Full text of Working Papers (Federal Reserve Bank of Chicago) : Earnings Mobility in the US : A New Look at Intergenerational Inequality, Working Paper 2001-18

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

Federal Reserve Bank of Chicago

Earnings Mobility in the US: A New Look
at Intergenerational Inequality

Bhashkar Mazumder

WP 2001-18

Earnings Mobility in the US: A New Look at Intergenerational Inequality

Bhashkar Mazumder
Federal Reserve Bank of Chicago

December, 2001

Abstract
This study uses a new data set that contains the Social Security earnings histories of
parents and children in the 1984 Survey of Income and Program Participation, to
measure the intergenerational elasticity in earnings in the United States. Earlier
studies that found an intergenerational elasticity of 0.4 have typically used only up
to five-year averages of fathers’ earnings to measure fathers’ permanent earnings.
However, dynamic earnings models that allow for serial correlation in transitory
shocks to earnings imply that using such a short time span may lead to estimates that
are biased down by nearly 30 percent. Indeed, by using many more years of fathers'
earnings than earlier studies, the intergenerational elasticity between fathers and
sons is estimated to be around 0.6 implying significantly less mobility in the U.S.
than previous research indicated. The elasticity in earnings between fathers and
daughters is of a similar magnitude. The evidence also suggests that family income
has an even larger effect than fathers’ earnings on children's future labor market
success. The elasticity of earnings is higher for families with low net worth,
offering some empirical support for theoretical models that predict differences due
to borrowing constraints. Some evidence of a higher elasticity among blacks is
found but the results are not conclusive.
I am grateful to David Card, David Levine, Ken Chay, John DiNardo, Michael Reich, Nada
Eissa, Mike Clune and seminar participants at Berkeley, Illinois, The Chicago Fed, Cornell,
UMass, BLS, Census and The Santa Fe Institute for their helpful advice and comments. I greatly
appreciated the help of Andrew Hildreth, Julia Lane and Susan Grad in helping me gain access to
the data. This research was conducted while the author was an employee of the Social Security
Administration. The help of the staff at SSA, especially Minh Hyunh, is also gratefully
acknowledged. The views presented here do not reflect the views of the Federal Reserve System.

I.

Introduction
How economically mobile is America? Do all individuals have the same opportunity to

achieve success in the United States labor market irrespective of their economic circumstances at
birth? Is there an economic underclass that is essentially trapped in poverty for generations? The
answers to these questions undoubtedly have a bearing on whether America should be viewed as
an equal opportunity society and whether additional policies are needed to address long-term
inequities. Despite the obvious importance of economic mobility as a basis for public policy,
economists have only recently begun to gain access to the data and develop the tools that might
allow for a clearer understanding of the dynamics of inequality among families over generations.
In recent years a growing body of research has used the regression coefficient relating a
son's log earnings to his father's, as a summary measure of the degree of intergenerational
mobility in society. 1 A high intergenerational elasticity is indicative of a rigid society, since it
implies that an individual’s position in the earnings distribution is largely a reflection of his
parents’ position in the previous generation. In contrast, a low intergenerational elasticity
suggests a relatively mobile society in which an individual's lifetime income is largely
independent of his or her parent’s economic standing. In fact, one minus the regression
coefficient provides a measure of the degree to which earnings “regress” towards the mean.
One useful way to illustrate the significance of this measure is to imagine what it implies
about the evolution of the black-white wage gap in the United States under a set of simplifying
assumptions. An intergenerational elasticity of 0.2, for instance, implies that only 20 percent of
any earnings gap between groups would remain after a generation (say 25 years).2 Using this
logic, the black-white weekly wage differential that stood at about 25 percent for men of age 25

1

Recent studies include Solon (1992), Zimmerman (1992), Altonji and Dunn (1991), Peters (1992), Shea
(2000), Mulligan (1997) and Corak and Heisz (1999). A full survey can be found in Solon (1999).
2
This example also assumes a common intergenerational elasticity for both groups and no other groupspecific effects. For example, a number of factors such as skill-biased technical change or declining
unionism could affect each group differently and temporarily widen the gap further.

2

to 40 in 1980 3 would be reduced to just 5 percent by 2005 for similarly aged men if all other
shocks were ignored. If instead, the intergenerational coefficient was 0.6, then the black-white
wage gap would still be a sizable 15 percent in 2005.
Another way to highlight the importance of the intergenerational elasticity is to consider
its potential implications on the long-term effects of public policy. If the intergenerational
elasticity is sizable and if it represents a causal link that can be exploited by policy makers, then
actions taken to improve the fortunes of individuals in one generation might have a large effect on
future generations as well.
The results from several studies from the 1990s (e.g. Solon 1992, Zimmerman 1992)
have pointed to an intergenerational elasticity in the U.S. of about 0.4, a figure twice as high as
what researchers had previously thought and suggestive of a far less mobile society than was
earlier believed. 4 All of the recent studies on the U.S., however, come from just two surveys, the
Panel Study on Income Dynamics (PSID) and the National Longitudinal Surveys (NLS), both of
which have relatively small sample sizes, and suffer from considerable attrition when
constructing intergenerational samples.5 In addition, because of data limitations, researchers
using these data must estimate fathers’ permanent earnings using only a few years of earnings.
Using a proxy for permanent earnings based on a short-term average, however, is likely to be
flawed since many studies on earnings dynamics have shown that transitory shocks to earnings
are highly serially correlated. In fact, using parameter estimates derived from previous studies on
earnings dynamics, it is apparent that even five-year averages of earnings yield estimates of the
intergenerational elasticity that are biased down by close to 30 percent. This implies that the true
intergenerational elasticity may be closer to 0.6.

3

See Smith and Welch (1989).
Solon (1999) presents a summary of findings of other studies with similar results.
5
For example, in the samples that use five year averages of fathers’ earnings, Solon (1992) using the PSID
has only 290 father-son pairs. Zimmerman (1992) using the NLS has only 192 when using a four year
average.
4

3

This analysis uses a new data source, the 1984 Survey of Income and Program
Participation (SIPP) matched to Social Security Administration's Summary Earnings Records
(SER) to produce new estimates of the transmission of earnings inequality across generations.
Although this data set has some drawbacks, it provides the long-term earnings histories for both
parents and children without any problem of sample attrition. In addition, the data provides
significantly larger samples and richer measures of income and wealth for the parents.
The key result of this study is that the intergenerational elasticity in earnings between
fathers and sons is estimated to be 0.6 or higher, a figure substantially above previous estimates
and indicative of a relatively immobile society. The higher estimate is largely attributed to the
availability of many more years of earnings data on fathers which eliminates the substantial
downward bias stemming from transitory shocks to earnings that exists in previous studies.
Indeed, the results when fathers' permanent earnings are based on shorter time horizons closely
track the findings from previous research.
This study also generates a number of new findings concerning the persistence of
earnings across generations. The intergenerational elasticity between fathers and daughters is
similar to that found between fathers and sons. The father-daughter relationship has received
scant attention in most of the existing literature on intergenerational mobility. 6 Using data on
both parents and using measures of non-earnings income, leads to higher estimates of the
intergenerational elasticity. This provides further evidence that previous estimates of
intergenerational mobility that were based on short-term averages of fathers earnings may have
understated the degree of intergenerational persistence in economic status.
This study also presents evidence that is consistent with theoretical models that
emphasize borrowing constraints as a source of intergenerational inequality (Becker and Tomes
1986, Mulligan 1997). Using detailed information on wealth from the SIPP, the intergenerational

6

This is probably due, in part, to the fact that marriage (coupled with higher average earnings for men)
weakens the reliability of daughters' earnings or income as a measure of economic standing.

4

elasticity is estimated to be significantly higher for families with low net worth and is negligible
for those in the top quartile of net worth. These results suggest that policies that target borrowing
constrained families may play an important role in reducing inequality over the long-term. The
estimates in this study also show a higher intergenerational elasticity among black families than
white families, particularly when both parents' earnings are included. The findings, however, are
not precise enough to justify a strong conclusion.
A methodological contribution of this study is that careful attention is paid to sample
selection rules to address Couch and Lillard's (1998) criticism that past studies may have
inappropriately dropped observations if sons or fathers report zero earnings. In this study various
exclusion rules are used to analyze the effects of including years of zero earnings for both fathers
and their children, and the results are not highly sensitive to these variations.
The paper proceeds as follows: Section II describes the measurement issues involved in
studies of intergenerational income mobility. In particular, this section demonstrates how the
measures of permanent income in the existing literature that use averages over just five years can
substantially underestimate the intergenerational elasticity. In Section III the construction of the
matched dataset is explained and a number of strategies are outlined to deal with some
shortcomings in the data. Section IV presents the methodology used in the study and describes
the main results. In addition, a variety of alternative approaches are presented that deal with
possible criticisms of the research. Section V presents extensions of the research. This includes
an analysis of the effects of family income on children's earnings, how borrowing constraints
might influence the intergenerational transmission of inequality and results concerning
differences in mobility by race. Section VI concludes.

5

II.

Measurement Issues
There is a long tradition dating back to Sir Francis Galton in 1877 that has examined the

rate of regression to the mean of different characteristics across generations. Sociologists were
the first to apply this type of model in analyzing the transmission of inequality across generations
by measuring the correlation of various measures of economic status across generations.7 The
first major economic model to analyze the inheritability of income across generations was by
Becker and Tomes (1979). They proposed a utility maximizing framework in which parents
choose between current consumption and investment in their children's human capital. Under a
set of simplifying assumptions they derived a straightforward result that son’s income is a linear
function of father’s income —suggesting a similar statistical approach as the Galton regression
model. A major contribution of their model was their emphasis on human capital as a primary
channel by which income inequality is transmitted. On the other hand, as Goldberger (1989)
pointed out, it is not clear that the human capital model offers any more empirical content
compared to earlier “mechanical” approaches to studying income transmission. More recently,
Mulligan (1999) has found only mixed evidence in support of the human capital model versus the
Galton model. 8
In any case, empirical studies undertaken by economists have typically used the
following regression model to measure the intergenerational elasticity between fathers and sons:
(1)

y1i = α + ρy0i + β1 Age 0i + β2 Age 20i + β3 Age1i + β4 Age21i + εi

Here y1i represents a measure of economic status such as the log of annual earnings of the son in
family i, while y0i is the corresponding measure for the father. The only additional right hand side
variables that are generally included are age and age squared, in order to account for the effects of

7
8

An early example is Duncan (1961).
In section V some results are presented that are consistent with the human captial model.

6

the lifetime profile of earnings for both the father and son. 9 Ordinary Least Squares (OLS) is
generally used to estimate the equation. The coefficient of interest, of course, is ρ, which
measures the intergenerational elasticity. 10
As might be expected, the earliest datasets that contained detailed intergenerational
information on income used relatively obscure samples.11 These studies used only single-year
measures of fathers’ income or earnings and found the intergenerational correlation to be less
than 0.2. On the basis of these results and other international studies, Gary Becker in his 1988
address to the American Economics Association, asserted that “In all these countries, low
earnings as well as high earnings are not strongly transmitted from fathers to sons…”.12
As carefully documented by Solon (1989, 1992), there are several problems with using
only single-year measures of economic status as a proxy for permanent status that will have the
effect of understating the true parameter estimate.13 These are illustrated in the following
statistical framework:
(2)

y0is = y 0i + w0is + v 0is

(3)

y1it = y 1i + w1it + v 1it
y1i = ρy0i + ε

(4)

9

Other covariates have generally not been included in these studies since the goal is to obtain a summary
measure of all the factors related to income that are transmitted over generations. Therefore, ρ should not
be given a causal interpretation.
10
If earnings are age adjusted and the variance in log earnings is the same for both generations then ρ is
also the intergenerational correlation. The intergenerational correlation has been emphasized in the
sociology literature on intergenerational mobility. The two measures are roughly comparable even if the
variance in permanent earnings differs across generations as shown by Solon (1992). The intergenerational
correlation is more susceptible to bias from mis-measurement of children’s earnings compared to the
regression coefficient. Bowles and Gintis (2001) have also argued that the regression coefficient is a
preferred measure since it does not confound changes in cross-sectional inequality with the association in
earnings across generations.
11
For example, Behrman and Taubman (1985) used a sample of white male twins who served in the armed
forces. Sewell and Hauser (1975) used a sample of high school seniors in Wisconsin who were no longer
in school seven years later.
12
See Becker (1988).
13
Bowles (1972) first pointed out some of the problems with using single year measures of income as a
proxy for permanent income.

7

In this setup, y0is represents the father's log earnings in year s, while y1it is the earnings of his son
in year t.14 Equation 2 breaks down the father's earnings in a particular year into three
components: y0i , a permanent component that reflects the true long-term earnings capacity; w0is , a
component that captures any transitory shocks that might affect that particular year's earnings;
and finally, v0is , a term that captures any errors due simply to mismeasurement such as an
inaccurate report of earnings.15 Equation 3 is the analogous decomposition for the son.
Equation 4 is the relationship of interest between the father's permanent earnings and the
son's permanent earnings. In actuality, researchers with access to only one year's measure of the
father and son's earnings will not be able to estimate (3) but instead, will regress the father's
measured earnings from a single year on the son's measured earnings also from a single year. If
we assume that the transitory shocks and the measurement error are independent of the true
permanent earnings, then the estimate of ρ, ρ̂ , will be biased towards zero by an attenuation
factor. It is easily shown that:
plim ρ̂ = ρλ,

(5)

2

σ y0

where λ =
σ 2 +σ 2 +σ 2
w0
v0
 y0


,



is an "attenuation" coefficient" arising from the mismeasurement of father’s permanent income.16
The first source of downward bias is generated through σ2 w0 term, the variance of transitory
fluctuations. Second, there is bias due to measurement error in the father’s earnings, which is
captured by σ2 v0, the variance of the measurement error term. Finally, many of the studies use
relatively homogeneous samples of fathers, which has the effect of reducing the "signal" in the
14

For simplicity, earnings are assumed to be measured as deviations from the mean and are adjusted for
age and age squared.
15
For the moment, both the transitory component and the measurement error component are viewed as
white noise.
16
In the regression context, any errors in measuring the son’s permanent earnings may lead to less precise
estimates but should not lead to biased coefficients. Errors in fathers’ earnings, in contrast, will bias the
coefficient.

8

data because σ2 y0, is relatively low. Unless the use of a homogeneous sample also happens to
reduces the noise, the downward bias will be exacerbated. The severity of these biases may be
quite substantial. By some estimates, the transitory component and measurement error term
account for about half of the total variance in a single year’s earnings.17
Several studies in the early 1990s used either the Panel Study of Income Dynamics
(PSID) or the National Longitudinal Surveys (NLS) —longitudinal datasets that were nationally
representative and allowed for multiple year measurements—to address these problems. 18 By
averaging the father's earnings over several years they were able to reduce the bias from
transitory income shocks and measurement error.19 The results in nearly all cases were
significantly higher than the 0.2 coefficient from the early literature and instead pointed to an
intergenerational elasticity of around 0.4.
These studies, however, overlooked the fact that averaging earnings over a short time
span might still result in considerable attenuation bias if there is persistence in transitory
fluctuations. In fact, it is well established from many error-component models of long-term
earnings profiles that the transitory component of income is highly serially correlated. 20 The
implications of this finding on past econometric results that used multiyear averages can best be
seen by extending the statistical framework to incorporate serial correlation in the transitory
component. 21 Specifically if we model w0is , the transitory component of earnings as a stationary
autoregressive process,
(6)

w0is = δw0is-1+ ξ is

17

See Card (1994) and Hyslop (2001). Solon’s (1992) survey of several studies suggested that noise may
account for about 30 percent of the variance in single-year earnings.
18
Solon (1999) identifies 15 different studies using these surveys. Probably the most widely cited are
Solon (1992) and Zimmerman (1992).
19
Additional techniques such as instrumental variables were also used, though these estimates often
introduced a positive bias and could only provide an upper bound estimate.
20
Some examples are Lillard and Willis (1978), MacCurdy (1982), Card (1994) and Hyslop (2001).
21
While both Solon (1992) and Zimmerman (1992) present formulas on the bias when incorporating serial
correlation in the transitory component, they do not pursue the implications of this on their results.

9

where δ represents the autoregressive parameter, then the attenuation coefficient when averaging
over T years, λT, can be expressed as follows:
(7)

λT =

σ y20
σ 2y 0 +

1
1
ασ w2 + σ v2
T
T

(

)


 1− δ T  
T − 
(1 − δ )  


where, α = 1 + 2δ 

 T (1 − δ ) 




In the absence of serial correlation in transitory fluctuations (i.e. δ = 0), the coefficient α
= 1 in equation (7), and it is clear that averaging lowers the noise relative to the signal. With
serial correlation, however, the α term creates an offsetting factor. Indeed, the larger δ is, holding
the other parameters constant, the larger the overall attenuation bias will be.22 In order to get a
sense of the possible implications, some simulations using plausible values for δ, and for the
fraction of total variance in one year's earnings that is due to transitory factors, permanent factors
and measurement error were undertaken.23 Using one set of estimates for these parameters from a
recent study by Hyslop (2001), the attenuation coefficient when averaging earnings over five
years was found to be 0.66. 24
A limitation with this approach, however, is that a very high persistence in transitory
shocks might effectively be considered “permanent” if the effects do not die off over the course
of an individual’s working life. This problem can be addressed by assuming that the typical
father will work for 45 years (from age 20 to age 65) and that what is really of interest for the

22

It should be noted that σ2 w and δ are related by σ 2w =

σ ξ2

. In the simulations that follow, assumptions
1−δ 2
are made regarding σ2 w and δ and σ2 ξ is assumed to adjust to satisfy this relationship.
23
If the numerator and denominator of (7) are divided by σ2 yt, then these are all the parameters that are
required to simulate the attenuation coefficients for any given value of T.

10

analysis is the fathers’ average earnings over this 45 year period. If we use the same assumptions
on the other parameters as before, then the the attenuation coefficient increases to 0.74. 25 A full
set of simulation results is shown in Table 1 where the value of δ is either 0.5 or 0.8 under three
different assumptions about the breakdown in the variance of single-year earnings. These results
suggest that estimates of the intergenerational elasticity of 0.4 using five-year averages may still
be biased down by about 25 to 30 percent.
As mentioned earlier this bias may be further compounded if the sample is more
homogenous which is a distinct possibilty in the PSID and NLS due to the high rates of sample
attrition. Solon (1992) for example, uses less than 60 percent of the original cohort of sons and
acknowledges evidence of greater homogeneity in the resulting sample.26
Recent research has also found that estimates of the intergenerational elasticity may be
sensitive to “lifecycle biases”.27 If the variance of the transitory component of earnings changes
considerably over the course of the lifecycle, then short-term averages of earnings taken at a time
when earnings are considerably noisy may lead to further bias. Indeed, since researchers don’t
have earnings information from before the starting point of longitudinal surveys, the father’s age
at the time earnings are measured may be quite high. 28 Several studies have found that the
transitory component of earnings follows a “U-shaped” pattern over the lifecycle. 29 This
suggests that measures of earnings around age 40 may have less attenuation bias than those taken
at age 30 or 50.

24

This assumes that δ = 0.8, that share of the variance in earnings accounted by permanent factors is 0.5, by
transitory factors is 0.3, and by measurement error is 0.2. These are also precisely the same estimates
found by Card (1994) and Mazumder (2001).
25
The procedure is described in detail in the appendix.
26
See Solon (1992) p 398.
27
See Jensen (1987) and Grawe (2000).
28
The average age of fathers is 42 for Solon (1992) and is 50 for Zimmerman (1992).
29
See Gordon (1984), Baker and Solon (1999) and Mazumder (2001).

11

III.

Data Issues

Overview of SIPP and SER
This analysis uses the 1984 Survey of Income and Program Participation (SIPP) matched
to Social Security Administration's (SSA) summary earnings records (SER). The 1984 SIPP was
a nationally representative longitudinal survey, which started with over 50,000 individuals in
nearly 20,000 households.30 Interviews took place every four months and resulted in highly
detailed data on employment, income and government program participation. 31 The survey began
in October 1983 and continued until July 1986 covering the period from June 1983 to June 1986.
Respondents were asked to provide the social security number of their family members and the
SSA subsequently attempted to match individuals who entered the SIPP in one of the first three
waves to their SER via their social security numbers. The resulting file contains the individual’s
SIPP identifiers along with annual taxable earnings from 1951 to 1998. 32

Matching Issues
This matched file allows for intergenerational analysis of families where children were
living with their parents between June 1983 and June 1984 and where the children had social
security numbers that were provided to SIPP interviewers.33 That information would allow

30

Unlike later SIPPs, there was no oversample of low income households.
There are also a variety of topical modules in each interview wave that provide rich information.
32
An additional set of variables is also available in the file including date of birth, sex, race, selfemployment status, agricultural status, military status and a number of variables related to social security
coverage. It should also be noted that the SER file can only be used to gather information on earnings and
not other forms of income (e.g., asset income and transfers) that are available in the SIPP and may have
been used in previous studies on income mobility.
33
There are some difficulties in matching children to their fathers using the 1984 SIPP. An explicit
description of family relationships does not take place until the eighth wave (January to March 1986) at
which point the sample size significantly declines due to budget cuts and attrition. Therefore, in order to
use a representative sample, the sons and daughters are matched to their father in the first wave using a
roundabout procedure. Although children are directly linked only to their mother (when one exists in the
household), they can be linked through the mother to the mother's spouse. This is consistent with the
existing literature which has largely not been concerned about whether the matches are to the biological
31

12

researchers to link data on parents contained in the 1984 SIPP to their children's earnings as
adults up to 1998. In order to go a step further and also access the full social security earnings
history for the parents, it is necessary that the parents also provided social security numbers.
Therefore an analysis of both children's and parents' full history of earnings requires that both be
successfully matched to their SER earnings.
The universe selected for analysis in this study includes children born between 1963 and
1968 who were coresident with either or both parents or living away from home while at college
during the first wave of the 1984 SIPP (June-September 1983). The age range was limited to
those 15 or older in 1983 because of the poor match rate for younger children. 34 This lower
bound on age also ensures that the sons and daughters are at least 27 years old when their
earnings are observed in the years 1995 through 1998. The sample was also restricted to those
who were age 20 or under in 1983 to ensure that the sample did not over-represent those who
stayed at home until a late age.35 The possible selection biases that could result from these rules
are addressed in Section IV.
There are a total of 4072 child-parent pairs in which both the child and at least one parent
are successfully matched to the SER file, representing an overall match rate of 87 percent.36 In
3158 cases, sons or daughter and their fathers are both successfully matched to their earnings
records. Of these, 1663 represent father-son pairs while 1495 cases are father-daughter pairs. An
father, arguing that what is being investigated is the broad effect of family background and not just genetic
influences.
34
In the early 1980s social security numbers were not nearly as universal among children as they are today.
The key factors that determined whether someone would have a social security number were if he or she
worked, had a bank account, owned stocks or received any form of government assistance. An
econometric analysis of whether a 15 to 20 year old in the SIPP was matched to the SER showed all of
these factors to be significant. This suggests that the sample used here over-represents both poor and rich
households. Weighting the sample by the inverse of the probability of being matched has a minor effect on
the results as shown in section IV.
35
Earlier studies such as Solon (1992) and Zimmerman (1992) have used 18 years of age as an upper age
cutoff for kids living at home. In the SIPP, however, sons and daughters living away while attending
college were considered living at home and are included. The percent of 19 and 20 year olds still living at
home or at college in the 1984 SIPP is over 70 percent.

13

alternative approach is to use SIPP income or earnings data from 1984 and 1985 for the parents
instead of matching them to the SER file. A major drawback, however, is that because of
attrition, budget cutbacks and nonresponse to earnings questions, there is a much smaller sample
with complete SIPP earnings data —only 912 father-son pairs and 809 father-daughter pairs.

SER Data Problems
In this study, the use of SER data introduces three key concerns. The first is that
although instances of zero annual earnings may reflect non-working, they could also be due to
employment in a job that is not covered by social security. 37 Although about 90 percent of jobs in
the U.S. are now covered, in the early 1980s the figure was somewhat lower. However, if even
10 percent of the sample is incorrectly classified as zeroes, this presents a significant problem if
regression results are sensitive to sample selection rules around zero earnings. A second problem
is that because earnings are only taxed for Social Security up to the taxable maximum for the
year, the SER file "topcodes" earnings at this cutoff. This is further compounded by the fact that
there have been large changes in the real value of the taxable maximum over the last forty years
resulting in large changes in the fraction of the sample who are topcoded, as shown in Figure 1.
Finally, even among those with positive earnings, a large number of individuals have
both covered and non-covered earnings.38 This is illustrated in Figure 2 which uses the full
sample of adults in the 1984 SIPP-SER and plots SER earnings on the x-axis and SIPP earnings
on the y-axis. If there was random reporting error, the graph would show a random scattering of
points around the forty-five degree line. Instead, there is a large fraction of people who report
dramatically higher earnings in the SIPP than are actually taxed for social security purposes.
36

The match rate within the pairs are as follows: fathers alone are matched at a 93.5 percent rate, mothers
alone are matched at a 93.2 percent rate, sons alone are matched at a 88.8 percent rate and daughters alone
are matched at a 88.2 percent rate.
37
Many federal, state and local government workers are not covered by social security. In addition,
workers in the underground economy or workers in certain occupations are paid outside of the tax system.

14

Each of these three issues may present econometric problems not only because they affect the
dependent variable, children's adult earnings, but because they also affect the independent
variable, parents' earnings. Because of the differences in available information for the children
compared to their parents, each of these issues are addressed separately for the two groups.

Data Solutions: Children’s Earnings
Distinguishing instances of zero covered earnings among the sons and daughters that are
due to lack of work rather than resulting from employment in the non-covered sector is perhaps
the most difficult issue of all. The problem is that there are no available data on hours worked for
the children in the sample as of the late 1990s. Approximately 12 percent of the sons and 21
percent of daughters had zero covered earnings in 1996.
What turns out to be useful, however is the use of another confidential dataset, the 1996
SIPP-SER, which matches a completely different set of individuals to their social security
earnings. In particular, this matched dataset contains detailed earnings from the SIPP for the
years 1996 and 1997 along with the social security earnings histories over the years 1951 to 1998.
This allows one to identify individuals who had zero social security earnings but reported positive
earnings in the SIPP in 1996 or 1997. Focusing exclusively on the 1963-1968 cohort, a series of
models were estimated that would allow for classifying men and women as either “non-working”
or “non-covered” in each year.39 These models are then applied to the children in the 1984 SIPP
to classify them into these groups for the years 1995 through 1998. Those identified as noncovered are then either dropped from the analysis or their earnings are imputed using the mean
level of log earnings for the analogous group from the 1996 SIPP. Similarly, those identified as
non-employed are assigned the mean level of log earnings for that group. Based on the within
38

This may be due to having more than one job, tax avoidance or a desire to maintain social security
eligibility if one’s main job is not covered.

15

sample forecasting results, the procedure appears to do a remarkably good job in classifying men
into the two groups. The details of the methodology and the statistical results are shown in the
Appendix.
The second problem, topcoding at the taxable maximum, is much easier to handle for the
children than for the parents. For example, only 6 percent of the sons and 2 percent of the
daughters were topcoded in 1996. There are two approaches that are used to address this
problem. The first is to estimate tobit models rather than Ordinary Least Squares (OLS) as will
be discussed in the next section. The second approach is to use the 1996 SIPP-SER, once again,
to impute earnings for those topcoded in 1995 through 1998. 40 Results using both approaches
will be shown in section IV.
The implications of the fact that some children will have both covered and non-covered
earnings are not entirely clear. Essentially, it means that for a fraction of the children, observed
earnings from the SER will under-represent actual earnings. To the extent that this measurement
error in the dependent variable is random it will not bias the intergenerational elasticity
coefficient although it will enlarge the standard errors. On the other hand, if this error is
correlated with fathers’ earnings, then the results would be biased. It is not obvious why sons or
daughters whose SER earnings under-represent their true earnings would tend to have fathers
with lower average earnings.41 In any event, there is no simple way to solve the measurement

39

Non-covered are defined as those with zero covered earnings but who worked in each month of the year
that they are surveyed by the SIPP. Non-workers are classified as those with zero covered earnings who
worked between 0 and 2 months. See the Appendix for an explanation of this categorization.
40
Specifically, the mean value of SIPP earnings of those in the cohort with SIPP earnings above the social
security taxable maximum was calculated for 1996 and 1997. There was no significant difference between
the imputed values for men and women. The 1995 imputation value simply used the 1996 value converted
to 1995 dollars using the CPI. Similarly, the 1998 value used the 1997 inflation adjusted value.
41
One way that this could arise is if fathers who have some non-covered earnings, typically have lower
total earnings, and also have children who are more likely to have non-covered earnings and hence, lower
observed earnings. In this case the bias would be upwards. This can be seen in the following example: If
y1i = y1i* + τ, where y1i * is the actual child’s earnings and τ is the measurement error, then, plim of ρ “hat”
= ρ + (Cov(τ,y0i )/Var(y0i ). If errors are larger in magnitude (more negative) at low values of fathers’
earnings, y0i , then Cov(τ,y0i ) will be positive. While there is evidence that there is a positive
intergenerational correlation in self-employment status (Dunn & Holtz-Eakin, 1996) it is not clear that this
translates into a sizable correlation in overall non-covered status. In addition, there is no clear evidence

16

error problem for the dependent variable given the lack of direct survey data on the children in
their adult years.

Data Solutions: Parent Earnings
The problems with using SER data are considerably easier to deal with for the parents
because of the rich information available in the 1984 SIPP. In addition, it is also possible to use
the parents' earnings data directly from the SIPP for the years 1984 and 1985 as an additional
check on the results obtained using the SER, keeping in mind of course, the limitations of using
just 2 years of data. While this strategy results in smaller samples, they are still significantly
higher than those using the PSID and NLS.
The SIPP survey questions are particularly useful with respect to the first problem, that
zero earnings might reflect non-covered status. For 1984 and 1985 there is very detailed
information on labor force status and pay in each month and therefore it is quite easy to identify
whether individuals who had zero SER earnings also reported no paid weeks of employment for
the full year.42 For earlier years, a topical module from the second wave on labor force history is
used to classify fathers with zero earnings in each year as either non-covered or non-workers.43
Those classified as non-covered may either be dropped from the analysis or have their earnings
imputed using the SIPP earnings.

that the distribution of earnings among the self-employed is different from the overall population. A
second possibility is that the same form of measurement error exists for both children’s earnings and
fathers’ earnings. This might be the case if both generations’ earnings are measured using data from the
SER file and if non-covered status is correlated across generations. In this case the measurement error in
children’s earnings may be correlated with measured fathers’ earnings. If this correlation is large enough,
it might result in larger coefficients when SER data is used to measure fathers’ earnings than when SIPP
earnings are used. It turns out the opposite is true as is shown in section IV.
42
For individuals who are not in the SIPP for the full year of 1984 or 1985, the criteria are modified based
on whatever survey information is available in order to classify zero earners.
43
Specifically, the questionnaire asks individuals a series of questions about recent employment
experiences such as tenure and time between jobs, that enables one to construct instances of year long
unemployment spells, reasonably well. Because of evidence of poor recall into the distant past, the process
is only used to classify non-workers going back to 1979.

17

The issue of topcoding is far more severe for the fathers since the taxable maximum
affected a higher share of the sample in earlier decades. Specifically, for the sample of fathers,
the topcode rate was above 50 percent during the early 1970s falling to about 20 percent by the
mid-1980s. The approach taken to correct for this is to divide the fathers into 6 groups by race
and education level and to impute annual earnings based on information from the full sample of
the 1984 SIPP-SER or from the March Supplements to the Current Population Survey (CPS).
The procedure is described in greater detail in the Appendix.
The problem of measurement error due to fathers with both covered and non-covered
earnings is handled through the use of the "class of worker" variable in the 1984 SIPP. This
variable identifies those who worked for the government or who were self-employed at any point
that they were in the SIPP. These two categories comprise the vast majority of workers who have
some non-covered earnings. In addition to removing downward bias due to measurement error
this procedure has the additional advantage of reducing the possible bias arising from the joint
mismeasurement of fathers’ and children’s earnings, as was discussed earlier. One drawback of
this approach is that it reduces the sample size by roughly a third. In addition, because there is no
information on class of worker for years before the 1984 SIPP, the classification based on 1984
and 1985 must be imposed when averaging fathers' earnings over many years.

18

IV.

Methodology and Main Results

SIPP Results
This study begins by estimating the intergenerational elasticity in earnings between
fathers and their children using the SIPP earnings data for fathers. Although the SIPP is limited
to just two years of earnings and necessitates a smaller sample, it serves as a useful benchmark
for the main analysis that uses the SER data. The econometric approach follows the recent
literature and estimates the following equation:
(8)

y1i = α + ρy0i + β1 Age 0i + β2 Age 20i + β3 Age1i + β4 Age21i + ε

Specifically, y0i , the father's earnings, will be the log of the average annual earnings of
fathers over 1984 and 1985. This includes earnings from up to two jobs and two businesses. In
all aspects of this analysis, earnings are converted to 1998 dollars using the CPI.44 Only those
fathers with earnings that are not imputed by the Census Bureau due to nonresponse are included.
The father's age, Age 0i , and age squared, Age2 0i , are measured in 1984. The son's or daughter's
earnings, y1i , is the log of average annual earnings over the years 1995 to 1998. These years are
chosen so the kids are no younger than 27 in any of the years that their earnings are measured,
thereby giving a more reasonable picture of lifetime earnings.45 Each year's earnings for the sons
and daughters are first adjusted using the procedure described in section III to identify and then
impute the earnings of non-covered and non-workers. The children's age measures, Age1i and
Age2 1i , use their age in 1998. Table 2 presents the key sample statistics. Unlike some previous

44

Specifically this is the Bureau of Labor Statistics headline series (BLS code “CUUR0000SA0”) for all
urban consumers.
45
Solon (1999) has argued that studies with young samples have found lower correlations because of mean
reversion in the transitory income component, i.e. those with higher permanent income have lower
transitory incomes at a young age, thereby inducing an attenuation bias. The average age of the kids in this
study is 31 which is similar to the average age of 29.6 reported by Solon (1992) and 33.8 reported by
Zimmerman (1992). Averages are taken over several years for the children to address the criticism by
Couch and Lillard (1998) that Solon and Zimmerman both omit years of zero earnings among the children
in their work.

19

studies, if more than one child is matched to a father, all father-child cases are used and the
standard errors are corrected for within family correlation. 46
The model is estimated in two ways to deal with the issue of topcoded earnings of the
sons and daughters. One way is to simply use OLS, but adjust the dependent variable using the
imputed earnings calculated from the 1996 SIPP-SER when sons or daughters have been
topcoded. The second way is to set up a tobit model with an individual specific right-censoring
point, as follows:
(9)

y*1i = ρy0i + β1 Age0i + β2 Age20i + β3 Age1i + β4 Age2 1i + εi
(10)
(11)

y1i = y* 1i if y1it < topt ∀ t
y1i = k i , if y1it ≥ topt in some t

Here y1i is the observed level of permanent earnings which is equal to the actual
permanent earnings level, y*1i , only if annual earnings each year is below topt , the taxable
maximum earnings in each year. If earnings are topcoded in any one year, then the actual
permanent earnings are treated as right-censored at the observed point k i .47 The disturbance term
is assumed to be normally distributed and maximum likelihood estimation is used to estimate the
intergenerational elasticity. 48 In the case of the daughters, there is likely to be little difference
between the OLS and tobit estimates since few women in the sample are censored at the taxable
maximum.
In the first set of results, three different sample selection rules are used. First, fathers
who do not have positive earnings in both 1984 and 1985 are dropped. This has been the
common practice in previous research. Given that there are only two years of earnings, allowing

46

The effects of restricting the sample to only the oldest child in a family is shown later in the section.
A problem with this approach is that it treats individuals the same regardless of the number of times they
were censored over the four years. Ideally, one would want to estimate a tobit model for each year using a
standard human capital earnings function and then average the predicted earnings over the four years for
the censored observations. Given the lack of survey data for the sons and daughters as adults, this was not
possible.
48
The "intreg" command in STATA is used which allows for a variable censoring point for each
observation and for clustered standard errors.
47

20

zero earnings in any year is likely to add considerable noise. The other two exclusion rules drop
fathers who have earnings below a cutoff point in either year. The cutoffs used are $1000 and
$3000 in 1998 dollars.
The results are shown in Table 3. Without using any earnings cutoff, the father-son
elasticity which has been the focal point of the literature, is estimated at 0.342 using OLS and a
bit higher at 0.384 using the tobit specification. The elasticity between fathers and daughters is
also quite similar. The tobit estimate is 0.360, which is only slightly higher than the OLS
estimate of 0.341. The difference between OLS and Tobit should be quite small since only about
2 percent of the daughters are topcoded. The results for the daughters might be biased upwards if
the high incidence of non-working among daughters is due to other factors such as child-bearing
which in turn, is correlated with parent earnings. Using an earnings cutoff does not appear to
change the results appreciably. In these cases, the father-son earnings elasticity appears to drop
slightly while the father-daughter elasticity remains remarkably stable. A reasonable summary of
Table 3 is that the intergenerational elasticity is about 0.35 and is not significantly different
between sons and daughters.
It should be kept in mind that these results are based only on two-year averages of
fathers' earnings. The comparable result from Solon (1992) is 0.385 and from Zimmerman
(1992) is 0.481 49 . Couch and Lillard (1998) using selection rules similar to those employed by
Zimmerman on the same data, find the elasticity to be 0.37 when using a four-year average. This
suggests that simply using the two-year averages from the SIPP gives results similar to those
obtained using the PSID and NLS. At a minimum, this adds further confirmation to the argument
that the early studies that found elasticities of 0.2 or less did not accurately reflect the degree of
earnings mobility in the U.S.

49

This estimate for Solon is the average of the results found in Table 2, column 2 of Solon (1992). The
estimate for Zimmerman is from Table 6 column 2 of Zimmerman (1992).

21

SER Results
The second stage of the study uses the SER earnings data for the fathers. This not only
significantly enlarges the sample, since SIPP nonresponse and attrition is eliminated, but it also
allows for averaging fathers' earnings over many more years. This longer time period should
largely eliminate the problem of attenuation bias stemming from measurement error and
transitory fluctuations in earnings. Once again the earnings elasticity is estimated separately for
sons and daughters and also with both groups pooled. In this exercise all the results are based on
the tobit specification using the same dependent variable as in the prior analysis. Fathers'
earnings are progressively averaged more years beginning with the two-year average of 1984 and
1985 as was done with the SIPP earnings. Additional estimates are based on averages of four
years, seven years, ten years and sixteen years. In all cases the averages are taken over the range
of years ending in 1985.
Table 4 presents results using the SER data. There are two broad categories of selection
rules on fathers' earnings that are used in this analysis. In the top panel of the table, fathers’
earnings must be positive in each year. In the lower panel, some years of zero earnings are
allowed. Within each panel, there are three additional selection rules: non-covered fathers are
dropped; non-covered fathers’ earnings are imputed; and government and self-employed fathers
and non-covered fathers are dropped. In the first set of results in the top panel (row 1 of Table 4),
it is not necessary to actually identify covered status, since all fathers with years of zero earnings
are dropped. Therefore, it is possible to construct averages that include years prior to 1979.
Under the second rule (estimates in row 2), in contrast, averages can only be constructed going
back to 1979 since it is difficult to identify covered status in prior years. Under the third rule
(row 3), those identified as government or self-employed workers at any time during the 1984
SIPP survey period are dropped.
The results from using the two-year average with SER data are clearly lower than what
was found using the SIPP. The highest coefficient is 0.289 when non-covered fathers are

22

dropped from the analysis. The fact that many fathers have non-covered earnings (in addition to
covered earnings), that are not captured in the SER data is the obvious explanation for the greater
attenuation using the SER data. In fact, when non-covered fathers are dropped and earnings are
required to be at least $3000 in each year, thereby eliminating many of those whose covered
earnings severely misrepresent their true earnings, the estimated coefficient rises to 0.334 (not
shown) which is comparable to the SIPP results from Table 3. This suggests that the results
based on the SER may, in fact, be biased down by even more than would be the case with
comparable survey data. It also suggests that the possibility of upward bias from correlated
measurement error between fathers and children when using SER data is more than offset by the
overall attenuation bias.
Another finding that is readily apparent from Table 4 is that the estimated elasticity is
only slightly lower when the imputed non-covered fathers are added to the sample. In fact, when
fathers' earnings are averaged over short time horizons the results are sometimes larger with this
adjustment.
The most striking finding is that the elasticity rises dramatically as the fathers' earnings
are increasingly averaged over more years. Indeed, the estimated father-son elasticity is 0.613
when the fathers' earnings are averaged over 16 years. The father-daughter elasticity is a bit
lower at .570. When the sample of fathers is restricted to private sector, non self-employed
workers, however, the father-daughter elasticity is estimated at 0.754. Such a high degree of
transmission is rather surprising and may be due to the correlation between fathers' earnings and
daughters' labor force participation.

Does Excluding Years of Non-Employment Matter?
The estimates in the lower panel of Table 4 also suggest that the results are not sensitive
to the inclusion of years of zero earnings. For example, when averaging earnings from 1979 to
1985, allowing as many as four years of zero earnings to be averaged in, has almost no effect.

23

When non-covered fathers are dropped, the father-son elasticity estimate falls slightly from 0.445
to 0.434. However, when non-covered fathers are imputed, the coefficient actually rises, from
0.376 to 0.403. While the choice of how many years of zero earnings to include is somewhat
arbitrary, as long as one positive year of earnings is required, the estimated elasticity is raised
substantially from the results that allow for zero earnings in all years.50 To illustrate this,
Appendix Table A2 shows the effects of varying the number of years of zero fathers' earnings
that are included, on the father-son intergenerational elasticity. It seems reasonable to conclude
that the results are not very sensitive to this variation. 51 Given that children who are not working
are also not excluded from the analysis, the criticism by Couch and Lillard (1998) that high
estimates of the intergenerational elasticity are based on exclusion rules are not supported by this
dataset.

The Effects of Topcoding
A possible problem when using the SER data for fathers' earnings is topcoding of the
independent variable. In the absence of any correction, this would result in an upward bias in the
elasticity coefficient. Imputing the topcoded fathers with the mean level of earnings for those
topcoded, ideally, should correct this problem. 52 A simple way to check the robustness of the
results of this procedure is to simply drop the topcoded fathers. Table 5 presents the results of
this exercise when fathers who are topcoded in any year over the relevant time horizon are

50

In cases where fathers' earnings are zero in all years, obviously, the log-log specification is untenable.
Following Couch and Lillard (1998), cases with zero average earnings are recoded as $1 so that the log
would be zero rather than negative infinity. It is not clear that this is a reasonable approach since zeroes on
a log-scale may significantly alter the results due to the leverage of such observations. In other
specifications (not shown) recoding zero earnings in a range from $500 to $3000 (6.2 to 8.0 on the log
scale) substantially raises the coefficient compared to what is shown in Table A2.
51
It should be noted that the results from the last row of Table 4 which average over 10 years and 16 years,
include years of zero earnings that are due to non-covered status. In these cases, more restrictive rules are
used. It was decided that fathers must have positive earnings in about 70 percent of the years in these
cases. This was chosen because under this rule, the results for a seven-year average when non-covered
zeroes are included is similar to the results when only zeroes due to non-working are allowed.
52
Of course, this assumes that true statistical model is a linear relationship between fathers' earnings
children's earnings, which itself, is the subject of inquiry in Section V.

24

dropped from the sample. The results are shown for sons and daughters pooled, in order to try to
keep the sample as large as possible. For the most part it appears that dropping these fathers
lowers the estimates of the elasticity. When using the seven-year average, however, the results
are still quite similar. Including topcoded fathers results in an estimate of 0.472 while dropping
these observations fathers results in an estimate of 0.439. Averaging fathers' earning using years
before 1979 is particularly troublesome because the taxable maximum in real terms was so much
lower during that time. As a result so many of the observations are topcoded, and hence,
dropped, that it is not clear that the results are meaningful. In fact, the average over 1970 to 1985
has a sample too small to even precisely estimate the coefficient.

The Role of Persistent Transitory Earnings
Given the high estimated elasticity in this sample, a natural question is what explains the
difference between the results presented here and the earlier literature? Since estimates taken
over shorter time periods match the results from previous studies it does not appear that there is
anything especially different about the sample, data or cohort that was chosen. The analysis from
section II suggests that it might be the case that short-term fluctuations in earnings are highly
persistent and are not adequately “averaged away”, especially when averages are taken over short
time periods. The simulation exercise presented in section II suggested that the intergenerational
elasticity calculated with a five-year average may be biased down by twenty five to thirty percent
under plausible assumptions. Taking this approach a step further, the entire "path" of the
attenuation factor can be plotted as the average of fathers earnings are taken over more years.
This can then be compared to the empirical results in this study under various assumptions on the
true intergenerational elasticity. Figure 3 shows this comparison using assumptions based on

25

results found by Card (1994) and Hyslop (2001) and assuming that the true intergenerational
elasticity is 0.6 as the results of Table 4 indicate.53
The simulated attenuation bias declines but a slowing rate as more years are used in the
averaging process. The results from Table 4, in contrast, show a more linear increase in the
estimated coefficient. In fact, the results when fathers' earnings are averaged over 16 years,
appear to be somewhat higher than what would be predicted using the simulated model when the
true intergenerational elasticity is assumed to be 0.6. A likely explanation is that transitory
fluctuations vary over an individual's lifespan, whereas in this simulation they are treated as
constant. Gordon (1984) and Baker and Solon (1999) for example, have shown that the transitory
variance follows a "U-shape" over an individual's lifetime. If this is indeed the case, then the
attenuation factor, λ*t , should be somewhat higher than the simple simulation predicts when the
fathers' earnings are averaged using years when their age is closer to forty. The longer term
averages may result in higher estimates because they average in years when there is less
transitory noise.
In the final analysis, this exercise is merely suggestive. The estimates of the
intergenerational elasticity are subject to sampling error and we certainly do not know the “true”
parameter values of the statistical model. Analysis of other matched datasets that may become
available in thecoming years and which do not suffer from some of the problems present in this
study may help to resolve these measurement issues. Still, highly serially correlated transitory
earnings along with lifecycle bias appears to be a reasonable explanation for why the results from
five-year averages might be so different from that found using a sixteen-year average.

Other Sample Selection Issues

53

The procedure for the simulation is described in detail in the Appendix.

26

There are some issues related to the construction of the matched dataset that can
potentially bias the results. First, children must have been coresident with their parents or living
away at college at the beginning of the survey. Second, in order to have been matched, they must
also have a social security number that was provided to the interviewer. To handle the first
problem, the sample of all individuals born between 1963 and 1968 in the 1984 SIPP were
divided into 24 groups by year of birth, sex and race. The rate of "living at home" was calculated
for each group. The inverse of these rates could then be used to weight the children in the
intergenerational samples used in this study. For the problem of matches based on social security
numbers, a probit analysis was done to predict the likelihood that individuals from the cohort in
the SIPP would be matched to their fathers. The inverse of the predicted probabilities can also be
used to weight the father-son pairs in the analysis. Table 6 shows the effects of incorporating
these weights on the estimated elasticities using the SIPP-based sample of fathers. The first row
simply presents the earlier estimated results from the bottom row of Table 3. The second row
weights the observations by the inverse of the probability that they are both living at home and
have provided a social security number. The overall elasticity when sons and daughters are
pooled is identical at 0.365 but rises slightly for sons and falls slightly for daughters.
Other variations are also attempted in Table 6. Restricting the sample to only the oldest
child in each family has a small but insignificant effect on sons and virtually no effect on
daughters. Dropping those aged 19 or 20 in 1983 lowers the elasticity to 0.283. The difference is
still within the sampling error but might indicate some effect. The result is consistent with the
observation by Solon (1999) that using the earnings of sons when they are observed at a younger
age can bias the results downwards. It is probably not due to the fact that older kids living at
home are more similar to their parents since many of those aged 19 or 20 are actually attending
college. The final two rows of Table 6 use different sample selection rules on children. Dropping
those children identified as non-covered rather than imputing them has almost no effect for sons
but a significant positive effect on daughters. Finally, it might be the case that outliers due to

27

extremely low values of children’s average earnings have influenced the parameter results. The
approach used to correct for this is possibility is to drop those children who are identified as nonworkers in more than two of the four years. This rule appears to have no effect.54

54

There is still some possibility that children with positive, but very low covered earnings, come from
families whose fathers, on average, have lower earnings, introducing an upward bias as described in section
II. Unfortunately there does not appear to be any way to correct for this possibility with the available data.

28

V.

Further Extensions

Family Income
An interesting finding in some previous intergenerational studies is that family income is
more highly correlated across generations than is fathers' earnings.55 Most of these studies,
however, have not discussed this result in much detail. 56 While this study is limited to the use of
earnings as an outcome for children it can examine the effects of other measures of parental
economic status. The use of family income provides a broader measure that includes not only the
mother but also incorporates other forms of non-earnings income into the analysis. Although the
SER data does not have data on other forms of income, the SIPP is particularly useful because it
provides a very detailed breakdown of sources of income that can be used for the parents. Table
7 provides the results of an analysis that substitutes income for earnings in the model and also
looks separately at two parent families, single mother families and both types of families pooled
together.57 In all cases, only parents whose income measure exceeds $3000 in 1998 dollars in
1984 and 1985 are included. Using fathers' income rather than earnings raises the
intergenerational elasticity quite a bit. For sons the estimate increases from 0.349 to 0.518.
Using income rather than earnings also appears to raise the elasticity sharply when two parents
are used and if only single mothers are examined. Adding mothers to the analysis also appears to
raise the elasticity, particularly for daughters. For example, looking at both parents' income
instead of just the fathers' income, raises the elasticity with daughters earnings from 0.496 to

55

These include Mulligan (1997), Shea (1997), Solon (1992), Altonji and Dunn (1991), Corak and Heisz
(1999) and Peters (1992).
56
The exception is Mulligan (1997) who argues that this result makes perfect sense in a standard
intergenerational permanent income model. In such a model under certain assumptions earnings mobility is
dictated by regression to the mean in ability which might be relatively rapid. Income mobility, however,
might be much slower because of financial asset transfers from parents to children irrespective of
investment in children's human capital. That analysis, however, does not explain differences between the
effects of parents' earnings as compared to parents' income on children's earnings.
57
Single-mother families are simply those where there is no spouse identified for the mother. Obviously,
this will miss unmarried couples and other living arrangements where there might be additional sources of
income.

29

0.708. The comparable increase for sons is from 0.518 to 0.553. Looking at single mothers only,
the estimated elasticities are dramatically lower, and in most cases, statistically insignificant.
This is no doubt due to poor classification of families and therefore, significant mis-measurement.
What might explain the higher results from parental income? For one thing, income may
be a less noisy measure of economic status than earnings. This is likely to be particularly true at
the low end of the parents' earnings distribution where individuals may receive income at times
when they receive virtually no earnings due to unemployment, e.g. unemployment insurance or
workers compensation. This may result in a higher estimated elasticity when parents' income is
used rather than earnings because of a smaller attenuation bias due to measurement error or
transitory shocks. In addition, there appears to be a sample selection effect. If the
intergenerational elasticity is higher at the low end of the distribution, and if more fathers are
dropped from the earnings analysis because of exclusion rules on earnings, then including these
individuals by using income rather than earnings might raise the elasticity. In fact, if the same
sample that is used to estimate the elasticity with fathers earnings in row 1 is also used to estimate
the elasticity with fathers' income, then the latter estimate falls from 0.518 to 0.385 (not shown).
In any case, it appears that using income rather than earnings for parents may give a more
accurate reading of intergenerational mobility, especially when only a few years of parents
earnings data are available.

Borrowing Constraints
Theoretical models of intergenerational mobility have emphasized borrowing constraints
as a key factor in the transmission of earnings inequality. Becker and Tomes (1986) and
Mulligan (1997) have argued that if parents can borrow from their children's future earnings, then
all parents will invest the optimal amount in their children's human capital. If earnings are
determined by human capital, and human capital is a function of ability, then the intergenerational
elasticity in earnings will only be positive if earnings and ability are correlated and will depend

30

on the rate at which ability regresses to the mean. With borrowing constraints, however, parents
with low income and able children will not invest the optimal amount in their children's education
inducing a higher intergenerational elasticity in earnings. Mulligan (1997) has attempted to test
this hypothesis using the PSID and by splitting the sample by those who expect to receive an
inheritance. He found no significant difference in elasticities between the two groups. One
problem with this approach is that it does not directly measure parents ability to finance
schooling for their children at the time that such an investment is made. Mulligan’s measure also
does not capture intervivos transfers. The model focuses solely on an intergenerational budget
constraint and does not analyze parents’ potential inability to borrow from their own future
income.
There are several advantages that this dataset can bring to this question. First, with a
larger sample it is possible to split the sample along some dimension that directly reflects fathers'
ability to access capital, and still estimate the parameters reasonably well. 58 Second, the topical
module from wave 4 of the 1984 SIPP can be used to gather more detailed information on
household balance sheets to more accurately classify families by their ability to invest fully in
their children's human capital. It was decided to use net worth to classify fathers as either
borrowing constrained or not borrowing constrained. This measure captures the ability of
individuals to borrow against their current wealth or to draw down assets in order to finance
human capital acquisition. One problem with this approach, of course, is that the measure is from
1984, when kids are aged 16 to 21 while the relevant period to measure borrowing constraints is
arguably at an earlier point in the child's educational career. In addition, since net worth and
income are highly correlated, any nonconstancies in the intergenerational income elasticity may
also be reflected in differences in ρ by levels of net worth that may or may not be due to
borrowing constraints.

58

Another approach is to include nonlinearities in fathers’ earnings. Experimentation with this approach
did not yield any statistically significant results.

31

Table 8 shows the results of this exercise. First, using the SIPP for parents' earnings, and
dividing the sample by the median level of net worth ( about $65,000 in 1984 dollars) the results
point to a sharp difference between those below the median and those above. The elasticity is
0.422 for those with lower than median net worth but only 0.140 for those above the median
level. While the difference is large, one could not reject the null hypothesis of equality at the 5
percent significance level. The second set of results compares those at or below the first quartile
of net worth with those at the top quartile. In this case the difference is even more dramatic and
is statistically significant. In fact for the top quartile, there appears to be zero elasticity. Indeed,
the permanent income model would predict this result if income is uncorrelated with ability.
Similar attempts were less conclusive using SER data for fathers' earnings as the bottom half of
Table 8 shows. A possible explanation for this result is that the high topcoding rate of fathers
compresses the fathers earnings distribution and given the strong correlation between net worth
and earnings, the full variation in the intergenerational elasticity is also compressed.

Differences by Race
One of the key comparisons that has not been explored in previous studies is whether
there are significant differences in mobility between blacks and whites. A higher elasticity
among blacks might suggest that even if overall mobility is high, economic progress for blacks
might be more difficult for other reasons such as borrowing constraints, neighborhood effects or
discrimination. Once again, obtaining reasonable sample sizes for such a comparison has been
virtually impossible in previous datasets. Table 9 shows the difference in estimates for blacks
and whites. Using the seven-year average of fathers' earnings from the SER, the elasticity among
blacks (0.487) was found to be nearly twice as high as the elasticity among whites (0.271) but the

32

difference was not significantly greater than zero at the five percent level. 59 In order to keep the
sample size as large as possible, the SER results imputed non-covered fathers and used all fathers
with positive earnings in any year.60 Additional results were attempted using the SIPP for parent
earnings. The comparison of fathers' earnings by race yields a very similar result to what was
found using the SER. The difference in elasticities is estimated at 0.222 which is nearly identical
to the 0.216 obtained using the SER sample, but in this case the smaller sample leads to a far less
precise estimate. The comparison of fathers' income elasticities leads to a larger difference,
though it is still estimated imprecisely. Looking at combined two parent earnings and income,
however, leads to incredibly large estimates for blacks that exceed 1. If taken seriously, it implies
no regression to the mean. The difference in estimates when using two parents is on the border of
significance at the five percent level. The results are similar, though less precise, when the
samples include only low net worth families (not shown) suggesting that the racial difference is
not simply due to borrowing constraints.
One difficulty in these comparisons lies in family composition. A much higher
percentage of black families are headed by single mothers where the estimated elasticities are
substantially lower (see Table 7). The small sample size of single mother families, however, does
not permit a breakdown by race. While further research is clearly needed, the results presented
here are suggestive of less mobility among blacks. Some plausible explanations for the higher
persistence might lie in employment discrimination, borrowing constraints, neighborhood effects,
inferior schools or disparities in home ownership.

59

The seven-year average was used because that is the longest average over which there is still a
classification of social security coverage status among fathers. This allows inclusion of zero earning years
that reflect non-employment but not non-coverage.
60
Using more restrictive exclusion rules raises the estimated correlation for whites slightly and lowers the
correlation for blacks slightly. The difference remains large but insignificant.

33

VI.

Conclusion
The study uses a new nationally representative intergenerational sample and finds strong

evidence that there is far less intergenerational mobility in the United States than was previously
thought. The unique advantage of this dataset is the availability of long-term earnings histories of
fathers. It appears that it is precisely this characteristic of the data, which results in the higher
estimates. Indeed, estimates based on short-term averages of father earnings closely track the
existing literature. Averages of fathers' earnings taken over long periods of time, however,
appear to be less sensitive to transitory fluctuations that many studies have shown are highly
persistent. Short-term proxies for permanent income may also be susceptible to lifecycle bias due
to the fact that the variance of the transitory component of earnings varies considerably by age.
Overall, the results point toward an intergenerational elasticity of about 0.6. If accurate, this
suggests that many well-documented wage gaps may persist for several generations.
The results appear to be fairly robust to sample selection rules, the match process, and to
the problems that are inherent in the use of social security earnings data. Ideally, future research
should attempt to verify the results here using long-term measures of permanent earnings from
other sources that do not require the kind of imputations that were necessary in this study. It may
be difficult, however, given that existing public use longitudinal data sets suffer from attrition and
lifecycle bias and have significantly smaller samples. What may be required in the future is
access to other administrative datasets that overcome these data problems.
The use of highly detailed survey data on income from the SIPP from just two years also
appears to bolster the main findings. The elasticity of parent income on children’s future earnings
is estimated to be in the 0.5 to 0.6 range.
While this study provides new descriptive evidence of the extent of mobility in the U.S.
there is still a tremendous amount that is not understood about how the transmission process
works. To what extent is the high estimate of the intergenerational elasticity truly a reflection of
the importance of financial resources as opposed to less tangible characteristics that cannot be

34

influenced by public policy? While far from conclusive, new evidence is provided suggesting
that intergenerational inequality may be related, in part, to access to capital. This offers a
potential avenue by which greater mobility may be fostered through public policy.
Some suggestive evidence also points to less mobility among blacks, a minority group
that has struggled to achieve full economic parity many decades after the end of slavery. This
suggests that the black-white wage gap may take considerably longer to equalize than
discrepancies among other groups.

35

References
Altonji, Joseph G. and Thomas A. Dunn (1991), "Relationships among the Family Incomes and
Labor Market Outcomes of Relatives," Research in Labor Economics, 12:269-310.
Atkinson, A.B., A.K. Maynard and C.G. Trinder (1983), Parents and Children: Incomes in Two
Generations (Heinemann, London).
Baker, Michael and Gary Solon (1999), "Earnings Dynamics and Inequality Among Canadian
Men, 1976-1992: Evidence From Longitudinal Tax Records," NBER Working Paper 5622,
National Bureau of Economic Research.
Becker, Gary S. and Nigel Tomes (1979), "An Equilibrium Theory of the Distribution of Income
and Intergenerational Mobility," Journal of Political Economy, 87:1153-1189.
Becker, Gary S. and Nigel Tomes (1986), "Human Capital and the Rise and Fall of Families,"
Journal of Labor Economics, 4:S1-S39.
Becker, Gary S. (1988), "Family Economics and Macro Behavior," American Economic Review,
78:1-13
Bowles, Samuel (1972). “Schooling and Inequality from Generation to Generation.” Journal of
Political Economy 80: S219-251.
Behrman, Jere R. and Paul Taubman (1985), "Intergenerational Earnings Mobility in the United
States: Some Estimates and a Test of Becker's Intergenerational Endowments Model," Review of
Economics and Statistics, 67:144-151.
Card, David (1994), "Intertemporal Labor Supply: An Assessment", in Christopher A. Sims (ed.)
Advances in Econometrics, Sixth World Congress, Vol. 2, Cambridge University Press.
Cambridge.
Chay, Kenneth Y. (1995), "Evaluating the Impact of the 1964 Civil Rights Act on the Economic
Status of Black Men Using Censored Longitudinal Earnings Data," October 1995, mimeo.
Chay, Kenneth Y. and Bo E. Honoré (1998), "Estimation of Semiparametric Censored Regression
Models: An Application to Changes in Black-White Earnings Inequality During the 1960s,"
Journal of Human Resources 33(1):4-38.
Corak, Miles and Andrew Heisz. (1999). "The Intergenerational Earnings and Income Mobility of
Canadian men: Evidence from Longitudinal Income Tax Data." Journal of Human Resources
34(3):504-533.
Couch, Kenneth A. and Dean R. Lillard (1998). "Sample Selection Rules and the
Intergenerational Correlation of Earnings." Labour Economics 5: 313-329.
Dearden, Lorraine, Stephen Machin, and Howard Reed (1997), "Intergenerational Mobility in
Britain," Economic Journal, 107:47-66.
Duncan, Greg, Johanne Boisjoly and Timothy Smeeding (1996), "Economic Mobility of Young
Workers in the 1970s and 1980s," Demography, 33:497-509.

36

Dunn, Thomas and Douglas Holtz-Eakin (1996), "Financial Capital, Human Capital and the
Transition to Self-Employment: Evidence from Intergeneration Links", NBER Working Paper
5622, National Bureau of Economic Research
Goldberger, Arthur S. (1989), "Economic and Mechanical Models of Intergenerational
Transmission", American Economic Review, 79:504-513.
Gordon, Roger H.. "Differences in Earnings and Ability," Garland, New York.
Hyslop, Dean (2001), "Rising U.S. Earnings Inequality and Family Labor Supply: The
Covariance Structure of Intrafamily Earnings," American Economic Review, 91:755-777.
Krueger, Alan (1995), "The Legacy of Separate and Unequal Schooling" Paper presented at
conference of Southern Economic Association, November 18-20, New Orleans, LA.
Levine, David I. (2000). "Choosing the Right Parents: Changes in the Intergenerational
Transmission of Inequality Between the 1970s and the Early 1990s", manuscript, University of
California, Berkeley.
Lillard, Lee A. and Robert J. Willis (1978), "Dynamic Aspects of Earning Mobility,"
Econometrica 46:985-1012
Mazumder, Bhashkar (2001), “Earnings Mobility in the U.S.,: A New Look at Intergenerational
Inequality”, PhD. Dissertation, University of California, Berkeley.
MaCurdy, Thomas E. (1982), “The Use of Time Series Processes to Model the Error Structure of
Earnings in a Longitudinal Data Analysis.” Journal of Econometrics, 18:83-114.
Minicozzi, Alexandra L. (1997), "Nonparametric analysis of intergenerational income mobility",
PhD dissertation (University of Wisconsin).
Mulligan, Casey B. (1997), Parental Priorities and Economic Inequality. University of Chicago
Press, Chicago.
Peters, Elizabeth H. (1992). "Patterns of Intergenerational Mobility in Income and Earnings,"
Review of Economics and Statistics, 74:456-466.
Sewell, William H. and Robert M. Hauser (1975), Education, Occupation and Earnings:
Achievements in the Early Career. Academic Press, New York.
Smith James P. and Finis R. Welch (1986), "Closing the Gap: Forty Years of Economic Progress
for Blacks.", The Rand Corporation, Santa Monica CA.
Solon, Gary (1989), "Biases in the Estimation of Intergenerational Earnings Correlations" Review
of Economics and Statistics, 71:172-174
Solon, Gary (1992), "Intergenerational Income Mobility in the United States," American
Economic Review, 82:393-408

37

Solon, Gary (1994) "Comments on 'Sample Selection Rules and the Intergenerational Correlation
of Earnings: A Comment on Solon and Zimmerman'". Unpublished manuscript..
Solon, Gary (1999), "Intergenerational Mobility in the Labor Market," Handbook of Labor
Economics, Elsevier
Zimmerman, David J. (1992), "Regression Toward Mediocrity in Economic Stature," American
Economic Review, 82:409-429

38

Table 1: Simulation Results on Attenuation Bias when Using Multiyear Averages
Attenuation Coefficient if….

Number of
Years Averaged
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

δ = 0 .5

0.641
0.733
0.783
0.817
0.843
0.863
0.879
0.892
0.904
0.913
0.921
0.928
0.935
0.940
0.945
0.949
0.953
0.957
0.960
0.963
0.966
0.968
0.970
0.973
0.974
0.976
0.978
0.980
0.981
0.982

δ = 0.8

0.670
0.735
0.767
0.790
0.808
0.823
0.837
0.849
0.859
0.869
0.878
0.887
0.895
0.902
0.908
0.915
0.921
0.926
0.931
0.936
0.940
0.945
0.948
0.952
0.956
0.959
0.962
0.965
0.968
0.971

δ = 0.5

0.519
0.630
0.693
0.737
0.772
0.799
0.821
0.840
0.856
0.869
0.881
0.891
0.900
0.908
0.915
0.922
0.927
0.933
0.937
0.942
0.946
0.950
0.953
0.956
0.959
0.962
0.964
0.967
0.969
0.971

δ = 0 .8

0.554
0.637
0.680
0.710
0.734
0.754
0.772
0.788
0.802
0.815
0.827
0.839
0.849
0.859
0.868
0.877
0.885
0.892
0.899
0.906
0.912
0.918
0.924
0.929
0.934
0.939
0.943
0.947
0.952
0.955

δ = 0.5

0.526
0.619
0.677
0.720
0.754
0.782
0.806
0.826
0.843
0.857
0.870
0.882
0.892
0.900
0.908
0.916
0.922
0.928
0.934
0.939
0.943
0.947
0.951
0.955
0.958
0.961
0.964
0.967
0.970
0.972

δ = 0 .8

0.572
0.629
0.662
0.687
0.709
0.728
0.746
0.762
0.777
0.792
0.805
0.817
0.829
0.840
0.850
0.860
0.869
0.877
0.886
0.893
0.901
0.908
0.914
0.920
0.926
0.932
0.937
0.942
0.947
0.952

Note: Simulation is based on equation 14 (See appendix). In the first pair of columns, the share of
single year variance in earnings accounted for by permanent factors is 0.7. In the last two pairs of
columns the share is assumed to be 0.5. Within each pair of columns, assumptions are made about
the share of transitory variance in the variance of a single year of earnings and the auto correlation
coefficient. The assumptions based on Hyslop (2001) are shown in bold.

Table 2: Summary Statistics for Fathers and Children

Samples using 1984 SIPP for fathers' earnings
Variable
Father's Age in 1984
Log Average Father's Earnings 84-85
Son's Age in 1998
Log Average Son's Earnings 95-98
Daughter's Age in 1998
Log Average Daughter's Earnings 95-98

N
796
796
796
796
719
719

Mean
46.9
10.4
32.4
10.0
32.5
9.1

S.D.

Mean
47.1
10.5
10.6
10.7
10.7
10.7
32.4
10.0
32.5
9.1

S.D.
6.3
0.9
0.7
0.6
0.5
0.4
1.7
1.2
1.7
1.8

6.2
0.8
1.7
1.2
1.7
1.7

Minimum Maximum
28
71
6.1
11.9
30
35
2.5
11.1
30
35
4.1
11.1

Samples using SER for fathers' earnings
Variable
Father's Age in 1984
Log Average Father's Earnings 84-85
Log Average Father's Earnings 82-85
Log Average Father's Earnings 79-85
Log Average Father's Earnings 76-85
Log Average Father's Earnings 70-85
Son's Age in 1998
Log Average Son's Earnings 95-98
Daughter's Age in 1998
Log Average Daughter's Earnings 95-98

N
1262
1262
1218
1160
1111
1063
1262
1262
1178
1178

Minimum Maximum
27
69
4.0
11.5
6.5
11.5
7.3
11.5
7.7
11.3
8.1
11.2
30
35
2.5
11.1
30
35
3.1
11.1

Note: All earnings are converted to 1998 dollars using the CPI. Children's earnings are imputed for
those predicted to be non-covered or non-workers as described in text. The SIPP sample pertains to
those shown in row 1 of Table 2. Fathers in SIPP sample must be present for all of 1984 and 1985
and have no instances of nonresponse to earnings questions. The samples for the SER pertains to
those shown in row 1 of Table 3. Fathers' age in the SER sample is for the sample used when
earnings are averaged over 1984-1985. SER earnings of those topcoded are imputed as described in
text. For both SIPP and SER samples, father statistics correspond to the relevant father-son
samples.

Table 3: Intergenerational Elasticities Using SIPP for Fathers' Earnings
elast.
(s.e.)
N

Dependent Variable
is Log Avg. Earnings, 1995-1998

Sons

Daughters

Pooled

Fathers

Tobit

OLS

Tobit

OLS

Tobit

OLS

Log Avg. 84-85
Father Earnings >0
each year

0.384
(0.091)
796

0.342
(0.085)
796

0.360
(0.106)
719

0.341
(0.103)
719

0.357
(0.074)
1515

0.322
(0.070)
1515

Log Avg. 84-85
Father Earnings >1000
each year

0.337
(0.080)
788

0.293
(0.072)
788

0.367
(0.117)
713

0.346
(0.112)
713

0.339
(0.074)
1501

0.300
(0.069)
1501

Log Avg. 84-85
Father Earnings >3000
each year

0.349
(0.078)
767

0.292
(0.070)
767

0.361
(0.128)
702

0.337
(0.122)
702

0.365
(0.080)
1469

0.315
(0.074)
1469

Note: For the dependent variable, probit models based on the 1996 SIPP matched to SER were
used to determine if zero earnings reflected noncoverage or non-worker status and were imputed
accordingly. In the case of OLS specification, topcoded children are imputed based on the
earnings distribution in 1996 SIPP-SER. Fathers must have been present for all interview months
and have no cases of nonresponse to earnings questions. Standard errors are adjusted for within
family correlation when more than one sibling is present.

Table 4: Intergenerational Elasticities Using SER for Fathers' Earnings
elast.
(s.e.)
N
Fathers
Log Avg. Earn.

Dependent Variable is Children's Log Avg Earnings, 1995-1998
All results use tobit specification

84-85

82-85

Sons
79-85

76-85

70-85

84-85

82-85

Daughters
79-85 76-85

70-85

84-85

82-85

Pooled
79-85 76-85

70-85

Father Earnings Must be Positive Each Year
Drop
Non-Covered
Fathers

0.253 0.349 0.445 0.553 0.613
(0.043) (0.059) (0.079) (0.099) (0.096)
1262
1218
1160
1111
1063

0.363 0.425 0.489 0.557 0.570
(0.065) (0.087) (0.110) (0.140) (0.159)
1178
1124
1070
1031
982

0.312 0.385 0.472 0.570 0.624
(0.041) (0.056) (0.071) (0.088) (0.099)
2440
2342
2230
2142
2045

Impute
Non-Covered
Fathers

0.289 0.313 0.376
(0.050) (0.052) (0.062)
1485
1462
1433

0.350 0.395 0.422
(0.062) (0.081) (0.096)
1360
1339
1310

0.323 0.358 0.406
(0.041) (0.051) (0.059)
2845
2801
2743

Drop
Government &
Self-Employed

--

--

0.273 0.419 0.474 0.533 0.652
(0.060) (0.082) (0.096) (0.111) (0.135)
844
825
801
779
746

--

--

0.526 0.563 0.635 0.750 0.754
(0.089) (0.137) (0.150) (0.173) (0.192)
782
758
736
719
690

--

--

0.394 0.487 0.557 0.659 0.727
(0.062) (0.084) (0.094) (0.109) (0.128)
1626
1583
1537
1498
1436

Allow Some Years of Zero Father Earnings*
Drop
Non-Covered
Fathers

0.234 0.334 0.434
(0.043) (0.057) (0.069)
1295
1268
1227

--

--

0.312 0.423 0.506
(0.060) (0.065) (0.091)
1201
1168
1127

--

--

0.264 0.372 0.474
(0.037) (0.046) (0.059)
2496
2436
2354

--

--

Impute
Non-Covered
Fathers

0.238 0.342 0.403
(0.042) (0.057) (0.059)
1534
1550
1571

--

--

0.295 0.384 0.474
(0.055) (0.061) (0.080)
1394
1406
1424

--

--

0.260 0.357 0.438
(0.035) (0.044) (0.052)
2928
2956
2995

--

--

Drop
Government &
Self-Employed

0.242 0.355 0.441 0.523 0.575
(0.059) (0.080) (0.084) (0.101) (0.109)
874
869
862
895
917

0.400 0.504 0.600 0.731 0.847
(0.084) (0.083) (0.113) (0.130) (0.145)
803
794
785
825
831

0.294 0.417 0.519 0.626 0.704
(0.051) (0.064) (0.072) (0.086) (0.094)
1677
1663
1647
1720
1748

Note: See text for how children's and fathers' earnings are constructed. Standard errors are adjusted for multiple siblings. *Required years of pos.
earnings are: 1 for 2-yr. averages; 2 for 4-yr. averages; 3 for 7-yr. averages; 7 for 10 yr.-averages and 11 for 16-yr. averages.

Table 5: Effects of Top-Coded Fathers on Intergenerational Elasticities

elast.
(s.e.)
N
Fathers
Log Avg. Earn. Over…

Dependent Variable is Children's Log Avg Earnings, 1995-1998
All results use tobit specification

84-85

Pooled (Sons & Daughters)
82-85
79-85
76-85

70-85

Positive Earnings
Each Year

0.312
(0.041)
2440

0.385
(0.056)
2342

0.472
(0.071)
2230

0.570
(0.088)
2142

0.624
(0.099)
2045

Positive Earnings
Each Year
Drop Topcoded dads

0.245
(0.049)
1713

0.317
(0.074)
1530

0.439
(0.121)
1144

0.451
(0.182)
784

0.295
(0.237)
343

Note: For the dependent variable, probit models based on the 1996 SIPP matched to
SER were used to determine if zero earnings reflected noncoverage or non-worker
status and were imputed accordingly. For fathers, SER earnings for those identified
as non-covered are dropped. In row 1, earnings for those topcoded are imputed using
March CPS data for 1970-80 and using 1984 SIPP for 1981 to 1984. In row 2 fathers
topcoded in any year over the relevant period are dropped.

Table 6: The Effects of Sample Selection Using the SIPP for Fathers' Earnings

elast.
(s.e.)
N

Dependent Variable
is Log Avg. Earnings, 1995-1998
All results use tobit specification

Fathers

Sons

Daughters

Pooled

Log Avg. 84-85
Father Earnings >3000 each year

0.349
(0.078)
767

0.361
(0.128)
702

0.365
(0.080)
1469

Weighted for
Match Likelihood
& Prob Living at home

0.375
(0.086)
767

0.339
(0.128)
702

0.365
(0.084)
1469

Eldest Kids Only

0.386
(0.095)
548

0.357
(0.147)
506

0.358
(0.092)
1054

Aged 15 to 18 only

0.283
(0.085)
542

0.400
(0.155)
486

0.367
(0.095)
1028

Non-Covered
Children are Dropped

0.362
(0.094)
644

0.473
(0.113)
498

0.409
(0.074)
1142

Require 2 years of Positive*
Children's Earnings

0.358
(0.080)
736

0.363
(0.130)
687

0.369
(0.082)
1423

Note: For the dependent variable, probit models based on the 1996 SIPP matched to
SER were used to determine if zero earnings reflected noncoverage or non-worker
staus and were imputed accordingly (except where otherwise indicated). Fathers'
earnings from 1984 SIPP required that the father be present for all interview months
and have no cases of nonresponse to earnings questions. Standard errors are
adjusted for within family correlation when more than one sibling is present.
*Really, this means children cannot be classified as non-workers in more than two
years.

Table 7: Intergenerational Elasticity of Parents' Income on Children's Earnings

elast.
(s.e.)
N

Dependent Variable
is Log Avg. Earnings, 1995-1998
All results use tobit specification
Sons

Daughters

Pooled

Father Earnings
Log Avg. 84-85

0.349
(0.078)
767

0.361
(0.128)
702

0.365
(0.080)
1469

Father Income
Log Avg. 84-85

0.518
(0.102)
871

0.496
(0.119)
773

0.499
(0.088)
1644

Two Parent Earnings
Log Avg. 84-85

0.385
(0.075)
776

0.491
(0.118)
719

0.444
(0.073)
1495

Two Parent Income
Log Avg. 84-85

0.553
(0.103)
842

0.708
(0.118)
768

0.635
(0.086)
1610

Single Mother Earnings
Log Avg. 84-85

0.215
(0.170)
161

0.357
(0.306)
145

0.239
(0.178)
306

Single Mother Income
Log Avg. 84-85

0.362
(0.151)
231

0.287
(0.183)
219

0.320
(0.123)
450

All Family Earnings
Log Avg. 84-85

0.322
(0.060)
959

0.502
(0.098)
879

0.406
(0.058)
1838

All Family Income
Log Avg. 84-85

0.478
(0.067)
1105

0.558
(0.080)
1006

0.523
(0.056)
2111

Note: Probit models based on the 1996 SIPP matched to SER were used to determine
if children's zero earnings reflected noncoverage or non-worker status and were
imputed accordingly. For SIPP parent measures, parent must be present for all
interview months and have no cases of nonresponse to earnings questions. All
parent measures require earnings greater than $3000 in 1998 dollars in 1984 and 1985.
Standard errors are adjusted for within family correlation when more than one sibling
is present.

Table 8: Intergenerational Elasticity by Level of Net Worth

elast.
(s.e.)
N

Dependent Variable
is Log Avg. Earnings, 1995-1998
All results use tobit specification
Pooled (Sons and Daughters)

Overall

High
Net Worth

Low
Net Worth

Diff.

t-stat

SIPP Results
Father Earnings
Log Avg. 84-85
Low is <=median
High is >median

0.358
(0.074)
1514

Father Earnings
Log Avg. 84-85
Low is <=25th percentile
High is >=75th percentile

0.146
(0.108)
757

0.412
(0.109)
757

0.265
(0.153)

1.729

-0.022
(0.140)
374

0.450
(0.136)
379

0.472
(0.195)

2.414

SER Results
Father Earnings
Log Avg. 79-85
Low is <=median
High is >median
Father Earnings
Log Avg. 79-85
Low is <=25th percentile
High is >=75th percentile

0.482
(0.072)
2186

0.286
(0.114)
1111

0.467
(0.097)
1075

0.181
(0.149)

1.212

0.193
(0.124)
559

0.471
(0.143)
532

0.278
(0.189)

1.467

Note: For the dependent variable, probit models based on the 1996 SIPP matched to SER were
used to determine if zero earnings reflected noncoverage or non-worker status and were imputed
accordingly. Fathers must have positive earnings in each year. When fathers' earnings are from
the 1984 SIPP, they must be present for all interview months and have no cases of nonresponse
to earnings or income questions. Only those fathers succesfully matched to their wave 4
questionnaire are kept in the sample. Standard errors are adjusted for within family correlation
when more than one sibling is present.

Table 9: Intergenerational Elasticity by Race

elast.
(s.e.)
N

Dependent Variable
is Log Avg. Earnings, 1995-1998
All results use tobit specification
Pooled (Sons and Daughters)
Overall

Father Earnings
Log Avg. 79-85

0.328
(0.046)
3077

White

Black

Diff.

0.271
(0.048)
2726

SER Results
0.487
0.216
(0.136)
(0.144)
255

t-stat

1.500

SIPP Results
Father Earnings
Log Avg. 84-85

0.357
(0.074)
1515

0.312
(0.074)
1362

0.534
(0.359)
108

0.222
(0.367)

0.605

Father Income
Log Avg. 84-85

0.364
(0.084)
1690

0.254
(0.082)
1498

0.620
(0.277)
134

0.366
(0.289)

1.265

Two Parent Earnings
Log Avg. 84-85

0.464
(0.082)
1573

0.343
(0.066)
1409

1.013
(0.339)
117

0.670
(0.345)

1.942

Two Parent Income
Log Avg. 84-85

0.518
(0.109)
1674

0.381
(0.100)
1488

1.109
(0.413)
130

0.728
(0.425)

1.713

Note: For the dependent variable, probit models based on the 1996 SIPP matched to SER
were used to determine if zero earnings reflected noncoverage or non-worker status and were
imputed accordingly. For SER results, fathers must have positive earnings in at least one year
and fathers who are classified as non-covered are imputed. When fathers' earnings are from
the 1984 SIPP, they must be present for all interview months and have no cases of
nonresponse to earnings or income questions. They must also have positive earnings in each
year. Standard errors are adjusted for within family correlation when more than one sibling is
present.

19
70
19
71
19
72
19
73
19
74
19
75
19
76
19
77
19
78
19
79
19
80
19
81
19
82
19
83
19
84
19
85
19
86
19
87
19
88
19
89
19
90
19
91
19
92
19
93
19
94
19
95
19
96
19
97
19
98

Percent

Figure 1: Percent of Sample Topcoded

70

60
Fathers

50

40

30

20

10
Sons

0

Year

Figure 2: 1984 SIPP vs. SER Comparison

lnearn84

lnd84

Log 1984 SIPP Earnings

15

10

5

45 degree line

0
0

5
lnd84
Log 1984 SER Earnings

10

Figure 3: Simulation and Actual Estimates from Averaging Fathers’ Earnings
0.7

0.6

Elasticity

0.5

0.4

0.3

0.2

0.1

0
1

2

3

4

5

6

7

8

9

10

No. of Years Averaged
Simulation

Actual Estimate

11

12

13

14

15

16

Appendix

Description of simulation to calculate attenutation coefficients shown in Table 1
In order to account for the fact that transitory shocks that persist over the course of an
individual’s working life are effectively “permanent”, a formula analogous to (7) is derived for
calculating attenuation coefficients when using multi-period averages of fathers’ earnings. As
before, the earnings process for fathers and sons is assumed to follow equations (2) and (3). It is
now assumed, however, that instead of (4) the equation of interest is:

y1i = ρy 45,i + ε

(12)

where y 45, i is the average of fathers’ earnings over the 45 years of his working life. Let t index
the years from 1 to 45. If a T year average of earnings beginning in year s, yT , s , is used as a
proxy for y 45, i , the attenuation factor λ* T,s is the following:
λ* T,s =

(13)

cov( y T ,s , y 45 )
var( yT , s )

While the denominator of this expression will be exactly the same as in (7), the
numerator is more complicated. The covariance between any multiyear average of fathers’
earnings and the entire lifetime average of earnings will depend not only the number of years that
are averaged but also on exactly which years are used in the average.1 Equation (14) provides the
exact formula for calculating λ* T,s .

σ 2 y0 +
(14)

λ* T,s =

σ y20

1

s + T −1

1 45 t− s 2
∑δ σ w
45 t =1
s
1
1
+ ασ w2 + σ v2
T
T

1
T

∑

The autoregressive structure implies that the correlations between the transitory components will depend
on the distance in time between the years used for the short-term average and the full 45 year average. For
example, a five-year average taken at the very beginning or end of one’s life will be less correlated with
lifetime earnings than a five-year average taken during the middle of one’s life.

(

)


 1− δ T  
T − 
(1 − δ )  


where, as before, α = 1 + 2δ 

 T (1 − δ ) 




Dividing the numerator and denominator by the variance in single year earnings, σ2 yt and then
using estimates for δ and the share of the variance of single year earnings accounted for by the
permanent component, transitory component and measurement error, enables one to calculate
λ* T,s for all possible values of T and s. In order to get a summary measure of the degree of
attenuation bias that is only a function of T, we can simply average the λ* T,s over all possible s for
a given value of T. Table 1 presents the values of λ* T as averages are taken over progressively
more years using three different sets of assumptions on the parameter values.

Procedure for assigning covered status among children
The 1996 SIPP-SER was utilized to classify those born in 1963-1968 with zero earnings
in 1996 or 1997, as either non-workers or non-covered. In this sub-sample, about 57 percent of
the men with zero SER earnings were employed for the full year and are classified as noncovered while 32 percent worked for only zero to two months of the year and are called nonworkers.2 Those working in the non-covered sector are primarily government workers or selfemployed. The comparable rates for the daughters were 21 percent and 71 percent, respectively.
These numbers suggest two conclusions. First, most of those with zero SER earnings are either
non-workers or full-time workers in the non-covered sector. Only about 10 percent of zero
earners fall in the gray area of having zero earnings and working part-year. Second, the problem
of non-covered workers is particularly important for men.

2

The universe is restricted to those who remained in the 1996 SIPP through the end of 1996. Those who
are considered employed for the whole year may have worked for 10 to 12 months. Because of the rotation
group structure of the SIPP some individuals may have only joined the survey starting in February or
March of 1996.

Because of the clear dichotomy among those with zero covered earnings, probit models
were used to predict the probability that individuals with zero covered earnings will have actually
worked a full year as a function of all available information contained in the 1996 SER file as
well as any basic demographic information that can be determined by adolescence.3 This
function is then applied to the sample of sons and daughters from the 1984 SIPP-SER to obtain
predicted probabilities for each individual that they were non-covered. A second set of probit
models were also estimated to predict the likelihood that someone with zero covered earnings
worked no more than two months of the year. The estimated function is then applied to the sons
and daughters to obtain a second set of predicted probabilities. Each of these probit models were
estimated separately for men and women, and for both 1996 and 1997. 4 The estimates from the
probit models were then combined in order to classify each son or daughter as either a nonworker or as non-covered for each year.5 Those identified as non-covered are then either dropped
from the analysis or their earnings are imputed using the mean level of log earnings for the group
from the 1996 SIPP. Similarly, those identified as non-employed may be assigned the mean level
of log earnings for those who worked between zero to two months.6
The results of the two probit models for men in 1996 are shown in Appendix Table A1. 7
Among the key variables that are significant are: having attended college; the number of years of
zero earnings during the late 1990s; total lifetime covered earnings; annual earnings in specific

3

For the most part, survey information from the 1996 SIPP is deliberately omitted from this analysis since
such information is obviously unavailable for the sample of sons and daughters from the 1984 SIPP. The
exceptions are some basic demographic information and whether individuals ever attended a college. For
the sons and daughters, data on whether they ever attended a college over the period of the 1984 SIPP (June
1983- June 1986) can then be exploited.
4
These were the only years from the 1996 SIPP for which annual earnings were available at the time the
research was permitted.
5
Specifically, individuals are classified based on the category in which they have a higher predicted
probability. This is equivalent to assigning them based on the sign of the difference in predicted
probabilities. The results for 1996 are used to classify those with zero covered earnings in 1995 and,
similarly the 1997 results are applied to those with zero covered earnings in 1998.
6
This strategy allows those children with zero SER earnings in all four years not to be entirely dropped
from the analysis
7
Results for women and for 1997 are available on request.

years; a flag indicating an active earnings discrepancy;8 being 29 years old; never having positive
covered earnings; being Mexican and being self-employed interacted with 1995 earnings. The fit
of these models is quite high as measured by the Pseudo R2 . The within sample forecasting
record is also very impressive. For men in 1996, over 90 percent of the true classifications of
non-covered and non-workers were correctly predicted. In terms of the entire sample of the
cohort of men in the 1996 SIPP-SER, this implies that less than 1 percent of the sample was
incorrectly classified. The error in forecasting women's status is higher and implies that about 3
percent of the sample is incorrectly classified. While it is impossible to know how well this
model predicts the correct classification of earnings for the sons and daughters in the 1984 SIPP,
the low forecast errors in the 1996 SIPP sample suggest that we can have a high level of
confidence in the results.
Procedure for handling topcoding among fathers
Fathers with topcoded earnings are divided into six race-education cells: by white or
black and by those with less than 16 years of schooling, exactly 16 years of schooling and more
than 16 years of schooling. For each year from 1981 to 1985 the full sample of the 1984 SIPPSER dataset is used to create imputed values for each group. Specifically, the mean value of
SIPP earnings in 1984 for each topcoded group is calculated and used for imputation. 9
For the years 1970 to 1980, the imputation values are derived from each year's March
Current Population Survey (CPS) instead of the 1984 SIPP. Given the well-documented change
in the earnings distribution from the 1970s to the 1980s, it is clearly inappropriate to use the 1984

8

These are cases where an individual has contested what they believe to be inaccurate reports of their
earnings with SSA and where the dispute has not yet been resolved.
9
Only topcoded individuals for whom SIPP earnings in 1984 is greater than or equal to 1984 SER earnings
and who are in the SIPP for all 12 months of 1984 are used in the calculations. For the years 1981 to 1983,
and 1985, calculating the imputations involves an added step. The percentile to use as a cutoff for
calculating the imputed values for each year is determined by using the percent topcoded in that year based
on the SER data for all the sample members in the 1984 SIPP-SER dataset (not just the fathers). For
example, in 1980, 8.8 percent of those with positive earnings in the full sample of the 1984 SIPP-SER
matched dataset, had topcoded earnings. The strategy then, was to use the top 8.8 percent of the SIPP
earnings distribution in 1984 to calculate the imputed values for each of the 6 groups for 1980. Of course,
the 1984 dollar values were then converted to 1980 dollars using the CPI.

earnings distribution to calculate the imputed values during the 1970s. For these years, the actual
taxable maximum published by the Social Security Administration is used as a cutoff point for
the CPS analysis. The mean value of earnings above the taxable maximum for each group is used
to impute earnings for those who were topcoded during these years.10

10

An attempt was also made to use information in the SER data file on the quarter of the year in which full
coverage was achieved. For years before 1978 this variable could be used to estimate full year earnings for
those topcoded. The results, however, were no different using this strategy.

Table A1: Probit Results on Predicting Non-Covered vs. Non-Worker, 1996 SIPP-SER
Men in 1996

Dependent Variable
Non-Covered

Variable
black*
college*
Years of 0 Earn 81-90
Years of 0 Earn 91-94
Years of 0 Earn 95-98
self-employed*
agricultural*
total quarters of coverage
earnings 1981
earnings 1982
earnings 1983
earnings 1984
earnings 1985
earnings 1986
earnings 1987
earnings 1988
earnings 1989
earnings 1990
earnings 1991
earnings 1992
earnings 1993
earnings 1994
earnings 1995
earnings 1997
earnings 1998
earnings discrepancy flag*
military*
age29*
age30*
age31*
age32*
age33*
first year of earnings
last year of earnings
total earnings to date
quarters of coverage 1990
quarters of coverage 1991
quarters of coverage 1992
quarters of coverage 1993
quarters of coverage 1994
quarters of coverage 1995
quarters of coverage 1997
quarters of coverage 1998

dF/dx
times 100
6.72
21.98
-10.30
2.77
48.64
-79.99
15.51
-2.78
-0.03
-0.02
-0.02
-0.01
-0.01
-0.02
-0.02
-0.02
-0.01
-0.01
-0.02
-0.02
-0.02
-0.02
-0.01
-0.02
-0.02
-90.26
6.79
29.15
0.28
15.58
5.09
-28.45
-1.13
7.58
0.02
0.92
-2.25
9.97
-3.01
-2.63
4.68
7.94
6.86

Non-Worker

z-stat
0.47
2.35
-2.13
0.2
2.19
-1.59
0.92
-1.77
-2.19
-2.14
-2.38
-1.68
-1.78
-2.12
-2.62
-2.36
-1.9
-1.97
-2.76
-2.2
-2.11
-2.24
-1.31
-2.35
-2.32
-5.97
0.38
2.42
0.02
1.05
0.3
-1.22
-0.63
3.1
2.35
0.2
-0.37
1.43
-0.55
-0.41
0.6
0.94
1.11

dF/dx
times 100
0.35
-10.73
1.97
-4.12
-10.12
33.34
-8.15
0.95
0.02
0.00
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
59.04
0.94
-8.75
-7.82
-7.49
6.69
15.28
-0.78
-2.92
-0.01
-3.11
-0.41
-6.86
1.58
-0.37
2.72
0.12
-0.35

z-stat
0.05
-1.91
0.75
-0.55
-1.21
0.87
-0.92
1.15
2.31
0.81
2.39
1.68
1.94
2.2
2.71
2.15
1.92
2.33
2.54
2.45
2.05
2.35
1.72
2.15
2.46
6.49
0.09
-1.19
-1.06
-0.95
0.67
1.22
-0.68
-2.52
-2.38
-1.14
-0.13
-1.87
0.52
-0.11
0.67
0.03
-0.1

Mean
0.13
0.35
5.22
2.17
3.16
0.22
0.14
24.83
556.23
749.30
1254.62
1703.05
2318.21
3208.63
3981.88
4162.30
5328.86
6030.20
5964.49
5756.44
5153.96
3450.91
2099.37
2532.74
5006.78
0.29
0.07
0.15
0.20
0.16
0.19
0.16
1722.25
1732.53
60245.30
2.02
2.02
1.81
1.48
1.23
0.72
0.93
1.19

Table A1: Probit Results on Predicting Non-Covered vs. Non-Worker, 1996 SIPP-SER (cont.)
Men in 1996

Dependent Variable
Non-Covered

Variable
never covered earnings
newly posted credit earn*
mexican*
mexican american*
hispanic*
earnings 1998 X self-emp.
0 earnings 1995 X self-emp.*
0 earnings 1997 X self-emp.*
0 earnings 1998 X self-emp.*
0 earnings 1995 X agr.*
# of 0's 95-98 X 1995 earn.

dF/dx
times 100
100.00
-14.11
25.88
-20.86
-7.56
0.01
41.91
-36.59
28.55
-25.60
-0.01

Non-Worker

z-stat
2.46
-0.55
2.86
-0.71
-0.31
2.46
1.84
-0.98
1.71
-0.84
-3.27

dF/dx
times 100
-100.00
16.10
-11.44
3.44
-9.18
0.00
-18.20
13.84
-3.45
31.93
0.00

z-stat
-2.49
1
-1.63
0.25
-1.21
-1.79
-1.31
0.57
-0.2
1.32
2.23

Mean
0.13
0.06
0.07
0.05
0.05
981.06
0.16
0.14
0.11
0.10
3760.91

Observations
258
258
Pseudo R squared
0.60
0.53
* indicates a dummy variable, dF/dx shows the effect of a discrete change in the variable from 0 to 1.
Note: Sample is from 1996 SIPP matched to SER for cohort born in 1963 to 1968 with zero SER
earnings. Sample is restricted to those who are interviewed for at least ten months of 1996 SIPP.
"Not Covered" have zero SER earnings but at least 10 paid months of work. Unemployed have zero
SER earnings and between 0 to 2 months of paid work.

Table A2: The Effects of Varying the Exclusion Rule on Years of Fathers' Zero Earnings
elast.
(s.e.)
N

Dependent Variable
Son's Log Avg Earnings, 1995-1998

Require the following number of years of positive earnings…
0

Fathers'
Log Avg. Earnings
1979-1985

1

2

3

4

5

6

7

0.160
0.290
0.336
0.434
0.468
0.462
0.440
0.445
(0.034) (0.062) (0.058) (0.069) (0.074) (0.076) (0.075) (0.079)
1299
1256
1245
1227
1212
1201
1181
1160

Note: For the dependent variable, probit models based on the 1996 SIPP matched to SER were
used to determine if zero earnings reflected noncovered status and if so were imputed. For
fathers, earnings for those identified as non-covered are dropped. Earnings for those topcoded
are imputed using March CPS data for 1970-80 and using 1984 SIPP for 1981 to 1984.