black women, that progress may even be reversed, given the con spicuous lack of growth in highincome workers and the equally promi nent increase in lowwage earners. At the very least, our analyses have shown that the experience of black women has been distinct on a number of levels from white women's. Research in this area must therefore recog nize that race forms an important independent dimension of gender in equality and that it is no longer defensible to conduct analyses of all women combined without any race distinctions.
Finally, the methods developed in this article represent an important contribution in their own right. Recent years have seen an increase in the use of distributional approaches to research on earnings in labor
l4 For example, the restructuring of department store jobs during the 1960s and 1970s resulted in an emerging highwage managerial layer filled mainly by men and a low wage semiskilled sales layer made up of women (Bluestone et al. 1981).
economics and sociology. Often, however, the measures used are rudi mentary (Smith and Welch 1989), or they do not fully distinguish the separate elements of change outlined in this article (Blau and Kahn 1994). The relative distribution and its decomposition afford the researcher a more comprehensive set of tools for analyzing inequality between groups, be they defined by race, age, skill level, industry, or occupation. More generally, these tools are part of an emerging framework in which the earnings of distinct labormarket groups are linked to complex changes in the overall structure of wages.
APPENDIX
The Relative Distribution
The notation used throughout this article has been designed for simplicity and brevity. The precise definition of the relative distribution unpacks this notation and makes explicit the rescaling from the original earnings scale to the percentile rankings that form the basis of the relative distribu tion. Let Yo and Yt be random variables representing earnings in the reference group and comparison groups, respectively. We define the cumulative relative distribution of Y,relative to Yoas
for 0 < r r 1, where Fo1(r) = inf,{y :F,(y) rr}, the inverse cumulative distribution function of Yo, and r is the percentile rank of y in the earnings distribution, F,. Note that G is defined even when the earnings distribu tions are not continuous. The relative distribution in equation (1) is then the corresponding probability density function defined as
Defining the Additive Decomposition
The goal is to take the total change in the relative distribution from year 0 to year t, namely, Rd(wt, m,) Rd(w,, m,) , and decompose it into the additive change contributed by each component. Using the general notation for the multiplicative coefficients, ki(m,) , the additive effect for each component is most naturally defined as
C[ki(m,)l = Rd(wo,mo) x [ki(mt) 11.
In this form, C[k,(m,)] represents the percentage point change in the initial Gender Gap in Earnings
relative distribution at earnings level m, that is due to component i at time t. To form a complete decomposition, the effects must sum to the total change from the initial to the current relative distribution:
The three main effects defined by equation (9) sum instead to
The difference is a residual "interaction effect" due to the fact that each coefficient multiplies the effects of the others (see eq. [9]). This interaction effect is most easily defined by subtraction:
REFERENCES
Blau, Francine D., and Andrea H. Beller. 1992. "BlackWhite Earnings over the 1970s and 1980s: Gender Differences in Trends." Review of Economics and Statis tics 74 (2): 27686.
Blau, Francine D., and Lawrence Kahn. 1992. "The Gender Earnings Gap: Learning from International Comparisons." American Economic Review (Papers and Pro ceedings) 82 (2): 53338.
. 1994. "Rising Wage Inequality and the U.S. Gender Gap." American Eco nomic Review (Papers and Proceedings) 84 (2): 2328.
Bluestone, Barry, Patricia Hanna, Sarah Kuhn, and Laura Moore. 1981. The Retail Revolution: Market Transformation, Investment, and Labor in the Modern Depart ment Store. Boston, Mass.: Auburn House.
Danziger, Sheldon, and Peter Gottschalk, eds. 1993. Uneven Tides: Rising Inequality in America. New York: Russell Sage.
Even, William, and David A. Macpherson. 1993. "The Decline of PrivateSector Unionism and the Gender Wage Gap." Journal of Human Resources 28 (2): 29396. Handcock, Mark. 1995. "Imputation Methods for Coarse Income Data." Technical report. New York University, Department of Statistics and Operations Research.
Harrison, Bennett. 1994. Lean and Mean: The Changing Landscape of Corporate Power in the Age of Flexibility. New York: Basic.
Harrison, Bennett, and Barry Bluestone. 1988. The Great UTurn: Corporate Restruc turing and the Polarizing of America. New York: Basic.
Institute for Women's Policy Research. 1992. The Wage Gap: Women's and Men's Earnings. Briefing paper. Institute for Women's Policy Research, Washington,
D.C.
Juhn, Chinhui, Kevin M. Murphy, and Brooks Pierce. 1991. "Accounting for the Slowdown in BlackWhite Convergence." Pp. 10743 in Workers and Their Wages, edited by Marvin Kosters. Washington, D.C.: American Enterprise Institute Press.
Karoly, Lynn. 1993. "The Trend in Inequality among Families, Individuals, and Workers in the United States: A TwentyFive Year Perspective." Pp. 1997 in Uneven Tides, edited by S. Danziger and P. Gottschalk. New York: Russell Sage.
Kuttner, Robert. 1983. "The Declining Middle." Atlantic Monthly 252:6072. Lorence, Jon. 1991. "Growth in Service Sector Employment and MSA Gender Earn ings Inequality: 19701980." Social Forces 69 (3): 76383. Marini, Margaret M. 1989. "Sex Differences in Earnings in the United States." American Sociological Review 15:34380. Montgomery, Edward, and William Wascher. 1987. "Race and Gender Wage In equality in Services and Manufacturing." Industrial Relations 26 (3): 28490.
Morris, Martina, Annette Bernhardt, and Mark Handcock. 1994. "Economic In equality: New Measures for New Trends." American Sociological Review 59: 20519.
Nasar, Sylvia. 1992. "Women's Progress Stalled? Just Not So." New York Times (October 18), pp. 1, 10.
Nomani, Asra. 1994. "Labor Letter." Wall Street Journal (October ll), p. Al.
O'Neill, June, and Solomon Polachek. 1993. "Why the Gender Gap in Wages Nar rowed in the 1980s." Journal of Labor Economics 11:20528.
Pfeffer, Jeffrey, and James N. Baron. 1988. "Taking the Workers Back Out: Recent Trends in the Structuring of Employment." Research in Organizational Behavior 10:257303.
Smith, James P., and Finis Welch. 1989. "Black Economic Progress after Myrdal." Journal of Economic Literature 2751964. Smith, Shelley A. 1991. "Sources of Earnings Inequality in the Black and White Female Labor Forces." Sociological Quarterly 32 (1): 11 738.
Tienda, Marta, Shelley A. Smith, and Vilma Ortiz. 1987. "Industrial Restructuring, Gender Segregation, and Sex Differences in Earnings." American Sociological Re view 52:195210.
Wellington, Allison J. 1993. "Changes in the MaleIFemale Wage Gap, 197685." Journal of Human Resources 28:83141.
Comments