They worked out, but we revised the method to better test the population structure.

rs17522122 has a negative loading (see Table 3). The strange thing is that the GWAS seems to find the right SNPs, but sometimes gets the direction wrong. I don't understand it. Still, this leaves 5 out of 6 with positive loadings, which is ok, but not totally convincing.

The preliminary paper is here: https://docs.google.com/document/d/1Sb68X4TImRK5_QhUkUftokBpKrXyO8H5cT2Ou8H75xk/edit

Piffer figured out how to use the package that can calculate Fst's based on entire chromosomes (we use Chr21 because it is the shortest). Then he calculated all the Fst differences and the SNP GI factor differences. If population structure is responsible for the correlations, then in multiple regression, the Fst's should have a strong beta, while the SNP GI factor should not. However, we found the opposite:

> cor(newdata)
IQdistances fst_21 gfdistances
IQdistances 1.0000000 0.5885368 0.7764679
fst_21 0.5885368 1.0000000 0.7855911
gfdistances 0.7764679 0.7855911 1.0000000
> fit <- lm(IQdistances ~ fst_21 + gfdistances, data=newdata)
> lm.beta(fit)
fst_21 gfdistances
-0.05602615 0.82048156

The above is with the 4 replicated SNPs.

With 6 the betas are:
Fst gdist
0.005 0.673

So a bit worse, as expected if rs17522122 is a false positive.

With the ADHD SNPs:

Fst adhddist
0.168 0.484

Lots of more population structure here. The ADHD SNPs are not quite as low p value as the others, so there are probably a number of false positives in the hits.

---

All in all, I'm not convinced yet that this genomic evidence irrefutable, but the initial results mostly point towards the genetic model in my opinion.

Most of the 'new' SNPs were in a close region with those already found or reported in the same study. Thus, they are not independent signals, their correlations are |.99 to 1.00|. We used only those not in strong linkage disequilibrium.

What's the p values of those obesity SNPs? They may be mostly noise (like the ADHD ones are I think).

Obesity is special. Within western countries, the correlation between obesity and GI is negative. But between countries, the correlation is positive. See: http://openpsych.net/ODP/2014/09/the-international-general-socioeconomic-factor-factor-analyzing-international-rankings/ For this reason I would not try obesity SNPs using the Piffer methods between countries. One might argue still within countries (e.g. USA). If your results are real and the SNPs are true hits, then we have a problem.

Well, try the Piffer method on the top 13 SNPs using data from 1000 genomes. The mean loading should be high and loadings should be all positive if there is/was strong selection and they are true hits.

There is a way for a SNP to load negatively while still being a true hit and the selection model being correct. This is if it only has a positive effect together with some other gene only found in lower scoring populations (epistasis), or alternatively has a negative effect due to some other gene in the higher scoring populations.

Or if a new mutation which has a positive effect arose in a low scoring population, this would happen too. The latter is detectable in that the frequencies in the other populations should be 0 if there has not been gene flow for this SNP recently.

Can you also run the usual analyses on the new hit? I'm thinking of these:
- New SNP factor x national IQ.
- New SNP factor x national IQ where population structure is removed from the SNP factor (semi-partial correlation) or used in multiple regression together.
- The mean factor loading of the new SNP factor compared with the mean factor loading of random SNP factors.