The Paracetamol Question That Two Million Babies Cannot Quite Answer

The packet is on the shelf. You are pregnant, you have a fever, and paracetamol (acetaminophen) is the one painkiller everyone agrees is safe to take. Your midwife said so. The NHS leaflet said so. You take it, and you feel better, and you do not think about it again.

For roughly half of all pregnant women in Taiwan, that is more or less what happened. A nationwide study that tracked over two million births has now spent years trying to work out whether those women’s children grew up any differently. The answer it has reached is, in the most rigorous possible sense, inconclusive.

That conclusion might sound like failure. It isn’t. The Taiwan study, published this week in JAMA Pediatrics, is the latest and largest entry in one of the more quietly contentious debates in prenatal medicine: does paracetamol taken during pregnancy raise the risk of attention-deficit/hyperactivity disorder or autism in the child? The study doesn’t settle the question. What it does instead is illuminate, with unusual precision, why the question is so hard to settle, and what that difficulty tells us about the limits of the methods we’ve been using to answer it.

The numbers, on the face of it, look damning. Among the 2,092,926 singleton births captured in Taiwan’s National Health Insurance database between 2004 and 2015, just over 48 per cent were born to mothers who had received at least two paracetamol prescriptions during pregnancy. Pei-Chen Lee at National Cheng Kung University and colleagues compared these children to those whose mothers had used it less or not at all, and found a clear association: children in the exposed group were about 12 per cent more likely to be diagnosed with ADHD, with a similar, if somewhat smaller, pattern for autism. Higher doses, or more frequent prescriptions, were associated with higher risk. The classic shape of a dose-response relationship.

But association is not causation, and epidemiologists have been arguing about confounding here for years. The problem is obvious once you think about it. Women who take paracetamol during pregnancy often do so because they are in pain, running a fever, or dealing with an infection. Pain, fever, and infection are themselves associated with stress on the developing foetus. Perhaps it isn’t the paracetamol at all. Perhaps it’s what the paracetamol was treating.

To get around this, researchers have increasingly turned to a design that feels almost elegantly simple: compare siblings. If you look only at families where one child was exposed in utero and another wasn’t, you control automatically for all the shared factors (genetics, household income, maternal health, family history of neurodevelopmental conditions) that a standard cohort study might miss. Surely, if the paracetamol itself is driving the risk, the exposed sibling should still show higher rates even within the same family.

In the Taiwanese data, they don’t. Among the 1.2 million children in the cohort who had at least one sibling in the database, the association between prenatal paracetamol and ADHD or autism entirely disappeared when the analysis was restricted to sibling comparisons. A Swedish study published in JAMA two years ago found exactly the same thing, as did a Japanese study, though it too took a different approach. The consistency across countries, healthcare systems, and study designs might suggest the matter is closed: it’s confounding all the way down, and paracetamol is off the hook.

Except then the Taiwan team looked more closely at their sibling data, and something strange emerged. They split the sibling pairs by birth order, asking separately: what happens when only the older sibling was exposed, and what happens when only the younger one was? The answers flatly contradicted each other. When the older sibling alone had been exposed in utero, that child was 33 per cent more likely to have ADHD than their unexposed younger sibling, and 75 per cent more likely to be diagnosed with autism. When only the younger sibling had been exposed, the picture reversed completely: that child was actually less likely to have ADHD or autism than their unexposed older sibling. The same drug, the same analytical design, opposite results depending on which child you were looking at.

There is no plausible biological story in which paracetamol protects a younger child’s neurodevelopment while harming an older one. Which means the reversal is not about paracetamol at all. It’s about something systematic in the sibling comparison design that the researchers have not yet been able to identify: perhaps changes in diagnostic practice over time (younger children in the cohort were diagnosed later in the study, when criteria had shifted), perhaps what epidemiologists call carryover effects (where one pregnancy’s exposure or outcome influences the next), perhaps amplified measurement error. Probably some mixture.

The implication is uncomfortable but important. Sibling comparison designs have gained considerable credibility in observational epidemiology precisely because they seem to dissolve confounding by inheritance and shared environment. The Taiwan data suggests that credibility may be partly illusory, at least for questions like this one, where exposures and outcomes both shift across time and birth order. The design controls for some things and not others, and may introduce its own biases in the process.

None of this means paracetamol is definitively safe. It means the methods we have, applied to the data we have, cannot get us to a definitive answer. The full cohort analysis says there is an association. The sibling analysis says there isn’t. And the bidirectional breakdown of the sibling analysis says the sibling analysis itself cannot be fully trusted. Three methods, three different signals, all from the same two million births.

What would actually resolve this, Lee and colleagues suggest, is a different kind of study entirely: one that assesses neurodevelopmental outcomes directly, using standardised tools applied at the same age to all children, rather than relying on diagnosis codes in an administrative database. Registry data has real strengths: no recall bias, enormous scale, no dropout. But it also captures diagnoses only when families seek care and clinicians apply a label. Both of those things change over time.

For now, the guidance from major health bodies remains what it has been for years: use the lowest effective dose for the shortest time, avoid prolonged use, and weigh whatever discomfort you’re treating against the theoretical risk. The precautionary principle, applied gently to something that two million pregnant women have already done without obvious consequence, while researchers keep looking.

doi:10.1001/jamapediatrics.2026.0071

Frequently Asked Questions

Does this study prove paracetamol during pregnancy causes ADHD or autism? No and it is notably cautious about saying so. The full dataset of two million births does show a statistical association between prenatal paracetamol prescriptions and both conditions, but the same data analysed differently (comparing siblings within families) makes that association disappear. The researchers conclude that causality has not been established, and that both analytical methods have limitations severe enough to prevent a firm conclusion either way.

Why do the results change so dramatically depending on how you analyse the data? The core issue is confounding: women who take paracetamol during pregnancy are often doing so because of pain, fever, or infection all of which may independently affect foetal development. Sibling comparisons were designed to control for this by holding family background constant. But the Taiwan team found their sibling data produced contradictory results depending on birth order, which suggests the sibling design itself may introduce biases (related to changing diagnosis rates over time and how one pregnancy’s circumstances influence the next) that are difficult to untangle.

Should pregnant women stop taking paracetamol? Major health agencies have not changed their guidance as a result of this or similar studies. The current recommendation is to use the lowest effective dose for the shortest time necessary, which has been the standard advice for some years. The uncertainty in the research reflects a genuine gap in what observational epidemiology can establish not a hidden danger that regulators are ignoring.

Why is this question so difficult to answer definitively? Conducting a randomised controlled trial of paracetamol in pregnancy the gold standard for establishing causation is ethically impossible. Researchers are therefore left with observational data, which can only ever establish association. The difficulty is separating the effect of the drug from the effect of whatever condition prompted its use, and doing so reliably across populations where diagnostic practices and healthcare-seeking behaviour differ. The Taiwan study makes clear that even the most sophisticated workarounds for this problem have limitations that are hard to quantify.