[BioC] Filtering out tags with low counts in DESeq and EgdeR?

[Martin Morgan]

On 05/21/2011 07:07 AM, Wolfgang Huber wrote:
> Hi Xiaohui
>
> to follow up on the filtering question:
>
> - the filter that Xiaohui applied is invalid, it will distort the
> null-distribution of the test statistic and lead to invalid p-values.
> This might explain the discrepancy.
>
> - the filter that Simon suggested is OK and should provide better results.
>
> - I'd also be keen to hear about your experience with this.
>
> A valid filtering criterion does not change the null distribution of the
> subsequently applied test statistic (it can, and in fact should, change
> the alternative distribution(s)). In practice, this means choosing a
> filter criterion that is statistically independent, under the null, from
> the test statistic, and in particular, that it does not use the class
> labels. Details in the below-cited PNAS paper.

Was wondering whether, since the dispersion parameter is estimated from 
the data, in some strict sense the filtering and testing procedures are 
not independent under the null anyway? For the same reason that one 
would not want to use a variance filter before a limma-style analysis, 
if I understand correctly.

Martin

>
> Best wishes
> Wolfgang
>
>
>
>
>
> Il May/21/11 11:02 AM, Simon Anders ha scritto:
>> Hi Xiaohui
>>
>> I agree thatit is worrying to get so different results from your two
>> approaches of using DESeq. Here are a few suggestion how you might
>> investigate this (and I'd be eager to hear about your findings):
>>
>> - Bourgen et al. (PNAS, 2010, 107:9546) have studied how pre-filtering
>> affects the validity and power of a test. They stress that it is
>> important that the filter is blind to the sample labels (actually: even
>> permutation invariant). So what you do here is not statistically sound:
>>
>> > filter=dat[rowSums(dat[,group1]>= 8) | rowSums(dat[,group2]>= 8), ]
>>
>> Try instead something like:
>>
>> filter=dat[rowSums(dat) >= 16, ]
>>
>> - How does your filter affect the variance functions? Do the plots
>> generated by 'scvPlot()' differ between the filtered and the unfiltered
>> data set?
>>
>> - If so, are the hits that you get at expression strength were the
>> variance functions differ? Are they at the low end, i.e., where the
>> filter made changes?
>>
>> - Have you tried what happens if you filter after estimating variance?
>> The raw p values should be the same as without filtering, but the
>> adjusted p values might get better.
>>
>> To be honest, I'm currently a bit at a loss which one is more correct:
>> Filtering before or after variance estimation. Let's hear what other
>> people on the list think.
>>
>>> 2. For EdgeR
>>
>> DESeq and edgeR are sufficiently similar that any correct answer
>> regarding filtering should apply to both.
>>
>>> 2) I got 800 DE genes with p.value<0.1, but got 0 DE genes after
>>> adjusting p.value, is this possible? Then, can I used the *unadjusted*
>>> p.value to get DE genes?
>>> To adjust pvalue, I used: nde.adjust=sum(p.adjust(de.p, method =
>>> "BH")< 0.05)
>>
>> Of course, this is possible. (Read up on the "multiple hypothesis
>> testing problem" if this is unclear to you.) Not also, though, that you
>> used an FDR of .1 in your DESeq code but of .05 here.
>>
>> Simon
>>
>> _______________________________________________
>> Bioconductor mailing list
>> Bioconductor at r-project.org
>> https://stat.ethz.ch/mailman/listinfo/bioconductor
>> Search the archives:
>> http://news.gmane.org/gmane.science.biology.informatics.conductor
>
>


-- 
Computational Biology
Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N. PO Box 19024 Seattle, WA 98109

Location: M1-B861
Telephone: 206 667-2793