Hi all,

I am performing differential expression (DE) analysis using a multi-level experimental design comprising 12 samples from two statuses and two cohorts. I am also incorporating sample weights (arrayWeights).

When I estimated the duplicate correlation using duplicateCorrelation(), I obtained a negative value (-0.15). In the limma User's Guide, in the context of a two-color experiment, the authors mention that when the estimated duplicate correlation is negative, there is no need to account for the correlation as a paired design.

My question is: does this mean that I should specify the design matrix as a paired design, for example:

# Multi-level experimental design
Group <- factor(paste(metadata$Cohort, metadata$Status, sep = "_"))
design <- model.matrix(~ 0 + Group)
colnames(design) <- levels(Group)

versus:

# Paired design
cohort <- factor(metadata$Cohort)
patient <- factor(metadata$Patient.ID)
design <- model.matrix(~ 0 + patient + cohort)

Is the second approach what the authors mean by treating the data as a paired design?

Or should I keep the original multi-level design matrix and simply omit block = metadata$Patient.ID and correlation = corfit$consensus.correlation from the lmFit() step because the estimated duplicate correlation is negative?

Thank you very much for your advice!