[BioC] Error when running limma on low variance data

[Gordon K Smyth]

> Date: Tue, 8 Sep 2009 12:28:28 -0400
> From: qq bioconductor <qq.bioconductor at gmail.com>
> Subject: [BioC] Error when running limma on low variance data
> To: bioconductor at stat.math.ethz.ch
> Content-Type: text/plain
>
> Dear list member,
>
>  I got an error message when running limma on my array dataset, which
> contains a total of four Affy arrays from two different conditions.
>
> Error in if (evar > 0) { : missing value where TRUE/FALSE needed
> In addition: Warning message:
> In fitFDist(var, df1 = df) :
>  More than half of residual variances are exactly zero: eBayes unreliable
>
> I found the variances of most probesets in my dataset is 0. Therefore
> median(fit$sigma^2) is 0.
> The error is because in my case the "evar" variable in function fitFDist is
> NaN, which is because some elements in the "e" variable are equal to -Inf
> and "emean" is -Inf.
> (    emean <- mean(e)
>    evar <- mean(n/(n - 1) * (e - emean)^2 - trigamma(df1/2))
> )
>
>
> Does anyone know how I can fix the error? Thank you!

Yes. Stop trying to analyse data when all the residual variances are zero. 
It makes no sense to attempt a statistical analysis of such data, and it 
is a sure sign of an underlying problem with the data.  At very least, you 
might filter out the probes with zero variances.

Best wishes
Gordon

> Sincerely,
> Qianqian