Dear Sergios,
A false discovery rate is simply the ratio of false positive results to all significant results, i.e. the proportion of incorrect significant results you can live with. Unfortunately, most FDR procedures are not exactly applicable to pairwise relative rate tests. There are two reasons: RR tests are not necessarily independent and identically distributed, and FDR procedures assume that a certain proportion of the observation actually came from the null model (i.e. the rates in two lineages were the same) - this may or may not be the case.
That said, I would recommend (for its ease of use), the Benjamini-Hochberg procedure, which is very simple to apply. Say you have N p-values, ranked from smallest to largest, p_1 <= p_2 <= ... <= p_N. Under fairly general conditions, you can ensure that the false discovery rate is no greater than 'q' (e.g 0.05), by rejecting the first i hypotheses, for which p_i <= (i/n) q. For example, for 10 hypotheses, with ranked p-values 0.001, 0.005, 0.01, 0.02, 0.05, 0.1, 0.2, 0.3, 0.5, 0.7, to enforce FDR <= 0.05, you would reject the first four hypotheses, because 0.001 < 1/10 * 0.05, 0.005 < 2/10 * 0.05, 0.01 < 3/10 * 0.05, 0.02 < 4/10 * 0.05, but 0.05 > 5/10 * 0.05.
The code in HyPhy to implement BH could be like this (assuming all the p-values are stored in a column vector PV with n rows)
Code:PV = PV % 0; /* sort */
q = 0.05;
for (k=0; k<Rows(PV); k=k+1)
{
if (PV[k] <= (k+1)/Rows(PV) * q)
{
fprintf (stdout, "Reject hypothesis ranked ", k+1, " with p = ", PV[k], "\n");
}
else
{
break;
}
}
The paper on the method is here Multimedia File Viewing and Clickable Links are available for Registered Members only!! You need to
.
For a much more technical pFDR paper take a look at Multimedia File Viewing and Clickable Links are available for Registered Members only!! You need to
HTH,
Sergei