I'm trying to figure out if I should be filtering out GWAS hits that have high standard error and I'm not quite sure what to do. It seems like it might not matter, because the standard error is used to calculate the t-statistic, which is then used to calculate the p-value. So in a way it's already built in. But reporting SNPs that have very high standard error doesn't seem quite right. What's the best way to handle this?
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
3
I am curious about the relationship between your p-value (or effect sizes) and standard error. I would expect the significant signals to have smaller stand error compared to the non-significant, background signals. If this is the case, there is no need to do filter.
-
$\begingroup$ That hasn't been my observation in looking at the raw UK Biobank GWAS results. There may be some correlation between effect size and standard error, but it's not a universal rule. I have seen results with large standard error, and also large effect sizes. $\endgroup$– gringer ♦Commented May 12, 2020 at 14:41
-
$\begingroup$ I agree. But if you look at genome-wide significant SNPs, are their std really large? Do you have rigorous filtering criteria before GWAS? I am also wondering what is your purpose of doing the GWAS? If it is just a standard GWAS or genomic correlation or LDSC I think it is fine without filtering? I do not remember seeing any paper doing that. $\endgroup$ Commented May 12, 2020 at 21:19
-
1$\begingroup$ I'm not doing the GWAS, I'm only looking at other people's results. It's not typical for people to report SNPs with large SE in their tables, because that means the p-value will be quite high (and most people concentrate on p-value above all else). $\endgroup$– gringer ♦Commented May 13, 2020 at 1:08
$\begingroup$
$\endgroup$
I think the effect size is more of an issue than the standard error.
If the standard error suggests that the effect direction could change sign, then it might be a good idea to filter it out. Otherwise, if the effect size remains large (or at least positive) even after accounting for a large error, then [to me] that would attach more weight to the likelihood that the association is valid.
I did a very quick analysis of the results of one UK Biobank study on Twitter, where I used the statistic of "the least extreme value in a 95% confidence interval" in a Manhattan plot instead of the more commonly-used p-value:
