I have a question: Being absolutely a layman here, has there been similar research into the correlation between IQ (or similar) and family structure? As is, does single-parenting have an effect? Does multigenerational cooperation in upbringing have an effect? Etc.
Single partner and gay parents households would indeed be part of shared environment, but kids have never been placed in such families, so their effect cannot be inferred from research results.
This type of analysis was already done(specifically for years of education). There was a different impact, with PGS scores going up as a function of genetic distance, and EA going down(also found for height btw). Especially with the finding of ~0 relative predictive validity for africans in the UK(unless the Hou et al. 2023 approach is used for EA and shows significant validity), i don't know if these results are too informative(also do you think africans have 10x rates of schizophrenia compared to the difference between europeans with schizophrenia and controls,(pfiffer 2023 found the same results) or like a 5sd difference in HRS(https://journals.lww.com/psychgenetics/FullText/2018/10000/Polygenic_risk_score_for_schizophrenia_is_more.2.aspx & https://wyclif.substack.com/p/schizophrenia-a-coda). Still think we need better polygenic scores without this bias, or a local admixture/sibling admixture study to be more informative about causes. I took a brief look, and apparently it's like 3x higher in the UK, and roughly the same in the US(psychotic disorder: ethnicity facts and figures, Fearon et al 2006, Schwartz and Blackenship 2014). But it seems unless there's a way of adjusting for this non-linear LD bias(the method mentioned in pffifer 2023 seems promising but wasn't persued further, and shows significant european bias for EA in almost every comparison(from 10d to 2d for one EUR-AFR comparison) I don't think this grouped polygenic score method will prove much.
Interesting post but I do have a few(more critical questions).
1. I know you've said that these PGS still have validity in unadmixed populations, but apparently in the UKBB there's almost a 100% falloff in predictive validity for EA in the unmixed african and caribbean populations (https://twitter.com/Legitimate_Yam/status/1715819108120318429). Isn't this a problem despite the local admixture results(that didn't include EA), showing significant validity in african populations).
2. I think your own results for african and caribbean immirgrants might support this, given that these groups seem to be a significant outlier in being closer to african Americans in their PGS(despite signficantly higher measured IQ scores, consistent with other data(Fuerst 2014, seems to be closer to 0.7d with more recent data)). This difference is likely even higher accounting for language bias(in your regression results with the ABCD sample you found a positive affect of being US-born and of being from an immigrant family across samples). Given that there is ample evidence for immigrant selection for these populations, even for predomiantely refugees, and acknowledged by hereditarians as a (Model, 2008; Easterly and Nyarko, 2008, https://humanvarieties.org/2015/11/05/the-measured-proficiency-of-somali-americans/,https://humanvarieties.org/2015/10/28/using-surnames-to-assess-ethnic-aptitude/), shouldn't the finding of a similar PGS be evidence of near zero validity(along with point 1). If you want to say that selection bias has an effect on these results as you said in your reply, you would be disagreeing with yourself fairly recently(https://www.emilkirkegaard.com/p/in-defense-of-hereditarianism-a-reply), and also the Indian and East Asian results would need another explanation(given that the Indian and East Asian iq is usually given as ~80 and ~105(Lynn 2015, lasker 2019(conference poster using MCGFA), but immigrants to the countries score more like 110 because of selection and have comparable PGS in the ABCD sample).
3. I looked at the IQ estimates from Pfiffer et al. and immediately there's some problems with the IQs(that would have an effect on the correlations), vietnam is definitely not 101, just going on IQ results it's 91.4 based on malloy's results(and the same from the Young Lives survey which is representative of vietnam), according to Becker it's even lower(77.84). The AQ data is probably higher but there's a significant complication from the very high shares not in school in PISA data(PISA says it was almost 50% in 2012 and 2015, but this is due to a mistake with the census data, it's slightly over 1/3 according to Dang et al. 2023). They also produced corrected PISA scores using data from the Young Lives study for math and reading for all 15 year olds, which using the means and sd for White americans in the same subjects and years, gives a ~97 and ~93.5 AQ, averaging to 95(for comparison the black white difference would be 83.6 using the same composite and averaged, both 3rd+ generation). I also know of another dataset that has PIAAC scores, and vietnam was ~95 on that(urban only). You already talked about the Palestinian scores and said it was closer to 75 and becker estimated ~80 for the gaza strip, instead of 85 in Pfiffer 2023. The Uyghur results are also probably incorrect(if they are only for ethnic uyghur and not for xinjiang province), according to Lynn & Cheng(2014) the Uyghur IQ is 86.4 compared to a Han chinese IQ of 105.9(Lynn 2012).
I do also have some additional data for Colombia and Kenya, but I'm saving it for a future blog post(and haven't done a full analysis yet). Unfortunately the CFPS doesn't have data for Uyghurs specific, but does have other ethnic groups in china(https://www.sebjenseb.net/p/ethnic-differences-within-china). Also because they distribute the data into deciles(instead of a continuous distribution), when calculating the adjustment for those not in school using the YL data, they assume that a person in school and not in school within each decile will score the same(which obviously given the difference in performance and the nature of normal distributions, is not true). This means the results are lower for vietnam adjusted for those in school(i'm sure you could simulate it). Also no idea what the source for the Kalash IQ is(99.4 according to Pfiffer), not mentioned in the spreadsheet either. According to Lynn & Cheng(2014), the IQ for Dai is 93.9(no IQ given by Pfiffer), and if only ethnic Russians are included in the russian genetic sample it should be more like 100(admittedly nitpicking,https://rpubs.com/EmilOWK/ethnic_Russian_IQ_meta). Haven't looked to much into the other datapoints(but I do have some criticism of chinese iq/aq estimates). Also there's some height data for different chinese ethnicities in the CFPS.
Viets, Dais and even Uyghurs look more intelligent than northern europeans in pgs analysis. I think this may related with bias towards east asian-like admixture. As for as i know Uyghurs are half caucasoid half mongoloid. Finnish people have also siberian ancestry.
Kinh may be ~100 genotypically(given significant admixture from southern china because of sinicization, so probably with a north south gradient, as well as the fact that vietnam is only like 85% Kinh, and some data from the young lives survey and STEP indicating an increase with age), but Dai and Uyghurs are significantly lower(maybe related to not going through selection for iq according to the clark-unz thesis). Finnish siberian admixture isn't too much(roughly the sample as northern russians, maybe 10% at the high end), and given the mids 90s for siberian populations(yakuts are 97 based on Lynn & Shibaev 2016(other groups like inuits are more mixed and are ~91(Lynn 2015), and saami are only marginally below finns despite ~25% admixture(Dutton 2014), it probably doesn't change much. I wish Pfiffer did the correlations between polygenic scores(adjusted for LD bias) and IQs, because he only calculated them for a few comparisons(maybe saving for a future paper).
One of articles Gusev links says that height PGS extrapolated from European pop decreases with increase of genetic distances from Europeans, yet plots given by Kirkegaard and Piffer show correct avg pop height. What is correct?
"The failure of 100s of billions of dollars spent on remedial, compensatory, early education the last 70 years suggest the causes aren't environmental. In fact, the gaps have remaining basically stable for 100 years..."
I would like to disagree with you that the money spent has been a failure. I may end up agreeing with you only if you convince me. What is the data on the Flynn-effect for U.S. African-Americans? For this spending to be a failure two things must be true: (1) the Flynn-effect would have happened irrespective of any money spent on further the conditions of AA/Black people, (2) the net effect on GDP from reduced crime and increase productivity minus the billions spent has to have been negative.
I am saying that because IQ differences are somewhat environmental, it is worth every penny to try to cause even the smallest secular increases. As an American living in a major city, I wish we would throw 10x as much money at the problem. Completely closing the IQ gap is unrealistic via education, health care, and social support but we should still try.
In modern african samples(comparing age heaping to mathematics tests) and late 18th-early 19th age heaping internationally and within italy the correlation is like 0.7 if you use beta regression to adjust for ceiling effects.
But that rarely works. You are treating the symptom not the disease. Liberals are correct in identifying crime as a symptom of a social illness. What they refuse to admit is that the cause is low IQ. But IQs can be increase slightly through better environment. What I am calling for here is a cost benefit-analysis of increased IQ-reduced-social-problems. I suspect many smart liberals secretly know this but can't admit it, which is why they are so hell bent on social programs.
I'm not sure you answered your own question either. The claims that LD-decay would distort the PGIs in random directions sounds plausible, so, even in a fully-hereditarian model, shouldn't we expect to *actually* get more noise for populations farther away from the training set?
Because that's not what we observe: the noise doesn't seem to widen much. For example, East Asians and Hispanic have similar heights, but there's no reason to expect them to have same average height PGI, yet it's almost exactly the same. Is that a coincidence?
The environmental model doesn't predict such a perfect correlation, but neither does the hereditarian one. Am I getting something wrong?
If we plotted all individuals from everywhere on the same plot, I'd expect the cloud of points to follow some general "flow" whose topology would match the phylogeny of humanity.
I see no reason why the population means should be perfectly calibrated, even when the within-population slopes are randomly changing. For example, look at the grey dots on the top of the left plot. Imagine we were to divide that population in the middle into two subpopulations. Then, the population means would depart from the diagonal. Why doesn't this happen when we split, for instance, white people into catholics and atheists?
The genetic distances between large human races aren't very much variable. European - East Asian distance is similar to Europen - Hispanic. And maybe noise increase is as square root of distance, then you won't see much.
This may be more of a rhetorical question. You mentioned a paper that questioned the genetic origin of height. I'm sure that the authors had some kind of model to support their idea, but does the reality of animal breeding ever enter into the conscious awareness of these researchers? Do they believe socio-economic factors making one line of cows or pigs larger than another? You would think that if their ideas were so sound and thorough that the agricultural industry would be very interested in all the money they could save on prime bull seed and and tracking heredity of their animals.
Forgive, that's not a very mathematically sophisticated argument or analysis, but it something that frequently crosses my mind when confronted with these ideas.
Reading this article is giving me flashbacks to argue heredity with almost everyone. If there is a gap in genetics wide enough to permit the passage of a single flea, suddenly that is the source of all differences between people. Someone should do a study trying to quantify the aversion people have to admitting natural biological difference, both between individuals and groups. I am aware from personal experiences of the motivating factors that quite consciously emanate from academia and professional activists, but even in right wing or christian spheres this tendency is quite strong.
I have a question: Being absolutely a layman here, has there been similar research into the correlation between IQ (or similar) and family structure? As is, does single-parenting have an effect? Does multigenerational cooperation in upbringing have an effect? Etc.
These things go under shared environment (C) in twin models, and C is usually close to 0, so those won't have any noticeable effect.
OK. Thank you for answering.
Single partner and gay parents households would indeed be part of shared environment, but kids have never been placed in such families, so their effect cannot be inferred from research results.
Parents increasing children phenotypical EDU by making their homework and thesis? Almost certaintly exists.
https://pubmed.ncbi.nlm.nih.gov/37198491/#&gid=article-figures&pid=extended-data-fig-8-uid-12
This type of analysis was already done(specifically for years of education). There was a different impact, with PGS scores going up as a function of genetic distance, and EA going down(also found for height btw). Especially with the finding of ~0 relative predictive validity for africans in the UK(unless the Hou et al. 2023 approach is used for EA and shows significant validity), i don't know if these results are too informative(also do you think africans have 10x rates of schizophrenia compared to the difference between europeans with schizophrenia and controls,(pfiffer 2023 found the same results) or like a 5sd difference in HRS(https://journals.lww.com/psychgenetics/FullText/2018/10000/Polygenic_risk_score_for_schizophrenia_is_more.2.aspx & https://wyclif.substack.com/p/schizophrenia-a-coda). Still think we need better polygenic scores without this bias, or a local admixture/sibling admixture study to be more informative about causes. I took a brief look, and apparently it's like 3x higher in the UK, and roughly the same in the US(psychotic disorder: ethnicity facts and figures, Fearon et al 2006, Schwartz and Blackenship 2014). But it seems unless there's a way of adjusting for this non-linear LD bias(the method mentioned in pffifer 2023 seems promising but wasn't persued further, and shows significant european bias for EA in almost every comparison(from 10d to 2d for one EUR-AFR comparison) I don't think this grouped polygenic score method will prove much.
Interesting post but I do have a few(more critical questions).
1. I know you've said that these PGS still have validity in unadmixed populations, but apparently in the UKBB there's almost a 100% falloff in predictive validity for EA in the unmixed african and caribbean populations (https://twitter.com/Legitimate_Yam/status/1715819108120318429). Isn't this a problem despite the local admixture results(that didn't include EA), showing significant validity in african populations).
2. I think your own results for african and caribbean immirgrants might support this, given that these groups seem to be a significant outlier in being closer to african Americans in their PGS(despite signficantly higher measured IQ scores, consistent with other data(Fuerst 2014, seems to be closer to 0.7d with more recent data)). This difference is likely even higher accounting for language bias(in your regression results with the ABCD sample you found a positive affect of being US-born and of being from an immigrant family across samples). Given that there is ample evidence for immigrant selection for these populations, even for predomiantely refugees, and acknowledged by hereditarians as a (Model, 2008; Easterly and Nyarko, 2008, https://humanvarieties.org/2015/11/05/the-measured-proficiency-of-somali-americans/,https://humanvarieties.org/2015/10/28/using-surnames-to-assess-ethnic-aptitude/), shouldn't the finding of a similar PGS be evidence of near zero validity(along with point 1). If you want to say that selection bias has an effect on these results as you said in your reply, you would be disagreeing with yourself fairly recently(https://www.emilkirkegaard.com/p/in-defense-of-hereditarianism-a-reply), and also the Indian and East Asian results would need another explanation(given that the Indian and East Asian iq is usually given as ~80 and ~105(Lynn 2015, lasker 2019(conference poster using MCGFA), but immigrants to the countries score more like 110 because of selection and have comparable PGS in the ABCD sample).
3. I looked at the IQ estimates from Pfiffer et al. and immediately there's some problems with the IQs(that would have an effect on the correlations), vietnam is definitely not 101, just going on IQ results it's 91.4 based on malloy's results(and the same from the Young Lives survey which is representative of vietnam), according to Becker it's even lower(77.84). The AQ data is probably higher but there's a significant complication from the very high shares not in school in PISA data(PISA says it was almost 50% in 2012 and 2015, but this is due to a mistake with the census data, it's slightly over 1/3 according to Dang et al. 2023). They also produced corrected PISA scores using data from the Young Lives study for math and reading for all 15 year olds, which using the means and sd for White americans in the same subjects and years, gives a ~97 and ~93.5 AQ, averaging to 95(for comparison the black white difference would be 83.6 using the same composite and averaged, both 3rd+ generation). I also know of another dataset that has PIAAC scores, and vietnam was ~95 on that(urban only). You already talked about the Palestinian scores and said it was closer to 75 and becker estimated ~80 for the gaza strip, instead of 85 in Pfiffer 2023. The Uyghur results are also probably incorrect(if they are only for ethnic uyghur and not for xinjiang province), according to Lynn & Cheng(2014) the Uyghur IQ is 86.4 compared to a Han chinese IQ of 105.9(Lynn 2012).
I do also have some additional data for Colombia and Kenya, but I'm saving it for a future blog post(and haven't done a full analysis yet). Unfortunately the CFPS doesn't have data for Uyghurs specific, but does have other ethnic groups in china(https://www.sebjenseb.net/p/ethnic-differences-within-china). Also because they distribute the data into deciles(instead of a continuous distribution), when calculating the adjustment for those not in school using the YL data, they assume that a person in school and not in school within each decile will score the same(which obviously given the difference in performance and the nature of normal distributions, is not true). This means the results are lower for vietnam adjusted for those in school(i'm sure you could simulate it). Also no idea what the source for the Kalash IQ is(99.4 according to Pfiffer), not mentioned in the spreadsheet either. According to Lynn & Cheng(2014), the IQ for Dai is 93.9(no IQ given by Pfiffer), and if only ethnic Russians are included in the russian genetic sample it should be more like 100(admittedly nitpicking,https://rpubs.com/EmilOWK/ethnic_Russian_IQ_meta). Haven't looked to much into the other datapoints(but I do have some criticism of chinese iq/aq estimates). Also there's some height data for different chinese ethnicities in the CFPS.
Viets, Dais and even Uyghurs look more intelligent than northern europeans in pgs analysis. I think this may related with bias towards east asian-like admixture. As for as i know Uyghurs are half caucasoid half mongoloid. Finnish people have also siberian ancestry.
Kinh may be ~100 genotypically(given significant admixture from southern china because of sinicization, so probably with a north south gradient, as well as the fact that vietnam is only like 85% Kinh, and some data from the young lives survey and STEP indicating an increase with age), but Dai and Uyghurs are significantly lower(maybe related to not going through selection for iq according to the clark-unz thesis). Finnish siberian admixture isn't too much(roughly the sample as northern russians, maybe 10% at the high end), and given the mids 90s for siberian populations(yakuts are 97 based on Lynn & Shibaev 2016(other groups like inuits are more mixed and are ~91(Lynn 2015), and saami are only marginally below finns despite ~25% admixture(Dutton 2014), it probably doesn't change much. I wish Pfiffer did the correlations between polygenic scores(adjusted for LD bias) and IQs, because he only calculated them for a few comparisons(maybe saving for a future paper).
Wouldn't error of transferring PGS function to anotheer population diminish as number of tag SNPs grows?
One of articles Gusev links says that height PGS extrapolated from European pop decreases with increase of genetic distances from Europeans, yet plots given by Kirkegaard and Piffer show correct avg pop height. What is correct?
He said he addressed this: https://twitter.com/SashaGusevPosts/status/1715628184291426504
"The failure of 100s of billions of dollars spent on remedial, compensatory, early education the last 70 years suggest the causes aren't environmental. In fact, the gaps have remaining basically stable for 100 years..."
I would like to disagree with you that the money spent has been a failure. I may end up agreeing with you only if you convince me. What is the data on the Flynn-effect for U.S. African-Americans? For this spending to be a failure two things must be true: (1) the Flynn-effect would have happened irrespective of any money spent on further the conditions of AA/Black people, (2) the net effect on GDP from reduced crime and increase productivity minus the billions spent has to have been negative.
I am saying that because IQ differences are somewhat environmental, it is worth every penny to try to cause even the smallest secular increases. As an American living in a major city, I wish we would throw 10x as much money at the problem. Completely closing the IQ gap is unrealistic via education, health care, and social support but we should still try.
Flynn effect has been similar by race(https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2907168/), there's more datasets showing basically the same thing, also when you do a meta-analysis of these results there's literally no trend going back 150 years(with old literacy and numeracy measures, https://miro.medium.com/v2/resize:fit:720/format:webp/1*vGEaHYd9A3DpckXjZ1vbfw.png). See also https://www.cremieux.xyz/p/the-deep-roots-of-intelligence-differences to confirm this for tons of other differences(ashkenazi > sephardi > spanish christians > new christians and muslim iq differnences, north-south diffference in italy(actually going back even further to 500 years according to https://www.cambridge.org/core/journals/journal-of-economic-history/article/div-classtitlequantifying-quantitative-literacy-age-heaping-and-the-history-of-human-capitaldiv/57B3C54D7B2EF7D11CC70D60F1F4B3C6), even stuff such as the east asian advantage in numeracy(https://ideas.repec.org/p/cpr/ceprdp/9991.html, although according to the flynn effect east asian countries only recently went higher, suggesting measurement invariances in old results). MCGFA is also confirmed in old data and gives the same results data(https://menghu.substack.com/p/g-explains-black-white-but-not-sex ), but apparently not for east asians, see previous note on flynn effect(also language bias would presumably effect results in 1st gen).
In modern african samples(comparing age heaping to mathematics tests) and late 18th-early 19th age heaping internationally and within italy the correlation is like 0.7 if you use beta regression to adjust for ceiling effects.
If you want reduced crime, maybe concentrate on it, and not throw money at IQ?
But that rarely works. You are treating the symptom not the disease. Liberals are correct in identifying crime as a symptom of a social illness. What they refuse to admit is that the cause is low IQ. But IQs can be increase slightly through better environment. What I am calling for here is a cost benefit-analysis of increased IQ-reduced-social-problems. I suspect many smart liberals secretly know this but can't admit it, which is why they are so hell bent on social programs.
Often social illness is lack of proper law enforcement, which has nothing to do with low IQ
I'm not sure you answered your own question either. The claims that LD-decay would distort the PGIs in random directions sounds plausible, so, even in a fully-hereditarian model, shouldn't we expect to *actually* get more noise for populations farther away from the training set?
Because that's not what we observe: the noise doesn't seem to widen much. For example, East Asians and Hispanic have similar heights, but there's no reason to expect them to have same average height PGI, yet it's almost exactly the same. Is that a coincidence?
The environmental model doesn't predict such a perfect correlation, but neither does the hereditarian one. Am I getting something wrong?
https://emilkirkegaard.dk/en/2023/08/polygenic-score-validity-and-group-differences/
I get that, but it still doesn't explain why the population averages would be exactly aligned on one line. Here's a sketch to explain what I mean:
https://imgur.com/HdoHXYc
If we plotted all individuals from everywhere on the same plot, I'd expect the cloud of points to follow some general "flow" whose topology would match the phylogeny of humanity.
I see no reason why the population means should be perfectly calibrated, even when the within-population slopes are randomly changing. For example, look at the grey dots on the top of the left plot. Imagine we were to divide that population in the middle into two subpopulations. Then, the population means would depart from the diagonal. Why doesn't this happen when we split, for instance, white people into catholics and atheists?
The genetic distances between large human races aren't very much variable. European - East Asian distance is similar to Europen - Hispanic. And maybe noise increase is as square root of distance, then you won't see much.
This may be more of a rhetorical question. You mentioned a paper that questioned the genetic origin of height. I'm sure that the authors had some kind of model to support their idea, but does the reality of animal breeding ever enter into the conscious awareness of these researchers? Do they believe socio-economic factors making one line of cows or pigs larger than another? You would think that if their ideas were so sound and thorough that the agricultural industry would be very interested in all the money they could save on prime bull seed and and tracking heredity of their animals.
Forgive, that's not a very mathematically sophisticated argument or analysis, but it something that frequently crosses my mind when confronted with these ideas.
Reading this article is giving me flashbacks to argue heredity with almost everyone. If there is a gap in genetics wide enough to permit the passage of a single flea, suddenly that is the source of all differences between people. Someone should do a study trying to quantify the aversion people have to admitting natural biological difference, both between individuals and groups. I am aware from personal experiences of the motivating factors that quite consciously emanate from academia and professional activists, but even in right wing or christian spheres this tendency is quite strong.
How much of the phenotypic b-w iq gap is currently predicted by pgs_iq?
Accounted for, maybe like 20%. https://www.researchgate.net/publication/371695470_A_Genetic_Hypothesis_for_American_RaceEthnic_Differences_in_Mean_g_A_Reply_to_Warne_2021_with_Fifteen_New_Empirical_Tests_Using_the_ABCD_Dataset
Have not read this article, so irrelevant comment but
What are your views and thoughts about autistic people, ADHD, and furries?