Belated links

6 comments:

  1. The PCA plots look very strange in the 'missing heritability' paper. Not sure if this points to a more significant error in methodology (they use PCA at several points in their analysis). I don't see how the data sets they used could have given the plots they show in the supplementary materials. And it's hard to explain by a cock-up while running the EIGENSOFT software.

    ReplyDelete
  2. Dasein,

    I have not gone through the formality of actually reading the paper, but I'm assuming they used the PCA results to control for population structure by excluding ethnic outliers. I'm further assuming the samples consisted overwhelmingly of Northern European-derived individuals. Incorrectly excluding some Northern Europeans would merely slightly reduce the power of the study. Incorrectly including non-Northern Europeans could potentially introduce error, but I doubt population structure alone could explain their result (assuming the sample is still overwhelmingly N. Euro and inter-group differences in height are small relative to the range of heights represented in the sample).

    ReplyDelete
  3. n/a,

    Yes, that was the purpose of the Africans and Eurasians PCA. The Europeans-only PCA appears to have been used for their variance estimates. That plot (also in the supplementary materials) doesn't look problematic. The problematic plot may just raise concerns that there were some problems with their data processing (which might affect other parts of the analysis as well). I think it's unlikely that there's anything here that will affect the paper's conclusions, like you say, but I can imagine that they will want to doublecheck parts of their analysis (assuming, of course, that the plot is in fact problematic, which I am fairly certain it is, and that it's not a trivial error, which it might well be).

    ReplyDelete
  4. Dasein,

    Looking at supplementary table 1, the inclusion or non-inclusion of European-only PCs as covariates does not appear to make any appreciable difference.

    I would expect any issues with data processing (outside of outright fraud) to generally decrease rather than increase the amount of variance explained. Anyway, you could try emailing the authors. I'm not inclined to be skeptical in this case mostly because the result makes intuitive sense to me, but I'm sure additional evidence one way or the other will be forthcoming before too long.

    ReplyDelete
  5. Those are good points, n/a. I have it on good authority that the plots are in fact in error (though how the correct axis labels ended up on them remains a mystery).

    Like you say, additional evidence should be forthcoming very soon. I don't see why their approach cannot be applied to other existing GWAS data sets to come up with estimates of how much variability can be explained by common SNPs.

    What I also find interesting about this paper is Dienekes' response to it (also described at link in my first comment). His interpretation (more specifially, his rather stubborn, uncritical acceptance of, and even praise for) the PCA plots makes me wonder how expert he really is. I don't know enough about many of the things he comments on to pass judgement, but in this case it seems he's more just presenting results than understanding them. This is also evident in another post in which he refers to the Yang paper:

    This genome-wide study suggests that common variation is not related to neuroticism, and that individual loci have at most a 2% share in its heritability. It is interesting to compare this study with the recent one on height, where a larger number of individuals (~4k) were studied on a smaller number of SNPs (~300k). Are there many SNPs with effects on neuroticism that may have been detected if a larger number of individuals were sampled? I doubt it for two reasons (i) the number of individuals sampled here is not that lower than in the height study, and (ii) neuroticism is a made-up construct that can be imperfectly measured with psychometric testing, while height is a physical reality that can be measured objectively. So, while neuroticism does have a correspondence to brain function, it is much more difficult to discover what it is.

    He doesn't seem to understand the meaning of the Yang paper, and thinks the methodologies in the papers are comparable, when in fact one is assessing the heritability from individual SNPs, and the other considers them as a whole. What's your impression of Dienekes? I imagine you know a lot about the stuff he follows. No doubt if someone blogs a lot on a complex topic, he's likely to get some stuff wrong, but the things mentioned above seem egregious enough, to me at least, to suggest that he may not be the expert many think he is.

    ReplyDelete
  6. "What's your impression of Dienekes?"

    He certainly shouldn't be treated as an authority. He does make mistakes and has his biases. "Expert"? I don't know. He's an intelligent amateur. On some issues (e.g., Y-STR mutation rates and demographic history), he's more right than most of the "professionals". On the other hand, he probably has at most a superficial understanding of some of the other topics he posts on and the example you cite reflects this.

    ReplyDelete