Griliches (1976) - Wages of Very Young Men
Griliches (1976) - Wages of Very Young Men
Griliches (1976) - Wages of Very Young Men
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
The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access to
Journal of Political Economy
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
Wages of Very Young Men
Zvi Griliches
Harvard University
I. Introduction
Over the past decade there has been much interest in and a large amount
of work done on estimating the economic returns to formal schooling and
on trying to untangle such returns from the contributions of native
ability, family background, discrimination, and nepotism. The rather
large literature that has emerged has been discussed and surveyed by a
number of authors (Jencks 1972; Welch 1975; Griliches 1975a; Rosen
1975; among others). Different methodologies and different sets of data
have produced very little agreement. It is not the purpose of this paper to
review and revive all the debates again. Instead, it will attempt to repli-
cate the results of an earlier study of "Education, Income, and Ability"
(Griliches and Mason 1972) on a new set of data, the National Longi-
tudinal Survey of Young Men, focusing on the estimation of the economic
returns to schooling in the presence of individual differences in ability.
The NLS data base is of interest because it is the most representative
data set combining information on earnings, schooling, and measures of
ability. It contains data on two measures of ability: IQ scores collected
from the high schools attended by the respondents and scores on a test of
"knowledge of the world of work" (KWW) administered at the time of
the initial interview in 1966. Data were also collected on parental back-
ground, wage rates (rather than just total income or earnings), and work
experience. These data are also of interest because of the availability of
repeated observations on the same individuals and the ability to match
family members across surveys. I shall not pursue, however, the last two
This is an abridged and extensively revised version of an earlier paper (1974). This work
has been supported by grants from the National Institute of Education (NE-6-00-3-020)
and NSF (SOC73-05374-AO1). I am indebted to Bronwyn Hall, Ruth Helpman, and
Stephen Messner for research assistance.
[Journal of Political Economy, 1976, vol. 84, no. 4, pt. 2]
(? 1976 by The University of Chicago. All rights reserved.
S69
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
S70 JOURNAL OF POLITICAL ECONOMY
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
WAGES OF VERY YOUNG MEN S71
and to allow for the possibility that IQ may not be a perfect measure of
the relevant ability concept, we define
Assuming that either IQor KWW are errorless measures of the relevant
concepts amounts to assuming that e2 and e4 have zero variance. If one
were to use KWW as a measure of ability, one would get .(substituting
[2] in [1] and ignoring constant terms)
2 Actually, KWW depends on schooling in 1966 (SC66) rather than on SC69. But
in the not-enrolled portion of our sample, SC66 and SC69 are too closely intercorrelated
to distinguish between them. In a separate study using expectational variables for income
and schooling, this distinction is made (see Griliches 1975b).
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
S72 JOURNAL OF POLITICAL ECONOMY
Thus, unless schooling does not affect KWW (/2 = 0), introducing
KWW into the wage equation will eliminate some of the "ability bias" in
the schooling coefficient but in its turn introduce another downward
bias (-f62/72) into the estimated schooling coefficient. If KWW i
subject to error of measurement (o2 # 0), the above conclusion will be
attenuated but will retain its qualitative implications. If both IQ and
KWW are subject to independent measurement error, equation (1) can
be estimated using instrumental variable methods. If one considers the
possibility that schooling itself may be determined on the basis of expected
income and that it too may be measured with error, this would imply a
correlation of e1 with S and suggest the use of instrumental variables for
it too.
Note that I have not discussed family background variables explicitly.
They are contained in the definition of X and allowed to be correlated with
the true A. In what follows I shall assume, for the most part, that family
background variables such as mother's education or father's occupation
do not affect wages directly (i.e., do not enter eq. [1] separately) but only
indirectly via their effects on ability and schooling. This is a testable
restriction on the model. For some purposes, however, one is more inter-
ested in the "total" effect of family background or ability on wages and
then in their "net" effects holding schooling constant. Such "total" effects
can be discerned from the reduced-form version of (1), derived by sub-
stituting (3) into it:
The data used in this study are based on a national sample of the civilian
noninstitutional population of males who were 14-24 years old in 1966.4
Blacks were oversampled in a 3:1 ratio. The original sample consisted of
5,225 individuals of whom 3,734 were white. By 1969 about 23 percent of
the original sample was lost, 13 percent of it only temporarily (to the
army). Data are currently available from the 1966-69 surveys of these
3For a more detailed discussion and modeling of such a system of equations, see
Griliches (1975b) and Chamberlain (1976).
4 See U.S. Department of Labor (1970-74, vols. 1-4) for more details on the sample.
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
WAGES OF VERY YOUNG MEN
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
JOURNAL OF POLITICAL ECONOMY
TABLE 1
(3.1) (2.8)
XBT ...................... ... ... 0.70 0.72
(0.27) (0.18)
SMSA ..................... ... ... 0.61 0.65
RNS ....................... 0.32 0.33 0.41 0.33
NOTE.-LW69 = logarithm of hourly earnings (in cents) on the current or last job in 1969; KWW =
score on the "knowledge of the world of work" test administered in 1966; IQ = score on IQ-type tests
collected from the high school last attended by the respondent; FOMY14 = occupation of father or head of
household when respondent was 14 scaled by the median earnings of all U.S. males in this occupation in 1959;
culture = index based on the availability of newspapers, magazines, and library cards in the respondent's
home; EXP69 = postschool work experience estimated on the basis of the work record (in weeks) since
1966 and the date of first job after school and the date stopped school (in years; truncated at age 14 if re-
spondent started working earlier); XBT = e- 0. EXP69; SMSA = respondent in SMSA in 1969; and
RNS = respondent in South in 1969.
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
WAGES OF VERY YOUNG MEN
for almost all of the sample. It is not an "intelligence test" but, rather, an
"occupational information test":
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
S76 JOURNAL OF POLITICAL ECONOMY
two ways. For the major variables of interest, such as wages, schooling,
and IQ, we restricted the sample only to those for whom complete and
reasonable data were available. For variables of subsidiary interest, such
as parental background, we imputed a mean value to the missing ob-
servations and added dummy variables corresponding to each set of
missing observations for each of the independent variables.9
The youngness of the sample requires much more careful attention to
the experience variable, since there is much on-the-job training and
search at this age and, hence, the wages of different youths of the same
age cannot be taken as reflecting the same amount of human capital if
they differ in their labor market experience. In this study, we construct
the experience variable directly from the work history of the individual;
hence, it is not tautologically equal to age minus schooling.'0 Since it is
likely that on-the-job training declines with time, the functional form
chosen to represent this variable (borrowed from Mincer [1974]) reflects
this. I have also used a more general functional form (cubic) for the
experience variable, but because later on we will be examining the pos-
sibility that schooling and hence also experience should be treated as
endogenous variables in a more complete achievement model, I will
concentrate attention on a single summary measure of it (XBT).
9 This corresponds to a procedure that estimates jointly the mean missing value for
each of these variables. This procedure, however, is neither consistent nor fully efficient.
Because of our special interest in the IQvariable, we also tried a more complete treatment
of the missing-value problem, using a regression of observed IQ scores on all the relevant
variables (including schooling and other tests) to estimate the missing values individually.
The results were not strikingly different and are not reproduced here (see Griliches
[1974] for details). See Dagenais (1973) for a discussion of the econometrics of such
"extrapolation" procedures and references to the literature on missing observations.
IO Age and schooling account for only 80 percent of the variance in the XBT variable.
11 Age is not statistically "significant" when added on top of the experience variables.
This differs from the results reported in the earlier version of this paper, which were based
on a cruder measure of experience.
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
WAGES OF VERY YOUNG MEN S77
TABLE 2
COEFFICIENTS
(AND t-RATIOS) OF:
OTHER VARIABLES R
EQUATION NUMBER SC69 KWW IQ IN EQUATIONS (SEE)
A: N = 2,062:
(Al) ......... 0.034 ... ... CL, age .356
(10.6) (0.342)
(A2) .. ..... 0.022 0.0070 ... CL, age .367
(6.1) (6.1) (0.339)
(A3) ........ 0.063 ... ... CL, cubic in EXP69, AFEX .383
(18.9) (0.335)
(A4) . ...... 0.050 0.0059 ... CL, cubic in EXP69, AFEX .391
(12.2) (5.3) (0.333)
(A5) ... ..... 0.051 0.0060 ... CL, XBT, AFEX .385
(12.5) (5.3) (0.334)
(A6) .. ..... 0.059 ... 0.0017 CL, XBT, AFEX, DIQ .380
(15.5) (2.6) (0.336)
(A7) ... ..... 0.062 ... ... CL, XBT, AFEX, CS .438
(17.0) (0.320)
(A8) . ...... 0.056 ... 0.0020 CL, XBT, AFEX, CS .442
(15.3) (3.3) (0.319)
(A9) ......... 0.056 ... 0.0016 CL, XBT, AFEX, DIQ,BKG .382
(13.6) (2.4) (0.336)
B: N = 1,362:
(Bi) ......... 0.065 ... ... CL, XBT, AFEX .309
(13.2) (0.332)
(B2) ...... 0.053 0.0054 ... CL, XBT, AFEX .316
(9.1) (3.8) (0.330)
(B3) ......... 0.059 ... 0.0019 CL, XBT, AFEX .313
(10.7) (2.8) (0.331)
(B4) ......... 0.058 ... 0.0026 CL, XBT, AFEX, black, CS .382
(11.0) (3.2) (0.315)
(B15) ......... 0.053 ... 0.0020 CL, XBT, AFEX, CS, BKG .386
(9.5) (2.8) (0.314)
NOTE.-Variables and variable sets: SC69 = schooling completed in 1969 (in years); KWW = score
on test of "knowledge of world of work"; IQ score on an IQ-type test; DIQ = dummy variables for
missing IQ score; XBT = e - o. I * EXP69 (EXP69 = cumulated work experience in 1969, in years); AFEX =
service in the armed forces (in years); DIQ = IQ missing; CL = current location: RNS, BRNS, SMSA
(RNS, region now South; BRNS, black in South now; SMSA, in SMSA in 1969); CS = current situation:
health 68, union, MRT (health 68, health impaired in 1968; union, member of a union in 1969; MRT,
married); BKG = background: black, siblings, culture, MED, DME, FOMY14, DF014 (siblings, number
of siblings; culture, index of home culture [library card, magazines, etc.]; MED, mother's education; DME,
MED missing; FOMY14, father's occupation when R14, scaled by the 1959 media earnings of males in the
particular occupation; DF014, FOMY14 missing); SEE = standard error of estimate (estimated standard
deviation of the residuals).
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
S78 JOURNAL OF POLITICAL ECONOMY
and [B3]). IQ-based estimates of the "ability bias" are only on the order
of 10 percent (holding experience constant). These estimates are robust
to the inclusion of additional variables, such as marital or union status of
the respondent.
4. The estimated schooling coefficients move between about .05 and
.06, somewhat lower than the .06-.08 reported by Mincer (1974) and
others but not out of line with other estimates limited to younger ages
and the earlier portion of the age-earnings profile. They are statistically
very significant and quite robust to the introduction or deletion of other
variables (except to the choice of age vs. experience).
5. The contribution of the ability measures (IQ or KWW) to the fit
of the various equations is miniscule. Their introduction leads to some
realignment in the estimated coefficients of schooling, but rarely does it
lead to a substantive reduction in the unexplained variation of wage
rates (though given our large samples, the actual reduction is often
"statistically" significant). The standard deviation of the residuals is
rarely changed by more than in the third decimal place, and their vari-
ance is reduced by less than 2 percent. Whatever their merit in reinter-
preting the role of schooling, the available ability measures do not
noticeably improve our ability to explain the observed dispersion in wage
rates or earnings, at least in this age group.12
6. Family background variables are not statistically significant as a
group when added on top of the schooling and ability variables. (Compare
the estimated standard errors in [A6] and [A7] or [B4] and [B5].) This
is consistent with our assumption that the background variables work
primarily via schooling and measured ability and have no substantive
direct contribution of their own in the wage equation.
7. There is little difference in the results between the larger sample and
the "real IQ only" subset. Apparently the selectivity of the missing IQ
values (concentrated among low-schooling individuals and blacks) does
not bias our results seriously.
Table 3 presents all the coefficients of one of the better fitting equations
(B4) and illustrates some of the additional results of this study. Note that
the three most "significant" variables are schooling, experience, and
union membership. The first years of experience appear to be quite valu-
able, roughly on the order of the contribution of a year of schooling,
starting at .06 per year at zero experience and dropping slowly to .03
after 7 years of experience and to only .015 per additional year after 14
years of experience.
While the number of those with military service is relatively small in
this sample (about 18 percent among those not enrolled in school) and the
estimated effect therefore not very precise, it appears that a year of mil-
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
WAGES OF VERY YOUNG MEN S79
TABLE 3
itary service has a significantly smaller effect on wages (1.6 percent) than
a year of work experience in the civilian sector. This differs from but is
not greatly inconsistent with the results of Griliches and Mason (1972),
who found no effect of differences in the length of military service on the
income of veterans in a sample where everybody had served in the armed
forces.
Black-white wage and income differentials are not a major focus of our
analysis. They are dealt with at some length in the current work of
Freeman (1974), and I shall not poach extensively on his territory. The
following major facts do emerge though from our analysis: with slightly
over a quarter (27 percent) of our sample black, we cannot really detect
a statistically significant wage differential except for blacks still located
in the South. Holding schooling completed constant, we find no evidence
of any significant wage discrimination against young black males outside
of the South.'3 In the South there persists about a 15 percent wage
differential in favor of whites.
In general, there are rather strong regional and city size effects in the
sample. Being in a metropolitan area adds about 11 percent to one's
wage rate. Being in the South currently subtracts about 6 percent, while
not having been healthy in 1968 leads to a 6 percent lower wage rate and
up to 20 percent lower earnings in 1969. One of the "strongest" variables
in these equations (in terms of t-ratios, fi coefficients, or partial correla-
tions) is being married, which is associated with both a higher wage rate
and expanded labor force participation. Married men have about a 10
percent higher wage rate per hour and earn about a third more per year
13 These estimates are based on equations estimated across the whole sample. When
estimated separately for blacks and nonblacks, the resulting wage and income equations
appear quite similar except that blacks seem to gain more from formal schooling and
less from work experience or the sheer passage of time (age) than whites do in this same
age range.
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
S8o JOURNAL OF POLITICAL ECONOMY
V. Reduced-Form Estimates
In the debates about the relative role of schooling and ability in deter-
mining the ultimate economic success of an individual, the net contribution
of ability (holding schooling constant) is often confused with its total
contribution (including that via schooling). In econometric terminology,
sometimes one is interested in the parameters of the "structural" equation
while at other times one is interested in those of the "reduced form," the
equation in which all of the other endogenous (casually subsequent)
variables have been solved out to show the total contribution of the
remaining exogenous variables.
Table 4 presents a number of estimates of such reduced- (and semi-
reduced-) form equations for the major variables of interest and for
several subsets of our sample. The first three equations deal with schooling.
Since a large fraction of our population have not yet finished their
schooling, we introduce a new variable expected ultimate level of school-
ing (EXSC) and compare it with actual schooling attained in 1969. By
and large, the results are similar: parental background, region of origin,
and IQ account for about a third of the variance of schooling. Mother's
education appears to be a somewhat stronger variable than father's
occupation but not by a great margin. Later ability, as measured by the
KWW test, is affected about equally by schooling and by early ability.
Number of siblings has a significant negative effect on late ability and on
achieved schooling. For given parental status and region of origin, black
youngsters do score lower on both early (IQ) and later (KWW) tests, on
the order of two-thirds and one-third of a standard deviation for IQ and
KWW (holding IQ scores constant), respectively. In spite of this, they
have a higher schooling attainment (on the order of half a year) than
white young men with equal parental background and test scores.
With regard to the estimated log wage equations, several important
points emerge: (1) There is a sizable net effect of schooling on wages and
income net of family background and measured IQ. (2) This effect is
significantly larger, by 50 percent or so, when we adjust for the lower
work experience of those with more schooling. (3) There are significant
though small negative effects of being black on wage rates (net of family
background, IQ, and schooling), particularly for those of southern origin
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
_s 4 cs 4 Uc Ln L Ln) o c-4 dt U o cs0 Ci C, - iC CO O-C
CO - C) -: Q0> -5 Q -t C) U) 1- 1r)C-) G -7-)C
C; CA o cs m ce Ln dz o: cs c c cr) c~r coz cl> cl)
En CU n o
C1 C13 C3 C3 C3 C1 O
;~~~~~C - - 4; - -
? z U + S + + + + + + ?
ryO cn OOOo
? ? d^^ O t O O O O ~~~~~O O .
C- ( e t . boboto SSSV
VO tr0 <t t ~~~~~~<-)< t t U
O P; v P; z m m m^; m m m ._ n~~~~~~7
X MS X X X X Mt X X co ) C<- C)
C3 * - * l * c * O - * O- r
C; s L) C; O l C: mOOss C;s Ol C; O~ 3 -2 b
dz (4~~~~~~~
m o cs s Ln o ce ?: > m ? UO~~~
cl dzi ^6 Cb ^ ^mOoO
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
S82 JOURNAL OF POLITICAL ECONOMY
who stayed in the South and especially in the larger, economically poorer
sample.'4 (4) Parental background variables have little direct effect on
wages once schooling and IQ are allowed for.
One way to look at these results is to consider what they imply for two
youngsters who are 1 standard deviation apart on each of the listed family
background variables and IQ."5 Equations (1) and (5) would predict
that they would find themselves, other things equal, about 0.75 and 0.25
of a standard deviation apart on schooling and wages, respectively, imply-
ing a rather strong regression toward the mean. Now, if a youngster with
lower family background and IQ managed somehow to acquire 4 extra
years of schooling (e.g., went on to and completed college), which would
equal an additional 1.5 standard deviation units of schooling, it would
more than wipe out his original handicap (using the schooling coefficients
of eq. [6]).16 If, in fact, he had an equal IQto start with, he would need
only 2 more years of schooling to compensate him for his lower social
class start.
Thus, while schooling by itself does not appear to be a variable that
accounts for a great deal of the observed variance in wages or income, the
model does predict that in the United States in the 1960s additional
schooling could be used to overcome social class handicaps. Compen-
satory education would not eliminate much of the observed income
inequality, the residual standard errors are on the order of a third to a
half in the wage and earnings equations, respectively, and more than half
of the observed variance is left unaccounted for by our variables, but it
could be used to compensate for and eliminate some of the systematic
sources of the observed differences in wages and earnings.
14 They are larger for earnings than for wages, but much of that is due to the lower
work experience of blacks. This disagrees with Adams and Nestel (1973), who do not
find a "Southern rural origin" effect for blacks but do find a negative effect of having
grown up black in a non-South city. We tried such southern rural and northern city
origin dummies but found them not particularly statistically significant though they did
occasionally have sizable coefficients. In the various versions of our model we do get,
from time to time, sizable negative estimated effects of having grown up (residence at
14) black in the rural South but very little consistent additional effect of having done so
in a northern city. The difference in results may be due to our use of a later year, a longer
list of other included variables, and our control for experience. There is some indication
in our data that young black men of northern city origin worked 5-7 percent less in
1968-69 than those of southern rural origin. Since Adams and Nestel do not control for
migration status, their major effect might be interpreted as a premium on having moved
from the South to the metropolitan North and could be related to the selectivity of
migration.
15 Since the various family background and IQ scores are not highly intercorrelated,
being 1 standard deviation apart on each of these measures implies that they are quite
a bit farther apart on a more inclusive measure of the distribution of social and genetic
inheritance.
16 Since he would be predicted to have about 2 years less schooling than his more
fortunate counterpart, he would need 2 years more than was predicted for him to achieve
equal schooling and 2 additional years to wipe out the remaining family background and
IQ deficits.
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
WAGES OF VERY YOUNG MEN S83
One of the objections that can be raised against the results reported above
is the use of rather poor test scores as a proxy for true ability. Both the IQ
and KWW measures may be subject to significant errors, which could bias
their coefficients downward and the schooling coefficient upward. Since
we have two test scores, we can use one of them as an instrument for the
other (assuming that their errors are independent of each other) together
with the parental background variables to estimate the "true" ability
coefficient by instrumental variables (TSLS) methods. Such estimates are
presented in rows 1, 3, and 5 of table 5. They indicate that allowing for
errors in the test measures raises their coefficients significantly while
reducing the schooling coefficient only by a bit more (about another 10
percent from a level of approximately .06). The results of equations (3)
and (5) compared with table 2, equations (B2) and (B3), imply that about
half of the independent variance of IQ (or about 30 percent of the total
IQvariance) and almost 60 percent of the independent variance of KWW
(or 35 percent of the total) are due to errors of measurement.
Given such high estimates of error variances in the tests, one is hard
pressed to maintain the assumption that such "errors" are uncorrelated
with schooling or with each other. These errors are computed by finding
that part of the test score variance that is correlated jointly with the other
test, parental background, schooling, and wages. The rest is then labeled
"error." If the rest is large, it is quite likely that it may be related to suc-
cess in school (e.g., test wiseness) even if it does not contribute directly to
earnings. This would imply that schooling cannot itself be treated as
independent of the disturbance in the wage equation once an erroneous
test measure is introduced instead of the "true" underlying ability dimen-
sion. It too must be treated as endogenous.
Moreover, since the schooling decision itself is made, at least in part, in
anticipation of economic returns, the piece of the disturbance in the wage
equation which is anticipated by the individual (even though unobserv-
able to the researcher) will affect the outcome of this decision. This may
result in a correlation between the unobserved variables in the wage
equation (the disturbance) and completed schooling, suggesting again the
use of simultaneous-equation methods in estimating the schooling
coefficient.' 8
1 See Griliches (1975b) and Chamberlain (1976) for more detailed discussion of the
issues raised in this section and for additional empirical results.
58 There are three reasons why schooling might be endogenous: (1) errors of measure
ment, (2) correlation between the disturbances in the income and schooling equations
(since schooling is optimized with respect to expected income), and (3) test wiseness:
the presence of another "ability" component in the test scores which is correlated across
tests and with schooling, inducing a correlation between the "errors" in these equations.
The procedure discussed in the text and the results presented in table 5 deal more or
less adequately with reasons (1) and (2). Reason (3) would prevent us from using one of
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
S84 JOURNAL OF POLITICAL ECONOMY
TABLE 5
Other Variables in
Equations and
Additional
SC69 KWW IQ Instruments SEE
A: N = 2,062:
1. KWW
endogenous ... 0.046 0.0086 ... CL, XBT, AFEX; 0.335
(6.4) (2.9) instruments: IQ,
DIQ, BKG, DIQP
BKG, age, age squared
2. KWW, SC69,
and XBT
endogenous ... 0.076 0.0044 ... Same as above 0.339
(5.7) (0.1)
B: N= 1,362:
3. IQendogenous. 0.052 ... 0.0038 CL, XBT, AFEX; 0.332
(7.0) (2.4) instruments: KWW,
BKG, age, age squared
4. IQ, SC69,
XBT endogenous 0.096 ... -0.0016 CL, XBT, AFEX; 0.339
(5.7) (0.6) instruments: BKG,
age, age squared
5. KWW
endogenous ... 0.036 0.0130 ... CL, XBT, AFEX: 0.334
(3.8) (3.6) instruments: BKG, IQ,
age, age squared
6. KWW, SC69,
XBT endogenous 0.078 0.0027 ... CL, XBT, AFEX; 0.333
(2.7) (0.4) instruments: IQ,
BKG, age, age squared
the tests as an instrument for the other. The model is still identified but not very pre-
cisely. Chamberlain and I are analyzing a mole general two-factor model, encompassing
all three reasons, using data on brothers from the same surveys. Adding information on
family structure allows one to relax and test some of the more dubious assumptions (such
as no correlation of the test errors) imposed in this paper.
19 See Griliches (1975b) for a more detailed discussion of these results. A downward
bias might arise not only from errors of measurement but also if "ability" is taken to
reflect the initial level of human capital. Then it is easy to show that the optimized
level of schooling will be related negatively to such a measure or component of ability.
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms
WAGES OF VERY YOUNG MEN S85
The studies most comparable to the results of Sections III and IV of this paper
are those of Kohen (1973), Bulcock, Fagerlind, and Emanuelsson (1974), and
Sewell and Hauser (1974). Kohen estimates a similar model based on the 1966
data for these same young men. His results are not too different, though
substantially weaker. In his equations, IQ contributes a bit more and schooling a
bit less than in ours, but schooling still outperforms IQ by a significant margin.
The main drawback, in our view, of his study is the omission of the very important
age-experience variables and the elimination of the lower quarter of the sample
due to missing IQ scores in that portion. Also, we are observing the same
individuals 3 years later, giving them a bit more time to mature and find their
way in the labor marketplace.
Another important study of the labor market success of young men is that of
Sewell and Hauser and their associates (1974) in which the 1957 cohort of
Wisconsin high school seniors has been followed for 10 years. The results of their
studies are too rich to summarize here except to note that they put more emphasis
on the effects of parental income on the subsequent success of sons and that they
get higher effects for ability and lower effects for schooling than we do. The main
drawback of their study is the restricted range of their sample. By focusing on
one cohort of high school seniors, they cut out a third or more of the total schooling
distribution and reduce the estimated effect of schooling thereby. Also, they do
not allow for the differential experience of individuals in the labor market and
underestimate thereby the ultimate effect of schooling on income.20 As a result
of this restriction of range and the omission of experience and current location
variables, the overall fit of their models is quite poor (the R2s for the income
equations are on the order of .05-.07 vs. .4-.6 for our samples and models).
Bulcock et al. (1974) have been reanalyzing the extended and updated Malmo
(Sweden) data set which had been previously analyzed by Husen (1969),
Griliches (1970), Hause (1972), and de Wolff and Slijpe (1973), among others.
The data set originates with a 1938 sample of 10-year-olds in the city of Malmo.
These have been followed up through 1972. Data are available on childhood IQ,
parental background, adult IQ (army test scores), schooling completed, and
occupation and income in 1971, for approximately 500-700 males. The main
difference between their study and ours (and the Sewell-Hauser one) is that the
Malmo respondents were significantly older in 1971 (about 43 years old). This
leads to a much higher estimated effect of "occupation" on income than was
found either by us or by Sewell and Hauser. They find a larger effect of "late
ability" measures on current income than we do, but they too do not allow for
differences in the length of work experience, which tends to result in an under-
estimate of the net effect of schooling. In spite of this, their estimated total effect
of schooling, both via occupation and the adult ability measure, is quite high.
Our study is the first, however, to deal seriously with the problem of errors,
of both measurement and concept, in the available measures of ability and
schooling.
20 Given that they are following a single age cohort, they cannot really allow separately
for the effects of experience, since in their sample it is almost perfectly negatively colinear
with schooling.
This content downloaded from 192.76.8.3 on Mon, 10 Oct 2016 09:20:19 UTC
All use subject to http://about.jstor.org/terms