When trying to adjust data for batch effects using the respective limma (v3.34.9) function removeBatchEffect() I got strange values (at least to my understanding).

My question: In the example below wouldn't one expect row 1 to show the same value distribution as row 3, i.e. the mean value of the respective row? What am I missing here?

library(limma)

> a = cbind(1,4,c(1,4,NA),c(NA,1,1))
> a
     [,1] [,2] [,3] [,4]
[1,]    1    4    1   NA
[2,]    1    4    4    1
[3,]    1    4   NA    1

> removeBatchEffect(a,1:4)
       1   2   3   4
[1,] 1.0 1.0 1.0  NA
[2,] 2.5 2.5 2.5 2.5
[3,] 2.0 2.0  NA 2.0
Warnmeldung:
Partial NA coefficients for 2 probe(s)

According to your code, every observation belongs to a different batch. The whole concept of batch removal doesn't make any sense in this context. Why are you doing this?

The reason why the returned value isn't always the mean of the non-NA values is that removeBatchEffects doesn't reorthogonalize the design matrix when some of the batch effects become non-estimable. The function isn't designed for situations when one of the batches is entirely missing. But it hardly matters. You're removing all the signal from the data, so the value that remains doesn't matter from a DE point of view.