Coarsening-at-random conditions for scRNA-seq zero inflation: a reliability-theoretic perspective

The single-cell RNA-sequencing (scRNA-seq) community has long debated whether observed zeros reflect biological absence of expression or technical dropout. Jiang et al. (2022) formalize this as a glass ceiling: existing methods cannot distinguish the two sources from observed counts alone. We show that this glass ceiling is an instance of a known identifiability problem from masked-data reliability statistics: the scRNA-seq zero-inflation problem maps directly onto the masked-cause series-system framework, with ERCC spike-ins playing the role of singleton candidate sets that restore identifiability. Importing the C1, C2, C3 coarsening conditions, we derive four theorems: cell-total consistency for the ZINB MLE; form-robustness of spike-in correction across dropout-function families; heterogeneity-robustness under cell-type mixtures; and an explicit first-order bias bound as a function of the spike-in to endogenous capture-efficiency gap. A closed-form expression for the bias coefficient under NB-plus-logistic dropout is derived; its first-order prediction matches simulation at moderate gap magnitudes to within 1 percentage point. A simulation portfolio confirms each theorem. Cell-total consistency holds to numerical precision; matched spike-in correction recovers per-gene means within 2 percent relative bias, mismatched-form within 3 to 6 percent.