I've written the first part of a chapter on the heritability of IQ scores. Focusing on what IQ is attempting to measure. I highlight multiple paradoxical findings demonstrating IQ is not just "one innate thing".
First, a few reasons to write this. 1) The online IQ discourse is completely deranged. 2) IQists regularly invoke molecular heritability as evidence for classic behavioral genetics findings while ignoring the glaring differences (ex: from books by Ritchie and Haier/Colom/Hunt).
Thus, molecular geneticists have been unwittingly drafted into reifying IQ even though we know that every trait is heritable and behavior is highly environmentally confounded. 3) IQ GWAS have focused on crude factor models that perpetuate the "one intelligence" misconception.
So what is an IQ test and the "g" factor? In short, test takes are asked questions related to pattern matching, memory, verbal/numeric reasoning, and general knowledge (some examples below). The weighted average of their scores is then the IQ score.
It turns out people who do poorly on one test tend to do poorly on other tests, which produces weak correlations known as the "positive manifold". These correlations can be summarized with factor analysis, producing a "general" factor (g) that explains 25-45% of total variance.
There's nothing special about the g score: it's just the IQ score computed with a different weighting (the "g loadings", more on these later).
But what could explain these correlations? In fact, many different theories can produce the same exact positive manifold and g patterns:
For example, Thomson's "sampling" theory, where IQ subtests sample from a large number of partially overlapping processes, can produce exactly the pattern of test correlations, leading factor, and factor loadings observed in the UK Biobank. Even though no actual g exists!
Or mutualism theory, where underlying processes interact dynamically over time, together with environmental inputs, to produce an apparent positive manifold. Again, g is not a causal variable, but an emergent statistical byproduct of these mutualistic relationships.
Finally, g/factor theory: where we treat the latent factors as measures of the true causal process itself. One factor doesn't fit the data well, so all sorts of more complex factor models have been proposed. Including a synthesis of sampling + factors in Process Overlap Theory.
I'm stressing the fact that many different theories fit the data in part because a causal/biological "g" is often taken as a given. But also because getting the theory right is critical to understanding what is actually being measured, test bias, and effective tests.
Ok, I promised paradoxical findings so here are five:
1) What are the weights used to compute g? They are highly correlated with how culturally specific a given subtest is (e.g. high for vocab, low for digit memory). So g is just IQ rescaled to emphasize cultural knowledge.
2) Ability + Age differentiation: The *highest* test correlations are among individuals with the *lowest* IQ, yet test correlations also increase with age. So IQ is measuring something different at the low/high ends and is also dynamic through development.
3) There is no "Matthew effect": In study after study, individuals with higher starting IQ do not acquire knowledge/skills faster -- in fact they converge! IQ also can't predict cognitive decline. This means IQ is not a measure of "processing speed" but of baseline knowledge.
4) Socioeconomic status (SES), in contrast, *is* associated with divergence: No SES/IQ differences are observed in kids at 10 months, yet low SES kids were 6 pts behind at age 2, and 15-17 pts (>1SD) behind by age 16. Thus IQ is confounded by SES from the start and throughout.
5) The Flynn Effect: Mean IQ in the population has been increasing over generations, with some studies showing the increase happening on more g/culture loaded subtests and among lower scorers and even within families. Environmental/cultural factors thus reshape IQ over time.
Clearly multiple dynamic phenomena are at play, so how does this fit with theory? In fact, longitudinal data strongly support mutualism, with gains in one cognitive domain translating into gains in other domains. Recently, even shown in RCTs [Stine-Morrow et al. 2024].
The same is observed in cross-sectional analyses, where mutualist/network models consistently fit IQ data better than factor models. [Knyspel + Plomin 2024] even applied networks to twin data and showed reversed relationships between twin "heritability" and g loading!
Finally, while neuroscience suffers from the same construct validity issues as IQ research, the one consistent finding is that there is no "neuro g": g correlates with many different structural/functional patterns, is better explained by network models, and does not replicate.
In short, IQ is indexing a bundle of different and often confounded processes, including individual and cultural shifts. IQ/g scores should be modeled as a dynamic network with environmental interactions, and there's absolutely no reason to treat them as "real" or "innate".
Now that we have a handle on what we are/aren't estimating with IQ, next time I will discuss the heritability and molecular genetic findings. My previous thread on the heritability of Educational Attainment is linked below. /x
Oof. Polygenic scores for IQ lose 75% of their explained variance when adding family controls, even worse than the attenuation for Educational Attainment. These are the scores Silicon Valley is using to select embryos 😬.
The TEDS cohort used here is a very large study with high-quality cognitive assessments collected over multiple time points. It is probably the most impressive twin study of IQ to date. That means very little room for data quality / measurement error issues.
It is important to highlight surprising null results. Just last week we were hypothesizing that large IQ score attenuation could be a study bias or an artifact of the Wilson Effect. Now we see it replicate in an independent study with adults.
Racism twitter has taken to arguing that observed racial differences must be "in part" explained by genetic differences, though they demure on how much. Not only is this claim aggressively misleading, it is completely unsupported by data. A 🧵:
Genetic differences between any two populations can go in *either* direction, matching the phenotypic differences we observe or going against them. Genes also interact with the environment, which makes the whole notion of "explaining" differences intractable.
The mere fact that a trait is heritable within populations tells us nothing about the explanatory factors between populations. See: Lewontin's thought experiment; Freddie de Boer's analogy to a "jumping contest"; or actual derivations (). pubmed.ncbi.nlm.nih.gov/38470926/
James Lee and @DamienMorris have an interesting perspective paper out describing "some far-reaching conclusions" about the genetics of intelligence. This type of "where are we now" paper is very fun and more people should write them! So, where are we now? 🧵
It's a short paper and it surveys three core findings from the past decade of intelligence genetics. These sections follow a structure that I would cheekily call ... "make a bold claim in the title, then walk it back in the text".
First up, they address the concern that associations with intelligence may actually be mediated by functionally irrelevant traits like physical appearance or pigment. The argument is that IQ GWAS has demonstrated enrichments for CNS/brain structure gene sets. This is true!
The SAT/meritocracy debate has always been a bit odd to me when the test makers themselves have studies showing self-reported high-school GPA is a consistently better predictor of college GPA and always adds on top of SATs.
Clearly SATs are neither the only nor even the best measure we have of college success and "holistic" admissions can be "meritocratic". It's up for debate whether the additional <10% predictive variance SATs give you are worth the high-school testing industrial complex.
A challenge with all of these analyses is they are measured after selection on the predictor variables themselves, which can induce biased estimates through range restriction. The raw correlations are even lower, and it is hard to know whether correcting is appropriate.
Hanania advocated passionately against "race mixing" for years, so he knows what he's talking about here. But it's worth adding that race-IQ obsessives also tend to make very poor predictions about the future. Let's review ...
The Bell Curve, published at the peak of the 80-90's crime wave, predicted a coming dystopian urban hellscape with a "cognitive underclass" living in state-managed facilities. Not only did all this fail to materialize, but crime rates collapsed.
Charles Murray has nevertheless spent the following 30 years predicting vindication for his claims was just around the corner ... each time pointing to a new corner.
Nice! Here we have an interesting paper using genetic ancestry to classify race/ethnicity in modern data and algorithms. Let's take a look at what this paper found: 🧵
First, I don't want to get too hung up on language, but TCB's tweet starts talking about "ethnicity", then shifts to "continental ancestries", and then entirely omits the largest ethnic group in the US: Hispanics. These terms have distinct definitions (). nap.nationalacademies.org/catalog/26902/…
Anyway, how well can this paper actually impute ethnicity from genetic ancestry in a large cancer population ()? ~17% of the time it gets Hispanic classification completely wrong or a no-call! worldscientific.com/doi/10.1142/97…