1
$\begingroup$

I am not a statistician so apologies if my terminology is wrong

I have a dataset of >14 million p values derived from Fisher's exact testing on genome scale sequencing data.

Benjamini-Hochberg correction of these p-values turns out only ~1500 significant p-values so I have been attempting to use qvalue R package to relax the FDR and therefore gain a greater number of significant p-values.

Applying the qvalue package with default options returns a $\pi_0$ of 1

qobj<-qvalue(p.fish.unadj)
qsummary(qobj)
pi0:    1

I presume this is a "good thing" and means that:

  1. I have good power in my dataset to detect true null results and therefore true alternative results also
  2. $\pi_0=1$ is an approximation because my true alternatives i.e. significant tests is such a small number compared to the overall size of the dataset

Am I right?

$\endgroup$
  • $\begingroup$ "...so I have been attempting to use qvalue R package to relax the FDR and therefore gain a greater number of significant p-values." Danger, Will Robinson! What is the point of gaining a greater number of significant $p$-values here? $\endgroup$ – cardinal Feb 6 '13 at 12:53
  • $\begingroup$ @cardinal - let me revise my language. The Fisher test is used in my work to assess significance of a difference in a score relating to individual bases in a genome. If groups of certain scores occur together this may be biologically significant. Although this could be looked at using a Stouffer-Liptak-Kechris correction I wondered if 'relaxation' of the FDR could be used to inform about any spatial correlation of significant p-values not necessarily significant at original FDR $\endgroup$ – plumb_r Feb 6 '13 at 13:12
1
$\begingroup$

It is actually $\pi_{0}(\lambda)=1$ (i.e. a function of $\lambda$ which controls the fraction of p-values used for the null-distribution). Recall that when the null-hypothesis is true, p-values will have a uniform distribution.

In other words, $\pi_{0}(0)=1$ "...is usually going to be much too conservative in genomewide data sets, where a sizable proportion of features are expected to be truly alternative. However, as we set $\lambda$ closer to 1, the variance of $\pi_{0}$ increases..."[1]

I suggest you re-run your qvalue() estimation with a fixed lambda of 0.5 to indeed be confident your analysis.

[1]Storey, John D., and Robert Tibshirani. "Statistical significance for genomewide studies." Proceedings of the National Academy of Sciences 100.16 (2003): 9440-9445.

$\endgroup$

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.