Everyone wants their kids to be smarter. What if there was something you could do—something 100% natural—in the first months of their life that would boost their intelligence, setting them up to do better in school, and work, and life in general?
Obviously I’m not talking about playing them Mozart piano sonatas. I’m talking about breastfeeding, the effects of which have been the subject of ultra-contentious debate for decades. Last month my friend Tom Chivers wrote a post on the benefits—or otherwise—of breastfeeding. It was sparked by the baby formula shortages in the US, and the fact that some people downplaying the shortage were arguing that mothers should “just breastfeed instead” – and heavily implying that this would be better for their children.
Tom is very sceptical about the effects of breastfeeding on any important outcome – including its effect on IQ scores. I’m quoted in his piece on that point, since I suppose cognitive testing and so on is my specialist subject. Here, I thought it might be interesting to go into more detail. Much more detail.
So, here we go. This is an epic-length post including everything I know—and everything you should know—about breastfeeding and intelligence, as measured by IQ-style cognitive tests. Even if you aren’t interested in the debate around breastfeeding per se, there’s still plenty to get stuck into: this question has all sorts of weighty methodological issues, like observational versus experimental research; confounding; blinding; meta-analysis; inconsistent and ambiguous studies; hype and overstatement; and much more.
Those methodological issues are where we’ll begin.
Why this is really hard to study
The vast, vast majority of breastfeeding-and-IQ studies are observational. That is, they collect data on whether, and how long, mothers breastfed their kids, and correlate this with how well the kid does on an IQ test. This can either be a prospective study (they gather data on babies who are currently being breastfed, then wait a few years to test their abilities when they’re, say, 5 or 10 years old), or a retrospective one (they test kids’ IQs and then ask their mothers to recall what they did about breastfeeding when their kid was a baby). There’s no intervention here: the scientists aren’t artificially trying to increase rates of breastfeeding or anything like that – they’re just letting people do what they’d do naturally, and seeing what happens.
The problem with correlational research like this is that it’s subject to confounding: even though you might find an association between breastfeeding and IQ, it could be that the former doesn’t cause the latter, but that they’re both caused by some third factor. Some possible third factors are:
Socioeconomic status. In the developed world at least, wealthier, better-off mothers are more likely to breastfeed their kids, and are more likely to do so for longer. They’re also more likely to have kids with higher IQs, perhaps because they can provide them with a better-quality environment as they grow up (better nutrition, a neighborhood with less pollution, maybe more opportunities to develop and practice their cognitive abilities – that kind of thing).
Parental intelligence. Smarter mothers are more likely to breastfeed, and do so for longer (perhaps because they read about all its benefits while researching their pregnancy) and have kids with higher IQs due to making more money and being able to provide the better environment I mentioned in the last bullet point (in other words, some of the socioeconomic effect might actually be a maternal intelligence effect). There could also be a paternal intelligence effect: maybe smarter dads are more encouraging and supportive of breastfeeding, because, again, they’ve read about how important it is, and they also provide a richer environment for their kids’ cognitive development.
Genetics. All the explanations I’ve given above are entirely environmental – but another explanation has to do with genetics. As we saw above, smarter mothers are more likely to breastfeed (same thing for the smarter fathers who support breastfeeding), and have kids with higher IQs. But parents—biological parents, anyway—aren’t just giving the kids breastmilk – they’re also giving them genes. Smarter parents give their kids genes that help to make them smarter, and this might be at least part of the explanation for the apparent breastfeeding-IQ link.
These confounding factors aren’t mutually-exclusive, by the way: the genetic effect could be operating as well as the environmental ones, and the genetic effect might partially cause the environmental ones too, if genes make parents smarter and then being smarter has an environmental effect on IQ.
Other confounders. There are other potential confounders that are probably less important, but still worth checking out. Maternal health is one: if the mother is, say, obese or less healthy in general, they might find it harder to breastfeed and might have a less healthy baby, with less healthy brain development: it wouldn’t be the breastfeeding that would be the causal factor here but the overall health of the mother. You can probably think of several other factors here (whether you smoke or drink while pregnant; maternal age; how premature the baby was, or its birth-weight) which might or might not be part of the equation.
The point is: in an observational study, it might look like breastfeeding is having an effect on a child’s IQ, but actually that child would’ve had a higher IQ anyway, due to these confounding factors. Note that I’m not saying any of the above explanations is definitely the case – it’s just that whenever you see a correlational study of breastfeeding and IQ, you can’t necessarily rule any of them out.
“But wait”, you might be thinking. “Don’t studies control for these confounding factors?” The answer is: they try to! They try to take into account some of these possible confounders by including them as covariates in their statistical models. Normally when they add these variables into the model, the relation between breastfeeding and intelligence gets smaller – but the question is: does an association remain, even after including the relevant controls? If it does, these researchers argue, there’s likely a causal effect of breastfeeding on IQ.
But there are two big problems here:
Sometimes you don’t have the right control variables. A lot of studies of breastfeeding come from pre-existing observational studies. That is, they’re from big datasets where people have collected a whole bunch of measures on a wide variety of health-related outcomes, and made the data available for researchers to study. They weren’t deliberately set up to examine breastfeeding in particular, so might not have all the relevant variables. For instance, maybe they just measured the kids’ IQs and not those of the parents (it’s rare to see the mother’s IQ measured, and even rarer to see the father’s – getting data on all those tests is a major hassle). So you simply can’t control for that – you might use a proxy like parental education. And although parental education and parental IQ are correlated (smarter people get more qualifications on average), they’re far from the same thing. If you just control for education, you could miss some of the variation that’s really due to intelligence – to use the technical term, there would still be “residual confounding” left over.
So: if you had more of the relevant controls, the effect size you found would probably be smaller. And it’s rare that a study includes all of them.
It’s extremely hard to properly control for stuff. This follows on from point 1. Imagine you want to control for maternal IQ. At the extreme, you could do what we just discussed, and use maternal education – that’s a very noisy measure of maternal IQ, since IQ and education are only correlated a bit. But let’s say you do have a measure of IQ – the mother took a test. It’s still going to be a noisy measure – it won’t be an exact, perfect representation of the mother’s cognitive ability (maybe she didn’t get enough sleep the day before the test; maybe she got distracted while answering some of the questions; maybe she got really lucky on a few of them and guessed the correct answer – all of these would add random noise to the measurement). It’s not just intelligence, either: it’s also tricky to measure people’s socioeconomic status: is it just their income? What job they have? What kind of home or neighborhood they live in? Some combination of all of the above? The same applies to almost any control variable you can name.
And when you use a noisy variable as a control in your study, you’re not fully controlling for that variable, because the noise is obscuring its “true” variance. The best you can say in most circumstances is that you partly controlled for some of the variation that’s due to maternal intelligence, or to socioeconomic status, or whichever variable it is. You can read about this problem at three different levels of difficulty: a non-technical explanation at Slate Star Codex, a long and detailed (but enlightening) discussion in a paper from 2016 (which even includes a little app that you can play around with to get a feel for the problem), and a briefer, but maths-heavy treatment from the economist Miles Kimball.
So: if you had better measures of the controls, the effect size you found would probably be smaller. And it’s rare that a study includes extremely high-quality, low-noise measures of all its possible confounders.
We don’t know all the analyses that were run. This is the problem of so-called “researcher degrees of freedom”, where scientists can choose any one of very many potential analyses to run—in our case, which of very many potential covariates to include—and they might be biased towards the analysis that gives them the result they want. It’s perfectly possible, for instance, to include only the control variables that produce an overall effect of breastfeeding – even if this might be spurious. We can’t just assume that every study included every possible covariate, or that they reported every analysis they ran, unless the authors can prove it to us somehow. They could do that, to some degree at least, by using pre-registration (writing out your planned analysis beforehand and committing to this being your only one) and multiverse analysis (running every possible analysis and comparing the results) but I didn’t see a single instance of either of these in any of the observational studies I’ll discuss below.
So: if you paid more attention to transparency, we could have a lot more confidence in your analysis. And it’s rare—in fact, unheard-of in this field—that a study gives us this information.
Of course, one way around all these problems with controls and confounding in observational studies is to… not do observational studies, and to do randomised controlled trials instead. And that has been done – a tiny bit. We’ll come to the randomised studies—okay, study, singular—towards the end of this post. For now though, bearing in mind the big problems of confounding, let’s take a look at what evidence we have from observational research.
The state of the art in 2015
This research goes back a long way: the first study on breastfeeding and intelligence (and a bunch of other factors) was published as long ago as 1929(!). Happily though, there have been some review studies since then that have collected together all the best studies, so I don’t have to.
The review study I’m most familiar with, and the most-cited one by far (with over 600 citations), is a meta-analysis from 2015. They looked for observational studies that compared breastfed to non-breastfed children, or compared longer to shorter periods of breastfeeding, and where the child had an IQ test at some later point. They found 18 studies that fit the bill, and meta-analysed to find the average effect. They found:
breastfed subjects achieved a higher IQ [mean difference: 3.44 points (95% confidence interval: 2.30; 4.58)]
An average effect of 3.44 IQ points isn’t massive on an individual level, but it’s still considerable (if any stats fans want it as a standardised measure, it’s just under a quarter of a standard deviation: 3.44/15 = 0.23). Put it this way: would you be happy taking a pill where the side effect was that it knocked off nearly three-and-a-half of your IQ points? Probably not! So being three-and-a-half points higher is probably desirable.
But what did these studies control for? From the meta-analysis it’s hard to tell: the authors made sure they all controlled for “home stimulation” of the child, which will cover part of socioeconomic status (we might expect richer kids to have more interaction and stimulation from their parents) but not for socioeconomic status itself. About half of the studies controlled for maternal IQ, though, and here’s what happened:
Studies that controlled for maternal IQ showed a smaller benefit from breastfeeding [mean difference 2.62 points (95% confidence interval: 1.25; 3.98)]
Ah. So we’re down to 2.62 IQ points if you add in that very important control – and remember, if it was measured with better fidelity, we’d probably be down even lower than that. I’m sure some of these studies controlled for some kind of socioeconomic measure too, but perhaps some didn’t (annoyingly, this meta-analysis doesn’t report it), so we might have to revise the number down a little bit further still. And let’s not forget paternal IQ – I don’t think this was included in any of the studies, but it’s clearly relevant as we saw above, so we can revise down again.
And then, take a look at the data on the age at which the child did their IQ test (this is in Table 2 of their paper). The meta-analysts found that, in younger groups (age 1-9 years) the average effect of breastfeeding was 4.12 IQ points – but in older groups (age 10-19), it was only 1.92 points. This looks rather like the “fade-out effect” that’s often seen in studies of cognitive development: interventions have a larger effect earlier in life, but it peters out, and the groups (in this case, breastfed vs. not) end up basically the same some years later. If that’s the case, it undermines even further the idea that breastfeeding has enduring benefits for kids’ cognitive development.
The death by a thousand controls continues: what about genetics? It’s even rarer that studies control for this in some way, and again, the meta-analysis isn’t specific about which studies did. I know that one of the studies they included did – an analysis of the National Longitudinal Study of Youth dataset from 2006 (conflict of interest: it was co-authored by my previous boss, Ian Deary). In that study, they were able to find siblings where one was breastfed and one wasn’t, and where one was breastfed for a longer time than the other. Now, this method isn’t without its flaws: perhaps sibling pairs where this happens are quite rare and unusual, so they might not be representative. But it does hold a lot of things constant all within one analysis – a good chunk of the genetic effect, and any family environment that the siblings shared. It’s a really nice way to control for lots of confounders at once. Here’s what they said:
There were 332 sibling pairs discordant for breastfeeding status and 545 discordant for duration of breast feeding. [We tested] mean differences between both groups of siblings for [IQ] scores. None of them was significantly different from zero.
That said, the meta-analysis did also include another sibling-control study from 2005 where the effects of breastfeeding on everything except IQ were wiped out in the within-family analysis; but they didn’t include a more detailed sibling-control study from 2014 where all the effects on IQ were completely wiped out when sibling controls were used.
There’s one more thing to say about this meta-analysis, and it has to do with publication bias. “Breastfeeding improves kids’ IQs!” is a good-news story: people want to hear more of it. They probably don’t want to hear “Breastfeeding has little or no effect on kids’ IQs!”. That could affect which studies are published: the good-news ones are more likely to get an easy ride through peer-review.
Here, the meta-analysts argue that it’s unlikely there’s publication bias because they found about the same-sized effect of breastfeeding in studies with less than, and greater than, 500 participants. The (unstated) logic, I think, is this: bigger studies are more reliable, and because they’re big, it’s hard to avoid publishing them in some form. Small studies are easier to shove into the file-drawer if they don’t find the “desired” effect. But if small studies and big studies give the same size of effect, it’s more likely that all the relevant studies are finding their way to publication.
But this is a very crude, arbitrary way of checking for publication bias: who decided 500 was the relevant number, and why? Also, they report doing a “funnel plot and Egger test”—two of the standard ways to detect publication bias—but conspicuously don’t report these anywhere (whoever peer-reviewed the meta-analysis must not have noticed). That’s where I came in – I got the relevant data, and did a few different statistical publication-bias analyses in a letter I got published in the same journal after the meta-analysis came out (and to which the authors responded).
None of the methods I used was foolproof, and they aren’t 100% knockdown evidence of publication bias. But my conclusion—I’ll spare you all the details here, but please do read the letter if you like that kind of thing—was that there was at least circumstantial evidence of publication bias in the meta-analysis. And what would happen if we adjusted for publication bias? You guessed it: we’d reduce the overall effect size of breastfeeding even more, since the bias made it seem bigger than it really was.
The 2015 meta-analysis is a nice example of how a review paper can appear to support—and is written up as if it supports—a particular position, but where it all starts to crumble when you dig into the evidence a little. With every extra control, every tighter study design, and every bias-reduction method, the effect dwindles. Would it dwindle completely to nothing in the perfect study with all confounders accounted for? It’s impossible to say for sure – but the already-quite-modest effect is certainly under threat from all these potential confounders.
But 2015 was a long time ago. In the next section, we’ll ask: has anything happened in the world of observational research since 2015 to make us change our minds?
Newer research… and a twist in the tale
An Irish study from 2017 is a paradigm example of the idea of covariates chipping away at the effect size. They had data from around 8,000 kids who had taken several IQ tests (along with other measures) at ages 3 and 5, and information on whether they had been breastfed as infants. When the researchers just looked at the basic associations, breastfeeding was correlated with everything good: higher cognitive abilities, fewer behavioural problems. But when they used 14 control variabes, including socioeconomic ones and maternal education, but not including maternal IQ, it rendered every IQ result non-statistically-significant – and removed all the others too, except for one small effect on hyperactivity at age 3 (though there was nothing at age 5).
But what about this similarly-sized study (n = about 9,000) from the US, published in 2021? Here, the kids did their IQ tests at about age 9-10, and the researchers controlled for a long list of covariates, though again, not for maternal IQ. Even after including all these controls, there was a statistically significant dose-response relation between breastfeeding and general intelligence: longer breastfeeding duration, higher child IQ. With breastfeeding for more than 12 months, the estimate was about 4.5 IQ points – though we might expect that to be chopped down a bit if they had maternal IQ or a sibling-control design.
A 2022 study from the UK had, again, a ballpark-similar sample size (n = about 7,000), and in this one they did have a measure of the mother’s IQ. After controlling for this and a wide range of other covariates, they found a weird and inconsistent pattern of results across different ages. First, as usual, the effects of breastfeeding were a lot smaller after all the controls. Then, the effect came and went: there wasn’t a significant association with IQ at age 5, but there was at age 7, which disappeared (pretty much) by age 11, but came back at age 14. At all ages, though, the effect size was pretty “modest” (as the authors put it), so would be at risk of disappearing altogether if there was residual confounding – which, let’s face it, there probably is!
Those are a few of the biggest post-2015 studies that I found: the ones that would move the needle most if we were updating our view. And here’s a selection of studies that give interesting other angles on the question:
Is it to do with the IQ test? A UK study from 2015 (n = 4200ish children) aggregated together lots of different IQ tests into one “latent” general intelligence measure, and still found effects even after control for maternal IQ. I’m a fan of using latent variables, because it reduces measurement errors that might be associated with any one individual test. But I do think there’s something odd about the fact that the individual IQ subtests (verbal memory, maths, reading, etc.) were all resoundingly unrelated to breastfeeding (like, in a very large sample, the associations have the following p-values after adjustment: .945, .833, .858, .743, .940), but when made into a latent variable, there was magically a substantial and significant association with breastfeeding. It’s certainly possible! But I’d be interested in a closer look at that dataset.
Does breastfeeding influence IQ change over time? Even if breastfeeding doesn’t influence the initial level of kids’ IQs, maybe it gives a boost to the change, the maturation, of cognitive abilities as the kids go through adolescence? Well: not according to this 2015, n = 11,000ish study from the UK that had IQ data all the way from age 2 to age 16. They barely found an effect on the level, and found nothing on the change. The one downside is that they used different tests at each age, making it trickier to estimate the “growth curve” of change over time.
Can we see what’s changing in the brain? Using the same sample as the 2021 study we saw above, where the kids had also had brain scans, a 2022 preprint reports correlations between breastfeeding and larger volumes in some tiny parts of the brain at age 9-11ish years (n = 7,800). They don’t provide any effect sizes at all in the paper, so it does seem a bit much for them to claim that “breastfeeding seems to be a major factor in the maturation of the brain”. They later do some “mediation” analyses, where they do report effect sizes, but they’re absolutely tiny, again indicating that the quoted statement might be a little OTT.
If the above was all we had, I’d be pretty confident in betting on the “residual confounding” explanation. That is, (a) very few studies control for all the relevant confounders and (b) due to measurement error, the confounders aren’t controlled-for perfectly – and both of those points mean that the effect could easily be either an illusion due to confounding or just very very small indeed.
But here’s a study that gives me pause. It’s actually by the same authors of the not-very-good meta-analysis we encountered above. It’s a prospective study from the Brazilian Pelotas Cohort, and it came out just after the meta-analysis was published, also in 2015. The fascinating thing about this study (which had just under 3,500 participants) is that the relation of breastfeeding to socioeconomic status is… not there. Almost every mother tried breastfeeding at some point in this sample (compared to the UK where people from poorer socioeconomic backgrounds are less likely to try it at all) and the relation between income and breastfeeding for three months is flat (compared to the UK where richer people are more likely to still be breastfeeding at this point).
This isn’t just a Brazilian thing – it’s the case in other non-“Western” countries too. See, for example, this paper which surprisingly argues that we need interventions to encourage higher-income people in Uganda to exclusively breastfeed more so we can “address inequalities in breastfeeding”.
So the logic of confounding (“maybe it’s just social class that explains breastfeeding and higher IQs”) doesn’t work at all here. And yet, in the Brazilian study, they found higher IQs in the breastfed children at age thirty years – and yes, not thirty months, thirty years. The effects had lasted that long, and also seemed to translate into doing better at school and earning more money. Even after controls (which, remember, won’t work the same way in this sample because socioeconomic status doesn’t show much relation to breastfeeding), the breastfed 30-year-olds were 3.76 IQ points higher than their non-breastfed counterparts.
Here are the possible explanations:
Breastfeeding really does have an effect, beyond all the confounding factors, and if we ran the perfect study (or set of studies; one is never enough) in the West we’d find this;
Breastfeeding really does have an effect – in Brazil. Perhaps due to different diets, different makeup of infant formula, or some other factors that differ between Brazil and the other countries we’ve discussed above, it really does matter whether you breastfeed if you’re in some countries (like Brazil) but not others (like the UK or the US). We really need to see more studies from non-Western countries to get more data on this;
The Brazilian study is confounded in some idiosyncratic, non-obvious way and breastfeeding isn’t actually having an effect. I must admit I find the sheer size of the effect (3.76 IQ points) surprising: larger than the meta-analytic estimate (3.44 IQ points), and the meta one came mainly from IQ tests of children, where this one is from 30-year-olds. If anything we’d expect to see the effect fading out as the participants get older, as we discussed above. So perhaps there are some other factors at play in this study.
Could one of those other factors be genetics? Well, the authors appear to have this covered: they wrote a follow-up paper where they controlled for genetics, and argued that “the genetic confounding hypothesis seems unlikely”. Except we can dispense with this one pretty easily: they didn’t use the sibling control method to control for genetics – instead, they used a polygenic score for education. The polygenic score only picks up on a small portion of genetic variation, so—whereas it’s maybe better than nothing—it’s really not “controlling for genetics” in any substantial sense. It would be like saying you’ve “controlled for education” when you controlled for someone’s performance on the first question on their age-16 school Science exam: sure, these things are related, but you’re missing out on an awful lot of information.
At any rate, the comparison between the Brazilian study and the Western ones is the strongest evidence we’ve seen so far that there might be something going on with breastfeeding and IQ.
But to really answer this question, we can’t just read the tea-leaves of all these observational studies. We need some experimental, causal studies. And that’s where we’re going next.
Causal studies (and sort-of causal ones)
Maybe I shouldn’t have bothered writing all the stuff above: after all, one good randomised controlled trial (RCT) is worth a hundred biased, confounded correlational studies. The Brazilian meta-analysts seem to agree: in the Abstract of their meta-analysis, they cheekily point to the RCT evidence, even though it wasn’t a part of their specific analysis at all.
For breastfeeding and IQ, there really is just one RCT. That’s the PROBIT trial from Belarus, which first reported its IQ results in 2008. PROBIT stands for “Promotion of Breastfeeding Intervention Trial”, because that’s what they did: it’s not like you can randomise mothers to breastfeed or not—for obvious practical and ethical reasons—so instead they randomised hospitals to make their environment friendlier towards breastfeeding (midwives, doctors, and nurses went on intensive courses in how to support breastfeeding mothers). In the hospitals randomised to be in the control group, it was business as usual. The intervention really worked: breastfeeding rates became much higher in the “breastfeeding-friendly” hospitals. Then the researchers followed up an amazing 13,889 of the kids when they got to age 6, and gave them four different IQ tests.
Here’s where things get a bit tricky. As is standard, the four tests were split into “Verbal IQ” (a vocabulary test and one where you have to work out similarities between words), “Performance IQ” (two pattern-spotting tests – one where you have to arrange 3D blocks and one where you have to work out a picture sequence), and “Full-Scale IQ” (an average of all four tests). There was no statistically significant effect on Full-Scale IQ. Nor was there one on Performance IQ. Only Verbal IQ showed a statistically significant effect, though it was a massive one: 7.5 IQ points (they also had the kids’ teachers rate them for reading, writing, and maths, and found no effects there at all).
But even that Verbal IQ result is shaky. Because the researchers couldn’t properly “blind” the whole study (the doctors testing the kids’ IQs knew whether they had been working at a breastfeeding-friendly hospital or not), they selected a subsample of participants to be “audited” and tested by psychologists who really were blind to each child’s breastfeeding status. In this audit subsample, the effects were much smaller (2.9 points for Verbal IQ), and none were statistically significant. Is it yet another case of “more controls, smaller effect” (or, in this case, “more blinding, smaller effect”)? You can take a look at a critical letter that was published at the time for this point and a couple of other issues with the study.
These results might all be moot anyway. Remember how I mentioned fade-out, above? The idea that intervention effects on IQ taper off over time? Well, it seems as if we have a classic case here: the PROBIT team followed up their participants and published a new study in 2018, by which time they had had their IQs tested again at average age 16. This time they gave them 10 different IQ tests (different ones from when they were 6-year-olds) which covered 7 different abilities. There were no significant effects, except on verbal ability, where the breastfed group had a 1.4 IQ point advantage – there’s your fade-out effect right there! On the one hand, the fact that it’s the same cognitive domain as the one that was improved in childhood (verbal ability) should make you trust it more; but on the other, the fact that it’s one of many tests and there was no correction for multiple comparisons should make you trust it less.
My overall interpretation is—and I’m sorry about this becuase it’s such an unsatisfying conclusion—that this RCT, massive as it is, leaves open many questions about its very ambiguous effects, despite the way it’s described in the scientific literature as if it completely clinched the question of breastfeeding’s effect on IQ. We’re going to need more, similar studies if we want to really nail the answer.
But alas, that’s really it, as far as RCTs go. In an older study from 1992, there’s a passing mention to some randomised studies, but they’re followed with the dreaded term in parentheses that you never want to see in scientific papers: “(unpublished)”. I couldn’t find any other mention of them anywhere.
However, we do have three more studies to mention. They’re not RCTs, but they’re examples of those clever “quasi-experimental” studies that economists do where they manage to magic causality out of observational data. They use an “instrumental variables” approach, where randomness is introduced into the study not by the deliberate design of the researcher, but by some other variable that occurs naturally. For example, some instrumental-variable studies use oil prices—which go up and down due to all sorts of unpredictable world events—to see how national income (which is affected by oil prices) impacts things like education or fertility (which are only affected by oil prices via the effect of oil prices on income).
To put that another way, the point of an instrumental variable (or “instrument”) is that it has to affect the outcome of interest (in our case, intelligence) only via its effect on the exposure of interest (in our case, breastfeeding) – it can’t have any other effect on intelligence via any other means. If you can sustain this argument, it’s evidence that the exposure causes the outcome (in our case, that breastfeeding causes higher intelligence), because there would be no other reason for variation in the instrumental variable to relate to variation in the outcome. But if there’s another way for the instrument to affect the outcome, the causal argument breaks down. So which instruments might fit the bill for breastfeeding?
The first one, from a 2010 paper, is caesarean sections. Here, the researchers argued that the choice, or need, to have a caesarean is unrelated to confounding factors that might affect intelligence, but that it does have an effect on breastfeeding, by making it harder to do so. So, via breastfeeding, caesareans can impact intelligence. The actual instrumental variables analysis finds no causal effect of breastfeeding on intelligence.
They go on, rather oddly, to make a convoluted argument that actually the instrumental variables results mean that the non-causal, standard, correlational results on this question are more likely to be correct “once a suitably rich set of confounders are included”. But then they note that in their analysis they didn’t control for maternal IQ… which brings us right back to square one.
The second study, from 2012, is a bit more like the RCT, in that it compared kids who were born in hospitals that had extra breastfeeding support (as part of a global UNICEF programme) to those who weren’t. It’s just that, unlike the RCT, whether a hospital had extra support wasn’t randomised, so there’s always the creeping suspicion that there was something different about the extra-breastfeeding hospitals that might confound the study (maybe these hospitals were better in some way that improved the kids’ IQs regardless of whether they were breastfed).
The authors go to some lengths to argue that it was effectively random, however, giving a big list of factors where they found no differences between hospitals with and without the breastfeeding initiative. They end up finding really big effects of breastfeeding on children’s IQ. But it’s still a major concern whether there might be some unmeasured confounder (see this critique of the instrumental-variables approach, for example). In other words: the same “residual confounding” criticism that we made of the purely observational studies still applies to this quasi-experimental one.
The third study, hot off the press in 2022 (though around as a working paper for some time before that), uses a different instrument altogether: the day of the week a child is born. The idea is that maternity hospitals have lower-quality breastfeeding support at weekends—it costs more to employ staff for weekend hours—so babies who are born at the weekend are less likely to have mothers who breastfeed effectively and consistently. Indeed, they found that weekend birth was linked to a lower likelihood of breastfeeding. They went on to show that, whereas there were no effects of breastfeeding on health or non-cognitive outcomes like emotional or behavioural problems, there were big effects on the IQ tests. They note that:
The magnitude of the effects are around 65% [of a standardard deviation; 9.75 IQ points], which are large but quite plausible given the estimates obtained by a randomized trial in Belarus…
Here, as with the observational studies, there’s an attempt to triangulate these instrumental-variables results with the RCT… but alas, we saw above that the RCT results are far from solid either. And 9.75 IQ points is a gigantic effect, relative to the other observational studies, and to things that we know impact IQ in general. As is common in instrumental-variables studies, the estimates are much bigger than we’d see in a normal model, with very wide confidence intervals that mean they come with a great deal of uncertainty. This always makes me gut-level suspicious of these kinds of estimates.
All three of these instrumental-variables studies, by the way, used the same dataset, the UK Millennium Cohort Study (the first one also used the National Child Development Study). The fact that the latter two used different instruments and still found the same effect is a plus, but of course it’s inconsistent with the null effect from the first one.
Ultimately what we need is more RCTs. RCTs where the intervention is deliberately randomised. RCTs in different countries (are results from Belarus really generalisable to other countries? We don’t know). RCTs that have long follow-ups. RCTs that collect lots of different cognitive outcomes. And so on.
Studies that lact evidence
It’s depressing that the answer to the question “does breastfeeding make kids smarter?” is still so murky. It’s not through want of studies: given all the research we’ve seen above, we should have a better idea by now.
Indeed, the literature on breastfeeding and intelligence is a nice example of a field where you can pile up reams of studies, and where very many of them—a clear majority—appear to point in one direction, of positive effects of breastfeeding. But when you dig in a bit, and when you consider just how difficult a question it is to answer, you end up far more confused: the observational studies are almost always missing a piece of the puzzle to avoid confounding, and the randomised trials we need are almost entirely lacking.
My instinct is still to lean towards the “confounding” explanation: that there are problems in the studies that make them spuriously point towards effects of breastfeeding on IQ. It’s certainly true that the best studies—the ones with maternal IQ, or sibling control, or both—are less likely to show an effect, or if they do it’s a smaller one in general. And the effects in the only RCT we have are undermined by both its own blinded sub-study which found no effects, and the subsequent fade-out in the follow-up research.
But it would be obstinate not to say that a few of the other studies—like the Brazilian one with the non-Western pattern of confounding, and a couple of the quasi-experimental ones—don’t make me update at least a little towards there being a real effect. They’re still not enough to overcome my general scepticism, though: we’d need more studies for that. The fact that my overall conclusion is “we need more studies” really is a serious indictment on the quality of decades of attempts to answer this (really important!) question.
There are three final things to say. The first is about our priors. When I recently tweeted sceptically about the benefits of breastfeeding, a few people made the following argument (I’m paraphrasing across three or four tweets):
Breastfeeding has been around for a long time. It’s a costly thing to do, in an evolutionary sense, so it would be weird if it had evolved and it didn’t have any particular benefits. Breastmilk has all sorts of nutrients and compounds in it that formula milk doesn’t; we’re only at the start of our understanding of how all of these work.
So: our prior should be that breastfeeding is the right thing to do, and it’ll take quite a lot of high-quality studies to move us away from that position.
I definitely appreciate the logic here. But there are a very large number of claims about the benefits of breastfeeding—not just for IQ—and they can’t all be just as likely to be true. Of course, given how strong your priors are it’ll take different degrees of evidence to budge them, and I hope this post has given you a decent enough summary of the evidence for you to update in whichever direction that’s relevant based on your previous beliefs.
But this argument actually reminds me of another important point, which is a further difficulty for this whole field of research: over the years, infant formula has evolved too. It’s had several new ingredients, like nucleotides and polyunsaturated fatty acids, added to it, making it even more like breast milk. In very many of the studies we’ve seen above, the babies were born well before these new ingredients were added to infant formula. So it’s harder to compare older studies (even quite recent ones; the polyunsaturated fatty acids were added in 2000) to babies born today or in recent years. This should reduce your confidence in the studies showing unique effects of breastfeeding a little bit more, in my view.
The second thing to say is about what I’m not arguing in this piece. I’m not arguing against breastfeeding, by any means – the only thing I’m aguing against is drawing strong conclusions from shaky, contradictory scientific research. If you want to breastfeed, and you can do so, nobody is saying you shouldn’t or that you’ve been duped into it. Heaven knows IQ is only one of a million things new parents are concerned about, and “I breastfeed because it’s nice and helps me bond with my child” is a perfectly valid argument in favour of doing so.
On the other hand, we should be sceptical about the pressure that’s put on so many mothers about breastfeeding by their midwives, doctors, and society in general. Given how strongly breastfeeding is pushed, and how guilty mothers who don’t or can’t breastfeed are made to feel, you’d imagine there would be super-solid evidence supporting its benefits. As we’ve seen for IQ at least, that’s really not the case.
The third and final point is that I’ve focused entirely on the IQ studies in this article. Although the Tom Chivers piece that I mentioned gives some reasons to be sceptical of many of the other assumed benefits of breastfeeding, and although the exact same concerns about confounding and causality apply to breastfeeding research that doesn’t involve IQ, I simply haven’t dug into the quality of the evidence for breastfeeding’s effects on the immune system, or on obesity, or on social bonding, or any of the other claimed benefits. Maybe I’ll look into that in a future post.
For now though, that’s everything you need to know about breastfeeding and intelligence (er, unless I’ve missed anything big – in which case do let me know). At the risk of sounding like a broken record, it’s another area where—in the morass of confusion and overstatement, studies that vary wildly in quality, and failure to use the optimal research designs—science has let us down. We know, to a high degree of certainty, the answers to so many important medical questions. Why don’t we know the answer to this one?
Image credits: Getty