Fig. 6
From: Analysis and correction of compositional bias in sparse sequencing count data
Ignoring zeroes can introduce bias in normalization, when zeroes predominantly arise from under-sampling. An artificial example with 10 features and two groups (“controls” and “cases”), when one of the features undergoes a roughly 50X expansion (a log2 fold change of 5.64) in cases compared to controls. This drives the relative abundances of the rest of the 9 features relatively low in the case group. As a result features that are largely present in the controls are not observed in the case group at moderate sequencing depths. Scaling normalization strategies that derive scales based only on the positive count values, can underestimate compositional changes as shown