Question: Multiple Testing Correction
0
4.2 years ago by
ea140220
United States
ea140220 wrote:

Hi,

I am running couple of different differential expression analysis on my data. I have a quick question that confuses me.

Suppose I am using a lfc threshold of x, when correcting for multiple should I use only the genes that have a lfc greater than my lfc threshold or should I use all the genes regardless of the lfc change? How is this handled in the ebayes with limma?

Thanks

rnaseq limma moderated t-test • 821 views
modified 4.2 years ago • written 4.2 years ago by ea140220
3
4.2 years ago by
United States
James W. MacDonald51k wrote:

If you are using a fold change criterion, you should not use multiple corrections simultaneously (well, you can if you want, but using a post hoc adjustment along with multiple comparisons makes the interpretation of the p-values or FDR impossible). If you want to incorporate a fold change, see ?treat.

Hi James,

thanks for the reply for the limma analysis I am using the treat function. However, I have my own method of calculation p-values for the significance of each comparison. However, considering my data I would like to say a change is statistifically significant whether the fold change is above some threshold. What I was confused was whether to prefilter the data using a fc threshold before multiple correction or do multiple correction first then call the a change statistifically significant if the adj.p.value<0.01 and fc<\theta. I hope i can express the question clearly

If you're not using the p-values from the treat function, then how are you using the treat function? My understanding is that the p-values are the main output from treat.

2

Ryan's quite right. The value of the chosen lfc threshold in treat is already incorporated into the p-value that is reported by the function. More specifically, the p-value represents the evidence that the observed log-fold change is significantly different from the specified lfc. There's no need for filtering on the log-fold change before or after the multiple testing correction; just use the reported FDR (i.e., adjusted p-values) directly from topTreat.