[BioC] Understanding limma, fdr and topTable
Kevin R. Coombes
krcoombes at mdacc.tmc.edu
Wed Jul 9 14:37:22 CEST 2008
James MacDonald wrote:
> aaron.j.mackey at gsk.com wrote:
>>> I would add that removing those genes that are unchanged in any
>>> sample will also help reduce the multiplicity problem. Regardless of
>>> the expression level, those genes that never change expression are
>>> uninteresting by default, so e.g., if beta-actin is highly expressed
>>> at the same level in all samples we don't really care to test for
>>> differential expression for that gene since it apparently is not
>>> differentially expressed.
>> This doesn't make sense. How can I choose to filter out "unchanged"
>> probesets without fitting a model of some sort, and making a
>> probabilistic decision for each probeset about whether it is
>> "unchanged" or not. Every probeset (save those below the detection
>> limit) will exhibit variance (though the variance may be below the
>> precision of the instrument to measure), right? You're not suggesting
>> that there are some probesets with zero variance?
> I don't really understand your point here. First, I never suggested
> fitting a model of any kind to select unchanged probesets, unless
> computing the variance is some kind of newfangled model fitting that I
> don't understand.
When you compute the variance and decide to eliminate probes from
consideration if the variance is below some value, you are performing a
statistical test. Implicitly, you are assuming a vague sort of model
that suggests that "if the variance is small enough, then the gene
cannot be differentially expressed". This does not mean that this
particular statistical test is either efficient or powerful. But it is,
nevertheless, a test of differential expression, and so should not
really be ignored when accounting for multiple testing.
> In addition, are you really claiming that a probeset that is 'below the
> detection limit' (whatever that means) will _not_ have any variance? I
> would say that doesn't make any sense. All expression values will
> exhibit some level of variance regardless of whether you might think
> they are 'below the detection limit'.
>> It seems to me that this approach leads to a false/erroneous reduction
>> in the multiplicity problem, as you've just moved the hypothesis
>> testing into a separate "phase" of the analysis. And it also would
>> mess up pooled variance estimates such as those used in eBayes-based
>> methods (e.g. limma).
> So yes, if I had actually advocated fitting a model you would be
> correct. However, simply deciding to exclude probesets that have a low
> variance will not affect the hypothesis testing. Although it could have
> an effect on the computation of the pooled variance estimates if you
> remove too many probesets as the pooled variance might increase.
> But the same can be said for any filtering method. If you remove a lot
> of probesets of low intensity (say all those with an absent call) then
> you very well could be removing probesets with a higher variance and
> then mess up the estimate of the pooled variance as well.
> As with all statistics there are tradeoffs and assumptions that are
> being made regardless of what you do.
>> So, while I might be willing to filter out known "dead" probesets
>> (that I never see above detection threshold over many hundreds of
>> assays), I'm in the camp that the statistics are corrupt if you filter
>> without regard to its affect on multiplicity corrections.
> I don't really know what you mean by 'detection limit'. Has someone
> published something somewhere that says a probeset with an expression
> value below X means the mRNA for that gene has not been detected?
> I am not sure how the filtering step will affect multiplicity
> corrections. If one were to use a two-stage modeling procedure that you
> seem to think I am advocating then of course the p-values themselves
> would be questionable as assumptions would have been violated. But I
> don't know where multiplicity correction comes into the equation.
> But personally I am not that much of a purist about multiplicity anyway.
> I have been known to select probesets based on adjusted p-value and a
> fold change criterion as well, which completely invalidates the meaning
> of the adjusted p-values.
>> As an aside, it should be possible to fit some of the models using
>> truncated/censored distributions (wherein the statistical model gets
>> to know that there were X number of probesets with values < threshold,
>> but doesn't pretend that those values are real). That's an idea for
>> the model developers to ponder ...
More information about the Bioconductor