Probably need to add a bit to the Supplementary Materials about power (also add a connecting sentence in the main text where power is first considered):
The preservation of detection power in the summation approach is driven by several factors. Firstly, as previously discussed, the estimate of the DE log-fold change is the same regardless of whether the average expression in each group is computed over cells or over plates. Secondly, the conditional variance of the count sum is substantially lower than that of the individual counts. This offsets any loss of power from the decrease in the number of samples when plates are considered instead of cells. While this effect is most obvious in situations with equal numbers of equally sized cells on each plate, power is still preserved in simulations involving varying numbers or sizes of cells (Supplementary Figure SX1). Finally, methods like edgeR and limma share information between genes to estimate the dispersion or variance. This mitigates the effect of reduced residual d.f. with fewer samples (though similar results are still observed in situations where EB shrinkage is limited -- see Supplementary Figure SX2).
And we need to add two figures containing ROC curves. One figure contains several subplots (one for each alternative simulation), and the other contains subplots for limited shrinkage.
LTLA
commented
8 years ago
Probably need to add a bit to the Supplementary Materials about power (also add a connecting sentence in the main text where power is first considered):
And we need to add two figures containing ROC curves. One figure contains several subplots (one for each alternative simulation), and the other contains subplots for limited shrinkage.