How bad are Richard Lynn's 2002 national IQ estimates?
They correlate at r = 0.93 with our current best estimates.
It turns out that they're really not bad, and they don't provide evidence of systematic bias on his part🧵
In this data, Lynn overestimated national IQs relative to the current best estimates by an average of 0.97 points.
The biggest overestimation took place in Latin America, where IQs were overestimated by an average of 4.2 points. Sub-Saharan Africa was underestimated by 1.89 pts.
Bias?
If you look at the plot again, you'll see that I used Lynn's infamously geographically imputed estimates.
That's true! I wanted completeness. What do the non-imputed estimates look like? Similar, but Africa does worse. Lynn's imputation helped Sub-Saharan Africa!
If Lynn was biased, then his bias had minimal effect, and his much-disdained imputation resulted in underperforming Sub-Saharan Africa doing a bit better. Asia also got a boost from imputation.
The evidence that Lynn was systematically biased in favor of Europeans? Not here.
Fast forward to 2012 and Lynn had new estimates that are vastly more consistent with modern ones. In fact, they correlate at 0.96 with 2024's best estimates.
With geographic imputation, the 2012 data minimally underestimates Sub-Saharan Africa and once again, whatever bias there is, is larger with respect to Latin America, overestimating it.
But across all regions, there's just very little average misestimation.
Undo the imputation and, once again... we see that Lynn's preferred methods improved the standing of Sub-Saharan Africans.
There's really just nothing here. Aggregately, Lynn overestimated national IQs by 0.41 points without imputation and 0.51 with. Not much to worry about.
The plain fact is that whatever bias Lynn might have had didn't impact his results much. Rank orders and exact estimates are highly stable across sources and time.
It also might need to be noted: these numbers can theoretically change over time, even if they don't tend to, so this potential evidence for meager bias on Lynn's part in sample selection and against in methods might be due to changes over time in population IQs or data quality.
It might be worth looking into that more, but the possibility of bias is incredibly meager and limited either way, so putting in that effort couldn't reveal much of anything regardless of the direction of any possible revealed bias in the estimates (not to imply bias in estimates means personal biases were responsible, to be clear).
Some people messaged me to say they had issues with interpreting the charts because of problems distinguishing shaded-over colors.
If that sounds like you, don't worry, because here are versions with different layering:
• • •
Missing some Tweet in this thread? You can try to
force a refresh
0. The sample for the population result is *everyone* with ≥1/16 and ≤15/16 African admixture, the within-family result is for the siblings among them. I didn't plot between-families results, but they're pulled from the same distribution as the population result, so when I have those computed I'll replot, but they shouldn't be any different (will take a few hours, maybe be tomorrow, w/e).
1. Yes, datasets will be merged and sharper results will be obtained for the article on this that's coming out ~shortly. Can't be done for all phenotypes in the UKBB, but can be done for IQ at least. For example, we don't have lipoprotein(a) (lp(a)) measurements in some of the samples of young American kids, some of them don't have objective skin color measurements, etc. As a side note, this result holds up with brain size (not shown) in the UKBB, but I'm unsure if it holds up in the young samples that'll be merged with it, as they're still developing in most cases.
Anyway, the IQ p-value is p = 0.008 (two-tailed) within families and extremely low between them. The within-family result is robust, will get sharper with more data, and is also sharpened by using a latent variable instead of a score, and by correcting for measurement error (not needed with the LVM). Score shown for simplicity and ease of others with UKBB access replicating this. Also, accounting for error in the admixture computation, the p-value would drop a bit further, but not by very much since error is very small.
2. The "IQ per unit of admixture" is statistically indistinguishable between the population and within-family results, and yes, it explains most of the Black-White difference in IQ. I just wanted comparably-scaled results for all the traits here, so you're seeing r's. It's pleasant that the within-family variance reductions aren't enormous for siblings, which is what we expect even with quite high heritabilities given their genetic relatedness. It's the same result we've seen with American data, and it's also nice to see that in the case of this trait, the global admixture result *can* be interpreted like the within-family one. Presumably this only holds with measurement invariance, as we see in the U.K. when comparing Whites and Blacks there. Since we see this in the U.S. too, it's likely that the previous, already-published within-family null—which had a sizable effect in the correct direction which also could not be distinguished from the global r—was just a false-negative.
3. This result replicates with other degrees of relatedness, but we might lose the causal interpretation with those ones because the estimand for, say, a cousin test is different given the identity of the "C" variance component shifts for that comparison.
4. What is architectural sparsity and why is it relevant? Consider this table from nature.com/articles/s4159… (cc: @hsu_steve):
Basically, sparsity refers to the number of variants involved in a trait. It also refers to their effect size distribution. So, for Alzheimer's, for example, the trait is highly polygenic, but APOE explains more variance than the entire rest of the PGS, so while being under highly polygenic control, it remains moderately sparse.
If you're still not grokking what I mean here, consider some distributions of cumulative effects across the chromosomes. Here's the result for lp(a), which is already known to be influenced by essentially one gene. Guess which chromosome it's located on:
Consider, as a comparison, creatinine, which is considerably less sparse, and thus has effects distributed across all chromosomes:
Now, consider what this does to within-family admixture assessments. Skin color, for example, is controlled by only a handful of genes. This means that global admixture shouldn't tag it very much between siblings. The same is even more true for lp(a). And in fact, now we have confirmation of this!
But, we know from other methods that lp(a) is effectively entirely genetically-caused between populations. This is just accepted within the medical community because it is an obvious fact that follows from its strong control by a single gene. You can also figure this out using local ancestry estimates. Basically, the correlation between genome-wide ancestry and ancestry at a causal loci is what we want to get at, and if you know the causal loci, you gain power by restricting your analysis to that area. This is what we find with lp(a) (not shown, but use your brain. The obviousness of this fact is why the author used lp(a) as an example).
Also, in some sense, the frequency of that locus between ancestries gives you what you need without doing all this within-family stuff, if you're confident it's causal. The effect estimate might still be biased by population structure though, so that's worth keeping in mind.
There will be a post soon with more details and the expanded set of results with the additional datasets, robustness tests, and plenty of other fun things to look at.
TL;DR: this is a spoiler, and it shows that, yes, you can explain the Black-White IQ difference in Britain mostly genetically, and the global admixture result that I've posted here before is equivalent to the within-family one. Woohoo for things that should hold up, in fact, holding up!
As a sort of replication of the Young paper, you cannot explain the difference in educational attainment (as years of education) in this way. Why? Well, hard to say. Compensatory factors like I found with the GCSEs? (Haven't read that post yet? Go check it out here: cremieux.xyz/p/explaining-a…). Poor phenotype quality? Very plausible, because education really is a huge garbage heap, but why would that be in the general population and not within families? Maybe it has to do with what other traits admixture tags? Maybe it does replicate, but we just can't see it, because the precision is too poor (possibly the lp(a) story too). Who knows!
Q: Will the combination with more datasets allow us to fix this educational attainment result?
A: No, because most of the other datasets involve young people, not people who have almost all completed their educations, as in the UKBB. That plus the generational change and international incomparability in the definition of educational attainment makes it too poor as a phenotype. Sorry!
Any questions?
Link to a fun previous post from the same dataset, showing a result that *does* hold up within families: x.com/cremieuxrecuei…
Link to another post mentioning the forthcoming article earlier: x.com/cremieuxrecuei…
I want to ping an old post that I've also posted some replications for.
Basically, parents are inequality averse, and they try to compensate for when one sibling is less gifted than another, reducing the ancestry/PGS effect on education within families.
Ever wondered why advertisements heavily feature Black actors when they're just 12-14% of the population?
I might have an explanation:
Black viewers have a strong preference for seeing other Blacks in media, whereas Whites have no racial preferences.
These results are derived from a meta-analysis of 57 pre-2000 and 112 post-2000 effect sizes for Blacks alongside 76 and 87 such effect sizes for Whites.
If you look at them, you'll notice that Whites' initial, slight preference declined and maybe reversed.
It's worth asking if this is explained by publication bias.
It's not!
Neither aggregately (as pictured), nor with results separated by race.
You're on trial, and the jury can't make up their minds. The decision is a coin flip: 50/50, you either get it or you don't.
Your odds of a given verdict depend on the "peers" making up your jury.
If you're Black and they're Black, your odds are good; if you're White, pray.
Though White jurors have, on average, no racial bias, the same can't be said for Black jurors.
Where the White jury gives you approximately the coin flip you deserve, the Black jury's odds for a verdict are like a coin rigged to come up heads 62% of the time.
Once you get to sentencing, things get even worse.
The White jury is still giving you a coin flip on a lighter or a harsher sentence, but the Black jury is giving lenient sentences to Blacks about 70% of the time.
It shows that the gender wage gap is mostly about married men and their exceptional earnings.
In this thread, I'm going to explain why married men earn so much more than everyone else🧵
The question is:
Does marriage maketh man?
Or
Are all the good men married?
That is, does marriage lead men to earn more, or do men who earn more get married more often?
To answer this question, we have to work through the predictions of different theories.
For example, one of my favorite papers on the subject looked into three different hypotheses to explain the "marriage premium" to wages, and they laid out a few testable predictions:
I've seen a lot of people wondering why America has such a high incarceration rate.
If you weren't even aware that it does, consider this graph from Prison Policy:
To understand why America is like this, consider that, when Stalin died, his secret police chief Lavrentiy Beria released more than a million non-political prisoners and the result was a massive crime wave.
This is not the only instance of this happening in history. Plenty of places have done large-scale prisoner releases, and they nearly universally have the same effects wherever they happen: crime goes up.
One of my favorite examples comes from Italy.
On July 31st, 2006, the Italian Parliament passed the Collective Clemency Bill. This bill reduced the sentences of eligible inmates convicted prior to May 2nd, 2006 by three years, effective August 1st, 2006. As a result, thousands of inmates were released immediately. In fact, 83% of all releases through December, 2007, happened in August, 2006.
The pardon was motivated by the activism of the Catholic Church, including personal involvement from Pope John Paul II. The Catholics argued that prisons were overfilled, holding people in crowded conditions was inhumane, and a release was needed. They also had historical precedent on their side: after the second World War, there were regular collective pardons in Italy, but they stopped in 1992 after a parliamentary change, making the 2006 pardon the first of its kind in fourteen odd years.
Researchers Buonanno & Raphael documented what happened when the pardon went into effect. First, take a look at the incarceration rates over time:
Prior to the pardon, incarceration rates were trending up fairly slowly.
Afterwards, they trended up at a much more rapid rate!
In fact, the incarceration rate converged back to roughly where the whole thing started after less than three years. By December 2008, it had reached a rate of 98 again, compared to 103 in August of 2006.
The reason why the incarceration rate rapidly returned to the level it was initially at isn't terribly shocking: it's because crime increased!
In response to a major increase in crime, police had to start arresting more people. Recidivists and those otherwise driven to crime by the release of so many criminals needed to be arrested or the crime spree would have carried on.
In other words, incarceration incapacitates criminals, and when you shock the incarceration rate by releasing tons of criminals from a state of being incapacitated, crime goes up until they're put back in jail.
Well, unless you're fine living with a higher crime rate. If you are, then the incarceration rate can remain at a lower level.
There's a tradeoff here: if country A has a population that tends to commit few crimes regardless of policy, they can have low incarceration rates. But if country B has a population that tends to commit many crimes regardless of policy, they'll have to settle for having higher incapacitation rates if they want to realize crime levels like country A.
The populations differ in terms of antecedents of crime, so the treatment of those populations has to differ if they're going to achieve the same results.
This clears up why America has such a high incarceration rate: it's because Americans are relatively violent people!
This also tells us why El Salvador's efforts have been such a success. But before being explicit about that, here's another result from Buonanno & Raphael.
Leveraging cross-province differences in the numbers of people pardoned, they found that incapacitation effects on crime were larger when the province had a lower pre-pardon incarceration rate! Or in other words, there were diminishing returns to increased incarceration!
The reason for this is that the population is constantly in flux. There's growth, there's immigration and emigration, there's death—people come and go. There'll always be someone who is going to commit another crime. If we're lucky, there'll also always be someone there to catch them.
Some people commit more crimes than others. If you lock up all of the worst offenders, you can seriously reduce crime. For example,
- In Sweden, 1958-1980, a rigorously enforced three-strike law could have halved violent crime (x.com/cremieuxrecuei…). In this example, it was found that 1% of the Swedish population did 63% of their violent crimes.
- In America, the vast majority of people admitted to state prisons, 2009-2014, were repeat offenders (x.com/cremieuxrecuei…)
- In cities like Chicago, Atlanta, D.C., Portland, and basically everywhere else, homicide victims and offenders tend to have long rap sheets (cremieux.xyz/p/minority-rep…)
These are fairly universal findings! Crime is very concentrated: within regions, within cities, along streets, among a few people, within a few ages. The further down you go, the greater the concentration of crime perpetration in general.
The reason higher pre-pardon incarceration rates meant smaller incapacitation effects was because the worst offenders tended to be locked up already in those areas. Accordingly, if you lock up the marginal offender in a high incarceration area, you prevent fewer crimes from happening compared to if you lock up Vincenzo Megamurderer who has a rap sheet longer than a foot race.
And this replicates!
- Vollaard found that a 2001 law passed in the Netherlands that handed down ten times longer sentences to prolific offenders reduced rates of theft by 25%. This was subject to diminishing returns: as municipalities dipped deeper into the pool of repeat offenders in applying repeat offender sentence enhancements, the incapacitation effect got smaller.
- Johnson & Raphael found that between 1978 and 1990 in the U.S., each additional prison year served prevented 14 serious crimes. At the time, the average incarceration rate was 186 per 100,000. In the period 1991 through 2004, each additional prison year served prevented was just 3, and 2.6 of those being property crimes. In this period, the average incarceration rate was 396 per 100,000. America had hit the point of diminishing returns.
In elasticity terms, the Italian collective pardon revealed a crime-prison elasticity of -0.4, and with dynamic adjustment, they were as high as -0.66.
- Johnson & Raphael found crime-prison elasticities of -0.43 for property crime and -0.79 for violent crime for the 1978-1990 period.
- Levitt used prison overcrowding litigation as an instrument to estimate the crime-prison elasticity with data from the late-1970s through to the early-1990s, and he found elasticities of -0.38 to -0.42 for violent and -0.26 to -0.32 for property crime.
- A year after Buonanno & Raphael's study, Barbarino & Mastrobuoni published their own analysis of Italian collective pardons for the eight pardons laid out in the period 1962-1990. They found an elasticity of total crime ranging between -0.17 and -0.30.
- Buonanno et al. found that, in a comparison of the U.S. and Europe, the crime-prison elasticity was -0.40. They were able to do this estimation because, modern Europe at the time had developed higher property and violent crime rates than the U.S. (excluding homicide), so they exploited panel data on the reversal of misfortunes that implied.
So, back to El Salvador: they currently have the lowest homicide rates in the western hemisphere.
Some people claim they've been on this path since 2015, but it's hard to make this case, when their reversion from that year's peak was consistent with regression to the mean, and regression to the mean does not tend to make things better than ever before. It was very likely the massive lockup of people who were confirmed criminals that has brought El Salvador this level of unprecedented peace.
To put a pin in this: incarceration rates are endogenous!
Different places have different incarceration rates because they have different underlying rates of crime, different levels of and population support for and cooperation with policing, and different tolerances for keeping criminals locked up or set free. Places are in different equilibriums for numerous reasons, which is why comparisons of incarceration, policing, and crime rates are often facially meaningless. It simply makes no sense to make an unqualified statement like 'Incarceration doesn't work - just look at Louisiana, which has both high incarceration and high crime!'
To really understand the linkage between incarceration and crime requires causally informative research like the wonderful work I cited on Italy's collective pardons. To really grasp the thorny issue of crime in general requires plying your counterfactual reasoning skills so that you don't make a silly mistake like saying:
Alaskans wear bigger, puffier, more insulating coats than Floridians, yet they suffer more hypothermia deaths. Therefore, we can reject that view that coats help people to stay warm.
Bonus! An earlier thread on counterfactual reasoning about a different topic: x.com/cremieuxrecuei…
Bonus post:
You can plausibly trade off policing and incarceration. If you choose to police more, you can incarcerate less with the same crime rate and vice-versa.
America's at a place where it should be doing more of both policing and incarcerating.
The homicide rate, rates of physical violence and theft against fellow Americans—it's all too common and America should get on a trajectory like Germany or Japan.