I have a set of genes from particular species. Metagenome samples were mapped to this database of genes' sequences. As a result I had a presence/absence matrix for all the possible genes and species, which looked like this:

         sample_1 | sample_2 | ... | sample_n

gene_1       1         0                1          
gene_2       1         1                0
gene_m       0         0                0

0 stands for absence of gene, 1 for presence

Some of the samples were taken from patients with particular disease, some from healthy people. So they are divided into two groups: case and control For each gene I hypothesize that it has same proportion of presence/absence for both case and control groups. Alternative is that number of genes which are present and absent is unevenly distributed. I assume I can generate contingency tables for each gene:

               Case   |   Control
Gene_i present    a   |   b
Gene_i absent     c   |   d

I thought I can use FET or Chi2 test depending on number of observations in cells. If minimum is less than 10, I use FET. Otherwise Chi2.

I have doubts whether assumption about marginal totals for FET is met, because there's no fixed number of times gene is present.

If minimum value in a table is more than 10, Chi2 test is performed, otherwise Fisher exact test is performed. Is it appropriate to use one procedure for p-values adjustment if they are obtained in different tests or p-values from these 2 tests should be treated separately?

  • $\begingroup$ I added more information about what I'm trying to do. It has nothing to do with GWAS $\endgroup$
    – stranger
    Aug 22 '17 at 8:51
  • $\begingroup$ I've updated the answer, but it's essentially the same. Think about point 2. $\endgroup$
    – NULL
    Oct 19 '17 at 22:54

Given the information you provided here are some factors you'd need to think about:

  • What are the features and assumptions of Fisher's test? Are you fine with those? Take a look at this, this, and this
  • Do you know that if you use Fisher's test for large number it most often get's approximated using a test like Chi2 without people knowing it?

So, given the last item in the list, the answer to your question is rather clear.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.