Hello everyone,
I calculated some pairwise FST estimates with hierfstat (pairwise.fstWC function) in R on my SNP dataset and then calculated confidence intervals (not p-values) on those estimates with the same package (function boot.ppfst). Is there a need to correct my CI for multiple testing?
As far as I know, I am estimating parameters and not testing hypotheses (p-value) so I wouldnt need to correct for false positives. Am I correct? Can anyone explain this to me?
Thanks,
Giulia
You are correct in your understanding. When you calculate pairwise FST estimates and derive confidence intervals (CIs) using the bootstrapping technique, you estimate parameters rather than conduct hypothesis tests. Therefore, there is no need to correct the CIs for multiple testing.
Multiple testing correction is typically applied when conducting hypothesis tests that involve multiple comparisons (e.g., testing for significant differences between groups). The purpose of correction methods, such as the Bonferroni correction or false discovery rate (FDR) control, is to control the overall false positive rate when testing multiple hypotheses simultaneously.
In your case, you are estimating parameters (pairwise FST values) and constructing confidence intervals around these estimates to quantify the uncertainty. The CIs provide a range of plausible values for the parameters, and they reflect the precision of your estimates based on the bootstrapping procedure. These intervals indicate how reliable your estimates are and allow you to draw appropriate conclusions from your data.
In summary, since you are not conducting hypothesis tests and focusing on parameter estimation instead, there is no need to correct the confidence intervals for multiple testing.
A CI is (very typically) just the interval of hypotheses that cannot be rejected by the test. So using and interpreting CIs actually is a form of testing. As soon as you try to interpret your estimate in relation to some fixed hypothesis (e.g. zero, to conclude that the unknown parameter is either positive or negative, e.g. if the SNP is over- or under-represented) is a test, no matter if you know the confidence/significance or not.
On omics data, a singe test (a single hypothesis) may not be of great interest. Instead, a more global pattern may be more instructive. This global patter can be composed of estimates. But then, the pattern should be judged, not the individual estimates (and their CIs).
If, however, you are interested in individual SNPs, directions may be off and you should have some kind of safety-belt to keep you form interpreting meaningless estimates. You can do this by tests - corrected for multiple testing - or with CIs - also corrected for multiple testing.
Ma'Mon Abu Hammad and Samuel Oluwaseun Adeyemo thank you very much for your help and explanation. It is what I thought. I don't necessarily understand what you mean, I need to use my whole SNP dataset and not individual SNPs Jochen Wilhelm
Ma'Mon Abu Hammad wrote:
In summary, since you are not conducting hypothesis tests and focusing on parameter estimation instead, there is no need to correct the confidence intervals for multiple testing.
The many authors who have worked on methods for computing simultaneous CIs for multinomial proportions (see the reference lists on the sites shown below) clearly take a different view! And before anyone jumps to the wrong conclusion, I am not suggesting that Giulia Trauzzi's CIs are for proportions--I have no idea what FST estimates are, in fact. I am just using multinomial proportions as another example where one might wish to make some kind of adjustment to multiple CIs.
- https://blogs.sas.com/content/iml/2017/02/15/confidence-intervals-multinomial-proportions.html
- https://rdrr.io/cran/DescTools/man/MultinomCI.html
HTH.
- Badges
- Science topic
- Similar topics
- Correction