residuals.survfit: IJ residuals from a survfit object.

Description

Return infinitesimal jackknife residuals from a survfit object, for the survival, cumulative hazard, or restricted mean time in state (RMTS).

Usage

# S3 method for survfit
residuals(object, times, 
    type="pstate", collapse=FALSE, weighted= collapse, data.frame=FALSE,
    extra = FALSE, ...)

Value

A matrix or array with one row per observation or cluster, and one column for each value in times. For a multi-state model the three dimensions are observation, state, and time. For cumulative hazard, the second dimension is the set of transitions. (A competing risks model for instance has 3 states and 2 transitions.)

Arguments

object: a survfit object
times: a vector of times at which the residuals are desired
type: the type of residual, see below
collapse: add the residuals for all subjects in a cluster
weighted: weight the residuals by each observation's weight
data.frame: if FALSE return a matrix or array
extra: return extra information when data.frame=FALSE. (This is used internally by the psuedo function.)
...: arguments for other methods

Details

This function is designed to efficiently compute the per-observation residuals for a Kaplan-Meier or Aalen-Johansen curve, also known as infinitesimal jackknife (IJ) values, at a small number of time points. Common usages are the creation of psuedo-values (via the pseudo function) and IJ estimates of variance. The residuals matrix has a value for each observation and time point pair. For a multi-state model the state will be a third dimension.

The residuals are the impact of each observation or cluster on the resulting probability in state curves at the given time points, the cumulative hazard curve at those time points, or the expected sojourn time in each state up to the given time points. For a simple Kaplan-Meier the survfit object contains only the probability in the "initial" state, i.e., the survival fraction. In this case the sojourn time, the expected amount of time spent in the initial state, up to the specified endpoint, is commonly known as the restricted mean survival time (RMST). For a multistate model this same quantity is more often referred to as the restricted mean time in state (RMTS). It can be computed as the area under the respective probability in state curve.

The program allows any of pstate, surv, cumhaz, chaz, sojourn, rmst, rmts or auc for the type argument, ignoring upper/lowercase, so users can choose whichever abbreviation they like best.

When collapse=TRUE the result has the cluster identifier (which defaults to the id variable) as the dimname for the first dimension. If the fit object contains more than one curve, and the same identifier is reused in two different curves this approach does not work and the routine will stop with an error. In principle this is not necessary, e.g., the result could contain two rows with the same label, showing the separate effect on each curve, but this was deemed too confusing.

Examples

Run this code

fit <- survfit(Surv(time, status) ~ x, aml)
resid(fit, times=c(24, 48), type="RMTS")