We now show how the choice of the reference cell type can influence the results of scCODA. As an example, we use the ulcerative colitis Lamina propria data from Smillie et al.1, comparing healthy and non-inflamed samples. We applied scCODA to this data 37 times, setting each cell type as the reference once (FDR level 0.05). Supplementary Fig. 8 shows the credible effects and effect size for each reference. For reference cell types that were mostly unchanged, i.e., were almost never found to be differentially abundant in the other runs, the found credible effects are largely consistent. On the other hand, cell types that were assigned a large negative effect (CD4+ activated Fos-lo, plasma cells) found significantly less credible effects when used as the reference, as the null level for the change is already negative. Taking epithelial cells, the only increasing cell type, as the reference led to the largest number of credible negative effects in other cell types. This shows that the reference cell type can have a large impact on the results of scCODA and should therefore be chosen with care.
