I have miRNA-seq data, and as part of the QC steps, I want to calculate the coefficient of variation between the QC samples (including inter-plate, intra-plate, and commercial universal controls) using DESeq-normalized counts. For these calculations, do I have to normalize only those specific QC samples together, or do I have to normalize these QC samples together with the rest of the other test samples? Note that not all the samples have the same miRNAs (some samples have 0 counts for specific miRNAs)

I tried both methods (normalizing only the QC samples and normalizing them with other test samples), and the CVs were different. So, I am wondering which is the correct way to give me the true variation between the QC samples?

Normalize together.

When you compute counts(dds, normalized=TRUE) these are scaled counts, where the scaling is toward the middle of the samples.

So if you have samples in the millions of reads, say 60-80 millions of reads, then the scaled counts are the values as if all samples were in the middle, say 70 million reads.

When you treat the datasets separately, the scaled counts won't be on the same scale necessarily.