INSTITUTE WORKING PAPER

No. 41

Are Marriage-Related Taxes and Social

Security Benefits Holding Back Female
Labor Supply?
October 2020

Margherita Borella
University of Torino and CeRP-Collegio
Carlo Alberto

Mariacristina De Nardi
University of Minnesota, Federal Reserve
Bank of Minneapolis, CEPR, and NBER

Fang Yang
Louisiana State University

DOI: https://doi.org/10.21034/iwp.41
JEL classification: E21, H2, J22, J31

The views expressed herein are those of the authors and not necessarily those of the Federal Reserve Bank of
Minneapolis or the Federal Reserve System.
 Are marriage-related taxes and Social Security
benefits holding back female labor supply?
Margherita Borella, Mariacristina De Nardi, and Fang Yang∗

September 21, 2020

Abstract
In the United States, both taxes and old age Social Security benefits depend
on one’s marital status and tend to discourage the labor supply of the secondary
earner. To what extent are these provisions holding back female labor supply?
We estimate a rich life cycle model of labor supply and savings for couples
and singles using the method of simulated moments (MSM) on the 1945 and
1955 birth-year cohorts and use it to evaluate what would happen without
these provisions. Our model matches well the life cycle profiles of labor market
participation, hours, and savings for married and single people and generates
plausible elasticities of labor supply. Eliminating marriage-related provisions
drastically increases the participation of married women over their entire life
cycle, reduces the participation of married men after age 60, and increases the
savings of couples in both cohorts, including the later one, which has similar
participation to that of more recent generations. If the resulting government
surplus were used to lower income taxation, there would be large welfare gains
for the vast majority of the population.

∗
Margherita Borella: University of Torino and CeRP-Collegio Carlo Alberto, Italy. Mariacristina
De Nardi: University of Minnesota, Federal Reserve Bank of Minneapolis, CEPR, and NBER.
Fang Yang: Louisiana State University. De Nardi gratefully acknowledges support from the ERC,
grant 614328 “Savings and Risks.” Yang gratefully acknowledges MRRC grant number 08098401,
pursuant to a grant from the U.S. Social Security Administration (SSA) funded as part of the
Retirement Research Consortium through the University of Michigan Retirement Research Center
Award RRC08098401. We thank Joe Altonji, Richard Blundell, Monica Costa Dias, Zvi Eckstein,
Joan Gieseke, Rasmus Lenz, Derek Neal, and Jon Skinner for useful comments and suggestions.
The views expressed herein are those of the authors and do not necessarily reflect the views of the
National Bureau of Economic Research, MRRC, the SSA, the CEPR, any agency of the federal
government, the Federal Reserve Bank of Minneapolis, or the Federal Reserve System.

1
1 Introduction
After increasing robustly from 1962 to the early 1990s, the labor force participa-
tion of women in the United States has been stagnating. Black, Schanzenbach, and
Breitwieser (2017) write, “The U.S. economy will not operate at its full potential un-
less government and employers remove impediments to full participation by women
in the labor market. The failure to address structural problems in labor markets—
including tax and employment policy—does more than hold back women’s careers
and aspirations for a better life. In fact, barriers to participation by women also act
as brakes on the national economy, stifling the economy’s ability to fully apply the
talents of 51 percent of the population.”
Barro and Redlick (2011) stress the importance of government policy and taxation
in determining the aggregate performance of an economy and suggest an important
role for the disincentives to female labor supply. More specifically, they study the ef-
fects of U.S. marginal tax rates over time and find that increases in average marginal
tax rates have significant negative effects on aggregate output, with the notable ex-
ception of the 1948 tax cut, which was followed by the 1949 recession. Notably, this
tax cut was in large part due to the introduction of joint filing for married couples
and implied that the marginal tax rate for the secondary earner increased drastically.
In this paper, we ask, to what extent does the dependence of taxes and old-age
Social Security benefits on marital status discourage female labor supply and affect
welfare? The mechanisms are the following. First, since couples file taxes jointly,
the secondary earner faces a higher marginal tax rate, which tends to discourage
their labor supply. Second, married and widowed people can claim Social Security
spousal and survivorship benefits under their spouses’ past contributions rather than
their own. Hence, their reduced labor supply does not necessarily imply lower Social
Security benefits. Since women have historically been the secondary earners, both
provisions tend to discourage female labor supply, but to what extent are these dis-
incentives holding it back? And to what extent do they affect cohorts with different
female participation?
To answer this question, we develop and estimate a rich life cycle model with
single and married people in which single people meet partners and married people
might get divorced. Every working-age person experiences wage shocks and every
retiree faces medical expenses and life span risk. People in couples face the risks of

2
both partners. Households can self-insure by saving and by choosing whether and how
much to work (for both partners if in a couple). Consistent with the data, we allow
for human capital to affect wages. We explicitly model Social Security with survival
and spousal benefits, the differential tax treatment of married and single people, the
progressivity of the tax system (including the earned income tax credit or EITC), and
old-age means-tested transfer programs such as Medicaid and Supplemental Security
Income (SSI). We also model the changes in the tax and Social Security systems that
our two cohorts face over time.
We estimate our dynamic structural model using the method of simulated mo-
ments (MSM) and data from the Panel Study of Income Dynamics (PSID) and from
the Health and Retirement Study (HRS) for the cohort born in the period 1941-1945
(referred to here as the “1945” cohort). That cohort has by now completed a large
part of its life cycle and is covered by these two data sets, which provide excellent
information over their working and retirement periods, respectively. Then, taking
the estimated preference parameters from that cohort as given, we also estimate our
model for the 1951-1955 cohort (referred to here as the “1955” cohort), which had
much higher participation of married women (and closer to that of more recent co-
horts) and for which policy and welfare implications might thus be very different.
Our estimated model matches the life cycle profiles of labor market participa-
tion, hours worked by the workers, and savings for married and single people for
both cohorts very well. It also generates elasticities of labor supply by age, gender,
and marital status that are consistent with those previously estimated by others.
The latter provides an additional test of the reliability of our model and its policy
implications.
For the 1945 cohort, we find that Social Security spousal and survivor benefits
and the current structure of joint income taxation provide strong disincentives to
work to married women and single women who expect to get married, and strong
incentives to work for married men after age 60. For instance, the elimination of
all of these marriage-based rules raises participation at age 25 by over 20 percentage
points for married women and by 5 percentage points for single women. At age 45,
participation for these groups is, respectively, still 15 and 3 percentage points higher
without these marital benefits provisions. In addition, the elimination of marriage-
based rules decreases the participation of married men starting at age 60, resulting
in a participation rate that is 8 percentage points lower by age 65. Finally, for this

3
cohort, the elimination of marital provisions increases the savings of married couples
by 20.3% at age 66.1 In terms of welfare, abolishing these marital provisions would
benefit most couples, all single men, and over one-third of single women and, thus,
over 90% of the people in this cohort.
Given that the labor supply of married women has been increasing rapidly over
time for cohorts born before the 1970s, a natural question that arises is whether the
effects of these marital provisions are also large for more modern cohorts in which
married women are more likely to work. To shed light on this question, we study a
cohort that is 10 years younger than our reference cohort (that is, the 1955 cohort),
for which we still have a completed labor market history and whose labor market
behavior is close to that of more recent cohorts. By way of comparison, the labor
market participation of married women at age 25 is just over 50% for our 1945 cohort,
whereas it is over 60% for our 1955 cohort.
To estimate our model for the 1955 cohort, we assume that their preference pa-
rameters are the same as the ones we estimate for the 1945 cohort, but we give the
1955 cohort their observed marriage and divorce probabilities, number of children,
initial conditions for wages and experience, and returns to working. We then esti-
mate the child care costs, available time, and participation costs that reconcile their
labor supply and saving behavior with the observed data. Finally, we run the policy
experiment of eliminating the marriage-related provisions for both taxes and Social
Security. We find that the effects for the 1955 cohort on participation, wages, earn-
ings, and savings are large and similar to those in the 1945 cohort, thus indicating
that the effects of marriage-related provisions are also large for cohorts in which the
labor participation of married women is higher. We also find that abolishing these
marriage-related provisions for this cohort at age 25 would also benefit most couples,
all single men, and over two-thirds of single women. In addition, the welfare benefits
to those gaining would be much higher, and the welfare costs of those losing would
be very small, because the human capital of women in the cohort is already higher
than that in the previous cohort at age 25.
Our paper provides several contributions. First, it is the first estimated structural
model of couples and singles that allows for participation and hours decisions of both
1
While our model takes marriage and divorce behavior from the observed data from each cohort,
we show in Section 8 that the empirical evidence finds small effects of these provisions on marriage
and divorce and that our results are robust to large changes in marriage and divorce behavior.

4
men and women, including those in couples, in a framework with savings. Our results
show that, in addition to lowering the participation of women, these marriage-related
policies significantly reduce the savings of couples and increase the participation of
married men later in life, and decrease welfare for the vast majority of the population.
Second, it is the first paper to study all marriage-related taxes and benefits in a
unified framework. Third, it does so by allowing for the large observed changes
in the labor supply of married women over time by studying two different cohorts.
Fourth, our framework is very rich along dimensions that are important in the study
of our problem. For instance, allowing for labor market experience to affect wages
(of both men and women) is important in that it captures the endogeneity of wages
and their response to policy and marital status changes. Carefully modeling survival,
health, and medical expenses in old age, and their heterogeneity by marital status
and gender, is crucial to evaluate the effects on labor supply and savings of Social
Security payments during old age and their interaction with taxation and old-age
means-tested benefits such as Medicaid and SSI, which we also model. By modeling
one-year periods, it gives people the ability to change their labor supply and savings
in a more flexible and realistic way. Finally, our model fits the data for participation,
hours worked, and savings, the estimated labor supply elasticities over the life cycle
for single and married men and women, and thus provides a valid benchmark to
evaluate the effects of the current marriage-related policies.

1.1 Related literature

We build on the literature on female labor supply over the life cycle. Within
this literature, Attanasio, Low, and Sánchez-Marcos (2008) and Eckstein and Lifshitz
(2011) point to the importance of changing wages and child care costs in explaining
increases in female labor supply over time. Eckstein, Keane, and Lifshitz (2019)
examine the changes over time in the selection and determinants of married women
working. Hubener, Maurer, and Mitchell (2016) study the effects of exogenous family
dynamics and endogenous labor supply on portfolio choice and retirement.
The structural papers in this branch of the literature typically assume that male
labor supply is exogenously fixed and/or that the choice of hours of both partners
is limited to full-time or full-time and part-time, and/or abstract from savings. We
also add to this literature by quantifying the disincentive effects of the U.S. Social

5
Security and tax code on the labor supply of women.
We contribute to the small body of literature studying policy reforms in environ-
ments that include couples. Guner, Kaygusuz, and Ventura (2012a) study the switch
to a proportional income tax and a reform in which married individuals can file taxes
separately and find that these reforms substantially increase female labor participa-
tion. Nishiyama (2017), Kaygusuz (2015), and Groneck and Wallenius (2017) find
that removing spousal and Social Security survivor benefits would increase female
labor participation, female hours worked, and aggregate output. Bick and Fuchs-
Schundeln (2018) focus on a simpler static model of married couples and find that
income taxes are an important factor driving differences in the labor supply of married
women across countries.
More generally, our paper differs from the previous literature in focus, method-
ology, and important model elements. In terms of focus, previous papers have only
studied the effects of removing marriage-related rules that pertain to either Social
Security or taxes and thus cannot answer the question as to what extent these pro-
visions jointly hold back female labor supply, which is the focus of our paper. In
terms of methodology, we not only estimate our model but also make sure that our
model’s inputs and outputs are consistent with the PSID and HRS data for the work-
ing and retirement periods, respectively. As a result, for instance, we estimate the
accumulation of human capital on the job from the data and allow the tax structure
to vary over time for each cohort (and estimate our tax functions from the PSID as
a function of cohort, year, and marital status). Thus, we take this variation into
account when we estimate our model. In terms of important model elements, none of
the previous papers models health shocks and medical expenses in retirement, which
are important to understand savings and the role of Social Security in insuring both
mortality and medical expense risks, nor do they have flexible labor supply of both
men and women, including in hours worked, over all of the working period. As we
show, the labor supply of men also changes as a result of the reforms, thus allowing
that of women to adjust differently than it would have if the labor supply and hours
of men had been fixed.

6
2 Background on marriage and U.S. taxes and old-
age Social Security benefits
Many countries tax the income of married people by making them file as if they
were single (individual taxation). As a result, when the secondary earners in couples
work, their marginal tax rate is based on their own income rather than on the sum
of their partner’s income and their own.
The United States, instead, taxes the income of married couples jointly (joint
taxation) and uses a different tax schedule for married and single people. The combi-
nation of joint taxation and a progressive tax system typically implies that a married
secondary earner faces a higher marginal tax rate than a single earner.
The question of when and why we ended up with such a system in the U.S. is an
interesting one. Our reading of the literature is that joint taxation was implemented
in 1948 with the goal of eliminating differences between community-property and
common-law property states. In community property states, all income received by
a married couple is considered jointly earned and owned. Thus, their residents felt
legally entitled to pay taxes, including at the federal level, on the average income
of each spouse. This was not possible for residents in common-law property states,
where the couples with a main earner (most of them at the time) thus ended up facing
a much higher average marginal tax rate. The 1948 reform was meant to eliminate
this source of inequality and essentially imposed that all couples have to file jointly
(under a different tax bracket system).
To illustrate the secondary earner’s disincentive to work, we use the effective tax
rates that we estimate from the PSID in 1988, a time period during which the earned
income tax credit (EITC) program is already active and people in our 1945 cohort
are still of working age (in 1988 the median woman in our 1945 cohort is 45 years
old). The details of our tax computations are at the end of Appendix B.
The left panel of Figure 1 illustrates the incentives to work for these single and
married women by plotting four marginal tax rates as a function of women’s earnings:
the marginal tax rates of single women and those of married women with husbands
at three different percentiles of earnings. A single woman earning $500 a year faces
a marginal tax rate of -10%, while a married woman earning the same amount faces
a marginal tax rate of 14%, 18%, and 21%, respectively, if she is married to a man
in the 25th, 50th, and 75th income percentiles (which correspond to, respectively,

7
 Women's marginal tax rates Empirical cumulative distribution
0.3 1

0.9
0.25

0.8
0.2
0.7

Cumulative probability
Marginal tax rate 0.15
0.6

0.1 0.5

0.4
0.05

0.3
0
Married to 75th earner 0.2
Married to 50th earner
-0.05
Married to 25th earner 0.1
Single Women

-0.1 0
0 1 2 3 4 5 6 7 8 9 10 -0.2 -0.1 0 0.1 0.2 0.3
Women's income 10 4 Non-working wives' marginal tax rate

Figure 1: Left panel: 45-year-old women’s marginal tax rate when single (starred red
line) or married to men at the 25th (dashed orange line), 50th (dotted orange
line), and 75th (circled orange line) income percentiles, as a function of women’s
earnings in 2016 dollars (minimum value $500). Right panel: cumulative density
function (cdf) of 45-year-old non-working wives’ marginal tax rates.

$43,090, $68,995, and $113,288 in 2016 dollars). Our estimated negative tax rate at
low income levels illustrates the impact of the EITC.
While this graph tells us that married women typically face a higher marginal tax
rate than single women, it does not tell us the distribution of marginal tax rates for
married women who are not working. Thus, the right panel of Figure 1 displays the
distribution of marginal tax rates for 45-year-old men whose wives are not working;
this marginal tax rate is also that of their wives, should they start working. Com-
paring the marginal tax rate of non-working wives with that of non-working single
women reveals that single women who are starting to work face a -10% marginal tax
rate, while 80% of married women face a marginal tax rate of 10% or higher, because
of their husband’s earnings and joint taxation. These graphs thus suggest that mak-
ing married people file as single rather than jointly could have large incentives for the
labor market participation of married women.
Social Security for a single person is a function of one’s average lifetime earnings.
Social Security for a married person is the higher between one’s own benefit entitle-
ment and half of the spouse’s entitlement while the other spouse is alive (spousal ben-
efit) and the higher between one’s own benefit entitlement and the deceased spouse’s
after the spouse’s death (survival benefit).
We use data from the PSID for 66-year-old couples in our 1945 cohort and Social
Security rules to generate Figure 2, which illustrates the magnitude of Social Security
spousal benefits. The left panel of Figure 2 plots household Social Security benefits

8
 Social Security Benefit Survivor Benefit
50000 25000
W/ m. benefit
W/O m. benefit

45000
20000

Average Wife's Survivor Benefit

Average Household Benefit 40000
15000

35000

10000
30000

W/ m. benefit
5000 W/O m. benefit
25000

20000 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Wife's own benefit decile Wife's own benefit decile

Figure 2: Left panel: Average household Social Security benefits at age 66 by wife’s own
Social Security benefit decile, with (circled blue line) or without (crossed red
line) marital benefit, in 2016 dollars. Right panel: Average survivor benefit
by wife’s own Social Security benefit decile, with (circled blue line) or without
(crossed red line) marital benefits, 2016 dollars.

while the husband is alive. It takes married women at retirement age and, based
on the deciles of their own Social Security entitlement, plots their average household
yearly Social Security benefits with (circled line) and without (crossed line) marital
benefits. For instance, the number 1 on the x -axis represents 66-year-old married
women in our 1945 cohort that are in the lowest decile of their own Social Security
contributions. At that decile, household Social Security benefits for those women and
their husbands are $32,000 under marital benefits and about $22,000 without marital
benefits. The comparison of the two lines in this picture reveals that about 50% of
married households benefit from Social Security marital benefits while their husband
is alive and that these benefits can be very large.
The right panel of Figure 2 takes the same married women and plots what their
yearly Social Security benefits would be after their husband’s death with and without
survivor’s benefits. For instance, once a widow, a 66-year old married woman at
the lowest 10% of Social Security contributions would receive less than $500 a month
based on her own contributions only, whereas she would receive $22,000 thanks to her
husband’s contributions and survivorship benefits. The picture shows that because
most women have lower potential wages than men’s, participate less, and work fewer
hours, survivorship benefits are large for over 80% of married women in this cohort.
This last set of graphs highlights that Social Security marital benefits are large and
can also reduce married women’s incentives to work.

9
3 Life cycle patterns for single and married men
and women in our cohorts
We pick the 1945 cohort because their entire adult life is first covered by the PSID,
which starts in 1968 and has rich information for the working period, and then by
the HRS, which starts covering people at age 50 in 1994 and has rich information for
the retirement period, including on medical expenses and mortality. Thus, we have
excellent data for this cohort over their entire life cycle. We pick our 1955 cohort to
be as young as possible to maximize changes in their participation, conditional on
having an almost complete working period for the same cohort.2
Figure 3 displays participation and average annual hours worked by workers. The
top panels refer to the 1945 cohort.3 The top left panel shows that married men have
the highest participation rate and only slowly decrease their participation starting
from age 45, whereas single men decrease their participation much faster. The par-
ticipation of single women starts about 10 percentage points lower than that of single
men and gradually increases until age 50. Married women have the lowest participa-
tion rate. It starts around 50% at age 25, increases to 78% between ages 40 and 50,
and gradually declines at a rate similar to that of the other three groups. The top
right panel highlights that married men on average work more hours than everyone
else. Women not only have a participation rate lower than men on average but also
display lower average hours, even conditional on participation.
The middle panels display the analogous information for the 1955 cohort. Com-
paring the top and bottom panels shows a large increase in participation by married
women across these two cohorts. Conditional on working, average annual hours have
also increased for married women. Finally, annual hours worked by married men con-
ditional on working are lower, which underscores the importance of modeling men’s
labor supply, in addition to that of women’s.
Because the availability of asset data in the PSID is limited (available only every
five years until 1999 and every other year afterward) and our 1955 cohort has not yet
2
Appendix A provides details about our computations and also shows that the majority of men
and women are married in both cohorts and that the fraction of married people goes down only
slightly across these two cohorts. Appendix I validates our labor market outcomes from the PSID
with those from the Current Population Survey (CPS) for both cohorts and shows that they are
very similar.
3
These profiles are obtained from the data by fitting a fourth-order polynomial in age fully
interacted with marital status and cohort dummies, separately for each gender.

10
 Labor Participation Average Working Hours (Workers)
1 2400

0.9
2200
0.8

0.7
2000

0.6

1800
0.5

0.4 Single men Single men

Single women 1600 Single women
0.3 Married men Married men
Married women Married women

0.2 1400
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age
Labor Participation Average Working Hours (Workers)
1 2400

0.9
2200
0.8

0.7
2000

0.6

1800
0.5

0.4 Single men Single men

Single women 1600 Single women
0.3 Married men Married men
Married women Married women

0.2 1400
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

105
6
Single men
Single women
5 Couples
Average Household Asset

0
30 40 50 60 70 80
Age

Figure 3: Life cycle profiles by gender and marital status for the 1945 cohort (top two
graphs), 1955 cohort (middle two graphs), and both cohorts (bottom graph),
PSID data

retired, we use the same asset profiles for both cohorts. The bottom panel in Figure 3
shows that average assets increase until age 70 for all groups, with single women
accumulating the lowest amount and showing no sign of a slowdown in accumulation
before age 75.

4 The model
Our model period is one year long and there are three stages in one’s life: a working
stage (ages 25 to 61), an early retirement stage (ages 62 to 65), and a retirement stage
(age 66 to the maximum age of 99).
During the working stage, single and married people choose how much to work

11
and save and face wage shocks. Married people face divorce shocks, and single people
might meet partners and get married.4
Wages are a function of one’s human capital (which is endogenously accumulated
while working) and are affected by shocks. We measure human capital at a point in
time as a person’s average accumulated earnings at that point in time. Thus, human
capital is a function of one’s past wages and labor supply (and of one’s education, to
the extent that education influences one’s wages).
We model (and estimate) available time to be split between working and leisure
and we allow it to depend on one’s gender and marital status. We interpret it as
net of home production, child care, and elderly care that one has to perform whether
working or not (and that is not easy to outsource). All workers have to pay a fixed
cost of working, which depends on their age, gender, and marital status. It represents
the cost of commuting, getting ready for work, making arrangements for being able
to go to work, and so on.
Single women and married people have children, and the number of their children
depends on maternal age and marital status. We allow for both time costs and
monetary costs of raising children. The time costs affect one’s available time for
working and enjoying leisure. The monetary costs enter our model in two ways.
First, they affect consumption through an adult-equivalent scale family size. Second,
working mothers have to pay child care costs that depend on the age and number of
their children, and on their own earnings. We thus assume that child care costs are
a normal good: women with higher earnings pay for more expensive child care.5
During the early retirement stage, people still experience wage shocks but
single people don’t get married anymore and couples no longer divorce.6 If they
claim Social Security, they can no longer work. Couples claim Social Security at the
same time.
During the first year of the retirement stage, those who have not already claimed
4
For tractability, we assume that people survive to retirement for sure. Although the death
of a spouse is a big shock for the households experiencing it (Fadlon and Nielsen, 2015), it is a
low-probability event in the data.
5
Introducing home production and child care choices is infeasible given the complexity of our
framework. The main caveat with our assumptions is that we do not allow these choices to vary
when policy changes.
6
In the HRS data, we observe our 1941-1945 birth cohort between ages 62 and 72. Over that
period, only 1% of couples get divorced and 4% of singles get married. Thus, the implied yearly
probability of marriage and divorce is very small.

12
Social Security do so and stop working. People face out-of-pocket medical expenses
and the risk of death. Thus, each married person faces the risk of his or her spouse
dying, in addition to their own. Mortality risk and medical expenses depend on
gender, age, health status, and marital status.
Given that we explicitly model labor participation and hours of husbands and
wives, savings, and medical expenses in old age, our model is computationally very
intensive (See Appendix D for more details.) For tractability, we make the following
additional assumptions. First, people who are married to each other are the same
age. Second, fertility is exogenous, and women have an age-varying number of children
that depends on their age and marital status and that we estimate from the data.
Lastly, we assume that marriage and divorce are exogenous processes that we also
estimate from the data. Thus, our results should be interpreted as holding marriage
and divorce patterns fixed at those historically observed for this cohort. We discuss
the empirical literature on the responses of changes in marriage and divorce rates
to policy changes and evaluate the robustness of our findings to this assumption in
Section 8.

4.1 Preferences
Let t be age ∈ {t0 , t1 , ..., ..., td }, with t0 = 25 and td = 99 being the maximum
possible life span. For simplicity of notation, think of the model as being written
for one cohort, so age t also indexes the passing of time for that cohort. We solve
the model for the two cohorts separately and make sure that each cohort has the
appropriate time- and age-varying inputs.
Households have time-separable preferences and discount the future at rate β.
The superscript i denotes gender, with i = 1, 2 being a man or a woman, respectively.
The superscript j denotes marital status, with j = 1, 2 being single or in a couple,
respectively.
Each single person has preferences over consumption and leisure, and the period
flow of utility is given by the standard CRRA utility function

((ct /ηti,1 )ω lt1−ω )1−γ − 1

v i (ct , lt , ηti,1 ) = ,
1−γ

where ct is consumption and ηti,j is the equivalent scale in consumption (which is a

13
function of family size, including children) and ηti,1 corresponds to that for singles.
The term lti,j is leisure, which is given by

lti,j = Li,j − nit − Φi,j

t Init , (1)

where Li,j is available time endowment, which can be different for single and married
men and women and should be interpreted as available time net of home production.
It is a convenient way to represent activities that require time and cannot easily be
outsourced. Leisure equals available time endowment less nit , hours worked on the
labor market and the fixed time cost of working. That is, the term Init is an indicator
function that equals 1 when hours worked are positive and zero otherwise, while the
term Φi,j
t represents the fixed time cost of working.
The fixed cost of working should be interpreted as including commuting time, time
spent getting ready for work, and so on. We allow it to depend on gender, marital
status, and age because working at different ages might imply different time costs for
married and single men and women. We assume the following functional form, whose
three parameters we estimate using our structural model:

exp(φi,j i,j
0 + φ1 t + φ2 t )
i,j 2
Φi,j
t = i,j 2 .
1 + exp(φi,j
0 + φ i,j
1 t + φ 2 t )

We assume that couples maximize their joint utility function

((ct /ηti,j )ω (lt1 )1−ω )1−γ − 1 ((ct /ηti,j )ω (lt2 )1−ω )1−γ − 1
w(ct , lt1 , lt2 , ηti,j ) = + .
1−γ 1−γ

Note that for couples, ηti,j does not depend on gender and that j = 2.

4.2 Environment
People can hold assets at at a rate of return r. The timing is as follows.
At the beginning of each working period, each single person observes his/her cur-
rent idiosyncratic wage shock, age, assets, and accumulated earnings. Each married
person also observes their partner’s wage shock and accumulated earnings. At the
beginning of each early retirement period, single people may claim Social Security
benefits and couples may claim retirement benefits jointly.

14
 At the beginning of each retirement period, each single person observes his/her
current age, assets, health, and accumulated earnings. Each married person also
observes their partner’s health and accumulated earnings.
Decisions are made after everything has been observed and new shocks hit at the
end of the period after decisions have been made.

4.2.1 Human capital and wages

We define human capital, ȳti , as one’s average past earnings at each age. Thus,
our definition of human capital implies that it is a function of one’s initial wages and
schooling and subsequent labor market experience and wages.7
Wages have two components. The first is a deterministic function of age, gender,
and human capital: eit (ȳti ). The second component is a persistent earnings shock it
that evolves as follows:

ln it+1 = ρiε ln it + υti , υti ∼ N (0, (συi )2 ).

The product of eit (·) and it determines an agent’s units of effective wage per hour
worked during a period.

4.2.2 Marriage and divorce

During the working period, a single person gets married with an exogenous prob-
ability that depends on his/her age, gender, and wage shock. The probability of
getting married at the beginning of next period is

νt+1 (·) = νt+1 (i, it ).

Conditional on meeting a partner, the probability of meeting with a partner p

with wage shock pt+1 is
ξt+1 (·) = ξt+1 (pt+1 |it+1 , i). (2)

Allowing this probability to depend on the wage shock of both partners generates
assortative mating. We assume random matching over assets at+1 and average accu-
p
mulated earnings of the partner ȳt+1 , conditional on the partner’s wage shock. Thus,
7
It also has the important benefit of allowing us to have only one state variable keeping track of
human capital and Social Security contributions.

15
we have
θt+1 (·) = θt+1 (apt+1 , ȳt+1
p
|pt+1 ). (3)

A working-age couple can be hit by a divorce shock at the end of the period that
depends on age and the wage shock of both partners,

ζt+1 (·) = ζt+1 (1t , 2t ).

If the couple divorces, they split the assets equally, and each of the ex-spouses becomes
single and moves on with half of the assets, their own wage shock, and own Social
Security contributions. Since we do not distinguish the previously divorced from the
singles, these two groups have the same number of children. We also abstract from
alimony in the case of divorce.

4.2.3 Costs of raising children and running a household

We keep track of the total number of children and children’s age as a function of
mothers’ age and marital status. The total number of children by one’s age affects
the economies of scale of single women and couples.
The number of children between ages 0 to 5 and 6 to 11 determines the child care
costs of working mothers (i = 2). The term τc0,5 is the child care cost for each child
ages 0 to 5, where that number of children is f 0,5 (i, j, t), while τc6,11 is the child care
cost for each child ages 6 to 11, which are f 6,11 (i, j, t). We use our structural model
to estimate these costs.

4.2.4 Medical expenses and death

At age 66, we endow people with a distribution of health that depends on their
marital status and gender. After that, they face survival, medical expenses, and health
shocks. Survival si,j
t depends on one’s age, gender, and marital status. Health status
ψti can be either good or bad and evolves according to a Markov process πti,j (ψti ) that
depends on age, gender, and marital status. Medical expenses mi,j i
t (ψt ) are a function
of age, gender, marital status, and health status.

16
4.2.5 Initial conditions

We take the fraction of single and married people at age 25 and their distribution
over the relevant state variables from the PSID (that is, assets, human capital, and
wage shocks, with the latter two being for each of the spouses in the case of a married
couple) for each of our two cohorts. We define notation for all of our state variables
in Section 4.4.

4.3 Government
Each cohort in our model faces the effective time-varying tax rates that it expe-
rienced in the data. As in Benabou (2002) and Heathcote et al. (2014), we adopt a
functional form that allows for negative tax rates (and thus incorporates the EITC),
and we allow our effective tax rates to depend on marital status, gender, and age for
each cohort (and thus time). Taxes paid are thus given by

i,j
T (Y, i, j, t) = (1 − λi,j
t Y
−τt
)Y. (4)

We estimate these functions using the PSID.

The government also uses a proportional payroll tax τtSS on labor income, up to a
Social Security cap yet , to help finance old-age Social Security benefits. We also allow
the payroll tax and the Social Security cap to change over time for each cohort as in
the data. We thus assume that the tax changes were anticipated by the households.
We use human capital ȳti (computed as an individual’s average earnings at age t,
up to the cap yet ) to determine both wages and old-age Social Security payments.
The insurance provided by Medicaid and SSI in old age is represented by a means-
tested consumption floor, c(j), as in Hubbard, Skinner, and Zeldes (1995).8

4.4 Recursive formulation

We define and compute nine sets of value functions: the value function of working-
age singles, the value function of singles during the early retirement stage, the value
function of retired singles, the value function of working-age couples, the value func-
tion of couples during the early retirement stage, the value function of retired couples,
8
Borella, De Nardi, and French (2018) discuss Medicaid rules and observed outcomes after re-
tirement.

17
the value function of an individual who is of working age and in a couple, the value
function of an individual who is at early retirement stage and in a couple, and the
value function of an individual who is retired and in a couple.

4.4.1 The value function of working-age singles

The state variables for a single person during one’s working period are age t,
gender i, assets ait , the persistent earnings shock it , and human capital ȳti . The
corresponding value function is

W s (t, i, ait , it , ȳti ) = max v i (ct , lti,j ) + β(1 − νt+1 (i))Et W s (t + 1, i, ait+1 , it+1 , ȳt+1
i
)+
ct ,at+1 ,nit
!
h i
c i p i p i p
βνt+1 (i)Et Ŵ (t + 1, i, at+1 + at+1 , t+1 , t+1 , ȳt+1 , ȳt+1 ) ,

(5)

subject to equation (1) and

Yti = ei,j i i i
t (ȳt )t nt , (6)

τc (i, j, t) = τc0,5 f 0,5 (i, j, t) + τc6,11 f 6,11 (i, j, t), (7)

T (·) = T (rat + Yt , i, j, t), (8)

ct + at+1 = (1 + r)ait + Yti (1 − τc (i, j, t)) − τtSS min(Yti , yet ) − T (·), (9)
i
ȳt+1 = (ȳti (t − t0 ) + (min(Yti , yet )))/(t + 1 − t0 ), (10)

at+1 ≥ 0, (11)

nit ≥ 0. (12)

The expectation of the value function next period if one remains single integrates
over one’s wage shock next period. When one gets married, we not only take a similar
expectation but also integrate over the distribution of the state variables of one’s
partner: (ξt+1 (pt+1 |it+1 , i) is the distribution of the partner’s wage shock defined in
equation (2), and θt+1 (·) is the distribution of the partner’s assets and human capital
defined in equation (3)).
The value function Ŵ c is the discounted present value of the utility for the same
individual, once he or she is in a married relationship with someone with given state

18
variables, not the value function of the married couple, which counts the utility of
both individuals in the relationship. We discuss the computation of the value function
of an individual in a marriage later in this section.
Equation (10) describes the evolution of human capital, which we measure as
average accumulated earnings (up to the Social Security earnings cap yet ) and which
we use as a determinant of future wages and Social Security payments after retirement.

4.4.2 The value function of singles during the early retirement stage

Let tr denote the age at which someone first claimed Social Security. The recursive
problem for an individual that has claimed Social Security at age tr can be written
as !
S s (t, i, ait , ȳri , tr) = max v i (ct , Li,j ) + βEt S s (t + 1, i, ait+1 , ȳri , tr) , (13)
ct ,at+1

subject to equations (8), (11), and

Yt = SS(ȳri , tr) (14)

ct + at+1 = (1 + r)at + Yt − T (·). (15)

The term SS(y¯r i , tr) is a function of the income that the single person earned during
his or her working life, ȳri , and claiming age tr.
Let N s (t, i, ait , it , ȳti ) denote the value function of a person during the early retire-
ment period who has not yet claimed benefits,
!
N s (t, i, ait , it , ȳti ) = max v i (ct , lti,j ) + βEt V s (t + 1, i, ait+1 , it+1 , ȳt+1
i
) , (16)
ct ,at+1 ,nit

subject to equations (1), (6), (8), (10), (11), (12), and

ct + at+1 = (1 + r)ait + Yti − τtSS min(Yt , yet ) − T (·). (17)

Let V s (t, i, ait , it , ȳti ) denote the value function for a person during the early re-
tirement stage who has not yet retired. At the beginning of each period, that person
chooses whether to claim Social Security benefits, and Dti is an indicator function for
that decision which maximizes

19
 !
V s (t, i, ait , it , ȳti ) = max
i
(1 − Dti )N s (t, i, ait , it , ȳti ) + Dti S s (t, i, ait , ȳti , t) . (18)
Dt

4.4.3 The value function of retired singles

The state variables for a retired single are age t, gender i, assets ait , health ψti ,
average realized lifetime earnings ȳri , and Social Security claiming age tr. His or her
recursive problem can be written as
!
Rs (t, i, at , ψti , ȳri , tr) = max v i (ct , Li,j ) + βsi,j i s i i
t (ψt )Et R (t + 1, i, at+1 , ψt+1 , ȳr , tr) ,
ct ,at+1

(19)
subject to equations (8), (11), (14), and
n o
B(at , Yt , ψti , c(j)) = max 0, c(j) − [(1 + r)at + Yt − mi,j
t (ψt
i
) − T (·)] (20)

ct + at+1 = (1 + r)at + Yt + B(at , Yt , ψti , c(j)) − mi,j i

t (ψt ) − T (·) (21)

at+1 = 0, if B(·) > 0. (22)

The term si,1 i

t (ψt ) is the survival probability as a function of age, gender, marital
status, and health status. The function B(at , Yti , ψti , c(j)) represents old-age means-
tested government transfers (such as Medicaid and SSI) that ensure a minimum con-
sumption floor c(j).

4.4.4 The value function of couples during the working period

The state variables for a married couple during the working stage are (t, at , 1t , 2t , ȳt1 , ȳt2 ),
where 1 and 2 refer to gender, and the recursive problem for the married couple (j = 2)
can be written as

W c (t, at , 1t , 2t , ȳt1 , ȳt2 ) = max1 w(ct , lt1,j , lt2,j )

ct ,at+1 ,nt ,n2t

+ (1 − ζt+1 (·))βEt W c (t + 1, at+1 , 1t+1 , 2t+1 , ȳt+1

1 2
, ȳt+1 ) (23)
2 
!
X 
+ ζt+1 (·)β Et W s (t + 1, i, at+1 /2, it+1 , ȳt+1
i
) ,
i=1

20
subject to equations (1), (6), (7), (10), and

T (·) = T (rat + Yt1 + Yt2 , i, j, t) (24)

ct +at+1 = (1+r)at +Yt1 +Yt2 (1−τc (2, 2, t))−τtSS (min(Yt1 , yet )+min(Yt2 , yet ))−T (·) (25)

at ≥ 0, n1t , n2t ≥ 0. (26)

The expected value of the couple’s value function is taken with respect to the condi-
tional probabilities of t+1 given the current value of the t for each of the spouses (we
assume independent draws). The term ζt+1 (·) represents the probability of divorce
that we defined in Section 4.2.2. The expected values for the newly divorced people
are taken using the appropriate conditional distribution for their own wage shocks.

4.4.5 The value function of couples during the early retirement period

For tractability, we assume that during the early retirement stage, couples can no
longer divorce. The recursive problem for couples that have claimed Social Security
at age tr can be written as
!
S c (t, at , ȳr1 , ȳr2 , tr) = max w(ct , L1,j , L2,j ) + βEt S c (t + 1, at+1 , ȳr1 , ȳr2 , tr) , (27)
ct ,at+1

subject to equations (8), (15), (11), and

n
1 2 3 1 2
o
Yt = max (SS(ȳr , tr) + SS(ȳr , tr), max(SS(ȳr , tr), SS(ȳr , tr)) (28)
2

In equation (28), the variable Yt represents the spousal benefit from Social Security,
which gives a married person the right to collect the higher amount between one’s
own benefit and half of their spouse’s benefit.
Let N c (t, at , 1t , 2t , ȳt1 , ȳt2 ) denote the value function of a couple that has not yet
claimed benefits,

N c (t, at , 1t , 2t , ȳt1 , ȳt2 ) = max1 w(ct , lt1,j , lt2,j )

ct ,at+1 ,nt ,n2t
! (29)
c
+ βEt V (t + 1
1, at+1 , 1t+1 , 2t+1 , ȳt+1 2
, ȳt+1 ) ,

21
subject to equations (1), (6), (10), (24), (26), and

ct + at+1 = (1 + r)at + Yt1 + Yt2 − τtSS (min(Yt1 , yet ) + min(Yt2 , yet )) − T (·). (30)

Let V c (t, at , 1t , 2t , ȳt1 , ȳt2 ) denote the value function for a married couple during
the early retirement stage that has not yet claimed Social Security benefits. At the
beginning of each period, a couple chooses whether to claim Social Security benefits,
that is Dt = 1. The early claiming decision maximizes
 
V c (t, at , 1t , 2t , ȳt1 , ȳt2 ) = max (1 − Dt )N c (t, at , 1t , 2t , ȳt1 , ȳt2 ) + Dt S c (t, at , ȳt1 , ȳt2 , t) .
Dt
(31)

4.4.6 The value function of couples during retirement

During retirement, each of the spouses faces health shocks ψti and survival shocks
si,2 i
t (ψt ). We assume that the health shocks of each spouse are independent of each
other and that the death shocks of each spouse are as well. During each period, the
married couple’s recursive problem (j = 2) can be written as

Rc (t, at , ψt1 , ψt2 , ȳr1 , ȳr2 , tr) = max w(ct , L1,j , L2,j )+
ct ,at+1

βs1,j 1 2,j 2 c 1 2 1 2
t (ψt )st (ψt )Et R (t + 1, at+1 , ψt+1 , ψt+1 , ȳr , ȳr , tr)+

βs1,j 1 2,j 2 s 1 ¯1
t (ψt )(1 − st (ψt ))Et R (t + 1, 1, at+1 , ψt+1 , ȳr , tr)+
!
βs2,j 2 1,j 1 s 2 ¯2
t (ψt )(1 − st (ψt ))Et R (t + 1, 2, at+1 , ψt+1 , ȳr , tr) ,

(32)

subject to equations (8), (11), (22), (28), and

ȳ¯ri = max(ȳr1 , ȳr2 ), (33)

n o
B(at , Yt , ψt1 , ψt2 , c(j)) = max 0, c(j)− (1 + r)at + Yt − m1,j 1 2,j 2

t (ψt ) − m t (ψt ) − T (·)
(34)
1 2 1,j 1 2,j 2
ct + at+1 = (1 + r)at + Yt + B(at , Yt , ψt , ψt , c(j)) − mt (ψt ) − mt (ψt ) − T (·) (35)

In equation (33), the variables ȳ¯ri , i = 1, 2 represent that the survivor collects

22
benefits based on the higher amount between their own contributions and those of
their deceased spouse.

4.4.7 The value functions of individuals in couples during working age

and retirement

We have to compute the joint value function of the couple to appropriately com-
pute joint labor supply and savings under the married couples’ available resources.
However, when computing the value of getting married for a single person, the rel-
evant object for that person is his or her discounted present value of utility in the
marriage. We thus compute this object for person of gender i who is married with a
specific partner.
Let ĉt (·), ˆlti,j (·), ât+1 (·), and D̂t (·) denote, respectively, the optimal consumption,
leisure, saving, and claiming decision for an individual of gender i in a couple with a
given set of state variables. During the working period, we have

Ŵ c (t, i, at , 1t , 2t , ȳt1 , ȳt2 ) = v i (ĉt (·), ˆlti,j )+

β(1 − ζ(·))Et Ŵ c (t + 1, i, ât+1 (·), 1t+1 , 2t+1 , ȳt+1
1 2
, ȳt+1 )+ (36)
βζ(·)Et W s (t + 1, i, ât+1 (·)/2, it+1 , ȳt+1
i
).

During the early retirement period, we have

N̂ c (t, i, at , 1t , 2t , ȳt1 , ȳt2 ) = v i (ĉt (·), ˆlti,j )

(37)
+ βEt V̂ c (t + 1, i, ât+1 (·), 1t+1 , 2t+1 , ȳt+1
1 2
, ȳt+1 )

Ŝ c (t, i, at , ȳr1 , ȳr2 , tr) = v i (ĉt (·), Li,j ) + βEt S c (t + 1, i, ât+1 (·), ȳr1 , ȳr2 , tr) (38)

V̂ c (t, i, at , 1t , 2t , ȳt1 , ȳt2 ) = (1 − D̂t (·))N̂ c (t, i, at , 1t , 2t , ȳt1 , ȳt2 ) + D̂t (·)Ŝ c (t, i, at , ȳr1 , ȳr2 , t).
(39)
During the retirement period, we have

R̂c (t, i, at , ψt1 , ψt2 , ȳr1 , ȳr2 , tr) = v i (ĉt (·), Li,j )+
βsi,j i p,j p c 1 2 1 2
t (ψt )st (ψt )Et R̂ (t + 1, i, ât+1 (·), ψt+1 , ψt+1 , ȳr , ȳr , tr)+

βsi,j i p,j p s i ¯i
t (ψt )(1 − st (ψt ))Et R (t + 1, i, ât+1 (·), ψt+1 , ȳr , tr),

(40)

23
where sp,j p
t (ψt ) is the survival probability of the partner of the person of gender i.

5 Estimation
We estimate our model on our two birth cohorts separately. For each cohort, we
adopt a two-step estimation strategy, as done by Gourinchas and Parker (2002) and
De Nardi, French, and Jones (2010 and 2016). We extend their approach to match
the life cycle profiles of labor market participation and hours (in addition to savings).
In the first step, for each cohort, we use data on the initial distributions at age
25 for our model’s state variables, and we estimate or calibrate those parameters
that can be cleanly identified outside our model. For example, we directly estimate
from the data the probabilities of marriage, divorce, and death, as well as the wage
processes while working and medical expenses during retirement.
In the second step, we use the method of simulated moments to estimate the
remaining model parameters. For the 1945 cohort, we estimate 19 model parame-
ters (β, ω, (φi,j i,j i,j 0,5 6,11
0 , φ1 , φ2 ), (τc , τc ), Li,j ).9 For the 1955 cohort, we assume that
the households of the 1955 cohort have the same discount factor β and weight on
consumption ω as the 1945 cohort, and we estimate the remaining 17 parameters.
To perform the estimation, for each cohort, we use the model to simulate a repre-
sentative population of people as they age and die, and we find the parameter values
that allow simulated life cycle decision profiles to “best match” (as measured by a
GMM criterion function) the data profiles for that cohort. The data that inform the
estimation of the parameters of our model are composed of the following 448 moments
for each cohort:

1. To better evaluate the determinants of labor market participation and their

responses to changes in taxes and transfers, we match the labor market par-
ticipation of four demographic groups (married and single men and women)
starting at age 25 and up to age 65 (41 time periods for each group).

2. To better evaluate the determinants of hours worked and their responses to

changes in taxes and transfers, we match hours worked conditional on par-
ticipation for four demographic groups (married and single men and women)
starting at age 25 and up to age 65 (41 time periods for each group).
9
We normalize the time endowment for single men.

24
 3. Because net worth, together with labor supply, is essential to smooth resources
during the working period and finance retirement, we match net worth for three
groups (couples and single men and women) starting at age 26 and up to age 65
(40 time periods for each group).10 Because people save to self-insure against
shocks and for retirement, matching assets by age is essential to evaluate the
effects of policy instruments and other forces, not only on saving but also on
participation and hours.

The mechanics of our MSM approach draw heavily from De Nardi, French, and
Jones (2010 and 2016) and are as follows. We discretize the asset grid, and, using value
function iteration, we solve the model numerically (see Appendix D for details). This
yields a set of decision rules that allows us to simulate life cycle histories for assets,
participation, and hours. We keep track of a large number of artificial individuals,
which are initially endowed with a value of the state vector drawn from the data
distribution for each cohort at age 25, generate their histories, and use them to
construct moment conditions and evaluate the match using our GMM criterion. We
search over the parameter space for the values that minimize the criterion. We repeat
the estimation procedure for each cohort.
Appendix E contains a detailed description of our moment conditions, the weight-
ing matrix in our GMM criterion function, and the asymptotic distribution of our
parameter estimates.

5.1 First-step estimation

Table 12 (in Appendix B) and Table 20 (in Appendix C) summarize our first-
step estimated or calibrated model inputs. The procedures for estimating wages as a
function of age and previous experience and earnings are new, as are the estimates of
the probability of marriage and divorce by age, gender, and wage shocks. Appendix B
details all of these inputs and reports additional first-steps inputs for both of each
cohorts.

5.1.1 Wages

We assume that wages are composed of a persistent stochastic shock and a com-
ponent that is a function of age, gender, and human capital. We measure human
10
Net worth at age 25 is an initial condition.

25
capital at a given point in time as one’s average realized earnings up to that time.
Thus, we allow past wages (and education, to the extent that it affects wages) and
labor market experience to affect one’s wage today. We estimate this relationship
from the PSID data.11

40 Single Men 40 Single Men

Single Women Single Women
Married Men Married Men
35 Married Women 35 Married Women
Hourly wage

Hourly wage
30 30

25 25

20 20

15 15
30 40 50 60 30 40 50 60
Age Age

Figure 4: Wage profile for single and married men and women at the average level of
human capital by age and subgroup. Left panel: 1945 cohort. Right panel,
1955 cohort. PSID data

Figure 4 displays the average age-efficiency profiles computed from the estimated
wage process that we estimate for men and women, evaluated at the average values
of human capital or average accumulated earnings at each age, (ȳt ). It shows that,
consistent with the evidence on the marriage premium, the wages of married men are
higher than those of single men. In contrast, the wages of married women are lower
than those of single women in our 1945 cohort, but this gap shrinks for our 1955
cohort because the average wage of married women has increased, while the average
wage for single women has stagnated. This is due to a combination of both different
11
Human capital, measured as average past earnings, soaks up more heterogeneity in wages than
education. Hence, we do not miss much by ignoring education when we take human capital into
account. To see this, consider the following. For our baseline specification, we estimate a fixed
effect regression of potential wage on age and human capital and their interactions with cohort and
gender. As shown in Table 13 in Appendix B, it yields an R-square of 0.103. We have also run an
alternative specification in which we run fixed effects regressions of potential wage on a polynomial
in age, interacted with gender and education. The resulting R-square is 0.067. Thus, the variability
in the wage data as measured by the R-square indicates that our measure of human capital explains
more of the variability in the data than a typical measure of education. The economic intuition
is that, conditional on years of education, types of major and quality of college imply much more
variation in wages than the variation that is implied by our measure of human capital. In addition,
we have also estimated a fixed effect regression which adds interactions with education for all of the
variables already included in our baseline specification (human capital, age, cohort, and gender).
This specification delivers an R-square of 0.116, which is only slightly higher than the one for our
base case.

26
returns to human capital and accumulated human capital levels. The stagnation of
men’s wages that we observe for our two cohorts is consistent with findings on wages
over time reported by Acemoglu and Autor (2011) and Roys and Taber (2017).
Table 15 in Appendix B reports our estimates for the earnings shock processes.
They imply that men and women face a similar persistence and earnings shock vari-
ance and that the initial variance upon labor market entry for men is a bit larger
than that for women.

5.1.2 Marriage, divorce, spousal wage shocks, spousal assets, and Social
Security benefits

We use the PSID to estimate the probabilities of marriage and divorce. Figure 14
in Appendix B displays our estimated benchmark probabilities of marriage for both
cohorts. Men with higher wage shocks are more likely to get married, but this gap
shrinks with age. In contrast, the probability of marriage for women displays little
dependence on their wage shocks. The comparison with the 1955 cohort shows that
the probability of getting married is smaller for the 1955 cohort, for both men and
women. Figure 15 in Appendix B reports results for our benchmark estimation of
divorce probabilities and shows that married men with lower wage shocks are more
likely to get divorced. The probability of divorce decreases with age, as does the
gap in the probabilities of divorce as a function of wage shocks. The probability of
divorce for women displays less dependence on the wage shock. The comparison with
the 1955 cohort shows that divorce rates are a bit lower in our more recent cohort,
once we condition on age and wage shocks.
We also estimate the joint distribution of (the logarithm of) the wage shocks of
new husbands and wives12 by age and we assume it is lognormal. We find that the
correlation of the logarithm of initial wage shocks between spouses is 0.27 in the 25-34
age group, 0.39 in the 35-44 group, and 0.45 after age 45. Because of these initial
correlations and the high persistence of shocks that we estimate at the individual
level, partners tend to have positively correlated shocks even after getting married.
Appendix B reports spousal assets and spousal Social Security earnings by spousal
wage shocks in case of marriage next period for both of our cohorts.
12
We assume it to be the same for both cohorts because the number of new marriages after age
25 is small during this time period.

27
5.1.3 Children

Figure 16 in Appendix B displays the average total number of children and average
number of children in the 0-5 and 6-11 age groups by parental age. It shows that the
number of children has decreased for married women and, to a smaller extent, for
single women in the 1955 cohort compared to the 1945 cohort.

5.1.4 Health, mortality, and medical expenses

Health, survival, and medical expenses in old age interact in an important way
to determine old-age longevity and medical expense risks. These risks, in turn, are
affected by the structure of taxation and Social Security rules. For these reasons, it
is important to capture the key aspects of health, mortality, and medical expenses to
evaluate the effects of these programs.
We take these data from the HRS, and because we have no data after age 65 for
the 1955 cohort, we assume that the 1955 cohort faces the same risks as the 1945
cohort in terms of health, mortality, and medical expenses.
Based on self-reported health status, we assume that health takes on two values,
good and bad. Figure 17 in Appendix B reports our estimated health transition ma-
trices by gender, age, and marital and health status. Women, married people, and
healthy people have longer life expectancies (Figure 18 in Appendix B displays the
survival probabilities by gender and marital and health status). Figure 19 in Ap-
pendix B displays the importance of medical expenditures after retirement. Average
medical expenses climb fast past age 85 and are highest for single and unhealthy
people.

5.2 Second-step estimation

Table 1 presents our estimated preference parameters for both cohorts.13 For
the 1945 cohort, our estimated discount factor is 0.990, the same value estimated
by De Nardi, French, and Jones (2016) on a sample of elderly retirees, and our
estimated weight on consumption is 0.406. We assume that the 1955 cohort shares
these preference parameters. While we normalize the total weekly time endowment of
single men to 5,840 hours a year, and thus 112 hours a week, for our 1945 cohort, we
13
Table 21 in Appendix F reports all of our estimated parameters for both cohorts and their
standard errors.

28
 Estimated parameters 1945 cohort 1955 cohort

β: Discount factor 0.990 Same as 1945 cohort

ω: Consumption weight 0.406 Same as 1945 cohort
L2,1 : Time endowment (weekly hours), single women 107 112
L1,2 : Time endowment (weekly hours), married men 107 101
L2,2 : Time endowment (weekly hours), married women 88 88
τc0,5 : Prop. child care cost for children ages 0-5 30% 25%
τc6,11 : Prop. child care cost for children ages 6-11 7% 19%
Φi,j
t : Partic. cost Fig. 23 Fig. 23

Table 1: Second-step estimated model parameters

estimate that single women have a total weekly time endowment of 107 hours a week.
We interpret this as single women having to spend five more hours a week managing
their household and rearing children (they have fewer children than married women
but still more than single men) or taking care of elderly parents. The corresponding
time endowments for married men and women are, respectively, 107 and 88 hours.
This implies that people in the latter two groups spend 5 and 24 hours a week,
respectively, running households, raising children, and taking care of aging parents.
Our estimates of non-market work time are remarkably similar to those reported
by Aguiar and Hurst (2007), who find that, in the 1985 American Time Use Survey
(ATUS) data set (when our 1945 cohort was 42 years old), men and women spent 14
and 27 hours a week, respectively, engaging in non-market work. Using more recent
data, Dotsey, Li, and Yang (2014) find that, similarly to Aguiar and Hurst (2007),
people spend 17 hours per week on average on activities related to home production.
It should be noted that, even for a working woman, 24 hours can amount to, for
example, spending seven hours each day on Saturday and Sunday and two hours a
day the other five days by parenting, cooking, doing laundry, cleaning, organizing
one’s house, and taking care of one’s parents. Thus, the data and model estimates
are very consistent in the way households spend time running their households and
providing care.
Our estimates for the 1945 cohort imply that the per-child child care cost of having
a child ages 0-5 and 6-11 are, respectively, 30% and 7% of a woman’s wage. In the
PSID data, child care costs are not broken down by age of the child, but per-child

29
child care costs (for all children in the age range 0-11) of a married woman are 31%
and 20% of her earnings at ages 25 and 30, respectively. Computing our model’s
implications, we find our corresponding numbers for a married woman are 23% and
18% of her earnings, respectively, at ages 25 and 30. Thus, our model infers child
care costs that are similar to those in the PSID data.
For the 1955 cohort, we notice two main changes compared with the 1945 cohort.
First, to help reconcile the lower hours worked by married men in this cohort, the
model estimates that their available time to work and enjoy leisure decreases by six
hours a week. Second, to help reconcile the slopes of hours and participation over the
life cycle by married women in the presence of fewer children, the model estimates that
the per-child child care costs of having younger children goes up, while that of having
older children goes down. While decomposing the effects of changing labor supply
between the two cohorts is very interesting (see, for instance, Attanasio, Low, and
Sánchez-Marcos (2008) and Eckstein and Lifshitz (2011)), we abstract from analyzing
it further because of space constraints.
Figure 23 in Appendix F reports the age-varying time costs of working by age
expressed as fraction of the time endowment of single men that are necessary to
reconcile the labor market participation of our four groups of people in each cohort.
Our estimated participation costs are relatively high when people are younger and,
with the exception of single men, increase again after 45. The time costs of going to
work might include factors other than commuting time. For instance, they might be
higher when children are youngest because, for instance, during that period parents
might need additional time to get their children back and forth from day care. They
also show that, conditional on all aspects of our environment, the participation costs
of married women are the lowest ones because married women face lower wages, have
a smaller time endowment (because of time spent engaging in home production and
child care), and tend to have higher-wage husbands who work.

5.3 Model fit

Figures 5 and 6 report our model-implied moments, as well as the moments and
95% confidence intervals from the PSID data for our 1945 cohort for the moments
that we target in our estimation procedure. They show that our model matches
participation, hours conditional on participation, and asset accumulation for all of

30
 Married women Single women
1 1
Labor Participation

Labor Participation
Model
Data
0.5 0.5
Model Data, upper bound
Data Data, lower bound
Data, upper bound
Data, lower bound
0 0
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age
Married men Single men
1 1
Labor Participation

Labor Participation
Model Model
Data Data
0.5 0.5
Data, upper bound Data, upper bound
Data, lower bound Data, lower bound

0 0
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

(a) Participation, couples (b) Participation, singles

Married women Single women

2500 2500
Hours among workers

Hours among workers

2000 2000

1500 1500
Model Model
Data Data
1000 1000
Data, upper bound Data, upper bound
Data, lower bound Data, lower bound
500 500
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age
Married men Single men
2500 2500
Hours among workers

Hours among workers

2000 2000

1500 1500
Model Model
Data Data
1000 Data, upper bound 1000 Data, upper bound
Data, lower bound Data, lower bound
500 500
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

(c) Hours for workers, couples (d) Hours for workers, singles

Figure 5: Model fit for participation (top graphs) and hours (bottom graphs) and average
and 95% confidence intervals from the PSID data

31
our demographic groups.

#105 Couples #105 Single women

Average Household Asset

8 5
Model
Model
4 Data
Data
7 Data, upper bound
Data, upper bound
3 Data, lower bound
Data, lower bound
6 2
Average Household Asset

1
5
0
25 30 35 40 45 50 55 60 65
4 Age
#105 Single men

Average Household Asset

5
3 Model
4 Data
Data, upper bound
2 3 Data, lower bound

2
1
1

0 0
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

(a) Assets, couples (b) Assets, singles

Figure 6: Model fit for assets and average and 95% confidence intervals from the PSID
data

Figure 25 in Appendix G compares additional model implications for couples with

those in the data for our 1945 cohort for moments that we do not target in estimation.
They show that our estimated model also matches the fraction of couples with two
workers, with only the husband working, with only the wife working, or with none
working by age. They also display that our model produces reasonable implications
for the hours worked over the life cycle for each type of couple. Our model fits the
data well for the 1955 cohort too (to save on space, we show the graphs for the 1955
cohort in Appendix G).
Thus, our parsimoniously parameterized model reproduces all of these features
of the data well, including those that are not matched by construction, which is
remarkable given that it is tightly parameterized. In fact, we estimate 19 parameters
and 448 targets for the 1945 cohort and 17 parameters and 448 targets for the 1955
cohort.

5.4 Identification
The fixed cost of participation by age and subgroup (Φi,j
t ) especially affects par-
ticipation by subgroup over the life cycle. The available time endowment (Li,j ) has
first-order effects on hours worked by workers. Child care costs have a larger effect
on hours than participation and especially affect hours worked by women when they

32
have young children. This effect is especially large for married women, as they have
more children than single women.
The discount factor (β) has large effects on savings. The weight on consumption
(ω) affects the intratemporal substitution between consumption and leisure and thus
affects hours worked at all ages. Because the wage is increasing with human capital
(and past hours worked), a high ω increases the value of consumption at all ages but
has a larger impact on the hours of older workers relative to younger workers.

6 Model validation in terms of elasticities

To help build confidence in our model’s responses to policy changes, we report its
labor supply elasticities. Table 2 shows the (compensated) elasticities of participation
and hours among workers with respect to an anticipated change to their own wage.14
It shows that, first, the elasticity of participation of women is larger than that of men,
for both married and singles. Second, it shows that married men have the lowest
elasticity of participation. Third, it shows that the elasticity of participation for all
groups is largest around retirement age, a finding that confirms that of French (2005)
for men. Fourth, our elasticities are consistent with those in Liebman, Luttmer, and
Seif (2009), which uses HRS data for people over age 50 and variation stemming from
Social Security rules. Their results imply that the yearly elasticity at the extensive
margin is 0.7 for the sample of men and women, 1.1 for women, and 0.2 (but not
statistically significant) for men. At the intensive margin, their elasticity is 0.4 for
men and women, 0.7 for men, and -0.3 (but not statistically significant) for women.
Thus, their estimated labor supply elasticities at the extensive and intensive margins
are consistent with those in our 50 and 60 age groups. More generally, our model’s
implied elasticities at all ages are in line with those in the vast existing literature, as
surveyed in Blundell and MaCurdy (1999) and more recently estimated by Attanasio
et al. (2018).
While important to compare with the empirical estimates, the compensated wage
elasticities are not necessarily indicative of how participation and hours would change
as a result of a wage change that permanently affects all of the population and which
is more similar to that implied by a permanent tax change at all income levels.
14
For this computation, we temporarily increase the wage for only one age and one group at a
time (married men, married women, single men, or single women) by 5%.

33
 Participation Hours among workers
Married Single Married Single
W M W M W M W M
30 1.0 0.0 0.5 0.2 0.2 0.3 0.4 0.3
40 0.7 0.1 0.4 0.2 0.3 0.5 0.5 0.5
50 0.6 0.2 0.4 0.5 0.5 0.5 0.8 0.5
60 1.1 0.8 1.4 2.0 0.4 0.2 0.5 0.3

Table 2: Model-implied elasticities of labor supply

To help shed light on what we should expect from our policy experiments, here we
report the effects of a permanent increase of 5% in the wage schedule of married
women when the wage structure of the three other demographic groups remains the
same. The left panel of Figure 7 shows that a permanent wage increase for married
women implies a much higher, and U-shaped, elasticity of participation for married
women, which peaks at 2.5 at age 25. It also reports the cross-elasticities of the other
groups to changes in the wages of married women. The right panel highlights that
permanent wage changes can lead to high increases in married women’s participation,
with participation being 4-7 percentage points higher over all of their life cycle. It
also shows that the participation of single women rises because they expect to get
married and obtain higher wages (and higher returns to their accumulated human
capital) upon marriage. There is little response in the participation of single men.
In contrast, married men’s participation after age 40 decreases when women’s wage
schedule increases. This shows that modeling men’s labor supply is important to
assess the effects of reforms affecting the wages of married women in a long-lasting
way.

7 Policy experiments: eliminating marital Social

Security benefits and joint taxation
We now turn to evaluating the effects of various policy reforms. We first show the
labor supply and savings responses resulting from the elimination of various marital
policies, and we then evaluate their welfare implications.

34
 Raise wages of married women by:5% Raise wages of married women by:5%
3 0.08
Single Men Single Men
Single Women Single Women
2.5 Married Men Married Men
0.06
Married Women Married Women
Elasticity in participation

Change in participation
0.04

1.5
0.02
1

0
0.5

-0.02
0

-0.5 -0.04
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

Figure 7: Elasticity of participation (left graph) and change in participation (right graph)
for a 5% permanent increase in the wage schedule for married women. Effect
on all four demographic groups. Model implications.

7.1 Outcomes
For each policy counterfactual, we compute two sets of results. The first one
balances the government budget constraint by adjusting the proportional component
of the income tax, while the second one keeps the government budget constraint
unbalanced. Because of space constraints, we report the effects of the latter set of
experiments in Appendix H.

7.1.1 Eliminating spousal Social Security benefits, 1945 cohort

According to the current Social Security rules, one’s spouse can receive half of his
or her partner’s contribution while their partner is alive and all of the benefits of their
deceased spouse. This provision potentially has three effects. First, it discourages
the labor supply of the secondary earner, given that he or she can benefit from
spousal benefits. Second, it encourages the labor supply of the main earner, who
is also working to provide Social Security benefits to the secondary earner spouse.
Third, it reduces retirement savings because it raises the annuitized income flow of
the secondary earner or non-participant.
When eliminating both spousal Social Security benefits, the government runs a
budget surplus and can cut the proportional component of the income tax from 4.0%
to 1.8%. The left panel of Figure 8 shows that the participation of married women is,
respectively, 10, 11, and 4 percentage points higher at ages 25, 55-60, and 65 without
spousal Social Security benefits. In contrast, married men decrease their participation
starting at age 55, and their participation is 6 percentage points lower by age 65. A

35
model in which married men cannot change their participation or can do it only after
a certain age would miss this effect. The participation of single women at ages 25-30
increases (by 3 percentage points) because, should they marry, they now expect no
Social Security benefits coming from their spouse’s labor supply. As they age, the
probability that they marry becomes negligible, and the effect on their participation
of the elimination of spousal benefits fades.
No spousal or survival benefit in SS No spousal or survival benefit in SS
0.3 10000
Single Men Single Men
Single Women Single Women
0.25 8000
Married Men Married Men

Change in average Labor Income

Married Women Married Women
0.2 6000
Change in participation

0.15 4000

2000
0.1

0
0.05

-2000
0
-4000
-0.05
-6000
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

Figure 8: Changes in participation (left panel) and labor income (right panel) after the
elimination of all spousal Social Security benefits when the income tax is reduced
to balance the government budget.

An important reason why these reforms have such large effects on the labor supply
of married women resides in the initial distribution of potential wages of men and
women at age 25. Table 3 shows that, in the 1945 cohort, 60% of women and only
20% of men belong to the bottom two quintiles of wages at age 25. Thus, most women
have low wages and tend to be secondary earners in this cohort. For this reason, they
react strongly to the elimination of spousal benefits.

Wage quintile
1 2 3 4 5
Men 7.9 % 12.4 % 21.3 % 28.2 % 30.2 %
Women 32.3 % 27.8 % 18.7 % 11.6 % 9.6 %

Table 3: Distribution of men and women across potential wage quintiles at age 25, 1945
cohort, PSID data.

Groneck and Wallenius (2017) and Kaygusuz (2015) study the effects of mari-
tal Social Security benefits in simpler models than ours in which, for instance, men

36
cannot change their labor supply and women can do so to a limited extent.15 They
report that, over all of the working period, their model implies an increase in the par-
ticipation by married women of 6.4 and 6.1 percentage points, respectively. Because
we also allow men to adjust their labor supply, and they choose to reduce it in older
ages, and because women (as in the data) have more flexibility in their hours worked,
we find effects that are a bit larger but in the same ballpark.
The right panel of Figure 8 reports changes in labor income for our four demo-
graphics groups. Married women work more, accumulate more human capital, and
earn more as a result of the reform. Married women’s labor income is about, respec-
tively, 18%, 12%, and 11% higher at ages 25, 55-60, and 65. The labor income of
married men drops by about 13% by age 65.

Couples Single men Single women

Savings, balanced government budget 14.9% 7.8% 11.2%

Table 4: Change in savings at age 66, in percentages, as a result of removing spousal Social
Security benefits when the income tax is reduced to balance the government
budget.

Table 4 shows the resulting changes in assets at retirement time. The reform
increases savings by reducing government payments to spouses and widows during
retirement, and assets at retirement go up by 14.9%, 7.8%, and 11.2% for couples,
single men, and single women, respectively.

7.1.2 Eliminating joint income taxation, 1945 cohort

Figure 9 displays the effects on participation of having everyone file as singles (the
married men file as single men and the married women as single women) and reducing
the proportional component of the income tax to balance the government budget
(from 4.0% to 3.5%). As a result of this policy, the participation of married women
increases by more than 20 percentage points until age 35 and by 10 percentage points
between ages 45 and 60. The participation of single women also increases slightly
until age 60. The right panel of Figure 1 provides the key intuition for this result:
15
Their models are also less rich along other important dimensions and are calibrated rather than
estimated.

37
 No marital differential tax
0.3
Single Men
Single Women
0.25 Married Men
Married Women
0.2

Change in participation
0.15

0.1

0.05

-0.05

25 30 35 40 45 50 55 60 65
Age

Figure 9: Changes in participation after the elimination of joint income taxation when the
income tax is reduced to balance the government budget.

the marginal tax rates for married women working are much lower when they do not
file jointly with their husbands.
Guner, Kaygusuz, and Ventura (2012a) study the switch from current U.S. taxa-
tion to single filer taxation in a calibrated model of a steady state and find that the
labor supply of married women goes up by 10-20 percentage points. Our effects on
the labor supply of married women are thus close to theirs.

7.1.3 Eliminating spousal Social Security benefits and joint income tax-
ation, 1945 cohort

This policy change implies a reduction in the proportional component of the in-
come tax from 4.0% to 2.0%. Figure 10 displays the participation profiles in our
Benchmark All policies change
1 1

0.9 0.9

0.8 0.8
Labor Participation

Labor Participation

0.7 0.7

Single Men
0.6 0.6
Single Women
Married Men
0.5 0.5 Married Women

Single Men
0.4 Single Women 0.4
Married Men
Married Women
0.3 0.3
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

Figure 10: 1945 cohort: Participation in the benchmark (left panel) and reformed econ-
omy (right panel) after the elimination of all spousal Social Security benefits
and joint income taxation when the income tax is reduced to balance the gov-
ernment budget.

benchmark economy and this counterfactual economy. Eliminating spousal Social

38
Security benefits and joint income taxation has a large effect on the participation of
married and single women. To make magnitudes clearer, the left panel of Figure 11
plots the differences in participation between the benchmark and this counterfactual
for each group of people by age. It shows that the participation of married women
is 16-30 percentage points higher until age 62 in the no-marital-provisions economy.
The participation of single women is about 5 percentage points higher until age 40.
The participation of married men is higher in their middle age, reaching a peak of
2 percentage points higher than in the benchmark, but is 8 percentage points lower
than in the benchmark at age 65. Thus, the timing of their participation changes
over their life cycle. This highlights the importance of also modeling their labor sup-
ply behavior over their life cycle, in addition to that of their wives when we change
provisions that affect both members in the household.
All policies change All policies change
0.3 10000
Single Men
Single Women
0.25 8000
Married Men
Change in average Labor Income

Married Women
0.2 6000
Change in participation

0.15 4000

2000
0.1

0
0.05

-2000
0 Single Men
Single Women
-4000
Married Men
-0.05
Married Women
-6000
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

Figure 11: Changes in participation (left panel) and labor income (right panel) after the
elimination of all spousal Social Security benefits and joint income taxation
when the income tax is reduced to balance the government budget.

Couples Single men Single women

Savings, balanced government budget 20.3% 8.8% 14.8%

Table 5: Change in assets at age 66, in percentages, as a result of removing spousal Social
Security benefits and joint income taxation when the income tax is reduced to
balance the government budget.

Table 5 displays the effects on savings at retirement time. Couples now save 20.3%
more for retirement, while single men and women save, respectively, 8.8% and 14.8%
more.

39
7.1.4 Eliminating Marital Social Security benefits and joint taxation for
the 1955 cohort

We now turn to studying the effects of marriage-related taxes and Social Security
benefits for the 1955 cohort. In the interest of space, we report results only for the
case in which we eliminate all three marriage-related provisions at the same time.
The left panel of Figure 12 displays the difference in the participation profile. This
graph shows that eliminating all marital-related provisions also has large effects for
the 1955 cohort, in which labor supply participation is much higher to start with.
Thus, the effects of these policies on a relatively younger cohort with a much higher
participation of married women continue to be very large.
The effects of increased labor market experience on wages are similar to those
in the 1945 cohort, and, as for the 1945 cohort, increased wages and participation
(hours increase little for the workers) imply higher average earnings of $5,000-6,000
per year for married women and $3,000 for single women for most of their life cycle.
Average earnings of married men start dropping earlier for this cohort, that is, at age
50, compared with age 55 for the 1945 cohort, but their drop is smaller by age 65 (see
right panel in Figure 12).
All policies change, fix G All policies change, fix G
0.3 10000
Single Men
Single Women
0.25 8000
Married Men
Change in average Labor Income

Married Women
6000
Change in participation

0.2

4000
0.15
2000
0.1
0

0.05
-2000
Single Men
0 Single Women
-4000
Married Men
Married Women
-0.05 -6000
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

Figure 12: 1955 cohorts: Changes in participation (left panel) and labor income (right
panel) after the elimination of all spousal Social Security benefits and joint
income taxation when the income tax is reduced to balance the government
budget.

Table 6 displays the effects on savings at retirement time. Couples now save 19.7%
more for retirement, while single men and women save, respectively, 8.4% and 14.9%
more.
Comparing Tables 3 and 7 highlights that the fraction of women in the lowest wage
quintile has decreased and the fraction of women in the highest one has increased from

40
 Couples Single men Single women
Savings, balanced government budget 19.7% 8.4% 14.9%

Table 6: Change in assets at age 66, in percentages, as a result of removing spousal Social
Security benefits and joint income taxation when the income tax is reduced to
balance the government budget.

Wage quintile
1 2 3 4 5
Men 10.4 % 14.0 % 19.1 % 26.0 % 30.4 %
Women 28.1 % 25.1 % 20.6 % 15.1 % 11.1 %

Table 7: Distribution of men and women across potential wage quintiles at age 25, 1955
cohort, PSID data.

the 1945 to the 1955 cohort, but it is still the case that, even in the 1955 cohort, most
women tend to have lower wages and thus be secondary earners in this cohort, and
therefore respond strongly to the elimination of marital provisions.

7.2 Welfare
To evaluate welfare changes, we calculate the asset compensation required for each
household at age 25 to stay in the benchmark economy and report it as a fraction of
average income in the benchmark economy. Thus, negative asset compensations mean
that households are better off in the benchmark economy. The first three columns
in Table 8 report the average welfare gains or losses conditional on one’s marital
status at age 25. We also report the fraction of households gaining and losing and
the average gains and losses among each of these groups.
The top panel of counterfactuals refers to the 1945 birth cohort. The first set of
results in the top panel compares our benchmark economy with one in which there are
no marital Social Security benefits and taxes remain unchanged despite the resulting
government surplus. On average, couples would need to be compensated by a onetime
asset transfer at age 25 that is equivalent to 0.25 of average earnings in the economy,
while single women would require 0.23 average earnings, as they expect to marry and
potentially lose these benefits after marriage. While sizable, these welfare costs are
not very large because, as of age 25, people already know that these benefit changes

41
 All Winners Losers
Couples SW SM Couples SW SM Couples SW SM
1945 cohort
(1). Remove Social Security spousal benefits, unbalanced budget
Average -0.25 -0.23 0.31 0.00 0.00 0.31 -0.25 -0.23 -0.02
Percentage 0.0 0.0 100.0 100.0 100.0 0.0
Percentage (No marriage prob.) 0.0 0.0 0.0 100.0 0.0 0.0
(2). Remove Social Security spousal benefits, balanced budget
Average 0.71 0.20 1.30 0.71 0.22 1.30 0.00 -0.04 0.00
Percentage 100.0 93.4 100.0 0.0 6.6 0.0
(3). Remove joint income taxation, unbalanced budget
Average 0.13 -0.19 1.04 0.35 0.07 1.04 -0.18 -0.20 0.00
Percentage 58.4 4.5 100.0 41.6 95.5 0.0
(4). Remove joint income taxation, balanced budget
Average 0.33 -0.10 1.25 0.45 0.11 1.25 -0.09 -0.15 0.00
Percentage 78.5 17.9 100.0 21.5 82.1 0.0
(5). Remove all marital related policies, balanced budget
Average 0.83 0.03 2.24 0.84 0.31 2.24 -0.04 -0.13 0.00
Percentage 98.9 35.8 100.0 1.1 64.2 0.0
1955 cohort
(6). Remove all marital related policies, balanced budget
Average 0.75 0.21 1.31 0.77 0.31 1.31 -0.05 -0.05 -0.02
Percentage 97.2 70.9 100.0 2.8 29.1 0.0

Table 8: Asset compensation required for staying in the benchmark economy, normalized
as a fraction of average income in the benchmark economy. SM: single men, SW:
single women. Top line for each experiment: average welfare gain or loss. Bottom
line for each experiment: fraction in that group gaining or losing welfare.

will take place at retirement time and they have many years to work and save to
make up for the potential loss in benefits. In contrast, single men benefit from this
policy change because their wives will work more, earn more, and accumulate more
human capital after they marry, and single men do not take into account their future
wife’s disutility from working more. The remaining columns in the table distinguish
the effects for winners and losers. The three “winners” columns show that all couples
and single women, on average, lose from this policy, whereas all single men gain from
it. The third row in this panel reports the percentage of people winning and losing for
the same policy experiment but in an economy in which people no longer marry after

42
age 25. It clarifies the role of marriage expectations in driving our welfare calculations
and highlights that this benefit removal would have no effects on single people, who
don’t expect to receive this benefit anyway if they do not expect to get married.
The second set of results removes marital Social Security provisions and balances
the government budget by reducing the proportional component of the income tax
from 4.0% to 1.8%. The first three columns display large welfare gains: couples would
be willing to pay, on average, an asset amount that corresponds to 0.7 times average
income in welfare terms, single women 0.2 times, and single men 1.3 times. The
second set of three columns shows that all couples, 93.4% of single women, and all
single men would benefit from these changes. The last set of three columns shows that
the 6.6% of the single women who lose would face very small losses. These are women
whose initial human capital is very low and are heavily relying on marital benefits.
Thus, this counterfactual suggests that eliminating these benefits while reducing the
income tax would benefit the vast majority of the young population and would have
only small welfare costs for a small fraction of single women.
The third set of results makes everyone file as single people. The first line in
this panel does not balance the government budget constraint and shows that the
willingness to pay for this policy, measured as a onetime asset amount as a fraction of
average income, is equal to 0.13 and 1.04 for couples and single men, respectively. In
contrast, single women would lose and require an asset compensation of 0.19 average
income. This happens because they know they will be working more and enjoying
less leisure in the future, and especially so after marriage. The winners and losers
columns reveal that 58.4% of couples, 4.5% of single women, and 100% of single men
would favor this policy and that both gains for the winners and losses for the losers
would be sizable.
The fourth set of results balances the government budget constraint by reducing
the income tax (from 4.0% to 3.5%) and generates more winners with larger welfare
gains and fewer losers, who also experience smaller welfare losses than in the previous
experiment. For instance, 78.5% of the couples would be willing to give up an asset
amount corresponding to 0.84 average income to live under this policy, while the
compensation for the remaining 21.5% would amount to assets equal to 0.09 of average
income.
The fifth set of results for the 1945 cohort eliminates all of the marriage-related
policies that we consider and balances the government budget constraint by reducing

43
the income tax, which goes down from 4.0% to 2.0%. This policy change generates
the largest aggregate welfare gains among the set that we consider for the 1945 co-
hort: 0.83, 0.03, and 2.24 times average income for couples, single women, and single
men, respectively. Among couples, 98.9% would gain, compared with 35.8% of single
women and 100% of single men. The bigger losers coming out of this policy are 64.2%
of single women, who lose, on average, 0.13 of average income.
The results in the last panel refer to the 1955 cohort and show that there would
also be large aggregate gains from removing marriage-related provisions and reducing
the income tax and that single women in this cohort would be less disadvantaged by
this policy than single women in the 1945 cohort: only 29.1% of them would lose,
compared with 64.2%. In addition, their loss would be much smaller (0.05 average
income, compared with 0.13 in the 1945 cohort). In both cohorts, only a minority of
couples would lose and would experience a small welfare loss.
Overall, our policy experiments thus indicate that removing marriage-related taxes
and Social Security benefits would increase female labor supply and the welfare of the
majority of the populations, whereas the rest would only bear small welfare costs.

8 Changes in marriage and divorce patterns in re-

sponse to policy
Because we study labor supply and savings responses to the elimination of two
important marriage-based policies, the question of the robustness of our findings to
changes in marriage and divorce patterns naturally arises. To address this question,
we first turn to summarizing previous empirical findings on the effects of changes
in Social Security rules and income taxes on marriage and divorce patterns. Then,
to evaluate the robustness of our results, we perform policy experiments in which
marriage and divorce exogenously change at the same time as we eliminate marriage-
based taxes and Social Security benefits by more than what has been found in the
empirical literature.

44
8.1 The effects of Social Security and income taxes on mar-
riage and divorce in the empirical literature
Before 1977, Social Security spousal benefits were available to the secondary earner
in case of divorce after 20 years of marriage. After that date, the threshold for
eligibility became 10 years of marriage before divorce.16 Dickert-Conlin and Meghea
(2004) examine the 1977 U.S. policy switch (using data from the U.S. National Vital
Statistics System, the 1980 census, and the Current Population Survey) and conclude
that it had no effect on divorces and remarriages. Goda, Shoven, and Slavov (2007)
also find no impact of the 10-year eligibility discontinuity on divorces (using data
from the PSID Marital History File). Dillender (2016) confirms that these rules have
no effects overall and small ones on a very small number of people: those who married
late. Thus, previous literature indicates that the effects of Social Security benefits
kinks are negligible in the United States.
Turning to the effect of income taxes on marriages and divorces, Alm and Whit-
tington (1995) use time series data from 1947 to 1988 and argue that “the magnitude
of this impact is quite small. This result suggests that some individuals respond to tax
incentives in their marriage choices, but that for many individuals taxes do not affect
these decisions.” Alm and Whittington (1997) use data from the PSID and estimate
a discrete-time hazard model of the probability of divorce from the first marriage.
They conclude that “couples respond to tax incentives in their decision to divorce,
although these responses are typically small.” Alm and Whittington (1999) utilize
the same data to estimate a discrete-time hazard model of the time to first marriage
from 1968 to 1992, and uncover that the income tax has no effect on the marriage
decisions of men and only a small effect on the marriage decisions of women. They
thus conclude that, in the context of the United States, “In general, the impacts of
the income tax variables, even when statistically significant, are small.”
Thus, for the United States, previous empirical studies on the impact of income
taxation and Social Security benefits on marriage and divorce find either no significant
effects or very small effects that apply to tiny groups of people.
Looking into welfare programs, Low et al. (2018) study the U.S. subpopulation of
low-education mothers on welfare and the 1996 welfare reform, which was meant to
16
We do not model this part of the benefits because the fraction of people divorcing after 10 years
of marriage is small, and this addition would add great computational complexity to our framework.

45
encourage labor supply by welfare recipients and reduce marital disincentives. They
use the Survey of Income and Program Participation (SIPP) data and document
that the reform greatly reduced welfare recipience and increased labor participation
of mothers, but had no effects on marriage and fertility. They do find an effect on
divorce rates, which declined from 0.9% before the reform to 0.7% after the reform.
This is a non-trivial drop as a fraction of divorces, but in absolute terms, the reform
reduced a very small number to a tiny one and refers to a small population.
More broadly on the effects of welfare programs, a survey by Moffitt (1998) con-
cludes, “Most find that the majority of studies show either no significant effects of
AFDC and other welfare programs, effects that are statistically significant but small
in magnitude, a set of mixed effects indicating some that are favorable and some un-
favorable, or effects that occur only for some specific types of programs. Although the
research reviewed in these chapters does not support a finding of no effect whatsoever
of welfare programs on demographic behavior, it would be difficult to argue that the
research often indicates very sizable or stable effects.”
Persson (2017) studies the elimination of marital survivorship benefits that took
place in 1989 in Sweden and infers larger effects than those found for the United
States. In terms of comparison with our work, her main finding is that the divorce
rate increased by 10% as a result of the elimination of marital survivorship benefits.
Although this effect is sizeable in percentage terms, it is a small change from the
standpoint of the overall population because the divorce rate is small.
In comparing findings for the United States and Sweden, it is also important to
keep in mind that cultural and religious factors are important reasons why people
marry and stay married, and that marriage is much more widespread in the United
States than in Sweden. For instance, United Nations (Department of Economic and
Social Affairs, 2012) reports that in 1985, a time period preceding the 1989 Swedish
marital benefits reform, only 35.8% of the 25- to 29-year-old Swedish women were
married, compared with 64.3% in the United States. In addition, in the 1980s un-
married women in the United States accounted for 18% of live births, compared to
40% in Sweden (Sorrentino, 1990).

46
8.2 Robustness of policy results to large changes in marriage
and divorces
In this subsection, we compare the effects of a policy experiment in which we
eliminate joint income taxation of couples and Social Security marital benefits (for
the 1945 cohort) for given marriage and divorce patterns with the effects of the same
policy when there are also two possible alternative changes in marriage and divorce
patterns. In the first robustness exercise, the policy decreases marriage rates by 20%
and increases divorce rates by 20%. Alternatively, in the second robustness exercise,
the policy increases marriage rates by 20% and decreases divorce rates by 20%. In
both, we also balanced the government budget.
All policies change All policies change, lower marriage and higher divorce rates All policies change, higher marriage and lower divorce rates
0.3 0.3 0.3
Single Men Single Men Single Men
Single Women Single Women Single Women
0.25 Married Men 0.25 Married Men 0.25 Married Men
Married Women Married Women Married Women
0.2 0.2 0.2
Change in participation

Change in participation

Change in participation
0.15 0.15 0.15

0.1 0.1 0.1

0.05 0.05 0.05

0 0 0

-0.05 -0.05 -0.05

25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age Age

Figure 13: Differences in participation after the elimination of all spousal Social Security
benefits and joint income taxation for the 1945 cohort. Left panel: benchmark
economy with unchanging marriage and divorce after the policy change. Middle
panel: 20% lower marriage probability and 20% higher divorce rate after the
policy change. Right panel: 20% higher marriage probability and 20% lower
divorce rate after the policy change.

Figure 13 highlights several important findings. First, all changes in participation

of the four groups are very similar whether marriage and divorce patterns change or
not. Second, comparing the left panel (no marriage and divorce changes) and the
middle panel (decreased marriage and increased divorce probability) shows that a
reform that lowers the probability of marriage and raises the probability of divorce
makes women more self-reliant on their own labor supply and human capital. Married
women work more (accumulating more human capital) to edge against divorce risk.
Single women are less likely to marry and also work more (accumulating more human
capital). Comparing the left panel (no marriage and divorce changes) and the right
panel (increased marriage and decreased divorce probability) highlights that increas-
ing the marriage rate and lowering the divorce rate has the opposite effect, but that

47
this effect is small and does not change the conclusions reached in our benchmark
policy experiment.

Couples Single men Single women

Benchmark 20.3% 8.8% 14.8%
Low marriage, high divorce 20.3% 7.6% 14.7 %
High marriage, low divorce 21.1% 11.7% 15.9 %

Table 9: Change in assets at age 66, in percentages, as a result of removing spousal Social
Security benefits and joint income taxation.

Table 9 displays the effects on savings at retirement time for the three experiments.
The second and third rows show that the effects on savings of couples, who make
up the vast majority of the population, are also very robust to changes in expected
marital patterns. Our results are thus robust to large changes in marriage and divorce
rates.

9 Conclusions
We estimate a model of labor supply and savings for single and married people
that allows for a rich representation of the risks that people face over their entire life
cycle and for the important provisions of taxes and Social Security for singles and
couples. We do so for both the 1945 and the 1955 birth cohorts, and we show that
our model fits the data very well, including along important dimensions that it was
not meant to match by construction, such as the elasticities of labor supply. We find
that the fact that young women entering the labor market have much lower wages
than those of men and the time and monetary costs that children imply are important
determinants in the labor supply of single and married men and women.
We use our model to evaluate the effect of marriage-based Social Security benefits
and the marriage tax bonus and penalty. We find that these marriage-based provisions
have a strong disincentive effect on the labor supply not only of married women, but
also of single young women who expect to get married. This lower participation
reduces their labor market experience, which, in turn, reduces their wages over their
life cycle. These provisions also induce married men to work longer at their careers
and depress the savings of couples. Our findings are robust to changes in marriage

48
and divorce patterns. These effects are very similar for the 1945 and 1955 birth
cohorts, even though the labor market participation of young married women in the
1955 cohort is over 10 percentage points higher than that of the 1945 cohort. We
also show that, if the government surplus resulting from the elimination of marriage-
related provisions were used to lower income taxation, there would be large welfare
gains for the vast majority of the population and the few losing would experience
small welfare losses.
Our paper provides several contributions. First, it is the first estimated structural
model of couples and singles that allows for participation and hours decisions of both
men and women, including those in couples, in a framework with savings. Second, it is
the first paper to study all marriage-related taxes and benefits in a unified framework.
Third, it does so by allowing for the large observed changes in the labor supply of
married women over time by studying two different cohorts. Fourth, our framework is
very rich along dimensions that are important in the study of our problem, including
labor market experience affecting wages and carefully modeling survival, health, and
medical expenses in old age, and their heterogeneity by marital status and gender.

49
References
[1] Acemoglu, Daron, and David Autor. 2011. “Skills, Tasks and Technologies: Im-
plications for Employment and Earnings.” in D. Card and O. Ashenfelter, ed.,
Handbook of Labor Economics, volume 4, (North Holland), 1043-1171.

[2] Aguiar, Mark, and Erik Hurst. 2007. “Measuring Trends in Leisure: The Al-
location of Time Over Five Decades.” Quarterly Journal of Economics 122(3),
969-1006.

[3] Alm, James, and Leslie A. Whittington. 1995. “Income Taxes and the Marriage
Decision.” Applied Economics 27(1), 25-31.

[4] Alm, James, and Leslie A. Whittington. 1997. “’Til Death or Taxes Do Us Part:
The Effect of Income Taxation on Divorce.” The Journal of Human Resources
32(2), 388-412.

[5] Alm, James, and Leslie A. Whittington. 1999. “For Love or Money? The Impact
of Income Taxes on Marriage.” Economica 66, 297-316.

[6] Altonji, Joseph G., and Lewis M. Segal. 1996. “Small-Sample Bias in GMM Esti-
mation of Covariance Structures.” Journal of Business and Economic Statistics,
14(3), 353–366.

[7] Attanasio, Orazio, Hamish Low, and Virginia Sánchez-Marcos. 2008. “Explaining
Changes in Female Labor Supply in a Life-Cycle Model.” American Economic
Review, 98(4), 1517-1552.

[8] Attanasio, Orazio, Peter Levell, Hamish Low, and Virginia Sánchez-Marcos.
2018. “Aggregating Elasticities: Intensive and Extensive Margins of Women’s
Labor Supply.” Econometrica, 86(6), pp. 2049-2082.

[9] Barro, Robert J., and Charles J. Redlick, 2011. “Macroeconomic Effects of Gov-
ernment Purchases and Taxes.” Quarterly Journal of Economics, February 2011,
51–102.

[10] Benabou, Roland. 2002. “Tax and Education Policy in a Heterogeneous-Agent

Economy: What Levels of Redistribution Maximize Growth and Efficiency?”
Econometrica, 70(2), 481–517.

50
[11] Bick, Alexander, and Nicola Fuchs-Schundeln. 2018. “Taxation and Labour Sup-
ply of Married Couples across Countries: A Macroeconomic Analysis.” Review
of Economic Studies, 85(3), 1543-1576.

[12] Black, Sandra E., Diane Whitmore Schanzenbach, and Audrey Breitwieser. 2017.
“The Recent Decline in Women’s Labor Force Participation ” in The 51%: Driv-
ing Growth through Women’s Economic Participation, October. Brookings.

[13] Blundell, Richard, and Tom MaCurdy. 1999. “Labour Supply: A Review of
Alternative Approaches”in Handbook of Labor Economics, O.Ashenfelter and
D. Card (eds.), North-Holland, December 1559-1695.

[14] Blundell, Richard, Jack Britton, Monica Costa Dias, and Eric French. 2017. “The
impact of health on labour supply near retirement.” Mimeo.

[15] Borella, Margherita, Mariacristina De Nardi, and Eric French. 2017. “Who Re-
ceives Medicaid in Old Age? Rules and Reality.” Fiscal Studies, forthcoming.

[16] Browning, Martin, and Pierre-André Chiappori. 1998. “Efficient Intra-Household

Allocations: A General Characterization and Empirical Tests.” Econometrica,
66(6), 1241-78.

[17] Casanova, Maria. 2012. “Happy Together: A Structural Model of Couples’ Joint
Retirement Choices.” Mimeo.

[18] Chiappori, Pierre-André. 1988. “Rational Household Labor Supply. ” Economet-

rica, 56, 63-89.

[19] Chiappori, Pierre-André. 1992. “Collective Labor Supply and Welfare. ” Journal
of Political Economy, 100(3), 437-67.

[20] De Nardi, Mariacristina, Eric French, and John B. Jones. 2010. “Why Do the
Elderly Save? The Role of Medical Expenses,” Journal of Political Economy,
118, 39-75.

[21] De Nardi, Mariacristina, Eric French, and John B. Jones. 2016. “Medicaid In-
surance in Old Age,” American Economic Review, 106(11), 3480-3520.

[22] De Nardi, Mariacristina, Svetlana Pashchenko, and Ponpoje Porapakkarm. 2017.

“The Lifetime Costs of Bad Health ” NBER working paper 23963.

51
[23] Dickert-Conlin, Stacy and Christian Meghea. 2004. “The Effect of Social Se-
curity on Divorce and Remarriage Behavior,” Working paper no. 9, Center for
Retirement Research.

[24] Dillender, Marcus. 2016. “Social Security and Divorce.” B.E. Journal of Eco-
nomic Analysis and Policy,” 16(2), 931-971.

[25] Dotsey, Michael, Wenli Li, and Fang Yang, 2014, “Consumption and Time Use
over the Life Cycle.” International Economic Review, 55(3), 665-692.

[26] Duffie, Darrell, and Kenneth J. Singleton. 1993. “Simulated Moments Estimation
of Markov Models of Asset Prices.” Econometrica, 61(4), 929–952.

[27] Eckstein, Zvi, and Osnat Liftshitz. 2011. “Dynamic Female Labor Supply.”
Econometrica 79(6), 1675-1726.

[28] Eckstein, Zvi, Michael Keane, and Osnat Liftshitz. 2019. “Career and Family
Decisions: Cohorts Born 1935–1975.” Econometrica, 87(1), 217-253.

[29] Fadlon, Itzik, and Torben H. Nielsen. 2015. “Family Labor Supply Responses to
Severe Health Shocks.” NBER Working Paper 21352.

[30] Fitzgerald, John, Peter Gottschalk, and Robert Moffitt. 1998. “An Analysis of
Sample Attrition in Panel Data: The PSID.” Journal of Human Resources, 33(2),
251-299.

[31] French, Eric. 2005. “The Effects of Health, Wealth, and Wages on Labor Supply
and Retirement Behavior.” Review of Economic Studies, 72(2), 395-427.

[32] Gourinchas, Pierre-Olivier, and Jonathan A. Parker. 2002. “Consumption Over

the Life Cycle.” Econometrica 70(1), 47–89.

[33] Goda, Gopi Shah, John Shoven and Sita Nataraj Slavov. 2007. “Social Security
and the Timing of Divorce,” NBER Working Paper 13382.

[34] Groneck, Max, and Johanna Wallenius. 2017. “It Sucks to Be Single! Marital
Status and Redistribution of Social Security. Working Paper.

[35] Guner, Nezih, Remzi Kaygusuz and Gustavo Ventura. 2012a. “Taxation and
Household Labour Supply.” Review of Economic Studies, 79, 1113-1149.

52
[36] Guner, Nezih, Remzi Kaygusuz and Gustavo Ventura. 2012b. “Income Taxation
of U.S. Households: Basic Facts.” CEPR Discussion Paper 9078.

[37] Heathcote, Jonathan, Kjetil Storesletten and Giovanni L. Violante. (2014). “Con-
sumption and labor supply with partial insurance: An analytical framework.”
The American Economic Review, 104(7), 2075-2126.

[38] Hubbard, R. Glenn, Jonathan Skinner, and Stephen P. Zeldes, 1995. “Precau-
tionary Saving and Social Insurance.” Journal of Political Economy, 103(2), 360-
399.

[39] Hubener, Andreas, Raimond Maurer, and Olivia S. Mitchell. 2016. “How Family
Status and Social Security Claiming Options Shape Optimal Life Cycle Portfo-
lios.”Review of Financial Studies, 29(4), 937-978.

[40] Kaygusuz, Remzi. 2015. “Social Security and Two-Earner Households.”Journal

of Economic Dynamics and Control, 59, 163-178

[41] Liebman, Jeffrey B., Erzo F.P. Luttmer, and David G. Seif. 2009. “Labor Supply
Responses to Marginal Social Security Benefits: Evidence from Discontinuity.”
Journal of Public Economics, 93(11-12), 1208-1223.

[42] Low, Hamish, Costas Meghir, Luigi Pistaferri, and Alessandra Voena. 2018.
“Marriage, Labor Supply and the Social Safety Net.” Mimeo.

[43] Moffitt, Robert. 1998. Welfare, The Family, And Reproductive Behavior, Re-
search Perspectives, ed. by Robert A. Moffitt. National Academies Press. Wash-
ington D.C.

[44] Moffitt, Robert, and Sisi Zhang. 2018. “Income Volatility and the PSID: Past
Research and New Results.” AEA Papers and Proceedings, 108, 277-280.

[45] National Research Council, 1995, Measuring poverty: a new approach, Washing-
ton, DC: The National Academies Press.

[46] Nishiyama, Shinichi. 2017. “The Joint Labor Supply Decision of Married Couples
and the Social Security Pension System.” Review of Economic Dynamics, 31,
277-304.

53
[47] Pakes, Ariel, and David Pollard. 1989. “Simulation and the Asymptotics of Op-
timization Estimators.” Econometrica, 57(5), 1027–1057.

[48] Persson, Petra. 2017. “Social Insurance and the Marriage Market.” Mimeo.

[49] Roys, Nicolas and Christopher Taber. 2017. “Skills Prices, Occupations and
Changes in the Wage Structure for Low Skilled Men.” Mimeo.

[50] Sorrentino, Constance. 1990. “The Changing Family in International Perspec-

tive.” Monthly Labor Review, 113(3), 41-58.

[51] United Nations, Department of Economic and Social Affairs Population Di-
vision, Fertility and Family Planning Section. 2012. “World Marriage Data,”
http://www.un.org/esa/population/publications/WMD2012/MainFrame.html.

54
Appendix A. Data: The PSID and the HRS
We use the Panel Study of Income Dynamics (PSID) to estimate the wage pro-
cess, the marriage and divorce probabilities, the initial distribution of couples and
singles over state variables, taxes, and the sample moments that we match using our
structural model.
The PSID is a longitudinal study of a representative sample of the U.S. population.
The original 1968 PSID sample was drawn from a nationally representative sample of
2,930 families designed by the Survey Research Center at the University of Michigan
(referred to here as the SRC sample), and from an over-sample of 1,872 low-income
families from the Survey of Economic Opportunity (referred to here as the SEO
sample). Individuals have been followed over time to maintain a representative sample
of families.
We study the two cohorts born in the periods 1941-1945 and 1951-1955. More
specifically, we select all individuals in the SRC sample who are interviewed at least
twice in the sample years 1968-2013, select only heads and their wives, if present,
and keep individuals born between 1931 and 1955. The resulting sample includes
5,129 individuals ages 20 to 70, for a total of 103,420 observations. In general, to
gather the information we need, we control for birth cohort effects in our estimates
(we use 5-year-of-birth windows) and use the results relative to the cohorts of interest.
Table 10 details our PSID sample selection.

Selection Individuals Observations

Initial sample (observed at least twice) 30,587 893,420
Heads and wives (if present) 18,304 247,203
Born between 1931 and 1955 5,153 105,985
Age between 20 and 70 5,129 103,420

Table 10: Sample selection in the PSID.

Table 11 shows that the majority of men and women are married and that the
fraction of married people goes down only slightly across these cohorts.
We use the Health and Retirement Study (HRS) to compute inputs for the re-
tirement period because this data set contains a large number of observations and
high-quality data for this stage of the life cycle. In fact, the HRS is a longitudinal
data set collecting information on people ages 50 and older, including a wide range

55
 Born in 1941-1945 Born in 1951-1955
Gender Age 25 Age 40 Age 55 Age 25 Age 40 Age 55
Men 0.87 0.90 0.88 0.82 0.86 0.84
Women 0.86 0.84 0.79 0.83 0.81 0.76

Table 11: Fraction of married men and women by age and cohort, PSID data.

of demographic, economic, and social characteristics, as well as physical and mental

health, and cognitive functioning.
The HRS started collecting information in 1992 on individuals born between 1931
and 1941, the so-called initial HRS cohort, which was then reinterviewed every two
years. Other cohorts were introduced over the years. Our data set is based on
the RAND HRS files for the period 1995-2012, to which we add the EXIT files to
include information on the wave right after death. Our sample selection is as follows.
Of the 37,317 individuals initially present, we drop individuals for whom marital
status is not observed (1,548 individuals). This yields 35,769 individuals and 185,255
observations. We then select individuals in the age range 66-100 born in 1900 to
1945, obtaining a sample of 16,118 individuals and 81,246 observations. As we cannot
observe individuals born after 1945 and older than age 66 in the HRS, for the 1955
cohort we use the same estimates obtained for the 1945 one.

Appendix B. First step estimation, methodology

Table 12 summarizes our estimated model inputs.

Wages
Because we allow our initial conditions, assortative matching in marriage, and
marriage and divorce probabilities to depend on the realized values of wage shocks,
we need to estimate not only wages as a function of human capital, age, and gender,
and the stochastic process for the wage shocks, but also the realized wage shocks for
all men and women of working age in our sample (whether they are working or not).
To do so, we proceed as follows. First, we impute potential wages for individuals
who are not working, so that we are able to construct potential wages as actual
wages for participants and potential wages for non-participants. Second, we estimate
potential wages as a function of age, gender, and human capital. Third, we estimate

56
 Estimated processes Source
Wages
ei,j
t (·) Endogenous age-efficiency profiles PSID
i
t Wage shocks PSID
Demographics
si,j
t (ψt )
i Survival probability HRS
ζt (·) Divorce probability PSID
νt (·) Probability of getting married PSID
ξt (·) Matching probability PSID
θt (·) Partner’s assets and earnings PSID
f 0,5 (i, j, t) Number of children ages 0-5 PSID
f 6,11 (i, j, t) Number of children ages 6-11 PSID
Health shock
mi,j
t (ψt )
i Medical expenses HRS
i,j i
πt (ψt ) Transition matrix for health status HRS
Government policy
λjt , τtj Income tax See text

Table 12: First-step estimated inputs summary.

the persistence and variance of its unobserved component and the realized wage shocks
using Kalman filtering.
Missing wages imputation. The observed wage rate is computed as annual
earnings divided by annual hours worked. Gross annual earnings are defined as labor
income during the previous year. Annual hours are given by annual hours spent
working for pay during the previous year.17
We impute missing wages by using coefficients from fixed effects regressions that
we run separately for men and women. To avoid endpoint problems with the polyno-
mials in age, we include individuals ages 22 to 70 in the sample. Define the observed
n ^
wage for labor market participants as ln wagekt = Ikt ln wagekt , where k denotes an
n
individual and t is age. The term Ikt is an indicator for participation (which is equal
to 1 if the individual participates in the labor market and has no missing hours or
earnings) and ln^ wage is the potential wage that we wish to estimate and do not ob-
serve. We estimate the unobserved potential wage by running the following regression
17
Wages may be missing both because an individual has not been active in the labor market
and because (s)he may have been active, but earnings or hours (or both) are missing. In addition,
because estimated variances are very sensitive to outliers, we set to missing observations with an
hourly wage rate below half the minimum wage and above $368 (in 2016 values). We use the same
imputation procedure for all these cases.

57
 0
on observables as ln wagekt = Zkt βz + fk + ςkt , where the dependent variable is the
logarithm of the observed hourly wage rate, fk is an individual-specific fixed effect,
and ςkt is an error term. We include a rich set of explanatory variables in Zkt : a
fifth-order polynomial in age, a third-order polynomial in experience (measured in
years of labor market participation), marital status (a dummy for being single), fam-
ily size (dummies for each value), number of children (dummies for each value), age
of youngest child, and an indicator of partner working if married. As an indicator of
health, we use a variable recording whether bad health limits the capacity of working
(this is the only health indicator available in the PSID for all years). Because this
health indicator is not collected for wives, we do not include it in the regression for
married women. Both regressions also include interaction terms between the explana-
tory variables. Variables that do not vary over time are captured by the individual
effect fk .
Using the estimated coefficients, we take the predicted value of the wage to be the
potential wage for observations with missing wages. Hence, we define potential wage
as 
ln wage n
kt if Ikt =1
ln wagekt =
Z 0 β̂ + fˆ if I n = 0
kt z k kt

Wage function estimation. We model wages as a function of human capital,

age, and gender, and we measure human capital as average realized earnings accrued
up to the beginning of age t (ȳt ).
To estimate the wage profiles, we proceed in two stages. First, we run the following
fixed effect regression for the logarithm of potential wages

G
X
i
ln wagekt = dk + f (t) + βg Dg ln(ȳkt + δy ) + ukt , (41)
g=1

on a gender-specific fifth-order polynomial in age f i (t), gender-cohort cells g, and

gender-cohort dummies Dg .18 The shifter δy is set equal to $5,000 to avoid taking
the logarithm of values that are too small.19 We also experimented by adding marital
18
Instead of following our general methodology of defining 5-year-of-birth cohorts, to estimate the
cohort-specific effect of human capital on wages in equation (41), we take two broader windows: the
1940s cohort includes the generations born in 1931-1945, while the 1950s cohort includes those born
in 1946-1955. We do so because we do not observe the complete age profile for the wages of the 1955
cohort.
19
While we use earnings subject to the Social Security cap to compute average earnings (the state

58
status dummies to capture the effect of changing marital status on wages, but they did
not turn out to be statistically different from zero, conditional on average earnings.
Second, to fix the constant of the wage profile for our cohorts of interest, we regress
the sum of the residuals and fixed effects dk + ukt+1 ≡ wkt+1 on cohort dummies
to compute the average effects for the cohorts born in 1941-1945 and in 1951-1955,
respectively. Table 13 reports the coefficients of the estimated equation from the first
stage, the fixed effect regression, while Table 14 reports those from the second stage.

Coefficient Standard Error

ln(ȳt + δy ) 0.307*** (0.0216)
ln(ȳt + δy )*female 0.0419 (0.0277)
ln(ȳt + δy )*born in 1950s 0.118*** (0.0265)
ln(ȳt + δy )*born in 1950s*female -0.0398 (0.0334)
Age -0.567*** (0.177)
Age2 /(102 ) 2.679*** (0.861)
Age3 /(104 ) -6.135*** (2.029)
Age4 /(106 ) 6.908*** (2.321)
Age5 /(108 ) -3.095*** (1.033)
Age*female 0.343 (0.220)
Age2 /(102 )*female -1.695 (1.070)
Age3 /(104 )*female 3.955 (2.526)
Age4 /(106 )*female -4.433 (2.895)
Age5 /(108 )*female 1.947 (1.291)
Constant 2.044** (0.851)
N 93363
R-sq 0.103

Table 13: Coefficients from fixed effects estimates. Dependent variable: logarithm of the
potential wage. PSID data. Robust standard errors in parentheses, clustered
at the individual level. * p<0.10, ** p<0.05, *** p<0.01

The estimated potential wage profiles, computed at average values of ln(ȳt ), are
shown in the main text. The shock in log wage is modeled as the sum of a persistent
component plus white noise, which we assume captures measurement error:

w̃kt+1 = ln kt+1 + ξkt+1 (42)

ln kt+1 = ρ ln kt + vkt+1 , (43)

where w̃kt+1 are the residuals from the second stage, and ξkt+1 and vkt+1 are inde-
pendent white-noise processes with zero mean and variances equal to σξ2 and σv2 ,
variable in our model), estimating this wage regression by using uncapped previous average earnings
yields very similar estimates.

59
 Men Women
Born in 1931-35 0.0178 -0.0537*
(0.0406) (0.0318)
Born in 1936-40 -0.00663 -0.0537*
(0.0385) (0.0319)
Born in 1946-50 -1.265*** -0.776***
(0.0277) (0.0239)
Born in 1951-55 -1.314*** -0.795***
(0.0273) (0.0227)
Constant 2.227*** -0.953***
(0.0231) (0.0192)
N 45366 47996
R-sq 0.584 0.382

Table 14: Second stage: coefficients from OLS estimates. Dependent variable: residuals
from fixed effects estimates. PSID data. Robust standard errors in parentheses,
clustered at the individual level. * p<0.10, ** p<0.05, *** p<0.01

respectively. We estimate these process separately for each gender.20

To estimate the realized value of the wage shocks, we estimate the system com-
posed by equations (42) and (43) by maximum likelihood, which can be constructed
assuming that the initial state of the system and the shocks are Gaussian, and using
standard Kalman filter recursions. With that, we can estimate both the parameters
in (42) and (43) and the entire state, that is, ln kt , t = 1, ...T .
Table 15 reports our estimates for the AR component of earnings.

Parameter Men Women

Persistence 0.947 0.945
Variance prod. shock 0.026 0.016
Initial variance 0.112 0.098

Table 15: Estimated processes for the wage shocks for men and women, PSID data.

Average realized earnings and accumulated Social Security

contributions
In the model we keep track of average accumulated earnings for a person (ȳkt )
subject to Social Security cap that is applied to yearly earnings and is time varying. To
20
For this, we limit the age range between 25 and 65 and, because we rely on residuals also taken
from imputed wages, we drop the highest 0.5% residuals for both men and women. This avoids
large outliers to inflate the estimated variances (however, the effect of this drop on our estimates is
negligible).

60
do so, we assume that individuals start working at age 22, and we compute individual-
level capped average earnings. This computation requires taking a stand on people
who appear in our data after age 22. Some individuals (5%) enter the sample after
turning 22 either because in 1968, the first year the PSID was collected, they were
older or because they entered as spouses or descendants and might thus be older than
22. Among people in this group, 46 enter the sample before turning 27: for those
individuals we assume average accumulated earnings at entry are equal to zero. For
the remaining 189 individuals, we use an imputation procedure to recover average
realized earnings at entry and then update the value following each individual over
time. We run a regression of capped earnings on a fourth-order polynomial in age fully
interacted with gender, education dummies, interactions of education and gender,
marital status, and race dummies also interacted with gender. Cohort dummies are
also included. We use the predicted values of this regression as the entry value for
individuals entering the sample after turning 27. Average earnings are then updated
for each individual following his/her observed earnings history (as done in the model).
For the purposes of imputing missing values of wealth, we also compute uncapped
average realized earnings using the same methodology for missing values of accumu-
lated earnings at entry as above.

Wealth
We define wealth as total assets (defined as all asset types available in the PSID)
plus home equity. Wealth in the PSID is only recorded in 1984, 1989, 1994, and then
in each (biennial) wave from 1999 onward. We rely on an imputation procedure to
compute wealth in the missing years, starting in 1968. This imputation is based on
the following fixed effect regression:

0
ln(akt + δa ) = Zkt βz + dak + wakt , (44)

where k denotes the individual and t is age. The parameter δa is a shifter for assets
to have only positive values and to be able to take logs, and the variables Z includes
polynomials in age, also interacted with health status, and with average earnings
(uncapped), family size, and a dummy for health status. The term dak is the indi-
vidual fixed effect and wakt is a white-noise error term. Equation (44) is estimated
separately for single men, single women, and couples, as wealth is measured at the

61
household level, on an enlarged sample of individuals born between 1931 and 1965.
We then use the imputed as well as the actual observations to estimate the wealth
profiles used as target moments and to parameterize the joint distribution of initial
assets, average realized earnings, and wage shocks for single men, single women, and
couples.

Distributions upon entering the model and for prospective

spouses
For single men and women, separately, we parameterize the joint distribution
of initial assets, average realized earnings, and wage shocks at each age as a joint
lognormal distribution:
   i 
ln(ait + δai ) µat + δai
i 
 ln(ȳti )  ∼ N  µiȳt , Σst  , (45)
  

ln it µit

where Σs is a 3x3 covariance matrix. We estimate its mean and variance as a function
of age t. For the mean, we regress the logarithm of assets plus shift parameter, average
earnings, and productivity shock ln ˆit on a third-order polynomial in age and cohort
dummies. The predicted age profile, relative to cohorts born in 1945 and 1955, is
the age-specific estimate of the mean of the lognormal distribution. Taking residuals
from the above estimates, we can estimate the elements of the variance-covariance
matrix by computing the relevant squares or cross-products. We regress the squares
or the cross-products of the residuals on a third-order polynomial in age to obtain,
element by element, a smooth estimate of the variance-covariance matrix at each age.
For couples, we compute the initial joint distribution at age 25 of the following
variables:
   
ln(a + δa ) µ a + δa
 ln(ȳ 1 )   µȳ1
   

   
 ln(ȳ 2 )  ∼ N  µȳ2 , Σc  , (46)
   
1
 ln( )   µ1
   

2
ln( ) µ2
where Σc is a 5x5 covariance matrix computed on the data for married or cohabiting

62
couples.

Marriage and divorce probabilities

We model the probability of getting married, νt+1 , as a function of gender, age
and the wage shock and perform the estimation separately for men and women using
PSID data. Our estimated equation is

i
νt+1 = P rob(M arriedt+1 = 1|M arriedt = 0, Zt ) = F (Zt0 βm ),

where F denotes the standard logistic distribution and Zt includes a polynomial in

age, cohort dummies, the logarithm of the wage shock, and the after-1997 dummy21 .
Using the estimated coefficients on the cohort dummies, we then adjust the probability
for the 1945 and 1955 cohorts, respectively.
Similarly, we estimate the probability of divorce as

ζt = P rob(Divorcedt+1 = 1|M arriedt = 1, Zt ) = F (Zt0 βd ),

where F denotes the standard logistic distribution and Zt includes a polynomial in

age, husband’s wage shock, wife’s wage shock, cohort dummies, and an indicator for
biennial waves.
Men Women Men Women
0.25 0.25 0.25 0.25
Lowest Lowest Lowest Lowest
2nd 2nd 2nd 2nd
3rd 3rd 3rd 3rd
0.2 4th 0.2 4th 0.2 4th 0.2 4th
Highest Highest Highest Highest
Prob. of marriage

Prob. of marriage

Prob. of marriage

Prob. of marriage

0.15 0.15 0.15 0.15

0.1 0.1 0.1 0.1

0.05 0.05 0.05 0.05

30 40 50 60 30 40 50 60 30 40 50 60 30 40 50 60
Age Age Age Age

Figure 14: Marriage probabilities by gender, age, and one’s wage shock for the 1945 cohort
(left panel) and 1955 cohort (right panel), PSID data.

21
The PSID goes from a yearly to a biennial frequency in 1997. To take this into account, we
include an indicator variable taking a value of one from 1997 on in the regression, which we then
abstract from when constructing the yearly probabilities.

63
 Single Men Single Women Couples
Marriage Marriage Divorce
Age 0.0386 -0.0424 0.0469
(0.0310) (0.0324) (0.0294)
Age2 /102 -0.109*** -0.0324 -0.114***
(0.0404) (0.0415) (0.0368)
ln kt 0.308** 0.0578 -0.399***
(0.131) (0.143) (0.101)
Spouse’s ln kt -0.116
(0.120)
I(year > 1997) 0.588*** 0.378* 0.505***
(0.208) (0.208) (0.167)
Born in 1931-35 0.144 -0.329 -0.259
(0.269) (0.229) (0.173)
Born in 1936-40 -0.0219 -0.586** -0.115
(0.238) (0.238) (0.141)
Born in 1946-50 -0.198 -0.292* -0.122
(0.181) (0.160) (0.111)
Born in 1951-55 -0.364** -0.352** -0.0762
(0.174) (0.156) (0.118)
Constant -1.545** -0.135 -3.537***
(0.606) (0.634) (0.574)
N 5064 8860 32071
pseudo-R2 0.026 0.060 0.023

Table 16: Estimated coefficients from logistic regressions. Column 1: Marriage of single
men; column 2: marriage of single women; column 3: divorce of couples. PSID
data. Robust standard errors in parentheses, clustered at the individual level.
* p<0.10, ** p<0.05, *** p<0.01

Men Women Men Women

0.06 0.06 0.06 0.06
Lowest Lowest Lowest Lowest
2nd 2nd 2nd 2nd
0.05 3rd 0.05 3rd 0.05 3rd 0.05 3rd
4th 4th 4th 4th
Highest Highest Highest Highest
0.04 0.04 0.04 0.04
Prob. of divorce

Prob. of divorce

Prob. of divorce

Prob. of divorce

0.03 0.03 0.03 0.03

0.02 0.02 0.02 0.02

0.01 0.01 0.01 0.01

0 0 0 0
30 40 50 60 30 40 50 60 30 40 50 60 30 40 50 60
Age Age Age Age

Figure 15: Divorce probabilities by gender, age, and one’s wage shock for the 1945 cohort
(left panel) and 1955 cohort (right panel), PSID data.

Figures 14 and 15 report the resulting marriage and divorce probabilities for both
cohorts.
Conditional on meeting a partner, the probability of meeting a partner p with
wage shock pt+1 is ξt+1 (·) = ξt+1 (pt+1 |it+1 , i). Using our estimated wage shocks and
partitioning households in age groups (25-35, 35-45, 45-65), we compute the variance-
covariance matrix of newly matched partners’ wage shocks by age groups. We then

64
derive the conditional distribution of meeting a partner assuming lognormality. In
the whole sample we observe only 750 new marriages in the age range 25-65, therefore
we do not allow this probability to depend on cohort.

Number of children
To compute the average number of children by age group, we use the individ-
ual information in the PSID and classify as children of the family in the following
categories: sons or daughters of the head, stepsons or stepdaughters of the head,
sons or daughters of the cohabiting partner but not of the head, foster sons or foster
daughters (not legally adopted), and children of the first-year cohabitor but not of
the head. Having done that, we add up the number of children in each age category
(0 to 5, 6 to 11, or 0 to 17 for the total number of children) and run a regression
on a fifth-order polynomial in age of the mother, interacted with marital status, and
cohort dummies to construct the average age profile of children in each age group for
single and married women. We use the profiles for the cohorts of mothers born in
1941-1945 and in 1951-1955.

Children ages 0-17 Children ages 0-17 Children ages 0-17 Children ages 0-17
2.5 Children ages 0-5 2.5 Children ages 0-5 2.5 Children ages 0-5 2.5 Children ages 0-5
Children ages 6-11 Children ages 6-11 Children ages 6-11 Children ages 6-11

2 2 2 2

1.5 1.5 1.5 1.5

1 1 1 1

0.5 0.5 0.5 0.5

0 0 0 0
30 40 50 60 30 40 50 60 30 40 50 60 30 40 50 60
Age of married woman Age of single woman Age of married woman Age of single woman

Figure 16: Number of children for married and single women for the 1945 cohort (left
panel) and 1955 cohort (right panel), PSID data.

Health status at retirement

We define health status on the basis of self-reported health. In the HRS, this
variable can take five possible values (excellent, very good, good, fair, poor). As
standard, we take health to be a dichotomous variable equal to 1 if self-reported

65
health is fair or poor and 0 otherwise.22 We estimate the probability of being in bad
health at age 66, using the observed frequencies for the 1941-1945 cohort, which is
the youngest cohort that we can observe at age 66+ in the HRS data. All the inputs
estimated from the HRS correspond to the 1941-1945 cohort. For lack of better data,
we also use them for our 1951-1955 cohort. For singles, we compute the sample
fraction of single men and single women in bad health in the age range 65-67, which
ensures that the sample size is big enough. For couples, we define the first member
in the couple as the husband and the second as the wife, and compute the sample
frequencies for the four possible health states in the couple as (good, good), (good,
bad), (bad, good), and (bad, bad).

Health dynamics after retirement

As before, we use the HRS data, and we define the health status variable ψ equal
to 1 if self-reported health at time t is equal to “fair” or “bad” and 0 otherwise. We
model the probability of being in bad health during retirement as a logit function:

exp(Xtψ0 β ψ )
πψt = P rob(ψt = 1 | Xtψ ) = ,
1 + exp(Xtψ0 β ψ )

which we then use to construct the transition matrix at each age, gender, and marital
status. The set of explanatory variables Xtψ includes cohort dummies, a second-order
polynomial in age, previous health status, gender, marital status, and interactions be-
tween these variables when they are statistically different from zero. As the HRS data
are collected every two years, we obtain two-year probabilities and convert them into
one-year probabilities. Table 17 reports our estimated coefficients, while Figure 17
displays the health transition matrix by gender, age, marital status, and health status
that we estimated.
22
Looking at labor supply behavior around retirement time, Blundell et al. (2017) show that this
measure of self-reported health captures health well and about as well as more involved measures
such as using large numbers of objective measures to predict health.

66
 Singles Couples
1 1

Prob. of staying in the same health

Prob. of staying in the same health

0.9 0.9

0.8 0.8

0.7 0.7

0.6 Men bad health 0.6 Men bad health

Men good health Men good health
Women bad health Women bad health
Women good health Women good health
0.5 0.5
70 80 90 70 80 90
Age Age

Figure 17: Health transition probabilities for singles and couples by age. HRS data.

Coefficient SE
Age 0.0934*** (0.0305)
Age2 /102 -0.0463** (0.0190)
Healtht−1 6.754*** (0.265)
Healtht−1 *Age -0.0547*** (0.00335)
Male 0.476** (0.242)
Male*Age -0.00551* (0.00313)
Age*Married -0.0262*** (0.00329)
Age2 /102 *Married 0.0307*** (0.00420)
Born in 1936-40 0.166*** (0.0610)
Born in 1931-35 0.221*** (0.0609)
Born in 1926-30 0.349*** (0.0698)
Born in 1921-25 0.392*** (0.0692)
Born in 1916-20 0.519*** (0.0720)
Born in 1900-15 0.677*** (0.0788)
Constant -6.200*** (1.215)
N 58547
Pseudo-R2 0.236

Table 17: Health dynamics over two-year periods. Logistic regression coefficients, depen-
dent variable: health status. HRS data. Robust standard errors in parentheses,
clustered at the individual level. * p<0.1, ** p<0.05, *** p<0.01

Survival probabilities

We model the probability of being alive at time t as a logit function:

exp(Xts0 β s )
st = P rob(Alivet = 1 | Xts ) = .
1 + exp(Xts0 β s )

which we estimate using the HRS data. Among the explanatory variables, we include
a fourth-order polynomial in age, gender, marital status, and health status in the
previous wave, as well as interactions between these variables and age, whenever they
are statistically different from zero. As the HRS is collected every two years, we
transform the biennial probability of surviving into an annual probability by taking

67
the square root of the biennial probability. Table 18 reports estimated coefficients,
and Figure 18 displays the implied survival probability by age, gender, and marital
and health status.
Coefficient SE
Age -19.03*** (6.826)
Age2 /102 34.66*** (12.53)
Age3 /104 -27.96*** (10.17)
Age4 /106 8.377*** (3.081)
Healtht−1 -3.816*** (0.307)
Healtht−1 *age 0.0313*** (0.00370)
Male -1.213*** (0.332)
Male*Age 0.00836** (0.00405)
Married 1.302*** (0.375)
Married*Age -0.0128*** (0.00465)
Born in 1936-40 0.161 (0.129)
Born in 1931-35 0.0817 (0.128)
Born in 1926-30 -0.00885 (0.138)
Born in 1921-25 0.0434 (0.139)
Born in 1916-20 0.0363 (0.142)
Born in 1900-15 0.0644 (0.145)
Constant 395.6*** (138.8)
N 63746
Pseudo-R2 0.171

Table 18: Logistic regression coefficients, dependent variable: survival over a two-year
period. HRS data. Robust standard errors in parentheses, clustered at the
individual level. * p<0.1, ** p<0.05, *** p<0.01

Singles Couples
1 1

0.95 0.95

0.9 0.9

0.85 0.85
survival probability

0.8 0.8

0.75 0.75

0.7 0.7

0.65 0.65

0.6 Men bad health 0.6 Men bad health

Men good health Men good health
0.55 Women bad health 0.55 Women bad health
Women good health Women good health
0.5 0.5
70 80 90 70 80 90
age age

Figure 18: Survival probability by age, gender, and marital and health status, both co-
horts. HRS data.

Out-of-pocket medical expenditures

Out-of-pocket medical expenses are defined as the total amount that the individual
spends out of pocket in hospital and nursing home stays, doctor visits, dental costs,

68
outpatient surgery, average monthly prescription drug costs, home health care, and
special facilities charges. They also include medical expenses in the last year of life,
as recorded in the exit interviews. In contrast, expenses covered by public or private
insurance are not included in our measure, as they are not directly incurred by the
individual. The estimated equation is

m0 m
ln(mkt ) = Xkt β + αkm + um
kt ,

where explanatory variables include a fourth-order polynomial in age fully interacted

with gender and current health status, and we include these interactions whenever
they are statistically different from zero. We estimate the equation on the HRS data
using a fixed effects estimator, which takes into account all unmeasured fixed-over-
time characteristics that may bias the age profile, such as differential mortality (as
discussed in De Nardi, French, and Jones (2010)). Marital status (also interacted
with other variables) does not turn out to be significantly different from zero in the
first step. We then regress the residuals and fixed effects from this equation on cohort,
gender, and marital status dummies to compute the average effect for each group of
interest. Table 19 reports estimated coefficients, while Figure 19 displays medical
expenditure by age, gender, and marital and health status.
Finally, we model the variance of the shocks by regressing the squared residuals
from the regression in logs on a third-order polynomial in age fully interacted with
gender and current health status, and on cohort, gender, and marital status dummies
and use it to construct average medical expenses as a function of age by adding half
of the variance to the average in logs before exponentiating.

104 Singles 104 Couples

2.5 2.5
Men bad health Men bad health
Men good health Men good health
Women bad health Women bad health
Women good health Women good health
2 2
Determinstic health cost in 2016$

Determinstic health cost in 2016$

1.5 1.5

1 1

0.5 0.5

0 0
70 80 90 70 80 90
Age Age

Figure 19: Medical expenditure by age, gender, and marital and health status. HRS data.

69
 Coefficient SE
Age 9.770*** (2.416)
Age2 /102 -18.63*** (4.612)
Age3 /104 15.68*** (3.893)
Age4 /106 -4.901*** (1.226)
Bad health 3.819*** (1.012)
Bad health*Age -0.0961*** (0.0263)
Bad health*Age2 /102 0.0624*** (0.0169)
Male*Age -9.160*** (3.793)
Male*Age2 /102 17.76*** (7.261)
Male*Age3 /104 -15.14*** (6.147)
Male*Age4 /106 4.792*** (1.942)
Constant -109.9*** (36.32)
Second stage
Male 174.9*** (0.0263)
Married 0.330*** (0.0195)
Male*Married -0.0469** (0.0316)
Born in 1936-40 -0.000573 (0.0302)
Born in 1931-35 -0.0534** (0.0296)
Born in 1926-30 -0.118*** (0.0358)
Born in 1921-25 -0.0954*** (0.0338)
Born in 1916-20 -0.143*** (0.0344)
Born in 1900-15 -0.309*** (0.0378)
Constant -73.77*** (0.0287)
N 65917
R2 first stage 0.017
R2 second stage 0.99

Table 19: Estimates for the logarithm of medical expenses, first stage (fixed effects) and
second stage (OLS). HRS data. Robust standard errors in parentheses, clustered
at the individual level. * p<0.1, ** p<0.05, *** p<0.01

Spousal assets and Social Security benefits

We assume random matching over asset and lifetime income of the partner condi-
tional on partner’s wage shock. Thus, we compute θt+1 (·) = θt+1 (apt+1 , ȳt+1
p
|pt+1 ) using
sample values of assets, average capped earnings, and wage shocks. More specifically,
we assume θt+1 is lognormally distributed at each age with mean and variance com-
puted from sample values. Assets include a shifter as described for the computation
of the joint distribution at age 25 (see Wealth subsection in this Appendix).
Figure 20 reports spousal assets by spousal wage shocks in case of marriage next
period. Both panels show that both women and men marrying early on in life expect
their partner to have relatively low assets on average, even conditional on the various
wage shocks. In contrast, those who marry later experience a much larger variation
in their partner’s assets conditional on their partner’s wage shocks. The gradient in
average assets by wage shocks increases especially fast for male partners and thus
exposes women to much more variability in their partner’s resources as they marry

70
 #10 5 Men #10 5 Women #10 5 Men #10 5 Women

Mean assets at marriage conditional on productivity

Mean assets at marriage conditional on productivity

Mean assets at marriage conditional on productivity

Mean assets at marriage conditional on productivity

4 4 4 4
Lowest Lowest Lowest Lowest
3.5 2nd 3.5 2nd 3.5 2nd 3.5 2nd
3rd 3rd 3rd 3rd
4th 4th 4th 4th
3 Highest 3 Highest 3 Highest 3 Highest

2.5 2.5 2.5 2.5

2 2 2 2

1.5 1.5 1.5 1.5

1 1 1 1

0.5 0.5 0.5 0.5

0 0 0 0
30 40 50 60 30 40 50 60 30 40 50 60 30 40 50 60
Age Age Age Age

Figure 20: Spousal assets by spousal wage shocks in case of marriage next period for the
1945 cohort (left panel) and 1955 cohort (right panel), PSID data.

later and later. The patterns are very close for the two cohorts.
#10 4 Men #10 4 Women #10 4 Men #10 4 Women
7 7 7 7
Lowest Lowest
Average ybar conditional on productivity

Average ybar conditional on productivity

Average ybar conditional on productivity

Average ybar conditional on productivity

2nd 2nd
6 6 6 6
3rd 3rd
4th 4th
Highest Highest
5 5 5 5

4 4 4 4

3 3 3 3

2 Lowest 2 2 Lowest 2
2nd 2nd
3rd 3rd
1 1 1 1
4th 4th
Highest Highest
0 0 0 0
30 40 50 60 30 40 50 60 30 40 50 60 30 40 50 60
Age Age Age Age

Figure 21: Spousal Social Security earnings by spousal wage shocks in case of marriage
next period for the 1945 cohort (left panel) and 1955 cohort (right panel),
PSID data.

Figure 21 reports spousal Social Security earnings by spousal wage shocks in case
of marriage next period. Given that male wage shocks are higher on average, Social
Security earnings for men are higher than those for women at all levels of the wage
shocks.

Taxes
We model taxes T on total income Y as T (Y ) = Y − λY 1−τ , where τ captures
the degree of progressivity and λ captures the average level of taxation in the system.
Since this specification implies (Y − T (Y )) = λY 1−τ and ln(Y − T (Y )) = ln(λ) +
(1 − τ )ln(Y ), we estimate τ and λ by regressing the logarithm of after-tax household
income on a constant and on the logarithm of pre-tax household income by cohort,

71
year, and household type (single man, single woman, couple).
We use PSID data from 1968 to 2015 (tax years 1967-2014) to estimate cohort-
and time-specific tax functions. Information about federal taxes paid is provided
directly by the PSID up to 1991. After that year it is gathered using TAXSIM, the
NBER simulation program computing taxes. In particular, we build on and extend
the program written by Kimberlin et al. (2015) to prepare the input needed by
TAXSIM.23
Before-tax household income is defined as the sum of all money income received by
the spouses (or by the individual if single) in a given tax year. It therefore includes
income of the head and the wife (if present), that is, labor income, the asset part
of income from farm, business, roomers, and so on, plus income from rent, interest
dividends, and so on, and wife’s income from assets, plus transfer income, that is,
Social Security, pension, annuities, other retirement income, welfare, aid to dependent
children, unemployment or worker’s compensation, help from relatives, alimony, or
child support. After-tax household income is defined as before-tax income minus
the federal income tax liability (including capital gains rates, surtaxes, AMT, and
refundable and non-refundable credits, as computed by TAXSIM).
To keep the number of observations large while keeping different tax regimes
separate, we follow a slightly different procedure for couples and singles. For couples,
we define two five-year cohorts (one born in 1941-1945, one in 1951-1955) and estimate
the tax functions over two- or three-year intervals. For single men and women, the
1945 cohort includes individuals born in 1938-1947, while the 1955 one includes those
born in 1948-1957. Then, we estimate yearly tax functions, using data relative to
a moving five-year window for each function, to have enough observations and to
capture relevant changes in the legislation.
All the inputs needed by TAXSIM are gathered directly from the PSID, for the
sample years 1992-2015. However, for years prior to 1999, medical expenses and
charitable contributions are not available and need to be imputed, as they may be
deducted from gross income (if the household chooses to itemize). Hence, we impute
23
The program by Kimberlin et al. (2015) prepares the input for TAXSIM for the PSID years
1999-2011 following Butrica and Burkhauser (1997). It differs from more simplified PSID TAXSIM
interface approaches in that multiple tax units are identified within each PSID family unit; thus,
cohabiting couples are treated as two separate tax units, with children assigned to the appropriate
tax unit (head or cohabitor) using relationship codes. We extend their program to include all years
between 1992 and 2015.

72
them by regressing the sum of the two items for pooled years 1999-2015 and predicting
the value using the estimated parameters (out-of-sample prediction). The included
explanatory variables are demographic and income variables, such as family size,
employment status of the head and the spouse if present, state of residence, wages,
pensions, other incomes, education, number of children, age, and marital status.
Then, we add an error term to that prediction to tackle the attenuation in the variance
of the distribution of the imputed values, following the procedure in David et al.
(1986), and French and Jones (2011). More precisely, the procedure is as follows.
First, we regress the sum of the two items on the vector of observables for the sample
of heads who choose to itemize, deduci = zi β + i . Second, for each household i
\ i = zi β̂ and the
for which deduc is observed, we calculate the predicted value deduc
residual êi = deduci − deduc
\ i . Third, we sort the predicted value deduc
\ i into deciles
and keep track of all values of êi within each decile. Next, for every individual j
\ j = zj β̂. Then we impute êj for households with
with missing deduc we impute deduc
missing deduc by finding a random individual i in the non-missing sample with a
\ i in the same decile as deduc
value of deduc \ j and set êj = êi . The imputed value of
\ j + êj .
deduc is deduc

Appendix C. Calibrated model parameters

Table 20 summarizes our first-step calibrated model inputs. We set the interest
rate r to 4% and the utility curvature parameter, γ, to 2.5. The equivalence scales
are set to ηti,j = (j + 0.7 ∗ fti,j )0.7 , as estimated by National Research Council (1995).
With the equivalence scale, $1 spending in a household of a couple with two kids gives
each household member a consumption of 0.42 cents. The term fti,j is the average
total number of children for single and married men and women by age.
The most recent paper estimating the consumption floor during retirement is the
one estimated by De Nardi, French, and Jones (2016) in a rich model of retirement
savings with endogenous medical expenses. In their framework, they estimate a utility
floor that corresponds to consuming $4,600 a year when healthy. However, they note
that SSI recipients are guaranteed a minimum income of $6,670. As a compromise, we
use $5,900 as our consumption floor for elderly singles, which is $8,687 in 2016 dollars,
and the one for couples to be 1.5 the amount for singles, which is the statutory ratio
between benefits of couples to singles. The retirement benefit at age 66 is calculated to
mimic the Old Age and Survivor Insurance component of the Social Security system.

73
 Calibrated parameters Source
Preferences and returns
r Interest rate 4% De Nardi, French, and Jones (2016)
γ Utility curvature parameter 2.5 see text
ηt Equivalence scales PSID
Government policy
SS(ȳri ) Social Security benefit See text
τtSS Social Security tax rate See text
yet Social Security cap See text
c(1) Minimum consumption, singles $8,687, De Nardi, French, and Jones (2016)
c(2) Minimum consumption, couples $8,687*1.5 Social Security rules

Table 20: First-step calibrated inputs summary

Social Security benefits

The Social Security benefit at age 66 is calculated to mimic the Old Age and
Survivor Insurance component of the Social Security system:
 
 0.9ȳr ,
 ȳr < 0.1115; 

SS(ȳr ) = 0.1004 + 0.32(ȳr − 0.1115), 0.1115 ≤ ȳr < 0.6725;
 
0.2799 + 0.15(ȳr − 0.6725), 0.6725 ≤ ȳr < ytcap
 

The marginal rates and bend points, expressed as fractions of average household
income, come from the Social Security Administration.24 The Social Security tax and
Social Security cap shown in Figure 22 have been changing over time. We also allow
them to change over time for the households in our cohorts.
#10 4
0.065 13

12
0.06
11
Social Security tax rate

Social Security tax cap

0.055 10

9
0.05
8

0.045 7

6
0.04
5

0.035 4
1960 1970 1980 1990 2000 2010 2020 1960 1970 1980 1990 2000 2010 2020
Year Year

Figure 22: Social Security tax and Social Security cap over time (expressed in 2016 dollars)
24
Social Security Administration, “Benefit Formula Bend Points,”
https://www.ssa.gov/oact/cola/bendpoints.html. We use their values for 2009.

74
Appendix D. The solution algorithm
This appendix describes the solution algorithm. We first solve the value functions
and policy functions. Then we simulate our model economy using the inputs and
estimate parameters following the procedures that we describe in the next section.
We optimize over six value functions over multiple time periods, compute three
more value functions, and have six continuous state variables. In addition, there
can be kinks in the value functions because both husbands and wives choose their
participation. Thus, to have reliable solutions, we compute them brute force on a
grid. To get a sense of dimensionality, the value function for working couples has the
following dimensions in terms of state variables: age (41 periods, as we have yearly
periods), assets, earnings shocks for each spouse, and human capital for each spouse.
Over these grids, we evaluate choices for consumption, savings, and labor supply of
both household members and compute all of the relevant expected values at each and
marital status for each of the value functions.
Even parallelizing our model in C on high-end workstations, the model requires 37
minutes for each set of parameter values to be solved. Estimating the model for one
cohort implies solving it thousands of times, which thus requires at least three or four
weeks each time. We reestimate our model for each cohort many times to check for
local minima, robustness, and so on. The computation time required is substantial.
During the retirement stage, single people do not get married anymore; hence,
their value function can be computed independently of the other value functions.
The value function of couples depends on their own future continuation value and
the one of the singles, in case of death of a spouse. Then there is the value function
of the single person being married in a couple, which depends on the optimal policy
function of the couple, taking the appropriate expected values. We compute them as
follows:

1. Compute the value function of the retired single person for all time periods after
retirement by backward iteration starting from the last period.

2. Compute the value function of the retired couple for all time periods after
retirement, which uses the value function for the retired single person in case of
death of one of the spouses by backward induction starting from the last period.

3. Compute the value function of the single person in a marriage for all time
periods after retirement.

75
 During the early retirement stage, single people do not get married, and married
individuals do not divorce or die; hence, the value function of the single person and
that of the couple can be computed independently. We compute them as follows:

1. Compute the value function of the single person for all time periods by backward
iteration starting from the last period in the early retirement stage.

2. Compute the value function of the couple for all time periods by backward
iteration starting from the last period in the early retirement stage.

3. Compute the value function of the single person in a marriage for all time
periods in the early retirement stage.

During the working age, the value functions are interconnected; hence, we solve
each of them at time t, working backward over the life cycle, at each period:

1. Take as given the value of being a single person in a married couple for next
period and the value function of being single next period, which have been
previously computed, and compute the value function of being single this period.

2. Given the value function of being single, compute the value function of the
couple for the same age.

3. Given the optimal policy function of the couple, use the implied policy functions
to compute the value function for a person in a couple.

4. Keep going back in time until the first period.

Appendix E. Moment Conditions and Asymptotic Distribu-

tion of Parameter Estimates
In this appendix, we review the two-step estimation strategy, the moment condi-
tions, and the asymptotic distribution of our estimation. To simplify notation, we do
not include a separate indicator for each of the two cohorts.
In the first step, we estimate the vector χ, the set of parameters than can be
estimated without explicitly using our model. In the second step, we use the method
of simulated moments (MSM) to estimate the remaining parameters, which are con-
tained in the M ×1 vector ∆. For the 1945 cohort, the elements of ∆ are the 19 model
parameters (β, ω, (φi,j i,j i,j 0,5 6,11
0 , φ1 , φ2 ), (τc , τc ), Li,j ).25 For the 1955 cohort, we assume
25
We normalize the time endowment of single men.

76
that the households have the same β and ω as the 1945 cohort, and we thus estimate
the remaining 17 parameters. Our estimate, ∆, ˆ of the “true” parameter vector ∆0 is
the value of ∆ that minimizes the (weighted) distance between the lifecycle profiles
found in the data and the simulated profiles generated by the model.
From ages 25 to 65, we match average assets for single men, single women, and
couples, as well as working hours and participation for single men, single women,
married men, and married women. For the generic variable z equal to hours (H),
i,j
participation (In), and assets (a), we denote zk,t the sample observation relative to
person k, of gender i, marital status j, and age t. Denoting zti,j (∆, χ) the model-
predicted expected value of z for age i, gender i, and marital status j, where χ is the
vector of parameters estimated in the first step, we write the moment conditions as

E[ai,j i,j
k,t − at (∆0 , χ0 )] = 0, ∀t = 2, ..., 41 (47)
i,j
E[Hk,t − Hti,j (∆0 , χ0 )] = 0, ∀t = 1, ..., 41 (48)

E[Ini,j i,j
k,t − Int (∆0 , χ0 )] = 0, ∀t = 1, ..., 41. (49)

Note that assets for couples, ai,j

k,t , do not depend on gender when marital status
is j = 2. Also, as assets at age 25 (t = 1) is an initial condition, it is matched by
construction. Thus, we have a total of J = 448 moment conditions. In practice,
we compute the sample expectations in equations (47), (48), and (49) conditional
on a flexible polynomial in age. More specifically, we regress each variable z on a
fourth-order polynomial in age and on a set cohort of dummies, fully interacted with
marital status and separately for each gender. We then compute the conditional
expectations for each cohort in turn using the estimated marital- and gender-specific
polynomial in age as well as coefficients relative to that cohort. These average age
profiles, conditional on gender, marital status, and cohort, are those shown in the
figures in the main text.
Suppose we have a data set of K persons that are each observed at up to T sepa-
rate calendar years. Let ϕ(∆; χ0 ) denote the J-element vector of moment conditions
described immediately above, and let ϕ̂K (.) denote its sample analog.
Letting Wc K denote a J ×J positive definite weighting matrix, the MSM estimator
∆ˆ is given by
argmin ϕ̂K (∆; χ0 )0 W
c K ϕ̂K (∆; χ0 ). (50)
∆

77
Note that we also estimate χ0 . For tractability reasons, and following much of the
literature, we treat it as known.
Under the regularity conditions stated in Pakes and Pollard (1989) and Duffie
and Singleton (1993), the MSM estimator ∆ ˆ is both consistent and asymptotically
normally distributed:
√  
ˆ − ∆0
K ∆ N (0, V), (51)

with the variance-covariance matrix V given by

V = (D0 WD)−1 D0 WSWD(D0 WD)−1 , (52)

where S is the variance-covariance matrix of the data;

∂ϕ(∆; χ0 )
D= (53)
∂∆0 ∆=∆0

is the J×M gradient matrix of the population moment vector; and W = plimK→∞ {W c K }.
When W = S−1 , V simplifies to (D0 S−1 D)−1 .
c K converges to S−1 ,
The asymptotically efficient weighting matrix arises when W
the inverse of the variance-covariance matrix of the data. However, as Altonji and
Segal (1996) point out, the optimal weighting matrix is likely to suffer from small
sample bias. We thus use a diagonal weighting matrix that is the same as S along the
diagonal and has zeros off the diagonal of the matrix. We estimate D and W with
their sample analogs.

Appendix F. Parameter Estimates

Figure 23 reports the age-varying time costs of working by age expressed as a
fraction of the time endowment of single men that are necessary to reconcile the
labor market participation of our four groups of people in each cohort.

Appendix G. Model fit, additional information

We do not match savings after age 66 because the asset data become very noisy
after that. However, the model does fit them well. Figure 24 shows the full profile of
assets generated by the model and those in the data for the 1945 cohort.
Figure 25 compares additional model implications with those in the data for cou-
ples. The top panels display participation patterns within married couples from the
model and the PSID data. While we match the participation of married men and

78
 Cohort 1945 Cohort 1955
Coeff. Std. Err. Coeff. Std. Err.
β 0.9898 (0.00025)
ω 0.4057 (0.00121)
Participation costs:
φ1,1
2 0.0001 (0.00003) 0.0008 (0.00003)
φ1,1
1 -0.0044 (0.00054) -0.0312 (0.00149)
φ1,1
0 -1.3769 (0.02518) -0.9631 (0.02210)
φ2,1
2 0.0006 (0.00003) -0.0001 (0.00001)
φ2,1
1 -0.0170 (0.00095) 0.0024 (0.00029)
φ2,1
0 -1.4778 (0.03245) -0.8859 (0.01233)
φ1,2
2 0.0007 (0.00002) 0.0016 (0.00004)
φ1,2
1 -0.0187 (0.00087) -0.0655 (0.00164)
φ1,2
0 -1.5621 (0.03095) -1.4657 (0.02473)
φ2,2
2 0.0033 (0.00008) 0.0064 (0.00012)
φ2,2
1 -0.1345 (0.00394) -0.2723 (0.00706)
φ2,2
0 -2.1433 (0.04556) -1.8571 (0.04541)
Time endowments:
F L2,1 -3.0381 (0.08251) -5.5347 (0.40068)
F L1,2 -3.0197 (0.07796) -2.1473 (0.03101)
F L2,2 -1.2955 (0.01272) -1.2654 (0.00768)
Childcare costs:
τc0,5 0.3002 (0.01316) 0.2502 (0.01527)
τc6,11 0.0683 (0.00718) 0.1885 (0.01007)

Table 21: Estimates of parameters. Standard errors in parentheses. We estimate F Li,j

L
and time endowment in the model is given by Li,j = 1+exp(F Li,j )
, where we
normalize L to 112 hours a week.

women by estimation, we do not match the fraction of couples with no earners or

with only women earners. The bottom panels show hours worked by husbands whose
wife is not working, husbands whose wife is working, and wives whose husband is
working. We report them by age over the whole working period. This, also, is not a
target that our estimation procedure seeks to match. Both sets of graphs reveal that
the model also reproduces these aspects of the data well. This is remarkable given
that our model is tightly parameterized compared with the number of targets that it
matches.
Figures 26 and 27 report our model-implied moments as well as target moments

79
 0.35 0.35
SM
SW
0.3 MM 0.3
MW

0.25 0.25

Participation cost
Participation cost

0.2 0.2

SM
0.15 0.15
SW
MM
0.1 0.1 MW

0.05 0.05

0 0
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

Figure 23: Estimated life cycle labor participation costs expressed as a fraction of the time
endowment of single men. SM: single men; SW: single women; MM: married
men; MW: married women. Left panel: 1945 cohort. Right panel: 1955 cohort.
Model estimates

#10 5 Couples #10 5 Single women

Average Household Asset

8 5
Model
Model
4 Data
Data
7 Data, upper bound
Data, upper bound
3 Data, lower bound
Data, lower bound
6 2
Average Household Asset

1
5
0
25 30 35 40 45 50 55 60 65 70 75 80
4 Age
#10 5 Single men
Average Household Asset

5
3 Model
4 Data
Data, upper bound
2 3 Data, lower bound

2
1
1

0 0
25 30 35 40 45 50 55 60 65 70 75 80 25 30 35 40 45 50 55 60 65 70 75 80
Age Age

(a) Assets, couples (b) Assets, singles

Figure 24: 1945 cohort. Model fit for assets and average and 95% confidence intervals
from the PSID data

and 95% confidence intervals from the PSID data for our 1955 cohort. They show
that our parsimoniously parameterized model also fits the data for the 1955 cohort
well.

80
 Fraction Among Couples Fraction Among Couples
1
0.9 Both working Both working
Men working only 0.9 Men working only
0.8 Women working only Women working only
Neither working 0.8 Neither working
0.7
0.7
0.6
0.6
0.5
0.5
0.4 0.4
0.3 0.3

0.2 0.2

0.1 0.1

0 0
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

(a) Participation, model (b) Participation, data

Average Working Hours (Workers) Average Working Hours (Workers)
3000 3000
Men, Both working Men, Both working
2800 Men, Men working only 2800 Men, Men working only
Women, Both working Women, Both working
2600 2600

2400 2400

2200 2200

2000 2000

1800 1800

1600 1600

1400 1400

1200 1200

1000 1000
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

(c) Hours for workers, model (d) Hours for workers, data

Figure 25: 1945 cohort. Participation and worker’s hours patterns for people in couples.
Model and PSID data comparison

Married women Single women

1 1
Labor Participation

Labor Participation

0.5 Model 0.5 Model

Data Data
Data, upper bound Data, upper bound
Data, lower bound Data, lower bound
0 0
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age
Married men Single men
1 1
Labor Participation

Labor Participation

Model Model
Data Data
0.5 0.5
Data, upper bound Data, upper bound
Data, lower bound Data, lower bound

0 0
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

(a) Participation, couples (b) Participation, singles

Married women Single women
Hours among workers

Hours among workers

2500 2500

2000 2000

1500 Model 1500 Model

Data Data
1000 Data, upper bound 1000 Data, upper bound
Data, lower bound Data, lower bound
500 500
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age
Married men Single men
Hours among workers

Hours among workers

2500 2500

2000 Model 2000

Data
1500 1500 Model
Data, upper bound
Data
1000 Data, lower bound 1000 Data, upper bound
Data, lower bound
500 500
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

(c) Hours for workers, couples (d) Hours for workers, singles

Figure 26: 1955 cohort. Model fit for participation (top graphs) and hours (bottom
graphs) and average and 95% confidence intervals from the PSID data

81
 #10 5 Couples Average Household Asset Average Household Asset #10 5 Single women
8
Model
Data 4
7 Model
Data, upper bound
Data
Data, lower bound
Data, upper bound
Average Household Asset

6 2
Data, lower bound

5 0
25 30 35 40 45 50 55 60 65
4 Age
#10 5 Single men
3 Model
4 Data
2 Data, upper bound
Data, lower bound
2
1

0
0 25 30 35 40 45 50 55 60 65
25 30 35 40 45 50 55 60 65
Age
Age

(a) Assets, couples (b) Assets, singles

Figure 27: 1955 cohort. Model fit for assets and average and 95% confidence intervals
from the PSID data

82
Appendix H. Policy experiments results without balancing
government budget for both cohorts
No spousal or survival benefit in SS No spousal or survival benefit in SS
0.3 10000
Single Men Single Men
Single Women Single Women
0.25 8000
Married Men Married Men

Change in average Labor Income

Married Women Married Women
0.2 6000
Change in participation

0.15 4000

2000
0.1

0
0.05

-2000
0
-4000
-0.05
-6000
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

No marital differential tax No marital differential tax

0.3 10000
Single Men Single Men
Single Women Single Women
0.25 8000
Married Men Married Men

Change in average Labor Income

Married Women Married Women
6000
Change in participation

0.2

4000
0.15
2000
0.1
0

0.05
-2000

0 -4000

-0.05 -6000
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

All policies change All policies change

0.3 10000
Single Men
Single Women
0.25 8000
Married Men
Change in average Labor Income

Married Women
0.2 6000
Change in participation

0.15 4000

2000
0.1

0
0.05

-2000
0 Single Men
Single Women
-4000
Married Men
-0.05
Married Women
-6000
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

Figure 28: 1945 cohorts: Changes in participation (left panels) and labor income (right
panels), unbalanced government budget. Top panels: after the elimination of
all the spousal Social Security benefits; middle panels: after the elimination
of joint income taxation; bottom panels: after the elimination of all marital-
related policies.

83
 Couples Single men Single women
Removing spousal Social Security benefits 9.7% 1.7% 4.3%
Removing all marital-related policies 15.3% 3.3% 8.5%

Table 22: 1945 cohorts: Change in assets at age 66, in percentages, unbalanced govern-
ment budget.

All policies change All policies change

0.3 10000
Single Men
Single Women
0.25 8000
Married Men
Change in average Labor Income

Married Women
6000
Change in participation

0.2

4000
0.15
2000
0.1
0

0.05
-2000
Single Men
0 Single Women
-4000
Married Men
Married Women
-0.05 -6000
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

Figure 29: 1955 cohorts: Changes in participation (left panel) and labor income (right
panel) after the elimination of all spousal Social Security benefits and joint
income taxation. Unbalanced government budget

Couples Single men Single women

Savings, unbalanced government budget 15.6% 3.8% 9.5%

Table 23: 1955 cohorts: Change in assets at age 66, in percentages, as a result of re-
moving spousal Social Security benefits and joint income taxation, unbalanced
government budget

84
Appendix I. Comparing PSID and CPS data
Starting in 1968, the PSID has excellent data for the cohort of people we want
to study. Its design allows the sample to remain representative of the US popula-
tion. Despite attrition, it has maintained its cross-sectional validity, as discussed by
Fitzgerald et al., 1998, and Moffitt and Zhang, 2018. Nonetheless, in this appendix,
we compare the key moments from the PSID with the corresponding ones that we
compute from the Current Population Survey (CPS) which does not have a panel
dimension, and hence does not allow us to compute many of the inputs that we need,
but has a relatively larger sample size.
Labor Participation Labor Participation
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 Single men 0.4 Single men

Single women Single women
0.3 Married men 0.3 Married men
Married women Married women

0.2 0.2
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age
Average Working Hours (Workers) Average Working Hours (Workers)
2400 2400

2200 2200

2000 2000

1800 1800

Single men Single men

1600 Single women 1600 Single women
Married men Married men
Married women Married women

1400 1400
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

Figure 30: Life-cycle profiles by gender and marital status for the 1945 cohort in the PSID
(left-hand-side panel) and CPS (right-hand-side panel) data

85
 Labor Participation Labor Participation
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 Single men 0.4 Single men

Single women Single women
0.3 Married men 0.3 Married men
Married women Married women

0.2 0.2
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age
Average Working Hours (Workers) Average Working Hours (Workers)
2400 2400

2200 2200

2000 2000

1800 1800

Single men Single men

1600 Single women 1600 Single women
Married men Married men
Married women Married women

1400 1400
25 30 35 40 45 50 55 60 65 25 30 35 40 45 50 55 60 65
Age Age

Figure 31: Life-cycle profiles by gender and marital status for the 1955 cohort in the PSID
(left-hand-side panel) and CPS (right-hand-side panel) data

86