NBER WORKING PAPER SERIES

USING GEOGRAPHIC VARIATION IN

COLLEGE PROXIMITY TO ESTIMATE
THE RETURN TO SCHOOLING

David Card

Working Paper No. 4483

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
October, 1993

I am grateful to Charles Thomas and Norman Thurston for outstanding research assistance, and
to Michael Boozer, Alan Krueger, and Cecilia Rouse for comments. This research was funded
by the Industrial Relations Section of Princeton University. This paper is part of NBER's
research program in Labor Studies. Any opinions expressed are those of the author and not
those of the National Bureau of Economic Research.
 NBER Working Paper #4483
October 1993

USING GEOORAPHIC VARIATION IN

COLLEGE PROXIMITY TO ESTIMATE
THE RETURN TO SCHOOLING

ABSTRACT

A convincing analysis of the causal link between schooling and earnings requires an

exogenous source of variation in education outcomes. This paper explores the use of college

proximity as an exogenous determinant of schooling. Analysis of the NLS Young Men Cohort

reveals that men who grew up in local labor markets with a nearby college have significantly

higher education and earnings than other men. The education and earnings gains are concentrated

among men with poorly-educated parents -- men who would otherwise stop schooling at

relatively low levels. When college proximity is taken as an exogenous determinant of schooling

the implied instrumental variables estimates of the return to schooling are 25-60% higher than

conventional ordinary least squares estimates.

Since the effect of a nearby college on schooling attainment varies by family background

it is possible to test whether college proximity is a legitimately exogenous determinant of

schooling. The results affirm that marginal returns to education among children of less-educated

parents are as high and perhaps much higher than the rates of return estimated by conventional

methods.

David Card
Department of Economics
Princeton University
Princeton, N.J. 08544
and NBER
 One of the most important "facts" about the labor market is
that better-educated workers earn higher wages. Hundreds of
studies in virtually every country show earnings gains of 5-15
percent (or more) per additional year of schooling.1 Despite this
evidence, most analysts are reluctant to interpret the earnings gap
between more and less educated workers as a reliable estimate of
the economic return to schooling. Education levels are not
randomly assigned across the population; rather, individuals make
their own schooling choices. Depending on how these choices
are made, measured earnings differences between workers with
different levels of schooling may over-state or under-state the
"true" return to education. 2
A convincing analysis of the causal link between education and
earnings requires an exogenous source of variation in education
choices. In this paper I argue that geographic differences in the
accessibility of college are a potential source·of such exogenous
variation.3 Using data from the Young Men Cohort of the

1
Studies of the United States are reviewed in Rosen (1977),
and Willis (1986). A survey of international studies is presented
in Psacharopoulos (1985).
2
See Griliches (1977) for an overview of the issues.
3
A similar idea is used by Kane and Rouse (1993) to control
for the endogeneity of choice between a four-year college and a
two-v., ear colle .ge.
Mallar (1979) used proximity to a training site to estimate the
effect of the Job Corps program.
 2
National Longitudinal Survey I find that men who were raised in
local labor markets with a nearby 4-year college have
significantly higher levels of education and earnings. This
differential persists even after controlling for regional and family
background factors (including parental education and family
structure). The effects of a nearby college are largest for men
with the lowest predicted levels of schooling attainment,
suggesting that the presence of a local college lowers the costs
and/or raises the perceived benefits of education among children
with relatively poor family backgrounds.
When college proximity is taken as an exogenous determinant
of schooling the implied instrumental variables estimates of the
return to education are 25-60 percent higher than the
corresponding ordinary least squares estimates. Contrary to
widespread belief (e.g. Ehrenberg and Smith (1991, pp. 320-
322)) but consistent with a growing number of studies of
endogenous school choice, these findings suggest that the cross-
sectional earnings gap between more- and less-educated workers
may under-state the economic return to schooling for some groups
of workers. 4

4
See e.g. Angrist and Krueger (1991a), Ashenfelter and
Krueger (1992), Kane and Rouse (1993) and Butcher and Case
(1993). All four of these studies report instrumental variables
estimates of the return to schooling that exceed the conventional
ordinary least-squares estimate in the same dat;l set
 3
Since the effect of a nearby college on schooling attainment
varies with family background it is possible to test whether-
college proximity is a legitimately exogenous determinant of
schooling -- i.e., whether growing up near a college has a direct
effect on earnings or only an indirect effect through the education
decision. Specifically, one can include college proximity in the
earnings equation and use the interaction of college proximity
with a indicator for low parental education as an instrumental
variable for education. This identification strategy relies on the
extra boost to education and earnings among children with poor
family backgrounds. The resulting estimates are still substantially
higher than the ordinary least squares estimates, and provide no
evidence against the hypothesis that college proximity is an
exogenous determinant of schooling.

Preliminary Analysis of Earnings and Schooling in the NLS

Young Men Cohort
The data in this paper are 9rawn from the National
Longitudinal Survey of Young Men (NLSYM). The NLSYM
began in 1966 with 5525 men age 14-24 and continued with
follow-up surveys through 1981. Some descriptive statistics for

Interestingly, Griliches (1977) concluded that ordinary least

squares estimates of the return to education were probably !lQt
downward-biased, once measurement error in schooling was taken
into account.
 4

the original sample and two subsamples are presented in Table 1.

Like other longitudinal surveys initiated in the mid-1960s, the
NLSYM was not a random sample of the U.S. population: rather,
men from neighborhoods with a high concentration of non-white
residents were over-sampled.5 As shown in column (1), the
NLSYM sample contains a relatively high fraction of men from
the Southern· region (41% versus approximately 32% for a
nationally representative sample) and a high fraction of blacks
(28% versus approximately 10% for a nationally representative
sample).
In the baseline interview individuals were asked the
composition of their family when they were age 14: 77 percent
lived with both their father and mother; 12 percent lived with
only their mother; the remainder lived with other relatives or at
least one step-parent (row 5). Individuals were also asked their
father's and mother's education, although a relatively large
fraction of the sample report missing values for these variables
(22% are missing father's education, 11% are missing mother's
education). For observations with missing data I have assigned
the overall mean of father's or mother's education. The statistical
models reported below include dummies indicating whether either
parent's education level is imputed.

5
See Hall and Turner (1970).
 5
The 1966 interview also included a 28 item test of
6
"Knowledge of the World of Work" (see row 7). The overall
score on this test is correlated with completed education and wage
rates in later waves of the survey, and the test has been used as
a measure of "ability" in several previous studies of education and
earnings (e.g. Griliches (1976, 1977)).
Finally, the NLSYM data set contains a number of
characteristics of the respondent's local labor market in 1966.7
Among these is an indicator for the presence of an accredited 4-
year college in the local labor market (row 8).8 About 70
percent of individuals lived in a labor market area with a nearby
college. The college proximity rate varies by region (lower in
the South and Mountain regions), by urban versus rural location
(higher for individuals living in a Standard Metropolitan
Statistical Area), and is correlated with race and parental
education (see below).

6
The test items were questions on the job activities of 10
specific occupations, the education requirements for these 10
occupations, and the relative earnings of 8 different pairs of
occupations.
7
These are based on the county of residence in 1966.
8
An indicator for the presence of a 2-year college is also
included in the NLSYM, but this variable turns out to be only
weakly correlated with education or earnings. See below.
 6

Like other longitudinal surveys the NLSYM is affected by

sample attrition. Approximately 20 percent of the sample
dropped out in the first 3 years of the survey, and only 65
percent of the original sample were interviewed in the final
(1981) wave. In selecting a cross-section from the NLSYM there
is evidently a tradeoff between response rates and the age of the
respondents. Earlier waves have higher response rates but
relatively young sample members whereas later waves have lower
response rates but older sample members. I compromise by using
labor market information from the 1976 interview. In 1976 the
youngest respondents are 24 years of age and the available sample
is still relatively large (3694 observations or 71 percent of the
original sample). An important advantage of the 1976 data is that
all respondents were directly asked their educational attainment
as of the 1976 interview.
Column (2) of Table 1 reports the characteristics of individuals
who were interviewed in 1976 and who provided valid education
responses. These men have the same age and regional
distributions as the original NLSYM sample but are slightly less
likely to be black. The mean level of reported education in 1976
is 13.2 years. One-third of the sample report exactly 12 years of
schooling, 23% report some college, and 27% report 16 or more
years of education.
 7
Eighty-three percent of men interviewed in 1976 report a valid
wage observation. The characteristics of this working subsample
are reported in column (3) of Table 1. Comparisons with the
mean characteristics in columns (1)-(3) show few differences
between the original sample, the subsample of 1976 interviewees,
and the subsample with 1976 wages.
To begin an investigation of the returns to schooling in the
NLSYM Table 2 presents a variety of conventional earnings
functions estimated by ordinary least squares (OLS). All models
include a linear education term, a quadratic function of potential
experience (age-education-6), a race indicator, and dummies for
residence in the South and in a metropolitan area (SMSA) in
1976. The specification in column (2) adds 8 indicators for
region of residence in 1966 and another for residence in an
SMSA in 1966. The models in columns (3)-(5) add an increasing
set of family background characteristics: measures off ather' s and
mother's education (column 3); interactions of father's and
• mother's education (column 4); and indicators for family structure
at age 14 (column (5)). As shown by the test statistics in row 13,
the full set of family background variables are never jointly
significant, although the family structure indicators are marginally
significant by themselves. The estimated education coefficient (in
row 1) is remarkably stable across specifications and implies a
 8

7.3-7.5% earnings advantage for each additional year of

education, controlling for experience and other factors.9
Despite their stability across specifications the estimated
education coefficients in Table 2 may give a biased estimate of
the true economic return per year of education. To facilitate
discussion of the econometric issues involved, consider a simple
two-equation system describing schooling (S) and log wages (yi)
for individual i (in 1976):
(1) 𝑆𝑖 = 𝑋𝑖 𝛾 + 𝑣𝑖
l

(2) 𝑦𝑖 = 𝑋𝑖 𝛼 + 𝑆𝑖 𝛽 + 𝑢𝑖

Here Xi is a vector of observed attributes (with E(Xiu) = E(Xivi)

= 0) and 𝛽 has the interpretation of the "true" return to
education. 10 A conventional earnings equation estimated by OLS
gives a consistent estimate of 𝛽 if and only if ui and vi are
uncorrelated (i.e. if Si is econometrically exogenous in (2)).

9
Note that the estimated coefficient of a linear education
variable is only strictly interpretable as a "rate of return" to
schooling under very rigid conditions (see Mincer (1974)). I use
the terminology "rate of return to schooling" to refer to the
education coefficient in conventional human capital model.
10
If the return to education varies across individuals then the
coefficient {3 in equation (2) should be interpreted as the average
return to education. Specifically, suppose Yi = Xia + Si{3i +
Ei, where /3i is the marginal return to education for i. Then
equation (2) holds with 𝛽 = E(𝛽) and ui = 𝜖𝑖 + 𝑆𝑖 (𝛽𝑖 − 𝛽).
 9
There are a variety of reasons why schooling may be
correlated with the unobserved component of earnings. One that
has received considerable attention in the literature is "ability
bias" (see e.g. Griliches (1977)). Suppose that some individuals
have an unobserved characteristic ("ability") that enables them to
earn higher wages at any level of education. If these individuals
acquire higher-than-average schooling then the OLS estimate of
(3 will be upward-biased. The fact that individuals with higher
test scores (on IQ or achievement tests) tend to have higher
earnings and more schooling is often interpreted as evidence of
ability bias.
Another important source of correlation between ui and vi is
measurement error in schooling. Measurement error induces a
negative correlation between the error components of earnings
and observed schooling, leading to a downward bias in OLS
estimates of (3 (see Griliches (1977)).11 A similar negative bias
arises if the true return to schooling varies across the population
and if individuals with lower levels of schooling have higher
returns to schooling. Such a negative correlation is implied by a
model of school choice in which individuals with different

11
Estimates in the literature (cited by Griliches) suggest that
10% of the variance in measured education is due to measurement
error. In this case the OLS estimate of the return to education is
downward biased by 10-15 percent, depending on what other
covariates are included in the model.
 10

discount rates invest in schooling until the marginal return to

schooling equals the discount rate (see Card (1993) and Lang
(1993)).
A consistent estimate of the true return to education can be
obtained if there is a component of the vector Xi that affects
12
schooling but not earnings. If schooling were randomly
assigned, for example, then the realization of the randomizing
process could be used as to estimate equation (2) by instrumental
variables (IV).13 In the absence of "pure random assignment,
11

however, one needs to identify a causal determinant of schooling

that can be legitimately excluded from the earnings equation. The
presence of a nearby college may be such a variable. Students
who grow up in an area without a college face a higher cost of
college education, since the option of living at home is
precluded.14 One would expect this higher cost to reduce

12
If the true rate of return to education varies across the
population then one can obtain a consistent estimate of the
average return to education for some subset of the population.
See Angrist and Imbens (1993).
13
Something like this idea is used by Angrist and Krueger
(1991b), who use draft-lottery status as an instrument for
schooling of men who could have served in the Vietnam war.
14
Tabulations of the October 1973 Current Population Survey
show that in the early 1970s 34% of college students age 18-24
lived with their parents while attending school. The fraction is
higher (39 %) for black students.
 11
investments in higher education, at least among children from
relatively low-income families. 15
To check this basic insight I fit a linear model to years of
completed schooling (in 1976) for the subset of men who grew up
in local labor markets without an accredited 4-year college. The
determinants of schooling include region and urban/rural
indicators (measured as of 1966), age and race dummies, and
family background factors (family structure and parental
16
education). I then divided the overall sample into quartiles of
predicted education in the absence of a nearby college and
calculated the mean levels of education by quartile of predicted
education for men who grew up in areas with and without a local
college. Figure 1 plots the mean levels of education. In every
quartile the mean level of education is higher for those who grew
up near a college. For men in the three highest predicted
quartiles of education the effect of college proximity is modest
(0.2 to 0.4 years). For men in the lowest quartile, however, the
difference in mean education is 1.1 years. As expected, the
presence of a nearby college has its strongest effect on men with
lowest propensities to continue their education (e.g. men from

15
See Anderson, Bowman, and Tinto (1972) for a review of
the sociological literature on the effects of college accessibility on
attendance probabilities.
16
The R-squared of the regression is 0.30.
 12
single-headed families with low parental education 1n rural
Southern areas).

Instrumental Variables Estimates of the Return to Education

Table 3' presents a series of reduced form education and
earnings equations and the corresponding structural estimates of
the return to education, using college proximity as an
instrumental variable for completed education. Columns (1) and
(2) show the coefficients of an indicator for college proximity in
models for years of schooling. Columns (3) and (4) show the
coefficients of the college proximity variable in reduced form
wage equations (i.e. models that exclude education). Finally,
columns (5) and (6) report the IV estimates of the return to
education: these are simply the ratios of the corresponding
reduced form coefficients in the earnings and schooling equations.
The models in columns (1), (3) and (5) exclude parental education
and family structure variables while the models in columns (2),
(4), and (6) include these variables.
Two alternative specifications are reported in the upper and
lower panels of the table. The models in the upper panel (Panel
A) include the conventional measures of experience and
experience-squared constructed from observed age and education.
If schooling is measured with error, however, then experience is
also mismeasured -- suggesting possible biases in the reduced
 13
form models in Panel A. By the same token, if education is truly
endogenous in the earnings equation, then so is experience, since
experience is mechanically related to education. Therefore, in the
lower panel (Panel B) I have estimated models that instrument
experience and experience-squared with age and age-squared.
Regardless of the inclusion or exclusion of family background
variables, and irrespective of the treatment of experience, the
conclusions from Table 3 are similar. Growing up near a college
has a strong positive effect on both education (0.32 to 0.38 years
of schooling) and earnings (4.2 to 4.8 percent). The use of
college proximity as an exogenous determinant of schooling yields
IV estimates of the return to education in the range of 0.12 to
0.14. These estimates are 50-60 percent higher than the
corresponding OLS estimates -- about the same relative ratio as
reported by Butcher and Case (1993), Kane and Rouse (1993),
and Angrist and Krueger (1993). Nevertheless, the standard
errors of the IV estimates are relatively large, and one cannot
reject the hypothesis that differences between the IV and OLS
estimates are due to sampling error.17

17
Under the null hypothesis that the OLS estimates are
consistent the variance of the difference between the IV and OLS
estimates of the return to education is the difference in their
variances, which is approximately equal to the variance of the IV
estimate.
 14

Table 4 presents a series of alternative specifications designed

to probe the robustness of the estimates in Table 3. The top row
of the table contains OLS and IV estimates of the return to
education for the "basic specifications" in Tables 2 and 3 (OLS
from column (5) of Table 2; IV from the lower panel of column
(6) in Table 3). Row 2 presents estimates from the same
specifications, using as a dependent variable the logarithm of
1978 wages for the subset of men with wages in that year.18
The OLS estimate of the return to education is slightly lower in
1978 than 1976: otherwise, the estimated coefficients and overall
fit of the wage equation are similar in the two years. As in the
1976 data, the use of college proximity as an instrument raises
the estimated return to education by over 50%.
Row 3 presents OLS and IV estimates of the return to
education when a direct measure of "ability" -- the "Knowledge
of the World of Work" (KWW) score -- is included in the model.
In the OLS model the KWW score is a significant determinant of
·earnings (t-statistic = 6.9): a 1-standard deviation increase in
KWW is associated with a 6.6 percent increase in earnings. The
addition of the KWW score leads to a 25% attenuation in the
return to education relative to the basic OLS estimate. Since
education and the KWW test score are highly correlated,

18
Education, experience, and the current location variables are
all defined as of the 1978 survey.
 15

however, some of this attenuation is potentially attributable to the

presence of measurement errors in education.19 When college
proximity is used as an instrument for education (row 3 column
2) the estimated return to education rises and the estimated
coefficient of the KWW test falls to a small and statistically
insignificant value.
A potential criticism of this specification is that the KWW
score is treated as an error-free measure of "ability". To address
this criticism, the IV specification in row 4 treats both education
and the KWW score as "endogenous" (or measured with error)
and uses a measure of IQ (taken from school records for a subset
of NLSYM respondents) to instrument the KWW score.20 This
lowers the IV estimate of the return to education slightly but
raises the standard errors of the education and KWW coefficients
to the point where neither is statistically different from 0.
The IV estimates presented in rows 5 and 6 of Table 4 use
two alternative measures of college proximity as instruments for

19
Assuming that measurement errors account for 10 percent
of the cross-sectional variance in observed _schooling (Siegel and
Hodge (1968)), and that the true effect of KWW on earnings is
0, the expected attenuation of the schooling coefficient when
KWW is added to the model is about 5 % .
20
Note that one could include IQ in the earnings equation and
use the KWW score as an instrument. This has no effect on the
conclusions from Table 4.
 16
schooling. In row 5 college proximity is defined as living in a
local labor market with a public 4-year college.21 Proximity to
a public college has a slightly smaller reduced form effect on
education (0.31 years versus 0.32 for proximity to any ccllege)
and a slightly larger reduced form effect on earnings (6.2 %
versus 4.2 %). Thus the implied IV estimate of the return to
college is higher. than the IV estimate using proximity to any
college, although the standard error is again relatively large.
The IV estimation in row 6 combines 2 college proximity
variables: one for any accredited 4-year college, another for any
accredited 2-year college. In the reduced form equations the
presence of a nearby 2-year college has small positive effects on
schooling and earnings (whether or not an indicator is included
for proximity to a 4-year college). Using both ·indicators as
instruments leads to an estimated rate of return to education of
0.12, and a very slight improvement in the standard error of the
estimate relative to the baseline estimate in row 1.
One difficulty with these college proximity measures is that
they pertain to the place of residence in 1966 rather than the
place of residence at age 18 or 19, when the college enrollment
decision is typically made. By the time of the 1966 interview

21
Among the men who grew up in local labor markets with
accredited 4-year colleges, 73% were in labor markets with a
public 4-year coilege.
 17

some of the older NLSYM respondents could have already moved

to be closer to a college, giving rise to a reverse causation
between college proxi_mity and schooling attainment.22 A simple
check is to exclude the oldest respondents in the sample (e.g.
those over age 19 in 1966). An important caveat to this
exclusion is that the narrowing of the age range of the sample
makes it more difficult to separately identify the effects of
education and experience. Row 7 presents OLS and IV estimates
based on the subsample of men age 14-19 in 1966. The OLS
estimate of the return to education for the subsample is similar to
the baseline estimate. The IV estimate is above the corresponding
OLS estimate, although at the low end of the range of IV
estimates (24 % above the OLS estimate).
The results of these specification checks confirm the two main
conclusions from Table 3. First, IV estimates of the rate of
return to schooling based on college proximity are uniformly
higher than OLS estimates. Second, although the IV estimates
are imprecise, the range of the point estimates is 25-60 percent
above the corresponding OLS estimates.

22
For unmarried respondents enrolled in college and living
away from home in 1966 the place of residence was defined as
the place of residence of their parents. Thus there should be no
reverse-causation for these individuals.
 18
Is College Proximity a Legitimate Instrument?
For college proximity to serve as a legitimate instrument for
completed education it must affect individual schooling decisions
but have no direct effect on earnings. There are at least three
reasons why men who grew up near a college may have higher
earnings than other men, controlling for education, geographic
information, and parental background. First, families that place
a strong emphasis on education may choose to live near a college.
Children of these families may have higher "ability" or may be
more highly motivated to achieve labor market success. Either
factor could induce a positive correlation between college
proximity and the unobserved determinants of wages (i.e. un in
equation (2)). Second, the presence of a college may be
associated with higher school quality at nearby elementary and
secondary schools. Card and Krueger (1992) show that higher
school quality is associated with higher earnings. The omission
of direct information on the quality of schools attended by men
in the NLSYM may then lead to an error component'in wages
that is correlated with college proximity. Finally, if only
imperfect indicators are available for the place of residence in
1976, and if men who grew up in areas with a nearby college
tend to live in higher-wage areas, then college proximity may be
correlated with unobserved geographic wage premiums.
 19

The interpretation of college proximity as a factor that lowers

the cost of higher education suggests that growing up near a
college should have a bigger effect on the education outcomes of
children from poorer families. The pattern of education
differentials in Figure 1 confirms this notion. Letting X1i denote
the components of Xi other than college proximity, the implied
model for schooling is:

(lb) 𝑆𝑖 = 𝑋1𝑖 𝛾1 + 𝐶𝑖 𝛿0 + 𝐶𝑖 ∗ 𝑃𝑖 𝛿1 + 𝑣𝑖 ,

where Ci is an indicator for growing up near a college, Pi is an

indicator for low family income, and the coefficients o0 and o1
are both positive. In this case, even if Ci is included directly in
the earnings equation:

(2b) 𝑦𝑖 = 𝑋1 𝛼1 + 𝐶𝑖 𝛼0 + 𝑆𝑖 𝛽 + 𝑢𝑖

the interaction Ci*Pi of college proximity and poor family

background can be used as an instrumental variable for education.
The maintained assumption in this identification strategy is that
the direct earnings effects of living near a college (e.g.,
unobserved geographic wage differentials) do not vary by family
background.
Table 5 presents reduced form and structural estimates of the
return to education based on equations (lb) and (2b). Low family
background is defined by neither parent graduating from high
 20
school.23 The reduced form coefficients in columns (1) and (2)
confirm that the effects of living near a college are bigger for
men with poorly-educated parents. The corresponding IV
estimate of the return to education is presented in column (3),
along with the direct earnings effect of living near a college. The
estimated return to schooling is slightly smaller than the IV
estimates in Table 3, and the estimated standard error is slightly
larger. On the other hand, the point estimate of the direct
earnings effect of college proximity is small and insignificantly
different from 0. Although imprecise, these estimates provide no
evidence against the assumption that college proximity is an
exogenous determinant of schooling.
One potential criticism of the specification in columns (1)-(3)
is the arbitrary classification of family backgrounds into only 2
categories. An alternative is to interact college proximity with a
broader set of parental education indicators. The results in
column (4) use interactions of college proximity with indicators
for 8 parental education classes (the same indicators used in the
earnings models in Tables 2-4). The expansion of the instrument

23
This definition of low family background was derived by
comparing mean education levels of men in the 8 parental
education classes used in the models in Tables 3 and 4. The
means show a discrete drop for men from the two lowest parental
education categories. I therefore combined the two categories as
a "low family background indicator.
11
 21
set has the effect of lowering the standard error of the IV
estimate, while raising the point estimate slightly. An over-
identification test for the mutual consistency of the available
instruments is insignificant (p-value =0.28). As in column (3),
the estimate of the direct earnings effect of living near a college
is small and statistically insignificant.
Another alternative is to interpret predicted education in the
absence of a nearby college (i.e. the predicted education level
used to generate the quartiles in Figure 1) as a continuous
indicator of "family background". Using the interaction of
predicted education and college proximity as an instrument, and
including college proximity directly in the earnings equation, the
IV estimate of the return to education is 0.122, with a standard
error of 0.075.
Regardless of the method of classifying family background, IV
estimates based on the interaction of family background and
_college proximity are similar to IV estimates based on college
proximity alone. Furthermore, estimates of the direct effect of
college proximity on wages are uniformly small and statistically
insignificant. Assuming that college proximity can be excluded
from the earnings equation, both college proximity and its
interaction with family-background indicators can be used as
instruments for schooling. For example, using 9 parental
education indicators interacted with college proximity as
 22

instruments for schooling, the IV estimate of the return to

schooling is 0.115, with a standard error of 0.034.24 Although
this estimate is 57% above the corresponding OLS estimate, the
Hausman-Wu statistic is 1.24 -- not large enough to reject the
hypothesis of no simultaneity bias at conventional significance
levels.
Discussion of Results
Although imprecise, the instrumental variables results
presented in Tables 3-5 suggest that a conventional OLS
estimation strategy yields a downward-biased estimate of the
"true return to education. This finding echoes the conclusion
11

reached in a number of recent studies of endogenous schooling

(cited above), and seems directly at odds with the widely accepted
notion that individuals with higher education would have above-
average earnings at any level of education. One possible
explanation for the positive gap between IV and OLS estimates of
the return to education is that the latter are downward-biased by
measurement error in schooling. In light of the estimated
reliability of survey measures of education, however, the potential
downward bias in the OLS estimates is on the order of 10-15
percent.The differences between the IV and OLS estimates in

24
The over-identification test statistic for this estimate (with
8 degrees of freedom) has a probability value of 0.38.
 23
this paper and in other recent studies is substantially above this
range.
An alternative possibility, discussed in some detail in Card
(1993), is that the "true" rate of return to education varies across
the population, and that the increase in education associated with
college proximity occurs for individuals with relatively high rates
of return to schooling. Algebraically, the IV estimate of the
return to schooling is the ratio of the differences in average wages
and average education between individuals who grew up in labor
markets with and without a nearby college.25 If the presence of
a nearby college affects only the education decisions of men with
poor family backgrounds, then the IV estimate depends only on
the marginal return to schooling in this subset of the population.
Thus one explanation for the relatively high IV estimates of the
return to education in Tables 3-5 is that the marginal return to
education among men with poor family backgrounds is relatively
high.
Why do men with poorly educated parents have high returns
to schooling? According to the simplest economic model of

25
Specifically, let y1 and y2 represent mean wages of
individuals who grew up in labor markets with and without a
nearby college (adjusted for other covariates), and let S1 and S2
represent mean years of schooling for the same 2 groups (again,
adjusted for other covariates). Then the IV estimate of the return
to schooling is (y1-y2)/(S1-S2).
 24

school choice (Becker (1967)), individuals have decreasing

marginal returns to schooling and invest in education until the
marginal return to the last year of schooling equals their marginal
discount rate.26 If most of the variance in education outcomes is
attributable to differences in individual-specific discount rates then
on average the less-educated population will be mainly composed
of individuals with high discount rates. Since low-income
families presumably face higher interest rates than high-income
families, this line of reasoning suggests that marginal returns to
schooling are highest for the children of poor families. In effect,
many less-educated workers stopped their schooling "too soon"
because they faced high marginal costs of funds for further
education.
This interpretation of the less-educated labor force stands at
odds with a more conventional view that the less-educated are less
"able" or have low benefits of schooling. At a minimum, the
results in Tables 3-5 suggest that marginal returns to schooling
among the less-educated are as high as typical OLS estimates of
the return to schooling. Taken in combination with the results in
other recent studies of endogenous education -- all of which find
downward bias in the OLS estimates -- the results here suggest

26
This is a condensed version of the argument developed in
Card (1993). See also Lang (1993).
 25
that the economic value of education for many children may be
significantly understated.

Conclusion
Any credible analysis of the causal link between education and
earnings requires an exogenous source of variation in education
choices. In this paper I explore the use of college accessibility as
an exogenous determinant of schooling. An analysis of education
and earnings outcomes for men in the NLS Young Men Cohort
shows that men who grew up in areas with a nearby 4-year
college have significantly higher schooling and significantly
higher earnings. These effects are concentrated among men with
poorly-educated parents -- men who would otherwise stop
schooling at relatively low levels. The implied instrumental
variables estimates of the earnings gain per year of additional
schooling (10-14 %) are substantially above the earnings gains
estimated by a conventional ordinary least squares procedure
(7.3%).
These inferences are robust to minor changes in specification,
including the addition of measured test scores to the earnings
model and changes in the definition of college proximity.
Nevertheless, they rely on the restrictive assumption that living
near a college has no effect on earnings apart from the effect
through education. To test this assumption I use the fact that
 26
college proximity has a larger impact on the schooling choices of
men with poorer family backgrounds. Thus, an interaction of
college proximity and low family background can be used as an
instrumental variable for observed schooling even in earnings
models that include a direct college proximity effect. The results
of this test give rise to estimates in the same range as the simpler
instrumental variables estimates based on college proximity alone.

While none of the instrumental variables estimates of the return

to education is very precise, they all point toward relatively high
returns to schooling for children of poorly-educated parents. This
pattern is consistent with a simple economic model of endogenous
schooling in which differential access to funds leads to relative
under-investment in schooling among children of lower-income
families.
 27

References
Anderson, C. Arnold, Mary Jean Bowman, and Vincent Tinto.
Where Colleges Are and Who Attends. New York:
McGraw Hill, 1972.

Angrist, Joshua D. and Guido W. Imbens. "Two-Stage Least

Squares Estimation of Average Causal Effects in Models
with Variable· Treatment Intensity." Unpublished
Discussion Paper, Harvard University Department of
Economics, 1993.

Angrist, Joshua D. and Alan B. Krueger. "Does Compulsory

Schooling Affect Schooling and Earnings?" Quarterly
Journal of Economics 106 (November 1991): 979-1014.
(1991a)

Angrist, Joshua D. and Alan B. Krueger. "Estimating the Payoff

to Schooling Using the Vietnam-Era Draft Lottery."
Princeton University Industrial Relations Section Working
Paper #290, August 1991. (1991b)

Angrist, Joshua D. and Alan B. Krueger. "Split Sample

Instrumental Variables". Unpublished Discussion Paper,
Princeton University Department of Economics, July
1993.

Ashenfelter, Orley and Alan B. Krueger. "Estimates of the

Economic Return to Schooling for a New Sample of
Twins." Princeton University Industrial Relations Section
Working Paper #304, July 1992.

Becker, Gary S. Human Capital and the Personal Distribution of

Income. Ann Arbor: University of Michigan Press, 1967.

Butcher, Kristin F. and Anne Case. "The Effect of Sihllng

Composition on Women's Education and Earnings."
 28
Unpublished Discussion Paper, Princeton University
Department of Economics, June 1993.

Card, David and Alan B. Krueger. "Does School Quality

Matter? Returns to Education and the Characteristics of
Public Schools in the United States." Journal of Political
Economy 100 (February 1992): 1-40.

Card, David. "Earnings Schooling and Ability Revisited".

Unpublished Paper, Princeton University Department of
Economics, August 1993.

Ehrenberg, Ronald and Robert Smith. Modem Labor Economics

(4th edition). New York: Harper Collins, 1991.

Griliches, Zvi. "Wages of Very Young Men. 11

Journal of
Political Economy 84 (1976): 569-586.

Griliches, Zvi. "Estimating the Returns to Schooling: Some

Econometric Problems. 11
Econometrica 45 (January
1977): 1-22.

Hall, George E. and Anthony Turner. "Sampling, Interviewing,

and Estimating Procedures". Appendix B in Career
Thresholds: A Longitudinal Study of the Education and
Labor Market Experience of Male Youths. United States
Department of Labor Manpower Administration.
Washington DC: USGPO, 1970.

Kane, Thomas J. and Cecilia E. Rouse. "Labor Market Returns

to Two- and Four-Year Colleges: Is a Credit a Credit and
Do Degrees Matter?" Princeton University Industrial
Relations Section Working Paper #311, January 1993.
 29
Lang, Kevin. "Ability Bias, Discount Rate Bias, and the Return
to Education." Unpublished Discussion Paper, Boston
University Department of Economics, May 1993.

Mallar, Charles D. "Alternative Econometric Procedures for

Program Evaluations: Illustrations from an Evaluation of
Job Corps." 1979 Proceedings of the Business and
Economic Statistics Section, American Statistical
Association. Washington DC: American Statistical
Asociation, 1979, pp. 317-321.

Mincer, Jacob. Schooling, Experience, and Earnings. New

York: National Bureau of Economic Research, 1974.

Psacharopoulos, George. "Returns to Education: A Further

International Update and Implications." Journal of
Human Resources 20 (Fall 1985): 583-604.

Rosen, Sherwin. "Human Capital: A Survey of Empirical

Research." In Ronald Ehrenberg, editor, Research in
Labor Economics volume 1. Greenwich Connecticut: JAI
Press, 1977.

Seigel, Paul and Robert Hodge. "A Causal Approach to the

Study of Measurement Error." In Hubert Blalock and
Ann Blalock, editors, Methodology in Social Research
New York NY: McGraw Hill, 1968.

Willis, Robert J. "Wage Determinants: A Survey and

Reinterpretation of Human Capital Earnings Functions."
In Orley Ashenfelter and Richard Layard, editors,
Handbook of Labour Economics. New York NY: North
Holland, 1986.
 Mean Years of Education
By Quartile of Predicted Education
16

15
C
.
3 14
"O
w
(II 13
L
0
a,
>
0
2

2 3
Quartile of Predicted Education

-e- No College Nearby College Nearby

Note: prediction equation is fit to subsample with no college nearby

Figure 1
Table 1: Semple Characteristics for Overall Sample and 1976 Subset
of National longitudinal Survey of Young Men

overall Subset Interview in 1976;

NLS·YM Valid Va 11 d Wage &
Sample Education Education

,. Age Distribution In 1966:

Age 14• 15 (X) 25.9 25.3 25.5
Age 1 6 • 1 7 24.9 23.8 24. 1
Age 18·20 23.1 24. 1 24.6
Age 21·24 26. 1 26.7 25.8

2. Regional Distribution in 1966:

Northeast (X) 20.2 20.0 20.7
Midwest 25.4 26.3 26.0
South 41 . 1 41 .3 41 .4
IJe st 13.3 12.5 11.9

3. Lived in SMSA 1966 (X) 66.0 64.3 65.0

4. Lived Near 4•year College 69.2 67.8 68.2

In 1966 (X)

5. Fam I l y Structure at Age 1 4:

Mother & Father (X) 76.8 79.2 78.9
Mother Only (X) 11. 8 10.0 1 0. 1

6. Average Parental Education

Mother's Education (yrs> 10.3 10.4 1 0. 3
Father's Education (yrs) 9.4 1 0. 0 10.0

7. Percent Black 27.5 23.0 23.0

8. Average Score on K WIJ Test 33.0 33.5 33.5

9. Interviewed in 1976 ( X) 70.7 100.0 1 00. 0

, 0. Hean Education in 1976 13.2 1 3.2 1 3. 3

1 1 . Live in South in 1976 ( X > 39.6 40.0 40.3

12. Sample Size 5225 3613 3010

Notes: Means are based on all available val Id observation• in any

subsample.
Table 2: Estimated Regression Models for Log Hourly Earnings

(1) ( 2) ( 3) ( 4) ( 5)

1. Education 0.074 0.075 0.073 0,074 0.073

(0.004) (0.003) (0.004) (0.004) (0.004)

2. Experience 0.084 0,085 0,085 0,085 0.085

(0.007) (0,007) (0.007) (0.007) (0.007)

3. Experience·Squared ·0.224 ·0.229 ·0.230 ·0.226 ·0.229

/100 (0.032) (0.032) (0.032) (0.032} (0.032)

4. 8 l II Cle Indicator • 0.190 • 0. 199 • 0.194 ·0.194 -0.189

(0.017} (0.018) (0.019) (0.019} (0.019)

5. Live in South -0.125 ·0.148 ·0.146 ·0.145 ·0.146

(0.015) (0.026) (0.026} (0.026) (0.026)

6. Li Ve in Sl<SA 0. 1 61 0.136 0.136 0. 137 0.138

(0.015) (0.020) (0.020) (0.020} (0.020)

7. Region in 1966 no yes yea yes yes

(8 indicators)
8. Live in SMSA In 1966 no . yes yes yes yes
II
9. Parental Education no no yes yes yes
(main effects)
10. Interacted Parente no no no yes yes
Education Classes
C
1 1 . Fam i I y Structure no no no no yes
( 2 indicators)

12. R·squared 0.291 0.300 0.301 0.303 0.304

13. P·value for family 0.235 0.462 0.16 5

background effects

Motes: Standard errors in parentheses. Sample size Is 3010. The dependent

variable In all cases Is the tog of hourly wages In 1976. The
mean and standard deviation of the dependent variable are 6.262 and
0.444.
1Variables representing years of education of mother and father, plus
indicators for missing mother's or father's education.
blndicators for 8 classes of mother's and father's education.
clndicators for father and mother present at age 14, and single
mother at age 14.
Table 3: Reduced Form and Structural Estimates of Education and
Earnings Models

Reduced Form Models: Structural Models

Education Earnings of Earnings

(1) (2) ( 3) (4) C 5) C 6)

A: Treat Experience and Experience Squared as Exo9enous

1. Live Near 0.320 0.322 0.042 0.045

College In (0.066) (0.063) (0.016) (0.016)
1966

2. Education 0.132 0.140

(0.055) (0.055)

3. Family no yes no yes no yes

Background
a
Variables

b/
B: Treat Experience and Experience Squared as Endogenous

4. Live Near 0.362 0.365 0.047 0.048

College in (0.114) (0.105) (0.019) (0.019)
1966

5. Education 0.122 0.132

(0.046) (0.049)

6. Fam i l y no yes no yes no yes

Background
a
Variables

Motes: standard errors in parentheses. Sample size is 3010. The dependent

variable In columns 1 and 2 is completed education In 1976 (mean
and standard deviation: 13.263 and 2.677). The dependent variable In
columns 3·6 is the log of hourly wages In 1976 (mean and standard
deviation: 6.262 and 0.444). All models include a black indicator,
indicators for southern residence and residence In an SMSA in 1976,
indicators for region In 1966 and living in an SMSA in 1966, as
well as experience and experience squared.
8
14 variables representing mother's and father's education,
indicators for missing father's or mother's education,
interactions of mother's and father's education, and dummies for
family structure at age 14.
b
In these models, experience is treated as endogenous. Instruments
for experience and experience squared are age and age squared.
Table 4: OLS and Instrumental Variables Estimates of the Return to
Educetfon: Alternative Specifications

OLS IV
Estimate Estimate •

1. Basic Specification 0.073 0. 132

(N,.3010) (0.004) (0.049)

2. Use 1978 II ages and Education 0.066 0. 117

(Nz2639 w i th 1978 data) (0.006) (0.061)

3. Include I( 1111 Test Score 0.055 0. 136

(N:2963 w i th valid I( II II ) (0.004) (0.078)

4. Include I( II II Test Scoreb 0.061 0.089

Instrument I( 1111 w Ith I 0 (0.005) (0.085)
(Nz2040 with Valid 1(1/1/ and I 0)

5. Use Proximity to Public College as in 0.194

as instrument for education row 1 (0.059)

6. Use Proximities to 2•year and as in 0.117

4·year colleges as instruments row 1 (0.047)
for education

7. Use Subsample Age 14• 19 In 1966 0.076 0.094

(Nz2037) (0.006) (0.064)

Notes: The dependent variable in row and rows 3·7 ls the log of hourly
wages in 1976. The dependent variable in row 2 is the log of
hourly wages in 1978. Reported estimates are coefficients of
linear education variable in models th t also include a black
indicator, Indicators for southern residence and residence in an
SMSA in 1976, indicators for region in 1966 and living in an SMSA
in 1966, expe;lence and experience squared, and 14 variables
representing mother's and father's education, indicators for missing
father's or mother's education, interactions of mother's and
father's education, and dummies for family structure at age 14.

1
1n these models education and experience are treated as endogenous.
Instruments for experience and experience squared are age and age
squared. Instrument for education is proximity to a 4•year college
unless otherwise noted.

bin the IV estimation KIii/ is treated as endogenous and 10 score is

added to the Instrument list.
Table 5: Instrumental Variables Estimates of the Return to Education
Based on Interaction of Parental Education and Proximity to
College

Reduced Form Models: Structural Models

Education Earnings of Earnings

(1) (2) (3) (4)

1. Live Near 0.154 0.029 0.o, 5 0. 013

College in (0.135) (0.024) (0.029) (0.024)
1966

2. Live Near 0.462 0.043

College* Low CO. 186) (0.032)
Parental
8
Education

3. Educationb 0.093 0.097

(0.065) (0.048)

4. Family yes yes yes yes

Background
Variablesc

Notes: standard errors in parentheses. Sample size Is 3010. The dependent

variable in all models Is the log of hourly wages In 1976. The mean
and standard deviation of the dependent variable ■re 6.262 and
0.444. All models include a black indicator, indicators for
southern residence and residence in an SMSA in 1976,
indicators for region in 1966 and living in an SMSA in 1966, as
well as experience and experience squared. Experience and
experience squared are treated as endogenous, with age and age
squared used as instruments.
a
Interaction of indicator for living near a college in 1966 and
indicator for both parents having less than high-school education.

b
In column 3 the instrument for education is an Interaction of
an Indicator for low parental education with an indicator for
living near a college in 1966. In column 4 the instruments are
interactions of 8 parental education class indicators with an
indicator for living near a college in 1966.
C
14 variables representing mother's and father's education,
indicators for missing father's or mother's education,
interactions of mother's and father's education, and dummies
for family structure at age 14.