Closed joshscurll closed 1 year ago
Never mind -- I read the methods in the GSEA PNAS paper and realized that the FDR is not computed from the nominal p-values but is rather estimated directly when computing enrichment scores. I assume the same is true in GSEApy.
Yes, FDR q are not from nominal p-values
I'm wondering how GSEApy computes FDR q-val and FWER p-val for the gsea module. I have found that FDR q-values do not preserve the ordering of p-values. Also, the FDR q-values are larger than the FWER p-values much more often than I would expect considering that controlling the FWER is much more stringent than controlling the FDR. The observation that FDR changes the p-value ordering is immediately apparent from the dataframe head in code cell 40 in Section 2.4.1 of the GSEApy documentation ("GSEApy Example"). Calculating FDR q-values and FWER p-values directly from "NOM p-val" myself using statsmodels.api.multipletests does not present these issues.
I am using GSEApy v1.0.4.