Coinciding with the launch of his book, Human Diversity: The Biology of Gender, Race, and Class (see my thoughts here), Charles Murray had an opinion piece in The Wall Street Journal where, among other things, he discussed the importance of polygenic scores in making predictions.

Polygenic scores are scores derived from DNA variants. Bits of DNA that are correlated with a phenotype, such as height, can be assigned a value, and then these values can be added together to produce a predicted height score. That predicted height score is called a polygenic score.

For height, this is mildly interesting, but now polygenic scores are available for educational attainment, socioeconomic status, and risk for a variety of mental health diagnoses. Almost all polygenic scores for psychological traits or behavioral outcomes are not ready for practical use at this time, but it is likely that they will be eventually.

In his Wall Street Journal piece and in Human Diversity, Murray (2020) sees polygenic scores as a game changer for the social sciences:

Polygenic scores are revolutionary because they are causal in only one direction. They don’t drop because tests make you nervous or rise because you grew up rich. They’re impervious to racism and other forms of prejudice. Socioeconomic and cultural environments can play an important role in how those bits of DNA are expressed, but they don’t change the codes themselves. That means polygenic scores will offer social scientists something they’ve never had before: a secure place to stand in assessing what is innate and what is added by the environment.

Charles Murray, The Wall Street Journal, January 27, 2020.

I agree with most of what Murray says here. My biggest qualm is that studies that generate polygenic scores do not necessarily identify the segments of DNA that lead to traits or outcomes. When I discuss genetics in my upcoming book (In the Know: Debunking 35 Myths About Human Intelligence), I am careful not to say that identified DNA variants/genes cause intelligence differences. Instead, I specify that these DNA variants are correlated with intelligence.

But the sentence about about how polygenic scores prompted a response (see full version here) from Michelle N. Meyer and Patrick Turley responded, saying, in part,

It is true that one’s genome is fixed at conception and generally doesn’t change regardless of environmental stimulus. However, polygenic scores only reflect the particular group of people studied in a particular time and place in which they lived. This means that polygenic scores can and do reflect racism, sexism or other prejudices, as well as more benign environmental factors. [PARAGRAPH BREAK] For example, in a society where people of color are denied access to childhood enrichment programs or adequate nutrition, a polygenic score for IQ might reflect genetic variants associated with skin pigmentation. This polygenic score would indeed moderately predict the IQ of people on average, but much of that predictive power would simply reflect social choices, not innate biology. For this reason, the notion of “genetic potential” is deeply problematic and misleading.

Michelle N. Meyer, & Patrick Turley, The Wall Street Journal, February 3, 2020

No one disputes that polygenic scores are useful for predicting outcomes in a similar environment as the environment that the scores were generated in. That’s true with any predictor. A hunter-gatherer in the jungles of Papua New Guinea with a high IQ would be predicted to attend and graduate from college if he or she lived in the United States. But living in an environment where that person can’t attend college means that the prediction is wrong. That doesn’t make IQ useless or invalidate its predictive power for people living in places where college is available. The same is true of polygenic scores. Fine. I’m glad we killed that straw man.

But Meyer and Turley engage in more seriously flawed logic that goes like this:

  1. Polygenic scores correlate with life outcomes, but may not be the biological causes of those outcomes.
  2. A purely environmental or social cause–such as racism or discrimination–may unfairly make some people unable to fulfill their genetic potential.
  3. Therefore, polygenic scores “. . . can and do reflect racism, sexism or other prejudices . . .”

Point #1 is correct, unquestionably. Point #2 is also correct. Where the logic breaks down is moving from Point #2 to Point #3. Just because something may occur does not mean that it does. In other words, even though Meyer and Turley can imagine a society where discrimination leads to spurious, environmentally-caused correlations with polygenic scores does not mean that these correlations are environmentally driven in the real world.

Maybe 21st century America is discriminatory or racist enough to make some polygenic scores correlate with life outcomes or psychological traits for purely social reasons. Or maybe not. But by merely stating that this scenario is possible does not mean that Meyer and Turley have proved that it actually does happen.

Lewontin’s bait-and-switch

I propose that we call this strategy Lewontin’s bait-and-switch. In 1970, biologist Richard Lewontin argued that within-group individual genetic differences in IQ did not contribute to between-group differences.

In short, Lewontin imagined a person planting two random handfuls of seeds into two different types of soil: one being rich, fertile soil, and the other being dry, barren soil. Because the seeds are selected randomly, there are no genetic differences between them. As a result, all between-group differences must be environmentally caused. Conversely, within a group, all the seeds share the same environment, so all within-group differences must be genetic in origin. He applied this analogy to humans to say support his belief that between-group differences in IQ were completely environmentally caused.

People find Lewontin’s seed analogy very convincing (Warne, Astle, & Hill, 2018). But it has a flaw in the logic: human subpopulations are not formed randomly. In fact, by definition, genetic differences between populations exist (see, for example, The 1000 Genome Project Consortium, 2015). Lewontin’s bait-and-switch is generalizing from an imaginary scenario to reality without showing empirically that the two resemble one another.

It’s telling that in both of these examples, Lewontin’s bait-and-switch is being used to undermine acceptance of genetic influences on behavior. It’s a rhetorical strategy–not an empirical argument, and it is not unusual among blank slatists or other people who resist genetic explanations of human behavior. (I have seen it multiple times in peer review of my manuscripts.)

Who to believe?

So, who to believe? Neither Murray nor Meyer and Turley brought data to the table. (And indeed, the opinion pages of a newspaper probably aren’t the place for that.) When neither side has data, there are two heuristics that one can use to determine which side to believe.

First, one can make a decision based on one’s own prior beliefs about the more plausible conclusion. The second option is to base a decision on parsimony: the simplest explanation that fits the data should be preferred.

Applying the first heuristic is completely subjective. Maybe you think America is racist, sexist, or discriminatory to create artificial correlations of polygenic scores for different groups of people. Or maybe you don’t. In the end, this heuristic only encourages people to continue believing what they already think, and it does not bring anyone closer to the truth.

Parsimony, on the other hand, is simple, elegant, and not reliant on subjective beliefs. Murray’s position is that polygenic scores, more or less, tell us exactly what they seem to do: the relationship between genetic heritage and traits/outcomes. The Meyer-Turley position is that polygenic scores do describe a relationship between genes and outcomes, but that it is heavily moderated through an environment that creates gene x environment interactions across racial, ethnic, and sex groups. Murray’s position is more parsimonious and should be the default belief.

The best way to break the stalemate, though, is with data. Until then, a tentative belief in the parsimonious explanation–that polygenic scores do reflect at least a partially biological origin of psychological traits and life outcomes–is best. The fact that someone can imagine a scenario that would disprove the parsimonious belief is not relevant (especially if it posits unproven complexities to a theory) unless there is data to support it.

Note on Influence Population Structure

The above discussion ignores how population structure impacts polygenic score prediction accuracy, mostly because neither Murray nor Meyer and Turley mentioned it. In short, as they currently stand, polygenic scores are much more accurate when predicting traits in Europeans than in other racial groups (see Domingue, Belsky, Conley, Harris, & Boardman, 2015, for a good example).

This is one of the reasons why polygenic scores for psychological traits are not ready for practical use right now. But this problem will be fixed once the genetic databases become more diverse. For the time being, though, polygenic scores are less accurate for non-Europeans–but this is at least partially due to a genetic artifact. It may or may not also be due to discrimination in the environment. One critical piece of data to this debate will be whether polygenic score predictions for non-Europeans converge with the predictions made for Europeans when the databases that generate the scores become more diverse.

Update: Another example

After I published the original version of this blog post, someone on social media reminded me of another example of Lewontin’s bait-and-switch. Jencks et al. (1972) gave the example of an imaginary society where people with red hair were denied an education. This would create a purely environmentally-caused correlation between the genes for red hair and academic achievement.

Yep! Such a society would create a correlation between polygenic score and a social outcome that is not purely genetically caused. But a hypothetical situation that does not resemble reality tells us little or nothing about how to interpret data in the real world.


The 1000 Genomes Project Consortium. (2015). A global reference for human genetic variation. Nature, 526, 68-74. doi:10.1038/nature15393

Domingue, B. W., Belsky, D. W., Conley, D., Harris, K. M., & Boardman, J. D. (2015). Polygenic influence on educational attainment: New evidence from the National Longitudinal Study of Adolescent to Adult Health. AERA Open, 1(3), 1-13. doi:10.1177/2332858415599972

Jencks, C., Smith, M., Acland, H., Bane, M. J., Cohen, D., . . . & Michelson, S. (1972). Inequality: A reassessment of the effect of family and schooling in America. New York, NY: Basic Books.

Lewontin, R. C. (1970). Race and intelligence. Bulletin of the Atomic Scientists, 26, 2-8. doi:10.1080/00963402.1970.11457774

Murray, C. (2020). Human diversity: The biology of gender, race, and class. New York, NY: Twelve.

Warne, R. T., Astle, M. C., & Hill, J. C. (2018). What do undergraduates learn about human intelligence? An analysis of introductory psychology textbooks. Archives of Scientific Psychology, 6, 32-50. doi:10.1037/arc0000038