1 Abstract

For RNA-sequencing analyses, the estimated variance used to test differential expression of genes is assumed to be equal between different conditions for most published methods, including “gold-standard” methods such as that of limma-voom, edgeR and DESeq2. However, there are cases where variability between conditions are distinctly different to each other and it is inappropriate to assume equal variances between conditions. Demissie et al. (2008) proposes a moderated Welch test which performs better than the moderated t-test when group variances are unequal. We instead handle heteroscedasticity between groups by carrying out group-specific modelling of variances. We show that this achieves greater power than modelling with a pooled-trend.