Reminder: The Thursday Class is only for those interested in studying uncertainty. I don’t expect all want to read these posts. Pease don’t feel like you must. Yet, I have nowhere else to put them. Your support makes this Class possible for those who need it. Thank you. MATH ALERT!
I don’t (and you shouldn’t) love GWAS tests of genes/SNPs and personalities; the evidence is thin and the results are based on weak and wrong techniques. Genes are crucial: it is our understanding of how which is questionable.
Video
Links: YouTube * Twitter – X * Rumble * Bitchute * Class Page * Jaynes Book * Uncertainty
HOMEWORK: Read the paper!
Lecture
The paper is “Robust inference and correlates from genetic associations with personality” by Ted Schwaba and many others in Nature. A portion of the Abstract (I still believe the word novel ought to be banned from all scientific publishing):
Personality traits describe stable differences in how people think, feel and behave, and how they interact with and experience their social and physical environments1,2. Many questions remain unanswered about associations between DNA and personality traits, such as their robustness, their generalizability and the biological and social pathways through which they act. Here we meta-analyse data across 46 cohorts comprising 611,037 to 1.14 million participants with European-like and African-like genomes for genome-wide association studies (GWAS) of the Big Five personality traits (extraversion, agreeableness, conscientiousness, neuroticism and openness to experience), and data from up to 50,725 participants for within-family GWAS. We identify 1,260 lead genetic variants associated with personality, including 824 novel variants3. …Polygenic prediction, genetic correlation and Mendelian randomization analyses indicate that personality traits have widespread, potentially causal associations with consequential behaviours and life outcomes. Overall, we find that the genetic architecture of personality is robustly generalizable, minimally confounded and widely relevant to human experience.
My argument is this: while it is obvious that genes are important and can account for some behavior, the methods to show this at personality levels are crude, and because these analyses accept the machine metaphor of life as a premise, which I say is false, and because these analyses add layers of model atop model, applied on disparate datasets, use suspect assumptions, they produce findings that are over-certain and are or tend to be misleading.
I won’t here reargue against the machine metaphor, having done so many times before (here and here). Suffice to say that organisms are wholes, and not mechanistic slaves to “selfish genes”, and that bodies have many ways to accomplish the same task rather than relying wholly and always on singular genes. Methods that correlate this or that gene with scored behavioral measures are thus bound to be incomplete, when not wrong.
We have already examined attempts to quantify the unquantifiable, and that the unquantifiable often includes human behavior, especially in those attempts to categorize people by personality types. We know of no thing more complex than people, and people do the damndest things at the oddest times. Thus, correlations of scored complex behaviors or personality “types” with certain variants of genes, is bound to be like any other correlation analysis where direct cause is not measured.
Now to the paper itself.
If you have to, please first review the lectures on regression (right way, wrong way), as I’ll take these as given. (I mean this warning, now.)
The authors’ idea in rough form is this: in several different groups, some with European ancestry and others with African, various SNPs (single nucleotide polymorphisms, spots on the genome that differ by some set amount between people) were identified, and then questionnaires with scored answers were given; and finally statistical tests of differences in means of the answers were calculated for each SNP, and compared in the two groups using our old favorite, the P-value and via meta-analyses.
The authors at least make a modest claim of parametric (i.e. inside-the-model-but-not-predictive) variance explained, of about 4%. If we added in predictive uncertainty, and accounted for attempts at quantifying the unquantifiable, the real-world predictive value is small indeed.
But not zero. I do not argue zero. Here we must not confuse coarse genetics analyses with the obvious signal of different behaviors in different peoples: all see large noticeable differences, even if these are not always measurable to exact degree. All history and observation is behind this contention. Equality is false. And, of course, genes play a large role in the differences between peoples. My criticism is much more subtle. It is that ascribing behavior (to a stated degree) is tied to this or that gene, or its variant, is badly over-certain. Which does not imply “always false”; though it does imply “sometimes false.”
I will not plow through every detail in the paper, a tedious and off-putting exercise. I will only cover those aspects of their modeling technique which we have discussed before to see how this particular analysis fits into our larger scheme. I’m leaving out much. In particular, I’ll skip crucial details necessary to recreate (and even comprehend all) their work. If you want these, you’ll have to suffer through their paper yourself.
Incidentally, like many papers these days, all the real material is buried in Supplementary material, a frustrating practice. For instance, that’s where the picture of the model below comes from. Be sure to download all of the paper if you’re following along.
Their Model
Notice that SNP can be 0 (the base or most frequent variant) and whether the location has one copy of the different allele or two (on both pairs of chromosomes; e.g. the base is AA, one person has AG, and a third GG). They don’t say it here, but often this value is not measured exactly and is instead an estimate,; i.e. a number with uncertainty between 0 and 2 is put in, an estimate derived from another regression. This uncertainty should be carried forward, but isn’t.
This model also assumes that the effect of the allele is linear in difference, when a better and more realistic model would put beta_1 I(SNP = 1) + beta_2 I(SNP = 2). This is likely to be a small difference in practice, I admit, but who knows, and there is no good reason to do it the wrong way (which all do). Yet they are forced into this wrong way because, as above, SNP is often estimated and not measured.
You see they stick in age squared. When they have it. Not a grand idea by any means, as you know. And then come the “10-20 genetic principle components” (PCs) of other measures slammed into that sum. Ten to twenty PCs. Those measures themselves are strongly inconsistent across the different databases, so these geographic/population models are to a degree apples to oranges. And, as you recall, PCs themselves are the result of yet another regression-like model, which creates linear combinations of those other measures, with the restriction that the linear combinations are orthogonal. Because the databases differ, the PCs cannot be compared. The uncertainty in the PCs is also ignored and not carried forward, resulting in more over-certainty.
This is not real control—you will recall—but statistical “cram it all into the model” control. Which because they cram different things in for different databases, and things like that screwy age-squared, means their beta_1’s, the very point of the study, are not really comparable across their populations.
The outcomes, the Ys of the equations, are the scored personality traits. Which, beyond the base criticism that these are attempts to quantify the unquantifiable, these scores are from all manner of peoples. How well do they compare, so that a given number on a scored trait in one people (say American whites) means exactly the same for that same scored trait in another people (say Congolese blacks)? Not that the peoples would score differently, which I presume they would, but are the scores themselves, like a yardstick would be cross-culturally, directly comparable? I doubt that.
The results are ranked by our old friend, the mighty wee P-value. On those beta_1’s. This is, of course, nuts. You have heard me rail about scientists waving their wee Ps in people’s faces long enough, so I won’t repeat any of that here, except to emphasize one well known but always ignored criticism, a complaint, like confirmation bias, scientists always think happen to the other guy. It is that as sample size increases, the P grows weer. The beta_1 can be as small as one likes, but not equal to 0. It can have negligible effect in Reality. But as the size of the data grows, the P shrinks, shrinks, shrinks! And becomes even more meaningless, as it does here.
Here’s how the results look. Note the log scale (y-axis) for the wee Ps.
The scored personality traits are each represented, and the SNPs/genes with the weest Ps are highlighted. You see the enormous sample sizes, which guarantee wee Ps, as said. See the horizontal red lines? Lines which stick out above them are “significant”. Which means most things are. If we throw a bag of rice on the ground, or rather many bags, we’d get similar signals with many patterns of grains. The genetic correlations are anemic.
So. Are these spikes coincidences? Do they bespeak of actual causation? Inside a gene, which SNP variants are causing the differences? Are we conflating measures of race with behavior? Tell us: just how does the causation work? They sure do believe in these mysterious causes. Here is just one of their many causal statements:
Conscientiousness in particular was genetically correlated with reduced substance use (across substances, mean rg = 0.21), greater sports frequency (rg = 0.35) and greater preference for low-calorie foods (rg = 0.17), as well as the likely downstream consequences of these behaviours…
They conclude “Genetic and environmental factors, along with their correlations and interactions, contribute to variation in personality”, which is again causal language.
Are they entitled to this language? Are we remembering all the sources of uncertainty mentioned? Or are we seduced by scientific results?
Enter Probability
Why not do this instead:
Pr(Y = score | SNP, other new data, old data, assumptions)?
This “integrates” out the nonexistent fictional not-found-in-real-life there-only-because-of-the-model parameters. And it directly gives us the probability a score takes a particular value given our evidence and for whatever variant of SNP we’re considering. Wonderful!
So why don’t people do this?
The “assumptions” are the same as they already make, such as for age-squared, the use of PCs, and all the other stuff in their paper I did not mention. The old data is the old data. That part is easy. So is specifying the SNP variant we’re interested in (it’s only equal to 0, 1 or 2, after all).
The hard part, and the reason probability has not replaced all the weird ad hoc techniques (like P-values) that plague the field, is that that “other new data” have to be specified. The model is built assuming their presence, therefore they have to be given definite values to get a probability. Here that includes all those PCs; specific individual values, I mean, for all PCs “inside” a person, and all the other stuff (the other demographic, biological measures). See the Video for clarification. Which means we get things like this:
Pr(Y = score | SNP, PC1 = p1, PC2=p2, …, PC20 = p20, Demo1 = d1, etc., old data, assumptions).
You get one of these for every combination of all that new data. And that’s just for Y = this score. You get the same batch again for Y = that score. And so on. That’s a lot! The math (at our level) is trivial. There is no problem with implementation, either: code is simple. There are, as you recall, ways to simplify this (see the Video), but probability requires much greater care to interpret. A mort of work is involved. Worst, the brimming positive career-advancing grant-guaranteeing certainty the ad hoc methods give you crumbles and evaporates. And that is intolerable.
Of course, what probability buys you is better accord with Reality, and a proper understanding of (ahem) Uncertainty. It lets you avoid the terrible and Deadly Sin of Reification, too, so it is far better for your soul.
All my work is free and user-supported. Here’s how to help:
- Subscribe at Substack (paid or free!)
- Cash App: $WilliamMBriggs
- Zelle: use email: matt@wmbriggs.com
- Buy me a coffee
- Paypal
- Other credit card subscription or single donations
- Hire me
- Subscribe at YouTube
- PASS POSTS ON TO OTHERS
Discover more from William M. Briggs
Subscribe to get the latest posts sent to your email.
