RSS
As you know, this blog moves at internet speed, not the glacial creep imposed by academic publishers, with their lucrative frustration of intellectual discourse. No sooner do I write about the work of Piffer, and his use of the new findings of Sniekers et al. (2017) than a senior author on that latter paper, Prof Danielle Posthuma, writes to me about Piffer’s calculations.
Piffer uses genetic variants that explain differences between individuals in a European population, to explain differences between European and non-European populations. Our results are not informative in explaining between population differences. (see e.g. http://www.cell.com/ajhg/fulltext/S0002-9297(17)30107-6).
There are basically three issues:
1 Causal alleles need not be the same in European and non-European populations. We do not know that now. Piffer implicitly assumes they are, but there is no evidence that the alleles we found can also explain part of the genetic variance in non-European populations.
2 Linkage Disequilibrium structure and allele frequencies are not necessarily the same across different populations, and therefore beta’s estimated in one population cannot be transferred to another population.
3 The only thing Piffer shows is that alleles that have been associated with educational attainment or IQ in an European population have different frequencies in other populations. This is true for a lot of alleles, irrespective of whether they are associated with Educational Attainment/IQ and is due to historical/geographical divergence, and thus does not have any relevance in explaining between group phenotypic differences. If our discovery sample were to be non-European we would probably also predict the European population to have a lower intelligence, simply because the alleles that explain differences with the non-European population may have a lower frequency in the European population, due to historical reasons. Thus, that prediction is meaningless for explaining between group differences.
Also, if we would apply the same to GWAS results for height, we would predict Africans are much shorter than Europeans, which is not the case. This is a very deep-rooted misunderstanding and misuse of GWAS results, and especially unfortunate for intelligence and the racial issues in that context.
Let me give you my lay responses.
Of course, the alleles need not be causal in the same way in Europeans and non-Europeans. Several recent papers have suggested that they are different. I find that rather odd, but there could be many ways to wire a brain.
The linkage disequilibrium structures do not have to be the same, and are probably not the same.
You say that Piffer’s approach shows sample of discovery bias. However, he overcame my scepticism by having a random SNP control condition. Does that not work?
I have not looked at how Piffer created the random SNP set. However, I do not think the question is whether the method is good. The question is whether the outcome makes sense (and can be interpreted in the way he does), and that is where I think he goes in the wrong, as the within group (within EUR) variance is not informative for explaining between group differences.
That is as far as our conversation went at this stage, and I hope this helps in further comment and discussion.
Thanks for posting! The link is broken (trailing “).” is the problem). Here is another
http://www.cell.com/ajhg/fulltext/S0002-9297(17)30107-6
The response feels like FUD to me. The height GWAS paper is certainly reason for skepticism (and, especially compared to the recent informal writings of Turkheimer, both it and Prof. Posthuma’s response at least have the virtue of looking both reasonable and correct). But I don’t think it is sufficient to dismiss Piffer’s results out of hand based on that. The remarkable thing about Piffer’s results IMHO is that the correlation is that high. I think any “rebuttal” of his work also has to have a credible explanation for that finding (AFAICT it was not due to data dredging which would be the obvious suspect). Whether or not his polygenic score is a good predictor is a separate question (as fun as it may be to speculate).
As I asked earlier, has anyone tried Piffer’s method with height? I think that would cast some light on this criticism.
Does the data used to generate the box plots in your earlier post also have educational attainment data for individuals? If so it would be interesting to look at how well the polygenic score predicts EA for both groups and individuals. And in particular how good the predictions are for the individual populations (e.g. Africans vs. Europeans). As I noted before, I am especially interested in the Yoruba given the right skew seen in the polygenic score distribution.
/jthompson/comments-on-piffer-from-prof-posthuma/#comment-1898059
Thanks. Piffer will very probably be replying.
While I’m coming at this question from a place of considerable ignorance regarding Piffer’s work, it does seem to me that even on a very crude level there’s a good case for his overall claim regarding the connection between the genes found for, say, IQ, and those present in various races.
The genes so found appear to rank 4 races, East Asian, White, American, and SubSaharan African in the correct order determined by IQ testing.
But what is probability of getting the correct order by chance?
1/(4!), or 1/24, or p=.042
Even at this crude level, the result is already statistically significant.
As the approach gets more refined, the statistical significance should only get better from there.
After thinking about this more I wanted to run some ideas by people here and get feedback.
I think Piffer’s methodology is best viewed as capturing an indirect measure of strength of selection for IQ on a given population (i.e. not as a direct predictor of IQ). As long as the SNPs involved are causal for IQ in each population (which may not be true for all due to linkage disequilibrium, but to me seems more likely than not, has anyone quantified how likely correctly detecting a causal SNP is?) then it does not matter if there are additional SNPs in other populations (cf. Pygmy height SNPs). The selection measure by the polygenic score will still capture the overall selection pressure which will affect those additional SNPs as well.
This implies that the polygenic measure will work for individuals to the degree that it
1. Identifies the selection pressure on their ancestral population(s).
2. Captures some small percent of variance for the SNPs themselves.
There will be noise here (say someone gets a good draw for the polygenic SNPs, but bad for others relevant), but it seems to me the polygenic score might be a better predictor than we would expect by naively looking at % variance explained by the SNPs. I think this will be especially true if significantly different subpopulations are present.
That brings us to the question of why the differing selection pressure? There has been much discussion about the selection pressure for IQ (e.g. cold survival hypothesis), but I have not seen much for selection pressure against IQ. Given that the +IQ alleles have not gone to fixation I think it reasonable to posit that many are subject to countervailing pressure.
The obvious candidates would be things like disease resistance, metabolic cost, head size, etc. This suggest some possible causes for differing selection on IQ based on both environment (e.g. disease frequency, food availability) or the individuals/groups themselves (e.g. birth canal size, which raises the question of what drives that).
Thoughts?
P.S. It is worth reiterating a comment I made in another thread. A SNP being causal for IQ in a population is not the same as being detectable by a GWAS in that population. Sufficiently large or small frequency may render a causal SNP difficult to detect.
Danielle Posthuma sounds like a sensible and sober (did not drink Ulster cool-aid) person:
The only thing Piffer shows is that alleles that have been associated with educational attainment or IQ in an European population have different frequencies in other populations. This is true for a lot of alleles, irrespective of whether they are associated with Educational Attainment/IQ
Thus, that prediction is meaningless for explaining between group differences.
I have not looked at how Piffer created the random SNP set. – Actually he got exceptionally high correlation r=0.74 for his random set and he did try that many random sets.
However, I do not think the question is whether the method is good. The question is whether the outcome makes sense (and can be interpreted in the way he does), and that is where I think he goes in the wrong – He badly wanted to get this result so he got it.
Random PS 0.246 PS 161 GWAS 0.804
Random PS 0.148
I can’t say that I am surprised but Posthuma’s response is disingenuous. She says:
No, that’s not the only thing he demonstrates! He also demonstrates that those different frequencies observed between populations correlate very highly with the known IQ values. Yes, the putative IQ-linked genes don’t have to be the identical in different populations. But Piffer’s results suggest very strongly that many of them are similarly IQ-linked in different populations! Why? Because from the statistical point of view, if they are not IQ-linked in different populations, the probability is very low that the polygenic scores will be correlated so highly to IQ. The probability further becomes something pretty close to zero when one gets essentially identical population rankings using two very different sets – particularly when the sets contain as many as 20-50 alleles.
Unless Piffer fabricated results, he has indeed demonstrated that populational IQ is largely determined by genetic loci. And yes, it also means that as more alleles are added, the finer discrimination would be possible.
I am not seeing it. Looking at the preprint’s Table 2. Beta values for multiple regressions between alleles and IQ values:
PS 9 GWAS hits 0.826
Random PS 0.246
PS 161 GWAS 0.804
Random PS 0.148
page 48
Thanks for starting (and publishing) this conversation.
Random PS 0.246 PS 161 GWAS 0.804
Random PS 0.148
I am not seeing it.
page 48
page 48
Jeez, you clearly didn’t even bother reading. This is a result of autocorrelation (which is entirely expected). The author straight up says so and then deals with it by doing multiple regressions. For which I quoted the results above: huge difference bewteen IQ-linked and random alleles.
Thank you for correcting me. I need to be more careful.
A new paper by Piffer has just been published: https://www.preprints.org/manuscript/201706.0039/v1
I was wondering if this paper (or the papers citing it in Pubmed) might be relevant: Patterns of linkage disequilibrium in different populations: implications and opportunities for lipid-associated loci identified from genome-wide association studies
P.S. A typo on page 10: "Skiekers"
sensible
Thanks! I noticed the discussion of linkage disequilibrium, but did not fully follow it. Any thoughts on how the speed of selection for these SNPs compares to the “typical”?
I was wondering if this paper (or the papers citing it in Pubmed) might be relevant: Patterns of linkage disequilibrium in different populations: implications and opportunities for lipid-associated loci identified from genome-wide association studies
P.S. A typo on page 10: “Skiekers”
I was wondering if this paper (or the papers citing it in Pubmed) might be relevant: Patterns of linkage disequilibrium in different populations: implications and opportunities for lipid-associated loci identified from genome-wide association studies
P.S. A typo on page 10: "Skiekers"
Thanks! I am working on a paper that will directly tackle the LD decay phenomenon.