Earnings, Education, Genetics, and Environment
Author(s): Paul Taubman
Source: The Journal of Human Resources, Vol. 11, No. 4 (Autumn, 1976), pp. 447-461
Published by: University of Wisconsin Press
Stable URL: http://www.jstor.org/stable/145426
Accessed: 20-02-2018 12:15 UTC
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
University of Wisconsin Press is collaborating with JSTOR to digitize, preserve and extend
access to The Journal of Human Resources
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
EARNINGS, EDUCATION,
GENETICS, AND ENVIRONMENT
PAUL TAUBMAN
ABSTRACT
A major and well-recognized difficulty in estimating the effects of education
on earnings is that the more educated are likely to be more able, irrespective
of education. If ability also determines earnings and is not controlled,
ordinary least squares will yield biased estimates of the education coefficient.
In this study, we use data on identical twins to control for differences
in ability that arise from genetic endowments and family environment.
Not controlling for genetics and family environment may cause a
large bias, up to two-thirds of the noncontrolled coefficient.
I. INTRODUCTION
Much research in economics has been concerned with the sources of the
inequality in earnings. While the human capital model has focused attention on
the role of education, on-the-job training, and, more peripherally, innate and
acquired cognitive skills, other economists have examined such diverse factors as
risk aversion, healthiness, and compensating differences for nonpecuniary rewards.l
While the development of better data, estimating techniques, and
economic theories have all led to an improved understanding of the sources of
inequality in earings, our empirical knowledge is unsatisfactory on a number of
important issues.
A widely recognized problem in prior studies of the impact of education on
earnings is that people with more education are likely to be more "able," net of
any additional ability produced by education. Unfortunately, since "ability" can
encompass many particular skills including cognitive, affective, and psychoThe
author is Professor of Economics, University of Pennsylvania.
[Manuscript submitted September 1975; accepted March 1976.]
1 For a summary and extension of the human capital model, see Mincer [8]. For a
discussion of risk aversion and other factors, see Taubman [12].
The Journal of Human Resources . XI . 4
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
448 I THE JOURNAL OF HUMAN RESOURCES
logical, few if any of which are measured adequately, it is difficult to pin
the effects of education. This study uses data on twins to attack this abil
problem from a different angle. When we regress differences in brothers' ear
on differences in schooling, we hold constant (by eliminating) those abilitie
are common to the brothers. For fraternal twins, we eliminate skills produc
the common or family environment, while for identical twins we elimin
common environment and skills based on genetic endowments.
Data on twins can also be used to explore the extent to which the vari
in earnings is attributable to the sum and the separate effects of ge
endowments and common (family) environment and the extent to which
variance is attributable to noncommon environment.2 The verb explore was
in the preceding sentence both because this is the first study in economi
examine these issues and because there are major difficulties or unproven
assumptions associated with the model used in this paper. However, w
undertaken subsequent to that reported here gives hope that many of
assumptions can be tested and, if valid, replaced with more tenable version
[1]).
In this paper I will use a version of the human capital model that is fairly
general and not particularly rigorous. Our basic assumption will be that a person
will be paid a real wage rate (the money wage rate divided by the price level,
which will be set at 1) equal to his marginal product. A person's marginal
productivity depends on a variety of skills and attributes. Unfortunately, about
the only potential skill that has been widely examined is "cognitive" ability.3
Economists and sociologists have related earnings to education, years of
experience, and other variables on the grounds that if schooling produces one or
more skills that are reimbursed in the marketplace, then the more educated
should receive higher earnings. We will generalize this model by assuming that
skills are produced by a combination of genetic endowments and environment,
where the latter is defined to encompass "everything else." If years of schooling
happens to be the only input into skills that we measure, we can write:
(1) InY=aS + bG + cN + u
where InY is the log of earnings, S is years of schoolin
ments, N is environment other than schooling, and u is
2 The common environment will include neighborhood and ot
will exclude some family effects. See below for details.
3 It is difficult to get a clear picture of the effect of this skill b
different concepts and measures of cognitive ability and stu
years of labor market experience. See, for example, Taubma
and Mason [6], and Sewell and Hauser [9]. The latter study
cognitive skills begin to have an effect only after five years
See Taubman [121 for an attempt to measure the effects of
4 If measures of other skill producers are available, they can
redefinition of G and N.
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
Taubman I 449
This paper will attempt to shed light on some of the influences of genetics,
family, and other environment for earnings for white men aged about 50 in
1973. We will use both regression and variance analyses to study several different
but related issues.
II GENETICS AND ENVIRONMENT
By genetic endowments (G), we mean the innate capabilities that are based on
person's genes, half of which are contained in the egg and the other half in the
sperm.5 Environment (N) includes all other systematic and nonsystematic determinants
of skills, including prenatal development. While environment is "everything
else," some particular aspects that are usually thought to be important
include family, peer group, on-the-job training, schooling, and military. As this
list indicates, an individual's environment includes elements over which he has
both little and much choice.
Twin Types
Males and females normally have 23 pairs of chromosomes. The genes are
contained in the chromosomes. Each gene contains two halves that may be the
same or different, for example, AA or AB. We will assume that each skill or trait
is influenced by many genes. Only a randomly determined half of each gene and
of each chromosome is contained in the egg or sperm, each of which is a gamete
of one parent. But once the egg is fertilized, that is, the two gametes combine,
the two halves of each gene are merged to form the individual's genes.
There are two types of twins-monozygotic (MZ) and dizygotic (DZ). The
MZs, often known as "identical," are the result of the splitting of an already
fertilized egg, while the DZs, or "fraternal twins," are the result of two different
eggs fertilized by two different sperm. Thus, DZ pairs do not have the same
genetic composition, although they will be more alike than randomly drawn
individuals. The MZ pairs, however, have the same genetic makeup because each
piece of the split fertilized egg contains all and only the genetic information of
the original fertilized egg (barring mutations).6
III. THENAS-NRC TWINSAMPLE
In this study, we will use the NAS-NRC twin sample, which is described in more
detail in the Appendix. Briefly, however, for this study we have a maximum of
5 We are ignoring mutations which occur very rarely, i.e., about once in 100,000 or less.
For discussion of the biological and statistical aspects of genes, see [4].
6 For a more complete discussion of the biological aspects, see [4].
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
450 I THE JOURNAL OF HUMAN RESOURCES
2,478 pairs, where each brother answered a questionnaire mailed in 1974.7 Most
of these pairs also answered several earlier questionnaires to which we have
access. To be included in the mailing, the twins had to be born between 1917
and 1927, to be white, to have served at some point in the military, and to be
alive in 1974. These last restrictions suggest an underrepresentation of people
with low intelligence or education or with poor mental and physical health, as
compared to the corresponding white male cohort. Even compared to a population
of veterans, the respondents to our questionnaire have more earnings and
education.8'9 Regression estimates from stratified samples with nonpopulation
weights still yield unbiased coefficients over the sample space.10
In some of our analysis, we assume that MZ and DZ pairs in our sample are
random drawings from the same population.1 The DZ pairs may have different
distributions of genes and/or environment because DZ pairs as a percentage of
births occur more frequently among older women. The corresponding percentage
for MZ pairs is independent of mother's age and SES class.12
The means and variances of earnings, schooling, and several other variables
by twin types are given in Table 1. It is evident that in our sample, the DZ twins
come from families in which the parents have a bit less education and occupational
status, and in which the number of siblings and older siblings are
somewhat greater, although the differences are not statistically significant. The
religious distributions are also very similar for MZs and DZs, which is a bit
surprising to me since I would have expected Catholics to be a larger portion of
7 We eliminate pairs if either does not have earnings whose derivation is described in
Taubman [11]. About 100 pairs whose zygosity is unknown are not used in this
analysis. If a person's education is unknown, we set it equal to his brother's since even
for DZ pairs, the brothers' correlation for education is almost .5. If neither brother
responded to the education question, we used the mean of 13. Less than 20 cases were
adjusted.
8 The average 1973 earnings and education in our sample are $18,000 and 13 years. In
the population as a whole, the corresponding figures for white veterans of the same
cohort are about $12,000 and 12 years. About one quarter of the differential can be
eliminated if we reweight by parental education and region of birth so as to produce the
average of white males born during the period 1917-27 on these variables.
9 While our sample is not representative of the population, it seems likely that we have
over- and underrepresented various population groups rather than excluded all their
members. For example, in Stauffer and others [ 10], it is indicated that there were huge
differences in disqualification for mental problems by induction camp, ranging from .5
to 50 percent.
10 However, since less than 5 percent of the sample had less than a ninth grade education,
our results may not be appropriate for those with low education.
11 In our statistical analysis, it is necessary to distinguish between MZ and DZ twins. For
the most part, the twins' zygosity is determined by their answers to: "As children, were
you and your twin alike as 'two peas in a pod' or of only ordinary resemblance?" This
simple question assigns pairs accurately almost 95 percent of the time. See [11], App.
A, for details.
12 See [11], App. A, for discussion and references.
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
Taubman I 451
TABLE 1
SOME SUMMARY STATISTICS FOR INDIVIDUALS IN THE NAS-NRC SAMPLE
(calculated separately for MZs and DZs)
MZs DZs
Mean Variance Mean Variance
1973 annual earnings 18.4a 150b 18.la 166b
In 1973 annual earnings 9.67a .28 9.64a .32
1967 or 1972 occupational
scored 50.4 472 49.8 445
Years of schooling 13.5 9.1 13.3 9.8
Initial full-time civilian
occupationcd 36.7 610 35.0 590
Age 51.0 8.4 51.2 8.8
Mother's education years 10.0 9.6 9.7 11.9
Father's education years 9.3 12.6 9.1 14.8
Father's occupational statusd 29.6 532 28.6 503
% Catholic 26 19 23 18
% Jewish 4 4 5 5
% other non-Protestant 2 2 3 3
Number of siblings alive 1940 2.6 4.9 3.0 5.6
Number of older siblings
alive 1940 1.6 3.3 2.1 3.7
Number of pairs 1019 907
Note: Calculations are for those for whom earn
variables, if one brother answered and the ot
brother. If both did not answer, both are se
whole tape, no responses were .5, 10, and 18 p
1967 occupation, respectively. For mother and f
of responses is used.
aThousands of dollars.
bMillions of dollars.
CAs recalled in 1974.
dCensus occupational status score.
the DZ pairs. The means of schooling, initial and later occupational status, and
earnings are nearly the same across twin type, although the variances differ by
up to 10 percent.13
The conclusion that MZs and DZs are random drawings from the same
population is further strengthened by a comparison of the simple correlations
given in Table 2.14 The left-hand portion of that table treats both brothers as
13 For samples of this size, the 5 percent level of significance in an F test is about 1.2.
14 See also the comparison below of our regression results with those based on Census
data.
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
452 I THE JOURNAL OF HUMAN RESOURCES
TABLE 2
INDIVIDUAL AND CROSS-SIB CORRELATIONS
Individuals Cross-Sib
S OC67 lnY S OC67 lnY
MZs
S 1 .54 .44 .76 .44 .40
OC67 1 .35 .43 .27
InY 1 .54
DZs
S 1 .51 .44 .54 .29 .29
OC67 1 .35 .20 .19
lnY 1 .30
individuals
consists of
asY' is the
earnings. T
ones for in
comparable
IV. REGRESSION ANALYSIS
In previous studies of the effects of education on earnings, it has not be
possible to control completely for the other determinants of earnings that migh
be correlated with schooling. With twins, however, it is possible to eliminate
genetic differences for the identical twins and common background for both
types by studying the within-pair differences in earnings.
To understand what can and what cannot be done with twins, it is
necessary to compare the estimates obtained when using individuals and withinpair
differences. As an aid in making this comparison, let us order individuals
within each pair randomly, for example, alphabetically, and denote the withinpair
difference by A. We can rewrite equation (1) as:
(la) AlnY = aS + bAG + cAN + Au
We can estimate both (1) and (la) using OLS and compare the
Denote the estimate from equation (1) as al and that from (la) a
standard methods, it can be shown that:
(2) plim(dl ) = a + [plim cov(S, bG + cN + u)] /plim var(S)
(3) plim(42) = a + [plim cov(AS, bAG + cAN + Au)] /plim va
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
Taubman 1 453
TABLE 3
INDIVIDUAL AND WITHIN-PAIR REGRESSIONS FOR In Y
(t-statistics in parentheses)
Based on individuals:
1. lnY= 9.019 + .0789S - .0084Age R2 =.201
(65.1) (32.2) (3.2)
2. lnY= 8.58 + .0795S R=.199
(262.5) (32.5)
MZ within-pair
3. AlnY= -0064+ .0270AS R2 .012
(.2) (3.6)
DZ within-pair
4. AlnY = -0090 + .0594AS R2 = .069
(.3) (8.2)
As is well known, (2) yields biased estimates if plim
nonzero, which is generally thought to be the case fo
model.
For MZ twins AG is zero. Making the usual assump
related with AS, plim a2 will be unbiased provided eit
correlated with AS. The first condition means that the dif
environments have no direct effect on earnings, thoug
ing. The second condition means that the differen
determine earnings are not correlated with schooling.
because of genetics or common environment, 02 will b
For DZ pairs, AG is not zero and a2 will not be con
bAG) is not zero. Let us assume that the sibling envir
the same across twin types.15 Then we can determ
controlled for by comparing MZ and DZ within equat
note that the bias in a2 based on DZ data can be larger
both the numerator and denominator in the plim expr
from levels to within-pair differences.
Much recent empirical work on earnings has employ
form such as:
(4) InY = d + aS + e(Age - S) = d + (a - e)S + e Age
We choose to use the same form partly to maintain comparability and partly
because a variety of tests suggested that the semilog form was statistically better
than double log or linear forms.
15 The same test can be used if the following somewhat weaker conditions prevail. Let N
be divided into common and specific components as Nij = mi + hip, where i indicates
family and j sib. Then AN = Ahij. As long as Ah is uncorrelated with AS, the test is
valid.
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
454 I THE JOURNAL OF HUMAN RESOURCES
Table 3 contains four regressions for 1973 earnings. The first two
estimated from observations on individuals; the third and fourth are esti
from within-pair differences. Since the men in the sample range in age fr
to 55 in 1973, it is not surprising that age has a small and even neg
coefficient and that the education coefficient is barely changed when
included. In both equations (1) and (2), in this table, the coefficient
education is about .08 and highly significant. In studies using Census data
semilog form, Mincer [8, p. 92] also estimates the coefficient on education
about .08.
Next let us examine the within-pair regressions. In the DZ within-pair
version, our estimate of a is .059, while in the MZ within-pair version it is .027.
Mincer [8] and others have shown that the difference in earnings by education
level to be greatest around age 50. Thus, the MZ within-pair results suggest that
years of schooling per se have only a small, though statistically significant, effect
on earnings.
We have used the analysis of covariance (Chow test) to test the null
hypothesis that the two within-pair equations are the same. With a calculated F
of about 4.8 and with 2 and 1932 degrees of freedom, we can reject this null
hypothesis at the 5 percent level. Thus, in this sample it appears necessary to
control for genetic endowments when estimating the returns to schooling. Since
the DZ within-pair estimates are less than the individual estimates, it is also
necessary to control for family environment.
The combined bias from not controlling for genetics and common environment
can be estimated as .051/.079, or about 65 percent. (Below we consider
the possible importance of measurement error on this estimate.) Prior studies
have tried to estimate the bias by including measures of IQ and/or family
background. To the best of my knowledge, no one has found such a large bias.
Of course, as Griliches [5] has pointed out, the bias depends on the effect of the
omitted variable on earnings and the slope coefficient of the omitted variable on
schooling. The latter coefficient may vary by cohort.16 But the NBER-TH
sample, which has yielded some of the larger estimates of the bias from omitting
IQ and family background, is drawn from about the same cohort.17 Yet the bias
in that sample does not approach two-thirds. Thus, it appears that regressions
that control for ability by using IQ, parental education and occupation, and
other often available measures will not adequately control for ability.
While we do not currently have measures of IQ available, we do have a
fairly wide list of measures that often have been used as proxies for environment
and/or genetic endowments.18 We have reestimated equation (2) including many
16 Taubman and Wales [13] demonstrate that the IQ education slope has changed by
cohort for those with at least a high school diploma.
17 See, for example, [14].
18 The information available will not let us determine if a variable is a proxy for
environment and/or genetic endowments.
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
Taubman 1 455
of these variables. The resulting equation is:
(5) lnY = 8.98 + 069S- .0074Age + .0052(SMother) + .0059(S Father)
(63.9)(25.8) (2.9) (1.8) (2.2)
+ .00066(OC Father) - .0078Sibs + .0SlCath + .321J
(2.0) (2.3) (2.8) (8.8)
+ .0068(Born South) - .0231 (Rised r
(.3) (1.4)
The education of the mother and f
nomic) status are generally included in s
in family income, child-rearing pract
offspring.19 In our equation the var
significant and that for mother's educat
findings are in rough accord with thos
Duncan [3] .20 Number of siblings (ali
included primarily to represent the redu
each child as family size increases. As w
the coefficient is negative and significa
primarily to represent different family
may be associated with various religious
the 1920s and 1930s. We find that Je
more than Protestants and others. Both
accord with those in the NBER-TH samp
1917 and 1924 (see [11]). Dummies for
having been born in the South are inclu
differences. These variables may also ca
dummies are insignificant here, though
of schooling.
Our primary interest in including thes
genetics is to determine their effect
include these proxies, this coefficient d
Thus, as suggested earlier, these vari
variables for which they are intended to
John Bishop and several other econom
previous statements on bias must be qua
is well known, if our true variable i
measurement error of v, then the bias fr
uncorrelated with Y and with s) depend
equations, the bias depends on ov
19 See [11, ch. 1] for a further discussion. The occupational status measure is
Duncan's SES score.
20 See, however, Sewell and Hauser [9] for a different picture.
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
456 I THE JOURNAL OF HUMAN RESOURCES
TABLE 4
ESTIMATES OF THE TRUE EFFECT OF SCHOOLING ON InY,
FOR VARIOUS VALUES OF MEASUREMENT ERROR VARIANCE
From MZ From
Within-Pair Equation Individual Equation True Bias
If:
a2= .05a2 .032 .084 62%
o2=.10a2 .048 .088 45%
a2 =.15a2 .070 .091 23%
U2 =.20oa .121 .096 -26%
greater than 21/a2 since the bro
brother's measurement error ei
wrongly reported numbers or w
is expanded to include quality (
Welch has suggested.21
However, we can calculate the
the estimates from within and b
tion that each estimate would b
given in Table 4 assume that th
dent.22 As is evident in the tabl
schooling is no greater than 10
comparisons, the MZ within-pair
Thus far we have shown that
family environment when we
the next section we will attem
environment to earnings.
V. ANALYSIS OF VARIANCE
Suppose we assume that schooling also is a function of genetic endowments and
environment (though not necessarily the same G and N as in the earnings
equation (1)). Then we can write a reduced-form equation for earnings in terms
21 aos = 2a - 2ass', where prime indicates a sib's brother.
22 If measurement error arises because of differences in quality of schooling, which
presumably are correlated across brothers, the MZ within-pair biases will be smaller
than those shown in Table 4.
23 See Bishop [2]. The Census-CPS match may overstate measurement error since in the
Census in some cases wives provide data for their husbands. This source of error is not
found in our study.
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
Taubman 1 457
of a single genetic index, which we will continue to label G, and a single
environmental index, which will include the random errors and which will still
be labeled N. Since both G and N are unobserved, we can dimension them such
that their coefficients are each 1 in our earnings equation. Let Y stand for the
log of earnings. Then our reduced-form equation is:24
(6) Y=G+N
Using (6), we can write
~(7) oa4 = oG +
Let a brother be denoted
calculate the cross-sib cova
(8) Oay' = oGG' + 2oGN' + OkN'
For MZ twins, aGG' = ao and OGN' = OGN because G is the same for the s
Thus, for MZ twins
(9) U}2 -Y = NN()y Uyy UrN - (N'
The term oNN', which is the covariance of the brothers' environment, can be
thought of as their common environment. This common environment is broad
than family environment to the extent that neighborhood and peer group effec
occur, but less than family environment to the extent that parents treat the
twins differently.25
For many purposes we would wish to include the neighborhood and peer
group effects with the family effect. Thus, we can treat the variance in (9) as
upper bound to the contribution of nonfamily environment to the variance o
the log of Y. We can estimate (o2 - ory,)/o2 from Table 2 by taking th
complement of the cross-sib correlation for an lnY. Using this approach,we fin
that noncommon environment accounts for at most 45 percent of the to
variance of lnY. For comparison, noncommon environment at most accounts fo
60 percent of occupational status and 25 percent of years of schooling.26
It is also possible to calculate the cross-sib covariance for DZ brothers. Th
expression is, of course:
(10) oyy' = oGG' + 2CGN' + uNN'
To be able to identify any additional parameters, it is necessary to mak
24 In (6) we have adapted a linear and additive form for InY. Taubman [11] has siho
that this form is acceptable for In earnings but not for earnings, though the results th
follow are the same for In Y and Y.
25 The twin literature on other subjects emphasizes the different treatment of twins.
26 When we redid the occupational status calculations after excluding the no-responses,
the results were hardly changed.
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
458 I THE JOURNAL OF HUMAN RESOURCES
number of assumptions which may or may not be valid. One set of assumpt
proposed in Taubman [11] is that all genetic effects are additive, ther
random mating, one brother's specific environment is not correlated with
sib's genes, and NNv' = TN' . Even with these strong assumptions, the mode
underidentified. Restricting the analysis to the cases where all the estimate
variances and covariances are nonnegative and less than a, Taubman [1
indicates that genetics and common environment would account for from 5
50 percent and from 18 to 5 percent of the total variance in Y, respectively.
Other assumptions made to identify the model could give different results, b
our assumptions seem at least as plausible as any others.28
In assessing the results in this section or more definitive ones based on m
complex techniques, it is important to keep in mind certain limitations
caveats. First, the results may not be robust to changes in model specificat
Second, the twin data yield estimates that are applicable to the population o
which the twins are representative, which presumably is white males b
between 1917 and 1927 who served in the military.29 In World War II, m
were rejected for service because of mental and physical defects that arose fr
a variety of genetic and environmental causes. It is not clear if our res
generalize to all white males about 50 years old in 1973, because it is not kn
if rejections from the service were more severe with respect to genetic end
ments or environment.
There are several other caveats that can best be understood if we rewrite
equation (6) with the coefficients on G and N treated as prices and not
standardized to equal 1. That is:
(6b) Yt = PGtGt + PNtNt
In order for our results to apply to white males at all times
necessary that for all t, PGt = P, PNt = PN, Gt = a, ot
27 There is also a covariance between G and N.
28 Behrman, Taubman, and Wales [1] have recently shown that embedding a twin model
in a latent variable framework may allow us to distinguish between various models and
to identify the separate contributions of genetics and family environment.
29 This assumes that there are no biases arising from differential response rates. Average
earnings in our sample exceed 1971 earnings of veterans of the same age by $7,000,
which is more than can be accounted for by inflation. The education level of our twins
also exceeds that of veterans. Moreover, in this sample, the average earnings of those
with an eighth grade education or less exceeds that of high school graduates for both
MZ and DZ pairs. In [ 11 ] it is shown that for our responders, the allocation of variance
is approximately the same for occupational status and earnings. But we have information
on occupational status for many of our nonresponders. The results for occupational
status are comparable for responders and nonresponders, although genetics is
accorded more emphasis in the responder analysis.
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
Taubman 1 459
aGN. Certainly prices can change if supply or demand for any relevant skills
shifts. IfPGt and PNt do not change proportionately, the contributions of G and
N to oa will alter. Even with prices fixed, the distribution of environment can
change; for example, the distributions of schooling, of family size, and of size of
city of upbringing have altered during this century. Similarly, the distribution of
genetic endowments can still be changing because of "recent" successful mutations,
changes in environment that alter the advantage of particular gene combinations,
or migration. These three reasons suggest that it would be desirable to
study twins from other cohorts.
VI. CONCLUSIONS
In this paper we have used a new and very rich data source, the NAS-NRC tw
sample, to examine the relationship of schooling and some of the effects
genetic endowments and family environment to earnings.
In our regression analysis, we found that our sample, though it is not a
random drawing of the population, yields results quite similar to those obtained
with Census data when we do not control for genetics and family environment
However, once we control for G and N, we find that the coefficient of schoolin
declines by two-thirds. It seems likely that only a modest portion of th
decrease is due to measurement error. Not controlling for genetics appears to
account for some of the decline, but part is probably also attributable to not
controlling for family environment.
In other studies people have tried to control for genetic endowments and
family environment by including a variety of proxy variables. When we use su
proxies as parental education, father's occupation, number of sibs, religio
region of birth, and rural upbringing, we find many of the variables significan
but that the education coefficient declines by only 12 percent. Thus, this list o
proxies is not complete enough to represent fully genetic endowments an
family environment.
Of course, G and N could influence earnings even if neither caused a bias on
the education coefficient. We calculate that noncommon environment, which
the complement of common environment and genetic endowments, accounts fo
about half of the variance in log Y around age 50. We have tried to separate t
other half into its genetic and family environment components, but have had
only limited success.
In summary, we can conclude that it is very important to control fo
genetics and family environment when studying the effects of schooling on
earnings, that the types of proxies for G and N generally available in Census
studies are far from adequate, and that a large proportion of the variance in
earnings at age 50 is accounted for by a combination of family environment an
genetic endowments.
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
460 I THE JOURNAL OF HUMAN RESOURCES
APPENDIX: THE SAMPLE
One of the major and recognized problems in earlier medical and psychological
research based on twins was that the sample of twins studied was not a random
drawing of the population of twins.30 Partly for this reason and partly to have
available a large enough group to study relatively rare diseases, a group of
geneticists and medical researchers decided to assemble a "random" sample of
white male twins. The sample construction and techniques are described in
Jablon and others [7], from which the following quotation is taken:
In 1955 experiments were initiated to explore methods of identifying
twins who served in the Armed Forces during World War II. The method
settled on was to obtain from the various state and city vital statistics
offices in the U.S. copies of the birth records of all white male twins born
in the years 1917-1927 and to match the names thus obtained against the
VA Master Index (VAMI) to determine which twins survived with both
entering military service. About 99% of all World War II veterans are
represented in VAMI.
Cooperation of 42 vital statistics offices was obtained (all of the
continental U.S. except Arizona, Connecticut, Georgia, Maine, Missouri,
New Orleans, Utah, and Vermont). Over 54,000 eligible pairs were found
by the participating offices, and, of these, 16,000 pairs were identified by
the VA as both having served in the armed services. For 15,000 pairs, one
member only was identified as a veteran, and for 23,000 pairs neither was
identified. Thus, 108,000 names were searched against the VAMI, of
which 47,000, or 43.5% were matched.
It is not possible to tell just why the proportion of matches was so
low. For a white male cohort born in 1920, about 86% survived in 1942.
About 80 percent of the survivors served in the military forces in World
War II, so that we might have expected to match about 69% rather than
43.5%. Possible reasons for the discrepancy include higher mortality of
twins than in singletons born in the same year, higher rates of rejection for
physical disability, and failures to match correctly at VAMI because of
changes in name or inaccurate birth dates shown on the VAMI index card.
It must be realized that the VAMI file is ordinarily searched only when a
military identifying number is known, thus assuring correct identification.
The file clerks, when searching for twins, had to rely on name and date of
birth, and there probably were many failures to match men for whom
cards were in fact on file.31
We mailed our survey on April 15, 1974, to 12,500 pairs for whom we had
recent addresses and who had cooperated with recent studies. As of August, we
30 For example, if the twins were selected because one had a particular illness, doctors are
likely to be more thorough in examining the other brother for the illness.
31 I have been told that inaccuracies in the VAMI index are no longer considered a major
reason for the low match rate. Infant mortality was much higher for twins than for
single births in the relevant time period. See Woodworth [15].
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms
Taubman 1 461
had 6,600 replies out of a possible 12,000 people with up-to-date addresses, even
though on the third mailing we did not try to contact the nearly 6,500 people
where neither brother had previously responded. The 6,600 replies contain
2,468 matched pairs and 1,600 unmatched individuals.
REFERENCES
1. J. Behrman, Paul Taubman, and Terence Wales. "Controlling for
Measuring the Effects of Genetics and Family Environment in Equat
for Schooling and Labor Market Success." University of Pennsylv
1976, mimeo.
2. John Bishop. "Biases in Measurement of the Productivity Benefits of
Human Capital Measurement." University of Wisconsin, 1974, mimeo.
3. Peter M. Blau and Otis Dudley Duncan. The American Occupational Structure.
New York: John Wiley and Sons, Inc., 1967.
4. L. Cavalli-Sforza and W. Bodmer. The Genetics of Human Populations. San
Francisco: W.E. Freeman and Co., 1971.
5. Zvi Griliches. "Estimating the Returns to Schooling-Some Econometric
Problems." Econometrica (1976).
6. Zvi Griliches and William Mason. "Education, Income and Ability." Journal
of Political Economy (May/June 1972 supp.).
7. S. Jablon and others. "The NAS-NRC Twin Panel: Methods of Construction
of the Panel, Zygosity Diagnosis and Proposed Use." American Journal
of Human Genetics (1967).
8. Jacob Mincer. Schooling, Experience and Earnings. New York: Columbia
University Press, 1974.
9. William Sewell and Robert Hauser. Education, Occupation and Earnings:
Achievement in the Early Career. New York: Academic Press, 1975.
10. S. Stauffer and others. The American Soldier, vol. 4. Princeton, N.J.:
Princeton University Press, 1950.
11. Paul Taubman. "The Determinants of Earnings: Genetics, Family and
Other Environments." American Economic Review (forthcoming).
12. _ . Sources of Inequality of Earnings. Amsterdam: North Holland Publishing
Co., 1975.
13. Paul Taubman and Terence Wales. Mental Ability and Higher Educational
Attainment in the 20th Century. New York: McGraw-Hill Book Co., 1972.
14. ,. Higher Education and Earnings. New York: McGraw-Hill Book Co.,
1974.
15. R. S. Woodworth. Heredity and Environment. New York: Social Science
Research Council, 1941.
This content downloaded from 147.251.185.127 on Tue, 20 Feb 2018 12:15:20 UTC
All use subject to http://about.jstor.org/terms