We also performed sensitivity analysis by the receiver-operating characteristic and precision-recall curve (Fig. 2a, b and Supplementary Table 1). To allow for a fair comparison of frequentist and Bayesian methods, we only considered the case of more than one sample per group for all methods, since frequentist tests are not applicable in the one-sample case. Furthermore, we excluded the standard Dirichlet-Multinomial model from the comparison due to problematic thresholding. In both metrics, scCODA outperformed all other tested methods (AUC = 0.99; average precision Score=0.94). Most other compositional methods also showed adequate ability to accurately recover the true effects, while non-compositional methods were among the worst-performing methods.
