Cuffdiff producing q-values > 1

[ypark28]

hope this help ->

http://brainder.org/2011/09/05/fdr-corrected-fdr-adjusted-p-values/

[ocs]

Thanks for that script, severin.

I have a question concerning the use of the cummin function. In this case it ensures that the q-value is not getting bigger than 1 (thats nice) but also that if once a q-value is reached, the following values can't get bigger but would be set to the lowest value known to this time point (actually this happens in reversed order). This means that some q-values are not the appropriate q-value for the according p-value. And it means also that they are ascending. At first glance this seem not to be legal. So for example if I have q-values (in p-value order):

0.01, 0.02, 0.06, 0.04, 0.07

this procedure would result in

0.01, 0.02, 0.04, 0.04, 0.07

.

Having a cutoff of 0.05, in line 1, the third would not be accepted but the fourth. In line 2 the third is accepted (but due to correction it wouldn't be). This is what is confusing me: What happens to q-values that are again lower as one preceding value? Why is it okay to accept them anyway? I looked at the paper, but its not clear to me.

[shilez]

Thank you severin!
Now I can generate q value based on the cufflinks pvalue output, I think there is a bug in cufflinks when generating q values, possibly that they miss the line qval[ord]=rev(cummin(rev(qval[ord]))), which results in a >1 q value.

[severin]

R script for fdr

Here is an R script for false discovery rate based on that paper.

#FALSE DISCOVERY RATE
bh.fdr=function(p){
#
#This function computes q-values using Benjamini and Hochberg's (1995)
#approach for controlling FDR. Courtesy of Dan Nettleton
#
m = length(p)
k = 1:m
ord = order(p)
p[ord] = (p[ord] * m)/k
qval = p
qval[ord]=rev(cummin(rev(qval[ord])))
return(qval)
}

[ocs]

Hello severin, thanks for the paper. I looked at it, but they always correct the cutoff for the p-value and not the p-value itself. They don't even mention the q-value. So the paper may be useful for the background but I can't see that it explains my inital question.

[severin]

fdr

Benjamini, Y. & Hochberg, Y., 1995. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society. Series B

The equation does the correction and it is up to the user to take appropriate values. ie >0.05

there may be a more resent paper too but this is the original.

this paper should clear it up

[shilez]

I got the same above 1 q value, would also want to know how q value is generated

[ocs]

Hello at all,

I use TopHat and Cufflinks to process some human RNA-seq data with an annotation gft-file for known genes (this means my pipeline is: TopHat -> Cuffdiff).

In the resulting gene_expr.diff file I obtain q-values greater than 1:

> range(expr$q_value)
[1] 0.00000 1.08198

Since I thought that a q-value is a probability (fdr corrected p-value) it should not be greater than 1.

Its clear to me that when the p-value is corrected there can be q-values greater than 1 but I thought its limited to 1 in this cases. I am confused about the values, 1.08 is a very small value assuming the q-value is not a probability anymore (then I would expect higher values) but its to big for a rounding error I think.

Maybe I got there something wrong. Could someone explain it to me?

Thanks in advance,
Oliver