367 - Tylenol, pregnancy, and autism: What recent studies show and how to interpret the data

KEYPOINTS

• 01

  Peter's four-step framework for any 'X causes Y' headline.

  First, confirm there really is a statistical association. Second, and hardest, estimate how likely it is to be causal, using sensitivity analyses and falsification tests, because 'there are no proofs in biology'. Third, if it's probably causal, judge the effect size, since a real but tiny effect may not change behavior. Fourth, be ready to update as new data arrive. Peter adds that complex conditions like autism almost never have a single cause, and that humans are 'not wired to think scientifically'.

• 02

  Autism prevalence rose about fivefold, which makes it easy to find spurious culprits.

  CDC figures went from 6.7 per 1,000 children in 2000 to 32.2 per 1,000 recently. Searching countless variables runs into the multiple comparisons problem. Peter's example: with a 7-out-of-10 coin-flip 'psychic test' (about a 5% chance by luck), test 100 people and about five will 'pass'. The Spurious Correlations website finds a 98.5% correlation between margarine consumption and Maine's divorce rate.

• 03

  The August review in Environmental Health overstated the evidence.

  The review, led by Andrea Baccarelli, was not a meta-analysis. It said six autism studies 'consistently' showed a positive link with dose-response in four of five, but two showed no significant association and only three directly tested dose. When Peter's team pooled the risk ratios themselves, the overall association was about a 5% relative increase, and the pooled result essentially matched the Swedish study, which had more than 10 times as many participants as all the others combined.

• 04

  The study with the strongest signal (Ji, 2020) has major limitations.

  It sorted about 1,000 Boston Birth Cohort children into thirds by acetaminophen in umbilical cord blood. But acetaminophen clears within roughly 24 hours, so the measure mainly shows use just before delivery, like estimating a year's doughnuts from your birthday. About 70% of samples had no detectable metabolites. Only around 3,000 of nearly 9,000 pairs stayed in follow-up, and autism prevalence in the analyzed group was about 11%, versus about 3% in the general population. Peter sees that selective retention as the single biggest problem.

• 05

  In the two largest cohorts, comparing siblings eliminated the association.

  Sweden (Ahlqvist et al., 2024; about 2.5 million children, median 13.5 years of follow-up) found a hazard ratio of 1.05 (1.02–1.08), an absolute increase of 0.09% at 10 years. Among siblings who differed in exposure, the effect disappeared. The authors called it 'non-causal'. A recent Japanese cohort of about 220,000 children, roughly 40% of them exposed (closer to US use than Sweden's about 7.5%), found a non-significant 6% (0.98–1.15) that also vanished with sibling control. Peter says the review's claim that the Swedish siblings were selected as doubly discordant would create collider bias, and the senior author says that's not how it was done.

• 06

  Most of the Bradford Hill criteria come out weak or against causation.

  Strength is weak: 1.05 against the 1.5 threshold for pharmacoepidemiology, and 10x for smoking and lung cancer. It's weaker than red meat and diabetes (1.10) or exercise and prostate cancer (1.05), both likely confounded. Consistency is moderate. Specificity is weak. Temporality holds, though Baccarelli and William Parker disagree on the critical window. The dose gradient is inconsistent. The prostaglandin mechanism is 'hand-wavy'. Analogy argues against it: aspirin, another prostaglandin inhibitor, showed about a 13% lower autism risk in the Swedish sibling analysis.

• 07

  Genetics and changes in diagnosis explain most of the rise.

  Heritability of autism is about 80–90%, versus about 80% for height and 60% for body weight, estimated from identical vs fraternal twin comparisons. Expanded criteria (DSM-III-R in 1987, the PDD family, then the single ASD umbrella in DSM-5 in 2013) are estimated to account for 40–60% of the increase, and greater awareness for 20–30%. Older parents account for 5–15%: average paternal age rose from 27.6 to 31.1 and fathers over 40 went from 4.1% to 8.9%. Maternal obesity, preterm birth and PM2.5 air pollution cover much of the rest.

• 08

  Fever and infection are themselves risks, and they may explain part of the link.

  Maternal infection is associated with a 25–40% higher autism risk and maternal fever with up to 200%. Fever, especially in the first trimester, is associated with 25–200% higher risk of cleft palate and neural tube defects. Since acetaminophen is the safest fever reducer in pregnancy (NSAIDs are category D in the third trimester because they can close the ductus arteriosus too early), its use may simply be marking infection.