the e ects of human capital depreciation on occupational ... · the e ects of human capital...
TRANSCRIPT
The E�ects of Human Capital Depreciation on
Occupational Gender Segregation
Hsueh-Hsiang (Cher) Li∗
January 21, 2013
Abstract
This paper analyzes how human capital depreciation a�ects occupational gender
segregation. Prior studies are generally biased because, given an occupational de-
preciation rate, female workers endogenously choose the duration of leave. I address
this problem by proposing an alternative depreciation measure utilizing involun-
tary job displacement shocks. Using this depreciation proxy along with additional
pecuniary and non-pecuniary occupational attributes, I estimate a conditional logit
model of occupational choices separately for male and female college graduates. My
results show that women have a stronger distaste than men for occupations with
high human capital depreciation.
Keywords: human capital, occupational segregation, occupational choices, gender in-
equality, job interruptions, female labor supply.
∗Department of Economics, University of Wisconsin-Madison, 1180 Observatory Drive, Madison, WI53706. Email: [email protected]
1 Introduction
Despite high labor force participation rates, career interruptions are still common among
women. Over 41 percent of college-educated women were out of work for more than 6
months at some point, and 23.4 percent had out-of-work spells that totaled two years
or more in the �fteen years after receiving a baccalaureate degree. Only 14.1 percent
of college-educated men were out of work for at least 6 months at some point, and
3.1 percent had accumulated out-of-work spells of more than 2 years (Goldin, 2006).1
Because women are more likely than men to take career breaks to care for their families,
they may choose occupations that have lower human capital depreciation to mitigate
the wage losses from job interruptions. This consideration is likely to be important
for occupations with frequent knowledge-updating requirements, such as the sciences,
engineering, and �nance.2 The leave patterns and depreciation consideration likely a�ect
the dispersions of men and women distributions across occupations. Since occupational
segregation results in persistent wage disparities between men and women, it is crucial
to understand the driving forces of segregation (Groshen, 1991; Blau and Kahn, 2000;
Bayard et al., 2003).3
Polachek (1981) provides a theoretical framework and empirical evidence to show that
women, anticipating spending time out of the labor force, rationally choose occupations
with lower wage losses during home-time. England (1982) argues that Polachek's (1981)
human capital depreciation hypothesis does not explain occupational gender segregation
1Table 1 of Goldin (2006) delineates the out-of-work pattern of the entering class of 1976 from theCollege and Beyond dataset of the Andrew W. Mellon Foundation.
2McDowell (1982) estimates how the durability of knowledge in�uences female career decisions inacademia using the the age pro�le of cited works, i.e. the durability of prior published research papersas cited in current research. He �nds that the costs of career interruptions to be largest in physics andchemistry, and lowest in history and English; and the proportion of women in a �eld is directly relatedto the durability of the research knowledge in the areas.
3Groshen (1991) �nds that after controlling for other forms of segregation, occupational segregationaccounts for up to 26 percent of the gender wage gap in services industry. Blau and Kahn (2000) �ndthat occupation, industry and unionism together explained 29 percent of the total gender gap. Bayardet al. (2003) �nd that male-female wage gap is lowered by 25 percent when they control for broadoccupation and industry categories.
1
for two reasons. First, she �nds that wage losses during home-time is similar for both
male- and female-dominated occupations. Second, occupational choices are similar for
women with continuous employment histories and women who interrupt their careers.
The Polachek-England debate led to several follow-up studies. For instance, following
England's (1982) speci�cation, Robst and VanGilder (2000) estimate an insigni�cant
coe�cient on the interaction between the percent of female workers in a given occupation
and home-time. However, their results change when they replace the continuous home-
time variable with an indicator of recent home-time in the prior year. They �nd that
the home-time penalty is higher in occupations with low female representation. Recent
research that uses German data also shows contradictory results. Kunze (2002) focuses
on workers who participated in apprenticeship training and �nds that the depreciation
rate is higher in female occupations. In contrast, Görlich and de Grip (2009) �nd that
the short-run depreciation rate is higher in male occupations than in female occupations
in the high-skilled labor market.
The inconclusive �ndings are likely to occur because these studies neglect occupa-
tional sorting on other attributes. Gender di�erences in the preferences for occupational
attributes, for instance, can also contribute to occupational segregation. Zafar (2009) and
Wiswall and Zafar (2011) elicit subjective expectation information from college students
and �nd that the pecuniary incentive is a more important determinant of college major
choices for men than for women, and non-pecuniary outcomes (e.g. enjoying coursework)
dominate female major choices. Fortin (2008) also �nds that gender di�erences in valuing
money and family play a modest but signi�cant role in accounting for the gender wage
gap. Hence, sorting on other occupational attributes may mask the e�ect of sorting on
human capital depreciation if these e�ects cancel each other out.
Another important issue missing from prior studies is the endogeneity between de-
preciation and home-time. All of these studies take the occurrence or the length of job
interruptions of a worker as a given and regress wages on job interruptions to obtain the
2
human capital depreciation rates. Although low depreciation occupations may attract
women who anticipate long career interruptions, it could also be the case that being in a
low depreciation occupation induces one to take a longer leave. Selection bias also arises
because observations of female wage changes are only available for women who return
to the labor force.4 If human capital depreciation rates are in fact higher for female-
dominated jobs than for male-dominated jobs, leavers from these female-dominated jobs
are likely to remain non-working because higher wage penalties push wage o�ers below
their reservation wages. Therefore, the human capital depreciation estimates may be
biased downward for female-dominated jobs without accounting for selection.
Although researchers �nd ample evidence showing that women's shorter work histo-
ries explain a substantial proportion of the gender wage gap (Mincer and Polachek, 1974;
Beblo and Wolf, 2002; O'Neill, 2003; Hotchkiss and Pitts, 2007; Bertrand, et al., 2010),
little research analyzes the e�ect of human capital depreciation on the gender wage gap.
My research �lls these gaps and contributes to the literature in three primary re-
spects. First, I propose an alternative human capital depreciation measure that uses
wage losses from unemployment following involuntary job displacements of male work-
ers. Although involuntary job terminations are arguably exogenous, any bias in my
depreciation measure due to the selection of male displaced workers (e.g. plants closed
in declining industries) is unlikely to cause the result of gender di�erences in responses to
depreciation. Second, I estimate the e�ect of human capital depreciation on occupational
choices by gender while controlling for selection on other occupational characteristics,
including wage levels, wage growth, work hours, and occupational tasks (e.g. cognitive
4Polachek (1981) estimates atrophy as the coe�cient of home-time in a �rst-di�erenced wage equationand regress the atrophy rates on home-time in the second stage. He recognizes the simultaneity issuein the second stage and addresses the problem with the following exclusion restrictions for home-time:potential labor market experience, number of children less than 18, an urban area dummy, a good orexcellent health dummy, the age of the youngest child, and an index of labor market demand. Validityof some of these instruments may be questionable�for instance, fertility (number and age of children)can be endogenous to occupational human capital depreciation rates since women in high depreciationoccupations may choose to delay fertility to lower her demand for home-time. More importantly, usinginstruments in the second stage does not address the issue of selection bias in human capital depreciationestimation from the �rst stage.
3
tasks, manual tasks). The inclusion of additional occupational attributes helps recon-
cile the seemingly contradictory results presented in previous studies because I allow
multi-dimensional occupational sorting. Third, I quantify to what extent human capital
depreciation explains the gender wage gap.
This paper investigates how human capital depreciation contributes to occupational
gender segregation in the high-skill labor market. The causes of gender segregation
likely di�er across labor market segments. For instance, occupational gender segregation
in the low-skill labor market may be largely explained by male comparative advantages
in physical strengths, which may be much less pertinent in the high-skill market. To
focus on a more homogeneous group, I restrict my sample to college graduates. My
empirical approach is to estimate occupational choices in two stages. In the �rst stage,
I estimate occupation-speci�c human capital depreciation rates as wage changes due to
unanticipated work termination shocks by using the DisplacedWorkers Survey (DWS). In
the second stage, using the �rst stage depreciation estimates and additional occupational
attributes, I analyze the discrete occupational choices by college graduates from the
National Longitudinal Survey of Youth 1979 (NLSY79) by using a conditional logit
model (McFadden, 1974; DeLeire and Levy, 2004).5
My results provide evidence that women are more likely than men to choose occu-
pations with low depreciation rates. This result is robust to alternative human capital
depreciation measures. My counterfactual analysis shows that human capital depreci-
ation explains 7 percent of the initial gender wage gap upon labor market entry, and
35 percent of the gender wage gap ten years later. In Section 2, I provide a heuristic
theoretical framework. Section 3 delineates my empirical strategies, followed by data de-
scriptions in Section 4. Section 5 presents my estimation results, followed by sensitivity
analyses and a counterfactual analysis in Section 6 and 7. I discuss the results and my
5The conditional logit model (McFadden, 1974) is a generalized multinomial model widely used toestimate how characteristics of options a�ect individual choices. For instance, DeLeire and Levy (2004)use this model to estimate the willingness to trade safety hypothesizing that risk averse individuals sortinto safer jobs.
4
�ndings in Section 8. Section 9 concludes.
2 Theoretical Framework
To understand how human capital depreciation in�uences individual occupational choices,
I consider a simple three-period model. Suppose there are only two occupations�A and
B�in a perfectly competitive labor market, and skills are nontransferable between these
two occupations. Occupation A has a higher human capital depreciation rate than does
occupation B. Individuals face occupational choices in the beginning of the �rst period.
In the �rst period, individuals choose one occupation and receive its corresponding
wage (i.e. wA or wB). In the second period, individuals may be hit by a random fertility
shock and the probability of having children is p. If individuals have children, they
withdraw from the labor market for a period. If individuals do not have children, they
continue working and receive the wage o�er wj + gj in the second period, where gj
indicates the wage growth in one period from the occupation j = {A,B}. In the last
period, children leave, and individuals work again in the same occupation they initially
chose. If individuals leave the labor market in the second period, their wage o�ers upon
returning to work in the third period are wA(1− δ) + gA for occupation A, and wB + gB
for occupation B, where δ represents human capital depreciation for job A and δ ∈ (0, 1].
Note that the depreciation is normalized to 0 for occupation B.
Assume that individuals are risk-neutral and choose an occupation to maximize their
discounted life-time earnings. With the real interest rate at r, the expected net present
value of the life-time earnings E(WA
)from occupation A is:
E(WA
)= wA + (1− p)
[1
1 + r
(wA + gA
)+
1
(1 + r)2(wA + 2gA
)]+p
{1
(1 + r)2[(1− δ)wA + gA
]}. (1)
5
The net present value of the life-time earnings from occupation B is:
E(WB
)= wB + (1− p)
[1
1 + r
(wB + gB
)+
1
(1 + r)2(wB + 2gB
)]+p
{1
(1 + r)2[wB + gB
]}. (2)
Individuals solve this optimization problem by choosing occupation j∗ that satis�es
this condition:
j∗ = argmaxj∈{A,B}{E(WA
), E(WB
)}. (3)
The decision rule is that individuals choose occupation A (i.e. j∗ = A) if the human
capital depreciation δ is below the threshold δ∗, and B otherwise. The decision rules are:
j∗ = A, ifδ < δ∗
j∗ = B, ifδ > δ∗, (4)
where
δ∗ =1
pwA[(1 + r)2 + (1− p) (1 + r) + 1
] (wA − wB
)+
1
pwA[(1− p) (1 + r) + (2− p)]
(gA − gB
), (5)
Equation 5 shows that a higher wage premium or wage growth for occupation A in-
creases the threshold δ∗. As a result, the likelihood of individuals choosing occupation A
increases (i.e. the cumulative probability function F (δ < δ∗) is increasing in(wA − wB
)and
(gA − gB
)). An equilibrium of non-zero employment for both occupations requires
that occupation A has either a higher wage level or higher wage growth (or both) to
guarantee that some workers choose occupation A.
The probability of fertility shocks also a�ects individuals' occupational decisions. The
6
partial derivative of the depreciation threshold δ∗ with regards to the fertility probabil-
ity is negative (i.e. ∂δ∗
∂p< 0), indicating that a higher fertility probability lowers the
depreciation threshold δ∗. As a result, the likelihood of individuals choosing occupation
A declines as fertility probability increases.
There are two empirical implications. First, if men are una�ected by the fertility
shocks (i.e. p = 0 for men), they continue working in the second period and the de-
preciation drops out from their decisions rules. Second, among women, if the fertility
expectations vary across individuals, low fertility women are more likely than high fertil-
ity women to choose occupation A. In other words, two otherwise identical women may
choose di�erent occupations if their subjective fertility expectations are di�erent. The
e�ects of fertility across and within gender groups may contribute to segregation.
3 Empirical Strategies
3.1 First Stage: Human Capital Depreciation
In order to examine how human capital depreciation a�ects men and women's occupa-
tional choices di�erently, the �rst step is to acquire depreciation estimates that are not
endogenous to female workers' decisions. I address this issue by proposing an alternative
measure using data from the Displaced Workers Survey (DWS). I estimate wage losses
incurred by involuntary work terminations as a proxy for human capital depreciation.
Human capital may depreciate because individuals become rusty at work after job dis-
ruptions or their skills are outdated due to the technological advancement. I broadly
de�ne human capital depreciation during job interruptions to include both skill deterio-
ration due to non-use and knowledge obsolescence due to technological changes (Rosen,
1975; De Grip and Van Loo, 2002). I consider the total e�ect of human capital depre-
ciation without distinguishing the mechanism because workers consider the total wage
losses during their separation from employment regardless of the causes.
7
My estimation relies on displaced male workers because most men maintain contin-
uous work histories. Even in the case of job terminations, they quickly return to work
when a new match is formed. Displaced female workers, on the other hand, are much
less ideal candidates for depreciation estimation because they, anticipating future leave,
may choose low depreciation jobs. Therefore, the fraction of women who return to work
after displacement (and hence are observable) may be highly correlated with depreciation
by occupation. This female selection may bias depreciation estimates, and therefore, I
restrict my sample to male workers in the DWS.
One problem with estimating depreciation by using the data from displaced workers
is that wage losses tend to be higher for displacement than for a voluntary leave because
employers may perceive displacement as a signal of low productivity. If the probability of
getting displaced di�ers across occupations, being displaced from an occupation with a
low displacement rate may send an even stronger signal of low productivity, which could
be associated with larger wage losses. The DWS and the Current Population Survey
(CPS) data show that, with a few exceptions, the probability of displacement is similar
across occupations. Computer analysts and construction workers have a higher proba-
bility of being displaced, while managers, social workers, educators, protective service
providers, and administrative sta� face a lower likelihood of displacement. However,
wage losses do not appear to be correlated with the propensity for displacement (i.e.
the correlation is -0.01), indicating that the stigma e�ect is likely to be universal across
occupations. Although this stigma e�ect cannot be separately identi�ed, it does not
distort the relative human capital depreciation rates, and does not bias my occupational
choice estimates.
Displaced workers in the data provide wage information from their pre-displacement
jobs and their current jobs, as well as the year and duration of the displacement. The
wage changes between these two jobs encompass two components: wage losses during
displacement and wage growth from experience accumulation in the new job. Figure 1
8
Figure 1: Wage Changes for Displaced Workers
provides a graphic illustration of wage changes after job displacement. Human capital
depreciation is the rate of wage losses, represented by δ, during the period of displacement
(denoted as �leave� on the �gure). Estimating human capital depreciation rates recovers
the slope (δ) of wage losses during job interruptions.
The duration between wage observations varies across individuals in the data from
the DWS. Assume that education is completed before an individual started working.
The wage change between the displacement shock (at time t) and the observation from
the new job (at time t+ k, when the wage information in the new job is available) is:
∆kwijt = −δjLeaveij + ∆kgj(Expijt) + β∆kZit + ∆kεit, (6)
where Leaveij is the duration of individual i′s job displacement from occupation j,
∆kgj(.) is an occupation-speci�c function of wage growth in experience, Expijt is the
accumulation of experience, ∆kZit is a vector of additional covariates including indi-
cators for industry switches (Neal, 1995; Sullivan, 2010), loss of union status (Topel,
1990; Stevens, 1997), and the year of displacement to account for year-speci�c e�ects of
9
displacement, and ∆kεit is the change in i.i.d. random shocks.6
My primary interest is to recover δj from the wage observations of workers who work
in the same occupation both before and after the displacement.7 Since Leaveij and
Zijt are observable, the parameter of interest, δj, can be identi�ed if ∆kgj(Expijt) is
also known. French, Mazumder, and Taber (2006) estimate the return to experience of
low-skilled workers by using a linear approximation. It is well justi�ed in their analy-
sis since their research focuses on young workers. However, a linear approximation is
inappropriate for this study because the age distribution of the workers in my data is
disperse and the wage pro�le is highly non-linear. Since the return to experience is not
the focus of this paper, I do not specify the functional form for the wage growth func-
tion ∆kgj(Expijt), instead allowing it to be �exible. In order to consistently estimate
∆kgj(Expijt), I need data from workers who work continuously in the same occupation
without job interruptions. However, the DWS does not have information on continuous
workers' experience and wages prior to the survey date, it can not be used to estimate
∆kgj(Expijt). Alternatively, I estimate wage growth by using the matched Current Pop-
ulation Survey March Supplements (CPS). I exploit the CPS �outgoing rotation groups�
design which allows me to observe the same individuals across a one-year period and cre-
ate one-year panels of continuous workers across the years that match the DWS sample
period.8 In this �outgoing rotation groups� design, the CPS interviews each household in
the sample for four consecutive months, drops them out for the next eight months, and
6The importance of industry-speci�c human capital is found in prior research (Neal, 1995; Sullivan,2010). Evidence from previous papers (Topel, 1991; Stevens, 1997) also show that unionized workerswho lose their union status incur additional wage losses.
7The selection problem arises because I focus on workers who were re-employed in the same occupa-tion. In an alternative speci�cation, I include a probit selection equation using workers' pre-displacementindustries, occupations, age, race, reasons and the year of displacement to obtain the Inverse Mills ratioas an additional regressor in the wage equation. The depreciation estimates do not change substan-tially. Because the exclusion restrictions in the selection equation are weak and the results suggest thatthis selection bias is not the main driving force behind the wage loss results, I report the unadjusteddepreciation estimates.
8I used the CPS March Supplement from 1991 to 2008 to match the DWS sample period. The DWSsurveyed respondents' displacement events up to three years prior to the survey year, and hence theearliest displacement observations for 1994 DWS data is in 1991.
10
then interviews them again for four more months. Therefore, a substantial proportion of
individuals in the sample are observed twice in the same month across two consecutive
years. By matching individuals across years, the sample size is large enough to estimate
non-parametrically age-speci�c wage growth rates by occupation.9
Once ∆kgj(Expijt) is estimated, and Leaveij and ∆lZijt are observable, δj is identi-
�ed.10 I assume that wage losses are linear in displacement durations since the maximum
span of displacement from this sample is no longer than three years and the sample size
is too small to non-linearly estimate occupation-speci�c depreciation.
3.2 Second Stage: Discrete Occupational Choices
In the second stage, I use the depreciation estimates obtained from the �rst stage to inves-
tigate how human capital depreciation a�ects men's and women's occupational choices.
An individual faces occupational choices among J alternatives in each of T time
periods. The utility that individual i derives from choosing the alternative j in time
t is Uijt = γYjt + eijt, where Yjt is a vector of observed occupational characteristics
(including wage levels, wage growth rates, work hours, tasks required by occupations),
and eijt follows the i.i.d. extreme value distribution. The probability that individual i
chooses alternative k in period t is:
Likt(γ) =eγYkt∑j e
γYjt(7)
Note that any individual characteristics that are invariant by choice alternatives drop
out from the above equation.
Let j∗ denote the alternative that person i chose in period t. The probability of
individual i's observed sequence of choices is the product of standard logits∏
t Lij∗t.
9I use the second-order Epanechnikov kernel with the Silverman Rule-of-Thumb bandwidth for thisnon-parametric estimation of age-speci�c wage growth as a proxy for experience-speci�c wage growth.
10The linear assumption is due to data restrictions. If a large sample is available, this assumption canbe relaxed and age- or experience-speci�c estimates would be feasible.
11
The log likelihood function of the sample is the summation of the individual log
likelihood:
LL =∑i
ln∏t
Lij∗t, (8)
and this model can be estimated by maximum likelihood. Since I am interested in
how human capital depreciation and other job characteristics di�erently a�ect male and
female occupational choices, I estimate the model separately for men and women, using
the work histories from the National Longitudinal Survey of Youth 1979 (NLSY79).
Since the second stage estimation is based on a generated regressor from the �rst
stage, the second-stage standard errors are biased downward without accounting for es-
timation errors from the �rst stage. To correct for the additional source of the variance
from the �rst stage, I re-estimate both stages with bootstrap samples. I �rst repeatedly
estimate depreciation rates in the �rst stage with bootstrap samples from the DWS.
For each depreciation estimate from the �rst-stage bootstrap, I draw a block bootstrap
sample that preserves individual work sequences and re-estimate the second-stage con-
ditional logit model. Standard errors are calculated by using the parameter estimates
over these bootstrap samples.11 The corrected standard errors are expected to be larger
after accounting for the �rst-stage estimation errors.
4 Data
I use data from four di�erent sources for the analysis. In the �rst stage, I obtained
pecuniary and non-pecuniary occupational characteristics, including wage levels, wage
growth rates, usual work hours, depreciation, and tasks required at work from the Current
Population Surveys March Supplements (CPS), the Displaced Workers Survey (DWS),
and the Dictionary of Occupational Titles (DOT). In the second stage, I use variables
obtained from the �rst stage to model individual occupational choices by using the
11I bootstrap 1,000 replications for both stages.
12
Table 1: Summary of Data SourcesOccupational Characteristics Dataset Year Selection Criteria
Stage 1
Wage Growth by Age Merged CPS 1991-2008 Full-time male continuous
1-Yr Panel workers, same Job
HC Depreciation Rates DWS 1994-2008 Full-time male,
same Occupation
Wage level, Usual Hours CPS 1979-2008 All full-time
Occupational Attributes DOT91 1991 All full-time
Stage 2
Individual Work History NLSY79 1979-2008 Complete bachelor
degrees by 1990
National Longitudinal Survey of Youth 1979 (NLSY79). Table 1 summarizes the data
sources.
4.1 Human Capital Depreciation
My primary interest in the �rst stage is to obtain human capital depreciation rates by
occupation. The most comprehensive source of information on the costs of job loss in the
U.S. is the Displaced Workers Survey (DWS). I use the DWS biennial surveys from 1994
to 2008 to estimate wage losses during job displacement as a proxy for human capital
depreciation. These surveys are supplements to the Current Population Survey (CPS),
and they collect information on workers who were displaced from their jobs. Displaced
workers are de�ned as persons 20 years of age and older who lost their jobs because their
plants or companies closed or moved, there was insu�cient work for them to do, or their
positions or shifts were abolished. Respondents are asked if they were displaced at any
point during the three calendar years prior to the survey date. These surveys also collect
wage information from the pre-displacement job and the current job along with the
duration of unemployment following displacement. However, because the information on
work hours is unavailable for most observations to create reliable hourly wage rates, I use
the weekly earnings of full-time workers for estimation. By restricting my sample to male
13
workers who work full-time in the same occupation after displacement, I obtained 2,389
observations for the human capital depreciation estimation.12 Depreciation estimates are
based on 22 broad occupational categories (excluding the military) following the 2010
Standard Occupational Classi�cation.13
Before estimating wage losses during job interruptions, I need to parse out the wage
growth component. Since wage growth rates vary substantially at di�erent career stages,
consistent estimates of experience-speci�c (or alternatively using age as a proxy for ex-
perience for male workers) wage growth rates are required. The small size of the DWS
subsample does not allow precise estimates of wage growth by age, and hence I use the
CPS data instead. I create one-year panels of continuous workers and use their one-year
wage changes to estimate annual wage growth by occupation.
Madrian and Lefgren (1999) point out that matching the CPS observations across
periods is challenging for reasons including non-response, mortality, migration, and re-
porting errors. Since the CPS sampling is based on household levels, the same identi�-
cation codes may be assigned to di�erent individuals across the times when individuals
move out and new residents move into the household. To ensure correct matching, I
use observations from wave 1-4 for the base year, and match the same individuals in
wave 5-8 in the following year. I match individuals across a one-year period by using
their household ID, personal line number, sex, and race. I am able to identify 83,920
males who worked full-time and full-year in the same occupation over a one-year period
12This DWS sample includes all male workers across educational levels. The small sample size pro-hibits me from estimating occupation-speci�c depreciation for college graduates. When I include inter-action terms between displacement by occupations and an indicator for college graduates, the coe�cientestimates of the interaction terms are not statistically di�erent from zero, indicating no systematic dif-ferences in depreciation across educational groups. The average annual wage loss across occupationsis 11.7 percent for college graduates, similar to the estimate of 11.4 percent for all male workers inmy sample. The invariance of depreciation estimates across education levels allows me to use all maleworkers in the sample for the �rst stage estimates. In the second stage model, I turn to focus on collegegraduates' occupational choices.
13Source: http://www.bls.gov/soc/major_groups.htm. Note that a few adjustments are made toallow comparability across di�erent occupational codes across four decades. These adjustments includecategorizing �rst-line supervisors and managers in food, sales, building maintenance occupation to themanagerial occupation since most of these workers are classi�ed as managers according to Census 1970occupational codes that are used by NLSY79 up to year 2000.
14
between 1991 and 2008 to match the displacement years from the DWS.14 I use the
one-year panels to estimate age-speci�c annual wage growth rates by occupation. After
teasing out this wage growth component, I am able to identify the depreciation rates by
occupation.
4.2 Occupational Characteristics
Workers do not consider depreciation rates in isolation. Although workers may prefer a
low depreciation occupation, they may still choose a high depreciation job if the �nancial
incentive is large enough to o�set the losses from high depreciation. The �nancial rewards
come from two primary sources: wage levels and future wage growth. Suppose there is
a trade-o� between entry wage levels and future wage growth. Workers' choices may
depend on their expected work patterns. If a worker anticipates working brie�y and
leaving the labor force permanently afterward, an occupation that o�ers a high entry
wage with a low growth prospect may be optimal. However, if the worker temporarily
withdraws from the employment, he or she may choose between a high entry wage and
rapid wage growth, depending on their relative magnitudes.
Wage data from the CPS show a positive 0.08 correlation between entry wages and
early wage growth for college graduates.15 My baseline depreciation estimates are weakly
correlated with entry wage levels (0.09), but the depreciation estimates are strongly
correlated with early wage growth (0.42). These correlations indicate that higher earnings
are associated with higher human capital depreciation. Particularly, high depreciation
occupations tend to o�er greater wage growth prospects. In order to investigate how
individuals choose occupations considering these pecuniary outcomes, I test my models
14The �rst wave of the DWS used in this paper is from 1994, and it collects retroactive displacementdata back to 1991.
15Using the CPS March Supplement from 1979 to 2008, I calculate the average entry level wage ofnew male workers in the age between 22 and 25, who completed at least four year college education butdid not obtain any graduate degrees beyond the bachelor level. Early wage growth by occupation comesfrom the wage di�erences between these new entrants and male workers with 10 years of experiences(use age as a proxy) who meet the same educational criteria.
15
with two sets of measures. The �rst set of wage measures uses entry-level wages and the
10-year wage growth by occupation. This measure emphasizes the initial wage o�er and
the wage growth in the early career stage, during which most workers gain substantial
wage increases. For the second set of wage measures, I use experience-speci�c wages
and one-year wage growth that evolves as individuals gain experiences. Therefore, this
second measure allows individuals with di�erent work histories to face di�erent wage
conditions over time.
Individuals also sort into di�erent occupations based on heterogeneous tastes and
preferences for a variety of occupational attributes. Particularly, women di�er from men
in work arrangements. For instance, working women on average work fewer hours than
men do on paid work. Bertrand, Goldin, and Katz (2009) show that male and female
MBA graduates have nearly identical incomes in the beginning of their careers, but the
earnings soon diverge primarily due to female career interruptions and shorter weekly
hours when children are present.16 Working mothers may be willing to trade wages for
shorter working hours to care for their children.17 If predominantly female occupations
coincidentally have low depreciation rates and the low work-hour intensities, the e�ect
of depreciation rates on occupational choices will be overstated. To account for sorting
on work-hour intensities, I use the average usual weekly work hours of full-time workers
from the CPS between 1979 and 2008.
In order to control for other occupational attributes, I obtain job characteristic mea-
sures from the DOT. The data are constructed by the U.S. Department of Labor to
provide standardized occupational information across industries. The information on
job characteristics is primarily obtained via on-site observations; for jobs that are di�-
cult to observe, information is collected from professional and trade associations. The
16Their sample consists of MBAs from the Booth School of Business of the University of Chicago from1990 to 2006.
17The trade-o� between wages and hours is explained in Altonji and Paxson (1988). There is alsoevidence of increasing work hours for highly educated and highly paid men from 1979 to 2006 (Kuhnand Lozano, 2008).
16
latest version of the DOT, released in 1991, lists 63 attributes, including worker functions
(data, people, things), vocational training required to perform the job, aptitudes, tem-
peraments, interests, physical demands, and work environments. In order to reduce the
dimensions of occupational characteristics, I conduct a factor analysis of these 63 DOT
variables to obtain �ve orthogonal factors that account for 74 percent of the variation in
occupational characteristics.18 These �ve factors can be primarily categorized by: cog-
nitive tasks, motor tasks, interpersonal tasks, physical demands, and hot or cold work
environments. The factor loadings of these �ve variables are provided in Table 2. These
measures are rescaled to follow the standard normal distribution for the interpretation
convenience.19 Table 3 summarizes these job characteristics by occupation.
4.3 National Longitudinal Survey of Youth 1979
In the second stage, I estimate individual occupational choices by using data from the
NLSY79, along with occupational characteristics obtained from the �rst stage. The long
panel of NLSY from 1979 to 2008 is suitable for this study since it covers almost the
entire fecundity period for the female sample. Since women are more likely to interrupt
careers for family reasons, an ideal dataset for this analysis should cover the period
when changes of family formation (e.g., getting married, having children) are likely to
occur.20 Furthermore, in the early waves, the NLSY79 surveys individuals' subjective
expectations for future work status at age 35. The subjective expectations for future
18Following Yamaguchi (2012), I use the April 1971 Current Population Survey augmented by thefourth edition of the DOT with the updated occupational characteristics from the revised fourth versionin 1991.
19The original scales di�er substantially across measures. For instance, the aptitude measures are: 1�The top 10% of the population�, 2 �The highest third exclusive of the top 10% of the population�, 3 �Themiddle third of the population�, 4 �The lowest third exclusive of the bottom 10% of the population�,and 5 �The lowest 10% of the population�. Physical demands and work environments are indicatorof the presence of such characteristics. Prior to conducting the factor analysis, I convert the originalmeasures to a uniform range between 0 and 1, with 1 representing higher requirements of such tasks.The �nal factor analysis results are rescaled to follow the standard normal distribution for interpretationconvenience.
20Respondents between 14 and 21 years old were sampled in 1979. By 2008, the age of this cohort isbetween 43 and 50 years old.
17
events are important if individuals, given di�erent human capital depreciation rates
associated with occupations, choose early occupations based on such expectations.
The NLSY79 is a nationally representative sample of 12,686 young men and women
who were 14-22 years old when they were �rst surveyed in 1979. These individuals were
interviewed annually through 1994 and were interviewed on a biennial basis thereafter.
To focus primarily on college graduates with continuous educational history, I restrict
my sample to high-skilled workers who obtain their bachelor's degrees as their termi-
nal degrees prior to 1990 to allow su�cient observations of post-college work histories.
I exclude individuals with graduate degrees to construct a more homogeneous sample
with the same level of educational attainment. Initially, 1,203 college graduates meet
the sample selection criteria, but 79 observations are dropped due to the lack of work
information throughout the panel. The �nal sample consists of 1,124 individuals, ap-
proximately half of whom are male. By 2008, due to funding constraints and sample
attrition, my sample has 755 individuals, including 380 men and 375 women.
To remove temporary jobs, I restrict work histories to post-college occupations af-
ter individuals make a full-time occupational transition, which is de�ned as working in
an occupation for at least 1,500 hours a year over a one-year period.21 The summary
statistics of the demographics, occupational attributes and distributions of the �rst job
are detailed in Table 4. Women and men do not di�er in terms of race and ability mea-
sured by the AFQT scores, but women are 0.2 years younger than men at their �rst job.
However, the occupational distributions are distinct for men and women. Predominantly
male occupations include construction, installation and repair, transportation, farming,
legal services, and engineering. Women are highly concentrated in health-care support,
health-care practicing, personal services, education, and administrative support.
The labor force participation rate remained high for both men (97%) and women
(85.5%) in 2008, although the majority of the workers experienced certain non-working
21I construct the work histories by using the weekly work status and the work hours variables.
18
spells (81% for men, and 94% for women).22 Job interruption patterns also diverged
for men and women and are detailed in Table 5. Although a substantial fraction of
the men experienced job interruptions, 49 percent had total spells of job interruptions
less than one year, and only 6.3 percent had an interruption history for over three
years. In contrast, only 34.3 percent of women interrupted their work for less than
one year, while 38.9 percent left for longer than three years. By 2008, the average
non-working duration among those with job interruptions were 1.28 and 4.85 years for
men and women, respectively. The pooled sample includes 14,874 individual-year work
observations. Repeated occupational choices by individuals throughout the panel are
useful in accounting for unobserved persistent individual preferences.
5 Results
Human capital depreciation estimates from Equation (6) are between -.56 to 1.65 log
points. The depreciation point estimates are summarized in Table 6. The distribution of
observations by occupations is uneven from the DWS sample, and the small sample size
for some occupations renders the OLS coe�cients imprecise.23 Outliers are likely to be
in�uential when the sample size is small. To alleviate the e�ect of outliers, I also run a
median regression for the wage equation. The OLS and median regression results are both
reported in Table 6. Although the estimates vary substantially for some occupations,
the relative rankings of human capital depreciation rates are mostly preserved. I use
22Percentage of individuals with non-working spells is high because the non-working spells include alljob interruptions regardless of the length of spells. Hence, if an individual leaves his or her old jobs,and takes an one-week break before starting working at the new jobs, this one-week break during thejob transition is counted towards the non-working spells.
23Since there is only one male observation for health-care support occupation, separately estimatingthe depreciation rate for this occupation is infeasible. Hence, I apply the depreciation rate for �health-care practitioners� occupation to �health-care support service� occupation. Health-care support serviceinclude occupations such as registered nurses, and medical assistants; while health-care practitionersinclude physicians, dentists, pharmacists, and medical technicians. If the wage penalty rates di�er forthese two groups, it is more likely that the wage depreciation rate is larger for health practitioners,and hence this approximation would bias my estimates towards 0. In this case, the e�ect of the humancapital depreciation is more likely to be underestimated in my analysis.
19
the mean human capital depreciation rates from the OLS as the primary measure in the
second stage.
The depreciation estimates are inversely correlated (-.22) to the fraction of women
in the occupations. Occupations with high human capital depreciation estimates from
both the OLS and median regressions include: management, social services, legal ser-
vices, sales, administrative support, and transportation. The high depreciation rates for
this set of occupations suggest that wage losses are not entirely driven by technological
advancement. For instance, losing business connections is more likely than technological
changes to cause the large wage losses among managerial, sales, and legal occupations.
On the other hand, computers and new technology are likely to replace routine jobs (e.g.
administrative services), and hence cause drastic wage reductions. Autor et al. (2003)
�nd similar results of the decreasing wage rates of routine manual jobs. Business and
�nance, architecture and engineering, and arts and media show moderate wage reduc-
tions. Occupations that do not su�er wage losses, but are instead associated with wage
gains (i.e. δj estimates are negative) during job interruptions, include computer analysis,
education, health care, protective services, food services, and farming occupations. Most
of these wage gain estimates are not statistically signi�cant. However, the large (-.56
log points) wage gain associated with job interruptions from the education occupation
is statistically signi�cant at the .10 level.
This statistically signi�cant wage gain for educators is puzzling. If educators lose
their jobs because of school closings, they may need to move to other geographical areas
to �nd similar jobs. If that is the case, the wage gain may simply be an artifact from
the wage changes due to migration. To account for migration, I include an indicator for
migration and re-estimate depreciation rates. The results show even larger wage gains
for educators after controlling for migration. It indicates that migration is unlikely to
be the cause of the wage gains.
Another potential explanation is that wages increase because workers switch to sectors
20
that pay higher wages. All of the individuals in the education occupation were displaced
from the public sector. After displacement, one-third of the individuals switched to
the private sector. After including an indicator for sector switches, the depreciation
estimate also shows larger wage gains for educators. Further tests, such as trimming
outliers and removing short displacement, change the magnitude of the depreciation
estimate for the education occupation; regardless, education remains at the bottom of
the depreciation estimates. It is likely that teachers' salaries are protected against losses
from job displacement because their pay scales are tied to experience. If the wage
cannot decrease, the likelihood of observing positive wage gains after displacement would
increase.
Since results from the sensitivity analyses indicate that education is among the oc-
cupations with the lowest human capital depreciation rates, I keep the original OLS
estimates without rescaling. As long as the relative di�erences of depreciation estimates
by occupation are robust, the qualitative results of the occupational choice estimates will
not be a�ected in the logit framework. The average annual wage loss across all occu-
pation is 11 percent, which is in line with �ndings from the previous literature (Ruhm,
1991; Jacobson et al., 1993; Farber, 1993).24
Conditional logit estimates in the second stage show evidence of sorting on human
capital depreciation rates by gender. Results with the �rst set of wage measures, includ-
ing new entrants' wages and 10-year wage growth, are presented in the upper panel in
Table 7. The coe�cients on the entry wage level are positive for both men and women,
indicating that pecuniary incentives are important determinants of occupational choices.
The coe�cient estimate is much larger for men (2.93) than for women (.99). This result
is consistent with prior �ndings that pecuniary motivations are stronger for men than
24Ruhm (1991) �nds that the weekly earnings of displaced workers were 16 percent lower than thoseof non-displaced workers in the year following displacement, and still 14 percent lower four years afterdisplacement. Jacobson et al. (1993) �nd that after �ve years upon the separation, average quarterlyearnings losses stood at 25 percent for displaced workers who were separated from long-tenured jobsin Pennsylvania. Farber (1993) �nds that relative to non-displaced workers, displaced workers su�eredwage losses of 11 percent during the two years immediately following displacement.
21
for women (Fortin, 2008; Zafar, 2009; Wiswall and Zafar, 2011). Men and women di�er
in their preferences for wage growth. Average male workers have a positive coe�cient on
wage growth (.4), although the estimate is not statistically signi�cant. Women, on the
other hand, have a negative coe�cient on wage growth (-.95). This puzzling coe�cient
estimate does not necessarily suggest that women dislike wage growth. This might be a
result of omitted variables, such as higher stress levels associated with wage growth and
the strong correlation between wage growth and depreciation.
Both men and women have a distaste for high human capital depreciation rates.
However, the average negative response to human capital depreciation rates is stronger
for women (-1.042) than for men (-.629). It is worth noting that correcting the standard
errors by accounting for the additional variance from the �rst stage is important. Par-
ticularly, the standard error of the depreciation coe�cient increases by four to �ve times
after the adjustment. Regardless, the women's negative coe�cient on depreciation re-
mains statistically signi�cant, but the men's coe�cient estimate is no longer statistically
di�erent from zero.
If women anticipate future family responsibilities and choose low depreciation occupa-
tions, there should be distinct patterns for women with and without such responsibilities.
In order to understand if human capital depreciation rates have heterogeneous e�ects
within gender, I allow the human capital depreciation rates to interact with the marital
status and parenthood. The conditional logit estimates show that the coe�cients on
the interaction terms are negative for women and positive for men. The diverging pat-
terns across gender indicate that married women and mothers are less likely than single
and childless women to choose high depreciation occupations. However, married men
and fathers are more likely than single and childless men to choose high depreciation
occupations. This di�erences imply that family plays di�erent roles for men and women.
If individuals, given di�erent family roles, have accurate expectations for their future
leave, continuous workers and job interrupters are expected to choose di�erent occupa-
22
tions. I examine whether individuals who take substantial leave (de�ned as accumulated
leave that is longer than six months) have a stronger distaste than continuous workers
for human capital depreciation. The coe�cient on the interaction between leavers and
depreciation is very close to zero. This result implies that there are no signi�cant di�er-
ences between female job disruptors and continuous workers. One potential explanation
is that the realization of an outcome may not always align with the individual's expec-
tation. Uncertainties about family formation and fertility may result in ex-post results
that di�er from ex-ante expectations. One key feature of the NLSY79 is that, in early
waves (from 1979 to 1984), it directly asked respondents if they expected to be working
at age 35. Using this information as a proxy for individual expectations for their future
leave, I compare the realized duration of job interruptions and the depreciation rates
of the �rst job chosen by individuals with di�erent expectations.25 The results show
that women who expected not to be working at age 35 took a similar duration of leave
as those who expected to work continuously. However, the women who expected not
to be working at age 35 chose a lower depreciation (.023) occupation as their �rst job.
In contrast, women who expected to be working at age 35 chose an initial occupation
with a higher depreciation rate (.092). The di�erence is statistically signi�cant at the
10-percent level. This di�erence in the initial occupational choices between women with
di�erent work continuity expectations provides evidence of sorting on expectations rather
than on realizations of career interruptions.
Men and women also show di�erent preferences for other occupational attributes.
Compared with men, women are less likely to work in occupations with long weekly
hours. In addition, women have stronger preferences than men for cognitive and inter-
personal tasks. Both high-skilled men and women prefer less motor-intensive jobs, and
this distaste is stronger among women. Moreover, women also show a strong distaste
25There are 87 women indicating that they expected not to be working at age 35, and 487 womenexpected to be working at age 35. This analysis focuses on women because only one male respondentindicated that he did not expect to be working at age 35.
23
for physically demanding occupations, while men hold a positive attitude about physical
demands. Both men and women prefer not to be working in a hot, humid, or cold work
environment.
The results of the conditional logit estimation based on the second set of wage mea-
sures (i.e. experience-speci�c annual log wages and annual wage growth rates) are re-
ported in the lower panel in Table 7. The coe�cient estimates change slightly for most
covariates, but the relative gender di�erences remain. The coe�cient on wage growth
changes substantially because this new measure replaces the �rst 10-year wage growth
with the experience-speci�c annual wage growth. The men's coe�cient on wage growth
remains insigni�cantly di�erent from zero, while the women's negative coe�cient is sta-
tistically signi�cant even after adjusting for the second-stage standard error.
6 Sensitivity Analyses
As I discussed earlier, wage losses during displacement may not necessarily be the same
for voluntary job terminations. However, as long as the relative dispersion of depreciation
estimates by occupation is preserved, the qualitative results of the occupational choice
estimates will not be a�ected. Hence, the exact point estimates of the depreciation play
a much less important role than their relative di�erences in this context.
One primary concern for my human capital depreciation estimates is related to the
insigni�cant coe�cient estimates for several occupations. To ensure that the gender
di�erences in attitudes towards human capital depreciation are not sensitive to these
insigni�cant coe�cient estimates, I change all insigni�cant coe�cients to zeros and re-
estimate the conditional logit model. The gender di�erence with regard to human capital
depreciation remains. This result indicates that this estimation di�erence across gender
is not driven by the insigni�cant depreciation estimates from the �rst stage.
In spite of my e�orts to verify the robustness of the wage gain for the education
24
occupation, the substantial wage gain during job interruptions is still alarming. To test
if the estimates are driven by these wage gains, I replace all the negative depreciation
estimates with zeros and re-estimate the conditional logit model. The gender di�erence
in the depreciation coe�cient maintains. A large wage gain estimate for the education
occupation does not appear to propel this gender disparity.
Another issue is that the sample size is small for some occupations, and the estimates
might be skewed by in�uential outliers. To address this concern, I estimate human capital
depreciation rates using median regression. The range of depreciation estimates from
the median regression is slightly narrowed. However, most of the relative rankings of
the depreciation rates are preserved. Using the depreciation estimates from the median
regression to run the conditional logit estimation results in a large gender di�erence in
preferences for human capital depreciation.
Results from the sensitivity analyses show that women are more likely than men
to avoid occupations with high human capital depreciation, and the results are robust
across a variety of alternative measures.
7 Counterfactual
In this section, I conduct a counterfactual analysis to investigate how much of the gender
gap can be explained by human capital depreciation. The CPS data from 1979 to 2008
show that college graduates' initial gender gap in earnings is 13 percent upon entering
the labor market. This gender gap widens to 28 percent in the ten years after college
graduation. To decompose the e�ect of human capital depreciation on wage disparities
across gender, I assume that wage levels and wage growth rates are the same within
occupations for both men and women.
I conduct this analysis by using NLSY79 data to estimate the e�ect of human capital
depreciation on the gender wage gap at two points: at labor market entry and 10 years
25
later. In the counterfactual scenario, I assign a zero depreciation rate to all occupations.
The choice of zeros is innocuous since a variable with any choice-invariant values will
drop out from the logit likelihood function. It shows that women would increase their
likelihood of choosing management, business and �nance, engineers, sales, legal services,
and administrative support as their occupations and reduce their likelihood of choosing
education and health-care occupations. Under the same wage o�er assumption, the
predicted gender wage gap at labor market entry is 3.51 percent when the human capital
depreciation is present. Removing the human capital depreciation reduces the initial
gender wage gap to 3.27 percent, equivalent to a 6.6 percent reduction of the initial wage
gap.
Assume that individuals do not change their work histories. The gender wage gap
is projected to be rising to 17.9 percent ten years later if human capital depreciates. In
the case without human capital depreciation, the gender wage gap is 11.6 percent. The
absence of depreciation narrows the wage gap by 34.9 percent. Among the 6.3 percentage
point wage gap reduction, 4.9 percentage points come from the lower wage reduction due
to the absence of depreciation, and 1.4 percentage points are attributed to changes in
occupational choices. Women's intermittent labor supply plays a much more important
role in explaining the gender wage gap than does the gender occupational sorting on
depreciation.26
8 Discussion
The missing piece from the Polachek-England debate is the fact that workers sort on
other occupational attributes. My results show that occupational sorting di�ers by gen-
der in a variety of dimensions. Although women are more likely than men to choose low
26Similarly, Adda et al. (2011) �nd that, compared to women without children, women with fertilityincur 64.4 percentage losses in their lifetime earnings. Labor supply di�erences contribute 49 percentagepoints to the losses. For the remaining 16 percentage points from wage factors, only 1.9 percentage pointsare attributed to depreciation.
26
depreciation occupations, women also prefer occupations with low physical demands.
Hence, although labor intensive occupations tend to have low depreciation rates, women
are still unlikely to choose these occupations. This is a plausible explanation for Eng-
land's (1982) non-�nding of a monotonic relationship between female representation and
depreciation rates across occupations.
England (1982) further argues that women with continuous work histories should be
indi�erent to human capital depreciation rates since they always stay in the labor force.
As a result, continuously working women should have a higher probability of choosing
high-depreciation (most likely predominantly male) occupations. However, she �nds that
continuously working women are no more likely than female job interrupters to choose
presumably high depreciation (male-dominated) occupations. My results also �nd that
the realizations of female job interruptions are not correlated with di�erent occupational
choices based on human capital depreciation. However, further evidence suggests that
it is expectations instead of realizations of job interruptions a�ecting female initial oc-
cupational choices. Women who anticipated not to be working in their mid-30s chose
initial occupations with lower depreciation rates, despite the fact that the realizations of
their job interruptions were not longer than those taken by women who expected to be
working continuously. This result suggests that individuals choose occupations based on
their ex-ante expectations that may appear suboptimal based on the ex-post outcomes.
The e�ects of human capital depreciation on female occupational choices are not
only found in the U.S. A recent study by Adda et al. (2011) focuses on western German
women with vocational training and �nds that human capital depreciation rates di�er
across three broadly de�ned occupations. They also �nd that women who expect to
be childless are more likely to choose occupations with higher depreciation rates. They
show that fertility reduces women's life-cycle earnings through reduced labor force par-
ticipation, and it also a�ects wages through career choices, depreciation rates, and less
accumulation of human capital. Although my research di�ers from theirs in method-
27
ology, data, samples, and the institutional context (e.g., more generous maternity and
parental leave policies in Germany than in the U.S.), my qualitatively similar �ndings of
female selection on human capital depreciation rates indicate that the e�ect of human
capital depreciation rates on female occupational choices is general, across borders and
working classes.
9 Conclusion
The high-skill labor market remains segregated by gender, in spite of increasing female
educational attainment and labor force attachment. Women are highly concentrated in
education and health-care professions but are underrepresented in scienti�c and engi-
neering �elds. The most salient di�erence between men and women in the labor market
is the time they spend out of the labor force. Women are still more likely than men
to take time o� from employment. These job interruptions not only reduce women's
earnings signi�cantly through foregone wages during the period of not working, but they
also a�ect female occupational choices through their expectations of future leave. An-
ticipating intermittent work, women may choose occupations with low human capital
depreciation to mitigate the wage losses during job interruptions.
Early studies typically take female job interruptions as a given, and estimate the
wage losses during the interruptions. Although women may choose low depreciation oc-
cupations because they expect to take longer lengths of leave, the observed results may
also be explained by reverse causality: women take longer time o� because the costs of
job disruptions are low. Furthermore, the likelihood of women returning to work after
job interruptions may also di�er across occupations; hence, estimating human capital
depreciation based on the censored observations may be biased. I propose an alterna-
tive depreciation measure using male workers' wage changes after displacement shocks
from the Displacement Workers Survey. This new measure is unlikely to be correlated
28
with female occupational selection on depreciation. My estimates show that education,
health-care professions, protective services, food services, and farming occupations are
associated with wage gains during job interruptions. However, management, social ser-
vices, legal services, sales, administrative support, and transportation occupations are
connected with high wage losses during job displacement. These relative human capital
depreciation estimates are robust against alternative measures.
Given the human capital depreciation estimates obtained from the �rst stage, I ana-
lyze a discrete occupational choices model. In this second stage, I also include additional
occupational attributes, such as wage levels, wage growth rates, usual work hours, and
the tasks required to perform the occupations. The conditional logit estimation results
show stronger female aversion to high depreciation occupations, and such distaste is par-
ticularly profound for married women and mothers. Exploiting female expectations for
future job interruptions at their young age, my results also show that the expectations of
future job discontinuities have stronger predictive powers than the realizations of career
disruptions for selection on human capital depreciation.
Finally, my counterfactual analysis indicates that, in the absence of human capital
depreciation, the gender wage gap narrows. This e�ect is larger in the 10 years after
college graduation than immediately upon labor market entry. However, the reduction
of the gender wage gap primarily comes from the direct wage e�ects in the absence of
depreciation. Occupational resorting contributes a smaller share in closing the wage gap.
My results bridge the seemingly contradictory empirical �ndings from prior studies. I
present evidence to show that women are more likely than men to choose low depreciation
jobs as the theoretical framework predicts. But after accounting for gender sorting
on human capital depreciation, a substantial gender wage gap remains unexplained.
My results suggest that human capital depreciation is one piece of the complex gender
disparity puzzle that provides fertile ground for future research.
29
References
[1] Adda, Jérôme, Christian Dustmann, and Katrien Stevens. (2011) �The Career Costs
of Children,� IZA Discussion Papers, No. 6201
[2] Autor, David, Frank Levy, and Richard Murnane. (2003) �The Skill Content of
Recent Technological Change: An Empirical Exploration,� Quarterly Journal of
Economics, 118(4), 1279-1334.
[3] Bayard, Kimberly, Judith Hellerstein, David Neumark, and Kenneth Troske. (2003)
"New Evidence on Sex Segregation and Sex Di�erences in Wages from Matched
Employee-Employer Data," Journal of Labor Economics, 21 (4), 887-922.
[4] Beblo, Miriam and Elke Wolf. (2002) �Wage Penalties for Career Interruptions: An
Empirical Analysis for West Germany,� ZEW Working Paper, 02-45
[5] Bertrand, Marianne, Claudia Goldin, and Lawrence F. Katz. (2010) �Dynamics
of the gender gap for young professionals in the �nancial and corporate sectors,�
American Economic Journal: Applied Economics, 2 (3), 228-255.
[6] Blakemore, Arthur E., and Stuart A. Low. (1984) �Sex di�erences in occupational
selection: The case of college majors,� The Review of Economics and Statistics, 66
(1), 157-163.
[7] Blau, Francine D., and Lawrence M. Kahn. (2000) �Gender di�erences in pay,� The
Journal of Economic Perspectives, 14 (4), 75-99.
[8] De Grip, Andries. and Jasper Van Loo. (2002). �The Economics of Skills Obsoles-
cence: A Review,� In A. De Grip, J. Van Loo, & K. Mayhew (Eds.), The Economics
of Skills Obsolescence, Research in Labor Economics 21, (pp.1-26). Amsterdam: JAI
Press.
30
[9] DeLeire, Thomas and Helen Levy. (2004) �Worker Sorting and the Risk of Death on
the Job,� Journal of Labor Economics, 22 (4), 925-953
[10] England, Paula. (1982) �The failure of human capital theory to explain occupational
sex segregation,� The Journal of Human Resources, 17 (3), 358-370.
[11] Farber, Henry S. (1993) "The Incidence and Costs of Job Loss, 1982-91. " Brookings
Papers: Microeconomics, 73-132.
[12] Fortin, Nicole. (2008) "The Gender Wage Gap among Young Adults in the United
States: The Importance of Money vs. People," Journal of Human Resources, 43,
886-920.
[13] French, Eric, Bhashkar Mazumder and Christoper Taber. (2006) "The Changing
Pattern of Wage Growth for Low Skilled Workers," in Blank, Danzinger, and Schoeni
(eds) Working and Poor: How Economic Challenges and Policy Changes Have A
ected Low Wage Workers, New York: Russell Sage Foundation.
[14] Gladden, Tricia and Christopher Taber. (2009) "The Relationship between Wage
Growth and Wage Levels," Journal of Applied Econometrics , 24 (6), 914-932.
[15] Goldin, Claudia. (2006) �The Quiet Revolution That Transformed Women's Em-
ployment, Education, and Family,� The American Economic Reviews, 96 (2), 1-21.
[16] Görlich, Dennis, and Andries de Grip. (2009) �Human capital depreciation during
hometime,� Oxford Economic Papers, 61, i98-i121.
[17] Groshen, Erica L. (1991) �The structure of the Female/Male wage di�erential: Is it
who you are, what you do, or where you work?� The Journal of Human Resources,
26 (3), 457-72.
31
[18] Hotchkiss, Julie L. and M. Melinda Pitts. (2007) "The Role of Labor Market In-
termittency in Explaining Gender Wage Di�erentials," The American Economic
Review, 97 (2), 417-421
[19] Jacobson, Louis S., Robert J. LaLonde, and Daniel G. Sullivan. (1993) �Earnings
Losses of Displaced Workers,� The American Economic Review, 83 (4), 685-709.
[20] Kambourov, Gueorgui, and Iourii Manovskii. (2009) �Occupational speci�city of
human capital,� International Economic Review, 50 (1), 63-115.
[21] Kimberly Bayard, Judith Hellerstein, David Neumark, and Kenneth Troske. (2003)
�New evidence on sex segregation and sex di�erences in wages from matched
Employee-Employer data,� Journal of Labor Economics, 21 (4), 887-922.
[22] Kunze, Astrid. (2002) �The Timing of Careers and Human Capital Depreciation,�
IZA Discussion Paper, 509, 1-24.
[23] Jacobson, Louis, Robert LaLonde, and Daniel Sullivan. (1993) "Earnings Losses of
Displaced Workers." American Economic Review, 83 (4), 685-709.
[24] Madrian, B. and L. J. Lefgren (1999). "A Note on Longitudnally Matching Current
Population Survey Respondents." NBER Working Paper T0247.
[25] McDowell, John M. (1982) �Obsolescence of knowledge and career publication pro-
�les: Some evidence of di�erences among �elds in costs of interrupted careers,� The
American Economic Review, 72 (4), 752-768.
[26] McFadden, Daniel. (1974) "Conditional Logit Analysis of Qualitative Choice Behav-
ior," in P. Zarembka (ed.), FRONTIERS IN ECONOMETRICS, 105-142, Academic
Press: New York.
32
[27] Mincer, Jacob, and Soloman W. Polachek. (1974) �Family Investments in Human
Capital: Earnings of Women,� Journal of Political Economy, 82 (2), Part 2, s76-
s108.
[28] Neal, Derek. (1995) �Industry-speci�c human capital: Evidence from displaced work-
ers,� Journal of Labor Economics, 13, 653�77.
[29] O'Neill, June. (2003) �The Gender Gap in Wages, circa 2000,� The American Eco-
nomic Review, 93 (2), 309-314.
[30] Peter Kuhn, and Fernando Lozano. (2008) �The expanding workweek? understand-
ing trends in long work hours among U.S. men, 1979�2006,� Journal of Labor Eco-
nomics, 26 (2), 311-343.
[31] Polachek, Solomon W. (1981) �Occupational self-selection: A human capital ap-
proach to sex di�erences in occupational structure,� The Review of Economics and
Statistics, 63 (1), 60-69.
[32] Robst, John, and Jennifer VanGilder. (2000) �Atrophy Rates in Male and Female
occupations," Economic Letters, 69 (3), 407-417.
[33] Rosen, Sherwin. (1975) Measuring the Obsolescence of Knowledge. In: F.T. Juster
(ed) Education, Income and Human Behaviour (pp. 199-232). New York: Carnegie
Foundation.
[34] Ruhm, Christopher J. (1991) "Are Workers Permanently Scarred by Job Displace-
ments?" American Economtic Review, 81 (1), 319-24.
[35] Stevens, Ann Hu�. (1997) �Persistent E�ects of Job Displacement: The Importance
of Multiple Job Losses,� Journal of Labor Economics, 15 (1), Part 1, 165-188
[36] Sullivan, Paul. (2010) �Empirical evidence on occupation and industry speci�c hu-
man capital,� Labour Economics, 17 (3) (06): 567-80.
33
[37] Topel, Robert. (1991) �Speci�c Capital, Mobility, and Wages: Wages Rise with Job
Seniority,� Journal of Political Economy, 99 (1), 145-176.
[38] Wiswall, Matthew and Basit Zafar. (2011) "Determinants of college major choice:
identi�cation using an information experiment," Sta� Reports 500, Federal Reserve
Bank of New York.
[39] Yamaguchi, Shintaro. (2012) �Tasks and heterogeneous human capital,� Journal of
Labor Economics, 30 (1), 1-53.
[40] Zafar, Basit. (2009) �College major choice and the gender gap,� Federal Reserve
Bank of New York, 364.
34
Table 2. DOT Factor Analysis, Factor Loadings
VariablesCognitive Task
Motor TaskInterpersonal Task
Physical Demands
Hot and Cold Environment
Data 0.8101 0.3578 0.1371 -0.0909 0.101People 0.7043 -0.0415 0.3016 0.1707 0.007Things -0.1075 0.704 -0.3754 0.0045 -0.1057Reasoning 0.854 0.3761 0.1078 0.0306 0.0003Math 0.786 0.4221 0.0703 -0.1413 0.0322Language 0.8819 0.279 0.1082 -0.0062 -0.0205Specific Vocational Preparation 0.7152 0.5622 0.151 -0.1259 0.0505Learning 0.8202 0.3066 0.1607 -0.0189 -0.0273Verbal 0.8993 0.1396 0.0423 -0.0331 -0.0081Numerical 0.7346 0.3137 -0.0555 -0.1094 -0.0459Spacial 0.0515 0.7658 0.0903 0.0011 -0.0065Form Perception 0.1803 0.7042 -0.2257 -0.0407 0.037Clerical Perception 0.7401 0.0234 -0.2404 0.0558 -0.092Motor Coordination -0.0961 0.5392 -0.4846 0.1487 -0.1854Finger Dexterity 0.0396 0.5106 -0.6067 0.0081 -0.069Manual Dexterity -0.4052 0.5779 -0.2798 0.0647 -0.0102Eye-Hand-Foot Coordination -0.4463 0.3594 0.3658 0.301 -0.1472Color Discrimination -0.1587 0.503 0.0192 0.3876 0.3365Directing 1 0.0179 0.4449 -0.0968 0.0728Repetitive Work -0.5794 -0.3211 -0.1222 0.0039 -0.1049Influencing 0.3977 -0.2051 0.2409 0.1779 -0.0167Variety 0.1782 0.2612 0.1498 -0.1403 0.1567Expressing 0.1281 0.0506 0.0275 0.0742 -0.0278Alone -0.0356 0.0418 0.0406 0.086 -0.0724Stress -0.0064 0.1477 0.0652 0.2928 0.1473Tolerances -0.1739 0.4939 -0.3782 -0.2057 -0.0934Under Instructions -0.0794 -0.0952 -0.1647 -0.0431 -0.0905Dealing with People 0.6401 0.0179 0.4449 -0.0968 0.0728Judgments 0.5357 0.482 0.1734 -0.0725 0.0986Strength -0.658 0.2405 0.4254 -0.0685 0.1566Climbing -0.4579 0.3887 0.4774 -0.1726 -0.242Balancing -0.3531 0.379 0.3388 -0.1872 -0.1144Stooping -0.6018 0.1707 0.3624 -0.1446 -0.0162Kneeling -0.4832 0.3633 0.386 -0.3247 -0.0976Crouching -0.5655 0.2574 0.3569 -0.2789 -0.0766Crawling -0.2034 0.3306 0.2502 -0.089 0.0186Reaching -0.5485 0.1542 -0.3595 0.1781 0.1548Handling -0.47 0.121 -0.3567 0.1817 0.1981Fingering 0.0188 0.2397 -0.5748 0.0773 -0.038Feeling -0.0857 0.3658 -0.0843 0.131 0.2823Talking 0.6816 -0.2356 0.2052 0.2849 0.023Hearing 0.6416 -0.2215 0.1668 0.296 0.0076Tasting -0.0488 0.0638 -0.047 0.0957 0.3379Near Acuity 0.273 0.3369 -0.4239 0.0765 -0.1207Far Acuity -0.1763 0.0929 0.3796 0.6105 -0.2598Depth Perception -0.4225 0.57 0.1653 0.1325 -0.2049Accommodation 0.2501 0.3405 -0.3596 0.0368 -0.1969Color Vision -0.2057 0.4735 0.0946 0.5032 0.2074Field of Vision -0.1613 0.0611 0.3993 0.5955 -0.3363Weather -0.3761 0.2255 0.461 0.0001 -0.2083Cold -0.19 -0.0605 0.0437 -0.0354 0.3067Hot -0.2632 -0.0067 0.0128 0.0205 0.4933Humid -0.2766 -0.0122 0.0893 -0.0557 0.3779Noise -0.4532 0.3228 0.1298 -0.0241 -0.1242Vibration -0.0543 -0.0032 0.0124 0.0069 -0.0249Atmospheric -0.3131 0.3149 0.1186 -0.1104 0.1189Moving mechanical parts -0.0981 0.2531 -0.0672 -0.1379 -0.0129Electric Shocks -0.0487 0.2819 0.0839 -0.0442 0.0607High Exp Places -0.143 0.3323 0.1764 -0.0327 0.1034Radiation 0.0265 0.0403 -0.0269 0.0205 0.0155Explosive -0.0632 0.1343 0.1509 0.1444 0.2413Toxic Chemicals -0.1083 0.2105 0.1054 0.0375 0.216Other Environment Conditions -0.334 0.2318 0.2802 -0.0133 0.1695Note: The factor loadings show how the variables are weighted for each orthogonal factor. In this analysis, I reduced the occupational characteristics to five orthogonal factors: cognitive task, motor task, interpersonal task, physical demands, and hot or cold work environment. Each factor is rescaled to followed the standard normal distribution.
Table 3. Occupational Characteristics
Occupation GroupsNew
Entrant's Log Wage1
Wage Growth in
the First 10 Years (Δ Log
point)2
log Wage Annual Wage
Growth
Cognitive Task
Motor Task
Interpersonal Task
Physical Demands
Hot and Cold Environment
Managers 10.49 0.621 11.14 0.029 45.76 0.98 -0.38 -0.84 -0.14 -0.23Business and Financial Professionals 10.63 0.424 11.04 0.017 42.75 1.01 -0.28 -1.05 -0.24 -0.36Computer Analysts and Mathematicians 10.74 0.374 11.17 0.021 42.59 1.69 -0.52 -1.43 0.34 -0.27Architect and Engineers 10.75 0.341 11.15 0.023 43.18 1.46 -0.49 0.72 0.85 -0.13Life, Physical, and Social Scientists 10.41 0.454 10.96 0.027 43.46 2.06 -0.36 -0.28 1.14 -0.17Social Services Providers 10.19 0.335 10.43 0.015 45.20 0.83 -0.33 -0.20 -0.84 -0.04Legal Professionals 10.65 0.641 11.10 0.033 45.76 2.30 -0.21 -1.51 -1.03 -0.17Educators and Librarians 10.33 0.376 10.71 0.026 43.49 1.17 -0.14 -0.50 -0.65 0.78Arts, Sports, and Media Professionals 10.41 0.464 10.71 0.017 43.26 1.04 -0.40 0.62 -0.05 0.50Healthcare Practitioners 10.62 0.494 11.08 0.029 43.64 1.39 -0.29 1.50 -0.35 -0.44Healthcare Support Providers 9.99 0.623 10.28 0.020 41.47 -0.85 0.03 0.12 -1.05 -0.58Protective Services Providers 10.54 0.316 10.83 0.011 44.41 -0.22 0.03 -0.41 -0.43 2.38Food Services Providers 9.93 0.236 10.13 0.010 41.78 -0.78 -0.37 -0.32 -0.39 -0.30Building Maintenance Providers 10.25 0.172 10.10 0.007 41.38 -1.01 0.92 -0.47 0.12 -0.32Personal Care and Services Providers 10.04 0.716 10.44 0.019 42.43 -0.98 -0.56 -0.16 -0.88 0.22Sales 10.52 0.584 10.93 0.020 43.50 -0.02 -0.32 -0.99 -1.00 -0.27Office and Administrative Staff 10.34 0.401 10.66 0.018 40.95 -0.05 -0.58 0.73 -0.24 -0.37Farming, Fishing, and Forestry Workers 10.05 0.386 10.09 0.007 46.00 -0.13 1.31 -0.38 0.81 0.29Construction and Extraction Workers 10.40 0.340 10.60 0.014 42.51 0.18 1.34 0.41 0.32 0.08Installation and Repair Workers 10.37 0.419 10.72 0.014 43.02 0.36 1.09 0.96 0.49 -0.31Production Workers 10.36 0.401 10.67 0.016 42.22 -0.69 -0.07 0.44 0.87 -0.30Transportation Workers 10.25 0.390 10.54 0.011 43.85 -0.83 0.43 -0.27 -0.06 1.57Notes:
4. DOT tasks are rescaled to follow the standard normal distribution.
Wage Measure 1 Wage Measure 23
1. New entrants' log wage is the mean log wage of male workers aged between 22 and 25 and completed at least four years of college education, but did not pursue any post-college education.2. Wage growth in the first ten years is the log wage difference between new entrants (male college graduates aged 22-25) and workers with 10 years of experiences (male college graduates aged 32-35).3. Log wage and wage growth from Measure 2 are age-specific based on male college graduates . The values shown here are the arithmetic mean values across 40 years of experience groups.
DOT Tasks
Average Full-Time Weekly
Hours
Table 4. Summary Statistics of NLSY79Men Women
At the first job:% Non-White 10.0 11.1 -1.0AFQT 75.7 74.6 1.0Age at First Full-time Job 24.1 23.9 0.2 +Obs 554 570Weighted (%) 51.3 48.7In 2008: Obs 380 375Weighted (%) 51 49Percent working (%) 97 85.5 11.5 **
Dist. Of First Jobs Within Gender (%)Percent Female
N % N % %Management 93 17.59 67 12.38 40.09Business and Financial Operations 31 4.96 52 8.85 62.91Computer Analysts and Mathematicians 41 7.19 20 5.1 40.30Architecture and Engineers 54 9.58 10 1.68 14.32Life, Physical, and Social Science 12 1.73 5 0.94 34.15Community and Social Services 9 1.62 11 1.85 52.06Legal 3 0.51 1 0.07 11.79Education, Training, and Library 26 3.88 78 12.22 74.97Arts, Design, Sports, and Media 25 4.74 24 6.09 54.96Healthcare Practitioners 12 1.45 60 9.5 86.15Healthcare Support 2 0.12 17 2.47 94.98Protective Service 16 2.22 4 0.46 16.36Food Preparation and Serving 17 3.43 9 2.29 38.84Building Maintenance 1 0.29 1 0.23 43.62Personal Care and Service 5 0.6 12 2.51 79.96Sales 73 14.16 43 7.82 34.43Office and Administrative Support 64 11.36 146 23.71 66.49Farming, Fishing, and Forestry 10 2.2 1 0.26 10.16Construction and Extraction 18 3.93 0.00Installation, Maintenance, and Repair 4 0.61 0.00Production 18 3.59 7 1.23 24.56Transportation and Material Moving 20 4.23 2 0.32 6.74Observations 100 554 100 570Weighted % 51.3 48.7** p<0.01, * p<0.05, + p<0.1
Within Male Within Female
Difference
Table 5. Leave Patterns by Gender, NLSY79 (2008 wave)Men Women
Observations in 2008 380 375% %
Percent w/o Job Interruptions 20.00 6.67Percent w/ Job Interruptions 80.00 93.33
Interruptions <= 1 yr 48.95 34.301 yr < Interruptions <= 3 yrs 24.74 20.00Interruptions > 3 yrs 6.32 38.93Duration of Interruptions Years Years
Mean 1.28 4.85Std. Dev. 1.88 5.75
Table 6. Wage Penalty Estimates
Occupation Groups N % Log Point N %Coef
(Log Point)Standard
ErrorCoef
(Log Point)Standard
Error
Managers 15,302 18.23 0.075 293 12.22 0.444** (0.0694) 0.408** (0.0387)Business and Financial Professionals 3,222 3.84 0.069 87 3.63 0.157 (0.147) 0.0126 (0.0823)Computer Analysts and Mathematicians 2,602 3.10 0.082 139 5.80 -0.0941 (0.0816) 0.0432 (0.0455)Architect and Engineers 4,591 5.47 0.077 138 5.76 0.121 (0.101) 0.103+ (0.0563)Life, Physical, and Social Scientists 1,100 1.31 0.034 21 0.88 -0.0558 (0.135) 0.0193 (0.0754)Social Services Providers 1,157 1.38 0.002 9 0.38 1.194 (2.061) 1.266 (1.150)Legal Professionals 917 1.09 0.090 24 1.00 1.649** (0.541) 1.009** (0.302)Educators and Librarians 2,794 3.33 0.061 24 1.00 -0.556+ (0.304) -0.433* (0.169)Arts, Sports, and Media Professionals 984 1.17 0.024 56 2.34 0.159 (0.170) 0.116 (0.0951)Healthcare Practitioners 1,908 2.27 0.083 39 1.63 -0.154 (0.206) -0.173 (0.115)Healthcare Support Providers 283 0.34 0.021 1 0.04 -- --Protective Services Providers 3,643 4.34 0.045 29 1.21 -0.379 (0.461) -0.375 (0.257)Food Services Providers 1,516 1.81 0.008 56 2.34 -0.117 (0.222) -0.175 (0.124)Building Maintenance Providers 2,000 2.38 -0.013 25 1.04 -0.377 (0.501) -0.450 (0.280)Personal Care and Services Providers 391 0.47 -0.014 6 0.25 0.109 (0.421) 0.0793 (0.235)Sales 5,018 5.98 0.071 126 5.26 0.230+ (0.117) 0.194** (0.0655)Office and Administrative Staff 5,118 6.10 0.036 63 2.63 0.149 (0.155) 0.126 (0.0866)Farming, Fishing, and Forestry Workers 1,200 1.43 0.003 35 1.46 -0.143 (0.195) -0.242* (0.109)Construction and Extraction Workers 4,771 5.69 0.036 480 20.03 0.0443 (0.0683) 0.00507 (0.0381)Installation and Repair Workers 6,546 7.80 0.047 185 7.72 0.209* (0.0961) 0.0862 (0.0536)Production Workers 9,924 11.83 0.030 249 10.39 -0.0126 (0.0704) 0.0899* (0.0393)Transportation Workers 8,933 10.64 0.048 312 13.02 0.217** (0.0796) 0.182** (0.0444)Total 83,920 100.00 2,397 100.00Note:Standard errors in parentheses** p<0.01, * p<0.05, + p<0.11. Wage growth rates are estimated nonparametrically by age and occupation group. The average wage growth rate is an arithmetic mean of wage growth rates of all continuous male workers at least 20 years old from the pooled matched one-year CPS panels between 1991 and 2008.2. Human capital depreciation estimates are based on male workers who worked in the same occupations after displacement from the DWS 1994-2008.
Wage Growth Rate EstimatesOLS Regression Median Regression
Human Capital Depreciation Rate EstimatesDWS Observations2CPS Observations1
Table 7. Conditional Logit Estimates , NLSY 1979-2008
Unadjusted Adjusted Unadjusted AdjustedMeasure 1:New Entrant lnWage 0.992 (0.204) ** (0.514) + 2.925 (0.261) ** (0.350) **10-Yr Wage Growth -0.950 (0.212) ** (0.698) 0.395 (0.371) (0.489)HC Depreciation -1.042 (0.111) ** (0.551) + -0.629 (0.107) ** (0.463)Weekly Hours 0.143 (0.046) ** (0.142) 0.501 (0.035) ** (0.073) **Cognitive Tasks 0.105 (0.073) (0.209) -0.751 (0.067) ** (0.086) **Motor Tasks -2.366 (0.169) ** (0.627) ** -1.342 (0.162) ** (0.205) **Interpersonal Tasks 0.084 (0.040) * (0.216) -0.267 (0.042) ** (0.112) *Physical Demands -0.894 (0.088) ** (0.271) ** 0.375 (0.071) ** (0.164) **Hot and Cold Environment -0.519 (0.079) ** (0.444) -0.580 (0.069) ** (0.160) *Measure 2:lnWage 1.120 (0.178) ** (0.436) * 2.831 (0.229) ** (0.294) **Annual Wage Growth -0.430 (0.125) ** (0.167) * -0.055 (0.126) (0.118)HC Depreciation -0.968 (0.118) ** (0.532) + -0.479 (0.118) ** (0.439)Weekly Hours 0.058 (0.046) (0.135) 0.385 (0.032) ** (0.062) **Cognitive Tasks 0.149 (0.070) * (0.191) -0.685 (0.061) ** (0.090) **Motor Tasks -2.137 (0.160) ** (0.611) ** -1.010 (0.142) ** (0.170) **Interpersonal Tasks 0.133 (0.040) ** (0.205) -0.175 (0.045) ** (0.119)Physical Demands -0.876 (0.079) ** (0.257) ** 0.270 (0.064) ** (0.157) +Hot and Cold Environment -0.373 (0.073) ** (0.406) -0.461 (0.065) ** (0.144) **Standard errors in parentheses** p<0.01, * p<0.05, + p<0.1
Coefficient CoefficientStandard Erorr Standard Error
NLSY, MenNLSY, Women