residuals.survreg {survival}R Documentation

Compute Residuals for ‘survreg’ Objects

Description

This is a method for the function residuals for objects inheriting from class survreg.

Usage

## S3 method for class 'survreg'
residuals(object, type=c("response", "deviance","dfbeta","dfbetas",
"working","ldcase","ldresp","ldshape", "matrix"), rsigma=TRUE,
collapse=FALSE, weighted=FALSE, ...)

Arguments

object

an object inheriting from class survreg.

type

type of residuals, with choices of "response", "deviance", "dfbeta", "dfbetas", "working", "ldcase", "lsresp", "ldshape", and "matrix".

rsigma

include the scale parameters in the variance matrix, when doing computations. (I can think of no good reason not to).

collapse

optional vector of subject groups. If given, this must be of the same length as the residuals, and causes the result to be per group residuals.

weighted

give weighted residuals? Normally residuals are unweighted.

...

other unused arguments

Value

A vector or matrix of residuals is returned. Response residuals are on the scale of the original data, working residuals are on the scale of the linear predictor, and deviance residuals are on log-likelihood scale. The dfbeta residuals are a matrix, where the ith row gives the approximate change in the coefficients due to the addition of subject i. The dfbetas matrix contains the dfbeta residuals, with each column scaled by the standard deviation of that coefficient.

The matrix type produces a matrix based on derivatives of the log-likelihood function. Let L be the log-likelihood, p be the linear predictor X %*% coef, and s be \log(σ). Then the 6 columns of the matrix are L, dL/dp,ddL/(dp dp), dL/ds, ddL/(ds ds) and ddL/(dp ds). Diagnostics based on these quantities are discussed in the book and article by Escobar and Meeker. The main ones are the likelihood displacement residuals for perturbation of a case weight (ldcase), the response value (ldresp), and the shape.

For a transformed distribution such as the log-normal or Weibull, matrix residuals are based on the log-likelihood of the transformed data log(y). For a monotone function f the density of f(X) is the density of X divided by the derivative of f (the Jacobian), so subtract log(derivative) from each uncensored observation's loglik value in order to match the loglik component of the result. The other colums of the matrix residual are unchanged by the transformation.

References

Escobar, L. A. and Meeker, W. Q. (1992). Assessing influence in regression analysis with censored data. Biometrics 48, 507-528.

Escobar, L. A. and Meeker, W. Q. (1998). Statistical Methods for Reliablilty Data. Wiley.

See Also

predict.survreg

Examples

fit <- survreg(Surv(futime, death) ~ age + sex, mgus2)
summary(fit)   # age and sex are both important

rr  <- residuals(fit, type='matrix')
sum(rr[,1]) - with(mgus2, sum(log(futime[death==1]))) # loglik

plot(mgus2$age, rr[,2], col= (1+mgus2$death)) # ldresp

[Package survival version 2.44-1.1 Index]