comp.risk: Competings Risks Regression

Description

Fits a semiparametric model for the cause-specific quantities : $$ P(T < t, cause=1 | x,z) = P_1(t,x,z) = h( g(t,x,z) ) $$ for a known link-function $h()$ and known prediction-function $g(t,x,z)$ for the probability of dying from cause 1 in a situation with competing causes of death.

Usage

comp.risk(
  formula,
  data = parent.frame(),
  cause,
  times = NULL,
  Nit = 50,
  clusters = NULL,
  est = NULL,
  fix.gamma = 0,
  gamma = 0,
  n.sim = 0,
  weighted = 0,
  model = "fg",
  detail = 0,
  interval = 0.01,
  resample.iid = 1,
  cens.model = "KM",
  cens.formula = NULL,
  time.pow = NULL,
  time.pow.test = NULL,
  silent = 1,
  conv = 1e-06,
  weights = NULL,
  max.clust = 1000,
  n.times = 50,
  first.time.p = 0.05,
  estimator = 1,
  trunc.p = NULL,
  cens.weights = NULL,
  admin.cens = NULL,
  conservative = 1,
  monotone = 0,
  step = NULL
)

Value

returns an object of type 'comprisk'. With the following arguments:

cum: cumulative timevarying regression coefficient estimates are computed within the estimation interval.
var.cum: pointwise variances estimates.
gamma: estimate of proportional odds parameters of model.
var.gamma: variance for gamma.
score: sum of absolute value of scores.
gamma2: estimate of constant effects based on the non-parametric estimate. Used for testing of constant effects.
obs.testBeq0: observed absolute value of supremum of cumulative components scaled with the variance.
pval.testBeq0: p-value for covariate effects based on supremum test.
obs.testBeqC: observed absolute value of supremum of difference between observed cumulative process and estimate under null of constant effect.
pval.testBeqC: p-value based on resampling.
obs.testBeqC.is: observed integrated squared differences between observed cumulative and estimate under null of constant effect.
pval.testBeqC.is: p-value based on resampling.
conf.band: resampling based constant to construct 95% uniform confidence bands.
B.iid: list of iid decomposition of non-parametric effects.
gamma.iid: matrix of iid decomposition of parametric effects.
test.procBeqC: observed test process for testing of time-varying effects
sim.test.procBeqC: 50 resample processes for for testing of time-varying effects
conv: information on convergence for time points used for estimation.

Arguments

formula: a formula object, with the response on the left of a '~' operator, and the terms on the right. The response must be a survival object as returned by the `Event' function. The status indicator is not important here. Time-invariant regressors are specified by the wrapper const(), and cluster variables (for computing robust variances) by the wrapper cluster().
data: a data.frame with the variables.
cause: For competing risk models specificies which cause we consider.
times: specifies the times at which the estimator is considered. Defaults to all the times where an event of interest occurs, with the first 10 percent or max 20 jump points removed for numerical stability in simulations.
Nit: number of iterations for Newton-Raphson algorithm.
clusters: specifies cluster structure, for backwards compability.
est: possible starting value for nonparametric component of model.
fix.gamma: to keep gamma fixed, possibly at 0.
gamma: starting value for constant effects.
n.sim: number of simulations in resampling.
weighted: Not implemented. To compute a variance weighted version of the test-processes used for testing time-varying effects.
model: "additive", "prop"ortional, "rcif", or "logistic".
detail: if 0 no details are printed during iterations, if 1 details are given.
interval: specifies that we only consider timepoints where the Kaplan-Meier of the censoring distribution is larger than this value.
resample.iid: to return the iid decomposition, that can be used to construct confidence bands for predictions
cens.model: specified which model to use for the ICPW, KM is Kaplan-Meier alternatively it may be "cox"
cens.formula: specifies the regression terms used for the regression model for chosen regression model. When cens.model is specified, the default is to use the same design as specified for the competing risks model.
time.pow: specifies that the power at which the time-arguments is transformed, for each of the arguments of the const() terms, default is 1 for the additive model and 0 for the proportional model.
time.pow.test: specifies that the power the time-arguments is transformed for each of the arguments of the non-const() terms. This is relevant for testing if a coefficient function is consistent with the specified form A_l(t)=beta_l t^time.pow.test(l). Default is 1 for the additive model and 0 for the proportional model.
silent: if 0 information on convergence problems due to non-invertible derviates of scores are printed.
conv: gives convergence criterie in terms of sum of absolute change of parameters of model
weights: weights for estimating equations.
max.clust: sets the total number of i.i.d. terms in i.i.d. decompostition. This can limit the amount of memory used by coarsening the clusters. When NULL then all clusters are used. Default is 1000 to save memory and time.
n.times: only uses 50 points for estimation, if NULL then uses all points, subject to p.start condition.
first.time.p: first point for estimation is pth percentile of cause jump times.
estimator: default estimator is 1.
trunc.p: truncation weight for delayed entry, P(T > entry.time | Z_i), typically Cox model.
cens.weights: censoring weights can be given here rather than calculated using the KM, cox or aalen models.
admin.cens: censoring times for the administrative censoring
conservative: set to 0 to compute correct variances based on censoring weights, default is conservative estimates that are much quicker.
monotone: monotone=0, uses estimating equations $$ (D_\beta P_1) w(t) ( Y(t)/G_c(t) - P_1(t,X))$$ montone=1 uses $$ w(t) X ( Y(t)/G_c(t) - P_1(t,X)) $$
step: step size for Fisher-Scoring algorithm.

Author

Thomas Scheike

Details

We consider the following models : 1) the additive model where $h(x)=1-\exp(-x)$ and $$ g(t,x,z) = x^T A(t) + (diag(t^p) z)^T \beta $$ 2) the proportional setting that includes the Fine & Gray (FG) "prop" model and some extensions where $h(x)=1-\exp(-\exp(x))$ and $$ g(t,x,z) = (x^T A(t) + (diag(t^p) z)^T \beta) $$ The FG model is obtained when $x=1$, but the baseline is parametrized as $\exp(A(t))$.

The "fg" model is a different parametrization that contains the FG model, where $h(x)=1-\exp(-x)$ and $$ g(t,x,z) = (x^T A(t)) \exp((diag(t^p) z)^T \beta) $$ The FG model is obtained when $x=1$.

3) a "logistic" model where $h(x)=\exp(x)/( 1+\exp(x))$ and $$ g(t,x,z) = x^T A(t) + (diag(t^p) z)^T \beta$$

The "logistic2" is $$ P_1(t,x,z) = x^T A(t) exp((diag(t^p) z)^T \beta)/ (1+ x^T A(t) exp((diag(t^p) z)^T \beta)) $$ The simple logistic model with just a baseline can also be fitted by an alternative procedure that has better small sample properties see prop.odds.subist().

4) the relative cumulative incidence function "rcif" model where $h(x)=\exp(x)$ and $$ g(t,x,z) = x^T A(t) + (diag(t^p) z)^T \beta $$

The "rcif2" $$ P_1(t,x,z) = (x^T A(t)) \exp((diag(t^p) z)^T \beta) $$

Where p by default is 1 for the additive model and 0 for the other models. In general p may be powers of the same length as z.

Since timereg version 1.8.4. the response must be specified with the Event function instead of the Surv function and the arguments. For example, if the old code was

comp.risk(Surv(time,cause>0)~x1+x2,data=mydata,cause=mydata$cause,causeS=1)

the new code is

comp.risk(Event(time,cause)~x1+x2,data=mydata,cause=1)

Also the argument cens.code is now obsolete since cens.code is an argument of Event.

References

Scheike, Zhang and Gerds (2008), Predicting cumulative incidence probability by direct binomial regression,Biometrika, 95, 205-220.

Scheike and Zhang (2007), Flexible competing risks regression modelling and goodness of fit, LIDA, 14, 464-483.

Martinussen and Scheike (2006), Dynamic regression models for survival data, Springer.

Examples

Run this code


data(bmt); 

clust <- rep(1:204,each=2)
addclust<-comp.risk(Event(time,cause)~platelet+age+tcell+cluster(clust),data=bmt,
cause=1,resample.iid=1,n.sim=100,model="additive")
###

addclust<-comp.risk(Event(time,cause)~+1+cluster(clust),data=bmt,cause=1,
		    resample.iid=1,n.sim=100,model="additive")
pad <- predict(addclust,X=1)
plot(pad)

add<-comp.risk(Event(time,cause)~platelet+age+tcell,data=bmt,
cause=1,resample.iid=1,n.sim=100,model="additive")
summary(add)

par(mfrow=c(2,4))
plot(add); 
### plot(add,score=1) ### to plot score functions for test

ndata<-data.frame(platelet=c(1,0,0),age=c(0,1,0),tcell=c(0,0,1))
par(mfrow=c(2,3))
out<-predict(add,ndata,uniform=1,n.sim=100)
par(mfrow=c(2,2))
plot(out,multiple=0,uniform=1,col=1:3,lty=1,se=1)

## fits additive model with some constant effects 
add.sem<-comp.risk(Event(time,cause)~
const(platelet)+const(age)+const(tcell),data=bmt,
cause=1,resample.iid=1,n.sim=100,model="additive")
summary(add.sem)

out<-predict(add.sem,ndata,uniform=1,n.sim=100)
par(mfrow=c(2,2))
plot(out,multiple=0,uniform=1,col=1:3,lty=1,se=0)

## Fine & Gray model 
fg<-comp.risk(Event(time,cause)~
const(platelet)+const(age)+const(tcell),data=bmt,
cause=1,resample.iid=1,model="fg",n.sim=100)
summary(fg)

out<-predict(fg,ndata,uniform=1,n.sim=100)

par(mfrow=c(2,2))
plot(out,multiple=1,uniform=0,col=1:3,lty=1,se=0)

## extended model with time-varying effects
fg.npar<-comp.risk(Event(time,cause)~platelet+age+const(tcell),
data=bmt,cause=1,resample.iid=1,model="prop",n.sim=100)
summary(fg.npar); 

out<-predict(fg.npar,ndata,uniform=1,n.sim=100)
head(out$P1[,1:5]); head(out$se.P1[,1:5])

par(mfrow=c(2,2))
plot(out,multiple=1,uniform=0,col=1:3,lty=1,se=0)

## Fine & Gray model with alternative parametrization for baseline
fg2<-comp.risk(Event(time,cause)~const(platelet)+const(age)+const(tcell),data=bmt,
cause=1,resample.iid=1,model="prop",n.sim=100)
summary(fg2)

#################################################################
## Delayed entry models, 
#################################################################
nn <- nrow(bmt)
entrytime <- rbinom(nn,1,0.5)*(bmt$time*runif(nn))
bmt$entrytime <- entrytime
times <- seq(5,70,by=1)

bmtw <- prep.comp.risk(bmt,times=times,time="time",entrytime="entrytime",cause="cause")

## non-parametric model 
outnp <- comp.risk(Event(time,cause)~tcell+platelet+const(age),
		   data=bmtw,cause=1,fix.gamma=1,gamma=0,
 cens.weights=bmtw$cw,weights=bmtw$weights,times=times,n.sim=0)
par(mfrow=c(2,2))
plot(outnp)

outnp <- comp.risk(Event(time,cause)~tcell+platelet,
		   data=bmtw,cause=1,
 cens.weights=bmtw$cw,weights=bmtw$weights,times=times,n.sim=0)
par(mfrow=c(2,2))
plot(outnp)


## semiparametric model 
out <- comp.risk(Event(time,cause)~const(tcell)+const(platelet),data=bmtw,cause=1,
 cens.weights=bmtw$cw,weights=bmtw$weights,times=times,n.sim=0)
summary(out)

Run the code above in your browser using DataLab