2
$\begingroup$

This a conceptual question.

I ran 10 tests and sorted the p-values in an increasing manner.

I calculate the FDR as p-value * (10 / rank)

Then, in this case, the larger the p-value, the lower the FDR:

p-value rank    FDR
0.16    1   1.60
0.17    2   0.85
0.18    3   0.60
0.19    4   0.48
0.2 5   0.40
0.21    6   0.35
0.22    7   0.31
0.23    8   0.29
0.24    9   0.27
0.25    10  0.25

If I set a significance threshold of p-value=<0.25, and an FDR threshold of =< 0.25, should I consider that the test with a p-value = 0.25 has only a probability of 25% of being a false positive, while the other tests, although having lower p-values, have higher probabilities of being false positives, just because they are higher in the ranking?

How should this be interpreted?

Thanks!

$\endgroup$
2
$\begingroup$

It is common practice to keep original p-values and p-values resulting from FDR (or any other correction) in the same order (possibly with ties).

So, if for original p-values, you have $p_1 < p_2 $, and after FDR correction, you get $FDR(p_1) > FDR(p_2)$, you can simply take $FDR(p_1) = FDR(p_2) = min\{FDR(p_1) , FDR(p_2)\}$.

This is what, for example, R function p.adjust does:

p.adjust((16:25)/100, method='fdr')
0.25 0.25 0.25 0.25 0.25 0.25 0.25 0.25 0.25 0.25
$\endgroup$
2
  • $\begingroup$ Thanks for your answer. Would you consider equally significant the p-value=0.16 and the p-value=0.25, given that both of them have the same FDR? $\endgroup$
    – Miguel
    May 14 '20 at 10:46
  • $\begingroup$ Yes, that would be mine interpretation of this result. $\endgroup$ May 14 '20 at 10:49

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.