The Relationship Between Education and Adult
               Mortality in the U. S.
                                 Adriana Lleras-Muney∗
                                        May 2, 2001


                                           Abstract
          Prior research has uncovered a large and positive correlation between educa-
      tion and health, but there are difficulties in determining whether this relation-
      ship is causal. In this paper I reexamine whether education has a causal impact
      on health. I follow synthetic cohorts using successive U.S. censuses to estimate
      the impact of educational attainment on mortality rates. I then use compulsory
      education laws from 1915 to 1939 as instruments to obtain a consistent causal
      estimate of this effect. While least squares estimates suggest that an additional
      year of education lowers the probability of dying in the next 10 years by ap-
      proximately 1.3 percentage points, results from the IV estimation show that the
      effect is in fact much larger, at least 3.6 percentage points. Overall, one more
      year of education increases life expectancy at age 35 by 1.2 years. These results
      provide evidence of a causal effect from education to health and suggest that
      the income returns to education substantially underestimate the overall returns
      to education.
       JEL: I12, I20, J10, J18, N32, N42

   ∗ Email: al277@columbia.edu. Department of Economics, Columbia University, New York, NY

10027. I would like to thank Ann Bartel, Francisco Ciocchini, Ana Corbacho, Rajeev Dehejia, Phoe-
bus Dhrymes, William Gentry, Kenneth Leonard, Manuel Lobato, Alexander Peterhansl, Nachum
Sicherman and the seminar participants at the Federal Reserve Board, Michigan State University,
University of Montreal, UNC-Greensboro, Princeton University, UC Berkeley, UC Davis, UC Irvine,
University of Illinois at Urbana Champaign for their comments and suggestions. I am especially
grateful to my advisor Sherry Glied who fully supported me throughout this project. All errors are
mine. This research was partially funded by Columbia University’s Public Policy Consortium, and
the Bradley Foundation.
1. Introduction
Access to health care insurance,1 expenditures on health care,2 and even income
levels3 have been shown to have little effect on health. On the other hand, there is a
large and positive correlation between education and health (Grossman and Kaestner
1997). This correlation is strong and signiÞcant even after controlling for different
measures of socio-economic status, such as income and race, and regardless of how
health is measured (morbidity rates, self-reported health status or other measures
of health). Given that the measured effects of education are large, investments in
education might prove to be a cost-effective means of achieving better health,4 if
education indeed helps us to be healthier. But prior research has not ascertained
whether the relationship between education and health is causal.
    The purpose of this paper is to determine whether education has a causal effect on
health, in particular on mortality. Recent studies5 suggest that the value of a healthy
life is very large. The negative relationship between education and mortality, the
most basic measure of health, has become well established since the famous Kitagawa
and Hauser (1973) study, which found signiÞcant differences in mortality rates across
educational categories for both sexes. More recent studies (e.g. Feldman et al.,
1989, Pappas et al., 1993, Christenson and Johnson, 1995, Deaton and Paxson, 1999)
conÞrm these Þndings. Elo and Preston (1996) control for a variety of other mortality
factors such as income, race, marital status, region of residence, and region of birth.
Rogers et al. (2000) further control for access to health care, insurance, smoking,
exercise, occupation, and other factors. Figures 1 and 2 document this relationship
using consecutive census data for the US: in all cohorts, those who survive have higher
education than those who do not.
    The existing literature has explained this correlation in three ways. One contro-
versial hypothesis is that education increases health, either because education makes
people better decision-makers (Grossman 1975) and/or because more educated people
have better information about health (Kenkel 1991, Rosenzweig and Schultz 1981).6
Another possibility is that poor health results in little education (Perri 1984, Curry
and Hyson 1999). Finally, this correlation could be caused by a third unobserved
variable that affects both education and health, for example genetic characteristics or
parental background. Many studies have attempted to include these factors.7 How-
   1 See Newhouse (1993).
   2 For example see Filmer and Prichett (1997).
   3 Grossman (1975) shows that income levels beyond a certain minimum did not have an impact

on health outcomes.
   4 That education might be a more cost-effective means to increase health than increasing medical

care expenditures was Þrst suggested by Auster et al (1969). Their Þndings suggest that the elasticity
of mortality rates with respect to education is twice as large as that of medical expenditures.
   5 For example Nordhaus (1999).
   6 Kenkel (1991) and Gilleskie and Harrison (1998) Þnd evidence to support both Grossman’s and

Kenkel’s hypotheses.
   7 Some studies suggest that adding controls will eliminate the observed relationship. Behrman et

al (1998) Þnd that using a random effects model to control for unobserved heterogeneity, the effect
of education on mortality disappears. Other examples include Wolfe and Behrman (1987), Duleep
(1986) and Menchik (1993).




                                                  2
ever, Fuchs (1982) argued that discount rates (which no study controls for) would
also explain the correlation: people who are impatient invest little in education and
health, while people who are patient invest a lot in both.8 Of course, these theories
are not necessarily mutually exclusive.
    In this paper I address this issue using a unique quasi-natural experiment: between
1915 and 1939, at least 30 states changed their compulsory schooling laws and child
labor laws. If compulsory schooling laws forced people to get more schooling than
they would have chosen otherwise, and if education increases health, then individuals
who spent their teens in states that required them to go to school for more years
should be relatively healthier and live longer.
    This natural experiment allows me to use these laws as instruments for education
and thus to identify the causal relationship between education and mortality. These
laws are valid as a natural experiment for various reasons. The 1914-1939 period
saw the largest historical increases in the number of students graduating from high
school. There were also a signiÞcant number of changes in the laws at that time.
Historians suggest that the laws were enforced during this time, and therefore were
likely to have affected many individuals. These laws were external to individuals;and
there were no systematic differences between children in the states where the laws
required (for example) 8 years of education and children in the states which required
9 years.9 Finally, the laws are very likely to be uncorrelated with health. Nonetheless,
I control for state-of-birth, cohort year, state expenditures on education in state-of-
birth, and a host of other state factors that might be correlated with the laws and
health outcomes.
    The intuition that compulsory education laws provide a natural experiment was
put forward Þrst by Angrist and Krueger (1991). They argued that because compul-
sory education laws forced individuals to stay in school until a certain age, those born
in later quarters would stay in school longer. Although they were criticized for their
choice of quarter of birth as an instrument,10 the underlying principle is appealing
and implementable. Other researchers have successfully used compulsory education
laws as instruments in the context of the returns to education for other countries.11
    No other papers have used natural experiments to measure the effect of education
on mortality. A few studies (Berger and Leigh 1988, Sander 1995, and Leigh and Dhir
1997) have used instrumental variable (IV) estimation with other measures of health,
such as blood pressure, smoking or exercise.12 But these studies are inconclusive
    8 Fuchs (1982) tried to measure discount rates (through a telephone survey), and use them to

predict education and health. His results were mixed. Farrel and Fuchs (1982) examined the issue
again, but the evidence they provide is indirect. Munasinghe and Sicherman (2000) however provide
evidence which suggests that time preference plays an important role in the determination of smoking
and earnings growth.
    9 See Lleras-Muney (2001) for a more detailed analysis of the effect of these laws on educational

attainment.
  1 0 See Bound, Jaeger and Baker (1995) and Bound and Jaeger (1996).
  1 1 Harmon and Walker (1995) look at the effects of the laws in the UK. Meghir and Palme (1999)

used Swedish data. Acemoglu and Angrist (1999) used compulsory education laws in the US as
instruments for average education at the state level to determine the size of the social returns to
education.
  1 2 Berger and Leigh (1988) estimate the effect of education on blood pressure using the NHANES I.




                                                 3
because each paper’s choice of instrument is questionable. For example, all of these
studies use parents’ background/education as instruments, but we know these are
correlated with children’s health,13 and furthermore, we know that health shocks
during childhood or gestation have persistent health effects into adulthood.14 Income
and education expenditures in state-of-birth could serve as instruments (Berger and
Leigh 1988), but again they might be correlated with state expenditures on health,
state industrial composition and other state characteristics that affect health.
    Using the 1960, 1970 and 1980 Censuses of the US, I select those individuals who
were 14 years of age between 1915 and 1939. I then construct synthetic cohorts and
follow them over time to calculate their mortality rates. I then match cohorts to the
compulsory attendance and child labor laws that were in place in their state-of-birth
when they were 14 years old. The census data have not been used to calculate mortal-
ity rates before in economic analyses,15 but this methodology has many advantages.
Because census data are very extensive and go back well into the 19th century, this
method could be used to analyze mortality experiences in periods where no other data
are available.
    Several IV estimations are presented, including an original two-stage procedure
for grouped data that can be applied when the Þrst stage can be estimated at the
individual level but the second stage can only be estimated at the aggregate level.
This procedure, inspired by the traditional two-stage least squares (2SLS) method,
can easily be applied to other cases as well. Comparison of these results with efficient
Wald estimates and standard 2SLS estimates conÞrms that the procedure is valid.
    The results provide evidence that suggests there is a causal effect of education on
mortality and that this effect is larger than the previous literature suggests. While
GLS estimates suggest that an additional year of education lowers the probability of
dying in the next 10 years by approximately 1.3 percentage points, my results from
the IV estimation show that the effect is in fact much larger: at least 3.6 percentage
points.
    This paper is organized as follows. Section 2 describes the data used in this
project, including a description of how the census is used to obtain mortality rates.
They use state-of-birth, income and education expenditures per capita from year-of-birth to age 6 in
state-of-birth, and dummies for ancestry as instruments for education. They also estimate the effect
of education on disability with NLS data, using IQ and family background measures as instruments.
In both cases schooling is signiÞcant. Using a sample of older persons from the 1986 PSID, Leigh
and Dhir (1997) use parental education, background, and state-of-residence at age 16 to instrument
for education in regressions for disability and exercise. Alternatively, they include direct measures
of time preferences and risk aversion. Education was not always signiÞcant. Finally Sander (1995)
examines the effect of schooling on the odds of quitting smoking using the General Social Survey. He
uses parental schooling as an instrument for schooling and Þnds that the effect of schooling is quite
large for whites.
  1 3 Many development studies show that family background affects children’s health. For a thorough

review of these studies see Strauss and Thomas (1995). IQ measures (Berger and Leigh, 1988) suffer
from the same problem.
  1 4 For examples see studies that looked at the consequences of the Dutch famine on the health

of adults conceived during the famine, such as Hoek, Brown and Susser (1998) or Roseboom et al
(2000).
  1 5 However this methodology is used in epidemiology. For example see the work by Haines and

Preston (1996).



                                                 4
Section 3 shows that compulsory attendance and child labor laws had an impact
on the educational attainment of individuals, and presents evidence that these laws
are good instruments. Section 4 presents the general econometric framework used
in the mortality analysis. The framework includes least squares as well as two IV
estimations. The results are presented and discussed in Section 5, and conclusions
are given in section 6.

2. Data
I use the U.S. censuses of 1960, 1970 and 1980, which are one percent random samples
of the population.16 The census provides information on age, sex, race, education, ur-
ban/rural residence, marital status, state of residence and state of birth. My samples
include all white persons born in the 48 states,17 that were 14 years of age between
1914 and 1939, with no missing values for completed years of education.18
    I use the censuses to follow “synthetic cohorts.” Although I do not observe the
same individuals over time (so I cannot observe individual deaths), I do observe the
same groups over time, which allows me to estimate group death rates. I aggregate
the censuses into groups deÞned according to their gender/cohort and state-of-birth
(descriptive statistics in Table 1). I follow 25 cohorts, born in 48 states. Using the
1960, 1970 and 1980 censuses, I can calculate two 10-year death rates for each group:
one for 1960-1970, and another for 1970-1980. For example, the 1960-1970 death rate
for a group is the number of people alive in 1960 (N60 ) minus the number of people
alive in 1970 (N70 ) divided by the population in 1960 (N60 ):

                                           N60 − N70
                                              N60
    One issue that arises in estimating death rates by groups is measurement error.
As Figure 3 shows, because of random sampling the number of deaths will be over-
estimated about half the time and underestimated half the time for all cohorts. As a
result, some estimated death rates are negative. In the data, we observe more nega-
tive death rates for younger cohorts and fewer negative death rates for older cohorts
(see Figure 4a); this is a pattern we should expect. As we can see in Figure 3B,
with a zero death rate (no change in the population), two successive samplings of the
same population result in a negative death rates half the time. When the death rate
increases (as the population ages), the likelihood that the second sample will contain
more observations than the Þrst falls, resulting in fewer negative death rates. We also
observe fewer negative death rates for states with large population (Figure 4b), which
is also to be expected since the sampling error is smaller for larger populations.
  1 6 The data come from the IPUMS 1960 general sample, the 1970 Form 2 State sample (originally

15% state sample), and the 1980 1% Metro sample (originally B sample). These data sets were
downloaded from the following web site: http://www.ipums.umn.edu
  1 7 Hawaii and Alaska were not then part of the Union.
  1 8 For consistency across censuses, I recoded completed years of education to be a maximum of 18

years instead of 20 in 1980.




                                                5
    The negative death rates are not a source of concern for two reasons. First,
the estimated death rates will result in consistent estimates of the true death rates.19
Second, average cohort death rates from the censuses are very similar to those obtained
from individual data from the NHEFS described below (see Figure 4c).Note that the
graph suggests there is evidence of age heaping: for ages that are multiples of 10, the
death rates fall, because individuals tend to over-report their age and chose a multiple
of ten when doing so.
    I also used the National Health and Nutrition Examination Survey I Epidemiologic
Follow-up Study, 1992 (hereafter NHEFS). This survey followed 14,407 individuals
who were between 25 and 74 years of age when interviewed for the Þrst National Health
and Nutrition Examination Survey (NHANES I) between 1971 and 1974. The NHEFS
collected information on individuals in four subsequent waves (1982-84, 1986, 1987 and
1992). The sample is composed of whites20 who were born in the 48 states between
1901 and 1925 and who were followed successfully, with no missing observations for
years of completed education. The sample is further restricted to those who were
alive in 1975 (N=4554). The NHEFS followed individuals and recorded whether they
had died by 1985. Table 2 shows the summary statistics for this data.
    The data on compulsory attendance and child labor laws come from a number of
sources. There are eight years of state-level data (1915, 1918, 1921, 1924, 1929, 1930,
1935 and 1939) on these laws,21 and some additional information for other years. I
imputed missing observations by using the older values. The information was not
recorded consistently by a single agency; in cases of conßicting pieces of data, I used
the newer information to correct the data. (See the Appendix for tabulations and
trends of the laws) I also collected data on state-level factors that contributed to the
growth of secondary education from 1915 to 193922 or that could affect mortality.
These include state expenditures on education, number of school buildings per acre,
percent of the population that was living in urban areas, percent of the white popu-
lation that was foreign born, percent of the population that was black, percent of the
population employed in manufacturing, average annual wages in manufacturing per
worker, average value of farm property per acre, and number of doctors per capita
(See Lleras-Muney 2001 for information on data sources).
    Each individual is matched to the laws and state characteristics that were in place
in their state-of-birth when they were 14 years old. I choose this age because it is the
lowest common drop-out age across states.23 This procedure assumes that individuals
  1 9 Also note that IV estimates are only consistent, not unbiased, estimates of structural parameters.

A consistent estimate of the dependent variable is sufficient for the IV estimators to be consistent.
  2 0 Other researchers have suggested that blacks had signiÞcantly different school experiences during

the begining of the century. See Card and Krueger (1992). Also preliminary work on my part
suggests that compulsory schooling laws and child labor laws did not affect blacks. The laws are
never signiÞcant. It is unclear why. See Lleras-Muney (2001).
  2 1 Acemoglu and Angrist (1999) have gathered similar data. The data for this project was collected

independently.
  2 2 The state-level variables were suggested by the work of Goldin (1994) and Goldin and Katz

(1997).
  2 3 Schmidt (1996) tested this assumption and found that the effect of the laws was larger when

matching at this age. Also, because grandfather clauses are common, it is reasonable to think that
the laws in place at age 14 were the laws that would be binding for individuals even when they were



                                                   6
went to school in their state-of-birth. Inevitably some individuals were mismatched.
However, Card and Krueger (1992) show that mobility was low during this period and
that this assumption results in a small error, roughly 10 percent. Furthermore, if such
an error exists, it likely will be uncorrelated with laws on compulsory attendance and
with child labor laws, because these laws were probably not the reason why individuals
moved across states. Indeed the data suggests these laws cannot explain mobility once
we control for education.24

3. Did Compulsory Attendance and Child Labor Laws affect
   schooling? First Stage
The validity of the methodology proposed in this paper rests on the crucial assumption
that compulsory attendance laws and child labor laws can be used as instruments.
This section estimates the Þrst stage, showing that the laws are good predictors of
educational attainment both at the individual and aggregate level. These results will
then be used in the two-stage (IV) estimations in Sections 4 and 5. I also provide
additional evidence here that the laws are good instruments.

3.1. What do we know about Compulsory Attendance and Child Labor
     Laws?
Since their inception in Massachusetts in 1852, compulsory attendance laws have been
complex. They specify a minimum and a maximum age between which school atten-
dance is required; a minimum period of attendance; penalties for non-compliance;
and the conditions under which individuals could be exempted from attending school,
including achievement of a certain level of education (for example the completion of
eighth grade), mental or physical disability, distance from school, and so on. The
most common exemption was for work. Work permits were available even for young
children, generally even younger than the minimum dropout age speciÞed by compul-
sory education laws. Child labor laws, which extensively regulated the employment
of minors, also included several conditions for the granting of such permits and for
exemptions.
    Child labor laws and compulsory attendance laws often were not coordinated.
Each stipulated different requirements for leaving school. For example, in 1924 in
Pennsylvania, the ages for compulsory attendance were 8 to 16, but the child labor
laws allowed 14 year-olds to get work permits and leave school.25 Continuation school
15 or 16 years old.
  2 4 I regressed mobility between state-of-birth and state-of-residence in1960 as a function of educa-

tion, compulsory education laws and all other covariates used in this paper. The F statistic of joint
signiÞcance of the laws has a value of 1.17 (p value of 0.3151), suggesting the laws cannot explain
mobility. Also Lleras-Muney (2001) shows that restricting the sample to those that are still living in
their state-of-birth yields estimates of the effect of the laws that are statistically identical to those
presented here.
  2 5 Work permits evolved over time and today they can be obtained for part-time employment which

does not involve dropping out of school. During this early period however, work permits effectively
allowed children to leave school. See Woltz (1955).



                                                   7
laws, which forced children at work to continue their education on a part-time basis,
were the only laws that attempted to bridge this gap. Compulsory attendance laws
and child labor laws were in place in all states by 1918, and were modiÞed frequently
thereafter.
    There is little agreement regarding the effectiveness of these laws. Landes and
Solomon (1972) analyze the impact of the laws on attendance from 1880 to 1910. They
Þnd that compulsory education laws did not contribute to the increase in enrollments
during this period. They further suggest that states with higher enrollments were
more likely to pass more restrictive laws than other states. Eisenberg (1988) shows
that attendance levels and expenditures per school-aged child were important factors
in explaining the passage of the compulsory attendance laws from 1870 to 1915. Stigler
(1950) and Edwards (1978) look at the impact of the laws on enrollments from 1940
to 1960 and conclude that they were not effective.
    There are a number of studies that support a different conclusion. Margo and
Finnegan (1996) use the 1900 census and Þnd that the compulsory education laws
did have an impact, but only when including measures of child labor laws as well.
Schmidt (1996) Þnds large effects of compulsory education laws on the probability of
high school completion between 1920 and 1934. Lang and Kropp (1986), using data
from 1908 to 1970, show that compulsory education laws affect enrollments, even for
groups not targeted by the laws. Angrist and Krueger (1991) test whether the laws
affected enrollments in 1960, 1970 and 1980 with a difference-in-difference estimator
(by state, quarter of birth and law) and Þnd signiÞcant effects. Lastly, Acemoglu and
Angrist (2000) Þnd that the effects of the laws on educational attainment are positive
and signiÞcant, but note that the effect of child labor laws is larger.26
    Previous studies (including my own27 ) suggest that only three of the many aspects
of these laws had an impact on individual educational attainment: the age at which a
child had to enter school (enter age), the age at which the child could get a work permit
and leave school (work age), and whether or not the state required children with work
permits to attend school on a part-time basis (contsch). Following Acemoglu and
Angrist (1999), I combine the age at which a child had to enter school and the age
required for work permit into a single variable, childcom, deÞned as:

                              childcom = work age − enter age

This variable is the implicit number of years that a child had to attend school, given
that the entering age and the work permit age were enforced. This variable takes the
values of 0, 4, 5, 6, 7, 8, 9, or 10.28 The other variable, contsch, takes the value of 1
if continuation school laws were in place. National trends and tabulations describing
these laws throughout the period of study are shown in Appendix E.
    The period from 1915-1939 is when compulsory education laws (hereafter I refer
to both compulsory attendance laws and child labor laws as “compulsory education
  2 6 Fora detailed review of these studies see Lleras-Muney (2001).
  2 7 SeeLleras-Muney (2001), Angrist and Acemoglu (1999) and Schmidt (1996).
  2 8 Note that there are only a few cases when childcom was 0 (9 states, sometime before 1920). This

occurs only at the beginning of the period if the state had no law that deÞned either entering age or
work permit age.



                                                 8
laws”) are more likely to have affected many individuals.29 Secondary schooling was
experiencing remarkable growth, especially in the Þrst 40 years of this century. Goldin
and Katz (1997) show that the percentage of young adults with high school degrees
increased from 9 percent in 1910 to more than 50 percent in 1940. Also, it has been
suggested by other social sciences that, in the previous period (up to 1915), these
laws were perceived as ineffective; most studies seem to conÞrm that view.30 But
social scientists agree that the laws were enforced by the 1920s31 and Schmidt’s work
(1995)–the only study to concentrate on this period–conÞrms it. Also, note that
Goldin and Katz (1997) show that high school graduation rates were unusually low
in the Second World War years due to the high wages that inexperienced workers
could command. Stigler (1950), Edwards (1978), and partially Angrist and Krueger
(1991)32 suggest that the laws declined in importance after the 1940s. So the Þrst
part of the 20th century provides the perfect window of opportunity for using the
laws as instruments. Finally, from a technical point of view, this period is interesting
because states were constantly changing their compulsory education and child labor
laws, and there is a sufficient amount of variation over time.

3.2. The effect of Compulsory Attendance and Child Labor Laws on edu-
     cational attainment
As preliminary evidence of the effect of these laws on education, I graph the average
education by childcom for the entire sample (Figure 5) and by cohort, for every 5th
cohort in the data (Figure 6). Both graphs show that average education is higher
for those in states where more education was compulsory. In order to add further
controls, I turn to regression analysis.
    Pooling individual data from the 1960 and 1970 census, I estimate the following
model:
                  Eics = b + CLcs π + Xics β + Wcs δ + γ c + αs + εics
The dependent variable is years of completed education for individual i of cohort c
born in state s. CL is a set of dummies for compulsory education laws in place in
state s when the individual was 14, Xics are individual characteristics such as gender
and place of current residence, Wcs is a set of characteristics of individual i’s state-of-
birth at age 14 (such as manufacturing wages, expenditures in education, per capita
doctors, etc.), γ c are cohort dummies, αs are state-of-birth dummies. The regression
also includes interactions between region-of-birth and cohort, an intercept (b) and a
  2 9 Schmidt  (1996) conÞrms this intuition.
  3 0 Katz  (1976) and Ensign (1921) suggest that in the 19th century and early 20th century laws
were created but not enforced. Many state laws did not even provide enforcement mechanisms, and
if they did, there were often insufficient means to enforce them, especially in rural areas.
   3 1 Truant officers had become commonplace. They were in charge of making sure that students

of age were effectively in school, and they could penalize the parents in cases of non-compliance.
Also, expenditures on education increased. Although it would be hard to argue that this increase
was solely the result of the passage of compulsory education laws, it is certainly true that the laws
required that the states provide public schools and pay for enforcement agencies. See Tyack (1974),
Katz(1976).
   3 2 They Þnd that the impact of the laws is about 4% in 1960, 2% in 1970 and 0.5 in 1980.




                                                 9
dummy for 1970. I also estimate the model by aggregating the data at the state-
of-birth/cohort and gender level.33 Both estimations will be used in Section 4 (Þrst
stage).
    Table 3 shows the results. The Þrst column estimates the relationship including
only state effects, cohort effects, a female dummy, and a set of dummies for the
laws. The coefficients are fairly robust to the addition of other controls (see column
2).34 The last column shows the results from estimating the equation using the
data aggregated at the state-of-birth, cohort and gender level. The estimations show
that the laws increased the educational attainment of individuals. As expected, all
dummies for the laws are positive and signiÞcant and they generally increase as the
number of compulsory years increases. Overall, the implied increase in educational
attainment due to childcom is around 4.8 percent.35 The effect is identical if the
sample is restricted to include only those observations for which childcom is not 0.36
This estimate is similar to those reported by Acemoglu and Angrist (2000), who
report an increase between 1 and 6 percentage points; by Eisenberg (1988), who Þnds
an effect of about 2 percent; and by Angrist and Krueger (1991), who Þnd that the
impact of the laws was about 4 percent in 1960.37 Also, the continuation school
dummy is positive.38
    Before turning to the effect of education on mortality, I present evidence that the
laws are good instruments. At the bottom the Table 3, I report the F-test of joint
signiÞcance of the laws; it shows that the laws are always jointly signiÞcant at the
5% level for both speciÞcations. Additionally the F-statistic is greater than (or very
close to) 5, which suggests that the instruments are strong. I also report the partial
R-squared coefficient, another measure of the instruments’ strength.39 It has a value
of 0.0001 or higher, which compares favorably to those reported by Bound, Jaeger,
and Baker (1995).
    It is also worth pointing out that the changes in the laws that took place during
  3 3 The   model estimated would be:
                         Egcs = b + CLcs π + Xgcs β + Wcs δ + γ c + αs + εgcs

  where Egcs is the average education in a given state, cohort and gender; and Xgcs are the average
characteristics of that group. The number of individuals in each group are used as weights.
  3 4 Inclusion of other variables, such as income, immigrant status of parents, and so on, has no

impact on them. Also, identical regressions by region-of-birth or by gender yield similar results. See
Lleras-Muney (2001) for these results.
  3 5 This was calculated by replacing the set of dummies by the continuous variable childcom. See

Lleras-Muney (2001).
  3 6 See Lleras-Muney (2001).
  3 7 Schmidt (1996) Þnds much larger effects, about 20% for her analysis of New York State.
  3 8 Continuation school is not signiÞcant in this sample, but previous work (see Lleras-Muney, 2001)

showed that this law affected white males and individuals born in the north and south of the U.S.
Therefore I include it.
  3 9 There is a large literature on the problem of weak instruments. Bound, Jaeger, and Baker (1995)

suggested that the researcher evaluate the quality of the instruments by looking at two statistics.
First, the F statistic on the excluded instruments in the Þrst stage should be statistically signiÞcant
and large. Staiger and Stock (1997) further suggests that a value of less than 5 could signal weak
instruments (this is a rule of thumb). Second, the partial r-square (obtained by regressing schooling
on the instruments, once the common variables have been partialled out) should be high. Following
their suggestion, these two statistics are reported here.


                                                  10
this period appear to have been exogenous to individuals. Although different states
might have had different tastes for education, the regressions here include a very large
set of controls (cohort dummies, state-of-birth dummies and region-of-birth*cohort
interactions are included) which should capture this effect. Also note that the addition
of controls (compare column 1 and 2 of Table 3) has little effect on the coefficients
of the laws, suggesting that any excluded state-of-birth/cohort level variables are
not correlated with the laws. Furthermore, Lleras-Muney (2000) presents evidence
consistent with exogenous laws: her results suggest that the laws impacted only the
lower end of the distribution of education. She rejects the hypothesis that changes in
the laws during this period resulted from (rather than caused) increases in education,
using a test inspired by Landes and Solomon (1972).40
    A Þnal concern is that the laws must affect individual health only through their
effect on education. There is no evidence that the laws included any clauses or re-
strictions that would have affected health independently. For example, there were no
lunch programs provided as part of school attendance. Also the states that led in
education during this period (the prairie states41 ) were not the same states that led
in health (northeastern states).42 But again, the controls included here are meant to
rule out this possibility. Finally, exogeneity tests are performed in the IV estimation
(see next section).
    Overall the results show that the laws did have an impact on educational achieve-
ment, and that their predictive power is large. The important implication is that
compulsory education laws can be used as instruments. Therefore I turn now to the
question of the effect of education on mortality.

4. Health and Education: Econometric model
4.1. Least Squares Estimation
The econometric model for the relationship between education and health can be
written as a linear system of simultaneous equations:

                                  Hi    = X1i β 1 + Ei π1 + ε1i                                   (4.1)
                                  Ei    = X2i β 2 + Hi π2 + ε2i                                   (4.2)

    H i is individual i’s health stock, E is his education level, X 1 is a vector of in-
dividual characteristics that affect health, such as smoking, and genetic factors. X 2
is a vector of individual characteristics that determine education, such as ability. X 1
  4 0 The test consists of matching individuals to the laws in place in their state-of-birth when they

were 17, 18 and up to 26 years of age, when these laws should no longer have affected them. Lleras-
Muney (2000) Þnds that future laws cannot explain educational attainment, whereas laws at age 14
can.
  4 1 See Goldin and Katz (1997).
  4 2 Starr’s 1982 book provides anecdotal evidence that the northern states lead in a variety of health

aspects. My own data supports this conclusion. For example, the north had the highest number of
doctors per capita throughout the period. And the number of doctors per capita in the north did
not decline from 1915 to 1939 but did decline in the rest of the country. The north also had the
highest declines in infant mortality rates during this period. (Results available upon request.)



                                                  11
and X 2 may contain common factors. This general speciÞcation allows for causality
to run from education to health and vice-versa.
    The purpose of this paper is to determine only whether or not education affects
health (i.e. π1 = 0?). Therefore I only estimate the health equation (equation 3.1).
Although health is unobserved, mortality is observable. Following Grossman’s (1972)
model of health, death occurs when the stock of health falls bellow a certain threshold.
In a less deterministic model, H i is proportional to the underlying probability (index)
of being alive, and death is the observed result. This is the usual limited-dependent-
variable set-up.
    This mortality equation can be estimated at the individual level using the NHEFS
but not the census. If individuals could be followed from the 1960 census to the 1970
census (or from 1970 to 1980), then (based on the previous discussion) the following
individual linear probability model could be estimated:
                  Dt,ics = b + Eics π + (Xt−1 )ics β + Wcs δ + γ c + αs + εics                   (4.3)
where Dti is equal to one if the individual is deceased at time t. E ics is i’s education
(measured by completed years of education), Xt−1 are other individual characteristics
measured as of (t-1) (including gender), Wcs is a set of characteristics of individual
i’s state-of-birth at age 14, γ c is a set of cohort dummies, αs is a set of state-of-birth
dummies, b is an intercept, and ε is the error term.
    Using the census, individuals cannot be tracked over time, but I can track groups
that are constant over time, and calculate their death rates by aggregating the data.
I aggregate by gender, cohort, and state-of-birth. This aggregation level uses all of
the available individual characteristics that are time invariant (except for education),
and therefore it maximizes the number of observations in the aggregate data.
    The aggregate model is derived from the individual model by averaging over indi-
viduals in a given gender/cohort/state-of-birth group as follows:
                  Dtgcs = b + E gcs π + (Xt−1 )gcs β + Wcs δ + γ c + αs + εgcs                   (4.4)

where Dtgcs represents the proportion of individuals that died in a given group or
the death rate for that group, and Xt−1 gcs represents the average characteristics of
that group at (t-1 ) (for example, the the percentage of people in that group living in
urban areas).43
    Note that I use a linear probability model for the estimation. The existence of
negative death rates makes it impossible to use a non-linear model such as a Logit
or Normit. However, since the dependent variable (the death rate) is not censored
below by 0, the linear probability assumption is less problematic in this case than in
general.44
    In the linear probability model, the error term is heteroskedastic and has the
following variance:
                                        Dtgcs (1 − Dtgcs )
                              var(ε) =                                            (4.5)
                                               ngcs
  4 3 Including
              a dummy for gender.
  4 4 Furthermore,in the next section, I will test this assumption by comparing results from the census
to those obtained from individual data.


                                                  12
where ncse is the number of individuals in that group and Dtgcs is the observed prob-
ability (death rate) for that group. A standard estimation procedure (the minimum
chi-square method45 ) in this case is to run weighted least squares, where the weights
                   p
are given by 1/ var(ε). Again, due to random sampling and the error it gener-
ates, the observed probabilities can be negative, so this estimation is not possible.46
In order to address the heteroskedasticity problem I estimate the equation by GLS
(weighted least squares) using the number of individuals in the group as weights. To
correct for further heteroskedasticity, I use White’s sandwich estimator.47
    It is intuitive that GLS estimates at the aggregate level will be biased and incon-
sistent, since the correlation between the error term and education (at the individual
level) will carry over when calculating group means.48 Now I turn to the Instrumental
Variables (IV) estimation which will yield consistent estimates of the causal effect of
education on mortality.

4.2. Efficient Wald estimates
One obvious solution to correct for the bias in the GLS coefficient is to use Instru-
mental Variables (IV). Given that many instruments are available, Two Stage Least
Squares (2SLS) would be the preferred estimation method. At the individual level,
the 2SLS model is:
               Dti     = b + Eics π + (Xt−1 )ics β + Wcs δ + γ c + αs + εi
               Eics    = b + CLcs π + (Xt−1 )ics β + Wcs δ + γ c + αs + εics
where D is equal to one if the individual is deceased at time t. E is i’s education
(measured by completed years of education), Xt−1 are other individual characteristics
measured as of (t-1) (including gender), Wcs is a set of characteristics of individual
i’s state-of-birth at age 14, γ c is a set of cohort dummies, αs is a set of state-of-birth
dummies, b is an intercept, and ε is the error term, which is assumed to be normal
N(0,σ 2 I). CL is the set of compulsory education laws that serve as instruments to
       1
identify the education equation. This model can be estimated using the individual
NHANES data but not with the census.
    Since the census data can be used only as grouped data, the Wald estimator is an
alternative estimator for the effect of education. Angrist (1991) showed that the Wald
estimator for grouped data is efficient and in fact equivalent to 2SLS using individual
level data. In the case of many explanatory variables the efficient Wald estimator is
found by GLS estimation of the following equation:
                                 Dcrl = E crl π + γ c + δ r + εcrl                              (4.6)
  4 5 See Maddala p. 29, Green p. 895.
  4 6 An  idea is to convert the negative weights into 0. Notice though this procedure will introduce
bias in the results: I observe both underestimated and overestimated death rates, but by converting
the negative (underestimated) ones into 0, I am “Þxing” the problem only for half of the observations.
Alternatively, one can run weighted least squares only using the observations for which the observed
death rate is positive, but again this will introduce the same kind of bias.
  4 7 In all the estimations, including the IV estimations, where state-of-birth characteristics are

included, the standard errors are also clustered at the state-of-birth and cohort levels.
  4 8 Proof available upon request.




                                                 13
where Dcrl is the death rate for individuals born in cohort c in region r under compul-
sory law l, and E csl is the average education of individuals born in cohort c in region
r under compulsory law l. The weights are given by the population in each group.
In other words, Wald is estimated by grouping the data by gender/cohort/region-of-
birth and compulsory education law. This procedure is equivalent to 2SLS at the
individual level, where cohort dummies γ c and region dummies δ r serve as their own
instruments (since they are exogenous), and compulsory education laws serve as in-
struments for education, the endogenous variable. The estimates are referred to as
the efficient Wald estimates.
    Note that because compulsory education laws are deÞned at the state-of-birth
and cohort level, I cannot control for both state-of-birth and cohort when using this
estimator. This is a drawback of the Wald estimator, especially if one thinks that
state-of-birth and the laws are correlated. In order to alleviate this problem, I control
instead for region-of-birth. But region-of-birth may not be a good proxy for state-of-
birth. Furthermore, other individual (Xt−1 ) and state-of-birth characteristics (Wcs )
cannot be included in this speciÞcation.

4.3. Two-Stage Least Squares with Aggregate Data
Alternatively, I can estimate the 2SLS model at the data that has been aggregated at
the state-of-birth/cohort and gender level. Estimation at the aggregate level results in
less efficient estimates (see Green pp. 433-434) but all the covariates (especially state-
of-birth) can be included. Using the aggregate data 2SLS is obtained by estimating
the following model:

                Dgcs     = b + E gcs π + X gcs β + Wcs δ + γ c + αs + εgcs
                E gcs    = b1 + CLcs π + X gcs β + Wcs δ + γ c + αs + εgcs

where now Dtgcs is the proportion of individuals who died in a given gender/cohort
and state-of-birth, E gcs is the average education of that group and (Xt−1 )gcs are other
average characteristics. Again, the weights are given by the number of observations
in each cell, and the excluded instruments from the mortality equation are the com-
pulsory education dummies, CLcs . The Þrst stage (estimation of E gcs ) was shown in
the previous section.

4.4. Mixed Two-Stage Least Squares Estimation
The census allows me to estimate the Þrst stage using individual data. The intuition
behind Mixed-2SLS is that it might be possible to take advantage of this fact and
gain efficiency (relative to the previous 2SLS) by estimating the Þrst stage at the
individual level (as done in the previous section) and then aggregating the data by
gender/cohort/state-of-birth. (See Dhrymes and Lleras-Muney, 2001.)49
  4 9 Dhrymes and Lleras-Muney (2001) compare 2SLS and Mixed 2SLS estimators. The question of

which estimator has lower variance turns out to be data dependent, but it is possible for Mixed 2SLS
to be more efficient.




                                                14
    Mixed 2SLS is obtained by estimating the following equation through weighted
least squares:

                                b
                     Dgcs = b + E gcs π + (X)gcs β + Wcs δ + γ c + αs + εgcs

    Again Dgcs represents the proportion of individuals who died in a given gen-
der/cohort and state-of-birth group, X gcs represents the average characteristics of
                                 b
the group, but now I include E gcs , the average predicted education for that group
from the Þrst stage.50 The weights are given by the number of observations in each
cell. The excluded instruments from the mortality equation are the compulsory ed-
ucation dummies. The only difference between standard 2SLS with aggregated data
and Mixed-2SLS is in the predicted education term. 2SLS uses predicted average
education whereas Mixed-2SLS uses average predicted education.
    More formally, deÞne H as the matrix that transforms the data into group means
                                                                                  b
and weights each group mean by the number of individuals in the group, and let X con-
tain all the same variables as in the GLS estimation, but with education replaced by
                                                                    h                   i
                                                                 b    b
the predicted level of education from the Þrst stage regression (X = E | Xt−1 | γ c | αs .
Then the estimator β Mixed can be expressed as:
                                         ³            ´−1
                                           b        b
                                β Mixed = X 0 H 0 H X     b
                                                          X 0 H 0 HD

  This procedure also results in consistent estimates. As usual the variance-covariance
matrix needs to be corrected.

5. Results
5.1. Least Squares Results
Although we have good reason to believe that GLS produces biased estimates, I report
them here as the benchmark for comparison with the IV results. Using the census, I
estimate the GLS model described above. The results are in the Þrst column of Table
4. The estimated coefficient of the effect of education on the death rate is about
-0.012. The coefficient is highly signiÞcant and is is robust to the inclusion of more
controls.51 It is a well known fact that there exist persistent differences in mortality
rates by gender. I therefore repeat the analysis by gender (Table 5). The effect of
education is positive and signiÞcant, but this is probably due to the small sample size.
    The validity of the aggregation procedure rests on the assumption that the ag-
gregate data can be understood as coming from unobserved individual data. It is
important that this intuition be conÞrmed, so I compare aggregate results from the
census with those obtained with the NHEFS individual data. This comparison allows
me to check the validity of the linear probability assumption and helps me to interpret
the aggregate results.
 5 0 The   expression for the Þrst stage was given in the previous section.
 5 1 Results  available upon request.



                                                   15
    Using individual NHEFS data, I estimate a linear probability model and a probit
model, where the dependent variable is a dummy indicating whether or not the person
died between 1975 and 1985. Then I aggregate the NHEFS data by gender, state-
of-birth, and cohort and again estimate the same linear model estimated with the
census data. Because of the small number of observations in the NHEFS aggregating
by gender, state-of-birth and cohort results in very few observations per cell, so I also
reproduce the results only aggregating by state-of-birth and cohort. The results are
shown in Table 4.
    Comparing the results from LS regressions from the census with results from the
NHEFS shows that the census data gives extremely accurate estimates of the effect
of education. The census LS estimates are very similar to those obtained using the
NHEFS aggregated data, which in turn are similar to those obtained at the individual
level, using either LS or probit estimations.
    These results suggest that sampling (and the measurement error it generates) does
not signiÞcantly affect the estimates for education, that there is no aggregation bias
and that the linear model is a good approximation of the education-death rate rela-
tionship. The comparison is also useful in terms of interpretation: a -0.012 coefficient
for education means that increasing the education of a given cell by one year lowers
its death rate by 1.2 percentage points. This coefficient also implies that increasing
an individual’s education by one year will lower his probability of dying between 1960
and 1970 (or between 1970 and 1980) by 1.2 percentage points. This latter interpre-
tation is more intuitive and useful. Note again that the OLS effect is quite large: at
the mean, this result implies that a 10 percent increase in education lowers mortality
by about 11 percent, therefore an elasticity of about -1.

5.2. IV results
The Þrst column in Table 6 presents the 2SLS results using the NHEFS. This estima-
tion is done at the individual level and using standard 2SLS. The estimate is positive
(the effect of education is about -0.02) but not signiÞcant: because this sample is
small, the Þrst stage estimation52 is poor. Nonetheless, although the standard er-
rors are high, the estimates from this sample are also larger than the GLS estimates
obtained from the same data.
    The second column shows the results from the Wald estimation. The Wald esti-
mate of the effect of education is about -0.037 and signiÞcant at the 5 percent level.
The third column presents the results of 2SLS estimation using aggregated data at
the gender/state-of-birth-and cohort level. The effect of education is about -0.045
and signiÞcant at the 10 percent level. The Mixed 2SLS results (last column) show
that the coefficient on education is approximately -0.059 and is signiÞcant at the 5
percent level. All of the previous estimates are signiÞcant at the 5% level using a one-
tailed test (the null hypothesis is that education is negative) which is perhaps more
appropriate in this set-up.53 Overall the results suggest that increasing education by
  5 2 Results not shown here but available upon request. In the Þrst stage estimation using the

NHEFS, only two of the dummies for compulsory education laws were signiÞcant at the 10% level,
and the set of dummies was jointly insigniÞcant.
  5 3 I thank Michael Grossman for this insight.




                                              16
one additional year lowers the 10-year death rate by at least 3.6 percentage points.54
    For the last two estimators I perform a test of overidentifying restrictions. The
χ2 statistic for the aggregate 2SLS model is 2.42 and 1.49 for the Mixed 2SLS model.
This statistic tests the hypothesis that the model is well speciÞed. It is calculated
as the sample size times the R2 from a regression of the residuals from the second
stage on all exogenous variables, including the instruments. In both cases the overi-
dentifying restrictions are not rejected at a 5 percent level (critical value 14.06). This
test in conjunction with earlier results from the Þrst stage suggests that compulsory
education laws are legitimate instruments.
    As a last attempt to address the potential endogeneity of the laws, I repeat the
estimations above using a larger set of instruments that include quarter of birth, com-
pulsory attendance and child labor laws, and the interactions of quarter of birth and
the laws. Presumably these individual-level instruments will increase the efficiency of
the estimates, and they are perhaps less likely to be endogenous.55 The results (Table
7A) are identical to those presented above.
    In table 7B I present the reduced form estimates, i.e. the direct effect of the laws
on mortality. The results are consistent with previous estimations: if the effect of
childcom on education is about 5 percent, and the effect of education on mortality is
about 6 percent, then the direct effect of the laws on mortality should be about 0.3
percent, which is approximately what the reduced form result shows.
    In table 7C I present the results excluding ages 40, 50 and 60 since the data showed
evidence of age heaping. This is a potential problem if age heaping is correlated with
education. The IV results are very similar to the previous results.
    The results by gender from the Census are presented in Table 8. The coefficient
on education is somewhat smaller for females than for males, conÞrming the Þndings
in the literature that the effect of education is larger for males. These results are
interesting for other reasons. First, they suggest that World Wars I and II did not
results in signiÞcant selection bias for men. Also, in these estimations the effect of
marriage is negative as the literature suggest, whereas the effect is positive in the
joint estimations. This is a composition effect: males are both more likely die and to
be married.
    This section has presented four different estimates of the effect of education on
mortality. Three different estimators, using two different data sets and three different
levels of aggregation, were used. Although each estimate has weaknesses, all estimates
point to the same conclusion: the effect of education is causal and in fact larger than
OLS suggests. Given this variety of estimates, this result is very robust.
   5 4 Estimates by region are comparable in size to those presented here except that they are generally

not signiÞcant.
   5 5 Note however that the use of these instruments might be questionable (see papers in footnote

10). Also Lleras-Muney (2001) shows for example that the laws affected whites but not blacks.
However quarter of birth does appear to affect blacks’ educational attainment. This again raises the
issue of whether quarter of birth has an independent effect on education unrelated to compulsory
attendance laws.




                                                  17
5.3. Discussion
The results are surprising for two reasons. The Þrst is that the IV estimates are larger
than the LS estimates. The second is that the effect of education is quite large. In
this section I discuss these two issues.
    In all the IV estimations presented here, the effect of education is much larger
than the LS estimates suggest. The Mixed 2SLS estimates suggest the effect is as
large as -0.058, whereas Wald estimates imply a coefficient of about -0.036. All IV
estimates are larger than LS estimates. At Þrst, this could seem to be a surprising
result: the a priori expectation was that LS estimates would be too large. However,
in the vast literature devoted to the earnings returns to education, researchers have
come to similar conclusions: OLS estimates of the effect of education on earnings are
too small.56
    One explanation is that the omitted variable bias is smaller than the bias that
results from measurement error in education.57 The health literature has not been
concerned with this potential problem although there is evidence of measurement error
in education (Card 1995). If the measurement error is random, then IV estimate will
be larger than the OLS estimate. 58
    Another explanation is the choice of instrument. Card (2000) suggests that one
possible reason why IV estimates of the return to education are generally larger than
OLS is that most instruments are based on policy interventions that affect the edu-
cation choices of individuals with low levels of education. Under the assumption that
different individuals face different returns to education due to unobserved character-
istics, IV estimates reßect the marginal rate of return of the group affected by the
policies (Imbens and Angrist, 1994; Angrist, Imbens and Rubin, 1996). If individuals
choose low levels of education because they face high costs (rather than low returns)
then the 2SLS estimates accurately measure those higher returns. In the context of
the health returns to education, the results suggest that the individuals affected by
the laws also face higher health returns to education than the rest of the population.
This would not be surprising if one believes that the health returns to education are
larger at lower levels of education.
    To Þnd suggestive evidence to support this claim, I re-estimate the model including
a quadratic term for education. If the health returns to education are decrease as
education increases then this quadratic term should be positive (since the effect of
education on mortality is negative). Table 9 shows the results. The quadratic term
is postive, suggesting that indeed the health returns to education are larger for lower
levels of education. Also note that I cannot reject the hypothesis that education
is exogenous using a Hausman test, which also supports the hypothesis that IV is
measuring the effect for the bottom half, and that effect is causal.
    Larger IV returns can also be explained if there exist health externalities from
education. Lleras-Muney (2001) shows that compulsory education and child labor
  5 6 Fora survey of these studies, see Card (1995).
  5 7 As Card (1995) mentions, the idea that measurement error bias could be just as serious as the
omitted variables bias in the returns to education was Þrst noted by Griliches (1977).
  5 8 Note however that if measurement error is not classical (non random) then IV estimates can also

be biased (see Hyslop and Imbens, 2000 and Kane, Rouse and Staiger, 1999)



                                                 18
laws affected those at the lower end of the distribution of education decreasing in-
equality in education. There is a large literature that indeed suggests that inequality
affects health.59 Again, to Þnd suggestive evidence for this hypothesis, I estimate the
OLS model only for those with more than 12 years of schooling,60 including now the
standard deviation of the distribution of education in their state-of-birth and cohort
(Table 10). The effect of the standard deviation of education is positive and signiÞ-
cant, so that higher inequality results in higher death rates (lower inequality results
in lower death rates).
    The second issue is that the effect of education is quite large, and it is important
to understand why. Education provides individuals with critical thinking skills, which
in turn might affect understanding of health risks (Grossman’s hypothesis). If this
is the case, then the interaction of education with a variety of factors is relevant.
For example, although access to information alone cannot explain health differences
across education groups (Kenkel, 1991), information available to the more educated
will result in greater beneÞts for them if they can understand it better, or can under-
stand its relevance.61 Another possibility is that the more educated might be more
likely to adopt and implement new medical technologies.62 Since both the availability
of information and the rate of medical innovation dramatically increased in the last
century, it is reasonable to think that the more educated were able to capture very
high returns during this period. At the same time, this would explain why individuals
did not voluntarily acquire education, in spite of the large returns: the increases in
medical technology and information were not foreseeable at the time they made their
education choices. Other direct mechanisms are documented in the cognitive psychol-
ogy literature: lack of education is correlated with stress, depression and hostility, all
of which have been shown to adversely affect health (Adler et al, 1994).
    There are a few other indirect mechanisms through which education might affect
health which are also consistent with the results in this paper. One obvious one is that
being in the classroom is less of a health risk than working, especially while growing
up. Also note that education gives you access to a higher income and different types
of jobs, both of which affect health. For example, only high school graduates in the
Þrst half of the century had access to white collar jobs, which provided healthier work
environments than manufacturing or agriculture.63 Controlling for income (or occu-
pation) does not change the results in this paper. But, since income is endogenous,
it is not possible (given that I have no instruments for income) to distinguish the
direct effect of education on health from its indirect effect through income. However,
a few pieces of evidence suggest that income alone might not be the sole mechanism.
  5 9 Deaton  and Paxson (1999) review the existing literature in their paper.
  60 I restrict the sample to differentiate the impact of inequality from that of own education. Note
that the sample restriction however is not necessarily appropriate.
  6 1 As a consequence they might seek care earlier, get more medical care, get more preventive care,

be more willing to use newly developed medical procedures/medicines, and so on.
  6 2 This idea was Þrst postulated by Nelson and Phelps (1966). Bartel and Lichtemberg (1987)

provide evidence at the plant level that highly educated workers have a comparative advantage with
respect to the adjustment to and implementation of new technologies.
  6 3 Different/better jobs might provide access to health insurance. Note however that health insur-

ance has not been proven to impact health (see introduction).




                                                 19
Grossman (1975) showed that the effects of income on health disappear once a certain
level of income has been reached, while the same is not true for education. Standard
results suggest that the returns to education are about 10% and that the elasticity
of mortality with respect to income is about -0.3.64 If the sole effect of education is
through income, one more year of education should decrease mortality by 0.0033 (for
average mortality of 0.11), which is a much smaller effect than what was estimated
here.
    Finally, the results in this paper do not imply that time preferences do not affect
health and education choices nor that there is no reverse causality from health to
education. They simply show that there is a causal effect of education on health,
and that this effect is not due to time preferences. However, as Becker and Mulligan
(1997) argue, education could lower the discount rate, making people more patient.65
This is yet another indirect mechanism that could explain my results.

6. Conclusion
This paper has shown that there is a large causal effect of education on mortality.
In fact, this effect is underestimated by OLS. Instrumental variables estimates show
that one more year of education decreases the probability of dying within 10 years by
at least 3.6 percentage points. To better understand the impact of education, using
the coefficient from the Wald estimation, I calculate how this effect translates into
life expectancy gains. I Þnd that in 1960, one more year of education increased life
expectancy at age 35 by at least 1.2 years. This is a very large increase.
     A few notes of caution on how to interpret these results for public policy purposes
are necessary. First, in order to make policy recommendations, we need to know
more about the speciÞc mechanisms by which education affects health. This paper
analyzes the effects of increasing education from relatively low initial levels. It is
unclear what the effects would be at higher initial levels of education. The average
education level for white Americans born in 1901 was at most 8.87 years.66 Today
many developing countries, including most Latin American countries,67 have average
levels of education that are similar. This paper implies that more aggressive education
policies could dramatically increase adult longevity in such countries. But cost beneÞt
analysis of such policies are extremely complex, since for example we do not know
what the cost of increasing education would be, or its effectiveness. Questions such
as these are beyond the scope of this paper. But the results presented here suggests
that the beneÞts of education are large enough that we need to consider education
policies more seriously as a means to increase health, especially in light of the fact
that other factors, such as expenditures on health, have not been proven to be very
effective. Finally the results also suggest that the measured effect of technology on
  6 4 Deaton   and Paxson (1999).
  6 5 This point is cited by Grossman (1999)
  6 6 This is the average education level of that cohort in 1960. Data for the entire population in 1901

does not exist for the US, but the average was probably much lower.
  6 7 Average education level of 25 year-olds in many Latin American countries was between 6 and

9: Bolivia, 8; Chile, 8.79; Ecuador, 7.12, Mexico, 6.23; Panama, 8.68; Peru, 7.2; Uruguay, 8.02 and
Venezuela, 7.15 (Source: IDB, 1998).


                                                  20
health might actually reßect the effect of increased education rather than the effect
of technology.
    This evidence that education increases life expectancy implies that the returns to
education, measured only in terms of earnings increases, substantially underestimate
the true returns to education. In view of the large magnitude of the effect of education
on health, it is clear that more attention needs to be devoted to the pathways of
inßuence. Existing models of the relationship between education and health are very
imprecise about the mechanisms through which education operates on health. It
is crucial that we understand these mechanisms better, so that we can implement
effective programs to increase the health of our population.

References
 [1] Acemoglu, Daron and Joshua Angrist, “How Large are the Social Returns to
     Education? Evidence from Compulsory schooling Laws,” NBER Working Paper
     No. W7444, December 1999
 [2] Adler, Nancy E. et al, “Socioeconomic Status and Health, the Challenge of the
     Gradient,” American Psychologist, vol 49, No. 1January 1994
 [3] Angrist, Joshua D., “Grouped Data Estimation and Testing in Simple Labor
     Supply Models,” Journal of Econometrics, February-March 1991
 [4] Angrist, Joshua D. and Alan B. Krueger, “Does Compulsory School Attendance
     Affect Schooling and Earnings?,” Quarterly Journal of Economics, November
     1991
 [5] Angrist, Joshua D., Guido W. Imbens and Donald B. Rubin, “IdentiÞcation of
     Causal Effects Using Instrumental Variables,” Journal of the American Statistical
     Association, June 1996
 [6] Auster, Richard, Irving Leveson and Deborah Sarachek, “The Production of
     Health, An Exploratory Study,” Journal of Human Resources 4, 1969
 [7] Bartel, Ann P. and Frank R. Lichtenberg, “The Comparative Advantage of Ed-
     ucated Workers in Implementing New Technology,” Review of Economics and
     Statistics, February 1987
 [8] Becker, Gary S. and Casey B. Mulligan, “The Endogenous Determination of
     Time Preference,” Quarterly Journal of Economics, August 1997
 [9] Berger, Mark C. and J. Paul Leigh “Schooling, Self Selection and Health,” Jour-
     nal of Human Resources 24, 1989
[10] Behrman, Jere R., Robin C. Sickles, and Paul Taubman , Causes, Correlates and
     Consequences of Death among Older Adults: Some Methodological Approaches
     and Substantive Analysis, 1998



                                          21
[11] Bound, John, David A. Jaeger and Regina Baker, “Problems with Instrumental
     Variables Estimation when the Correlation Between the Instruments and the
     Endogenous Explanatory Variables is Weak,” Journal of the American Statistical
     Association, 90, June 1995
[12] Bound, John, and David A. Jaeger, “On the Validity of Season of Birth as an
     Instrument in Wage Equations: A Comment on Angrist and Krueger’s “Does
     Compulsory School Attendance Affect Schooling and Earnings”,” NBER Work-
     ing Paper 5835, November 1996
[13] Card, David, “Estimating the Returns to Schooling: Progress on some Persistent
     Econometric Problems,” NBER Working Paper no.W7769, June 2000
[14] Card, David, “Earnings, Schooling, and Ability Revisited,” Research in Labor
     Economics, vol. 14, 1995
[15] Card, David and Alan Krueger, “Does School Quality Matter? Returns to Edu-
     cation and the Characteristics of Public Schools in the United States,” Journal
     of Political Economy 100, January 1992
[16] Christenson, Bruce A., and Nan E. Johnson, “Educational Inequality in Adult
     Mortality: An Assessment with Death CertiÞcate Data from Michigan,” Demog-
     raphy 32. May 1995
[17] Currie, Janet and Rosemary Hyson, “Is the Impact of Health Shocks Cushioned
     by Socioeconomic Status? The Case of Low Birth weight,”American Economic
     Review, May 1999
[18] Deaton, Angus and Christina Paxson, “Mortality, Education, Income and In-
     equality among American Cohorts,” NBER Working Paper 7140, May 1999
[19] Dhrymes, Phoebus and Adriana Lleras-Muney (2000), “Estimation of Models
     with Group Data by Means of ‘2SLS’,” mimeo, Columbia University, 2001
[20] Duleep, H.O., “Measuring the Effect of Income on Adult Mortality Using Longi-
     tudinal Administrative Record Data,” Journal of Human Resources 21, 1986
[21] Edwards, Linda N., “An Empirical Analysis of Compulsory Schooling Legislation,
     1940-1960,” Journal of Law and Economics, April 1978
[22] Eisenberg, M.J. (1988) “Compulsory Attendance legislation in America, 1870 to
     1915,” Ph.D. Dissertation, University of Pennsylvania
[23] Elo, Irma T. and Samuel H. Preston, “Educational Differentials in Mortality:
     United States, 1979-85,” Social Science and Medicine 42(1), 1996
[24] Ensign, Forest Chester, “Compulsory School Attendance and Child Labor,” Iowa
     City, IA: The Athens press, 1921
[25] Farrell, P. and Victor R. Fuchs, “Schooling and Health: The Cigarette Connec-
     tion,” Journal of Health Economics, 1982

                                        22
[26] Feldman J. D. Makuc, J. Kleinman and J. Cornoni-Huntley, “National Trends in
     Educational Differences in Mortality,” American Journal of Epidemiology, 1989
[27] Filmer, Deon and Lant Prichett, “Child Mortality and Public Spending on
     Health: How Much does Money Matter?,” World Bank Policy Research Working
     Papers 1864, December 1997
[28] Fuchs, Victor R., “Time Preference and Health: An Exploratory Study,: in Vic-
     tor Fuchs, Ed., Economic Aspects of Health, Chicago: The University of Chicago
     Press, 1982
[29] Gilleskie, Donna B. and Amy L Harrison, “The Effect of Endogenous Health
     Inputs on the Relationship between Health and Education,” Economics of Edu-
     cation Review, June 1998
[30] Goldin, Claudia, “How America Graduated From high School: 1910 to 1960,”
     NBER Working Paper No 4762, 1994
[31] Goldin, Claudia and Lawrence Katz, “Why the United States Led on Education:
     Lessons from Secondary School Expansion, 1910 to 1940,“ NBER Working Paper
     6144, August 1997
[32] Green, William H., Econometric Analysis, 3rd Edition, Prentice Hall, New Jersey
     1997
[33] Griliches, Zvi, “Estimating the Returns to Schooling: Some Econometric Prob-
     lems ,” Econometrica, January. 1977
[34] Grossman, Michael, “The Human Capital Model of the Demand for Health,”
     NBER Working Paper 7078, April 1999
[35] Grossman, Michael, “The Correlation between Health and Schooling,” in House-
     hold Production and Consumption, Ed N. E. Terleckyj, Studies in Income and
     Wealth, Vol. 40, Conference on Research in Income and Wealth. New York:
     Columbia University Press for the National Bureau of Economic Research, 1975
[36] Grossman, Michael, “The Demand for Health: A theoretical and Empirical In-
     vestigation,” New York: Columbia University (for the NBER), 1972a
[37] Grossman, Michael, “On the Concept of Health Capital the Demand for Health,”
     Journal of Political Economy 80, 1972b
[38] Grossman, Michael and R. Kaestner “Effects of Education on Health,” in J.R.
     Berhman and N. Stacey Eds. The Social BeneÞts of education, University of
     Michigan Press, Ann Arbor, 1997
[39] Haines, Michael R. and Samuel H. Preston, “The Use of the Census to Estimate
     Childhood Mortality: Comparisons from the 1900 and 1910 United States Census
     Public Use Samples,” Historical Methods, Vol 30, no.2, Spring 1997


                                        23
[40] Harmon, Colm, and Ian Walker, “Estimates of the Economic Return to School-
     ing for the United Kingdom,” American Economic Review, Volume 85, Issue 5,
     December 1995
[41] Hoek, H. W., A.S. Brown and E. Susser, “The Dutch Famine and Schizophrenia
     Spectrum Disorders,” Social Psychiatry and Psychiatric Epidemiology, Volume
     33, Issue 8, 1998
[42] Hyslop, Dean R. and Guido W. Imbens, “Bias from Classical and Other Forms
     of Measurement Error”NBER Technical Working Paper, August 2000
[43] Inter-American Development Bank. Facing up to Inequality in Latin America:
     Economic and Social Progress in Latin America. 1998-1999 Report. Distributed
     by John Hopkins University Press, Washington D.C., 1999
[44] Imbens, Guido W. and Joshua D. Angrist, “IdentiÞcation of Local Average Treat-
     ment Effects,” Econometrica, Volume 62, Issue 2, March 1994
[45] Kane, Thomas J, Cecilia Elena Rouse and Douglas Staiger, “Estiamting Returns
     to Schooling when Schooling is Misreported,” NBER Working Paper 7235, July
     1999
[46] Katz, Michael, “A History of Compulsory Education Laws,” Phi Delta Kappa
     Educational Foundation 1976
[47] Kenkel, Donald, “Health Behavior, Health knowledge and Schooling,” Journal of
     Political Economy, Volume 99, Issue 2, April 1991
[48] Kitagawa and Hauser, Differential Mortality in the United States: a Study in
     Socioeconomic Epidemiology. Cambridge, MA: Harvard University Press, 1973.
[49] Landes William and Lewis C. Solomon, “Compulsory Schooling Legislation: An
     economic Analysis of Law and Social Change in the Nineteenth Century,” Journal
     of Economic History, March 1972
[50] Lang, Kevin and David Kropp, “Human Capital versus Sorting: The Effects of
     Compulsory Attendance Laws,” Quarterly Journal of Economics, August 1986
[51] Leigh, J. Paul and Rachna Dhir, “Schooling and Frailty Among Seniors,” Eco-
     nomics of Education Review, Volume 16, No. 1, 1997
[52] Lleras-Muney, Adriana, “Were State Laws on Compulsory Education Effective?
     An analysis from 1915 to 1939,” mimeo, Columbia University, 2001
[53] Maddala, G.S. Limited Dependent and Qualitative Variables in Econometrics,
     Econometric Society Monographs No. 3, Cambridge University Press, 1997
[54] Margo, Robert A. and T. Aldrich Finegan, “Compulsory Schooling Legislation
     and School Attendance in Turn-of-The Century America: A ‘Natural Experi-
     ment’,” Economics Letters, October 1996


                                        24
[55] Meghir, Costas and Marten Palme, “Assessing the Effect of Schooling on earnings
     Using a Social Experiment,” Unpublished Working Paper, University College
     London, 1999
[56] Menchik, Paul L., “Economic Status as a Determinant of Mortality among Black
     and White Older Men: Does Poverty Kill?,” Population Studies, November 1993
[57] Munasinghe, Lalith and Nachum Sicherman, “Why do Dancers Smoke? Time
     Preference, Occupational Choice, and Wage Growth,” NBER Working Paper
     7542, February 2000
[58] Nelson, Richard R. and Edmund S. Phelps, “Investment in Humans, Technolog-
     ical Diffusion, and Economic Growth,” American Economic Review, Volume 56,
     Issue1/2, March 1966
[59] Newhouse, Joseph P., Free for All? Lessons from the Rand Health Insurance
     Experiment. Cambridge: Harvard University press. 1993
[60] Nordhaus, William D., “The Health of Nations: The Contribution of Improved
     Health to Living Standards,” mimeo, Yale University, November 1999
[61] Pappas, Gregory, Susan Queen, Wolber Hadden and Grail Fisher, “The Increas-
     ing Disparity in Mortality Between Socioeconomic Groups in the United States,
     1960 and 1986,” The New England Journal of Medicine, 1993,
[62] Perri, Timothy J., “Health Status and Schooling Decisions of Young Men,” Eco-
     nomics of Education Review, 1984
[63] Rogers, Richard G., Robert A. Hummer and Charles B. Nam, Living and Dying
     in the USA, Academic Press 2000
[64] Roseboom, T.J. et al, “Coronary heart Disease after Prenatal Exposure to the
     Dutch Famine, 1944-45,” Heart, Volume 84, Issue 6, 2000
[65] Rosenzweig M.R and T.P. Schultz, “Education and Household Production of
     Child Health,” In Proceedings of the American Statistical Association (Social
     Statistics Section) Washington, DC: American Statistical Association, 1991
[66] Steven Ruggles and Matthew Sobek et. al., Integrated Public Use Microdata Se-
     ries: Version 2.0 Minneapolis: Historical Census Projects, University of Min-
     nesota, 1997
[67] Sander, William, “Schooling and Quitting Smoking,” Review of Economics and
     Statistics, 77, 1995
[68] Schmidt, Stefanie, “School Quality, Compulsory Education Laws, and the
     Growth of American High School Attendance, 1915-1935,” MIT Ph.D. Disserta-
     tion 1996
[69] Staiger, Douglas and James H. Stock, “Instrumental Variables Regression with
     Weak Instruments,” Econometrica;65(3), May 1997

                                        25
[70] Starr, Paul, “The Social Transformation of American Medicine,” New York: Ba-
     sic Books, 1982
[71] Stigler, George, “Employment and Compensation in Education,” NBER Occa-
     sional Paper. Number 53. 1950
[72] Strauss, John and Duncan Thomas, “Human Resources: Empirical Modelling of
     Household and Family Decisions,” in Handbook of Development Economics, Vol.
     III, Edited by J. Behrman, T.N. Srinivasan. Elsevier Science 1995
[73] Tyack, David One Best system: A History of American Urban Education. Har-
     vard University Press. Cambridge, Massachusetts, 1974
[74] Wolfe, Barbara and Jere R. Behrman, “Women’s Schooling and Children’s
     Health: Are the Effects Robust with Adult Sibling Control for the Women’s
     Childhood Background?,” Journal of Health Economics 6 , 1987
[75] Woltz, Charles, K.,“Compulsory Attendance at School,” in Law and Contempo-
     rary Problems, School of Law, Duke University Vol. 20 Winter 1955




                                       26
Appendix: Trends for Compulsory education and Child Labor laws

Compulsory Attendance Laws

Age at which must enter school (enter age)

         States 1915    States 1928    States 1939
 6             0              2              2
 7            16             28             33
 8            25             17             13
 9             1              1              0
Total         42             48             48


Child Labor Laws

Minimun age to get work permit (work age)

        States 1915    States 1928    States 1939
 12            2
 13            1
 14           38            42               32
 15            4            4                 4
 16            0             2               12
Total         45            48               48


Continuation School Laws

Have Continuation School Laws

      States 1915       States 1928    States 1939
 0         36                20             19
 1         12                28             29
Total      48                48             48


Constructed Variable: Implicit number of years had to attend school

Childcom = work age - enter age

              States 1915    States 1928     States 1939
     0              8
     4              1
     5              2             1
     6             21            15                9
     7             14            26               23
     8              2             5                7
     9                                             8
    10                           1                 1
   Total          48             48               48




                                                       27
Time trends for the laws


                 Age at which child must enter school.
                  National average among states with
          8                     Laws
        7.5

          7
               15 17 19 21 23 25 27 29 31 33 35 37 39
                                Year



                 Average age required for work permit
                       among states with laws
         15
        14.5
         14
        13.5
         13
                15 17 19 21 23 25 27 29 31 33 35 37 39
                                 Year




                  Proportion of states that required
              attendance at part time or evening school
          0.8
          0.6
          0.4
          0.2
            0
                15 17 19 21 23 25 27 29 31 33 35 37 39

                                       Year


                                  28
                   TABLE 1: SUMMARY STATISTICS- AGGREGATED CENSUS DATA

Variables                                                    Mean        Std. Dev.       Min           Max

Individual      10-year death rate                           0.106         0.136           -7          0.875
characteristics Years of completed education                10.697         1.020         4.818          18
                1970 Dummy                                  0.471          0.499           0             1
                Female                                       0.517         0.500           0             1
                Married                                      0.818         0.096           0             1
                Live in North                                0.255         0.369           0             1
                Live in West                                 0.285         0.351           0             1
                Live in South                                0.159         0.227           0             1
                Live in an urban area                        0.685         0.122           0             1
                Age                                         50.366         8.482          35            69
                Born in 1901                                0.029          0.167           0             1
                Born in 1902                                0.025          0.157           0             1
                Born in 1903                                0.028          0.166           0             1
                Born in 1904                                0.029          0.169           0             1
                Born in 1905                                0.031          0.174           0             1
                Born in 1906                                0.032          0.177           0             1
                Born in 1907                                0.033          0.180           0             1
                Born in 1908                                0.036          0.186           0             1
                Born in 1909                                0.036          0.187           0             1
                Born in 1910                                0.038          0.191           0             1
                Born in 1911                                0.039          0.193           0             1
                Born in 1912                                0.040          0.195           0             1
                Born in 1913                                0.042          0.200           0             1
                Born in 1914                                0.043          0.202           0             1
                Born in 1915                                0.044          0.205           0             1
                Born in 1916                                0.044          0.205           0             1
                Born in 1917                                0.044          0.206           0             1
                Born in 1918                                0.046          0.209           0             1
                Born in 1919                                0.047          0.213           0             1
                Born in 1920                                0.048          0.213           0             1
                Born in 1921                                0.048          0.214           0             1
                Born in 1922                                0.050          0.217           0             1
                Born in 1923                                0.049          0.216           0             1
                Born in 1924                                0.049          0.215           0             1
                Born in 1925                                0.050          0.217           0             1

State-of-Birth % Urban                                       53.523        21.279         12.300        97.500
Characteristics % Foreign                                    11.737         8.523          0.400        31.300
                % Black                                      8.983         11.901         0.010         54.200
                % Employed in manufacturing                   0.067         0.039          0.003         0.283
                Annual Manufacturing wage                  7161.911       1368.253       713.030      12095.160
                Value of farm per acre                      540.048        276.353        47.700       1802.575
                Per capita number of doctors                  0.001         0.000          0.000         0.003
                Per capita education expenditures            96.474        42.142          5.372       601.391
                Number of school buildings per sq. mile      0.174          0.090         0.002         0.474
      Number of observations: 4795, corresponding to cells defined at the gender, state-of-birth, and cohort.
      All means calculated using weights, where the weights are given by the number of observations in each
      cell. Monetary values are in 1982-84 dollars




                                                      29
                             TABLE 2: SUMMARY STATISTICS- NHEFS


Variables                                                 Mean       Std. Dev.    Min        Max

Individual      Died between 1975 and 1985               0.254        0.435        0          1
characteristics Years of completed education             10.360       3.326         0         17
                Female                                   0.540        0.498        0           1
                Married                                  0.755        0.430        0          1
                Live in North                            0.214        0.410        0          1
                Live in West                             0.250        0.433        0          1
                Live in South                            0.269        0.444        0          1
                Live in an urban area                    0.526        0.499        0          1
                Age                                      62.941       7.561        50         74
                Born in 1901                             0.039        0.193        0          1
                Born in 1902                             0.054        0.226        0          1
                Born in 1903                             0.056        0.230        0          1
                Born in 1904                             0.056        0.230        0          1
                Born in 1905                             0.061        0.239        0          1
                Born in 1906                             0.068        0.251        0          1
                Born in 1907                             0.055        0.227        0          1
                Born in 1908                             0.042        0.200        0          1
                Born in 1909                             0.025        0.155        0          1
                Born in 1910                             0.026        0.160        0          1
                Born in 1911                             0.027        0.161        0          1
                Born in 1912                             0.028        0.165        0          1
                Born in 1913                             0.028        0.165        0          1
                Born in 1914                             0.031        0.174        0          1
                Born in 1915                             0.033        0.178        0          1
                Born in 1916                             0.032        0.177        0          1
                Born in 1917                             0.034        0.182        0          1
                Born in 1918                             0.035        0.184        0          1
                Born in 1919                             0.041        0.198        0          1
                Born in 1920                             0.037        0.188        0          1
                Born in 1921                             0.038        0.192        0          1
                Born in 1922                             0.039        0.194        0          1
                Born in 1923                             0.034        0.182        0          1
                Born in 1924                             0.044        0.206        0          1
                Born in 1925                             0.036        0.187        0          1

State-of-Birth % Urban                                    49.846        20.734     12.3       97.5
Characteristics % Foreign                                  11.489        8.434      0.4       31.3
                % Black                                   10.108        13.652     0.01       53.8
                % Employed in manufacturing                0.065         0.040     0.003     0.283
                Annual Manufacturing wage                6971.696     1380.099   713.030   11007.230
                Value of farm per acre                    549.371      292.371    48.484    1802.575
                Per capita number of doctors              0.0013        0.0003    0.0002     0.0026
                Per capita education expenditures          86.305       44.411     5.372    601.391
                Number of school buildings per sq. mile    0.173         0.092     0.003     0.474
   Number of observations: 4554. Monetary values are in 1982-84 dollars




                                                   30
                              Figure 1: Number of observations per cohort

   25000


   20000


   15000                                                                                               1960 census
                                                                                                       1970 census
   10000                                                                                               1980 census

    5000


        0
            1901 1903 1905 1907 1909 1911 1913 1915 1917 1919 1921 1923 1925
                                               Birth year




                             Figure 2: Average years of education by cohort

    12.00




    11.00
                                                                                                        1960 census
                                                                                                        1970 census
                                                                                                        1980 census
    10.00




     9.00
            1901 1903 1905 1907 1909 1911 1913 1915 1917 1919 1921 1923 1925
                                                Birth year

Note: Figures 1 and 2 follow the same cohorts from the 1960 census up to the 1980 census.
In Figure 1 we can observe that that 10-year mortality increases with age: for older cohorts the number
of individuals observed in 1980 is much smaller than in 1960 or 1970.
In figure 2 we can see that the average level of education is higher in 1980 than in 1960 for all cohorts,
suggesting that those who died in each cohort had bellow average levels of education.



                                                    31
Figure 3: Calculating Death Rates with the Census




                                               The 1960 and 1970 census are 1/100 random
                                               samples of the population, therefore the number
                                               of individuals in any given group is always
                                               observed with error. Because of this sampling
                                               error the death rates for any given group are over-
                                               estimated 50% of the time and underestimated
                                               50% of the time. However, since the sampling is
                                               truly random, the observed death rates are
                                               consistent estimates of the true death rates.




Figure 3B: An example for a young cohort: 0 death rate



                                               If the true death rate is 0 then I observe 50%
                                               negative death rates. As cohorts age, the death
                                               rate increases (see example above) the number
                                               of negative death rates falls.




                                      32
                                                       Figure 4A: Percentage of negative Death Rates per Cohort

0.7

0.6

0.5

0.4                                                                                                                        1960-1970 Death Rate
0.3                                                                                                                        1970-1980 Death Rate

0.2

0.1

                 0
                                          1901       1904   1907   1910     1913      1916      1919   1922   1925
                                                                      Birth Year




                                                 Figure 4B: Percentage of negative death rates by average state size

                                          900
average number of observations in state




                                          800
                                          700
                                          600
                                          500
                                          400
                                          300
                                          200
                                          100
                                            0
                                                 0           0.1          0.2            0.3           0.4           0.5       0.6         0.7
                                                                                percentage of negative death rates




                                                                                               33
                    Figure 4C: Observed 10-year death rates by age

0.6


0.5


0.4


0.3                                                                  1960 Census
                                                                     1970 Census
0.2                                                                  1975 NHEFS


0.1


  0
       35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65 67 69 71 73
-0.1

                                   Age




                                      34
            Figure 5: Average education level by years of compulsory education

12


11


10


 9


 8


 7
        0           4           5           6            7          8       9    10
                                    Years of compulsory education




             Figure 6: Average education level by compulsory education for
                                   selected cohorts

13.00

12.00

11.00
                                                                                  1915
                                                                                  1920
10.00
                                                                                  1925
                                                                                  1930
 9.00
                                                                                  1935
 8.00                                                                             1939


 7.00

 6.00
            0       4       5           6            7       8          9   1
                             Years of compulsory education




                                                35
                  TABLE 3: EFFECT OF COMPULSORY EDUCATION LAWS ON EDUCATION

Variables                                                          Individual data   Individual data    Aggregate data
Dependent Variable    Education
Education Laws        Childcom Category 4(a)                           0.347**            0.355**          0.323**
                                                                      (0.075)            (0.088)           (0.077)
                      Childcom Category 5                              0.323**            0.260**          0.321**
                                                                      (0.090)            (0.099)           (0.098)
                      Childcom Category 6                              0.274**              0.302**          0.266**
                                                                      (0.069)              (0.087)             (.074)
                        Childcom Category 7                            0.385**              0.408**          0.369**
                                                                      (0.070)              (0.087)            (0.074)
                        Childcom Category 8                            0.416**              0.398**          0.315**
                                                                      (0.075)              (0.089)            (0.078)
                        Childcom Category 9                            0.580**              0.512**          0.470**
                                                                      (0.076)              (0.092)            (0.084)
                        Childcom Category 10                           0.325**              0.328**          0.318**
                                                                      (0.080)              (0.095)            (0.095)
                        Continuation School Required (=1)              0.027                 0.017              0.027
                                                                      (0.026)              (0.028)            (0.031)
Individual              Female                                         0.116**              0.140**            -0.012
characteristics                                                       (0.014)              (0.014)            (0.014)
                        Married                                                             0.433**          -1.111**
                                                                                           (0.011)            (0.097)
                        Live in an urban area                                               0.968**          0.278**
                                                                                           (0.019)            (0.138)
State-of-Birth          % Urban                                                             0.021**          0.028**
Characteristics                                                                            (0.004)            (0.005)
                        % Foreign                                                           -0.002              0.004
                                                                                           (0.008)            (0.010)
                        % Black                                                             0.024**          0.020**
                                                                                           (0.009)            (0.010)
                        % Employed in manufacturing                                         -0.350           -1.220**
                                                                                           (0.513)            (0.621)
                        Annual Manufacturing wage                                            0.000              0.000
                                                                                           (0.000)            (0.000)
                        Value of farm per acre                                               0.000              0.000
                                                                                           (0.000)            (0.000)
                        Per capita number of doctors                                      150.018**         188.615**
                                                                                          (71.840)             (66.2)
                        Per capita education expenditures                                  0.001**              0.000
                                                                                           (0.000)            (0.000)
                        Number of school buildings per sq. mile                             -0.359             -0.166
                                                                                           (0.289)            (0.373)
                        3 region of residence dummies                     No                  Yes                Yes
                        Region of birth*cohort dummies                    No                  Yes                Yes
                        R-Squared                                       0.0811             0.1052               0.888
                        F-statistic on instruments                       14.93**             8.37**           4.49**
                        Partial R-squared                               0.0003             0.0001             0.0108
         * significant at 10% ** significant at 5%. All regressions include a dummy for the 1970 census, state-
        of-birth dummies, cohort dummies and an intercept. For the individual-level regressions (1 and 2)
        N=814805 and the standard errors (in parenthesis) are clustered at the state-of-birth and cohort level.
        (a) Childcom=work permit age - entry age. Childcom=0 is the excluded category.
        (b) Data aggregated by gender/cohort/state-of-birth. Robust standard errors. N=4792

                                                           36
               TABLE 4: EFFECT OF EDUCATION ON MORTALITY-LEAST SQUARE RESULTS

Variables
Data                                                     Census           NHEFS     NHEFS          NHEFS            NHEFS
                                                                                             (a)
Method                                                    WLS              OLS      Probit           WLS             W LS
       (d)                                                         (b)                                      (b)
Level                                                  Aggregate         Individual Individual Aggregate          Aggregate(c)
Dependent                                                10-year            died      died         death rate      death rate
Variable                                                death rate         75-85     75-85           75-85           75-85

Individual      Education                               -0.012**         -0.012**   -0.011**       -0.017**        -0.013**
characteristics                                          (0.004)          (0.002)    (0.002)        (0.004)         (0.005)
                Female                                  -0.048**         -0.147**   -0.151**       -0.137**        -0.139**
                                                         (0.004)          (0.013)    (0.013)        (0.015)         (0.030)
                  Married                                0.227**         -0.044**     -0.053         -0.005          -0.015
                                                         (0.030)          (0.016)    (0.015)        (0.030)         (0.037)
                  Live in an urban area                 -0.136**         0.037**     0.039**       0.056**         0.080**
                                                         (0.044)          (0.015)    (0.015)        (0.024)         (0.030)

State-of-Birth % Urban                                     0.000           -0.003     -0.003         -0.002         -0.002
Characteristics                                           (0.001)         (0.005)    (0.005)        (0.005)         (0.005)
                % Foreign                                  0.000            0.005     0.012           0.005           0.005
                                                          (0.002)         (0.007)    (0.008)        (0.007)         (0.007)
                  % Black                                  0.000          -0.014*     -0.012         -0.014          -0.014
                                                          (0.002)         (0.008)    (0.008)        (0.009)         (0.009)
                  % Employed in manufacturing              -0.075          -0.085     -0.060         -0.091          -0.100
                                                          (0.105)         (0.590)    (0.563)        (0.621)         (0.640)
                  Annual Manufacturing wage                0.000            0.000     0.000           0.000           0.000
                                                          (0.000)         (0.000)    (0.000)        (0.000)         (0.000)
                  Value of farm per acre                   0.000            0.000     0.000           0.000           0.000
                                                          (0.000)         (0.000)    (0.000)        (0.000)         (0.000)
                  Per capita number of doctors             -2.043           1.058    17.451         -0.762          -2.139
                                                         (14.384)        (48.228)   (39.746)       (49.833)        (52.857)
                  Per capita education expenditures         0.000           0.000      0.000          0.000           0.000
                                                          (0.000)         (0.000)    (0.000)        (0.000)         (0.000)
                  # of school buildings per sq. mile       0.045         0.712**    0.744**        0.725**         0.758**
                                                          (0.064)         (0.345)    (0.334)        (0.360)         (0.380)

                  N                                        4792          4554      4554          1557           942
                  R-Squared                               0.3685       0.1736                  0.3952          0.5219
        All regressions include 24 cohort dummies, 47 state of birth dummies, region-of-birth * cohort, region
        of residence dummies and an intercept. Standard errors (in parenthesis) are clustered at the state-of-birth
        and cohort level. The census regressions also include a dummy for the 1970 census.
        (a) The reported coefficients are the mean marginal effects. The standard errors are calculated using
             the Delta Method.
        (b) Data are aggregated at the cohort/gender and state-of-birth level.
        (c) Data aggregated at the cohort and state-of-birth level only.
        (d) All regressions at the aggregate level are weighted by the number of observations in the original
             cell

        * significant at 10% ** significant at 5%.




                                                           37
    TABLE 5: EFFECT OF EDUCATION ON MORTALITY-LEAST SQUARES RESULTS BY
                                  GENDER

     Variables                                                      Males            Females
     Dependent       10-year death rate
     Variable

     Individual      Education                                       0.008             0.011
     characteristics                                               (0.007)           (0.008)
                     Married                                      -0.293**          -0.324**
                                                                   (0.081)           (0.057)
                     Dummy for 1970                               0.017**           -0.063**
                                                                   (0.005)           (0.008)
                     Live in an urban area                          -0.080            -0.086
                                                                   (0.058)           (0.065)

     State-of-Birth % Urban                                         -0.001              0.000
     Characteristics                                               (0.002)            (0.002)
                     % Foreign                                      -0.002              0.001
                                                                   (0.004)            (0.004)
                     % Black                                        -0.001             -0.002
                                                                   (0.004)            (0.004)
                     % Employed in manufacturing                    -0.023             -0.072
                                                                   (0.313)            (0.261)
                     Annual Manufacturing wage                       0.000              0.000
                                                                   (0.000)            (0.000)
                     Value of farm per acre                          0.000              0.000
                                                                   (0.000)            (0.000)
                     Per capita number of doctors                  11.124             -18.978
                                                                  (24.294)           (28.178)
                     Per capita education expenditures               0.000              0.000
                                                                   (0.000)            (0.000)
                     Number of school buildings per sq. mile        -0.063              0.096
                                                                   (0.144)            (0.151)

                      N                                                  2397           2395
                      R-Squared                                         0.4668         0.2297
All regressions include 24 cohort dummies, 47 state-of-birth dummies, region-of-birth * cohort
interactions, region-of-residence dummies and an intercept. All regressions are weighted by the number
of observations in the original cell. Standard errors (in parenthesis) are robust.

* significant at 10% ** significant at 5%.




                                                    38
                    TABLE 6: EFFECT OF EDUCATION ON MORTALITY-IV RESULTS
Variables
     Data                                                    NHEFS(b)     Census(a)(c)   Census(a)(b)(c)   Census(a)(b)(c)
   Method                                                      2SLS          Wald            2SLS          Mixed 2SLS
    Level                                                    Individual   Aggregate       Aggregate         Aggregate
Dependent                                                      Died        10-year         10-year           10-year
Variable                                                     1975-1985    death rate      death rate        death rate


Individual      Education                                     -0.020       -0.037**         -0.045*          -0.059**
characteristics                                               (0.054)       (0.006)         (0.026)           (0.027)
                1970 Dummy                                                   0.003          0.012**           0.021**
                                                                            (0.004)         (0.005)           (0.007)
                Female                                       -0.142**      -0.071**        -0.048**          -0.040**
                                                              (0.030)       (0.004)         (0.004)           (0.005)
                Married                                        -0.040                       0.190**           0.266**
                                                              (0.027)                       (0.041)           (0.031)
                Live in an urban area                          0.046                       -0.126**            -0.080
                                                              (0.055)                       (0.045)           (0.054)

State-of-Birth % Urban                                         -0.002                        0.001             0.001
Characteristics                                               (0.005)                       (0.001)           (0.001)
                % Foreign                                      0.005                         0.001             0.001
                                                              (0.007)                       (0.002)           (0.002)
                % Black                                        -0.014                        0.001             0.001
                                                              (0.008)                       (0.002)           (0.002)
                % Employed in manufacturing                    -0.089                        -0.118            -0.080
                                                              (0.605)                       (0.113)           (0.137)
                Annual Manufacturing wage                      0.000                         0.000             0.000
                                                              (0.000)                       (0.000)           (0.000)
                Value of farm per acre                         0.000                         0.000             0.000
                                                              (0.000)                       (0.000)           (0.000)
                Per capita number of doctors                   7.298                         6.078             6.675
                                                             (62.347)                      (15.337)           (17.31)
                Per capita education expenditures               0.000                        0.000             0.000
                                                              (0.000)                       (0.000)           (0.000)
                Number of school buildings per sq. mile      0.698**                         0.051             0.044
                                                              (0.350)                       (0.066)           (0.075)

                State-of-birth Dummies                           Yes            No            Yes            Yes
                Region of Birth Dummies                          No            Yes             No            No
                Cohort Dummies                                   Yes           Yes            Yes            Yes
                Region-of-birth*cohort                           Yes            No            Yes            Yes
                Region of residence dummies                      Yes            No            Yes            Yes
                N                                               4554           1396          4792           4792
     All regressions include an intercept.
     (a) Regressions are weighted by the number of observations in the original cell.
     (b) Standard errors (in parenthesis) are clustered at the state-of-birth and cohort level and have been
          corrected in the second stage.
     (c) Note: P2SLS and aggregate 2SLS use data aggregated at the gender/cohort/state-of-birth. Wald
          uses data aggregated at the gender/cohort/region-of-birth/compulsory education laws level.

     * significant at 10% ** significant at 5%.


                                                        39
                              TABLE 7: ADDITIONAL ESTIMATIONS

                     A: Quarter of birth, laws and interactions used as instruments

                                                           2SLS            Mixed 2SLS
                Variables
                Dependent        10-year death rate
                Variable

                Individual       Education                 -.067**           -.062**
                                                           (.0260)            (.024)


                                     B: Reduced form Results (OLS)

                         Variables
                         Dependent       10-year death rate
                         Variable

                                         Childcom                    -0.0027**
                                                                      (0.0013)
                                         Continuation school          -0.0032
                                                                       (0.005)


                             C: Age Heaping: Exclude ages 40, 50, and 60

                                                           2SLS            Mixed 2SLS
                Variables
                Dependent        10-year death rate
                Variable

                Individual       Education                  -.040                 -.052
                                                           (.026)                (.026)

All regressions include the same controls as in Table 6.

* significant at 10% ** significant at 5%.




                                                      40
             TABLE 8: EFFECT OF EDUCATION ON MORTALITY-IV RESULTS BY GENDER

                                                                2SLS        2SLS      Mixed 2SLS Mixed 2SLS
Variables                                                       Males      Females       Males       Females
Dependent      10-year death rate
Variable

Individual      Education                                        -0.047      -0.044      -0.077        -0.054
characteristics                                                 (0.051)     (0.063)     (0.040)       (0.039)
                Married                                        -0.292**    -0.311**    -0.263**      -0.297**
                                                                (0.082)     (0.059)     (0.087)       (0.065)
               Dummy for 1970                                    0.025     -0.057**    0.032**       -0.052**
                                                                (0.009)     (0.011)     (0.009)       (0.011)
               Live in an urban area                             -0.096      -0.102      -0.002       -0.036
                                                                (0.059)     (0.067)     (0.079)       (0.078)

State-of-Birth % Urban                                           0.002       0.001        0.001         0.001
Characteristics                                                 (0.003)     (0.002)     (0.002)       (0.002)
                % Foreign                                        -0.001      0.001       -0.001         0.001
                                                                (0.004)     (0.004)     (0.003)       (0.003)
               % Black                                           0.000       -0.001       0.000        -0.001
                                                                (0.004)     (0.004)     (0.003)       (0.003)
               % Employed in manufacturing                       -0.095      -0.128      -0.056        -0.103
                                                                (0.328)     (0.264)     (0.182)       (0.194)
               Annual Manufacturing wage                         0.000       0.000        0.000         0.000
                                                                (0.000)     (0.000)     (0.000)       (0.000)
               Value of farm per acre                             0.000      0.000        0.000         0.000
                                                                (0.000)     (0.000)     (0.000)       (0.000)
               Per capita number of doctors                     22.270       -6.219     27.940         -5.860
                                                               (24.002)    (32.441)    (22.782)       (26.29)
               Per capita education expenditures                  0.000       0.000       0.000         0.000
                                                                (0.000)     (0.000)     (0.000)       (0.000)
               Number of school buildings per sq. mile           -0.018      0.078       -0.057         0.094
                                                                (0.148)     (0.155)     (0.103)       (0.106)

                N                                                     2397         2395    2397         2395
     All regressions include 24 cohort dummies, 47 state-of-birth dummies, region-of-birth * cohort
     interactions, region-of-residence dummies and an intercept. All regressions are weighted by the number
     of observations in the original cell. Standard errors (in parenthesis) are robust.

     * significant at 10% ** significant at 5%.




                                                         41
  TABLE 9: TESTING FOR THE FUNCTIONAL FORM OF THE EFFECT OF EDUCATION ON
                                 MORTALITY

           Variables
                   Method                                                               OLS
           Dependent Variable      10-year death rate



           Individual              Education                                          -0.064**
           characteristics                                                             (0.029)
                                   Education Squared                                   0.003*
                                                                                       (0.001)
                                   1970 Dummy                                           0.010
                                                                                       (0.005)
                                   Female                                               -0.047
                                                                                       (0.004)
                                   Married                                              0.224
                                                                                       (0.030)
                                   Live in an urban area                                -0.138
                                                                                       (0.044)

           State-of-Birth          % Urban                                              0.000
           Characteristics                                                             (0.001)
                                   % Foreign                                            0.000
                                                                                       (0.002)
                                   % Black                                              0.000
                                                                                       (0.002)
                                   % Employed in manufacturing                          -0.058
                                                                                       (0.107)
                                   Annual Manufacturing wage                            0.000
                                                                                       (0.000)
                                   Value of farm per acre                               0.000
                                                                                       (0.000)
                                   Per capita number of doctors                         -4.526
                                                                                      (14.546)
                                   Per capita education expenditures                    0.000
                                                                                       (0.000)
                                   Number of school buildings per sq. mile              0.041
                                                                                       (0.064)

All regressions include 24 cohort dummies, 47 state-of-birth dummies, region-of-birth * cohort
interactions, region-of-residence dummies and an intercept. All regressions are weighted by the number
of observations in the original cell. Standard errors (in parenthesis) are clustered at the state-of-birth and
cohort level. Estimated using census data aggregated at the gender/state-of-birth/cohort level. N=4792




                                                     42
 TABLE 10: EFFECT OF THE DISTRIBUTION OF EDUCATION ON MORTALITY: ARE THERE
                                EXTERNALITIES?


     Variables
        Method                                                                           OLS
     Dependent     10-year death rate of those with more than 12 years of
     Variable      schooling


                   Average Education (12 years of schooling or more)                     .010
                                                                                       (0.011)
                   Standard deviation of education of the entire distribution of       -.092**
                   education (education from 0 to 17)
                                                                                        (.025)

All regressions include the covariates in previous table plus 24 cohort dummies, 47 state-of-birth
dummies, region-of-birth * cohort interactions, region-of-residence dummies and an intercept. All
regressions are weighted by the number of observations in the original cell. Standard errors (in
parenthesis) are clustered at the state-of-birth and cohort level. Data aggregated at the gender/state-of-
birth/cohort level. N=4792




                                                   43