I know that the GWAS association p-value threshold is 1e-8. This makes sense because in GWAS, you make ~1 million hypothesis (i.e. use that many SNPs in the association test). However, let's say I do not do genome-wide analysis but only use the 100 SNPs from a specific gene to check for association with my phenotype. Then, do you think a raw p-value of 1e-4 (which becomes 1e-2 after multiple hypothesis correction) is a reliable p-value to come to a conclusion that there is an association between that SNP and the checked phenotype? Obviously, if I checked all 1 million SNPs instead, this p-value would be insignificant after correction, but I am not checking 1 million SNPs, as I mentioned.
1 Answer
I know that the GWAS association p-value threshold is 1e-8
This may be a common threshold of statistical significance that is used, but it's definitely not an absolute value. It's a hack to try to work around many issues with GWAS associated with testing millions of SNPs. Unfortunately, the most relevant issue in GWAS (for spurious significance) is unexpected shared genetic structure in the cases or controls, and this cannot be excluded by a p-value threshold, regardless of how low it is set.
In your 100-SNP example, you are correct that the bonferroni correction would be a threshold of 1e-4 (for a desired significance threshold of 0.01), but it's important to be cautious about GWAS results, even when the p-value indicates otherwise. Trust no one, especially yourself. P-values should never be used as an indication of importance (which suggests that most GWAS results should probably be reconsidered), and are better used in conjunction with other evidence for identifying relevant / significant tests.
My recommendation for GWAS (or any similar bulk genotyping study, as in your 100-SNP subset) is that at least a bootstrap sub-sampling process is carried out, to make sure that observations found within the comparison of interest are at least well replicated when comparing smaller sub-groups of the cases and controls. More information on that can be found in the poster I presented at Queenstown Research Week last year, and detail about the implementation in my draft preprint (which will probably never be properly published, given how old the research is).
-
$\begingroup$ Thanks for the reply, those make a lot of sense. A followup question: I was also told by a GWAS expert that when the minor allele frequency is low, one would not trust the p-value even if it is significant. He said so when I showed him the p-value of 1e-4 for a SNP with a 2% minor allele frequency. Do you think it is due to the need for the additional bootstrap analysis you suggested? I guess a low p-value would not be replicated in bootstrap analysis in case of rare variants and small overall sample size. Could you share your suggestions specifically for analyzing rare variants? $\endgroup$– user5054Sep 9, 2017 at 21:31
-
$\begingroup$ Whether or not it will be internally replicable will depend on the sample size of the subsampled population. It would be possible to specifically subsample to increase the frequency of single SNPs, but that might cause problems with associated interaction effects. $\endgroup$– gringer ♦Sep 10, 2017 at 2:44
-
$\begingroup$ Certainly, you need to take more care about rare variants. Most tests will provide both a p-value and an effect size (which is better than the p-value for ranking SNPs). If the effect size is large, then it lends more weight to an association near the SNP locus. It may also be a good idea to do a haplotype association test within the haplotype block(s) that include the SNP, as any linked SNP could be causative. Bear in mind that GWAS are only testing for association, and there are other follow-up studies that should be done to support findings of association. $\endgroup$– gringer ♦Sep 10, 2017 at 2:55