33 ms·

taking all of the associations into account and making a polygenic predictor score "does a lousy job predicting the outcome for any specific individual, but it

taking all of the associations into account and making a polygenic predictor score "does a lousy job predicting the outcome for any specific individual, but it can explain 11 percent of the population-wide variation in years of schooling." only 11 percent. this tells us that the remainder is more likely to be attributable to environmental factors. in other words, even with a view of the genes, we would need to encourage education more strenuously if we wanted to improve educational outcomes for society.

deugtniet 8y ago
Full disclosure: I have attended multiple presentation on this paper by some of these authors, and I'm doing a PhD in quantitative Genetics. So I like to think I have something to add. But I'm not an expert on educational attainment, so correct me if I'm wrong. The 11% of variance that can be predicted using the current information of a fairly bad genetic predictor. The predictor is bad because many cohorts used in this study do not share all their information, but only share so called summary statistics. Which are combined into a prediction model. If all the individual level data is available, it becomes possible to get a much better prediction model. In addition there are ways to estimate how much can be predicted from genetic information, a term called heritability. Heritability can be estimated from so called twin studies. Where twins who are genetically identical (monozygotic) are compared to twins who are not genetically identical (dizygotic). In this way you can measure the environmental effect (the variation seen in the monozygotic twins) and the genetic effect (the variation that is seen in the dizygotic twins compared to the monozygotic twins). This is, of course, a gross oversimplification, but can provide reasonable measures of an upper bound of what can be explained genetically For most traits this upper bound is around 0.5 R^2, and I believe for educational attainment a bit higher, but I could not find it in the paper or a quick google search. Another factor in this case is that there are unmeasured genotypes which can explain some extra variation. There are also solutions for this. Which are able to explain the upper bound of what can be predicted from the data at hand. For instance meaning that you can find 80% of all the genetic variation using the genetic measurements that you have. But you often need individual level data for this. Something which cohorts are not often prepared to share. Another complicating factor is the heterogeneous nature of the phenotype: educational attainment is different per country. So it's hard to say what's really being measured. It's likely very different in France, compared to say The Netherlands. Due to the differing nature of their educational systems. Personally I'm morally conflicted by these studies which measure intelligence. It's very interesting to see the practical outcomes from the paper. I've seen an example where it's well powered to stratify high educational attainment and low educational attainment individuals for social science studies. But I'm of the opinion that this may be a slippery slope towards misinterpretation, as we currently see in the whole IQ business. As well as too strong an emphasis on being 'smart' in certain social groups. We know Intelligence (or whatever is being measured) is heritable. But I'm not fond of the implications which people attach to these studies. Even though I know people are going to act on it. For the earliest examples of this, I'm going to look at certain east Asian countries where ethics may be very different compared to the western views.
- millermp12 8y ago
  "Personally I'm morally conflicted by these studies which measure intelligence." Are you, or your cohort fighting on the side of the angels, as morally conflicted by the treatment received by Arthur Jensen, Charles Murray et al? It's less a "moral conflict" than the dawning realization that 50 years of New Age psychobabble (aka New Left "social sciences") are coming to an end.
- oldgradstudent 8y ago
  > In addition there are ways to estimate how much can be predicted from genetic information, a term called heritability. Heritability can be estimated from so called twin studies. Where twins who are genetically identical (monozygotic) are compared to twins who are not genetically identical (dizygotic). In this way you can measure the environmental effect (the variation seen in the monozygotic twins) and the genetic effect (the variation that is seen in the dizygotic twins compared to the monozygotic twins). This is, of course, a gross oversimplification, but can provide reasonable measures of an upper bound of what can be explained genetically For most traits this upper bound is around 0.5 R^2, and I believe for educational attainment a bit higher, but I could not find it in the paper or a quick google search. Just wondering, have you ever encountered any of the serious critiques of twin studies. The comparison of dizygotic (DZ) and monozygotic (MZ) twins relies on the Equal Environment Assumption (EEA) - which assumes that the environment experienced by MZ twins is the same as that of DZ twins. This assumption has been shown empirically to be false - parents, siblings, teachers, friends, and the twins themselves treat MZ twins very differently than they treat DZ twins. There are good reasons to think that heritability results derived from twin studies are highly inflated.
  - danieltillett 8y ago
    No it doesn’t at least for g. MZ twins reared together show as much variation as those reared in different families (at least by adulthood).
gwern 8y ago
> only 11 percent. this tells us that the remainder is more likely to be attributable to environmental factors. No, it doesn't. It merely tells us that with the data up to now, we've reached 11% (up from, incidentally, ~0.3% just 5 years ago). It provides a very loose lower bound on genes. Saying that is like saying, 'our software testing up to now has only revealed bugs in 11 places, therefore, every other line of code must be bug-free!' Even if you knew nothing about the actual heritability estimates, one would think that finding so much so quickly would make one suspect that there's a lot more to find...
- the8472 8y ago
  Plus they focus on SNPs, not other types of gene variants.
- cryoshon 8y ago
  there's always more to find, but consider that this model is leaving out the environment entirely and its predictive value is even less than you might think at first glance. we know the environment impacts gene transcription, and we know that genetic determinism (as far as outcome phenotypes) is not a thing. likewise, there may be epigenetic factors at play which wouldn't be captured in a GWAS yet still might (or might not) have an impact on outcomes. it isn't a "the other lines of code are bug free" situation so much as a "we expect there to be an abundance of bugs with undefined consequences outside of the subset which we have examined which most likely outweigh the stuff we found"
toasterlovin 8y ago
I wouldn't write off genetics. GWASes, as a technique, are still pretty new and the number of genomes available for GWASes is still relatively small. Plus, twin studies point to traits like educational attainment being >11% heritable.
- cryoshon 8y ago
  GWAS are great, don't get me wrong. but there's always a lot of jumping to conclusions that is done whenever a GWAS finds something interesting. afterwards, you find the people wishing that they could CRISPR themselves to fit the pattern found in the GWAS, etc. you can see that in the article the authors of the study specifically went out of their way to try to caution people against doing that.

deugtniet 8y ago

Full disclosure: I have attended multiple presentation on this paper by some of these authors, and I'm doing a PhD in quantitative Genetics. So I like to think I have something to add. But I'm not an expert on educational attainment, so correct me if I'm wrong. The 11% of variance that can be predicted using the current information of a fairly bad genetic predictor. The predictor is bad because many cohorts used in this study do not share all their information, but only share so called summary statistics. Which are combined into a prediction model. If all the individual level data is available, it becomes possible to get a much better prediction model. In addition there are ways to estimate how much can be predicted from genetic information, a term called heritability. Heritability can be estimated from so called twin studies. Where twins who are genetically identical (monozygotic) are compared to twins who are not genetically identical (dizygotic). In this way you can measure the environmental effect (the variation seen in the monozygotic twins) and the genetic effect (the variation that is seen in the dizygotic twins compared to the monozygotic twins). This is, of course, a gross oversimplification, but can provide reasonable measures of an upper bound of what can be explained genetically For most traits this upper bound is around 0.5 R^2, and I believe for educational attainment a bit higher, but I could not find it in the paper or a quick google search. Another factor in this case is that there are unmeasured genotypes which can explain some extra variation. There are also solutions for this. Which are able to explain the upper bound of what can be predicted from the data at hand. For instance meaning that you can find 80% of all the genetic variation using the genetic measurements that you have. But you often need individual level data for this. Something which cohorts are not often prepared to share. Another complicating factor is the heterogeneous nature of the phenotype: educational attainment is different per country. So it's hard to say what's really being measured. It's likely very different in France, compared to say The Netherlands. Due to the differing nature of their educational systems. Personally I'm morally conflicted by these studies which measure intelligence. It's very interesting to see the practical outcomes from the paper. I've seen an example where it's well powered to stratify high educational attainment and low educational attainment individuals for social science studies. But I'm of the opinion that this may be a slippery slope towards misinterpretation, as we currently see in the whole IQ business. As well as too strong an emphasis on being 'smart' in certain social groups. We know Intelligence (or whatever is being measured) is heritable. But I'm not fond of the implications which people attach to these studies. Even though I know people are going to act on it. For the earliest examples of this, I'm going to look at certain east Asian countries where ethics may be very different compared to the western views.

millermp12 8y ago
"Personally I'm morally conflicted by these studies which measure intelligence." Are you, or your cohort fighting on the side of the angels, as morally conflicted by the treatment received by Arthur Jensen, Charles Murray et al? It's less a "moral conflict" than the dawning realization that 50 years of New Age psychobabble (aka New Left "social sciences") are coming to an end.
oldgradstudent 8y ago
> In addition there are ways to estimate how much can be predicted from genetic information, a term called heritability. Heritability can be estimated from so called twin studies. Where twins who are genetically identical (monozygotic) are compared to twins who are not genetically identical (dizygotic). In this way you can measure the environmental effect (the variation seen in the monozygotic twins) and the genetic effect (the variation that is seen in the dizygotic twins compared to the monozygotic twins). This is, of course, a gross oversimplification, but can provide reasonable measures of an upper bound of what can be explained genetically For most traits this upper bound is around 0.5 R^2, and I believe for educational attainment a bit higher, but I could not find it in the paper or a quick google search. Just wondering, have you ever encountered any of the serious critiques of twin studies. The comparison of dizygotic (DZ) and monozygotic (MZ) twins relies on the Equal Environment Assumption (EEA) - which assumes that the environment experienced by MZ twins is the same as that of DZ twins. This assumption has been shown empirically to be false - parents, siblings, teachers, friends, and the twins themselves treat MZ twins very differently than they treat DZ twins. There are good reasons to think that heritability results derived from twin studies are highly inflated.
- danieltillett 8y ago
  No it doesn’t at least for g. MZ twins reared together show as much variation as those reared in different families (at least by adulthood).

gwern 8y ago

> only 11 percent. this tells us that the remainder is more likely to be attributable to environmental factors. No, it doesn't. It merely tells us that with the data up to now, we've reached 11% (up from, incidentally, ~0.3% just 5 years ago). It provides a very loose lower bound on genes. Saying that is like saying, 'our software testing up to now has only revealed bugs in 11 places, therefore, every other line of code must be bug-free!' Even if you knew nothing about the actual heritability estimates, one would think that finding so much so quickly would make one suspect that there's a lot more to find...

the8472 8y ago
Plus they focus on SNPs, not other types of gene variants.
cryoshon 8y ago
there's always more to find, but consider that this model is leaving out the environment entirely and its predictive value is even less than you might think at first glance. we know the environment impacts gene transcription, and we know that genetic determinism (as far as outcome phenotypes) is not a thing. likewise, there may be epigenetic factors at play which wouldn't be captured in a GWAS yet still might (or might not) have an impact on outcomes. it isn't a "the other lines of code are bug free" situation so much as a "we expect there to be an abundance of bugs with undefined consequences outside of the subset which we have examined which most likely outweigh the stuff we found"

toasterlovin 8y ago

I wouldn't write off genetics. GWASes, as a technique, are still pretty new and the number of genomes available for GWASes is still relatively small. Plus, twin studies point to traits like educational attainment being >11% heritable.

cryoshon 8y ago
GWAS are great, don't get me wrong. but there's always a lot of jumping to conclusions that is done whenever a GWAS finds something interesting. afterwards, you find the people wishing that they could CRISPR themselves to fit the pattern found in the GWAS, etc. you can see that in the article the authors of the study specifically went out of their way to try to caution people against doing that.