Breastfeeding and IQ: Association Is Not the Whole Answer

0 The Short Answer

Pooled across the published literature, breastfed groups score about 3.4 IQ points higher than non breastfed groups, and that gap falls to about 2.6 points among the studies that measured and adjusted for the mother's own intelligence. Those two numbers come from a 2015 systematic review by Bernardo Horta, Christian Loret de Mola and Cesar Victora in Acta Paediatrica, which combined 17 studies yielding 18 estimates. The headline figure was a mean difference of 3.44 points (95% confidence interval 2.30 to 4.58). Among the subset that controlled for maternal IQ, the pooled difference was 2.62 points (95% confidence interval 1.25 to 3.98).

The interesting part of this topic is not the size of the gap. It is what happens to the gap when one variable enters the model. Mothers who breastfeed differ from mothers who do not in ways that have nothing to do with milk: education, income, working conditions, health, support at home, and measured cognitive ability. When researchers put maternal IQ into the equation, roughly a quarter of the apparent advantage disappears. That single adjustment is a live demonstration of how confounding works, and once you have watched it happen here, you can spot the same mechanism in every child health headline you will ever read.

The second thing worth understanding immediately is scale. A difference of two or three points is a difference between two population averages. It is not a property of any child. On a scale with a mean of 100 and a standard deviation of 15, two distributions separated by 2.62 points overlap almost completely, which means that if you picked one breastfed and one formula fed child at random, the formula fed child would score higher roughly 45 times out of 100. The average moved. Almost nothing about an individual moved with it.

1 What the 2015 Meta-Analysis Actually Reports

The Horta review is the standard reference on this question, and it is worth reading precisely rather than through the headlines it generated. The authors searched Medline, LILACS, SCIELO and Web of Science, and they applied two filters that most popular coverage never mentions. They excluded studies restricted to infants, because a cognitive score taken in the first year of life is a weak predictor of anything later. They also excluded estimates that had not been adjusted for stimulation or interaction in the home, on the grounds that a household that reads to a child and a household that does not are not comparable, whatever anyone was fed. That filter is the one most childhood enrichment literatures skip: years of music lessons and test scores show the same tidy dose gradient, and a world with no causal effect at all would produce it anyway, since sustaining a decade of lessons requires the income, the parental education and the household stability that predict scores on their own.

What survived those filters was 17 studies with 18 usable estimates, pooling to a 3.44 point advantage. The confidence interval, 2.30 to 4.58, is the honest range: the data are compatible with an effect just above two points and with one approaching five. Anyone quoting 3.44 as a fixed quantity is quoting the midpoint of a range and dropping the range, which is the most common way a real result becomes a false one.

The authors then did the analysis that makes the paper useful. They isolated the studies where maternal intelligence had been measured and included as a covariate, and re-pooled just those. The estimate dropped to 2.62 points, with a wider proportional interval of 1.25 to 3.98. Notice that the lower bound moved from 2.30 down to 1.25. The adjusted analysis is not simply a smaller effect. It is a less certain one, because fewer studies bothered to measure the mother's cognitive ability in the first place.

The authors' own conclusion was that breastfeeding is associated with better performance on intelligence tests and that the pattern, including results from randomized promotion trials, is consistent with a causal contribution. That is a defensible reading of their data. It is also a much narrower claim than the one that circulated afterward, which tended to compress into a slogan about breastfeeding making children smarter. A pooled group difference of two to three points, with a confidence interval that includes values close to one point, is not a slogan. It is a small effect measured under difficult conditions.

2 What This Research Must Not Be Used For

This section sits near the top rather than at the bottom because it is the part most likely to matter to a reader who arrived here worried. Three uses of this literature are simply invalid, and each is invalid for a reason that comes from the statistics themselves, not from politeness.

• It cannot predict any individual child. Every number on this page describes a difference between group averages. Individual scores vary enormously around those averages, and the variation within each feeding group dwarfs the distance between the groups. Knowing how a child was fed tells you close to nothing about how that child will score, and it tells you nothing at all about what that child will become.
• It cannot support shame, blame or judgment. Many parents cannot breastfeed. Supply, medication, medical conditions, surgery, adoption, mental health, work that does not accommodate it, and plain personal choice all determine feeding, and none of them is a moral fact. Research findings about population averages are not a scoreboard for parents, and using them that way misreads the arithmetic as badly as it misreads the people.
• It cannot be turned into medical advice by a website. Nothing here tells anyone what to do. Infant feeding involves nutrition, health history, growth, mental health and practical constraints that no observational study captures, and those trade-offs are properly worked through with a pediatrician, midwife, lactation consultant or family doctor who knows the specific situation.

There is also a fourth misuse worth naming, because it runs the other way. Some readers, reacting against the guilt this literature has caused, want the effect to be exactly zero. That is not what the evidence shows either. The most defensible summary is a small average difference of uncertain size, partly and possibly entirely explained by everything else that travels alongside feeding. Overstating and erasing are both distortions, and the rest of this page tries to do neither.

3 What Two or Three Points Looks Like on the Scale

Most of the confusion in this topic is unit confusion. IQ points feel large because the familiar landmarks are large: 130 for gifted programs, 100 for the population median. But the relevant comparison here is not between landmarks, it is between two nearly identical bell curves shifted slightly apart. Converting the pooled findings into standard deviations, using the modern scale of mean 100 and standard deviation 15, makes the size visible.

Those overlap and probability figures are arithmetic, derived from the pooled estimate under standard normal assumptions rather than measured directly. They are worth internalizing anyway, because they are the translation step that popular coverage skips. If you would like to see how score differences map onto rank in the population generally, the IQ percentile chart shows how much of the distribution sits between any two points, and the page on average IQ explains why the mean is fixed at 100 by construction rather than discovered.

The practical upshot is uncomfortable for headline writers. An effect can be real, statistically solid, replicated across countries, and still be too small to notice in a classroom, a family or a life. Population averages and individual outcomes are different objects, and a difference that matters at the level of a national statistic can be invisible at the level of a person.

4 Maternal IQ, the Confounder That Ate a Third of the Effect

The drop from 3.44 to 2.62 is not a rounding artifact. It is the signature of a confounder, and the clearest study of that particular confounder is the 2006 analysis by Geoff Der, David Batty and Ian Deary, published in the BMJ using the 1979 US National Longitudinal Survey of Youth.

Their first finding is the one that reframes the entire literature. Maternal IQ predicted whether a mother breastfed better than her race, her education, her age, her poverty status, her smoking, the quality of the home environment, the child's birthweight or the child's birth order. A one standard deviation advantage in the mother's measured intelligence more than doubled the odds that she breastfed. Feeding method, in that dataset, was among other things a measurement of the mother.

Since maternal cognitive ability is itself associated with child cognitive ability through both inheritance and the environment mothers create, any raw comparison of fed groups is partly a comparison of families. Der and colleagues found roughly a four point unadjusted advantage in their data, and adjustment for maternal intelligence accounted for most of it. Their sibling pair analysis, discussed further below, pushed the residual toward zero, and they concluded that breastfeeding has little or no direct effect on childhood intelligence.

Two research groups looking at overlapping questions therefore reached different conclusions: Horta and colleagues read the pooled evidence as consistent with a small causal contribution, Der and colleagues read theirs as consistent with essentially none. That disagreement is not a scandal and it is not a reason to dismiss both. It is the normal state of a field where the exposure cannot be randomly assigned to individuals, and where the strength of your conclusion depends on how well your particular dataset measured the things that travel with the exposure. Understanding what IQ tests measure and what they inherit from family background is most of what you need to see why.

5 Everything Else That Travels With Feeding

Maternal IQ is the most instructive confounder, not the only one. Feeding method in most countries is bundled with a stack of advantages that no statistical adjustment fully removes.

The UK Millennium Cohort Study analysis by Reinaldo Pereyra-Elias, Maria Quigley and Claire Carson, published in PLOS ONE in 2022, followed 7,855 singletons born in 2000 to 2002 up to age 14 and tried to quantify exactly this. Their result cuts both ways with unusual clarity: adjusting for socioeconomic position alone roughly halved the association, and after adjusting for both socioeconomic position and maternal cognitive ability, a modest association still remained. Half the raw signal was social, and something survived the scrub.

The debate did not stop there. A 2024 comment in the same journal argued that the surviving association could still reflect residual confounding, because maternal cognitive ability is measured with error, and a confounder measured imprecisely is only partly controlled. Adjusting for a noisy version of a variable removes a noisy fraction of its influence. This is a general and underappreciated problem: statistical control is only as good as the measurement behind it, and "we adjusted for it" is not the same claim as "we removed it."

Beyond income and cognition, the list of plausible travelling companions is long: parental time available at home, paid leave, birth weight and gestational age, maternal health, smoking, the language environment in the house, and the education of both parents. Some studies adjust for a few of these. None adjusts for all of them well. This is why the sections that follow move away from observational adjustment entirely and look at designs that try to sidestep confounding by construction rather than by regression.

6 Sibling Designs: Same Mother, Different Feeding

If the worry is that breastfeeding families differ from formula feeding families, one elegant response is to stop comparing families. Sibling designs compare children raised by the same mother in the same household, where one sibling was breastfed and another was not. Anything stable about the family, including the mother's intelligence, education, income bracket and parenting style, is held constant automatically, without needing to be measured at all. What the design does not hold constant is how the siblings differ from each other genetically, and what heritability actually means is the guide to that portion: an estimate of how much of the variation between people in one population tracks genetic differences, never a percentage that applies to a single child.

The most widely cited application is the 2014 study by Cynthia Colen and David Ramey in Social Science and Medicine, which used 25 years of panel data from the National Longitudinal Survey of Youth to examine children aged 4 to 14 across 11 outcomes, including measures of vocabulary, reading comprehension, math ability and memory based intelligence alongside physical health indicators. In conventional between family models, breastfed children looked better on 10 of the 11 outcomes, exactly the pattern the wider literature reports. When the analysis was restricted to siblings with within family fixed effects, the estimates for all but one outcome dropped sharply and lost statistical significance. The authors concluded that much of what gets attributed to breastfeeding may reflect selection into breastfeeding rather than the feeding itself.

Der and colleagues had already reported the same pattern eight years earlier in their sibling pair analysis, which is part of why their conclusion was so flat. Two independent teams, working with the same American panel, found that the effect largely evaporates when the comparison happens inside families.

Sibling designs are not a magic solution and it would be dishonest to present them as one. Within a family, feeding decisions differ for reasons: a difficult first experience, an illness, a return to work, a change in circumstances between births. Those reasons can themselves affect the child, so discordant siblings are not a randomized pair. Sample sizes also shrink dramatically, which widens confidence intervals and makes a real small effect harder to detect. The correct reading is not that sibling studies prove the effect is zero. It is that the effect is substantially smaller inside families than between them, which tells you a great deal about where the between family difference was coming from.

7 PROBIT: The Closest Thing to an Experiment

You cannot randomly assign babies to be breastfed. You can randomly assign hospitals to promote breastfeeding, and that is what the Promotion of Breastfeeding Intervention Trial did. Led by Michael Kramer, PROBIT enrolled 17,046 healthy breastfeeding mother and infant pairs in Belarus between June 1996 and December 1997, and randomized the maternity hospitals and their affiliated clinics to either a support program modeled on the WHO and UNICEF Baby-Friendly Hospital Initiative or to continued usual practice.

The intervention worked as an intervention. Exclusive breastfeeding at three months ran at 43.3 percent in the intervention arm against 6.4 percent in the control arm, and any breastfeeding at twelve months at 19.7 percent against 11.4 percent. That contrast is what makes the trial informative: two large groups, randomly formed, that ended up feeding very differently.

At the 6.5 year follow-up, reported by Kramer and colleagues in Archives of General Psychiatry in 2008 with 13,889 children assessed on the Wechsler Abbreviated Scale of Intelligence, the intervention arm scored higher, with cluster adjusted mean differences of 7.5 points on verbal IQ (95% confidence interval 0.8 to 14.3), 2.9 points on performance IQ (confidence interval crossing zero, from minus 3.3 to 9.1) and 5.9 points on full scale IQ (also crossing zero, from minus 1.0 to 12.8). Teachers rated the intervention children higher in reading and writing.

Read that carefully, because the pattern is more informative than the headline number. Only the verbal difference reached significance, and its confidence interval spans nearly fourteen points, so the trial is compatible with a verbal advantage of about one point and with one of fourteen. The distinction between fluid and crystallized ability is directly relevant here: the domain that moved was the language loaded one, the domain most sensitive to verbal environment and interaction, while performance IQ barely budged. Critics also noted that examiners and pediatricians were not blind to group assignment in a country where the intervention was locally visible, which is a real limitation on the verbal measures in particular.

8 PROBIT at Sixteen Years: What Survived

Trials that stop at age 6 leave the most important question open, because early cognitive differences frequently fade. PROBIT followed its cohort into adolescence, and the 16 year results, published by Seungmi Yang and colleagues in PLOS Medicine in 2018, are the most valuable data point on this entire page. Reading them means holding the domains apart rather than averaging them, because the language loaded scales and the figural ones are not substitutes for each other: tests built around matrices and shapes mostly sample fluid reasoning and visual spatial processing, and a score from one is neither culture free nor a stand in for a full battery.

Of the original 17,046 participants, 13,427 were assessed, a retention rate of 79.5 percent that is remarkable for a two decade study. Cognition was measured with a computerized NeuroTrax battery covering memory, executive function, visual spatial perception, verbal function, attention, information processing and fine motor skills, plus a global score.

There was no significant difference in overall neurocognitive function between the arms. Verbal function was 1.4 points higher in the intervention group by intention to treat (95% confidence interval 0.3 to 2.5), memory was 1.2 points higher (0.01 to 2.4, a lower bound essentially touching zero), and an instrumental variable analysis estimating the effect of exclusive breastfeeding for at least three months produced a 3.5 point verbal advantage (0.9 to 6.1). The remaining domains showed nothing statistically distinguishable from zero. The authors concluded that the long term effects on neurocognitive development shrink with age and are, by adolescence, limited to verbal function.

Compare the two follow-ups and the arc is clear. At 6.5 years, a verbal difference of 7.5 points with a wide interval. At 16 years, a verbal difference of 1.4 points and a global difference indistinguishable from zero. The best randomized evidence available says the effect is domain specific rather than general, and that it attenuates as children accumulate a decade of schooling, reading and everything else that shapes measured ability. This is the pattern early intervention research keeps finding across many different exposures, and it belongs in any honest summary of breastfeeding and cognition.

9 Pelotas at Thirty: The Longest Look

The other landmark study runs longer than PROBIT and answers a different question. The 1982 Pelotas birth cohort in southern Brazil enrolled 5,914 newborns and has tracked them ever since. Cesar Victora and colleagues reported the 30 year assessment in Lancet Global Health in 2015, with data on both IQ and breastfeeding duration available for 3,493 adults tested between June 2012 and February 2013. Cohorts of that length are what built cognitive epidemiology, the field where a roughly 15 point advantage in childhood score is associated with something like a 20 to 24 percent lower risk of death over decades of follow up, most of it running through health behavior, education and circumstances rather than through the number.

Participants breastfed for 12 months or more scored 3.76 points higher on IQ than those breastfed for less than one month (95% confidence interval 2.20 to 5.33). They had completed 0.91 more years of education (0.42 to 1.40) and earned 341 Brazilian reais more per month (93.8 to 588.3). A mediation analysis suggested that IQ accounted for 72 percent of the income difference, which is a striking claim about the pathway even if mediation analysis carries assumptions worth treating cautiously.

Two features of Pelotas make it more than another cohort. First, it is a dose comparison rather than a binary one, contrasting long duration against almost none, which is the contrast where an effect should be largest if one exists. Second, and this is the part that gives the study unusual leverage, breastfeeding in Brazil in 1982 was not concentrated among the wealthy and educated in the way it is in the United States or the United Kingdom. It was distributed fairly evenly across social classes. The confounding structure that dominates Anglophone datasets was much weaker in this population, so a persistent association there is harder to explain away with socioeconomic position alone.

The findings are still observational, and they still carry the standard caveats: 30 year recall of feeding duration is imperfect, adult IQ at 30 was measured once, and a cohort cannot randomize anything. But a result that holds in a population where the usual confounders point in different directions is stronger evidence than the same result in a population where feeding tracks straight along class lines. Pelotas is the main reason serious researchers have not concluded that the entire association is social sorting.

10 Why the Biological Story Is Plausible and Still Not Decisive

Behind the epidemiology sits a mechanistic hypothesis that has driven this field for decades: human milk contains long chain polyunsaturated fatty acids, including docosahexaenoic acid, which are structural components of neural tissue during a period of rapid brain growth. If those compounds are limiting in early development, supplying them could plausibly affect later cognition. The hypothesis is coherent and is the reason the topic was studied so intensively in the first place.

Coherence is not confirmation, and the mechanistic story has two problems that the statistics cannot resolve. The first is that formula composition has changed enormously across the decades the pooled studies span. A child born in 1975 who was fed formula received a very different product from one born in 2015, after manufacturers began adding these fatty acids routinely. When a meta-analysis combines studies across that span, the comparison group is not a fixed thing, and effects estimated on older cohorts may not describe current ones.

The second is that breastfeeding is not only a nutrient delivery mechanism. It involves physical closeness, feeding on demand, a particular rhythm of interaction, and a set of behaviors around the infant that differ in ways nobody has fully isolated. If the causal pathway runs partly through interaction rather than composition, then the exposure being measured is not exactly the exposure being theorized. The domain pattern in the PROBIT follow-ups, where verbal function outlasted everything else, is at least suggestive on this point, since language is the domain most sensitive to interaction and least explained by fatty acid supply.

None of this settles the mechanism. It does explain why a field with a plausible biological story, decades of data, one large randomized promotion trial and a 30 year cohort has still not converged on a confident number. When the exposure is a bundle, the comparison group is a moving target and the confounders are the strongest predictors in the dataset, small honest effects are the most any method can deliver.

11 How to Read the Next Headline About Infant Feeding and IQ

The point of walking through this literature in detail is transferable. Here is the checklist that this topic teaches, applicable to almost any claim about a childhood exposure and later ability. The same reading habits defuse the height and IQ correlation, a real but tiny association that says nothing about any person in front of you.

• Ask what the effect size is in the original units. "Significantly higher IQ" tells you a p value survived. It does not tell you whether the gap is 2 points or 20. Demand the number, then convert it to standard deviations by dividing by 15.
• Ask for the confidence interval, not the point estimate. A verbal difference reported as 7.5 points reads very differently once you see the range runs from 0.8 to 14.3. The interval is the finding; the midpoint is a summary of it.
• Ask which confounders were measured and how well. The specific question that matters here is whether maternal cognitive ability was measured at all, and with what instrument. Studies that skip it will report larger effects, reliably.
• Ask whether the design compares families or children within a family. Between family comparisons contain everything that distinguishes those families. Sibling and randomized designs remove much of it, and their estimates are consistently smaller.
• Ask how long the follow-up ran. Effects measured at age 5 and effects measured at age 16 answer different questions, and in this literature the older measurement is much less impressed than the younger one.
• Ask whether the finding is general or domain specific. An advantage confined to verbal function is a different scientific claim from a general increase in ability, even though both get written up as "higher IQ."

Run those six questions on the breastfeeding literature and you get the summary this page has been building toward: a small average difference, largest in unadjusted between family comparisons, smaller with maternal IQ in the model, smaller again inside families, concentrated in verbal domains, and fading with age. Every step of that shrinkage is informative, and none of it makes the underlying research bad. It makes it normal. Careful research on hard questions usually looks like this, and readers who understand what a given IQ score actually signifies are much harder to mislead with a two point gap.

12 Verdict: A Real, Small, Shrinking Effect

Stated for the record, the defensible position on breastfeeding and IQ has four parts. There is a consistent association across many countries and decades, with the best pooled estimate around 3.44 points unadjusted and 2.62 points once maternal intelligence is accounted for. That association shrinks further under stronger designs, dropping toward non significance in sibling comparisons and settling at roughly a point and a half on verbal function by adolescence in the only large randomized promotion trial. Some effect probably survives all of the scrubbing, with Pelotas offering the strongest reason to think so. And the surviving effect is small enough that it describes populations rather than people.

Anyone who tells you this literature proves breastfeeding raises intelligence is overreading it. Anyone who tells you the entire association is social class is underreading it, and ignoring a cohort where social class did not track feeding. Sitting with a modest, uncertain, partly confounded effect is less satisfying than either slogan, and it is what the evidence actually supports.

The wider lesson may be more useful than the specific finding. Adult cognitive ability is the accumulated product of genetics, health, schooling, language exposure, opportunity and a long list of things that happened over decades. Searching for the single early input that sets it is a natural instinct and a poor model of how development works. Small effects from many sources, most of them entangled with each other, is the honest picture, and it is why any single factor, including this one, moves the average by a couple of points at most. If you are interested in how measured ability relates to outcomes later in life, the page on IQ and job performance covers effect sizes with the same attention to what the numbers can and cannot support.

If reading this made you curious about the structure of cognitive ability rather than its origins, the measurable part is your own current profile across domains. A structured battery reports separate results for verbal comprehension, fluid reasoning, visual spatial processing, working memory, processing speed and quantitative reasoning, each with an interval around it, which is a far more informative object than a single number about anyone.

13 Frequently Asked Questions

14 Sources Behind This Page

The relationship discussed here comes from published research, and the honest reading includes its limits. These are the primary sources behind the numbers on this page.

• World Health Organization. The WHO topic page on breastfeeding, the public health baseline behind the debate.
• Plomin, R. & Deary, I.J. (2015). Genetics and intelligence differences: five special findings. Molecular Psychiatry. Open access. The standard modern review of what twin and DNA evidence does and does not show.
• Nisbett, R.E. et al. (2012). Intelligence: new findings and theoretical developments. American Psychologist, 67(2). The broad APA review of what moves measured intelligence and what does not.
• Schmidt, F.L. & Hunter, J.E. (1998). The validity and utility of selection methods in personnel psychology. Psychological Bulletin, 124(2). The meta-analysis behind general ability as the strongest single predictor of job performance.
• Strenze, T. (2007). Intelligence and socioeconomic success: a meta-analytic review of longitudinal research. Intelligence, 35(5).
• Deary, I.J., Strand, S., Smith, P. & Fernandes, C. (2007). Intelligence and educational achievement. Intelligence, 35(1). A five year cohort of over 70,000 English schoolchildren.
• Kuncel, N.R., Hezlett, S.A. & Ones, D.S. (2004). Academic performance, career potential, creativity, and job performance: can one construct predict them all? Journal of Personality and Social Psychology, 86(1).
• Whalley, L.J. & Deary, I.J. (2001). Longitudinal cohort study of childhood IQ and survival up to age 76. BMJ, 322(7280). Open access.
• Batty, G.D. et al. (2007). IQ and mortality: a population based cohort study. International Journal of Epidemiology, 36(3), 538-553.
• Zagorsky, J.L. (2007). Do you have to be smart to be rich? The impact of IQ on wealth, income and financial distress. Intelligence, 35(5).