event (version 1.1.2)

survkit: Weibull and Cox Models with Random Effects

Description

survfit was written in Fortran by Dr. V. Ducrocq (INRA, France: [e-mail]) and Dr. J. Soelkner (Vienna: [e-mail]) to fit Weibull and Cox proportional hazards models with random effects for very large data sets. This is a cut-down version adapted to R. The full Survival Kit, including the manual, can be obtained from http://www.boku.ac.at/nuwi/popgen.

Usage

survkit(times, censor=NULL, ccov=NULL, tvcov=NULL,
	strata=NULL, id=NULL, model="Weibull", baseline=FALSE,
	residuals=FALSE, survival=NULL, svalues=NULL, valrho=NULL,
	constraints=NULL, impose=NULL, dist=NULL, random=NULL,
	estimate=NULL, moments=FALSE, rule=NULL, pedigree=NULL,
	integrate=NULL, jointmode=FALSE, within=NULL, converge=1.e-8,
	iterlim=100)

Arguments

times

Vector of times (events, right-censoring, change in time-varying covariate, left-truncation).

censor

Corresponding vector of censoring indicators. 1: event; 0: censored; -1: change of time-varying covariate; -2: left-truncation time.

ccov

Model formula for time-constant covariates. These may have one value per individual or one per time. Because of the way factor variables are handled, interactions must be coded as new variables.

tvcov

Model formula for time-varying covariates with one value per time. There can only be one change-point per individual. Again, interactions must be coded as new variables.

strata

A factor variable specifying stratification. With the Weibull model, different intercepts and power parameters are calculated for each stratum. For the Cox model, a different baseline curve is fitted.

id

A variable giving individual identification numbers (starting at one). If not supplied, all times are assumed to refer to different individuals.

model

Weibull or Cox model, or Kaplan-Meier estimates.

baseline

If TRUE, the baseline values are calculated for the Cox model.

residuals

If TRUE, calculate residuals (only for Cox model).

survival

Calculate values of the survival function at quantiles,or at equally-spaced, specific, or all observed times.

svalues

A vector of quantile values (between 0 and 100), spacing and maximum for equally-spaced, or specific times for survival.

valrho

A fixed value of the Weibull power parameter if it is not to be estimated.

constraints

By default, the category of each factor variable with the largest number of events is taken as baseline. Other options are none which gives values around the mean and find. See also, impose.

impose

A list of a vector of variable names and a corresponding vector of their baseline category numbers. Any factor variables not given will have their first category as baseline.

dist

The distribution of the random effect: loggamma, normal, or multivariate (normal).

random

A factor variable specifying the random effect.

estimate

One fixed value for the mode of the variance of the random effect or three values if the mode is to be estimated: lower and upper bounds, and precision.

moments

Estimate the first three moments of the random effect as well as the mode.

rule

For the multivariate normal random effect, the genetic relationships: usual, mgs (sire or father model), or sire.dam (father and mother).

pedigree

A matrix with four columns required for the multivariate normal random effect, containing the individual id, the sex, the father's category, and the mother's category.

integrate

A factor variable to integrate out as the log-gamma random effect in a Weibull model. (Not available for the Cox model.)

jointmode

If TRUE, the log-gamma variance parameter is estimated simultaneously with the other parameters using the information in estimate. Otherwise, a fixed value, given in estimate is assumed.

within

A second factor variable (within the integrate variable) to integrate out.

converge

The convergence criterion, by default 1.e-8.

iterlim

Maximum number of iterations.

Author

V. Ducrocq, J. Soelkner, and J.K. Lindsey

See Also

coxre, kalsurv.

Examples

# y <- trunc(rweibull(20,2,20))
y <- c(6,22,43,16,7,6,15,35,10,9,18,34,7,13,10,17,14,19,11,13)
# cens <- rbinom(20,1,0.9)
cens <- c(1,1,1,1,0,1,1,1,1,1,1,1,0,1,1,1,1,1,1,1)
id <- gl(2,10)
# x <- rnorm(20)
x <- c(1.82881379,1.06606868,0.70877744,-0.09932880,-0.60626148,-0.75371046,
  0.23884069,0.51199483,-0.73060095,-0.93222151,2.27947539,-0.73855454,
 -0.36412735,-0.89122114,-0.05025962,-0.10001587,1.11460865,-1.87315971,
 -0.11280052,-1.6880509)
# Kaplan-Meier estimates
survkit(y, censor=cens, model="Kaplan")
# null Weibull model
survkit(y, censor=cens)
# one time-constant covariate
survkit(y, censor=cens, ccov=~x)
# stratify
survkit(y, censor=cens, ccov=~x, strata=id)
# estimate a normal random effect
survkit(y, censor=cens, ccov=~x, random=id, dist="normal",
	estimate=c(0.1,10,0.01), moments=TRUE)
# try a fixed value for the normal random effect
survkit(y, censor=cens, ccov=~x, random=id, dist="normal",
	estimate=1.3)
# estimate a log-gamma random effect
survkit(y, censor=cens, ccov=~x, random=id, dist="loggamma",
	estimate=c(0.1,10,0.01))
# estimate a log-gamma random effect by integrating it out
if (FALSE) {
survkit(y, censor=cens, ccov=~x, dist="loggamma", estimate=1.3,
	integ=id, jointmode=TRUE)
# try a fixed value of the log-gamma random effect, integrating it out
survkit(y, censor=cens, ccov=~x, dist="loggamma", estimate=1,
	integ=id)
}
#
# Cox model with one time-constant covariate
print(z <- survkit(y, censor=cens, ccov=~x, model="Cox", residuals=TRUE,
	baseline=TRUE))
residuals(z)
baseline(z)
# obtain the quantiles
print(z <- survkit(y, censor=cens, ccov=~x, model="Cox",
	survival="quantiles", svalues=seq(10,90,by=10)))
survival(z)
# estimate a log-gamma random effect
survkit(y, censor=cens, ccov=~x, model="Cox", random=id,
	dist="loggamma", estimate=c(0.1,10,0.01))