Predicting Polygenic Selection for IQ
This is a linkpost for an essay by Ryan Beck appearing in the Metaculus Journal. Readers are encouraged to read the essay in the Journal where they can engage directly with the author and make their own forecasts.
Creating embryos in a petri dish, examining the genes of the resulting embryos, and then implanting the embryo with the highest probability of creating the most intelligent child sounds like something out of science fiction. But it’s closer to becoming scientific reality than some might expect. A Metaculus question asks whether genetic engineering techniques will be commercially available by 2050 that can raise IQ by 10 points. Polygenic embryo selection seems to be the most promising technology with a chance of achieving this by far. In the following essay I review the current state of the art of polygenic embryo selection and use this information to provide a forecast for this question.
Disclaimers and Acknowledgements
This analysis is informed by a fairly superficial literature review. I’ve spent a decent amount of time trying to understand the information I present below, but I have no background in genetics or biology and my understanding of the following information is still pretty weak, and I’ve only read some of the papers I’ve linked, while skimming others. While I tried to cover the general consensus and differing views related to these topics, it’s possible the following information is not representative of the current state of the art and consensus of the field. Critiques and corrections are welcome.
This originated as a Metaculus comment which was motivated by the Impactful Forecasting Prize, a really cool initiative to incentivize sharing forecasting rationales on a set of Metaculus questions judged to be impactful.
This Metaculus Journal post is an updated version of the original comment and corrects at least one error that significantly impacts my forecast. See the footnote at the end for a description of the error and see my original comment to see the prior version of this analysis with the error included.¹
I’ve also updated this analysis to incorporate feedback from @alwaysrinse and @gene_smith. Thanks to both of them for sharing their thoughts. I recommend checking out the original comment thread for their thoughts and perspectives.
Before digging into the literature my forecast on the above question was 25% based on a skepticism of high heritability estimates for IQ, a general impression that polygenic scores are still far off from having strong predictive power, and the general view that it was too optimistic to expect things to develop so rapidly. The following analysis may be biased toward these priors.
Background and Primers
This section provides primers that I think are the most helpful for anyone without previous exposure or with limited previous exposure to the topic wanting to understand the background information more thoroughly. This can be skipped for anyone who is already familiar with the subject or is primarily interested in the broader takeaways.
IQ, explained in 9 charts—If you’re not sure how important IQ really is or whether it measures something meaningful this is a great article to start with. The takeaway is that higher IQs are linked with many other positive life outcomes, such as increased wealth and reduced mortality.
Polygenic Selection of Embryos—A nice Metaculus Journal post by @galaga which gives a valuable overview of the process of selecting embryos based on polygenic scores and its use for reducing disease risk, as well as potential for increased adoption of this technology in the future.
Twin studies to GWAS: there and back again—An excellent overview of what genome-wide association studies (GWAS) are and what information they provide as well as their advantages and limitations, as well as why twin studies are used and what information they can provide.
Heritability 101: What is “Heritability”? - The basics of heritability, what it is and a broad overview of what it represents. If you’re familiar with the concept of heritability you can skip to the next one.
Heritability 201: The Types of Heritability and How We Estimate It—Excellent overview of how heritability is measured, the different types of heritability, and the differences between the heritability estimates from different sources and how they relate to one another.
Estimating Trait Heritability—Additional short reading on heritability, may be useful for additional clarity.
r² or R² — When to Use What—Really good overview of what the correlation coefficient and coefficient of determination means, this helps to get a better understanding of some of what is being measured with some of these heritability and polygenic score measurements. As far as I can tell the heritability and polygenic score measurements are the coefficients of determination of a regression on how much of the variance in the trait being measured is explained by genetic factors. However, I’m still a little hazy on what measures are being used so I advise against trusting my description in the previous sentence.
Genotype—The genetic make-up of an organism.
Phenotype—The observable characteristics or traits of an organism.
Genome—All of the genetic information of an organism.
Single Nucleotide Polymorphism (SNP) - A single genetic variation at a specific location in the genome. The variation in many different SNPs impact gene function and result in large effects on the phenotype.
Allele—One of the possible states that can occur at a variant location (for example, you could think of an SNP as a location that can have either a value of 0 or 1, and an allele is the actual value).
Genome-wide association study (GWAS) - A study that examines associations between numerous SNPs to estimate how much of the variance in phenotype can be explained by the variance in SNPs.
Heritability (h²) - The proportion of variation in a trait explained by inherited genetic variants. There are different types of heritability (broad and narrow) and the symbol used to represent it may vary by these types and by what measure of heritability is being used (for example, this paper uses hₛₙₚ² as the estimated amount of heritability that could be explained by SNPs across the population studied).
Polygenic Scores (PGSs) - A score constructed from GWAS results that “quantif[ies] an individual’s overall genetic propensity or ‘risk’ for a disorder or trait”.
Cognitive Ability/Performance, Intelligence, and Intelligence Quotient (IQ) - My understanding is cognitive ability or performance, intelligence, and IQ can be used somewhat interchangeably. Cognitive ability/performance is measured through tests thought to correlate well with intelligence, such as the SAT or a variety of other tests. Intelligence quotient (IQ) specifically refers to a test of intelligence that produces a score on a standardized scale and is referred to as a person’s IQ.
State of the Art on Heritability and Polygenic Scores
Before genome-wide association studies (GWAS), the most reliable method for predicting heritability was twin studies. Researchers would use identical (monozygotic) and fraternal (dizygotic) twins and measure how alike they were in certain traits. Twins are valuable for this because they’re born at the same time, which means each child lives in a very similar environment. Researchers can exploit the fact that identical twins have all the same genes while fraternal twins only share about half the same genes in order to estimate how much of their phenotype comes from genes. If you assume identical and fraternal twins have a similar amount of shared environmental influence, and then you compare how alike identical twins are in certain traits to how alike fraternal twins are in the same traits, you can estimate how heritable those traits are.
Twin studies have produced estimates of the heritability of IQ ranging from around 0.6 to around 0.8, meaning genetics explains about 60 to 80% of the variance in IQ. Twin studies are still used, but now they’re supplemented with other methods.
Initially, there was a lot of interest in candidate gene investigations, where specific genes were thought to correlate strongly with traits. However, these have fallen out of favor, as it’s been discovered that most traits are polygenic, meaning they’re influenced by many variations in the genome instead of just one gene variation.
Genome-wide association studies look at common single nucleotide polymorphisms (SNPs), and estimate an association between these SNPs and observed traits. These have grown in popularity as genotyping has fallen in cost and become popular. Genetic testing from companies like 23andMe have enabled large sample sizes for use with GWAS studies. As sample sizes grow larger, the estimates of heritability from GWAS grow even larger. These GWAS estimates can also be used to create polygenic scores, which can be used to predict traits that a sequenced genome will exhibit.
The largest GWAS for cognitive ability I’m aware of was based on a sample of 269,867 people, and reported a polygenic score that could explain about 5.2% of the variance in cognitive ability using an out-of-sample prediction (this means how well the polygenic score explains the variance in cognitive ability in samples it wasn’t constructed from). The paper also reported an SNP heritability of 19% for cognitive ability. As described in this article linked in the background, SNP heritability is an estimate of the true heritability based on the SNPs used in the GWAS, so it will be less than the heritability estimated by twin studies and depends on the sample size and number of SNPs examined.
Another study has reported a 9.7% polygenic score for cognitive ability, currently the highest polygenic score for cognitive ability I’m aware of. This study had a slightly smaller sample for cognitive ability (257,841 people), but used a recently developed statistical method called multi-trait analysis of GWAS (MTAG) to combine correlated phenotypes to increase the predictive power. This study used years of education, cognitive performance, self-reported math ability, and highest math class taken in combination in order to provide a better polygenic score for cognitive ability. Since these phenotypes are highly correlated, SNPs identified for one phenotype, such as math ability, are likely to be predictive of the other phenotypes, such as cognitive performance. The statistical method effectively increases the sample size by combining multiple correlated phenotypes to produce a polygenic score for cognitive ability with greater predictive power. The predictive power of these polygenic scores is expected to grow larger as sample sizes become larger.
The theoretical upper bound on the explanatory power of polygenic scores is the estimate of 60% to 80% from the twin studies. Notably, height is estimated to have a heritability of about 80% based on twin studies. A recent GWAS of height with a sample size of around 5 million people found that the GWAS SNP heritability estimate has topped out at around 50%. A polygenic score using out-of-sample estimation had a predictive power around 40%.
From the results on height, we might expect the polygenic scores for cognitive ability or IQ to grow substantially as IQ GWAS sample sizes grow, since both have a heritability estimate from twin studies topping out around 80%, and the best IQ polygenic score can explain 9.7% of the variance out-of-sample while the best for height can explain about 40%. However, IQ is much more difficult to measure than height, which could limit the speed at which the accuracy of IQ polygenic scores increase.
To date, the largest GWAS study of educational attainment, which is correlated with IQ, used 1.1 million subjects and found polygenic scores that could explain about 11% of the variance in educational attainment, an increase from the previous largest study which could explain about 3.2% of the variance using a study of just under 300 thousand subjects. The large educational attainment study is typically referred to among geneticists as “EA3”. This is also the study mentioned above that reported a 9.7% polygenic score for cognitive ability.
Caveats and Missing Heritability
The above heritability estimates aren’t perfect, and there are some important things to understand about what the differences are and why twin studies estimate much higher heritability than GWAS does. First, some basic limitations for GWAS studies:
GWAS studies typically only look at common SNPs. Common typically means SNPs that occur in 1% or more of the population. There are variants rarer than that that likely contribute to heritability.
They typically only estimate additive genetic influences. These are likely to make up the bulk of the effect, but GWAS may underestimate the heritability somewhat by ignoring non-additive effects.
They don’t perform well on different ancestries. The majority of the largest samples are based exclusively on people of European descent. The educational attainment GWAS study on 1.1 million people lost about 85% of its predictive power when tested on a sample of African Americans, and this is fairly typical.
Their polygenic scores don’t work as well within families, such as between siblings. Siblings share maybe half of their genes, while GWAS studies are a correlation across many people with different genes. Additionally, couples aren’t selected at random, they often seek out similar traits (this is called assortative mating). So the predictive power when trying to explain variance within siblings is reduced, and the reduction can be large. EA3 found within-family effects to be roughly 40% of the between-family effects for educational attainment.
They can pick up the effects of “genetic nurture”. Imagine that there are some genes that make you more likely to read to your child every day. A polygenic score on that child will pick up those genes, and if reading makes you more intelligent that polygenic score will show that those genes for reading are associated with greater intelligence. But that’s not the traditional interpretation of heritability, those genes aren’t acting through the child, the score is picking up on the genetic influence in the parent’s genes that make them more likely to read to the child. The prevalence of this effect has been estimated by constructing polygenic scores on parents for alleles that weren’t transmitted to the child. This estimate found that the effect of non-transmitted alleles on educational attainment were about 30% of the size of the effect of the transmitted alleles.
The effects described above have mixed impacts on GWAS results. The focus on common SNPs and additive influences can result in an underestimate of heritability, while effects of genetic nurture may mean they overestimate heritability. And the inability to predict well on other ancestries also poses a challenge. Overall GWAS results are very useful, but there is work to be done to quantify the impact of genetic nurture and to explain why GWAS results produce much lower estimates of heritability than twin studies. This gap between GWAS heritability and twin study heritability is commonly referred to as the “missing heritability”.
For more information this article gives an excellent overview of the different types of heritability estimates. And this section of Gwern’s overview of embryo screening for IQ describes one of the common ways to estimate SNP heritability (genome-wide complex trait analysis, GCTA), which is typically an upper bound on the polygenic scores with the current state of the art approaches. There is some debate about what explains this missing heritability. Some studies have estimated that much of the missing heritability can be explained by rare SNP variants. Others think that while rare SNP variants explain some of the missing heritability, estimates that attempt to account for these rare variants overstate the case, and that some of the missing heritability is explained by twin studies overestimating the heritability. Twin studies rely on several assumptions for accurate estimates, such as equal shared environments between identical and fraternal twins. But it’s possible that identical twins have environments that encourage similarity more than fraternal twins, which could bias twin study estimates to overestimate heritability.
The information above is important to consider in order to interpret estimates of IQ gain from embryo selection.
Estimates of IQ Gain from Embryo Selection
Three estimates have been produced so far on how much average IQ gain can be expected from embryo selection. One by Shulman and Bostrom (2014), another by Gwern (2016 initially but seems to have been occasionally updated), and one by Karavani et al. (2019). Each estimate produced smaller gains than the former. The initial estimate by Shulman and Bostrom made some very simple assumptions, such as that genetic variation can account for 50% of the variance in IQ, and then adjusted downward for within family effects. Their estimate produced an average gain in IQ of 11.5 points when selecting from ten embryos and 18.8 points when selecting from 100.
Gwern’s estimate was lower because he used SNP heritability measured from a study, which explained 33% of the variance, and he produced another estimate based on future gains in measurement accuracy in IQ testing used with GWAS studies which produced a 44% estimate. The former produces an average gain of 9 IQ points out of 10 embryos, and the latter a gain of about 11 points on average.
Karavani et al. provide a formula to theoretically estimate the gain in IQ based on polygenic score and number of embryos.¹ This formula produces similar results to Gwern’s estimate (with a polygenic score of 33% and 10 embryos they would find a gain of about 10 points, and with a polygenic score of 44% and 10 embryos they would find a gain of 11.6 points). But they also compare these theoretical results to simulated results, using actual and randomly selected couples to simulate a number of embryos. Their simulated results produce somewhat lower estimates of gain in IQ than the theoretical results. For 10 embryos and a polygenic score of 4.3% the formula would estimate an average gain of 3.6 points while the simulated results found a gain of around 3 points. With the simulated results they estimate that quadrupling the polygenic score to 17.2% and selecting out of 10 embryos would produce an average gain in IQ of about 6 points (compared to a theoretical gain from the formula of about 7.3 points).
They also provide a comparison between the expected gain from the formula and simulation for number of embryos (see Figure 2 in their paper). This estimate shows a convergence between the theoretical and simulated results as the number of embryos grows to more than 20.
Notably, Figure S3 in their paper provides an estimate of the IQ gains with larger sample sizes in future GWAS studies. It predicts saturation, or where gains in sample size no longer produce additional improvements in heritability estimates, to yield around a 7 point IQ gain when selecting from 10 embryos. However, they note that this estimate depends on the estimate of SNP heritability used, which they used as 19% from the paper mentioned previously. This is much lower than twin study heritability, which might suggest this 7 point IQ gain may not be a true upper bound at 10 embryos because estimates of SNP heritability for IQ may grow in the future.
Additionally, Karavani et al. estimate gains when selecting for multiple traits. They find that selecting on multiple traits reduces the gain of the individual traits, which makes sense. They provide a formula to estimate the reduction in gain as well. When selecting for two traits that have a correlation of 50% the gain of each trait would be about 87% of if you were selecting for individual traits. The better correlated the traits you’re selecting for are, the closer you get to having the same gain as selecting for one trait.
I generated this figure below from their approximate gain formula for IQ. It shows the average gain in IQ predicted by their formula based on the number of embryos and polygenic score. This figure can be useful for estimating the average gain in IQ possible based on assumptions about future polygenic scores and number of embryos available to select from. The 10 IQ point threshold used for this Metaculus question is shown in bold purple.
Theoretical Gain in IQ using Karavani et al. Approximate Formula
Both Gwern and Karavani suggest that the prospects for increasing the amount of viable embryos available through in vitro fertilization (IVF) in the future are not promising. They report that the actual median number of embryos created is 5, so the use of 10 embryos in these studies is already optimistic in the case of IVF. As @gene_smith points out in a comment, we’re really interested in achievable births, since some embryos don’t result in live births. If polygenic selection for IQ is done on 10 embryos, but the top scoring embryo doesn’t result in a live birth then the expected gain would be reduced. However, number of embryos has been used throughout this essay as a shorthand for the number of viable embryos assuming all embryos would result in a live birth.
The real case for optimism in the number of embryos available is from in vitro gametogenesis. This is a process where eggs could be developed from stem cells in a petri dish, which has the potential to make a large amount of eggs available for embryo selection. In vitro gametogenesis could potentially even make iterated embryo selection possible, where eggs are generated to create embryos, the optimal embryo is selected using a polygenic score for a desired trait or traits, and then stem cells are harvested from the optimal embryo to generate new eggs and another round of embryos. This could lead to a large increase in gains as multiple rounds of selection could make additive improvements before an embryo even develops into a child.
I’ve done significantly less research on this topic, but as far as I can tell the current state of the art is that in vitro gametogenesis has been successfully performed in mice for the first time in 2021 (Takashi Yoshino et al 2021). Previously stages of the in vitro gametogenesis process had been completed in mice, but to successfully develop the generated germ cells into eggs they had to be implanted into a live mouse. But the latest paper shows that the entire process was completed in vitro for the first time, generating embryos that were then implanted into mice and birthed into offspring who survived to adulthood.
Based on the above information and ignoring political and social factors for the time being, the factors that matter are how large the largest polygenic scores are and the number of embryos that can be used during the process. The paper by Karavani et al suggests that polygenic scores for IQ could become saturated at about 7 points of average gain. However, this may not be a realistic upper bound on the contribution of polygenic scores because estimates of SNP heritability for cognitive ability may become larger in the future. Additionally, whole genome sequencing and identification of rare variants could potentially eliminate the missing heritability and allow polygenic scores to approach the twin heritability threshold.
To be able to achieve 10 IQ points of average gain a sufficient combination of polygenic score and embryos available to select from has to be available. I’ve selected three scenarios, shown below added on to the figure I previously provided. In my view these scenarios are not independent at all, so the probabilities shouldn’t be added together. They’re simply estimates of the progress of technology in embryo production and polygenic scores given the information I described above. Therefore I’m taking the highest probability of these scenarios as the technological portion of my forecast, or the portion not accounting for political and social factors.
Theoretical Gain in IQ using Karavani et al. Approximate Formula, Annotated with Scenarios
Scenario 1: No additional gains in the number of embryos available to select from materialize, but polygenic scores have eliminated most of the missing heritability and are approaching the estimates from twin heritability, explaining around 50% of the variance in cognitive ability.
Polygenic scores for IQ have reached about 50% predictive power (40%) AND
Selection from at least 5 viable embryos is possible (100%, this is the status quo)
Scenario 2: Some technology exists that allows for selection from 10 embryos, and polygenic scores are able to predict just under 35% of the variance in IQ.
Polygenic scores for IQ have reached about 35% predictive power (80%) AND
Some technology is available that provides at least 10 embryos to select from, an increase from the current median of 5 (90%).
Scenario 3: Polygenic scores for IQ have reached about 20% predictive power, which is near the current estimates for SNP heritability, and in vitro gametogenesis or some other technology is able to provide around 45 embryos or more to select from.
Polygenic scores for IQ have reached about 20% predictive power (90%) AND
In vitro gametogenesis has to be able to generate healthy human embryos by 2050 (90% - deferring to the crowd on this question) AND
Must be able to generate 45 embryos to select from (90%)
Scenario 1 works out to 40% chance, scenario 2 works out to a 72% chance, and scenario 3 works out to a 73% chance. For scenario 1 I’m at 40% on polygenic scores reaching 50% predictive power. I’m mostly just skeptical that polygenic scores reach such a high level of predictive power, polygenic scores for height reached saturation at about 40% predictive power, and while further gains may be made it’s still quite a significant distance from the 80% heritability estimate of height from twin studies. Estimates for IQ will continue to lag those of height due to the comparative difficulty in getting large samples. Additionally, I find the argument that twin studies may be overestimating twin heritability of cognitive ability somewhat persuasive.
However in scenario 2 I’m at 80% chance of predictive power reaching 35%. This seems likely due to the example of polygenic scores for height reaching 40%, twin heritability estimates are in the same ballpark for height and cognitive ability, so within the next few decades I think it’s likely that cognitive ability polygenic scores reach the saturation polygenic score estimated for height. But I have included some uncertainty because the supplemental material to the paper estimating a 5.2% polygenic score for IQ says the following on page 34:
Surprisingly, the proportional increases in number of associated loci found with this sample size relative to previous studies (205 versus 18 loci with sample N=269,867 versus 78,303) did not translate to a similarly large increase in variance explained by PGS (r² =5.2% versus 4.8%).
Hard to know why the gain in polygenic score is so small from a pretty substantial increase in sample size, but it leads me to deduct a bit from the possibility of reaching saturation. But multi-trait analysis of GWAS (MTAG) and other new approaches likely to be developed in the future will probably continue to increase the amount of variance explained even if larger sample sizes alone aren’t sufficient.² In scenario 2 I also think there’s a 90% chance that by 2050 it’ll be possible to select from at least 10 embryos, whether that’s due to in vitro gametogenesis or just smaller incremental improvements in IVF processes.
For scenario 3 I’ve assigned a 90% chance of polygenic scores reaching 20% predictive power, 90% chance that in vitro gametogenesis will be a viable method to create human embryos, and a 90% chance that in vitro gametogenesis will be able to generate 45 embryos conditional upon it being a viable method to create embryos. I think 20% predictive power is very achievable given the higher polygenic scores for height and the fact that the field has several decades to advance. My forecast for a live birth from in vitro gametogenesis is only slightly more pessimistic than the crowd, and I don’t feel that I have any particular insight here so I’ve deferred to the crowd’s forecast of around 90% chance by 2050. And lastly I think 45 embryos is doable if in vitro gametogenesis can produce viable embryos, since I’m ignoring cost and just focusing on technology for this portion of my forecast I think there’s around a 10% chance that, conditional on in vitro gametogenesis producing viable embryos, it’s either too time consuming or challenging to produce 45 embryos for use with polygenic prediction or it isn’t developed enough to be available for the general public.
The maximum probability from each of the scenarios was 73% from scenario 3, so I’m using 73% as my forecast that a 10 point gain in IQ from embryo selection will be possible by 2050 from a purely technological perspective.
Political and Social Factors
For political and social factors, there are several concerns. One is the substantial decrease in predictive power when polygenic scores are extended to other ancestries. If polygenic scores for non-European ancestry remain significantly behind those for European ancestry due to smaller sample sizes I have a very hard time seeing companies offering selection for IQ. I imagine that offering large gains in IQ only to couples of European ancestry would not sit well with the public. However, I think there’s probably a 90% chance that GWAS data from other ancestries will be able to provide polygenic scores close enough to European polygenic scores to make little difference by 2050.
Another factor is cost. We may have the technology to use in vitro gametogenesis to create a few hundred or thousand embryos in the lab, but how expensive will that be? Additionally, each of those embryos must be genotyped in order to assess their polygenic score, which even if costs of genotyping come down, these could add up. Gwern provides an estimate of the costs and benefits and finds a fairly sizeable positive benefit to large IQ gains, but I’m not sure the incentives align for the parents. Yes your child may benefit financially from an increased IQ, and society as well, but will anyone besides rich parents think that’s worth the cost?
Lastly is political and social acceptance. This is hard to estimate, but I’ll go with about 80% chance that at least one country would be willing to allow this if the science was available.
Trying to combine all those political and social factors would either be impossible or very complicated, so I’m just going to roughly say there is a 70% chance of the political and social factors I mentioned above aligning so that if the science existed it would be adopted — while cost-effective enough to satisfy this question.
Another factor I haven’t included above is multiple trait selection. Selecting solely on high gains to IQ may be discovered to adversely select for certain diseases or other undesirable conditions. And prospective parents are likely to find other traits desirable. I think it’s fairly likely that any services for trait selection that include IQ would also select for several other traits, reducing the gains to IQ somewhat. I’d say I’m 75% confident in that. This is a social factor in the sense that it depends on how society views the costs and benefits of large gains in IQ compared to selecting for other traits, but it’s a social factor that limits the technological gain so I’ll account for it separately.³
For scenario 3 I came up with a 73% probability. I’m going to round that down to 70% to roughly account for the effect of multiple selection reducing the gains to IQ. Combining that with the 70% political and social factors estimate produces a final forecast of 49%.
I think there’s a 49% chance that technology offering a 10 point average gain in IQ using genetic engineering will be commercially available by 2050. Thirty years is a fairly long time, but it can be easy to overestimate the speed at which the science may develop, particularly for scientific developments with significant social impact. For example, we seem to have the technology to clone humans, but that has yet to happen.
Still, my forecast suggests the chances of this happening by 2050 are about the same as a coin flip, which is a pretty significant increase over my initial 25% forecast before researching. If polygenic selection for IQ becomes commercially available and is adopted by even a small fraction of the population, it could significantly increase the number of eminent scientists and innovators and potentially lead to higher levels of economic growth. And even if this technology isn’t in use by 2050 it’s easy to imagine it happening sometime this century, leaving open the possibility of a transformative impact from polygenic selection in the not too distant future.
¹ I made an error in my initial comment where I, being a dumb engineer who typically sees the notation log = log₁₀ and ln = natural log, assumed that the log in Karavani et al.’s formula for expected gain in IQ was log base 10, but in reality it was the natural log. I had initially noticed my estimates were slightly below theirs, but it seemed pretty minor so I assumed they might be using their more precise formula (they provided an approximate formula and said they also used a more precise estimate). After looking again I noticed their theoretical estimates from the formulas were higher than the simulated estimates and higher than what I was calculating to the point that I figured I must be doing something wrong, so I dug through their background material and found that I was supposed to be using the natural log. For the sake of transparency I haven’t edited my original comment except to call out in bold text that an error was impacting my analysis and to provide some brief additional info. If you want to look back and see how that mistake impacted my forecast and what I’ve changed since then you can compare this notebook to my original comment. The primary effect was that my table of average IQ gains underestimated the gain, so I replaced it with the charts in this analysis which more clearly shows the relationship and provides a lower threshold to achieve a gain of 10 IQ than my previous table would have suggested.
² The next educational attainment GWAS is expected this year and has increased the sample size from 1.1 million to 3 million. This will likely shed quite a bit of light on how much predictive power and heritability for cognitive ability may increase as sample sizes increase. Some preliminary discussion of the results is available here.
³ My read of the resolution criteria is that a company will probably have to be able to demonstrate a 10 point average gain in IQ, something I think it would be unlikely to be able to do if in practice no one ever selects solely for IQ. However, I’m not very confident in this interpretation so my reduction for this effect in my forecast is pretty minor.