GWAS of Personality

What could I be thinking?

What Turkheimer Doesn’t Think

OK, I am back from my travels, more or less recovered from jet-lag, and ready to express my reservations about the latest GWAS of the Big Five personality model, published by

Michel Nivard

and Ted Schwaba. Michel (cordially as ever) calls me out for my skepticism, expressed some time ago on the old site, which I no longer have access to:

What could I have meant by saying that I have “low expectations”?

Did I think—

• The idea that there is genetic variation in personality is wrong, that in fact human personality is a matter of family environment?

• That if someone tried to find correlations between SNPs and personality variables they would be stymied because there is nothing there to find?

• That any attempt to talk about the genetics of personality amounts to reprehensible genetic determinism?

• That GWAS of personality shouldn’t be funded?

• That whatever genetic associations might be found would turn out to be random, with no correspondence to what is known about personality?

• That the well-known correlations of personality with other variables would have nothing to do with genetics?

No. Hopefully it is obvious to anyone who knows anything about me that I don’t think any of those things, as I have described many times. If the first law of behavior genetics had been formulated in 2021 instead of 1991, it would have said, “All human behavioral traits are correlated with a large collection of SNPs.” I guess that is more or less what Chabris et al’s Fourth Law wound up saying.

What isn’t in Schwaba et al

Another way to go about this is to observe what isn’t in Schwaba et al. I emphasize that most of the following are virtues:

Any nature-nurture content whatsoever. When Eysenck (1951) conducted the first twin study of neuroticism, he found a heritability around .7, and concluded that “a neurotic predisposition is to large extent hereditarily determined.” This kind of talk continued until quite recently. See Bouchard (1994) among many others. It isn’t just that SNP heritabilities have turned out to be smaller than the old twin heritabilities, as I will discuss below; the whole idea that we are conducting behavior genetics of personality to try and quantify or “disentangle” the effects of genes and environment has been abandoned.

Any heritabilities over .1. Critics sometimes suggest that I don’t emphasize enough that modern heritabilities are SNP heritabilities, and as such don’t directly contradict the twin heritabilities of yesteryear. Of course, that’s true; indeed it is the main problem. But the Eysencks and Bouchards of the previous generations were willing enough to take the big twin heritabilities as evidence of genetic supremacy in the explanation of individual differences, so one way or another the small ones today have to be taken seriously as well. If the world had worked out the opposite way, say if there were no such thing as identical twins and SNP heritabilities were the first indication anyone had received about the genetics of personality, the whole modern nature-nurture debate would never have happened. I think if you were to tell the average person that about 10% of differences in personality were attributable to genetics, they would shrug. Seems about right. (By the way, I don’t really buy inflating the heritabilities by correcting them for unreliability. This is an old method that Arthur Jensen used to use a lot to produce more impressive looking heritabilities. What I was taught in grad school is that it makes sense to “correct” a measure if there is some plausible possibility of actually measuring it more accurately in the future— you are predicting a future result, not attributing reality to some idealized abstraction. The heritability of personality is what it is.)

Any serious talk about identifying the “genes for” different personality traits. The authors count out their genome wide significant SNP hits and observe their consistency across samples, but don’t put any serious effort into identifying genes. Once again, it is easy to forget that this was until very recently the point of the exercise. Plomin and Caspi (1998), titled, DNA and Personality, opens with the sentence, “The purpose of this paper is to describe new genetic approaches that are beginning to identify genes for personality and to consider how these findings will affect personality research in the future.” de Moor et al (2010), basically an early version of Schwaba et al, states, “Thus, gene-finding efforts for the major personality dimensions may yield important insights into the genetic etiology of psychiatric disease.” That’s fine; it didn’t work. But to understand what the genetics of personality does mean, you have to be transparent about what it doesn’t mean. Behavioral GWAS has too often distracted readers from noticing that old claims didn’t work out by distracting with new claims based on the latest technology. Those new claims are always less specific and causally weaker, and they have a way of not working out themselves, to be replaced yet again by the next shiny new toy.

Any concerted effort to explain the etiology of personality traits. Once the field had given up on the idea of finding individual genes related to complex behavioral traits, the next line of defense was what was called “biological annotation.” DNA is differentially expressed throughout the body, and scientists can look to see if some tissues are “enriched” in DNA with associations with the trait. Once again, there is nothing wrong with this, except that, just as in gene-finding, it over-promises. What have we learned about the development of complex human behavioral differences from all this technology? Generally, it is something along the lines of, “Human behavior appears to have something to do with the brain.” I always remember this Figure from EA2, which was presented rapturously at a BGA meeting.

Ah, now I see. It has been almost ten years since EA2. What have we learned about the biology of genetic contributions to educational attainment in that time? This is what we get here: vague associations with the brain, leading to promises that very soon we will know something more specific.

Single cell gene expression data integrated with our GWAS suggests there is no straightforward reductive mapping from broad, multifaceted personality traits to specific neuron types. However, the availability of these high powered GWAS combined with ever-expanding access to temporal and spatial brain gene expression data will enable future developmental and system-specific analysis of neural personality etiology.

Useful polygenic scores. In 2022, Robert Plomin, as always showing the courage to make specific predictions in his theoretical papers, said (he was talking about psychopathology):

In the next 10 years, polygenic scores for psychopathology will predict more than 10% of the variance and will be used in childhood to predict profiles of adult vulnerabilities.

In Schwaba et al, the polygenic scores accounted for between 1 and 4 percent of the variance, which was significantly different from zero and was not attenuated when examined within sibling pairs. But once again, if the discussion of behavioral genomics from the outset had been along the lines of, “One day, we may be able to account for 3 or 4 percent of the variance in personality using polygenic scores,” would anyone have been surprised? Would Plomin have been able to get away with all that game-changer nonsense? Sure, there may well be social science applications for such weak predictors, using them as instruments or control variables, but no one is going to use them to select embryos or assign kids to classrooms. The sensible conclusion would be that genetic data is not very useful for making predictions about individual people’s personalities. And remember that the heritabilities are under 10%, so it is no longer possible to point to some golden future of mega-samples (this one is already pretty damn big) when polygenic scores are really going to blossom.

What Do We Get? Correlations

So given that we don’t find the genes for personality, we don’t get impressive heritabilities, we don’t get biological explanations of personality differences, and we don’t get useful polygenic scores, what do we get? We get correlations. Individual difference personality is a correlational enterprise. The five factor model is defined by factor analysis of correlations among traits, and the construct validity of personality traits is established by correlating them with other things. And yes, Schwaba et al find that personality traits are correlated with other things along genetic pathways in sensible ways, just as you would expect.

The “just as you would expect” is the problem, and requires a bit of a digression in an already too-long post. The three laws of behavior genetics are null hypotheses; the point of the Three Laws paper was that testing whether h² = 0 was pointless, because h² was never equal to zero. After that paper sunk in, the field moved on from demonstrating over and over again that h2 > 0 to twin-based multivariate behavior genetics, arguing that even if the value of twin studies couldn’t be found in establishing the heritability of things, maybe it could be found in understanding the multivariate structure of phenotypes in the genetic and environmental covariance matrices.

In 2014, along with Erin Horn and Erik Pettersson, I wrote a paper called A Phenotypic Null Hypothesis for the Genetics of Personality that extended the three laws to the multivariate case. Personality provides the perfect example of what the phenotypic null hypothesis is about. Back in the nineties, people like John Loehlin got to wondering what would happen if, instead of factor analyzing raw self-report personality items (which yields the FFM) you collected twin data, estimated separate ACE covariance matrices, and factor analyzed them one at a time. The idea was to find independent genetic and environmental structures of personality, in the hope that it would lead us to real structural hypotheses, instead of endless re-assertions of, “It’s heritable.”

But a funny thing happened. When the genetic covariance matrix was factor analyzed, it yielded…. the Five Factor Model. Same for the shared and nonshared environmental matrices. Eventually Costa and McCrae came to call it, “The Puzzle of Parallel Structure.” Loehlin and Nick Martin concluded that the structure of personality is phenotypic in origin. (nb. Not environmental, but phenotypic.) The phenotypic null hypothesis states that in general, the multivariate structure of genetic and environmental data is the same, leading to the conclusion that there is no genetic and environmental structure. There is just phenotypic structure, and because genes and environments are correlated with all phenotypes, they go along for the ride. Behavioral variation causes genetic variation, not the other way around. See for example this old post.

In twin models, the ACE covariance structures are estimated from the phenotypic structures and zygosity, so it is possible to test the hypothesis that the ACE structures differ from each other and reject the phenotypic null model. Ironically, the Number 1 feature of multivariate GWAS (as advertised by Michel) is that you can estimate genetic covariances between phenotypes that have never been measured in the same data. That means you can no longer test the hypothesis that the genetic correlations are any different from the phenotypic ones, but you can still eyeball it. Ultimately, it’s a simple question: Is the genetic covariance you have found between X and Y surprising, something you would not have predicted from the phenotype? My standard example is when people proudly announce that they have found a genetic correlation between schizophrenia and bipolar disorder. No shit. There is a phenotypic correlation between schizophrenia and bipolar disorder. Finding the same thing in the genetic variance is a nice sanity check, but nothing more.

So what do we get here?

Genetic correlations estimated using LDSC indicate widespread genetic sharing between personality traits and health-relevant daily behaviors. Conscientiousness in particular was linked to reduced substance use, greater sports participation, and greater preference for low-calorie foods, as well as the downstream consequences of these behaviors: fewer spells in the hospital, healthier aging, and lower BMI (Figure 3; Supplementary Tables S31-S32). Personality was also genetically correlated with fluctuations in accelerometer-measured behavior across the day, with openness to experience linked to increased night-time activity and conscientiousness to increased activity during the day (Figure 2D; Guerreiro et al., 2024).

Or,

Those with PGIs reflecting lower neuroticism and higher extraversion, agreeableness, and openness to experience were perceived to have a more attractive personality, and those with PGIs reflecting higher agreeableness and conscientiousness were perceived to be more well-groomed.

Again, nice sanity check, but this isn’t exactly news.

Conclusion

I want to be clear at the end (before allowing myself a critical paragraph to follow) that the study is beautifully executed and for the most part humbly and realistically interpreted. I appreciate that it avoids hereditarian interpretations of the results, and in contrast I am not at all an “environmentalist” about personality. Picking apart family relations as a way to explain personality differences would produce even less news than the genetic approach. Human behavioral differences are hard to explain.

I think post-hereditarian GWAS represents a new kind of science. Behavior genetics was badly burned by the candidate gene debacle. Thoughtful investigators like these are rightfully embarrassed by hereditarian overstatement about genetic hegemony in explaining behavior. As an antidote to hereditarianism and the replication crisis, they developed a new, very cautious, paradigm. Here is an analogy. Imagine a world in which a new kind of telescope has been invented. This telescope is capable of looking far into the heavens, but unfortunately the scientists of the time don’t have any major astronomical hypotheses to investigate, and are moreover very sensitive to the possibility that any bold claims they might make would fail to replicate. So instead of generating and investigating new and untested hypotheses, the astronomical community divides the sky up into little regions, and assigns teams of investigators to point the telescope at their designated region and write down what they see according to a set of widely accepted rules. “Aha, we took a close look at the personality quadrant, and found 2,768 stars and 52 galaxies. Spectral analysis showed that there is hydrogen gas in the nebulae.” Doing this doesn’t answer any actual scientific questions that people had, but no one has to worry that it won’t replicate, and once it has been written down, some future scientist who does have something to investigate may find all these carefully recorded parameters useful.

Fine. I confess to being old enough to be dissatisfied with this kind of science. I still like daring hypotheses from independent investigators. I know those findings often don’t replicate, but…. that is a topic for another post.

Finally, I need to live up to my original promise to write a grumpy response. I just said that the authors’ interpretation of their results was mostly humble. Then there is the discussion section, which begins,

In a major consortium effort assembling data from 46 cohorts covering 611K-1.14M EUR and AFR participants, we conducted highly-powered GWAS of each of the Big Five personality traits, producing dramatic gains in the number of discovered loci, validating powerful and robust polygenic indices, and comprehensively characterizing genetic architecture, confounding, assortative mating, and widespread genetic associations with socially relevant behaviors, health, and important life outcomes. These results overhaul the state of scientific knowledge on the genetic etiology of variation in human personality, establishing a rigorous basis for genetic inference and a fundamental role of personality genetics in the human condition.

The whole section continues this breathless tone. I tell my students: Don’t editorialize, don’t advertise. Why describe your own effort as “major”, your own results as “dramatic”, “powerful”, and “robust”? (Since when is 3% of the variance powerful and robust, anyway?) Why give yourself credit for “overhaul[ing] the state of scientific knowledge?” If your work deserves all those encomiums it will eventually get them; it only makes you look insecure to say it about yourself. The fact is, I think contemporary behavioral genomics is insecure, warding off generally tiny effect sizes, smaller-than-expected heritabilities and elusive etiological explanations. The risk is that overstated conclusions of that kind get picked up by other researchers who are not as careful as these authors about avoiding hereditarianism or misplaced real-world applications of soft and small genetic results. Over-enthusiastic science can do unintended harm.