BACKGROUND AND OBJECTIVES: There is mixed evidence from correlational studies that breastfeeding impacts children’s development. Propensity score matching with large samples can be an effective tool to remove potential bias from observed confounders in correlational studies. The aim of this study was to investigate the impact of breastfeeding on children’s cognitive and noncognitive development at 3 and 5 years of age.
METHODS: Participants included ∼8000 families from the Growing Up in Ireland longitudinal infant cohort, who were identified from the Child Benefit Register and randomly selected to participate. Parent and teacher reports and standardized assessments were used to collect information on children’s problem behaviors, expressive vocabulary, and cognitive abilities at age 3 and 5 years. Breastfeeding information was collected via maternal report. Propensity score matching was used to compare the average treatment effects on those who were breastfed.
RESULTS: Before matching, breastfeeding was associated with better development on almost every outcome. After matching and adjustment for multiple testing, only 1 of the 13 outcomes remained statistically significant: children’s hyperactivity (difference score, –0.84; 95% confidence interval, –1.33 to –0.35) at age 3 years for children who were breastfed for at least 6 months. No statistically significant differences were observed postmatching on any outcome at age 5 years.
CONCLUSIONS: Although 1 positive benefit of breastfeeding was found by using propensity score matching, the effect size was modest in practical terms. No support was found for statistically significant gains at age 5 years, suggesting that the earlier observed benefit from breastfeeding may not be maintained once children enter school.
- DHA —
- PSM —
- propensity score matching
- SDQ —
- strengths and difficulties questionnaire
- SEM —
- structural equation modeling
What’s Known on This Subject:
The medical benefits of breastfeeding for mother and child are considered numerous, yet the effect of breastfeeding on cognitive abilities remains largely debated given selection into breastfeeding. The effect on behavior is even less well understood.
What This Study Adds:
In applying quasi-experimental techniques which mimic random assignment, this study supports limited positive impacts of breastfeeding for children’s cognitive and noncognitive development. Although significant, the effect of breastfeeding on noncognitive development is small in practical terms.
The medical benefits of breastfeeding for both mother and child are considered numerous and well documented.1–5 Yet the effect of breastfeeding on general cognitive abilities has been a topic of debate for nearly a century.6 The mechanism argued to be responsible for these effects is the nutrients found in breast milk.7,8 Two specific types of long-chain polyunsaturated fatty acids, namely docosahexaenoic (DHA) and arachidonic acid, have been implicated in both visual and neural development and functioning through neural maturation, which is important for cognitive abilities, such as problem solving.9–11
The link with nutrients may also impact specific cognitive abilities like language development. For example, language abilities, such as vocabulary, are highly dependent on working and long-term memory given the consolidation and retrieval processes needed during acquisition.12,13 In rats, deficiency of fatty acids, such as DHA, during lactation resulted in poor memory retention during learning tasks, whereas supplementation of DHA had reversal effects.14 If the hypothesized “causal” mechanism of superior nutrition in breast milk is true, coupled with the specific impact of DHA on memory, breastfeeding should also impact language abilities. To date, ∼20 studies have investigated this association and all but 115 examined a combined measure of language (receptive and expressive) or receptive language only. There remains debate as to whether expressive and receptive language in early childhood form distinct modalities of language,16,17 raising the question of whether breastfeeding would be equally beneficial to each modality in the case of a 2-factor language model.
Less studied is the impact of breastfeeding on behavior. Breastfeeding may lead to reduced behavioral problems as a result of early skin-to-skin contact, which helps form a secure mother-infant bond.18 Any effects of breastfeeding on cognitive and language development could also prevent the development of behavior problems. The absence of early behavior problems has social, economic, and medical value to society through reduced prevalence of delinquency, incarceration rates, and substance abuse,19–21 making this an important area of research. With few exceptions, there remains a dearth of high-quality studies examining behavior,22–25 and among them, consensus is not evident.
Without randomization of mothers to breastfeeding and formula conditions, it is challenging to confirm the causal impact of these hypotheses. One study randomized the provision of a breastfeeding intervention, modeled on the Baby-Friendly Hospital Initiative, and found that the children of mothers in the intervention group had higher intelligence scores compared with controls at age 6 years.26 The strongest effects were for verbal intelligence. This study offers the best support to date for a causal link between breastfeeding and cognitive development. However, it is the only cluster randomized trial on human lactation.
The majority of studies in this field are observational, thus the causal implications of breastfeeding are questionable given the inherent difficulty in controlling for selection into breastfeeding. For example, initial associations with cognitive development are often reduced after adjustments for confounders, such as parental education/IQ (ie, from an average 5-point to 3-point difference27), and, in some cases, the associations are no longer statistically significant.28 A variety of observational studies now apply quasi-experimental methods to better address the issue of selection bias, making inroads toward a better understanding of potential causal paths. The techniques used include propensity score matching (PSM), instrument variables, and sibling pair models. This study uses PSM because the sibling pair model limits the available pool of participants and instrument variables are extremely sensitive to the validity of the chosen instrumentation, which should be associated with the exposure but not with the outcome except for via the exposure.
Using a large longitudinal population sample, we applied PSM, which mimics random assignment, in an effort to investigate the potential impacts of breastfeeding on children’s cognitive ability, expressive vocabulary, and behavior problems. Both breastfeeding duration and intensity were examined. Significant advantages for children who were breastfed, after matching, were expected for all outcomes. Grounded in the recommendations of the World Health Organization,29 it was expected that larger effect sizes would be observed for children who were fully breastfed and for longer durations.
Participants included families enrolled in the Growing Up in Ireland infant cohort. Families with infants born between December 2007 and May 2008 were identified from the Child Benefit Register and randomly selected to participate. The overall recruitment response rate was 65% (N = 11 134). A detailed description of the study design can be found elsewhere.30 We used data collected at 9 months and 3 and 5 years of age. Only families with complete data for all confounders when children were 9 months and children who were born full term were included (N = 9854; 88.5% of the initial sample). Boys represented 50.6% (N = 4991) of the sample. Attrition across waves reduced the sample size to 8715 children at 3 years and 8032 at 5 years. Some children had missing data on the cognitive and vocabulary scales, resulting in 8535 and 8241 children respectively at age 3 and 7972 and 7942 children respectively at age 5. Additionally, missing teacher reports for behavior at age 5 years resulted in 7478 children being included in these analyses. Demographic characteristics of the families and rates of breastfeeding engagement can be found in Table 1 and Fig 1. Ethics approval was obtained from the Research Ethics Committee, Department of Children and Youth Affairs Ireland, and written consent was collected from parents/guardians before data collection.
Children’s cognitive abilities and expressive vocabulary were measured by using 2 scales from the British Abilities Scale31. The pictures similarities scale assessed problem-solving skills and the naming vocabulary scale assessed expressive vocabulary. The construct validity of each scale was derived by using the Wechsler Preschool and Primary Scale of Intelligence-Revised (r = 0.74 and 0.83, respectively).31 Standardized scores that adjusted for performance as compared with other children of the same age, with a mean of 50 and a SD of 10, were used. Age was adjusted in 3-month age bands.
The Strengths and Difficulties Questionnaire (SDQ32) was used to assess children’s problem behaviors. The parent version was used at age 3 years and both the parent and teacher versions were used at age 5 years. The SDQ is comprised of 5 scales (emotional symptoms, conduct problems, hyperactivity/inattention, peer relationship problems, and prosocial behavior) with ratings of applicability of behaviors on a 3-point scale. A total difficulties scale is included, combining the 4 problem scales, to yield an overall difficulties score. We used the conduct problems, hyperactivity/inattention, and difficulties scales given our focus on externalizing problems. Validation of the SDQ has been extensively documented.33Table 2 reports the correlations between parent and teacher SDQ reports and the means and SDs for all child outcomes.
Breastfeeding information was collected retrospectively when infants were 9 months old via maternal report. Support for the reliability of recall in previous breastfeeding studies has been established.34 However, given the lower reliability regarding the timing of the introduction of additional fluids/solids, Labbok and Krasovec’s definition of full (ie, exclusive or almost exclusive) and partial breastfeeding are used.35 Two breastfeeding variables were created to assess whether the infant was fully or partially breastfed and the duration of each. Mothers were asked 4 questions: “Was <baby> ever breastfed,” “How old was <baby> when he/she completely stopped being breastfed,” “Was <baby> ever exclusively breastfed,” and “How old was <baby> when he/she completely stopped being exclusively breastfed?” First, infants were grouped by breastfeeding status, both full and partial (5940) and never breastfed (3914). Of those who had ever been breastfed, 4795 had full breastfeeding at some point. Next, breastfeeding duration was grouped into 3 intervals; breastfed up to 31 days, 32 to 180 days, and ≥181 days. Each category of duration was treated as mutually exclusive, dummy coded, and compared against infants who had never been breastfed for the purpose of matching.
Confounders have been suggested in part to account for the associations found between breastfeeding and child outcomes. We matched groups (breastfed, never breastfed) on 14 of the most pertinent factors. At the child level, factors included sex (boy/girl), birth weight (≥2500 g), and having neonatal intensive care (yes/no). At the maternal level, factors included age (≤24 years, 25–29 years, 30–34 years, or ≥35 years), highest level of education (primary level/no education, second level, or third level), working status before pregnancy (yes/no), ethnicity (Irish, any other white background, African or any other black background, Asian background, or other, including mixed background), depression (a score of ≥11 on the Center for Epidemiologic Studies Depression Scale), and type of delivery (vaginal or caesarean). Family-level factors included having a partner in the residence (yes/no), social class (professional/managerial, other nonmanual/skilled manual, or semiskilled/unskilled), medical card status (free medical care, free general practitioner care, or no free medical care), total number of household members who smoked during the pregnancy (none, or ≥1), and whether the cohort infant had siblings living in the household.
PSM reduces selection bias by matching children who were breastfed to children who were not, but who had a similar probability of being breastfed based on their measured characteristics. We used PSM logit models with nearest neighbor 1:1 matching techniques. In nearest neighbor matching, the sample is randomly ordered with matching occurring sequentially between the treatment (breastfed) and control (not breastfed) group based on participants’ propensity scores. Typically, the pair is then removed from the list and the next match is created. To ensure optimal matches, we imposed a caliper so that pairs could only be matched if the propensity score was within a tenth of a SD of the other. We also allowed matching with replacement given the low rates of longer durations and full breastfeeding in this cohort. Although matching with replacement has been argued to increase variance in the data, it also arguably reduces bias in the sample by ensuring better quality of matches.36 Balance checks in all models revealed substantial reductions of bias between matched groups on all individual confounders (ie, 0%–13.9% remaining bias in partial breastfeeding models, 0%–18.1% remaining bias in full models; data available on request). The remaining overall mean bias across models ranged from 3.2% to 8.5%. The ≤20% remaining bias has been suggested as the acceptable cutoff after matching.37 Thus, we concluded that the analytic matching technique resulted in good matches between conditions. Matching resulted in all participants falling within the area of common support. The average treatment effect on those who were treated (ie, children who were breastfed) is reported. Adjustments were made for multiple hypothesis testing by using the Holmes-Bonferroni method. All statistical analyses for PSM were conducted by using Stata version 13 software (Stata Corp, College Station, TX).
To note, although PSM is advantageous in mimicking random assignment, a drawback is the challenge in evaluating a linear dose-response association, which has previously been found. Structural equation modeling (SEM) offers an alternative approach to examining this dose-response association. Additionally, SEM uses the full sample and has greater power. Thus, the data were also modeled by using SEM, where confounders were treated as correlated exogenous variables, the duration of breastfeeding was treated as a continuous mediating variable, and child outcomes were treated as correlated, which could be influenced by both breastfeeding and confounders. These results can be found in the Supplemental Material.