Journal of Health Economics 30 (2011) 753–763

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

The causal relationship between education, health and health related behaviour:
Evidence from a natural experiment in England
Nils Braakmann ∗
Newcastle University, Business School – Economics, Ridley Building, Claremont Road, Newcastle upon Tyne, Tyne and Wear, NE1 7RU, United Kingdom

a r t i c l e i n f o a b s t r a c t

Article history: I exploit exogenous variation in the likelihood to obtain any sort of educational qualification between
Received 22 November 2010 January- and February-born individuals for 13 academic cohorts in England. For these cohorts compulsory
Received in revised form 19 May 2011 schooling laws interacted with the timing of the CSE and O-level exams to change the probability of
Accepted 25 May 2011
obtaining a qualification by around 2–3 percentage points. I then use data on individuals born in these
Available online 6 June 2011
two months from the British Labour Force Survey and the Health Survey for England to investigate the
effects of education on health using being February-born as an instrument for education. The results
JEL classification:
indicate neither an effect of education on various health related measures nor an effect on health related
I12
I20
behaviour, e.g., smoking, drinking or eating various types of food.
© 2011 Elsevier B.V. All rights reserved.
Keywords:
Education
Health
Socio-economic gradient
Education gradient
Compulsory schooling

The social determinants of health have been a major focus of
interest in recent years (see, e.g., Adams et al., 2003; Commission on
Social Determinants of Health, 2008). A robust correlation has been
found between individual education and individual health (see, e.g.,
Grossmann, 2006, for a survey). Recent research (e.g., Cutler and
Lleras-Muney, 2010) has started to investigate the channels driving
this observed correlation.
In general, there are two broad explanations why education
and health might be correlated: the first is that the observed
positive correlation is spurious and in fact caused by underlying
third variables like parental or family background, parental invest-
ments into their children or differences in non-cognitive traits or
time preferences. A related argument would be a possible reverse
causality stating that people who expect to have better health
are willing to invest more into education as they expect to live
longer giving them more time to reap the returns to that invest-
ment. The second strand of arguments gives reasons for a possible
causal link between education and health. A first potential expla-
nation is a higher productivity of higher educated individuals that
directly transfers into a higher level of health production given the
∗ Tel.: +44 191 222 8145.
E-mail address: nils.braakmann@ncl.ac.uk
same inputs (Grossman, 1972; Michael, 1973). This argument can
be seen as an analogy to the well-known relationship between
education and wages. Some evidence on this relationship is pro-
vided by Spandorfer et al. (1995) who show that low literacy goes
hand in hand with a poor comprehension of hospitals’ discharge
instructions and by Goldman and Smith (2002) who find a relation-
ship between education and compliance with medical treatments.
A second argument brought forth by Grossmann (1972) is that
higher educated people might be better at allocating inputs such as
time over health-relevant activities, e.g., through better informa-
tion about medical treatments (see Glied and Lleras-Muney, 2003).
Finally, higher educated individuals earn more than lower educated
individuals, which may allow them to buy more expensive medical
treatments, healthier foods or live in healthier regions.
Most recent papers investigating the second strand of argu-
ments and trying to establish whether there is indeed a causal
link between education and health have used changes in compul-
sory schooling laws. While comparisons of these studies are not
easy as most papers use different outcome variables, results can
generally considered to be very mixed: Spasojevic (2003) finds pos-
itive effects on a health index and BMI for 50-year olds in Sweden,
Lleras-Muney (2005) finds large decreases in the 10-year mortal-
ity of the same age group in the USA. Similar results are found for
70-year olds in the USA by Glied and Llreas-Muney (2003). Finally,

0167-6296/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhealeco.2011.05.015
754 N. Braakmann / Journal of Health Economics 30 (2011) 753–763
Oreopoulos (2006) finds positive effects on self-rated health and
the occurrence of activity limiting disability for 25–84-year olds in
the UK and the USA. However, an almost equal number of studies
fail to find any effect using the same identification strategy: Arendt
(2005) finds no effect on self-rated health, smoking and body mass
index in Denmark, while Albouy and Lequien (2009) reach the
same conclusion regarding mortality at 50 and 80 in France. Using
two changes in compulsory schooling laws in Britain in 1947 and
1972 that led to large changes in the educational attainment of the
affected cohorts, Clark and Royer (2010) find no evidence for any
effect on either health outcomes or health related behaviour. On a
somewhat related topic, Doyle et al. (2005) find no effect of parental
education on the health of 8-year-old children in the UK. Finally,
Adams (2002) finds mixed evidence for the USA depending on the
outcome used.1
In this paper I exploit a natural experiment in England – first
used by Anderberg and Zhu (2010) in the context of matching in
the marriage market – leading to differences in the likelihood of
having obtained any educational qualification between individuals
born in January and February in the same birth cohort (see Sec-
tion 1 for a brief description and Anderberg and Zhu, 2010, for
additional institutional details). Specifically, for the birth cohorts
1957 to (roughly) 1970 regulations regarding the date individu-
als reaching the minimum school leaving age could actually leave
school interacted with the timing of the exams for the first quali-
fication that could be obtained in England (the “O-levels” and the
“CSE”). In these cohorts group, individuals born in February have
a 2–3% higher probability of having obtained a qualification. I will
also provide some evidence that this effect only exists for the lowest
possible qualification, i.e., CSE vs. no qualification, while no differ-
ences exist for the probability of having passed O-levels or A-levels
or having a university degree.
In what follows I will focus mainly on individuals being born
in either January or February. A dummy for being born in Febru-
ary can then be used as an instrument for having obtained any
type of qualification, i.e., CSE or above, in regressions using vari-
ous health outcomes and health related behaviours as outcomes.
It is important to stress that this instrument does not suffer from
the same problems as the famous quarter of birth instrument used
by Angrist and Krueger (1991). Firstly, as we will see, the instru-
ment is generally much stronger, leading to an almost 3% increase in
the likelihood of having obtained any qualification and passing all
usual weak instrument tests. Secondly, using January vs. February
born as an instrument also avoids some of the potential endo-
geneity problems associated with quarter of birth instruments.
To recall these briefly: there is recent evidence that the charac-
teristics of women giving birth differ over the year, which may
lead to unobserved differences in parental background for children
born in different quarters (Buckles and Hungerman, 2008). Further-
more, the evidence presented against the validity of the quarter
of birth instrument in the seminal paper by Bound et al. (1995)
documents small differences in school performance, mental and
physical health as well as family income for individuals born in
different seasons as well as regional differences in seasonal birth
patterns. These problems can be expected to be much smaller when
looking only at individuals born in two adjacent months. Firstly,
while families can – at least to some extent – plan the season they
want to give birth in, this is far less possible with respect to the
1
There are also a variety of studies for various countries using other identifica-
tion strategies or instruments, e.g., Berger and Leigh (1989), Kenkel (1991), Arkes
(2003), DeWalque (2003, 2004, 2007), Auld and Sidhu (2005), Cipollone et al. (2006),
Kenkel et al. (2006), Groot and Maassen van den Brink (2007) and Braakmann (2010).
However, the picture that emerges from these studies is in no way clearer.
exact month of birth. In fact, as I will show in Section 3, there
are no statistically significant differences in parental background
between individuals born in either January or February. Secondly,
while differences in maternal nutrition, weather conditions, sun-
light exposure, etc. may play a role for explaining differences in
mental or physical health for children born in different seasons,
these factors can be expected to be more or less equal for children
born in either January or February. Taken together, these argu-
ments suggest that the instrument is much stronger and much
more likely to be truly exogenous than the well-known quarter
of birth instrument.
Using data from the Labour Force Survey, I show that the higher
likelihood of having obtained at least a CSE did not lead to dif-
ferences in various subjective and objective health outcomes. Note
that even though the individuals in the sample are still quite young,
they are of a similar age than those considered in the studies by
Berger and Leigh (1989), Spasojevic (2003), Lleras-Muney (2005),
Kenkel et al. (2006), or Cipollone et al. (2006), who all found effects
of education on measures of health. I also use data from the Health
Survey for England to take a look at health related behaviour. How-
ever, the results also show no differences in the likelihood to smoke,
drink heavily or eat various more or less healthy foods. Finally, I
present some explanation for this possible lack of effect, specifi-
cally that the income and labour market situation of individuals
nudged into completing their CSE qualification was not better than
that of individuals leaving school without any qualification.
The following section presents the institutional background
in greater detail. Section 2 describes the data and the general
econometric approach. Results for objective health outcomes are
presented in Section 3, results for health-related behaviour fol-
low in Section 4. Section 5 provides a quick view on differences
in income and labour market outcomes. Section 6 concludes.
1. Institutional background
As already outlined in the introduction, the natural experiment
in this paper arises through the interplay of compulsory schooling
laws and the timing of the O-level and CSE exams – the first exams
leading to a (possible) terminating qualification in the UK – for
individuals born between September 1957 and the early 1970s who
were not yet affected by introduction of the GCSE exams in the late
1980s. In the following I briefly sketch the institutional background
as necessary for the understanding of the paper. A more complete
description of the education system in England can be found in
Anderberg and Zhu (2010).
In England, children are admitted into school in the academic
year they turn 5.2 Academic years begin on September 1st and
run until August 31st of the following year. Each academic year
is divided into three terms beginning in September, January and
April respectively. The minimum school leaving age for compul-
sory education was changed twice in recent times. The first change
was due to the 1944 Butler Education act and changed the mini-
mum school leaving age from 14 to 15 in April 1947. The second
change was from 15 to 16 years, introduced in the Raising of School
Leaving Age Order of 1972. It came into effect by September 1973,
affecting children born from September 1957. The earlier increase
2
The exact date of admission for children turning 5 during an academic year varies
between local authorities (see Crawford et al., 2007, p. 13 for an overview of the
different admission policies used in England). However, children born in January and
February would generally be admitted at the same time regardless of the system that
is locally operated, i.e., there is no difference in the duration of schooling between
these two groups that is caused by a different beginning of their respective school
careers.
 N. Braakmann / Journal of Health Economics 30 (2011) 753–763 755

was used in the study by Oreopoulos, 2006, while both of them tive to earlier cohorts is a direct consequence of the increase in the
were used in the paper by Clark and Royer (2010). minimum school leaving age and the associated higher likelihood
In contrast to the US, children reaching the minimum school of pupils taking the exams at 16.
leaving age in the UK may not leave school immediately. From Fig. 1 suggests that the difference between January and Febru-
1962 to 1997 children born between September 1st and January ary born individuals exists only for the proportion of individuals
31st were allowed to leave school at the end of the Spring term, with CSE qualification, i.e., the lowest possible school leaving qual-
i.e., directly before Easter. Children born between February 1st and ification. This observation is consistent with the institutional setup
August 31st, however, had to stay in school until the Friday before described above, specifically that some individuals who would have
the last Monday in May. preferred to leave school earlier were nudged into taking the CSE
As pointed out by Anderberg and Zhu (2010), these regulations examinations, while more academically inclined individuals tak-
create two discontinuities in the compulsory duration of schooling. ing the higher-level qualifications were unaffected. It also suggests
The first occurs between individuals born in August and Septem- that none of the additional CSE holders went on to take higher-level
ber, the second between individuals born in January and February. qualifications.
The August–September-discontinuity occurs between academic
cohorts, making it less useful for comparisons. The second discon-
2. Data and general approach
tinuity, however, occurs within academic cohorts as children born
in January and February will enter school at the same time, but dif-
2.1. Data
fer in the earliest date they are allowed to leave school. Focusing
on the second discontinuity allows to control for possible differ-
I use data from two datasets representative of the English popu-
ences in the content of education between academic cohorts as
lation: The British Labour Force Survey (LFS) and the Health Survey
well as possible effects of the age at school entry, while still allow-
for England (HSE). The former provides a larger sample size and
ing for a full control of birth cohort effects. There is also a known
will be used for the analysis of the relationship between education
difference between August-born and other children, analyzed by
and various objective measures of health, like specific diseases. The
Crawford et al. (2007), which could invalidate the analysis if the
HSE will in turn be used for some complementary investigations
August–September-cut-off was used, while no such differences
on health related behaviour. I also provide some supplementary
exist between January- and February-born children.
descriptive analysis using wave 13 (or m) of the British Household
In general, the discontinuities outlined above would only
Panel Study (BHPS), which was collected in 2003/4.4
change the (compulsory) duration of education by one term (or
The LFS is a survey conducted by the Office of National
roughly two months). However, for birth cohorts up to the early
Statistics since 1973. The data are representative for the pop-
1970s, whose first possible qualifications were the CSE or the –
ulation of households living at private addresses or National
academically more demanding – O-levels, the school leaving date
Health Service institutions. Data collection takes place quar-
interacted with the timing of the exams that were taken at the age
terly since spring 1992. From 1992 to May 2006 data collection
of 16 at the end of the summer term (see Anderberg and Zhu, 2010).
took place in a seasonal pattern with surveys being conducted
Note that the exams were also open for students who left after the
in winter (December–February), spring (March–May), summer
spring term. These students could return to school for the exams
(June–August) and autumn (September–November). The current
without having to attend school for the last term.
sample size is approximately 50,000 responding households in
Like Anderberg and Zhu (2010), I focus on the cohorts still fac-
Great Britain with an additional 2000 being added from North-
ing the old CSE/O-level system, which was abandoned with the
ern Ireland resulting in coverage of 0.1% of the target population.
1988 introduction of the GCSE, as these groups are generally older,
Each household is surveyed in five consecutive quarters in a rotat-
which makes it more likely that I will observe any health effects
ing panel design. As roughly one-fifth of the respondents enter
(the oldest cohorts facing the GCSE would still be in their twen-
and leave each quarter there is an 80% overlap between two adja-
ties at the end of the observation period) and mixing groups facing
cent quarters. Like Anderberg and Zhu (2010) I only use the first
different education systems could create other unknown problems
observation for each individual.
and biases.
The LFS provides information on the labour market status and
For students born before September 1957, who could leave
personal situation of individuals living in the UK during a reference
school at the age of 15, the interaction with the timing of the
period, usually a specific week. The questionnaire therefore encom-
exams is non-existent as individuals leaving school at the earliest
passes information on employment, including information on the
occasion would leave school one year before the CSE and O-level
current employer, socio-demographic characteristics, education,
exams. However, for individuals born between September 1957
and wages as well as information on the respective household.
and the early 1970s, the combination of the variation in school
I use data from the first quarter of 1998 to the last quarter of
leaving dates and the timing of the exams creates large disconti-
2002. Until 1997 health related data was only collected if it affected
nuities in the likelihood of having obtained any qualification. In
an individual’s work in some way. From that date onwards data on
all these cohorts individuals born in February are generally about
all health problems was collected. Month of birth, which is crucial
2–3% more likely to leave school with a qualification than individ-
to construct the instrument, is contained in the data only until the
uals born one month earlier. This is illustrated in Fig. 1, which plots
last quarter of 2002.
the share of individuals with a certain educational qualification in
The HSE is an annual survey conducted since 1991 by the Joint
the respective age cohort along with non-parametric regressions
Health Surveys Unit of the National Centre for Social Research and
for both groups.3 Note that the large increase in individuals with
the Department of Epidemiology and Public Health, Royal Free and
a CSE or O-level qualification from cohorts born around 1957 rela-
University College Medical School, London on behalf of the Depart-
ment of Health. Sample sizes vary between 12,000 and 20,000
individuals depending on the year. The survey involves a question-
3
The non-parametric estimates are obtained using Stata’s lowess command
which provides locally weighted scatter plot smoothing, in this case using a tricube
weight and the default bandwidth. Fig. 1 uses only the LFS data. A similar pattern
4
also exists in the HSE. Additional figures are available from the author on request. See http://www.iser.essex.ac.uk/survey/bhps for details on the BHPS.
756 N. Braakmann / Journal of Health Economics 30 (2011) 753–763

Fig. 1. The relationship between month of birth and qualification obtained, only January and February born individuals. Panel (a): Share of individuals with CSE qualification,
1950–1970, Labour Force Survery. Panel (b): Share of individuals with O-level qualification, 1950–1970, Labour Force Survery. Panel (c): Share of individuals with A-level
qualification, 1950–1970, Labour Force Survery. Panel (d): Share of individuals with university degree, 1950–1970, Labour Force Survery.

naire with a series of core questions as well as questions focusing on
one specific topic each year and is accompanied by a nurse visit to
the respondent’s home for further medical tests. To be comparable
with the LFS sample, I again use data from 1998 to 2002.
I also use some data from wave 13 (collected in 2003/4) of the
BHPS to compare the parental background of January and Febru-
ary born individuals to show that the two groups do not differ in
observable characteristics that are uninfluenced by individual edu-
cation. The BHPS is an annual survey conducted by the ESRC UK
Longitudinal Studies Centre within the Institute for Social and Eco-
nomic Research at the University of Essex. The target population
of the first wave of the survey were adult members of households
with a domestic residence in England, Scotland south of the Caledo-
nian Canal and Wales. In later waves, households that moved into
Scotland north of the Caledonian Canal or from domestic residence
into institutions (excluding prisons) were followed.
In this paper I focus on individuals born between September
1957 and 1970 in the UK and currently living in England. The aca-
demic cohort entering school in September 1957 was the first to
face the new minimum school leaving age of 16, which creates the
interaction between month of birth and the CSE/O-level exams. For
individuals born after 1970 the data increasingly show individuals
having taken the GCSE. These later cohorts are dropped for reasons
of homogeneity.
Applying these restrictions leads to a sample of 55,154 individ-
uals of which 8971 are born in either January or February for the
LFS. 22,270 are men (3621 born in January or February) and 32,884
(5350) are women. Using the same restrictions as above on the HSE
leads to a sample of 15,822 individuals of which 2683 are born in
either January or February. 7033 are men (1179 born in January or
February) and 8789 (1504) are women. The BHPS samples are quite
small with between 132 and 185 observations per gender/month
of birth cell.5
5
The gender imbalance in the LFS estimation sample relative to the raw data,
where the share of women is 51.67%, can be explained as follows: (a) women are less
likely to have missing values in the question regarding any lasting health problem.
Dropping individuals with missing values raises the proportion of women to 55.46%.
(b) Restricting the sample to individuals born between September 1957 and 1970
changes the proportion of women from 55.46% to 59.62%. The reason for this latter
effect is unknown. However, similar shares are found in the HSE and the BHPS.
As there does not seem to be any imbalance between January and February born
individuals, however, one can be confident that the main results are unaffected by
these effects.
 N. Braakmann / Journal of Health Economics 30 (2011) 753–763 757

Table 1
Descriptive statistics, estimation sample.

Variable All individuals Only January and February born

Observations Mean Standard deviation Observations Mean Standard deviation

Labour Force Survey

Has at least a CSE qualification (1 = yes) 55154 0.7774 0.4160 8971 0.7787 0.4151
Born February–August (1 = yes, vs. September–January) 55154 0.5853 0.4927 8971 0.4861 0.4998
Born in February (1 = yes, vs. born in January) 8971 0.4861 0.4998 8971 0.4861 0.4998
Age (years) 55154 35.66 4.03 8971 35.87 3.94
University degree (1 = yes) 55154 0.1623 0.3688 8971 0.1574 0.3642
A levels (1 = yes) 55154 0.1013 0.3017 8971 0.1026 0.3034
O levels (1 = yes) 55154 0.3217 0.4671 8971 0.3252 0.4684
CSE (1 = yes) 55154 1921 0.3939 8971 0.1936 0.3952
Ever had health problem longer than 12 months (1 = yes) 55154 0.2553 0.4361 8971 0.2634 0.4405
Health problem limited/limits activity (1 = yes) 55154 0.1335 0.3401 8971 0.1342 0.3409
Problems with hands, legs, back or neck (1 = yes) 55154 0.1043 0.3057 8971 0.1077 0.3100
Difficulty in seeing or hearing (1 = yes) 55154 0.0181 0.1332 8971 0.0194 0.1379
Disfigurement, skin conditions, allergies (1 = yes) 55154 0.0242 0.1536 8971 0.0265 0.1607
Chest/breathing problems, asthma, bronchitis (1 = yes) 55154 0.0633 0.2435 8971 0.0635 0.2439
Heart, blood pressure, blood circulation problems (1 = yes) 55154 0.0259 0.1588 8971 0.0273 0.1630
Stomach, liver kidney, digestive problems (1 = yes) 55154 0.0313 0.1741 8971 0.0321 0.1763
Diabetes (1 = yes) 55154 0.0082 0.0902 8971 0.0082 0.0905
Depression, bad nerves, anxiety (1 = yes) 55154 0.0352 0.1844 8971 0.0398 0.1955
Epilepsy, mental handicap, mental illness (1 = yes) 55154 0.0274 0.1631 8971 0.0284 0.1662
Male (1 = yes) 55154 0.4038 0.4907 8971 0.4036 0.4907

Health Survey for England

Has at least a CSE qualification (1 = yes) 15822 0.8420 0.3648 2683 0.8259 0.3792
Born February–August (1 = yes, vs. September–January) 15822 0.5925 0.4914 2683 0.4801 0.4997
Born in February (1 = yes, vs. born in January) 2683 0.4801 0.4997 2683 0.4801 0.4997
Age (years) 15822 35.62 3.93 2683 35.72 3.97
Smoker (1 = yes) 15822 0.3061 0.4609 2683 0.3097 0.4625
Drinks over weekly limits (1 = yes) 15822 0.2220 0.4156 2683 0.2203 0.4145
Eats fried food 6 times a week (1 = yes) 15822 0.0111 0.1049 2683 0.0134 0.1151
Eats fried food at least 3 times a week (1 = yes) 15822 0.0610 0.2393 2683 0.0663 0.2489
Eats fruit or vegetables 6 times a week (1 = yes) 15822 0.1857 0.3889 2683 0.1789 0.3833
Eats fruits and vegetables at least 3 times a week (1 = yes) 15822 0.2167 0.4120 2683 0.2042 0.4032
Eats chocolate, biscuits or crisps 6 times a week (1 = yes) 15822 0.1215 0.3267 2683 0.1096 0.3124
Eats chocolate, biscuits or crisps at least 3 times a week (1 = yes) 15822 0.2241 0.4170 2683 0.2117 0.4086
Male (1 = yes) 15822 0.4445 0.4969

Similar to Anderberg and Zhu (2010), my main variable of Hi(ct) = ˛ + ˇ1 ∗ agei(ct) + ˇ2 ∗ age2i(ct) + ˇ3 ∗ age3i(ct)
interest is a dummy variable for having completed any sort of qual-
ification, i.e., at least a CSE. In the following section I will also briefly + c + ıt +  ∗ qual(ct) + εi(ct) , (1)
use a more detailed measure of education distinguishing between
where Hi(ct) is the respective health measure of individual i from
CSE/O-levels, A-levels and university degrees. Finally, I will present
cohort c observed in year t, c and ıt are birth cohort and year effects
some robustness checks using only individuals with at most a CSE
respectively and qual(ct) is a dummy indicating whether individ-
qualification.
ual i has completed at least a CSE qualification. In cases where the
From the LFS, I take a variety of measures on objective health
regression is based on a pooled sample a gender dummy is added as
conditions, e.g., a dummy indicating whether an individual has
an additional control. Note that Hi(ct) may be a dummy in which case
any long-lasting health problem, whether this problem limits the
Eq. (1) is a linear probability model. This fact, however, is not par-
activities the individual can do as well as information on a num-
ticularly problematic in this case as the instrument, the variable of
ber of specific diseases. Furthermore, I use some data on labour
interest as well as almost all control variables are dummy variables,
market outcomes to investigate one potential channel between
which attenuates concerns regarding the linearity assumption typ-
education and better health, specifically differences in income. I
ically leading to differences between linear probability models and
also use a number of variables from the HSE on health related
non-linear models such as Probit (see Angrist, 2001). The only
behaviour, i.e., whether an individual smokes, drinks more than the
exception is age, which is entered as a high-order polynomial. All
recommended limit or regularly eats various types of more or less
estimates are weighted using the weights provided with the data.
healthy foods.
However, results are essentially identical when not weighting the
Table 1 contains basic descriptive statistics on all HSE and LFS
estimates.
variables used in the analysis. Note that there are a consider-
As outlined in the introduction qual(ct) might be correlated with
able number of individuals with long-lasting health problems even
εi(ct) if there are, for instance, common genetic or family back-
though the sample is quite young on average.
ground related factors influencing both an individual’s health and
its propensity to complete a first qualification. This in turn would
2.2. General estimation approach and evidence on instrument bias the estimate for  in Eq. (1). To overcome this problem, I rely
validity on the institutional features outlined in the preceding section and
use a dummy for being February-born as an instrument for hav-
Similar to Anderberg and Zhu (2010), who focus on matching in ing completed any qualification. Most of the analysis will focus on
the marriage market, the main analysis consists of regressions of individuals born either in January or February (henceforth called
the form: the discontinuity sample), although I will also present estimates
758 N. Braakmann / Journal of Health Economics 30 (2011) 753–763

Table 2
Comparison of background variables between January and February born individuals, BHPS.

Variable Mean Std. dev Mean Std. dev p-Value means different

January born February born

Men
Lots of books in household (1 = yes) 0.2587 0.4395 0.2927 0.4569 0.5392
Mother has some qualification (1 = yes) 0.4846 0.5017 0.4732 0.5015 0.8602
Mother has no qualification (1 = yes) 0.5154 0.5017 0.5268 0.5015 0.8602
Father has some qualification (1 = yes) 0.5630 0.4979 0.5565 0.4990 0.9190
Father has no qualification (1 = yes) 0.4370 0.4979 0.4435 0.4990 0.9190
Individual is first born or only child (1 = yes) 0.4514 0.4994 0.4146 0.4947 0.5474
Observations 153 132

Women
Lots of books in household (1 = yes) 0.3801 0.4868 0.4171 0.4945 0.4833
Mother has some qualification (1 = yes) 0.4937 0.5015 0.5122 0.5014 0.7406
Mother has no qualification (1 = yes) 0.5063 0.5015 0.4878 0.5014 0.7406
Father has some qualification (1 = yes) 0.5490 0.4992 0.5484 0.4993 0.9911
Father has no qualification (1 = yes) 0.4510 0.4992 0.4516 0.4993 0.9911
Individual is firstborn or only child (1 = yes) 0.4046 0.4922 0.4134 0.4938 0.8674
Observations 180 185

based on the whole sample for comparison purposes. These esti- Table 3
Probability of obtaining at least a CSE qualification by month of birth, Labour Force
mates use being born between February and August (inclusively)
Survey, OLS.
as an instrument.
For reasons already outlined in the introduction, it should be Born in.. Women Men
kept in mind that the instrument is much more likely to be truly February 0.0478*** 0.0242*
exogenous when focussing only on January- and February-born (0.0114) (0.0140)
individuals. Firstly, while families can – at least to some extent March 0.0343*** 0.0271**
(0.0113) (0.0135)
– plan the season they want to give birth in, which may lead to
April 0.0437*** 0.0430***
differences in parental characteristics for individuals born in dif- (0.0114) (0.0136)
ferent times of the year (Buckles and Hungerman, 2008), this is far May 0.0443*** 0.0332**
less possible with respect to the exact month of birth. Secondly, (0.0113) (0.0136)
June 0.0167 0.0194
while differences in maternal nutrition, weather conditions, sun-
(0.0118) (0.0142)
light exposure, etc. may play a role for explaining differences in July 0.0195 0.0379***
mental or physical health for children born in different seasons (see (0.0119) (0.0140)
the discussion in Bound et al., 1995), these factors can be expected August 0.0216* 0.0264*
to be more or less equal for children born in either January or (0.0122) (0.0143)
September 0.0138 −0.0085
February.
(0.0122) (0.0147)
Table 2 presents some evidence that January and February born October 0.0002 0.0048
individuals indeed do not differ in a number of characteristics that (0.0125) (0.0150)
are uninfluenced by individual education decisions (see Anderberg November 0.0005 0.0041
(0.0129) (0.0154)
and Zhu, 2010, for further robustness checks that paint the same
December 0.0239* −0.0180
picture). This table uses data from the 2003/4 wave of the BHPS (0.0127) (0.0156)
as neither the LFS nor the HSE contains any information on fam-
Observations 32884 22270
ily background. As we can see from the table there are practically
no differences in parental education, the number of books in the Coefficients, robust standard errors in parentheses. */**/***Statistical significance on
the 10%, 5% and 1% level. All estimations control for a cubic polynomial in age and
household – broadly interpreted as evidence on the valuation of
a full set of year of birth and year dummies.
education by the parents – and whether the child was a first-born
or only child. This evidence is reassuring that the institutional setup
described in Section 1 indeed creates a quasi-experimental situa- low first-stage F-values and insignificant and often small point esti-
tion. mates for the dummy indicating that an individual was born in the
Table 3 shows further evidence on the validity of the instru- respective later month.
ment. It regresses a full set of months of birth dummies along with
the other controls on the likelihood of having completed at least 3. Education and health outcomes
a CSE. If the institutional setup described in Section 1 creates the
variation in CSE completion rates seen in Fig. 1, we would expect Table 5 presents first stage results for the February-born instru-
to see higher probabilities of completion for individuals born from ment using the LFS data. For almost all specifications, we observe a
February to August and no real differences between individuals positive influence of being born after the January cut-off on the like-
born between September and January. The evidence in Table 3 is lihood of having obtained a qualification. Similar to the evidence in
broadly consistent with this expected pattern. Fig. 1, the results generally indicate that individual born after the
Finally, if the instrument is valid there should be no difference cut-off raises the probability of having at least a CSE by between 2
in completion rates between individuals born in adjacent months and 4.5%.
other than January/February and August/September. Table 4 plots Restricting the sample to individuals born in January and Febru-
first stage results using a number of other cut-off dates, specifi- ary reduces the statistical power of the analysis. However, with
cally December vs. January, February vs. March and March vs. April. one exception, the relationship between being February-born and
As we can see the first stage is essentially non-existent with very having a qualification becomes stronger, which is the result to be
 N. Braakmann / Journal of Health Economics 30 (2011) 753–763 759

Table 4
First stage results for various pseudo-cut-offs, Labour Force Survey, dependent variable: has at least CSE qualification (1 = yes).

December vs. January February vs. March March vs. April

Men
Born in resp. second month (1 = yes) 0.0142 0.0023 0.0151
(0.0321) (0.0136) (0.0131)
N 3723 3734 3870
r2 0.0164 0.0076 0.0111
F (excluded instruments) .022628 .025402 1.31731

Women
Born in resp. second month (1 = yes) −0.0447 −0.0134 0.0102
(0.0292) (0.0110) (0.0109)
N 5452 5467 5598
r2 0.0104 0.0083 0.0069
F (excluded instruments) .09184 1.46938 .881919

Coefficients, robust standard errors in parentheses. */**/***Statistical significance on the 10%, 5% and 1% level. All estimations control for a cubic polynomial in age and a full
set of year of birth and year dummies.

Table 5
First stage results, Labour Force Survey, dependent variable: has at least CSE qualification (1 = yes).

Men and women Men Women

All individuals Only January All Only January and All Only January and
and February individuals February born individuals February born
born

Born February–August (1 = yes) 0.0284*** 0.0334*** 0.0247***

(0.0039) (0.0061) (0.0050)
Born in February (1 = yes) 0.0363*** 0.0228 0.0471***
(0.0090) (0.0142) (0.0117)
Observations 55,154 8971 22,270 3621 32,884 5350
R2 0.0075 0.0090 0.0111 0.0127 0.0055 0.0109
F (excluded instruments) 53.68 15.86 30.30 2.46 24.23 15.86

Coefficients, robust standard errors in parentheses. */**/***Statistical significance on the 10%, 5% and 1% level. All estimations control for gender (where appropriate), a cubic
polynomial in age and a full set of year of birth and year dummies.

expected when the institutional explanation outlined in Section
1 is responsible for this relationship. Additionally, the first stage
values of the F statistics generally confirm the absence of weak
instrument problems. The one exception is the male discontinu-
ity sample. However, even here the (insignificant) point estimate
for being February-born indicates a large effect on the likelihood
of having obtained any qualification. The relative weakness of the
instrument in that specification, which is also indicated by the low
F statistics, does not need to be problematic as Angrist and Pischke
(2009) argue that weak instruments do not need to be a major prob-
lem in just identified models like the one used here. However, the
potential problems in this sample should be kept in mind when
discussing the main results.
I conducted some additional analyses on the changes the instru-
ment causes in the educational distribution by running regressions
of the respective instrument and the control variables from Eq.
(1) on dummy variables for various qualifications, specifically for
having completed university, A-levels, O-levels or a CSE.6 I also
looked at F-values obtained by treating these regressions as first
stages for the respective dependent variable. As can be expected
for an instrument keeping individuals in school just long enough
for them to take the first possible exam, we only see an influence on
the probability of having completed the CSE. Here, being February-
born raises the probability of having completed that qualification
by between 1 and 4% with again weaker and insignificant results
being found for the male discontinuity sample. The changes in the
probability of having completed A-levels or a university degree,
however, are close to zero in all samples and consequently always
insignificant. For O-levels some differences can be seen when using
the whole sample. Point estimates for the discontinuity sample,
however, are always small and insignificant. The first stage F-
values point a very similar picture: the instrument only has power
when looking at CSE-completion, while it is generally weak for the
higher-level qualifications. These results strengthen the idea that
the differences in educational attainment between January- and
February-born individuals are indeed caused by the institutional
setting described in Section 1.
Now consider the main analysis whose results are displayed in
Table 6. As results were very similar for men and women, only the
results for the pooled sample are presented. Results for the other
estimations are available on request. Note first that the OLS results
in the samples using all individuals and in the discontinuity sam-
ples are always very similar, which is a sign that individuals born in
January and February are not that different from other individuals
when it comes to the relationship between education and health.
As one would expect the estimates support a positive relationship
between education and health: individuals with any qualification
are always much less likely to have a health problem and to be lim-
ited by it or to have any of the specific diseases that are considered
in the analysis. These effects are also often economically large and
always highly significant.
This picture changes when looking at the IV-results: the pattern
of point estimates in all samples becomes more erratic, suggesting a
more or less random pattern of positive and negative relationships
between education and the various health measures.7 Additionally,

6
Tables for these regression results are available from the author on request.
Anderberg and Zhu (2010, Table 6) present similar evidence using a slightly different
classification of educational qualifications.
7
These results are also confirmed when plotting the health outcomes of January-
and February-born individuals over birth-cohorts similarly to Fig. 1.
 760 N. Braakmann / Journal of Health Economics 30 (2011) 753–763

Table 7

Each cell is from a different regression. Coefficients, robust standard errors in parentheses. */**/***Statistical significance on the 10%, 5% and 1% level. All estimations control for gender, a cubic polynomial in age and a full set
Epilepsy, mental

First stage results, Health Survey for England, dependent variable: has at least CSE
mental illness

qualification (1 = yes).

−0.0374***

−0.0426***
handicap,

(0.0551)

(0.0988)
(0.0022)

(0.0057)
Men and women

0.0979*

55,154

0.0971
55154

8971

8971
All individuals Only January and
February born
Depression, bad
nerves, anxiety

Born February–August (1 = yes) 0.0209***

(0.0063)
−0.0384***

−0.0405***
Born in February (1 = yes) 0.0339**

−0.1744

(0.1164)
(0.0024)

(0.0603)

(0.0062)
55,154
0.0183
55154

(0.0145)

8971

8971
Observations 15,822 2682
R2 0.0229 0.0474
−0.0046*
Diabetes

Cragg–Donald Wald F statistic 11.26 5.18

−0.0012

(0.0552)
(0.0304)

(0.0027)
(0.0010)

0.0531*

55,154

0.0168
55154

Kleibergen–Paap Wald rk F statistic 11.02 5.22

8971

8971
Coefficients, robust standard errors in parentheses. */**/***Statistical significance on
the 10%, 5% and 1% level. All estimations control for gender, a cubic polynomial in
age and a full set of year of birth and year dummies.
Stomach, liver

−0.0122***
problems
digestive

−0.0070
(0.0558)

(0.1026)
(0.0048)
(0.0020)
kidney,

55,154

0.1135
0.0362
55154

8971

8971
all estimates are insignificant. Using Anderson–Rubin-tests that are
robust to weak instruments does not change that picture, in fact p-
pressure, blood

values are generally almost identical. Note that these results are
Heart, blood

−0.0198***
−0.0170***
circulation

not simply a result of large standard errors rendering otherwise

problems

−0.0486
(0.0517)

(0.0928)
(0.0019)

(0.0049)

sensible estimates insignificant. To the contrary, the results show

55,154
0.0427
55154

8971

8971

an almost equal number of positive and negative point estimates.

Note that the difference between the OLS and the IV results is
Chest/breathing

not surprising in itself as these two techniques estimate differ-

ent effects. In particular, the IV-estimates are LATE-estimates for
−0.0288***

−0.0341***
bronchitis
problems,

−0.0554

(0.1379)
(0.0788)
(0.0028)

(0.0071)

those individuals who changed their educational status due to the

asthma,

55,154

0.0149
55154

8971

8971

instrument, in other words the changes in health due to some indi-

viduals being nudged into completing a first qualification by them
being born in February. However, the IV results certainly do not
skin conditions,
Disfigurement,

provide much support for a causal relationship between education

−0.0057***

and health. These results are similar to a number of other studies,

−0.0019
allergies

(0.0492)

(0.0917)
(0.0017)

(0.0041)
55,154

0.0973
0.0008
55154

e.g., Arendt (2005), Doyle et al. (2005), Albouy and Lequien (2009)
8971

8971

and Clark and Royer (2010), using changes in compulsory school-

The impact of education on health outcomes, Labour Force Survey, OLS and IV results, men and women.

ing laws as well as to some of the studies using other identification

strategies.
Difficulty in

−0.0114***

−0.0160***

To sum up the current results: while the institutional setting

seeing or

−0.1187
−0.0294

(0.0444)

(0.0828)
(0.0016)

(0.0043)
hearing

55,154

described in Section 1 creates large discontinuities in education

55154

8971

8971

between January- and February-born individuals, there do not

seem to be comparable discontinuities in various health measures.
hands, legs, back

How can these results be explained? A first explanation is that there

Problems with

is indeed no causal relationship between education and health and

−0.0556***

−0.0648***

−0.0329
−0.0092

(0.1761)
(0.0985)
(0.0035)

(0.0090)

that all observed health differences between individuals with dif-

or neck

55,154
55154

8971

8971

ferent levels of education are caused by third factors like genetic

endowments or family background.
A second and related explanation is that while there might a
Health problem
limited/limits

causal relationship between education and health, that relationship

−0.0941***

−0.1059***

simply does not operate on the no qualification/low qualification

−0.2809

(0.1967)
(0.1105)

(0.0102)
(0.0040)
activity

55,154
0.0216
55154

8971

8971

margin. In fact, Cutler and Lleras-Muney (2010, p. 3) point out that

the relationship between education and health becomes stronger
OLS estimates: only January and February born

as one moves up the educational distribution. As the instrument

than 12 months
Ever had health
problem longer

used here is only informative about changes in the lower end

IV results: only January and February born
−0.1176***
−0.1020***

of the educational distribution, this explanation seems possible.

−0.1390
(0.1418)

(0.0121)

(0.2506)
(0.0048)

55,154
0.0686
55154

I tested this explanation with a restricted sample containing only

8971

8971

of year of birth and year dummies.

individuals with at most a CSE, i.e., only individuals without any

OLS estimates: all individuals

qualification or with a CSE. The OLS results generally indicate con-

At least CSE qualification

At least CSE qualification

At least CSE qualification

At least CSE qualification

IV results: all individuals

siderable health advantages for individuals with a CSE relative to

those without any qualification. The IV-estimates, however, show
the now familiar pattern of results and again indicate no significant
relationship between having a CSE and health outcomes.8
Observations

Observations

Observations

Observations
(1 = yes)

(1 = yes)

(1 = yes)

(1 = yes)
Table 6

8
Detailed results are available on request.
 N. Braakmann / Journal of Health Economics 30 (2011) 753–763 761

Each cell is from a different regression. Coefficients, robust standard errors in parentheses. */**/***Statistical significance on the 10%, 5% and 1% level. All estimations control for gender, a cubic polynomial in age and a full set
4. Education and health related behaviour

Eats chocolate, biscuits

and crisps at least 3
Table 7 presents first stage results for the relationship between
month of birth and education in the HSE. Given the smaller sample

times a week
size I refrain from splitting the sample by gender and present only

0.0364***

0.0401**

−0.2667
(0.2723)

(0.0171)
(0.0072)

(0.3800)
results for the whole sample. The results are generally very similar

15,822

0.3555

15,822

2682

2682
to the ones obtained using the LFS. Due the smaller sample sizes the
first stage F-values are smaller and in fact slightly problematic in
the discontinuity sample. This potential weak instrument problem
should be kept in mind when looking at the results from the main

biscuits or crisps 6
analysis.

Eats chocolate,

times a week
Turning to the analysis of interest in Table 8, we see the

−0.0009
(0.2316)

(0.0144)

(0.3170)
(0.0062)
expected positive correlation between education and health

15,822

15,822
0.0963

0.0812
0.0070

2682

2682
related behaviour in the OLS estimates: individuals with a qual-
ification are less likely to smoke or to drink excessively, eat less
fried food and more vegetables and fruits than individuals with-
out a qualification. The only health related behaviour where the

vegetables at least
higher-educated fare worse is in their higher frequency to consume

3 times a week
Eats fruits and
chocolate, biscuits and sweets.

0.0526***

0.0598***

−0.0469
(0.3761)
(0.2710)

(0.0186)
(0.0078)
Now consider the IV-estimates for both samples. The picture

15,822

0.1987

15,822

2682

2682
that emerges is again very similar to the one obtained in the pre-
vious section: the point estimates show again an erratic pattern of
positive and negative results and are always insignificant.9 In other
words, the results show again no support for a causal relationship

vegetables 6 times
between education and health, although the caveats mentioned in
the previous section should be kept in mind.

Eats fruit or

0.0347***

0.0560***

−0.2132
−0.1033
(0.2608)

(0.0175)

(0.3809)
(0.0074)
a week

15,822

15,822

2682

2682
5. Why are there no effects?

As outlined in the introduction, there are three broad reasons

why we might expect a causal relationship between education and
Eats fried food at
least 3 times a

health. In the following I will present some reasoning why there

−0.0383***

−0.0482***
might not be an effect on health for the particular quasi-experiment

−0.3017

−0.0710
(0.2754)
(0.2061)

(0.0150)
(0.0060)

considered in this paper—even though it changes the likelihood of

15,822

15,822
week

2682

2682
having a qualification considerably.
Note first that February born individuals gain only a relatively
The impact of education on health related behaviour, Health Survey for England, OLS and IV results.

small amount of additional time in education, specifically one term.

This fact suggests that the productivity-based explanations appear
Eats fried food 6

to be less relevant in this particular case. It is also the main dif-

times a week

−0.0165***

−0.0237***

ference to papers looking at increases in compulsory schooling,

−0.1627*

−0.0305
(0.1324)
(0.0979)
(0.0032)

(0.0088)

as these are primarily changes in the duration of education. This

15,822

15,822

2682

2682
experiment in contrast essentially helps February born individuals
in acquiring a signal that might be valued in the labour market. The
main causal channel for a potential effect on health would thus be
weekly limits

income differences.
Drinks over

Table 9 present regression estimates using employment and two

0.0573***

0.0871***

−0.4112
(0.3469)

(0.5171)
(0.0083)

(0.0190)
15,822

0.3582

15,822

measures of wages as outcomes. The results are qualitatively iden-

2682

2682

tical when restricting the sample to individuals with at most a CSE

and looking at differences between CSEs and no qualifications at all.
The OLS estimates in each case suggest considerable returns to hav-
−0.1763***

−0.1366***
OLS estimates: only January and February born

ing any qualification as well as to having a CSE relative to having

−0.1377
−0.2068
(0.3752)

(0.5289)
(0.0107)

(0.0250)
Smoker

15,822

15,822

no qualification. The IV estimates, however, are always insignifi-

2682

2682
IV results: only January and February born

cant with again mixed point estimates. How can these results be
reconciled? Note first that the individuals who take and pass their
At least CSE qualification (1 = yes)

At least CSE qualification (1 = yes)

At least CSE qualification (1 = yes)

At least CSE qualification (1 = yes)

CSE exams due to the institutional setup considered here are very
of year of birth and year dummies.
OLS estimates: all individuals

likely marginal passes. This in turn might imply that employers

consider them as being closer to individuals without a qualifica-
IV results: all individuals

tion than to individuals who passed their CSE with better marks.
As the OLS estimates do not distinguish between these two groups,
while the IV estimates would give the effects for the marginal cases,
Observations

Observations

Observations

Observations

the difference can be explained on these grounds—in particular as

Table 8

9
This result is again robust to using Anderson–Rubin-tests.
762 N. Braakmann / Journal of Health Economics 30 (2011) 753–763

Each cell is from a different regression. Coefficients, robust standard errors in parentheses. */**/***Statistical significance on the 10%, 5% and 1% level. All estimations control for gender, a cubic polynomial in age and a full set
employers are very likely able to observe marks. As the individuals
who were nudged into completing the CSE – the compliers from

17710.0000

17710.0000
the IV estimates – apparently do not enjoy a labour market advan-

2875.0000

2875.0000
Ln(hourly

0.4243***
0.4303***

−0.2667
(0.4389)

(0.4424)
(0.0226)
(0.0090)
tage over individuals without a qualification, it seems plausible that

0.4282
wage)

there is also no effect on health.

6. Conclusion
weekly pay)

17777.0000

17777.0000

2887.0000

2887.0000
0.5998***

0.5703***
In this paper, I used a natural experiment in England that cre-
Ln(gross

−0.4396
(0.0173)

(0.7046)

(0.0441)

(0.7306)
0.5540
ated exogenous variation in the likelihood to obtain any sort of
qualification between January and February born individuals for
13 academic cohorts in England. For these cohorts compulsory
schooling laws interacted with the timing of the CSE and O-level
exams to change the probability of obtaining a qualification by
32884.0000

32884.0000

5350.0000

5350.0000
Employed

0.2516***

0.2527***

around 3 percentage points. Using data from the Labour Force Sur-
−0.1239

−0.0298
(0.2254)

(0.0161)

(0.2685)
(0.0065)
(1 = yes)
Women

vey and the Health Survey for England, I then showed that these
within-cohort differences in education did not transform into cor-
responding differences in various objective health measures or in
health related behaviour like smoking or drinking. While OLS esti-
mates show the expected influence between having a qualification
13691.0000

13691.0000

2214.0000

2214.0000
Ln(hourly

0.3919***
0.4092***

and the outcomes in samples using all individuals as well as in a dis-

(0.3272)

(1.2582)
(0.0259)
(0.0103)

0.3378

1.3057
wage)

continuity sample using only individuals in January and February,

this relationship disappears in both samples when instrumenting
education by being February-born. The results consequently do not
show support for a causal link between education and health—at
least not for the individuals being affected by the particular inter-
weekly pay)

13744.0000

13744.0000

2222.0000

2222.0000
0.3636***

0.3460***

vention considered here. It is important to stress though that the

Ln(gross

(1.2095)
(0.0121)

(0.3505)

(0.0305)
0.3356

1.0105

results do not rule out a causal link between higher forms of edu-
cation and health. As the institutional setting considered here only
affect individuals at the margin of completing a first qualification,
no statements can be made regarding changes in the higher end of
the educational distribution. However, the results are in line with
22270.0000

22270.0000

3621.0000

3621.0000
Employed

0.1894***

0.1833***

some of the previous evidence using changes in compulsory school-

−0.0408
(0.1458)

(0.4816)
(0.0165)
(0.0067)
(1 = yes)

0.2511

ing laws (Arendt, 2005; Doyle et al., 2005; Albouy and Lequien,
Men

2009; Clark and Royer, 2010) and contradict some other studies
using the same identification strategy (Spasojevic, 2003; Glied and
Lleras-Muney, 2003; Lleras-Muney, 2005; Oreopoulos, 2006). In
sum, the question whether there is a causal link between education
31401.0000

31401.0000

5089.0000

5089.0000
Ln(hourly

0.4210***

0.4096***

and health seems to be open.

(0.2613)

(0.0171)

(0.4510)
(0.0068)

0.6588
0.0724
wage)

Author note
Evidence on labour market outcomes, Labour Force Survey, OLS and IV results.

All analyses used Stata 11.1. Do-Files are available from the
weekly pay)

31521.0000

31521.0000

author on request. The data used in this study can be obtained

5109.0000

5109.0000
0.4614***
0.4906***
Ln(gross

(0.3675)

(0.0271)

(0.6220)
(0.0109)

from the Economic and Social Data Service, see www.esds.ac.uk/

0.5984
0.0053

for details. All analyses and opinions expressed in this paper as

well as any possible errors are under the sole responsibility of the
author.
Men and women

55154.0000

55154.0000

8971.0000

8971.0000
Employed

References
0.2258***

0.2230***
OLS estimates: only January and February born
−0.0864
(0.1343)

(0.0116)

(0.2360)
(0.0047)
(1 = yes)

0.0411
IV results: only January and February born

Adams, S.J., 2002. Educational attainment and health: evidence from a sample of
older adults. Education Economics 10 (1), 97–109.
Adams, P., Hurd, M.D., McFadden, D., Merrill, A., Ribeiro, T., 2003. Healthy, wealthy,
and wise? Tests for direct causal paths between health and socioeconomic sta-
At least CSE qualification (1 = yes)

At least CSE qualification (1 = yes)

At least CSE qualification (1 = yes)

At least CSE qualification (1 = yes)

of year of birth and year dummies.

tus. Journal of Econometrics 112 (1), 3–56.

OLS estimates: all individuals

Albouy, V., Lequien, L., 2009. Does compulsory education lower mortality? Journal
of Health Economics 28 (1), 155–168.
IV results: all individuals

Anderberg, D., Zhu, Y., 2010. The effect of education on marital status and part-
ner characteristics: evidence from the UK. In: CESifo Working Paper 3104,
June 2010.
Angrist, J.D., 2001. Estimation of limited-dependent variable models with dummy
Observations

Observations

Observations

Observations

endogenous regressors: simple strategies for empirical practice. Journal of Busi-

ness and Economic Statistics 19 (1), 2–28.
Angrist, J.D., Krueger, A.B., 1991. Does compulsory school attendance affect school-
Table 9

ing and earnings? Quarterly Journal of Economics 106 (4), 979–1014.

Angrist, J.D., Pischke, J.S., 2009. Mostly Harmless Econometrics—An Empiricist’s
Companion. Princeton University Press.
 N. Braakmann / Journal of Health Economics 30 (2011) 753–763
Arendt, J.N., 2005. Education effects on health: a panel data analysis using school
reform for identification. Economics of Education Review 24 (2), 149–160.
Arkes, J., 2003. Does Schooling Improve Health? Working Paper. RAND Corporation.
Auld, M.C., Sidhu, N., 2005. Schooling, cognitive ability, and health. Health Economics
14 (10), 1019–1034.
Berger, M.C., Leigh, J.P., 1989. Schooling, self-selection, and health. Journal of Human
Resources 24 (3), 433–455.
Braakmann, N., 2010. A note on the causal link between education and
health—evidence from the German short school years. University of Lueneburg,
Working Paper in Economics 176.
Bound, J., Jaeger, D.A., Baker, R., 1995. Problems with instrumental variables esti-
mation when the correlation between the instruments and the endogeneous
explanatory variable is weak. Journal of the American Statistical Association 90
(430), 443–450.
Buckles, K., Hungerman, D.M., 2008. Season of birth and later outcomes: old ques-
tions, new answers. In: NBER Discussion Paper 14573.
Cipollone, P., Radicchia, D., Rosolia, A., 2006. The effect of education on youth mor-
tality. In: Working Paper. Bank of Italy.
Clark, D., Royer, H., 2010. The Effect of Education on Adult Health and Mortality:
Evidence from Britain. NBER Working Paper No. 16013.
Commission on Social Determinants of Health, 2008. Closing the Gap in a Genera-
tion: Health Equity through Action on the Social Determinants of Health. World
Health Organization Final Report, Geneva.
Crawford, C., Dearden, L., Meghir, C., 2007. When you are born matters: the impact
of date of birth on child cognitive outcomes in England. In: CEE Discussion Paper
93.
Cutler, D., Lleras-Muney, A., 2010. Understanding differences in health behaviors by
education. Journal of Health Economics 29 (1), 1–28.
DeWalque, D., 2003. How does education affect health decisions? The cases of
smoking and HIV/AIDS. Ph.D. Dissertation, University of Chicago, Department
of Economics.
DeWalque, D., 2004. Education, information, and smoking decisions, evidence from
smoking histories 1940-2000. In: World Bank policy research working paper
3362.
763
DeWalque, D., 2007. Does education affect smoking behaviors? Evidence using the
Vietnam draft as an instrument for college education. Journal of Health Eco-
nomics 26, 877–895.
Doyle, O., Harmon, C., Walker, A., 2005. The impact of parental income and education
on the health of their children. In: IZA Discussion Paper 1832.
Glied, S., Lleras-Muney, A., 2003. Health inequality, education and medical innova-
tion. In: NBER Working Paper 9738.
Goldman, D.P., Smith, J.P., 2002. Can patient self-management help explain the
SES health gradient? Proceedings of the National Academy of Science 99 (16),
10929–10934.
Groot, W., Maassen van den Brink, H., 2007. The health effects of education. Eco-
nomics of Education Review 26 (2), 186–200.
Grossman, M., 1972. On the concept of health capital and the demand for health.
Journal of Political Economy 70 (2), 223–255.
Grossmann, M., 2006. Education and nonmarket outcomes. In: Eric Hanushek, Finis
Welch (Eds.), Handbook of the Economics of Education, vol. 1. Elsevier Science,
Amsterdam: North-Holland, pp. 577–633.
Kenkel, D.S., 1991. Health behavior, health knowledge, and schooling. Journal of
Political Economy 99 (2), 287–305.
Kenkel, D.S., Lillard, D.R., Mathios, A.D., 2006. The roles of high school completion
and GED receipt on smoking and obesity. Journal of Labor Economics 24 (3),
635–660.
Lleras-Muney, A., 2005. The relationship between education and adult mortality in
the United States. Review of Economic Studies 72 (1), 189–221.
Michael, R.T., 1973. Education in nonmarket production. Journal of Political Econ-
omy 81 (2), 306–327.
Oreopoulos, P., 2006. Estimating average and local average treatment effects of
education when compulsory schooling laws really matter. American Economic
Review 96 (1), 152–175.
Spandorfer, J.M., Karras, D.J., Hughes, L.A., Caputo, C., 1995. Comprehension of dis-
charge instructions by patients in an urban emergency department. Annals of
Emergency Medicine 25 (1), 71–74.
Spasojevic, J., 2003. Effects of education on adult health in Sweden: results from a
natural experiment. Ph.D. Thesis, City University of New York Graduate Center.