Covariates and Paired Designs

Covariates and paired designs

Covariate adjustment

With a design file containing more than one factor, the moderated t method can adjust for additional factors. The model per gene is

log₂ expression ~ group + covariate₁ + covariate₂ + …

and the test is on the group coefficient. Variance moderation uses the residual degrees of freedom of the adjusted model.

Covariate type	Encoding
Categorical	Indicator columns
Numeric with more than two distinct values	Continuous slope
Numeric with two values	Indicator

Covariate-adjusted results match limma lmFit with the corresponding design matrix and eBayes.

Paired and blocked designs

Supply the subject (or block) identifier as a categorical covariate. This fits the subject as a fixed-effect block, which is the standard analysis for paired samples.

Refused designs

benchsiDE refuses to fit, and states the reason, when:

• a covariate is completely confounded with the groups being compared,
• a covariate is constant within the contrast, or
• the adjusted model has no residual degrees of freedom.

Other methods

voom and Welch t results are unadjusted. When a covariate is selected with these methods, the results carry an explicit warning.