[Sérieux] Ethno-différencialisme, race-realism, génétique et courbe en cloche

[Pegase]

Très complet, merci Lancelot.

La vidéo revient aussi sur l'opposition entre généticiens du comportement et généticiens moléculaires en arguant que toutes les affirmations des premiers sont toujours devenues de plus en plus difficiles à prouver dans le temps (le fameux missing heritability problem).

 

Vous avez des sources sur cette question ? On m'a conseillé Robert Sapolsky.

[Adrian]

The heritability of IQ test performance I:What does IQ measure?

 

Citation

6.0 | Summary

• An IQ test score is a weighted average of scores across tests of memory, vocabulary, numeric reasoning, and general knowledge. A consistent observation from IQ test data is that individuals who do poorly on one test tend to do poorly on multiple tests – producing a correlation across scores known as the “positive manifold”. Motivated by these correlations, IQ is often summarized with a “general factor” (or g) score, which is simply a different re-weighted average of IQ subtest scores.
• Much research has gone into developing theories of IQ test correlation and the positive manifold. Theories are critical to understanding what the test is measuring, how to design accurate/unbiased tests, how to use tests to understand interventions, and how to draw broader conclusions about the mind. Common theories that can explain the positive manifold include:

• Bonds/sampling theory: Cognition is formed by a very large number of processes, with each IQ subtest sampling some subset of processes. The positive manifold is merely a statistical artifact of the incidental overlap between the processes across subtests.
• Dynamic mutualism: Cognition is formed by the “mutualist” interaction between multiple underlying processes where development of one process leads to growth in other processes. These relationships induce a positive manifold over time even though no general factor exists as a causal entity.
• g/Factor theory: Cognition is formed by the general factor itself (akin to a “mental energy”), which drives performance across all subtests in addition to some independent test-specific abilities. g theory has been expanded to include multiple factors or factor hierarchies.
• Process Overlap Theory: A synthesis of sampling and factor theories where subtests sample from processes that are either domain-specific (with three named domains) or domain-general. The positive manifold is an emergent property of the domain general process overlap but is not itself a causal entity.

• Five paradoxical findings regarding the measurement of IQ and g:

• The general factor weights (or “loadings”) of subtests are nearly perfectly correlated with the culturally-specificity of each subtest (e.g. higher for vocabulary, lower for numeric memory) (Kan et al. 2013). Thus a g score is merely an IQ score re-weighted by acculturation.
• The positive manifold increases with lower IQ, meaning that low IQ individuals tend to do poorly on all tests whereas higher IQ individuals tend to do well on only some tests. Paradoxically, the positive manifold also increases as children develop and decreases as adults decline. These paradoxical observations are incompatible with a single g-factor process across all individuals and likely necessitate a dynamic model.
• Multiple longitudinal studies have found that IQ is not associated with an increased rate of cognitive or academic growth in young age through adolescence (Larsen and Little 2023), nor with faster acquisition of job skills in adults (Schmidt et al. 1988), nor with the rate of cognitive decline in the elderly (Gow et al. 2011). This yields a third paradox: if g is a measurement of “general processing speed”, why is it not associated with faster learning (nor with slower cognitive decline)?
• Socioeconomic status (SES) is significantly associated with IQ gains: on average, children in low SES families started at a 6 IQ point deficit at age 2, and had a 15-17 IQ point deficit by age 16 (von Stumm and Plomin 2015). The effect of SES on academic achievement remained significant even when adjusting for future IQ measurements, indicating that SES can drive an environmental feedback loop between IQ and schooling. This yields a fourth paradox: if g is causal and largely independent of the shared environment, as is often claimed, why does the shared environment (SES) correlate with IQ gains while IQ does not?
• The Flynn Effect: IQ has increased an average of ~2 points per decade, with the largest gains in teenagers (Wongupparaj et al. 2023). The gains were not homogenous: increasing scores were observed on the more g loaded subtests in lower IQ individuals and decreasing scores were observed on less g loaded subtests in higher IQ individuals (Colom et al. 2023). The heterogenous increase in IQ suggests substantial, culturally induced changes acting on different processes. This yields a fifth paradox for g, as the underlying construct itself appears to be changing rapidly over time or else implies that the average individual born in the 19th century was severely mentally handicapped.
• Test-retest reliability of a general factor in the UK Biobank was high (r = 0.82) for a 30 day gap (Fawns-Ritchie and Deary 2020). Test ranking generally stabilizes in adolescence (r>0.7) (Tucker-Drob and Briley 2014) with moderate correlations (r = 0.45) observed from age 11 to age 90 after removing outliers (Deary, Pattie, and Starr 2013).
• There is meta-analytic evidence that IQ test results can be substantially impacted by motivation and the stronger the motivation the larger the effect (though studies have generally been limited and small) (Duckworth et al. 2011).

• Multiple lines of evidence – longitudinal measurements, cross-sectional model fitting, neurological, and interventional – refute a simple general factor theory of intelligence:

• Dynamic / mutualist models have been supported by cross-sectional and longitudinal studies, where performance in one cognitive domain is associated with improvement in other domains (Kievit, Hofman, and Nation 2019). Network models also often provide a better fit to IQ subtest data in cross-sectional studies (Kan et al. 2020). A recent large-scale genetic analysis revealed higher-order negative relationships between subtests, and a substantially different relationship between familial correlations and g loadings from that of a conventional factor model (Knyspel and Plomin 2024).
• Neurological measurements exhibit highly complex structure and do not support a single “neuro g” factor: structural data does not form a unidimensional latent factor mapping to IQ (Kievit et al. 2012) nor do structural correlations from different test batteries strongly overlap (Haier et al. 2009); network models generally fit the relationship between connectivity data and IQ tests best (Anderson and Barbey 2023; Soreq et al. 2021); and studies of focal brain injuries generally show local rather than general effects (Protzko and Colom 2021).
• One of very few theory-motivated experimental studies found unexpectedly large gains from mutualistic models: participants who trained on a diverse set of tasks had more substantial gains on a different specific task than participants who trained only on that task (Stine-Morrow et al. 2024). This preliminary finding highlights how experimental studies could be used to probe cognitive theories and produce surprising results.

• In sum: IQ scores clearly do not behave as a single latent process and are best seen as a bundle of processes with multiple sources of within-/between- individual and temporal heterogeneity. G-factor scores do not mitigate these issues and may in fact exacerbate them, as they are simply a reweighting of subtests to favor those with higher cultural specificity. While IQ scores may be a convenient data summarization technique, they also quickly become obsolete due to the Flynn Effect, are critically confounded by SES, and do not capture the process we are often most interested in: cognitive growth and decline. There is no reason to think that IQ scores are fundamentally more real or biological than Educational Attainment, they are simply an index of a different mix of processes which remains just as poorly understood.

 

[Paperasse]

il y a 41 minutes, Rincevent a dit :

il y a coévolution entre gènes et culture.

 

Un truc que je souhaitais évoquer sur le forum, sur cette question, c'est qu'on peut lire qu'Engels est le premier à avoir introduit la théorie de "coévolution gènes cultures" (dans son essai non terminé The Part Played by Labour in the Transition from Ape to Man, 1876).

 

Kangal, K. (2020). Engels’s emergentist dialectics :

Citation

Gould was particularly impressed by Engels’s view that the human “hand is not only the organ of labor, it is also the product of labor.… As humans learned to master their material surroundings, Engels argues, other skills were added to primitive hunting — agriculture, spinning, pottery, navigation, arts and sciences, law and politics." Elsewhere, Gould asserted that all human evolution stands and falls with gene-culture coevolution and “the best nineteenth-century case for gene-culture coevolution was made by Friedrich Engels in his remarkable essay of 1876."

 

Si c'est vrai, ça me semble un peu en contradiction avec cette idée d'associer le socialisme à la "tabula rasa"...

 

J'avoue que j'ai pas encore ultra creusé cette question. De ce que j'ai l'intuition, c'est que c'est un peu une sur-interprétation de considérer cet ouvrage comme l'origine de la "coévolution gènes cultures"

 

(je place cette remarque ici et pas dans le fil "Mes lectures du moment" pour pas partir en hors sujet et casser le débat sur le IQ)

[Rincevent]

il y a une heure, Paperasse a dit :

Gould

J'ai arrêté de lire là. 

[Adrian]

Twin Studies Exaggerate IQ Heritability

 

Citation

Heritability estimates are inflated by gene-environment correlation among other factors.


While twin/adoption studies had found that the heritability of IQ was around 80%, the largest GWAS analysis of intelligence, which examined the effect of more than nine million gene variants, found that only 2 to 5% of variation in IQ could be predicted from genes.1 Another high-quality GWAS found a heritability for cognitive ability of 14%. Whereas twin/adoption studies found that the heritability of IQ and height was quite similar, geneticist Sasha Gusev has shown, in a highly illuminating discussion of GWAS, that IQ is much more environmentally sensitive than height.
 

 

[Rincevent]

il y a 2 minutes, Adrian a dit :

Twin Studies Exaggerate IQ Heritability

Ok, j'allume le bat-@Lancelot-signal.

[Lancelot]

Lancelot a le regret de vous informer qu'il n'a pas le loisir ni particulièrement l'envie de lire tous les liens au pif vers des articles de blog balancés sans contexte comme une carte pokémon. Il vous invite cependant à lire son message quelques posts plus haut seulement où il s'intéresse précisément à la littérature critiquant les twin studies, et à revenir vers lui quand vous aurez des questions ou arguments précis.

[Bézoukhov]

T’as de la chance ça pourrait être une vidéo.

 

En substance le gus est un spécialiste des GWAS et dit qu’il y a sûrement quelque chose qui cloche dans les twin studies parce que les GWAS ne montrent pas d’aussi gros coefficients.

[Sloonz]

Discuté au 51e paragraphe de https://www.astralcodexten.com/p/links-for-november-2024

[Lancelot]

2 hours ago, Bézoukhov said:

T’as de la chance ça pourrait être une vidéo.

 

En substance le gus est un spécialiste des GWAS et dit qu’il y a sûrement quelque chose qui cloche dans les twin studies parce que les GWAS ne montrent pas d’aussi gros coefficients.

Et il a tout à fait le droit de penser ce qu'il veut.

[Bézoukhov]

Ce que j’en pense c’est que j’y connais rien en GWAS mais qu’avec la bonne base de données et ce genre de méthodes statistiques, je peux vous montrer presque ce que je veux.

 

(sauf qu’il y a une plus de deux sexes, on peut pas non plus faire n’importe quoi)

[Adrian]

 

 

 

 

 

 

 

Citation

Relative to most other animals, including primates, humans are unusually genetically homogenous (1, 71–73). Fewer than 1 in 1,000 nucleotide bases differ between any randomly chosen pair of humans (72). What genetic variation does exist is not structured into races. Lewontin (74) first reported that the proportion of genetic variation among socially defined human races is low, and suggested that racial taxonomy for our species is biologically meaningless, a conclusion supported by many subsequent studies (e.g., refs. 75–78). Approximately 93% of human genetic variation is found within populations, and only 4% among continents, with the remaining 2 to 3% among populations within continents (76). Most alleles are widely distributed around the world, with few (c. 7%) private to individual continents, and those typically at low frequency; thus there are no distinctive alleles present in all members of one region but absent from individuals outside the region (79). Even for complex traits affected by multiple loci, population genetic theory shows that “groups are not unduly likely to differ on traits that are determined by many loci, even when the loci influencing the trait would provide a sufficient basis for accurate classification” (80).

 

Within evolutionary biology the term “race” has a precise scientific meaning, broadly equivalent to “subspecies,” but human populations are not distinct enough to qualify as races. Subdivision is commonly quantified using a statistic called F_ST, which relates the genetic variation within and between populations. F_ST ranges from 0 (no population subdivision) to 1 (complete subdivision), with an F_ST exceeding 0.25 potentially indicating the existence of genetic subgroups (1). An authoritative estimate from the 1000 Genomes Project (81) of the global human F_ST was 0.052 to 0.083, revealing that the human population exhibits relatively little genetic subdivision; sharp genetic boundaries and distinct evolutionary lineages of “races” do not exist (1, 82). Rather, human biological diversity is clinal: It changes continuously with geographical separation, as groups merge into each other (83, 84). There may be discontinuities caused by geographic barriers (85), but these barriers are not insurmountable (10). As modern humans migrated from Africa, continuous contact and gene flow meant that they did not form subspecies (1, 86), but rather a “nested pattern of genetic structure inconsistent with the existence of independently evolving biological races” (87).

 

Humans who migrated to higher latitudes experienced a reduction in sunlight, leading to selection that favored lighter skin at several loci (10, 39, 88). However, the distribution and evolution of genes for skin color are not representative of the rest of the human genome, most of which offers little evidence that selection has played a dominant role, with random genetic drift and serial founder effects thought to be more important (79, 89). The genes underlying skin color do not provide a reliable index of genetic differentiation between groups (90, 91).

 

Edwards (92) claimed that Lewontin (74) had committed a “fallacy” by failing to consider correlations between genes, which he argued would allow individuals to be assigned to racial groups. Subsequent studies have established that large sets of loci can contain a great deal of information about population membership and permit ancestry inferences (76). However, it does not follow that Lewontin’s conclusions about race were fallacious (1, 93). This can be seen by considering the hypothetical case that all genetic variation is geographically continuous and nonclustered; geographical origin is informative but there are no biological races. This scenario is actually a reasonable approximation to reality. Most humans from the same geographic region are fractionally more similar to one another genetically than to individuals from a distant region. However, because of our history of migration and interbreeding, human genetic variation tends to be spatially distributed in a continuous fashion. If a fallacy existed, it was Edwards’ confounding geographical origin with race (17).

 

Another way to address these issues is to examine what genetic substructure does exist and ask whether it maps onto human races. While the mean genetic difference for two individuals from the same population is almost as large as that for two individuals chosen from any two populations (79), if enough loci are considered, humans can be organized into groups by genetic similarity. Methods exist for partitioning species into clusters of genetically similar individuals, and model-based clustering is a popular technique for quantifying the genetic ancestry of humans and other organisms (Box 1). The methods are generally applied to neutral genetic variation, since most DNA variants that differ in frequency between groups are neutral, functionally insignificant, and probably of little relevance to phenotypic differences between populations (94). For reliable assignment of individuals to continents, over a hundred genetic variants are required, and more for fine-grained population subdivision (91). This reasoning cannot be reversed to predict accurately the genotype of an individual from knowledge of their region of origin; knowing an individual’s continental ancestry only slightly improves the ability to predict their genotype (91).

 

Statistically estimated clusters using human DNA variants are inconsistent with socially defined races (Box 1). Even those willing, with minimal evidence, to assume that ancestry analyses support the existence of some human races, would be forced to conclude that “races” are not those implied by traits such as skin color or eye shape. Ancestry estimated from DNA variants, while itself an ambiguous concept (95), nonetheless provides a far more subtle and complex description of an individual’s genetic makeup than “race” (72, 90). No human populations are “pure” in a genetic sense; individuals do not fall neatly into one of the categories usually defined as “races,” and perceived boundaries between so-called “races” are arbitrary.

 

Et Concepts: race and genetic ancestry que j'avais déjà posté ailleurs.

[Rincevent]

il y a 12 minutes, Adrian a dit :

 

Etape 1 : cette idée est complètement à côté de la plaque, pas la peine de s'y intéresser.

Etape 2 : cette idée est dangereuse, il ne faut pas y toucher.

Etape 3 : cette idée est perverse, il faut largement l'amender.

Etape 4 : cette idée est mainstream, de toute façon on l'a toujours défendue.

 

Intéressant de voir que les questions de biodiversité humaine percolent peu à peu dans l'étape 4...

[Mathieu_D]

9 hours ago, Rincevent said:

Etape 1 : cette idée est complètement à côté de la plaque, pas la peine de s'y intéresser.

Etape 2 : cette idée est dangereuse, il ne faut pas y toucher.

Etape 3 : cette idée est perverse, il faut largement l'amender.

Etape 4 : cette idée est mainstream, de toute façon on l'a toujours défendue.

 

Intéressant de voir que les questions de biodiversité humaine percolent peu à peu dans l'étape 4...

Heu on serait plutôt à l'étape 5, l'idée est générée par l'IA de legorafi et personne ne sait si c'est du lard ou du cochon et à ce point peu importe.

[Bézoukhov]

J'ai rien compris à la phrase. L'impression d'être dans un sketche des inconnus.

[Adrian]

 

https://threadreaderapp.com/thread/1916562799393906889.html

Puisqu'on voit souvent passer des ACP censées démontrer la pertinence des races... Il développe plus ses exemples ici

[Daumantas]

Pas certain que ce soit le bon sujet mais c'est moi ou depuis quelques mois il y a de plus en plus de nouvelles entreprises/jeunes pousses dans le secteur de la sélection d'embryon (Nucleus, HeraSight, Heliospect Genomics) voire dans la modification desdits embryons sans doute au stade du zygote (je pense à Manhattan Project). Est-ce que c'est juste moi qui me fais des idées à cause de la couverture médiatique plus importante ou il y a un vrai démarrage de l'industrie ?

[Adrian]

C'est souvent discuté sur X entre scientifiques, quidam et blogueurs, c'est peut-être pour ça que tu as l'impression qu'il y a des nouveautés. La sélection d'embryon pour le QI, par exemple (au hasard hein), plaira sûrement aux gens de la Silicon Valley. La dernière pousse (il me semble) HeraSight, a été créée récemment et un généticien prof en Californie qui a travaillé avec eux parle ici du whitepaper et explique que la méthode des autres (Nucleus et autres) n'est pas ouf.

 

Ces méthodes avaient déjà été critiquées par d'autres généticiens pour qui Genomic prediction of IQ is modern snake oil ou Science fictions are outpacing science facts for polygenic embryo selection ou How embryo selection exploits our flawed intuitions about risk ; et que prof en Californie avait répondu (sur le dernier lien) ici

 

 

 

 

[Adrian]

 

https://threadreaderapp.com/thread/1977927672371069023.html

[Daumantas]

Je ne sais pas trop d'où il sort son 2 % car les derniers modèles expliquent 15 % de la variance totale du QI (donc si on coupe la poire en deux et qu'on estime que le QI c'est moitié génétique, moitié environnemental on parvient déjà à expliquer environ 30 % de la variance détectable).

[Rincevent]

il y a 59 minutes, Daumantas a dit :

Je ne sais pas trop d'où il sort son 2 % car les derniers modèles expliquent 15 % de la variance totale du QI (donc si on coupe la poire en deux et qu'on estime que le QI c'est moitié génétique, moitié environnemental on parvient déjà à expliquer environ 30 % de la variance détectable).

Oui. Qui plus est, il y a une différence énorme entre "on n'explique pas encore cet énorme bousin" et "il n'y a rien à expliquer, salaud de malpensant".

[Lancelot]

Le contexte du pari, j'imagine, est que Murray croyait à la hype des GWAS.

[Daumantas]

Il y a 17 heures, Rincevent a dit :

Oui. Qui plus est, il y a une différence énorme entre "on n'explique pas encore cet énorme bousin" et "il n'y a rien à expliquer, salaud de malpensant".

Pour rajouter sur le sujet, si on en croit Gene Smith (qui semble respecté dans ce milieu), on peut déjà obtenir environ 2SD dans l'hypothèse où l'on parviendrait à éditer quelques dizaines de paires de base (cf. How to make superbabies), alors tenir la position qu'en fait tout ça c'est du flan et qu'il vaudrait mieux ne pas s'intéresser à ces sujets qui fâchent...

[Daumantas]

.

[Daumantas]

Est-ce que Gusev et la clique habituelle vont s'en remettre ?

tl; dr : nouveau prédicteur qui obtient un r d'environ 0,5 que ce soit en dehors ou au sein d'une même famille.

[Daumantas]

Herasight a publié un nouveau papier où ils expliquent pouvoir à présent utiliser leurs prédicteurs avec un simple PGT-A sans perte d'efficacité là où auparavant il fallait un PGT-P (interdit dans pas mal de pays dont l'Europe et assez coûteux). Ça permet dans les faits de démocratiser à l'ensemble de la population du monde développé ce genre d'outils ce qui est excellent chose.

[Adrian]

Du nouveau sur la «missing heritability question »

 

Même si, de ce que j'ai compris, ces méthodes n'ont pas été appliquées sur le QI ou "l'education attainment" ou que les données sont de faible qualité (sur le QI par exemple).

[Rincevent]

Il y a 6 heures, Adrian a dit :

Du nouveau sur la «missing heritability question »

 

Même si, de ce que j'ai compris, ces méthodes n'ont pas été appliquées sur le QI ou "l'education attainment" ou que les données sont de faible qualité (sur le QI par exemple).

Ce n'est pas @Lancelot qui disait que les GWAS c'est de la merde ?

[Lancelot]

11 hours ago, Rincevent said:

Ce n'est pas @Lancelot qui disait que les GWAS c'est de la merde ?

Je ne serais pas aussi péremptoire 

Quand tu lis Human Diversity de Murray, tu en ressors avec l'impression que cette technique est LE truc qui va tout révolutionner ("we will understand IQ genetically. I think most of the picture will have been filled by 2025" pour citer le tweet qu'Adrian a posté le mois dernier et qui résume l'ambiance). Or force est de constater qu'il y a des complications. Ça ne veut pas dire qu'on doit abandonner la technique (ou l'audace de faire des prédictions).