Do Male-Female Wage Dierentials Reect Dierences in the Return to Skill?

Cross-City Evidence From 1980-2000

Abstract Over the 1980s and 1990s the wage dierentials between men and women (with similar observable characteristics) declined signicantly. At the same time, the returns to education increased. It has been suggested that these two trends may reect a common change in the relative price of a skill which is more abundant in both women and more educated workers. In this paper we explore the relevance of this hypothesis by examining the cross-city co-movement in both male-female wage dierentials and returns to education over the 1980-2000 period. In parallel to the aggregate pattern, we nd that male-female wage dierentials at the city levels moved in opposite direction to the changes in the return to education. We also nd this relationship to be particularly strong when we isolate data variation which most likely reects the eect of technological change on relative prices. We take considerable care of controlling for potential selection issues which could bias our interpretation. Overall, our cross-city estimates suggest that most of the aggregate reduction in the male-female wage differential observed over the 1980-2000 period was likely due to a change in the relative price of skill that both females and educated workers have in greater abundance.

Comments welcome. The authors acknowledge the helpful comments of seminar participants at the University of British Columbia, the Federal Reserve Bank of San Francisco, and Dartmouth College. Mark Doms was heavily involved in the early stages of this project, and we are indebted to him for his substantial input. All errors are our own.

Introduction

Over the 1980-2000 period the US wage gap between men and women with similar characteristics decreased signicantly. Since 2000 this wage gap has stayed quite stable. There are many proposed explanations for the observed decrease in the gender wage gap, including increased positive selection of women into the labor market (Mulligan and Rubinstein, 2008), an improved match between actual and potential measures of experience (ONeill and Polacheck, 1993) among other consequences of greater labor force attachment potentially induced by increased access to birth control (Bailey, Hershblein, and Miller, 2012) and decreased discrimination.1 One especially intriguing observation is that the gender wage gap has mirrored rather closely movements in the return to education over this period (see Figure 1), with the gender gap decreasing when the return to education increased.2 This has had lead some researchers, most notably F. Welch (2000), to conjecture that the two patterns may be driven by a common underlying force. According to this conjecture, men and women (with similar education) bring to the market dierent bundles of skills, as do more educated relative to less educated individuals. When methods of production change drastically such as with the introduction and diusion of the PC this changes the price of dierent skill attributes which in turn induces movements in the gender wage gap and education wage gaps, since both gaps compare the value of skill bundles. To be more precise, if individuals are viewed as bringing to the market both soft-cognitive skills (or interpersonal skills) and hard-motor skills (or brawn), and if women and more educated workers are both relatively more endowed is such softer skills, then an increase in the relative price of soft versus hard skills should cause the male-female wage gap to decrease at the same time as the return to education increases. While some evidence lends support to this view, the literature on gender gaps appears to remain skeptical of its relative importance in explaining the decrease in the wage gap observed over the 1980s and 1990s, since the claim is backed-up foremost by time-series evidence (which could easily be spurious) or else does not directly analyze gender wage gaps.3 The object of this paper is to more thoroughly esh out the empirical content of
ONeill and Polacheck (1993) also attribute some of the decline to a decrease in blue-collar wages. Gender discrimination is often mentioned, but, as it is dicult to quantify, its importance is not usually empirically assessed. For example, Blau and Kahn(2006) say that the larger unexplained portion of the decline in the wage gap in the 1980s compared to the 1990s may be evidence of a contribution of decreased discrimination, though they also consider other interpretations. 2 The regression of the gender wage gaps in Figure 1 on the college-high school wage gap gives -0.625 for the high school level gender gap and -0.787 for the college level gender gap. 3 The evidence in support of this view includes the time series correlation between the male-female wage gap and the returns to education (such as in Figure 1; also Welch, 2000, and Fortin and Lemieux, 2000) the
1

this conjecture and then explore its implications for cross-city observations on gender gaps and returns to education. The idea that a change in the relative price of soft/cognitive skills versus hard/motor skills could be behind some or all of the decrease in the gender wage gap observed over the 19802000 period is a rather straightforward proposition. However, empirically evaluating its relevance is hindered by the fact that the relevant relative price is not directly observable. As we will show using a simple two attribute model, when female workers and more educated workers have relatively more soft-cognitive versus hard-motor skills than other types of workers, then the relative price of soft versus hard skills plays the role of a common latent factor that drives in opposite directions the gender gap and the return to education. The main idea of this paper is to examine the quantitative relevance of this unobserved factor by examining how city level measures of the gender gap and the return to education change in response to forces that likely caused changes in the local price of these skills. If this common factor is present, then the gender gap and the return to education should systematically comove in opposite directions in response to forces that aect the relative price of soft versus hard skills. The main challenge in implementing this simple idea is nding city level factors that plausibly aects city level relative prices of skill.4 To make headway in identifying such factors, we refer to the capital-skill complementarity literature which argues that 1980 to 2000 was a period where the introduction of new capital equipment most notably the PC drastically changed the relative price of certain skills since it acted as a substitute for many motor or hard skills, and a complement for more cognitive skills.5 Based on this view, one may consider regressing
decline in wages in jobs which require motor skills and the rise in wages in jobs which require cognitive skills (Baccolod and Blum, 2010), and the correlation across industries between the change in female employment share and the adoption of computers, especially in blue collar jobs (Weinberg, 2000). Black and Spitz-Oeners (2010) nding that a majority of womens relative wage increases in Germany between 1979 and 1999 can be accounted for by a large relative shift away from routine cognitive tasks (which Autor, Levy, and Murnane (2003) found was associated with computerization) is also consistent with the decline in the wage gap having a technological origin. 4 To identify the relevant co-movement, these factors must simultaneously be uncorrelated with other residual factors which may aect either the gender wage gap or the returns to skill. 5 One view of recent technological change is that the PC is a revolutionary technology (Caselli, 1999) of discretely higher skill intensity than previous technology; its adoption is therefore depends on comparative advantage: the relative price of (and therefore supply of) skill. This is empirically supported in Beaudry and Green, 2003, 2005; and Beaudry, Doms, and Lewis, 2010. Autor, Levy and Murnane (2003) also model computer adoption as responding to skill ratios, and a version of their model will be the main model of production we consider in this paper. Another view is that PCs are the latest example of ongoing improvements in the quality of capital that favor skilled workers which perhaps goes back as much as a century (Goldin and Katz, 2008).

city level observations of both changes in gender wage gaps and in returns to education on measures of technological adoption such as the local use of PCs and see if these two wage dierentials move in opposite directions. While we will report the results of such an exercise, such an approach is potentially problematic as technological adoption is itself an endogenous process. For this reason, we also use insight from the endogenous technological adoption literature (similar to that used in Beaudry, Doms & Lewis, 2010, hereafter BDL.) to illustrate how local labor market conditions prior to the introduction of the PC can be used as instruments for technology adoption. The main idea is that, if PCs complement soft skills and substitute for hard skills, then they should be adopted most strongly in localities where soft skills were initially relatively abundant. Hence this idea suggests that the city level change in the gender wage gap and the return to education should respond in opposite directions to measures of a citys pre-PC-era relative abundance of these skills. Using data from the 1980 and 2000 Censuses of Population, we begin by examining whether the gender wage gap and the return to education moved in opposite directions across cities during this period. Interestingly, we nd that the cross-city evidence echoes the time series pattern of strong negative co-movement, thereby providing initial evidence supporting the conjecture that the gender wage gap and the return to education may be driven by a common latent force. However, based on OLS estimates, the cross-city co-movement is substantially weaker than that found in the time series. We then exploit the insights of the endogenous technological adoption literature in the presence of capital skill complementarity to explore instrumental variable estimates of this relationship. Again, this class of models suggest that cities where soft skills were most abundant pre-PC (in 1980) should be the same cities where PCs are adopted most rapidly, inducing a greater increase in the return to education and a greater decrease in the gender gap. We report evidence in support of each element of this process. Our main nding is that, once we isolate variation in the wage data more likely driven by changes in the relative prices of unobserved skills, we nd that the cross-city link between gender wage gap and the return to education to be very similar to that found in the time series. We then use the cross-city evidence to help answer our initial time series question. Our results suggest that most of the aggregate reduction in the male female wage dierential observed during the 1980-2000 period can be attributed to the change in a latent relative price of a skill which is more abundant in both female and more educated workers. The remaining sections of the paper are structured as follows. In Section 2 we present the simple theoretical structure which guides our approach to the data. The theory encompasses

two elements: rst, we clarify how gender gaps and returns to education are related in a two attribute model of wage through a common latent factor. Then we use insights from the literature on capital-skill complementarity and endogenous technological adoption to discuss the identication of the eects of the common latent factor. Section 3 presents the data used to examine determinants of the gender wage gaps and discusses implementation and identication issues. Section 4 presents our main results and Section 5 discusses their implication for aggregate changes. In the empirical analysis we examine in depth issues of selection that may bias our results as such issues are thought to be potentially very important in the behavior of the gender wage gap over our period of interest (see Mulligan and Rubinstein, 2008). Finally Section 6 oers concluding comments.

Theory

The idea we want to evaluate is whether changes in the gender wage gap and in the return to education may be driven by a common underlying force reecting the price of a skill which is relatively more abundant among women and more educated workers. In particular, the main hypothesis is that the role of this common factor driving the gender wage gap became most evident during the 1980-2000 period when technological change as reected in the diusion of PCs considerably changed relative prices of skills. The object of this section is to present a simple theoretical structure which will clarify how we can use crosscity variation in wage outcomes to examine the issue. As noted in the introduction, there are two distinct components which underly the theory. On the one hand, there is the notion that wages reect payments to bundles of skills. On the other hand, there is the idea that the diusion of PCs tended to increase the relative price of cognitive-soft skills because of its complementarity to such factors, while it acts as a substitute for more routine-hard skills. We now present each of these elements in turn in order to derive estimating equations and associated instrumental variable strategies.

2.1

The gender wage gap and the returns to education in a two attribute model

To begin, consider an environment where each worker brings to the market a two dimensional vector of skills. The two components will be referred to as cognitive-soft skill (denoted S )

and raw labor (denoted L). Individuals dier in the amount of each skill they possess.
S represent the amount of cognitive skill embodied in a a worker with education e Let eg L {e1 , e2 , ..., eN } and gender g {m, f }, and let eg represent the amount of raw labor embodied

in the same individual. For an individual in city c at time t his wage will be given by
S S L L Wegct = (eg wct + eg wct )egct ,

(1)

S L where wct and wct are the local prices of the cognitive skill and of raw labor respectively at

time t, and egct combines any systematic discrimination egt (that can potentially vary over time by gender and education) and a pure measurement error term egct ( egct = egt + egct ). For now, we need not focus on why people with dierent skills may cluster more in some locations than others. Instead, we can take the cross-city distribution of worker types as given and postpone a discussion of the related endogeneity issues. The main diculty with using equation (1) is that none of the right hand side terms are directly observable. Nonetheless, we can pursue some of its empirical implications by examining wage gaps across individuals. We begin with the male-female log wage gap at education level e, which we denote M F dif fect . From (1), this can be expressed as

M F dif fect = ln Wemct ln Wef ct S S L em wct = ln em + ln 1 + L L wL ef em ct ln

ln 1 +

S S ef wct L L ef wct

+ ln emct ln ef ct

S wct + ln emct ln ef ct , L wct

or, to simplify the notation, we can express it as:

(2)

1 1 em where e = ln , e = L
ef

S em L em

S ef L ef

S , Pct =

S wct , L wct

and ect = ln emct ln ef ct . Equation

(2) says that the cross-city dierences in the male-female wage gap depend on a common education group eect and varies across cities because of dierences in the relative price of 6

S skills, Pct . Similarly, the within gender wage gap between education levels ej and ei can also

be expressed as a function of the relative price of skills.

(3)

Eji dif fgct = ln Wej gct ln Wei gct

2 2 = ln Lj and jig = with jig

L e g

S e jg L e jg

S e ig L e ig

and eij gct = ln ei gct ln ej gct .

Equations (2) and (3) illustrate that in a two attribute model, both the gender wage gap and the returns to education are linked by the relative price of skills which acts as a latent common factor. Moreover, if we are willing to assume that at a given level of education men
1 have a smaller ratio of soft skills to raw labor (which implies e < 0) and that within gender 2 > 0), more educated workers have have relatively more soft skills (so if ej > ei , then jig

then we see that change in the relative price of skills P S will cause the gender wage gap and the returns to education to move in opposite directions. Although (3) and (2) still cannot be directly estimated, part of what we will exploit in estimation is changes in the prices of skills during the era of PC diusion. So taking dierences of (3) and (2) we get:
1 M F dif fec e PcS + ec

(4)

(5)

2 Eji dif fgc jig PcS + eji gc

Note that the error terms in (5) and (4) may not have a zero mean as they potentially contain changes in systematic discrimination (changes in egt ). If we substitue (5) into (4) to eliminate the unobserved skill price, we get the following relation between the gender wage gap and the returns to education:
1 1 e e E Dif f + eji gc ji gc ec 2 2 jig jig

(6)

M F dif fgec

1 2 Our conjectures about skill endowments imply e /jig < 0.

One of our main goals will be

to estimate (6) consistently, as this will be necessary to help evaluate the role of changes in skill price for explaining aggregate changes in the gender wage gap. However, even leaving aside the potential endogeneity of cross-city wage variation in (6), the fact that wage gaps are
1 2 likely measured with error implies OLS estimates of e /jig will be substantially attenuated.

In order to estimate (6) consistently, we will therefore need instruments for the returns to education that reect changes in the relative price of skill P S . For this reason, we now discuss how we can use insights from the capital-skill complementarity literature to nd such instruments.

2.2

Production and Endogenous PC Adoption

In order to discuss factors determining the relative price of skills, consider an environment where there is only one produced good and where prices reect marginal products. Initially, the good is produced using only the skills of dierent workers. Then we consider the introduction of a new capital good which is meant to capture the introduction of PCs. Our aim is to highlight how this aects the price of the two dierent skill attributes. For expositional simplicity, we follow Autor, Levy , and Murnane (2003) (hereafter, ALM) and model the economy after the arrival of computing technology with the following Cobb-Douglass production structure:
C Qc = A P c P Cc + Lc 1 Sc

(7)

where Qc represents aggregate output, Lc is the aggregate level of raw labor supplied by the dierent individuals hired in market c, Sc represents the aggregate amount of soft skilled hired in market c, P Cc represents the use of personal computers, and (0, 1). The only
C way in which this production function diers from ALM is the factor loading P c , which

we include to capture potential city-specic productivity dierences in the use of PCs.7 The important element of this technology is that PCs substitute for raw labor and complement the soft-cognitive skill. The results we exploit in what follows relies on this assumption but not on the particularly restrictive functional form given by 7.8
< 0 may vary across education groups, which we will allow for in some of our estimates, but we will restrict it to be constant in our main discussion. 7 In ALM, this production function represented many industries, each with dierent s. 8 See Beaudry, Doms and Lewis (2010) for a more general discussion.
6 1 2 e /jig

Now, consider a period before the arrival of PCs, which we will call t = 1980 to match our
1 C . This implies that = 0, so Qc = AL empirics below. We model this by setting P c Hc c

before the arrival of PCs, the relative price of soft skills versus raw labor is given by:
S Pc, 1980 = ln S wc, Sc,1980 1980 = ln ln L wc,1980 1 Lc,1980

(8)

= ln

lnsc,1980 1

where sc,1980 =

Sc,1980 . Lc,1980

Equation (8) simply indicates that the relative price of soft skills

prior to the introduction of PCs was negatively related to the local abundance of soft skills.9 After the arrival of the P C , which we assume is available in all localities at the same price (denoted P P C ), we can express the change in the relative price of skills in two dierent manners depending on whether or not we use the optimally condition for the determination of P Cs which is given by:
C PC ln P c A (1 ) ln c

P Cc + 1 + (1 ) ln sc = ln P P C Lc

If we use this optimally condition in conjunction with the marginal product conditions for each skill, we can express the change the in the price of skill as:
C ln P ln (A) ln P P C c = + ln sc,1980 + 1 1

(9)

ln PcS

Equation (9) indicates that the change in the relative price of skill at the city level will be greatest where its relative supply is initially most abundant (i.e. where sc,1980 is greatest). This property reects the capital-skill complementary of the arrival of the PC.10 Before the arrive of the PC, regions with more soft-cognitive skills have a relatively low price for this skill and a high price for hard skills. This makes the adoption of PCs very attractive in such a market. Therefore, PCs should be adopted more aggressively in such market causing the relative price of soft skill to increase most where soft skills are initially more abundant. If we use (9) to replace P S in 4 and 5 we get
C ln P ln (A) ln P P C c 1 1 + e ln sc,1980 + e + ec 1 1

1 (10) M F dif fec = e

This property holds for a wide variety of production setups and we can easily generalize the structure as not to obtain a unit elasticity. 10 This property does not rely on the particular functional form of the production function but does depends on the arrival a new technology where the PC is a complement to soft skills and a substitute to hard skills.

(11) Eji dif fgc =

2 ln (A) jig

C ln P ln P P C c 2 2 + jig ln sc,1980 + jig + eji gc 1 1

The interesting aspect of these two equations is that they now contrast how cross-city movements in the gender wage gap and the returns to education will respond dierently to a potentially observable aggregate factor: the relative supply of skills. In particular, these equations indicate that following the introduction of the PCs, we should see the gender wage gap fall most in cities where soft skill were most abundant prior to the arrival to the PC
1 (since, again, e < 0). Moreover, it indicates that we should simultaneously see the greatest

increase in the return to education precisely in these same cities. If we have a measure of these relative skills, then in principle we can estimate (10) and (11) consistently by OLS under the assumption that the local skill supply prior to the arrival of PC was not anticipating which cities would be best at using PCs (ie, the skill supply in 1980 is not systematically
C related to P c ). These two equations also suggest that one way of estimating equation (6)

is to use measured skills in 1980 as an instrument for EDij dif fc . This should also allow for consistent estimates under the assumption that skill supply in 1980 did not forecast PC eciency across cities. While equation (9) oers a simple and useful way of linking changes in the relative price of skill and initial skill supplies, it hides much of the mechanism underlying the the model. In particular, by using the optimality condition for the adoption of PCs, we have somewhat obscured the fact that it is the adoption of the new technology that according to the capitalskill complementarity view is causing the opposite movements in the gender wage gap and the returns to education. In order to see these intermediate forces more explicitly, it is useful to express the change the relative price of skill using only marginal product conditions for each skill. In this case, we can express the change in the relative price of skill as (12) PcS P Cc C ln sc + P c Lc

Now using 12 to replace the price of skill in 4 and 5, we obtain

P Cc 1 1 PC ln sc + e c + ec e Lc P Cc 2 2 C jig sc + jig P + eji c c Lc 10

2 (14) Eji dif fgc jig

Equations (13) and (14) now makes more explicit the relationship between the gender wage gap, returns to education and technological change. In particular, these equations indicate that greater PC adoption should be associated with a greater reduction in the gender wage gap and a greater increase in the return to education. Moreover, it suggests that a faster increase in the relative supply of skills should be associated with increases in the gender wage gap and decreases in the returns to skill. The diculty with these two equations, relative to equations (10) and (11), is that they are much more prone to endogenity. In
C 11 particular, observed adoption of PCs will be correlated with local eciency of PC use (P c ).

To address the endogeneity of P C adoption, we use 1980 measures of skill supply as the instruments, under the assumption that pre-PC-era skill supplies were not anticipating PC eciency. The formulation given by equations (13) and (14) also highlight the potential use of PC adoption as an instrument for estimating equation (6). Although PC adoption is
C endogenous and correlated with P c , this terms does not enter the error term in equation

(6) and hence is potentially a valid candidate as an instrument. In summary, our model of capital-skill complementarity in a two attribute model has highlighted dierent factors that should cause opposite movements in the gender gap and the returns to education. Furthermore, the model has provide insight regarding what instruments are potentially admissible for exploring these relationships. In the empirical section we will examine these implications to show that the ideas behind this simple model nd considerable support in the data. Once this is shown, we will discuss how the estimated relationships based on cross-sectional observation can be used to evaluate the role of skill price changes in explaining the decrease in the aggregate gender wage gap observed over the 1980-2000 period.

Data and Empirical Methods

Our empirical investigation will focus on estimating the main relationships described in the previous section (these are equations (6), (10),(11), (13) and (14)) using data on aggregate outcomes drawn from U.S. metropolitan areas. We will begin this section by discussing the data and then we will further discuss implementation and identication issues.
The change in the local supply of skill is also possibly correlated with the error terms in (13) and (14), something we will discuss further below.
11

11

3.1

Data and Measures

Our goal is to estimate three types of relationships across 230 U.S. metropolitan areas. First, the relationship between male-female and education wage gaps as described in (6). Second, the relationships between the pre-PC era supply of skill and each of the two wage gaps the gender gap and the returns to education as described in (10) and (11). And nally the relationship between the use of PCs and the two wage gaps as described in (12) and (13). To this end, we compute wages and skill supplies pre-PC using the 5% public-use version of the 1980 Census of Population (Ruggles et al., 2010), and post-PC using the 2000 Census of Population. Skill mix was constructed using only data on those aged 16-65 with positive (potential) work experience (age - years of schooling - 6 > 0), not living in group quarters. Hourly wages were constructed for the further subsample with positive wage and salary earnings and hours worked in the past year, without any self-employment earnings, currently employed and not in school. Hourly wages were Windsorized to be between two and 200 dollars in 1999 dollars. Male-female wage gaps are constructed separately for ve education groups that can be consistently identied across censuses: high school dropouts, high school graduates, those with some college education (but less than four years), four-year college graduates, and advanced degrees.12 Wages are regression adjusted, separately by gender, education group, and year, for a quartic in potential work experience and dummies for foreign-born, black, Hispanic, and being born after 1950 (where Lemieux, 2006, describes a cohort break in trends in returns to schooling.) To account for heterogeneity in years of education among workers in the dropout, some college, and advanced degree groups, we also includes a linear control for years of education and its interaction with the dummy for being born after 1950 for these groups.13 To make the means interpretable, adjusted wages are centered on the predicted values for the average female characteristics (in our whole sample of metropolitan areas) in each year.14
In the 1980 Census, high school and college workers are dened as those who have completed exactly 12 years and 16 years of schooling, respectively, and in the 2000 Census, are those who report being in the category high school graduate and Bachelors degree. 13 In neither census is there literally a years of education variable, but categories of years (1980) or degrees (2000). Within these three education groups with heterogeneous education, the grouping of education is quite dierent in the two censuses. In both cases, we impute years from the midpoint of the categories in the group. 14 In equation form, we estimate ln Wiegct = aegct + egt Xiegct + uiegct , where ln Wiegct is the natural log hourly wage of person i of education group e and gender g living in city c in year t, which is regressed on xed eects, aegct , and the adjustment variables, Xiegct , mentioned above. This is evaluated at the national
12

12

The main education wage gap we use on the right hand side is the simple average of male and female adjusted college-high school wage gaps (which come from the same adjustment procedure). We use the average to avoid any possibility of there being a mechanical relationship between the left- and right-hand side wage gaps. Our empirical implementation of the quantities of soft-cognitive skill and raw labor consists of the following. First, raw labor input is assumed to vary only with gender and not
l l with education, and we normalize f = 1, so m can be written as 1 + , for some > 0,

which will be estimated from average male-female wage gaps (described below). Letting and
F ct

M ct

represent aggregate hours worked by men and women, respectively, in local labor

market c in year t, we therefore dene (15) Lct = (1 + )

F ct .

Next, we impose that cognitive skill is linearly increasing in years of education above ten years (and is at below that) by the same amount for both genders, which roughly describes the relationship between wages and education by gender (see Figure 2). That is, we dene (16) Sct =
e

max[(e 10, 0)]

ect ,

where

ect

represents the aggregate hours worked of all persons with e years of schooling

living in metropolitan area c and year t, and is the return to schooling. To obtain estimates of and we use the wage sample with more than 10 years of education in the 1980 Census to estimate an individual level regression of ln(hourly wage) on years of = 0.077, and on the male schooling and a male dummy. The coecient on schooling is = 0.423. Substituting these into (15) and (16) generates our estimates of L ct dummy is ct and our skill mix measure, s ct /L ct . Note that this translates education and and S ct = S hours worked into human capital using xed coecient in all cities and years. (The choice , in particular, is immaterial.) of This skill mix measure may appear unusual, and it imposes the extreme assumption that gender wage gaps are entirely driven by gender dierences in raw labor input per hour.15
female mean Xef t , and thus the adjusted male-female wage gap in education group e, city c and year t is emt ef t ) Xef t . given by yect = aemct aef ct + ( 15 For example, controlling for a gender-specic quartic in potential experience reduces the estimate of to 0.12.

13

However, Figure 3 shows that, in both 1980 and 2000, it is highly correlated with a more conventional skill mix measure used in studies of the eect of computerization on skill demand: the natural log of the ratio of college equivalent to high school equivalent hours.16 In that gure, each circle represents one of the 230 metropolitan areas in our nal sample. Circle sizes are proportional to the 1980 population, the weights used throughout the analysis. Our analysis will also construct standard errors to be asymptotically robust to arbitrary error correlation within metropolitan area (across skill groups) and to heteroskedasticity. Our measure of computerization is personal computers per worker at the average employer in the metropolitan area, adjusted for three-digit (SIC) industry crossed with size category dummies. The underlying data are rm level data collected by the marketing rm HarteHanks in 2000 and 2002. For simplicity, we will refer to this as 2000 data. (Weighted) summary statistics on our metropolitan-level wage gap and skill mix measures are shown in Table 1. In each year there are 1,150 observations on the male-female wage gap from 230 metropolitan areas and ve education groups. As has been documented elsewhere, the male-female wage gap declined over this period, by about 12 log points in our data. This decline was largest between less-educated men and women. Table 1 also shows there is something to be explained - there is variation in the level and change in the gender wage gap across labor markets, even within education group, which itself is perhaps a new fact. We now ask whether it is related to skill mix in the way the model suggests.

3.2

Further Identication issues

There are at least three additional issues we want to discuss as they could bias our inferences. First, there is Mulligan and Rubinsteins (2008) (MR) claim that rising returns to skill induced dierential entry of high skill women into the labor market, shrinking the gender wage gap through change in the selection of working women from negative to positive. Second, there is Mincer and Polachecks (1974) view that the gender wage gap results from women interrupting their careers to have children, or put in labor economists terms, because they had low actual work experience for a given amount of potential work experience compared
16 As in Card(2009), Figure 3 denes college equivalents as those with a four year degree plus 0.4 share of those with 1-3 years college. High school equivalents are 0.6 share of those with 1-3 years college, plus all of those with exactly 12 years, plus 0.7 share of those without a high school degree. (The fractional divisions derive from the workers in an education group supplying less than one eciency unit per hour and/or dividing their labor supply between college and high school tasks.)

14

to men. This gap in actual work experience may have diminished over time as women had more continuity in their careers, shrinking the wage gap.17 Third, there is the possibility that a decline in discrimination just happened to occur at the same time as the diusion of PCs. In our view, the most important source of potential bias comes from the MR story. In our model, education and male female wage gaps are negatively correlated because of common unobserved skill prices, but in the MR story they are related for a dierent reason: high returns to skill dierentially induce high skill women to enter the workforce. In this view, the arrival of computers could still reduce the male-female wage gap, but through changes in selection (rather than through changes in the quality constant male-female wage gap).18 We do three things to address this type of selection. First, we will include estimates that control for selection in a manner similar to the way MR did. We will use an inverse mills ratio transformation of the predicted probability of female employment, using the presence of kids under age six in the household (whose eect is allowed to vary by marital status) as an instrument for female employment. Further details on this estimation strategy are below. In keeping with MR, this selection correction is allowed to vary by year, and in our case, regionally. Second, also like MR, we will examine groups of women whose employment is likely less sensitive to the wage structure, including single women without kids. Third, and more simply, we control for female employment rates. As these could be endogenous outcomes of wages, however, we use and interpret these controls cautiously. As for other potential sources of decline in the male-female wage gap, while we are not a priori ruling out that systematic dierences in work experience or gender discrimination bias our estimates, there is no clear reason why these should be correlated with our right-hand side variables. Nevertheless, we want to address these potential biases to the extent possible. To minimize the inuence of unobserved dierences in actual (rather than potential) female work experience across markets and over time, i.e., to address the Mincer-Polachek type story, all of our wage gaps are adjusted for for gender (x education x year)-specic potential
In addition, the diusion of new birth control technology may have driven some of these other purported changes (Goldin and Katz, 2002; Bailey, 2006). In particular, Bailey, Hershblein, and Miller (2012) provide evidence that early legal access to the birth control pill induced women to increase the total hours worked by a given age, among other investments in skills. 18 MR do not take a stand on what is generating the changes in the return to skill. Note that even in the MR story, selection may or may not bias our estimates: our male-female wage gaps condition on education (in addition to other observables, described below), so the MR-type bias will arise only if residual wage gaps are also larger in more high skilled cities. In addition, we are attempting to identify factor-supply, rather than demand, driven dierences in education wage gaps.
17

15

experience proles. To help control for changes in gender discrimination, we will include estimates that control for state xed eects, which capture the eect of any state-level legislation improving the rights of women. A fourth, more general, critique of our estimation approach is that there may be systematic demographic or compositional dierences correlated with skill mix and changes in the relative wages of females. One plausible source of potential bias is dierences in industry mix across metro areas. Olivetti and Petrongolo (2011) nd that dierences in industry mix can account for a substantial portion of cross-country dierences in the gender gap. In addition, during this time period the decline in the wages of less-skilled men in manufacturing could have simultaneously lowered male-female wage gaps and raised college-high school wage gaps in manufacturing-intensive locales (which is correlated with being a less-skilled locale).19 While some of this might be due to technological change, some of it might be due to other forces (like a decline in union power). Therefore we will control for measures of industry mix, and manufacturing share in particular (described below). To address the possibility that other types of compositional dierences drive the results, we will also evaluate the sensitivity of our estimates to controls for demographic mix (e.g., black share, immigrant share) and other city characteristics.

4
4.1

Results
Ordinary Least Squares

Columns (1)-(7) of Table 2 shows OLS estimates of the relationship in equation (6), between changes in the male-female and college-high school wage gaps between 1980 and 2000, with various sets of controls. The controls have been demeaned so that the intercept, which is shown, can be interpreted as the counterfactual change in the male-female wage gap in a location with the average value of controls but no increase in the returns to college. (See Section 5, below.) The estimates also pool together gender wage gaps for our ve dierent education categories (and, again, standard errors are calculated to be robust to error correlation across education groups in a metro area). Controlling only for education group eects (which by denition make no dierence to the
ONeill and Polachek (1993) nd that the decline in blue collar wages accounts for a quarter of the decline in the male-female wage gap in the 1980s.
19

16

point estimates, since the education wage gap does not vary across education groups) produces a coecient of -0.231, or that a one percentage point increase in the return to college is associated with a 0.231 percentage point decline in the male-female wage gap. Despite having (what is likely a very) noisy right-hand side variable, this relationship is highly signicant. Other columns of Table 2 add controls. In light of Olivetti and Petrongolo (2011), we believe that it may be important to control for industry mix. So in columns (2)-(4) we explore three dierent versions of industry mix controls. To begin with, we control for durable and nondurable manufacturing employment shares, measured in 1980, whose impact is allowed to vary by the two, what we will call broad, education categories that Olivetti and Petrongolo (2011) used: (1) workers with some college or below or (2) four years of college or more.20 As expected, this lowers the coecient, as the decline in male wages in manufacturing-intensive locales lowers both male-female and raises college-high school wage gaps. These controls do not, however, account for all of the OLS relationship. Dierences in the size of manufacturing may be the most plausible source of industry-mix driven bias, but they may not be the only source.21 However, with only 230 metro areas, adding detail to the industry mix controls can quickly make estimates imprecise and will also tend to make measurement error attenuation worse. So we have tried to nd parsimonious ways of controlling for detailed industry mix. In column (3) we use an alternative wage adjustment (for the dependent variable) which controls for a full set of census industry dummies.22 These estimates are larger than the ones in column (2). This may mean that detailed industry controls make little additional dierence, though the estimates in column (3) are also conceptually dierent they are within industry gender wage gaps, and so do not capture any broader eects of industry mix on wage gaps of men and women not in a particular industry. To account for this, in column (4) we control for the manufacturing shares as before, and add an index which measures the average womanpower requirements of the local detailed industry mix (in 1980), as in for broad education group b, and
20
j

fjb

jc

j jc

, where fjb represents the female

share of total hours worked in industry j (in our entire sample of 230 metropolitan areas)
jc

is total hours worked in industry j and city c. This is

Dening broad education groups this way is also consistent with evidence suggesting that workers of dierent education levels within these broad groups are near perfect substitutes (e.g., Goldin and Katz, 2008). 21 We found that adding other two-digit sector employment shares as controls has little additional impact on the point estimate, though it does make the standard errors larger. 22 We use the approximately 200 industry categories that are harmonized to 1990 industry categories (Ruggles et al., 2010). These controls are in addition to the other controls included in the wage adjustment, described in the previous section, which is again separately estimated by gender, education group, and year.

17

calculated separately for the same two broad education groups as before.23 These estimates are similar to column (2), but with the smallest standard errors of any of the approaches. This reinforces the value of using a parsimonious set of controls. Throughout the rest of the paper we will use these as our controls for industry mix. Regional dierences in the extent of gender discrimination might aect our estimates. These are very dicult to quantify. To at least try to capture the eects of state policies which might aect the male-female wage gap, we control in column (5) for state dummies. The coecient is larger with this control, suggesting such policies do not work in the same direction as our results.24 Column (6) adds a few other controls which might have a compositional impact. These include the share foreign-born, black, and Hispanic, and the unemployment rate and the natural log of the the citys labor force. The latter two attempt to capture any dierences in sensitivity of male-female wage gaps to the business cycle or agglomeration eects. The others might shift skill and gender ratios where they settle (recall that wage gaps are already adjusted for nativity, race, and ethnicity at the individual level). Column (7) controls for female employment rates by broad education. This is meant to address a possible concern related to MRs evidence that working women were historically negatively selected. If this negative selection were particularly strong in high skill markets which it might be because returns to skill were low in these markets then this would be another force which could account for the positive relationship between skill share and male-female wage gaps. In fact, these controls have only a little impact. As this control is potentially an endogenous outcome of the male-female wage gap, we added it last, and below, we examine other methods for correcting for selection. Column (8) replaces the independent variable, the college-high school wage gap, with (similarly regression-adjusted and averaged over men and women) estimated linear return to schooling (above 10 years), multiplied by four. The estimates using this measure are quite similar, though less precise. They may be less well measured because of the need to interpoTo account for productivity dierences across education groups within these broad education categories, the female share of hours is calculated for college and high school equivalents (Card, 2009, denition). The college equivalent female demand index is applied to the top two education groups, and the high school equivalent one is applied to the bottom three. 24 Indeed, male-female wage gaps were (perhaps unexpectedly) highest in 1980 in highly educated markets like San Francisco, Minneapolis, and Boston where it is likely that there was more widespread support for equal treatment of women: the Equal Rights Amendment, for example, was ratied in California, Minnesota, and Massachusetts, among other states in their regions.
23

18

late education education categories into a linear years of education, which is constructed dierently in 1980 and 2000 because of the change in how education is coded. (See Figure 2.) We believe these coding changes are likely less of a problem for measuring the wage gaps between college and high school workers. So we will continue using college-high school wage gaps as our main measure of education wage gaps.

4.2

Relationship to Skill Mix

Now we turn to reduced form estimates of the relationship between wage gaps and our proxy for the initial relative abundance of soft-cognitive skills, s c,1980 , which we will call human capital/raw labor. As outlined in the theory section, this is expected to have oppositesigned relationships with changes in male-female and changes in college-high school wage gaps. In particular, as (10) and (11) described, initial human capital is expected to have a negative relationship with the change in male-female wage gap and a positive relationship with the change in college-high school wage gap. Table 3 shows the relationship of the male-female (in Panel A) and college-high school (in Panel B) with skill in 1980. As before, column (1) controls only for education group dummies. The coecient on s is negative in Panel A and positive in Panel B, as the theory predicts. For the most part, the controls have little impact on these relationships, though the industry controls, added in column (2) enlarge the magnitude of both relationships a bit. The high stability of the estimates in panel B is particularly reassuring for the validity of the approach we are taking, because Panel B estimates are also the rst stage relationship for the main IV estimates of the relationship between changes in male-female and college-high school wage gaps (below). Figure 4 shows residual plots of the bivariate relationship between the changes in wage gaps, corresponding to the rst four columns of Table 3. Reassuringly, the relationship does not appear to be driven by any particularly inuential points.

4.3

PC Adoption

Tables 4a shows the relationship between PCs per worker in 2000 and the 1980-2000 change in male-female wage gap, while Table 4b shows the relationship with the 1980-2000 change in the college-high school wage gap. PCs per worker in 2000 is a proxy for the intensity of computer adoption over 1980-2000, in light of the fact that PCs per worker was zero in 1980.

19

As per equation (13) and (14), all regressions control for the change in skill mix over the period (although this turns out to make little dierence). Panel A of Table 4a shows least squares estimates. The coecient on PCs in column (1) with limited controls is -0.204, which says a 0.1 increase in PCs per worker (a little more than a standard deviation Table 1) is associated with a 2.4 percentage point decline in the growth of the male-female wage gap. As expected, an increase in skill supply, s is associated with an increased male-female wage gap.25 The industry controls have little eect on the relationship with PCs, while controls added in other columns diminish the relationship. As discussed in the theory section, OLS estimates are expected to be biased towards zero by unobserved factors which make PCs more productive. Panel B shows instrumental variables estimates, where the personal computer variable is treated endogenous, and the instrument is 1980 human capital/raw labor. First stage F-stats are shown below Panel B. Without controls other than education eects, the rst stages F-stat is 136, and with all of the controls it is 63, both quite strong. Figure 5 also shows the bivariate relationship between PCs per worker in 2000 and skill mix in 1980 is strong and not driven by outliers, consistent with other work showing a relationship between local skill mix and PC adoption (Caselli and Coleman, 2001; Doms and Lewis, 2006). As expected, the point estimates in Panel B are larger. In addition, unlike OLS estimates, they also do not generally diminish with the addition of controls. The point estimate here suggests a 0.1 increase in PCs per worker lowers the male-female wage gap roughly four percentage points. The change in skill mix is treated as exogenous in Panel B, though it may not be. To address this, we tried a couple of things. First, we used the size of immigrant enclaves in 1970 to predict changes in skill mix 1980-2000, exploiting the fact that there was a boom in immigration over this period, and that immigrants tend to cluster near immigrants of the same origin. Though this variable does a reasonable job of predicting changes in skill mix, it does not have enough power to do so within state, so the approach does not work once state eects are controlled for. Nevertheless, using this approach, the estimates without state eects are quite similar to the estimates in columns (1) and (2) of panel B.26 In addition, we tried dropping the change in skill mix as a control. As can be seen in Appendix Table A1a,
1 In equation (13), recall, the parameter e is expected to be negative. In particular, the coecient estimate (standard error) on PCs per worker with just education group eects is -0.242(0.064) and with industry mix controls is -0.339(0.080) when we add the ethnic enclave instrument and treat changes in skill mix as endogenous. The corresponding coecient on skill mix changes are 0.796(0.227) and 0.862(0.267). 26 25

20

this has little eect (on either the OLS or IV estimates). Table 4b performs a parallel set of estimates where the dependent variable is the change in the college-high school wage gap. The pattern of estimates is quite similar to Table 4a, except, as expected, with coecients of the opposite sign. The alternative instrumental variables strategies produce similar estimates in this case as well.27 In summary, U.S. labor markets with greater PC adoption tend to experience both faster increases in the college-school wage gap and faster declines in the male-female wage gap.

4.4

IV estimates of the co-movement between the gender gap and the returns to college

Table 5 puts together the male-female and education wage gap results into instrumental variables estimates (which are the ratio of coecients in panels A and B of Table 3). All of the estimates show a negative signicant relationship between the wage gaps, with a magnitude roughly around 0.6-0.7. This is larger in magnitude than the OLS estimates in Table 2, consistent with OLS estimates being attenuated. It is, however, approximately the same magnitude of comovement found in the aggregate. In particular, a regression using aggregate annual data for 1980-2000, stacking adjusted male-female gaps (for our same ve education groups) on college-high school gaps produces a coecient (standard error) of -0.710 (0.0452).28 In light of the signicant relationship between wage gaps and PC adoption, another approach is using PC adoption as an instrument.29 PC adoption is, after all, our proxy for the treatment: the adoption of computers is what we hypothesize is driving the changes in the wage structure. In addition, recall that the error term of the (rst stage) relationship between education wage gaps and PCs, (14), is, according to the theory, uncorrelated with the error in the wage gap relationship we are trying to identify, (6). In practice, the IV
With the ethnic enclave instrument added, the coecient (standard error) on PCs in column (1) would be 0.425(0.129) and in column (2) would be 0.458(0.179). The point estimates on the change in skill mix are -2.049(0.480) and -2.238(0.600) for columns (1) and (2), respectively. Estimates dropping the change in skill mix control can be found in Appendix Table A1b. 28 Standard error clustered on year. Wage series constructed using merged outgoing rotation groups of the Current Population Survey, using only data on those aged 16-65 with positive (potential) work experience (age - years of schooling - 6 > 0), positive wage and salary earnings and hours worked in the past year, and currently in the labor force. Hourly wages were Windsorized to be between two and 200 dollars in 1999 dollars. Some of the wage series used in the regressions are shown in Figure 1. 29 The rst stage corresponding to this IV specication is shown in panel A of Appendix Table A1b.
27

21

estimates we get using our PC adoption variable as the instrument are a bit more sensitive to controls (see Appendix Table A2), though they are in the same ballpark.30 Next, we turn to two important robustness checks: (1) to what extent are our estimates driven by changes in the selection of women into cities workforces? And, (2) does the timing of wage changes match the timing of the arrival of PCs, or were there similar changes in wage structure occurring before the arrival of PCs?

4.5

Is it Selection?

In a prominent recent paper, MR argue that rising returns to skill have induced the selection of women into work to become more positive. This suggests an alternative mechanism for our results so far: rather than being driven by common unobserved skill prices, the negative relationship between changes in college-high school and male-female wage gaps might be due to dierences in selection. To investigate this possibility, in Table 6 we apply MRs selection correction methods to our data. Some of these methods are data intensive, and so in this section we restrict our sample to a set of 159 larger metropolitan areas where these methods are feasible.31 Column (1) of Table 6 shows the estimated relationship between male-female and collegehigh school wage gaps in this subsample, with OLS estimates in Panel A, and instrumental variables estimates in Panel B. Compared to column (6) of Table 2 (for OLS) and column (4) of Table 5 (for IV), these estimates are quite similar: the change of sample has little impact on the estimates. Columns (2) and (3) compute male-female wage gaps only using female demographic groups with a high probability of working. MR argued that this allowed for identication at innity: intuitively, women with a high probability of working are plausibly less sensitive to the wage structure in their employment decisions, and so their wages are likely less biased by selection. One way in which we identied women with a high probability of working is as follows. We estimated, separately for each education group, the probits for being in the wage sample, that is

(17) P r(wgobsic ) = (ac + Xic + Zic +

ic ) ,

Coecient estimates (standard errors) range from -0.940(0.491) to -0.425(0.310). Specically, we limit the sample to metro areas with at least 100 wage observations in all ve education groups in both 1980 and 2000, and for which all of the probits for female employment converged within 10 iterations.

22

where P r(wgobsic ) represent the probability that woman i in metro area c meets our criteria for being in the wage sample (see Data section). The probit includes metro area xed eects, ac ; a vector of adjustment controls, Xic , that are identical to what is used in to adjust wages in earlier estimates but also includes dummies for marital status and their interaction with black, Hispanic, and foreign-born; and a set of instruments Zic , used by MR, that are twoway interactions between marital status and presence of children under age six, which we further interact again with black, Hispanic and foreign-born.32 (Although the validity of these instruments is questionable, something MR acknowledge, we nevertheless would like to probe the sensitivity of our results to using them.) We evaluated the predicted values in the average, metro area, that is putting in a in place of the vector of estimated xed eects. We dene women with a high probability of working as those with at least a 0.6 probability of working. To avoid compositional changes, we estimate this probit using the 1980 data alone, and apply the same estimates to dene such women in 2000.33 Estimates using this subsample of women are shown in column (2). Comparing to column (1), there some sign that this diminishes the magnitude of the estimates. In column (3) we take a simpler approach with the same motivation, looking only at native-born non-Hispanic white women without kids under age 6. The estimates in this subsample tend to be, if anything, larger in magnitude. Another approach we can take is to use the full sample of women but control for selection. In performing the wage adjustments on women, we control for an inverse mills ratio transformation of the estimate of (17) (now allowed to vary by year) which accounts for selection under the assumption of normal errors.34 The estimates, shown in column (4) are similar to those in column (1). Not shown is the fact that the mean of our selection adjusted male-female wage gap replicates the result of MR: there is no longer any average decline in male-female wage gap once this adjustment is made. Nevertheless this adjustment does not eliminate the correlation between changes in wage gaps across metro areas. One of the key points MR raise is that the selection function itself may depend on the
MR limit the sample to white non-Hispanics. We have tried this as well. It has little eect on the estimates, but, predictably, leads the standard errors to be a bit larger. 33 By 2000, the probability of women working had shifted up, so the women who met this threshold in 2000 also would have at least a 0.6 probability of working. 34 In the wage adjustment step, we add to the list of controls described in the Data section marital status controls interacted with dummies for black, Hispanic, and foreign born; and, for females, the interaction of the inverse mills ratio with dummies for black, Hispanic, and foreign born. The adjusted wages are evaluated at the national mean of female characteristics and with the inverse mills ratio set to zero.
32

23

wage structure. So rather than just estimate a single probit for each year x education group, in column (5) we allow the probit estimates to vary by metro area. If this made the relationship between wage gaps diminishes in magnitude, it would cast doubt on the view that the relationship was being driven by unobserved skill prices. In fact, there is little systematic sign of this in column (5). It should be said that estimating the adjustment separately for each MSA clearly pushes the data beyond its limits, leading to some very imprecise estimates that include zero in the condence interval.35 So it perhaps fair to say that these estimates do not completely rule out that selection is driving our results. However, the point estimates here, combined with earlier results which control for female employment rates, seem to suggest selection only explains part of the observed comovement of wage gaps.

4.6

The Timing of the Changes

As an additional test of the model, in Table 7 we examine more closely the timing of the changes in the wage relationships. Our key empirical fact is that changes in the male-female and college-high school wage gap moved in opposite directions in the 1980-2000 period. We argued that this was due to the arrival of PCs during this period, which resulted in a shift in the production and wage structure. Because PC adoption is likely endogenous, we used 1980 skill mix to identify the relationship. If we nd similar relationships prior to the introduction of PCs, therefore, it would cast serious doubt on the interpretation that the relationship was being driven by the introduction of PCs. To see if this is the case, in Table 7 we examine the same relationships between changes in wage gaps and between the gender wage gap and skill supply before and after the PC is introduced. There are some minor data issues that must be addressed in order to carry out this estimation, due to the fact that 1970 census is both much smaller and has much less detailed measures of geography and hours worked than later censuses. (These issues are detailed in the notes to Table 7 and in the Data Appendix.) Table 7 therefore examines a sample of 137 large metropolitan areas which can be consistently identied in 1970 and in later censuses. It turns out that only with the 1970, 1980, and 1990 censuses and not with the 2000 census can we construct wage measures entirely consistently.36 To insure that
In many metro area-education group-year cells in the probit, the coecients on the Z s are also not jointly signicant, making the estimates only identied o of functional form. 36 Specically, the 2000 Census does not include the hours worked last week variable that is used in this construction. The 1970 data come from the two one percent public-use county-group les, and the 1990 data are from the ve percent public use data (both Ruggles et al., 2010).
35

24

any dierences in results in the 1970s come from the change of decade and not of methods, Table 7 also examines the post-PC 1980-1990 period, using data construction methods that are available in the 1970 census. Panel A shows the OLS relationship between changes in male-female and college-high school wage gaps, like in Table 2, but with the smaller sample. Column (1) shows the relationship in the pre-PC (1970-1980) period. Unlike in Table 2, there is no signicant relationship. Column (2) makes clear that this not because of the change of sample or methods, because 1980-1990, using the same methods, there is a negative signicant relationship between the changes in wage gaps, which is, if anything, larger in magnitude that comparable estimates (in column (6)) in Table 2. There is also a signicant relationship in this sample 1980-2000 using our original wage construction methods in this sample (column (3)). Panel B looks at the relationship between the change in the wage gaps and the instrument, 1980 human capital/raw labor. Column (1) of Panel B shows that although the reduced form relationship is negative 1970-1980, it is not signicant. The point estimate is also much larger after 1980 than before. The size of the standard errors in column (1) are unfortunate, because it means we cannot totally rule out similar pre-PC trends in the male-female wage gap. But the relationship appears to at least be much weaker before the arrival of PCs.37

Implications for Aggregate Outcomes

We started this paper with the observation that the wage dierential between men and women with similar levels of education decreased substantially between 1980 and 2000, while at the same time the return to education increased substantially. As suggested by, among others, Welch (2000), this decrease in the gender wage gap could reect a change in a the relative price of skill which is more abundant among women and more educated workers. We have now shown that this idea nds considerable support also in cross-city data: the male female wage dierentials and the return to education at the city level appear to react in opposite directions to factors that likely inuence the relative price of soft-cognitive skills
It is also possible that a weak relationship exists in the 1970s because at least some similar technological change occurred before the arrival of PCs. For example, Autor, Levy, and Murnane nd evidence of shift in the skill-biased shift in the task content of the economy in the 1970s that is similar in nature and smaller in magnitude than later decades. Goldin and Katz (2008) argue that technological change has been skillbiased throughout the twentieth century, but only recently has the growth in the supply of skills failed to keep up with demand, generating increased education wage gaps.
37

25

versus hard-raw motor skills. In this section we want to discuss how our cross-city estimates can be used to infer about the potential role of changes in skill prices in explaining the decrease in the aggregate male-female wage gap between 1980 and 2000. There are two ways of using our cross-city estimate to evaluate the role of skill prices changes in explaining the gender wage gap. First, if we are ready to assume that the aggregate movement in the return to college reects mainly a change in the relative price of skill, then we can use our estimates of
1 2

obtained from the IV estimation of equation (6) to calculate

the contribution of increase in skill price on the the gender gap. In particular, over this period we observed the return to college increase by 19.2%. If we multiply this by our estimates of
1 2

reported in Table 5, we get a predicted eects on the male female wage dierential

ranging from -10.2% to -15.9%. Since the decrease in the male-female wage dierential over this period was approximately 12.4%, this exercise implies that essentially all of the decrease in the gender wage gap can be explained by the change in the relative price of skills. The OLS estimates of the relationship (Table 2) can account for about one-quarter. If we are not willing to assume that the change in the aggregate return to education over this period was driven only by a change in a relative prices, there is a second way to proceed. Starting from equation (6), consider taking a weighted sum of each term where the weights are the relative population weights of the cities. The weighted sum of the error terms (including any estimated intercept for this regression) then gives us an estimate of the change in the male-female wage dierential which cannot be attributed to the change in the skill price. If we do this exercise based on the IV estimates in Table 5, we again nd something very close to zero. In fact, this can be seen by directly be examining our estimate of the intercept of this regression.38 When we estimate this relationship between OLS as in Table 2, we nd that the intercept is signicantly negative. In contrast, when we estimate this relationship by instrumental variables, we nd that the intercept is insignicantly dierent than zero, indicating that on average across cities the predicted local increase in the returns to college combined with the estimate of
1 2

is sucient to explain the changes in the gender wage

gap over the period.39 Hence, we believe that that these two pieces of evidence point in the same direction: Most of the decrease in the gender wage gap over the 1980-2000 period can be attributed to a change in the relative price of a skill which we have referred to as a soft-cognitive skill that is more abundant among women and more educated workers.
38 39

Recall that all of the controls are demeaned, and all regressions are weighted by 1980 population. We also nd this when using PC use as the instrument (Appendix Table A2).

26

Conclusion

Motivated by the simultaneous decline in male-female and rise in education wage gaps in recent decades, this paper has asked whether both trends were driven by a change in the relative price of an an unobserved skill which both women and educated workers have in abundance we called it soft-cognitive skills induced by skill biased technological change. This idea has been suggested before. But by exploring cross city variation in wages and skill mix between 1980 and 2000, we are able to move beyond the aggregate relationships and, for the rst time, directly examine the relationship between the return to education and male-female wage gaps. Consistent with the idea that females are relatively abundant in soft skills, we nd that after the arrival of PCs, markets that experienced faster increases in the college-high school wage gap saw bigger drops in the male-female wage gap. This relationship remains strong when controlling for industry mix, or when examining dierences in education wage gaps induced by our proxy for the supply of human capital: consistent with a standard models of skill-biased technological change, the decline in male-female and increase in education wage gaps was largest in initially human capital intensive markets. As robustness checks, we show that our estimates survive attempts to account for cross-city dierences in the selection of women in the workforce, including focusing on female subgroups with high propensities to work, as well as controlling for an estimate of the selection bias. In addition, we show that there was no signicant trend in the relationship between education and male-female wage gaps in the 1970s, before the introduction of PCs. Overall, our estimates are consistent with a substantial role for changing skill prices in accounting for the decline in the male-female wage gap between 1980-2000. Even applying our OLS estimates, which we have reason to believe are substantially attenuated, suggests that the rise in the return to education between 1980 and 2000 accounted for one-fourth of the decline in the male-female wage gap over the period. Our IV estimates can account for all of the increase. This does not mean that other forces cannot inuence gender equality in earnings but it does suggest the historically dramatic decline in the male-female wage gap in the 1980s may have been largely driven by technological forces unique to that period. The relative stagnation of the male-female wage gap in more recent years despite continuing changes in education wage gaps (Figure 1) may reect that.

27

References Acemoglu, Daron and David Autor. Skills, Tasks and Technologies: Implications for Employment and Earnings. in Orley Ashenfelter and David Card, eds., Handbook of Labor Economics, Volume 4. Amsterdam: Elsevier-North Holland, 2011, pp. 1043-1171. Altonji, Joseph and David Card. The Eects of Immigration on the Labor Market Outcomes of Less-Skilled Natives. in John M. Abowd and Richard B. Freeman, eds., Immigration, Trade and the Labor Market. Chicago: University of Chicago Press, 1991, pp. 201-34. Autor, David H., Frank Levy, and Richard J. Murnane. The Skill Content of Recent Technological Change: An Empirical Exploration, Quarterly Journal of Economics, 118(4): November 2003, pp. 1279-1334. Bacolod, Marigee, and Bernardo S. Blum. Two Sides of the Same Coin: U.S. Residual Inequality and the Gender Gap. Journal of Human Resources 45(1): Winter 2010, pp. 197-242. Bailey, Martha J. More Power to the Pill: The Impact of Contraceptive Freedom on Womens Lifecycle Labor Supply. Quarterly Journal of Economics 121(1): February 2006, pp. 289-320. Bailey, Martha J., Brad Hershblein. and Amalia R. Miller. The Opt-In Revolution? Contraception and the Gender Gap in Wages. American Economic Review: Applied Economics : May 2012. Beaudry, Paul, Mark Doms, and Ethan Lewis. Should the PC be Considered a Technological Revolution? Evidence from US Metropolitan Areas. Journal of Political Economy 118(5): October 2010, pp. 988-1036. Beaudry, Paul and David Green. Wages and Employment in the United States and Germany: What Explains the Dierences? American Economic Review 93(3): June 2003, pp. 573-602. . Changes in U.S. Wages, 1976-2000: Ongoing Skill Bias or Major Technological Change? Journal of Labor Economics 23(3): July 2005, pp. 609-648. Black, Sandra and Alexandra Spitz-Oener. Explaining Womens Success: Technological Change and the Skill Content of Womens Work. The Review of Economics and Statistics 92(1): February 2010, pp. 187-194. 28

Caselli, Francesco. 1999. Technological Revolutions. American Economic Review 89(1): May 1999, pp. 78-102. Caselli, Francesco and John Wilbur Coleman II. Cross-Country Technology Diusions: The Case of Computers. American Economic Review Papers and Proceedings 91(2): May 2001, pp. 328-335. Doms, Mark and Ethan Lewis. Labor Supply and Personal Computer Adoption. Federal Reserve Bank of San Francisco Working Paper #2006-15, June 2006. Fortin, Nicole M. and Thomas Lemieux. Are Womens Wage Gains Mens Losses? A Distributional Test. The American Economic Review Papers and Proceedings 90(2): May 2000, pp. 456-460. Goldin, Claudia; Katz, Lawrence F. The Power of the Pill: Oral Contraceptives and Womens Career and Marriage Decisions. Journal of Political Economy 110 (4): August 2002, pp. 730-70. Goldin, Claudia and Larry Katz. The Race Between Education and Technology. Cambridge and London: Harvard University Press, Belknap Press, 2008. Lemieux, Thomas. The Mincer Equation Thirty Years after Schooling, Experience, and Earnings. in S. Grossbard-Shechtman, ed., Jacob Mincer, A Pioneer of Modern Labor Economics, Springer Verlag, 2006. Lewis, Ethan. Immigration, Skill Mix, and Capital-Skill Complementarity. Quarterly Journal of Economics 126(2): May 2011, pp. 1029-1069. Minnesota Population Center. National Historical Geographic Information System: Pre-

release Version 0.1. Minneapolis, MN: University of Minnesota, 2004; available at http://www.nhgis.org. Mulligan Casey, B. and Yona Rubinstein. Selection, Investment, and Womens Relative Wages. Quarterly Journal of Economics 123(8): August 2008, pp. 1061-1110. Olivetti, Claudia and Barbara Petrongolo. Gender Gaps across Countries and Skills: Supply, Demand and the Industry Structure. National Bureau of Economic Research Working Paper #17349, August 2011. ONeill, June and Solomon Polachek. Why the Gender Gap in Wages Narrowed in the 1980s. Journal of Labor Economics 11(11): January 1993, pp. 205-228. 29

Ruggles, Steven, J. Trent Alexander, Katie Genadek, Ronald Goeken, Matthew B. Schroeder, and Matthew Sobek. Integrated Public Use Microdata Series: Version 5.0 [Machine-readable database]. Minneapolis: University of Minnesota, 2010. Weinberg, Bruce. Computer Use and the Demand for Female Workers. Industrial and Labor Relations Review 53(2): January 2000, pp. 290-308. Welch, Finis. Growth in Womens Relative Wages and in Inequality among Men: One Phenomenon or Two? The American Economic Review: Papers & Proceedings 90(2): May 2000, pp. 444-449.

30

Data Appendix

To be completed.

31

Figure 1. Male-Female, College-HS Wage Gap*, 1979-2010

.2 1980

.3

.4

.5

1990

2000

2010

*Adjusted for gender x year specific quartic in potential experience (age-years ed-6), evaluated at the female average over 1979-2010. **Simple average of male and female adjusted college-high school wage gaps.

Figure 2. Ln(Wage) by Years of Education 1980 and 2000, by Gender

B. 2000
Females

Mean ln(Wage), 1999$ 2.5 3

10 15 Years of Education

20

2 0

Mean ln(Wage), 1999$ 2.5 3

10 15 Years of Education

20

Figure 3. Human Capital/Raw Labor vs. ln(College/HS Equivalent Hours), 1980 and 2000
1980
.3 .3

2000

.25

.2

.15

.1

-1.5

-1 -.5 0 ln(Coll/HS Equiv Hrs),1980

.5

.1 -1.5

.15

.2

.25

-1 -.5 0 ln(Coll/HS Equiv Hrs),2000

.5

Figure 4. Residual Plots: Change in Wage Gaps, 1980-2000 vs. 1980 Human Capital/Raw Labor
Male-Female
.6
Education group effects (column 1)

College-High School
.2
Education group effects (column 1)

.4

.2

-.4

-.2

-.1 .6

-.05

0
+ industry mix (2)

.05

.1 .2

-.2 -.1

-.1

.1

-.05

0
+ industry mix (2)

.05

.1

.4

.2

-.4

-.2

.6

+ state effects (3)

.2

-.1

-.05

.05

.1

-.2 -.1

-.1

.1

-.05

0
+ state effects (3)

.05

.1

.4

.2

-.4

-.2

.6

+ city controls (4)

.2

-.1

-.05

.05

.1

-.2 -.1

-.1

.1

-.05

0
+ city controls (4)

.05

.1

.4

.2

-.4

-.2

-.1

-.05

.05

.1

-.2 -.1

-.1

.1

-.05

.05

.1

Figure 5. 2000 Adjusted* PCs/Worker vs. 1980 Human Capital/Raw Labor

SF Bay Area Boston Minneapolis Denver Houston NYC Seattle DC Chicago St. Louis CharlotteLas Vegas Los Angeles Honolulu

El Paso Scranton, PA Hickory, NC .1

Fresno, CA

.15 .2 Human Capital/Raw Labor, 1980

.25

Table 1. Descriptive Statistics 1980 Mean (1) Adjusted Male-Female Wage Gap All Education Levels High School Dropouts High School Graduates >0 and <4 years college 4 Years College Advanced Degree Adjusted College-HS Wage Gap Human Capital/Raw Labor Adjusted PCs/Worker Number of Metro Areas 230 0.376 0.461 0.445 0.380 0.337 0.257 0.257 0.188 0.098 0.075 0.063 0.053 0.055 0.068 0.045 0.021 0.254 0.261 0.283 0.248 0.236 0.239 0.449 0.231 0.497 230 0.054 0.054 0.054 0.047 0.048 0.054 0.046 0.023 0.055 230 -0.123 -0.200 -0.162 -0.132 -0.102 -0.018 0.192 0.043 0.082 0.068 0.041 0.038 0.052 0.067 0.046 0.014 Stand Dev (2) Mean (3) 2000 Stand Dev (4) Change 1980-2000 Mean (5) Stand Dev (6)

Notes: Raw data sources are the 1980 (Ruggles et al., 2010) and 2000 public use 5% censuses of population for the w age and human capital variables, and 2000 and 2002 surveys by Harte Hanks for PCs per w orker, both of w hich have been collapsed (using sample w eights for Census variables) to a metropolitan area-average level dataset w hose descriptive statistics are show n in this table. Sample used to compute human capital/raw labor consists of w orkers age 16-65 w ith positive potential w ork experience (age - years of education -6) and hours w orked last year, and not residing in group quarters. This variable is close to an hours-w eighted average years of schooling above 10 years (multiplied by the estimated return to schooling in 1980, 0.077), w ith an adjustment for the share of hours w orked by females -- see text for details. Wage sample further limited to those w ho are currently employed, w ith positive w age earnings but zero business and farm earnings and not currently enrolled in school. Wage adjusted, separately by gender and education (and year), for a quartic in potential experience, linear returns to education (for high school dropouts, some college, and advanced degree categories) and dummies for foreign-born, black, Hispanic, and being born after 1950. (The latter is also interacted w ith years education for the same three groups). PCs per w orker are regression adjusted for 3 digit industry x employer size dummies. The descriptive statistics in this table are w eighted by1980 population.

Table 2. Change in Male-Female Wage Gaps vs. Change in Schooling Wage Gaps, 1980-2000, Estimated by OLS (1) Adj Coll-HS Wage Gap, 1980-2000 4 x Adj (linear) Return to Schl, 1980-2000 Intercept Root MSE R-squared Observations Controls Education Group? Industry Mix x Broad Education?1 State Effects? City Controls?2 F Emp Rate x Broad Ed?3 -0.231 (0.059) (2) -0.128 (0.049) (3) -0.163 (0.047) (4) -0.141 (0.046) (5) -0.207 (0.057) (6) -0.167 (0.053) (7) -0.151 (0.054) -0.149 (0.080) -0.114 (0.005) 0.049 0.662 1,150 Y 2 Manuf +Index Y Y Y (8)

-0.078 (0.012) 0.054 0.577 1,150 Y N N N N

-0.098 (0.009) 0.052 0.613 1,150 Y 2 Manuf N N N

-0.066 (0.009) 0.046 0.491 1,150 Y Detailed (Individual) N N N

-0.095 (0.009) 0.051 0.617 1,150 Y 2 Manuf +Index N N N

-0.083 (0.011) 0.050 0.655 1,150 Y 2 Manuf +Index Y N N

-0.091 (0.010) 0.049 0.660 1,150 Y 2 Manuf +Index Y Y N

-0.094 (0.010) 0.049 0.664 1,150 Y 2 Manuf +Index Y Y Y

Data Source: 5% Public-use 1980 (Ruggles et al., 2010) and 2000 public use censuses of population, w hich are aggregated to 230 metropolitan areas, and, for the dependent variable, x 5 education groups (high school dropouts, graduates, some college <4 years, 4 years college, and advanced degree) that are the units of observation in the regressions. Sample consists of w orkers age 16-65 w ith positive potential w ork experience (age - years of education -6) and hours w orked last year, not residing in group quarters,w ho are currently employed, w ith positive w age earnings but zero business and farm earnings and not currently enrolled in school. Dependent variable is the gap betw een the average ln hourly w age of men and w omen w ith the same broad education level, adjusted in gender x education (x year) specific regressions for a quartic in potential experience, linear education (for dropouts, some college, and advanced degree) and dummies for black, Hispanic, foreign-born, and born after 1950 (w ith latter also interacted w ith years education for the same three groups). Independent variable is the average of (simililarly adjusted) college-high school w age gaps for men and w omen. All regressions in table w eighted by 1980 population. Standard errors, in parentheses under slope estimates, are computed to be robust to arbitrary forms of heteroskedasticity and error correlation w ithin metropolitan area. 1Columns (2) and (4)-(8) control for the share of employment in non-durable and durable manufacturing their interactions w ith a dummy for being in the "some college" or low er broad education group. Column (3) adds to the indvidual-level w age adjustments controls for detailed census of population industries that corresponds roughly to the level of industry detail in 1980. Columns (4) and higher control for a female demand "Index" w hich is the female share of employment obtained by averaging national female employment shares by detailed industry using local employment by industry as w eights. This index is constructed separately for "college equivalents" (w hich is is used as the control for broad education categories four years college and advanced degrees) and "high school equivalents" (w hich is is used as the control for broad education categories dropouts, high school and some college) but entered as a single control variable. 2City controls are share black, Hispanic, and foreign-born; and the unemployment rate and the natural log of the labor force. 3Female employment rates entered separately for those w ith some college or less and four years college or more.

Table 3. Change in Wage Gaps on Human Capital/Raw Labor, 1980 (1) (2) (3) (4) (5) -0.932 (0.315) 0.667 1.287 (0.389) 0.604 1,150 Y 2 Manuf +Index Y Y Y

A. Change in Adjusted Male-Female Wage Gap, 1980-2000 Human Capital/ -0.528 -0.804 -0.891 -1.048 Raw Labor, 1980 (0.142) (0.177) (0.252) (0.316) R-squared 0.578 0.630 0.661 0.666

B. Change in Adjusted College-High School Wage Gap, 1980-2000 Human Capital/ 0.993 1.248 1.258 1.267 Raw Labor, 1980 (0.276) (0.381) (0.289) (0.351) R-squared Observations Controls Education Group? Industry Mix x Broad Education?1 State Effects? City Controls?2 F Emp Rate x Broad Ed?3 0.207 1,150 Y N N N N 0.274 1,150 Y 2 Manuf +Index N N N 0.588 1,150 Y 2 Manuf +Index Y N N 0.603 1,150 Y 2 Manuf +Index Y Y N

Data Source: 5% Public-use 1980 (Ruggles et al., 2010) and 2000 Censuses of Population, w hich are aggregated to 230 metropolitan areas (and, for the dependent variable in panel A, x 5 education groups -- high school dropouts, graduates, some college <4 years, 4 years college, and advanced degree). Sample for constructing independent variable consists of w orkers age 16-65 w ith positive potential w ork experience (age - years of education -6) and hours w orked last year, not residing in group quarters. "Human capital" is the sum of the interaction of hours w orked w ith years of education beyond 10 times 0.077, and "raw labor" is 1.423 times the sum of male hours w orked plus the sum of female hours w orked. The w age sample is the subsample w ho are currently employed, w ith positive w age earnings but zero business and farm earnings and not currently enrolled in school. Dependent variable in panel A is the (change in the) gap betw een the average ln hourly w age of men and w omen w ith the same broad education level, w here the adjustment are gender x education (x year) specific regressions on a quartic in potential experience, linear education (for dropouts, some college, and advanced degree) and dummies for black, Hispanic, foreign-born, and born after 1950 (w ith latter also interacted w ith years education for the same three groups). Dependent variable in panel B is the average of (simililarly adjusted change in the) college-high school w age gaps for men and w omen. All regressions in table w eighted by 1980 population. Standard errors, in parentheses under slope estimates, are computed to be robust to arbitrary forms of heteroskedasticity and error correlation w ithin metropolitan area. 1Columns (2)-(5) control for the share of employment in non-durable and durable manufacturing their interactions w ith a dummy for being in the "some college" or low er broad education group, and for a female demand "Index" w hich is the female share of employment obtained by averaging national female employment shares by detailed industry using local employment by industry as w eights. This index is constructed separately for "college equivalents" (w hich is is used as the control for broad education categories four years college and advanced degrees) and "high school equivalents" (w hich is is used as the control for broad education categories dropouts, high school and some college) but entered as a single control variable. 2 City controls are share black, Hispanic, and foreign-born; and the unemployment rate and the natural log of the labor force. 3Female employment rates entered separately for those w ith some college or less and four years college or more.

Table 4a. Change in (Adjusted) Male-Female Wage Gaps, 1980-2000, vs. Personal Computer Adoption Intensity (1) Personal Computers/ Worker, 2000 Change in Hum Cap/ RL, 1980-2000 Root MSE R-squared (2) (3) (4) -0.115 (0.061) -0.192 (0.365) 0.049 0.659 (5) -0.089 (0.066) -0.355 (0.408) 0.049 0.663

A. Ordinary Least Squares -0.204 -0.207 -0.157 (0.048) (0.041) (0.052) 0.821 0.417 0.032 (0.180) (0.172) (0.338) 0.053 0.591 0.051 0.626 0.050 0.654

Personal Computers/ Worker, 2000 Change in Hum Cap/ RL, 1980-2000 Root MSE Observations First Stage F-Stat p-value Controls Education Group? Industry Mix x Broad Education?1 State Effects? City Controls?2 F Emp Rate x Broad Ed?3

B. Instrumental Variables -0.239 -0.363 -0.367 (0.063) (0.081) (0.095) 0.845 0.547 0.589 (0.188) (0.203) (0.367) 0.053 1,150 136.5 0.000 Y N N N N 0.051 1,150 107.9 0.000 Y 2 Manuf +Index N N N 0.050 1,150 93.8 0.000 Y 2 Manuf +Index Y N N

-0.482 (0.152) 0.371 (0.466) 0.051 1,150 74.1 0.000 Y 2 Manuf +Index Y Y N

-0.440 (0.158) 0.178 (0.479) 0.050 1,150 63.1 0.000 Y 2 Manuf +Index Y Y Y

Data Source: Dependent variable and human capital -- 5% Public-use 1980(Ruggles et al., 2010) and 2000 Censuses of Population, w hich are aggregated to 230 metropolitan areas, and, for the dependent variable, x 5 education groups (high school dropouts, graduates, some college <4 years, 4 years college, and advanced degree) that are the units of observation in the regressions. Sample used to compute human capital/raw labor, consists of w orkers age 16-65 w ith positive potential w ork experience (age - years of education -6) and hours w orked last year, and not residing in group quarters. "Human capital" is the sum of the interaction of hours w orked w ith years of education beyond 10 times 0.077, and "raw labor" is 1.423 times the sum of male hours w orked plus the sum of female hours w orked. The w age sample is the subsample w ho are currently employed, w ith positive w age earnings but zero business and farm earnings and not currently enrolled in school; dependent variable is the change in the adjusted w age gap betw een the average ln hourly w age of men and w omen w ith the same broad education level, w here the adjustment are gender x education (x year) specific regressions on a quartic in potential experience, linear education (for dropouts, some college, and advanced degree) and dummies for black, Hispanic, foreign-born, and born after 1950 (w ith latter also interacted w ith years education for the same three groups). PCs per w orker computed from stacked 2000 and 2002 marketing surveys by Harte Hanks, w hich are regression adjusted for 3 digit industry x employer size dummies. Instrument is 1980 human capital/raw labor. All regressions in table w eighted by 1980 population. Standard errors, in parentheses under slope estimates, are computed to be robust to arbitrary forms of heteroskedasticity and error correlation w ithin metropolitan area. 1Columns (2)-(5) control for the share of employment in non-durable and durable manufacturing their interactions w ith a dummy for being in the "some college" or low er broad education group, and for a female demand "Index" w hich is the female share of employment obtained by averaging national female employment shares by detailed industry using local employment by industry as w eights. This index is constructed separately for "college equivalents" (w hich is is used as the control for broad education categories four years college and advanced degrees) and "high school equivalents" (w hich is is used as the control for broad education categories dropouts, high variable. 2City controls are share black, Hispanic, school and some college) but entered as a single control and foreign-born; and the unemployment rate and the natural log of the labor force. 3Female employment rates entered separately for those w ith some college or less and four years college or more.

Table 4b. Change in Adjusted College-High School Wage Gaps, 1980-2000, vs. Personal Computers Adoption Intensity (1) Personal Computers/ Worker, 2000 Change in Hum Cap/ RL, 1980-2000 Root MSE R-squared (2) (3) (4) 0.210 (0.062) 0.850 (0.538) 0.030 0.598 (5) 0.204 (0.063) 0.938 (0.520) 0.030 0.602

A. Ordinary Least Squares 0.306 0.256 0.246 (0.089) (0.088) (0.063) -1.242 -1.117 0.748 (0.258) (0.277) (0.525) 0.041 0.227 0.040 0.253 0.030 0.584

Personal Computers/ Worker, 2000 Change in Hum Cap/ RL, 1980-2000 Root MSE Observations First Stage F-Stat p-value Controls Education Group? Industry Mix x Broad Education?1 State Effects? City Controls?2 F Emp Rate x Broad Ed?3

B. Instrumental Variables 0.463 0.526 0.467 (0.139) (0.185) (0.119) -1.349 -1.341 0.160 (0.288) (0.346) (0.544) 0.041 1,150 136.5 0.000 Y N N N N 0.042 1,150 107.9 0.000 Y 2 Manuf +Index N N N 0.032 1,150 93.8 0.000 Y 2 Manuf +Index Y N N

0.544 (0.148) 0.338 (0.521) 0.032 1,150 74.1 0.000 Y 2 Manuf +Index Y Y N

0.567 (0.175) 0.387 (0.525) 0.032 1,150 63.1 0.000 Y 2 Manuf +Index Y Y Y

Data Source: Dependent variable and human capital -- 5% Public-use 1980 (Ruggles et al., 2010) and 2000 Censuses of Population, w hich are aggregated to 230 metropolitan areas. Sample used to compute human capital/raw labor, consists of w orkers age 16-65 w ith positive potential w ork experience (age - years of education -6) and hours w orked last year, and not residing in group quarters. "Human capital" is the sum of the interaction of hours w orked w ith years of education beyond 10 times 0.077, and "raw labor" is 1.423 times the sum of male hours w orked plus the sum of female hours w orked. The w age sample is the subsample w ho are currently employed, w ith positive w age earnings but zero business and farm earnings and not currently enrolled in school; dependent variable is the change in the average of male and female ln hourly w age gaps betw een college and high school w orkers, w here the adjustments are gender x education (x year) specific regressions on a quartic in potential experience, linear education (for dropouts, some college, and advanced degree) and dummies for black, Hispanic, foreign-born, and born after 1950 (w ith latter also interacted w ith years education for the same three groups). PCs per w orker computed from stacked 2000 and 2002 marketing surveys by Harte Hanks, w hich are regression adjusted for 3 digit industry x employer size dummies. Instrument is 1980 human capital/raw labor. All regressions in table w eighted by 1980 population. Standard errors, in parentheses under slope estimates, are computed to be robust to arbitrary forms of heteroskedasticity and error correlation w ithin metropolitan area. 1Columns (2)-(5) control for the share of employment in non-durable and durable manufacturing their interactions w ith a dummy for being in the "some college" or low er broad education group, and for a female demand "Index" w hich is the female share of employment obtained by averaging national female employment shares by detailed industry using local employment by industry as w eights. This index is constructed separately for "college equivalents" (w hich is is used as the control for broad education categories four years college and advanced degrees) and "high school equivalents" (w hich is is used as the control for broad education categories dropouts, high school and some college) but entered as a single control variable. 2City controls are share black, Hispanic, and foreign-born; and the unemployment rate and the natural log of the labor force. 3Female employment rates entered separately for those w ith some college or less and four years college or more.

Table 5. Male-Female Wage Gaps vs. College-High School Wage Gaps, Estimated by Instrumental Variables (1) Adj Coll-HS Wage Gap, 1980-2000 Intercept Root MSE R-squared Observations Controls Education Group? Industry Mix x Broad Education?1 State Effects? City Controls?2 F Emp Rate x Broad Ed?3 -0.532 (0.206) -0.020 (0.038) 0.0555 0.549 1,150 Y N N N N (2) -0.644 (0.288) 0.001 (0.054) 0.0557 0.549 1,150 Y 2 Manuf +Index N N N (3) -0.708 (0.279) 0.014 (0.053) 0.0523 0.617 1,150 Y 2 Manuf +Index Y N N (4) -0.827 (0.383) 0.036 (0.073) 0.0535 0.601 1,150 Y 2 Manuf +Index Y Y N (5) -0.724 (0.365) 0.017 (0.069) 0.0522 0.620 1,150 Y 2 Manuf +Index Y Y Y

Data Source: 5% Public-use 1980 (Ruggles et al., 2010) and 2000 public use censuses of population, w hich are aggregated to 230 metropolitan areas, and, for the dependent variable, x 5 education groups (high school dropouts, graduates, some college <4 years, 4 years college, and advanced degree) that are the units of observation in the regressions. Sample consists of w orkers age 16-65 w ith positive potential w ork experience (age - years of education -6) and hours w orked last year, not residing in group quarters,w ho are currently employed, w ith positive w age earnings but zero business and farm earnings and not currently enrolled in school. Dependent variable is the gap (in panel A, or change in the gap, in panel B) betw een the average ln hourly w age of men and w omen w ith the same broad education level, w here the adjustment are gender x education (x year) specific regressions on a quartic in potential experience, linear education (for dropouts, some college, and advanced degree) and dummies for black, Hispanic, foreign-born, and born after 1950 (w ith latter also interacted w ith years education for the same three groups). Independent variable is the average of (simililarly adjusted) college-high school w age gaps for men and w omen. The instrument is 1980 "human capital/raw labor"; first stage and further description of this variable are found in the notes to tables 3a and 4a. All regressions in table w eighted by 1980 population. Standard errors, in parentheses under slope estimates, are computed to be robust to arbitrary forms of heteroskedasticity and error correlation w ithin metropolitan area. 1Columns (2)-(5) control for the share of employment in non-durable and durable manufacturing their interactions w ith a dummy for being in the "some college" or low er broad education group, and for a female demand "Index" w hich is the female share of employment obtained by averaging national female employment shares by detailed industry using local employment by industry as w eights. This index is constructed separately for "college equivalents" (w hich is is used as the control for broad education categories four years college and advanced degrees) and "high school equivalents" (w hich is is used as the control for broad education categories dropouts, high school and some college) but entered as a single control variable. 2City controls are share black, Hispanic, and foreign-born; and the unemployment rate and the natural log of the labor force. 3Female employment rates entered separately for those w ith some college or less and four years college or more.

Table 6. Correcting for Selection Female Subsample Predicted NatBn,NonEmp. Rates Hs Whites w/no kids<6 High1 (2) (3) Sel Corr: Adjusted for Inverse Mills Ratio Estimated Estimated Nationally By MSA (4) (5) -0.354 (0.318) 0.312 0.247

(1)

A. Change in Adjusted Male-Female Wage Gap, 1980-2000, OLS -0.150 -0.085 -0.335 -0.185 Adj Coll-HS Wage Gap, 1980-2000 (0.070) (0.079) (0.101) (0.072) Root MSE R-squared 0.047 0.684 0.075 0.512 0.096 0.422 0.050 0.895

B. Change in Adjusted Male-Female Wage Gap, 1980-2000, IV -0.610 -0.452 -0.924 -0.490 Adj Coll-HS Wage Gap, 1980-2000 (0.317) (0.276) (0.401) (0.282) Root MSE Observations Controls Education Group? Industry Mix x Broad Education?2 State Effects? City Controls?3 0.049 795 Y 2 Manuf +Index Y Y 0.075 795 Y 2 Manuf +Index Y Y 0.097 795 Y 2 Manuf +Index Y Y 0.050 795 Y 2 Manuf +Index Y Y

-0.884 (0.775) 0.313 795 Y 2 Manuf +Index Y Y

Data Source: 5% Public-use 1980 (Ruggles et al., 2010) and 2000 public use censuses of population, w hich are aggregated to 230 metropolitan areas, and, for the dependent variable, x 5 education groups (high school dropouts, graduates, some college <4 years, 4 years college, and advanced degree) that are the units of observation in the regressions. Sample consists of w orkers age 16-65 w ith positive potential w ork experience (age - years of education6) and hours w orked last year, not residing in group quarters,w ho are currently employed, w ith positive w age earnings but zero business and farm earnings and not currently enrolled in school. Dependent variable is the gap betw een the average ln hourly w age of men and w omen w ith the same broad education level, adjusted in gender x education (x year) specific regressions for a quartic in potential experience, linear education (for dropouts, some college, and advanced degree) and dummies for black, Hispanic, foreign-born, and born after 1950 (w ith latter also interacted w ith years education for the same three groups). Independent variable is the average of (simililarly adjusted) college-high school w age gaps for men and w omen. All regressions in table w eighted by 1980 population. Standard errors, in parentheses under slope estimates, are computed to be robust to arbitrary forms of heteroskedasticity and error correlation w ithin metropolitan area. The instrument, used in panels B, is 1980 "human capital/raw labor," w hich is discussed in notes to previous tables. 1Women w ith characteristics associated w ith a greater than 60 chance of w orking in the average metropolitan area (as of 1980). To determine this, a probit for being in the w age sample (among w omen age 16-65 w ith positive potential w ork experience not residing in group quarters) on the same variables as the w age adjustment plus dummies for marital status interacted w ith a dummy for children under six, w as separately estimated by gender and education using the 1980 data. The predicted values, p, from this probit (evaluated at the average of the metropolitan area effects) w as computed and w omen w ith p>0.6 w ere retained. 2 The share of employment in non-durable and durable manufacturing and their interactions w ith a dummy for being in the "some college" or low er broad education group, and a female demand "Index" w hich is the female share of employment obtained by averaging national female employment shares by detailed industry using local employment by industry as w eights. The index is constructed separately for "college equivalents" (w hich is is used as the control for broad education categories four years college and advanced degrees) and "high school equivalents" (w hich is is used as the control for broad education categories dropouts, high school and some college) 3City controls are share black, Hispanic, and foreign-born; and the unemployment rate and the natural log of the labor force.

Table 7. Change in Male-Female Wage Gap by Period, Large MSA subsample Analysis Period: 1970-1980 (1) 1980-1990 (2) 1980-2000 (3) -0.198 (0.071) 0.745 -1.088 (0.436) 0.749 685 1980-2000 Y 2 Manuf +Index Y Y

A. Versus Change in College-High School Wage Gap Change in College-High 0.126 -0.250 School Wage Gap (0.105) (0.061) R-squared Human Capital/ Raw Labor, 1980 R-squared Observations Wage Methods:1 Controls Education Group? Industry Mix x Broad Education?2 State Effects? City Controls?3 0.409 0.582

B. Versus 1980 Human Capital/Raw Labor -0.377 -1.380 (0.409) (0.295) 0.408 685 1970 Y 2 Manuf +Index Y Y 0.592 685 1970 Y 2 Manuf +Index Y Y

Data Source: 1970 public-use "county group," 1980 5% public use, (both Ruggles et al., 2010) and 1990 and 2000 5% public use censuses of population, w hich are aggregated to 137 metropolitan areas x 5 education groups (high school dropouts, graduates, some college <4 years, 4 years college, and advanced degree) that are the units of observation in the regressions. Sample for human capital/raw labor consists of w orkers age 16-65 w ith positive potential w ork experience (age - years of education -6) and hours w orked last year, not residing in group quarters; w age sample further limited to those w ho are currently employed, w ith positive w age earnings but zero business and farm earnings and not currently enrolled in school. Dependent variable is the gap betw een the average ln hourly w age of men and w omen w ith the same broad education level, adjusted in gender x education (x year) specific regressions for a quartic in potential experience, linear education (for dropouts, some college, and advanced degree) and dummies for black, Hispanic, foreign-born, and born after 1950 (w ith latter also interacted w ith years education for the same three groups). Independent variable in Panel A is the average of (simililarly adjusted) college-high school w age gaps for men and w omen. Independent variable in Panels B is "human capital" -- the sum of the interaction of annual hours w orked w ith years of education beyond 10 times 0.077 -- divided by "raw labor" --- 1.423 times the sum of male annual hours w orked plus the sum of female hours w orked. All regressions in table w eighted by 1980 population. Standard errors, in parentheses under slope estimates, are computed to be robust to arbitrary forms of heteroskedasticity and error correlation w ithin metropolitan area. 1"1970" w age methods employ categorical versions of "hours last w eek" and "w eeks last year" available in the 1970 Census, interpolated to continuous hours and w eeks (using means in 1980 census data) to compute annual hours w orked. Column (3), like the other tables, uses reports of (continuous) "usual hours per w eek" and "w eeks last year" to compute annual hours w orked. 2The share of employment in non-durable and durable manufacturing and their interactions w ith a dummy for being in the "some college" or low er broad education group, and a female demand "Index" w hich is the female share of employment obtained by averaging national female employment shares by detailed industry using local employment by industry as w eights. The index is constructed separately for "college equivalents" (w hich is is used as the control for broad education categories four years college and advanced degrees) and "high school equivalents" (w hich is is used as the control for broad education categories dropouts, high school and some college) 3City controls are share black, Hispanic, and foreign-born; and the unemployment rate and the natural log of the labor force.

Appendix Table A1a. Alternative Estimate of Change in (Adjusted) Male-Female Wage Gaps, 19802000, vs. Personal Computer Adoption Intensity, Dropping Change in Hum Cap Control (1) (2) (3) (4) (5)

Personal Computers/ Worker, 2000 Root MSE R-squared Personal Computers/ Worker, 2000 Root MSE Observations First Stage F-Stats Personal Computers p-value Controls Education Group? Industry Mix x Broad Education?1 State Effects? City Controls?2 F Emp Rate x Broad Ed?3

A. Ordinary Least Squares -0.170 -0.185 -0.156 (0.062) (0.045) (0.052) 0.054 0.573 -0.285 (0.080) 0.054 1,150 119.9 0.000 Y N N N N 0.051 0.623 0.050 0.654

-0.123 (0.062) 0.049 0.659 -0.464 (0.163) 0.051 1,150 68.9 0.000 Y 2 Manuf +Index Y Y N

-0.105 (0.066) 0.049 0.662 -0.429 (0.169) 0.050 1,150 58.3 0.000 Y 2 Manuf +Index Y Y Y

B. Instrumental Variables -0.405 -0.337 (0.091) (0.102) 0.052 1,150 111.2 0.000 Y 2 Manuf +Index N N N 0.050 1,150 160.3 0.000 Y 2 Manuf +Index Y N N

Data Source: Dependent variable and human capital -- 5% Public-use 1980(Ruggles et al., 2010) and 2000 Censuses of Population, w hich are aggregated to 230 metropolitan areas, and, for the dependent variable, x 5 education groups (high school dropouts, graduates, some college <4 years, 4 years college, and advanced degree) that are the units of observation in the regressions. Sample used to compute human capital/raw labor, consists of w orkers age 16-65 w ith positive potential w ork experience (age - years of education -6) and hours w orked last year, and not residing in group quarters. "Human capital" is the sum of the interaction of hours w orked w ith years of education beyond 10 times 0.077, and "raw labor" is 1.423 times the sum of male hours w orked plus the sum of female hours w orked. The w age sample is the subsample w ho are currently employed, w ith positive w age earnings but zero business and farm earnings and not currently enrolled in school; dependent variable is the change in the adjusted w age gap betw een the average ln hourly w age of men and w omen w ith the same broad education level, w here the adjustment are gender x education (x year) specific regressions on a quartic in potential experience, linear education (for dropouts, some college, and advanced degree) and dummies for black, Hispanic, foreign-born, and born after 1950 (w ith latter also interacted w ith years education for the same three groups). PCs per w orker computed from stacked 2000 and 2002 marketing surveys by Harte Hanks, w hich are regression adjusted for 3 digit industry x employer size dummies. Instrument is 1980 human capital/raw labor. All regressions in table w eighted by 1980 population. Standard errors, in parentheses under slope estimates, are computed to be robust to arbitrary forms of heteroskedasticity and error correlation w ithin metropolitan area. 1Columns (2)-(5) control for the share of employment in nondurable and durable manufacturing their interactions w ith a dummy for being in the "some college" or low er broad education group, and for a female demand "Index" w hich is the female share of employment obtained by averaging national female employment shares by detailed industry using local employment by industry as w eights. This index is constructed separately for "college equivalents" (w hich is is used as the control for broad education categories four years college and advanced degrees) and "high school equivalents" (w hich is is used as the control for broad education categories dropouts, high school and some college) but entered as a single control variable. 2City controls are share black, Hispanic, and foreignborn; and the unemployment rate and the natural log of the labor force. 3Female employment rates entered separately for those w ith some college or less and four years college or more.

Appendix Table A1b. Alternative Estimates of Change in Adjusted College-High School Wage Gaps, 1980-2000, vs. Personal Computers Adoption Intensity, Dropping Change in Hum Cap Control (1) Personal Computers/ Worker, 2000 Root MSE R-squared Personal Computers/ Worker, 2000 Root MSE Observations First Stage F-Stats Personal Computers p-value Controls Education Group? Industry Mix x Broad Education?1 State Effects? City Controls?2 F Emp Rate x Broad Ed?3 (2) (3) (4) 0.246 (0.079) 0.030 0.589 0.561 (0.143) 0.032 1,150 68.9 0.000 Y 2 Manuf +Index Y Y N (5) 0.246 (0.081) 0.030 0.591 0.592 (0.165) 0.032 1,150 58.3 0.000 Y 2 Manuf +Index Y Y Y

A. Ordinary Least Squares 0.255 0.197 0.287 (0.104) (0.096) (0.063) 0.044 0.095 0.042 0.171 0.031 0.576

B. Instrumental Variables 0.535 0.629 0.475 (0.162) (0.206) (0.108) 0.047 1,150 119.9 0.000 Y N N N N 0.047 1,150 111.2 0.000 Y 2 Manuf +Index N N N 0.032 1,150 160.3 0.000 Y 2 Manuf +Index Y N N

Data Source: Dependent variable and human capital -- 5% Public-use 1980 (Ruggles et al., 2010) and 2000 Censuses of Population, w hich are aggregated to 230 metropolitan areas. Sample used to compute human capital/raw labor, consists of w orkers age 16-65 w ith positive potential w ork experience (age - years of education -6) and hours w orked last year, and not residing in group quarters. "Human capital" is the sum of the interaction of hours w orked w ith years of education beyond 10 times 0.077, and "raw labor" is 1.423 times the sum of male hours w orked plus the sum of female hours w orked. The w age sample is the subsample w ho are currently employed, w ith positive w age earnings but zero business and farm earnings and not currently enrolled in school; dependent variable is the change in the average of male and female ln hourly w age gaps betw een college and high school w orkers, w here the adjustments are gender x education (x year) specific regressions on a quartic in potential experience, linear education (for dropouts, some college, and advanced degree) and dummies for black, Hispanic, foreign-born, and born after 1950 (w ith latter also interacted w ith years education for the same three groups). PCs per w orker computed from stacked 2000 and 2002 marketing surveys by Harte Hanks, w hich are regression adjusted for 3 digit industry x employer size dummies. Instruments is1980 human capital/raw labor. All regressions w eighted by 1980 population. Standard errors, in parentheses under slope estimates, are computed to be robust to arbitrary forms of heteroskedasticity and error correlation w ithin metropolitan area. 1Columns (2) and higher control for a female demand "Index" w hich is the female share of employment obtained by averaging national female employment shares by detailed industry using local employment by industry as w eights. This index is constructed separately for "college equivalents" (w hich is is used as the control for broad education categories four years college and advanced degrees) and "high school equivalents" (w hich is is used as the control for broad education categories dropouts, high school and some college) 2City controls are share black, Hispanic, and foreign-born; and the unemployment rate and the natural log of the labor force. 3Female employment rates entered separately for those w ith some college or less and four years college or more.

Appendix Table A2. Male-Female Wage Gaps vs. College-High School Wage Gaps, Estimated by Instrumental Variables (PCs/Worker as Instrument) (1) Adj Coll-HS Wage Gap, 1980-2000 Intercept Root MSE Observations Controls Education Group? Industry Mix x Broad Education?1 State Effects? City Controls?2 F Emp Rate x Broad Ed?3 -0.668 (0.279) 0.006 (0.052) 0.057 1,150 Y N N N N (2) -0.940 (0.491) 0.058 (0.092) 0.062 1,150 Y 2 Manuf +Index N N N (3) -0.542 (0.215) -0.018 (0.041) 0.051 1,150 Y 2 Manuf +Index Y N N (4) -0.500 (0.309) -0.027 (0.059) 0.050 1,150 Y 2 Manuf +Index Y Y N (5) -0.425 (0.310) -0.041 (0.059) 0.050 1,150 Y 2 Manuf +Index Y Y Y

Data Source: 5% Public-use 1980 (Ruggles et al., 2010) and 2000 public use censuses of population, w hich are aggregated to 230 metropolitan areas, and, for the dependent variable, x 5 education groups (high school dropouts, graduates, some college <4 years, 4 years college, and advanced degree) that are the units of observation in the regressions. Sample consists of w orkers age 16-65 w ith positive potential w ork experience (age - years of education -6) and hours w orked last year, not residing in group quarters,w ho are currently employed, w ith positive w age earnings but zero business and farm earnings and not currently enrolled in school. Dependent variable is the gap (in panel A, or change in the gap, in panel B) betw een the average ln hourly w age of men and w omen w ith the same broad education level, w here the adjustment are gender x education (x year) specific regressions on a quartic in potential experience, linear education (for dropouts, some college, and advanced degree) and dummies for black, Hispanic, foreign-born, and born after 1950 (w ith latter also interacted w ith years education for the same three groups). Independent variable is the average of (simililarly adjusted) college-high school w age gaps for men and w omen. The instrument is 2000 adjusted PCs/w orker. Standard errors, in parentheses under slope estimates, are computed to be robust to arbitrary forms of heteroskedasticity and error correlation w ithin metropolitan area. 1Columns (2)-(5) control for the share of employment in non-durable and durable manufacturing their interactions w ith a dummy for being in the "some college" or low er broad education group, and for a female demand "Index" w hich is the female share of employment obtained by averaging national female employment shares by detailed industry using local employment by industry as w eights. This index is constructed separately for "college equivalents" (w hich is is used as the control for broad education categories four years college and advanced degrees) and "high school equivalents" (w hich is is used as the control for broad education categories dropouts, high school and some college) but entered as a single control variable. 2City controls are share black, Hispanic, and foreign-born; and the unemployment rate and the natural log of the labor force. 3Female employment rates entered separately for those w ith some college or less and four years college or more.