In addition to common univariate statistical methods as well as RNAseq method (limma/edgeR/DESeq). MicrobiomeAnalyst support three methods designed for microbiome data - MaAsLin2, LinDA and ANCOM-BC2. All three are available for single-factor comparisons (Univariate tests) and for models with covariates (Covariate analysis), and all three fit one model per taxon with your group variable, covariates and an optional Block random effect. They differ in what they feed into that model and in how they deal with compositionality.
MaAsLin2 regresses log2 relative abundance (total-sum scaling with a pseudo-count; count-based negative binomial, zero-inflated and compound Poisson models are optional). Compositionality is not modelled, and the coefficient is roughly the log2 fold change of relative abundance.
LinDA regresses a centred log-ratio-type transform of the counts (adaptive pseudo-count, values winsorised at the 97th percentile) and then subtracts the shared shift that compositional closure imposes on all taxa. Its log2 fold change is closer to the change in absolute abundance, and it runs in well under a second.
ANCOM-BC2 also models absolute abundance, but estimates the per-sample sampling fraction (the compositional bias) explicitly with an iterative procedure, detects structural zeros (taxa absent from an entire group), and by default repeats the fit with several pseudo-counts to flag results that depend on how zeros were handled. It is the most thorough of the three and by far the slowest: in
Why they disagree — they test different hypotheses. Default MaAsLin2 tests changes in relative abundance: each taxon’s share of the sequencing reads. LinDA and ANCOM-BC2 both test changes in a bias-corrected proxy for absolute abundance — LinDA by removing the shared shift that compositional closure imposes on all taxa, ANCOM-BC2 by explicitly estimating each sample’s sampling fraction. When MaAsLin2 disagrees with either of them, it is usually because they are answering mathematically different questions: if a few abundant taxa truly increase, the relative abundance of everything else must fall, and MaAsLin2 reports that fall while the other two do not. LinDA and ANCOM-BC2 test the same hypothesis and usually agree on direction, but their lists still differ at the margin because they estimate the bias differently and handle zeros differently (winsorisation versus a pseudo-count sensitivity check and structural-zero detection); those differences show up near the significance cutoff rather than among the strongest signals. Effect sizes are on different scales across methods even though all are reported as log2 fold change, so compare ranked lists rather than numbers.