This is not recommended due to a few underlying problems with RNA-seq data that include compositional bias, library size differences, and correlation between mean and variance. I won't go into these concepts in detail but use software designed to account for these problems such as DESeq2 or edgeR. Their manuals will have more information on the above concepts.

Ah, the never ending question: I usually link this thread here from the limma/edgeR developer which discusses strategies if you only have normalized counts and need to deal with it: https://support.bioconductor.org/p/125144/#125156