Recently Davide Piffer and I had a disagreement on X over recent selection for IQ in Britain. The below chart someone posted on X shows British education polygenic scores expressed in standard deviation (SD) units over 1000 years :
..
He then elaborated further on his substack,
His point about polling is well taken which is precisely why I asked him in the above the X thread whether Harvard students would score just as high on a five minute version of the SAT. A pollster doesn’t have time to poll hundreds of millions of Americans so they sample a few thousand. Similarly, if we don’t have time to give all two hours of the SAT, one could give a five minute sample of all the SAT questions. So if anyone knows why Piffer didn’t find this example analogous, please enlighten me.
But Piffer also says:
So here Piffer is saying that if one of the causes of a phenotype has increased by X SD, then we might correctly assume the entire phenotype has increased by the same degree, even if the correlation between the part and whole is very imperfect among individuals. Okay! There’s been no increase in British brain size (a cause of intelligence) since Medieval times (excluding post-WWII nutritional gains related to the Flynn Effect), thus there’s been no increase in genetic IQ.
If Piffer wants to analogize to polls, that’s fine with me, after all brain size can be considered a sample of the full population of neurological traits (brain folds, myelination, nerve conduction speed, reaction time etc) that cause IQ, but what do pundits do when the polls conflict? They average them.
So let’s say the polygenic poll Piffer did implies a 0.6 SD increase in IQ, but a poll of neurological properties (with brain size being the sample) suggests a 0 SD increase in IQ. Averaging them together suggests a 0.3 SD increase in genetic IQ and generously assuming an adult IQ heritability of 0.8 (Jensen, 1998), and thus a genotype-phenotype correlation of 0.89 (the square root of heritability). Thus if we cloned Medieval Brits and raised them in today’s UK, they’d probably average 0.3 SD(0.89) = 0.27 SD (about 4 IQ points) below today’s white Brits (or they might even score higher than today’s white Brits if you believe Michael Woodley’s dysgenic theories).
Previous attempts to combine Education PGS and brain size have done a good job estimating Cro-Magnon IQ and contemporary group differences.
Pumpkin Person 
See what I mean? Implicitly, hereditarians believe it shows sufficient causation. Rule-following considerations, lack of psychophysical laws, the multiple realizablilty of psychological kinds, all refute these kinds of interpretations of GWAS PGSs.
Hereditarian claims about SNPs and IQ, variance explained, and PGS differences presuppose a psychogenetic law in the form of “G -> P.” So hereditarian interpretations commit a category error by treating corr kinds as law-governed causal kinds. The arrow they want is impossible. The correlations arise from population stratification, passive GxE correlations, reverse causation, measurement artifacts.
https://www.embopress.org/doi/full/10.15252/embr.201744140
Sufficient causation would mean they think everyone would have the same IQ in every environment. No HBDer believes that.
There’s still the presupposition that genetic differences are sufficient causes, or else the hereditarian wouldn’t even use these studies. “Variance explained is X percent due to G.” That’s presupposes “G -> P.”
Variance explained is just a figure of speech in statistics. It just means how much of the variance in Y would vanish if everyone were the same on X. It does not necessarily mean caused.
Yea, I think that’s a concept that has quite a few false assumptions (just like h2).
https://notpoliticallycorrect.me/2023/11/29/the-illusion-of-separation-a-philosophical-analysis-of-variance-explained/
Hahaha so hilarious how you went from being the most racist blogger on the internet to the least racist in just a few years.
True. It was only a year after I started that I changed my views though. Been blogging almost ten years, changed my views almost 9 years ago. Crazy how time flies.
lets say Hip ratio increased 0.5 SD between some two periods of time.
We could say that hips and height is correlated at 0.59
(((0.5 group difference over time * 0.59 individual difference)*height of pop 1 – height pop 2)*mean pop 1) = ?
but this mistakes what happens
We see that it increased by 0.5 for BOTH hips and height on the group level.
That means the coefficient of hips to height may or may not have changed within the time periods not between them.
Given we don’t compare individuals to multiple groups means and coefficients (we don’t compare individual Swedish man bob to Japanese population as a whole and then reduce that number back to the Swedish population), we are comparing Group A to Group B in what has changed we cannot use a individual coefficient to the group with a Group to Group change.
I hope I make sense?
Brain size is not the only factor in intelligence.
PGS is looking at education.
0.6 SD increase in educational attainment would make people 800 years ago less suited to reading and writing because there was no selection pleasures to do so. (The SAT pushed a huge pleasure on society recently).
So whatever genes makes it possible to be better at education took longer to develop a pressure, there was not available books or paper to use and ink was expensive.
The brain might not be bigger but it was used for other things besides book learning. Symbols syntax. Most people died in the plague in the country side where people were less hygienic oriented/educated.
In China they been under that pressure a long time. The imperial exams.
Romans made weapons but that’s a pressure too that might converge on education. look at Greek statues. After roman fell most people stopped education hundreds of years.
0.6 SD increase in educational attainment would make people 800 years ago less suited to reading and writing because there was no selection pleasures to do so.
It’s not a 0.6 SD increase in education. It’s a 0.6 SD increase in genes associated with education and IQ but they’ve only found a tiny percentage of the genes that are supposedly associated with either variable.
In China they been under that pressure a long time. The imperial exams.
Then we’d expect China to be smarter than the rest of Northeast Asia but we don’t see that.
“It’s not a 0.6 SD increase in education. It’s a 0.6 SD increase in genes associated with education and IQ but they’ve only found a tiny percentage of the genes that are supposedly associated with either variable.”
ok,
but then that means we cannot tell which genes are doing the heavy lifting.
we cannot know how much it increased intelligence.
because this is not a single person
if we made it 0.27 SD increase then wed need a confidence interval or something.
it could have increased by zero or it could have increased by 0.6 * 2 = 1.2 SD
the fuzziness is not about how much an individual would increase but the entire population .
you cannot reduce the number it must go up as well
0 to 0.6 to 1.2
those number would be the range
not just down to 3 points downward
but again the entire population means 0.6 has the most density of probability of how much it went up
normal regression does not apply
Regression still applies but the line of best fit would be predicting group IQ, not individual IQ, and predictions are much more accurate at the group level than at the individual level. The thing is you could make an even more accurate predictions using group brain size or even group skin color but as more genes are discovered, especially causal ones, Piffer’s numbers will only get better.
“Then we’d expect China to be smarter than the rest of Northeast Asia but we don’t see that.”
it was a local thing
they gathered all the smartest people in a region
then put them as the leaders
just like the SAT doesn’t matter at the state level but it still separated people in the usa
in Japan they did the same but with wars before the unification of the 1600s clan against clan
samurai swords are difficult to make
so the smartest people survived as the leaders
Japan still is separated by a clan structure hierarchy
so eugenics all around
Mongolians practiced primitive herder lifestyle yet they also enjoy high polygenic scores, even bigger brains and conquered the Chinese twice, so I don’t buy it.
you don’t “buy” assortative mating?
you don’t think the SAT separated people in the USA?
people don’t separate by SES status either?
“brain size can be considered a sample of the full population of neurological traits”
There’s a difference between a random sample of questions from all over the WAIS vs having everyone take the same subtest, even if that subtest was randomly chosen for the group, right? The first would show a normal Flynn Effect while the latter would have a trajectory that depends on the specific subtest.
And there’s a difference polling random individuals vs polling a randomly picked narrow subset of individuals(men, women, blacks, whites…)
Specifically, there could be different selection pressures on brain size and intelligence more broadly.
You’re not wrong but piffer used the example of height increasing to the same degree as torso length so by that logic brain size increases should equate to IQ increases
But piffer said it must be 0.6 SD increase of education genes because the group level mean is based on the sample size not one individual coefficient between sample A mean and sample B mean.
That means you cannot regress it by the wrong coefficient (individual) and get a correct answer.
lets say the mean of group A is 0.6 SD above Group B
The genes did change by that much.
The chart doesn’t have a regress line because variation changes together. Hair length, shoe size everything. Everyone in the group averaged together is 0.6 SD above what it was 800 years ago for education genes.
Why regress a sample mean of “intelligence” to 0.2 SD when that would work for the induvial but not the group?
In other words education genes increased for all the people in the sample averaged out, why regress at all?
Sample average is the mean and you are saying the mean is not based on that average? but on the individuals relationship to the mean?
But what I understand is that Given group A differs from Group B we should not do this?
We need apples to apples comparison.
So the mean of genes for education did change between the two groups and that increase did increase peoples education by what piffer said it did.
but all features go together
(all people average together to form a mean)
we cannot regress hips to height and be accurate
People 800 years ago would be 9 points less in genes for education because at the group level that averages out over all the people within each of the two samples and thus we can say 9 points less in “IQ” not 3 points less.
am I missing something?
to use regression you’d need to do the regress on each individual sample person in the group. then average the results.
you are averaging the group before the individual is being measured by a regression.
you must regress the individual genes of all the dead people in the sample first, that changes results of the average because some people are higher on the poly score not just lower in the samples of dead people.
If we discard the torso argument and the brain size argument then we’re left with Piffer’s other arguments I suppose.
His other argument is that is that if his sample of genes are a representative sample of all cognitive genes, they are a poll and should not be regressed to the mean because we don’t regress actual poll results to the mean. But we don’t regress actual poll results to the mean because (1) there’s no obvious mean to regress them to, and (b) poll results have such small error margins that the degree of regression would be trivial. But because of the huge missing heritability problem in Piffer’s field, his poll of the genome has a HUGE standard of error. Now one may argue that at the group level error cancels out which is true when it comes to scatter around the regression slope but false when it comes to the slope itself.
But ironically this makes Piffer’s results even stronger in that if you get an effect size of say X on a measure that unreliably measures the difference being selected, then the true effect size is much greater, not equal as Piffer implies. On the other hand, if the difference is not caused by selection but by genetic drift, then the true effect size will be much weaker than what Piffer reports. Since population differences are often a combination of both, the two phenomena largely cancel out so Piffer ends up being right but not by the logic he cites. However you could probably just use skin color genes and make equally accurate results.
Basically “if the signal is real, then the effect size is larger”, but “if it’s drift, then the observed effect is spurious and the true effect is near zero”? All of these scores have pretty large errors, no? And yea, the missing heritability problem is pretty devestating. Remember the Howe et al estimates? Remember the the admixture “European genes” estimates?
Basically “if the signal is real, then the effect size is larger”, but “if it’s drift, then the observed effect is spurious and the true effect is near zero”?
Well said.
All of these scores have pretty large errors, no?
Yes but there does seem to be an overall pattern, so in addition to sampling error and genetic drift, there was likely some selection in my opinion which they attribute to Gregory Clark’s theory that the poor went extinct every generation and were replaced by the rich.
By “real signal” do you mean a causal genetic signal? Have you read the responses to Clark? I’ll get them later.
By “real signal” do you mean a causal genetic signal?
I mean selection. Groups could also differ in some causal genes via genetic drift. Of course Piffer does not even claim to be using causal genes but claims to have found genes that are correlated with causal genes or something like that. He uses this to explain certain anomalies, like pygmies having relatively high PGS in one of his datasets. He says that’s because they diverged so long ago that his PGS don’t correlate with the causal genes among pygmies.
Have you read the responses to Clark? I’ll get them later.
The HBD-o-sphere worships Clark. Maybe a lot of them are ethnically British or British adjacent and want to portray their ancestors as a Master race that gave us the Industrial revolution and they did give us the Industrial Revolution, but they want it to be for genetic reasons.
Also what do you mean by “poll” here?
“if his sample of genes are a representative sample of all cognitive genes, they are a poll and should not be regressed to the mean because we don’t regress actual poll results to the mean”
Piffer’s argument is that even though they’ve only identified a tiny fraction of the many genes thought to affect IQ, it’s enough to tell how genetically smart a population is because just as a Gallup poll can accurately measure the opinions of 200 million U.S. adults with a sample size of only 500 people, he too just needs to look at a small sample of genetic variants, to approximate all the thousands of variants that affect IQ
What if the “real signals” are illusory and just confounds with no causal power?
“He uses this to explain certain anomalies, like pygmies having relatively high PGS in one of his datasets. He says that’s because they diverged so long ago that his PGS don’t correlate with the causal genes among pygmies.”
Sounds ad hoc.
I don’t think any genes have been found that affect any disease the GWAS hopefuls have been hoping—never mind for IQ.
Group A has a sample of individual people regressing the individual first
then averaged together to form X mean
Group B has a sample of individual people regressing the individual first
then averaged together to form Y mean
|X – Y| = 0.6 SD
that is what I think Piffer is saying
(A mean – B mean) * (regression value) is the wrong way to do it
Regression to the mean is just making the best guess with the info you have. Since most people have IQs resembling those related to them (today’s Brits) and resembling their PGS scores (-0.6 SD), your best guess is Medieval Brits had IQs somewhere between these data points.
However Piffer argues that since there was selection, genes that weren’t captured by the PGS would have increased too so we shouldn’t regress, but actually if Piffer really believes this then not only should he not be shrinking the difference, but he should be expanding it because a weak proxy like those polygenic scores will never show the full effect size.
If on the other hand, the differences are caused by genetic drift, they should indeed be regressed to the mean.
If the truth is a mix of both, Piffer will end up being right by accident.
Piffer doesn’t know that both genetic drift and selection pressures happen together?
I don’t think so, highly unlikely he in not aware of both.
maybe? as I see it regression is not needed exactly but a probability density of where the best number is / might be at which is why I think Piffer disagrees with you on effect sizes in his blog post dealing with square root(R^2)
each plot point already has an error bar
Regression down to 0.2 SD is redundant.
or it must regress up by the same amount to 1 SD
https://en.wikipedia.org/wiki/Error_bar
There are 12 plot points with twelve bars in the image.
Its why the density of probability should be between 1.2 SD and 0 SD as I said. (the middle is 0.6 SD)
unless I am wrong you are not considering the error bars are you?
I might be misunderstanding.
we still don’t know (Piffer is accidently correct anyway?)
Piffer doesn’t know that both genetic drift and selection pressures happen together?
I’m sure he knows but his argument is selection alone will result in a weak measure of geneotype having the same effect size as a strong measure would have shown. This is false. The noisier the measure, the weaker the signal, so whatever selection he’s measuring in his weak data is just a shadow of the much stronger selection he’d see if he had strong data
each plot point already has an error bar
But if he regressed he’d also have an error bar. It’s like on the WAIS when they say one’s score is 120, they’ll say you have an IQ of 117 to 123. however the manual also gives the option of regressing the IQ to the mean based on the test’s reliability, so 117 to 123 becomes 115 to 121
If Group A has a mean of 100 IQ and Group B has a mean of 120 IQ
Then because height is correlated at 0.3
Group A 0.3(100 – 100) = 0
Group B 0.3(120 – 100) = 6
(6 + 0) + 100 = 106
group B regressed to Group A at 106 in height?
that don’t make sense.
–
12 groups exist with 12 means
so we are regressing only to the group in the middle?
1250 A.D.
?
If 12 means exist you need to regress each mean to itself not the others.
You get something like this:
No you’d regress to (or progress away from) modern white Brits because they’re the only white Brits whose IQs are actually known.
You are confusing modern samples with samples at different time periods.
If each sample has an IQ level, you don’t regress that sample from modern times but from the time it was taken.
Its why the Flynn effect would not work if regression was used.
Take people 100 years ago and the IQ went up by 30 points.
So if we regress that it means (30 * 0.2) + 100 = 106
that means people 100 years ago had a mean of 94 not 70 – people today are 106 compared to 100 years ago?
see how that works?
by that logic the mean was “94” 100 years ago not 70 relative to today.
Scores went up 30 points not 6 points
same logic applies to poly scores
only individuals regress to the mean not group A 100 years ago regressing to Group B in modern times. Or even further back.
Something is wrong?
Piffer says 9 points over 800 years.
but if 100 years = 6 points decrease
100 – ((800 / 100)*6) = IQ 52
1250 AD people were not that stupid were they?
Is 30 points increase real or just 6 points for the Flynn effect?
You are confusing modern samples with samples at different time periods.
If each sample has an IQ level, you don’t regress that sample from modern times but from the time it was taken.
But modern IQ is the only one we know, Remember regression to the mean is simply a mathematical best guess based on what you know and all we know is their polygenic scores and the IQs of their descendants
Its why the Flynn effect would not work if regression was used.
It depends what you mean. The Flynn effect is not a genetic change.