Although being important drivers of biological processes such as in disease1, development2, aging3, and immunity4, shifts in cell-type compositions are non-trivial to detect using scRNA-seq. Statistical tests need to account for multiple sources of technical and methodological limitations, including the low number of experimental replications. The total number of cells per sample is restricted in most single-cell technologies, implying that cell-type counts are proportional in nature. This, in turn, leads to a negative bias in cell-type correlation estimation5 (Fig. 1a). For example, if only a specific cell type is depleted after perturbation, the relative frequency of others will rise. If taken at face value, this would lead to an inflation of differential cell types. Therefore, standard univariate statistical models that test compositional changes of each cell type independently may falsely deem certain population shifts as real effects, even though they were solely induced by the inherent negative correlations of the cell-type proportions (Fig. 1b). Yet, common statistical approaches currently applied in compositional cell-type analysis ignore this effect. For example, Haber et al.6 applied a univariate test based on Poisson regression,
