Prediction Strength in Race and IQ

In this post I’m going to argue that many people under estimate the strength of evidence favoring a genetic interpretation of racial differences in IQ because they don’t appreciate the degree to which evidence that can be explained by both genetic and non-genetic models tends to nonetheless favor genetic models. I’ll argue that this is due to some very basic rules about how probabilistic reasoning works which, for whatever reason, people tend to lose sight of in this debate.

A Confused View of Scientific Reasoning

It is possible to hold a view of scientific reasoning which says that we exclusively evaluate competing hypotheses by constructing tests which will definitely go one way if one hypothesis is true and definitely go another way if another hypothesis is true. For example, some theory of motion might make a prediction about the movement of planets which is certain to be true if the theory is correct but extremely unlikely to be true if the theory is wrong.

Holding this view, some say that we don’t have methods able to test the causes of racial differences because we don’t have any tests that singularly provide this sort of definite verification of any particular hypothesis. For instance, it is well established that gene variants which predict IQ differences between members of the same race differ in frequency between the races in a way that correlates with racial differences in IQ (Lasker et al., 2019; Fuerst et al., 2021). However, it is also known that people with gene variants that predict an increased IQ are more likely than others to be exposed to IQ boosting environments. So group differences in these genes could reflect a genetic or an environmental cause of racial IQ gaps. For this reason, testing whether groups differ in the frequency of these gene variants is not a test that will definitely give one result if a genetic model is correct and definitely give another result if it is not correct. A similar complaint can be made about many lines of evidence in the debate on race and IQ and for this reason we might conclude that the science justifies an agnostic view on the issue.

The Importance of Prediction Strength

The mistake in this view is that it assumes that evidence which can be explained by either of two conflicting hypotheses can be assumed to be explained equally well by either hypothesis. In reality, explanation exists on a continuum of quality and the property relative to explanation quality I want to focus on here is strength of prediction. By this I mean how probable a given observation is on a given hypothesis.

Consider contrasting a broadly scientific worldview with a broadly magical one. The scientific worldview is characterized by various equations which make very specific predictions about how the physical world should behave. It assigns very high probabilities to observations that follow from these equations such that the scientific theories would be cast into doubt if our observations did not conform to the equation’s predictions. By contrast, a broadly magical view of the world might not make any specific predictions about how the physical world will behave. Nonetheless, the magical worldview could produce a possible explanation for any given observed pattern of physics, for instance by simply positing that however physics turns out must be how some spirit intended it to turn out. Given this, there are no observations which are extremely improbable on the magical worldview and so no observations which are both highly probable on the scientific worldview but highly improbable on the magical worldview. Nonetheless, we consider the predictions of science to be extremely good evidence in its favor.

More formally, we can say that an observation increases the probability of one hypothesis relative to another to the degree that it is more probable on one hypothesis than another and this does not require that the hypothesis be highly improbable on either hypothesis. For instance, if some observation is 60% likely on hypothesis A and 90% likely on hypothesis B then it is evidence for B even though both hypotheses offer reasonably good explanations of the observation. Of course, if the observation were only 6% likely on A then this would be even greater evidence for B, but still it is evidence for B even when the observation is more likely than not to happen on both A and B.

Race and IQ

Returning to race and IQ, it is true that an environmentalist model can explain why IQ-linked gene variants would differ by race. However, we can tell this isn’t a very strong prediction on an environmentalist model because no one would say that serious doubt would be cast on the model if it turned out that IQ-linked gene variants did not differ by race. So the probability of them doing so on an environmental model can’t be that much greater than 50%. By contrast, because these gene variants surely pick up on real genetic causes of IQ even if they also correlate with environmental ones, it would cast doubt on a genetic model if IQ-linked gene variants that explain a significant amount of IQ variance within each race didn’t differ between races, so the probability of them doing so on the genetic model is clearly far greater than 50%. Thus, because the genetic model predicts this observation more strongly than does the environmental model, it is evidence for the genetic model even though either hypothesis can offer a somewhat plausible explanation for the finding.

It is worth noting that the typical environmentalist move here is to refocus on a different but closely related set of observations about which neither theory makes strong predictions. Specifically, environmentalists sometimes point to a paper which created a genetic score for IQ based on how well each gene variant predicted IQ differences specifically between siblings and found that the races didn’t differ in this score more than could be plausibly attributed to chance (Bird, 2021). This was in contrast to previous work which had found significant differences between blacks and whites using genetic scores constructed in this way (Pesta et al., 2020).

Constructing the genetic score in this way greatly reduced the amount of variance in IQ it could account for within race both because it was less confounded by the environment and because it was less able to detect genetic effects. Consequently, it’s been pointed out that a genetic model wouldn’t strongly predict that differences between races in this score would be statistically significant and so this null result isn’t very problematic for the model (Kirkegaard, 2022). This situation is fairly typical. Most research in the social and medical sciences is underpowered in the sense that a null result is expected even when the hypothesis being tested is true (Mallet et al., 2017; Szucs et al., 2017).

If the outcome had gone the other way environmentalists would have simply claimed that racial differences in IQ are partly caused by environments that differ between siblings (e.g. some genes lead to one sibling looking less black which leads to them being treated better by society). Because of this, until genetic scores constructed this way can account for more IQ variance within races these observations are not strong tests of either model.

Importantly, this doesn’t change anything about the reasons I gave for thinking that these genetic scores constructed in the normal way, based on how gene variants predict IQ differences between unrelated individuals, provide significant evidence in favor of a genetic model. These genetic scores are more predictive of IQ differences than are the sibling based scores partly because they capture more of the genetic causes of IQ differences and consequently their results are more strongly predicted by a genetic model and hence still evidence in its favor.

Importantly, this process is iterative so that each observation like this moderately increases the probability of a genetic model relative to an environmental one in a compounding way. For this reason I’ll point out another line of evidence that fits this pattern.

As a second example, consider the association between European genetic admixture and IQ among African and Hispanic Americans (Kirkegaard et al., 2019). This association persists after controlling for self identified race/ethnicity (SIRE) and parental socio-economic status so we know those aren’t the cause of this correlation. Relatedly, it’s also been shown that skin color does not correlate with IQ within sibling pairs of African Americans so discrimination differing by phenotypic proxies for racial admixture like skin color can’t explain why more genetically European African Americans have higher IQs either (Hu et al., 2019).

Still, an environmental theory might explain this by positing some pattern of non-random mating or differential discrimination based on cues of racial admixture more subtle than skin color. But no one would say that an environmental model would be cast into doubt if genetic racial admixture didn’t correlate with IQ. Actually, we know environmentalists would take such a finding to be evidence for their model because they did so back in the 1970s when very crude biological markers for racial admixture (blood type) produced misleading null results (Scarr et al., 1977).

By contrast, a genetic model would absolutely be called into question if racial admixture didn’t predict IQ. Thus, the genetic model once again makes a strong prediction consistent with the evidence where the environmental theory makes, at best, a weak one, and so this evidence increases the probability of a genetic model being true.

A Lack of Symmetry

Upon reading this it may occur to you that this may be true of the lines of evidence that proponents of the genetic model cite in their favor while the opposite situation is may be true of the lines of evidence proponents of the environmental model cite in their favor. Actually, environmentalists tend to cite observations which are not strongly predicted on either model. I’ll give a few examples to substantiate this claim.

First, environmentalists often cite the Flynn effect, or the fact that IQ scores rose significantly around the world in the 20th century, as evidence for their model. This is not a strong prediction of the environmental theory: if IQ scores hadn't increased much no one would think that somehow implied that racial differences in IQ were unlikely to be due to the environment. This is not just speculation since environmentalists, including James Flynn for whom the “Flynn effect” was named, did not drop did their model when they found out that the Flynn effect consists of a rise in specific cognitive abilities while the races differ in general cognitive ability which there was not a great rise in during the 20th century (Nijenhuis et al., 2013; Flynn, 2013). Secondly, notice that the view that the races differ in IQ due to genes obviously makes no prediction about whether the environmental improvements of the 20th century would improve the mean IQ within each race. By analogy, the hypothesis that different plan species within a garden differ in height due to genes doesn’t predict that the height of both species won’t greatly increase given an improvement in soil quality. The observation is just totally irrelevant.

A second line of argument which is interesting to analyze through this lens is rather implicit, namely the intuition that the very existence of the history of explicit institutional racism constitutes strong evidence favoring an environmental model. This isn’t often explicitly stated but I think it’s one of the most important intuitions behind egalitarianism. But on an environmental model it isn’t obvious that such a history is highly probable since when that explicit discrimination ended egalitarians lost no confidence in their view and immediately moved to positing secret and implicit discrimination in its place. If anything, environmentalists have become more confident in their model as distance from that history has increased. Similarly, such a history isn’t improbable on a genetic model, in fact if you think that humans will tend to react in discriminatory ways to the reality of genetic inequality then it might even be weakly predicted on a genetic model. For this reason, the “racist past” of the west shouldn’t be seen as evidence favoring racial equality.

The third point to make here gets at the heart of what is really wrong with the environmental model. One kind of evidence environmentalists point to is consists of racial gaps in specific environmental factors which are said to causally relate to IQ. However, when it is pointed out that, for instance, we spend more on the education of black students than white students, or that racial differences in lead exposure statistically account for less than a single point of the black-white IQ gap, or that differences in parental income, and therefore also all the effects of parental income, don’t cause differences in offspring IQ, environmentalists are unmoved (Last, 2020; Boyle et al., 2021). Instead the environmentalist simply claims that we haven’t measured the correct environmental factor and goes on to suggest a new list of factors we should test in the future. This is possible because the environmental model as it exists in the modern debate on race and IQ isn’t a hypothesis about any specific environmental explanation, rather, it’s an almost a priori commitment to the effect that while we don’t know what the exact correct explanation of racial IQ gaps is the one thing we do know is that the correct explanation doesn’t include genetics. When seen through this light it becomes very unsurprising that this is a hard theory to get strong predictions from.

Conclusion

If you read through the literature on race and IQ with this sort of thing in mind you’ll find that it features the testing of a combination of strong and weak predictions by the genetic theory. Proponents of the environmental model respond to these lines of evidence with some explanation that is not strongly predicted by their theory and for each line of evidence it’s a different explanation than the last. By contrast, while it is possible to imagine possible strong predictions of the environmental model, for instance it strongly predicts that if black and white Americans totally switched environments their IQ gaps would also reverse, it doesn’t make any strong predictions about hardly anything that we can currently observe. This allows it to claim consistency with all our observations despite strongly predicting none of them, but this exact situation also means that these observations imply a genetic model is more probable than an environmental model despite both models being able to explain all our observations.