Search
Question: Limma un moderated t-test
1
2.3 years ago by
chris86380
UCL, United Kingdom
chris86380 wrote:

Hi

The limma t test function is very fast compared with using the R function with a for loop for large data sets. I want to use this t test function on other data types we have that are not high throughput. So I want it to do a t test that is not moderated for these data. Does anyone know how to do this?

Thanks,

Chris

modified 2.3 years ago by Axel Klenk910 • written 2.3 years ago by chris86380
4
2.3 years ago by
Aaron Lun20k
Cambridge, United Kingdom
Aaron Lun20k wrote:

If your other data sets are not high-throughput, what's the harm in running the t.test function over a couple of loops? It probably wouldn't take too long for a small number of features. Besides, empirical Bayes is more versatile than you might think. Even though it's used in limma to share information across genes, it can also be used generally, e.g., across genomic windows, TF binding sites or baseball players. So, you might consider whether you could just use the moderated t-test for your data, even if that data is lower-throughput.

With all that said, if you want to compute unmoderated t-statistics, I guess you'd do something like this:

# where 'fit' is the output from lmFit() or contrasts.fit().
unmod.t <- fit$coefficients/fit$stdev.unscaled/fit$sigma pval <- 2*pt(-abs(unmod.t), fit$df.residual)

However, you better have a large sample size, otherwise your power to reject the null hypothesis will be low.

Thanks, that's useful. I think the power is fine because each group has sizes 40 - 70.

Oops, forgot to multiply by 2 to make it two-sided. Also note that pval is a matrix, so you'll need to subset it by column to get the test for your coefficient of interest. For two-group setups, this should be equivalent to t.test with var.equal=TRUE.

3
2.3 years ago by
Axel Klenk910
Switzerland
Axel Klenk910 wrote:

Or try functions rowttests() / fastT() from package genefilter.

Cheers,

- axel

very useful, this.

I am surprised that R did not already have a 'vectorized' t.test.