Parents' Education and Child Body Weight in France: The Trajectory of the Gradient in the Early Years

Full text

Parents’ education and child body weight in France:
The trajectory of the gradient in the early years
B´en´edicte H. Apouey, Pierre-Yves Geoffard
To cite this version:
B´en´edicte H. Apouey, Pierre-Yves Geoffard. Parents’ education and child body weight in
France: The trajectory of the gradient in the early years. PSE Working Papers n 2015-36.
2015. <halshs-01223321>
HAL Id: halshs-01223321
https://halshs.archives-ouvertes.fr/halshs-01223321
Submitted on 2 Nov 2015
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entific research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destin´ee au d´epˆot et `a la diffusion de documents
scientifiques de niveau recherche, publi´es ou non,
´emanant des ´etablissements d’enseignement et de
recherche fran¸cais ou ´etrangers, des laboratoires
publics ou priv´es.

WORKING PAPER N° 2015 – 36
Parents’ education and child body weight in France: The trajectory
of the gradient in the early years
Bénédicte H. Apouey
Pierre-Yves Geoffard
JEL Codes: I12
Keywords: Children; BMI-for-age z-score; Body Weight; Overweight;
Socioeconomic Status; Education
PARIS-JOURDAN SCIENCES ECONOMIQUES
48, BD JOURDAN – E.N.S. – 75014 PARIS
TÉL. : 33(0) 1 43 13 63 00 – FAX : 33 (0) 1 43 13 63 10
www.pse.ens.fr
CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE – ECOLE DES HAUTES ETUDES EN SCIENCES SOCIALES
ÉCOLE DES PONTS PARISTECH – ECOLE NORMALE SUPÉRIEURE – INSTITUT NATIONAL DE LA RECHERCHE AGRONOMIQUE

Parents’ education and child body weight in France:
The trajectory of the gradient in the early years
B´en´edicte H. Apouey, Pierre-Yves Geoffard
October 30, 2015
B´en´edicte H. Apouey (corresponding author)
Paris School of Economics - CNRS, 48, Boulevard Jourdan, 75014 Paris, France.
Phone: 33-1-43-13-63-07.
Fax: 33-1-43-13-63-55.
E-mail address: benedicte.apouey@psemail.eu; benedicte.apouey@gmail.com.
Pierre-Yves Geoffard
Paris School of Economics - CNRS, 48, Boulevard Jourdan, 75014 Paris, France.
E-mail address: pierre-yves.geoffard@psemail.eu.
1

ABSTRACT
This paper explores the relationship between parental education and offspring body
weight in France. Using two large datasets spanning the 1991-2010 period, we examine the
existence of inequalities in maternal and paternal education and child reported body weight
measures, as well as their evolution across childhood. Our empirical specification is flexible
and allows this evolution to be non-monotonic. Significant inequalities are observed for
both parents’ education – maternal (respectively paternal) high education is associated
with a 7.20 (resp. 7.10) percentage points decrease in the probability that the child is
reported to be overweight or obese, on average for children of all ages. The gradient with
respect to parents’ education follows an inverted U-shape across childhood, meaning that
the association between parental education and child body weight widens from birth to age
8, and narrows afterward. Specifically, maternal high education is correlated with a 5.30
percentage points decrease in the probability that the child is reported to be overweight
or obese at age 2, but a 9.62 percentage points decrease at age 8, and a 1.25 percentage
point decrease at age 17. The figures for paternal high education are respectively 5.87,
9.11, and 4.52. This pattern seems robust, since it is found in the two datasets, when
alternative variables for parental education and reported child body weight are employed,
and when controls for potential confounding factors are included. The findings for the
trajectory of the income gradient corroborate those of the education gradient. The results
may be explained by an equalization in actual body weight across socioeconomic groups
during youth, or by changes in reporting styles of height and weight.
HIGHLIGHTS
•We examine the gradient in parental education and reported child body weight across
childhood.
•Maternal high education is associated with a 7.20 percentage points decrease in child
overweight.
•Our specification allows the evolution of the gradient to be non-monotonic with child
age.
•The gradient follows an inverted U-shape with age: maternal high education is
correlated with 5.30, 9.62, and 1.25 percentage points decreases in child overweight at
ages 2, 8, and 17, respectively.
•The results suggest an equalization in body weight or a change in reporting style in
youth.
JEL classification: I12.
Keywords: Children; BMI-for-age z-score; Body Weight; Overweight; Socioeconomic Sta-
tus; Education.
2

1 Introduction
The overweight rate among youth is relatively low in France and it has remained stable
between 1990 and 2006. However, recent trends in overweight for boys are a public health
concern: their rate of overweight has increased by 2.8 percentage points between 2006 and
2010. In 2010, around 15% of children were overweight (OECD, 2014). Overweight reduces
quality of life in the physical, emotional, social, and school functioning domains (Schwim-
mer et al., 2003). In addition, overweight in childhood has adverse health consequences in
adulthood, including diabetes, cardiovascular diseases, some cancers, and premature mor-
tality (Reilly and Kelly, 2011). There is also some evidence that overweight is associated
with lower skills attainment, although this result is mixed (Murasko, 2015; Palermo and
Dowd, 2012). Finally, overweight in the early years is associated with poorer labor market
outcomes (Sargent and Blanchflower, 1994). Therefore, the importance of understanding
the determinants of child body weight is clear.
Family socioeconomic status (SES) tends to be inversely related to child body weight
(Apouey and Geoffard, 2014; Costa-Font and Gil, 2012; Murasko, 2009; Steckel, 1995).
This association may be explained by two mechanisms.1First, SES may have an impact
on body weight, for instance if low-SES families are more likely to buy high-fat, energy-
dense foods because they are less expensive (Drewnowski and Specter, 2004), or if low-SES
individuals live in neighborhoods in which access to healthy food is limited (Baker et al.,
2006). Second, the association between family SES and child body weight may be an
artefact due to the omission of third factors, such as genetics.
The evolution of the gradient in family SES and child body weight over childhood
is a rather unexplored topic. On the one hand, one could expect that the association
between SES and body weight increases with age. That will be the case if body weight
develops over long periods and reflects the cumulative effect of family SES. On the other
hand, the increase in social inequalities across childhood may not be irreversible. Indeed,
older children and adolescents may progressively detach themselves from their parents and
1Child body weight should only have a limited effect on family SES. Suppose that our SES variable of
interest is education. Note first that in France, people tend to have children after they leave the education
system – in 2010, parents have their first child at age 28 (INED, 2010), which is a long time after they
finish full time education. Note also that among the small proportion of young parents who stop going
to school, the share of those who do it because of their child weight must be limited. As a consequence,
we expect the impact of child body weight on parents’ education to be negligible. On a related issue,
the literature on the gradient in family income and child general health has established that the impact
of child general health on family income is statistically small and only explains a very small share of the
gradient (Apouey and Geoffard, 2013; Case et al. 2002).
3

spend more time with their peers. As a result, they may all be influenced by the same
social norms regarding ideal body weight irrespective of their family SES, and low-SES
children may be actively controlling their body weight to lose weight, which would lead
to a decrease in the gradient with age. Using data on children ages 2-19, Murasko (2009)
finds that the association between income and obesity is stable with age in the US. In
a sample of individuals transitioning from early to middle adulthood, Baum and Ruhm
(2009) find that the gradient in maternal education and offspring body weight widens with
offspring age in the US. This implies that the negative effect of poor maternal education
accumulates over child lives. This result is consistent with previous findings on child
general health by Case et al. (2002), who highlight that the correlation between family
income and child general health strengthens as children grow older in the US.
Our article asks whether family SES matters for child body weight, and how the
relationship between SES and child body weight changes as children get older in France.
Our data come from two surveys, the Survey on Health and Medical Care (“Enquˆete sur la
Sant´e et les Soins M´edicaux” – ESSM) for 1991-1992 and 2002-2003 on the one hand, and
the Health, Health Care, and Insurance Survey (“Enquˆete sur la Sant´e et la Protection
Sociale” – ESPS) for 1996-2010 on the other hand, which contain approximately 40,000
observations of children in total.
Family SES can be defined in a number of ways, using for instance education, income,
wealth, or profession. Education and income capture slightly different aspects of family
SES: education by itself may be related to social norms, awareness, knowledge, and access
to information, whereas income should capture the ability to pay. This explains why in the
published literature the trajectories of the education and income gradient are sometimes
different: in Case et al. (2002) for instance, the gradient in education and child general
health is stable as children grow older, while the gradient in income strengthens with age.
In this paper, we focus on parents’ education and we measure it in two different ways.
However, we also present results on income, which are consistent with those on education
but are less robust (see Section 4.2.1).2
There are two types of anthropometric data: in the objective approach, height and
2We focus on education rather than income for several reasons. First, even if our results on income
and education are consistent, the findings on education are clearer and more robust. This might be due to
measurement error in the income variables. Second, for space reasons. Third, although we are not aware
of any study that compares the role of income with that of education on child body weight in France, there
is some suggestive evidence that education matters more than income in explaining body weight. Indeed,
a report argues that education is a stronger determinant of diet quality than income in France (Recours
and Hebel, 2006).
4

weight are measured, whereas in the subjective approach, height and weight are self-
reported. Reported data are important because they capture how people perceive their
body, and perceived body weight will have an influence on weight control, eating behaviors,
and physical activity – individuals who do not perceive that they are obese won’t try to
lose weight. The ESPS and ESSM surveys contain reported data, and using these piece of
information, we compute child body mass index (BMI)-for-age z-score and a dichotomic
variable for whether the child is overweight or obese.
To uncover the evolution of the gradient across childhood, we start by investigating
the correlation between parents’ education and child body weight separately by child age
groups, using the ESSM data. The results enable us to develop a specification that ac-
counts for the shape of the gradient across childhood. We then examine whether this shape
is statistically significant using the large sample of children of all ages. Overall, our em-
pirical specification allows for the possibility that during childhood there may be periods
of convergence among children from diverse social environments, followed by periods of
divergence, or the contrary. In other words, we account for the fact that inequalities may
not monotonically increase with age, which means that they may not necessarily be cumu-
lative. In further analysis, we investigate whether our results are robust to the inclusion
of potential confounding factors (e.g. parents’ labor market status, family income, and
parents’ body weight). At the same time, we look at the trajectory of the income gradient.
We also re-estimate our model (for education) including cohort or child fixed effects, since
the age-profile of the gradient may be obscured by cohort or child characteristics. In a
subgroup analysis, we examine the gradient separately by child gender and survey year.
This paper contributes to the literature on social disparities in body weight and health
in the early years on several fronts. First, we contribute to the literature on health inequal-
ities in France. In 2000, Anne Tursz lamented that knowledge on the relationship between
social environment and child health in France was limited and we are under the impression
that only little progress has been made since then (Tursz, 2000). In particular, the existing
literature on the gradient in childhood uses data from specific child age groups or specific
regions (Currie et al., 2012; Guignon and Bad´eyan, 2002; Kaminski and Saurel-Cubizolles,
2000; Klein-Platat et al., 2003), with very few exceptions (Apouey and Geoffard, 2014,
2015). It is thus an open question whether the findings can be generalized to all French
children. In contrast, in this article, we study inequalities using two nationally representa-
tive samples of French children, which hopefully enables us to get general results. Second,
5

by focusing on France, we also complement the literature on the evolution of the social
gradient in child anthropometric measurements across childhood, which has solely focused
on the UK and the US so far. Indeed, the results for the UK and the US may not hold
in France because the health care system is different,3the prevalence of overweight and
obesity is smaller,4and social norms are probably stronger in France.5Third, to uncover
the evolution of the gradient across childhood, we use an econometric specification that
allows the impact of parental education on child outcomes to be non-monotonic with child
age. In contrast, the recent literature often assumes that the gradient in body weight is
a monotonic function of child age. Fourth, we study the impact of paternal education
on child body weight outcomes in addition to maternal education. Previous research has
generally investigated the role of the mother’s education only.
2 Background
A negative association between economic conditions and child body weight is found in
most developed countries (Classen and Hokayem, 2005; Murasko, 2009, 2011; Zhang and
Wang, 2006). In what follows, we only present the small number of papers that focus
on France. Using data on children ages 2 to 17 from the 1996-2010 waves of the ESPS,
Apouey and Geoffard (2014) highlight that family income is negatively correlated with
child BMI-for-age and overweight status. For a sample of 30,000 children ages 5 and 6 in
2000-2001, Guignon and Bad´eyan (2002) show that children living in poorer areas are more
likely to have weight problems. Employing a dataset containing 2863 adolescents aged 12
from the Department of Bas-Rhin in 2001, Klein-Platat et al. (2003) highlight that the
mother’s educational level and the family income tax level are negatively and significantly
correlated with the probability of overweight or obesity, whereas the father’s educational
level is not.6Finally, in the 2009-2010 Health Behaviour in School-Aged Children (HBSC)
3France has a mixed public-private social insurance system, and private insurance provides cover against
public sector co-payments on prescription medicines and dental care in particular; in the UK, the system
is public and tax-financed; and in the US, the system is predominantly private.
4The share of overweight or obese children is 13.1% for boys and 14.9% for girls in France (in 2006-
2007, for children ages 3-17, using the IOTF standards), 35.6% for boys and 36.3% for girls in Eng-
land (in 2013, for children ages 11-15, using the 85th percentile cut-off), 33.6% for boys and 27.4%
for girls in Scotland (in 2012, for children ages 2-15, using the 85th percentile cut-off), and 35.1 for
boys and 33.2 for girls in the US (in 2009-2010, for children ages 5-17, using the IOTF standards). See
http://www.worldobesity.org/site media/library/resource images/Child Global 10th June 2015 WO .pdf
5Indeed, for adults, ideal body weight is smaller in France than in the UK (for both men and women)
and underweight is considered in a more positive light in France than in the UK (for women) (De Saint
Pol, 2009b).
6This result is obtained in a model that controls for the child gender and the parents’ obesity status.
6

study, adolescents from less affluent families are more likely to be overweight or obese
(Currie et al., 2012).
The evolution of the gradient in family SES and body measures across childhood is
a rather unexplored topic, and the literature has only focused on the UK and the US
so far. Howe et al. (2013) use data on a cohort of British children born in 1991-1992
from the Avon Longitudinal Study of Parents and Children. They demonstrate that
maternal education differences in offspring total body fat mass (assessed by whole body
dual x-ray accelerometry scan) is significant and increases with age for females, while it
is insignificant and stable for males. Using cross-sectional data on children ages 2-19,
Murasko (2009) finds that the association between income (measured by the poverty-to-
income ratio) and obesity is stable with age in the US. Murasko (2013) exploits longitudinal
data on US children ages 6 to 14 from the Early Childhood Longitudinal Study, 1998-1999
Kindergarten Cohort (ECLS-K), to examine the gradient in BMI (and height). His results
suggest that the impact of (the log of) income on BMI decreases (linearly) with age.
Of particular relevance is also research by Baum and Ruhm (2009): using data from
the National Longitudinal Survey of Youth (NLSY) on individuals moving through early
adulthood, these authors show that disparities in obesity increase (linearly) with age in
the US, which is consistent with an age-pattern during childhood.
This literature provides evidence on the evolution of the gradient in the UK and the
US, and one can wonder whether the results are universal. In addition, some of these
articles use a model that regresses child outcomes on family SES, child age, an interaction
term between family SES and child age, and controls, which tests whether the association
between family SES and child outcomes monotonically increases or decreases with age
(Baum and Ruhm, 2009; Murasko, 2013).7In this paper, we complement the previous
literature by using data for France and allowing the evolution of the gradient in parental
SES and child outcomes to be non-monotonic.
3 Data and methods
3.1 The ESSM and the ESPS
Our primary data source is the ESSM for 1991-1992 and 2002-2003. In 1991-1992, the sur-
vey was carried out by the “Institut National de la Statistique et des Etudes Economiques”
7The model does not include an interaction term between family SES and higher order polynomials of
child age.
7

(INSEE) and the “Centre de Recherche, d’Etudes et de Documentation en Economie de
la Sant´e” (CREDES), and in 2002-2003, the survey was carried out by the INSEE. The
survey is cross-sectional and it is done every ten years approximately – the 1991-1992 and
2002-2003 surveys are the two most recent waves. The data are collected by face-to-face
interviews and self-completion questionnaires. To our knowledge, the ESSM data have
been used to study social inequalities in body weight in adulthood (De Saint Pol, 2009a;
Etil´e, 2014; Singh-Manoux et al., 2010), but not in childhood yet.
We complement the ESSM with the ESPS data between 1996 and 2010. The exact
survey years are 1996, 1997, 1998, 2000, 2002, 2004, 2006, 2008, and 2010.8The survey
is carried out by the “Institut de Recherche et Documentation en Economie de la Sant´e”
(IRDES) and the “Caisse Nationale de l’Assurance Maladie des Travailleurs Salari´es”
(CNAMTS). The survey is a rotating panel, and some households are re-interviewed ap-
proximately every four years. However the number of children that are actually followed
over time is small, and consequently we pool all the waves and do not take advantage of
the longitudinal nature of the data in our main specification. We exploit the longitudinal
aspect of the ESPS in a robustness check though. The data are collected by a combination
of phone interviews, face-to-face interviews, and self-completion questionnaires.
To our knowledge, these surveys provide the two largest samples containing children of
all ages matched with their parents for France. The samples are representative of French
households. In this paper, we define a child as an individual aged 0-17, who is the child of
the respondent or his partner. The ESSM sample contains 14,080 children, and the ESPS
sample, 24,908 observations.
Note that the ESSM contain less observations than the ESPS. However, the questions
on education are consistent across waves in the ESSM, allowing us to create for each parent
two education variables, whereas they are not in the ESPS, which explains why we are
only able to create for each parent one education variable that suffers from measurement
error (see Section 3.2.2). This is the reason why the ESSM data is our preferred dataset
in this analysis.
8We do not use the 1994 wave of the ESPS because there is no information on the month when the
child height and weight are reported.
8

3.2 Variables
3.2.1 Reported child body weight
Both surveys contain child reported height and weight. Height and weight are reported
by the parents in the ESSM data, and either by the parents or by the child (or by an
unknown household member) in the ESPS data.9
We use information on the child height and weight, gender, and age in months at
the time of the interview, to derive the gender- and age-adjusted BMI-for-age z-scores,
using the “zanthro” Stata function (Vidmar et al., 2013) with the WHO reference growth
chart. The Stata “zanthro” function allows to use the WHO, UK, or US growth reference
chart. We chose the WHO reference because our study is for France and so international
standards seem more appropriate. The WHO growth chart uses the “WHO Child Growth
Standards” and “WHO Reference 2007” composite data files as the reference data. The
“WHO Child Growth Standards” are the result of the Multicentre Growth Reference
Study, which focuses on children ages 0-5 years and contains growth data on infants
and young children from Brazil, Ghana, India, Norway, Oman, and the US. The “WHO
Reference 2007” contains information on children and adolescents ages 5-19.
In addition, we create two dummy variables for whether the child is overweight or
obese, using the cut-off points of the WHO and of the IOTF. To create the overweight or
obesity dummy using the WHO standards, we consider that a child ages 0-5 is overweight
if his z-score is greater than +2, and that a child ages 5-17 is overweight if his z-score
is greater than +1, following the guidelines of the WHO.10 To create the overweight or
obesity dummy using the IOTF standards, we use the “zbmicat” Stata command. The
IOTF standards are derived from BMI data from Brazil, Great Britain, Hong Kong, the
Netherlands, Singapore, and the US. The overweight or obesity status derived from the
IOTF reference is only defined for children older than 2 years.
Reported height and weight capture weight perceptions. They are different from mea-
sured/actual body measures. Adolescents tend to over-report their height and under-
report their weight (Chau et al., 2013; Robinson et al., 2014; Sherry et al., 2007). A
9Before 2008, there are no guidelines containing the age at which children are allowed to report their
height and weight. In 2010, children ages 16 and 17 are asked to report their own height and weight.
Between 1996 and 2010, 14.71% of children report their own height and weight. 4.3% of children less than
10 report their own height and weight (these may simply be mistakes in the answers to the question on
the identity of the respondent). At age 12, 19.5% of children report their own height and weight, at age
14, 29.5%, at age 16, 42.0%, and at age 17, 52.0%.
10See www.who.int/nutgrowthdb/about/introduction/en/index5.html for children under 5 and
http://www.who.int/growthref/who2007 bmi for age/en/ for children ages 5-17.
9

regression-based procedure to correct self-reported data has been developed for French
data, but it only applies to adolescents, while our sample also contains young children,
and it requires information on body shape perception, that is not available in our data
(Legleye et al., 2014). Consequently, the findings below are based on original self-reported
data.
3.2.2 Parental education
Our main SES variable is parental education, although in some analyses we examine the
role of family income. The ESSM data contain information on the parent’s highest degree,
which we recode in three categories: “low education,” which corresponds to having a degree
less than “baccalaur´eat,” “medium education,” which captures a “baccalaur´eat” diploma,
and “high education,” which corresponds to a diploma higher than “baccalaur´eat.” Using
this information, we create two dummies for parents’ education, that we will alternatively
use in our econometric models: (1) a dummy for whether the parent has a medium or
high education, we call this variable “medium or high edu,” and the reference category is
then that he has a low education; and (2) a dummy for whether he has a high education,
we call this variable “high edu,” and the reference category is then that he has a low or
medium education.11
In the ESPS, the questions on education are inconsistent across waves. Between 1996
and 2004, individuals are asked about the highest grade they attended. We know whether
individuals attended university or not, but we do not know their highest grade at university.
In 2006, individuals are asked about their highest grade and also their highest degree.
Finally, in 2008 and 2010, they are asked about their highest degree; and if they do not
have any degree, they are asked about their highest grade. Using this information, it is
impossible to construct a consistent variable for education in three categories like we did in
the ESSM data. Instead, we create a dichotomic variable “high edu”, which equals one if
the parent attended higher education (in 1996-2004) or if the parent received a university
degree (in 2006-2010). Since it is possible to attend higher education without getting any
degree, this variable is not perfectly consistent across waves.
11Another way to go would be to estimate a model that would regress a child health outcome on a dummy
for the parent’s high education and a dummy for the parent’s medium education (and controls) and use the
parent’s low education as the reference category. However, because we want to include interaction terms
between the parent’s education and child age and age square, this approach makes the results unnecessarily
difficult to interpret. See the method and results below.
10

3.2.3 Control variables
In our econometric models, we include the following list of basic controls: a dummy for the
child gender, dummies for the presence of the mother and father in the household, the ages
of the mother and the father interacted with their presence, and survey year dummies.
When we use the ESPS data, we also include dummies for the identity of the respondent
to the child height and weight questions. The questions on the child height and weight can
be answered either by the child himself, his parents, or an unknown person. We control
for the identity of the respondent to account for potential differences in reporting styles.
Our list of controls is very similar to the one used in the previous literature on the
gradient in health in childhood (Case et al., 2002; Currie and Stabile, 2003; Khanam et
al., 2009; Reinhold and J¨urgen, 2012).
3.3 Descriptive statistics
The definitions and summary statistics of the variables of interest are given in Table 1.
In the ESSM sample, 17.6% of children are overweight or obese according to the WHO
standards and 15.0% according to the IOTF reference. That the prevalence is higher
according to the WHO definition than to the IOTF definition has already been found in
the previous literature (Gonzalez-Casanova et al., 2013; Rangelova et al., 2014). Note
that mean BMI-for-age z-scores and overweight or obesity prevalence are similar across
surveys, which suggests that reported height and weight are reliable in the sense that
they are stable. The distributions of the education variables in the ESSM and ESPS are
somewhat different, because the education variables are not defined in the same way due
to data limitations.
[Insert Table 1 here]
Figure 1 displays the BMI-for-age z-score and the overweight or obesity status as a
function of age, separately for boys and girls, in the ESSM data. For the z-score and for
overweight or obesity status according to WHO standards, we use the following age groups:
0-2, 1-3, ..., 15-17. For overweight and obesity status according to the IOTF standards, the
age groups go from 2-4 to 15-17, since overweight is not defined for children less than 2 in
this approach. A non-linear pattern is found for all (proportional) body weight measures
for both genders. BMI-for-age and overweight increase or remain stable between birth (or
11

age 2) and age 8, and decrease afterward. Between ages 2 and 8, the increase in BMI-for-
age and overweight status (WHO) seems to be faster for boys than for girls. In contrast,
there is some evidence of a decrease in BMI-for-age and overweight (WHO and IOTF)
after age 8 for boys and girls. Moreover, in our data, the decrease is greater for girls than
for boys. Additional descriptive statistics show a decline in overweight prevalence from
21.3% at ages 8 to 6.3% at age 17, using the two waves of the ESSM.
We also find a decrease in overweight prevalence from 20.2% at ages 7-9 (in 2002-2003)
to 8.7% at ages 14-15 (in 2002-2003), for girls, using one single wave of the ESSM.12 We
observe a steep decline in overweight for girls in the ESPS as well, from 19.4% (in 2002-
2004) and 18.3% (in 2006-2008) at ages 7-9, to 11.1% (in 2002-2004) at ages 14-15.13 The
overweight prevalence at ages 14-15 is greater in the ESPS; given that the sample size is
larger in the ESPS than in the ESSM, the descriptive statistics from the ESPS might be
more reliable here. In any case, since we observe the same pattern in both the ESSM and
ESPS, we believe that the sharp decrease in overweight with age for girls is not statistical
noise.
This steep decline in reported proportional body weight for girls goes hand in hand
with a decrease in measured proportional body weight, but this decrease in smaller: using
measured data, it has been shown that approximately 19.5% of girls (in 2007) are over-
weight or obese at ages 7-9, versus 16.0% (in 2003-2004) at ages 14-15 (DREES, 2011).
Thus for girls ages 7-9, overweight prevalence in the ESSM and ESPS is consistent with the
prevalence from DREES (2011). But for girls ages 14-15, the prevalence is much smaller
in the ESSM and ESPS than in DREES (2011). This gap is due to difference in the type
of data: reported data (in the ESSM and ESPS) capture body weight perceptions, while
measured data (in DREES) reflect actual body weight. Reported data are relevant be-
cause perceptions likely determine weight-related control behaviors – in other words, one
cannot expect an individual who does not perceive he is overweight to be careful with his
weight.14
[Insert Figure 1 here]
12Here we use the survey year that is the nearest to the survey years from DREES (2011), for comparison
purposes.
13We chose the survey years to match those of the ESSM and of DREES (2011), for comparison purposes.
14Our model (1) is estimated separately by age groups, so it perfectly controls for any change in unob-
served variables with are correlated with age. We estimate this model using our reported body weight data
and find that the gradient in education is U-shaped with age. Our descriptive statistics highlight that the
gap between reported and measured body weight is a function of age. If that gap depends only on age,
then our findings imply that the gradient in education and measured body weight is U-shaped.
12

Figure 2 shows age-related changes in BMI-for-age and overweight, separately by
parental educational level, in the ESSM data. Both maternal and paternal education
levels are negatively associated with child body weight: children whose parents have a
high education have a lower BMI-for-age and are less likely to be overweight than those
whose parents have a low or medium education.
Overall, between ages 2 and 8, proportional body weight generally increases or re-
mains stable for both children whose parents have a low or medium educational level and
children whose parents have a high educational level. After age 8, proportional body
weight decreases. We find some evidence that before age 8, the speed of the increase in
(proportional) body weight is greater for children whose parents have a low and medium
education. Moreover, after age 8, the decrease in body weight is sharp for children whose
parents have a low or medium level of education, whereas it is at most slow for children
whose parents have a high level of education. As a result, the gap between the two curves
first increases and then decreases with age, which suggests that there is process of diver-
gence in body weight among children before age 8, followed by a process of convergence
afterward.
In the bottom left subfigure, the probability of overweight (IOTF) for children whose
mother has a low or medium level of education follows the same pattern: it first increases
and then decreases with age. But the trajectory of overweight for children whose mother
has a high level of education is different, since the probability of overweight continuously
decreases with age. In spite of this difference, the gap between children also strengthens
before age 8 and weakens afterward in this subfigure.
[Insert Figure 2 here]
3.4 Methods
To describe the evolution of the gradient across childhood and adolescence in a precise
manner, we use the following model, that we estimate separately for different age groups:
BC=α+SE SPβ+DAC+XC,P χ+v(1)
where BCdenotes a child body weight outcome, SE SPthe SES variable (i.e. parental
education or family income), DACa series of dummy variables for the child age, and XC,P
child- and parents-level control variables. The coefficient of interest is βwhich indicates
13

whether the gradient in SES and child body weight is positive and significant. The model
is estimated separately for children ages 0-2, 1-3, 2-4, ... and 15-17. Note that these age
groups overlap, in order to smooth our estimates of the gradient.
Like Figure 2 in the descriptive statistics, this multivariate model will highlight that the
association between SES and child body weight outcomes has a U-shape across childhood.
To describe this non-monotonic evolution of the gradient, we use the following model for
the total sample of children:
BC=α+SE SPβ+ACδ1+A2
Cδ2+ [SE SP×AC]φ1+ [SESP×A2
C]φ2+XC,P χ+v(2)
where ACand A2
Cdenote the child age and age squared; SESP×ACis an interaction
term between SES and child age; and SESP×A2
Cdenotes the interaction between SES
and child age square. If φ1is negative and significant and φ2is positive and significant,
then the inverted U-shape of the gradient is statistically significant.
When estimating equation (2), we generally include the control variables, without
interacting them with child age and age square. We thus assume that the impact of the
controls is stable with age. In some analyses, we relax this assumption and allow the
impact of the controls to change with age. Specifically, we estimate a fully interacted
model in which the controls are all interacted with age and age square.
Equation (2) is flexible enough to allow the relationship between body weight and age
to be different for low- and high-SES children. Indeed, equation (2) can be rewritten as:
BC=α+SE SP[β+ACφ1+A2
Cφ2] + [ACδ1+A2
Cδ2] + XC,P χ+v(3)
For low-SES children (for instance, parental education is low, i.e. SE SP= 0), the
association between body weight and age is captured by BSESP=0
C=α+ [ACδ1+A2
Cδ2] +
XC,P χ+v. In contrast, for high-SES children (SESP= 1), the association between body
weight and age is different and is captured by BSE SP=1
C=α+ [β+AC(φ1+δ1) + A2
C(φ2+
δ2)] + XC,P χ+v.
Note that the model does not assume that both children for whom SESP= 0 and for
whom SE SP= 1 have a peak in body weight during childhood. It does not assume either
that when there is a peak for both children for whom SESP= 0 and for whom SESP= 1,
this peak should occur at the same age. Indeed, for children for whom SESP= 0, the
14

peak will be at age −δ1
δ2,15 and there will be no peak if δ1= 0. For children for whom
SE SP= 1, the peak will be at age −φ1+δ1
φ2+δ2,16 and there will be no peak if if φ1+δ1= 0.
The models are estimated using OLS. When the child body weight outcome is a dummy
variable, the model is then a linear probability model (LPM). Although in this case we
could have used a logit or a probit model, we prefer to use a LPM because standard sta-
tistical software do not estimate the real magnitude of the interaction effects in nonlinear
models, especially when they are several interaction terms like in our models (Ai and Nor-
ton, 2003) and coefficients are easier to interpret in LPMs . We calculate robust standard
errors.
4 Results
4.1 Main results
To accurately depict the evolution of the gradient in body weight and SES across child-
hood, we start by estimating equation (1) separately for children ages 0-2, 1-3, 2-4, ...
15-17, using the ESSM data. We alternatively use child BMI-for-age and overweight or
obesity status as our dependent variables. Our main explanatory variables are either the
dummy for whether the mother has a high educational level or the dummy for whether
the father has a high educational level. In Figure 3, we graph the βs coefficients on the
parents’ education as a function of child age group.
For the top left subfigure, the sample contains children who live in the same house-
hold as their mother (i.e. children who live with their mother and father, or only with
their mother), the dependent variable is the child BMI-for-age, and the main explana-
tory variable is the mother’s high education. The dots represent the βcoefficients on the
mother’s high education, whereas the bars capture the confidence intervals. For any age
group, the correlation between the mother’s education and the child BMI is negative and
significant. In addition, the association follows a U-shape with child age: the correlation
between the mother’s education and child BMI first strengthens (in absolute values) be-
tween birth and age 5 approximately, and then weakens from age 5 to 17. This implies
that social differences in (reported) body weight first increase in early childhood and then
decrease at the end of childhood and during adolescence. The left subfigure on the second
15The algebra is: ∂BS ESP=0
C
∂AC= 0 ⇔AC=−δ1
δ2.
16The algebra is: ∂BS ESP=1
C
∂AC= 0 ⇔AC=−φ1+δ1
φ2+δ2.
15

row presents the gradient in the mother’s high education and child overweight or obesity
status using the WHO reference. Again, the association is negative and significant for
all age groups, which implies that there are social inequalities during all childhood years.
The gradient follows a (an inverted) U-pattern, meaning that social inequalities increase
and then decrease over childhood. These results are consistent with those found when the
dependent variable is child BMI. Finally, the bottom left subfigure shows that at any age
the mother’s education is negatively and significantly associated with the probability that
the child is overweight or obese according to the IOTF standards. Consistent with the
previous findings, the association follows a U-shape.
Previous research on child inequalities in childhood has mainly focused on the impact
of the mother’s education on child health, leaving the father’s education aside. One of our
contributions here is to explore the effect of the father’s education. The right hand side
subfigures in Figure 3 show our results. The findings are very similar to those obtained
for maternal education. A significant association between the father’s education and the
child BMI and overweight or obesity is found for most age groups. Moreover, the gradient
follows an inverted U-pattern for our three outcomes.
[Insert Figure 3 here]
The results presented so far are based on separate estimations of equation (1) by child
age groups, and suggest that the gradient reinforces and then weakens with age. To go one
step further, we specify this functional relationship between parents’ education and child
outcomes, in equation (2). This model is estimated using a sample of children of all ages
from the ESSM data and the findings are presented in Table 2. In the four panels, we use
alternative parents’ education variables. In columns (1) and (2), the dependent variable
is the child BMI-for-age; in columns (3)-(4), the overweight or obesity status according to
the WHO definition; and in columns (5)-(6), the overweight or obesity status using the
IOTF reference. Columns (2), (4), and (6) present the estimates of equation (2).
In panels A and B, we focus on the impact of the mother’s education, and the sample
contains children who live in the same household as their mother. Panel A, column (1),
presents the slope of the gradient in the mother’s high/medium education and child BMI-
for-age. The correlation is negative and significant. The model in column (2) adds the
interaction terms between the mother’s medium or high education and child age and age
square. The coefficient on the mother’s education now gives the relationship between
16

education and BMI for children aged 0. This coefficient is not significant, which means
that the slope of the gradient is not significant for children aged 0. The interaction term
between education and age is negative and significant, whereas that between education and
age square is positive and significant. This implies that the gradient has an inverted U-
shape, and that social inequalities in reported BMI rise and then decline across childhood.
The gradient is the largest around age 9.17 Columns (3) and (5) suggest that the mother’s
medium or high education is correlated with a 6.4-6.8 percentage points significant decrease
in the probability that the child is overweight or obese. Consistent with the results in
column (2), the interaction terms in columns (4) and (6) reveal that social inequalities in
reported overweight first widen and then narrow during childhood.
In panel B, the main explanatory variable is a dummy for whether the mother has high
education, rather than medium or high education. The results are qualitatively similar
to those in panel A. The association between the mother’s education and child BMI and
overweight is negative and significant. The interaction terms in columns (2), (4), and (6)
indicate that the association follows a U-shape over childhood. For instance, maternal
high education is correlated with a 5.30 percentage points decrease in child overweight
or obesity (IOTF) at age 2, but a 9.62 percentage points decrease at age 8, and a 1.25
percentage point decrease at age 17.18 The gradients in child BMI and overweight reach
their peak around the age of 8.
Panels C and D show that the father’s education is also negatively and significantly
correlated with the child BMI and overweight. The size of the gradient in the father’s
education is either similar or slightly smaller than the gradient in the mother’s education.
Like for maternal education, the gradient in paternal education generally follows a signif-
icant inverted U-shape with child age. For instance, paternal high education is correlated
with a 5.87 percentage points decrease in child overweight or obesity (IOTF) at age 2, but
a 9.11 percentage points decrease at age 8, and a 4.52 percentage points decrease at age
17.19
[Insert Table 2 here]
17We use estimates with four decimals, rather than just three like in the table. The second derivative
with respect to education and to age equals zero around age 9. The algebra is the following: −0.0571 +
2×0.0033 ×Age = 0 for Age = 8.65.
18Using the estimates with four decimals, the size of the gradient in maternal high education and child
overweight or obesity at each age equals −0.0210 −0.0182 ×Age + 0.0011 ×Age2.
19The size of the gradient in paternal high education and child overweight or obesity at each age equals
−0.0367 −0.0124 ×Age + 0.0007 ×Age2.
17

4.2 Robustness analysis
4.2.1 Additional control variables and the income gradient
We here investigate whether our result on the trajectory of the education gradient is robust
to the inclusion of additional controls – family income and interaction terms between
income and age polynomials in particular. By doing so, we also examine the trajectory of
the income gradient in childhood.
First, we include three additional explanatory variables in our model, namely family
income, maternal labor market status, and parental body weight. Indeed, parents’ edu-
cation may be a proxy for other SES variables, such as family income or parents’ labor
market status. To isolate the effect of parents’ education, we thus need to re-estimate our
model controlling for these variables. Moreover, parents’ body weight may be an omitted
variable in our model if it has an impact on both parents’ education and child body weight.
In particular, the recent literature provides evidence on the intergenerational transmission
of body weight (Classen, 2010; Classen and Hokayem, 2005; Monheit et al., 2009, for the
US; Costa-Font and Gil, 2012, for Spain).
In the ESSM, income is defined as monthly pre-tax family income. In the 1991-1992
wave of the data, income is given in ten brackets, and we thus compute the middle of the
brackets. Consequently there is some measurement error in the income variable. In the
2002-2003, the exact income level is available. Using the consumer price index, we adjust
income so that the variable is comparable across waves. In our specification, we take the
logarithm of income to account for the non-linearity in the relationship between income
and body weight.
We measure maternal labor market status using a dummy for whether the mother
participates actively in the labor market (i.e. she is employed or unemployed). Parental
body weight indicators are parents’ BMI and dummy variables for whether the parents
are overweight or obese.
Our findings are presented in Table 3. The models control for child age and age
square, the basic controls, and income. In addition, regressions in panels A control for
the mother’s education, the mother’s labor market status, and the mother’s body weight,
whereas regressions in panel B control for the father’s education and the father’s body
weight. In most specifications, the coefficients on the interaction terms remain highly
significant, which implies that our main results on the education gradient are robust to
18

the inclusion of additional controls. Moreover, the coefficients on income are negative and
generally significant, suggesting that income plays a protective role against overweight.
For instance, in panel A, column (2), when the logarithm of income increases by 1 (i.e.
when income is multiplied by 2.7), the probability that the child is overweight (WHO)
decreases by 2.9 percentage points.20
[Insert Table 3 here]
We now test whether our results on the trajectory of the education gradient are affected
by the inclusion of interaction terms between income and age polynomials, and investigate
the shape of the gradient in income at the same time. We start by estimating equation
(1) using the logarithm of income as our SES variable, separately by age groups. Here we
do not control for parents’ education. The results, which are presented in Figure 4, show
that for most age group family income is negatively and significantly associated with body
weight. Moreover, the gradient in income has an inverted U-shape across childhood.
[Insert Figure 4 here]
We then estimate equation (2), still using income as our SES proxy, for the sample
of children of all ages. We first include the basic controls, not interacted with age and
age square. The results (available upon request) show that the interaction terms between
income and age and age square are not significant, whether we control for parents’ edu-
cation or not. However, this model assumes that the effect of the control variables does
not depend on child age. Given that interaction terms between the control variables and
age could play the role of third common hidden factors in the equation, omitting these
variables from our model may bias our estimates on interaction terms between income
and age. We thus estimate a fully interacted model (that relies on weaker assumptions),
in which all the control variables are interacted with age and age square. The results are
shown in Table 4.
In panel A, we do not control for parental education. In both columns (1) and (2),
the gradient in income has a significant inverted U-shape. Column (1) implies that the
correlation between income and BMI-for-age is the strongest in absolute values at age 9.
20The results also show that maternal labor market status is not significantly associated with child body
weight and that parents’ body weight is positively and significantly correlated with child body weight. We
also estimate models that control for parents’ general health, using the 2004-2010 ESPS data. The results
are consistent with an inverted U-shape gradient in education.
19

In column (3), the signs of the coefficients are consistent with the U-shape trajectory, but
one of the interaction terms is not significant.21
In panel B, we estimate the fully interacted model using the mother’s high education
as our SES variable. We show the coefficients on the interaction terms, but not that on
the mother’s education, for space reasons. The shape of the education gradient clearly
emerges. Moreover, the coefficients on the interaction terms have the same magnitude
as in the model in Table 2, panel B (in which the controls are not interacted). Thus for
education, the results of the specification in which we do not interact the controls are
similar to those from the fully interacted model – for this reason, in Tables 5-9 in which
we will only focus on education, we won’t interact the control variables.
In Table 4, panel C, family SES is measured using both education and income. In
columns (1) and (3), the interaction terms for income are no longer significant. In con-
trast, in column (2), the evolution of the income gradient with age remains significant.
Interestingly, this panel highlights that the trajectory of the education gradient is robust
to the inclusion of interaction terms for income.
Finally, in panels D and E, we re-do our analysis using the ESPS data. Panel D
corroborates the findings on the trajectory of the income gradient from panel A. In panel
E, we add the mother’s high education. Column (1) still shows an (inverted) U-shape for
the income gradient. In column (2), the results on income are less clear. In column (3), the
interaction terms on income switch signs. This surprising result is due to multicollinearity
(i.e. the high level of correlation between income and education in this specific model, in
other words we “overcontrol” for family SES here), and does not mean that inequalities
in income and overweight (IOTF) decrease and then increase in childhood.
Table 4 thus provides some evidence that the gradient in income has the same shape as
the gradient in education. However, the results on the education gradient are clearer and
more robust, maybe because there is less measurement error in education than in income.
For this reason (and for space reasons), the rest of the paper only investigates the role of
education.
[Insert Table 4 here]
21We find that the difference in our results (between the model in which the controls are not interacted
and the fully interacted model) is mainly due to the inclusion of the interaction terms between survey year
dummies and age and age square. Indeed, there is a correlation between year dummies and income on the
one hand, and year dummies and body weight on the other hand. The correlation between year dummies
and income may be due to the change in the income variable over time (in 1991-1992, income is given in
brackets, whereas in 2002-2003, we have information on the exact family income level).
20

4.2.2 Cohort fixed effects
The variation in the education gradient with age that we show above will not quantify the
evolution of social inequalities in health over the life cycle in the presence of cohort effects.
A natural strategy is then to include cohort fixed effects in the model. We assume that
all children born in a 5-year interval belong to the same cohort. If we were to control for
cohort fixed characteristics in our model when we use the ESSM data, they would be very
highly correlated with child age and with the survey year fixed effects, because the ESSM
survey is only carried out every ten years (1991-1992 and 2002-2003). Consequently, we
would face a multicollinearity problem.
In contrast, it is possible to include cohort fixed effects in our model when we use the
ESPS data, because this survey is done every one or two years. We create a series of cohort
dummies for children born before 1984, between 1985 and 1989, between 1990 and 1994,
... and between 2005 and 2010, and re-run our model including these cohort dummies.
The standard list of controls is also included. The results, which are reported in Table 5,
show that the gradients in the mother’s education and child overweight or obesity, in the
father’s education and child BMI, and in the father’s education and child overweight or
obesity still take the form of an inverted U-shape.
[Insert Table 5 here]
To explore more in depth the role of cohorts’ characteristics, we compare the effect
of parental education on child body weight when we control for cohort fixed effects and
when we do not. We focus on the impact of the mother’s high education and child BMI-
for-age, for space reasons. We represent the effect as a function of age in Figure 5.22 The
figure underlines that accounting for cohort fixed effects only has a small influence on the
trajectory.
[Insert Figure 5 here]
22To get the curve for the model “not accounting for cohort fixed effects,” we proceed as follows: we
estimate equation (2) not including cohort fixed effects; using the estimated coefficients ˆ
β,ˆ
φ1, and ˆ
φ2, we
compute the effect EF of the mother’s education, which is the difference in body weight between children
whose mother has a high level of education (SE SP= 1) and children whose mother has a low level of
education (SE SP= 0), which writes EF =ˆ
BSESP=1
C−ˆ
BSESP=0
C=ˆ
β+ACˆ
φ1+A2
Cˆ
φ2; finally we graph
the curve representing the effect EF as a function of age. To get the curve for the model “accounting
for cohort fixed effects,” we follow the same steps but include the cohort fixed effects when we estimate
equation (2). The curves in Figure 5 are below zero because of the negative correlation between parents’
education and child BMI-for-age.
21

4.2.3 Child fixed effects
Our main results could also be biased because we do not account for child fixed char-
acteristics. To test the robustness of our findings, we re-estimate our model using the
ESPS data and including child fixed effects (which capture race or ethnicity for instance).
In that setting, the interaction terms between education and child age are still identified
because child age varies over time. The number of observations that we use is smaller than
in our main specification, because only a fraction of children are followed over time. In
particular, the observations from the 2010 wave are dropped due to the fact that the ESPS
sample was completely renewed in 2010. The results, which are presented in Table 6, are
consistent with the (inverted) U-shape of the gradient.
[Insert Table 6 here]
4.2.4 Other robustness tests
Our BMI z-scores are derived from the WHO Growth Charts. However, our results on
the (inverted) U-shape of the gradient seem to be robust to the use of alternate growth
references. Indeed, when we use z-scores derived from the British 1990 Growth Charts or
the 2000 CDC Growth Charts, the gradient either slightly strengthens or remains stable
from birth to age 8, and clearly weakens after age 8.
So far, our models have accounted for either the mother’s education or the father’s
education. We also estimate models in which we control for both maternal and paternal
education and all interaction terms. The results are consistent with the (inverted) U-
shape of the gradient. However, these results are not as clear as when we control for
either maternal or paternal education, since the interaction terms are smaller and some
of them lose their significance. The reason is that maternal and paternal education are
correlated, due to assortative mating in particular. It seems to us that maternal and
paternal education are too correlated to be included in the same model.
Note also that when we estimate the evolution of inequalities using our equation (2)
but controlling for child age dummies instead of child age and age squared, the results are
almost unaffected.
Moreover, there is no evidence that the U-shape is due to a change from parents’
reported height and weight to child reported height and weight. Indeed, when in the
ESPS data we only keep children whose body measures are reported by their parents, the
22

U-shape still clearly appears.
Finally, we also re-estimate our models dropping children ages 0-2. Indeed, evaluating
overweight/obesity at ages 0-2 is touchy. Clinicians disagree on whether there can be
an overweight problem in infancy. In addition, overweight in infancy is poorly defined
– the literature and clinical guidelines use different criteria to define overweight in early
childhood (Heinzer, 2005). The results, which are presented in Table 7, are consistent
with our previous findings: the interaction terms always have the expected sign, the two
interaction terms are significant in all the models for the mother’s education (panel A,
columns (2), (4) and (6)) and in one model for the father’s education (panel B, column
(4)).
We now compare Table 7 with Table 2, for which very young children were included in
the sample. For the mother’s high education, the significance level is 1% or 5% in Table 7,
whereas it was always 1% in Table 2, and the magnitude of the interaction terms is either
similar or slightly larger in absolute value. For the father’s high education, the interaction
terms are no longer significant in column (2), there is only one significant interaction term
in column (6) – the loss of significance could be due to the decrease in the sample size –
and the coefficients are slightly greater in absolute value in column (4).
[Insert Table 7 here]
4.3 Additional results
4.3.1 Analysis by gender
We now relax the assumption that there is the same evolution of the gradient with age
for boys and girls. This analysis is motivated by previous studies for the US highlighting
that the evolution of the gradient with age is different between genders (Wang and Zhang,
2006). To this end, we re-estimate our model separately for boys and girls and show
our estimates in Table 8. The results are in line with our previous results. Indeed, the
signs on the interaction terms are consistent with an inverted U-shaped gradient in all
models. In addition, the coefficients are significant in a number of models. Interestingly,
the relationship between parents’ education and child body weight is not sex-specific,
since we find it for mothers-daughters and fathers-sons, but also for mothers-sons and
fathers-daughters.
[Insert Table 8 here]
23

4.3.2 Analysis by survey year
We also re-estimate our model separately for 1991-1992 and 2002-2003, in the ESSM, and
for 1996-2004 and 2006-1010, in the ESPS, to see whether the shape has emerged over
time. Note that when we use the ESPS, we include several survey years to avoid small
samples. The results are presented in Table 9. In all panels, the signs on the interaction
terms are consistent with a U-shape. The U-shape can be observed as of 1991-1992, and
in each subsequent period.
[Insert Table 9 here]
5 Discussion
Using two datasets containing approximately 40,000 children observations, we investigate
the relationship between parental education and child reported body weight across child-
hood in France. Reported data capture body weight perceptions, which likely have an
influence on weight-related behaviors on their own. Our findings indicate that children
whose parents have a high level of education have a lower BMI-for-age and are less likely
to be overweight (Table 2). Our results for French children echo findings for French adults,
which are also based on reported height and weight data. Specifically, De Saint Pol (2009a)
highlights significant differences in BMI depending on occupation, standard of living, and
educational level, for both females and males, in 2003. Moreover, Etil´e (2014) uses the
recentered influence function technique to quantify the role of education in BMI and finds
that educational expansion decreased BMI inequalities between 1981 and 2003. Finally,
Singh-Manoux et al. (2010) highlight that inequalities in socioeconomic status (measured
by education or income) and height have remained significant and stable between 1970
and 2003.
The evolution of reported (proportional) body weight across childhood is non-monotonic
in our data. BMI-for-age and overweight status generally increase between ages 2 and 8
and decrease afterward (Figure 1). The increase in early childhood is consistent with the
adiposity rebound, which refers to the increase in BMI that starts (after a nadir) between
ages 3 and 8 in most countries. Note however that the adiposity rebound is assessed using
the unadjusted BMI (computed using measured height and weight) and body composi-
tion measures in the medical literature, and not BMI-for-age z-scores or overweight status
(computed using reported data). In our data, additional results show that the rebound
24

in unadjusted BMI happens between ages 4 and 6. Moreover, we show a sharp decrease
in overweight after age 8 for girls. This pattern is found in our two datasets, which sug-
gest that it is not due to statistical noise. In general, the whole trajectory of reported
body weight across childhood in our analysis is in line with international findings: for
instance, Costa-Font and Gil (2012) also use reported height and weight and find a peak
in overweight in childhood in Spain.
In the absence of any control variable, our data already suggest that the gradient
follows an inverted U-shape with age. Specifically, when we distinguish the trajectory of
body weight according to parental education, we observe that between ages 2 and 8 the
speed of the increase in body weight is greater for children whose parents have a low and
medium education than for children whose parents have a high educational level, and that
later on, the speed of the decrease in body weight is also larger for children whose parents
have a low or medium educational level (Figure 2). These differences in the trajectory
of body weight with age between children imply that the correlation between family SES
and child body weight strengthens in early childhood and then weakens for older children.
The econometric specification, that includes control variables, support this finding.
The gradient in parents’ education and child body weight has an inverted U-shape across
childhood, which implies that the gradient first widens and then narrows as children get
older. They reach their peak at the age of 8. This result holds when we estimate our
models separately by age group and when we use the full sample of children of all ages.
This results is also found when we account for maternal labor market status, family income,
parental body weight, cohort fixed characteristics, or child fixed traits. Consistent with
education, the income gradient is also U-shaped across childhood, although our results
may be less robust.
Our findings on the trajectories of the education and income gradients in child body
weight in France are somewhat different from related results from the literature. Indeed,
Murasko (2013) finds a strengthening of the gradient in family income and child (measured)
BMI in the US. Howe et al. (2013) show that the gradient in maternal education and
offspring fat mass increases with age for girls, and remains stable for boys, in the UK.
Moreover, the gradient in family income and child health becomes more pronounced as
children get older in France, which implies that the health of children from families with low
income erodes faster with age (Apouey and Geoffard, 2014). In addition, recent research
shows a stability of the gradient in general health in the UK (Apouey and Geoffard, 2013;
25

Howe et al., 2013), although this is still a debated conclusion (Currie et al., 2007; Case et
al., 2008; Propper et al., 2007). Taken together, these findings highlight that the trajectory
of the gradient in childhood depends on the country and on the child outcome of interest
(reported or measured anthropometric outcomes, general health, etc).
A possible explanation for the inverted U-shape of the gradient emphasizes the role of
socialization and schooling in youth, echoing the theory of West (1988, 1997). This author
focuses on the UK and argues that social health inequalities are strong in childhood,
that they decrease or virtually disappear in youth (starting age 12 in his approach), and
re-emerge in late youth. The empirical results of West have been recently discussed, as
several authors did not find any weakening of health inequalities during adolescence in the
UK (Apouey and Geoffard, 2013; Case et al., 2008). However, West’s intuition on the role
of socialization remains relevant in our context to explain the decrease in the education
gradient after age 8 (note that in West’s approach, inequalities start decreasing later,
around age 12). For West, one of the causes of the decrease in youth is that adolescents
progressively detach themselves from their parents and their parental home, and spend
more time at school and with their peers (West, 1988, 1997). In the context of our
study, socialization at school may have an influence on weight-related behaviors and body
weight perceptions. It is possible that starting age 8 this socialization effect becomes
more prominent while the effect of parents’ education starts decreasing. Consistent with
this intuition, another strand of literature has shown significant peer effects in primary
and secondary schools, related to body weight, body perception, physical activity, and
eating habits in Canada and the US (Ali et al., 2011; Fortin et al., 2011; Halliday et al.,
2009; Renna et al., 2008; Trogdon et al., 2008). Finally, (mandatory) physical education
courses in primary school may also lead to a weakening of the association between parents’
education and body weight.
The inverted U-shape in the gradient in reported body weight may reflect an inverted
U-shape in the gradient in measured body weight. In that case, social inequalities in body
weight are neither cumulative nor irreversible, since inequalities would start decreasing
around age 8. In this approach, a potential explanation of the shape of the gradient
has to do with “selective overweight.” Indeed, reported body weight follows an inverted
U-shape across childhood for both boys and girls in France (see Figure 1). In parallel,
there is some evidence that measured body weight also follows this shape, although the
26

age of the peak is uncertain and might be around 10-11.23 In age groups in which few
children are overweight, before age 4 and after age 14, there may be almost no selection
at play: only the most prone to overweight (because of genetic factors, not necessarily
related to parental characteristics) will be overweight. In that case, parental education
will only play a minor role and the gradient will be small. In contrast, in age groups
in which overweight is more prevalent (because child weight is particularly affected by
socioeconomic and environmental conditions), parental education will play an important
role. Hence, in age groups where overweight is more widespread, a selection process may
be at play, which may explain why a stronger gradient is observed. Overall, if there is
such a selection process, the equalization of body weight after age 8 will be due to the
“natural” decline in BMI and to the decline in selectivity in overweight with respect to
parental education.24
Given that reported anthropometric data differ from measured data (there is a so-
called “bias” or “reporting style” in reported data), the shape of the education gradient
across childhood may also partly capture a change in the association between parents’
education and the reporting style across childhood. More precisely, parents may become
worried about social norms regarding body weight around age 8, soon after the child enters
primary school. If children whose parents have a low level of education under-report their
true body weight starting age 8, then we will observe that the gradient has an inverted
U-shape although in reality the gradient increases with age (assuming that children whose
parents have a high level of education do not over-report their body weight, which is an
acceptable assumption in our opinion).25 In that case, the change in reporting style with
education and age will bias the measurement of the evolution of health disparities. Of
particular relevance would thus be research on the evolution of the bias with family SES
and age (taken together). As far as we are aware, this topic has not been explored so far.
Indeed, the literature that assesses the validity of self-reported body weight in the
23We did not find data on the evolution of the BMI-for-age z-score with age in France. A report based
on measured data shows that for girls overweight prevalence remains constant or slightly increases between
ages 7-9 and 10-11, and then clearly decreases between ages 10-11 and 14-15. The peak is then at ages
10-11. For boys, overweight prevalence increases between ages 7-9 and 10-11, and then decreases between
ages 10-11 and 14-15. The peak is also at ages 10-11 (DREES, 2011). However, these measured data on
overweight by age are not precise, since they are only calculated for some specific age groups (ages 5-6,
7-9, 10-11 and 14-15), and not at each age. Consequently, there is some uncertainty about the exact age
of the peak, which might in reality be before age 10.
24We are thankful to a referee for suggesting this explanation.
25Computing the extent of the bias that would change the U-shaped trajectory into an increasing tra-
jectory is difficult. It would be possible only if information on the coefficients of the model that regresses
measured body weight on child age, parents’ education, their interaction terms, and the control variables
was available.
27

early years studies the correlation between a number of factors and the reporting bias,
but does not take into account the evolution of this correlation with age. It suggests
that low-SES adolescents are more likely to under-report their body weight. For instance,
a study on French adolescents shows that family SES is associated with BMI under-
reporting (Chau et al., 2013): when the father has a manual occupation, under-reporting
of BMI is greater. In contrast, insufficient family income is not significantly associated
with BMI under-reporting. The study does not examine the role of parents’ education on
the reporting bias. Moreover, the literature highlights that BMI under-reporting is related
to overweight or obesity status, body dissatisfaction, and reporting capabilities, among
adolescents (Chau et al., 2013; Elgar et al., 2005; Krematsoulas et al., 2013; Sherry et al.,
2007; Robinson et al., 2014; Want et al., 2002).
We provide evidence that the gap between reported and measured body weight depends
on age (for girls in particular, see Section 3.3). If this gap only depends on age but not
on an interaction effect between family SES and age, then our estimates of the change
of the gradient in reported body weight will capture the actual changes of the gradient
in measured body weight, because our models do take age into account. Nevertheless, if
the gap depends on an interaction effect between age and family SES (in other words if
the association between parents’ education and the bias is not constant with age), then
our estimates using reported data will be different from results derived from measured
data. Therefore, future research may focus on estimating whether the correlation between
family SES and the reporting bias changes across childhood. This will help understand
our findings.
Our results should be interpreted in light of several limitations. First, parents’ educa-
tion is measured using relative broad categories, and there might be finer distinctions of
educational level which are not captured by our variables. In particular, it is possible that
within our educational categories, one more year of education matters for child health.
Moreover, within each category, differences in education quality may also matter (Jones
et al., 2011, for secondary schools in the UK). The quality of education may be related to
the university or school that the parent attended. Our education categories are thus likely
to be heterogeneous. Second, we focus on the role of parental education and family income
on child outcomes, but the role of other family SES indicators could be explored. Third,
this paper quantifies the correlation between family SES and child body weight outcomes,
but does not evaluate the causal impact of SES on outcomes (Apouey and Clark, 2015).
28

6 Conclusions
Recent studies for the UK and the US provide some evidence that the gradient in parents’
SES and child body weight strengthen as children age. Our analysis uses adjusted BMI-
for-age z-score and overweight status measures (derived from reported height and weight
data), allows the evolution of the gradient to be non-monotonic with child age, and finds
that the gradient in child BMI-for-age z-score and overweight follows an inverted U-shape
across childhood in France. This pattern is found in two large datasets, when alternative
parental education variables are used, when potential confounding factors are accounted
for, and when cohort or child fixed effects are included. In addition, it is unlikely that this
pattern is only biological since it is observed for both boys and girls. The shape is not a
new phenomenon, since it is even found in data collected at the beginning of the 1990s.
Further work is needed to understand this pattern; in particular, we need to know more
about the evolution of the association between family SES and measured body weight, as
well as between family SES and the bias in reported height and weight.
29

Conflict of interest
We wish to confirm that there are no known conflicts of interest associated with this
publication and there has been no significant financial support for this work that could
have influenced its outcome.
Acknowledgments
The ESPS data were supplied by IRDES (France), and the ESSM data, by INSEE
(France). The authors take responsibility for the integrity of this work. The authors
would like to thank the Editor John Komlos, six anonymous referees, Fabrice Etil´e, and
Claudia Senik for useful comments or discussions.
30

References
Ai, C., Norton, E. C. (2003). Interaction terms in logit and probit models. Economics
Letters, 80(1), 123-129.
Ali, M. M., Amialchuk, A., Heiland, F. (2011). Weight-related behavior among adoles-
cents: The role of peer effects. PLoS ONE, 6(6): e21179.
Apouey, B., Clark, A. E. (2015). Winning big but feeling no better? The effect of lottery
prizes on physical and mental health. Health Economics, 24(5), 516-538.
Apouey, B., Geoffard, P.-Y. (2013). Family income and child health in the UK. Journal
of Health Economics, 22(4), 715-727.
Apouey, B. H., Geoffard, P.-Y. (2014). Child health and access to health care in France:
Evidence on the role of family income. Revue d’Epid´emiologie et de Sant´e Publique,
62(3), 170-190.
Apouey, B. H., Geoffard, P.-Y. (2015). Le gradient et la transmission interg´en´erationnelle
de la sant´e pendant l’enfance. Economie et Statistique, 475-476, 113-133
Baker, E. A, Schootman, M., Barnidge, E., Kelly, C. (2006). The role of race and poverty
in access to foods that enable individuals to adhere to dietary guidelines. Preventing
Chronic Disease, 3(3), A76.
Baum, C. L., Ruhm, C. J. (2009). Age, socioeconomic status and obesity growth. Journal
of Health Economics, 28(3), 635-648.
Case, A., Lubotsky, D., Paxson, C. (2002). Economic status and health in childhood:
the origins of the gradient. American Economic Review, 92(5), 1308-1334.
Chau, N., Chau, K., Mayet, A., Baumann, M., Legleye, S., Falissard, B. (2013). Self-
reporting and measurement of body mass index in adolescents: refusals and validity,
and the possible role of socioeconomic and health-related factors. BMC Public
Health, 13(815).
Classen, T. J. (2010). Measures of the intergenerational transmission of body mass index
between mothers and their children in the United States, 19812004. Economics and
Human Biology, 8(1), 30-43.
Classen, T. J., Hokayem, C. (2005). Childhood influences on youth obesity. Economics
and Human Biology, 3(2), 165-187.
Costa-Font J., Gil, J. (2012). Intergenerational and socioeconomic gradients of child
obesity. FEDEA Working paper.
Currie, J., Stabile, M. (2003). Socioeconomic status and child health: Why is the rela-
tionship stronger for older children? American Economic Review, 93(5), 1813-1823.
Currie, C., Zanotti, C., Morgan, A. (2012). Social determinants of health and well-being
among young people. Health Behaviour in School-aged Children (HBSC) study:
International report from the 2009/2010 survey. Copenhagen.
De Saint Pol, T. (2009a). Evolution of obesity by social status in France, 1981-2003.
Economics and Human Biology, 7(3), 398-404.
31

De Saint Pol, T. (2009b). Surpoids, normes et jugements en mati`ere de poids : Com-
paraisons europ´eennes. INED, Population & Soci´et´es, 455.
DREES (2011). L’´etat de sant´e de la population en France. Suivi des objectifs annex´es
`a la loi de sant´e publique - Rapport 2011.
Drewnowski, A., Specter, S. E. (2004). Poverty and obesity: the role of energy density
and energy costs. American Journal of Clinical Nutrition, 79(1), 6-16.
Elgar, F. J., Roberts, C., Tudor-Smith, C., Moore, L. (2005). Validity of self-reported
height and weight and predictors of bias in adolescents. Journal of Adolescent
Health, 37, 371-375.
Etil´e F. (2014) Education policies and health inequalities: Evidence from changes in
the distribution of body mass index in France, 1981-2003. Economics and Human
Biology, 13, 46-65.
Fortin, B., Yazbeck, M. (2011). Peer effects, fast food consumption and adolescent weight
gain. CIRANO - Scientific Publications 2011s-20.
Gonzalez-Casanova, I., Sarmiento, O. L., Gazmararian, J. A., Cunningham, S. A., Mar-
torell, R., Pratt, M., Stein, A. D. (2013). Comparing three body mass index classifi-
cation systems to assess overweight and obesity in children and adolescents. Revista
Panamericana de Salud P´ublica, 33(5), 349-355.
Guignon, N., Bad´eyan, G. (2002). La sant´e des enfants de 6 ans `a travers les bilans de
sant´e scolaire. DREES, Etudes et R´esultats, 155, 1-8.
Halliday, T. J., Kwak, S. (2009). Weight gain in adolescents and their peers. Economics
and Human Biology, 7, 181-190.
Heinzer, M. M. (2005). Obesity in infancy: Questions, more questions, and few answers.
Newborn and Infant Nursing Reviews, 5(4): 194-202.
Howe, L. D., Lawlor, D. A., Propper, C. (2013). Trajectories of socioeconomic inequalities
in health, behaviours and academic achievement across childhood and adolescence.
Journal of Epidemiology and Community Health, 67, 358-364.
INED (2010). France 2009 : L’ˆage moyen `a la maternit´e atteint 30 ans. Population &
Soci´et´es, 465.
Jones, A. M., Rice, N., Dias, P. R. (2011). Long-term effects of school quality on health
and lifestyle: Evidence from comprehensive schooling reforms in England. Journal
of Human Capital, 5(3), 342-376.
Kaminski, M., Saurel-Cubizolles, M.-J. (2000). La sant´e p´erinatale. In D. Fassin, H.
Grandjean, M. Kaminski (Eds.), Les in´egalit´es sociales de sant´e (pp. 173-192). La
D´ecouverte, Recherches.
Klein-Platat, C., Wagner, A., Haan, M. C., Arveiler, D., Schlienger, J. L., Simon,
C. (2003). Prevalence and sociodemographic determinants of overweight in young
French adolescents. Diabetes/Metabolism Research and Reviews, 19(2), 153-158.
Kreatsoulas, C., Hassan, A., Subramanian, S. V., Fleegler, E. W. (2014). Accuracy of
self-reported height and weight to determine body mass index among youth. Journal
of Child and Adolescent Behaviour, 2(1), 1000126.
32

Legleye, S., Beck, F., Spilka, S., Chau, N. (2014). Correction of body-mass index using
body-shape perception and socioeconomic status in adolescent self-report surveys.
Plos One, 9(5), e96768.
Monheit, A., C., Vistnes J. P., Rogowski, J. A. (2009). Overweight in adolescents:
Implications for health expenditures. Economics and Human Biology, 7(1), 55-63.
Murasko, J. E. (2009). Socioeconomic status, height, and child obesity. Economics and
Human Biology, 7(3), 376-386.
Murasko, J. E. (2011). Trends in the association between family income, height, and
body mass index in US children and adolescents: 1971-1980 and 1999-2008. Annals
of Human Biology, 38(3), 290-306.
Murasko, J. E. (2013). Associations between household income, height, and BMI in
contemporary US schoolchildren. Economics and Human Biology, 11, 185-196.
Murasko, J. E. (2015). Overweight/obesity and human capital formation from infancy
to adolescence: evidence from two large US cohorts. Journal of Biosocial Science,
47(02), 220-237.
OECD (2014). Obesity and the economics of prevention: Fit not Fat. Key facts - France,
update 2014.
Palermo, T. M., Dowd, J. B. (2012). Childhood obesity and human capital accumulation.
Social Science & Medicine, 75(11), 1989-1998.
Propper, C., Rigg, J., Burgess, S. (2007). Child health: Evidence on the roles of family
income and maternal mental health from a UK birth cohort. Health Economics,
16(11), 1245-1269.
Rangelova, L., Petrova, S., Konstantinova, M., Duleva, V., Dimitrov, P. (2014). Over-
weight and obesity prevalence in Bulgarian schoolchildren: A comparison between
two international standards. International Journal of Biomedical and Advance Re-
search, 5(9), 454-458.
Rasmussen, M., Holstein, B. E., Melkevik, O., Damsgaard, M. T. (2013). Validity of
self-reported height and weight among adolescents: the importance of reporting
capability. BMC Medical Research Methodology, 13(85).
Recours, F., Hebel, P. (2006). Les populations modestes ont-elles une alimentation
d´es´equilibr´ee? CREDOC. Cahier de Recherche 232.
Reilly, J. J., Kelly, J. (2011). Long-term impact of overweight and obesity in childhood
and adolescence on morbidity and premature mortality in adulthood: Systematic
review. International Journal of Obesity, 35(7), 891-898.
Renna, F., Grafova, I. B., Thakur, N. (2008). The effect of friends on adolescent body
weight. Economics and Human Biology, 6(3), 377-387.
Robinson, L. E., Suminski, R. R., Perez, G., Palmer, K. K., Irwin, J. M. (2014). Accuracy
of self-reported height and weight in low-income, rural African American children.
Journal of Child & Adolescent Behavior, 2(3), 1000144.
33

Sargent, J. D., Blanchflower, D. G. (1994). Obesity and stature in adolescence and earn-
ings in young adulthood. Analysis of a British birth cohort. Archives of Pediatrics
and Adolescent Medicine, 148(7), 681-687.
Schwimmer, J. B., Burwinkle, T. M., Varni, J. W. (2003). Health-related quality of
life of severely obese children and adolescents. Journal of the American Medical
Association, 289, 1813-1819.
Sherry, B., Jefferds, M. E., Grummer-Strawn, L. M., (2007). Accuracy of adolescent
self-report of height and weight in assessing overweight status. A literature review.
Archives of Pediatrics & Adolescent Medicine, 161(12), 1154-1161.
Singh-Manoux, A., Gourmelen, J., Ferrie, J., Silventoinen, K., Gu´eguen, A., Stringhini,
S., Nabi, H., Kivimaki, M. (2010). Trends in the association between height and
socioeconomic indicators in France, 1970-2003. Economics and Human Biology,
8(3), 396-404.
Steckel, R. H. (1995). Stature and the standard of living. Journal of Economic Literature,
33(4), 1903-1940.
Trogdon, J. G., Nonnemaker, J. M., Pais, J. M. (2008). Peer effects in adolescent over-
weight. Journal of Health Economics, 27(5), 1388-1399.
Tursz, A. (2000). La sant´e de l’enfant. In D. Fassin, H. Grandjean, M. Kaminskidu
(Eds.), Les in´egalit´es sociales de sant´e (pp. 193-206). La D´ecouverte, Recherches.
Vidmar, S. I, Cole, T. J., Pan, H. (2013). Standardizing anthropometric measures in
children and adolescents with functions egen: Update. Stata Journal, 13(2), 366-
378.
Wang, Z., Patterson, C. M., Hills, A. P. (2002). A comparison of self-reported and
measured height, weight and BMI in Australian adolescents. Australian and New
Zealand Journal of Public Health, 26(5), 473-478.
Wang, Y., Zhang, Q. (2006). Are American children and adolescents of low socioeconomic
status at increased risk of obesity? Changes in the association between overweight
and family income between 1971 and 2002. American Journal of Clinical Nutrition,
84(4), 707-716.
West, P., Sweeting, H. (2004). Evidence on equalisation in health in youth from the West
of Scotland. Social Science & Medicine, 59(1), 13-27.
34

Tables
Table 1: Descriptive statistics (ESSM and ESPS data)
ESSM data ESPS data
1991-92/2002-03 1996-2010
Means Means
(Standard deviations) (Standard deviations)
Dependent variables
Child BMI-for-age z-score 0.075 (1.312) 0.079 (1.285)
Child overweight or obesity (WHO) 0.176 (0.381) 0.172 (0.377)
Child overweight or obesity (IOTF) 0.150 (0.357) 0.147 (0.354)
Main explanatory variables
Mother’s medium or high edu 0.392 (0.488) -
Mother’s high edu 0.244 (0.429) 0.335 (0.472)
Father’s medium or high edu 0.357 (0.479) -
Father’s high edu 0.239 (0.426) 0.286 (0.452)
Notes. BMI means body mass index; ESSM, “Enquˆete sur la Sant´e et les Soins M´edicaux,” and ESPS,
“Enquˆete sur la Sant´e et la Protection Sociale.” In the ESPS, the education variables are measured with
error.
35

Table 2: Parents’ education and child body weight (ESSM data)
(1) (2) (3) (4) (5) (6)
BMI-for-age Overweight or Overweight or
z-score obesity (WHO) obesity (IOTF)
Panel A. ESSM, 1991-92/2002-03. Effect of the mother’s medium or high education
Mother’s medium -0.175*** -0.021 -0.068*** -0.018 -0.064*** -0.002
or high edu (0.023) (0.068) (0.007) (0.014) (0.007) (0.027)
Mother’s medium -0.057*** -0.018*** -0.019***
or high edu ×Age (0.017) (0.004) (0.006)
Mother’s medium 0.003*** 0.001*** 0.001***
or high edu ×Age2(0.001) (0.000) (0.000)
Observations 14,080 14,080 14,080 14,080 12,744 12,744
Panel B. ESSM, 1991-92/2002-03. Effect of the mother’s high education
Mother’s high edu -0.203*** -0.066 -0.076*** -0.035** -0.072*** -0.021
(0.025) (0.069) (0.007) (0.014) (0.007) (0.028)
Mother’s high edu ×Age -0.055*** -0.017*** -0.018***
(0.018) (0.005) (0.007)
Mother’s high edu ×Age20.003*** 0.001*** 0.001***
(0.001) (0.000) (0.000)
Observations 14,080 14,080 14,080 14,080 12,744 12,744
Panel C. ESSM, 1991-92/2002-03. Effect of the father’s medium or high education
Father’s medium -0.153*** -0.037 -0.058*** -0.014 -0.064*** -0.039
or high edu (0.024) (0.069) (0.007) (0.014) (0.007) (0.028)
Father’s medium -0.041** -0.012*** -0.008
or high edu ×Age (0.018) (0.005) (0.007)
Father’s medium 0.002** 0.001** 0.000
or high edu ×Age2(0.001) (0.000) (0.000)
Observations 12,772 12,772 12,772 12,772 11,512 11,512
Panel D. ESSM, 1991-92/2002-03. Effect of the father’s high education
Father’s high edu -0.192*** -0.024 -0.059*** -0.000 -0.071*** -0.036
(0.025) (0.073) (0.007) (0.015) (0.007) (0.029)
Father’s high edu ×Age -0.059*** -0.018*** -0.012*
(0.018) (0.005) (0.007)
Father’s high edu ×Age20.003*** 0.000*** 0.000**
(0.001) (0.000) (0.000)
Observations 12,772 12,772 12,772 12,772 11,512 11,512
Notes. In panels A and B, the sample contains children who live with their mother (i.e. with their mother
and father, or with their mother only), and in panels C and D, it contains children who live with their
father (i.e. with their mother and father, or with their father only). In columns (1), (3), and (5), the
control variables are child age dummies and the basic controls. In columns (2), (4), and (6), the control
variables are child age, child age square, and the basic controls. Robust standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1.
36

Table 3: Estimates including additional controls (ESSM data)
(1) (2) (3)
BMI-for-age Overweight or Overweight or
z-score obesity (WHO) obesity (IOTF)
Panel A. ESSM, 1991-92/2002-03. Effect of the mother’s high education and income
Mother’s high edu 0.032 -0.005 0.010
(0.072) (0.015) (0.029)
Mother’s high edu ×Age -0.057*** -0.018*** -0.018***
(0.018) (0.005) (0.007)
Mother’s high edu ×Age20.004*** 0.001*** 0.001***
(0.001) (0.000) (0.000)
Ln(income) -0.032 -0.029*** -0.035***
(0.025) (0.007) (0.007)
Observations 13,532 13,532 12,246
Panel B. ESSM, 1991-92/2002-03. Effect of the father’s high education and income
Father’s high edu 0.031 0.022 -0.024
(0.077) (0.016) (0.030)
Father’s high edu ×Age -0.053*** -0.016*** -0.009
(0.019) (0.005) (0.007)
Father’s high edu ×Age20.003*** 0.001*** 0.001
(0.001) (0.000) (0.000)
Ln(income) -0.107*** -0.044*** -0.043***
(0.026) (0.007) (0.008)
Observations 11,954 11,954 10,773
Notes. In panel A, the sample contains children who live with their mother (i.e. with their mother and
father, or with their mother only), and in panel B, it contains children who live with their father (i.e.
with their mother and father, or with their father only). All models include child age, age squared, and
the basic controls. In addition, regressions in panel A control for family income and the mother’s labor
market status and body weight (in column (1) the mother’s BMI; in columns (2) and (3) the mother’s
overweight or obesity status). Regressions in panel B control for family income and the father’s body
weight (in column (1) the father’s BMI; in columns (2) and (3) the father’s overweight or obesity status).
Robust standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1.
37

Table 4: The gradients in education and income, using the fully interacted model (ESSM
data)
(1) (2) (3)
BMI-for-age Overweight or Overweight or
z-score obesity (WHO) obesity (IOTF)
Panel A. ESSM, 1991-92/2002-03
Ln(income) -0.0395 -0.0049 -0.0410
(0.0695) (0.0147) (0.0278)
Ln(income) ×Age -0.0431** -0.0172*** -0.0091
(0.0171) (0.0044) (0.0065)
Ln(income) ×Age20.0024*** 0.0009*** 0.0006*
(0.0009) (0.0003) (0.0003)
Observations 13,786 13,786 12,490
Panel B. ESSM, 1991-92/2002-03
Mother’s high edu ×Age -0.0531*** -0.0183*** -0.0194***
(0.0186) (0.0048) (0.0069)
Mother’s high edu ×Age20.0032*** 0.0011*** 0.0011***
(0.0010) (0.0003) (0.0003)
Observations 14,080 14,080 12,744
Panel C. ESSM, 1991-92/2002-03
Mother’s high edu ×Age -0.0383* -0.0122** -0.0147*
(0.0210) (0.0054) (0.0078)
Mother’s high edu ×Age20.0025** 0.0008** 0.0009**
(0.0011) (0.0003) (0.0004)
Ln(income) ×Age -0.0299 -0.0132*** -0.0052
(0.0192) (0.0050) (0.0073)
Ln(income) ×Age20.0015 0.0007** 0.0003
(0.0010) (0.0003) (0.0004)
Observations 13,610 13,610 12,316
Panel D. ESPS, 1996-2010
Ln(income) ×Age -0.0477*** -0.0123*** 0.0080
(0.0147) (0.0036) (0.0055)
Ln(income) ×Age20.0026*** 0.0006*** -0.0003
(0.0008) (0.0002) (0.0003)
Observations 22,903 22,903 20,519
Panel E. ESPS, 1996-2010
Mother’s high edu ×Age -0.0034 -0.0142*** -0.0149**
(0.0157) (0.0040) (0.0060)
Mother’s high edu ×Age2-0.0001 0.0007*** 0.0007**
(0.0008) (0.0002) (0.0003)
Ln(income) ×Age -0.0455*** -0.0067* 0.0140**
(0.0159) (0.0039) (0.0062)
Ln(income) ×Age20.0025*** 0.0003 -0.0006**
(0.0008) (0.0002) (0.0003)
Observations 22,353 22,353 20,006
Notes. In panels A and D, the sample contains children who live with either their mother or their father.
In panels B, C, D, and E, the sample contains children who live with their mother (i.e. with their mother
and father, or with their mother only). In all panels, we include chid age and age square and interaction
terms between the basic controls and child age and age square. Although the coefficients are not shown
for space reasons, in panels B, C, and E, we control for the mother’s high education, and in panels C, D,
and E, for the logarithm of income. Robust standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1.
38

Table 5: Estimation using the ESPS data and including cohort fixed effects
(1) (2) (3)
BMI-for-age Overweight or Overweight or
z-score obesity (WHO) obesity (IOTF)
Panel A. ESPS, 1996-2010. Effect of the mother’s high education
Mother’s high edu ×Age -0.018 -0.016*** -0.010**
(0.014) (0.003) (0.005)
Mother’s high edu ×Age20.001 0.001*** 0.001**
(0.001) (0.000) (0.000)
Observations 24,908 24,908 22,338
Panel B. ESPS, 1996-2010. Effect of the father’s high education
Father’s high edu ×Age -0.051*** -0.018*** -0.012**
(0.014) (0.003) (0.005)
Father’s high edu ×Age20.002*** 0.001*** 0.001**
(0.001) (0.000) (0.000)
Observations 22,956 22,956 20,502
Notes. In panel A, the sample contains children who live with their mother (i.e. with their mother and
father, or with their mother only), and in panel B, it contains children who live with their father (i.e. with
their mother and father, or with their father only). Parents’ education, child age and age squared, basic
controls, and cohort fixed effects, are included.
*** p<0.01, ** p<0.05, * p<0.1.
39

Table 6: Estimation using the ESPS data and including child fixed effects
(1) (2) (3)
BMI-for-age Overweight or Overweight or
z-score obesity (WHO) obesity (IOTF)
Panel A. ESPS, 1996-2008. Effect of the mother’s high education
Mother’s high edu ×Age -0.024 -0.017*** -0.015*
(0.022) (0.006) (0.008)
Mother’s high edu ×Age20.001 0.001** 0.001
(0.001) (0.000) (0.000)
Observations 14,367 14,367 13,261
Panel B. ESPS, 1996-2008. Effect of the father’s high education
Father’s high edu ×Age -0.063*** -0.022*** -0.020**
(0.024) (0.006) (0.009)
Father’s high edu ×Age20.003** 0.001*** 0.001**
(0.001) (0.000) (0.000)
Observations 13,467 13,467 12,394
Notes. In panel A, the sample contains children who live with their mother (i.e. with their mother and
father, or with their mother only), and in panel B, it contains children who live with their father (i.e. with
their mother and father, or with their father only). Child age and age squared, basic controls, and child
fixed effects, are included.
*** p<0.01, ** p<0.05, * p<0.1.
40

Table 7: Parents’ education and child body weight, for children ages 3-17 (ESSM data)
(1) (2) (3) (4) (5) (6)
BMI-for-age Overweight or Overweight or
z-score obesity (WHO) obesity (IOTF)
Panel A. ESSM, 1991-92/2002-03. Effect of the mother’s high education
Mother’s high edu -0.215*** -0.001 -0.080*** 0.002 -0.071*** 0.014
(0.026) (0.151) (0.007) (0.034) (0.007) (0.039)
Mother’s high edu ×Age -0.065** -0.024*** -0.025***
(0.032) (0.008) (0.008)
Mother’s high edu ×Age20.003** 0.001*** 0.001***
(0.001) (0.000) (0.000)
Observations 11,858 11,858 11,858 11,858 11,929 11,929
Panel B. ESSM, 1991-92/2002-03. Effect of the father’s high education
Father’s high edu -0.221*** -0.1249 -0.067*** 0.016 -0.073*** -0.033
(0.027) (0.159) (0.008) (0.035) (0.007) (0.038)
Father’s high edu ×Age -0.036 -0.021** -0.013
(0.034) (0.008) (0.008)
Father’s high edu ×Age20.002 0.001** 0.000*
(0.001) (0.000) (0.000)
Observations 10,688 10,688 10,688 10,688 10,751 10,751
Notes. In panel A, the sample contains children who live with their mother (i.e. with their mother and
father, or with their mother only), and in panel B, it contains children who live with their father (i.e. with
their mother and father, or with their father only). In columns (1), (3), and (5), the control variables are
child age dummies and the basic controls. In columns (2), (4), and (6), the control variables are child age,
age square, and the basic controls. Robust standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1.
41

Table 8: Estimates for boys and girls (ESSM and ESPS data)
(1) (2) (3)
BMI-for-age Overweight or Overweight or
z-score obesity (WHO) obesity (IOTF)
Panel A. ESSM, 1991-92/2002-03. Effect of the mother’s high education, for girls
Mother’s high edu ×Age -0.060** -0.021*** -0.020**
(0.024) (0.006) (0.009)
Mother’s high edu ×Age20.003** 0.001*** 0.001**
(0.001) (0.000) (0.000)
Panel B. ESSM, 1991-92/2002-03. Effect of the father’s high education, for girls
Father’s high edu ×Age -0.070*** -0.023*** -0.021**
(0.026) (0.006) (0.010)
Father’s high edu ×Age20.003** 0.001*** 0.001**
(0.001) (0.000) (0.000)
Panel C. ESSM, 1991-92/2002-03. Effect of the mother’s high education, for boys
Mother’s high edu ×Age -0.048* -0.013* -0.015
(0.026) (0.007) (0.010)
Mother’s high edu ×Age20.003** 0.001** 0.001**
(0.001) (0.000) (0.000)
Panel D. ESSM, 1991-92/2002-03. Effect of the father’s high education, for boys
Father’s high edu ×Age -0.048* -0.012 -0.005
(0.028) (0.008) (0.010)
Father’s high edu ×Age20.003** 0.001 0.000
(0.002) (0.000) (0.001)
Panel E. ESPS, 1996-2010. Effect of the mother’s high education, for girls
Mother’s high edu ×Age -0.013 -0.022*** -0.013*
(0.019) (0.005) (0.007)
Mother’s high edu ×Age20.000 0.001*** 0.001*
(0.001) (0.000) (0.000)
Panel F. ESPS, 1996-2010. Effect of the father’s high education, for girls
Father’s high edu ×Age -0.032 -0.019*** -0.018**
(0.020) (0.005) (0.008)
Father’s high edu ×Age20.001 0.001*** 0.001**
(0.001) (0.000) (0.000)
Panel G. ESPS, 1996-2010. Effect of the mother’s high education, for boys
Mother’s high edu ×Age -0.022 -0.011** -0.007
(0.019) (0.005) (0.007)
Mother’s high edu ×Age20.001 0.001** 0.000
(0.001) (0.000) (0.000)
Panel H. ESPS, 1996-2010. Effect of the father’s high education, for boys
Father’s high edu ×Age -0.068*** -0.016*** -0.005
(0.020) (0.005) (0.007)
Father’s high edu ×Age20.004*** 0.001*** 0.000
(0.001) (0.000) (0.000)
Notes. In panels A, C, E, and G, the sample contains children who live with their mother (i.e. with their
mother and father, or with their mother only), and in panels B, D, F, and H, it contains children who live
with their father (i.e. with their mother and father, or with their father only). Parents’ education, child
age and age squared, and basic controls are included. When the ESPS data are used, cohort fixed effects
are also included.
*** p<0.01, ** p<0.05, * p<0.1. 42

Table 9: Estimates of the gradient over time (ESSM and ESPS data)
(1) (2) (3)
BMI-for-age Overweight or Overweight or
z-score obesity (WHO) obesity (IOTF)
Panel A. ESSM, 1991-1992. Effect of the mother’s high education
Mother’s high edu ×Age -0.020 -0.000 -0.017
(0.036) (0.009) (0.012)
Mother’s high edu ×Age20.001 -0.000 0.001
(0.002) (0.001) (0.001)
Panel B. ESSM, 1991-1992. Effect of the father’s high education
Father’s high edu ×Age -0.067* -0.015 -0.026**
(0.036) (0.009) (0.013)
Father’s high edu ×Age20.004** 0.001* 0.001**
(0.002) (0.001) (0.001)
Panel C. ESSM, 2002-2003. Effect of the mother’s high education
Mother’s high edu ×Age -0.071*** -0.026*** -0.020**
(0.021) (0.005) (0.008)
Mother’s high edu ×Age20.004*** 0.002*** 0.001***
(0.001) (0.000) (0.000)
Panel D. ESSM, 2002-2003. Effect of the father’s high education
Father’s high edu ×Age -0.056** -0.021*** -0.006
(0.022) (0.006) (0.009)
Father’s high edu ×Age20.003** 0.001*** 0.000
(0.001) (0.000) (0.000)
Panel E. ESPS, 1996-2004. Effect of the mother’s high education
Mother’s high edu ×Age -0.009 -0.015*** -0.010
(0.017) (0.004) (0.006)
Mother’s high edu ×Age20.000 0.001*** 0.001
(0.001) (0.000) (0.000)
Panel F. ESPS, 1996-2004. Effect of the father’s high education
Father’s high edu ×Age -0.063*** -0.016*** -0.009
(0.018) (0.004) (0.007)
Father’s high edu ×Age20.003*** 0.001*** 0.001
(0.001) (0.000) (0.000)
Panel G. ESPS, 2006-2010. Effect of the mother’s high education
Mother’s high edu ×Age -0.038* -0.022*** -0.015*
(0.022) (0.005) (0.008)
Mother’s high edu ×Age20.002 0.001*** 0.001*
(0.001) (0.000) (0.000)
Panel H. ESPS, 2006-2010. Effect of the father’s high education
Father’s high edu ×Age -0.035 -0.021*** -0.018**
(0.023) (0.006) (0.009)
Father’s high edu ×Age20.001 0.001*** 0.001*
(0.001) (0.000) (0.000)
Notes. In panels A, C, E, and G, the sample contains children who live with their mother (i.e. with their
mother and father, or with their mother only), and in panels B, D, F, and H, it contains children who live
with their father (i.e. with their mother and father, or with their father only). Parents’ education, child
age and age squared, and basic controls are always included. When the ESPS data are used, cohort fixed
effects are also included. Robust standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1. 43

Figures
Figure 1: Evolution of the child BMI-for-age z-score and of the share of children overweight
or obese, with age (ESSM data)
−.2 0 .2 .4
BMI−for−age z−score
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Age
Girls Boys
ESSM data
Evolution of child BMI−for−age with age
.1 .15 .2 .25 .3
Overweight or obesity
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Age
Girls Boys
ESSM data
Evolution of child overweight or obesity (WHO) with age
.05 .1 .15 .2
Overweight or obesity
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Age
Girls Boys
ESSM data
Evolution of child overweight or obesity (IOTF) with age
44

Figure 2: Evolution of body weight with age, by parents’ educational level (ESSM data)
−.4 −.2 0 .2 .4
BMI−for−age
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Age
Mother’s low or medium edu Mother’s high edu
ESSM data
Child BMI−for−age and mother’s education
−.4 −.2 0 .2 .4
BMI−for−age
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Age
Father’s low or medium edu Father’s high edu
ESSM data
Child BMI−for−age and father’s education
0 .1 .2 .3
Overweight or obesity
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Age
Mother’s low or medium edu Mother’s high edu
ESSM data
Child overweight or obesity (WHO) and mother’s education
.05 .1 .15 .2 .25 .3
Overweight or obesity
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Age
Father’s low or medium edu Father’s high edu
ESSM data
Child overweight or obesity (WHO) and father’s education
.05 .1 .15 .2 .25
Overweight or obesity
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Age
Mother’s low or medium edu Mother’s high edu
ESSM data
Child overweight or obesity (IOTF) and mother’s education
.05 .1 .15 .2
Overweight or obesity
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Age
Father’s low or medium edu Father’s high edu
ESSM data
Child overweight or obesity (IOTF) and father’s education
45

Figure 3: The gradient in parents’ education and child body weight across childhood
(ESSM data)
−.4 −.3 −.2 −.1 0
Coefficient on education
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Child age group
Coeff 90% CI Smoothing
ESSM data
Effect of mother’s high edu on child BMI−for−age
−.4 −.3 −.2 −.1 0 .1
Coefficient on education
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Child age group
Coeff 90% CI Smoothing
ESSM data
Effect of father’s high edu on child BMI−for−age
−.15 −.1 −.05 0
Coefficient on education
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Child age group
Coeff 90% CI Smoothing
ESSM data
Effect of mother’s high edu on child overweight (WHO)
−.15 −.1 −.05 0
Coefficient on education
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Child age group
Coeff 90% CI Smoothing
ESSM data
Effect of father’s high edu on child overweight (WHO)
−.15 −.1 −.05 0
Coefficient on education
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Child age group
Coeff 90% CI Smoothing
ESSM data
Effect of mother’s high edu on child overweight (IOTF)
−.15 −.1 −.05 0
Coefficient on education
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Child age group
Coeff 90% CI Smoothing
ESSM data
Effect of father’s high edu on child overweight (IOTF)
Notes. The coefficients on the y-axis are obtained by estimating equation (1) separately by age groups.
The age groups are: 0-2, 1-3, 2-4, ..., 15-17. For the left-hand side subfigures, the sample includes children
who live with their mother (i.e. with their mother and father, or with their mother only). For the right-
hand side subfigures, the sample includes children who live with their father (i.e. with their mother and
father, or with their father only). The dots represent the estimated coefficients. The bars represent the
90% confidence intervals. The plain line represents a non-parametric smoothing.
46

Figure 4: The gradient in family income and child body weight across childhood (ESSM
data)
−.4 −.3 −.2 −.1 0 .1
Coefficient on ln(income)
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Child age group
Coeff 90% CI Smoothing
ESSM data
Effect of income on child BMI−for−age
−.15 −.1 −.05 0
Coefficient on ln(income)
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Child age group
Coeff 90% CI Smoothing
ESSM data
Effect of income on child overweight (WHO)
−.12 −.1 −.08 −.06 −.04 −.02
Coefficient on ln(income)
0−2 2−4 4−6 6−8 8−10 10−12 12−14 14−16
Child age group
Coeff 90% CI Smoothing
ESSM data
Effect of income on child overweight (IOTF)
Notes. The coefficients on the y-axis are obtained by estimating equation (1), using the logarithm of
income as our SES variable, separately by age groups. The age groups are: 0-2, 1-3, 2-4, ..., 15-17. The
dots represent the estimated coefficients. The bars represent the 90% confidence intervals. The plain line
represents a non-parametric smoothing.
47

Figure 5: The shape of the gradient with and without cohort-fixed effects (ESPS data)
−.24 −.22 −.2 −.18 −.16 −.14
Effect of the mother’s high edu (standard deviations unit)
0 2 4 6 8 10 12 14 16
Child age
Without cohort fixed effect With cohort fixed effects
ESPS data
Effect of mother’s high edu on child BMI−for−age
Notes. The curves represent the impact of maternal high education on child body weight, as a function
of child age, from equation (2). Child age and age square and basic controls are always included. Cohort
fixed effects are either included or not. The sample includes children who live with their mother (i.e. with
their mother and father, or with their mother only). The interpretation is the following: for children ages
0, when we do not control for cohort fixed effects, when the mother’s education increases from low or
medium to high, the BMI-for-age z-score decreases by 0.15 (the unit of BMI-for-age z-score is the standard
deviation unit).
48