[BioC] Limma and fitFDist()

[Gordon Smyth]

Dear Gerhard,

There have been only two changes to the empirical Bayes calculations 
in limma since the early days. These have been in limma versions 
2.4.0 and 2.4.13.

The change in 2.4.0 was introduced to better handle the possibility 
of residual standard deviations being exactly zero. The residual 
standard deviations were offset slightly away from zero. This in turn 
was prompted by the fact that rma() can produce a subset of 
probe-sets with identifical expression values across samples when the 
number of samples is small (three or four), as an artifact of the 
summarisation method. You can read the discussion from the 
Bioconductor mailing list at that time.

Later I became worried that the above change would make limma 
slightly more conservative and would change people's historical 
results. So in 2.4.13 I wound it back a bit so that limma would give 
the same results as it did historically for most data sets, while 
still giving some protection against zero residual variances.

The fact that you are seeing different results between 2.4.11 and 
2.4.13 suggests that you have a data set with some very small 
standard deviations. The more conservative results from 2.4.11 
(slightly larger standard deviations but more smoothing) are probably 
more reliable for your data. But even better would be to indentify 
why you have such extremely small standard deviations and seek to avoid them.

Best wishes
Gordon

At 04:30 AM 3/05/2006, Gerhard Thallinger wrote:
>Dear Gordon,
>
>   I am using limma regularly to analyze our microarray experiments
>and would like to thank you for providing such an invaluable tool.
>
>A while ago I analyzed an experiment with limma 2.4.11. It consists of
>20 single channel hybridizations of 10 samples (from 2 groups) before
>and after treatment. Limma identified 45 (out of 15000) DE genes
>(with a p-value < 0.01) between the 2 sample groups.
>
>Recently I reanalyzed the experiment after upgrading to limma 2.4.13.
>Using the same script the number of DE genes dropped to 15 only.
>
>Being puzzeled, I checked the changeLog() which mentions a change in
>fitFDist(). Replacing fitFDist() with the version from 2.4.11 brings
>back the 45 DE genes from the first analysis.
>
>Now I am wondering which of the results is more "reliable" and whether
>there is a way to compensate for the change in fitFDist() (like
>increasing the p-value threshold, ...).
>
>Any hints are highly appreciated
>
>Gerhard
>
>P.S.: I had the same results with R 2.3.0 and limma 2.6.0 (aka 2.4.15 ??).
>
>------------------------------------------------------------------------
>DI Gerhard Thallinger              E-mail:  Gerhard.Thallinger at tugraz.at
>Institute for Genomics and Bioinformatics   Web: http://genome.tugraz.at
>Graz University of Technology               Tel:        +43 316 873 5343
>Petersgasse 14/V                            Fax:        +43 316 873 5340
>8010 Graz, Austria                 Map: http://genome.tugraz.at/Loc.html