Tylenol and Autism: A Replication!
A Japanese sibling control study, probabilistic sensitivity analysis, and negative control study all-in-one paper confirms there's no link!
Last week, I reviewed a recent systematic review cited by the HHS in support of their argument that Tylenol (acetaminophen/paracetamol) is a factor involved in causing autism. That report was not credible; in fact, it was internally contradictory and its authors eschewed causal thinking and showed that they had a poor understanding of epidemiological methods. For more detail, go read my review of it:
Or, if you prefer a short-and-sweet rebuttal, go check out the earlier article I wrote on the topic:
If you would just like to understand the “autism epidemic”, as it were, then go see:
And finally, if you want to see how to end the “autism epidemic”, check out:
As luck would have it, more evidence has come out since that systematic review was published. This new evidence came online at the beginning of September and evidently everyone missed it, but it is considerable. Given what we already know—it is highly unlikely that acetaminophen has any role in autism—, it’s nothing new: it reinforces the knowledge that acetaminophen does not cause autism!
Japanese Evidence
The new study is a tour de force.
The new study takes place in Japan, far afield from the northern European countries all of the previously available causal evidence ultimately comes from. The authors’ dataset is a national administrative database linking mothers and their offspring between April 2005 and March 2022. This database contains data on a total of 16 million people covered by the Society-Managed Health Insurance System, including employees of medium-to-large companies in urban areas and their dependents. Overall, the database contains a representative slice of 13% of Japan’s total population.
The study ended up with a sample of 182,830 mothers who gave births, with a total sample of 217,602 of the resulting children. Usage of acetaminophen during pregnancy was identified from prescription dispensation records. Although acetaminophen is available over-the-counter in Japan, most usage comes from prescriptions1, and usage during pregnancy is generally found to be uncommon. The authors used any usage during pregnancy as their exposure measure and they used ADHD, autism, and intellectual disability diagnoses or corresponding treatment records as their outcomes.
These authors had a lot more covariates available to them than previous authors did. They had data on covariates ranging from maternal birth order to the results of her most recent health checkup to a wide array of different drugs they’ve used at various points in time. They used this data to do basic adjustments, to do stabilized inverse probability of treatment weighting (IPTW), and to do propensity-score matching (PSM). Stated differently, they had more controls and better control strategies than many other papers.
The authors were also able to do three other important things related to the wonderful data they had available to them. The authors were able to computed bounding factors to assess the strength of possible selection bias required to explain their results. They were able to use their PSM approach to calculate E-values, indications of the minimum strength an unmeasured confounder needs to have to explain their observed results. They were able to also do a negative control analysis using drug usage data from before, during, and after pregnancy. Finally, they were able to assess the effect of misclassification in a probabilistic sensitivity analysis.
Baseline Results
Maternal acetaminophen usage was related to teach outcome before any adjustments were made. Usage predicted a 34% greater crude risk of ADHD, a 17% greater risk of autism, and a 20% greater risk of intellectual disability—all highly statistically significant results.
Adjusted Results
After basic covariate adjustment, the results for ADHD and autism remained statistically significant, but were reduced to 22% and 8% increased risk. The result for intellectual disability dropped to 5% greater risk, but was no longer anywhere near significant.
PSM and IPTW
The PSM and IPTW results were directionally comparable to the crude and adjusted results, but significance was mixed. For the IPTW results, ADHD risk remained elevated by 32% and autism risk was marginally significantly elevated by 9%, whereas intellectual disability was not significantly elevated (+3%). For the PSM results, ADHD risk remained elevated by 22%, but the other results were rendered nonsignificant: +7% for autism and +3% for intellectual disability.
Confounders and Selection
The strengths of confounding and selection bias required to explain these results were pitifully small. The E-value computed for an overall composite neurodevelopmental outcome (+19% at baseline, +8% in PSM) based on the three diagnoses was an HR of 1.37, and the bounding factor required to explain the result for the same outcome (+9% in IPTW, +11% in IPCW) was 1.24. In the authors’ words:
an unmeasured confounder associated with both the exposure and outcome by a risk ratio of at least 1.37 could explain away the observed effect.
and
if the selection mechanism related to pregnancy loss were associated with both the maternal acetaminophen use and the composite outcome with risk ratios of approximately 1.24 each, the true association could be entirely attributable to bias.
To put these in perspective, translate them to correlations. HRs of 1.37 and 1.24 are on the order of Pearson correlations of 0.06 to 0.09. Very minimal biasing factors could explain these results, which are not unlike other published results in the systematic review, as we saw when I looked into it last time.
Negative Controls
Using NSAIDs or aspirin was associated with +8% or a nonsignificant +14% boost to composite neurodevelopmental risk. For acetaminophen use before pregnancy, there was a significant negative association with risk, -7%, whereas usage after pregnancy was marginally nonsignificantly associated with elevated risk, +5%.
In a previous study—which I’ve already discussed—done on the Nurses’ Health Study II cohort, acetaminophen usage before and after pregnancy were not significantly associated with neurodevelopmental outcomes, but they were directionally associated with elevated risk consistently before pregnancy and inconsistently after pregnancy, but always nonsignificantly. The association during pregnancy was directionally consistent, but sometimes nonsignificant or marginally significant.
In the Norwegian and Mother Child Cohort, usage before pregnancy was associated with marginally significantly elevated risk of communication problems, and no association with externalizing or internalizing symptoms. This risk elevation (+19%) was similar to the risk elevation for acetaminophen usage during all three trimesters (+18%), which was itself nonsignificant. In the Ahlqvist et al. study, there were no significant negative control associations except for one with aspirin use in their sibling comparison.
All of these results do not agree: “NSAIDs and aspirin use during pregnancy and acetaminophen use after pregnancy indicated the presence of bias away from the null, whereas acetaminophen use before pregnancy indicates the bias towards the null… These findings highlight the inherent limitations of negative control analyses, as there is no definitive standard, and the detected biases may not fully capture the true underlying bias structure.”
Frankly, it’s hard to take negative control results seriously in general because, as Ahlqvist et al. noticed, mothers move away from taking medications during pregnancy. That’s the smart thing to do, and it’s something we should expect to be selective, much as phenomena like SIDS and childhood obesity are selective with respect to parent knowledge and conscientiousness. The idea behind these analyses is that the bias structure will be shared before and during or during and after, or before, during, and after pregnancy, and this is just not realistic in a world where mothers shift their habits selectively when they become parents.
Misclassification Assessment
Finally, these authors found something I suggested was possible when I last wrote about this topic: that misclassification of acetaminophen use could bias results upwards. They note:
Our probabilistic bias analysis indicated that misclassification of acetaminophen exposure, particularly owing to unrecorded OTC use, may have led to an overestimation of the observed associations. This bias is reflected in the consistent attenuation of effect estimates towards the null across a range of plausible levels of OTC use. Although OTC use among pregnant women in Japan is generally low, our findings indicate that even minor misclassifications could partly explain the modest associations observed. Moreover, this limitation is not unique to prescription-based data; exposure misclassification can similarly occur in studies relying on self-reported data, where recall errors or underreporting may further obscure the true exposure status.
Wrapping Everything Together
The results of these methods are in agreement: there’s not much evidence for acetaminophen risk, using the best non-causal methods cited in the systematic review, with a large, representative sample. But the authors didn’t stop there. They went on to replicate Ahlqvist et al.’s sibling control results in a sample with 23,593 children with discordant acetaminophen exposure! Their replication was fairly exact, too. Where Ahlqvist et al. saw the direction of risk nonsignificantly reverse, so did these authors. Give the results a look:
Notice the presence of two sibling estimates. One of those is not adjusted for other covariates and the other is adjusted for them, and they agree with one another.
Ergo, there we have it! Every line of defense for the prior literature, and every reason cited to dismiss Ahlqvist et al. fails. Both their significant cross-sectional effect and their directionally opposite, nonsignificant sibling control results replicated outside of Sweden, in distant, high base rate, Japan; misclassification issues likely pushed results in the wrong direction for the literature to support an acetaminophen-autism link, and; the sibling result was not attributable to prior covariate adjustments.
Professional societies have responded appropriately to HHS’ announcement about the link between acetaminophen and autism, by rejecting it. With luck, providers will heed that advice and eventually the U.S. government will follow suit. That is how things should be, and this new causal evidence supplies us with ample reason to think the field’s earlier conclusions were correct: acetaminophen does not cause autism.
Japan’s 2024 OTC analgesics market was valued at $350.1M. A representative OTC APAP price from Daiichi Sankyo’s new Calonal-A (300 mg) is 36 tabs for ¥1,518 → ¥42.2/tab → ¥140.6 per gram of APAP. Set $1 to ¥150, that’s ~$0.94 per gram. Japan’s largest APAP brand, Ayumi, ramped up production of prescription Calonal (not Calonal-A) to 2.88 billion 200mg equivalent tablets per year, or 576 metric tons of API capacity per year and they’re responsible for 85-90% of prescription acetaminophen in Japan. Convert OTC APAP value to grams using the retail price-per-gram, as (OTC analgesics market * APAP share)/($/g) and if APAP is 30%, of OTC analgesics versus an assumed 500 metric ton Rx share, you get that 18% is OTC. If it’s half of the OTC analgesics market, you get 27%. If it’s 70%, then you get 34%.
With reasonable assumptions, the majority of acetaminophen in Japan must be prescription.