[BioC] Handingly outliers with voom [was: tagwise parameters for negative binomial in edgeR]

[Gordon K Smyth]

Dear Ina,

voom is already quite robust as it is (more so than un-robustified edgeR). 
However limma also has special options for either observation or 
dispersion outliers.

In the sequence

   v <- voom(y,design)
   fit <- lmFit(v,design)
   fit <- eBayes(fit)

replacing the second line with

   fit <- lmFit(v,design,method="robust")

is intended to robustify against observation outliers while (as Ryan has 
said) modifying the third line to

   fit <- eBayes(fit,robust=TRUE)

handles dispersion outliers.

In practice, my lab uses robustified eBayes() a lot but robustified 
lmFit() very little.  lmFit() has included the "robust" option since limma 
was first posted to Bioconductor in 2002, but it has been found to be 
seldom needed and so tends to be forgotten.

Best wishes
Gordon

> Date: Fri, 21 Mar 2014 10:27:18 -0700
> From: "Ryan C. Thompson" <rct at thompsonclan.org>
> To: Ina Hoeschele <inah at vbi.vt.edu>
> Cc: Bioconductor mailing list <bioconductor at r-project.org>
> Subject: Re: [BioC] tagwise parameters for negative binomial
> 	distribution in edgeR
>
> Hi Ina,
>
> I don't think voom has any special consideration for observation
> outliers, but limma's 'eBayes' function has a 'robust' argument which I
> believe has the same effect as the corresponding argument in edgeR's
> 'estimateDisp', i.e. dealing with outlier tags that have abnormally
> high (or low) variance.
>
> -Ryan
>
> On Fri 21 Mar 2014 08:48:23 AM PDT, Ina Hoeschele wrote:
>> Hi Mark,
>>    how would the presence of observation outliers potentially causing dispersion outliers be handled in voom?
>> Many thanks, Ina

______________________________________________________________________
The information in this email is confidential and intend...{{dropped:4}}