collectResiduals: Residual hooks and collection methods

View source: R/zlmHooks.R

collectResidualsR Documentation

Residual hooks and collection methods

Description

After each gene is fit, a hook function can optionally be run and the output saved. This allows extended computations to be done using the fitted model, without keeping it in memory. Here this is used to calculate various residuals, though in some cases they can be done using only the information contained in the ZlmFit-class.

Usage

collectResiduals(x, sca, newLayerName = "Residuals")

discrete_residuals_hook(x)

continuous_residuals_hook(x)

combined_residuals_hook(x)

deviance_residuals_hook(x)

fitted_phat(x)

partialScore(x, effectRegex)

Arguments

x

ZlmFit-class

sca

SingleCellAssay object to which the residuals should be added

newLayerName

character name of the assay layer

effectRegex

a regular expression naming columns of the design corresponding to Z_0. Generally these should be the treatment effects of interest.

Value

copy of sca with new layer

Functions

  • discrete_residuals_hook(): Hook to get the discrete residuals, ie, difference between expected probability of expression and observed

  • continuous_residuals_hook(): Hook to get the continuous residuals, ie, residuals for conditionally positive observations. If an observation is zero, it's residual is defined to be zero as well.

  • combined_residuals_hook(): Hook to get the combined residuals, ie, Y-E(U)*E(V)

  • deviance_residuals_hook(): Standardized deviance residuals hook. Computes the sum of the standardized deviance residuals for the discrete and continuous models (scaled to have unit variance). If the observation is zero then only the discrete component is used.

  • fitted_phat(): Hook to return p_hat, the predicted probability of expression.

  • partialScore(): Compute Y_i-E(V_i|X_i, Z_0)E(U|X_i, Z_0), where Z_0 is a treatment effect (being left in) and X_i is a nuisance effect (being regressed out).

Total residual types

Each component of the model contributes several flavors of residual, which can be combined in various fashions. The discrete residual can be on the response scale (thus subtracting the predicted probability of expression from the 0/1 expression value). Or it can be a deviance residual, revealing something about the log-likelihood.

Partial residuals

It's also possible to consider partial residuals, in which the contribution of a particular covariate is added back into the model.

See Also

zlm

Examples

data(vbetaFA)
svbeta <- subset(vbetaFA, ncells==1)
svbeta <- svbeta[freq(svbeta)>.4,]
window <- function(x1) lapply(assays(x1), function(x2) x2[1:3, 1:6])
#total residuals of the response
z1 <- zlm(~ Stim.Condition, svbeta, hook=discrete_residuals_hook)
window(collectResiduals(z1, svbeta))
z2 <- zlm(~ Stim.Condition, svbeta, hook=continuous_residuals_hook)
window(collectResiduals(z2, svbeta))
z3 <- zlm(~ Stim.Condition, svbeta, hook=combined_residuals_hook)
window(collectResiduals(z3, svbeta))
#partial residuals
colData(svbeta)$ngeneson <- colMeans(assay(svbeta)>0)
z5 <- zlm(~ Stim.Condition + ngeneson, svbeta)
partialScore(z5, 'Stim.Condition')

RGLab/MAST documentation built on Sept. 30, 2023, 1:08 p.m.