tagwise

Tagwise is a term used in statistics and high-throughput sequencing analysis to denote a per-feature dispersion parameter estimate, most commonly encountered in RNA-seq differential expression analysis within Bioconductor's edgeR package. It refers to gene-specific dispersion values used in negative binomial models of count data.

In count data from RNA-seq, variance often exceeds the mean, which is captured by a dispersion parameter in the negative binomial distribution. Dispersion can be modeled as common (the same for all genes), trended (dependent on mean expression level), or tagwise (gene-specific). Tagwise dispersion estimates are obtained by borrowing strength across genes through an empirical Bayes approach, stabilizing estimates for genes with low counts or few replicates by leveraging information from the broader dataset.

In practice, tagwise dispersion is integrated into the generalized linear model fitting process. After preparing the count data and design matrix, dispersion components are estimated in stages, typically starting with a common dispersion, then a trend, and finally tagwise dispersions. The resulting tagwise values influence the standard errors and test statistics used to identify differential expression, improving accuracy when gene-to-gene variability in dispersion is present.

Limitations and alternatives: Tagwise dispersion estimation requires multiple biological replicates and careful quality control; it can be sensitive to outliers or model misspecification, which may affect results. Alternatives and complementary approaches include DESeq2 with gene-wise shrinkage and limma-voom, which offer different strategies for handling dispersion and variance in count data. Tagwise dispersion remains a core component of many RNA-seq analysis pipelines for robust differential expression inference.