Methods

Note: mvabund is not included by default in the various functions, as it is slow. It can be included by including "mva" in the tests argument.

---

Overview of methods

	Abbr. (link)	Response	Paired/Blocked	Covars	Normalization	Transf.	Model
ANCOM-BC	abc	All	No	Yes	Bias-correction		Log-linear
ALDEx2	adx	Two-class	No	No	ILR		Poisson/Dirichlet
ANOVA	aov	Multi-class	No	Yes	TSS (2)		Gaussian
ANOVA - ALR	aov	Multi-class	No	Yes	ALR		Gaussian
ANOVA - CLR	aov	Multi-class	No	Yes	CLR		Gaussian
ANOVA log	lao	Multi-class	No	Yes	TSS (2)	Log (3)	Gaussian
ANOVA log #2	lao2	Multi-class	No	Yes	TSS	Log	Gaussian
baySeq	bay	Two-class	No	No	Quantile (4)		Negative Binomial
Correlation - Pearson	pea	Quantitative	No	No	TSS (2)		Bivariate Normal
Correlation - Spearman	spe	Quantitative	No	No	TSS (2)		Nonparametric
DESeq2	ds2x	Categorical	Yes, as covariate	Yes	RLE		Negative Binomial
DESeq2 man. geoMeans	ds2	Categorical	Yes, as covariate	Yes	RLE (5)		Negative Binomial
EdgeR - Exact test	ere	Two-class	No	No	TMM		Negative Binomial
EdgeR - Exact test2	ere2	Two-class	No	No	RLE		Negative Binomial
EdgeR - Quasi likelihood	erq	All	Yes, as covariate	Yes	TMM		Negative Binomial
EdgeR - Quasi likelihood2	erq2	All	Yes, as covariate	Yes	RLE		Negative Binomial
Friedman Rank Sum test	fri	Multi-class	Exclusively	No	TSS (2)		Nonparametric
GLM - Negative binomial	neb	All	Yes, as random effects	Yes	None (1)		Negative Binomial
GLM - Poisson	poi	All	Yes, as random effects	Yes	None (1)		Poisson
GLM - Quasi-poisson	qpo	All	No	Yes	None (1)		Quasi-poisson
GLM - ZI Negative Binomial	znb	All	No	Yes	None (1)		Zero-inflated Negative binomial
GLM - ZI Poisson	zpo	All	No	Yes	None (1)		Zero-inflated Poisson
Kruskal-Wallis test	kru	Multi-class	No	No	TSS (2)		Nonparametric
LIMMA	lim	All	Yes, as random effects	Yes	TSS (2)		Gaussian
LIMMA - ALR	lim	All	Yes, as random effects	Yes	ALR		Gaussian
LIMMA - CLR	lim	All	Yes, as random effects	Yes	CLR		Gaussian
LIMMA log	lli	All	Yes, as random effects	Yes	TSS (2)	Log (3)	Gaussian
LIMMA log #2	lli2	All	Yes, as random effects	Yes	TSS	Log	Gaussian
LIMMA voom	vli	All	Yes, as random effects	Yes	TMM	Voom	Gaussian
Linear regression	lrm	All	Yes, as random effects	Yes	TSS (2)		Gaussian
Linear regression - ALR	lrm	All	Yes, as random effects	Yes	ALR		Gaussian
Linear regression - CLR	lrm	All	Yes, as random effects	Yes	CLR		Gaussian
Linear regression log	llm	All	Yes, as random effects	Yes	TSS (2)	Log (3)	Gaussian
Linear regression log #2	llm2	All	Yes, as random effects	Yes	TSS	Log	Gaussian
MetagenomeSeq featuremodel	msf	Two-class	No	No	CSS		Zero-inflated Lognormal
MetagenomeSeq ZIG	zig	All	Yes, as random effects	Yes	CSS	Log	Zero-inflated Gaussian
Mvabund	mva	All	Yes, as covariate	Yes	None (1)		Negative Binomial
Permutation test	per	Two-class	Yes	No	TSS (2)		Nonparametric
Quade test	qua	Multi-class	Exclusively	No	TSS (2)		Nonparametric
SAMseq	sam	All	Yes	No	Resampling		Nonparametric
Welch t.test	ttt	Two-class	Yes	No	TSS (2)		Gaussian
Welch t.test - ALR	ttt	Two-class	Yes	No	ALR		Gaussian
Welch t.test - CLR	ttt	Two-class	Yes	No	CLR		Gaussian
Welch t.test log	ltt	Two-class	Yes	No	TSS (2)	Log (3)	Gaussian
Welch t.test log #2	ltt2	Two-class	Yes	No	TSS	Log	Gaussian
Welch t.test Rank-normalization	ttr	Two-class	Yes	No	Rank		(Gaussian)
Wilcoxon	wil	Two-class	Yes	No	TSS (2)		Nonparametric

Table descriptions:

• (1): Log of library sizes used as offset when relative = TRUE
• (2): None when relative = FALSE
• (3): Log transformation is done before normalization (as opposed to after)
• (4): This can be be changed to TSS or TMM
• (5): This version of DESeq2 uses manual geometric means which handle zeroes differently than the default. See more here
• TSS: Total Sum Scaling
• TMM: Trimmed Mean by M-value
• RLE: Relative Log Expression
• CSS: Cumulative Sum Scaling
• ILR: Isometric Log-Ratio
• ALR: Additive Log-Ratio
• CLR: Centered Log-Ratio

Paired permutation test

A paired permutation test is implemented specifically for this package. The test is similar to the original, but with a different test statistic and permutation scheme. The permutations are constrained in the paired version such that the predictor is only permuted within each level of the paired argument (e.g. subjects). The test statistic first finds the log-ratio between the two predictor levels (e.g. case and control) for each level of the paired argument and the final statistic is the mean of these log-ratios.

Is your favorite method missing?

Either add it yourself (see under 'Extra features'), or suggest it in a GitHub issue