Comparison of false positives found by different methodologies. Using real metagenomic data, we simulated features with no differential abundance by randomly dividing subjects from a single population into two subpopulations.

We found that for a stringent p-value threshold of 0.001, the negative binomial model (NB) resulted in a false positive rate 20 times higher than the other methodologies. The Log-t of Lu et al. resulted in the lowest false positive rate among the methods tested while Student's test and Metastats performed equally well.