THE ORIGINS OF INTERGENERATIONAL
ASSOCIATIONS: LESSONS FROM
SWEDISH ADOPTION DATA*
ANDERS BJO¨RKLUND
MIKAEL LINDAHL
ERIK PLUG
We use unique Swedish data with information on adopted children’s biological
and adoptive parents to estimate intergenerational mobility associations in
earnings and education. We argue that the impact from biological parents captures
broad prebirth factors, including genes and prenatal environment, and the
impact from adoptive parents represents broad postbirth factors, such as childhood
environment. We find that both pre- and postbirth factors contribute to
intergenerational earnings and education transmissions, and that prebirth factors
are more important for mother’s education and less important for father’s
income. We also find some evidence for a positive interaction effect between
postbirth environment and prebirth factors.
I. INTRODUCTION
A large body of evidence suggests substantial intergenerational
associations in earnings and education. Recent research
has documented some cross-national differences with weaker intergenerational
associations in Scandinavian countries and
stronger in the United States.1
The standard approach in this
literature relates children’s outcomes to those of their rearing
parents. But this approach does not allow one to sort out the
underlying contributions of prebirth factors (genetics and prenatal
environment) from postbirth environmental factors. Understanding
the origins of intergenerational persistence in economic
status is necessary for developing relevant theories and designing
social policies.
Adoption creates a potential “natural experiment” that generates
some (partially) independent variation in prebirth and
* We are very grateful for most constructive and detailed comments by
Lawrence Katz. We also thank Edward Glaeser, two anonymous referees, Helena
Holmlund, Sandy Jencks, Matthew Lindquist, and numerous seminar and conference
participants for useful suggestions. Financial support from the Swedish
Council for Working Life and Social Research (FAS) and the Dutch National
Science Foundation (NWO-VIDI 452.03.309) is gratefully acknowledged.
1. See Solon [1999] and a symposium in the Journal of Economic Perspectives
(Summer 2002) for surveys.
© 2006 by the President and Fellows of Harvard College and the Massachusetts Institute of
Technology.
The Quarterly Journal of Economics, August 2006
999
postbirth factors. If adopted children are (conditionally) randomly
assigned to adoptive parents and adoption takes place close to
birth, then regressions of child outcomes on adoptive parents’
characteristics and biological parents’ characteristics can provide
information on the relative importance of postbirth and prebirth
factors. Existing studies of adopted children find lower intergenerational
associations for adoptive families than for biological
families suggesting some role for prebirth factors; recent studies
include Sacerdote [2000, 2002, 2004], Plug and Vijverberg [2003,
2005], Plug [2004], and Bjo¨rklund, Lindahl, and Plug [2004].2
But
lack of data on the characteristics of the biological parents of
adoptees has limited the inferences that can be drawn from such
samples.
This paper uses a unique administrative data set on a large
sample of adoptees born in Sweden (from 1962–1966) that contains
information on the children’s educational and economic
outcomes and on the characteristics of both their adoptive and
biological parents. The intergenerational associations for this
sample of adoptees are compared with a representative sample of
all (nonadopted) children born in Sweden in this period and to
subsamples with similar characteristics of biological parents and
similar postbirth environments as the sample of adoptees.
The basic finding is that both pre- and postbirth factors
contribute to intergenerational earnings and education transmissions,
and that prebirth factors are more important for mother’s
education and less important for father’s income. We also find
some evidence for a positive interaction effect between postbirth
environment and prebirth factors.
The paper is organized as follows: Section II introduces the
extended intergenerational models that capture transmissions
from both biological and adoptive parents. Section III briefly
discusses the institutional setting of adoptions in Sweden in
1962–1966 and the administrative data used in the analysis.
Section IV reports our basic results. Sections V and VI present a
number of robustness tests and results using nonlinear models.
Section VII discusses our main findings and concludes.
2. See also Scarr and Weinberg [1978] and Das and Sjogren [2002] for work
along these lines.
1000 QUARTERLY JOURNAL OF ECONOMICS
II. INTERGENERATIONAL TRANSMISSION MODELS
II.A. Linear Models
In recent intergenerational research the prototypical model
used by economists can be expressed as
(1) Yj
bc
ϭ ␤0 ϩ ␤1Yj
bp
ϩ ␯j
bc
,
where Y denotes the logarithm of long-run income or a measure
of educational attainment at adult age, subscript j indexes the
family in which the child is born and raised, superscripts bc and
bp denote the biological child and parent,3
and ␯j
bc
is a childspecific
characteristic uncorrelated with Yj
bp
. The estimated intergenerational
coefficient ␤1 measures the strength of the intergenerational
association and represents the combined effect of
many different mechanisms, including the effects that come from
genetic inheritance, prenatal environment, and the environment
in which the child grew up.
Our data on adoptees and their biological and adoptive parents
allow us to decompose the estimated intergenerational coefficient
␤1 into two components—one component that measures
the contribution of genetics and prenatal environment (which we
label prebirth factors) and the other one that measures the contribution
of postbirth environment on the outcomes of children.
We model the transmission of an outcome from the biological
parent bp and the adoptive parent ap to the adopted child ac born
in family j and adopted and reared in family i as
(2) Yi
ac
ϭ ␣0 ϩ ␣1Yj
bp
ϩ ␣2Yi
ap
ϩ ␯i
ac
,
where ␯i
ac
represents unobserved child-specific characteristics uncorrelated
with Yj
bp
and Yi
ap
. It is important to note that a
parental Y does not only capture this parental characteristic but
also everything else that is correlated with it. This means that we
estimate parameters that capture broad prebirth factors (Yj
bp
)
and broad postbirth environment (Yi
ap
), and that we do not consider
␣1 and ␣2 as the causal impact of the parents’ Y. Prebirth
factors include genetic and prenatal environmental factors. Postbirth
factors include, for example, the impact of Y on the quality
3. Most studies—e.g., those that rely on U.S. PSID and NLSY data—have
used rearing parents, including biological, step, and adoptive parents. We use the
notation bp for expositional convenience.
1001THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
and quantity of time, goods, and money that parents devote to
their children.
We can further separate prebirth factors into impacts of
genetic and prenatal environmental factors by estimating the
effects of the biological mother and father in separate regressions.
Assuming that genetic transfers that come from the biological
mother and father are equally important, and that the father’s
behavior does not affect the child’s prenatal conditions in the
womb, we can interpret the impact of the father’s characteristics
as an estimate of genetic factors and the difference in impacts
between the biological mother’s and father’s characteristics as an
estimate of the impact of prenatal environmental factors.
The procedure to link the prototypical mobility model in (1) to
the adoption model in (2) is relatively simple. For children who
are not adopted but born and raised in the same family j, we
know that Yi
ap
and Yj
bp
are identical. If we replace Yi
ap
with Yj
bp
,
model (2) collapses to model (1) where the intergenerational
transmission coefficients ␤1 and ␣1 ϩ ␣2 are identical. Two considerations
are of interest here. The first one relates to whether or
not we can use adoptees to separate pre- from postbirth contributions.
The second one relates to whether or not we can infer the
relative contributions of pre- and postbirth factors for a representative
Swedish child from a sample of adoptees. These considerations
are explored below.
II.B. Identification Issues
Let us begin with the identification of pre- and postbirth
contributions using adoptees. To estimate and interpret ␣1 and ␣2
as pre- and postbirth contributions, we need to make two assumptions.
First, adoptees are randomly assigned to adoptive families,
or that the assignment is related to variables that we—as researchers—observe
and control for. And second, children move to
their adoptive parents immediately at birth. Both assumptions
are easily violated.
For one, the assignment process is not always random. In
fact, if adoption authorities have information on the children’s
biological parents, they may use it to match children to adoptive
families. If matching occurs and the characteristics of the adoptees’
birth and adoptive parents become correlated, adoption estimates
will be biased. This is a serious problem, one that is shared
with almost all adoption studies. If nonrandom assignment is
present, it is possible that part of the estimated prebirth effects
1002 QUARTERLY JOURNAL OF ECONOMICS
are picking up the effect of those postbirth factors that are either
unobserved and correlated with Yj
bp
, or observed and measured
with error. And vice versa for the estimated postbirth effects. In
Section V we formalize nonrandom assignment within an omitted
variables and measurement error framework and examine
whether nonrandom assignment itself is empirically important.4
Also, not all adoptees are adopted as babies. It is possible
that some adoptees spend time with their birth parents, or receive
institutional care between birth and adoption. In such
cases, the estimated prebirth effects are too high, and estimated
postbirth effects too low. We assess the sensitivity of our results
using a sample that is arguably more limited to children adopted
as babies.
To compare ␤1 and ␣1 ϩ ␣2, we need to assume that adopted
and own-birth children5
and their parents are randomly drawn
from one particular distribution of children and parents, and that
parents do not differentiate between own-birth and adopted chil-
dren.6
Again, these assumptions may prove to be difficult to
maintain. For the adoptees’ parents, for example, Yi
ap
and Yj
bp
are
distinctively different. Then how do the estimates for own-birth
children compare with those obtained for adoptees? Investigating
this is easier said than done. Whereas the estimates for own-birth
children come from a representative sample, the adoption results
come from a sample of children with, on average, disadvantaged
prebirth but favorable postbirth backgrounds. These opposite
sources of selection make it difficult—if not impossible—to come
up with a sample of own-birth children that is comparable to our
sample of adoptees. We deal with this by comparing own-birth
children with the adoptees in two different ways. First, we com-
4. Adoption authorities collect information on a wide array of factors. An
omitted variables approach applies to those factors that are exclusively observed
by adoption authorities (interviews and assessments of likely parenting skills and
seriousness). A measurement error approach applies to those factors that are
observed by the researcher, and possibly the authority, but that are imperfectly
measured in our data (long-run economic status).
5. We use the concept of own-birth children whenever children are raised by
their biological parents.
6. Previous adoption studies (such as Plug [2004] and Sacerdote [2004])
lacked data on biological parents of adoptees, and could therefore only estimate
the impact of postbirth factors. They could only indirectly infer the impact of
prebirth factors by taking the difference between the intergenerational estimate
for own-birth children and the estimate of the impact of postbirth factors. This
means that they had to assume that ␣1 ϭ ␤1 Ϫ ␣2. However, since adoptees are
different compared with own-birth children on a number of dimensions, it is not
a priori clear that this assumption holds. In this paper we estimate both ␣1 and ␣2
directly.
1003THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
pare adoptees with own-birth children who start their lives under
very similar conditions as adoptees. Second, we compare adoptees
with another sample of own-birth children who are reared in
environments similar to those of the adoptees. Different samples
are made comparable through subsampling and propensity score
matching.
Another issue is that children, who are given up for adoption,
may be different from other children because of the adoption
itself. If, for example, indications that adopted children reveal
more emotional problems than their classmates—see Bohman
[1970] for some Swedish evidence—reflect causal effects of adoption,
outcomes like educational attainment and earnings might
also be affected. In the empirical analysis, however, we allow for
adoptees and own-birth children to be different by running mobility
models on separate samples. As long as these differences
are unrelated to the parental schooling and earnings characteristics,
any real adoption effect is captured by the difference between
intercepts ␣0 and ␤0.
Note that if results are different, the implication is not necessarily
that adoption itself has real effects. Functional form
assumptions may be responsible as well. In particular, we focus
our attention on genes-environment interactions.
II.C. Nonlinear Models
Interaction between genes and environment is an issue that
has been given much attention in the recent literature.7
Empirical
tests, however, are only indirect, and there is hardly any
consensus about the presence of such interactions.8
Our data set
7. For discussions, see Cunha et al. [2005] and Ridley [2003]. Turkheimer et
al. [2003], an often cited study, use a little over 100 monozygotic and 200 dizygotic
twin pairs and variance decomposition techniques to estimate how much genes
and environment can explain the variation in IQ among seven-year-old children,
at different levels of socioeconomic background for their parents. They find genetic
(shared environmental) effects to be more (less) important the higher the childrens’
socioeconomic backgrounds.
8. Plomin, DeFries, and Fuller [1988] survey genes-environment interaction
studies in behavioral genetics. They find few statistically significant interactions
and state that interactions “are easily posed but rarely documented.” Bjo¨rklund,
Ja¨ntti, and Solon [2005] use sibling correlations for nine sibling types to decompose
earnings variation into genetic and environmental components. When they
extend a conventional model with a parameter that reflects interaction between
genes and environment, they get a negative but insignificant estimate. Further,
their model with an interaction effect is rejected by the data. Note also, while they
also are using Swedish data, their approach is different from ours. First, they use
sibling data, and we use intergenerational data. Second, they apply the variance
decomposition approach, whereas we estimate regression coefficients telling what
1004 QUARTERLY JOURNAL OF ECONOMICS
is especially suitable to test for such interactions. We have information
on the adoptive and biological parents for a very large
number of adoptees making it possible to estimate interaction
terms with reasonable precision. We can modify model (2) to
account for interaction effects of pre- and postbirth factors to
(3) Yi
ac
ϭ ␣0 ϩ ␣1Yj
bp
ϩ ␣2Yi
ap
ϩ ␣3Yj
bp
Yi
ap
ϩ ui
ac
.
The interaction coefficient ␣3 is positive if children with beneficial
prebirth background benefit relatively more from a good
postbirth environment, which would indicate that genetic and
environmental factors are complements in the production of life
success for the child. The corresponding model for parent and
child in a nonadoption family, where Yi
ap
and Yj
bp
are identical, is
written as
(4) Yj
bc
ϭ ␤0 ϩ ␤1Yj
bp
ϩ ␤2͑Yj
bp
͒2
ϩ uj
bc
.
We do not want to infer interactions from a positive ␤2 since
functional-form assumptions dictate that the interaction in model
(3) automatically leads to a nonlinear mobility model in (4). If, for
example, quadratic effects are present but not modeled in (3), ␤2
would capture both interaction and quadratic effects. Thus, we
also test for interaction effects in models where quadratic main
effects are added to (3). Note also that if the interaction in (3)
would look like ͌Yi
ap
⅐ ͌Yj
bp, we return to the original linear
model in (1).
III. INSTITUTIONS AND DATA
9
We use administrative register data from Statistics Sweden
on all legal adoptions, i.e., adoptions decided by public court and
notified in the Swedish population register. In particular, our
data set contains all persons who were born in Sweden between
1962 and 1966 and adopted by both parents. To show the usefulness
of this data set, we start this section by describing the
Swedish adoption institutions during this period of time. Then we
describe the data set and the variables in some detail.
a unit of parental education implies for offspring’s education and how a log point
of paternal income (earnings) is related to offspring’s log income (earnings). Third,
they do not analyze education.
9. The main sources for this section have been handbooks for social workers
dealing with adoptions, the last one being Kungl. Socialstyrelsen [1968].
1005THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
III.A. Adoptions in Sweden 1962–1966
The basic principle of Swedish adoption law has always been
that an adoption should be “in the best interests of the child.”
This means, for example, that the decision whether an adoption
should take place and the choice of adoptive parents should be
motivated by concern for the child and not for the couple who
wants to adopt. Economic compensation between the adoptive
and the biological parents was not allowed. Because the period we
consider was characterized as one with “excess demand” from
prospective adoptive parents, payments to biological parents
would probably have existed if allowed. An adopted child received
the same legal status, e.g., with respect to inheritance, as a
biological child. Further, all formal connections with the biological
parents were broken.
A social authority within the local government was responsible
for the process. Thus, mothers (and fathers) who wanted to
give a child up for adoption as well as families who wanted to
adopt contacted this authority. The legal adoption decision was
taken by public court after being advised by the social authority.
The vast majority of adoptions took place at an early age of
the child. In a typical case, an unmarried pregnant woman considered
adoption and therefore contacted the social authority. But
the mother could not decide to give up a child until she had
recovered from the delivery. The newborn child was therefore
initially placed in a special nursery home. An unmarried biological
father had no formal say in the adoption decision, but ought to
be contacted on the issue and allowed to give his opinion. Quite
often, however, the father was unknown. The population register
we use identifies 92 percent of the biological mothers and 58
percent of the biological fathers.
The child was first placed in a prospective adoptive family on
a trial basis as a foster child. Placements were recommended to
occur before the age of six months. If the trial period—lasting
some 3–6 months—turned out well, the next step for the prospective
adoptive parents was to apply to the court for a legal adoption
decision. The formal process by the court could take several
months so the adopted child could have spent quite a long time
with its new parents when the adoption was finally formalized.
In general, the biological mother was young, unmarried, and
poor. Table I (see below), reporting descriptive statistics for all
cases when both biological parents of adopted children in Sweden
1006 QUARTERLY JOURNAL OF ECONOMICS
born 1962–1966 were known, shows that both biological mothers
and fathers were younger than nonadopted children’s parents.
Nonetheless, there is a substantial age variation among biological
parents, and only 32 percent of the mothers and 12 percent of the
fathers were teenagers when the child was born.
Low income was a common reason to give a child up for
adoption, and lower social classes were clearly overrepresented
among biological mothers. It is notable, though, that socioeconomic
status before age 30 does not strongly correlate with longrun
status. Further, “shame” was also a common reason for
adoption. Thus, although we have reason to believe that most
mothers were of low socioeconomic status also in a long-run
perspective, they were not necessarily a very homogeneous group
in this respect.
The responsible social worker, assigned by the social authority,
was required to undertake a careful investigation of prospective
adoptive parents. The guidelines for adoptions emphasized
that, given a reasonably “stable” situation, economic resources
and social status were not most important. Although not a formal
requirement, it was expected that the mother could stay home to
care for the child. The guidelines said that good adoptive parents
should be tolerant, since an adopted child could encounter problems
and maybe not meet the expectations of its parents. “Normal
people” were considered the best adoptive parents. Nonetheless,
due to these considerations, one would expect adoptive parents to
be underrepresented among families with low socioeconomic
status.
The law required that adoptive parents were 25 years of age.
There was no upper age limit, but the adoptive parents were
supposed to be young enough to be able to be the biological
parents of the child. There was no requirement about the duration
of the marriage. But other requirements made it unlikely
that newly married couples would be able to adopt a child. For
example, the social worker was supposed to find out that the
prospective adoptive parents would not be able to have their own
biological children.
A crucial issue in adoption research is whether there likely is
selective placement so that there is a positive correlation in
important traits between biological and adoptive parents. Our
data—see Section V below—confirm such a positive correlation in
education and income. Bohman [1970], who studied adoptions in
1007THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
Sweden, showed that the actual behavior of social workers generated
such patterns.
Although the adoption case described above was by far the
most common one, there were also other cases. One case was
when a foster child had “grown into” the foster family so that the
foster parents wanted to adopt the child. Stepparent adoptions
were yet another kind of adoption, namely such as when the
spouse of one biological parent adopted the latter’s child. Our
data allow us to identify such cases; we found six cases and
eliminated them. Adoptions could also take place within families,
e.g., the parents of a young mother could adopt a child who would
have been their biological grandchild. Such cases would create
severe problems for our study. However, they were very rare
during our study period; see Nordlo¨f [2001].
III.B. The Data Set and Descriptive Statistics
Because of the adoption process described above, the Swedish
population register contains information about both adoptive and
biological parents. From the population register, we also get
access to the adopted children’s siblings, both on their biological
parents’ side (our data distinguish between full and half siblings)
and on their adoptive parents’ side. On the latter side, there could
be both siblings who also are adopted and siblings who are biological
children of the adoptive parents.10
Further, we use education and tax registers to get information
about parental characteristics and child outcomes at adult
age. The 1970 census and the 1990 version of Statistics Sweden’s
special education register provide information about biological
and adoptive parents’ education. The 1970 Census, upon which
the education register is built, contains detailed education information,
which is available in terms of detailed education classifications.
We infer years of schooling and a university dummy
from this information.11
We use tax-register data to get fathers’
earnings and income for 1970, 1975, 1980, 1985, and 1990. Earn-
10. Our data tell us in what census the adoptee first lived with its adoptive
parents, although the parents were not necessarily formal adoptive parents at this
point of time due to the trial period and delay with the court’s decision.
11. We use the education register dated 1990 to obtain parental education
measures. If not available, we use the 1970 Census instead. The reason for doing
so is that some parents were quite young in 1970. The 1970 Census, upon which
the education register is built, and the 1990 register contain identical educational
level classifications. We assign the following years of schooling to seven educational
levels: 7 for (old) primary school, 9 for (new) compulsory schooling, 11 for
short high school, 12 for long high school, 14 for short university, 15.5 for long
1008 QUARTERLY JOURNAL OF ECONOMICS
ings include income from work including self-employment and
sickness benefits. Income includes earnings, but also some taxable
benefits like unemployment insurance and pensions as well
as capital income and realized capital gains. The tax registers in
turn are based on compulsory reports from employers, and the
taxable benefits are similarly reported by the authorities responsible
for the schemes. We use father’s income averaged over a
twenty-year period running from 1970 to 1990.12
Because (due to
the adoption guidelines in the early 1960s) adoptive mothers were
expected not to work, we focus only on the intergenerational
mobility effects of father’s earnings and income. Our analysis
thus follows the empirical literature on income mobility that
focuses on the impact of long-run income of fathers.
We measure children’s outcome in 1999 when children are
33–37 years of age. At this age, children should have completed
their school, and their annual income is likely to measure longrun
(or lifetime) income quite well with only a classical measurement
error. This means that our intergenerational estimates will
not suffer from so called life-cycle bias.13
Education information
comes from the education register.14
Children’s income and earnings
data stem from the same administrative registers as the
ones for parents.
The number of adopted children born in Sweden in 1962–
1964 was about 1100 each year, and fell to 1000 in 1965 and 900
in 1966.15
The number of adoptees born in Sweden between
1962–1966 equals 5292. We work with a subsample of 2125
adopted children. The reduction of 3067 observations gives the
impression that nonresponse is serious. Of the 3067 adoptees who
university, and 19 for Ph.D. University education is a dichotomous variable that
indicates whether someone completed fifteen or more years of schooling.
12. More specifically, we first exclude those observations in which annual
income (or earnings) is missing, below 1000 dollars, or obtained when parents
were younger than 30 or older than 60. With annual income and earnings measures
measured in logarithms, we then take averages.
13. Using U. S. data, Haider and Solon [2006] examine how intergenerational
income estimates are affected by the age at which offspring’s income is observed.
They find that the bias is small when annual income is measured around age 35.
Bo¨hlmark and Lindquist [2006], who replicate and extend Haider and Solon on
Swedish data, arrive at the same conclusion.
14. We use the maximum level achieved in the 1990, 1993, 1996, or 1999
education registers. For children and their parents we use the same transformation
from levels to years.
15. With a falling number of children born in Sweden given up for adoption,
the number of foreign-born adoptees started to increase. For example, between
1962 and 1966 the number of international adoptions rose from 100 to 350.
1009THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
fell outside the sample, about 500 observations were lost because
of a set of age restrictions pertaining to the adopted child and
adoptive parents.16
One hundred and thirty-five adoptees had no
records on their own or adoptive parents educational classifications,
and about 570 adoptees were eliminated because information
on their biological mother was missing. The main problem
exists with the adoptees’ biological fathers. Almost 2000 adoptees
have biological fathers who are unknown.17
This leaves us with
2125 adoptees. The income analysis is based on fewer observations
than the education analysis: about 150 (335) adoptees had
annual income (earnings) values that were either missing or
below the Swedish equivalent of 1000 dollars.
Table I reports the means and standard deviations of the
main variables in our analysis. We report these statistics for our
adoptee sample and for a representative sample of all own-birth
children required to have lived with both of their biological parents
in the fall 1970 census; the latter are drawn from a 20
percent random sample of all nonadopted children born in Sweden
in 1962–1966.18
The table’s first panel shows that adopted
children are quite similar to the random sample of same-aged
own-birth children when it comes to the outcome variables. There
are some small differences, however, and they consistently show
that adopted children did slightly worse: years of schooling is
about 0.4 years lower, the fraction with university education is 8
percentage points lower, and earnings and income are 0.09 log
points lower, compared with own-birth children.
The second panel compares the characteristics of own-birth
and adopted children’s birth parents. These differences are
larger. Own-birth children’s fathers have 0.73 more years of
schooling than adopted children’s fathers. The corresponding difference
for mothers is 0.53. University education is also more
frequent among own-birth children’s birth parents than among
adopted children’s birth parents. The earnings and income differentials
are 0.29 log points for fathers. Further, adopted children’s
16. We exclude those who (i) did not live with adoptive parents in the
November 1970 census, (ii) died at age Ͻ26 (45 adoptees), (iii) had too young
adoptive parents (Ͻ25) or a too old adoptive parent (mother Ͼ47 or father Ͼ66).
17. The reduction of almost 2000 observations is substantial and may introduce
sample selection bias. Later on in the paper we test whether the nonresponse
is selective and muddles our results. We find that this is not the case.
18. We start off with 108,550 but work with 94,079 children. We lose almost
14,500 observations because of missing school classifications (more than 5,200)
and because of children who are raised in single-parent families (more than
9,200).
1010 QUARTERLY JOURNAL OF ECONOMICS
birth parents are about 3.5 years younger than own-birth children’s
biological parents. The fraction of teenage parents is also
considerably higher among adopted children.
The third panel reports characteristics of adoptive parents.
Comparing with adopted children’s birth parents, we find quite
TABLE I
MEANS AND STANDARD DEVIATIONS OF VARIABLES FOR CHILDREN AND PARENTS
Own-birth
children
Adopted
children
Children
Years of schooling 12.07 2.07 11.67 1.89
University education 0.32 0.47 0.24 0.43
Log earnings in 1999 7.54 0.67 7.43 0.72
Log income in 1999 7.62 0.56 7.53 0.58
Male 0.51 0.50 0.52 0.50
Age in 1999 35.29 1.42 35.49 1.42
Birth parents
Years of schooling, father 9.63 3.12 8.90 2.51
Years of schooling, mother 9.65 2.83 9.12 2.43
University education, father 0.16 0.36 0.07 0.26
University education, mother 0.16 0.37 0.08 0.28
Average log earnings 1970–1990, father 7.67 0.44 7.38 0.51
Average log income 1970–1990, father 7.69 0.43 7.40 0.46
Age when child is born, father 30.37 6.58 26.88 6.96
Age when child is born, mother 27.09 5.73 23.35 5.80
Teenage mother 0.09 0.29 0.32 0.47
Teenage father 0.02 0.14 0.12 0.33
Adoptive parents
Years of schooling, father 10.20 3.31
Years of schooling, mother 9.67 2.99
University education, father 0.20 0.40
University education, mother 0.18 0.39
Average log earnings 1970–1990, father 7.77 0.47
Average log income 1970–1990, father 7.81 0.44
Age when child is born, father 35.66 5.36
Age when child is born, mother 32.96 4.93
Number of observations 94,079 2,125
Standard deviations are shown in italics. The exceptions to the stated number of observations are for log
earnings in 1999: 87,490 for own-birth children and 1,827 for adopted children, for log income in 1999: 92,168
for own-birth children and 1,998 for adopted children. For average log earnings 1970–1999: 93,627 for birth
fathers of own-birth children, 2,078 for birth fathers of adopted children, and 1,981 for adoptive fathers of
adopted children. For average log income 1970–1999: 93,831 for birth fathers of own-birth children, 2,107 for
birth fathers of adopted children, and 2,120 for adoptive fathers of adopted children.
1011THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
substantial differences, especially for father’s characteristics. Average
years of schooling are 1.30 higher; incidence of university
education is 0.13 higher; earnings and income are around 0.40 log
points higher. The age differentials are as large as 8.8 years for
fathers and 9.6 years for mothers. These age differentials, in turn,
probably account for some of the earnings and income differentials.
In our subsequent analysis, we control for these age
differentials.
These differences, which are all statistically significant, give
the impression that children who are given up for adoption come
from poorer families but are placed in well-to-do families. Despite
mean differences between the parents, the standard deviations
show that there is a considerable overlap between the three
distributions. In some of the regressions using own-birth children
and parents, we use samples of own-birth parents that are
matched so as to mimic the adopted or biological parents. The
matching procedure and the characteristics we match on are
discussed in subsection V.D.
IV. BASIC RESULTS: LINEAR MODELS
Table II reports the intergenerational transmission estimates
for education and income using a linear model. We run
separate regressions on samples of own-birth and adopted children.
In the first panel we report the least square results of model
(1) for education and income on the sample of own-birth children.
All regressions include an intercept and individual controls for
the child’s gender, birth year and region of birth, and the respective
parent’s birth year. These estimates are not reported. The
regressions are typical of estimations in the previous intergenerational
mobility literature.
The estimated effects of parental schooling show—as expected—that
more highly educated parents raise their children’s
schooling years and university graduation chances. We find that
the schooling impacts of both parents are equally important. The
magnitude of these estimates suggests that four more years of
parental schooling—of either the father or mother—are associated
with one more year of schooling for the child, and that a
parent with a university degree is associated with a 34 percent
higher likelihood that his or her children also will obtain such a
degree. Another result, in columns (3) and (6), is that the coefficient
for either parent’s education falls quite substantially when
1012 QUARTERLY JOURNAL OF ECONOMICS
their partner’s education is brought into the equation. Assortative
mating on education lies behind this pattern.
The estimated intergenerational elasticities with respect to
father’s earnings and income are virtually identical and equal to
0.24. They are similar to those obtained by Bjo¨rklund and Ja¨ntti
[1997] and Bjo¨rklund and Chadwick [2003] for Sweden.19
In the second panel we report the results for specification (2)
where our four outcomes for adoptees are run on the same variables
for their adoptive and biological parents using the same
format as before. The first three columns report the estimates for
19. The previous mobility models are also estimated for sons and daughters
separately. The intergenerational transmissions for schooling are found to be very
similar. The intergenerational transmissions for earnings and income are, however,
somewhat larger for sons. A similar pattern is observed when we estimate
these models on our samples of female and male adoptees.
TABLE II
ESTIMATED TRANSMISSION COEFFICIENTS IN LINEAR MODELS
Years of schooling University Earnings Income
(1) (2) (3) (4) (5) (6) (7) (8)
Own-birth children
Bio father .240** .170** .339** .237** .235** .241**
(.002) (.002) (.004) (.004) (.005) (.004)
Bio mother .243** .158** .337** .246**
(.002) (.002) (.004) (.004)
Adopted children
Bio father .113** .094** .184** .148** .047 .059*
(.016) (.016) (.036) (.036) (.034) (.028)
Bio mother .132** .101** .261** .229**
(.017) (.017) (.034) (.034)
Adoptive father .114** .094** .165** .102** .098** .172**
(.013) (.014) (.024) (.026) (.038) (.031)
Adoptive mother .074** .021 .145** .097**
(.014) (.015) (.024) (.026)
Sum of estimates .227** .188** .349** .249** .145** .231**
for bio and (.019) (.029) (.040) (.059) (.049) (.040)
adoptive fathers
Sum of estimates .207** .122** .406** .326**
for bio and (.021) (.016) (.039) (.029)
adoptive mothers
Standard errors are shown in parentheses; * indicates significance at 5 percent level, and ** at 1 percent
level. All specifications include controls for the child’s gender, 4 birth cohort dummies for the child, 8 birth
cohort dummies for biological/adoptive father/mother, and 25 region dummies of where the biological/
adoptive family lived in 1965. The numbers of observations in the second panel for own-birth and adopted
children are 94,079/2,125 in columns (1)–(6), 87,079/1,780 in column (7) and 91,932/1,976 in column (8).
1013THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
years of schooling. In column (1) we begin with the father-child
link and find that the estimated effects of both the biological and
adoptive father’s years of schooling are statistically significant,
positive, and equally important. The estimate of 0.11 implies that
one additional year of the biological or adoptive father’s schooling
raises the child’s schooling by just over one-tenth of a year. In
column (2) we consider the mother-child link. We observe positive
and significant associations for both biological and adoptive
mothers. Strikingly, however, the slope for biological mother’s
schooling of 0.13 is almost twice as high as the one for adoptive
mothers.
Of course, the estimated transmission effects represent both
the direct effect of the given parent’s schooling and the indirect
effect that comes from the other parent’s schooling; the indirect
effect is due to assortative mating on schooling, or on something
else that correlates with schooling. In our sample, the years-ofschooling
correlation between parents who adopt equals 0.49.
Between adoptees’ biological parents, however, this correlation is
only 0.19. In column (3) we take the intergenerational effect of the
partner into account by including both mothers’ and fathers’
schooling. For biological parents the partial influences of both
parents’ years of schooling fall somewhat but remain statistically
significant, positive, and equally important. For adoptive parents,
the partial schooling effects also fall, most notably for mothers.
We find that the maternal schooling effect is no longer significant
and close to zero while paternal schooling remains positive and
significant. These results are in line with recent studies on intergenerational
transmission of schooling that control for inherited
ability and assortative mating and produce schooling effects for
fathers that are higher than the schooling effects for mothers
[Behrman and Rosenzweig 2002, 2005; Antonovics and Goldberger
2005; Plug 2004].
In columns (4)–(6) of Table II we switch the dependent and
independent variable to a dummy for university degree. With this
variable most of our findings are similar to those previously
reported. The coefficients imply (a) that children with adoptive
mothers and fathers with a university degree experience a significantly
higher chance of graduating from university themselves;
(b) that for the intergenerational transmission of schooling,
prebirth factors also matter for university education; and (c)
that the relative contributions of biological and adoptive parents
are very similar to the ones obtained for years of schooling edu-
1014 QUARTERLY JOURNAL OF ECONOMICS
cation, indicating that pre- and postbirth factors are equally
important for father’s schooling and that prebirth circumstances
are more important for mother’s schooling. There is one notable
difference. We find that the estimated university effect for adoptive
mothers is positive and statistically significant with her
partner’s schooling included. Thus, it seems that for university
education both adoptive parents’ education contributes to the
education of the next generation.
We now turn to earnings and income. In column (7) we begin
with the father-child link in earnings and find that higher earnings
of the adoptive and biological father are associated with
higher earnings of the child. The estimated elasticities are 0.05
for the biological father’s earnings and 0.10 for the adoptive
father’s earnings, and only the latter is statistically significant.
The results for income, reported in column (8), are in line
with what we observed for earnings. We find (a) positive and
significant estimates for biological fathers; (b) positive and significant
slopes for adoptive fathers; and (c) indications that adoptive
fathers are more important than biological fathers in explaining
the income link.20
In the bottom two rows of panel 2, we show the sum of the
estimated coefficients for adoptive and biological parents of the
adopted children. We find that, in most cases, the sum is only
marginally different from the mobility estimate for a random
sample of own-birth children (panel 1). Only for university education
for mothers is the difference statistically significant.
We can learn four lessons from these adoption results. First,
we find that biological parents matter. All our mobility specifications
show positive and almost always significant slopes for biological
fathers and mothers. The partial impacts we find for
biological mothers’ and fathers’ education appear to be quite
similar. This is exactly what we expected. As long as genes are
automatically passed on from father to child and similarly from
mother to child, the genetic effects should be identical. The small
20. We note here that the impact of the income of the adoptive father is much
larger than for earnings. The difference in sample size is not responsible. In an
analysis not reported in the paper, we find that income effects remain the same
when estimated on the smaller earnings sample. The explanation for the different
estimates is that postbirth factors are more important for the intergenerational
transmission of nonlabor income (mainly capital income for this sample) than for
earnings, and that the reverse is true for postbirth factors. Anyway, when we test
for equality of earnings and income coefficients, a t-test of 1.54 does not reject
equality ( p-value ϭ 0.123).
1015THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
but positive differences between the effects of biological mothers
and fathers further suggest that effects that run through the
prenatal environment are relatively small. Such environmental
effects are even smaller when we consider classification error as
one of the alternative explanations. Suppose that the partial
impacts of biological mothers and fathers are identical, but that
fathers are more likely to be misclassified as biological fathers.
We would then observe bigger effects for mothers. In adoption
samples, where it might be more difficult to uncover the true
identity of biological fathers, we would then expect classification
errors to be higher for fathers than for mothers.
Second, we provide evidence that adoptive parents matter as
well. In our schooling regressions, the positive and significant
associations found for both parents do indicate that better educated
parents provide a better environment for their children to
do better in school.
Third, on the basis of a comparison of biological and adoptive
parents, we find that most of the mother’s influence on children
takes place through prebirth factors. For fathers we find pre- and
postbirth factors to be equally important for education, whereas
postbirth factors are more important for earnings and income.
Fourth, the total impact of the adoptive and biological parents’
resources on the outcomes of adoptive children is remarkably
similar to the impact of the biological parent’s outcomes for
that of biological children. This indicates that adoption per se (the
break from the biological mother, the time at the nursery) has
almost no effect on the strength of the intergenerational schooling
association among parents and children.21
V. ROBUSTNESS OF BASIC RESULTS
While our estimates suggest that for adoptees both their
adoptive and birth parents matter, we should treat these estimates
with care. Several problems involved in using adoption
data could lead to misinterpretations. In this section we concentrate
on four of these, being (a) nonrandom placement of adoptees
to their adoptive families, (b) not all adoptees are adopted as
babies, (c) many adoptees have unknown birth fathers, and (d)
adoptees and adoptive parents are different from other children
21. Note that this does not hold for the estimated effects found for mothers
with university degrees.
1016 QUARTERLY JOURNAL OF ECONOMICS
and their parents. To examine the impact of each of these four
problems, we focus our attention on specifications that use years
of schooling and earnings for reasons of brevity.22
V.A. Nonrandom Assignment
In almost all adoption studies random assignment is assumed.
But in the presence of selective placement, this assumption
is violated, and adoption estimates are biased. Without information
on the adoptees’ birth parents—which applies to most
adoption studies—the random assignment assumption is practically
untestable. With our data on the adoptees’ biological background,
however, we can test whether our adoption results are
sensitive to selective placement.
We first investigate whether there is evidence of nonrandom
assignment of adoptees in Sweden and estimate the relationship
between the education and earnings characteristics of the adoptive
and biological parents of adoptees. Random assignment
would give us zero correlations. We, on the other hand, find
correlations that range from 0.091 for fathers’ earnings to 0.144
(0.140) for fathers’ (mothers’) years of schooling. These numbers
are quite high and suggest that nonrandom assignment among
adoptees and their adoptive parents is substantial.23
If we think of this particular assignment problem as an
omitted-variable bias problem, nonrandom assignment affects
our adoption estimates if Yi
ap
is correlated with unobserved prebirth
factors or Yj
bp
is correlated with unobservable postbirth
factors. To assess the robustness of our results against omitted
variables, we check how the estimates attached to the adoptive
parent’s schooling and earnings change when we (a) exclude the
biological parent’s controls for schooling and earnings and (b)
include as many background characteristics of the biological parents
as are available and measured around the time of adop-
tion.24
Small changes indicate robustness. In rows 2 and 3 in
Table III, we show that the coefficients that correspond to the
22. We have also compared the results with those obtained for university
education and income. We found no systematic differences. Sensitivity results for
university education and income are available upon request.
23. These correlations are not driven by age and region effects. When we
regress out age and region effects, the correlations remain virtually identical.
Also, note that the magnitude of the correlations is very similar to the correlations
in IQ between the biological and adoptive mother/father found in the well-known
Texas Adoption Study [Brody 1992].
24. Biological parents’ age, marital status, education, earnings, income, occupation,
and regional controls measured at 1965 or 1970.
1017THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
TABLE III
SENSITIVITY ANALYSES: ALTERNATIVE SAMPLES AND SPECIFICATIONS
Years of schooling Earnings
Fathers Mothers Fathers
Bio Adopt Bio Adopt Bio Adopt
Adopted children
(1) Baseline results: .113** .114** .132** .074** .047 .098**
(N ϭ 2125, 1780) (.016) (.013) (.017) (.014) (.034) (.038)
Other samples:
Nonrandom assignment
(2) exclude info on birth .126** .093** .095*
parents (N ϭ 2125, 1780) (.013) (.014) (.037)
(3) include info on birth .097** .055** .027
parents (N ϭ 2125, 1780) (.013) (.015) (.039)
(4) exclude info on adoptive .132** .150** .045
parents (N ϭ 2125, 1780) (.016) (.017) (.033)
(5) include info on adoptive .105** .117** .038
parents (N ϭ 2125, 1780) (.017) (.018) (.034)
Adoption age
(6) adopted between age 0 .109** .120** .124** .062** .079 .124
and 2 (N ϭ 638, 573) (.030) (.024) (.033) (.026) (.067) (.074)
Unknown fathers
(7) including those without .128** .083**
info on birth fathers (.012) (.010)
(N ϭ 4123)
Comparable samples
(8) raised with own-birth .059 .129** .114** .052 .123 .035
siblings (N ϭ 526, 435) (.034) (.024) (.035) (.030) (.064) (.077)
Own-birth children
(9) Baseline results: .240** .243** .235**
(N ϭ 94,079, 87,079) (.002) (.002) (.005)
Other samples:
(10) raised with adopted .285** .251** .280**
siblings (N ϭ 412, 381) (.031) (.035) (.080)
(11) with bio siblings .180** .106 .216
adopted out (.056) (.067) (.113)
(N ϭ 193, 160)
Matched samples:
(12) on adoptees, rearing .248** .254** .217**
parents (N ϭ 84,358, (.003) (.004) (.008)
78,229)
(13) on adoptees’ bio .199** .196** .182**
background (N ϭ 93,655, (.008) (.009) (.021)
86,703)
Standard errors are shown in parentheses; * indicates significance at 5 percent level, and ** at 1 percent
level. All specifications include controls for the child’s gender, 4 birth cohort dummies for the child, 8 birth
cohort dummies for biological/adoptive father/mother, and 25 region dummies of where the biological/
adoptive family lived in 1965. Also sample sizes are shown in parentheses. For instance, in row 1 (N ϭ 2125,
1780) means that number of observations is 2125 in columns 1–4 (for years of schooling), and 1780 in columns
5–6 (for earnings).
1018 QUARTERLY JOURNAL OF ECONOMICS
adoptive parent’s outcomes rise (fall) when we exclude (include)
the biological parents’ schooling or earnings, but not by much.
The only exception appears to be the estimate for the adoptive
father’s earnings in row 3, which is much smaller than the one we
observe in our baseline. Yet the difference is not significantly
different. By the same logic, it is useful to see what happens to the
coefficients of the adoptees’ birth parents when we exclude and
include information of the adoptive parents. The estimates attached
to the adoptees’ birth parents in rows 4 and 5 appear to be
even less sensitive to the inclusion (exclusion) of the adoptive
parents’ characteristics. The coefficients change, but they change
only in the margin.
If, instead, we think of this assignment problem as a measurement
error problem, adoption estimates are biased as well. In
addition to the standard argument that measurement error in
Yj
bp
and Yi
ap
will bias the estimated pre- and postbirth effects to
zero, there is also the combination of nonrandom assignment and
measurement error that is pushing the estimated pre- and postbirth
effects upwards. That is, the estimated influence of postbirth
factors will increase the greater are the errors in measuring
prebirth factors, and vice versa.
When we formalize these measurement error processes, we
are able to assess the magnitude of the bias. Suppose that the
true mobility model is given by
(5) yi
ac
ϭ ␣0 ϩ ␣1 yj
bp
ϩ ␣2 yi
ap
ϩ ␯i
ac
.
Instead of yj
bp
and yi
ap
we observe Yj
bp
and Yi
ap
. We assume
classical measurement error and define Yk
ϭ yk
ϩ εk
, where εk
(k ϭ bp, ap) represent the measurement errors that are uncorrelated
with yj
bp
, yi
ap
and with each other. If cov(Yj
bp
, Yi
ap
) ϭ CY,
and var(Yj
bp
) ϭ var(Yi
ap
) ϭ VY, the least-squares estimators in
our estimated model will have the following properties:
(6) p lim ␣ˆ1 ϭ ␣1 ϩ ͑1 Ϫ h͒
m␣2 Ϫ ␣1
1 Ϫ m2 ,
(7) p lim ␣ˆ2 ϭ ␣2 ϩ ͑1 Ϫ h͒
m␣1 Ϫ ␣2
1 Ϫ m2 ,
where m ϭ CY/VY and h ϭ Vy/VY represent nonrandom assignment
and reliability ratio.25
25. Note that the regression of Yj
bp
on Yi
ap
gives us an estimate of m.
1019THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
If we further assume reliability ratios h of 0.88 for our years
of schooling and log earnings measures and a correlation between
biological and adoptive parents schooling and earnings m of
0.15—in the upper end of what we find in our data—it is possible
to predict the magnitude of the bias.26
The estimates in Table III,
row 1, suggest that ␣ˆ1 and ␣ˆ2 are equal to 0.11, 0.13 and 0.05, and
0.11, 0.07 and 0.10, for fathers’ and mothers’ schooling and fathers’
earnings, respectively. Equations (6) and (7) then predict
that true values ␣1 and ␣2 should be 0.13, 0.15 and 0.05, and 0.13,
0.08 and 0.11. It turns out that the bias introduced by the combination
of selective placement and measurement error accounts
for no more than 13 percent of the estimated impact of the
parents’ characteristics. Note that the bias in the estimated prebirth
(postbirth) effects introduced by those observed postbirth
(prebirth) factors that are measured with error is much smaller
and accounts for at most 5 percent of the estimated impact of the
adoptive (biological) parents’ characteristics.27
In sum, we do not believe that selective placement is affecting
our results in a substantial way as results remain qualitatively
very similar whether we exclude or include variables (as in
the case of omitted variables) or when we assume reasonable
magnitudes of measurement errors.
V.B. Adoption Age
Our second problem is that some children are adopted at a
later age. If a significant number of adoptees are not adopted as
babies, we end up overestimating prebirth effects and underestimating
postbirth effects. So far, we have ignored adoption age
and implicitly assumed that adoption took place at birth. With
respect to the obtained effects for the adoptees’ birth parents, this
seems a reasonable assumption. In Sweden possible postbirth
effects that come from the adoptees’ birth parents do not exist for
most adoptees since most children who are registered for adoption
are placed in special nursery houses the moment they are
26. Our reliability ratios are based on those found in previous studies. Isacsson
[1999], using the same education data as we, estimates a reliability ratio for
years of schooling of 0.88. Reliability ratios in annual earnings measures are
smaller: about 0.64; see Bjo¨rklund [1993]. Reliability ratios, however, increase
with the number of measures. Would the reliability ratio be obtained with 5,
instead of 1, earnings measures, as we do, the ratio would rise to .89 in the
absence of serial correlation; see Solon [1992, footnote 17] for formulas. Serial
correlation is likely to be low when income is measured five years apart.
27. An example of this type of analysis in a different setting is contained in
Borjas [1992].
1020 QUARTERLY JOURNAL OF ECONOMICS
born. With respect to the obtained effects estimated for the adoptive
parents, we are not so sure whether the adoption-at-birth
assumption is likely to hold. The problem is that we do not have
a clean measure of adoption age. We can infer the age at adoption
only crudely from the time we observe whether children are
adopted or not in the censuses of 1965 or 1970. If we look at those
adoptees born exactly one year prior to the census date in 1965—
meaning those who are born between October and December
1964—we find that about 80 percent of all adoptees (94 out of
117) are adopted within a year. If this is typical for adoptees
during this time, it implies that over 80 percent of all our adoptees
are adopted within a year of their birth date. We therefore do
not worry too much about this issue. We still perform some
sensitivity analysis by restricting our adoption sample to all
children between 0 and 2 who are adopted at the time we observe
them in the census.28
With this subsample we can estimate the
effects that come from the adoptive parents more accurately
looking at adoptees who are more likely to be adopted as babies
and receive the full parental treatment. Results are reported in
row 6 of Table III. Other than the reduction in sample size, we
find that the estimates attached to the parental schooling and
earnings variables are very similar to the ones we observe for the
full sample. These results give no compelling reason to believe
that timing of adoption seriously affects our estimates.
V.C. Unknown Fathers
A third problem could arise because we have restricted our
sample of adoptees to those for whom we have information on
both biological parents. But for almost half of the original sample
the father is unknown. It is possible that children with unknown
birth fathers are different from other children in ways related to
their parental resources. To test whether absent information on
the father affects the estimates, we extend the current sample of
adoptees to adoptees for whom we have information only on birth
mothers. In row 7 of Table III we report schooling estimates for
birth and adoptive mothers that are almost identical to the ones
observed for the restricted sample (but with higher precision). We
therefore rule out this source of bias.
28. These children are born between November 1963 and November 1965,
and live with their adoptive parents—or parents who subsequently will become
their adoptive parents—at the time we observe them in the census in (November)
1965.
1021THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
V.D. Comparable Samples
Finally, we address the problem that adoptees and adoptive
parents are different from other children and their parents. As
discussed in subsection II.B, we do this in two different ways. We
consider own-birth children who are reared in a childhood environment
that is comparable with the post-childhood environment
of adoptees. We then consider own-birth children whose parental
characteristics are comparable with those of adoptees’ birth parents.
We look at comparable samples in two different ways, first
by using subsampling and then using propensity score matching.
We begin with limiting our own-birth sample to those families
who also have adopted children. We know that the adopted
and own-birth children in these families share the same childhood
environment, but not the prebirth experience. If estimates
for adoptees are to be informative about intergenerational associations
between own-birth children and their parents, we expect
estimates for own-birth children with adopted siblings to be similar
to the estimates for all own-birth children. This is not the case.
In rows 9 and 10 we see that the estimates for own-birth children
in adoptive families are somewhat bigger than the estimates for
the representative sample of own-birth children.29
We then limit
our birth sample to those children born in families in which at
least one child is given up for adoption. We know that these
own-birth children share similar genes and pre-childhood experiences
with adoptees, and that they therefore start their lives
under very similar conditions as adoptees do. When we estimate
previous intergenerational relationships using these particular
own-birth children, we find schooling and earnings estimates that
are smaller than those observed for all own-birth children.30
The disadvantage of looking at particular subsamples is that
we need to work with very small samples. We therefore also
present results for comparable samples using propensity score
29. We do not think that treatment differentials are driving these results. Of
course, these findings are consistent with the idea that parents may favor their
own offspring over their adopted children. Case, Lin, and McLanahan [2000]
propose selfish genes as one of the responsible mechanisms. Selfish genes, however,
would also predict that intergenerational transmissions for adoptees with
own-birth siblings are weaker than for other adoptees. Rows 1 and 8 indicate that
this is not the case. In fact, most of the estimates for adoptive children in these
families are quite similar to estimates for the sample of all adopted children.
30. But if parents (mostly mothers) could choose and rather put their problematic
child up for adoption, it is possible that we find lower correlations because
these children are less sensitive to parental treatments.
1022 QUARTERLY JOURNAL OF ECONOMICS
matching in rows 12 and 13,31
even though we acknowledge that
this may not eliminate all the differences between adoptees and
own-birth children. When we match the characteristics of the
own-birth children’s parents against the characteristics of the
adoptees’ rearing parents, we get slopes that are very similar and
in most cases statistically identical to those obtained for all ownbirth
children (row 12). When we make the samples more comparable
by matching the characteristics of own-birth children’s
parents against the characteristics of adoptees’ biological parents,
the results are in line with what we find earlier: all estimates
are smaller than those observed for all own-birth children
(row 13). In most cases the reduction is statistically significant.
Overall, we find that subsampling and matching generate similar
patterns.
We conclude that our results using more comparable samples
suggest that intergenerational associations of schooling, earnings,
and income for own-birth children are often stronger for
own-birth children who share postbirth backgrounds with adoptees
than those own-birth children who share prebirth backgrounds
with adoptees. If these findings also imply that intergenerational
associations are much stronger in families that are
better educated and generate more income, we might question
whether the linear specifications we estimate in equations (1) and
(2) are in fact linear. Our results indicate that they are not.
VI. NONLINEAR MODELS
We test for nonlinear intergenerational transmission by including
the square of parental schooling and the square of the
father’s earnings and income. In the first panel of Table IV, where
we report the new mobility estimates obtained from our repre-
31. The matching is done by regressing an indicator of adoption status (1 if
adopted, 0 if nonadopted) on the following variables: 4 birth year dummies for the
child, child’s gender, 5 educational level dummies of the father and mother in
1970, father’s and mother’s income in 1970, dummies for father and mother
having positive income in 1970, father’s and mother’s earnings in 1970, dummies
for father and mother having positive earnings in 1970, about 50 birth year
dummies for the father and mother, mother’s marriage status measured in 5
categories in 1970, 57 two-digit occupation dummies for both parent’s occupation
in 1970, and 25 region dummies of the mother in 1965. When we match on
adoptees’ family environment (panel 3) we use the characteristics of the adoptive
mother and father and when we match on adoptees’ biological background (panel
2) we use the characteristics of the biological mother and father. All of the
estimates were obtained using Leuven and Sianesi’s [2003] psmatch2 program for
Stata.
1023THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
TABLEIV
ESTIMATEDTRANSMISSIONCOEFFICIENTSINNONLINEARMODELSWITHINTERACTIONS
YearsofschoolingUniversityEarningsIncome
FathersMothersFathersMothersFathersFathers
(1)(2)(3)(4)(5)(6)(7)(8)(9)(10)
Own-birthchildren
BioparentϪ.009Ϫ.058**Ϫ.807**Ϫ.938**
(.015)(.017)(.075)(.064)
Bioparentsquared.011**.014**.069**.077**
(.001)(.001)(.005)(.004)
Adoptedchildren
Bioparent.050Ϫ.222Ϫ.055Ϫ.472**.199**.166**Ϫ.187Ϫ.403Ϫ1.164*Ϫ1.342*
(.051)(.127)(.055)(.139)(.045)(.041)(.108)(.502)(.525)(.670)
Bioparentsquared.015*.023**.017.015
(.006)(.006)(.037)(.034)
Adoptiveparent.061Ϫ.003Ϫ.097Ϫ.310**.170**.108**Ϫ.293*Ϫ.076Ϫ.995*Ϫ.998
(.043)(.090)(.050)(.121)(.025)(.026)(.125)(.648)(.501)(.710)
Adoptiveparent.004.012*Ϫ.003.003
squared(.004)(.005)(.043)(.035)
Bioparent‫ء‬Adoptive.006.003.018**.013*Ϫ.041.286**.043**.034**.156*.151*
parent(.004)(.005)(.005)(.005)(.074)(.071)(.015)(.010)(.067)(.068)
Standarderrorsareshowninparentheses;*indicatessignificanceat5percentlevel,and**at1percentlevel.Allspecificationsincludecontrolsforthechild’sgender,4birth
cohortdummiesforthechild,8birthcohortdummiesforbiological/adoptivefather/mother,and25regiondummiesofwherethebiological/adoptivefamilylivedin1965.Thenumbers
ofobservationsinthesecondpanelforown-birthandadoptedchildrenare94,079/2,125incolumns(1)–(6),87,079/1,780incolumn(7),and91,932/1,976incolumn(8).
1024 QUARTERLY JOURNAL OF ECONOMICS
sentative sample of own-birth children, we find that the estimates
attached to the quadratic terms are always positive and statistically
significant. This clearly suggests that the intergenerational
associations are stronger in families with higher education and
income. Our results correspond to the nonlinear intergenerational
effects found in earlier mobility studies [Behrman and
Taubman 1990; Solon 1992; Bjo¨rklund and Chadwick 2003].
The question why intergenerational transmissions are so
much stronger at the top than at the bottom of the schooling,
earnings, and income distribution is an important one. Some
authors have argued that interactions between nature and nurture
are very important [Dickens and Flynn 2001; Ridley 2003].
And indeed, if smart children would benefit relatively more from
having smart parents, the intergenerational transmission would
be greatest among highly educated and high-income families.
Finding credible evidence, however, is difficult. Our adoption
data offer a great opportunity to test whether this is the mechanism
at work.
To estimate that part of the transmission that comes from
the interaction between the postbirth environment (adoptive parents)
and genetic factors (biological parents), we include the interacted
effect between the adoptive and biological parents. This
is done in the second panel of Table IV. We find evidence of a
positive interaction for mother’s education and father’s earnings
and income, but not for father’s education. To test whether these
interacted effects are not picking up other nonlinear effects that
possibly exist between parents and their children, we also include
the square of parental education, earnings, and income. Our
results indicate that this is not the case. The interacted estimates
are not sensitive to the inclusion of higher order terms.
VII. CONCLUSIONS
In this paper we investigate the origins of intergenerational
education and income associations using data on Sweden-born
adoptees and their biological and adoptive parents. Our empirical
strategy is to decompose the intergenerational association into
prebirth and postbirth components, or combinations thereof. We
use the biological parent’s characteristics as an indicator of genetic
background and prenatal environment, and the adoptive
parent’s characteristics as an indicator of the child’s postbirth
environment. Our conclusions follow from regressions where we
1025THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
simultaneously include the biological and adoptive parent’s characteristics
in intergenerational mobility equations.
We find that both pre- and postbirth factors are important for
the child’s educational and economic outcomes. That is, for none
of the outcomes studied can these factors be said to be negligible.
The relative contributions fluctuate a bit. For mother’s education,
for example, we find that prebirth factors are more important
than postbirth environment, whereas for father’s long-run earnings
and income, the postbirth environment is more important
than prebirth factors. We also find evidence of slightly larger
intergenerational transmission coefficients for biological mothers
than for biological fathers. Because the impact of the biological
mother reflects both genetic and prenatal environmental factors,
whereas the impact of the biological father reflects only genetic
factors, we believe that prenatal environmental factors are small
in magnitude. This in turn allows us to discuss our estimated preand
postbirth effects using the classical nature and nurture labels.
For example, we can interpret our estimated interaction
effects between biological and adoptive parents’ status as naturenurture
interactions. Interestingly, such interactions are positive
and significantly different from zero, for mother’s education and
father’s income. These interactions further corroborate our conclusion
that both pre- and postbirth factors are important, but
also suggest that the importance of nurture varies across the
nature distribution.
What are the implications of these findings? First, we provide
evidence that both adoptive and biological parents matter, which
suggests that both nature and nurture components are important.
This implies that any comprehensive explanatory theory of
intergenerational mobility must incorporate both factors like genetic
heredity and factors in the rearing family. Any theory that
only focuses on one of these will be incomplete. These results also
help us to understand why specific policies have an impact on
intergenerational mobility. For instance, welfare policies can increase
mobility by improving the environment in which children
are raised. Further, antidiscrimination policies can increase mobility
by reducing the impact of physical characteristics that are
genetically determined. Second, our findings with respect to the
positive nature and nurture interactions also raise some interesting
issues. Both nature and nurture remain important, but if the
two operate together, it becomes very difficult to separate one
from another. If improving equality of opportunity is a policy
1026 QUARTERLY JOURNAL OF ECONOMICS
objective, then our results suggest that policies should be designed
to improve the conditions particularly for children raised
in low-educated and poor families. After all, in the presence of
interactions these children suffer from being less able, being
raised under poorer conditions as well as the interactions between
the two. But if, at the same time, these policies do not
recognize that genetically disadvantaged children benefit much
less from an improved environment than more able children do,
our results also suggest why policies like these have not been that
effective in the past.
SWEDISH INSTITUTE FOR SOCIAL RESEARCH (SOFI), STOCKHOLM UNIVERSITY
SWEDISH INSTITUTE FOR SOCIAL RESEARCH (SOFI), STOCKHOLM UNIVERSITY
DEPARTMENT OF ECONOMICS, TINBERGEN INSTITUTE, UNIVERSITY OF AMSTERDAM
REFERENCES
Antonovics, Kate L., and Arthur S. Goldberger, “Does Increasing Women’s Schooling
Raise the Schooling of the Next Generation? Comment,” American Economic
Review, XCV (2005), 1738–1744.
Behrman, Jere R., and Mark R. Rosenzweig, “Does Increasing Women’s Schooling
Raise the Schooling of the Next Generation?” American Economic Review,
XCII (2002), 323–334.
Behrman, Jere R., and Mark R. Rosenzweig, “Does Increasing Women’s Schooling
Raise the Schooling of the Next Generation? Reply,” American Economic
Review, XCV (2005), 1745–1751.
Behrman, Jere R., and Paul Taubman, “The Intergenerational Correlation between
Children’s Adult Earnings and their Parents’ Income: Results from the
Michigan Panel Survey of Income Dynamics,” Review of Income and Wealth,
XXXVI (1990), 115–127.
Bjo¨rklund, Anders, “A Comparison between Actual Distributions of Annual and
Lifetime Income: Sweden 1951–1989,” Review of Income and Wealth, XXXIX
(1993), 377–386.
Bjo¨rklund, Anders, and Laura Chadwick, “Intergenerational Income Mobility in
Permanent and Separated Families,” Economics Letters, LXXX (2003),
239–246.
Bjo¨rklund, Anders, and Markus Ja¨ntti, “Intergenerational Income Mobility in
Sweden Compared to the United States,” American Economic Review, LXXXVII
(1997), 1009–1018.
Bjo¨rklund, Anders, Markus Ja¨ntti, and Gary Solon, “Influences of Nature and
Nurture on Earnings Variation: A Report on a Study of Sibling Types in
Sweden,” in Samuel Bowles, Herbert Gintis, and Melissa Osborne, eds.,
Unequal Chances: Family Background and Economic Success (New York:
Russell Sage Foundation, 2005).
Bjo¨rklund, Anders, Mikael Lindahl, and Erik Plug, “Intergenerational Effects in
Sweden: What can we Learn from Adoption Data?” IZA Discussion Paper No.
1194, 2004.
Bo¨hlmark, Anders, and Matthew Lindquist, “Life-Cycle Variations in the Association
between Current and Lifetime Income: Replication and Extension for
Sweden,” Journal of Labor Economics (2006), forthcoming.
Bohman, Michael, Adopted Children and their Families (Stockholm: Proprius,
1970).
Borjas, George J., “Ethnic Capital and Intergenerational Mobility,” Quarterly
Journal of Economics, CVII (1992), 123–150.
Brody, Nathan, Intelligence, 2nd edition (San Diego, CA: Academic Press, 1992).
1027THE ORIGINS OF INTERGENERATIONAL ASSOCIATIONS
Case, Anne, I-Fen Lin, and Sara McLanahan, “How Hungry Is the Selfish Gene?”
Economic Journal, CX (2000), 781–804.
Cunha, Flavio, James J. Heckman, Lance Lochner, and Dimitry V. Masterov,
“Interpreting the Evidence on Life Cycle Skill Formation,” NBER Working
Paper No. 11331, 2005.
Das, Mitali, and Tanja Sjogren, “The Inter-generational Link in Income Mobility:
Evidence from Adoptions,” Economics Letters, LXXV (2002), 55–60.
Dickens, William T., and James R. Flynn, “Heritability Estimates versus Large
Environmental Effects: The IQ Paradox Resolved,” Psychological Review,
CVIII (2001), 346–369.
Haider, Steven J., and Gary Solon, “Life-Cycle Variation in the Association between
Current and Lifetime Earnings,” American Economic Review, 2006
forthcoming.
Isacsson, Gunnar, “Estimates of the Return to Schooling in Sweden from a Large
Sample of Twins,” Labour Economics, VI (1999), 471–489.
Kungl. Socialstyrelsen, “Råd och anvisningar i socialvårdsfrågor,” No. 209, 1968.
Leuven, Edwin, and Barbara Sianesi, “psmatch2: Stata module to perform full
Mahalanobis and propensity score matching, common support graphing, and
covariate imbalance testing,” http://ideas.repec.org/c/boc/bocode/s432001.html.,
version x.x.x., 2003.
Nordlo¨f, Barbro, “Svenska adoptioner i Stockholm (Swedish adoptions in Stockholm
1918–1973,” FOU-rapport 2001:8 Socialtja¨nstfo¨rvaltningen, Stockholm
Stad, 2001.
Plomin, Robert, John C. DeFries, and David W. Fulker, Nature and Nurture
During Infancy and Early Childhood (Cambridge, UK: Cambridge University
Press, 1988).
Plug, Erik, “Estimating the Effect of Mother’s Schooling on Children’s Schooling
Using a Sample of Adoptees,” American Economic Review, XCIV (2004),
358–368.
Plug, Erik, and Wim Vijverberg, “Schooling, Family Background, and Adoption: Is
It Nature or Is It Nurture?” Journal of Political Economy, CXI (2003),
611–641.
Plug, Erik, and Wim Vijverberg, “Does Family Income Matter for Schooling
Outcomes? Using Adoption as a Natural Experiment,” Economic Journal,
CXV (2005), 880–907.
Ridley, Matt, Nature via Nurture (New York: Harper Collins, 2003).
Sacerdote, Bruce, “The Nature and Nurture of Economic Outcomes,” NBER Working
Paper No. 7949, 2000.
——, “The Nature and Nurture of Economic Outcomes,” American Economic
Review Papers and Proceedings, XCII (2002), 344–348.
——, “What Happens When we Randomly Assign Children to Families?” NBER
Working Paper No. 10894, 2004.
Scarr, Sandra, and Richard Weinberg, “The Influence of Family Background on
Intellectual Attainment,” American Sociological Review, LVIII (1978),
674–692.
Solon, Gary, “Intergenerational Income Mobility in the United States,” American
Economic Review, LXXXII (1992), 393–408.
——, “Intergenerational Mobility in the Labor Market,” in Orley Ashenfelter and
David Card, eds., Handbook of Labor Economics, Volume III (Amsterdam:
Elsevier, 1999).
Turkheimer, Eric, Andreana Haley, Mary Waldron, Brian D’Onofrio, and Irving I.
Gottesman, “Socioeconomic Status Modifies Heritability of IQ in Young Children,”
Psychological Science, XIV (2003), 623–628.
1028 QUARTERLY JOURNAL OF ECONOMICS