ESSAYS IN LABOR ECONOMICS
By
Jin Young Lee

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
DOCTOR OF PHILSOPHY
Economics
2012

ABSTRACT
ESSAYS IN LABOR ECONOMICS
By
Jin Young Lee
This dissertation focuses on the interrelationship of socioeconomic outcomes of women. The
first chapter of the dissertation explores the fragility of estimated effects of unilateral divorce
laws on divorce rates and concludes that the impact of unilateral divorce laws remains unclear. I
also make the methodological point that identification in differences-in-differences research
becomes weaker in the presence of dynamics, especially in the presence of unit-specific time
trends. The second chapter examines the plateau in U.S. women’s labor force participation from
the early 1990s to the present. The series of shift-share analyses shows that, for the most part,
both the plateau and the earlier upward trend appeared within almost every category broken
down by education, marital status, and child-rearing. In the third chapter, I look at the role of
teenage women's anticipated future labor force attachment in explaining the upward trend in U.S.
women's college-going. Combined with the trend towards higher work expectations of young
women across birth cohorts, the results suggest that teenagers’ future work expectations may
account in part for the upward trends in women's college attendance and completion.

ACKNOWLEDGMENTS

This dissertation would not have been possible without the tremendous support and guidance of
my advisor, Gary Solon. He has been my inspiration of how to be an economist. I will be
grateful to him forever.
I am equally indebted to my committee members, Stacy Dickert-Conlin, Steven Haider,
and Songqing Jin. Their comments were extremely helpful in proceeding with every stage of my
dissertation research.
Without my family members, Mom, Dad, and my twin sister, I would not have made it
this far. I owe the most profound thanks to them for their endless love and encouragement.
Jongwoo Park, my love, provided abundant support throughout my time at Michigan
State University. His love made all the difference.
Finally, I dedicate this dissertation to my parents.

iii

TABLE OF CONTENTS

LIST OF TABLES ...........................................................................................................................v
LIST OF FIGURES ....................................................................................................................... vi
CHAPTER 1
THE FRAGILITY OF ESTIMATED EFFECTS OF
UNILATERAL DIVORCE LAWS ON DIVORCE RATES ..........................................................1
1.1 Introduction ........................................................................................................................1
1.2 Robustness Checks.............................................................................................................2
1.2.A Weighted and Ordinary Least Squares Estimates of Linear-in-Levels
Specifications .........................................................................................................2
1.2.B Alternative Functional Forms.................................................................................5
1.3 Summary and Discussion ...................................................................................................6
CHAPTER 2
THE PLATEAU IN U.S. WOMEN’S LABOR FORCE PARTICIPATION:
A COHORT ANALYSIS ..............................................................................................................12
2.1 Introduction ......................................................................................................................12
2.2 Data ..................................................................................................................................14
2.3 Cohort shifts in labor force participation and
related socioeconomic outcomes of women ...........................................................15
2.3.A Birth cohort and FLFP .......................................................................................15
2.3.B Educational attainment ........................................................................................16
2.3.C. Marriage ..............................................................................................................21
2.3.D Child rearing .......................................................................................................23
2.3.E The joint contribution of education, marriage, and child rearing ........................24
2.4 Summary and discussion..................................................................................................26
CHAPTER 3
THE UPWARD TREND IN WOMEN’S COLLEGE-GOING:
THE ROLE OF TEENAGERS’
ANTICIPATED FUTURE LABOR FORCE ATTACHMENT....................................................42
3.1 Introduction ......................................................................................................................42
3.2 The potential importance of work expectations for college-going decisions ..................44
3.3 Data ..................................................................................................................................48
3.4 Evidence on the association between work expectations and college-going ...................54
3.5 Summary ..........................................................................................................................59
REFERENCES ..............................................................................................................................76

iv

LIST OF TABLES

Table 1.1 Estimated Effects of Unilateral Divorce Laws
with No Controls for State-Specific Time Trends .........................................................9
Table 1.2 Estimated Effects of Unilateral Divorce Laws
with Controls for Linear State-Specific Time Trends..................................................10
Table 1.3 Estimated Effects of Unilateral Divorce Laws
with Controls for Quadratic State-Specific Time Trends ............................................11
Table 2.1 Decomposition in difference in FLFP rates between birth cohorts
using socioeconomic outcomes of women ..................................................................28
Table 3.1 Summary statistics .......................................................................................................60
Table 3.2 Average partial effects from weighted logit regressions using NLSYW68 ................62
Table 3.3 Average partial effects from weighted logit regressions using NLSY79 ....................63
Table 3.4 Average partial effects from weighted logit regressions using NLSY97 ....................64
Table 3.5 Estimates from unweighted logit estimation ...............................................................65
Table 3.6 Estimates from weighted least squares estimation of LPM .........................................66
Table 3.A1 Weighted averages of college outcome variables by sample....................................67
Table 3.A2 Coefficient estimates from weighted logit regressions using NLSYW68 ................68
Table 3.A3 Coefficient estimates from weighted logit regressions using NLSY79 ....................69
Table 3.A4 Coefficient estimates from weighted logit regressions using NLSY97 ....................70

v

LIST OF FIGURES

Figure 2.1: Labor force participation rates for men and women ...................................................29
Figure 2.2: Female labor force participation rates by age and cohort ...........................................30
A. 1934 - 1982 cohorts.............................................................................................................30
B. 1952 - 1958 cohorts .............................................................................................................30
Figure 2.3: Women's educational attainment by cohort, ages 25-2 ...............................................31
Figure 2.4: Female labor force participation rates by education ...................................................32
A. High school or less ..............................................................................................................32
B. Some college .......................................................................................................................32
C. College completion or more ................................................................................................33
Figure 2.5: Decomposition of cohort difference in FLFP using education ..................................34
A. Ages 35-39 ..........................................................................................................................34
B. Ages 25-29 ..........................................................................................................................34
Figure 2.6: Fraction of women ever-married .................................................................................35
A. By age group .......................................................................................................................35
B. By educational attainment and selective age group ............................................................35
Figure 2.7: Female labor force participation rates by marital status..............................................36
A. Ever-married women...........................................................................................................36
B. Never-married women.........................................................................................................36
Figure 2.8: Decomposition of cohort difference in FLFP using marriage .....................................37
A. Ages 35-39 ..........................................................................................................................37
B. Ages 25-29 ..........................................................................................................................37
Figure 2.9: Fraction of women who have children under 6 in the household ...............................38
Figure 2.10: Labor force participation rates for women with and without young children ...........39
A. Women with children under 6 in the household .................................................................39
B. Women without children under 6 in the household ............................................................39
Figure 2.11: Decomposition of cohort difference in FLFP using child-rearing ............................40
A. Ages 35-39 ..........................................................................................................................40
B. Ages 25-29 ..........................................................................................................................40
Figure 2.12: Decomposition of cohort difference in FLFP using joint variable ............................41
A. Ages 35-39 ..........................................................................................................................41
B. Ages 25-29 ..........................................................................................................................41

vi

CHAPTER 1
THE FRAGILITY OF ESTIMATED EFFECTS OF
UNILATERAL DIVORCE LAWS ON DIVORCE RATES

1

1.1 Introduction
An influential American Economic Review article by Leora Friedberg (1998) estimated
the impact on divorce rates from the adoption of unilateral divorce laws by most states during the
1970s. Using state/year data for 1968-1988, Friedberg estimated fixed-effects regressions of the
annual per-capita divorce rate on a dummy variable for unilateral divorce with controls for state
effects and year effects. In the simplest specification, she estimated almost exactly a zero effect
of unilateral divorce, but in her preferred specifications that also controlled for linear or
quadratic state-specific time trends, she estimated that adoption of unilateral divorce increased
the divorce rate by more than 0.4 divorces per 1,000 population. This estimated impact seemed
substantial relative to the sample mean divorce rate of 4.6, and Friedberg concluded (p. 608),
“The move towards unilateral divorce accounted for 17 percent of the increase in divorce rates
between 1968 and 1988.”
In a subsequent American Economic Review article, Justin Wolfers (2006) questioned
Friedberg’s assumption that unilateral divorce shifted the divorce rate by a permanent constant.
Instead, he posed the possibility that the impact might follow a distributed-lag pattern.
Expanding the sample period to 1956-1988, he estimated regressions similar to Friedberg’s
except that he replaced her single dummy variable for unilateral divorce with a series of

1

This chapter is co-authored with Gary Solon, Professor in Economics at Michigan State
University.

1

dummies indicating that unilateral divorce had been in place for up to 2 years, 3-4 years, 5-6
years, …, 13-14 years, or at least 15 years. Based on the results, he concluded (p. 1802) “that the
divorce rate rose sharply following the adoption of unilateral divorce laws, but that this rise was
reversed within about a decade.” In the present note, we check the robustness of Wolfers’s
estimates to some reasonable variations in estimation method and functional form.

2

1.2 Robustness Checks
1.2.A Weighted and Ordinary Least Squares Estimates of Linear-in-Levels
Specifications
Like Friedberg, Wolfers estimated his regressions by weighted least squares (WLS),
weighting each state/year observation by the state’s population. The results are shown in the first
columns of our tables 1.1, 1.2, and 1.3, respectively for models with no state-specific trends,
linear trends, and quadratic trends. Thanks to Wolfers’s generous posting of his data and code
on his website, it was easy to obtain exact replications of his results. In all three specifications,
unilateral divorce is estimated to increase the divorce rate initially by about 0.3 divorces per
1,000 population. The basis for Wolfers’s conclusion about dynamics is that the impact appears
to remain substantial throughout the first eight years since adoption, but beyond eight years there
is little evidence of a substantial positive impact.
Like Friedberg, Wolfers estimated his standard errors (shown in parentheses in the first
columns of our tables 1.1-1.3) on the assumption that the error term in each weighted regression
is homoskedastic and serially uncorrelated. Setting aside the homoskedasticity assumption for
2

It is understandable that Wolfers hewed closely to Friedberg’s estimation method and
functional form in an effort to explore the role of dynamics while holding these other factors
constant.

2

the moment, our analysis of the residuals from Wolfers’s regressions shows overwhelming
evidence of strong serial correlation. First-order autoregressions of the WLS residuals show
estimated autocorrelations of 0.93, 0.94, and 0.89 respectively for the three specifications.
Furthermore, the estimated autocorrelations remain substantial at higher orders. For example,
3

the respective estimates of the fifth-order autocorrelations are 0.73, 0.73, and 0.53.

Obviously,

neglect of such strong serial correlation could lead to serious bias in the estimation of standard
errors. We therefore have used Stata’s cluster option to implement Arellano’s (1987) method of
correcting standard error estimates for both serial correlation and heteroskedasticity. The
resulting standard error estimates, shown in brackets in the first columns of our tables 1-3, are
dramatically larger than those reported by Wolfers. Once it is recognized that the coefficient
estimates may be much less precise than they previously appeared, the statistical significance of
the coefficient estimates for lags up to eight years becomes less compelling.

4

An even more striking pattern emerges when we reconsider whether weighting by
population is an appropriate correction for heteroskedasticity. The common practice of
weighting by population leads to efficient coefficient estimation under the very strong
assumption that the error terms for individuals within the state are homoskedastic and
independent of each other. As explained in detail by Dickens (1990), however, in the likely case
that the individual error terms are positively correlated because they share a common state-level
3

As shown in Nickell (1981) and Solon (1984), these estimates are biased downward by the
incidental-parameters problem in fixed-effects estimation. With a time-series length of 33 years,
however, the bias is small.
4
This same point is made by Vogelsang (2008) with respect to the model with no state-specific
trends. We say “may be less precise” rather than “are less precise” because Vogelsang goes on
to find that, when he also applies his method for correcting for spatial correlation (correlation
among error terms across states), the standard error estimates come back down. In the end,
Wolfers’s assumption of homoskedasticity and no serial correlation clearly is incorrect, but it is
not so clear what the precision of the coefficient estimates really is.

3

error component, the original (unweighted) state-average error terms may be nearly
homoskedastic, and weighting by population size may induce a substantial heteroskedasticity
problem. In this case, ordinary least squares (OLS) may be considerably more efficient than
WLS. At minimum, then, it is good practice to report coefficient estimates from both methods,
along with heteroskedasticity-robust standard error estimates. Also, as emphasized by
DuMouchel and Duncan (1983), since both estimators are consistent if the model otherwise is
correctly specified, discrepancies between WLS and OLS estimates can be a basis for testing
model specification.
In the second columns of tables 1.1-1.3, we report the results from estimating by OLS
instead of WLS. Where the WLS coefficient estimates for lags up to eight years had appeared
substantially positive, this is much less true for the OLS coefficient estimates in all three
specifications. This same finding has been reported by Dröes and van Lamoen (2010).
Furthermore, in accordance with Dickens’s critique of population-weighting, although the robust
standard error estimates (shown in brackets) appear larger for OLS than for WLS in table 1, this
is less clear for the specifications in tables 1.2 and 1.3.
In any case, the substantial discrepancies between the WLS and OLS estimates are
suggestive of model misspecification. We will take up the issue of functional form in the next
sub-section. Here, we will close by noting another potential specification issue – the assumption
that the vector of distributed-lag coefficients is homogeneous across states may be incorrect. If
the impact of unilateral divorce is heterogeneous – i.e., if it interacts with other state
characteristics – then WLS and OLS estimates that do not explicitly account for those
interactions may identify different averages of the heterogeneous effects. By design, WLS
places more weight on the more populous states. If unilateral divorce tends to have larger effects

4

in those states, then the WLS estimates will tend to be larger than the OLS estimates. This
possibility is suggested by the results in the third columns of tables 1.1-1.3, which redo WLS
with California omitted from the sample. The coefficient estimates are much smaller than the
full-sample WLS results in the first columns. The discrepancy between the estimates in columns
1 and 2 may arise partly because California looms large in the full-sample WLS estimates, but is
weighted the same as any other state in the OLS estimates.5

1.2.B Alternative Functional Forms
All the regression models estimated by Wolfers, as well as Friedberg, are linear models
for the level of the divorce rate. This is not the only possible functional specification, and
arguably not the most natural. For example, given that the divorce rate is always positive, linear
models for the logarithm of the divorce rate are an obvious alternative. Indeed, since the divorce
rate is a fraction, linear models for the logit of the divorce rate (i.e., the log odds of divorcing)
also seem appealing. In practice, though, there is almost no difference in the present case
between the log and the logit.

With the divorce rate R expressed per capita, the logit of R is

log[ R /(1  R)]  log R  log(1  R) , which is almost identical to log R because log(1  R) is so

close to zero. In the rest of tables 1.1-1.3, we report results for only the log of R because the
logit results are virtually identical.
The last two columns in tables 1.1-1.3 report both WLS and OLS estimates of models in
which the dependent variable is the divorce rate in log rather than level form. We think this
sensitivity analysis is our most striking of all. Once the divorce rate is in log form, neither the
5

As noted by Friedberg (1998, footnote 20), Nevada is a major outlier with respect to divorce
rates. We have found, however, that, with state fixed effects accounted for, excluding/including
Nevada is much less consequential for the gap between WLS and OLS estimates than is the
treatment of California.

5

WLS nor OLS estimates show any indication of a positive effect of unilateral divorce on divorce
rates in any of the three specifications. Whereas Friedberg and Wolfers differed over whether
the positive impact of unilateral divorce on divorce rates is permanent or temporary, the log
results show no positive impact at all. The fragility of the previously published estimates does
not prove those estimates are wrong. Rather, the extreme sensitivity of the results to functional
form assumptions, as well as estimation methods, leaves unclear what the true impact of
unilateral divorce has been.

6

Three other points about functional form are worth noting. First, the discrepancy
between the levels results and the log (and logit) results is a vivid counter-example to the
frequent claim that functional form assumptions are of little consequence in the estimation of
7

models for binary, fractional, or otherwise limited dependent variables.

Second, compared to

the levels specifications, the log specifications show less discrepancy between the WLS and OLS
results. Third, with regard to Dickens’s analysis of weighting, the robust standard error
estimates for the log specifications suggest that, for those specifications, OLS is considerably
more efficient than WLS.

1.3. Summary and Discussion
6

It may be worth adding that none of the estimates in our article, Wolfers’s, or Friedberg’s
indicates that the adoption of unilateral divorce accounts for most of the rise in divorce rates
during the 1970s. As Stevenson and Wolfers (2007, p. 46) put the point, “Despite apparent
conflict in this literature, it is worth emphasizing the point of substantial agreement: each of
these authors finds that liberalized divorce laws had at most a small effect on divorce rates, and
these reforms explain very little of the rise in divorce over the past half century.”
7
For example, a lengthy passage on this topic in Angrist and Pischke (2009) concludes (p. 107),
“The upshot of this discussion is that while a nonlinear model may fit the CEF [conditional
expectation function] for LDVs [limited dependent variables] more closely than a linear model,
when it comes to marginal effects, this probably matters little. This optimistic conclusion is not
a theorem, but, as in the empirical example here, it seems to be fairly robustly true.”

6

Our main substantive finding is that estimates of the impact of unilateral divorce laws on
divorce rates are highly sensitive to variations in model specification and estimation method.
We do not take this to mean that the conclusions in Wolfers (2006) are necessarily wrong. We
take it to mean that the true impact of unilateral divorce laws remains unclear.
In addition, the Wolfers article appropriately emphasized broader methodological
lessons. Having found that Friedberg’s (1998) results seemed sensitive to allowance for dynamic
response, Wolfers concluded (p. 1802),
A major difficulty in difference-in-difference analyses involves separating out preexisting
trends from the dynamic effects of a policy shock. Her approach appears to confound the
two. This problem – that state-specific trends may pick up the effects of a policy and not
just preexisting trends – is quite general. Slight modifications to standard procedures
yield more directly interpretable results.
We fully agree that, when the impact of a policy change is not necessarily immediate and
constant, researchers should follow Wolfers’s example of exploring the dynamics of the
response. But the results we have reported suggest an additional cautionary note. The
differences-in-differences research design with unit-specific time trends is essentially a type of
regression discontinuity design, with time as the “running variable.” As in other regression
discontinuity designs, identification is based on the appearance and size of a “jump” in the
dependent variable at the discontinuity point – in this case, the time at which the new policy is
adopted. When the shift in the dependent variable may vary with the length of time since the
policy change, and especially when that complication is accompanied by other differences across
states in time trends, the sharpness of the identification strategy suffers. In such cases, we are
asking the data a much more nuanced question than just whether the dependent variable series

7

showed a constant discrete shift at the moment of policy adoption, and sometimes the answer
will be unclear. In that broader methodological light, it may not be so surprising that the
evidence on the impact of unilateral divorce laws is ambiguous.

8

Table 1.1 Estimated Effects of Unilateral Divorce Laws
with No Controls for State-Specific Time Trends
(1)

(2)

(3)

(4)

(5)

Divorce rate

Divorce rate

Divorce rate

Log of
divorce rate

Log of
divorce rate

WLS

OLS

WLS
excluding
California

WLS

OLS

First 2 years

0.267
(0.085)
[0.188]

-0.219
(0.192)
[0.292]

0.071
(0.093)
[0.131]

-0.136
(0.033)
[0.114]

-0.091
(0.028)
[0.051]

Years 3-4

0.210
(0.085)
[0.159]

-0.273
(0.194)
[0.423]

0.123
(0.093)
[0.168]

-0.206
(0.033)
[0.132]

-0.114
(0.029)
[0.065]

Years 5-6

0.164
(0.085)
[0.171]

-0.425
(0.198)
[0.490]

0.097
(0.092)
[0.184]

-0.249
(0.033)
[0.144]

-0.152
(0.029)
[0.076]

Years 7-8

0.158
(0.084)
[0.174]

-0.452
(0.200)
[0.477]

0.103
(0.091)
[0.190]

-0.270
(0.032)
[0.151]

-0.171
(0.029)
[0.081]

Years 9-10

-0.121
(0.084)
[0.163]

-0.703
(0.203)
[0.479]

-0.144
(0.090)
[0.178]

-0.331
(0.032)
[0.157]

-0.220
(0.030)
[0.083]

Years 11-12

-0.324
(0.083)
[0.180]

-0.741
(0.203)
[0.503]

-0.262
(0.090)
[0.187]

-0.376
(0.032)
[0.165]

-0.234
(0.030)
[0.085]

Years 13-14

-0.461
(0.084)
[0.199]

-0.845
(0.212)
[0.544]

-0.390
(0.091)
[0.211]

-0.419
(0.032)
[0.170]

-0.265
(0.031)
[0.088]

Dependent
variable:
Estimation
method:

-0.507
-0.776
-0.348
-0.493
-0.326
Years 15+
(0.080)
(0.208)
(0.090)
(0.031)
(0.031)
[0.233]
[0.472]
[0.224]
[0.188]
[0.098]
Notes: Divorce rate is number of divorces per 1,000 persons by state and year. Standard error
estimates in parentheses assume homoskedasticity and serial non-correlation. Standard error
estimates in brackets are robust to heteroskedasticity and serial correlation. All regressions
include controls for year and state fixed effects.

9

Table 1.2 Estimated Effects of Unilateral Divorce Laws
with Controls for Linear State-Specific Time Trends
(1)

(2)

(3)

(4)

(5)

Divorce rate

Divorce rate

Divorce rate

Log of
divorce rate

Log of
divorce rate

WLS

OLS

WLS
excluding
California

WLS

OLS

First 2 years

0.342
(0.062)
[0.196]

0.141
(0.096)
[0.112]

0.104
(0.066)
[0.095]

-0.022
(0.020)
[0.063]

-0.017
(0.018)
[0.026]

Years 3-4

0.319
(0.070)
[0.154]

0.211
(0.107)
[0.120]

0.170
(0.073)
[0.133]

-0.049
(0.023)
[0.063]

-0.014
(0.020)
[0.031]

Years 5-6

0.300
(0.077)
[0.174]

0.177
(0.121)
[0.180]

0.152
(0.080)
[0.153]

-0.051
(0.025)
[0.064]

-0.022
(0.023)
[0.034]

Years 7-8

0.322
(0.084)
[0.195]

0.250
(0.132)
[0.283]

0.172
(0.087)
[0.177]

-0.033
(0.028)
[0.065]

-0.013
(0.025)
[0.039]

Years 9-10

0.081
(0.091)
[0.206]

0.133
(0.143)
[0.370]

-0.049
(0.094)
[0.202]

-0.052
(0.030)
[0.067]

-0.030
(0.027)
[0.046]

Years 11-12

-0.102
(0.099)
[0.223]

0.144
(0.154)
[0.424]

-0.164
(0.101)
[0.240]

-0.051
(0.032)
[0.074]

-0.015
(0.029)
[0.052]

Years 13-14

-0.202
(0.107)
[0.236]

0.210
(0.168)
[0.484]

-0.270
(0.109)
[0.248]

-0.043
(0.035)
[0.077]

-0.005
(0.032)
[0.060]

Dependent
variable:
Estimation
method:

-0.210
0.311
-0.246
0.006
0.026
Years 15+
(0.119)
(0.187)
(0.121)
(0.039)
(0.035)
[0.263]
[0.540]
[0.275]
[0.084]
[0.073]
Notes: Divorce rate is number of divorces per 1,000 persons by state and year. Standard error
estimates in parentheses assume homoskedasticity and serial non-correlation. Standard error
estimates in brackets are robust to heteroskedasticity and serial correlation. All regressions
include controls for year and state fixed effects.

10

Table 1.3 Estimated Effects of Unilateral Divorce Laws
with Controls for Quadratic State-Specific Time Trends
(1)

(2)

(3)

(4)

(5)

Divorce rate

Divorce rate

Divorce rate

Log of
divorce rate

Log of
divorce rate

WLS

OLS

WLS
excluding
California

WLS

OLS

First 2 years

0.302
(0.054)
[0.169]

0.050
(0.075)
[0.089]

0.110
(0.059)
[0.093]

-0.021
(0.020)
[0.053]

-0.015
(0.018)
[0.021]

Years 3-4

0.289
(0.065)
[0.127]

0.062
(0.092)
[0.143]

0.198
(0.072)
[0.131]

-0.047
(0.025)
[0.054]

-0.014
(0.022)
[0.026]

Years 5-6

0.291
(0.079)
[0.141]

-0.036
(0.116)
[0.179]

0.203
(0.090)
[0.160]

-0.046
(0.030)
[0.056]

-0.022
(0.028)
[0.033]

Years 7-8

0.351
(0.097)
[0.165]

-0.026
(0.144)
[0.205]

0.257
(0.111)
[0.202]

-0.019
(0.037)
[0.061]

-0.008
(0.035)
[0.043]

Years 9-10

0.161
(0.117)
[0.172]

-0.210
(0.177)
[0.240]

0.076
(0.137)
[0.232]

-0.028
(0.045)
[0.069]

-0.018
(0.043)
[0.056]

Years 11-12

0.047
(0.142)
[0.190]

-0.270
(0.215)
[0.278]

0.006
(0.167)
[0.280]

-0.014
(0.054)
[0.088]

0.005
(0.052)
[0.070]

Years 13-14

0.031
(0.167)
[0.211]

-0.289
(0.257)
[0.328]

-0.047
(0.199)
[0.310]

0.011
(0.063)
[0.098]

0.029
(0.063)
[0.086]

Dependent
variable:
Estimation
method:

0.251
-0.226
0.074
0.104
0.087
Years 15+
(0.205)
(0.317)
(0.242)
(0.078)
(0.077)
[0.261]
[0.407]
[0.341]
[0.113]
[0.107]
Notes: Divorce rate is number of divorces per 1,000 persons by state and year. Standard error
estimates in parentheses assume homoskedasticity and serial non-correlation. Standard error
estimates in brackets are robust to heteroskedasticity and serial correlation. All regressions
include controls for year and state fixed effects.

11

CHAPTER 2
THE PLATEAU IN U.S. WOMEN’S LABOR FORCE PARTICIPATION:
A COHORT ANALYSIS

2.1 Introduction
One of the most remarkable changes in the U.S. labor market in the twentieth century
was a massive increase in the rate of female labor force participation (FLFP). The FLFP rate
was only 17 percent in 1890 (Goldin 2000, Table 10.1), but had more than tripled by a century
later, reaching 59 percent in 1994 (Figure 2.1). This trend seems even more dramatic when
1

contrasted with the gradual decline in men’s labor force participation.

Since 1994, however, an

important new pattern has emerged: after many decades of trending upward, the FLFP rate has
2

leveled off at around 60 percent.

This recent stagnation in the labor force involvement of U.S.

women is rare even in world economies (U.S. Bureau of Labor Statistics, table of women’s labor
force participation rates, selective countries, 1970-2009, available at
http://www.bls.gov/spotlight/2011/women/).
The end of the upward trend in U.S. women’s labor force participation is an important
and puzzling phenomenon. For example, labor economists traditionally have described the longrunning rise in women’s labor force participation as a movement along a positively sloped labor
1

See Juhn and Potter (2006) and Fallick and Pingle (2007) for the trend in U.S. men’s labor
force participation.
2
Although the 1994 Current Population Survey redesign has caused changes in the measurement
of many of the statistics derived from the CPS (Polivka and Miller, 1998), trends in the FLFP are
not sensitive to the redesign. All estimates presented in this paper are based on data unadjusted
for the1994 CPS redesign. Changes in the percent institutionalized also are not an issue because
the trend in the fraction of women institutionalized was stable over time. My calculation using
U.S. Censuses shows that the percentages of women institutionalized (correctional and mental
institutions and institutions for the elderly, handicapped, and poor) in 1980, 1990 and 2000 are
1.37, 1.49, and 1.34, respectively.

12

supply curve in response to real wage growth. The plateau in women’s labor force participation,
however, has occurred despite continued real wage growth (Eckstein and Lifshitz, 2011). A
simple labor supply story therefore cannot account for the trend shift unless it involves a credible
explanation for why the labor supply curve for women switched from quite positively sloped to
inelastic.
This paper uses March Current Population Survey data from 1968 to 2010 to assemble a
clear and detailed catalog of facts about the plateau in U.S. women’s labor force participation.
3

Following and extending earlier work by Goldin (2006) and Fallick and Pingle (2007), I find
that a great deal of what has happened to FLFP can be succinctly summarized in terms of trends
across birth cohorts. In addition to providing a more up-to-date cohort analysis, I conduct shiftshare analyses that document within-group trends and composition effects based on
4

disaggregations by educational attainment, marital status, and child care responsibilities. In the
end, I deliver a constellation of evidence about the trends in women’s labor force participation
and related socioeconomic outcomes that I hope will stimulate and guide further research into the
causes of the shifting trends.
Although numerous recent studies have analyzed FLFP in the United States, these studies
are less comprehensive than mine in that they cover shorter time periods, are restricted to
particular subpopulations, or provide no shift-share analyses using related socioeconomic
outcomes. For example, Boushey (2005) and Hoffman (2009) find that the negative effects of
3

Aaronson et al. (2006) and Percheski (2008) are other precedents.

4

Race is also an important demographic variable that may interact with women's labor supply,
but it plays a minor role in explaining the plateau. Since the early 1990s, the FLFP rate among
white women has shown a leveling-off, and the rate among black women has displayed a very
slight decline. My sample includes all race and ethnic groups. Results in this paper still hold
when the population is limited to white women only.

13

children on labor force participation for women aged 25 to 44 fell over time in the Current
Population Survey’s Outgoing Rotation Group data for 1984, 1989, 1993, 2000, and 2004.
Percheski (2008) investigates cohort differences in female employment rates using data from
decennial Censuses for 1960 through 2000 and the American Community Survey for 2005, but
she restricts her samples to professional women. Goldin (2006) explores how women's
employment, education, and family status have transformed together using various data sources,
with an emphasis on the period before the 2000s. Her conjecture that the plateau in FLFP was a
temporary phenomenon due to the recession of the early 2000s is testable (and refuted) by my
updating of the evidence through 2010. Macunovich (2010) presents the change in FLFP by
socioeconomic subgroup over time using March CPS data for 1976 through 2009, but does not
look at compositional changes of the subgroups. Fallick and Pingle (2007) do decompose the
changes in male and female labor force participation rates over time into group-share and withingroup changes, but the subgroups for their decompositions are based only on age.
My series of shift-share analyses uniquely provides the quantitative connection between
FLFP and related socioeconomic outcomes of women. It shows that both the rising FLFP and
the plateau can be explained mostly by trends within almost every group – in particular, within
every education and child-rearing group, and within the ever-married group.

2.2

Data
My data come from the Integrated Public Use Microdata Series (IPUMS) version of the

March Current Population Surveys (CPS) for 1968 through 2010. I restrict my sample to non-

14

5

institutionalized civilian women aged 16 – 64. I do not include observations before 1968
because a variable “MOMLOC,” which is used to link mothers to their own children, is not
available for those observations. Throughout this paper, I use single-year birth cohorts and fiveyear age bins (or a four-year age bin for the youngest age group). Thus, each cohort includes
women born in the reference year, and the whole sample is divided into ten age groups, from age
group 16-19 to 60-64. Birth year is calculated as calendar year minus age. In order to retain a
sufficient sample size for each age-cohort group, the figures are based on the cohorts 1932-1982.
To compute averages for each subpopulation, I use CPS-provided sampling weights. Under the
simplifying approximation that the covariance of sampling errors between any two age-cohort
groups is zero, most of my standard error estimates are less than one percent of the
corresponding estimates. Therefore, I do not clutter my figures with confidence bands.

2.3 Cohort shifts in labor force participation and related socioeconomic outcomes of
women
2.3.A Birth cohort and FLFP
Figure 2.2.A displays the life-cycle trajectories of labor force participation of women.
The figure presents only nine of the cohorts for expositional simplicity. (The trajectory for the
1982 cohort is hardly distinguishable from those for the cohorts 1964-1976 due to their
overlapping.) The choice of the cohorts is arbitrary and has no bearing on the conclusions. For
the cohorts up through the 1952 birth year, the life-cycle trajectories are strongly inverse-Ushaped, with labor force participation peaking in the late forties. For more recent cohorts, the

5

The exception is Figure 2.1, which is based on a sample of U.S. men and women age 16 or
older from 1948 to 2010.

15

trajectories are still concave, but less strongly so in the sense that labor force participation is
higher in the twenties than was the case for earlier cohorts.
The central pattern for purposes of this paper is that, comparing across cohorts, labor
force participation rose up through the 1952 cohort and stopped rising for subsequent cohorts.

6

This is a very strong pattern for women in their thirties or older. For example, the FLFP rate in
age group 35-39 was 60 percent for the 1940 cohort, 68 percent for the 1946 cohort, and 76
percent for the 1952 cohort, and then it essentially has held steady for subsequent cohorts: 76
percent for the 1958 cohort, 77 percent for the 1964 cohort, and 75 percent for the 1970 cohort.
The exception to this pattern is that labor force participation for younger women
continued to rise up through the 1970 cohort and then leveled off. Later in this paper, I will
examine the roles of delayed marriage and child rearing in accounting for that pattern.
To examine what happened between the intervening cohorts around the plateau, Figure
2.2.B presents the trajectories for the cohorts 1952-1958. The figure confirms that the rise in
FLFP for women in their thirties and older was brought to an end by the cohorts born after 1952.
For example, for women aged 35-39, the FLFP rates for the 1952, 1953, and 1954 cohorts were
76, 75, and 76 percent, respectively.

2.3.B Educational attainment
As noted by many writers (for example, Goldin, Katz, and Kuziemko, 2006), the rise in
women’s labor force participation was accompanied by a rise in women’s educational
attainment. In this section, I reaffirm the trend in women’s educational attainment, and I explore
its connection to the rise and plateau in women’s labor force participation.
6

Fallick and Pingle (2007) also point out that FLFP leveled off with the cohorts born in the early
1950s.

16

I divide the sample into three education categories: high school or less, some college, and
college completion or more. Women with high school or less education represent female high
school dropouts and graduates; women with some college education include those who had some
college experience but no bachelor’s degree; women with college completion or more denote
7

women with a bachelor's degree.

Figure 2.3 shows the fraction of women by educational attainment and birth cohort, for
age group 25-29. The figure confirms the familiar finding that women’s educational attainment
has risen dramatically across cohorts. While the high school fraction has declined significantly,
both fractions of some college and college completion or more have risen in parallel. For
example, 64 percent of women born in 1948 had not gone to college as of their late twenties.
Women born 33 years later, in contrast, showed a complete reversal: about 64 percent of women
had attended college by the same age. For the 1965 and subsequent cohorts, more than half of
women attended college in this age group. In terms of time period, this indicates that by the
early 1990s, when the plateau in FLFP appeared, the fraction of women with college education
began to exceed the fraction without.
Interestingly, unlike the FLFP rate, the fraction attending college continued to rise
beyond the cohorts born in the early 1950s. The fraction for women aged 25-29 continued to rise
until cohorts born in the early 1970s, and then it leveled off. An important topic for further
research is why the plateau in women’s college-going occurred 20 years after the plateau in labor
force participation. Lee (2012) suggests that the trend towards higher work expectations of

7

In order to construct a measure of educational attainment that is consistent over my sample
period, I adopt the methodology proposed by Jaeger (1997). Because of a 1992 change in CPS
coding of educational attainment, the trend in high school graduation before 1992 is not
comparable to the trend after 1992. I therefore merge the high school dropouts and graduates
into one category.

17

young women across birth cohorts may partly account for the upward trends in women's college
attendance and completion.
Figure 2.4 presents more information on the connection between the trends in FLFP and
women’s educational attainment. The three panels of Figure 2.4 display FLFP rates within each
education category. Comparing the three panels shows that FLFP rates regularly have been
greater for more educated women. Comparing cohorts within each panel shows that both the rise
and plateau in FLFP appear similarly in all education categories. The figure displays the
difference between younger age groups across education categories for cohorts born after 1952.
The participation rate was higher in the late twenties than in the early thirties for women with
college education, while the opposite was true for women without college education. In
combination, these comparisons suggest that trends in FLFP must be related to both the withineducation-category trends and the changing fraction of women with college education.
To quantify the connection between trends in FLFP and educational attainment, I proceed
to a “shift-share” analysis. Let

yd

be a dummy variable that equals one if a woman in group d

is in the labor force and zero otherwise. For now, d is a multiple-category variable that equals
one if a woman has high school education, two if she has college experience, and three if she has
a bachelor’s degree. (In a later section, d will have categories more than three involving various
combinations of college-going, marriage, and child rearing.) Note that the overall FLFP rate y
can be written as

 d yd , where  d denotes group d 's population share and y d
d

is group

d’s labor force participation rate. Then the change in the overall FLFP rate between cohorts s
and t can be decomposed as:

18

yt  ys 

  td  yt d    sd  ys d
d
d

  td  yt d    td  ysd    td  ysd    sd  ysd
d
d
d
d



 ysd  ( td   sd )
d





  td  ( yt d  ysd )
d

(1)

Composition effect Within-groupeffect

The first element in the final expression indicates the composition effect, which is the change in
labor force participation attributable to the change in the population share of each education
category. The second component represents the within-group effect, which is the difference
8

attributable to the changes in labor force participation within each education category.

Figure 2.5 summarizes the results of this decomposition for ages 35-39 and ages 25-29
since the decomposition results for women younger and older than 30 diverge substantially.
Each decomposition is performed in a way that the FLFP rate of a given cohort is compared to
the FLFP rate of a fixed base cohort. The base cohort is the earliest cohort observed in the data
for a given age group. Thus, in Figures 5, 8 and 12, the base cohorts for age groups 35-39 and
25-29 are the 1933 and 1943 cohorts, respectively. Each decomposition figure contains four
lines. The top solid line shows the FLFP rate by cohort, and the bottom solid line presents the
FLFP rate of the base cohort, which is by definition a horizontal line. The broken red line and
the solid blue line with a mark represent within-group and compositional changes, respectively.
The composition effect for a given cohort plots the composition effect in equation (1) holding the
8

Of course, there are numerous ways of decomposing. Alternatively, equation (1) can be

written as

yt  ys 

 yt d  ( td   sd )    sd  ( yt d  ysd ).
d
d

I have performed

decompositions using this alternative equation, and it turns out that the decomposition results are
robust to ways of decomposing.

19

within-group FLFP rates constant at the values for the base cohort. Thus, it shows what the
FLFP rate would have been following the base cohort if there were only compositional changes.
Similarly, the within-group change plots the within-group effect from equation (1) holding
constant the population share of each education category.
The top panel of Figure 2.5 shows that, for ages 35-39, the rise in FLFP up through the
cohorts of the early 1950s was driven almost entirely by the within-group effect, that is, by rising
FLFP among all those with high school or less, some college, and college or more education.
And the plateau for subsequent cohorts reflects the plateaus within all education categories. The
bottom panel shows that, for ages 25-29, the rise in FLFP up through the cohorts of the early
1970s was driven predominantly by the within-group effect, but also to some extent by the rising
shares of college attendees and graduates. The plateau after the early 1970s reflects leveling off
in both the within-group labor force participation rates and the fractions going to college.
To describe the quantitative magnitudes more precisely, I provide Table 2.1 as a
supplement to the decomposition figures. I pick up two arbitrary sets of cohorts, the 1946, 1958,
and 1970 cohorts for age group 35-39 and the 1956, 1968, and 1980 cohorts for age group 25-29,
and report the estimates for the cohorts shown in Figures 5, 8, 11, and 12 in Panels A, B, C, and
D of Table 2.1, respectively. Columns (1), (2), and (3) respectively present the change in FLFP
between the base cohort, the 1933 cohort, and each cohort from the former set; columns (4), (5),
and (6) respectively present the change in FLFP between the base cohort, the 1943 cohort, and
each cohort from the latter set. Column (3) in Panel A of Table 2.1 indicates that seven percent
(=1.75/26.01) of the total change in FLFP between the 1933 and 1970 cohorts for age group 3539 was due to the changes in educational composition of women, whereas column (6) in the
same panel implies that 16 percent (=5.02/30.64) of the change between the 1943 and 1980

20

cohorts for age group 25-29 was. Detailed within-group effects by category indicate that the
within-group effects reported in both columns came almost evenly from the three education
categories.

2.3.C. Marriage
Marriage patterns also have changed dramatically over the past half century (Isen and
Stevenson, 2010). Figure 2.6.A shows the fraction of women ever married (currently married,
9

divorced, separated, or widowed) by age and birth year.

The fraction plummeted for all age

groups. The particularly large declines for women under age 30 reflect a trend towards later
marriage. About 65 percent of women born in 1946 were ever married as of age 20-24. For
women of the same age but born 30 years later, only 28 percent had been married. Since this
sharp decline was accompanied by a rise in the share of college-educated women, one might
think that marital delay is associated with women's college-going behavior. That conjecture is
partly true because the share of ever-married women has been lower among those that attended
college for most cohorts. Figure 2.6.B, however, shows that the share of ever-married women
has not dropped more for women with college than for women without college. For cohorts
born in the late 1960s, the share of ever-married women was even higher among those that attend
college as of age 35-39. For cohorts born before 1970, the share declined more rapidly among
those that complete college as of age 25-29, but the share for each education category showed a
similar pattern of decline for the subsequent cohorts.
Figure 2.7 displays labor force participation rates by age and birth cohort disaggregated
between ever-married and never-married women. The figure shows that never-married women
9

The figure does not include the series for women aged 45 or older because it is not much
different from the series for women aged 40-44.

21

traditionally have participated in the labor force more than ever-married women, but that the
upward trend in FLFP up through the 1952 cohort was especially pronounced for ever-married
women. Again, the two figures in combination suggest that both within-group trends and
composition changes must have contributed to changes in FLFP.
Again, I perform a shift-share analysis to quantify the roles of within-group and
composition changes. The framework is still as shown in equation (1), but this time with a
binary categorization between ever-married and never-married women. The results are
summarized in Figure 2.8. The top panel shows that, for women aged 35-39, the upward trend in
labor force participation up through the cohorts of the early 1950s was driven almost entirely by
rising labor participation within categories, but the detailed within-group effect reported in Panel
B of Table 2.1 indicates that it is completely driven by ever-married women.
The lower panel shows that, for women aged 25-29, the upward trend in labor force
participation up through the cohorts of the early 1950s was mostly driven by within-group
trends, but also reflects the increasing share of women still not married by their late twenties.
The continuing rise in FLFP after those cohorts entirely reflects a continued trend away from
marriage by the late twenties. Column (6) in Panel B of Table 2.1 indicates that this trend
contributed 46 percent (=14.10/30.64) of the change in FLFP between the 1943 and 1980 cohorts
for age group 25-29.

22

2.3.D Child rearing
Women’s labor force participation is strongly related to the presence of young children in
10

the household, so I will define “child rearing” as the presence of children under the age of six.

Figure 2.9 shows age-specific trends across cohorts in this measure of child rearing. The striking
pattern is a trend towards later child rearing: child rearing in the twenties has declined, while
child rearing in the thirties has increased. Of course, this pattern is connected to the trend
towards later marriage. As women in more recent cohorts marry and have children later than
those in earlier cohorts, the contrast across age groups becomes smaller. For example, among
women born in 1946, 60 percent had young children while in their late twenties, compared to
only 22 percent while in their late thirties. For women born 24 years later, 41 percent had young
children while in their late twenties and 32 percent did in their late thirties.
Figure 2.10 shows labor force participation rates by age and cohort, disaggregated
between women who do and do not have children under the age of six. A comparison of the two
panels confirms that labor force participation is much lower among women with young children.
The lower panel shows that, among women without young children, the rise in labor force
participation leveled off after the 1952 cohort. In contrast, the upper panel shows that, among
women with young children, labor force participation continued to rise at least through the
cohorts of the early 1970s.
Figure 2.11 repeats the shift-share analysis, this time with a categorization between
women with and without young children. The lower panel shows that, for women aged 25-29,
the rise in labor force participation up through the cohorts of the early 1970s was driven mostly

10

Although the age of six is a conventional cutoff, it is admittedly arbitrary. However, results
using other age cutoffs (the ages of two, three, four, five and seven) do not change much from
those in Figures 8 and 9.

23

by upward within-group trends, but also by an increasing fraction of women without young
children. The point can be more clearly illustrated using estimates shown in Panel C of Table
2.1. The estimates in column (5) in Panel C imply that 72 percent (=21.97/30.42) of the change
in FLFP between the 1943 and 1968 cohorts was driven by the within-group effect, with a bigger
within-group effect from women with young children. The other 28 percent of the change was
driven by the change in women’s child-rearing composition between the two cohorts.
The upper panel in Figure 2.11 tells a different story about women aged 35-39. For them,
the rise in labor force participation up through the cohorts of the early 1950s was driven almost
entirely by rising labor participation within both groups, women with and without young
children. For subsequent cohorts, the within-group effect actually exceeds the overall change in
labor force participation. That is, the labor force participation rate would have been higher if not
for a negative composition effect due to the increasing share of women with young children in
this age group. The estimates in column (1) indicate that 87 percent (=16.51/19.01) of the
change in FLFP between the 1933 and 1946 cohorts was driven by the within-group effect. The
estimate of the composition effect in column (3) implies that the FLFP rate would have fallen
between the 1933 and 1970 cohorts by 0.4 percentage points, had it not been for the within-group
effect.

2.3.E The joint contribution of education, marriage, and child rearing
Of course, the separate shift-share analyses in the previous sections overlook the
possibility of interactions among the roles of education, marriage, and child rearing. This section
extends the decomposition in equation (1) by using twelve categories: every combination of the

24

three education categories, ever-married or never-married, and with or without children under the
age of six.
The results, shown in Figure 2.12, are much as one would guess from the separate
decompositions in the previous sections. For women aged 35-39, the rise in labor force
participation up through the cohorts of the early 1950s is driven almost entirely by rising labor
force participation within categories. And the plateau for subsequent cohorts reflects a leveling
off of labor force participation within categories. Column (3) in Panel D of Table 2.1 indicates
that 14 percent (=3.65/26.01) of the change in FLFP between the 1933 and 1970 cohorts was due
to the composition effect. Decomposition of the within-group effect for this age group, although
not reported in the table, shows that 74 percent of the within-group change in FLFP between the
1933 and 1946 cohorts was driven by married women without young children and 11 percent of
the within-group change was driven by college-educated married women with young children.
The upward within-group trend among the latter group of women is quantitatively more
important in explaining the change in FLFP between the 1933 and 1970 cohorts, with 32 percent
of the change driven by the latter group.
For women aged 25-29, the rise in labor force participation up through the cohorts of the
early 1950s is driven by both within-group trends and by composition effects, with the latter
mainly associated with the trends away from early marriage and childbearing. The subsequent
rise in labor force participation up through the cohorts of the early 1970s is mostly driven by the
continuation of the trends away from early marriage and childbearing. Column (6) in Panel D of
Table 2.1 indicates that 46 percent (=14.21/30.64) of the change in FLFP between the 1943 and
1980 cohorts was driven by the composition effect. Decomposition of the within-group effect
for this age group suggests that within-group change among high-school-educated married

25

women with young children is quantitatively the most important in explaining the within-group
change in FLFP between the 1943 and 1956 cohorts, with 39 percent of the within-group change
driven by this group. The rate drops to 20 percent for the within-group change in FLFP between
the 1943 and 1980 cohorts.

2.4 Summary and discussion
After many decades of trending upward, the female labor force participation rate in the
United States hit a plateau in the early 1990s. Using data from March Current Population
Surveys, I have documented salient aspects of this important social change. I have shown that,
to a first approximation, the plateau can be parsimoniously characterized as a leveling off of
labor force participation for birth cohorts from the early 1950s on. My series of shift-share
analyses shows that, both the plateau and the earlier upward trend appeared within every
category broken down by education, child rearing, and marital status except that the upward
within-group trend did not appear among never-married women.
The main qualification to that simple summary is that, for women under the age of 30,
labor force participation continued to rise up through the cohorts of the early 1970s. The shiftshare analyses show that the continued rise for younger women was intertwined with the trends
towards later marriage and childbearing. I also have shown that the trend towards more college
education among women continued beyond the cohorts of the early 1950s. The composition
effects from educational attainment, marriage, and child rearing respectively contributed 7, 11,
and -2 percent of the changes in FLFP between the 1933 and 1970 cohorts for age group 35-39.
The corresponding contributions to the change in FLFP between the 1943 and 1980 cohorts for
age group 25-29 are 16, 46, and 33 percent.

26

Of course, all of these findings beg the question of why these interrelated trends have
developed in the ways that they have. Goldin (2006), Juhn and Potter (2006), and Boushey
(2005) speculated that the plateau in women’s labor force participation may have been a
temporary response to the recession of the early 2000s, but my updating of the evidence refutes
that conjecture. To my knowledge, the most ambitious attempt to identify causes of the plateau
is the study by Fortin (2009), which argues that women in more recent cohorts have reverted to
more traditional gender role attitudes than women in preceding cohorts and that this change in
gender role attitudes explains at least a third of the plateau in FLFP.
Clearly, more research is needed. My hope is that, by providing a transparent summary
of the salient facts about recent developments in women’s labor force participation and related
socioeconomic outcomes, the present paper will stimulate and guide further research that will
delve more deeply into the causes of the shifting trends. For example, if real wage growth and
advances in home production technology are credited with explaining the long-running rise in
women's labor force participation, why did those forces become inoperative for cohorts born
after the early 1950s?

27

Table 2.1. Decomposition in difference in FLFP rates between birth cohorts using socioeconomic
outcomes of women
Age group
35-39
25-29
Cohorts between 1933
and

Cohorts between 1943
and

1946

1958

1970

1956

1968

1980

(1)

(2)

(3)

(4)

(5)

(6)

19.01

26.79

26.01

25.32

30.42

30.64

Composition effect

0.65

0.65

1.75

1.87

3.53

5.02

Within-group effect

18.35

26.15

24.26

23.44

26.88

25.62

10.08

10.15

7.11

12.73

9.93

8.88

Some college

4.69

10.30

8.53

6.54

10.18

9.67

College completion or more

3.59

5.69

8.63

4.17

6.77

7.06

19.01

26.79

26.01

25.32

30.42

30.64

Composition effect

0.25

1.76

2.76

5.47

9.84

14.10

Within-group effect

18.75

25.03

23.25

19.84

20.57

16.54

Ever-married

18.44

24.49

22.30

19.46

21.13

16.64

Never-married

0.31

0.55

0.95

0.39

-0.56

-0.10

19.01

26.79

26.01

25.32

30.42

30.64

Composition effect

2.50

0.66

-0.40

6.98

8.45

10.02

Within-group effect

16.51

26.13

26.41

18.34

21.97

20.62

4.58

10.40

12.09

11.40

13.75

13.33

11.92

15.73

14.32

6.93

8.22

7.29

19.01

26.79

26.01

25.32

30.42

30.64

Composition effect

3.06

2.55

3.65

8.63

11.40

14.21

Within-group effect

15.95

24.24

22.36

16.69

19.02

16.43

A. Decomposition using educational attainment
Total change in FLFP

High school or less

B. Decomposition using marital status
Total change in FLFP

C. Decomposition using child-rearing status
Total change in FLFP

With children under 6
Without children under 6
D. Decomposition using joint variable
Total change in FLFP

Note: The table reports selective estimates shown in Figures 5, 8, 11 and 12. The unit of the
estimates is percentage point.

28

Figure 2.1: Labor force participation rates for men and women
90

(%)

85
80
75
70
65
60
55
50

Men

45
40

Women

35
30
48

52

56

60

64

68

72

76 80
Year

84

88

92

96

00

04

08

Note: The figure shows civilian labor force participation rates of U.S. men and
women age 16 or older. For interpretation of the references to color in this and all
other figures, the reader is referred to the electronic version of this dissertation.
Source: BLS series report available at http://data.bls.gov/cgi-bin/srgate

29

Figure 2.2: Female labor force participation rates by age and cohort
A. 1934 - 1982 cohorts
(%)
85
80
75
70
65
60
55
50
45

1934 cohort

1958 cohort

1964 cohort

1970 cohort
35
16-19

1946 cohort

1952 cohort

40

1940 cohort
1976 cohort

1982 cohort

20-24

25-29

30-34

35-39 40-44
Age

45-49

50-54

55-59

60-64

Note: All figures throughout this paper use five-year age bins and single-year
birth cohorts.

81

B. 1952 - 1958
(%)

79
77
75
73
71
69
67
1952 cohort
1955 cohort
1958 cohort

65
63
61
20-24

25-29

30-34

1953 cohort
1956 cohort

35-39
Age

Source: 1968-2010 IPUMS-CPS

30

40-44

1954 cohort
1957 cohort

45-49

50-54

Figure 2.3: Women's educational attainment by cohort, ages 25-29

1
0.9
High school or less
0.8
Some college

0.7

College completion or more

0.6
0.5
0.4
0.3
0.2
0.1
0
1943

1947

1951

1955

1959 1963
Birth year

1967

1971

1975

1979

Note: High school or less category includes female high school dropouts and
graduates as well as GED recipients. Some college includes women with three
or less years of college education and/or women with an associate degree.
College completion or more includes women with four or more years of college
education and/or women with a bachelor's degree.
Source: 1968-2010 IPUMS-CPS

31

Figure 2.4: Female labor force participation rates by educational

90

A. High school or less
(%

85
80
75
70
65
60
55
50
45
40

1934 cohort

1940 cohort

1946 cohort

1952 cohort

35

1958 cohort

1964 cohort

1970 cohort

1976 cohort

30
25-29

30-34

35-39

40-44

45-49

50-54

55-59

60-64

Age

90

B. Some college
(%)

85
80
75
70
65
60
55
50
45
40
35
25-29

1934 cohort

1940 cohort

1946 cohort

1952 cohort

1958 cohort

1964 cohort

1970 cohort

1976 cohort

30-34

35-39

40-44

45-49
Age

32

50-54

55-59

60-64

Figure 2.4 (cont'd)

C. College completion or more
(%)
90
85
80
75
70
65
60
55
50
45
40
35
25-29

1934 cohort

1940 cohort

1946 cohort

1952 cohort

1958 cohort

1964 cohort

1970 cohort

1976 cohort

30-34

35-39

40-44

45-49
Age

Source: 1968-2010 IPUMS-CPS

33

50-54

55-59

60-64

Figure 2.5: Decomposition of cohort difference in FLFP using education

80

A. Ages 35-39
(%)

76
72
68

FLFP rate

64

Within-group change

60

Compositional change

56

1933 cohort

52
48
44
1933

1937

1941

1945

1949 1953
Birth year

1957

1961

1965

1969

Note: Decomposition is implemented on the FLFP change between base and
given cohorts using a multiple-category education indicator. The base cohorts
for age groups 35-39 and 25-29 are the 1933 and 1943 cohort, respectively.

B. Ages 25-29
(%)
80
76
72
68

FLFP rate

64

Within-group change

60

Compositional change

56

1943 cohort

52
48
44
1943

1947

1951

1955

1959 1963
Birth year

Source: 1968-2010 IPUMS-CPS

34

1967

1971

1975

1979

Figure 2.6: Fraction of women ever-married
A. By age group
1
0.9
0.8
0.7
0.6
0.5

Ages 45-49
Ages 40-44

0.4

Ages 35-39

0.3

Ages 30-34

0.2

Ages 25-29
Ages 20-24

0.1

Ages 16-19

0
1932 1936 1940 1944 1948 1952 1956 1960 1964 1968 1972 1976 1980
Birth year

Note: Ever-married women include currently married, divorced, separated
and widowed women.
Source: 1968-2010 IPUMS-CPS

B. By educational attainment and selective age group
1
0.9
0.8
0.7
0.6
Ages 35-39, high school or less
0.5
Ages 35-39, some college
0.4
0.3
0.2

Ages 35-39, college completion or
more
Ages 25-29, high school or less
Ages 25-29, some college

0.1
1933 1937 1941 1945 1949 1953 1957 1961 1965 1969 1973 1977 1981
Birth year

Note: Women with high school include both high school dropouts and
graduates. Women with college completion or more mean women with a
bachelor's degree.

35

85

Figure 2.7: Female labor force participation rates by marital
status
A. Ever-married women
(%)

80
75
70
65
60
55
50
45

1934 cohort
1952 cohort
1970 cohort

40
35
16-19

85

20-24

25-29

1940 cohort
1958 cohort
1976 cohort
30-34

35-39 40-44
Age

1946 cohort
1964 cohort
1982 cohort
45-49

50-54

55-59

60-64

55-59

60-64

B. Never-married women
(%)

80
75
70
65
60
55
50
45
40

1934 cohort
1952 cohort
1970 cohort

1940 cohort
1958 cohort
1976 cohort

1946 cohort
1964 cohort
1982 cohort

35
16-19

20-24

25-29

30-34

35-39

40-44

Age

Source: 1968-2010 IPUMS-CPS

36

45-49

50-54

Figure 2.8: Decomposition of cohort difference in FLFP using marriage

80

A. Ages 35-39
(%)

76
72
68

FLFP rate

64
Within-group change
60
Compositional change
56
1933 cohort
52
48
44
1933

1937

1941

1945

1949 1953
Birth year

1957

1961

1965

1969

Note: Decompositions are implemented on the FLFP change between base and
given cohorts using a dummy variable indicating whether a woman has been
ever-married. The base cohorts for age groups 35-39 and 25-29 are the 1933
and 1943 cohort, respectively.

80

B. Ages 25-29
(%)
FLFP rate

76

Within-group change

72

Compositional change

68

1943 cohort

64
60
56
52
48
44
1943

1947

1951

1955

1959 1963
Birth year

Source: 1968-2010 IPUMS-CPS

37

1967

1971

1975

1979

Figure 2.9: Fraction of women who have children under 6 in the household
0.7
0.6
0.5
0.4
0.3
Ages 20-24

0.2

Ages 25-29
Ages 30-34

0.1

Ages 35-39
0
1933 1937 1941 1945 1949 1953 1957 1961 1965 1969 1973 1977 1981
Birth year

Source: 1968-2010 IPUMS-CPS

38

Figure 2.10: Labor force participation rates for women
with and without young children
90

(%) A. Women with children under 6 in the household

85

1934 cohort

1940 cohort

1946 cohort

80

1952 cohort

1958 cohort

1964 cohort

1970 cohort

1976 cohort

1982 cohort

75
70
65
60
55
50
45
40
35
30
25
20
20-24

90

25-29

Age

30-34

35-39

(%) B. Women without children under 6 in the household

85
80
75
70
65
60
55
50
45
40
35

1934 cohort

1940 cohort

1946 cohort

30

1952 cohort

1958 cohort

1964 cohort

25

1970 cohort

1976 cohort

1982 cohort

20
20-24

25-29

Age

Source: 1968-2010 IPUMS-CPS

39

30-34

35-39

Figure 2.11: Decomposition of cohort difference in FLFP using child-rearing

80

A. Ages 35-39
(%)

76

FLFP rate
Within-group change

72

Compositional change

68

1933 cohort

64
60
56
52
48
44
1933

1937

1941

1945

1949 1953
Birth Year

1957

1961

1965

1969

Note: Decompositions are implemented on the FLFP change between base and
given cohorts using a dummy variable indicating whether a woman has children
under 6 in the household. The base cohorts for age groups 35-39 and 25-29 are
the 1933 and 1943 cohort, respectively.
B. Ages 25-29
(%)
80
FLFP rate
76
Within-group change
72
68

Compositional change
1943 cohort

64
60
56
52
48
44
1943

1947

1951

1955

1959 1963
Birth year

Source: 1968-2010 IPUMS-CPS

40

1967

1971

1975

1979

Figure 2.12: Decomposition of cohort difference in FLFP using joint variable

80

A. Ages 35-39
(%)
FLFP rate

76

Within-group change

72

Compositional change

68

1933 cohort

64
60
56
52
48
44
1933

1937

1941

1945

1949 1953
Birth year

1957

1961

1965

1969

Note: Decompositions are implemented on the FLFP change between base and
given cohorts using a variable which indicates women's educational, marital,
and child-rearing status jointly. The base cohorts for age groups 35-39 and
25-29 are the 1933 and 1943 cohort, respectively.

80

B. Ages 25-29
(%)
FLFP rate

76

Within-group change

72

Compositional change

68

1943 cohort

64
60
56
52
48
44
1943

1947

1951

1955

1959 1963
Birth year

Source: 1968-2010 IPUMS-CPS

41

1967

1971

1975

1979

CHAPTER 3
THE UPWARD TREND IN WOMEN’S COLLEGE-GOING:
THE ROLE OF TEENAGERS’
ANTICIPATED FUTURE LABOR FORCE ATTACHMENT

3.1 Introduction
In recent decades, U.S. women’s rates of college attendance and completion have
1

increased so rapidly that women have surpassed men in both rates. As Figure 3.1 shows, both
the attendance and completion rates increasingly favor women for birth cohorts from the early
1960s on. Although many studies have suggested reasons that women’s college-going exceeds
men’s, most of those studies are surprisingly unhelpful for explaining why the gender gap has
changed over time in the way that it has. In this paper, I focus on the trend in women’s collegegoing and show that changes in teenagers’ anticipated future labor force participation might
explain a substantial part of the upward trend in women's college-going.
For example, one strand of the literature on the gender gap in collegiate attainment
suggests that non-pecuniary costs of college may be lower for women than men. Becker,
Hubbard, and Murphy (2010) and Goldin, Katz, and Kuziemko (2006) argue that greater average
non-cognitive skills of women allow women to prepare better for college and to attend and
complete college with lower costs. High school graduation rates by gender also provide

1

The definitions of college attendance and completion used in the Current Population Surveys
are slightly different from those used in the National Longitudinal Surveys. For samples from
the Current Population Surveys, I define college attendance as at least one year of college
education and college completion as four or more years of college education. For samples from
the National Longitudinal Surveys, I define college attendance as any college enrollment and
college completion as attainment of a bachelor's degree.

42

evidence of women's better preparedness for college. Heckman and LaFontaine (2010) argue
that a decline in the high school graduation rate for men across birth cohorts, in contrast to the
stable trend for women, may explain part of the growing gender difference in college attendance.
However, the lower college cost (reflecting higher non-cognitive skills) for women
cannot explain the trend in college-going by gender because it is not a new phenomenon. As
Buchmann and DiPrete (2006) and others note, girls have outperformed boys in school for a long
time. Moreover, the high school graduation rate for women, which could be another indicator of
non-cognitive skills, has remained stable over the past 40 years (Heckman and LaFontaine,
2010).
Another strand of the literature has suggested that women’s college-going exceeds men’s
because women face a higher wage premium for college education (Charles and Luoh, 2003;
Chiappori, Iyigun, and Weiss, 2009; DiPrete and Buchmann, 2006; Dougherty, 2005). Again,
however, this observation does not explain the trend in college-going by gender because these
studies do not indicate that women’s wage premium for college has risen faster than men’s. In
any case, a recent study by Hubbard (2011) casts doubt on the argument that women’s wage
premium exceeds men’s by showing that the apparent difference may be merely an artifact of
top-coding bias.
So what can explain the trend in women’s college-going relative to men’s? The closest
precedent for the present study is an article by Goldin, Katz and Kuziemko (2006), which argues
that changes in young women's anticipated labor force participation may be an important factor
in the rise of women's college-going. Using the National Longitudinal Survey of Young Women
1968 (NLSYW68) and the National Longitudinal Survey of Youth 1979 (NLSY79), Goldin et al.
find that young women's anticipated participation rose steadily from 1968 to the early 1980s. In

43

addition, they compare college attendance and completion rates between women with future
work plans and women without and find that both rates are significantly higher for women with
work plans. In this paper, I present more up-to-date trends in young women's work expectations
using three National Longitudinal Surveys – the NLSYW68, the NLSY79, and the National
Longitudinal Survey of Youth 1997 (NLSY97). In addition, I further develop Goldin, Katz and
Kuziemko's insight by providing a simple theoretical model to illustrate the potential importance
of young women’s work expectations and by presenting a detailed set of regression analyses to
test the main implication of the model.
The next section of this paper presents the illustrative theoretical model and explains its
practical relevance and testable implications. Section 3.3 describes the NLS data and uses them
to document that teenage women’s expectations of future labor force attachment have indeed
risen across cohorts. Section 3.4 goes on to test empirically whether teenage women’s work
expectations are positively associated with college-going and uses the results to examine the
extent to which rising work expectations can account for the upward trend in women’s collegegoing. Section 3.5 summarizes the findings.

3.2 The potential importance of work expectations for college-going decisions
As noted in the previous section, many studies of gender differences in college-going
stress the potential role of the wage rate premium associated with a college education. Most of
those studies, however, overlook the interaction of the wage premium with anticipated work
hours. To put the point in an extreme way, a high hourly wage premium for a college education
would provide little incentive for college-going to a person who expected to work only one
nanosecond after college. In this extreme case, even with a high hourly wage premium, the

44

earnings payoff to college would be infinitesimal. The more general point that the expected
earnings payoff to college depends on anticipated work hours as well as the hourly wage
premium is illustrated by the following simple model.
Suppose woman i ' s hourly wage Wi is determined by

(1)

Wi  e

where

Di

  Di i

,

is equal to one if woman i goes to college for four years and zero if she does not go to

college. The heterogeneity term  i , which reflects idiosyncratic ability, motivation, etc., is
observed by woman i , but not by the econometrician. The parameter  is the wage rate
premium for college education. Assume that individuals can borrow at interest rate r and that
the wage rate premium for college education is large enough so that some women are willing to
attend college (   4r ).
When deciding whether to attend college or not, woman i compares the expected
present discounted value ( E [PDV] ) with and without college education:

(2)



E [PDV with college]   E [ H i ]  e  i ert dt  Ti
0

1
 ( ) E [ Hi ] e  4r i  Ti , and
r
(3)



E [PDV without college]   E [ H i ]  e i ert dt
0

1
 ( ) E [ H i ] e i ,
r

45

where

E [ Hi ] is woman i 's expected hours of work; Ti is her expected PDV of college costs

besides foregone earnings, such as tuition and net non-pecuniary costs of college; and lifetimes
are long enough that the PDV calculations are approximately correct with an assumption of
infinite lifetimes. Then, individual i will choose to attend college if and only if (2)  (3) , or
equivalently if and only if

(4)

i    ln Ti  ln r  ln(e 4r 1)  ln E [Hi ].

Thus, the probability that woman i attends college is

(5)

Pr[ i    ln Ti  ln r  ln(e 4r 1)  ln E [ Hi ]]
 1 F (  ln Ti  ln r  ln(e 4r 1)  ln E [ Hi ]),

where F (  ) is the cumulative distribution function of  i .
This result confirms the conjecture that the probability of going to college depends
positively on not only the hourly wage premium for college, but also on the individual’s expected
future hours of work. Of course, the model is exceedingly simple in many respects. To name
just a few examples, it abstracts from the possibility of starting college but not finishing; it
ignores the life-cycle evolution of wages and hours after the completion of education; it ignores
uncertainty about future wages; and it assumes perfect capital markets. There is no reason to
think, however, that incorporating any or all such complications would undo the result that

46

optimizing decisions about college-going should depend on anticipations of future work hours.
It follows that one possible reason for women’s increased college-going is a trend
towards greater expectations of future labor force attachment among more recent cohorts of
young women. At first blush, it seems highly plausible that teenage women’s anticipations of
future labor force attachment have risen because, as is well known, women’s labor force
participation rates actually did increase substantially for a long time. For example, Figure 3.2
shows that the female labor force participation rate at age 35 increased by about 30 percentage
points between cohorts born in the early 1930s and those born in the early 1950s. A large
literature explores many possible reasons for this trend, such as advances in home production
technology (Goldin, 2006, for example) and diffusion of more effective contraception (Bailey,
2006; Goldin and Katz, 2002).
Closer inspection of Figures 3.1 and 3.2, however, poses a puzzle. Figure 3.2 shows that
women’s labor force participation rates, after rising for a long time, reached a plateau for birth
cohorts from the early 1950s on. Although Figure 3.2 documents the plateau only for age 35,
Fallick and Pingle (2009) and Lee (2012) show that the plateau applies more broadly. On the
other hand, Figure 3.1 shows that women’s college-going has continued to rise beyond the
cohorts born in the early 1950s. So can changing expectations about future labor force
attachment really explain the continued rise in women’s college-going?
The answer is that it can if the trend across cohorts in expectations about future labor
force attachment has lagged behind those cohorts’ eventual realizations of labor force attachment
in adulthood. For example, Figure 3.2 shows that women born in 1960 and women born in 1975
have turned out to have similar labor force participation rates. But what if the women born in
1960 did not anticipate when they were teenagers that they were going to work for pay so much,

47

but the women born in 1975 did? Then the difference in their beliefs as teenagers about their
future work still could have contributed to a difference in college-going.
Fortunately, this story about changing work expectations does not have to be left to
speculation. Because various waves of the National Longitudinal Surveys have asked teenage
women about their expectations regarding future work, it is possible to track trends in teenagers’
beliefs across cohorts. The next section describes the NLS data in detail and demonstrates that
young women’s anticipations of future labor force attachment have indeed risen across cohorts.
It also is possible within each NLS to test my model’s prediction that those teenage women who
express stronger expectations of future labor force attachment are more likely to go on to college.
Section 3.4 will present the results of such tests.

3.3 Data
My data come from three National Longitudinal Surveys: the NLSYW68, which began
by surveying 5,159 young women aged 14 to 24 in 1968; the NLSY79, which began by
surveying 12,686 young men and women aged 14 to 22 in 1979; and the NLSY97, which began
by surveying 8,984 young men and women aged 12 to 16 in 1997. Those surveys provide
nationally representative data and collect detailed information on future work plans, educational
2

attainment, and family background of the respondents.

2

Moreover, since each survey represents

Unfortunately, data on future work plans of men comparable to those of women from the
NLSYW68 are not available. The National Longitudinal Surveys of Young Men 1966
(NLSYM66) asked about future job plans under a presumption that all men would be working
around age 30. To be specific, the question asked is “what kind of work would you like to be
doing when you are 30 years old?” The answer should be either a specified job code or a choice
between “same as present job” and “don't know.”

48

different generations, data from the three surveys allow me to look at long-run trends in
teenagers’ work expectations for a broad range of birth cohorts.
To examine the trend in women's early work expectations across cohorts, I create a
unified measure of their anticipated participation, a dummy variable indicating whether an
individual expects to be working or not at a later age. I construct the work expectation measure
based on information on their future work plans surveyed in early rounds of each survey. The
NLSYW68 and the NLSY79 ask whether a respondent expects to be working or not at age 35.
The original question is a multiple-choice question that reads: “what would you like to be doing
when you are 35 years old?” The choices are 1) working; 2) same as present job; 3) don’t know;
4) married, keeping house, raising family; 5) other. Although the code varies with each round,
working and getting married are always included in the choice set. The respondents who said
they planned to get married were asked about plans after marriage in a following question. The
subsequent question is: “sometimes women decide to work after they have been married for a
while. If you were to work, what kind of work would you prefer?” And the choices are 1) work
(different job from present or last); 2) work (same job as present or last); 3) don’t plan to work.
Because those respondents chose marriage over work in the first question and the following
question does not specify age at which they are working, I neglect the information from the
3

subsequent question.

3

It seems conceivable that the first question does not provide useful information because the
choices are not exclusive to each other. A respondent might expect to get married and to be
working at age 35, but the question forces her to choose either of them. To examine the
possibility, I compare women’s responses to men’s. While the fraction of women from the
NLSY79 who expressed an expectation of working at age 35 varies across cohorts, that of men
from the same sample was stable across cohorts with a rate higher than 0.9. Therefore, I believe

49

These data, however, are not available for the NLSY97. The NLSY97 asks instead the
anticipated percentage chance that a respondent will be working 20 or more hours per week at
4

age 30.

Figure 3.4 shows the amount of variation in the variable. The distribution is

considerably right-skewed with a weighted average of 94. Of the respondents, about 81 percent
at age 16 and 84 percent at age 17 answered that the chance would be greater than or equal to 90
percent. To construct an expectation variable similar to that from the NLSYW68 and the
NLSY79, I recode the percentage rate and create a dummy variable that equals one if the rate is
above 50 and zero otherwise. Due to the different context of the question and the arbitrariness of
the variable construction, the work expectation data from the NLSY97 are not quite comparable
to those from the NLSYW68 and the NLSY79. In any case, given the necessarily arbitrary
wording of the survey questions, the measure of work anticipation from each of the three surveys
is only a crude proxy for the young women’s beliefs about their future labor force attachment.
Figure 3.3 displays the data by age and sample. The vertical axis represents the share of
women who thought they would be working in their thirties. Note that point estimates for older
women from the NLSYW68 and the NLSY79 are based on more observations due to the panel
nature of the samples. For example, age 14 from the NLSYW68 contains women born in 1954
and no longer exists in the data after 1968. Age 19 from the same sample, however, includes
women born in 1949 to 1954 because 19-year-old women are observed simultaneously in 1968
through 1973. In the case of the NLSY97, the ages 16 and 17 contain women born in 1980 and
1981, respectively, since the NLSY97 asked the question only at the first round of the

the first question provides useful information unless women and men take the meaning of the
question differently.
4

The NLSY97 asks: “what is the percent chance that you will be working for pay more than 20
hours per week when you turn 30?”

50

5

interview.

The figure shows that there is a strong cohort effect, but not a noticeable age effect

in the trend. In all age groups, the fraction of young women who expect to be working at age 35
rose sharply across cohorts – around 0.45 for the early 1950s cohorts and 0.75 for the early
1960s cohorts. The estimate for the early 1980s cohorts from the NLSY97, which is weakly
comparable with those from the previous two cohort groups, is 0.95. This increasing pattern of
women's anticipated labor force attachment and the upward trend in their college-going are
consistent with the proposition that the two are positively related.
Figure 3.5 illustrates their positive relationship more clearly. It shows four trends: actual
labor force participation in the broken black line; teenagers’ anticipated labor force participation
at age 35 in the red line with circle marks; teenagers’ anticipated chance of working 20 or more
hours of work at age 30 in the grey line with triangle marks; and college attendance rates in the
6

solid blue line. The rates of both college attendance and anticipated participation have risen
across birth cohorts, while the actual participation rate has plateaued. The rise in women's
college-going and the plateau in labor force participation happened around the same time
because not all of the early work plans were realized. In fact, the discrepancy between the
anticipated and actual participation is quite large. Women's actual labor force participation can
be viewed as a realized version of young women's anticipated participation. The comparison of
5

The NLSYW68 and the NLSY79 surveyed future work plans repeatedly, from 1968 to 1987
for the NLSYW68 and from 1979 to 1984 for the NLSY79.
6

An odd feature of Figure 3.5 is that the anticipated participation rate for the 1950 cohort is
much lower than for any other cohort. It seems that this lower estimate is due to a mysterious
1968 survey effect. The anticipated participation rate from the 1968 wave is smaller than in any
other wave not only for the 18-year-olds, but for other ages as well. The regression results in the
next section are qualitatively robust to this survey effect. Excluding observations from the 1968
wave removes 370 observations from my regression sample, but the magnitude of the estimated
effect of the anticipation variable on college-going increases by around 0.04, so that the
significance level stays about the same.

51

the anticipated and actual participation reveals that while earlier cohorts of women considerably
underestimated how much they would turn out to work, recent cohorts of women overestimated
their future labor force participation.
The main thesis of this paper is that rising expectations of future labor force attachment
among teenage women may play an important role in women’s increasing college-going. This
story involves two empirically testable propositions: 1) that teenage women’s expectations of
future work have indeed increased across cohorts, and 2) that teenage women who express
stronger expectations of future work are more likely to attend and complete college. Having
used the NLS data in this section to verify the first proposition, in the next section I will use them
to test the second. Because the regression analyses in the next section will focus on the
relationship between young women’s anticipated labor force attachment and their college-going,
I will restrict the regression samples to young women with information on their work
expectations. For women from the NLSYW68 and the NLSY79, I use their work expectations
when they are 18 since the majority of young women decide whether to go to college at that
7

age.

For women from the NLSY97, I use their work expectations when they are 17 because

their anticipation at age 18 is not available. I also limit the college outcome variables to
enrollment by age 19 and completion by age 25 because the work anticipation when they are 17
or 18 is more likely to be relevant to college-going decisions made around age 18.

7

Among women with information on college enrollment by age 19, about 18 percent of women
from the NLSYW68 and 21 percent of women from the NLSY79 are already in college at age
18. Empirical results in the next section are robust to using their work expectations when they
are 17. Using the earlier work expectations measure, I lose at most 533 observations in my
regression sample and the magnitude of the estimated effect of the expectations diminishes by
about 60 percent although the effect still remains highly statistically significant.

52

I exclude observations with missing information on the college outcomes and family
background from my regression samples. Some supplementary regressions will include controls
for intelligence: IQ scores for the NLSYW68, percentile scores for the Armed Forces Qualifying
Test (AFQT) for the NLSY79, and percentile scores on the Armed Services Vocational Aptitude
Battery (ASVAB) for the NLSY97. However, the main regression samples will keep
observations with the intelligence measures missing because the deletion of those observations
leads to a sample that seriously over-represents college-goers. See Table 3.A1 for the details of
how each deletion changes the mean of each college outcome variable.
In the end, the regression samples from the NLSYW68, the NLSY79, and the NLSY97
contain 1,734, 2,377, and 731 observations, respectively. Table 3.1 provides weighted summary
statistics by sample. Women from the NLSY97 have the highest average college attendance and
completion rates, while women from the NLSY79 have higher average rates than those from the
NLSYW68, consistent with the national trends. However, as shown in Table 3.A1, all the
estimates are a bit higher than the comparable estimates from the Current Population Surveys
(CPS). Regarding college attendance, this discrepancy may come in part from the different
definitions used in the two data sets. In the NLS samples, individuals who have enrolled in
college are coded as college-goers, even individuals with less than one year of college
experience. In the CPS, however, individuals with less than one year of college education are
categorized as high school graduates. Because women who were 19 or older at the first
interview do not have information on their anticipated labor force participation when they were
18, they are dropped from the regression samples. As a result, each sample consists of younger
cohorts compared to the original sample. Whites are over-represented in the samples from
NLSYW68 and NLSY79, but are under-represented in the sample from NLSY97.

53

3.4 Evidence on the association between work expectations and college-going
To examine whether teenage women who express higher expectations of future labor
force attachment are more likely to attend and complete college, I estimate models of the form

(6)

where

Pr(Di  1)  F (a  bHi  X ic)

Di

is a binary college-going outcome for woman i , F (  )  0 ,

woman’s anticipated labor force attachment,

H i is a measure of the

X i is a vector of observed demographic and family

background variables that may affect her college-going decision, and a, b , and c are parameters.
The analysis in Section 3.2 predicts that b is positive. Due to data availability,
in two ways. In regressions based on the NLSYW68 or NLSY79,

H i is measured

H i is a dummy variable that

indicates whether the teenage woman expects to be working at age 35. In regressions based on
the NLSY97,

H i is the teenage woman’s anticipated chance of working 20 or more hours per

week at age 30. In order to retain a sufficient sample size, I use a relatively sparse set of control
variables. The vector

X i includes parents' years of schooling, dummy variables for white and

birth year, and in some regressions, intelligence measures.
Tables 3.2, 3.3 and 3.4 present results from estimating the logit version of equation (6).
The three tables are based respectively on samples from the NLSYW68, the NLSY79, and the
NLSY97. The tables report the estimated average partial effect (APE) for each explanatory

54

8

variable.

In columns (1)-(3), the dummy dependent variable is for attending college by age 19;

in columns (4)-(6), it is for completing college by age 25. The specifications in columns (1) and
(4) contain only

H i as an explanatory variable; those in columns (2) and (5) additionally

include parents' years of schooling and dummies for white and birth cohort; those in columns (3)
and (6) also control for an intelligence measure. In these tables, I report results based on the
weighted quasi-maximum likelihood estimator (Manski and Lerman, 1977). Tables 3.A2-3.A4
show the logit coefficient estimates along with robust standard error estimates. The standard
error estimates accompanying the estimated APEs in Tables 3.2-3.4 are calculated by the delta
method (Bartus, 2005).
In Table 3.2, the estimated effect of the work expectations dummy from every
specification is positive and statistically significant, which indicates that women who expect to
be working at a later age are more likely to attend and complete college. The estimates in
columns (2) and (5) mean that, other things equal, the college attendance and completion
probabilities are about 0.17 and 0.14 higher, respectively, when a young woman expects to be
working at age 35. Comparing across columns, once the demographic and family background
variables and IQ scores are added as controls, the estimate becomes smaller in magnitude.
Estimated effects of the control variables also show expected signs. As many studies point out,
9

parents' schooling has a positive association with college attendance and completion.

Whites

are less likely to invest in college education, other things equal, although the negative

8

The reported APE is the weighted sample mean of the estimated partial effects for each
individual in the sample.
9

For example, see Haveman and Wolfe (1995). However, there is an ongoing debate about the
positive association. Behrman and Rosenzweig (2002) and Plug (2004) argue that an increase in
the schooling of mothers would not have a positive impact on the schooling of children.

55

relationship is statistically significant only when controlling for IQ scores. The negative estimate
on white is fairly common in the literature (for example, Jepson, 2008; Light and Strayer, 2002).
Table 3.3 presents results very similar to those in Table 3.2. Young women's
expectations of working at a later age are highly correlated with their college-going outcomes in
the NLSY79, as well as in the NLSYW68. One difference from Table 3.2 is that the estimated
effects for the white dummy are always statistically significant, and even positive in the
regressions for college completion. With AFQT percentile scores as controls, however, the
estimate turns back to negative. Interestingly, the estimated effects for the work expectations
variable are smaller in magnitude for college completion than for college attendance. One
possible explanation is that the time gap between the moment at which expectations are reported
and the time of college graduation may weaken the relationship. In my empirical setup, the gap
could be seven years at maximum, which is a long enough period to change individuals'
circumstances. Unforeseen events arising during the period, such as economic hardship or
waning interest in school, may prevent some sample members from completing college.
Table 3.4 presents the results using the NLSY97 sample. The estimated effects of the
work expectations variable are, again, positive and usually statistically significant. For
interpreting the results in Table 3.4, on needs to recognize that the key explanatory variable is
now the anticipated percent chance that an individual is working 20 or more hours at age 30.
Unlike in Table 3.3, the estimated APEs are similar in magnitude across the college outcomes.
The estimates in columns (2) and (5) mean that a 10 percentage point increase in the anticipated
chance of work increases the college attendance and completion probabilities by about four
percentage points. Once ASVAB percentile scores are added as controls, the estimates become
less significant, maybe because of a small sample size caused by the deletion of observations

56

with the percentile scores missing. Except for the estimates for father's years of schooling,
estimates for other covariates are generally significant in a statistical sense. The magnitude and
statistical significance of estimates on white vary widely across the specifications.
Further analyses reported in Tables 3.5 and 3.6 check the robustness of the results to
weighting and specification of functional form. The weighted estimation may not be appropriate
for my samples, especially the samples from the NLSYW68 and the NLSY79, in which blacks
are under-represented, because the sampling weights provided by the National Longitudinal
Surveys are to adjust for over-representation of blacks. Thus, in Table 3.5, I also perform
unweighted estimation of the specifications from columns (2) and (5) of Tables 3.2-3.4. The
unweighted estimates reported in columns (1) to (6) in the table indicate that there is no
important difference between the weighted and unweighted estimates. The unweighted estimates
for work expectation variables are generally a little smaller than the weighted estimates, except
the one from the regression for college completion using the NLSYW68. The sign and
significance of the unweighted estimates, however, are almost identical to those of the weighted
estimates.
To check robustness to functional form, I estimate linear probability models of collegegoing, as an alternative to the logit specifications in the preceding tables. Columns (1) to (6) in
Table 3.6 present the results. The weighted least squares (WLS) estimates in general agree in
sign and significance with the logit APE estimates, with the WLS estimates being of a bit greater
magnitude.
Combined with the evidence in Section 3.3 that young women’s expectations of future
labor force attachment have increased across cohorts, this section’s evidence that such
expectations are positively associated with college-going suggests that rising expectations of

57

labor force attachment may account for at least some of the increase in women’s college
attendance and completion. But how much? To get a rough idea, I perform an Oaxaca-Blinder
decomposition. I use the following simple decomposition technique:

(7)

Y 97  Y 68  X 97b*97  X 68 b*68
 ( X 97  X 68 ) b*68  X 97 (b*97  b*68 ) ,
explained effect unexplained effect

where

Y j is a college attendance or completion rate of sample j, X j is an anticipated

participation rate of sample j, and b* j is a coefficient estimate from the linear regression of
college outcome on anticipated participation using sample j. For simplicity, I am emphasizing
the decomposition based on the linear probability model, but a more elaborate extension for the
logit specification (Fairlie, 2005) yields very similar results. One problem is that the
decomposition needs comparable explanatory variables across samples, which are not available
in my data sets. To estimate b97 , I use a binary variable indicating whether, at age 17, a woman
expects with at least 90 percent chance to be working 20 or more hours at age 30. Setting the
threshold as high as 90 percent stacks the deck somewhat against finding a large role for
changing expectations. Nevertheless, it turns out that the change in women’s early work
expectations between the NLSYW68 and the NLSY97 can explain 66 percent of the change in
10

their college attendance and 69 percent of the change in their college completion.

10

Of course,

The estimates of the explained effect are 0.074 (with standard error of 0.011) and 0.060 (with
standard error of 0.009), respectively, for the changes in college attendance (0.112) and
completion (0.087).

58

the Oaxaca-Blinder calculation may provide only a crude approximation of the connection
between the trend in women's early work expectations and the trend in their college-going
because it is based on crude measures of work expectations, which are not quite comparable
across samples.

3.5 Summary
In recent decades, U.S. women’s rates of college attendance and completion have
overtaken men’s, and are now considerably higher. This study follows up on a suggestion by
Goldin, Katz, and Kuziemko (2006) that rising expectations of future labor force attachment
among young women may account for part of the increase in women’s college-going.
In Section 3.2, I formalized the insight from Goldin et al. by presenting a simple
theoretical model illustrating the role of work expectations in optimal college-going decisions.
In Section 3.3, I used information on teenage women’s work expectations from a succession of
National Longitudinal Surveys to verify that work expectations have indeed risen across cohorts.
In Section 3.4, I used the NLS data to corroborate that young women with higher work
expectations are considerably more likely to attend and complete college. In combination, the
results indicate that young women’s rising work expectations may indeed be an important part of
the story behind their rising rates of college attendance and completion.

59

Table 3.1. Summary statistics
Women from

NLSYW68
Mean

NLSY79

(Std. dev.) Mean

NLSY97

(Std. dev.) Mean

(Std.dev.)

Dependent variable
1 if
attending college by age 19

0.475

(0.500)

0.497

(0.500)

0.587

(0.493)

completing college by age
25

0.215

(0.411)

0.237

(0.426)

0.302

(0.459)

Father's years of schooling

10.9

(3.78)

11.7

(3.42)

13.9

(3.33)

Mother's years of schooling

11.1

(2.78)

11.6

(2.63)

13.2

(2.99)

0.933

(0.161)

0.934

(0.249)

0.721

(0.462)

1.00

(0.00)

Independent and control
variables

Anticipated chance of work
when 17
1 if
expecting to work at age 35
when 18

0.458

(0.498)

0.698

(0.459)

expecting to work at age 30
when 17
white

0.899

(0.301)

0.826

(0.379)

1949

0.196

(0.397)

1950

0.184

(0.388)

1951

0.207

(0.405)

1952

0.231

(0.422)

1953

0.181

(0.385)

1960

0.173

(0.378)

1961

0.222

(0.416)

1962

0.199

(0.399)

1963

0.202

(0.402)

1964

0.204

(0.403)

birth cohort

1980

60

Table 3.1. (cont’d)
Women from
Number of observations

NLSYW68

NLSY79

NLSY97

1,734

2,377

731

Note: Custom weights offered by NLS Investigator are used. Parents' years of schooling greater
than 18 in the NLSY79 and the NLSY97 samples are recoded as 18, in the same fashion as the
NLSYW68. The work expectations variable from the NLSYW68 and the NLSY79 indicates
whether a respondent expects to be working at age 35, while that from the NLSY97 indicates
whether she expects to be working 20 or more hours per week at age 30 with a chance higher
than 50 percent.

61

Table 3.2. Average partial effects from weighted logit regressions using NLSYW68
Women from NLSYW68

Dependent variable
Attending college by age 19

Completing college by age 25

(1)

(2)

(3)

(4)

(5)

(6)

0.181**
(0.026)

0.172**
(0.022)

0.158**
(0.027)

0.147**
(0.022)

0.140**
(0.019)

0.133**
(0.025)

Father's years of schooling

0.033**
(0.004)

0.039**
(0.005)

0.024**
(0.003)

0.019**
(0.005)

Mother's years of schooling

0.055**
(0.005)

0.038**
(0.007)

0.022**
(0.005)

0.016**
(0.006)

Dummy variable for
expecting to work at age
35 when 18

IQ score

0.010**
(0.001)

0.009**
(0.001)

Dummy variable for
white
birth cohort
Number of observations

-0.038
(0.034)

-0.159**
(0.039)

-0.011
(0.027)

-0.149**
(0.049)

No

Yes

Yes

No

Yes

Yes

1,734

1,734

1,092

1,734

1,734

1,092

Note: Custom weights offered by NLS Investigator are used. Each point estimate is the average
of the partial effect for units across the sample distribution of each covariate. Robust standard
errors are in parentheses. * denotes p < 0.05 and ** represents p < 0.01.

62

Table 3.3. Average partial effects from weighted logit regressions using NLSY79
Women from NLSY79

Dependent variable
Attending college by age 19
(1)

(2)

Completing college by age 25

(3)

(4)

(5)

(6)

0.121**
(0.022)

0.117**
(0.022)

0.074**
(0.017)

0.054**
(0.015)

Dummy variable for
expecting to work at age 0.203** 0.156**
35 when 18
(0.027)
(0.023)
Father's years of schooling

0.036**
(0.004)

0.019**
(0.004)

0.026**
(0.003)

0.012**
(0.003)

Mother's years of schooling

0.045**
(0.006)

0.033**
(0.005)

0.030**
(0.005)

0.020**
(0.004)

Percentile score on AFQT

0.007**
(0.000)

0.005**
(0.000)

Dummy variable for
white
birth cohort
Number of observations

-0.075** -0.213**
(0.022)
(0.020)

0.039*
(0.017)

-0.085**
(0.020)

No

Yes

Yes

No

Yes

Yes

2,377

2,377

2,316

2,377

2,377

2,316

Note: Custom weights offered by NLS Investigator are used. Each point estimate is the average
of the partial effect for units across the sample distribution of each covariate. Robust standard
errors are in parentheses. * denotes p < 0.05 and ** represents p < 0.01.

63

Table 3.4. Average partial effects from weighted logit regressions using NLSY97
Women from NLSY97

Dependent variable
Attending college by age 19
(1)

Anticipated chance of work
at age 30 when 17

(2)

(3)

0.415** 0.372**
(0.124) (0.125)

Father's years of schooling

Completing college by age 25
(4)

(5)

(6)

0.233
(0.160)

0.459**
(0.148)

0.370**
(0.129)

0.285*
(0.128)
0.007
(0.006)

0.009
(0.006)

0.004
(0.006)

0.012*
(0.005)

0.033**
(0.007)

0.018**
(0.007)

0.028**
(0.007)

Mother's years of schooling

0.016*
(0.008)

Percentile score on ASVAB
0.008**
(0.001)

0.006**
(0.000)

Dummy variable for
white

Number of observations

0.161**
(0.039)
731

-0.035
(0.039)

731

593

0.191**
(0.031)
731

731

0.046
(0.037)
593

Note: Custom weights offered by NLS Investigator are used. Each point estimate is the average
of the partial effect for units across the sample distribution of each covariate. Robust standard
errors are in parentheses. * denotes p < 0.05 and ** represents p < 0.01.

64

Table 3.5. Estimates from unweighted logit estimation
APE from unweighted logit estimation
Dependent variable

Attending college by 19

Completing college by 25

NLS sample

68

79

97

68

79

97

(1)

(2)

(3)

(4)

(5)

(6)

0.164**
(0.022)

0.146**
(0.021)

0.143**
(0.019)

0.069**
(0.015)

Dummy variable for
expecting to work
at age 35 when 18
Anticipated chance of work at
age 30 when 17

0.372**
(0.001)

0.316*
(0.126)

Father's years of schooling

0.032**
(0.003)

0.029**
(0.003)

0.005
(0.005)

0.021**
(0.003)

0.024**
(0.003)

0.008
(0.005)

Mother's years of schooling

0.049**
(0.005)

0.035**
(0.004)

0.028**
(0.006)

0.022**
(0.004)

0.025**
(0.004)

0.025**
(0.006)

-0.020
(0.027)

-0.100**
(0.020)

0.171**
(0.037)

-0.030
(0.025)

0.036*
(0.016)

0.195**
(0.031)

Yes

Yes

No

Yes

Yes

No

1,734

2,377

731

1,734

2,377

731

Dummy variable for
white
birth cohort
Constant
Number of observations

Note: Robust standard errors are in parentheses. * denotes p < 0.05 and ** represents p < 0.01.

65

Table 3.6. Estimates from weighted least squares estimation of LPM
Weighted least squares estimation of LPM
Dependent variable

Attending college by 19

Completing college by 25

NLS sample

68

79

97

68

79

97

(1)

(2)

(3)

(4)

(5)

(6)

0.181**
(0.024)

0.168**
(0.025)

0.150**
(0.021)

0.094**
(0.021)

Dummy variable for
expecting to work
at age 35 when 18
Anticipated chance of work at
age 30 when 17

0.369**
(0.001)

0.298**
(0.086)

Father's years of schooling

0.032**
(0.004)

0.037**
(0.004)

0.009
(0.006)

0.023**
(0.003)

0.031**
(0.004)

0.012*
(0.005)

Mother's years of schooling

0.050**
(0.005)

0.040**
(0.005)

0.033**
(0.007)

0.021**
(0.004)

0.028**
(0.005)

0.030**
(0.007)

-0.029
(0.033)

-.086**
(0.023)

0.164**
(0.039)

-0.008
(0.024)

0.013
(0.018)

0.182**
(0.033)

No

Yes

Yes

Dummy variable for
white
birth cohort
Constant
Number of observations

Yes

Yes

No

-0.452** - 0.426** -0.438** - 0.304** -0.517** -0.672**
(0.047) (0.047) (0.146) (0.043)
(0.047) (0.126)
1,734

2,377

731

1,734

2,377

731

Note: Weighted least squares estimation of LPM uses custom weights offered by NLS
investigator. Robust standard errors are in parentheses. * denotes p < 0.05 and ** represents
p < 0.01.

66

Table 3.A1. Weighted averages of college outcome variables by sample
Sample of women

College attendance by 19
(CPS estimate)

College completion by 25 Observation count
(CPS estimate)
[attd/cmplt]

1. Before any deletion
NLSYW68

0.315 (0.355)

0.167 (0.175)

4,103/5,123

NLSY79

0.401 (0.402)

0.205 (0.220)

5,114/5,649

NLSY97

0.603 (0.580)

0.362 (0.329)

4,181/3,843

2. After deletion of observations with missing values of work expectation, college outcome, and
parents' education (Samples presented in Table 1)
NLSYW68

0.475 (0.373)

0.215 (0.204)

1,734/1,734

NLSY79

0.497 (0.410)

0.237 (0.226)

2,377/2,377

NLSY97

0.587 (0.580)

0.302 (0.329)

731/731

3. After the deletion of observations with missing values of work expectation, college outcome,
parents' education, and intelligence measure
NLSYW68

0.527 (0.373)

0.240 (0.204)

1,092/1,092

NLSY79

0.502 (0.410)

0.241 (0.226)

2,316/2,316

NLSY97

0.603 (0.580)

0.320 (0.329)

593/593

Note: The comparable estimates for the NLSYW68 (the NLSY79) in Panel 1 are based on
women born between 1942 and 1953 (1957 and 1964), while those for the NLSYW68 (the
NLSY79) in Panels 2 and 3 are based on women born between 1949 and 1953 (1960 and 1964).
The CPS estimates of college attendance are based on women aged 20-24 with some college
education and those of college completion are based on women aged 25-29 with a bachelor's
degree. The comparable CPS estimates for the NLSY97 are identical throughout the panels
because the estimates are based on a single birth cohort, the cohort 1980.

67

Table 3.A2. Coefficient estimates from weighted logit regressions using NLSYW68
Women from NLSYW68

Dependent variable
Attending college by age 19

Completing college by age 25

(1)

(2)

(3)

(4)

(5)

(6)

0.732**
(0.108)

0.962**
(0.131)

0.988**
(0.174)

0.879**
(0.134)

1.025**
(0.146)

0.974**
(0.185)

Father's years of schooling

0.188**
(0.025)

0.246**
(0.034)

0.176**
(0.027)

0.141**
(0.035)

Mother's years of schooling

0.312**
(0.035)

0.244**
(0.044)

0.160**
(0.035)

0.123**
(0.046)

Dummy variable for
expecting to work at age
35

IQ score

0.065**
(0.007)

0.066**
(0.008)

Dummy variable for
white
birth cohort
Constant
Number of observations

-0.218
(0.195)

-1.08**
(0.294)

-0.077
(0.193)

-1.01**
(0.313)

No

Yes

Yes

No

-0.438**
(0.075)

-5.67**
(0.411)

-11.5**
(0.846)

-1.75**
(0.102)

-5.48**
(0.423)

-10.9**
(0.855)

1,734

1,734

1,092

1,734

1,734

1,092

Yes

Yes

Note: Custom weights offered by NLS Investigator are used. Robust standard errors are in
parentheses. * denotes p < 0.05 and ** represents p < 0.01.

68

Table 3.A3. Coefficient estimates from weighted logit regressions using NLSY79
Women from NLSY79

Dependent variable
Attending college by age 19
(1)

Completing college by age 25

(2)

(3)

(4)

(5)

(6)

0.817**
(0.121)

0.744**
(0.138)

0.710**
(0.150)

0.662**
(0.157)

0.592**
(0.170)

Father's years of schooling

0.188**
(0.023)

0.119**
(0.024)

0.220**
(0.031)

0.129**
(0.033)

Mother's years of schooling

0.240**
(0.032)

0.203**
(0.033)

0.250**
(0.041)

0.211**
(0.042)

Dummy variable for
expecting to work at age
35

0.832**
(0.114)

Percentile score on AFQT

0.041**
(0.003)

0.053**
(0.004)

Dummy variable for
white
birth cohort
Constant
Number of observations

-0.400**
(0.120)
No

-1.41**
(0.157)

Yes

Yes

No

Yes

-5.14**
(0.392)

-1.69**
(0.133)

-7.77**
(0.536)

-8.30**
(0.577)

2,316

2,377

2,377

2,316

-0.600** -5.17**
(0.098)
(0.378)
2,377

2,377

0.336*
(0.149)

-0.874**
(0.208)
Yes

Note: Custom weights offered by NLS Investigator are used. Robust standard errors are in
parentheses. * denotes p < 0.05 and ** represents p < 0.01.

69

Table 3.A4. Coefficient estimates from weighted logit regressions using NLSY97
Women from NLSY97

Dependent variable
Attending college by age 19

Completing college by age 25

(1)

(3)

(4)

(5)

(6)

1.74**
(0.540)

1.71**
(0.591)

1.44
(1.00)

2.23**
(0.738)

2.37**
(0.837)

2.20*
(1.00)

0.043
(0.029)

0.027
(0.036)

0.074*
(0.033)

0.056
(0.043)

0.153**
(0.037)

Anticipated chance of work
at age 30 when 17

(2)

0.113**
(0.041)

0.180**
(0.046)

Father's years of schooling
Mother's years of schooling

Percentile score on ASVAB

0.124*
(0.063)
0.045
(0.005)

0.051**
(0.005)
Dummy variable for
white
0.713**
(0.171)
Constant
Number of observations

-1.28
(0.511)
731

-0.221
(0.250)

1.20**
(0.213)

0.353**
(0.279)

-4.35**
(0.777)

-5.01**
(1.169)

-2.96**
(0.711)

-7.52**
(1.16)

-8.21**
(1.45)

731

593

731

731

593

Note: Custom weights offered by NLS Investigator are used. Each point estimate is the average
of the partial effect for units across the sample distribution of each covariate. Robust standard
errors are in parentheses. * denotes p < 0.05 and ** represents p < 0.01.

70

Figure 3.1: College attendance and completion rates
0.7
College attendance (women, age 20-24)
0.6

College attendance (men, age 20-24)
College completion (women, age 25-29)

0.5

College completion (men, age 25-29)

0.4
0.3
0.2
0.1
0
1937 1941 1945 1949 1953 1957 1961 1965 1969 1973 1977 1981 1985
Birth year

Note: CPS-provided sample weights are used. College attendance indicates at
least one year of college education. College completion denotes four or more
years of college education and/or the attainment of a bachelor's degree.
Source: 1962-2011 IPUMS-CPS.

71

Figure 3.2: Labor force participation rate, age 35

0.95

0.85

0.75

0.65
Men
0.55
Women

0.45

0.35
1932 1936 1940 1944 1948 1952 1956 1960 1964 1968 1972 1976
Birth year

Note: CPS-provided weights are used.
Source: 1962-2011 IPUMS-CPS.

72

Figure 3.3: Women's work expectations

1
0.9
0.8
0.7

NLSY97 (birth cohorts 1980 - 1981)

0.6

NLSY79 (birth cohorts 1960 - 1965)

0.5

NLSYW68 (birth cohorts 1949 - 1954)

0.4
0.3
14

15

16

Age

17

18

19

Note: The NLSYW68 and NLSY79 data are based on a question of whether a
respondent expects to be working at age 35. The NLSY97 data are based on a
question asked about the percent chance of working at least 20 hours per week at
at age 30. I recode the chance and create a dummy variable which has 1 if the
percentage rate is above 50 or 0 otherwise.
Source: The NLSYW68, NLSY79, and NLSY97.

73

Figure 3.4: Anticipated chance of work for women from NLSY97
Frequency
count
600

500

400

300
Percent chance when age 16
200

Percent chance when age 17

100

0
0

3

15

25

35

45

60 70 78 83 87
Percent chance (%)

89

93

95

97

99

Note: Total numbers of observations are 912 for age 16 and 822 for age 17.
Source: The NLSY97

74

Figure 3.5: Women's college attendance
and anticipated and actual participation
1
0.9
0.8

Actual participation (age 35)
Anticipated participation at age 35 (age 18)
Anticipated chance of work at age 30 (age 16 or 17)
College attendance (ages 20-24)

0.7
0.6
0.5
0.4
0.3
0.2
0.1
1932 1936 1940 1944 1948 1952 1956 1960 1964 1968 1972 1976 1980 1984
Birth year

Source: IPUMS-CPS (actual participation, college attendance), the NLSYW68
and the NLSY79 (anticipated participation), and the NLSY97 (anticipated
chance of work)

75

REFERENCES

76

References
Aaronson, Stephanie, Bruce Fallick, Andrew Figura, Jonathan Pingle, and William Wascher.
2006. “The Recent Decline in the Labor Force Participation Rate and Its Implications for
Potential Labor Supply.” Brookings Papers on Economic Activity, 2006: 69-154.
Angrist, Joshua D., and Jörn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An
Empiricist’s Companion. Princeton: Princeton University Press.
Arellano, Manuel. 1987. “Computing Robust Standard Errors for Within-Groups Estimators.”
Oxford Bulletin of Economics and Statistics, 49(4): 431-4.
Bailey, Martha J. 2006. “More Power to the Pill: The Impact of Contraceptive Freedom on
Women's Life Cycle Labor Supply.” Quarterly Journal of Economics, 121(1): 289–320.
Bartus, Tamás. 2005. “Estimation of Marginal Effects using Margeff .” Stata Journal, 5(3): 309–
29.
Becker, Gary S., William H. J. Hubbard, and Kevin M. Murphy. 2010. “Explaining the
Worldwide Boom in Higher Education of Women.” Journal of Human Capital, 4(3):
203–41.
Behrman, Jere R., and Mark R. Rosenzweig. “Does Increasing Women's Schooling Raise the
Schooling of the Next Generation?” American Economic Review, 92(1): 323-34.
Boushey, Heather. 2005. “Are Women Opting Out? Debunking the Myth.”
Economic and Policy Research Report, Washington, DC.

Center for

Buchmann, Claudia, and Thomas A. DiPrete. 2006. “The Growing Female Advantage in College
Completion: The Role of Family Background and Academic Achievement.”American
Sociological Review, 71(4): 515–41.
Charles, Kerwin Kofi, and Ming-Ching Luoh. 2003. “Gender Differences in Completed
Schooling.” Review of Economics and Statistics, 85(3): 559–77.
Chiappori, Pierre-Andre, Murat Iyigun, and Yoram Weiss. 2009. “Investment in Schooling and
the Marriage Market.” American Economic Review, 99(5): 1689–1713.
Dickens, William T. 1990. “Error Components in Grouped Data: Is It Ever Worth Weighting?”
Review of Economics and Statistics, 72(2): 328-33.
DiPrete, Thomas A., and Claudia Buchmann. 2006. “Gender-Specific Trends in the Value of
Education and the Emerging Gender Gap in College Completion.” Demography, 43(1):
1–24.
Dröes, Martijn I., and Ryan C.R. van Lamoen. 2010. “Did Unilateral Divorce Laws Raise
Divorce Rates? A Reconciliation and New Results: Comment.” Discussion Paper No.
10-11, Tjalling C. Koopmans Research Institute, Utrecht School of Economics.

77

Dougherty, Christopher. 2005. “Why Are the Returns to Schooling Higher for Women than for
Men?” Journal of Human Resources, 40(4): 969–88.
DuMouchel, William H., and Greg J. Duncan. 1983. “Using Sample Survey Weights in
Multiple Regression Analyses of Stratified Samples.” Journal of the American Statistical
Association, 78(383): 535-43.
Eckstein, Zvi and Osnat Lifshitz. 2011. “Dynamic Female Labor Supply.” Econometrica, 79(6):
1675-1726
Fairlie, Robert W. 2005. “An Extension of the Blinder-Oaxaca Decomposition Technique to
Logit and Probit Models.” Journal of Economic and Social Measurement, 30: 305-16.
Fallick, Bruce, and Jonathan Pingle. 2009. “A Cohort-Based Model of Labor Force
Participation.” Finance and Economics Discussion Series, Federal Reserve Board,
Washington, D.C.
Friedberg, Leora. 1998. “Did Unilateral Divorce Raise Divorce Rates? Evidence from Panel
Data.” American Economic Review, 88(3): 608-27.
Goldin, Claudia, and Lawrence Katz. 2002. “The Power of the Pill: Oral Contraceptives and
Women’s Career and Marriage Decisions.” Journal of Political Economy, 110(4): 730–
70.
Goldin, Claudia, Lawrence Katz, and Ilyana Kuziemko. 2006. “The Homecoming of American
College Women: The Reversal of the College Gender Gap.” Journal of Economic
Perspectives, 20(4): 133–56.
Haveman, Robert, and Barbara Wolfe. 1995. “The Determinants of Children’s Attainments: A
Review of Methods and Findings.” Journal of Economic Literature, 33(4): 1829–78.
Heckman, James J., and Paul A. LaFontaine. 2010. “The American High School Graduation Rate:
Trends and Levels.” Review of Economics and Statistics, 92(2): 244–62.
Hoffman, Saul D. 2009. “The Changing Impact of Marriage and Children on Women's Labor
Force Participation.” Monthly Labor Review, 2009(2): 3-14.
Hubbard, William H. J. 2011. “The Phantom Gender Difference in the College Wage Premium.”
Journal of Human Resources, 46(3): 568–86.
Isen, Adam, and Betsey Stevenson. 2010. “Women's Education and Family Behavior: Trends in
Marriage, Divorce and Fertility.” Working Paper 15725, National Bureau of Economic
Research.
Jaeger, David A. 1997. “Reconciling the Old and New Census Bureau Education Questions:
Recommendations for Researchers.” Journal of Business & Economic Statistics, 15(3):

78

300-309.
Jepsen, Christopher. 2008. “Multinomial Probit Estimates of College Completion at 2-year and
4-year Schools.” Economics Letters, 98 (2): 155–60
Juhn, Chinhui, and Simon M. Potter. 2006. “Changes in Labor Force Participation in the
United States.” Journal of Economic Perspectives, 20(3): 27-46.
King, Miriam, Steven Ruggles, J. Trent Alexander, Sarah Flood, Katie Genadek, Matthew B.
Schroeder, Brandon Trampe, and Rebecca Vick. 2010. Integrated Public Use Microdata
Series, Current Population Survey: Version 3.0. [Machine-readable database].
Minneapolis, MN: Minnesota Population Center [producer and distributor for IPUMSCPS].
Lee, Jin Young. 2012. “The Plateau in U.S. Women's Labor Force Participation: A Cohort
Analysis.” Unpublished manuscript.
Lee, Jin Young. 2012. “The Upward Trend of Women’s College-going: The Role of Teenagers’
Anticipated Future Labor Force Attachment.” Unpublished manuscript.
Light, Audrey, and Wayne Strayer. 2002. “From Bakke to Hopwood: Does Race Affect College
Attendance and Completion?” Review of Economics and Statistics, 84(1): 34–44.
Macunovich, Diane, J. 2010. “Reversals in the Patterns of Women’s Labor Supply in the U.S.,
1976-2009.” Monthly Labor Review, 133(11):16-36.
Manski, Charles F., and Steven R. Lerman. 1977. “The Estimation of Choice Probabilities from
Choice Based Samples.” Econometrica, 45(8): 1977–88.
Nickell, Stephen. 1981. “Biases in Dynamic Models with Fixed Effects.” Econometrica, 49(6):
1417-26.
Percheski, Christine. 2008. “Opting Out? Cohort Differences in Professional Women's
Employment Rates from 1960 to 2005.” American Sociological Review 73(June): 497517.
Plug, Erik. 2004. “Estimating the Effect of Mother’s Schooling on Children’s Schooling Using a
Sample of Adoptees.” American Economic Review, 94(1): 358-68.
Polivka, Anne E., and Stephen M. Miller. 1998. “The CPS after the Redesign: Refocusing the
Economic Lens.” Labor Statistics Measurement Issues, John Haltiwanger, Marilyn E.
Manser, and Robert Topel (eds.), Chicago: University of Chicago Press, 249-286.
Ruggles, Steven, J. Trent Alexander, Katie Genadek, Ronald Goeken, Matthew B. Schroeder, and
Matthew Sobek. 2010. Integrated Public Use Microdata Series: Version 5.0 [Machinereadable database]. Minneapolis, MN: Minnesota Population Center [producer and
distributor for IPUMS-USA]

79

Solon, Gary. 1984. “Estimating Autocorrelations in Fixed-Effects Models.” Technical Working
Paper No. 32, National Bureau of Economic Research.
Stevenson, Betsey, and Justin Wolfers. 2007. “Marriage and Divorce: Changes and Their
Driving Forces.” Journal of Economic Perspectives, 21(2): 27-52.
U.S. Bureau of Labor Statistics. 2011. “Women at Work.” BLS Spotlight on Statistics.
Vogelsang, Timothy. 2008. “Heteroskedasticity, Autocorrelation, and Spatial Correlation Robust
Inference in Linear Panel Models with Fixed Effects.” Unpublished.
Wolfers, Justin. 2006. “Did Unilateral Divorce Laws Raise Divorce Rates? A Reconciliation
and New Results.” American Economic Review, 96(5): 1802-20.

80