Tylenol carries no added risk for autism

Neuroscience, epidemiology, and a damaging public statement

In my capacity as an expert in neuroscience, autism, and statistics, I spoke extensively with a reporter from the Harvard Crimson, Ella Niederhelman. She was reporting on a study cited by Donald Trump/Robert Kennedy Jr. in their press conference, in which they falsely claimed that maternal Tylenol use causes autism. The piece by her and Abigail Gerstein is excellent; go read it.

The Crimson’s headline focused on payments received by the senior author of the study, Harvard Dean of Public Health Andrea Baccarelli. I had missed this point. However, we do not need to get into claims of public corruption. Instead, we can simply review the actual scientific evidence.

Baccarelli once gave paid testimony in a federal lawsuit, and later expanded his testimony into a scientific review article, co-authored with Diddier Prada, Beate Ritz, and Ann Z. Bauer. That article is fairly transparent in its details, which made it possible to examine the original research that he and his co-authors relied upon. I have done that in a neutral manner. In contrast to them, I reach the following conclusion: maternal use of Tylenol/acetaminophen carries no measurable risk of autism spectrum disorder to the baby.

This does leave open the interesting question of how it was that Prada/Ritz/Bauer/Baccarelli were able to reach the opposite conclusion. They wrote: “Our analyses…support evidence consistent with an association between acetaminophen exposure during pregnancy and increased incidence of [neurodevelopmental disorders].” My inspection of their argument suggests that in the case of ASD, they took an effect that, were it to exist, would be so small as to be extremely hard to measure, and then “rounded up” the estimate where they could.

The published article appears to reflect previous testimony in a federal lawsuit by Baccarelli. The judge in that case found that Baccarelli’s testimony failed to meet standards of scientific expertise. She listed, in detail, ways in which he cherrypicked data. Go read her decision (do a ctrl-F search for “Baccarelli”). It is not flattering. Then again, for $150,000 one might put up with a lot.

OK, now let us get into the weeds.

Putting risk levels into perspective

Done well, modern epidemiology research is amazing. Using birth records in Scandinavia, Japan, or other countries where detailed information is available, it is possible to measure risks small enough that they would normally escape detection. Such research requires hundreds of thousands, or even millions, or subjects. It is a testament to modern statistical methods.

But the same research methods mean that we can now identify risks so small that, while statistically significant, they have no bearing on the practicalities of everyday life. If a risk is small enough, the question becomes a matter of asking: did the researchers choose a genuinely important topic, or one where they had the technical means to do the study?

At this point, the study of acetaminophen (“Tylenol”) for autism and other neurodevelopmental disorders has become such a topic. To understand why, let’s go back to an article I wrote years ago, “How To Think About The Risk Of Autism” (gift link here). It has held up well.

How To Think About The Risk Of Autism

Even though the article is 10 years old, the general understanding of what causes autism is unchanged: autism’s causes are mostly genetic, arising from combinations of genes, with some additional contributions from maternal stress-related factors in the third trimester. If you want to know what causes autism at a general level, you can stop here. No need for some federal report by charlatans.

Of course, there are lots of details. Which genes increase risk? Which environmental events contribute? What brain systems are affected? These are areas of active research.

Now, let’s get into the main evidence offered by Baccarelli - and how to work with this kind of data responsibly.

Thinking about risk ratios

Risk ratios can be eye-glazing. But here we go. Let’s just say that when 2.5 million kids are studied, and it’s still not possible to detect the added risk from taking acetaminophen...we can act like there’s no added risk.

The basic concept is that we can take the rate in the general population, and then use it as a benchmark to see what the *true* added risk is from a possible cause. As I wrote in 2014:

Over the last few years, we’ve seen an explosion of studies linking autism to a wide variety of genetic and environmental factors. Putting these studies in perspective is an enormous challenge. In a database search of more than 34,000 scientific publications mentioning autism since its first description in 1943, over half have come since 2008.

As a statistically minded neuroscientist, I suggest a different approach that relies on a concept we are familiar with: relative odds. As a single common measuring stick to compare odds, I have chosen the “risk ratio,” a measure that allows the bigger picture to come into focus.

For a variety of studies I asked the same question: How large is the increased risk for autism? My standard for comparison was the likelihood in the general population of autism spectrum disorder. Here’s an example. Start from the fact that the recorded rate of autism is now 1 in 68, according to a report released last week by the Centers for Disease Control and Prevention. If babies born in purple farmhouses have a rate of autism of 2 in 68, this doubling means that the purple farmhouse carries a risk ratio of 2. However, correlation is not causation, and there is no need to repaint that farmhouse just yet.

Here’s the key chart from that piece. Note the really important risks: genetic inheritance, third trimester stress, and factors that can affect maternal/paternal DNA.

But also note the studies where there is no identified risk: antidepressant use, sonograms, and vaccination. All of these are very well tested.

Now, let’s look at Baccarelli’s argument that Tylenol has some link to autism. He and his co-authors present a summary of 8 studies; I give their full article here for free. However, they have misstated the conclusion of the largest study, of 2.5 *million* kids by Ahlqvist et al. (2024). Baccarelli claims that they found a risk ratio of 1.05 (5% increased risk). Already, this is tiny: it’s far smaller than even the smallest risk in the chart, which is 1.30 (see the row that says “+ 30”).

However, that’s the risk before correcting for confounds. Once that correction is done, the risk goes to zero.

What’s a confound? Basically, people take acetaminophen for a reason, and those reasons might also be causes of later problems. In a disorder like autism that is known to be genetic and have environmental causes, many confounds are possible: gene defects correlated with the mother’s health problems, pain, and infection.

Baccarelli and co-authors admitted that they can’t rule out that “...there is, as always, a possibility for residual confounding.” Yet, when given the choice, they highlighted a risk ratio that was unadjusted for confounds.

To quote Mad Men: not great, Bob.

Here, the second entry shows what happened when Ahlqvist compared kids with their siblings, thus removing familial confounds. Adjusted, the risk ratio is now 0.98, not distinguishable from 1.00. In other words, a it’s 2% decreased risk (but that’s nominal - in fact it’s not significant). See the Ahlqvist article.

Really into the weeds: a funnel plot

One last foray into the weeds.

Baccarelli has misinterpreted great studies (Ahlqvist et al., 2024, Liew et al., 2016) and given equal credence to studies with as few as 215 subjects (seriously, go look at their Table 2). But when you put it all together, the data converge upon a negative result.

This graph is called a funnel plot. Where one study reported more than one risk ratio, I tried to choose what the authors thought was their most conclusive and general finding. The best true estimate of the risk is at the top of the triangle. And it’s super-close to 1.00.

A final note

Under normal conditions, this kind of work could be done by a federally-convened panel of experts. But when the Department of Health and Human Services, the Centers for Disease Control, and the National Institutes of Health all under attack from within, it became necessary to pitch in. Thank you for going on this deep dive with me.