In 2018, people taking DNA tests were measured on four proxies for IQ: education (n = 1.1 million), cognitive test, typically a super quick one ( n = 257,841), self-reported math ability (n = 564,698) and hardest math class completed (n= 430,445). Thousands of genetic variants (SNPs) were discovered and they implicated the brain and neuron to neuron communication (Lee et al., 2018). A total genetic score summarizing all four variables explained 7 to 10% of the variation in cognitive ability (implying a correlation of about 0.3), and 11 to 13% of the variation in education attainment (implying a correlation of 0.35) though that might be inflated by population stratification.

While none of the four phenotypes has a high g loading, you’d expect the total of all four (depending on how they were weighted) to have a reasonable g loading, so why the dismal correlation with genes? After all twin studies promised an adult heritability of 0.75, implying a genotype phenotype correlation of 0.87!

HBD deniers argue the twin studies gave inflated results because identical twins share prenatal effects and were not raised sufficiently apart. HBD proponents claim the twin studies are valid and the missing heritability is hiding in rare genetic variants not sampled by SNPs (single nucleotide polymorphisms).

While scientists are not yet able to predict an individual’s IQ from DNA with any useful certainty, they can predict the average IQ of a group from the average DNA of the group. That’s because error at the individual level cancels out at the group level. So while the above mentioned polygenic score (which Davide Piffer named EA3 PGS) correlates only 0.3 with IQ at the individual level, it correlates 0.985 at the ethnic group level (Piffer 2021).

Further, the correlation can not be dismissed as population stratification unless one believes that the same spurious correlations that hold within Whites (on which the polygenic scores were built) also hold between wildly different ethnic groups reared in the developed World. Of course with only seven data points, we can’t be overly confident in the strength of this correlation if expanded to a larger group of populations, but we can be 95% sure the correlation is at least 0.9.

In March 2024, Piffer and his colleague Emil O. W. Kirkegaard studied the EA3s of 2,625 “European-ish ancient genomes” and produced a most remarkable chart

Eyeballing the above chart, I estimated the EA3 PGSs of each of the historical eras. Unfortunately, I can’t simply apply the line of best fit from figure 9 shown above to these historical eras because Piffer calculates EA3s as Z scores relative to specific data set and thus are incomparable across datasets. Nonetheless, Piffer has confirmed that Upper Paleolithic Europeans have educ-cognitive scores similar to contemporary Africans (Piffer X account, June 6, 2024) On a scale where U.S. & British whites average 100, Europeans average 99 and it’s estimated Black Africans would average 80 if reared in the First World (Lynn 2006). If we thus equate the PGS of Upper Paleolithic Europeans to be equal to 80 and the PGS of contemporary Europeans (1 KG Eur) to 99, we get the following estimates of how each era would score in today’s environment:

Historical era (European-ish)Estimated average EA3 PGSAverage expected IQ if reared in today’s First World (based on estimated line of best fit)
Upper Paleolithic-1.680
Neolithic-0.491
Copper Age-0.590
Bronze Age-0.392
Iron Age+0.297
Imperial+0.195
Medieval+0.0595
Contemporary Europeans (from the 1000 genome study)+0.599

Of course minor fluctuations of a few IQ points here and there are to be expected given sampling errors and should not be over-interpreted. The big takeaway is that average European-ish genetic IQ has been relatively stable over the last few thousand years.

[Update 2025-03-20: An earlier version of this article incorrectly classified 1KG as an agent genome database when in fact it is contemporary]

[Update 2025-04-26: An earlier version of the article incorrectly estimated historical IQs from the line of best fit in figure 9, which can not be done because the historical PGS were calculated with respect to a different data set]