statLens Method Card
limma: linear models with empirical Bayes moderated statistics
v3.68.5STATLENS VERIFIEDFits a linear model to each feature and borrows information across features with empirical Bayes variance moderation, giving stable tests even with very few samples. Works on microarray and other continuous log-scale data, and on RNA-seq counts through voom precision weights.
Category
Regression / mixed models
Analysis task
Differential analysis, Association
Study designs
Cross-sectional, Case-control, Cohort, Paired, Longitudinal, Repeated measure, Clinical trial
Source
AI-drafted from the paper and package documentation; pending expert review
Scientific questions this method addresses
Which genes are differentially expressed between two or more conditions?
Which features change with a continuous covariate after adjusting for batch, sex or age?
Which genes respond to treatment in paired before/after samples from the same patients?
Which features change over time in a designed time-course experiment?
Statistical specification
E(y_g) = X\beta_g, \quad \tilde{s}_g^2 = \frac{d_0 s_0^2 + d_g s_g^2}{d_0 + d_g}
Model family
Linear model per feature with empirical Bayes moderated variances
Estimand
Log fold change (a contrast of β_g) per feature; H₀: cᵀβ_g = 0, tested with the moderated t-statistic
Multiple testing
Benjamini–Hochberg FDR
Assumptions
Log-scale expression is approximately normal given the covariates; Feature variances share a common prior (scaled inverse chi-square); Samples are independent, or within-block correlation is common across features (duplicateCorrelation); For RNA-seq: library sizes are normalized and the voom mean–variance trend is well estimated
Sample size
Usable from 2 replicates per group; power improves with more replicates
Best used when
Few replicates (2–3 per group), where per-feature variances are unstable
Complex designs: several factors, covariates, interactions and custom contrasts
Microarray, proteomics, methylation M-values and other continuous log-scale data
RNA-seq with moderate to high sequencing depth, via voom
Do not use when
Raw counts are passed directly without voom or a log transformation
Counts are extremely sparse, such as raw single-cell or low-depth microbiome counts
Relative abundances are analysed without a compositional transformation
The outcome is binary, censored or time-to-event
Relationships
Typed edges to other Method Cards. These are what let statLens assemble multi-step workflows rather than single methods.
alternative_toedgeRedgeR models counts directly with a negative binomial GLM; limma-voom uses precision-weighted linear models and is faster, with comparable performance in many benchmarks
alternative_toDESeq2DESeq2 models raw counts with a negative binomial GLM and shrinks fold changes
can_followComBatBatch-corrected data can be explored with limma, but adding batch to the design matrix is usually preferred for testing
Outputs
Mapped to the statLens standard statistical result, plus method-specific fields.
Standard fields
feature_idestimatetest_statisticp_valueadjusted_p_valueconfidence_interval_lowerconfidence_interval_uppercontrast