“f” LIBRARY l
M

1004: ichigan State
University

 

 

 

This is to certify that the
dissertation entitled

MAJOR LIFE EVENTS AND THE
ACCUMULATION OF WEALTH

presented by

DEBORAH KAYE FOSTER

has been accepted towards fulfillment
of the requirements for the

Doctoral degree in Economics

 

 

I

Mr Professor’s Signature

i’ (j 7 1k) ?c; %

 

Date

MSU is an affirmative—action, equal-opportunity employer

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

Fin::awazgi“009

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5/08 K.IProyAcc&PreleIRC/Dateoueindd

 

MAJOR LIFE EVENTS AND THE ACCUMULATION OF WEALTH
By

Deborah Kaye Foster

AN ABSTRACT OF A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
DOCTOR OF PHILOSOPHY
Department of Economics

2008

Professor Steven J. Haider

ABSTRACT
MAJOR LIFE EVENTS AND THE ACCUMULATION OF WEALTH
By
Deborah Kaye Foster

Much attention in the popular press and scholarly journals has been devoted to the
economic resources of the elderly and the divorced, particularly as their proportion of the
US. population continues to grow. Amid these rising concerns, my dissertation explores
some of the major risks confronting these groups.

The first chapter is motivated by the volume of recent literature examining
individual retirement savings behavior. An often-cited paper by Venti and Wise (2001)
investigates how much of the observed variation in retirement wealth is attributable to
differences in household lifetime earnings and life events. Their analysis suggests that
these factors have little effect and concludes that preferences are the predominant driver,
but their data suffer from several shortcomings. I reexamine the role of long-run earnings
and life events using the Panel Study of Income Dynamics (PSID) and find that these
factors account for over five times more of the variation in retirement wealth than
reported by Venti and Wise. Overall, almost half of the variation in retirement wealth
can be explained statistically by long-run earnings and life events.

In the second chapter, the focus shifts to the wellbeing of divorced individuals. It
is well documented that an individual’s income decreases substantially following divorce
and that the decline is much greater for women than for men; however, relatively little
work has examined the wealth effects of divorce. Using the PSID, I find that the wealth

effects of divorce are greater for women than men, and the same is true for total financial

 

.li‘lllill ill

resources. The gender gap grows when adjustments for family size are included. The
gender gap in wealth and total resources remains in the longer-term. Additionally, there
is no evidence that spouses with relatively low post—divorce income levels were
compensated with a larger share of the marital wealth. I also examine the effect of
divorce laws on post-divorce finances. Individuals who divorce in states with unilateral
divorce experience significantly smaller wealth declines than those who divorce under
mutual divorce. Men are better off under community property laws while women fare
better with equitable distribution. There is no evidence of compensatory behavior in
equitable distribution or community property states.

Economic analyses typically do not differentiate between marital separation and
divorce. Instead, the two are combined into one marital status category. A potential
motivation for combining the two statuses is that separation is a transitory state that
ultimately leads to divorce. The third chapter examines whether the transitory
assumption holds empirically. The findings indicate that approximately 75% of
individuals divorce within ten years of marital separation, but almost a quarter do not.
Thus, I conclude that separation is a transitory state for most individuals, but a non-trivial
portion of separated individuals do not divorce. The probability of exiting separation
varies by duration of separation, race, and several personal characteristics. The exit
probability declines steadily with duration. Whites are most likely and blacks are least
likely to exit separation. Older individuals and those with kids are more likely to remain

separated.

DEDICATION
This dissertation is dedicated to my husband, Adam, who has supported and
encouraged me every step of the way. Thank you for sharing this journey with me —

here’s to our next adventure together.

iv

ACKNOWLEDGEMENTS

I could not ask for a better dissertation advisor than Dr. Steven Haider. Many,
many thanks for your endless patience, encouragement, and guidance. You always
recognized my progress and challenged me to make my research even better. You have
been a true inspiration to me, and I am forever grateful.

I offer my sincere appreciation to committee members Dr. Stacy Dickert-Conlin
and Dr. Paul Menchik. You each brought a unique perspective to my research. I am very
thankful for your thoughtful comments and suggestions.

Many thanks to my fellow graduate students, particularly Peggy, Brian, and
Nicole. You understand the ups and downs of the graduate experience better than anyone.
I will always treasure the good times we’ve shared.

Most of all, I give my heartfelt appreciation to my husband, Adam. You have
always been my greatest supporter. You celebrated the small victories along the way and
helped give me the will to persevere when things were tough. It means so much to share

this accomplishment with you.

TABLE OF CONTENTS

LIST OF TABLES ............................................................................................................ vii
LIST OF FIGURES ........................................................................................................... ix
CHAPTER 1: Who Saves? The Role of Long-run Earnings and Life Events on
Retirement Wealth .............................................................................................................. 1
Introduction ..................................................................................................................... 1
Literature Review ............................................................................................................ 3
Data ................................................................................................................................. 5
Results ........................................................................................................................... 10
Conclusion and Discussion ........................................................................................... 23
Appendix ....................................................................................................................... 26
References ..................................................................................................................... 35

CHAPTER 2: Irreconcilable Differences: A Gender Comparison of Post-divorce

Economic Wellbeing ......................................................................................................... 37
Introduction ................................................................................................................... 37
Divorce Laws and Terminology ................................................................................... 38
Literature Review .......................................................................................................... 4O
Conceptual Framework ................................................................................................. 44
Data ............................................................................................................................... 49
The Economic Effects of Divorce ................................................................................. 55
Legal Regime ................................................................................................................ 68
Conclusion .................................................................................................................... 78
Appendix ....................................................................................................................... 80
References ..................................................................................................................... 82

CHAPTER 3: Is Marital Separation a Transitory State? Evidence from the P811) .......... 84
Introduction ................................................................................................................... 84
Data ............................................................................................................................... 85
Results ........................................................................................................................... 87
Conclusion .................................................................................................................... 95
References ..................................................................................................................... 96

vi

LIST OF TABLES

Table 1.1: Correlations between Long—run Earnings for Averages Ages 35-54 and

Select Age Ranges ..................................................................................................... 8
Table 1.2: Comparison of Long-run Earnings Deciles with Censored and Uncensored
Data ......................................................................................................................... 12
Table 1.3: Household Long-run Earnings and Wealth by Decile .................................... 14
Table 1.4: Regression Goodness of Fit by Long-run Earnings Measure ......................... 15
Table 1.5: Regression Goodness of Fit by Life Events Measure ..................................... 17
Table 1.6: Regression Goodness of F it Including Additional Factors ............................. 20
Table A1.1: Overall Regression Results ........................................................................ 29
Table A1.2: Comparison of Long-run Earnings Quintiles with Censored and Uncensored
Data ......................................................................................................................... 31
Table A1 .3: Regression Goodness of Fit with PSID Family Weights ............................. 32
Table A1.4: Regression Goodness of Fit Excluding Home Equity ................................. 33
Table A1.5: Regression Goodness of Fit with Family Equivalence Scales ..................... 33
Table Al.6: Sample Comparison of Median Wealth and Long-run Earnings Deciles ..... 34
Table 2.1: Summary Statistics ....................................................................................... 51
Table 2.2: Incidence and Magnitude of Spousal Support and Child Support for Divorced
Individuals .............................................................................................................. 53
Table 2.3: Changes in Income, Wealth, and Total Financial Resources .......................... 57
Table 2.4: OLS Regressions of Changes in Resources for Divorced Individuals ............ 60
Table 2.5: Longer-term Changes in Income, Wealth, and Total Financial Resources ..... 63

Table 2.6: OLS Regressions of Longer-term Changes in Resources for Divorced
Individuals .............................................................................................................. 64

Table 2.7: OLS Regressions of Income-Wealth Trade-off ............................................. 67

vii

Table 2.8: Changes in Total Financial Resources by Legal Regime ............................... 69
Table 2.9: OLS Regressions of Changes in Resources with Legal Regime ..................... 74

Table 2.10: OLS Regressions of Income-Wealth Trade-off by Property Division Type .77

Table A2.1: Divorce Laws by State, 1984-2007 ............................................................ 80
Table 3.1: Summary Statistics for Separated and Divorced Individuals .......................... 88
Table 3.2: Hazard Model Estimates ............................................................................... 94

viii

LIST OF FIGURES

Figure 3.1: Post-Separation Marital Status by Year Since Separation ............................. 91
Figure 3.2: Probability of Exit from Separation by Length of Separation ....................... 92

Figure 3.3: Probability of Exit fi'om Separation by Length of Separation and Race ........ 93

ix

Who Saves?

The Role of Long-run Earnings and Life Events on Retirement Wealth

1 . Introduction

The looming retirement of the baby boomers, the debate over the future of Social
Security, and high-profile pension bankruptcies have all heightened interest in personal
retirement savings in America. Despite the recognized importance of accumulating
sufficient wealth for retirement, the data show that wealth in the US. is very unevenly
distributed. For instance, Smith (1997) examines wealth for households in the AHEAD
in which at least one member is age 70 or older. He finds that the top 10‘h percentile of
households has 2,500 times as much wealth as the bottom 10th percentile. Such findings
have led researchers to investigate potential causes for this inequality.

In an often-cited paper, Venti and Wise (2001) examine whether the variation in
wealth can be attributed to differences in the opportunity to save. 1 They use the Health
and Retirement Survey (HRS) to determine how much of the observed variation in
retirement wealth can be attributed to factors that affect the amount of money that can be
allocated to savings. Specifically, they examine the role of household lifetime earnings
and life events (including inheritances, children, marital status, and health) on wealth at
retirement. Their analysis suggests that these factors have little effect, leading them to
conclude that “the bulk of the dispersion must be attributed to differences in the amount
that households choose to save” (57). In short, they attribute differences in wealth

accumulation to different saving preferences, not to differences in the ability to save.

 

I Citations of Venti and Wise (2001) are found, for example, in the Journal of Political Economy (2003),
Journal of Financial Economics (2004), Southern Economic Journal (2005), Journal of Finance (2006), and
Journal of Monetary Economics (2007).

Despite the attention their study receives, the data Venti and Wise use suffer from
several crucial shortcomings. First, their measure of lifetime earnings comes from the
Social Security Administration (SSA) earnings reports that are top-coded. Venti and
Wise report that over 25% of the sample exceeded the earnings cap in 1971, but only
4.8% of their sample was affected in 1991 (30). Even though the portion of the sample
affected by the earnings cap declines, it still has the potential to change conclusions.

Hi gh-earners typically have more wealth, and those are the exact individuals whose
lifetime earnings measure is affected by the censoring. Additionally, the demographic
information used to measure life events is based on the HRS baseline survey completed
in 1992, when the primary respondents are between the ages of 51 and 61. Although
some retrospective information is included, relatively little is known about the
participants’ lives prior to the survey.

The goal of this paper is to reexamine the role of long-run earnings and life events
on retirement savings using the Panel Study of Income Dynamics (PSID). The survey
includes earnings information that allows for more precise examination of the earnings-
wealth relationship. The same households are observed over an extended period of time,
and thus both the occurrence and timing of life events can be considered. The findings
indicate that the effects of earnings and life events are over five times greater than
previously estimated.

Although the difference in results is striking, my analysis suffers from the same
difficulties in interpretation as Venti and Wise’s analysis. Such a descriptive analysis
cannot untangle the extent to which the life events have a causal effect on wealth

accumulation or instead reflect underlying differences in preferences. For instance, it is

likely that individuals who place a high value on the future will both save more for
retirement and invest in human capital to obtain higher future earnings.

The remainder of the paper is organized as follows. Section 2 includes a brief
review of the relevant literature. In Section 3 the data used in the analysis are presented
(with further detail available in the Appendix). Results are found in Section 4, and I

summarize and discuss the findings in Section 5.

2. Literature Review

Wealth inequality in the United States is well-documented. Indeed, wealth is
much more unevenly distributed than income.2 To gain insight into why these disparities
exist, researchers have examined the correlates of wealth. Such studies are typically
cross-sectional and highlight differences in wealth holdings across different demographic
characteristics.

For example, Blau and Graham (1990) report that young black families only have
about 18% of the wealth held by young white families. Possible explanations include
differences in income, inheritances, and asset allocation.3 Gender differences are also
evident. Studies such as Schmidt and Sevak (2006) document that women have less
wealth than men. The gender wealth gap is often attributed to differences in income and
risk aversion, resulting in women investing differently than their male counterparts.4

Lupton and Smith (2003) examine wealth differences by marital status. They find

that married couples not only have more per capita wealth than divorced individuals, but

 

2 For example, see Hurst, Luoh, and Stafford (1998), Wolff (1998), and Cagetti and De Nardi (2005).

3 See Menchik and J ianakopolos (1997), Altonji, Doraszelski, and Sega] (2000), and Gittleman and Wolff
2000) for further detail.

Refer to Bajtelsmit and Bemasek (1996) for a survey of this evidence.

the duration of marital status matters, too; those married longer have more wealth, and
those divorced longer have less. Along similar lines, Wilmoth and Koso (2002) find that
wealth is affected not only by marital status, but by the sequence of marital events over a
person’s lifetime.

Researchers have also examined the effect of children on wealth. Schmidt and
Sevak (2006) find that younger children do not affect wealth levels, but households with
children between the ages of 18 and 24 have less wealth. They argue that these findings
are consistent with parents saving for their children’s education. On the other hand,
Lupton and Smith (2003) do not find a significant relationship between children and
wealth regardless of children’s age.

Hurst (2006) compares the behavior of households in the PSID who enter
retirement with relatively high wealth to those with relatively low retirement wealth. He
finds that the low-wealth group responds to predictable changes in income, but the high-
wealth group does not. From this he concludes that the low-wealth group is short-sighted
while the high-wealth group demonstrates consumption smoothing behavior.

Venti and Wise (2001) include many of the factors discussed above to determine
how much of the overall variation in retirement wealth can be explained by lifetime
earnings, a variety of life events, and investment choice. They find that, even for
households with similar lifetime earnings, substantial variation in retirement wealth exists.
Next, they control for age, marital status, children, inheritances, and health. The results
show that these factors only explain about 4% of the observed variation. The effect of

investment choice (about 8%) is also very small. From this, they conclude that the

primary cause of wealth dispersion is differences in preferences — some choose to save

while young and accumulate extensive wealth while others do not.

3. Data
A brief description of the data and variables used in the analysis are provided

below. Please see the Appendix for further detail.

3.1 The Sample

The data are from the PSID. The longitudinal survey, conducted by the Survey
Research Center at the University of Michigan Institute for Social Research, was
administered annually from 1968 to 1997. Beginning in 1999, it switched to a biennial
schedule. The most recent data available are from 2005. In all, there are 34 waves of
data spanning a period of 38 years.

My sample is composed of males (and their households) that meet two selection
criteria. The first requirement is a measure of retirement wealth. Specifically, the
household must complete a wealth supplement when the male is 55 (plus or minus two)
years old. Age fifty-five is younger than the standard retirement age, but the majority of
wealth accumulation should have occurred by that age and the earlier wealth measure
mitigates problems arising from early retirees who have already begun to deplete their

wealth.5 The second criterion is that the household is observed twenty years prior to the

 

5 Use of age 55 also makes the results more readily comparable to those of Venti and Wise whose sample
has an average age of 55.4 years.

measure of retirement wealth. This is necessary to calculate the measure of long-run
earnings. The final sample consists of 944 households.6

By construction, age 55 wealth information is compiled from the 1989, 1994,
1999, 2003, and 2005 wealth supplements.7 In the remainder of the paper, households
whose retirement wealth measure comes from the same wealth supplement are referred to

as cohorts.

3.2 Measuring Wealth

The wealth measure comes from the PSID wealth supplement.8 The survey’s
measure of total wealth is composed of eight asset categories: business/farm,
checking/savings, debt, real estate, stock, vehicles, home equity, and other savings. The
wealth measure does not include public or private pension wealth. Total wealth is
reported at the household level and cannot be allocated to individuals within the
household.9 1 convert all values to 2005 dollars using the Bureau of Economic Analysis’
Personal Consumption Expenditures (PCE) price index.

Self-reported wealth information is subject to severe non-response. Because the
total wealth measure is comprised of several different categories, failure to report any

single component means that a total wealth measure is not available for that household.

 

6 The results presented here exclude households in the SEO sub-sample. The analysis was also performed
including those sample members with the PSID-provided family weights, and the results (available in the
Appendix) are not substantially different than those presented here.

7 The 1984 supplement is excluded by the requirement that households are observed 20 years prior to the
wealth measure because the PSID did not begin until 1968. The 2001 supplement is not used because all
but two of those households are picked-up in the 1999 and 2003 supplements.

8 See Smith (1995) for discussion of the HRS wealth measure and how it compares to the PSID measure.

9 Although wealth cannot be assigned to individuals within the household, it is possible to adjust for family
size. Results employing the National Academy of Sciences family equivalence scale are similar to those
presented in the text and are available in the Appendix.

Higher-wealth households have complex holdings that can be difficult to value and are
therefore more likely not to be reported. If the non-response is not random (as in this
situation) then estimates based on the data will be biased.

The PSID wealth supplement addresses this issue with unfolding brackets.l0
Individuals are first asked to report the dollar amount of a given asset. If the respondent
does not state a specific dollar amount then the interviewer follows-up with a series of
questions (along the lines of “Would it amount to $50,000 or more?”) to determine the
appropriate bracket (range of values). Based on those responses, the PSID uses a three-

level hot deck procedure to impute the missing values.'1

3.3 Measuring Long-run Earnings

Ideally, the analysis would include lifetime earnings. Given that the PSID only
collects contemporaneous earnings, I am restricted to a measure of long-run earnings
with an inherent trade—off — the measure gets closer to actual lifetime earnings as more
years are included, but at the same time, the sample gets smaller.

I begin with twenty years of earnings, corresponding to approximate ages 35-54.
To gauge whether this choice is reasonable, I compare it with the earnings measure and
the resulting sample size for a variety of age ranges. The results are displayed in Table
1.1. The individual cells indicate the correlation between the chosen long-run earnings
measure (again, it is approximately ages 35-54) and long-run earnings using the indicated
age range. Sample sizes are noted in brackets. The effect of including additional years at

the beginning of the age range is evident when going from right to left across a row;

 

'0 See Juster and Smith (1997) for details on how brackets remedy non-response bias.

H Details regarding the PSID imputation process are provided in the PSID data documentation at
http://psidonline.isr.umich.edu/Data/Documentation/wlth.html.

moving down a column, one can see the effect of adding more years to the end of the age
span. It is immediately obvious that increasing the age range greatly reduces sample size.
For example, for a starting age of 35, increasing the ending age just five years from 54 to
59 decreases the sample size from 944 to 540. That extra five years eliminates over 40%
of the sample, but the measure including those extra years is highly correlated (0.96) with
the original measure. Based on this analysis, I conclude that use of earnings for ages 35-
54 yields a measure that is highly correlated with lifetime earnings while retaining a large

sample size.

 

Table 1.1: Correlations between Long-run
Earnings for Average Ages 35-54 and

Select Age Ranges
— Average Starting Age

 

 

 

20 25 30 35

q, 49 0.95 0.96 0.97 0.97
3:” [404] [634] [790] [944]
ED 54 0.98 0.99 1.00
'2 [634] [790] [944]
L; 59 0.96 0.96
g [386] [540]
E 64 0.98

[310]

 

 

 

Note: Sample size in brackets.
Source: Author's calculations from PSID

The PSID variable used to create the long-run earnings measure is total labor
income. It includes wages/salaries, bonuses, overtime, tips, commissions, professional
practice or trade, market gardening, additional job income, and miscellaneous labor
income. Since retirement decisions are typically made at the household level, the

variable of interest is household long-run earnings. To that end, the spouse’s total labor

income is included when applicable.'2 All earnings are converted to 2005 dollars using
the PCB and discounted using a 2% discount rate. The results are not sensitive to the
discount factor used.

Because the survey switched to a biennial schedule in 1997, earnings are not
available for all years. For the affected sample members, earnings for the missing years
are calculated as the simple average of earnings in the previous and following year. To
adjust for missing earnings data in years when the survey was conducted (which is
different from reported earnings of zero), I use the average of non-missing household

earnings as the long-run. earnings measure.l3

3.4 Measuring Life Events

The life events considered here are the same as those included in Venti and Wise:
inheritance, children, marital status, and health. The difference is, due to the panel
structure of the PSID, timing of these events can also be considered. Specifically, the life
event variables are divided into three time periods corresponding to approximate ages 35-
39, 40-49, and 50-54. That is roughly events occurring during the male’s 305, 403, and
505.

Inheritance data are from a question included in each survey year as to whether
any lump payments were received. It includes both inheritances and settlements from an
insurance company or lawsuit. The PSID collects the number of children living in the
household at the time of the survey (regardless of biological relationship with other

household members), and the maximum number reported for each age range is used. The

 

'2 The wife’s earnings are only available for years during which the couple is married.
'3 In the sample of 944 households, 72 are missing income for one or more years.

marital status variables included are indicators for those who remain married throughout,
get married, or whose marriage ends (through divorce, separation, or widowhood) during
the relevant period. The measure of health comes from a series of questions about work

missed due to own illness or illness of a family member.

4. Results

The results are organized as follows. First, I reexamine long-run earnings, life
events, and demographic characteristics as analyzed by Venti and Wise and discuss the
differences between our results and possible reasons for them. Then two additional
issues (earnings volatility and pro-existing wealth) are addressed.

The goal of this analysis is to gauge how well long-run earnings and life events
explain the observed variation in retirement wealth, and thus a measure of variation is
required. Common measures include the R-squared, adjusted R-squared, and root mean
square error (RMSE). The R-squared is the percentage of variation in the dependent
variable that is explained by the independent variables. The adjusted R-squared is similar,
but it makes a degrees-of-freedom adjustment which effectively penalizes the inclusion
of an additional explanatory variable. Because of that alteration, the adjusted R-squared
should no longer be interpreted as a percentage. An alternative measure is the RMSE.
Unlike the R-squared and adjusted R—squared, this measure is based on the residuals and
thus expresses the amount of variation not attributable to the regressors. Like the
adjusted R-squared, the RMSE includes a degrees-of-freedom adjustment that penalizes
the inclusion of more variables. The RMSE is linked to the R-squared and adjusted R-

squared through the identity that the total sum of squares (SST) is equal to the explained

10

sum of squares (SSE) plus the sum of squared residuals (SSR). The R-squared is the
ratio of SSE/SST, and the percentage change in RMSE is a function of SSR/SST. Thus,
as the R—squared and adjusted R-squared increase, the RMSE decreases. The increase in
one of the R-squared measures and the percent decrease in the RMSE essentially capture
the same thing.

I follow Venti and Wise and use the percent reduction in the RMSE. The findings
are similar regardless of which measure is used, but using RMSE makes my estimates
comparable to those of Venti and Wise. It should be noted, however, that all of these
measures have similar limitations. Measures of goodness-of-fit do not measure the
quality of the model, and they are within-sample measures only (not population
estimates).

Because the focus here is on overall explanatory power, individual coefficient
estimates are not included in the text of the paper. However, coefficients and standard

errors for the overall model are reported in the Appendix.

4.1 The Role of Long-run Earnings

To alleviate problems arising from the use of censored earnings data, Venti and
Wise rank their sample by lifetime earnings and divide it into deciles. They concede that
results for the lowest and highest deciles are less reliable but state that “the ranking by
Social Security earnings represents a good approximation to a ranking based on actual
total earnings, and that thus the deciles are a good approximation to actual lifetime

earnings deciles” (30).

ll

Because I have uncensored data, it is possible to examine the effectiveness of
their correction method. I begin by applying the SSA’s annual earnings maximums to the
annual earnings information collected in the PSID. If the reported earnings are below the
earnings maximum, actual earnings are used. When actual earnings exceed that
maximum, the maximum amount is substituted for actual earnings.l4 Once this
adjustment is made for each year of earnings for both the male and, when applicable, his
wife, I create a new variable called Social Security-adjusted long-run earnings. As with
the original long-run earnings variable, it is the average of annual earnings for the twenty
years prior to the retirement wealth measure, which is approximately ages 35-54.

Each household is assigned to a long-run earnings decile two times. The first is
based on censored earnings data, and the second is from the uncensored earnings data.

Table 1.2 provides a cross-tabulation of the censored and uncensored deciles.

 

Table 1.2: Comparison of Long-run Earnings
Deciles with Censored and Uncensored Data

Uncensored Long-run Earnings Decile
1 2 3 4 5 6 7 8 9 10

 

 

 

 

 

2 1 91 3 - - - - - - - -
g 2 3 83 6 1 - - 1 - - -
a. 3 — 8 76 7 1 1 - — 1 -
E 4 - - 12 63 10 6 3 - 1 -
513 5 - - - 24 46 10 - 6 3 6
E) 6 - - - — 38 28 10 8 8 3
g 7 - - - - - 46 20 10 11 8
g 8 - - - - - 4 55 16 11 8
g. 9 - - - - - - 6 48 17 23
c3 10 - - - - - - - 6 42 46

 

Note: Sample size is 944.
Source: Author's calculations from PSID

 

'4 In the sample, 38% of households have at least one year of earnings that exceeds the corresponding SSA
maximum.

12

If the SSA maximum had no effect (meaning households are assigned to the same decile
whether censored or uncensored data are used), all households would be along the
diagonal. It is evident looking at Table 1.2 that the effect of censored data is not
restricted to the last decile as Venti and Wise assume. Instead, the effect is spread
primarily between deciles 4-10. More than 52% of households in the censored fifth
decile change groups, and at least 70% of the seventh through ninth deciles change
deciles. Overall, 49% of the sample is assigned to a different decile when the SSA
maximum is imposed, and 20% of those move more than one decile.”

Table 1.3 compares long-run earnings and the distribution of retirement wealth for
censored and uncensored deciles.l6 Looking first at median long-run earnings, censored
and uncensored earnings measures are similar for the lower deciles, but large differences
emerge in the upper deciles. The difference between median earnings for deciles using
censored and uncensored data increases from 10% of the censored data for the fifth decile
up to 41% for the tenth decile. There is a similar pattern for median wealth; the lower
deciles are relatively similar, but the upper deciles differ considerably.

Because one purpose of the analysis is to determine how much of the observed
variation in wealth is attributable to earnings, differences in the distribution of wealth
within deciles are even more important than differences in median wealth levels across
deciles. As the coefficient of variation column demonstrates, the within-decile variation
is considerably different for the censored and uncensored data. For all but 3 deciles, the
degree of variation in wealth to be explained by earnings is overstated when censored

earnings are used.

 

‘5 See the Appendix for a similar analysis and discussion using long-run earnings quintiles.

16 See the Appendix for a comparison of Venti and Wise’s HRS wealth distribution with the PSID wealth
distribution.

13

 

Table 1.3: Household Long-run Earnings and Wealth by Decile

Censored Data

 

 

 

 

 

 

Lifetime Median Wealth
Income Household 25th 50th 75th Coefficient
Decile Long-run Earnings Percentile Percentile Percentile of Variation
1 $24,088.19 $4,354.15 $29,584.51 $83,320.67 176.43
2 40,315.95 29,224.41 71,529.20 147,817.94 309.28
3 49,124.05 34,039.46 100,563.66 216,446.16 210.50
4 57,757.16 78,652.89 186,854.92 355,603.75 238.43
5 65,205 .86 79,082.04 203,000.00 394,423 .22 287.51
6 74,982.73 88,690.13 180,000.00 330,795 .48 184.07
7 83,283.33 157,319.40 272,968.45 526,985.10 230.85
8 91,200.08 164,900.00 294,399.67 553,257.50 107.73
9 102,088.31 191,849.09 419,166.46 828,830.41 98.96
10 122,074.25 277,100.00 440,627.50 885,000.00 256.47
Uncensored Data
Lifetime Median Wealth
Income Household 25th 50th 75th Coefficient
Decile Long-run Earnings Percentile Percentile Percentile of Variation
1 $24,882.15 $4,354.15 $29,450.68 $79,984.73 167.14
2 40,281.15 29,224.41 70,398.75 138,038.72 153.1 1
3 50,544.33 40,500.00 100,563.66 202,818.42 284.01
4 60,728.24 76,500.00 176,715.51 328,000.00 192.69
5 72,037.76 73,908.44 141,951.80 252,344.55 222.60
6 82,735.51 95,869.81 216,073.03 379,295.57 260.63
7 93,350.46 162,254.74 238,811.55 351,936.54 203.36
8 106,222.00 175,268.60 323,612.47 549,034.16 101.57
9 124,054.20 264,645.02 432,373.67 849,500.00 151.23
10 171,514.48 422,776.43 681,995.80 1,387,848.71 184.93
Notes:

1 Total sample of 944 households that completed a wealth supplement when the male was near age 55 and were
observed 20 years prior to the wealth measure.

2 Household Long-run Earnings is the average of non-missing discounted male's real total labor earnings for 20
years prior to retirement wealth supplement and, when applicable, the wife's discounted real total labor earnings
in those same years.

3 . . . . .
Total wealth includes value of busrness/farm owned, checkrng/savrngs accounts, real estate, stock, vehrcle,
private annuity/IRA, home equity, and other personal savings.

4 All dollar amounts are deflated to 2005 dollars.
Source: Author's calculations from PSID

l4

The effect of censored earnings data is further evident in Table 1.4. The first two
rows verify that the subsequent findings are not driven by inherent differences between
the HRS and PSID samples; Venti and Wise report that lifetime earnings reduce the

RMSE by 5.1%, and the PSID censored long-run earnings yield a decrease of 1.6%.17

 

Table 1.4: Regression Goodness of Fit by Long-run Earnings Measure

 

 

 

 

Percent
Reduction
Specification N in RMSE
Venti & Wise Lifetime Earnings Deciles 3,992 5.05%
PSID Censored Long-run Earnings Deciles 808 1.57%
PSID Uncensored Long-run Earnings Deciles 5.35%
PSID Uncensored Long-run Earnings 13.62%
PSID Uncensored Long-run Earnings with Polynomial Terms
2nd 14.39%
3rd 21.08%
4th 25.98%
PSID Uncensored Long-run Earnings with Polynomial Terms 26.03%
and Cohort Indicators

 

Note: Wealth is the dependent variable for all regressions.
Sources: Venti and Wise (2001) and author's calculations from PSID

Using long-run earnings deciles based on uncensored annual earnings instead of censored
data reduces the RMSE by 5.4%. Removing the SSA maximums more than doubles the
explanatory power of long-run earnings. Use of the long-run earnings measure in place
of decile indicators more than doubles the percent reduction in RMSE. When uncensored

long-run earnings and four polynomial terms are included in the regression, the residual

 

‘7 Possible reasons for the difference between my results with censored earnings deciles and Venti and
Wise’s findings include sample selection (individuals must remain in the PSID for 20 years to be in my
sample), differences in the wealth measure (Venti and Wise’s measure includes pension wealth by the
PSID measure does not), and that the PSID sample has more censored earnings than the HRS sample.

15

standard deviation is reduced by 26.0%. I 8 These results represent a substantial departure
from Venti and Wise’s estimate of 5.1%.

Table 1.2 provides some insight to the causes for this large difference. Previous
literature establishes that wealth in the US. is concentrated at the top. In Table 1.2 it is
clear that those high-wealth households are also the ones most affected by the censored
data. Indeed, 25 of the 94 households in the uncensored tenth decile are included in
censored deciles five through eight. Use of censored earnings data inappropriately
allocates some of the highest wealth households to lower long-run earnings deciles,
obfuscating the correlation between wealth and long-run earnings.

Recall, this sample differs from that of Venti and Wise because it is composed of
five different cohorts. Each cohort was exposed to a different macroeconomic
environment which may affect their wealth accumulation. To control for any
macroeconomic effects, a series of cohort indicators is added to the regression in the final
row of Table 1.4. This specification is used as the baseline regression for the remainder

of the analysis.

4.2 The Role of Life Events

Venti and Wise identify inheritances, children, marital status, and health as factors
that could affect retirement savings, but their findings indicate that these life events
explain at most 4.0% of the observed variation in wealth. However, as was the case with
long-run earnings, the data used in their analysis limit the dependability of the results.

The demographic information used to measure life events is based on one-time measures

 

‘8 Further polynomial terms can be added, but the marginal effect on the explanatory power decreases and
their inclusion reduces the degrees of freedom. If the R-squared or adjusted R-squared is used to measure
explanatory power, long-run earnings explain approximately 37% of the observed variation.

16

collected when the primary respondents are between the ages of 51 and 61. Because the
PSID observes the same households across time, this analysis can also incorporate the
timing of these life events. As discussed in Section 3, this is done by controlling for

events that occurred when the male was roughly in his 30s, 40s, and 508.

 

Table 1.5: Regression Goodness of Fit by Life Events Measure

 

 

 

Percent
Reduction Incremental

Specification N in RMSE Effect
Venti & Wise One-time Life Event Measures 3,992 9.08% 4.03%
PSID One-time Life Event Measures 808 31.98% 5.95%
PSID Age-varying Life Events Measures 48.39% 22.36%
Age-varying Life Events Added Individually
Baseline Regression 842 26.00%

Plus Inheritance 31.81% 5.81%
Baseline Regression 882 24.04%

Plus Children 29.26% 5.22%
Baseline Regression 862 23.93%

Plus Marital Status 44.50% 20.57%
Baseline Regression 877 23.95%

Plus Health 24.26% 0.31%

 

Notes:

I Venti and Wise interact life event variables with earnings decile, so here life event variables are
interacted with long-run earnings.

2 Baseline regression is retirement wealth on long-run earnings, long-run earnings polynomial terms, and

cohort indicator dummy variables.
Sources: Venti and Wise (2001) and author's calculations from PSID
Table 1.5 highlights the effect of the improved life event measures. The first row
reports Venti and Wise’s finding of 4.0% of variation explained by life events. The next

row shows the results when life event variables at age 55 (the closest approximation to

Venti and Wise) are added to the baseline regression.l9 Replication of Venti and Wise’s

 

'9 Venti and Wise’s variables are: inheritances received before 1980, between 1980 and 1988, and after
1988, current marital status, number of children, and current health rating. This analysis uses inheritances

l7

analysis with the PSID sample yields very similar results. However, when age-varying
life events are employed, the explanatory power of life events increases dramatically
(from 6.0% to 22.4%).

The combined effect of earnings and life events is evident in Table 1.5. The total
percent reduction in RMSE when both factors are included is 48.4%. Thus, the total
explanatory power of long—run earnings and life events is over five times greater than
Venti and Wise’s estimate of 9.1%.

The lower panel of Table 1.5 shows the percent reduction in RMSE and
incremental effect for each life event individually. Due to missing values for some
variables, the sample size is different for each factor. The baseline regression is
performed for each sub-sample and then the indicated life event is added.

Because the life event variables are likely correlated with one another, the
individual regressions should be viewed with caution. However, considering each
variable one-at-a-time sheds light on which of the life events seem to have a larger effect.
The factor with the single largest effect is marital status. This is not surprising given
previous research, but it is remarkable to see such a large effect using a relatively simple
measure. The smallest effect comes from health and is most likely due to limited health
measures available in the PSID.

The coefficient estimates (reported in the Appendix) warrant mention. The
interaction term of inheritances received in the 305 and long-run earnings is statistically
significant, and it is jointly significant with inheritances received in the 30s. Mean long-

run earnings in the sample are $85,507, so the coefficients indicate that the correlation of

 

received between ages 35-39, 40-49, and 50—55, current marital status, maximum number of children ever
reported to the PSID between the ages of 35 and 55, and a current health rating.

18

inheritances is small for households with average earnings, positive for those with below-
average earnings, and negative for those with higher-than-average earnings. The
coefficients for inheritances in the 408 and 508 are not statistically significant, but they do
suggest that inheritances received in the 40s increase wealth for those with average
earnings, and inheritances in the 508 have virtually no effect for households with average
long-run earnings. An F-test indicates that the inheritance variables are jointly significant
at the 5% level.

The children-earnings interaction terms are significant for all three age groups,
and the individual and eamings-interacted terms are jointly significant for all ages. For
those with average earnings, the presence of children in the household when in their 30s
decreases wealth by about $50,000, but having kids in their 405 increases wealth by over
$70,000. Those who still have children in the household in their 505 have significantly
less wealth ($110,000 for those with average earnings).

As suggested by the large decrease in RMSE and verified by an F -test, the marital
status variables are highly significant. The effect of being married consistently from 35-
39 or 50-54 is not statistically significant, but for those with average earnings, being
married consistently in the 405 increases wealth by over $250,000. The other large effect
is for those whose marriage ends during their 305. Those with above average earnings
have more wealth, but for men with lower earnings, the effect is negative and large in
magnitude. The health variables for the 308 are jointly significant at the 10% level, but

their effect is relatively small. Overall, the health variables are not jointly significant.

19

4.3 Race and Education

Venti and Wise initially exclude race and education from their analysis because
those variables are likely to be correlated with saving preferences. After finding little
effect of long-run earnings, life events, and investment choice, they add both education
and race to the regression. Their results indicate that these variables explain very little of
the variation in retirement wealth.

Results including education and race in the baseline regression are in Table 1.6.
Consistent with Venti and Wise’s findings, the effect of both education and race is
negligible. Briefly returning to the coefficient estimates in the Appendix, none of the

race and education variables are statistically significant at the 10% level.

 

Table 1.6: Regression Goodness of Fit Including Additional Factors

 

 

 

 

Percent
Reduction Incremental
Control Variables N in RMSE Effect
Additional Demographics
Baseline Regression 943 16.57%
Plus Education 16.53% -0.04%
Baseline Regression 933 16.84%
Plus Race 16.78% -0.06%
Earnings Volatility
Baseline Regression 944 16.57%
Plus Standard Deviation of Annual Income 21.84% 5.28%
Plus Standard Deviation of Annual Income
Divided by Mean Long-run Earnings 21.80% -0.04%
Plus Standard Deviation of Annual Income
Interacted with Cohort Indicators 30.19% 8.39%
Pro-existing Wealth
Baseline Regression 404 19.41%
Plus Pre-existing Wealth 23.56% 4.15%

 

Note: Baseline regression is retirement wealth on long-run earnings, long-run earnings polynomial
terms, and cohort indicator dummy variables.

Source: Author's calculations from PSID

20

The lack of significance is intriguing because previous literature finds a
substantial race wealth gap that emerges quite early in life. As mentioned in the literature
review, explanations for the racial wealth gap include differences in income, inheritances,
and asset allocation. Afier controlling for income and a fairly small number of life events
(including inheritances), there are no significant differences in wealth by race. These
findings suggest that the majority of the racial differences are captured by income and the

select life events.

4.4 Other Financial Factors

In addition to allowing for a careful reexamination of role of long-run earnings
and life events, the annual earnings data and repeated wealth supplements in the PSID
invite further analysis. First, the uncensored annual earnings are used to consider
earnings volatility. Next, the repetition of the PSID wealth supplement is exploited to
examine the relationship between retirement wealth and wealth holdings earlier in life.
The results including these financial factors are in the lower panels of Table 1.6.
4.4.1 Earnings Volatility

Thus far, the analysis has focused on the relationship between wealth and total
earnings. However, it is possible that the timing of earnings is also important. For
example, the precautionary savings model suggests that, due to uncertainty about the
future, households increase saving in the current period to insure themselves against
unforeseen shocks (such as job loss or medical bills) in the future.20 It seems likely that

the precautionary savings motive would be greater for households with more variation in

 

20 See Kazarosian (1997) for empirical evidence of the precautionary savings motive.

21

annual earnings than for those with more stable earnings. As a descriptive way of
considering this, earnings volatility is added to the regression analysis.2|

The standard deviation in annual household earnings is used to represent earnings
volatility. As shown in Table 1.6, this variable decreases the RMSE by 5.3%. To allow
the magnitude of the precautionary savings motive to vary across earnings levels, the
volatility measure is divided by mean long-run earnings. This new term has no
incremental effect. Since earnings volatility is likely related to macroeconomic factors, it
seems possible that the effect will vary across cohorts. This is incorporated into the
analysis with an interaction of the standard deviation of earnings with the cohort indicator
dummy variables. This decreases the RMSE by an additional 8.4%.

The coefficient estimates (reported in the Appendix) are consistent with a
precautionary savings motive; higher variation in annual earnings increases wealth. The
variable is also highly significant. The earnings-adjusted term is not statistically
significant.

4.4.2 Pro-existing Wealth

The PSID administered the wealth supplement approximately every 5 years
beginning in 1984. It is therefore possible to observe household wealth 20 years prior to
the measure of retirement wealth for a portion of the sample. The last panel of Table 1.6

displays the results including this pre-existing wealth in the baseline regression. The

incremental effect of this variable is 4.2%. Wealth 20 years earlier has almost as much

 

2' In addition to earnings volatility, the timing of income may also be important. For instance, it makes
sense that those who earn most of their money earlier in life would be able to accumulate more retirement
wealth than others. To account for this possibility, a variable for the percentage of total long-run income
earned in the first five years is added to the regression. The results are not reported, but the variable is not
statistically different from zero and has a negligible effect on the RMSE. The same is true for variables
representing the percentage of income earned in the first ten years and last five years.

22

explanatory power as inheritances and children (see Table 1.5). The magnitude of the
effect is remarkable given the widening and churning of the wealth distribution in the
United States (Hurst, Luoh, and Stafford 1998).

The results suggest that some households are able to accumulate substantial
wealth early in life, and those are the very households that enter retirement with greater
wealth. Such findings provide further support that the opportunity to save is important in
determining wealth differences. It is doubtful that differences in saving preferences
would have a large effect at such a young age. It is more likely that differences so early
- in life reflect differences in the ability to save. For instance, individuals who pay their
way through college possibly take longer to complete their degree and have to take out
loans. They have less opportunity to save than those whose parents paid for college and

were therefore able to finish faster and without any financial obligations.

5. Conclusion and Discussion

The correlation between retirement wealth and both household long-run earnings
and a variety of life events is reexamined using the PSID. Previous research by Venti
and Wise suggests that the effect of these variables is minimal. Indeed, they conclude
that long-run earnings explain about 5.1% of the variation and only an additional 4.0% is
attributable collectively to inheritances, children, marital status, and health. Data
limitations, namely the use of censored earnings data and one-time measures of life

events taken. later in life, cast doubt on the dependability of their findings.

23

I perform an analysis similar to that of Venti and Wise and find strikingly
different results. In contrast to the meager 5.1% of variation they find is attributable to
long-run earnings, the analysis with uncensored earnings data indicates that household
long-run earnings explain 26.0% of the observed variation in wealth. The percent of
wealth variation explained by life events is 22.4% (compared to Venti and Wise’s
estimate of 4.0%). Overall, the findings suggest that the explanatory power of long-run
earnings and life events is over five times greater than previous estimates imply.

When long-run earnings, life events, and earnings volatility are included jointly,
the reduction in RMSE is 56.7%. These results indicate that the opportunity to save is
important. This conclusion is in stark contrast to Venti and Wise’s finding that the
variation in wealth is almost entirely attributable to preferences.

Although the magnitude of the results is substantial, a few words of caution are
required. First, there are several differences between Venti and Wise’s data and the PSID
sample used in this analysis. The requirement to observe the same individual over 20
years yields a much smaller sample than that used by Venti and Wise and makes it harder
to distinguish specific attributes. Also, the PSID sample includes only male-headed
households, and on average, men have higher earnings and wealth than women.
Government employees are excluded from Venti and Wise’s sample but are included here.
Finally, Venti and Wise’s wealth measure includes private pension wealth, but the PSID
measure does not.22 As mentioned above, such a descriptive analysis cannot completely
isolate ability from preferences. To the extent that preferences are reflected in the key

righthand side variables, the explanatory power of earnings and life events presented here

 

22 The HRS wealth measure is similar to that in the PSID, but Venti and Wise impute the value of Social
Security wealth and employer-provided pensions.

24

is over-stated. Despite the caveats, the overall conclusion is striking: individuals with
greater ability to save accumulate substantially more wealth than previous estimates

suggest.

25

Appendix
A1. PSID Variables

The PSID’s measure of total wealth includes the net value of business/farm,
checking/savings, debt, real estate, stock, vehicles, home equity, and other savings.
Checking/savings amounts include money in checking or savings accounts, money
market funds, CDs, government savings bonds, and treasury bills. Items included as debt
are credit card balances, student loans, medical or legal bills, and loans from relatives.
Mortgages and vehicle loans are explicitly excluded. The real estate category contains
real estate other than the main home, land, rental real estate, or money owed on a land
contract. Stock also includes mutual funds and investment trusts. Prior to 1994, it
included IRAs and private annuities. In 1994, IRAs and private annuities were
specifically excluded from the category, and beginning in 1999, they were reported
separately. Vehicle wealth includes all cars, trucks, motor homes, trailers, and boats.
Home equity is for the primary residence only. Other savings includes bond firnds, cash
value of a life insurance policy, a valuable collection for investment purposes, and rights
in a trust or estate.

For the sample members with missing data due to the biennial survey schedule,
the next survey is used. For example, the demographic data from the 1999 survey is used
for both 1998 and 1999. Any exceptions to that general rule are noted below.

The inheritance variable is an indicator of whether a lump payment of at least
$10,000 was received in the given year. The PSID collects information about the amount
of the inheritance, but that amount was capped at $10,000 for survey years 1968-1982.

For consistency, no dollar amounts are included in the analysis. In the later years when

26

the survey changes to a biennial schedule, the inheritance variable is created from a
question asking about gifts or inheritances received worth a total of $10,000 or more in
the last five years.23

The PSID collects the number of children living in the household at the time of
the survey. The maximum number of children reported over the time period is interacted
with a variable indicating whether or not the household has any children.

The current report is used as the primary source for marital status and any missing
values are supplemented with the marital history report. Marital status is categorized as
never married, married, widowed, divorced, or separated.

The relationship between wealth and marital status is complex.24 In an effort to
keep the analysis relatively simple but still capture the important dynamics, variables are
created for those who remain married throughout, get married, or whose marriage ends
during the relevant period. The base group is therefore those who are consistently
unmarried (including those who are never married, widowed, divorced, or separated).

Like Venti and Wise, 1 have limited information on health. They use the
individual’s current health status reported at the time of the survey when respondents are
approximately 55 years old. The PSID variable most similar to that comes from a
question that asks the individual to rate his or her health on a scale of one to five. This
question was added to the survey in 1984 and is therefore available at age 55 for most of

the sample. However, as Venti and Wise recognize, a subjective health rating given later

in life is not necessarily an indicator of health over the person’s lifetime.

 

23 The 2003 and 2005 surveys ask only about the previous two years.
24 For more details, see Wilmoth and Koso (2002).

27

The PSID provides a better alternative. The survey includes a series of questions
about missing work due to illness. The questions are: Did you miss any work in [the
previous year] because you were sick, how much work did you miss [because you were
sick], did you miss any work in [the previous year] because someone else in the family
was sick, and how much work did you miss [because someone else in the family was
sick]?25 Although it is not a direct measure of health, these questions give an indication
of the individual’s health during their lifetime. Also, because the survey asks about
missing work for family members, at least some information about the health of children
and spouses is available when applicable. That information is used to create a variable
equal to the total number of work weeks missed due to either own illness or that of a
family member.

Additional variables considered in the analysis are education and race. The PSID
collects the number of years of education completed to date. 1 use the highest level of
1.26

education reported. The race reported in the last survey is used for each individua

Possible race categories are white, black, Hispanic, and other.

 

25 For survey years 1968-1975, work missed due to own illness and work missed to illness of another
family member were combined.

6 The one exception is the second cohort. Race was not collected in the 1994 survey, so 1 use race from
the 1993 survey for those individuals.

28

A2. Regression Coefficients
Below are the regression coefficients and standard errors for the regression

including all of the righthand side variables discussed in the analysis.

 

Table A1.]: Overall Regression Results

 

 

 

 

Variable Coefficient Standard Error
Long-run Earnings 6.56 5.53
Long-run EamingsAZ -1.30E-04 4.30E-05
Long-run EamingsA3 4.61E-10 1.27E-10
Long-run Eamings"4 -4.44E-16 1.10E-16
Cohort 2 70,073.01 90,078.22
Cohort 3 14,397.58 82,878.23
Cohort 4 -38,116.65 81,169.76
Cohort 5 -103,268.30 97,439.50
Inheritance in 308 348,144.80 192,470.70
Inheritance in 308*Long-run Earnings -4.54 1.64
Inheritance in 408 206,562.00 150,031.90
Inheritance in 40s*Long-run Earnings -0.87 1.43
Inheritance in 508 222,878.30 161,233.10
Inheritance in 50s*Long-run Earnings -2.27 1.62
Children in 30s 130,225.80 66,424.47
Children in 305*Long-run Earnings -2.11 0.81
Children in 408 -101,886.00 73,633.51
Children in 403*Long-run Earnings 2.05 0.88
Children in 505 98,733.86 66,844.20
Children in 508*Long-run Earnings -2.45 0.72
Married in 305 808,756.60 507,622.50
Married in 305*Long-run Earnings -9.18 6.26
Married in 40s -2,100,093.00 349,129.60
Married in 40s*Long-run Earnings 27.81 3.83
Married in 503 810,425.80 462,516.40
Married in 503*Long-run Earnings -9.45 5.56
Get Married in 303 929,100.70 506,981.80
Get Married in 305*Long—run Earnings -12.64 6.26
Get Married in 403 -604,210.90 503,660.20
Get Married in 40s*Long-run Earnings 9.46 5.88
Get Married in 505 -53,254.87 401,380.50
Get Married in 50s*Long—run Earnings —0.31 4.95

29

 

Table A1.1 (cont'd).

 

 

 

Variable Coefficient Standard Error
End Marriage in 30s —1,121,914.00 279,345.90
End Marriage in 308*Long-run Earnings 14.50 3.26
End Marriage in 403 -981,752.00 530,221.80
End Marriage in 40s*Long-run Earnings 12.43 6.52
End Marriage in 50s 379,260.30 570,949.90
End Marriage in 503*Long—run Earnings -7.92 6.98
Missed Work in 303 -9,408.79 7,765.24
Missed Work in 303*Long—run Earnings 0.24 0.12
Missed Work in 405 41,476.85 4,792.58
Missed Work in 403*Long-run Earnings 0.02 0.06
Missed Work in 508 1,054.50 7,228.51
Missed Work in 505*Long-run Earnings 0.05 0.10
Black -102,602.80 1 10,565.60
Hispanic -472,740.80 298,870.10
Other Race 233,948.90 195,913.90
Education -5,694.99 12,675.72
Std. Dev. Of Earnings 19.29 2.60
Std. Dev. Of Eamings/Mean Earnings -15,549.97 174,533.60
Constant -90,746.88 275,593.90
Observations 806

R-squared 0.6180

 

Note: Dependent variable is retirement wealth.

Source: Author's calculations from PSID

A3. Alternative Analyses
Censored v. Uncensored Long-run Earnings

When examining the effect of censored earnings data, sample size will affect the
results. There will be fewer observations in each decile for smaller samples, making it
more likely that observations will change categories. Since my sample size is smaller
than Venti and Wise’s, I also follow the same procedure with quintiles. As shown in
Table A12, a substantial number of observations (257) change categories and the effect

is largely concentrated among the upper quintiles.

30

 

Table A1.2: Comparison of
Long-run Earnings Quintiles with
Censored and Uncensored Data

Uncensored Quintile
1 2 3 4 5

 

 

 

 

 

0

3g 1 181 7 — 1 -
5, 2 8 157 19 3 2
“,3 3 - 25 120 24 20
§ 4 - - 50 101 38
Q)

U 5 - - - 60 128

 

Note: Sample size is 944
Source: Author's calculations from PSID

SEO Sub-sample with Weights

The PSID initial sample of approximately 4,800 households comes from two
sources. The first 1,872 households are from the sample used in the 1966-1967 Survey of
Economic Opportunity, which was 3 Census study to analyze the effect of the War on
Poverty. That sub-sample is composed primarily of low-income households and is
referred to as the SEO sample. To make the sample nationally representative, a cross-
section sample of 2,930 households was added from the Survey Research Center’s
national sampling frame. That segment of the sample is called the SRC.

In each year of the survey, the PSID provides weights to appropriately combine
the two sub-samples and make the sample nationally-representative. However, given the
structure of the sample in which households are combined from different waves of the
survey, it is unclear which weights should be used. I therefore elect to present the results
excluding the SEO sample members. However, I also performed the same analysis using
the full sample. The table below includes results using the 1989 family weights for all

households as well as those using family weights for the last year the male is observed in

31

the sample (which is approximately age 55). Overall, the central finding remains
unchanged: substantially more of the variation in retirement wealth is attributable to long-

run earnings and life events when better measures are employed.

 

Table Al.3: Regression Goodness of Fit with PSID

 

 

 

 

 

 

 

Family Weights
Percent Reduction in RMSE

1989 Age 55
Control Variables Family Weights Family Weights
Long-run Earnings 27.60% 25.56%
Life Events 14.46% 7.33%
Percent of Variation Attributable to
Long-run Earnings and Life Events 42.06% 32.88%
Notes:

‘ Long-run Earnings regression is retirement wealth on long-run earnings,
long-run earnings polynomial terms, and cohort indicator dummy variables.

2 Life events include inheritance, children, marital status and health. All life
event variables are interacted with long-run earnings.

3 Sample size for 1989 family weights is 1,123; sample size for Age 55
family weights is 1,106.
Source: Author's calculations from PSID

Excluding Home Equity

The PSID total wealth measure used in the main analysis includes home equity,
and this is consistent with the wealth measure in Venti and Wise’s analysis. However,
the retirement wealth literature contains extensive discussion as to whether it is correct to
include home equity in retirement wealth (for instance, see Gale 1997). I therefore
complete the same analysis using a measure of total wealth that excludes home equity.
The results, shown in Table A1.4 below, verify that the explanatory power of long-run

earnings and life events is quite similar for the two wealth measures.

32

 

Table A1.4: Regression Goodness of Fit

 

 

Excluing Home Equity
Incremental
Specification Effect
Long-run Earnings 23.76%
Life Events 20.78%
Total 44.54%

 

Note: Specification of regressions is identical to
that presented in the main text of the paper.
Source: Author's calculations from PSID

Employing Family Equivalence Scale

Married households should save more since they must finance retirement for two
people instead of one. Some research (such as Wilmoth and Koso 2002) analyzes wealth
per person instead of household wealth, and poverty research frequently applies family
equivalence measures to adjust for family size. I use the National Academy of Sciences
adjustment of (A+PK)"F where A is the number of adults, K is the number of children, P
is the relative weight of children (recommended to be 0.7), and F is the family equivalent
(recommended to be between 0.65 and 0.75, so I use 0.7). Wealth is adjusted based on
family size at the time of the wealth measure. Those results are presented in Table A1.5
below. The total amount of variation explained is similar, but here more is attributed to

life events than to earnings.

 

Table A1.5: Regression Goodness of Fit
with Family Equivalence Scales

 

 

Incremental
Specification Effect
Long-run Earnings 16.08%
Life Events 26.11%
Total 42. 19%

 

Note: Specification of regressions is identical to
that presented in the main text of the paper.
Source: Author's calculations from PSID

33

A4. Wealth-Earnings Distribution Comparison

Table A1.6 below compares median wealth by long-run earnings decile for the
HRS sample used in Venti and Wise (2001) and the PSID sample used in this analysis.
The only clear difference between the two wealth measures is that the Venti and Wise
wealth amounts include pensions.

PSID wealth is considerably larger. Although it is not possible to trace the exact
cause for this difference, it is most likely due to the fact that Venti and Wise include
women in their sample. As discussed in the literature review, the wealth gap between
men and women is substantial. The larger wealth in my sample may also be contributing

to the larger results.

 

Table A1.6: Sample Comparison of Median Wealth
by Long-run Earnings Deciles

 

 

 

 

 

 

 

Long-run Median Wealth
Earnings Venti Censored Uncensored
Decile & Wise PSID PSID
1 $5,000 $28,750 $26,320
2 34,429 66,332 65,346
3 52,803 85,298 85,298
4 82,620 185,008 156,263
5 105,166 187,941 136,904
6 144,188 159,522 202,207
7 189,832 233,331 214,766
8 221,692 255,636 287,453
9 305,536 360,100 376,950
10 387,609 375,003 646,225
Notes:

1 Venti & Wise wealth includes business equity, personal financial assets, real
estate, personal retirement assets (including IRA and 401(k) balances),
vehicles, home equity, and pensions; PSID wealth includes business/farm
equity, checking/savings accounts, real estate, private annuity/IRA, stock,

2 PSID wealth values are given in 1992 dollars to be more comparable with
Venti and Wise.

Sources: Venti and Wise (2001) and author's calculations from PSID

34

References

Altonji, Joesph G., Ulrich Doraszelski, and Lewis Segal. 2000. “Black/white differences
in wealth.” Economic Perspectives, Federal Reserve Bank of Chicago, Winter 2000.

Bajtelsmit, Vickie L. and Alexandra Bemasek. 1996. “Why Do Women Invest
Differently Than Men?” Financial Counseling and Planning, 7: 1-10.

Blau, Francine D. and John W. Graham. 1990. “Black-White Differences in Wealth and
Asset Composition.” Quarterly Journal of Economics, 105(2): 321-339.

Cagetti, Marco and Mariacristina De Nardi. 2005. “Wealth inequality: data and models.”
Federal Reserve Bank of Chicago Working Paper 2005-10.

Citro, Constance F. and Robert T. Michaels, Editors. 1995. Measuring Poverty: A New
Approach. National Academy Press, Washington, DC.

Gale, William G. 1997. “The Aging of America: Will the Baby Boom Be Ready for
Retirement?” The Brookings Review, 15(3): 4-9.

Gittleman, Mary and Edward N. Wolff. 2000. “Racial Wealth Disparities: Is the Gap
Closing?” Working Paper.

Hurst, Erik. 2006. “Grasshoppers, Ants and Pre-Retirement Wealth: A Test of Permanent
Income Consumers.” NBER Working Paper 10098.

Hurst, Erik, Ming-Ching Luoh, and Frank Stafford. 1998. “Wealth Dynamics of
American Families: 1984-94.” Brookings Papers on Economic Activity, 1998(1): 267-
337.

Juster, F. Thomas and James P. Smith. 1997. “Improving the Quality of Economic Data:
Lessons from the HRS and AHEAD.” Journal of the American Statistical Association,
92(440): 1268-1278.

Kazarosian, Mark. 1997. “Precautionary Savings — A Panel Study.” The Review of
Economics and Statistics, 79(2): 241-247.

Lupton, Joseph P. and James P. Smith. 2003. “Marriage, Assets, and Savings.” In
Marriage and the Economy: Theory and Evidence from Advanced Industrial Societies,
edited by Shoshana A. Grossbard-Shechtman, 129-152. New York: Cambridge
University Press.

Menchik, Paul L. and Nancy Ammon Jianakopolos. 1997. “Black-White Wealth
Inequality: Is Inheritance the Reason?” Economic Inquiry 35(Apri1 1997): 428-442.

35

Schmidt, Lucie and Purvi Sevak. 2006. “Gender, Marriage, and Asset Accumulation in
the United States.” Feminist Economics, 12(1-2): 139-166.

Smith, James P. 1995. “Racial and Ethnic Differences in Wealth in the Health and
Retirement Survey.” The Journal of Human Resources, Special Issue on the Health
and Retirement Survey, 30(1995): S158-S183.

Smith, James P. 1997. “Wealth Inequality Among Older Americans.” The Journals of
Gerontology, Series B, 523(Special Issue): 74-81.

Venti, Steven F. and David A. Wise. 2001. “Choice, Chance, and Wealth Dispersion at
Retirement.” In Aging Issues in the United States and Japan, edited by Seiritsu Ogura,
Toshiaki Tachibanaki, and David A. Wise, 25-64. Chicago: The University of
Chicago Press.

Wilmoth, Janet and Gregor Koso. 2002. “Marital History and Wealth Outcomes.”
Journal of Marriage and Family, 64 (February 2002): 254-268.

Wolff, Edward N. 1998. “Recent Trends in the Size Distribution of Household Wealth.”
Journal of Economic Perspectives, 12(3): 131-150.

36

Irreconcilable Differences:

A Gender Comparison of Post-divorce Economic Wellbeing

1 . Introduction

It is well documented that an individual’s income decreases substantially
following divorce and that the decline is much greater for women than for men.27 Wealth
is a key component of economic wellbeing, but it is largely absent from the literature
examining the economic effects of divorce.28 In general, wealth can serve as a buffer in
times of decreased income, but there is added importance in the context of divorce.
Allocation of marital assets between spouses is central in divorce proceedings, and thus it
is possible that trade-offs between income and wealth are made among spouses.

Analyses of the economic effects of divorce that exclude wealth do not account for this
possibility and thus may misrepresent the economic effect of divorce and the related
gender inequality. No study to date has looked at whether the inclusion of wealth alters
the conclusions based on income alone—in other words, whether differential wealth
allocations are used to compensate divorcees with poor earnings prospects.

The goal of this analysis is to gain a better understanding of the overall economic
effect of divorce and the observed gender inequality. I use the Panel Study of Income
Dynamics (PSID) to quantify the income and wealth changes following divorce.
Repeated wealth supplements in the PSID enable examination of longer-term effects of
divorce to see whether individuals recover from divorce and whether the recovery differs

by gender. Given the central role of state divorce laws in the allocation of financial

 

27 See Holden and Smock (1991) for a review of this literature.

28 Exceptions include Lupton and Smith (2003), Schmidt and Sevak (2006), Wilmoth and Koso (2002),
and Zagorsky (2005).

37

resources upon divorce, differences in state divorce laws and their effect on measures of
wellbeing are also considered.

Overall, the findings indicate that studies based only on income understate the
economic effects of divorce and the related gender inequality. Although hindered by
small sample sizes, the analysis of longer-term economic effects of divorce suggests
divorced individuals are able to return to pre-divorce income and wealth, but they are
further behind married couples and the gender differential persists. Differences also exist
across legal regime. Individuals who divorce in states with unilateral divorce laws
experience smaller financial declines than those in mutual divorce states. Despite
intentions to yield fair outcomes in which spouses with lower income potential receive a
larger portion of marital assets, there is no evidence of compensatory wealth allocation in
states with equitable distribution laws. In general, the legal environment has a larger

effect on wealth than income.

2. Divorce Laws and- Terminology

A typical divorce settlement details the division of assets between spouses and
child custody arrangements (where applicable). It may also include spousal support (also
known as alimony) and/or child support payments. Spousal support is regular payments
from one spouse to the other that serve to supplement income. These payments may be
required indefinitely, for a set period of time, or until that spouse remarries.29 Child

support payments are made to the spouse who retains primary custody of the children

 

29 Whether marital misconduct is considered in determining alimony awards varies by state.

38

until the children reach age 18.30 A divorce settlement may be reached informally
between the spouses, through more formal arbitration proceedings, or by the court, but all
agreements must be formally approved by the court.

Divorce is legislated at the state level in the United States. Hence, each state has
its own laws regarding the requirements for divorce, which assets are subject to division,
and how those assets are allocated between spouses. Although each state law is unique,
there are common themes that permit broad classifications. Here I discuss three separate
types of laws: grounds for divorce, type of property division, and whether fault grounds
are considered in the division of property.

The grounds for divorce can generally be described as mutual or unilateral. States
with mutual divorce laws require both spouses to agree to divorce. In unilateral divorce
states, one spouse can attain a divorce without the other spouse’s consent. The
mutual/unilateral distinction is commonly used in the economics literature and is adOpted
here. Table A21 in the appendix indicates which states have unilateral divorce laws.

Another common classification scheme is fault vs. no-fault divorce. The terms
“no-fault” and “unilateral” are often used interchangeably, but they have different
implications. No-fault refers to laws that allow a divorce to take place without assigning
guilt to either party. However, the absence of guilt requirements does not imply
unilateral divorce. For instance, some states with no-fault laws still require a mutual
agreement to divorce or lengthy periods of separation prior to divorce. Unilateral divorce
implies both no-fault laws and the absence of such restrictions that prevent either party

from initiating a divorce without consent from the other party.

 

30 Child support can also occur when neither parent has custody. In this case, payments are made to the
third party caring for the child.

39

There are three primary types of property division laws in the United States:
community property, common law, and equitable distribution. Community property
states give both spouses equal ownership of marital property and divide assets 50-50
upon divorce. In common law states, division of the property is based on which spouse
holds the title. States with equitable distribution laws empower the courts to determine a
“fair” division of assets. Of course, “fair” does not necessarily mean equal. Table A2.]
in the Appendix gives the type of property division laws for each state and changes that
occurred during the sample period.

A closely-related issue is whether fault is relevant in property division. Even
states that eliminated fault-grounds for divorce may still allow fault to be considered in
dividing assets. For example, in the community property states of Nevada and Texas, the
party at fault surrenders his or her right to part or all of the marital property. In some
equitable distribution states, fault is considered in determining a fair allocation of assets.
Table A2.1 indicates whether states eliminated fault grounds from property division and

the year the change occurred, where applicable.

3. Literature Review

The literature on the effects of divorce is vast, so here the focus is restricted to the
two branches that are most relevant to this analysis. The first examines the correlation
between divorce and measures of financial resources. The second analyzes the effects of

divorce legislation on post-divorce finances.

40

3.1 Income and Wealth

Early studies investigating the income effects of divorce relied on cross-sectional
data. More recent studies use longitudinal data, and the sample is generally composed of
women who are married initially and separate or divorce during the relevant time period.
A variety of measures such as household income, income-needs ratio, per capita income,
and poverty rates are present in the literature. Some studies (such as Hoffman 1977 and
Duncan and Hoffman 1985) perform simple before and afier comparisons, but others
(including Jacob 1989 and McKeever and Wolfinger 2001) control for observable
characteristics such as education, race, age, length of marriage, and macroeconomic
indicators.

Holden and Smock’s 1991 review article provides an excellent summary of this
literature and identifies three central themes: women’s economic wellbeing is worse after
divorce, men fare substantially better than women following divorce, and the effect on
women is prolonged in the absence of remarriage. Specifically, the referenced articles
find a decline in women’s income of 22-71 percent following divorce. Differences in the
magnitude of the effect are primarily attributed to differences in sample composition
(such as how divorce is defined and whether those who remarry are included in the
sample).

Relatively little work has examined the wealth effects of divorce, largely due to
data availability. Until recently, most surveys either did not include wealth or measured
it poorly (Browning and Lusardi 1996). With the advent of reliable wealth surveys such
as the PSID wealth supplement and the Health and Retirement Study (HRS), researchers

began analyzing wealth differentials by marital status.

41

Lupton and Smith (2003) use the HRS and PSID to examine the relationship
between wealth and marital status. Using the baseline HRS survey, they find that
separated and divorced individuals have substantially less wealth than those who are
widowed, never married, and married. There is also evidence that the duration of marital
status matters; those who are married longer have even more wealth, and longer duration
of divorce is associated with even less wealth. Turning to the PSID and exploiting its
panel structure, Lupton and Smith find that, among married couples, those who
eventually divorce have less wealth than those who remain married, even after
controlling for differences in income.

Schmidt and Sevak (2006) use the PSID to examine cross-sectional wealth
differences by marital status. Their results are largely consistent with those of Lupton
and Smith, but they also find that there is no wealth gap by marital status (or gender) for
those between the ages of 25 and 39 in 2001. This suggests that either the wealth gaps
are disappearing in general or that they do not develop until later in life.

Wilmoth and Koso (2002) use the HRS to more closely examine the effects of
marital status. They utilize the survey’s marital history to include detailed marital status
categories instead of the current marital status used in most studies. Results indicate that
significant wealth gaps exist across marital status, but the findings regarding gender
differentials are mixed. They emphasize that researchers should include variables
indicating the sequence of marital events whenever possible.

Zagorsky (2005) investigates the correlates of wealth with the NLSY. He finds
that wealth begins to fall about four years prior to divorce, and that, on average, divorced

individuals have 77 percent less wealth than singles. Also, while in percentage terms

42

women are more affected by divorce than men, the dollar difference by gender is

relatively small.

3.2 Divorce legislation

Unlike the analyses discussed above that are limited to identifying correlations,
the literature regarding unilateral divorce attempts to uncover the causal effect of divorce
laws on financial resources. Researchers exploit the variation in timing of laws across
states to identify the effect of unilateral divorce legislation on a variety of outcome
measures. Most of the unilateral divorce literature is focused on identifying the effect of
these laws on divorce rates, but here the focus is on the financial effects.

The effect of unilateral divorce on divorce settlements is discussed in Weitzman
(1985) and Peters (1986). Weitzman’s analysis is a before and after comparison for
women in California. Peters uses the 1979 Current Population Survey (CPS) and exploits
differences across states. Both studies find that divorce settlements received by women
are smaller under unilateral divorce laws. Weiss and Willis (1993) come to the opposite
conclusion; they find that, although not statistically significant, the divorce transfer to the
wife is actually larger in the absence of fault grounds.3 1 They attribute the contradictory
findings to their inclusion of differences in property division laws across states. However,
as they note, only 5% of their sample divorced under no-fault laws, hindering the
dependability of their results. Their sample is also restricted to individuals in the
National Longitudinal Study of the High School Class of 1972 who divorce by 1986.

Thus, sample members are fairly young and have relatively short marriages.

 

3' Instead of the mutual/unilateral distinction used in much of the literature, Weiss and Wills divide states
into three categories: fault, no fault (including those that require a separation between 6 months and 3
years), and choice of fault grounds.

43

Gray (1996) examines women in the CPS who divorced between 1975 and 1979
and employs an interaction of unilateral divorce and property division laws. He finds that
women receive significantly smaller settlements in states with unilateral divorce and
common law property division, but unilateral divorce in states with community property
laws does not affect the settlement size. He concludes that, although researchers agree
that women are banned economically by divorce, “the extent to which no-fault divorce is
responsible remains controversial” (281).32

Jacob (1989) argues that overall income is of more interest than the size of
divorce settlements because so few women receive spousal support and most do not ever
receive the child support awarded to them by the court. He finds that women who
divorce under no-fault laws have significantly higher post-divorce income than those
under the fault regime. However, when the unilateral/mutual classification is employed,
the effect becomes negative (but statistically insignificant). He thus concludes that

changes in divorce laws had a negligible, if any, effect on women’s finances.

4. Conceptual Framework

A framework in which to think about divorce is helpful to understand the transfers
that occur with divorce and how those transfers vary by legal regime. To begin with, the
financial resources (R) available to each spouse (referred to generally as person 1 and

person 2) during marriage are given as
Rl|m=l = amiy(a1 + a2 + am +y1 + yz) (1)

and

 

32 Gray notes that he follows Peters (1986) and classifies states as having unilateral or mutual divorce laws,
but he uses the terms “unilateral” and “no-fault” interchangeably in the paper.

44

R21m=1 = 0lmz)’(ai + a2 + am “'10 + Y2) (2)
where a] and a2 represent the annuitized value of pre-marital assets of each spouse, am is
the annuitized value of marital property, y. and y; are earnings for each person, and y
denotes returns-to-scale from being married. During marriage, total resources are
allocated between spouses such that person 1 receives 01ml portion and person 2 gets am;
portion. Because of the presence of public goods (such as the home), the sum of am; and
(Imz may be greater than one.

If divorce occurs, the resources are divided such that

Rllm=0 = adam + a1 + B1311 + (1-Bz)y2 — c (3)
and
R21m=0 = (l'ad)am + a2 + (1'131)YI + 325’2 — C- (4)

All retums-to-scale are lost, and marital assets are divided such that ad is allocated person
1 and (1- ad) is allocated to person 2. It is assumed that all marital assets, including
public goods, must be liquidated and divided among spouses, so the two portions sum to
one. It is also assumed that individuals retain all of their personal assets. This is true in
most states, but the definition of personal assets varies across states, and some states
allow them to be divided upon divorce. The possibility of child and/or spousal support
payments is represented by [31 and [32. [3. is the portion of person 1’s earnings he or she
retains, and (l-Bz) is the portion of person 2’s earnings allocated to person 1 in the form
of support payments. If person 1 pays support to person 2, [31 is between zero and 1 and
B; is equal to 1. Conversely, if person 2 pays support to person 1, [3. is equal to 1 and [32

between zero and 1. In the absence of child and spousal support, both [3. and [32 equal 1.

45

Finally, individuals incur a cost of divorce, c, in legal fees and other related expenses.

For simplicity, it is assumed that cost is the same for both spouses.

4.1 Mutual/Unilateral Divorce

When one spouse prefers marriage and the other prefers divorce, the marital
outcome and post-divorce (if it occurs) financial resources of each spouse depends on the
applicable state divorce laws. For ease of exposition, it is assumed throughout the
analysis that person 1 wants a divorce and person 2 prefers to remain married.

If person 2 prefers marriage, divorce can occur under mutual divorce laws if a
transfer between spouses can entice that individual to divorce. In this case, the post-

divorce resources of each spouse are given by

Rl|m=0 : adam + a1 + BIYI +(1'B2)YZ — C _t (5)
and
R2|m=0 =(1-Old)am + a2 + (1'503’1 + 132Y2 — C +1 (6)

where t represents the transfer between spouses. The upper bound on t comes from the
amount of financial assets available to person 1. Those include a. and person 1’s portion
of marital assets. However, it is likely that person 1 will pay the minimum amount
required to entice person 2 to divorce, and that amount cannot be defined in this simple
framework that does not include non-monetary factors.

Adoption of unilateral divorce means that transfers to entice divorce are no longer
required; anytime divorce is preferable to marriage for either spouse, the couple will

divorce. The post-divorce resources of each spouse after divorce are given by

R1|m=0 = (163m + 31+BIY1+(1'132)Y2 — C (7)

46

and
R2|m=0 = (l'ad)am + 32 + (I'BI)Y1 + 1321’2 — C (8)

The difference in post-divorce resources between mutual and unilateral divorce is
the amount of the transfer, t. Thus, for couples where one spouse prefers marriage and
the other prefers divorce, the framework predicts that adoption of unilateral divorce will
affect the economic wellbeing of spouses following divorce. However, the direction of
that difference hinges on which spouse prefers divorce, not gender. Without knowledge
of such preferences, it is impossible to predict the presence or direction of a gender
differential. For couples where both spouses prefer divorce, divorce occurs under both
mutual and unilateral divorce laws and adoption of unilateral divorce laws does not affect
their post-divorce financial resources.

There is potentially one other difference between mutual and unilateral divorce.
Proponents of unilateral divorce claimed that it would diminish the adversarial nature of
divorce (Bahr 1983 and Weitzman 1985). If so, unilateral divorce may eliminate the
need for drawn-out legal battles, thereby reducing the legal costs associated with divorce.
This change would be incorporated in the conceptual framework as a reduction in c in

unilateral states, resulting in higher post-divorce financial resources for both spouses.

4.2 Property Division

Differences in state property division laws are captured in the framework by ad.
In equitable distribution states, the emphasis is on fairness instead of equality. Several of
these states explicitly direct judges to take into account differences in earnings power

(among other things) when allocating assets between spouses (Freed and Foster 1973).

47

This implies that spouses with relatively low post-divorce income levels should be
compensated with a relatively larger share of the marital wealth (referred to herein as
compensatory wealth allocation). Due to the compensatory nature of the law, the exact
value of Old will vary on a case-by-case basis.

In community property states, assets are to be allocated equally between spouses,
so ad=1/2. Because ad is specified by the law, compensatory wealth allocation is not
expected in those states. Marital property is assigned to the individual who holds the title
in common law states, so the exact value of ad is unknown. However, all but one state
(Mississippi) with common law property division rules adopted equitable distribution by
1985 (Weitzman 1985).

States also differ on whether fault is taken into account in the division of marital
assets. As states began removing the fault grounds for divorce, the question arose as to
whether to retain fault grounds in asset division. Some states removed fault from asset
allocation decisions at the same time they removed the fault requirement to obtain a
divorce, some states did so within a few years, and other states still retain fault
considerations in property division (Freed and Foster 1973 and Stevenson and Wolfers
2007). When person 1 is at fault, ad will be smaller in states with fault considerations,
but the opposite is true if person 2 is at fault. Because ad affects wealth but not income,

no-fault property division should have a larger effect on wealth than income.

4.3 Implications

It is useful to summarize what the framework predicts regarding post-divorce

financial resources. First, total financial resources will decline at divorce. Previous

48

research documents a larger decline in income for women than for men, but wealth must
be included to understand the total financial effects of divorce and determine whether a
gender differential exists.

The transition from unilateral to mutual divorce has two financial effects. First,
for couples where one spouse prefers divorce and one spouse prefers marriage, post-
divorce resources will change by the transfer, t, with unilateral divorce. Since the
framework does not predict whether men or women are more likely to prefer divorce, it is
not clear whether this change will have differential effects by gender. Second, if
unilateral divorce does in fact reduce the legal costs of divorce, the size of c will decrease
and financial wellbeing for all divorced individuals will be greater.

Property division laws affect ad, but the framework does not indicate whether
differences should exist by gender. The emphasis on equality in equitable distribution
states suggests more compensatory behavior is expected there than in other states.
Additionally, removal of fault grounds from asset allocation will help one spouse and

hurt the other, but whether there will be a gender differential is not known.

5. Data

This analysis utilizes the PSID, a longitudinal dataset that began in 1968.
Although it was initially an annual survey, it changed to a biennial schedule in 1997.
Data are currently available through 2005. The survey collects a broad range of social
and economic information from households.

Two components of the survey are of particular interest here. First, the PSID

maintains a marital history for each household. It includes date of marriage, current

49

status, and dates of separation, divorce, and widowhood where applicable for up to ten
marriages. Second, the PSID contains a supplementary questionnaire in the 1984, 1989,
1994, 1999, 2001, 2003, and 2005 surveys to collect comprehensive wealth information.
The supplement employs unfolding brackets to minimize potential bias from non-
response.

The primary sample includes all individuals who divorced between adjacent
wealth supplements. Because individuals who become part of the PSID sample through
marriage are no longer surveyed after divorce, it is not possible to follow both spouses
after divorce. Only the spouse who is originally a member of the PSID sample is
included. Thus, comparisons of outcomes for men and women rely on the sampling
frame to provide representative samples of both genders. For comparison, I create a
sample of couples who are married and living together consistently between adjacent
wealth supplements.

The divorce sample includes 928 individuals, of which 502 are women and 426
are men. The married sample consists of 12,406 couples. By construction, it is possible
that the same couple appears more than once in the married sample. For all analyses, I
use the PSID family sampling weights from 1984 to make the sample nationally
representative.33

Table 2.1 provides summary statistics for divorced individuals (overall and by

gender) and married couples. Married households are older, have higher average

household income, and have been. married longer. They are also less likely to have

 

33 The analysis was also completing using individual weights from 1984 and weights (both individual and
family) from the last year observed. The effect of different weights on the results is minimal.

50

children living in the household, most likely due to the age difference. Additionally,

there are slightly more minorities in the divorced sample.

 

Table 2.1: Summary Statistics
All Divorced Individuals (N=928) Married Couples (N=12,406)

 

 

 

Mean Std. Dev. Mean Std. Dev.
Female 0.52 0.50
Education 13.27 2.13 13.47 2.61
If Children 0.57 0.49 0.49 0.50
Age 36.70 9.57 50.07 14.59
Household Income 63,769 41,592 88,108 104,349
Black 0.07 0.26 0.06 0.23
Other Race 0.05 0.22 0.03 0.18
If Remarry 0.26 0.44
Length of Marriage 12.37 8.69 22.48 15.05
Unilateral 0.58 0.49 0.55 0.50
No-Fault Property 0.39 0.49 0.46 0.50
Common Law 0.02 0.12 0.01 0.10
Equitable Distribution 0.80 0.40 0.80 0.40

Divorced Females (N=502) Divorced Males (Nfl26)

Mean Std. Dev. Mean Std. Dev.
Education 13.37 2.08 13.16 2.18
If Children 0.71 0.45 0.42 0.49
Age 36.17 9.83 37.28 9.26
Household Income 61,932 40,426 65,773 42,783
Black 0.07 0.26 0.07 0.25
Other Race 0.06 0.24 0.03 0.18
If Remarry 0.23 0.42 0.29 0.45
Length of Marriage 12.69 9.04 12.02 8.28
Unilateral 0.61 0.49 0.56 0.50
No-Fault Property 0.39 0.49 0.39 0.49
Common Law 0.02 0.13 0.01 0.12
Equitable Distribution 0.79 0.41 0.81 0.40

 

Note: Results weighted with 1984 PSID family weights.
Source: Author's calculations from PSID

Overall, the divorced males and females are comparable. Women are more likely

to have children, and that is consistent with findings in the CPS that 90 percent of

51

children live with their mother (Bianchi and Spain 1986). Men are also slightly more
likely than women to remarry within the first few years following divorce.

Two aspects of my divorced sample warrant further discussion. As mentioned in
the Literature Review, estimates of the income effects of divorce vary widely, and that is
largely attributed to differences in sample composition. Two primary issues are how
divorce is defined and whether those who remarry are included. Much of the literature
includes all individuals whose relationship ends, be it through separation or divorce.
Because this analysis examines differences in state divorce laws, the sample is restricted
to those who formally divorce. The second issue concerns those who remarry. Much
research documents that remarriage greatly increases recovery, so some studies elect to
exclude individuals who remarry from their sample (Holden and Smock 1991). However,
I am interested in the effects of divorce for all divorced individuals, not just those who do
not remarry, so I include all divorced individuals in the sample.

The analysis employs three primary measures of economic wellbeing: wealth,
income, and total financial resources, all converted to 2005 dollars using the Bureau of
Economic Analysis’ Personal Consumption Expenditures (PCE) price index. The wealth
variable is household total net wealth from the PSID wealth supplement. It includes
business/farm assets, checking/savings accounts, debt, real estate, stock, vehicles, home
equity, and other savings. Because the wealth supplement is only available in some years,
wealth cannot be examined immediately before and after divorce. Recall, the sample is

composed of those who divorce between wealth supplements, so the pre-divorce wealth

52

supplement is used as the Before wealth measure, and the next wealth supplement is the
After measure.34

The income measure is total family income and is composed of all taxable and
transfer income. It includes amounts received from spousal support and child support
and excludes spousal support and child support payments made. Although income is
available in each year of the survey, income is used for the same years as the wealth

measures for consistency.

 

Table 2.2: Incidence and Magnitude of Spousal Support
and Child Support for Divorced Individuals
Paid Received
Number Amount Number Amount

 

 

 

 

 

Spousal Support

All Divorced Individuals 26 14,013 23 9,374
Women 3 26,941 23 9,374
Men 23 11,416 0 -

Child Support

All Divorced Individuals 185 7,322 166 5,068
Women 42 8,158 157 4,655
Men 143 7,059 9 9,491

Notes:

I Results weighted with 1984 family weights.

2 Amount is the average of those with non-zero values.
Source: Author's calculations from PSID

Table 2.2 displays the prevalence and amount of spousal support and child
support in the sample. Approximately 4.6 percent of women receive spousal support, and
31.3 percent of women receive child support. These numbers correspond to the national

averages reported by the Census Bureau. They report that, in 1989, 4.5 percent of

 

34 Specifically, for those who divorce between 1985 and 1988, the 1984 and 1989 supplements are the
before and afler measures, respectively. The 1989 and 1994 supplements are used for those who divorce
between 1990 and 1993, and the 1994 and 1999 surveys are used for those who divorce 1995-1998. Slight
adjustments are made in later years when more frequent wealth supplements are available. For those who
divorce in 2000-2002 the 1999 and 2003 surveys are used, and for those who divorce in 2003 or 2004 the
2001 and 2005 surveys are used.

53

divorced women were supposed to receive spousal support payments and 37 percent of
women living with their own children whose fathers were absent received at least part of
their child support awards (Lester 1989).

In addition to examining income and wealth individually, a joint measure of the
two is created to gauge the overall economic effects of divorce. Because income is a
flow and wealth is a stock, the two cannot sensibly be added together. To address this
issue, total wealth is annuitized.35 Life expectancy from the Social Security
Administration is applied, and it is assumed that wealth lasts until the end of life.36 I use
an annual interest rate of 2%. The joint measure, herein referred to as total financial
resources, is the sum of annual income and annuitized wealth.

In relation to the conceptual framework, the values for person 1’s annuitized
wealth are equivalent to (a. + a2 + am) before divorce and (01dam + a1 — c) after divorce. It
is not possible in the PSID to allocate wealth in the household to individual members, so
only the total amount is observed. The inability to assign wealth to individuals also
prevents observation of 01m and (rd, Annual income is equal to (yl + y;) prior to divorce
and (thy. + (l-Bz)y2 - t) afterwards, where t may be equal to zero if no transfer is required.
To the extent that divorces prior to adoption of unilateral divorce did not require transfers
to entice divorce, differences between mutual and unilateral divorce in the sample will be
smaller. Each of the components of income is available in the PSID, but only the total

amount is employed in the analysis.

 

35 The formula used is: annual payment amount=present value of wealth/[(1-(1/(1+i)"n))/i] where n is the
number of periods and i is the annual interest rate.

36 I performed the same analyses annuitizing wealth over 18 years to mitigate the importance of age at
divorce. This change does not affect the findings reported here.

54

The year of divorce and state of residency in that year are used to determine the
appropriate legal variables. Divorce laws by state are in the Appendix (Table A2.1).
Unilateral divorce data are from Gruber (2004). No-fault property division is from
Ellman and Lohr (1998). The type of property division is compiled fiom Gray (1998)

and Freed and Foster (1979 and 1981).

6. The Economic Effects of Divorce

This section analyzes changes in income, wealth, and total financial resources
following divorce. The initial analysis looks at the short-term effects of divorce. The
sample is then restricted to those observed four years after the original post-divorce
measure to provide some insight as to the longer-term financial effects. Finally, a

descriptive exercise is performed to look for evidence of compensatory behavior.

6.1 Short-term Changes

First, changes in income, wealth and total financial resources for both divorced
individuals and married couples are examined. The results are in Table 2.3. The Before
and After measures are as described in the Data section. The change in resources is
calculated for each individual, and the Change columns summarize that variable for the
sample. The Percent Change column shows the percent difference between the sample
Before and After measures. The final measure, Adjusted Percent Change, is less
common. It is similar to the percent change, but it uses the average of Before and After

as the denominator. This adjustment caps the change at plus/minus 200% to limit the

55

effect of outliers and is well-defined even if one of the two wealth quantities is zero.37 Its
unique properties make the Adjusted Percent Change the preferred measure. Both the
mean and median are presented for all measures.

The income disadvantage of divorce is immediately apparent. The mean adjusted
percent decrease for divorced individuals overall is 12%, and the decline for divorced
individuals is substantially greater for women (16%) than men (8%). These decreases are
even greater when compared to the 3% gain in income for married couples.38

The second panel of Table 2.3 displays changes in wealth. Divorced individuals
see a mean decline of over 23%. A gender differential is again evident — women’s wealth
declines by 34%, but men’s falls by only 12%. Not only do divorced individuals
experience a large decline while married couples enjoy a substantial increase, but
divorced individuals also have less initial wealth. This difference is consistent with
Lupton and Smith’s finding that couples who eventually divorce have less wealth than
those who stay married.

A similar pattern exists for total financial resources. The gender gap in the total
resources measure is even larger than for income. Women experience a loss of 16% and
men’s total resources fall by less than 1%. The widening of the gender gap with the
inclusion of wealth suggests that studies based only on income understate the economic

effects of divorce and the size of the gender differential.

 

37 The adjustment perfectly restricts all positive values to the plus/minus 200% range. Because wealth can
take negative values, a cap is imposed for observations where the calculation yielded a larger change.
Overall, this measure handles negative values and transitions between positive and negative values much
better than the traditional percent change.

38 The PSID collects wealth at the household level so it is not possible to consider married men and
married women separately.

56

 

 

 

 

 

 

god food .5; £605 Kmd wot? mfidw #342 03.3 mmode $3.2 “octane/H
6\°N.o- 6\6m.m- Roms- Xmm ow- vow; memdm mmmxé 83% 3%? omv :02
£167 $0.3- o\om.:- ficww- wmcd- omwfi- 2a.? mmmdo mmvdm wmmdo Now 5:8?
Axis. Read- $0.2- 8mg“- owof- afim- mmfifim ammo Rode omimo mam 3835—2
39.38% .22—«Em :33.
$93 $w.mm $6.; Rowan omodm mom.mw mnNWS mum—:4 8ch Kmem 35.2 BEBE
$5.3- .x.w.:- $34- 3N6- www.m- mmm.m_- Smdm @362 53.3 Kai; one 52
fiswwm- .xcoém- Ronda- .xcwdm- 3&5- 334m- 83% Egg 35.94 Qmfifl mmm 5:83
£95m- Axummm- Reg—N- £33.. NNOK- 33%.. Sme oomde ommdv SECS woo; 3835.14.
5.33
Rove Refim Xim €346 mwo.m 90% 3%: 3%.? 32.3 widw 36.2 BEE
fame- Roms- RON—T $0? ~34- vmm- 81mm gnaw www.mm «$5.3 one 32
£6de $5.2- $32- $95- 33$- 35“. omnfiv Emfim 8%? Nmm.$ mom 5803
Rafiw- RENT £3457 $6.? mmVH- SQN- Snow mom.$ macdm oohmo wmm 39839—3.
0E3:—
momnaau 33,—.
5602 .822 $632 :82 .8632 :32 5602 $34 .8602 :32 Z
oWSEO Eooeom A3,. owSEO Scuba owcmao cot/V. 880m

 

82:83— :32.2.3 .3:... ES :53 >9 653:— E mow—5:0 and ~35.

57

9mm 88m macaw—:28 $85364 ”850m

.5233 BNESSE cam 2:83 .3 Sam 05 8e $83832 BEES-m .80-w m

@853 CO Soto

05 :8: 8 $oo~ -\+ 3 owcmno 2t 38 98 835883 05 E Base. 98 880m .20 omega; 05 wow: owSEO :8ch woman}. N

same; base .32 55 BE? 838m _

 

 

 

 

 

”882
$w.w $o.w $24.5 $5.2 emoxn $56 23.? 39$ wwwfim m5o€m 36.2 BEBE
$34.5. $24.3 $N.mm $096. 306 35.2 945.3 wmmwv «33.3 39mm 024 5E
$m.m $m. m $02 $w.2 58.2 comm 5oN.5N 536m $5.3 815 Now 5:53
$ _ .2 $32 $o.o_ $wém vaoxu vofiw 25.2” mmm. 2 v vmw.5m www.mm wmm 3835 :<

89583— 3235.”— .83.
$3.3 $v.v~ $24.3 $w6m 35.3 woodv 241$ .56me @330 EQNS 35.2 onmE
$02 $m.m $26.2 $o.5m awed oofiom oooxum noofo 32.2 @9465 @524 20E
$mdm- $0.8- $5.2- $02- N5m.- m _ N; T $3.3 mmodm $2.52 omeo mmm c0803
$w.©- $09 $92 $m© 3.5- .334 312 ova. 2 5 25.52 85.50 woo; 38925 =<

5.33
$w. m $m .m $v.m $v. m coo; Ncm.m Stem comdv $0.? 30.? 03.2 BEBE
$m.vm $5. _N $m.wm $5.5m N5md $92 35.3“ 93% ooofim 2 fimm cmv 5E
$m.m $0. m $ _ .m $.31 cmw mm 2 .v ooodm Nmmxmm 85.3 2 dam Now 5:83
$ 2 .E $v.2 $w.52 $o©m 23:6. wmo.5 856m vwvdm owodm cumdm wma ©0235 =<
2:3:—
oumm .3:—a...— ueu 13255. 82:25
560E SSE 560E 58E 360E :8E 82on SSE BEBE :moE Z
owSEO “580m mac/w owgno 688m owSEO “8:, 880m

 

3.2—o8 m.~ «Ea-H

58

The divorced/married comparison discussed above is insightful, but it ignores
returns to scale. To correct for this, 1 use the National Academy of Sciences factor to
adjust for family size.39 The results adjusted for family size are in the bottom three
panels of Table 2.3. The same trend is evident in all three resource measures - the mean
adjusted percent change for women is considerably smaller in magnitude than the
unadjusted measures, but the gender gap is substantially larger with the family size
adjustment. For example, the mean adjusted percent change in income is -15.7% for
women and -7.9% for men, yielding a gender gap of approximately 7.8 percentage points.
However, once adjusted for family size, the income gender gap grows to 17.8 percentage
points. The widening of the gender gap is because women are more likely to live with
the children after divorce, so their family size only decreases by one (the husband). On
the other hand, men who leave a household with a wife and children to live on their own
experience a larger decline in family size. The widening gender gap may also indicate
that child support payments do not adequately compensate for the cost of caring for
children.

After adjusting for family size, the effect of divorce for men is quite small.
Indeed, the adjusted post-divorce income levels for men are very similar to those for
married couples. Divorced men still have substantially less wealth than married couples,
but total financial resources are similar for the two groups.

It is possible that the gender differences observed above are attributable to

differences in personal and/or household characteristics. For instance, the greater

 

39 The adjustment factor is (A+PK)"F where A is the number of adults, P is an adjustment for the relative
weight of kids, K is the number of children, and F is the scale economy factor that converts adult
equivalents into comparable units in terms of their efficient use of the family’s resources. The
recommended values of 0.7 for P and F are used (Citro and Michaels 1995).

59

reduction in women’s income may simply reflect that women have less education than
men. To allow for this possibility, I employ multiple regression analysis.40 Most of the
literature focuses on differences between married and divorced women, but the focus
here is on gender differences among divorced individuals. Table 2.4 displays results for
three dependent variables: change in income, change in annuitized wealth, and change in
total financial resources.“ Individual and household characteristics include education,
age, pre-divorce income, race, length of marriage, number of years between divorce and

the post-divorce measures, and cohort indicators.

 

Table 2.4: OLS Regressions of Changes in Resources for Divorced Individuals

 

 

 

Change in Change in Change in Total
Income Annuitized Wealth Financial Resources
Female -7,222 -1,472 -10,718
(3,174)M (1,256) (3,581)***
High School Diploma 17,975 -722 17,857
(5,268)*** (2,084) (5,944)***
College Diploma 28,904 921 31,616
(6,232)" * (2,465) (7,031)***
If Children 953 1,342 2,275
(3,346) (1,324) (3,776)
Age 256 873 910
' (1,021) (404)" (1,151)
Age Squared -9.73 -11.07 -18.67
(11.57) (4.58)M (13)
Pre-divorce Household Income -0.54 -0.08 -0.61
(0.04)*** (0.02)*** (0.05)***
Black -2,664 -780 -3,852
(6,006) (2,376) (6,776)
Other Race -1 ,252 1,134 -1,783
(7,180) (2,840) (8,101)

 

40 Results presented are for OLS regressions. I also performed median regressions, and the results are not
substantially different.

4] Results were comparable when the regression analysis was completed using the adjusted percent change
in income, annuitized wealth, and total financial resources as the dependent variables.

60

 

Table 2.4 (cont'd).

 

 

 

 

Change in Change in Change in Total
Income Annuitized Wealth Financial Resources
Married 6-10 Years -1,413 -1,926 -2,832
(4,339) (1,717) (4,896)
Married 11-15 Years 3,252 -3,371 769
(5,484) (2,170) (6,188)
Married over 15 Years 5,042 -2,849 3,891
(5,542) (2,192) (6,253)
1 Year Since Divorce -11,909 -2,543 -13,842
(4,674)“ (1,849) (5,274)" *
2 Years Since Divorce -8,708 -3,066 -10,703
(4,652)* (1,841)* (5,249)M
3 Years Since Divorce 2,842 -3,755 -554
(4,760) (1,883)" (5,371)
Cohort 2 -6,647 1,011 -6,457
(4,314) (1,706) (4,867)
Cohort 3 5,099 1,143 6,113
(4,437) (1,755) (5,006)
Cohort 4 9,224 56 9,088
(5,318)* (2,104) (6,000)
Cohort 5 4,280 4,116 9,016
(6,371) (2,520) (7,188)
Constant 24,335 -8,330 21,012
(20,226) (8,001) (22,820)
Observations 922 922 922
R-squared 0.22 0.05 0.22
Notes:

I Dependent variables are indicated in column headings.

2 Results weighted with 1984 family weights.

3 Standard errors are in parenthesis.

4 Significance at 10% indicated by *, significance at 5% indicated by **, and significance at 1%

indicated by ***.

Source: Author's calculations from PSID

The female coefficient is negative in all three specifications and is significant in

the income and total resources regressions. This indicates that the gender gap identified

in Table 2.3 is not attributable to the observable personal and household characteristics.

The education coefficients are strongly significant in the income and total resources

regressions and show that those with more education are less harmed by divorce. The
education coefficients are also likely capturing the effect of labor market experience,

which is not included in the regression.

6.2 Longer-term Changes

Relatively little is known about the longer-term effects of divorce.42 Some insight
is available from the repeated wealth supplements in the PSID. For a portion of the
sample, wealth and income four years later than the original post-divorce measure are
available. Thus, a portion of the sample is observed approximately six years after divorce.

The short- and longer-term changes in income, wealth, and total financial
resources are in Table 2.5. The upper panel displays the total changes, and the lower
panel shows the amounts adjusted for family size. Most divorced individuals recover to
or beyond pre-divorce levels, but they are still far below those of married couples. At the
mean, the gender income gap is almost gone while the gender wealth gap remains large.
Looking at the measures adjusted for family size, both divorced men and divorced
women recover beyond pre-divorce financial levels and the gender gap is smaller. The
effect of the family size adjustment on the gender gap is different in the long run most
likely because most men have remarried by the later measures and therefore have larger
households than immediately after divorce. Even though the gender gap is smaller, it is
still substantial. Longer-term changes in income and total financial resources for men

and married couples are very similar.

 

42 Studies examining the long-term effects of divorce are Duncan and Hoffman (1985), Stirling (1989), and
Peterson (1989).

62

 

Table 2.5: Longer-term Changes in Income, Wealth, and Total Financial Resources

 

 

 

 

 

 

 

 

 

 

 

Short-term Change Longer-term Change
N Mean Median Mean Median

Total Changes

Income

All Divorced 535 -3,331 -6,707 9,647 1,755
Women 288 -6,608 -10,1 10 8,413 -56
Men 247 102 -2,676 10,940 5,357

Married 5,649 1 1,245 5,606 13,648 5,817

Wealth

All Divorced 586 -28,01 1 -6,649 48,428 4,601
Women 307 -46,324 -14,762 9,668 1,447
Men 279 -9,939 -1 ,077 86,679 9,169

Married 5,774 71,002 23,697 166,009 51,705

Total Financial Resources

All Divorced 535 -2,931 -5, l 75 13,294 3,938
Women 288 -7,799 -1 1,041 9,701 -398
Men 247 2,171 398 17,059 7,087

Married 5,649 16,957 8,71 1 29,569 14,092

Changes Adjusted for Family Size

Income

All Divorced 535 4,455 1,378 8,849 4,874
Women 288 393 -2,224 6,920 2,651
Men 247 8,71 1 4,660 10,870 8,343

Married 5,649 6,112 3,005 8,160 4,160

Wealth

All Divorced 5 86 199 -1,269 46,683 6,482
Women 307 -15,335 -4,588 19,799 3,256
Men 279 15,530 2,962 73,214 7,879

Married 5,774 46,164 14,226 1 1 1,900 32,610

Total Financial Resources

All Divorced 535 4,741 1,625 1 1,287 6,020
Women 288 13 -l,720 7,981 2,720
Men 247 9,694 4,399 14,751 9,697

Married 5,649 10,063 4,990 15,827 7,91 1

Notes:

I Results weighted with 1984 family weights.
2 Short-term change is change between Before and After measures.

3 Longer-term change is change between Before and measure approximately 4 years
after original Afier measure.
Source: Author's calculations from PSID

63

 

 

$8.3 $8.8 :88 88.8 $8.; 8 $3.8
$3.8 8... N88. $8.7 8.: $8
$83 8 $88 $8.8 $88 88; 8 88.8
88 £3 8 $3- 88- 8.3
$8.8: $8.8 $8.8 $:..8 $8.8 88.8
83 8nd- _ a- 88.8 :3 $8.8
$8.5 $8; 8 $88 $_ 18 $8.88 .888
888- 5.8- 53- $8.8 $8; _- $8.8-
8:88 $8.8: $8.8 $8.8 3.8.28 88.8
:88- 88.8- $8.8 $8.7 88.8- 5.8-
.188 ._.._....$o8 :88 1:88 .188 $.88
one- 8.8- 8.8- 8.8- who- $2.
$288 $38 $3 8 $08 858 2.5:
8.8.- 238 8.8 82:- 8.:.- 8.8-
$8.8 $8.: 88.: $8 $8.8 $$.:
83 8.3 8_ :8 $3 8w
$8.8 $8.8 $88 $8.: $8.8 $8.8
8.8 8.3 New... 88$ 83 $3
.1883 1.2.88 $8.8 28.8 ._.......$8..88 .688
82.8 $8.8 88.8 83. $8.8 888
$8.3 $8.8 $8.8 $2.8 $8.5 $8.8
8.3.” 5.8 8% an $8.8 $8.8
.868 1:888 5:888 28.: $8.8 .....$8.8
82.8- 53.7 83- 84$- 83- 8.8.8-
«EB-Swag «EB-tonm SHE-cowgi— E..8-to:m Egg-homes HEB-tonm
moo-SOmDM among—mm 130B. 530% 05005

88> 2 5.6 wot-82

88> 2.: @0582
m58> 26 @2th

comm $50

885

2:85 3E8”— 8538-88
38:5 ow<

ow<

28250 .3

«8285 owe—.00

2535 688 .8:

28:6”—

 

233365 63.8.»:— .:¢ 39.583— 5 mow—=20 Egon-Lomgq he 8.382me mx—O 3." ~35.

 

64

9mm :88 8888328 8553.. ”888m

.1... 8 8.885 .8 a 8:888; 25 ....._. 8 8.885 .8 a 8:888; ._.. 8 2.2885 .82 a 85.2885 .

.8855qu E 0.8 88.5 Raw—Sum m

.3853 38.8 88 83 8282, £338 .

.mwfinmon 5828 E @8865 8a mpg—nag.» 858:0qu _

 

 

65

 

”8qu
8.8 8.8 8.8 8.8 8.8 $8 Essa-m
8m 8m 8m 8m 8m $8 825580
$3.28 $8.88 $8.88 $8.88 $8.88 $2.88
82 88.8 $82- 8.8 83 88$ 8988
$8.8 $8.8 288 $88 28.8 $8.8
$3 8% 85. 83 $8..” 888 m :28
:$8.8 $8.8 ......$...m.8 $8.8 $8.8 $8.8
88$ 88$- 8% $8 88.: 2.8.8 N :88
$2 2.8: $8.8 ......$$.8 $8.8 $8.8 2.8.8
8.8- 8.3 5.8 83- 82.8 85. 852m 8:5 83> m
$8.8: $8.8 .288 2.8.8 2.88 28.8
$8.2m- 8.8. 8.8.8 88$- 88.:- 888 8.35 8:8 88> N
.8858 $8.8 .8888 $8.8 $8.8 $8.8
$8.8- 888- 88$- 83a 8.;- 88$- oeoza 8% 30> _
8.28-Hogs HEB-82% 8.28-Ewes FEB-tonm stag-choq Egg-tog
mooSOmoM REFS—HE Each. 5302/ 08005

 

Apt—o3 efi «Ea-H

Regression results are in Table 2.6. The short-term columns are a replication of
Table 2.4 but with the restricted sample, and the longer-term columns use the changes
four years after those in the short-term columns. In the income regression, the gender
gap decreases with time (from almost $10,000 to approximately $7,500) and becomes
statistically insignificant. The opposite occurs in the wealth regressions — the female
coefficient becomes more negative (from -$2,400 to -$6,100) and significant in the
longer-term. The gender gap in total financial resources is significant in both the short-

and longer-term and is approximately $14,000 in both regressions.

6.3 Compensatory Wealth Allocation

Spouses who suffer a relatively large drop in income upon divorce may be
compensated with a relatively larger share of marital assets. I perform a descriptive
exercise to see if there is empirical evidence of this tradeoff. The simple correlation
between the change in income and change in wealth is not sufficient because it does not
control for pre-divorce income and wealth levels. Individuals with very large income and
wealth prior to divorce will likely have larger dollar declines than those with little pre-
divorce income and wealth.

To determine whether there is evidence of compensatory behavior, post-divorce
annuitized wealth is regressed on the post-divorce income and pre-divorce characteristics.
Although an unusual specification, it is appropriate for the situation at hand. The aim is
to consider whether an individual who experiences a relatively large decline in income
suffers a relatively smaller decline in wealth compared to others with the same pre-

divorce characteristics. In the absence of compensatory behavior, the post-divorce

66

income coefficient should be zero; a negative coefficient would indicate compensatory

behavior.

 

Table 2.7: OLS Regression of Income-Wealth Trade-off

 

 

Coefficient
Post-divorce Income 0.032
(0.007)***
Pre-divorce Annuitized Wealth 0.568
(0.040)***
Pre-divorce Annuitized Wealth Squared -9.590E-07
(0.000)***
Pre-divorce Income 0.022
(0.023)
Pre-divorce Income Squared -5.290E-08
(0.000)
Female -2,540.01
(704.533)***
Constant 2,063.67
(4,481.817)
Observations 928
R-squared 0.35

 

Notes:
I Dependent variable is post-divorce annuitized wealth.

2 Although not shown in the table, variables for pre-divorce
personal and household characteristics, cohort indicators, and time
since divorced are included in the regression.

3 Results weighted with 1984 family weights.
4 . .
Standard errors are in parenthesrs.

5 Significance at 10% indicated by *, significance at 5% indicated
by **, and significance at 1% indicated by ***.
Source: Author's calculations from PSID

The regression results are in Table 2.7. The post-divorce income coefficient is

statistically greater than zero. This suggests that, for individuals with the same pre-

divorce income, wealth, and other observable characteristics, the person who has the

larger decline in income also has a larger decline in wealth. These results are contrary to

compensatory behavior. However, it should be noted that these findings may be driven

67

by pre—marital assets (which are not observed in the PSID) and therefore do not rule out

the possibility of compensatory behavior.

7. Legal Regime

The framework presented in Section 4 indicates that post-divorce economic
resources are affected by legal regime. In this section, differences across laws are
examined to see if the empirical results verify the framework’s predictions and to gain
insight on issues where the framework cannot make predictions. Small sample sizes will
affect the ability to identify significant differences, but it is still useful to look at the
comparisons. Whether compensatory wealth allocation varies by legal regime will also
be investigated. For all legal analyses, the three observations whose divorce occurred in
the same year the state changed any of its divorce laws are eliminated from the sample to

avoid potential misclassification.

7.1 Short-term Changes
Table 2.8 gives the changes in total financial resources by legal regime. The top
panel compares mutual and unilateral divorce, the second panel displays the three types
of property division, and the bottom panel compares fault and no-fault property division.
The framework indicates that, for couples where one spouse wants a divorce and
the other prefers to stay married, unilateral divorce would be beneficial to the first and
harmful to the second.43 The decline in total financial resources is smaller in unilateral

states than in mutual states for divorced individuals overall (see Table 2.8), providing

 

43 Rasul (2006) finds evidence that adoption of unilateral divorce affected the marriage decision so that
couples who chose to marry after the adoption of unilateral divorce are better matched than those who
married under fault laws. This analysis ignores the potential compositional change.

68

 

 

69

 

$33 $.m.w $.Nd $.02 V8.5 :wJ; vwofiw avoffi wadmw wnod: momd ESE—ED
$cm.w $6.5 $cw6 $5.: mmfic 96.: www.mw vmfih: 5%.: whine Pod 33:2
83:00 Bram—2
$.17 $ow._- $66- $.56 mmw- :36 83% Eme 336 3:65 5mm ESE—ED
$.N.o- $.v6- $52- $.56- hm- wom.- omwdv 3N6: wofiom omnfio mom 33:2
:22
$.92- $.m.m_- $.v.m_- $.31 $0.:- va.m- 9%.? ~36: wood: vomdo mwm 138283
$.o.w_- $.32- $.m.w7 $5.3- Nomfi- omwd- 33% 3me v3.3 www.mo mNm 33:2
5803
$6.:- $L.w- $._.w- $.o.o www.m- a? 83% mwfimn www.mo Sofia: 3m .8828:
$3.3- $.02- $.m.m_- $.m.w- mica- www.m- Rim: wood: ocofiw E93 mm: 33:2
8835 =<
ESE—Egaasz
92on :32 :2on :82 $622 :32 5:52 :32 5:32 :32 Z
owSEU 3083 up}: owSEU :5ch owEEU :o¢< 880m

 

 

 

2:33— .:34 an 82:83— .:355m .83. E mow—:50 "wfi ~32.

 

 

 

$5.2 $.m.w $.w.w $5.2 $5.: 5Nc.w V5m.5m woo._5 35.? Sumac mo: 33 :oEEoU
$.w.w $.m.w $.22: $L.2 ~:.5 Nmm.w_ v8.55 mm5.wm_ @265 250.3: min .585 .8800
$5.5 $5.5 $5.5 $5.2 w5m6 own-am: 553% $5.52 $35 N363 35.5 .55 .2:me
3330 cor—.82
$.N.No $.m.mm $5.5m $5.9V 53.3. 55.5 wooJo m5m6© N36: ~53...V 2 3:..— :oEEoU
$.N.o- $5.7 $.95- $.m.3 ow- 55N.w 83% 55.3 53.25 «5:.w5 N5 .585 .8800
$.99 $.o.m- $._.5- $5.o- cow. :3. 2%.? £435 85.0 555.5 mmm .35 .23-n:
:02
$5.3”- $.w.vm- $5.2- $.N.w_- 3&5- va.5- 856m 53...?- . 505.mm 9.de mm 33 38:80
$5.5m- $.v.om- $5.Nm- $3.2- $5._N- 03.5:- 5mv.wv Suede www.m5 55.55 53 .505 .EEoU
$6.:- $.w.:- $5.2- $.v.m- 356- mm 1a. 815.: owimc oo_.5m $0.3 m5m .35 .2:?”—
3:83
$5.3- $.: $5.3 $5.2 v8.5- 356 omimv vadv $5.9m 53.? ov 33 :oEEoU
$5.3- $5.c_- $5.3- $5.5- w5w.w- waxw- owmfim NO5.N5 58.25 0535 SN .585 .8800
$.5.m- $.m.w- $5.3- $L.- www.m- N3}- _mw.mm $5.8 mom.5m 3.36 85 .55 .zsvm
39839 =<
«PC. 33>:— atone...—
:232 :32 3:32 :32 3:32 :32 5:32 :32 8:32 :32 Z
owSEU E335 .53.. owSEU 8335 “5:20 3a.... 883m

 

A9233 ”N «3.5.

70

amn— th 0:062:23 0.863.. “00::om
05 :8: 8 $666 -160 06:06“. 05 33 6:0 5338036 06: mm 00.3.. 6:0 0860mm 60 0680.6 05 000: 06.360 6:00:06 605:6}. N
£053 6.8.... £2 5.3 330.03 £30: _

 

71

 

 

.0082
$.66 $6.6 $6.2 $6.6— 666.6 666.2 626.66 666.62 665.66 6_ 16: 566.6 630.702
$.66 $.5.5 $L.5 $6.6_ ~6_.6 665.6_ K666 666.6: 656.65 656.662 66¢.5 2:06
60652
$5.7 $6.6- $.5. 7 $.66 @66- 666.6 666.56 666.65 666.66 666.65 562 6:06-02
$6.6- $6.6- $.66- $.6.6- ow- :1- 6v6.w6 666.65 665.26 655.65 _56 :36
:02
$6.67 $6.67 $6.67 $6.5- 656.6- 665.6- 26.66 3.6.66 . 666.66 62.66 6: 2:06.02
$6.67 $.6.57 $._.67 $6.6- 65.6. 25.6- 566.56 666.66 w6_.56 656.56 666 2:06
3:83
$6.5- $6.5- $._.v7 $6.7 666.6- 665- 66.66 656.66 62.66 666.65 666 2:06-02
$.66- $.6._ 7 $6.2 7 $.66- 5w5.v- 36.6- 666.66 656.66 666.66 53.66 666 :36
600035 3..
:ommgn b.5695 ==£éZE=fim
3:002 :32 3602 :32 3602 :32 :3602 :32 :0602 :32 Z
06:20 6:00:06 .6046 06:20 6:00:06 06:20 36¢. 0.8.60m

 

6.2.86 3 033

evidence that unilateral divorce reduces legal fees (represented as c in our conceptual
framework) and improves post-divorce wellbeing for all divorced individuals. It is
difficult to determine whether one gender was affected differently than another with the
simple means comparison.

The married sample is included for comparison. Those results indicate that the
findings must be viewed with caution. The married couples in unilateral states
experienced larger increases in resources than those in mutual states. This may indicate
heterogeneity across states or reflect that couples who married under unilateral divorce
laws are better matched.44

The second panel of Table 2.8 divides the sample into the three types of property
division: equitable distribution, community property, and common law. There are not
enough common law divorces to draw any conclusions about that type of property
division, but those observations are included in Table 2.8 for completeness.45

The differences between equitable distribution and community property are rather
large. Overall, those in equitable distribution states fare better than those in community
property states. It is possible that the difference between the two types of property
division reflects lower costs of divorce (denoted as “c” in the conceptual framework) in
community property states. Another explanation is that less wealth depletion occurs in
equitable distribution states. In community property states, the requirement that marital
property be divided equally between spouses may necessitate the sale of the family home.
Divorces in those states cannot be finalized until all assets are allocated, so couples may

sell their home (which is most likely their largest asset) at a lower price than they

 

44 See Rasul (2006) for a discussion of changes in the marriage market caused by liberalization of divorce
laws.

45 By 1985 (the first year in my sample), Mississippi was the only state with common law property division.

72

otherwise would. Such behavior would result in lower post-divorce wealth for both
spouses.

The third panel of Table 2.8 suggests that removing fault grounds from property
division was beneficial. Although the mean and median provide mixed results for men,
overall results and those for women indicate smaller reductions in resources with no-fault
property division. These findings suggest that removal of fault grounds in property
division may function much like removal of fault grounds for divorce proceedings — it
reduces the need for long, expensive legal battles and results in better post-divorce
wellbeing for both spouses.

As discussed in Section 6, it is possible that the gender differences evident in the
mean and median comparisons are attributable to personal and/or household
characteristics. To allow for this, legal indicators are added to the same regression
analysis presented in Table 2.4. The legal variable coefficients are in Table 2.9.

For each of the three dependent variables, regression results are presented for two
different specifications. The first includes the legal variables; the second adds
interactions between the gender and legal variables. The female coefficient is negative
and significant in the income and total financial resources regressions that do not include
the interaction terms. Once the female-legal interactions are added, the female
coefficient is no longer significant, but the female-unilateral term is significant in both
the income and wealth regressions, and the female-no-fault property term is significant in
the income regression. This indicates that the gender gap is a result of the differential

effects of legal regime.

73

 

 

gnawed 8233: 5.533 A2384; 2392.3 Amenitms
ammo:- Nofiwm $.83.- wmfi; 8.5.2- 33.5 8:355 29355
asfifis $38.3: $285 $3.26 agafldc @333:
o _ . is. 853. - mwezé- 3S3- ESS- ooemm €285 $9658
$2325 $2 «3.0 2538.: Layman: L: <85; €2.3th
mommmm £33- :me S. 83. wZSw 3.83 éaem nae-oz
amtwwma .........€Nw. _ moi 1.23% a 8.3 :38. _ 9.: $3.23 LES-$6
8.23 2.22: 2.83 2.?3 5.3» m3 5.» Ease:
328.5 ILSQEQQ canowte @333 Amwogwas 3.9%.de
$.83? 3.9.3 7 8.3a- 8%3- $53.2- 2.3%. 2E0;
mcomuomuouék’ mcomuogofih o\>) maomuoeopéki 2530885 OE? 253088555 8030885 OE»
moo-unemom Emoamcmn— Euch- 5—«03 08005

 

 

 

ofimuom Bus: 5? 89:58: E mow—:20 he 28:8..me mac 5d 935—.

 

74

9mm :83 macaw—:28 Mafia/x ”ooSom

.32.. 3 8:865 £1 E commommcwi was .3. .3 @8865 o\om E 038536 .._. .3 @8863 $2 “a 8305:me n

6883933 8m 3836 boss 08: 98 £888.: toga .mocmtouofiano 20:33: 98 3:8th benefit-2a 8.3 8:22.380 N

$6065ch 5 8m $8.5 @3985 a.

new? £5 £2 5:» BEE, 33QO m

.mwficmon 2828 E @8363 8a BEE-Hg E00539 _

 

 

”8qu
mg 8o 85 85 go so wanna-m
N3 NS Na N8 Na N? 823530

$8.838 $24.86 $383.8
$38.2 8.3% 3&2; Ea ascmbaaom
$3.358 $8.825 $33.3
2.0%.: 3.5.3 8.2? dam .aeooaaefi
$8.830 $3890 Lamewté
$.21? 8.23 $.81:- beaem gas-oz...2ae£
$3.8 18 ...$3.§.® $8.55
3 .33 3??- wweww; _ 8833 Eggbaeom
mco30803£§> $830835 33 wco308035§> mco30m-535 OE» mao30303£\3 $532235 03»

 

mooSomom EoSEE .98-

£83

0885
A9283 a." 039—.

75

The fi'amework suggests that unilateral divorce helps the spouse who prefers
divorce, hurts the spouse who prefers marriage, and helps both spouses by reducing legal
fees. The unilateral variables are positive and significant in the regressions without the
female-legal interaction terms. When the female-legal interactions are included, the
unilateral coefficient is only significant in the wealth regression. There unilateral divorce
has a positive effect for both men and women, but the effect for men is much greater.

The positive effect indicates that unilateral divorce reduces the costs associated with
divorce, and the larger reduction for men suggests that men prefer divorce more often
than women.

As expected, none of the property division type coefficients are statistically
significant. Only a very small portion of the sample divorced under common law
property division. This results in large standard errors for those variables.

The effect of no-fault property division is quite different in the income and wealth
regressions. Removal of fault grounds has a positive effect on income, but inclusion of
the female interaction term indicates the gains are to men and women have significantly
lower income without fault grounds. In contrast, the overall effect of no-fault property
division on wealth is negative, and there is no significant gender difference. The
significance of changes in property division in the income regression is unexpected

because no-fault property laws directly affect wealth but not income.

7.2 Compensatory Wealth Allocation

As discussed above, there is no evidence of compensatory wealth allocation in the

overall sample. However, the conceptual framework suggests that compensatory

76

behavior should be more prevalent in equitable distribution states. To consider this, the
same regression analysis is run separately for equitable distribution and community

property states.46

 

Table 2.10: OLS Regressions of Income-Wealth Trade-off
by Property Division Type

 

Equitable Distribution Community Property

 

 

 

Post-divorce Income 0.039 0.023
(0.009)*** (0.014)
Pre-divorce Annuitized Wealth 0.472 0.098
(0.052)*** (0.112)
Pre-divorce Annuitized Wealth Squared -8.620E-07 1.940E-06
(0.000)*** (0.000)***
Pre-divorce Income 0.014 0.002
(0.026) (0.054)
Pre-divorce Income Squared 4.040E-08 6.230E-08
(0.000) (0.000)
Female -2,504.37 -2,388.98
(782.038)*** (1,691.382)
Constant 1,165.86 1,654.13
(4,890.369) (11,861.579)
Observations 694 196
R-squared 0.3 1 0.63
Notes:

I Dependent variable is change in annuitized wealth.

2 Coefficients for pre-divorce personal and household characteristics, cohort indicators,
and time since divorced are suppressed.

3 Results weighted with 1984 family weights.

4 Standard errors are in parenthesis.

5 Significance at 10% indicated by *, significance at 5% indicated by **, and significance
at 1% indicated by ***.
Source: Author's calculations from PSID

 

46 The analysis was also completed for fault/no-fault property division laws, and the results (not reported)
are essentially the same. The coefficient on the post-divorce income term is 0.036 with no-fault property
and 0.033 when fault considerations are allowed.

77

Table 2.10 displays the regression coefficients. The results for equitable
distribution states are very similar to those for the entire sample — individuals with larger
declines in income also suffer relatively larger declines in wealth. As discussed in
Section 6, this does not rule out the possibility of compensatory behavior, but there is no
evidence of it in this analysis. In community property states, the coefficient is

statistically no different from zero as predicted in the conceptual framework.

8. Conclusion

This analysis uses the PSID to closely examine the changes in economic resources
surrounding divorce. Previous studies analyze changes in income upon divorce, but
relatively little is known about the changes in wealth. Economists typically think of
wealth as providing a buffer to assist in smoothing consumption when negative shocks
occur, but its importance is magnified in the context of divorce. Upon divorce, marital
assets are divided between spouses, and the legal environment should affect that
allocation. Furthermore, it is possible that different divorce laws promote compensatory
behavior more than others. Exclusion of wealth from studies of the economic effects of
divorce ignores this possibility and may misrepresent the economic impact of divorce and
the related gender inequality.

Both the income and wealth effects of divorce are analyzed. Consistent with
previous research, income declines for divorced individuals collectively, but women
experience a greater decline than men. The same is true for wealth and total financial
resources. Divorced individuals overall experience a decline in economic resources

following divorce, but the loss is much larger for women than for men. Although small

78

sample sizes limit the precision of the estimates, the longer-term changes in economic
resources indicate that the gender gap remains.

Given the central role of divorce laws in determining the post-divorce financial
resources of both spouses, differences in income, wealth, and total financial resources are
compared across legal regimes. Both unilateral divorce and no-fault property laws have a
more significant effect on wealth than income, as the simple framework of divorce
suggests. Additionally, unilateral divorce significantly increases post-divorce wealth for
men and women (consistent with such laws reducing the cost of divorce), but the gain is
significantly smaller for women. Such results suggest that, overall, men prefer divorce
more often than women. There is no evidence of compensatory wealth allocation in
equitable distribution states or community property states.

Overall, the results indicate that post-divorce financial resources are significantly
affected by the legal environment. This suggests a role for policy to reduce the harmful
economic effects of divorce and the related gender inequality. However, caution is
required. As demonstrated by the interaction of unilateral divorce and property division
type, the combination of laws affects the outcome. Hence, while there is a role for policy,

any law changes need to consider the overall legal regime in the state.

79

Appendix

 

Table A2.]: Divorce Laws by State, 1984-2007

 

 

Unilateral Property Division No-fault

State Divorce Type Property Division
Alabama Yes Equitable distribution No
Alaska Yes Equitable distribution Yes
Arizona Yes Community property Yes
Arkansas No Equitable distribution Yes
California Yes Community property Yes
Colorado Yes Equitable distribution Yes
Connecticut Yes Equitable distribution No
Deleware Yes Equitable distribution Yes
Florida Yes Equitable distribution 1986
Georgia Yes Equitable distribution No
Hawaii Yes Equitable distribution Yes
Idaho Yes Community property 1990
Illinois No Equitable distribution Yes
Indiana Yes Equitable distribution Yes
Iowa Yes Equitable distribution Yes
Kansas Yes Equitable distribution 1990
Kentucky Yes Equitable distribution No
Louisiana No Community property No
Maine Yes Equitable distribution 1985
Maryland No Equitable distribution No
Massachusettes Yes Equitable distribution No
Michigan Yes Equitable distribution No
Minnesota Yes Equitable distribution Yes
Mississippi No Common law No
Missouri No Equitable distribution No
Montana Yes Equitable distribution Yes
Nebraska Yes Equitable distribution Yes
Nevada Yes Community property No
New Hampshire Yes Equitable distribution No
New Jersey No Equitable distribution Yes
New Mexico Yes Community property Yes
New York No Equitable distribution No
North Carolina No Equitable distribution No
North Dakota Yes Equitable distribution No

80

 

Table A2.] (cont'd).

 

 

Unilateral Property Division No-fault

State Divorce Type Prggerty Division
Ohio No Equitable distribution No
Oklahoma Yes Equitable distribution Yes
Oregon Yes Equitable distribution Yes
Pennsylvania No Equitable distribution No
Rhode Island Yes Equitable distribution No
South Carolina No Equitable distribution No
South Dakota 1985 Equitable distribution No
Tennessee No Equitable distribution No
Texas Yes Community property No
Utah 1987 Equitable distribution 1987
Vermont No Equitable distribution No
Virginia No Equitable distribution No
Washington Yes Community property No
West Virginia No Equitable distribution No
Wisconsin Yes Community property No
Wyoming Yes Equitable distribution No

 

Sources:
Unilateral divorce from Gruber (2004)

Property division type from Gray (1998) and Freed and Foster (1979 and 1981)
No-fault property division data fiom Ellman and Lohr (1998)

81

References

Bahr, Stephen J. 1983. “Marital Dissolution Laws: Impact of Recent Changes for
Women.” Journal of Family Issues, 4(3): 455-466.

Bianchi, Suzanne M. and Daphne Spain. 1986. American Women in Transition. Russell
Sage Foundation, New York.

Browning, Martin and Annamaria Lusardi. 1996. “Household Saving: Micro Theories
and Micro Facts.” Journal of Economic Literature, 34(4): 1797-1855.

Citro, Constance F. and Robert T. Michaels, Editors. 1995. Measuring Poverty: A New
Approach. National Academy Press, Washington, DC.

Duncan, Greg J. and Saul D. Hoffman. 1985. “A Reconsideration of the Economic
Consequences of Marital Dissolution.” Demography, 22(4): 485-497.

Ellman, Ira Mark and Sharon L. Lohr. 1998. “Dissolving the Relationship Between

Divorce Laws and Divorce Rates.” International Review of Law and Economics, 18:
341-359.

Freed, Doris Jonas and Henry H. Foster, Jr. 1973. “Economic Effects of Divorce as of
June 1, 1973.” Family Law Quarterly, 7(3): 275-343.

Freed, Doris Jonas and Henry H. Foster, Jr. 1979. “Divorce in the F ifiy States: An
Overview as of 1978.” Family Law Quarterly, 13(1): 105-128.

Freed, Doris Jonas and Henry H. Foster, Jr. 1981. “Divorce in the Fifty States: An
Overview.” Family Law Quarterly 14(4): 229-284.

Gray, Jeffrey S. 1996. “The economic impact of divorce law reform.” Population
Research and Policy Review, 15( 1996): 275-296.

Gray, Jeffrey S. 1998. “Divorce—Law Changes, Household Bargaining, and Married
Women’s Labor Supply.” The American Economic Review, 88(3): 628-642.

Gruber, Jonathan. 2004. “Is Making Divorce Easier Bad for Children? The Long-run
Implications of Unilateral Divorce.” Journal of Labor Economics, 22(4): 799-833.

Hoffman, Saul. 1977. “Marital Instability and the Economic Status of Women.”
Demography, 14(1): 67-76.

Holden, Karen C. and Pamela J. Smock. 1991. “The Economic Costs of Marital

Dissolution: Why Do Women Bear 3 Disproportionate Cost?” Annual Review of
Sociology, 17(1991): 51—78.

82

Jacob, Herbert. 1989. “Another Look at No-Fault Divorce and the Post-Divorce Finances
of Women.” Law & Society Review, 23(1): 95-116.

Lester, Gordon H. 1989. “Child Support and Alimony: 1989.” Current Population
Reports: Consumer Income Series P-60, No. 173.

Lupton, Joseph P. and James P. Smith. 2003. “Marriage, Assets, and Savings.” In
Marriage and the Economy: Theory and Evidence from Advanced Industrial Societies,
edited by Shoshana A. Grossbard-Shechtman, 129-152. New York: Cambridge
University Press.

McKeever, Matthew and Nicholas H. Wolfinger. 2001. “Reexamining the Economic
Costs of Divorce for Women.” Social Science Quarterly, 82(1): 202-217.

Peters, H. Elizabeth. 1986. “Marriage and Divorce: Informational Constraints and Private
Contracting.” The American Economic Review, 76(3): 437-454.

Peterson, Richard R. 1989. Women, Work, & Divorce, State University of New York
Press.

Rasul, Imran. 2006. “Marriage Markets and Divorce Laws.” Journal of Law, Economics,
and Organization, 22(1): 30-69.

Schmidt, Lucie and Purvi Sevak. 2006. “Gender, Marriage, and Asset Accumulation in
the United States.” Feminist Economics, 12(1-2): 139-166.

Social Security Administration. 2007. Actuarial Publications: Period Life Table (updated
July 9, 2007).

Stevenson, Betsey and Justin Wolfers. 2007. “Marriage and Divorce: Changes and their
Driving Forces.” PCS Working Paper Series, PSC 07-04.

Stirling, Kate J. 1989. “Women Who Remain Divorced: The Long-Term Economic
Consequences.” Social Science Quarterly, 70(3): 549-561.

Weiss, Yoram and Robert J. Willis. 1993. “Transfers among Divorced Couples: Evidence
and Interpretation.” Journal of Labor Economics, 1 1(4): 629-679.

Weitzman, Lenore J. 1985. The Divorce Revolution: The Unexpected Social and
Economic Consequences for Women and Children in America. The Free Press.

Wilmoth, Janet and Gregor Koso. 2002. “Marital History and Wealth Outcomes.”
Journal of Marriage and Family, 64 (February 2002): 254-268.

Zagorsky, Jay L. 2005. “Marriage and divorce’s impact on wealth.” Journal of Sociology,
41 (4); 406-424.

83

ls Marital Separation a Transitory State?

Evidence from the PSID

1 . Introduction

Economic analyses typically do not differentiate between marital separation and
divorce. Instead, the two are combined into one marital status category. A potential
motivation for combining the two statuses is that separation is a transitory state that
ultimately leads to divorce. If that is the case, then combining the two groups seems
quite sensible. However, if separated individuals return to marriage or remain separated
without formally divorcing, these individuals are mis-categorized in analyses that
combine the two groups. This is particularly important for topics like government benefit
programs and health insurance where legal marital status can affect eligibility and
coverage. Thus, the goal here is to examine whether the transitory assumption holds
empirically.

The analysis employs the Panel Study of Income Dynamics (PSID) and begins
with a cross-sectional comparison of separated and divorced individuals. Next, I exploit
the panel nature of the PSID to follow individuals for ten years after an initial marital
separation to determine whether they return to marriage, remain separated, divorce, or
become widowed. The remainder of the analysis focuses on the propensity to exit marital
separation. I examine differences in the probability of exiting separation by duration of
marital separation and race, and then I perform a hazard analysis to see what

characteristics are associated with a greater likelihood of exit from separation.

84

1 find that substantial differences exist between the separated and divorced groups.
In general, the analysis suggests that separation is a transitory state for most individuals,
but not all. The likelihood of exiting separation declines steadily over time, but
considerable differences exist between racial groups. The hazard analysis verifies that
the racial differences are significant. Additionally, older individuals and those with

children (especially young children) are less likely to exit separation.

2. Data

The data for this analysis are from the PSID. It is a longitudinal survey conducted
at the Survey Research Center of the Institute for Social Research at the University of
Michigan. It began in 1968 with a sample of about 5,000 households. The survey was
conducted annually through 1997 and every-other-year since then. The most recent data

available are from the 2005 survey.

2.1 Defining the Analysis Samples
I use three different samples for my analysis. The first sample, the Cross-
Sectional Sample, is composed of all individuals who are currently separated or divorced
in the 1984, 1994, or 2005 surveys. The three cross-sectional groups are pooled into one
sample which includes 580 separated and 2,797 divorced individuals. Analyses with the
cross-sectional sample employ the PSID family weights from the respective survey year.
The second sample, the Ten Year Sample, includes all individuals who first report

their marital status as separated in any of the surveys between 1969 and 1995 and remain

85

in the PSID for at least ten years after separation.47 The long time period ensures that the
majority of exits from separation are observed and therefore allows me to examine post-
separation marital status transitions. The first year in which marital separation is reported
is used as the initial time of separation. I then group individuals into one of four
categories: returned to marriage, remained separated, divorced, or became widowed.48
All individuals in the sample are initially separated, and I follow them until their first
change in marital status or the end of the ten year period, whichever occurs first. If a
person’s marital status becomes married (without a divorce occurring first), that
individual is classified as returned to marriage. Those whose marital status is reported as
separated in all ten years are in the remained separated group, and those who divorce
(within ten years of separation) are classified as divorced. The final group includes all
individuals whose marital status changes from separated to widowed. The Ten Year
Sample includes 1,541 individuals. The PSID family weights for the year of marital
separation are used in all analyses utilizing this sample.49

The third sample, the Exit Sample, is similar to the Ten Year Sample without the
restriction that individuals remain in the PSID for ten years after separation. Thus, it
collects all individuals who experience a first marital dissolution between 1968 and
2005.50 Each year an individual is included in the sample is a separate observation (the
data are stacked), and once the individual either exits separation or leaves the PSID, no

further observations from that person are included. This sample yields itself to hazard

 

47 Individuals already separated in 1968 are excluded from the sample because those observations are
censored.

48 The categorical assignment of individuals is based only on the first change in marital status following
separation. Individuals who divorce, remarry, and become separated again are classified as divorced.

49 Results are not sensitive to the weights selected.

50 Since it is no longer necessary to observe individuals ten years after separation, individuals who
separated between 1996 and 2005 are included in the sample.

86

analysis where marital status transitions and attrition are allowed. The sample contains
9,935 observations on 3,065 individuals. Family weights from the year of marital

dissolution are used for all analyses with the marital dissolution sample.5 '

2.2 Individual and Household Characteristics

The individual and household characteristics included in the analysis are income,
wealth, whether children are present in the home (and, if so, the age of the youngest
child), age, gender, and race.52 Income is total family income including both taxable and
transfer income. Wealth is measured at the household level and includes the net value of
business/farm assets, checking/savings accounts, debt, real estate, stock, vehicles, home
equity, and other savings. All dollar values are converted to 2005 real dollars with the
Bureau of Economic Analysis’ Personal Consumption Expenditures (PCE) price index.
The race categories are white, black, Hispanic, and other (such as Asian and American
Indian). Due to the small number of Hispanic and other racial groups in the sample, they

are not included in all analyses.

3. Results

A cross-sectional comparison serves as the starting point of the analysis. I then
use the Ten Year Sample to examine the marital status changes following separation.
Next, I use the larger Exit Sample to identify characteristics associated with exit from

separation.

 

5' Results are not sensitive to the weights selected.

52 In 1968 and 1969, the PSID collects the age of the youngest child up to five years, but children between
ages six and 17 are classified only as ages 6-8, 9-13, and 14-17. I use ages 7, 11, and 15 for the three
groups, respectively.

87

3.1 Cross-sectional Comparison

The analysis begins with a cross-sectional comparison of separated and divorced
individuals. The motivation for this comparison comes from the cohabitation literature.
There the goal is to determine whether cohabitation is a stage in the marital process or if
it is a substitute for marriage. The parallel to my research question is straightforward: is
marital separation a stage in the divorce process or a substitution for divorce? As
highlighted in Smock’s 2000 review of the cohabitation literature, one branch of that

research compares characteristics of individuals in different marital statuses.

 

Table 3.1: Summary Statistics for Separated and Divorced Individuals

 

 

 

 

 

Separated (N=5 80) Divorced (N=2,797)

Mean Median Std. Dev. Mean Median Std. Dev.
Income 37,981 25,1 15 49,022 45,919 32,912 166,757
Wealth 60,046 2,581 204,452 144,655 28,603 557,907
Education 12.24 2.33 12.94 2.47
Age 42.93 12.38 47.48 13.67
If Kids 0.45 0.50 0.28 0.45
Number of Kids 0.83 1.12 0.48 0.91
Female 0.59 0.49 0.59 0.49
White 0.53 0.50 0.82 0.39
Black 0.42 0.49 0.15 0.36
Hispanic 0.04 0.19 0.01 ' 0.12
Other Race 0.01 0.08 0.02 0.13
Year Married 1982 14.28 1974 13.98
Marriage Number 1.36 0.63 1.31 0.59
Length of Marriage 10.631 8.95 11.618 9.03
Year of Dissolution 1993 10.95 1985 11.28
Length of Dissolution 6.07 7.08 10.40 9.66

 

 

Notes:
' Sample composed of pooled cross-sections from 1984, 1994, and 2005.
2 PSID family weights for the respective cross-sectional year are employed.

3 All values are in 2005 dollars.
Source: Author's calculations from PSID

Table 3.1 displays resource measures and demographic characteristics for

separated and divorced individuals. If separated and divorced individuals look similar,

88

this would provide support for the transitory assumption. However, large differences
between divorced and separated individuals are evident. Separated individuals have
lower income and wealth levels, are slightly younger, and are more likely to have
children. Additionally, the proportion of blacks and Hispanics is greater in the separated
sample.

How do these differences fit with the typical assumption that separation is a
transitory state? The higher proportion of separated individuals with children could
reflect the fact that divorces involving children are more complex and therefore take
more time to negotiate. However, it is also possible that individuals with children are
more reluctant to go through with divorce and instead separate while attempting to save
the marriage. Even if those marriages eventually end in divorce, individuals with
children may behave differently during separation than those individuals who are
separated only while waiting for the official divorce decree.

If separation was a transitory state into divorce, then the lower income and wealth
levels of separated individuals would indicate that the large decline in economic
resources occurs at separation and is followed by financial recovery. The differences in
duration of dissolution are consistent with this story; on average, the divorced sample has
been divorced for 10 years (and was likely separated prior to that), and the separated
sample has only been separated for about six years.53 However, the literature on the
economic effects of divorce on women casts doubt on that explanation. As summarized
in Holden and Smock’s 1991 review of that literature, “most longitudinal studies indicate

that women’s economic vulnerability is prolonged for at least five years, unless

 

53 When the separated sample is restricted to those separated for less than three years (results not shown
here), the gap between separated and divorced individuals is smaller than when all separated individuals are
included but remains large.

89

remarriage occurs.” Given that recovery from marital dissolution is limited, it is unlikely
that the difference between separated and divorced individuals observed here is
attributable entirely to post-divorce recovery.

The differences in racial composition between separated and divorced individuals
indicate that separation is not equally transitory across races. Whites appear much more
likely to divorce than blacks. This racial disparity is consistent with Sweet and Bumpass’
(1987) findings in the 1980 Census that, within three years of separation, 91 percent of
non-Hispanic white women were divorced, but only 55 percent of black women were

divorced.

3.2 Post-separation Marital Status Transitions

The cross-sectional analysis casts doubt on the assumption that marital separation
is a transitory state. To gain further insight into the issue, I use the Ten Year Sample to
see what marital status transitions occur after separation. Figure 3.1 shows the
cumulative distribution of individuals who are separated, divorced, returned to marriage,
and widowed over the ten year period.

Clearly, the majority of separated individuals get divorced within ten years. Just
one year after separation, over 50% have divorced, and that increases to over 75% by the
end of the ten year period. However, a non-trivial portion of the sample does not

divorce.54 Within ten years, 12% returned to marriage, 6% stayed separated, and 5%

 

54 Individuals who return to marriage or remain separated for over ten years may divorce eventually, but
that is not captured in this analysis because I include only the first marital status transition after separation.

90

became widowed.55 Thus, it appears that marital separation is a transitory state into

divorce for most individuals, but not all.

Figure 3.1: Post-Separation Marital Status by Years Since Separation

100%
80%

60%

Percent

40%

20% a

 

 

0%-

 

4 5 6 7
Years Since Marital Separation

I Separated I Divorced D Returned to Marriage Widowed
Notes:

' Sample composed of all individuals who report current marital status as separated in

any survey 1968-1995 and are present in the PSID ten years later (N=1.541).

2 PSID family weights from the year of separation are employed.

Source: Author's calculations from PSID
3.3 Exit from Separation

The previous analysis demonstrates that while most separated individuals

transition to divorce, others do not. The remainder of this study utilizes the Exit Sample
and examines the likelihood of exiting separation. All individuals who return to marriage,

get divorced, or become widowed are classified as exiting separation. Individuals who

drop out of the PSID are included in the sample until attrition.

 

55 Weaver (2000) finds evidence that divorced individuals report their marital status as widowed following
the death of their former spouse. The PSID cleans the data and makes efforts to verify status, but to the
extent they are unable to do so and individuals have incorrectly reported their marital status as widowed, it
is possible that some of those people are mis-categorized here.

91

One factor to consider is the duration of marital separation. Figure 3.2 shows the
exit probability by duration of separation. The likelihood of exiting separation declines
steadily over time. The probability of exit jumps around between eight and 13 years after
separation, but this is attributable to smaller sample sizes. Referring back to Figure 3.1,
less than 10% of the sample is still separated after seven years, resulting in small sample

sizes for the later durations. However, the overall trend is clearly a decline.

Figure 3.2: Probability of Exit from Separation
by Length of Separation

_o
\r

 

.9
ON
I

 

.9
Ur

p
.p.

_o
w

 

Exit Probability

0.2 »

0.1

 

 

 

 

l23456.789101112131415161718

Duration of Marital Separation in Years
Notes:

' Sample composed of all first marital dissolutions between 1968 and 2005.

2 PSID family weights from the year of separation are employed.
Source: Author's calculations from PSID

Previous research (for example, Bramlett and Mosher 2002) finds that the
probability of divorce conditional on separation is much lower for blacks than for whites.
However, it is possible that differences in the probability of divorce merely reflect

differences in how individuals exit separation and not differences in the exit rate.

92

Figure 3.3 displays the exit probability for whites and blacks.56 For the first five years
after separation, whites are considerably more likely to exit separation than are blacks.
Beyond five years the two racial groups look similar, but as with Figure 3.2, that is likely
due to the small sample sizes. Overall, these findings indicate that racial differences exist

not only in the likelihood of divorce, but for exiting separation at all.

Figure 3.3: Probability of Exit from Separation
by Length of Separation and Race

.0
\r

 

.o
O\

.o
Ut

 

.9
a.

 

Exit Probability
p
U.)

.0
N

 

0.1

 

 

 

 

 

1 2 3 4 5 6 7

Duration of Marital Separation in Years

+ White . -I— Black
Notes:

I Sample composed of all first marital dissolutions between 1968 and 2005.

2 PSID family weights from the year of separation are employed.
Source: Author's calculations from PSID

I perform a hazard analysis with the Exit Sample to determine if the differences
observed in Figures 3.2 and 3.3 are statistically significant and whether other personal
characteristics are significantly correlated with exit from marital separation. Table 3.2

displays the hazard analysis results.

 

56 Sample sizes are too small to include Hispanics and other racial groups.

93

 

Table 3.2: Hazard Model Estimates

 

 

 

Variable Coefficient Variable Coefficient
Pre-Separation Income 0.00E+00 Separated 1 Year 3.606
0 (0.084)***
If Kids Before Separation -1.276 Separated 2 Years 3.272
(0.070)*** (0.098)***
Age of Youngest Kid Before Separation 0.062 Separated 3 Years 2.821
(0.007)*** (0.120)***
Age Before Separation -0.061 Separated 4 Years 2.414
(0.002)*** (0.151)***
Black -l.160 Separated 5 years 2.280
(0.069)*** (0.186)***
Hispanic -0.36l Separated 6 years 2.006
(0.174)" (0.232)***
Other Race -0.514 Separated 7 years 1.689
(0.196)*** (0.288)***
Separated in 19805 -0.443 Separated 8 years 2.152
(0.062)*** (0.273)***
Separated in 19905 -0.278 Separated 9 years 0.812
(0.072)*** (0.501)
Separated in 20003 -0.267 Separated 10 years 2.104
(0.221) (0.334)***
Observations 9,935
Notes:

1 Sample composed of all first marital dissolutions between 1969 and 2005.

2 PSID family weights for the year of marital separation are employed.

3’ All values are in 2005 dollars.

4 Standard errors in parentheses.

5 Significance at 10% indicated by *, significance at 5% indicated by **, and significance at

1% indicated by ***.

Source: Author's calculations from PSID

The separation duration coefficients represent the baseline exit hazard in each

year after separation. The coefficients confirm what was evident in Figure 3.2 — the

likelihood of exiting separation declines steadily with duration. The significantly

negative coefficient for blacks verifies the racial differences evident in Figure 3.3.

Additionally, the exit probability for Hispanics and other racial groups is between those

for blacks and whites. The coefficients for the personal and demographic characteristics

94

indicate that older individuals are less likely to exit separation. Individuals with kids are
also less likely to exit separation, but those with older children are more likely to exit
than those with a young child. The probability of exiting marital separation has changed
over time. The base group in the regression is those who separated between 1969 and
1979. Compared to that group, individuals who separated in the 19805 and 19905 are
significantly less likely to exit separation, but the likelihood of exiting separation for

those who separated since 2000 is not significantly different from the 19705.

4. Conclusion

This analysis investigates whether marital separation is a transitory state en route
to divorce. A cross-sectional comparison of separated and divorced individuals suggests
that substantial differences between the two groups exists and provides suggestive
evidence that separation is not a transitory state. Examination of post-separation marital
status changes verifies this finding: approximately 75% of individuals divorce within ten
years of marital separation, but almost a quarter do not. Thus, I conclude that separation
is a transitory state for most individuals, but a non-trivial portion of separated individuals
do not divorce.

The probability of exiting separation varies by duration of separation, race, and
several personal characteristics. The exit probability declines steadily with duration.
Whites are most likely and blacks are least likely to exit separation. Older individuals
and those with kids (particularly young kids) are more likely to remain separated. The
likelihood of exiting separation was lower in the 19805 and 19905, but increased in the

20005 to levels similar to the 19705. These findings suggest that the considerable

95

differences between separated and divorced individuals identified in Table 3.1 reflect not
just that separation is a longer-term state for some, but also that separation is more

transitory for some groups than for others.

96

References

Bramlett, MD and WD Mosher. 2002. “Cohabitation, Marriage, Divorce, and Remarriage
in the United States.” National Center for Health Statistics, Vital Health Stat, 23(22).

Holden, Karen C. and Pamela J. Smock. 1991. “The Economic Costs of Marital
Dissolution: Why Do Women Bear a Disproportionate Cost?” Annual Review of
Sociology, 17(1991): 51-78.

Smock, Pamela J. 2000. “Cohabitation in the United States: An Appraisal of Research
Themes, Findings, and Implications.” Annual Review of Sociology, 2000(26): 1-20.

Sweet, JA and LL Bumpass. 1987. The Population of the United States in the 19805: A
Census Monograph Series. Russell Sage Foundation.

Weaver, David A. 2000. “The Accuracy of Survey-Reported Marital Status: Evidence

from Survey Records Matched to Social Security Records.” Demography, 37(3): 395-
399.

97

 

llllllllllllllll

1293 02956 6928

     

. ‘1: ‘., -