Hello:
Apart from the original paper that introduced limma-voom, are there any studies that use real data and show that limma-voom works better than limma-trend?
Regards,
Nik Tuzov
Hello:
Apart from the original paper that introduced limma-voom, are there any studies that use real data and show that limma-voom works better than limma-trend?
Regards,
Nik Tuzov
The "original paper" that you refer to introduced both methods, voom and limma-trend, and showed good performance for both. voom was shown to have the edge when the sequencing depths were very inconsistent between replicates but, otherwise, limma-trend was just as good.
As far as I am aware, third party comparison studies have tended to concentrate on voom and limma-trend has been somewhat neglected.
Guidance on when to use limma-trend or voom is given in the limma User's Guide.
Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
Thank you for replying.
On a related note, is there any downside to using voom when limma-trend would be appropriate? For example, is limma-trend theoretically more powerful if the sequencing depths are consistent?
If by "powerful" you mean "more likely to call a TP" then it's impossible to say. For a given fold change, the p-value can go up or down when you switch from "trend" to voom. Voom seems to have more variance parameters than "trend", so the downside of voom is possible overfitting of the model. How much one will pay for a wrong choice between these two options will depend on sample size and the signal-to-noise ratio in the dataset.
I don't think voom is actually estimating more parameters than eBayes with trend=TRUE. They're both estimating the relationship between abundance and variance by fitting a single lowess/loess curve, so I think that the same number of parameters is being estimated either way. They just differ in how they apply the trend to the data.
For a given gene, limma-trend extracts one variance parameter from that lowess curve, and limma-voom extracts N of them where N is the number of libraries. It is assumed that when the library sizes are equal all N parameters coincide and it's better to use limma-trend.
No, that is not correct. As Ryan says, voom and limma-trend both fit similar non-parametric curves. In both cases, the trend is equivalent to about 3-4 unknown parameters. The parameters are estimated from the whole data set, not for an individual gene, so neither method is subject to "over-fitting". Actually the curve that voom fits is very slightly smoother than the one that limma-trend fits, so voom is actually in effect estimating fractionally fewer parameters than limma-trend.
I think you might be confusing observation weights with parameters. The voom observation weights are computed from the variance trend and so are not independent unknown parameters.
Is there a compelling reason to use lowess in limma-voom vs splines in limma-trend?
If you examine the documentation to `contrasts.fit`, there is a warning that using weights with a non-orthogonal design matrix uses an approximation. Since voom uses weights on inidividual observations, and limma-trend does not, this could be an advantage to using limma-trend.