Assortative mating is when similar people marry and have children. Some people worry about assortative mating in Silicon Valley: highly analytical tech workers marry other highly analytical tech workers. If highly analytical tech workers have more autism risk genes than the general population, assortative mating could put their children at very high risk of autism. How concerned should this make us?
Methods / Sample Characteristics
I used the 2020 Slate Star Codex survey to investigate this question. It had 8,043 respondents selected for being interested in a highly analytical blog about topics like science and economics. The blog is associated with – and draws many of its readers from – the rationalist and effective altruist movements, both highly analytical. More than half of respondents worked in programming, engineering, math, or physics. 79% described themselves as atheist or agnostic. 65% described themselves as more interested in STEM than the humanities; only 15% said the opposite.
According to Kogan et al (2018), about 2.5% of US children are currently diagnosed with autism spectrum disorders. The difference between “autism” and “autism spectrum disorder” is complicated, shifts frequently, and is not very well-known to the public; this piece will treat them interchangeably from here on. There are no surveys of what percent of adults are diagnosed with autism; it is probably lower since most diagnoses happen during childhood and the condition was less appreciated in past decades. These numbers may be affected by parents’ education level and social class; one study shows that children in wealthy neighborhoods were up to twice as likely to get diagnosed as poorer children.
Given that respondents are likely wealthier than average, we might expect a rate of 2.5% – 5%. Instead the rate is noticeably higher than that, consistent with the hypothesis that this sample will be more autistic than average. About 4% of the SSC survey sample had a formal diagnosis of autism, but this rose to 6% when the sample was limited to people below 30, and to 8% below 20. This sample is plausibly about 2-3x more autistic than the US population. Childhood social class was not found to have a significant effect on autism status in this sample.
I tried to get information on how many children respondents have, but I forgot to ask important questions about age until a quarter of the way through the survey. I want to make sure I’m only catching children old enough that their autism would have been diagnosed, so the information below (except when otherwise noted) comes from the three-quarters of the sample where I have good age information. I also checked it against the whole sample and it didn’t make a difference.
Of this limited sample, 1,204 individual parents had a total of 2,459 children. 1,892 of those children were older than 3, and 1,604 were older than 5. I chose to analyze children older than 3, since autism generally becomes detectable around 2.
71 children in the 1,892 child sample had formal diagnoses of autism, for a total prevalence of 3.7%. When parents were asked to include children who were not formally diagnosed but who they thought had the condition, this increased to 99 children, or a 5.2% prevalence. Both numbers are much lower than the 8% prevalence in young people in the sample.
What about marriages where both partners were highly analytical? My proxy for this was the following survey question:
I’ll be referring to these answers as “yes”, “sort of”, and “no” from here on, and moving back to the full sample. 938 parents answered this question; 51 (5.4%) yes, 233 (24.8%) sort of, and 653 (69.4%) no. Keep in mind the effective sample is even smaller, since both partners in two-partners-read-SSC-families may have filled out the survey individually about the same set of children (though this should not have affected the “sort of” group). Here is the autism rate for each group, with 95% confidence interval in black:
There is little difference. If we combine the latter two groups, the confidence interval narrows slightly, to 2.7 – 6.5.
I asked respondents about the severity of their children’s autism.
People who hadn’t previously reported any children with autism gave answers other than N/A for this one, which was confusing. Instead of the 71 children we had before, now it’s up to 144 children. I’m not sure what’s going on here. Of these phantom children, 101 had mild cases, 31 moderate, and only 12 severe. Severe autism was only present in 0.6% of the children in the sample. There was no tendency for couples where both partners were highly analytical to have children with more severe autism.
Autism rates in this survey were generally low. Although the general rate of 3.7% was higher than the commonly-estimated US base rate of 2.5%, this is consistent with the slight elevation of autism observed in higher social classes.
There was no sign of elevated risk when both partners were highly analytical. The sample size was too small to say for certain that no such elevation exists, but it can say with 95% confidence that the elevated risk is less than three percentage points.
This suggests that the answer to the original question – does assortative mating between highly analytical people significantly increase chance of autism in offspring – is at least a qualified “no”.
Why should this be? It could just be that regression to the mean is more important in this case than any negative effects from combining recessive genes or mixing too many risk genes together. Or maybe we should challenge the assumption that being a highly analytical programmer is necessarily on a continuum with autism. It seems like p(highly analytical|on autism spectrum) is pretty high, but p(on autism spectrum|is highly analytical) might be much lower.
Obvious limitations of this survey include the small sample size of both-partners-highly-analytical couples, the weak operationalization of highly analytical as “member of the SSC, rationalist, and effective altruist communities”, and the inability to separate non-autistic children from children who are not yet diagnosed. Due to these limitations, this should only be viewed as providing evidence against the strongest versions of the assortative mating hypothesis, where it might increase risk by double, triple, or more. Smaller elevations of risk remain plausible and would require larger studies to assess.
I welcome people trying to replicate or expand on these results. All of the data used in this post are freely available and can be downloaded here.
If we take:
p(on autism spectrum) = 2.5%
p(highly analytical) = 10% (i.e. the top 10% most analytical people)
Then we get (via good ol' Bayes):
p(highly analytical|on autism spectrum) = 4 x p(on autism spectrum|is highly analytical)
For simplicity's sake, let:
p(highly analytical|on autism spectrum) = 80%
p(on autism spectrum|is highly analytical) = 20%
Which means that the probability that 2 randomly selected highly analytical people are both on the autism spectrum is 1 in 25 (compared to 1 in 1,600 in the general population).
1204 x 0.3 = 360 couples with children in the sample were both defined as highly analytical, so we have ~15 couples where both are on the autism spectrum. This may explain why any effect is too small to detect.
Obviously these are made up numbers and the actual boundaries of the categories are more fluid than that but this does seem like a plausible explanation.
Thanks. I was recently worrying a bit about this specific issue, and this is reassuring.