Environmental Factors in Autism, Ranked by Evidence Quality

What the review found

Among roughly a hundred environmental exposures with enough published evidence to assess, one is described by its reviewers as likely to be causal: valproate, a medication prescribed for epilepsy and bipolar disorder, taken during pregnancy. Everything else falls somewhere between association and no association at all.

The ranking comes from Amirhossein Modabbernia, Eva Velthorst and Abraham Reichenberg at the Icahn School of Medicine at Mount Sinai, writing in Molecular Autism in 2017. They searched PubMed through December 2016 for systematic reviews and meta-analyses of environmental risk in autism and kept 32 of them, nine qualitative reviews and 23 meta-analyses, covering nine families of exposure. The paper pools no new data of its own; it grades work already published.

Each meta-analysis was scored on precision, meaning how narrow the confidence interval around the estimate was, on consistency between studies, on whether exposure had been measured directly or inferred from population averages, and on magnitude. Magnitude ran on a fixed scale: a relative risk below 2.0 counted as small, 2.0 to 5.0 as medium, above 5.0 as large.

Three exposures are grouped as unrelated to risk. Pooled across ten studies, vaccinated children showed no higher rate of autism than unvaccinated children, and those estimates were precise and consistent; a separate analysis of three cohort studies of the MMR vaccine found a non-significant decrease. Thimerosal, a preservative formerly used in some vaccines, showed precise and consistent evidence of no association across nine studies, and fifteen studies of maternal smoking during pregnancy produced the same result.

The associations that do hold are mostly obstetric, and by the review’s scale they are small. Across 27 studies, each ten-year increase in maternal age was associated with an 18 percent higher rate in the offspring, and each ten-year increase in paternal age with 21 percent. In cohort studies, pregestational diabetes was associated with a 74 percent increase and gestational diabetes with 43 percent. Maternal overweight was associated with 28 percent and obesity with 36 percent. Birth complications involving physical trauma or interrupted oxygen supply showed the strongest links of any group.

Valproate stands apart from all of it. Prospective and retrospective studies of prenatal valproate exposure found a strong association with autism that scaled with dose and survived adjustment for seizures during pregnancy and for maternal IQ, and that combination is why the reviewers were willing to use the word causal.

What this means for a parent

None of these numbers came from an experiment. They came from observational research, which records what happened to families who differed from one another in many ways at once, and the review is direct about how far that limits any reading of cause. Confounding is the standing problem. A third factor travelling with both the exposure and the outcome will manufacture an association where no causal path exists, and it will do so no matter how many families are counted.

The caesarean section finding shows the mechanism plainly. Twenty-one studies produced a 36 percent higher rate of autism after caesarean delivery, an increase that shrank once the analyses adjusted for confounding factors and disappeared entirely in the cohort studies. The association was real, and its causal reading did not survive contact with better design.

Eight studies pooled to a 50 percent increase in autism among children whose mothers took an SSRI in pregnancy. When the authors compared those mothers only against other mothers with psychiatric conditions who had not taken the medication, the increase vanished. Part of what the original studies had measured was the illness for which the medication is prescribed.

Many genetic conditions associated with autism are also associated with difficult births. A complication at delivery may therefore register something already present in a child, and an epidemiological association can run in that direction just as easily as the other.

A figure like 18 percent describes rates across a whole population of pregnancies. It is a statement about how a large group differs from another large group, and it supports no inference at all about which pregnancy produced which child. Nothing in this paper licenses a parent to look backward through their own pregnancy and find the moment.

Studies of environmental risk in autism are, in the authors’ assessment, in their infancy next to the genetics, held back by small samples, retrospective designs and exposure measured indirectly. The one exposure they are prepared to call probably causal is a prescription medication, and even there the wording stops at probably. Any claim about the cause of a child’s autism that reaches you sounding more certain than that has outrun the evidence it stands on.