Fig. 2

miRglmm and miRglmm-Poisson outperform the aggregation methods in terms of mean squared error (MSE) for miRNA in which an effect is induced (panel A). DESeq2 provides the lowest MSE when there is no effect induced. The 95% confidence interval coverage proportion of miRglmm and miRglmm-Poisson are much higher than the coverage proportion of the aggregation methods when an effect is induced (panel B). miRglmm and miRglmm-Poisson provide more precise estimates of differential expression compared to aggregation methods (panel C). All methods, except Wilcoxon, perform similarly in terms of identifying significant differential expression when it exists (True Fold Change 0.5 or 2) and failing to reject the null when there is no difference (True Fold Change = 1) (panel D). edgeR does not provide SE estimates to allow calculation of coverage proportion, and Wilcoxon does not provide effect estimates to calculate MSE, coverage proportion or variance, so these methods are not present in those respective panels. Results are based on 100 independent simulated data sets