Social Science & Medicine 187 (2017) 101e108

Contents lists available at ScienceDirect

Social Science & Medicine

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

Maternal breastfeeding and children's cognitive development*

Kanghyock Koh*
Ulsan National Institute of Science and Technology (UNIST), Republic of Korea

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

Article history: Do children with lower test scores benefit more from breastfeeding than those with higher scores? In
Received 28 November 2016 this paper, I examine the distributional effects of maternal breastfeeding on the cognitive test scores of
Received in revised form 11,544 children who were born in 2000 and 2001 in the United Kingdom using a semiparametric quantile
30 May 2017
regression model. I find evidence that maternal breastfeeding has larger positive impacts on children
Accepted 11 June 2017
with lower test scores. Effects for children below the 20th percentile are about 2e2.5 times greater than
Available online 12 June 2017
those for children above the 80th percentile. I also find that these distributional effects are larger when
the duration of breastfeeding is extended. One policy implication is that a public policy aims at pro-
Keywords:
United Kingdom
moting breastfeeding might narrow a disparity in children's cognition.
Distributional effects © 2017 Elsevier Ltd. All rights reserved.
Maternal breastfeeding
Children's development
Semiparametric quantile regression

1. Introduction et al., 1981). Note that the compositional superiority of breast

Economic theories of human capital have emphasized the
importance of parental investment in children (Becker, 1981;
Cunha and Heckman, 2007). A large number of empirical studies
show that human capital developments in early childhood (e.g.,
cognitive ability) play a significant role in human capital de-
velopments later in life, as measured by educational attainment,
employment, wages, etc. (Heckman, 2006; Cunha and Heckman,
2007; Cunha et al., 2010; Almond and Currie, 2011a,b).
Maternal breastfeeding has been emphasized as an influential
factor in early childhood development (e.g., Kramer et al., 2001,
2008) based on biological mechanisms through which maternal
breastfeeding aids children's development. First, the composition
of breast milk is superior to that of formula. Breast milk contains
long-chain polyunsaturated fatty acids, such as docosahexaenoic
acid (DHA) and arachidonic acid (AA), which form the major
structures of neuronal membranes and play critical roles in nervous
system functioning by positively stimulating development of the
human brain (Fernstrom, 1999). Infants require sufficient amounts
of these acids during the first few months after birth (Clandinin
*
I thank Anna Aizer, Kenneth Chay, Hyojin Han, Hyunjoo Yang, and the editor
and four anonymous referees for valuable comments. Any remaining errors are
mine. This work was supported by the New Faculty Research Fund(1.160080) of
UNIST(Ulsan National Institute of Science & Technology).
* 50 UNIST-gil, Ulju-gun, Ulsan 44919, Republic of Korea.
E-mail address: kanghyock.koh@unist.ac.kr.
milk may be reduced because DHA and AA are now added to for-
mula. However, Fitzsimons and Vera-Hernandez (2012) reported
that DHA and AA were not included in formula in the UK (the
context of this paper) until most of the infants who are considered
in the empirical analysis were born. Deoni et al. (2013) provided
MRI evidence that breastfed infants exhibit better development in
specific brain areas associated with language and visual reception
abilities compared to those who are fed formula or a mixture of
breast milk and formula. Second, skin-to-skin contact between
mother and child stimulates maternal hormonal responses such as
the production of prolactin and oxytocin, which may indirectly
improve cognitive development (Del Bono and Rabe, 2012).
In spite of this clear biological mechanism, empirical evidence
for the effects of breastfeeding on children's cognitive development
has been conflicting. A comprehensive review by the Agency for
Healthcare Research and Quality (Ip et al., 2007) summarized 400
articles (out of 9000 abstracts) and found that breastfeeding has
few or small effects on children's cognitive ability. Many studies are
based on observational data, and it is difficult to infer causality due
to many potential confounding variables, such as duration of
breastfeeding, children's demographics and birth outcomes (such
as birth weight and gestational age), parental health-related
behavior, socioeconomic status (SES), intelligence, and family
characteristics (Anderson et al., 1999).
Researchers use various empirical strategies to address the
endogeneity issue. First, they estimate the probability of breast-
feeding using detailed information regarding the characteristics of

http://dx.doi.org/10.1016/j.socscimed.2017.06.012
0277-9536/© 2017 Elsevier Ltd. All rights reserved.
102 K. Koh / Social Science & Medicine 187 (2017) 101e108
children, parents, and families. They then estimate the effects of
breastfeeding on cognitive development by comparing children
with similar propensities of maternal breastfeeding (Jiang et al.,
2011; Belfield and Kelly, 2012; Borra et al., 2012; Rothstein, 2012;
Cesur et al., 2017). For example, Jian, Foster and Gibson-Davis
(2011) used the Panel Study of Income Dynamics in the U.S. and
found that the positive associations between breastfeeding initia-
tion and children's cognitive development measured by the
Woodcock Johnson Psycho-Educational Battery Revised (WJ-R) and
Wechsler Intelligence Scale for Children-Revised (WISC-R) were
significantly reduced when observational characteristics were
controlled for. The estimated associations are between one-tenth
and one-fifth of a standard deviation, which is small in magni-
tude. Second, some researchers use sibling or family fixed effects to
control for unobserved family characteristics (Der et al., 2006; Rees
and Sabia, 2009; Colen and Ramey, 2014; Cesur et al., 2017). Recent
research points out that maternal intelligence or IQ and home
environment explain breastfeeding status better than other SES
factors such as income or education, but few previous studies have
accounted for these covariates (Der et al., 2006). Many studies using
the propensity score approach or fixed effects analysis have found
that a positive association became smaller or disappeared once a
large number of confounding factors were controlled for. Re-
searchers pointed out these results as evidence for the endogeneity
issue in breastfeeding (Belfield and Kelly, 2012; Cesur et al., 2017).
Third, economists use exogenous sources of variation in breast-
feeding status as instrumental variables to study the causal rela-
tionship between breastfeeding and children's cognitive
development (Del Bono and Rabe, 2012; Fitzsimons and Vera-
Hernandez, 2012). For example, infant feeding support is pro-
vided by midwives and nurses in the U.K. Since staff working hours
are reduced during the weekend, this support is reduced as well.
Mothers are less exposed to support when they give birth on Friday
or Saturday. Fitzsimons and Vera-Hernandez (2012) use this insti-
tutional fact to estimate the effects of breastfeeding on children's
cognitive development among low-income mothers and their
children. Finally, Kramer et al. (2001, 2008) studied the causal ef-
fects of the Promotion of Breastfeeding Intervention Trial (PROBIT),
which randomly provided health care worker assistance for the
initiation and maintenance of breastfeeding, on breastfeeding
initiation and duration and consequences for children's health and
cognitive abilities in Belarus.
Previous literature mainly focused on the average effects of
breastfeeding on children's cognition. Many measures for cogni-
tion, such as test scores, are continuously distributed. These dis-
tributions could provide a deeper understanding of the effects of
breastfeeding on children's cognition by evaluating the effects at
different quantiles in addition to the average effects. For example,
children at lower quantiles could benefit more from breastfeeding
than average children or children at higher quantiles, even though
the average effects of breastfeeding are small or insignificant based
on the existing literature. However, there is little knowledge
regarding the effects of breastfeeding on the distribution of
cognition. To fill this gap, I investigated the effects of maternal
breastfeeding on the distribution of children's cognitive ability in
this study. More specifically, I estimated the effects of breastfeeding
at different quantiles of cognitive test scores using a semi-
parametric quantile regression model. To address the endogeneity
of breastfeeding, I used the propensity score as the inverse proba-
bility (Rosenbaum and Rubin, 1983; Hirano et al., 2003). Based on
rich information about children's characteristics, cognitive test
scores, and maternal breastfeeding from the UK Millennium Cohort
Survey, I compared cognitive test scores among children with
similar propensities for maternal breastfeeding based on observed
parental characteristics.
2. Method
2.1. Data
To investigate the effects of maternal breastfeeding on children's
development, I used data from the Millennium Cohort Study (MCS),
a longitudinal study of about 18,500 children who were born in
2000 and 2001 in the United Kingdom. Specifically, I used data from
four surveys (at ages 9 months, 3 years, 5 years, and 7 years) and
excluded data for multiple births and children who were not living
with their biological mothers at the time of the first interview
(n ¼ 293). Since I used secondary observational data, ethics
approval was not required.
I collected data on self-reported maternal breastfeeding status
from the first MCS survey, which included a question about the
duration of maternal breastfeeding. Mothers reported the age of
babies when breastfeeding stopped, as measured in days, weeks, or
months, which was used for calculating the duration of breast-
feeding. It was not feasible to obtain data on exclusive breast-
feeding from the MCS because survey questions related to infant
feeding were not exhaustive. Since the recommended duration of
breastfeeding in the UK was 4 months in 2000, I created a binary
breastfeeding variable indicating whether children were breastfed
for at least 4 months. To study the effects of breastfeeding with
different duration cutoffs, I examined the effects of initiation of
breastfeeding. Given current policy recommendations [e.g., from
the World Health Organization (WHO)], I also examined the effects
of extended breastfeeding (i.e., for at least 6 months).
To assess children's cognitive development, I used scores from
six British Ability Scale (BAS) tests, which are used to measure the
cognitive abilities of children aged 2.5e8 years old (Elliott et al.,
1997). Since these tests are individually administered by trained
interviewers, the scores provide a more accurate measure than
parents' self-reported measures (Fernald et al., 2009). Using the
age-adjusted scores from the MCS, I calculated average scores
within and across ages and created a summary index. I did this to
address a concern that one null hypothesis could be rejected simply
because I tested it with multiple null hypotheses (Kling et al., 2007).
This also yields a well-behaved continuous distribution of test
scores (Fig. 1).
I also collected information on maternal demographics,
maternal prenatal characteristics, and spouse and family charac-
teristics as control variables for the propensity score approach:
mother's age at birth, race, marital status, and education; planned
Fig. 1. Distribution of children's cognitive development. Data source: The MCS.
 K. Koh / Social Science & Medicine 187 (2017) 101e108 103

pregnancy; dummy variables for receiving any prenatal care and
attending any prenatal classes; timing of initial care; dummy var-
iables for any complications during pregnancy (including bleeding
or threatened miscarriage in early pregnancy; bleeding in later
pregnancy; pregnancy diagnosed as twins, triplets, or more;
persistent vomiting; elevated blood pressure; eclampsia/pre-
eclampsia or toxemia; urinary infection; diabetes; too much fluid
around the baby; suspected slow growth of baby; and other sus-
pected problems) and labor (including breech birth, shoulder first,
very long (or rapid) labor, fetal distress e heart rate sign, meconium
sign, and other complications); smoking status during pregnancy;
working status during and after pregnancy; frequency of alcohol
consumption; dummy variables for countries (England, Wales,
Scotland, and Northern Ireland); father's age at birth, education
level, and job status; dummy variables for joint job status of a
couple indicating if both in work, father (mother) in work but
mother (father) not in work, both not in work, father (mother) in
work or on leave but mother (father) not in work, and father
(mother) in work but mother (father) is unknown; number of
household members and smokers in the household; and dummy
variables for government support and income level. The mothers'
and fathers' education levels were categorized as NVQ level 1 to 5
and overseas. I defined a dummy variable of low education level if
the NVQ level was 3 or lower to construct interactions with other
covariates. Due to missing values in these control variables, I used
11,540 children in the main analysis.
2.2. Empirical strategy
If T is the treatment indicator, Y1 is the outcome if treated
(breastfed for more than 4 months), and Y0 is the outcome if un-
treated (breastfed for less than 4 months), then the average treat-
ment effect on the treated (ATT) is EðY1 jT ¼ 1Þ  EðY0 jT ¼ 1Þ. Let
the conditional probability of being treated based on observed
characteristics be the propensity score (Rosenbaum and Rubin,
1983), pðXÞ ¼ PrðTjX ¼ xÞ, where X is a vector of observed pre-
treatment characteristics such as maternal characteristics, parental
characteristics, and household environment. I include control var-
iables mentioned in the previous section. I also include polynomials
and interaction terms among them. I estimated the propensity
scores using the logit regression specification and used their inverse
probability weights to nonparametrically estimate the ATT (Hirano
et al., 2003). I used b p ðXÞ/(1  b p ðXÞ) as a weight for the control
group. Since the ATT may not be constant, I also estimated average
treatment effects (ATE) using 1/ b p ðXÞ and 1=ð1  b p ðXÞÞ as sampling
weights for treatment and control groups, respectively. For statis-
tical inferences, I calculated robust standard errors to correct for
heteroscedasticity in test scores by breastfeeding status. The pro-
pensity score approach has some advantages over the ordinary least
squares (OLS) approach with many control variables. First, I can
alleviate misspecification errors by not assuming any parametric
relationship between breastfeeding and children's cognitive
development (Zhao, 2008). Second, I can address the curse of
dimensionality issue associated with OLS when many control var-
iables are included (Dehejia and Wahba, 1999) by summarizing all
information from the control variables into a single index (i.e., the
propensity score). Third, I can graphically examine the nonpara-
metric relationship between maternal breastfeeding and children's
cognitive development by using the propensity score.
Then I investigated the distributional impact of maternal
breastfeeding using a semiparametric quantile regression model.
Following Firpo (2007), I estimated the quantile treatment effect for
quantile t using D b t ≡b
q 1;t  b
q 0;t (j ¼ 0; 1Þ; where b q j;t is defined as
PN
argminq i¼1 u b j $rt ðYi  qÞ, the check function rt ð$Þ at a real number
c is rt ðcÞ ¼ c$ðt  1½c  0Þ; and u b j is the inverse probability of the
estimated propensity score for the ATT, which are
u
b1 ¼ T and u
b0 ¼ 1T where N is the sample size. I esti-
N,b
p ðXÞ N,ð1b
p ðXÞÞ
mated the propensity scores using local logit regression. To esti-
mate treatment effects at different quantiles, the outcome variable
should be continuous and well behaved without spikes or gaps,
which is again confirmed by Fig. 1.
3. Empirical results
Table 1 shows summary statistics by breastfeeding status. The
first three columns show that in general, a statistically significant

Table 1
Baseline characteristics of control variables.

Variables No propensity score adjustment Propensity score adjustment

Breastfeeding Difference Breastfeeding Difference

Y N Y-N Y N Y-N

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

Maternal age at birth 30.97 28.31 2.66*** 30.97 30.97 0.00

Low education 0.39 0.67 0.28*** 0.39 0.39 0.00
Low household income 0.38 0.58 0.20*** 0.38 0.38 0.00
Working (after birth) 0.56 0.53 0.03*** 0.56 0.56 0.00
White 0.82 0.88 0.06*** 0.82 0.80 0.02
Married at pregnancy 0.97 0.95 0.02*** 0.97 0.97 0.00
Planned pregnancy 0.69 0.60 0.09*** 0.69 0.69 0.00
Prenatal care (ever) 0.98 0.97 0.01*** 0.98 0.97 0.01
Prenatal class (ever) 0.43 0.34 0.09*** 0.43 0.43 0.00
Smoking during pregnancy 0.10 0.26 0.16*** 0.1 0.09 0.01
Any complications during pregnancy 0.35 0.39 0.04*** 0.35 0.36 0.01
Alcohol consumption 4.82 5.12 0.31*** 4.82 4.82 0.00
Any government support 0.21 0.37 0.16*** 0.21 0.21 0.00
Number of smokers in household 0.05 0.14 0.09*** 0.05 0.05 0.00
Spouse's age at delivery 33.63 31.06 2.57*** 33.63 33.64 0.01
Low education (spouse) 0.41 0.66 0.25*** 0.41 0.41 0.00

Note: I report differences in selected characteristics by maternal breastfeeding status. The first three columns show average values and their differences without any
adjustment. The last three columns show average values and their differences after the propensity score adjustment. I used b
p ðXÞ/(1  b
p ðXÞ) as a weight for the control group.
***p < 0.01, **p < 0.05, *p < 0.10.
Data source: The MCS.
104 K. Koh / Social Science & Medicine 187 (2017) 101e108
proportion of children who were breastfed for less than 4 months
came from families with lower SES. The last three columns show
average values and their differences after the propensity score
adjustment. I used b p ðXÞ/(1  b
p ðXÞ) as a weight for the control
group. Compared to the unadjusted differences, the adjusted dif-
ferences are smaller and statistically insignificant, implying that
treatment and control groups appear to be comparable based on
observable characteristics. Fig. 2 shows the distributions of esti-
mated propensity scores by breastfeeding status, which are virtu-
ally the same. This also implies that the treatment and control
groups have similar observable characteristics conditional on the
propensity score.
In Fig. 3, I present a plot of the estimated propensity scores
against the actual proportion of breastfeeding in 100 equally sized
cells. The graph shows that the 100 cells are generally clustered
around the 45-degree line, which implies that the estimated pro-
pensity score predicts the actual proportion of breastfeeding fairly
well.
Fig. 2. Distribution of estimated propensity score by breastfeeding status. Data source:
The MCS.
Fig. 3. Estimated propensity score and actual probability of breastfeeding
(4 Months). Data source: The MCS. Note: This figure shows a plot of the average
estimated propensity scores against the actual probability of breastfeeding in 100
equal-sized cells. The red line is the 45-degree line. (For interpretation of the refer-
ences to colour in this figure legend, the reader is referred to the web version of this
article.)
To show the effects of maternal breastfeeding graphically, I
plotted the averages of children's cognitive test scores against the
estimated propensity scores separately in 20 equally sized cells by
treatment status (Fig. 4). The positive slopes indicate a general
positive relationship between maternal breastfeeding and chil-
dren's cognitive test scores. Among children with similar pro-
pensity scores, children who were breastfed generally had higher
test scores than their counterparts who were not breastfed. This
implies that maternal breastfeeding has a positive impact on
cognitive development among children with similar propensity
scores.
In panel A and B of Table 2, I present the estimated ATT and ATE.
Maternal breastfeeding increases cognitive test scores by 1.04 and
0.94 points, respectively. Compared to the standard deviation of the
dependent variable, the magnitudes of the impact of maternal
breastfeeding on children's cognitive development are around 10%
of the standard deviation. The estimated effects of maternal
breastfeeding are also robust under other regression specifications.
In columns (1) and (2) of Appendix Table 1, instead of nonpara-
metric estimations with inverse probabilities, I used b p ðXÞ and b
p ðXÞ2
as parametric controls and included an interaction term between
maternal breastfeeding and b p ðXÞ to capture differential relation-
ships between the propensity of maternal breastfeeding and chil-
dren's cognitive test scores, respectively. Finally, in column (3), I
added the control variables used in the propensity score estimation
to the parametric specifications.
In panel C of Table 2, I present the quantile regression coefficient
values for cognitive test scores. The impact of maternal breast-
feeding is around 1.1 to 1.2 points for children below the 20th
percentile of the cognitive test score distribution and around 0.5
points for children above the 80th percentile of the cognitive test
score distribution. The impacts of maternal breastfeeding on chil-
dren's development are greater among children with lower
cognitive test scores.
Table 3 shows estimated effects of breastfeeding with different
cutoffs. I used ever breastfeeding status and extended breastfeed-
ing status (i.e., whether a child was breastfed for at least 6 months).
I used the same regression specification for the propensity score
estimation. Columns (1) and (2) show ATTs, ATEs, and quantile
regression coefficient values. The estimated ATTs and ATEs on
cognitive test scores do not change. One thing to note is that the
Fig. 4. Children's cognitive development as a function of estimated propensity score.
Data source: The MCS. Note: This figure shows a plot of average cognitive test scores
against estimated propensity scores in 20 equal-sized cells by maternal breastfeeding
status.
 K. Koh / Social Science & Medicine 187 (2017) 101e108 105

Table 2 Table 3
Distributional effects of maternal breastfeeding on children's cognitive test scores. Distributional effects of different durations of maternal breastfeeding on children's
cognitive test scores.
Treatment status Breastfeeding  4 months
Treatment status Ever Breastfeeding Breastfeeding  6 months
A. ATT
1.04*** (1) (2)
(0.22)
A. ATT
B. ATE 1.26*** 0.96***
0.94*** (0.31) (0.24)
(0.27)
B. ATE
C. Quantiles 1.16*** 0.83**
0.1 1.08* (0.27) (0.33)
(0.62)
C. Quantiles
0.2 1.23***
0.1 1.33** 1.42*
(0.46)
(0.55) (0.82)
0.3 1.17***
0.2 1.5*** 1.58***
(0.38)
(0.50) (0.56)
0.4 1.23***
0.3 1.17*** 1.50***
(0.33)
(0.38) (0.43)
0.5 1.17***
0.4 1.17*** 1.00***
(0.29)
(0.37) (0.39)
0.6 1.00***
0.5 1.13*** 1.13***
(0.28)
(0.33) (0.36)
0.7 0.83***
0.6 1.33*** 1.10***
(0.26)
(0.31) (0.33)
0.8 0.50*
0.7 1.00*** 0.70**
(0.27)
(0.36) (0.29)
0.9 0.50*
0.8 1.00*** 0.27
(0.30)
(0.38) (0.29)
Mean (SD) of outcome 61.52 (10.05)
0.9 0.93** 0.33
Note: For treatment variable, I use the dummy variable if a child was breastfed for at (0.44) (0.35)
least 4 months. The cognitive test score measure is the average score of six British Mean (SD) of outcome 61.52 (10.05)
Ability Scales (BAS) for naming vocabulary, pattern construction, picture similarity,
Note: In each column, I estimate the effects of breastfeeding using different cutoffs.
and word reading. I used different sampling weights for ATTs and ATE:
For treatment variable, I use the dummy variable if a child was ever breastfed in
b
p ðXÞ/(1  b
p ðXÞ) as a sampling weight for the control group in panel A and C, and 1/
columns (1), and at least 6 months in column (2). The cognitive test score measure is
b
p ðXÞ and 1=ð1  b p ðXÞÞ as sampling weights for treatment and control groups in
the average score of six British Ability Scales (BAS) for naming vocabulary, pattern
Panel B. Robust standard errors are calculated and presented in the parentheses.
construction, picture similarity, and word reading. I used different sampling weights
***p < 0.01, **p < 0.05, and *p < 0.10.
for ATTs and ATE: b p ðXÞ/(1  b
p ðXÞ) as a sampling weight for the control group in
Data source: The MCS.
panel A and C, and 1/ bp ðXÞ and 1=ð1  b
p ðXÞÞ as sampling weights for treatment and
control groups in Panel B. Robust standard errors are calculated and presented in the
distributional effects become larger for extended duration of parentheses. ***p < 0.01, **p < 0.05, and *p < 0.10.
breastfeeding. The impacts of ever breastfeeding for those below Data source: The MCS.

the 20th percentile are about one and a half times larger than for
those above the 80th percentile. The impacts of extended maternal
breastfeeding for those below the 20th percentile are about five
times larger than for those above the 80th percentile.
Table 4 shows results of robustness checks. To begin with, I
included multiple birth sample because it has different character-
istics compared to the baseline sample (Appendix Table A2). Panels
A, B, and C of column (1) show ATT, ATE, and quantile regression
coefficient values. The estimated effects of breastfeeding did not
change much from the baseline results. Next, the propensity score
approach could not infer the causal effects of breastfeeding because
of unmeasured confounding (Nickel, 2015). To examine any bias
due to this, I included variables about children and maternal health
that were not accounted for from the baseline propensity score
estimation. First, I used birth weights and gestational age as mea-
sures for children's health because children's birth weights and
birth outcomes are important determinants of children's cognitive
development (Almond and Currie, 2011b) and may affect maternal
breastfeeding decisions (Cunha and Heckman, 2007). Children with
better birth outcomes as measured by birth weight and gestational
age are more likely to be breastfed (Appendix Table A3). Second,
maternal health status is also considered an important factor for
breastfeeding (Jiang et al., 2011). I used self-evaluated health status
and a dummy variable of having any chronic disease from the first
wave of the MCS as measures of maternal health. Since there was
no direct information on maternal health status during breast-
feeding, I used those variables as proxies. Panels A and B of columns
(2) and (3) show that the magnitudes of the ATTs and ATEs are
slightly smaller than my baseline regression results: the effects are
around 0.9 and 0.8 points, respectively, which again equals
approximately 10% of the standard deviation of the cognitive score
distribution. Panel C of both columns shows that the impacts of
maternal breastfeeding on children's development are still greater
among children with lower cognitive test scores: the impact of
maternal breastfeeding is around 1.2 (1.0) points for children below
the 20th percentile of the cognitive test score distribution and
around 0.3 (0.3) points for children above the 80th percentile of the
cognitive test score distribution.
4. Discussion
Previous literature focused on the average effects of breast-
feeding on children's cognition. This study first examined the
distributional impacts of maternal breastfeeding on children's
development using a semiparametric quantile regression model. I
found that the effects of breastfeeding on children's test scores
were greater among children with lower test scores. The effects
were around 12e13% of the standard deviation, which is considered
small in magnitude according to Cohen's effect size rule (Cohen,
1992) and the previous literature (Ip et al., 2007; Jiang et al.,
2011). However, the effects could become larger at the population
level as the disparity in educational achievement becomes larger
within a society (Rose, 1985). Given evidence that breastfeeding is
associated with a reduction in several health risks among children
and mothers (Ip et al., 2007; Eidelman et al., 2012), the benefits of
106 K. Koh / Social Science & Medicine 187 (2017) 101e108

Table 4 An economic theory may provide a possible explanation for the

Robustness checks. main findings of this study. Let's assume that children's human
Including Including birth Including both birth capital production function is concave against maternal breast-
multiple births outcomes outcomes and feeding. Low SES children usually have a low chance of being
maternal health breastfed and so achieve low test scores based on the underlying
(1) (2) (3) biological mechanism. Because they are less likely to be breastfed,
A. ATT marginal returns on breastfeeding could be larger than those in
1.06*** 0.90*** 0.91*** higher SES children due to the concavity of the human capital
(0.22) (0.23) (0.23) production function. These results have a policy implication for the
B. ATE promotion of breastfeeding among individuals with lower SES.
0.98*** 0.81*** 0.79*** Significant SES-related disparities in children's academic perfor-
(0.26) (0.26) (0.27) mance are well established in the UK (Banerjee, 2016). Government
C. Quantiles programs such as Sure Start have been introduced to narrow these
0.1 1.08* 1.00 0.83 gaps (Melhuish et al., 2008). Since a high proportion of children
(0.62) (0.65) (0.65) with lower test scores generally come from families with low SES
0.2 1.23*** 1.17*** 1.00**
(0.46) (0.45) (0.46)
and since there are observed SES-related disparities in maternal
0.3 1.17*** 1.17*** 1.00*** breastfeeding incidence and duration, government policies such as
(0.38) (0.39) (0.39) the UNICEF UK Baby Friendly Initiative aims at promoting breast-
0.4 1.23*** 1.17*** 1.16*** feeding may reduce SES-related gaps in children's academic
(0.33) (0.33) (0.34)
development. However, there are at least two reasons why this
0.5 1.17*** 1.00*** 0.83***
(0.29) (0.30) (0.30) policy implication should be interpreted with caution. First, the
0.6 1.00*** 1.00*** 0.92*** policy implication of this study is based on results with strong
(0.28) (0.28) (0.28) statistical assumptions. It is hard to draw causal inferences
0.7 0.83*** 0.73*** 0.73*** regarding the effects of breastfeeding. Second, the results may not
(0.26) (0.26) (0.26)
0.8 0.50* 0.33 0.33
be generalizable. Public agencies such as the American Academy of
(0.27) (0.27) (0.27) Pediatrics and the WHO recommend that mothers exclusively
0.9 0.50* 0.50* 0.50* breastfeed their infants for at least 6 months and ideally continue to
(0.30) (0.30) (0.30) breastfeed beyond 6 months. Under the Affordable Care Act in the
Mean (SD) of 61.52 (10.05)
United States, the initiation and extended duration of maternal
outcome
breastfeeding are promoted. However, this study is based on data
Note: In each column, I estimate the effects of breastfeeding using different speci- from the UK. The results should therefore be cautiously applied to
fications. I include multiple births in column (1), and children's birth weights and
gestational age in column (2), and maternal health status as well as birth weights
other countries owing to significant differences in health care and
and gestational age in column (3). For treatment variable, I use the dummy variable other social policy.
if a child was breastfed for at least 4 months. The cognitive test score measure is the As an avenue for future research, the distributional effects on
average score of six British Ability Scales (BAS) for naming vocabulary, pattern other benefits of breastfeeding, such as the health benefits for
construction, picture similarity, and word reading. I used different sampling weights
children and mothers, could also be considered. Previous literature
for ATTs and ATE: b p ðXÞ/(1  bp ðXÞ) as a sampling weight for the control group in
panel A and C, and 1/ b p ðXÞ and 1=ð1  b p ðXÞÞ as sampling weights for treatment and has mainly focused on the average effects of breastfeeding, but
control groups in Panel B. Robust standard errors are calculated and presented in the there is little evidence of the distributional effects on children's and
parentheses. ***p < 0.01, **p < 0.05, and *p < 0.10. maternal health. To draw a more convincing causal inference on the
Data source: The MCS. effects of breastfeeding, it would be useful to exploit a natural
experiment that quasi-randomizes breastfeeding status or
breastfeeding could be larger when considering such health ben- duration.
efits (Cesur et al., 2017).
The method and data suffer from some limitations. First, the
Appendix
propensity score approach is based on the ignorability assumption
that there is no other confounding factor once a large number of
covariates are controlled for (Rosenbaum and Rubin, 1983). This
assumption is too strong to infer causal effects of breastfeeding due Table A1
to unmeasured or unmeasurable confounding (Nickel, 2015). As an Effects of maternal breastfeeding on children's cognitive test scores.
example of unmeasured confounding, the MCS does not provide
(1) (2) (3)
information on mothers’ intelligence, such as IQ or test scores, or
home environment scores, which are considered important con- Breastfed  4 months 1.022*** 1.032** 1.090***
(0.207) (0.496) (0.214)
founding factors (Der et al., 2006; Belfield and Kelly, 2012; Cesur
et al., 2017). In addition, it is almost impossible from most obser- Parametric control of bp ðXÞ Y Y Y
vational data to obtain information on unmeasurable factors such Breastfed4mon b p ðXÞ Y
Control variables (XÞ Y
as maternal genes. Second, the data do not provide enough infor-
mation to construct an exclusive breastfeeding variable. Since the Data source: The MCS.
treatment group consisted of children who were breastfed at all, Note: Breastfed 4months indicates whether a child was breastfed for at least 4
months. The cognitive test score measure is the average of six British Ability Scales
this may underestimate the actual benefits of exclusive breast- (BAS) for naming vocabulary, pattern construction, picture similarity, and word
feeding (Jiang et al., 2011). Finally, breastfeeding duration was reading. In column (1), I use b p ðXÞ and b p ðXÞ2 as controls instead of using sampling
based on maternal recall. This implies that there may be errors in weights. In column (2), I add Breastfed4mon  b p ðXÞ to the regression specification of
treatment status and could cause additional bias in the estimates. column (1). In column (3), I add b p ðXÞ and bp ðXÞ2 and the control variables that were
used to estimate bp ðXÞ to the regression specification of column (1). Robust standard
Since it is unclear if this type of error is random or systematic
errors are calculated and presented in the parentheses. ***p < 0.01, **p < 0.05, and
(Promislow et al., 2005), it is hard to see the exact direction of the *p < 0.10.
bias.
 K. Koh / Social Science & Medicine 187 (2017) 101e108 107

Table A2
Characteristics of baseline and excluded sample.

Variables Baseline Excluded Excluded - Baseline

(1) (2) (3)

Maternal age at birth 28.29 30.81 2.52***

Low education 0.61 0.56 0.05*
Low household income 0.61 0.56 0.05
Working (after birth) 0.47 0.44 0.03
White 0.84 0.89 0.05**
Married at pregnancy 0.81 0.88 0.07***
Planned pregnancy 0.54 0.63 0.09***
Prenatal care (ever) 0.96 0.97 0.01
Prenatal class (ever) 0.34 0.36 0.02
Smoking during pregnancy 0.25 0.18 0.07***
Any complications during pregnancy 0.38 0.48 0.1***
Alcohol consumption 5.14 4.87 0.27
Any government support 0.43 0.43 0.00
Number of smokers in household 0.14 0.09 0.05**
Spouse's age at delivery 31.88 33.60 1.72***
Low education (spouse) 0.40 0.37 0.03

Data source: The MCS.

Note: I used the same variables that I used in Table 1. ***p < 0.01, **p < 0.05, and *p < 0.10.

Table A3
Maternal breastfeeding and children's birth outcomes.

Birth weight (grams) Pr (Birth weight  2500 g) Pr (Birth weight  1500 g) Gestational age (days) Pr (Gestational age  37 weeks)

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

Breastfed  4 months 77.607*** 0.030*** 0.005** 1.687*** 0.019***

(12.932) (0.006) (0.002) (0.286) (0.005)

Average of control group 3199.442 0.094 0.013 275.617 0.069

Observations 12,788 12,796 12,796 12,703 12,703

R-squared 0.005 0.003 0.001 0.004 0.002

Data source: The MCS.

Note: Breastfed  4 months indicates whether a child was breastfed for at least 4 months. I used b
p ðXÞ=1  b
p ðXÞ as a sampling weight for the control group. Robust standard
errors are calculated and presented in parentheses. ***p < 0.01, **p < 0.05, and *p < 0.10.

References
Almond, D., Currie, J., 2011a. Human capital development before age five. Handb.
Labor Econ. 4, 1315e1486.
Almond, D., Currie, J., 2011b. Killing me softly: the fetal origins hypothesis. J. Econ.
Perspect. 25 (3), 153e172.
Anderson, J.W., Johnstone, B.M., Remley, D.T., 1999. Breast-feeding and cognitive
development: a meta-analysis. Am. J. Clin. Nutr. 70 (4), 525e535.
Banerjee, P.A., 2016. A systematic review of factors linked to poor academic per-
formance of disadvantaged students in science and maths in schools. Cogent
Educ. 3 (1), 1178441.
Becker, G., 1981. A Treatise on the Family. Harvard University Press, Cambridge, MA.
Belfield, C.R., Kelly, I.R., 2012. The benefits of breastfeeding across the early years of
childhood. J. Hum. Cap. 6 (3), 251e277.
Borra, C., Iacovou, M., Sevilla, A., 2012. The effect of breastfeeding on children's
cognitive and noncognitive development. Labour Econ. 19 (4), 496e515.
Cesur, R., Sabia, J.J., Kelly, I.R., Yang, M., 2017. The effect of breastfeeding on young
adult wages: new evidence from the add health. Rev. Econ. Househ. 15 (1),
25e51.
Clandinin, M.T., Chappell, J.E., Heim, T., Swyer, P.R., Chance, G.W., 1981. Fatty acid
utilization in perinatal de novo synthesis of tissues. Early Hum. Dev. 5 (4),
355e366.
Cohen, J., 1992. A power primer. Psychol. Bull. 112 (1), 155.
Colen, C.G., Ramey, D.M., 2014. Is breast truly best? Estimating the effects of
breastfeeding on long-term child health and wellbeing in the United States
using sibling comparisons. Soc. Sci. Med. 109, 55e65.
Cunha, F., Heckman, J.J., 2007. The technology of skill formation. Am. Econ. Rev. 97
(2), 31.
Cunha, F., Heckman, J.J., Schennach, S.M., 2010. Estimating the technology of
cognitive and noncognitive skill formation. Econometrica 78 (3), 883e931.
Dehejia, R.H., Wahba, S., 1999. Causal effects in nonexperimental studies: reevalu-
ating the evaluation of training programs. J. Am. Stat. Assoc. 94 (448),
1053e1062.
Del Bono, E., Rabe, B., 2012. Breastfeeding and Child Cognitive Outcomes: Evidence
from a Hospital-based Breastfeeding Support Policy (No. 2012-29). ISER
Working Paper Series, 2012-29. http://hdl.handle.net/10419/91672.
Deoni, S.C.L., Dean III, D.C., Piryatinksy, I., O'Muircheartaigh, J., Waskiewicz, N.,
Lehman, K., Han, M., Dirks, H., 2013. Breastfeeding and early white matter
development: a cross sectional study. NeuroImage 82, 77e86.
Der, G., Batty, D., Deary, I.J., 2006. Effect of breast feeding on intelligence in children:
prospective study, sibling pairs analysis, and meta-analysis. Br. Med. J. 333,
945e948.
Eidelman, A.I., Schanler, R.J., Johnston, M., Landers, S., Noble, L., Szucs, K.,
Viehmann, L., 2012. Breastfeeding and the use of human milk. Pediatrics 129
(3), e827ee841.
Elliott, C.D., Smith, P., McCulloch, K., 1997. British Ability Scales (BAS II): Technical
Manual, second ed. NFER-Nelson, London, UK.
Fernald, L.C., Kariger, P., Engle, P., Raikes, A., 2009. Examining Early Child Devel-
opment in Low-income Countries. The World Bank, Washington, DC.
Fernstrom, J.D., 1999. Effects of dietary polyunsaturated fatty acids on neuronal
function. Lipids 34 (2), 161e169.
Firpo, S., 2007. Efficient semiparametric estimation of quantile treatment effects.
Econometrica 75 (1), 259e276.
Fitzsimons, E., Vera-Hernandez, M., 2012. The Causal Effects of Breastfeeding on
Children's Development. Institute for Fiscal Studies, London, UK.
Heckman, J.J., 2006. Skill formation and the economics of investing in disadvan-
taged children. Science 312 (5782), 1900e1902.
Hirano, K., Imbens, G.W., Ridder, G., 2003. Efficient estimation of average treatment
effects using the estimated propensity score. Econometrica 71 (4), 1161e1189.
Ip, S., Chung, M., Raman, G., Chew, P., Magula, N., DeVine, D., et al., 2007. Breast-
feeding and Maternal and Infant Health Outcomes in Developed Countries.
Evidence Report/technology Assessment No. 153 (Prepared by Tufts-New En-
gland Medical Center Evidence-based Practice Center, under Contract No. 290-
02-0022). AHRQ Publication No. 07-E007. Agency for Healthcare Research and
Quality, Rockville, MD.
Jiang, M., Foster, E.M., Gibson-Davis, C.M., 2011. Breastfeeding and the child
cognitive outcomes: a propensity score matching approach. Maternal Child
Health J. 15 (8), 1296e1307.
108 K. Koh / Social Science & Medicine 187 (2017) 101e108
Kling, J.R., Liebman, J.B., Katz, L.F., 2007. Experimental analysis of neighborhood
effects. Econometrica 75 (1), 83e119.
Kramer, M.S., Chalmers, B., Hodnett, E.D., Sevkovskaya, Z., Dzikovich, I., Shapiro, S.,
Collet, J.P., Vanilovich, I., Mezen, I., Ducruet, T., Shishko, G., 2001. Promotion of
breastfeeding intervention trial (PROBIT): a randomized trial in the Republic of
Belarus. J. Am. Med. Assoc. 285 (4), 413e420.
Kramer, M.S., Aboud, F., Mironova, E., Vanilovich, I., Platt, R.W., Matush, L.,
Igumnov, S., et al., 2008. Breastfeeding and child cognitive development: new
evidence from a large randomized trial. Archives General Psychiatry 65 (5),
578e584.
Melhuish, E., Belsky, J., Leyland, A.H., Barnes, J., National Evaluation of Sure Start
Research Team, 2008. Effects of fully-established Sure Start Local Programmes
on 3-year-old children and their families living in England: a quasi-
experimental observational study. Lancet 372 (9650), 1641e1647.
Nickel, N.C., 2015. Look before you leap: can we draw causal conclusions from
breastfeeding research? J. Hum. Lactation 31 (2), 209e212.
Promislow, J.H., Gladen, B.C., Sandler, D.P., 2005. Maternal recall of breastfeeding
duration by elderly women. Am. J. Epidemiol. 161 (3), 289e296.
Rees, D.I., Sabia, J.J., 2009. The effect of breast feeding on educational attainment:
evidence from sibling data. J. Hum. Cap. 3 (1), 43e72.
Rose, G., 1985. Sick individuals and sick populations. Int. J. Epidemiol. 14 (1), 32e38.
Rosenbaum, P.R., Rubin, D.B., 1983. The central role of the propensity score in
observational studies for causal effects. Biometrika 70 (1), 41e55.
Rothstein, D.S., 2012. Breastfeeding and children's early cognitive outcomes. Rev.
Econ. Statistics 95 (3), 919e931.
Zhao, Z., 2008. Sensitivity of propensity score methods to the specifications. Econ.
Lett. 98 (3), 309e319.