I am using fisher exact test to call for somatic mutation. Now, if I go for any multiple correction, it appears to me as overdone. It is removing almost everything. With, 0.01 cut off on and average, it is giving me ~100 somatic mutation per exome. Besides, if I choose the cutoff as 0.05, the number shoots up to 400. I would like your suggestions on that, and if possible, some references for citation.
Nobody will be able to seriousely answer this question. It depends on your aims. What will you control (test-wise error rate; family-wise error-rate; false discovery-rate; something else)? Why do you want to control it? What are the costs of false decisions, what are the benefits of correct decisions? Can you quantify this? Can you make a balance?
If you did a screening to continue the work on some candidates, then simply ask yourself how many candidates you can further investigate (say: k) and then take the top-k and proceed. Would be nonsense to control any kind of error rate in this case.
If you want to show an overall-"result" in a publication, show the complete distribution of the p-values (or of the odds-ratios or some other estimate). No need to make some arbitrary cut-off for this purpose.
I hope I need to have some cut off to name some variant Somatically changed. Probably, I want to see frequently mutated genes( somatically). And to define a gene as somatic, I need to define some cut off. P-value distribution or OR is fine. When someone should name an allele frequency change as somatic. Some cut off need to be defined for sure. In my case, I thought to define it through Fisher's Exact Test.
Fisher's exact test is first of all just a procedure. The actual decision is your part.(*) It must be based on your aims. Statistics does not tell you what a reasonable basis for your decisions will be. This depends, as I said, on the cost-value ratio, and this has to be specified by you (what is a very difficult if not impossible task). Therefore, most (published) decisions lack any rational basis.
Controling the FDR might be the most appropriate in your case. Reviewers typically accept FDRs up to 20%, but it may be as low as 1%. However, I would still put the cart before the horse and define the number of candidates I would like to consider, and then give the FDR for this list.
There is no possibility to "objectively" distinguish "somatic" or not by any statistic or cut-off or whatsoever. The only thing one can do is to just take a decision, depending on the knowledge from the data (ORs and/or p-values) and all other available and relevant boundary conditions, so that other experts would agree or accept your decision as beeing reasonably sensible in the given context. You wrote: " But it would be meaning less, to name any frequency change as somatic. Some cut off need to be defined for sure. " - most of the published results are meanigless, since they are based on the meaningless decision rule: "results are considered significant when p
- Badges
- Science method