shows descriptive statistics for the potential confounders, broken down by breastfeeding status. The association of each of them with breastfeeding status is shown as an odds ratio. Children who were breast fed had mothers with higher IQ and with more education and who were older, less likely to be in poverty or to smoke, and more likely to provide a more stimulating and supportive home environment. Hispanic mothers were less likely to breast feed their children and black mothers much less likely. The children who were breast fed were likely to be heavier at birth and earlier in birth order, although this could be a reflection of family size. There was no difference in gestation or the proportion of male infants who were breast fed.
Association of potential confounders with breast feeding
A one standard deviation advantage in mother's IQ more than doubled of the odds of breast feeding. Mother's education had a similar but slightly weaker effect. As gestation and sex were not significantly associated with breast feeding, they were not included in the remaining analyses. For the analyses of PIAT scores we had full data on 16 744 assessments of 5475 children born to 3161 mothers.
shows the effect of breast feeding on the cognitive outcomes, both unadjusted and adjusted individually for each of the confounders. Each row gives the effect of breast feeding on the outcome as a regression coefficient (B) with its standard error and P value, adjusted for the confounder listed. The attenuation of the effect size is shown as a percentage of the unadjusted figure. The unadjusted effects of breast feeding correspond to an advantage for those breast fed of 4.1 to 4.7. These are comparable with effects reported in other studies.5
Adjustment for mother's IQ reduced this advantage by 71% to 75%, and adjustment for mother's education by 34% to 42%. Family poverty, maternal race, maternal age, and HOME stimulation score were important confounders.
Effect of breast feeding on cognitive outcomes, unadjusted and adjusted singly for each confounder, in 3161 mothers, 5475 children, and 16 744 assessments. All significant at P<0.001 except where marked
shows the results of including all the confounders simultaneously. In this case, each row of the table gives the effect of the confounder on the outcome, adjusted for all the others and for the child's age at assessment. The fully adjusted effect of breast feeding averages slightly less than half a point with a range of 0.36 to 0.52. These are small effects, and none is significant even with the large numbers in these analyses. Maternal IQ has the largest independent effect, with maternal education, age, family poverty, HOME stimulation score, and birth order all making independent contributions for most outcomes.
Mutually adjusted effects of breast feeding and confounders on cognitive outcomes in 3161 mothers, 5475 children, and 16 744 assessments
To simulate a study without adjustment for maternal intelligence, but which was otherwise well controlled, we re-ran the models in omitting maternal IQ. The effects of breast feeding were then at least double those shown in and all were significant (). Although these effects were still small, they do show that omitting maternal intelligence can seriously overestimate the effect of breast feeding.
Supplementary analysis: omitting mother's IQ, and duration of breast feeding (only in those breast fed)
Because observational studies use dose-response relation to infer causality26
we repeated the analysis including data only on those who were breast fed and introducing duration of breast feeding in months (defined as four weeks) to the model (). Although the effects for reading comprehension and PIAT total score were significant, all effects were small considering that the median duration of breast feeding is three months and the 95th centile is 14 months. We repeated the analysis dividing those breast fed into quarters of duration of breast feeding and contrasted each of these with those not breast fed (). The results broadly concur with the previous analysis. There were significant differences for the PIAT total score and the maths and reading comprehension subscales, but only for those in the top quarter of duration of breast feeding. There were also some signs of a departure from linearity in the second quarter.
Sibling pairs analysis
There were 332 sibling pairs discordant for breastfeeding status and 545 discordant for duration of breast feeding. shows the mean differences between both groups of siblings for PIAT scores. None of them was significantly different from zero. In both cases, we also tested for any differences between the sibling pairs in the confounding factors used in the earlier analysis. None was significant (results not shown), suggesting that the method controls for all those factors and there is no need for further analysis. In short, the sibling pairs analysis concurs with the earlier analysis in offering no support for a cognitive advantage of breast feeding once confounding is taken into account.
Supplementary analysis—sibling pairs
Our search yielded 431 references, 73 of which reported original data for full term infants. In 19 reports the analysis explicitly controlled for parental IQ and in one further report10
it implicitly controlled for parental IQ via sibling comparisons. The 19 reports used data from 12 studies because studies that followed children prospectively and assessed them at different ages often repeated the analysis and published the results separately. When the published results did not present the data in a comparable format, we approached the corresponding author directly.
The figure summarises eight of the 12 studies,27-34
together with the results for PIAT total score from our study. We excluded four studies because the data were no longer available in a suitable form,35
“nearly 100% had some breast feeding”36
(T Vik, personal communication), or we could not contact the authors.37,38
When a study assessed children at more than one age we chose the oldest age available as childhood measures of cognitive ability increase in reliability with age.39
When more than one measure of ability was assessed we took the more general measure.
The studies with the biggest IQ advantage for breast feeding were those with smaller sample sizes and that controlled for fewer of the important additional confounders—namely, socioeconomic status, maternal education, maternal age, and HOME score. The study with the largest effect had a sample that was socially heterogeneous (D Hay, personal communication) and the next largest one that was “disadvantaged, consisting almost exclusively of members of social classes IV and V.”32
Two studies assessed the children at only 2 years of age when the available measures are “best characterized as indices of mental development, and not IQ.”8
The greatest support for an effect of breast feeding came from the study by Fergusson et al,27
which was based on the previous largest sample and from a birth cohort with an outstanding follow-up rate. It did not, however, control for maternal age, smoking, or the HOME score, and its measure of maternal education contained only the three categories: secondary education only; secondary school qualification; tertiary qualification. The study therefore may not have been adequately controlled, as the authors themselves acknowledge in their conclusion.
In the figure we have ordered the studies by decreasing standard error, and the asymmetric pattern of the effect estimates suggests publication bias. There is significant heterogeneity in their estimates of the effect of breast feeding (Q = 21.07, df = 8, P = 0.007), which is unsurprising given the minimal criteria for inclusion, the differences between the samples studied, and in particular the different numbers of additional covariates controlled for. This is consistent with the results from the longitudinal survey: maternal IQ explains most of the effect of breast feeding, but not all; other important confounders need to be taken into account. To adjust for this difference between studies we used the number of additional confounders in each study as a predictor of the effect of breast feeding in a meta-regression. Because this varies between studies but not within them it is a good candidate for meta-regression.40
Although a count is a somewhat simplistic summary measure, it did account for the heterogeneity (model Q = 7.41, df = 1, P = 0.006; residual Q = 13.66, df = 7, P = 0.058). The estimates (standard errors) were: intercept 3.37 (0.84); slope -0.40 (0.15), which equates to an IQ advantage of 3.37 points in a study controlling for IQ but no other confounders, and 0.16 of a point for a study with IQ and all eight additional confounders. With full control for covariates there is effectively no advantage to breast feeding. This result is not entirely due to the high weighting that this study has in the meta-regression; an unweighted meta-regression yielded an estimated advantage of breast feeding of -0.39 for a fully controlled study (that is, a slight disadvantage).
We analysed the results from the only other study to have used sibling comparisons10
separately and with the corresponding results from this study. (We used the method of Evenhouse and Reilly10
to obtain separate estimates for strict comparability.) The estimated standardised difference in means was 0.025 (SE = 0.041, P = 0.540) for breastfeeding status and 0.040 (0.036, P = 0.271) for duration of breast feeding. Thus, the evidence from the only two sibling pair studies to date, when taken together, offer no support for an advantage of breast feeding.