statLens Method Card
edgeR: empirical Bayes negative binomial models for count data
v4.10.7STATLENS VERIFIEDModels sequencing counts per feature with negative binomial or quasi-likelihood GLMs, sharing information across features to estimate dispersions. Designed for differential expression with few replicates and complex multi-factor designs.
Category
Differential abundance
Analysis task
Differential analysis, Association
Study designs
Cross-sectional, Case-control, Cohort, Paired, Clinical trial, Multi-centre
Source
AI-drafted from the paper and package documentation; pending expert review
Scientific questions this method addresses
Which genes or tags differ in abundance between groups in sequencing count data?
Which features are associated with experimental factors in a multi-factor count design?
Which features change between paired samples taken before and after treatment from the same subjects?
Statistical specification
y_{gi} \sim \mathrm{NB}(\mu_{gi}, \phi_g), \quad \log \mu_{gi} = x_i^\top \beta_g + \log N_i
Model family
Generalized linear model per feature (negative binomial or quasi-likelihood)
Distribution
Negative binomial
Estimand
Log fold change per feature; H₀: a contrast of β_g = 0 (quasi-likelihood F-test or likelihood ratio test)
Multiple testing
Benjamini–Hochberg FDR
Assumptions
Input is raw counts; library composition differences are handled by TMM normalization; Most features are not differentially abundant (TMM assumption); Dispersion is shared or trended across features with similar abundance; Samples are independent given the design; only fixed effects are modelled
Sample size
Works from 2 replicates per group; more replicates improve dispersion estimates
Best used when
Small numbers of replicates
Complex multi-factor GLM designs with custom contrasts
The quasi-likelihood pipeline (glmQLFit) is used for strict error-rate control
Other count assays such as ChIP-seq or CRISPR screens
Do not use when
Data are normalized, transformed or non-integer
Repeated measures need random effects
Counts are extremely sparse or zero-inflated and low-count features are not filtered
Most features are expected to change, which breaks the TMM assumption
Relationships
Typed edges to other Method Cards. These are what let statLens assemble multi-step workflows rather than single methods.
alternative_toDESeq2Same negative binomial family; differs in normalization (TMM vs median-of-ratios) and dispersion estimation
alternative_tolimmalimma-voom can take an edgeR DGEList directly and fits precision-weighted linear models instead
requireslimmaedgeR depends on limma for shared linear-model infrastructure
Outputs
Mapped to the statLens standard statistical result, plus method-specific fields.
Standard fields
feature_idestimatetest_statisticp_valueadjusted_p_value