model.edgeR.tagcom: Modeling NB Tagwise-Common Dispersion with the Adjusted...

In gu-mi/NBGOF: Goodness-of-Fit Tests and Model Diagnostics for Negative Binomial Regression of RNA Sequencing Data

Description

This function fits an NB regression model with tagwise dispersions using the adjusted profile likelihood estimator (APLE). In edgeR, this tagwise dispersion model lies between two extreme scenarios: entirely individual φ_i for each gene and entirely shared values (i.e. common dispersion). The function estimateGLMTagwiseDisp in edgeR shrinks the dispersion towards a common dispersion or a global dispersion trend. In our implementation, the function model.edgeR.tagcom shrinks the tagwise dispersions towards a common dispersion. See details below. The output of this function will be passed to the main GOF function nb.gof.m.

Usage

model.edgeR.tagcom(counts, x, lib.sizes=colSums(counts), prior.df = prior.df, method=method)

Arguments

counts	an m-by-n count matrix of non-negative integers. For a typical RNA-Seq experiment, this is the read counts with m genes and n samples.
x	an n-by-p design matrix.
lib.sizes	library sizes of an RNA-Seq experiment. Default is the column sums of the counts matrix.
prior.df	prior degrees of freedom to control the level of shrinkage (default is 10). See estimateGLMTagwiseDisp for more details. This argument is controlled by a higher level function nb.gof.m.
method	method for estimating the common dispersion, including "CoxReid", "Pearson" and "deviance". If NULL, then use the "CoxReid" method. See estimateGLMCommonDisp for more details.

Details

In this tagwise-common dispersion model, the dispersion parameter φ_i is estimated by maximizing a penalized log-likelihood APL_g(φ_g) plus a weighted shared log-likelihood. The weight denoted by G_0 controls the level of shrinkage applied to purely genewise dispersions towards a common dispersion. In the paper of McCarthy et al. (2012), G_0=20/df performs well over real RNA-Seq datasets. Here the numerator 20 is the prior degrees of freedom, which can be specified by the argument prior.df in the lower-level function dispCoxReidInterpolateTagwise in edgeR. The denominator "df" is the residual degrees of freedom (number of libraries minus the number of distinct treatment groups). We leave the default value of prior.df at 10. See the estimateGLMCommonDisp, estimateGLMTagwiseDisp, dispCoxReidInterpolateTagwise and glmFit functions in the edgeR package for more details.

Value

A list of quantities to be used in the main nb.gof.m function.

Author(s)

Gu Mi <[e-mail]>, Yanming Di, Daniel Schafer

References

See https://github.com/gu-mi/NBGOF/wiki/ for more details.

See Also

model.edgeR.genewise for the non-shrinkage, entirely genewise model, and model.edgeR.tagtrd for the tagwise model that shrinks towards the trended dispersion.

gu-mi/NBGOF documentation built on Oct. 25, 2020, 3:30 a.m.