National IQs Are Valid
National IQ estimates are robust, reliable, and realistic
If you follow me here or on Twitter/X, I’m sure you’ve seen a map like this, showing country-level differences in average IQs:
The figures in this map are derived from a raft of studies compiled by Richard Lynn and Tatu Vanhanen in their 2002 book IQ and the Wealth of Nations. The book itself is little more than a compilation and discussion of these studies, all of which are IQ estimates from samples located in different countries or based on diasporas (e.g., refugees) from those countries.
To get this out of the way, the estimates from IQ and the Wealth of Nations hold up. They are replicable and they are meaningful. At the same time, they are contentious. Lynn and Vanhanen’s estimates have many detractors, but virtually all of the negative arguments have one thing in common: they’re based on the idea that the estimates feel wrong, rather than with any actual inaccuracies with them.
Let’s review.
“Countries Can’t Have Mean IQs in the 70s!”
One of the most common arguments against Lynn and Vanhanen’s national IQ estimates is that it is simply impossible for whole countries’ mean IQs to be what people often consider to be so low that they’re considered prima facie evidence of mental retardation. This feeling is based on misconceptions about how mental retardation is diagnosed and defined, and misunderstandings about the meanings of very low IQs across populations. Let’s tackle definition first.
The belief that mental retardation is defined by an IQ threshold is similar to many of the arguments against the validity of national IQs in that it is based on a failure to give even a cursory thought to one’s own arguments. It’s a belief that cannot survive reading the latest version of the DSM, or for that matter, thinking of psychologists as competent people. If you open up the DSM-5 and turn to the section entitled Intellectual Disabilities, the first thing you see is the diagnostic criteria, which read as follows:
The following three criteria must be met:
Deficits in intellectual functions, such as reasoning, problem solving, planning, abstract thinking, judgment, academic learning, and learning from experience, confirmed by both clinical assessment and individualized, standardized intelligence testing.
Deficits in adaptive functioning that result in failure to meet developmental and sociocultural standards for personal independence and social responsibility. Without ongoing support, the adaptive deficits limit functioning in one or more activities of daily life, such as communication, social participation, and independent living, across multiple environments, such as home, school, work, and community.
Onset of intellectual and adaptive deficits during the developmental period.
There are four listed severity levels for mental retardation, Mild, Moderate, Severe, and Profound, and they are “defined on the basis of adaptive functioning, and not IQ scores, because it is adaptive functioning that determines the level of supports required. Moreover, IQ measures are less valid in the lower end of the IQ range.” It’s true that most mentally retarded people have IQs in the range of 55 to 70, so it’s easy to get misled into thinking that IQ is the defining factor for mental retardation. But an IQ of about 70 and below only indicates (but doesn’t diagnose) mental retardation because of what it tends to be caused by in certain populations. This is a roundabout way of saying that a low IQ indicates mental retardation because of what it means for a person's behavior, which, I'll explain, covaries with its causes.
IQs are mostly normally distributed, and IQs represent the influences of multiple different constructs and causes, but they primarily reflect differences in general intelligence. At the same time, there’s a major deviation from normality at the lower end of the scale. If you sample well enough, the picture you’d get would look like this:
The reason is that there are the normal-range causes that produce the traditional bell curve, and then there are extreme circumstances that produce extraordinarily low IQs. Contrarily, we don't know of anything that produces abnormally high IQs. There's no known one-off mutation that makes someone a genius, but there are several mutations that we now know can make a person extremely unintelligent. Consider these:
It's much easier to break a machine than to throw a wrench in it and make it work better. And that makes sense! Your wrench in the gears is likely to break something, not to increase efficiency. If you hit someone in the head hard enough, you can reduce their IQ score, but not in a million years will you turn them into von Neumann.
The reason someone's IQ drops after you bludgeon them is more singular and specific than the reasons IQ varies in the general population. A hit to the noggin may leave someone unable to flex their short-term memory, even if their visuospatial rotation capabilities are unaffected. There's some degree of isolation of function and compensation for deficits in the brain. We know this thanks to many observations, like on the effects of neural lesions.
When we say that someone with an IQ of ≤70 is mentally retarded, we're saying that they lack adaptive behavior—that they're not very bright, and so it makes living life hard. When someone with an IQ >70 is mentally retarded—which happens!—we’re saying the same thing. But if a population legitimately has a mean IQ of 70 (and some do), we'll notice that they're not drooling troglodytes who can't put on their shoes. This is because the reasons for their low IQ are not things that cause specific and extreme deficits, but instead, things that cause normal-range variation, which is far less severe in nature than something that causes massive, specific, discontinuously-caused deficits.
Supposing discontinuous causes that create major, specific deficits yields testable consequences. We can see them play out clearly by leveraging different countries’ population registers.
In large Israeli (B) and Swedish (A) datasets, we can see that when a person has a mild intellectual disability, their sibling tends to be less intelligent too. You also see this for heights, autism, or any other highly polygenic trait. This is expected with continuously distributed causes because siblings share portions of their genetic endowments.
But when a sibling has a severe—otherwise known as, "idiopathic"1—intellectual disability that's likely to have a discontinuous cause, the distribution of their siblings’ IQs is the same as the distribution for the general population. And of course they would, because the reason for the difference isn’t something that siblings should be expected to partially share.
The deficits in adaptive behavior that result from idiopathic intellectual disability are ones that make life hard to live in discrete and extreme ways. In some cases, you wouldn't even recognize them as being "retarded" because they have conditions like amnesia, where the person is clearly odd and unfortunate, and they score poorly on an IQ test, but they might sound totally coherent otherwise.
Normal-range causes lead to linear changes in adaptive behavior; idiopathic mental retardation, leads to a discontinuous decrease in adaptive behavior.
Accordingly, an IQ of 70 does not have the same meaning for members of different groups, since if a group with a mean IQ of 70 randomly pulls a person with an IQ of 70, they aren't likely to score like someone who's mentally retarded for idiopathic reasons. Arthur Jensen, incidentally, predicted this sort of result in his 1972 Genetics and Education based on some observations he made as an educator.
My student said he was looking for a good culture-free or culture-fair test of intelligence and had not been able to find one. All the tests he used, whether they were claimed to be culture-fair or not, were in considerable agreement with respect to children diagnosed as educationally mentally retarded (EMR), by which they were assigned to special small classes offering a different instructional program from that in the regular classes. To qualify for this special treatment, children had to have IQs below 75 as well as lagging far behind their age-mates in scholastic performance. My student, who had examined many of these backward pupils himself, had gained the impression that the tests were quite valid in their assessments of white middle-class children but not of minority lower-class children.
Many of the latter, despite IQs below 75 and markedly poor scholastic performance, did not seem nearly as retarded as the white middle-class children with comparable IQs and scholastic records. Middle-class white children with IQs in the EMR range generally appeared more retarded than the minority children who were in special classes. Using nonverbal rather than verbal tests did not appreciably alter the problem. I confirmed my students observations for myself by observing EMR children in their classes and on the playground and by discussing their characteristics with a number of teachers and school psychologists. My student’s observations proved reliable.
EMR children who were called ‘culturally disadvantaged’, as contrasted with middle-class EMR children, appeared much brighter socially and on the playground, often being quite indistinguishable in every way from children of normal IQ except in their scholastic performance and in their scores on a variety of standard IQ tests. Middle-class white children diagnosed as EMR, on the other hand, though they constituted a much smaller percentage of the EMR classes, usually appeared to be more mentally retarded all round and not just in their performance in scholastic subjects and IQ tests. I asked myself, how could one devise a testing procedure that would reveal this distinction so that it could be brought under closer study and not depend upon casual observations and impressions.
The distinction between types of mental retardation that are demarcated by their symptoms and, thus, by their causes, is important to understand. It's why an IQ of 70 for a Japanese person is likely to indicate an extraordinarily severe issue in need of attention, while for a Bushman, they won't have any trouble surviving. This is why the DSM-V notes in the Diagnostic Features section and bolded here: “The essential features of intellectual disability are deficits in general mental abilities (Criterion 1) and impairment in everyday adaptive functioning, in comparison to an individual’s age-, gender-, and socioculturally matched peers (Criterion 2).”
Keep that bolded part in mind and you’ll understand the importance of test norming and, knowing about discontinuous causes and the adaptive functioning deficit requirement for retardation diagnosis, you’ll never again be able to think that a mean IQ around 70 invalidates national IQs. But there’s another reason you shouldn’t, and it’s that you’re probably not an ultra-hereditarian.
In 2010, Richard Lynn pointed out that high IQs in Sub-Saharan Africa were incompatible with anything but the judgment that the group differences in the U.S. were entirely genetic in origin and even extremely poor environments have no effects on cognitive development. In response to an attempt to say that Sub-Saharan Africans had a mean IQ closer to 80 than 70, he wrote:
[The] assumption that [some of these samples of] children had IQs of 85 and 88 seems improbable. These are the IQs of blacks in the United States. It can hardly be possible that blacks in the United States who have all the advantages of living in an economically developed country, with high income, good health care, good nutrition and education, would have the same IQ as blacks in impoverished Nigeria. If this were so, we would have to infer that these environmental disadvantages have no effect whatever on IQs and even the most hard line hereditarians would not go that far.
Or, if you want to see this point made diagrammatically (courtesy of X user AnechoicMedia), here you are:
Low IQs are also predictable from national development, making them that much more realistic. Using the latest national IQ dataset, I’ll show this for Sub-Saharan Africa—a region often claimed to have invalid IQs precisely because they’re ‘too low’.
First, we’ll predict Sub-Saharan African national IQ from a regression of log(GDP PPP Per Capita) on national IQ estimates. Whether the regression is performed with or without Sub-Saharan Africa, the results are similar. The measured mean IQ of Sub-Saharan Africa is 71.96, the predicted IQ in the regression with Sub-Saharan Africa included is 74.86, and without them, it’s 76.78. Or in other words, it’s not very different.
Without the logs, the predicted IQs are 78.29 and 82.47, but that’s not an appropriate model, as you’ll see in a moment. But first, you might contest the latest national IQ dataset. The studies underlying its estimates and the methodology to assemble them are well-documented, so if you really want to contest them, you should start there. But if you dismiss them out of hand, you still can’t escape prediction by development, because we can just use the World Bank’s Harmonized Learning Outcomes (HLOs), an alternative national IQ dataset created by qualified researchers (including Noam Angrist), with good methods (test-score-linking), recently published (2021) in a respectable journal (Nature).2
Using HLOs, the observed mean IQ of Sub-Saharan Africa turns out to be 71.05. The predicted mean IQ with Sub-Saharan Africa in the regression is 72.25, and without it, 72.55. Without logs, those predictions become 77.35 and 81.85. But again, that model isn’t appropriate. The reason it’s not appropriate is due to nonlinearities that logging helps to handle. Take a look:
You can see this same nonlinearity elsewhere, such as in PISA scores:
So to overcome this issue, we can do a simple piecewise regression. Using a GDP PPP Per Capita of $50,000 as the cut-point—and feel free to go use whatever you want, it doesn’t really change the result so long as it’s reasonable—we get these regressions:
With this method, the predicted mean IQ of Sub-Saharan Africa is 76.12 with it included and 79.77 without. Using the World Bank’s HLOs instead, the results are 73.75 with Sub-Saharan Africa and 76.6 without it.
All of these estimates are pretty close to the ground-truth, and I suspect they would be even closer if all the data came from the same years instead of near years, and I know it’s closer if Actual Individual Consumption is used instead of GDP Per Capita, since that tends to iron out some of the issues with tax havens and oil barons. But regardless, it should now be apparent that low national IQs are
Not indications of widespread, debilitating mental retardation
Where they’re predicted to be given levels of national development
Thus not a reason to cast aside national IQ estimates
“Lynn Imputed Estimates, So The National IQs Are Fake!”
In his earlier national IQ datasets, Lynn ran into a lot of missing data. To get around this issue, he exploited the fact that there’s spatial autocorrelation in order to provide imputed national IQs. Some people think, however, that these imputed IQs are bad and make his data fake, failing to realize that imputation is normal and it doesn’t have meaningful effects on Lynn’s national IQ estimates.
In the most recent dataset Lynn created before he died, his imputation procedure was like so:
To calculate these imputations, Lynn and Becker (2019a, b) took advantage of the spatial autocorrelation that often exists in international data and identified the three countries with the longest land borders that had IQ means. A mean IQ, weighted by the length of the land border, was calculated and used as an estimate for the country’s missing estimated mean IQ. For island nations, the three closest countries with IQ data were identified, and an unweighted mean was calculated and imputed as an estimated mean IQ value for the missing country’s data.
Geographic imputation of this sort is the responsible thing to do when data is not missing at random and it can be predicted from other values in the dataset, because more data means more power and, given the nature of the missingness, less bias. So whether Lynn was being responsible or irresponsible is a matter of how well the imputations hold up. Thankfully, they do hold up.
The simplest way to check the robustness of Lynn’s national IQs is to compare his imputed national IQs to subsequently sampled national IQs. I’ll do this with his much maligned 2002 and 2012 national IQs. To assess the validity of the imputed IQs I’ll correlate them with our current best national IQs and the World Bank HLOs.
Lynn’s 2002 imputed national IQs correlate at r = 0.90 with our current best national IQs and 0.72 with HLOs. 102 countries were imputed which we now have estimates for. Compared to our current best national IQs, 72 were overestimates and the average estimation error was 1.47 points upwards. The average overestimation on Lynn’s part was 3.71 points, and the average underestimation was 3.75 points. Since these were imputed countries, they don’t really reveal anything about Lynn’s estimation biases in general, as they’re a select subset of generally poor or small countries. Lynn’s 2012 imputed national IQs correlate at r = 0.92 with our current best national IQs and 0.76 with HLOs. 66 countries were imputed which we now have estimates for. Compared to our current best national IQs, 37 were overestimates and the average estimation error was 0.72 points upwards. The average overestimation on Lynn’s part was 3.48 points, and the average underestimation was 2.61 points.3
I want to mention again that it should not be surprising that this procedure doesn’t really do much. It’s simple, straightforward, and scientifically uncontroversial, plus it’s theoretically sound, so it’s not shocking that it works. In a related domain, I found that this worked out just fine in the U.S.
I took the NAEP Black-White score gaps from the lower-48 and predicted them for each state from the observed gap for their immediate neighbors. The Michigan-Minnesota and New York-Rhode Island water borders were counted as neighboring state borders, and the result of imputing across state borders was a correlation of about 0.60 with the real gaps:
With a larger sample size and even more so with imputation just of observations that are missing for theoretically important reasons, this would almost certainly work better, but regardless, it’s fine: Imputation just works.
“Lynn Was Systematically Biased Against Non-Whites!”
This is easy to check, so you have to wonder why people believe it. If it were true, it would be easy to show rather than to merely assert.4 To check this, just compare Lynn’s estimated national IQs to independent collections of national IQ estimates and see if they vary systematically and meaningfully. I’ve done this, with both our current best national IQs as the reference and using World Bank HLOs as reference. The results are practically the same, but this should be unsurprising, since Lynn computed sets of national IQs based on IQ tests and based on achievement tests, and they were highly aligned. But I digress. Here’s what I found using the current best national IQs:
Compared to our current best estimates, Lynn’s original estimates were pretty close to the line, with a mix of under- and over-estimation. The degree of under- and over-estimation is minor, ranging between underestimating Sub-Saharan Africa by 1.89 points and overestimating Latin America by 4.21 points. Europe was overestimated by just 1.01 points, indicating no real evidence for the theory that Lynn favored Whites. If we drop imputed numbers, then we can see this even more clearly, because Sub-Saharan Africa without imputation was actually underestimated by an even larger margin of 2.40 points, but Europe without imputation was only overestimated by 0.18 points.
Fast forward to Lynn’s 2012 dataset and the results are even tighter, and the underestimation of Sub-Saharan Africa drops to 0.49 points, while the overestimation of Europe follows in lock-step, falling to 0.04 points. Without imputation, these numbers become 0.97 points of underestimation and, curiously, Europe is actually underestimated by 0.15 points.
In addition to producing replicable estimates, the estimates Lynn produced also weren’t off by much in general. I don’t think this should be surprising, but some people have responded to this with the argument that…
“Lynn’s Estimates Were Never Validated By Independent Data!”
This is a response to the above that is, simply put, a complete lie. The idea here is that the above validation of Lynn’s national IQ estimates is based on looking only at permutations of Lynn’s national IQ estimates, derived largely from the same sets of studies. But there is only one way to arrive at that view, and it’s to simply assume it’s true without looking at the data to see that it’s clearly not.
The wonderful thing about Lynn’s data is that he documents all of his decisions and you can peruse his national IQ database, cutting out studies you don’t like and adjusting all the estimates accordingly. Reasonable changes to Lynn’s adjustments and inclusions don’t make a difference. The data is publicly available, so you can go and confirm that yourself. But an even simpler way to show that Lynn’s estimates hold up is to just see if they hold up in data that’s entirely non-overlapping with Lynn’s data. So for that, I’ll use the World Bank’s HLOs:
The World Bank HLOs correlate at 0.83 with Lynn’s estimates with imputation and 0.89 without imputation. Compared to HLOs, Lynn underestimated Sub-Saharan Africa by a piddling 5.03 points, but he also underestimated Europeans by 0.61 points, and overestimated Oceania, Latin America, and South Asia by 5.71, 2.74, and 5.89 points, respectively.
This is just not consistent with a pattern of strong, European-favoring misestimation, and if we drop Lynn’s imputations that becomes even clearer because the underestimation of Sub-Saharan Africa increases to 5.44 points, while the overestimation of Oceania, Latin America, and South Asia shift to 2.68, 1.97, and 1.92 points. And if we look at Lynn’s later 2012 national IQ estimates, the patterns just aren’t that different, a fact that’s true whether we subset HLOs to be derived completely beyond the dates of Lynn’s data or not.
So to summarize, unless everyone shares Lynn’s biases, then his national IQ estimates do not suggest he was biased in favor of Europeans. The impacts of his imputation methods also suggest that he wasn’t biased in that direction either.
“The Samples Lynn Used Were Often Very Poor!”
A major argument that Lynn was in the wrong is that some of his samples were disadvantaged, underprivileged, or whatever other euphemisms you might wish to use for being poor, and that perhaps his samples from poor countries were worse in this respect, because he wanted them to look worse. This doesn’t stand up to scrutiny.
Lynn always caveated national IQs based on limited samples. Though you won’t see that if you just look at the resulting national IQ maps, you definitely see that if you look in his books describing what data was used, how it was gathered, and so on. It is true that some of Lynn’s referenced samples were very poor, but this always has to be considered in the light of representativeness. If poverty is the norm for a region, then samples should be poor to be representative. Similarly, if a health condition like anemia is the norm for a region—and in many places, it actually is!—then the best sample will have high rates of anemia, regardless of whether that’s ‘acceptable’ for a sample from a developed Western country.5
We can think about this in the context of a few sample means from poor places which we know to be psychometrically unbiased when compared with U.S. samples. In this case, I’ll reference samples Russell Warne provided data on from Kenya and Ghana.
The samples were from 1997 and 2015 and they used the WISC-III in Kenya and WAIS-IV in Ghana. The Kenyan test-takers came from grade 8 schools and the Ghanaian ones were public and private high schoolers from Accra, alongside a sample of university students.
The problem with this sampling is that education in Kenya in 1997 and Ghana in 2015 is nowhere near this level. In Kenya, most people didn't even reach grade 8 at the time, and in Ghana, most people don't reach university but university students were 57% of the sample! Furthermore, the Kenyan sample was from Nairobi, so it was at least sampled in what was, as of 2018, Kenya's 2nd-richest county and home to <9% of the population. The situation was even less representative in Ghana, where the sample was from Accra and thus also <9% of the population and a third of Ghana's GDP.
These samples were highly unrepresentative of their respective populations because they were relatively elite. The bias was not as bad in the Kenyan sample: they were one to two SDs above the country as a whole in socioeconomic status. But in Ghana, they were three to four SDs above the average. That's like comparing high school dropouts to college graduates.
Warne knew of this problem and wrote:
As eighth-graders, the members of this sample were more educated than the average Kenyan. In 1997, the average Kenyan adult had 4.7 years of formal schooling; by 2019, this average had increased to 6.6 years.
And:
Like the [Kenyan] sample, this [Ghanaian] sample is much more educated than the average person in the country. In 2015, the average Ghanaian adult had 7.8 years of schooling (which increased to 8.3 years by 2019), whereas 61% of [this sample] had 16 years of education or more.
Knowing this, mentally adjust their scores downward in accordance with how much you think being highly educated and relatively wealthy should overstate their scores relative to the general population. So, what were their scores? Here:
To put these scores into intuitive terms, we need to do two things: the factor correlation matrix and an assumption of equal latent variances. The correlation matrices were provided in the paper's figures 1 through 3. If we assume that the scale we want is based on a mean of 100 and an SD of 15 ("the IQ metric"), we just place the gaps on that scale (i.e., *15 + 100), subtract the number of factors times 100, divide that by the square root of the sum of the elements in the correlation matrix (we'll use the American values, since their sample was larger), and add back 100.
The Kenyan sample, which was 1-2 SDs above the Kenyan average in terms of education, had a mean IQ score of 79.08 versus the American average of 100. The Ghanaian sample, which was 3-4 SDs above the Ghanaian average in terms of education, had a mean IQ of 92.32 versus the American average of 100. And do note, the American average the Ghanaian sample would be compared to would be lower than the average the Kenyan sample was compared to due to demographic change over time.
These estimates are much higher than Lynn’s, and that makes sense, because these are clearly socioeconomically extremely well-off samples, perhaps not by Western standards, but certainly by their own national standards. But some people think these are the sorts of samples that should be used to represent poor countries. That’s wrong, but the view has its supporters, like Wicherts, Dolan, Carlson and van der Maas, who claimed as much in 2010. But Lynn called this out at the time too:
Wicherts, Dolan, Carlson & van der Maas (WDCM) (2010) contend that the average IQ in sub-Saharan Africa assessed by the Progressive Matrices is 78 in relation to a British mean of 100, Flynn effect corrected to 77, and reduced further to 76 to adjust for around 20% of Africans who do not attend school and are credited with an IQ of 71. This estimate is higher than the average of 67 proposed by Lynn and Vanhanen (2002, 2006) and Lynn (2006).
The crucial issues in estimating the average IQ in sub-Saharan Africa concern the selection of studies of acceptable representative samples, and the adjustment of IQs obtained from unrepresentative samples to make them approximately representative. Many samples have been drawn from schools but these are a problem because significant numbers of children in sub-Saharan Africa have not attended schools during the last sixty years or so, and those who attend schools have higher average IQs than those who do not.
Lynn then reviewed general population studies and those returned a Sub-Saharan African mean IQ in the 60s. He subsequently reviewed primary school studies and the data yielded a median of 71, which he adjusted to 69 to account for the fact that only about 80% of the population gets a primary school education in Sub-Saharan Africa. After that, he reviewed studies of secondary school students, and those returned a mean IQ in the 70s, but that is not an acceptable estimate for a few simple reasons.
(1) many adolescents in sub-Saharan Africa have not attended secondary school and tertiary institutes. For instance, Notcutt estimated that in South Africa in 1950 only about 25% of children aged 7-17 were in schools and “we cannot assume that those who are in school are a representative sample of the population”. (2) Entry to secondary school has generally been by competitive examination, resulting in those with higher IQs being selected for admission. Thus, “entry to secondary schools in the East African countries of Kenya, Uganda, and Tanzania is competitive… approximately 25 percent of the population complete the seven standards of primary school and there are secondary places for 10-12 percent of these”. Similarly, Silvey writing of Uganda around 1970 stated that at this time only 2% of children were admitted to secondary schools and entry is determined partly by a primary school leaving examination, and Heynman and Jamison writing of Uganda in 1972, note that admission to secondary school is based on “achievement performance on the academic selection examination… and there are secondary school places for only one child in 10”.
This is egregious, but this is typical for higher estimates for poor places. They tend to be based on samples that are, as in these cases, relatively privileged, pre-screened for IQ, etc., and despite that, still generally not that impressive in socioeconomic or IQ terms. These also aren’t the worst estimates like this that Lynn has noted. My favorite was this:
WDCM include a number of studies that cannot be accepted for a variety of reasons. Their samples of university students are clearly unrepresentative. The Crawford-Nutt sample consisted of high school students (IQ 84) in math classes admission to which “is dependent on the degree of excellence of the pupil’s performance in the lower classes” and described as “a select segment of the population”. The students were also coached on how to do the test and “Teaching the strategies required to solve Matrix problems yields dramatic short-term gains in score”. This is clearly an unrepresentative sample.
Curiously, people usually ‘get this’ sort of representativeness issue when it comes to things like China only sampling from rich, well-off areas to look good in international assessments like PISA and TIMSS, but they don’t get this when it applies to samples that look rich by one country’s standards and poor by the standards of the developed world. It’s precisely those sorts of large-scale examinations which bring me to the easiest way to vindicate Lynn’s sampling in a very general sense.
Large-scale international examinations like PISA have sampling frames, requirements to be met for a sample’s scores to be considered valid and comparable to those of other countries, and if countries meet them, then it’s likely their samples were sufficiently representative to make a statement about a country’s youth or, in the case of assessments like PIAAC, its adults. These large-scale assessments have standards for sampling, and they also have psychometric standards. Their samples end up representative—at least of developed countries (see above in this section)—and their test scores aren’t biased, and these still correlate highly with the results produced by Lynn and his colleagues. This is the basis for the World Bank’s HLOs, so by this point in this article, you’ve already seen this point made multiple times! And to be completely fair to Lynn, he made this point too, people just neglect that he did. It’s no doubt part of why he computed his own academic achievement test-based national IQs (see Footnote 3).
“But IQs Below 60 Are Unrealistic!”
Some people probably still think that some of Lynn’s estimates are too low. Lynn believed the same thing in cases where there just wasn’t enough data, which is why he Winsorized national IQ estimates and expressed his doubts about extremely low estimates.
recently noted that sub-60 IQs also aren’t empirically supported. Through reviewing additional evidence on countries listed as having sub-60 IQs, he found that each one ended up with either no estimate (due to missingness) or an IQ above 60.The unstated part of the argument that some national IQ estimates are unrealistically low is that somehow this disqualifies larger portions of the dataset, but that’s a non sequitur.
“But Lynn Disadvantaged Poor Countries By Using Child Samples!”
Some people have claimed that Lynn made poor countries appear to perform worse than they actually do by using samples of children instead of adults. This criticism reveals a lack of awareness that test scores have age-specific norms. But if we assume it’s a legitimate criticism, then it suggests Lynn unfairly advantaged poor countries since the existing not-so-strong evidence shows that those countries are more likely to have scores that decline relatively with age.
“National IQs Suffer From Psychometric Bias!”
During Lynn’s life, he didn’t focus much on psychometric bias. This is fair, because by the time it became a big focus for the field, he was already old. But in any case, as I’ve noted above, we have little reason to think it’s a big concern for most estimates. Furthermore, it’s not even clear that bias is systematic in general. Consider, for example, the comparison of Britons and South Africans I discussed here. In that comparison, there was bias, but it favored the lower-scoring South African group!
In general, when people find representative samples and no psychometric bias, the results aren’t different from what Lynn found, so unless someone wants to substantiate this concern, it’s little more than a waste of ink.6
“The Flynn Effect Means National IQs Are Liable to Change!”
The Flynn effect is widely misunderstood. I've written an article on this that goes into much greater depth about what I mean. But the important point when it comes to national IQs is that the Flynn effect is not about differences in intelligence, instead, it primarily concerns test bias. The existence of the Flynn effect also doesn’t imply there will be convergence between countries, it cannot be said to be the source of any convergence across countries without evidence that doesn’t currently exist, and the Flynn effect is explicitly adjusted for in national IQ computation. It just doesn’t have any relevance to the discussion because the evidence for larger Flynn effects during catch-up economic growth is extraordinarily poor.
But furthermore, increases in scores for cohorts over time will sometimes reflect bias rather than ability gains and relative cognitive performance across countries is generally very stable. There’s not really a reason to consider this argument, even for countries proposed—but not shown—to have major upward swings in their national IQ, like Ireland:
“Lynn Gets the Causality Backwards From IQ to GDP!”
We know that IQ differences are partially causally explained by differences in brain size. Because development seemingly minimally impacts brain sizes—and it theoretically should not anyway—, brain sizes can be used successfully to instrument for national IQs, allowing us to estimate the causal impact of national IQs on outcomes like growth, crime rates, and so on. This has been done and it works well. The same result also turns up using ancestry-adjusted UVR and numeracy measured in the 19th century, both of which also can’t be caused by modern development.
These results suggest that if Lynn is getting the causality backwards from IQ to measures of national economic success, he’s still dominantly correct that national IQs precede development. Not only that, but national IQs are, as mentioned, largely stable over time, despite the world experiencing a lot of development. We can see this very directly using the World Bank’s HLOs again. The paper introducing them includes this diagram, showing (a) percentages enrolled in primary education over time, and (b) HLOs over the same period. Notice the dramatic increase in the former and relative stability of the latter.7
Since we do know there’s a considerable degree of stability in measured national IQs, we can leverage stability as an assumption and see the curious result that, over time, it looks like Lynn’s estimates are vindicated more and more, because development measures have gotten more in line with his national IQs!
Ultimately, people who want to argue Lynn got causality backwards aren’t really taking issue with the national IQ estimates themselves, they’re just specifying an alternative claim that they hope sounds like it can invalidate Lynn’s estimates. But—and here’s the kicker—Lynn firmly believed that national IQs would increase with development; his estimates were point-in-time estimates, not the final letter forever and ever, and he fully expected them to change because he thought very poor places were environmentally disadvantaged. So this isn’t even really an argument against Lynn’s general views per se.
“What About [Fake Data]?”
People often reply to national IQ estimates with news articles, misinterpreted scientific articles, or nonsense sources that they allege show low-scoring countries actually do very well. Three examples that were brought up to me recently were Iran, India, and online IQ tests.
The example of Iran was a meta-analysis of different studies of Iranians. Someone brought this up to me to claim that Iran actually had a national IQ like America’s, at 97.12. Perplexed, I asked the simple question: Did they use American or British norms? And the answer is ‘no’, the test norms were Iranian, so if anything, these samples had an IQ below what’s expected—that is, a mean of 100. There’s really no need to ask further questions about these results, because being on different norms and not having the handbook handy to make the scores comparable means that they do not permit international comparisons. But this is a pretty standard sort of argument for people to make to contradict Lynn’s estimates.
The example of India was a news report about alleged testing of Indian students by Mensa’s India chapter. The report reads:
In the past couple of months, Mensa India, Delhi, administered its internationally recognized IQ test to over 4,000 underprivileged children in Delhi and NCR as part of a unique project aimed at identifying and mentoring poor children with high IQ. Of the 102 extremely bright children it selected, over a dozen, including Amisha, achieved an IQ score of 145-plus, which puts her in the genius category.
The others achieved IQ scores of 130-145, which puts them in the category of ‘very gifted’ children. The average score in Mensa India’s IQ test is between 85 and 115. Interestingly, all of these children are sons and daughters of labourers, rickshaw pullers, security guards, street vendors, etc.
Now, does this say anything in defiance of Lynn’s numbers? No, because we don’t know the norms. We don’t even know much about these statistics at all, we just have hearsay without accompanying statistics. This barely rises above an anecdote, but it is the sort of thing that people will misinterpret to mean Lynn was wrong, somehow.
The online IQ test argument is one of the silliest, but it gives people a lot of hope. They argue that test results such as these, produced by people self-selecting into taking tests of unknown and—when vetted—dubious validity suggest that national IQ estimates are way off the mark. But they don’t do that, because they’re based on selective rather than representative samples. Relatedly, people sometimes argue that national IQ estimates are wrong because international students taking American tests like the GRE perform better than, say, Americans, when the national IQ of their country of origin is lower than America’s. But once again, this test-taking is selective, making it less relevant than proponents of using that data to reject national IQs might hope.
Another less common strategy to reject national IQs is to just unreasonably ignore data. For example, today I encountered someone who plotted Haiti’s national IQ over time with two datapoints, one from the late-1940s and the other from the late-1970s. They alleged that Haiti had gotten much smarter over time, with their national IQ rising from around 60 to almost 100. But looking at all available data, this view cannot be supported. What they did to make their claim isn’t even a reasonable way to compute a national IQ. Getting a national IQ estimate requires looking at multiple lines of evidence and qualifying the inclusion of samples and whatnot, whereas their strategy was to act like there were just two studies to discuss and to conclude that the later one was the ground-truth regardless of its reliability or any other of its qualities.
“National IQs Don’t Measure Differences in Intelligence!”
This complaint takes two forms. The first is the more general rejection of intelligence tests measuring intelligence. That’s not relevant to national IQs and it’s poorly supported, but the arguments on that topic are familiar, so I’ll skip to the relevant argument.
The second form is that, for some reason, differences in national IQs are due to different factors than those that explain test performance within populations. This perspective is incompatible with measurement invariance, so it is necessarily wrong for any psychometrically unbiased comparisons. A sense in which the claim can be recovered is that for a given unbiased comparison, there might be mean differences in specific factors rather than in g—a sort of international versions of the contra hypothesis for Spearman’s hypothesis. This is not the case for the PISA tests or any other unbiased national IQ comparison of which I’m aware, so while it’s a possible perspective, it’s empirically contradicted at the moment.
Some people prefer achievement tests to national IQs, despite the positive manifold strongly suggesting that achievement tests and IQ tests both measure g8, and confirmatory factor modeling confirming that. The people with that preference also tend to make another, related argument: that achievement tests are not measures of g and cannot be treated as such. But like the above arguments, there are only negative reasons to think this is true for international examinations like the PISA tests, which have a strong general dimension, much like standard IQ tests.
Criticisms of National IQs Are Intellectually Lazy
People really want national IQ estimates to be debunked or, worse, to feel personally favorable. So they concoct a lot of bad arguments to make it sound like national IQs are off. To recap, here are some of the ways:
They confuse themselves and others about how mental retardation is defined and act as though it’s defined by IQ alone, so certain national IQs must be implausible. In doing so, they ignore the importance and existence of norms, as well as the modern definition of mental retardation itself.
They claim methodological choices like imputation are extraordinarily biasing when that is not the case and checking shows that the choice is actually not even biasing in the direction of national IQs being unfair to the groups national IQs have been proposed to illustrate bias against.
They claim that sampling is directionally biased in a certain way, when inspection of the data generally shows that, if anything, it’s biased in a way that leads to understated lower-tail cognitive differentiation.
They make claims that ‘sound right’ like that comparing children and adults is bad, even when we have age-based norms, so this cannot be a genuine criticism unless it is theoretically qualified that somehow children in certain countries are more disadvantaged than their adults and that this disadvantage translates to lower cognitive performance. I’m not aware of anyone who believes this theoretical qualifier and existing evidence speaks against it.
They make groundless extrapolations like ‘The Flynn effect means national IQs are worthless’ or ‘national IQs have changed a lot’ and they refuse to justify these inferences.
They look for odd references and outlier studies to justify throwing out a whole corpus of material.
Etc.
The common thread between different national IQ criticisms is the weaponization of ignorance. Critics toss out claims they think are right or which sound right, and they don’t check their work. A stand-out example is that people regularly say that Lynn was biased against Sub-Saharan Africans on the basis of his use of samples that are well-off by the standards of their countries purely because they are poor by the standards of the developed world. But this argument cannot stand. It has no merit, and it only serves to insult the reader and to convince them that the person making the argument has done some of the required work to dismiss Lynn’s estimates, when they’ve really only done the required work to say they’ve just barely cracked open the book!
After throwing out enough criticisms, people feel that national IQ estimates simply cannot stand, that they must be wrong, or so many criticisms wouldn’t be possible in the first place. But on this they’re wrong, and the very fact that they have so many criticisms is a strike against them, because the criticisms are so uniformly bad that they should embarrass the person making them. Making matters worse, critics seem to never go back when they’re shown to be wrong. Their attempted debunkings get roundly debunked and national IQ estimates remain reliable as ever, and they just keep making the same tired arguments that no right-minded person could still believe.
The defining feature of criticisms of national IQs is not all of this lazy argument though, it’s what comes next: insults. People like Lynn are taken to be ‘stupid’, people who believe in a given estimate that upsets people regardless of how well-supported it is are taken to be ‘morons’, and looking into and understanding national IQ estimates becomes less common because, after all, the only people who would look into them are the sorts of ‘stupid moron racists’ who actually bother to check their work.
Questions? Comments? Concerns?
Think I missed any big arguments? Want more details about something I said? Need something explained? Noticed a grammatical error, spelling mistake, or other triviality? Think I’m right or wrong about some claim? Have more data for me to look at?
Then tell me, because this is a living post that will be updated over time.
Now enjoy the most up-to-date national IQ map. It’s imputation-free!
These are technical terms. It should be clear that I am not saying that “idiopathic” is defined as “severe” or vice-versa, but some people have reading comprehension difficulties so I’m saying this explicitly.
You can also use the Human Capital Index, the Basic Skills Dataset, World Bank Test Scores, Learning-Adjusted Years of Schooling, Rosetta Stone Scores, and many more. They’re interchangeable with or without overlapping underlying data.
Lynn only imputed 51 national IQs in 2019 and there weren’t many entries in the imputed list shared with the HLO dataset, although the results for the current best national IQs were similar to the correlations with the 2002 and 2012 imputed national IQs (r = 0.73). Warne noted that these imputed national IQs had only modest or, in a case where the n = just 8, a negative correlation with other measures, so they might be unsuitable. In his own words:
The weak or negative correlations between these geographically imputed IQs and the achievement test data from some of those countries is a major cause of concern and makes it hard to identify any valid interpretation of these IQ scores.
So my recommendation if you’re unhappy with these national IQs is to just use the dataset without imputations. Lynn’s dataset includes documentation of all studies included, methodological decisions, and six different national IQs affected by different methodological choices, so this is easy. These are, in Warne’s words:
An unweighted mean of all adjusted IQs for each sample within a country,
A mean of all adjusted IQs weighted by the size of each sample within a country,
A mean of all adjusted IQs weighted by the size and quality of each sample within a country,
A mean IQ based on educational achievement data from the TIMSS, PIRLS, or PISA tests
An unweighted mean of the previous two estimates IQs (IQ + Achievement), and
A mean equal to either [5] or a mean imputed from neighboring countries that do have IQ data.
The adjustments mentioned are for the Flynn effect and the country a test is normed in.
To make this point they would also have to ignore that Lynn is the person who popularized the finding that Asians have higher average IQs than Whites.
The reality that poor countries will have poor samples if they’re going to have representative samples has to be caveated in a way that actually leaves Lynn looking good. This is the case because, in order to take part in, say, a verbal test, you have to be able to read, to have a modicum of education. Because education is not the norm, in many areas, that means that valid samples of the sort Lynn would use because they could produce plausible scores will necessarily overrepresent the skills of people in very poor places.
But, thankfully, this may not be too concerning in many places in the developing world. This is a point I’ve made before. Regardless, if it were concerning, it would be quantitatively concerning, and adjustments could be made to account for it.
It’s also something Lynn was concerned about, albeit from the angle that he should use tests more likely to be ‘culture-free’ like Raven’s matrices rather than, say, vocabulary tests.
Using the HLOs for 2010, 2017, 2018, and 2020, a cross-lagged panel model suggests minimal impacts of GDP PPP Per Capita on HLOs, but consistently finds impacts of HLOs on subsequent years’ GDPs. It’s not clear this is meaningful, because there are only four years, they’re close together, GDP should probably have its returns on further removed cohorts from the year of GDP measurement, and the last year in the sample is the start of the COVID pandemic, which disrupted economies and education systems alike. With this model and robust standard errors, I found just one significant GDP to HLO effect, from 2018 to 2020 (p = 0.005).
HLOCLPM <- '
# Regressions for HLO and GDP across time
HLO_Score_2017 ~ HLO_Score_2010 + GDP_2010
GDP_2017 ~ HLO_Score_2010 + GDP_2010
HLO_Score_2018 ~ HLO_Score_2017 + GDP_2017
GDP_2018 ~ HLO_Score_2017 + GDP_2017
HLO_Score_2020 ~ HLO_Score_2018 + GDP_2018
GDP_2020 ~ HLO_Score_2018 + GDP_2018
# Correlated residuals
HLO_Score_2010 ~~ GDP_2010
HLO_Score_2017 ~~ GDP_2017
HLO_Score_2018 ~~ GDP_2018
HLO_Score_2020 ~~ GDP_2020
'The distinction has to be qualified to be properly understood. To learn more about the distinction between cognitive testing and achievement testing, see here:
Explaining Anomalous GCSE Results
Britain’s Black African student population does much better on General Certificates of Secondary Education (GCSE) examinations than on IQ tests. The gap on IQ tests tends to be somewhere between 7 and 10 points, whereas lately there’s generally no gap or higher performance for Black Africans relative to White Britons when it comes to the GCSEs.
Thanks for the extensive, in-depth analysis of this highly maligned research.