The Effects of Breastfeeding Exclusivity on Early Childhood Outcomes

Abstract

Objectives. We examined the relationship between breastfeeding exclusivity and duration and children’s health and cognitive outcomes at ages 2 and 4 years.

Methods. We used the Early Childhood Longitudinal Study—Birth Cohort, a nationally representative sample of 10 700 children born in the United States in 2001. Parent interviews and child assessments were conducted in measurement waves at 9 months, 2 years, 4 years, and in kindergarten, with the focus on ages 2 and 4 years. We employed propensity scores as a means of adjusting for confounding involving observed characteristics.

Results. Outcome analyses using propensity scores showed some small effects of breastfeeding on key outcomes at age 4 years but not at age 2 years. Effects appeared to be concentrated in reading and cognitive outcomes. Overall, we found no consistent evidence for dosage effects of breastfeeding exclusivity. Our sensitivity analyses revealed that a small amount of unobserved confounding could be responsible for the resulting benefits.

Conclusions. Our study revealed little or no effect of breastfeeding exclusivity and duration on key child outcomes.

In terms of its public health profile, breastfeeding has few competitors. The American Academy of Pediatrics (AAP) declares that

Pediatricians and other health care professionals should recommend human milk for all infants in whom breastfeeding is not specifically contraindicated and provide parents with complete, current information on the benefits and techniques of breastfeeding.^1(p498)

By contrast, the scientific evidence on the benefits of breastfeeding offers mixed and weak support for the relationship between breastfeeding and key child outcomes.² Researchers recognize the methodological challenges of breastfeeding research, such as confounding by measured and unmeasured variables.³ Mothers self-select breastfeeding, and the characteristics that guide this choice likely affect child outcomes. Variability in attending to potential confounding may explain why the putative effects of breastfeeding vary across studies.

Another issue in uncovering the effects of breastfeeding is defining the exposure; breastfeeding can be measured by duration, exclusivity, or volume of breast milk consumed. Breastfeeding is believed most effective when mothers do so exclusively.^3,4 Reflecting this belief, the Healthy People 2010 goals include a target for exclusive breastfeeding at 6 months of 25%.⁵ Likewise, the AAP recommends that solid foods not be introduced before a child’s 4-month birthday.^1,6 Therefore, one explanation for the mixed effects of breastfeeding is that women who breastfeed vary in how they breastfeed, and children who are exclusively breastfed may benefit more than those who are nonexclusively breastfed.

In addition, further differentiating breastfeeding beyond simply breastfed or not complicates causal inference. Whether a woman breastfeeds a child at all is a matter of one behavior: breastfeeding initiation. Exclusivity, however, involves other processes, such as discontinuing breastfeeding and initiating solid foods. In general, the more processes shaping an exposure, the more possibilities for self-selection and unobserved confounding.⁷

We considered the previous research available on the effect of breastfeeding “dose” (exclusivity and duration) on child outcomes and then examined the issue using a large nationally representative panel data set. Relative to past research, our analyses adjusted for a rich set of covariates and employed improved methodology; notably, propensity scores (PSs) and analyses that considered the sensitivity of key findings to unobserved confounding.

PREVIOUS RESEARCH

Most research on the consequences of breastfeeding for mothers and their children involved comparisons of women who did and did not breastfeed. A small body of research examined the “dose” of breastfeeding—aspects such as duration, exclusivity, and volume. The definition of dosage varied extensively across studies.⁸

Raisler et al.³ defined dosage using the proportion of breast milk in the child’s overall diet determined by monthly maternal reports. Studies that did not distinguish between exclusive breastfeeding—meaning breastfeeding with no complement of water, other milk, or other foods—and nonexclusive breastfeeding measured dosage in different magnitudes of duration. In Howie et al.,⁹ the breastfeeding dosage was at least 13 weeks; Oddy et al.¹⁰ used thresholds of 4 and 6 months, and Der et al. used number of months and quarters of duration.¹¹ Others defined dose using both exclusivity and duration, but with varying magnitudes of duration.^12–18

Reflecting these differences in the exposure and other factors, the various studies differed in whether the dosage of breastfeeding influenced cognitive outcomes. A handful of studies found an association between breastfeeding duration and exclusivity and cognitive and achievement outcomes during childhood^10,19–21 and adulthood.²² No effect was apparent in other studies.^11,23–25

The relationship between breastfeeding dosage and child health was also equivocal. Beaudry et al.²⁶ examined the effect of breastfeeding duration on respiratory and gastrointestinal illnesses before 6 months and found a protective effect of breastfeeding on respiratory illness only; however, Howie et al. found a dose–response relationship with gastrointestinal disorders.⁹ Some studies found that compared with no breastfeeding, breastfeeding exclusively and mostly breastfed infants had lower odds ratios of some health outcomes during infancy, including diarrhea, wheezing, and vomiting,^14,15 and had fewer sick baby medical visits.³ However, there were inconsistent findings for hyperlipidemia, hypertension, type 2 diabetes, and atopic disease.^2,27–29 Likewise, childhood obesity, overweight, and body mass index (BMI) were related to breastfeeding in some studies but not in others.^{17,21,30–33}

Defining dosage was only 1 of several differences across studies that perhaps explained variation in findings. Most previous research assumed ignorability—that is, no confounding from unobserved variables. However, many studies included a limited set of covariates and omitted obvious confounders (such as maternal education). Recent work showed that maternal IQ^11,34 and the quality of the home environment³⁵ were key predictors of breastfeeding. In general, the more attentive researchers were to the selection of mothers breastfeeding, the smaller the estimated effects.^31,36

Recognizing this problem, other studies built on alternative methods to circumvent confounding. Kramer et al.^27,37,38 examined a breastfeeding promotion program in Belarus where participation was randomized. Although the program fostered exclusive breastfeeding for a longer duration, and children had slightly higher (one fifth of a SD) verbal IQ scores, the key finding was compromised by a major methodological problem: those rating the children were not blinded to intervention status.^39,40 The IQ differences were not robust across observers (teachers and pediatricians). Additionally, it was questionable whether effects observed in Belarus could be generalized to the United States given the differences in nutrition.

Other studies tried different approaches, such as Mendellian randomization.^41,42 Steer et al.⁴³ found that a polymorphism moderated the apparent effect of breastfeeding. However, because breastfeeding itself is not influenced by the polymorphism, this approach did not directly address the role of unobserved confounding involving breastfeeding.⁴³

Studies differed in other ways, such as sample restrictions. In their study of asthma, Kull et al.¹⁴ excluded children who began wheezing during lactation. Clearly, such a condition reflected unobserved factors as well as breastfeeding. Conditioning on a common outcome (wheezing) created a spurious relationship between breastfeeding and the unobserved determinants of wheezing. Similarly, Raisler et al.³ excluded children who were low birth weight.

METHODS

We analyzed the Early Childhood Longitudinal Study—Birth Cohort (ECLS-B), a probability sample of 10 700 children born in the United States in 2001. The ECLS-B provides information on the individual, family, and community factors that influence children’s health and development during the first 6 years of life. These data are nationally representative with the use of the survey weights. We used these data with permission through a license agreement with the Institute for Education Sciences, US Department of Education (a flow diagram of the study using the Consolidated Standards of Reporting Trials is available as a supplement to the online version of this article at http://www.ajph.org).

The data included parent interviews and child assessments at each measurement wave, which occurred at approximately 9 months, 2 years, 4 years, and at kindergarten. We chose to focus on health and cognitive outcomes at ages 2 and 4 years for brevity. Although disparate, these outcomes were measured using developmentally appropriate instruments and were key indicators of child well-being in terms of health and cognition that were examined in previous research. Outcomes at 2 years included cognitive and motor skills θ scores, and BMI. Age 4 years outcomes included θ scores for math, reading, and fine motor skills, as well as BMI. Trained administrators conducted these assessments during visits with the child and the primary caregiver in the family’s home. The θ scores were developed by the National Center for Education Statistics using item response theory (see Najarian et al.⁴⁴ for more detail on the procedures). Note that sample sizes for both the 2- and 4-year old measurement waves were lower than the sample size of the 9-month wave (8850, 7700, and 10 700, respectively).

Definition of Exposure

Breastfeeding status was determined from parental reports collected at the 9-month measurement wave that described when the child was first fed formula, food, finger food, or milk. Children were labeled “exclusively breastfed” through the month in which the parent identified the child as being first fed one of the other foods.

Children were assigned to 1 of 9 mutually exclusive categories: never breastfed, and 8 categories representing zero to 7 or more months of being breastfed exclusively (hereafter BFE). Children who were in the zero months of BFE category included children who were breastfed, but never exclusively. Sample size prevented further differentiating the highest levels of BFE (i.e., 8 and 9 months of BFE).

Using Propensity Scores for Causal Inference

To deal with confounding, we relied on inverse probability of treatment weights (IPTW), a form of propensity score (PS) methodology. (A description of the PS methodology and more details of our IPTW procedures are available as data as a supplement to the online version of this article at http://www.ajph.org.) We briefly describe our methodology here and situate it in the broader set of methods for causal inference.

Comparisons of outcomes can be adjusted across different levels of an exposure for potential confounding in a variety of ways. Previous research generally has relied on regression by adding the potential confounders as covariates. Regression has the advantage of producing estimates of adjusted between-group differences that are the most precise if the underlying assumptions are correct. Principal among these is the functional form of the regression model. Another is the assumed completeness of the covariates, known as the “ignorability” assumption or “no unobserved confounding.” Violation of either of these conditions may bias the estimated effect of the exposure.

As an alternative to regression, we used PS adjustment. The PS represents a summary of the covariates—in particular, a sum of the covariates weighted by their potential to confound comparisons of outcomes across levels of an exposure. Like regression, PS methods assume ignorability. However, PS methods have several potential advantages over standard regression. These are primarily diagnostic and include checking covariate balance (discussed in the following).

PS can be used in a variety of ways to adjust for confounding. We preferred the IPTW because of the intuition behind it and because of its ability to accommodate multileveled exposures. When the IPTW are used as sampling weights, the weighted data represent pseudopopulations in which breastfeeding is no longer associated with the variables used to calculate the scores. Then, any analysis can be performed using statistical software that allows for sampling weights.

The IPTW is calculated as the inverse of the predicted probability of receiving the exposure the child actually received. We estimated the probability of BFE in 2 steps. We first calculated the probability of any breastfeeding; we then calculated the probability of levels of exclusivity given any breastfeeding. The first stage used a logit model for any breastfeeding. The second stage used an ordered logit with an outcome of the number of months (zero to 7 or more) of exclusive breastfeeding (i.e., children who were in the zero months of BFE category were breastfed, but never exclusively); this stage of estimation was limited to the women who breastfed at all. Each stage generated a predicted probability, and these were combined to calculate the probability of the level of observed breastfeeding. This 2-stage process allowed the determinants of any breastfeeding to differ from those for the dosage of breastfeeding among women with at least some breastfeeding.

Children who were never breastfed were weighted by the inverse probabilities generated from the first stage only, and children who were ever breastfed received a combined IPTW from both stages (n = 3500 and n = 7100, respectively), where the P (level of breastfeeding observed) = P (any breastfeeding) * P (months of exclusive breastfeeding   any breastfeeding). Both stages included covariates that predicted breastfeeding from previous research (reviewed previously and available as data as a supplement to the online version of this article at http://www.ajph.org).

Once observations are weighted using the PS, the distribution of covariates should be independent of the treatment or exposure. This property, however, is asymptotic, and in finite samples, the extent to which it is realized depends on practical issues, such as functional form involving the relationship between the covariates and the exposure. For that reason, “balance checking” is critical⁴⁵ (details are available as data as a supplement to the online version of this article at http://www.ajph.org).

Analyses of Outcomes

All analyses were IPTW-weighted ordinary least-squares regressions conducted in Stata 10 (StataCorp, College Station, Texas) using SVY commands to account for the complex survey design. In addition to the dummy variables representing the BFE levels, the child’s age in months at the assessment date and 2 covariates that failed numerous balance checks (Special Supplemental Nutrition Program for Women, Infants and Children [WIC] use and days premature) were included. Regressions did not include any other covariates used to estimate the PS. We estimated the IPTW-adjusted models for each outcome and effect sizes for each BFE coefficient. Analyses included all children in the sample for whom there were measures of our dependent variables.

Sensitivity analyses.

As noted, the key assumption on which PS methods and regression depend is that of ignorability. For that reason, we also calculated γ statistics (explained in the following) to gauge the sensitivity of our estimates to unobserved confounding. We also conducted other supplemental analyses. Previous research indicated that BFE might have different effects for certain subpopulations.^46,47 Therefore, we also estimated separate models for children who were low birth weight and for mothers who were overweight. We also performed regression analyses to provide a comparison with our PS results. In addition, we tested whether errors were correlated across outcomes within children using seemingly unrelated regression,⁴⁸ and tested for chance findings using the Benjamini–Hochberg adjustment.^49,50

RESULTS

Table 1 presents mean characteristics for the full sample, the ever-breastfed sample, and for each BFE group. As in previous research, women breastfeeding exclusively for longer durations were older, had more than 1 child, were disproportionately White, more likely to have participated in the WIC program, and had higher levels of education compared with women who had ever breastfed and women in the entire sample.

TABLE 1—

Means and Proportions of Child and Parent Characteristics by Breastfeeding Status: Early Childhood Longitudinal Study—Birth Cohort, United States, 2001

			BF Sample by Months of Exclusive Breastfeeding
Characteristics	Full Sample (n = 10 700), % or Mean ±	Any BF Sample (n = 7150)	0 (n = 1800)	1 (n = 1400)	2 (n = 900)	3 (n = 850)	4 (n = 1050)	5 (n = 503)	6 (n = 600)	≥ 7 (n = 200)
% of full samplea	100	67	17	13	8	8	10	5	6	2
Covariates
Health occupation, %	6	11	5	7	7	9	9	6	4	0
Education
< high school, % (ref)	15	11	16	15	14	11	7	8	8	5
High school, %	27	23	27	29	26	23	20	15	16	23
Some college, %	30	31	32	31	28	38	31	29	26	21
≥ college, %	28	35	25	25	32	28	42	48	50	51
Race/ethnicity
Non-Hispanic White, % (ref)	62	64	47	61	64	62	71	72	70	63
Non-Hispanic Black, %	13	9	12	11	9	12	8	5	5	10
Hispanic, %	19	21	28	22	22	20	15	16	19	21
Asian, %	3	4	5	3	2	3	4	4	4	3
Other, %	3	2	2	3	3	3	2	3	2	3
Mother’s age, y, mean ±SD	28.59 ±6.27	29.18 ±5.98	28.69 ±6.27	27.89 ±6.02	28.71 ±6.27	28.26 ±5.51	29.53 ±5.45	30.66 ±5.58	31.51 ±5.69	31.66 ±5.89
Tobacco use during pregnancy	11	8	8	14	11	9	3	5	3	2
Multiparous, %	59	57	56	56	58	54	57	57	64	71
CESD, mean ±SD	1.15 ±1.49	1.11 ±1.52	1.02 ±1.90	1.28 ±1.01	1.05 ±1.75	1.02 ±1.88	1.17 ±1.13	1.04 ±1.61	1.19 ±0.99	1.00 ±1.99
LBW, %	7	6	9	7	8	6	3	4	4	11
Alcohol use during pregnancy, %	0	0	0	0	0	0	1	0	0	1
WIC use during pregnancy, %	38	31	40	42	33	35	21	19	18	29
Urban, %	85	88	91	85	84	85	89	92	88	89
Food insecurity, %	11	9	13	9	11	10	7	8	4	14
Premature, %	12	11	13	12	11	12	9	7	9	17
Outcomes
Age 2 years
Mental Θ score, mean ±SD (range)	51.09 ±9.79 (15.1–88.9)	51.93 ±9.18 (15.8–86.1)	49.51 ±9.66 (15.1–80.3)	51.08 ±10.24 (17.7–80.1)	51.57 ±9.64 (20.5–78.9)	51.05 ±9.10 (20.3–75.7)	53.80 ±9.54 (17.4–88.1)	54.01 ±10.03 (23.6–79.1)	54.59 ±9.31 (22.3–88.8)	52.05 ±8.33 (23.1–87.4)
Motor Θ score, mean ±SD (range)	50.52 ±9.81 (2.1–97.5)	50.64 ±9.80 (2.1–81.5)	49.24 ±9.85 (2.9–81.1)	51.05 ±10.66 (4.4–97.4)	51.13 ±9.90 (11.5–80.7)	49.61 ±9.69 (13.3–78.0)	51.36 ±9.18 (10.6–84.7)	51.36 ±9.58 (10.6–79.8)	50.69 ±9.61 (12.3–85.1)	52.52 ±7.64 (17.2–71.9)
BMI, kg/m2, mean ±SD	15.65 ±7.12	15.49 ±7.17	15.28 ±7.63	15.95 ±6.43	15.15 ±8.13	15.74 ±6.75	15.73 ±6.56	15.06 ±7.43	15.43 ±6.77	14.68 ±7.35
Age 4 years
Reading Θ score, mean ±SD (range)	−0.45 ±0.74	−0.38 ±0.74 (−2.5–2.6)	−0.51 ±0.76 (−2.5–2.2)	−0.49 ±0.74 (−2.4–2.4)	−0.44 ±0.75 (−2.4–1.9)	−0.49 ±0.72 (−2.5–1.7)	−0.24 ±0.71 (−2.3–2.1)	−0.23 ±0.77 (−2.3–1.9)	−0.22 ±0.72 (−2.1–2.6)	−0.25 ±0.73 (−2.0–2.3)
Math Θ score, mean ±SD (range)	−0.42 ±0.77	−0.34 ±0.75 (−2.8–2.3)	−0.43 ±0.80 (−2.8–2.1)	−0.41 ±0.74 (−2.6–2.1)	−0.44 ±0.72 (−2.8–1.8)	−0.45 ±0.74 (−2.7–1.7)	−0.21 ±0.72 (−2.8–2.4)	−0.19 ±0.77 (−2.8–1.6)	−0.20 ±0.76 (−2.8–2.4)	−0.28 ±0.72 (−2.3–1.8)
BMI, kg/m2, mean ±SD	16.70 ±2.40	16.59 ±2.21	16.71 ±2.35	16.65 ±2.18	16.91 ±2.42	16.54 ±1.94	16.54 ±2.33	16.40 ±2.32	16.28 ±1.61	16.65 ±1.96
Fine motor Θ score,b mean ±SD	3.42 ±1.53	3.54 ±1.55	3.55 ±1.57	3.47 ±1.53	3.48 ±1.54	3.36 ±1.60	3.68 ±1.54	3.70 ±1.52	3.44 ±1.50	3.87 ±1.66

Note. BF = breastfed; BMI = body mass index; CESD = Center for Epidemiological Studies Depression Scale; LBW = low birth weight; WIC = Women, Infants, and Children program. The “Any BF Sample” group includes ever breastfed children, including children who were never exclusively breastfed. The exclusivity group “zero” only includes children who were never breastfed exclusively, but were breastfed some. Health care occupation includes health care practitioners, health technologists, and other health care support occupations (Code # 11, 12 of the Standard Occupational Classification Manual of the Office of Management and Budget).

^aObservations are rounded to the nearest 50 in compliance with data security requirements.

^bRange for fine motor score is 0–7 for the entire sample.

We calculated the IPTW for each child as previously described. (The significant variables for the logit and ordered logit models are available as data as a supplement to the online version of this article at http://www.ajph.org.)

Breastfeeding Exclusivity Results

We tested the overall relationship between breastfeeding and all outcomes using seemingly unrelated regression. We rejected the null hypothesis that BFE categories were unrelated to all outcomes (as a group; χ² = 439.5; P < .01).

Final model estimates are presented in Table 2. Table 2 includes 1 column for each outcome reporting the IPTW-adjusted estimates with rows for each level of exposure. In each model, the reference group is children who were never breastfed. Effect sizes are reported for each coefficient with the standard error; these represent the effect of the exposure measured in terms of standard deviations.

TABLE 2—

Estimates of Effect of Breastfeeding Exclusivity: Early Childhood Longitudinal Study—Birth Cohort, United States, 2001

	Age 2 Years Outcomes			Age 4 Years Outcomes
	Mental (n = 8700)	Motor (n = 8500)	BMI (n = 8700)	Reading (n = 7450)	Math (n = 7450)	BMI (n = 7650)	Fine Motor (n = 7150)
0 mo
Estimate (95% CI)	3.372 (0.363, 6.381)	0.879 (−1.598, 3.356)	0.004 (−1.003, 1.011)	0.258 (0.166, 0.350)	0.431 (0.221, 0.641)	−0.085 (−0.571, 0.401)	0.334 (0.171, 0.497)
SE	1.535*	1.264	0.514	0.047*	0.107*	0.248*	0.083*
ES	0.345	0.09	0.001	0.351	0.562	−0.035	0.218
γ	0.037	0.344	−0.251	0.226	0.289	0.165	0.112
1 mo
Estimate (95% CI)	2.008 (0.922, 3.094)	0.872 (−0.145, 1.889)	0.207 (−0.485, 0.899)	0.173 (0.095, 0.251)	0.259 (0.175, 0.343)	−0.474 (−0.770, −0.178)	0.345 (0.174, 0.516)
SE	0.554*	0.519	0.353	0.04*	0.043*	0.151*	0.087*
ES	0.205	0.089	0.029	0.235	0.338	−0.197	0.225
γ	0.094	0.193	−0.068	0.129	0.228	−0.074	0.114
2 mo
Estimate (95% CI)	2.55 (1.466, 3.634)	1.094 (−0.276, 2.464)	−0.509 (−1.420, 0.402)	0.233 (0.141, 0.325)	0.231 (0.149, 0.313)	−0.299 (−0.617, 0.019)	0.426 (0.240, 0.612)
SE	0.553*	0.699	0.465	0.047*	0.042*	0.162	0.095*
ES	0.261	0.111	−0.072	0.317	0.301	−0.125	0.278
γ	0.15	0.25	−0.201	0.192	0.194	0.008	0.156
3 mo
Estimate (95% CI)	2.182 (1.149, 3.215)	−0.281 (−1.424, 0.862)	−0.624 (−1.565, 0.317)	0.172 (0.082, 0.262)	0.231 (0.143, 0.319)	−0.496 (−0.796, −0.196)	0.184 (0.002, 0.366)
SE	0.527*	0.583	0.48	0.046*	0.045*	0.153*	0.093*
ES	0.223	−0.029	−0.088	0.234	0.301	−0.207	0.12
γ	0.117	0.089	−0.221	0.111	0.186	−0.082	0.001
4 mo
Estimate (95% CI)	3.634 (2.531, 4.737)	0.64 (−0.520, 1.800)	0.077 (−0.531, 0.685)	0.293 (0.209, 0.377)	0.375 (0.287, 0.463)	−0.548 (−0.830, −0.266)	0.488 (0.296, 0.680)
SE	0.563*	0.592	0.31	0.043*	0.045*	0.144*	0.098*
ES	0.371	0.065	0.011	0.398	0.489	−0.228	0.318
γ	0.258	0.183	−0.076	0.284	0.374	−0.111	0.193
5 mo
Estimate (95% CI)	2.732 (1.156, 4.308)	0.955 (−0.742, 2.652)	−0.467 (−1.774, 0.840)	0.256 (0.123, 0.389)	0.385 (0.234, 0.536)	−0.199 (−1.030, 0.632)	0.495 (0.295, 0.695)
SE	0.804*	0.866	0.667	0.068*	0.077*	0.424*	0.102*
ES	0.279	0.097	−0.066	0.348	0.502	−0.083	0.323
γ	0.118	0.269	−0.251	0.167	0.305	0.264	0.193
6 mo
Estimate (95% CI)	3.767 (2.450, 5.084)	0.097 (−1.287, 1.481)	−0.484 (−1.401, 0.433)	0.315 (0.199, 0.431)	0.362 (0.260, 0.464)	−0.652 (−0.962, −0.342)	0.337 (0.078, 0.596)
SE	0.672*	0.706	0.468	0.059*	0.052*	0.158*	0.132*
ES	0.385	0.01	−0.068	0.428	0.472	−0.272	0.22
γ	0.25	0.153	−0.197	0.271	0.339	−0.143	0.051
≥ 7 mo
Estimate (95% CI)	2.186 (−0.268, 4.640)	0.442 (−2.075, 2.959)	−1.792 (−4.228, 0.644)	0.408 (0.139, 0.677)	0.442 (0.228, 0.656)	−0.566 (−1.411, 0.279)	0.648 (0.180, 1.116)
SE	1.252	1.284	1.243	0.137*	0.109*	0.431	0.239*
ES	0.223	0.045	−0.252	0.555	0.577	−0.236	0.423
γ	0.473	0.301	−0.595	0.19	0.298	0.116	0.117

Note. BMI = body mass index; CI = confidence interval; ES = effect size; IPTW = inverse probability of treatment weights; OLS = ordinary least squares. BMI was defined as weight in kilograms divided by height in meters squared. Results from OLS regression using survey estimation procedures. Each column presents results for the IPTW-adjusted model for 1 outcome. The rows display results across models for each level of breastfeeding exposure. In each model, the reference group is children who were never breastfed. Effect sizes are reported for each coefficient below the standard error and should be interpreted as a standard deviation change in the outcome variable. The γ statistics are reported as effect sizes and can be interpreted as the size of the effect of the hypothetical unobserved variable. Sample sizes were lower than the baseline panel (7950). Observations are rounded to the nearest 50 in compliance with data security requirements. The exclusivity group “zero” only includes children who were never breastfed exclusively, but were breastfed some.

*P < .05.

Age Two Years Results

Mental.

BFE had positive effects at all levels of the exposure (zero to 7 or more months) with small effect sizes (range = 0.20 to –0.38). Results did not indicate a dose–response pattern of BFE on mental ability at age 2 years.

Motor.

BFE for 1 month was positively associated with motor ability, but the effect size was very small (0.09).

Body mass index.

There were no significant effects of BFE on BMI.

Age Four Years Results

Reading.

There were positive effects of BFE at all levels of the exposure. Effect sizes ranged from small to moderate (0.23–0.55), whereas longer periods of BFE produced slight increases in effect sizes.

Math.

There were positive effects of BFE at all levels of the exposure. As with reading, effect sizes ranged from small to moderate (0.30–0.58), but there is no apparent dosage pattern.

Body mass index.

BFE for 1, 4, and 6 months negatively affected BMI, with a narrow range of very small effect sizes (−0.12 to −0.27) and no clear dosage pattern.

Fine motor.

There were positive effects of BFE at all levels of the exposure. There was no clear pattern of BFE dosage, and effect sizes were small to moderate (0.12–0.42).

Dosage Results

Figure 1 displays the empirical results for 5 models. This plot includes the outcome models where we found significant effects of BFE. The reference category was children who were never breastfed. The fluctuations in the plotlines appear erratic and illustrate the absence of BFE dosage effects. The math plotline is particularly illustrative; children who were never BFE fared almost exactly as well as children who were BFE for 7 months or more. For mental scores at 2 years, children who were BFE for 7 months or more actually fared substantially worse than children who were never BFE.

FIGURE 1—

Comparative line plots of the significant estimated effects of breastfeeding exclusivity (BFE) exposures on child outcomes at age 2 and 4 years: Early Childhood Longitudinal Study—Birth cohort, United States, 2001.

Note. BMI = body mass index. Each line represents the results from 1 outcome model; each point on the line is the estimated coefficient of a BFE exposure for that outcome. The reference category is children who were never breastfed. The x-axis delineates the duration of the BFE exposure, and the y-axis describes the magnitude of the effect size. Patterns of BFE dosage in the figure can be found by following the line for an outcome from left to right as BFE duration increases. This display only includes models with significant results.

Assuming a false discovery rate of 0.10, the Benjamini–Hochberg adjustment did not eliminate any significant effects in the adjusted models. The explanation is that the significance levels were much smaller than 0.05.

Sensitivity Analyses

Regression.

The regression analyses results analogous to those presented here (their interpretations are available as data as a supplement to the online version of this article at http://www.ajph.org, and are very similar to the results in Table 2).

Potential unobserved confounding.

As noted, a key assumption on which the PS methodology rests is no unobserved confounding. This assumption involves unobserved characteristics, and whether such confounding exists or not is unknown. However, how large the confounding would have to be to explain away our findings can be gauged using a γ statistic sensitivity analysis, as described by Rosenbaum.⁵¹

We illustrate the γ statistic calculation process in Figure 2 (further details are available as data as a supplement to the online version of this article at http://www.ajph.org). Gammas were calculated for all models, converted to effect sizes, and are displayed for each BFE estimate in Table 2 in the row below its estimated effect size. The values ranged from 0.04 to 0.59. These values revealed that a modest violation of the ignorability assumption or “no unobserved confounding” could be masking the true causal effect of our exposure.

FIGURE 2—

Illustration of the γ sensitivity factor calculation procedure and meaning of the γ statistic: Early Childhood Longitudinal Study—Birth cohort, United States, 2001.

Note. Building on Rosenbaum,⁵¹ we calculated γ statistics to assess the extent to which unobserved confounding could affect our estimates. The calculation proceeded in the following steps: (1) estimate potentially confounded effect using inverse probability of treatment weights (IPTWs) at determined level of significance based on critical value. Convert estimate to standardized coefficient (B); (2) multiply the standard error of the estimate by the critical value to determine the "threshold" value of standardized breastfeeding exclusively (BFE) coefficient needed for the effect to reach significance; and (3) calculate the difference between the potentially confounded estimated effect and the threshold value for a significant effect to determine the amount (in terms of effect size) by which the point estimate may be inflated or suppressed by unobserved confounding.

Key subgroups.

The effect pattern on outcomes at age 4 years was similar in the subgroup analyses for mothers who were overweight (n = 2450) and children who were low birth weight (n = 2100). For overweight mothers, effects appeared to be concentrated in reading and math. Child BMI results indicated small effect sizes and 1 significant effect (BFE for zero months) in fine motor skills. When significant, BFE effect sizes were small, with no strong evidence of dosage effects. For children who were low birth weight, the BFE effects were concentrated in reading. There were some small effects scattered across BMI, math, and fine motor skills, but no consistent effect pattern of BFE.

DISCUSSION

Our results add to the small existing literature that examine the extent to which a woman who breastfeeds her children affects their outcomes during early childhood. We examined the effects of exclusive breastfeeding on child health and cognitive outcomes using a large, nationally representative cohort study with detailed infant feeding data. This study had 3 key strengths: development of a detailed measure of breastfeeding exposure that captured exclusivity and duration, use of IPTW to address confounding, and sensitivity analyses that gauged the degree to which our results could be explained by possible unobserved confounding.

Overall, our findings suggested some small but inconsistent effects of exclusive breastfeeding on key outcomes. We did not find any evidence of a dosage effect of BFE, and instead detected rather haphazard pattern of effects in which children who were exposed to more exclusive breastfeeding did worse or equally as well as children who were never breastfed exclusively. The literature that highlighted the potential role of exclusivity presumed that more is better, and that earlier research hid the benefits of breastfeeding that occurred only when breast milk was a child’s sole source of nutrition or when breastfeeding lasted for several months. We did not see that pattern here. However, adjusting for observed confounding clearly reduced the effects of breastfeeding on child outcomes.

Furthermore, we found inconsistent results for children ages 2 and 4 years. The former revealed virtually no relationship. The effects of BFE at age 4 years were positive, and ranged between small and medium effect sizes (0.1–0.5). The effects on age 4 years cognitive outcomes presumably should be mediated by age 2 years outcomes. One possibility is that the age 4 year measurement was better than that at age 2 years. Furthermore, having entered preschool in many instances, children at age 4 years might face added cognitive challenges, providing opportunity to manifest the benefits of breastfeeding.

The key weakness of our study was that unobserved confounding remained a potential threat to the internal validity of our findings. We conducted sensitivity analyses that revealed that unobserved confounding of modest size could be hiding the true benefits of breastfeeding. However, this confounding would have to work in a counterintuitive direction: those women who breastfed more would have to be more disadvantaged (i.e., have children with otherwise lower IQ) than would other women. The direction of this bias would be the reverse of the expected relationship and would have to exist over and above the list of extensive covariates incorporated in our analyses.

Overcoming unobserved confounding in breastfeeding, as we showed here, was a significant statistical challenge. Public health and health care professionals should take caution regarding their interpretation of the scope of known effects of breastfeeding in practice. Although research could support some of the more immediate benefits of breastfeeding in terms of nutrition and immunity, it would be misleading to imply that there are definitive long-term benefits of breastfeeding for children—especially with respect to cognition and health. The challenges to determining causality in breastfeeding research must be considered before interpreting evidence to inform infant feeding guidelines or public health interventions.