Predicting event probabilities from product limit estimates

Evaluation of estimated survival or event probabilities at given times and covariate constellations.

# S3 method for class 'prodlim'
predict(
  object,
  times,
  newdata,
  level.chaos = 1,
  type = c("surv", "risk", "cuminc", "list"),
  mode = "list",
  bytime = FALSE,
  cause,
  ...
)

Arguments

object

A fitted object of class "prodlim".

times

Vector of times at which to return the estimated probabilities (survival or absolute event risks).

newdata

A data frame with the same variable names as those that appear on the right hand side of the 'prodlim' formula. If there are covariates this argument is required.

level.chaos

Integer specifying the sorting of the output: `0' sort by time and newdata; `1' only by time; `2' no sorting at all

type

Choice between "surv","risk","cuminc","list":

"surv": predict survival probabilities only survival models

"risk"/"cuminc": predict absolute risk, i.e., cumulative incidence function.

"list": find the indices corresponding to times and newdata. See value.

Defaults to "surv" for two-state models and to "risk" for competing risk models.

mode

Only for type=="surv" and type=="risk". Can either be "list" or "matrix". For "matrix" the predicted probabilities will be returned in matrix form.

bytime

Logical. If TRUE and mode=="matrix" the matrix with predicted probabilities will have a column for each time and a row for each newdata. Only when object$covariate.type>1 and more than one time is given.

cause

Character (other classes are converted with as.character). The cause for predicting the absolute risk of an event, i.e., the cause-specific cumulative incidence function, in competing risk models. At any time after time zero this is the absolute risk of an event of type cause to occur between time zero and times .

...

Only for compatibility reasons.

Value

type=="surv" A list or a matrix with survival probabilities for all times and all newdata.

type=="risk" or type=="cuminc" A list or a matrix with cumulative incidences for all times and all newdata.

type=="list" A list with the following components:

times

The argument times carried forward

predictors

The relevant part of the argument newdata.

indices

A list with the following components

time: Where to find values corresponding to the requested times strata: Where to find values corresponding to the values of the variables in newdata. Together time and strata show where to find the predicted probabilities.

dimensions

a list with the following components: time : The length of times strata : The number of rows in newdata names.strata : Labels for the covariate values.

Details

Predicted (survival) probabilities are returned that can be plotted, summarized and used for inverse of probability of censoring weighting.

See also

predictSurvIndividual

Author

Thomas Alexander Gerds <tag@biostat.ku.dk>

Examples

dat <- SimSurv(400)
fit <- prodlim(Hist(time,status)~1,data=dat)

## predict the survival probs at selected times 
predict(fit,times=c(3,5,10))
#> [1] 0.7823540 0.6134853 0.2906672

## NA is returned when the time point is beyond the
## range of definition of the Kaplan-Meier estimator:
predict(fit,times=c(-1,0,10,100,1000,10000))
#> [1] 1.0000000 1.0000000 0.2906672        NA        NA        NA

## when there are strata, newdata is required 
## or neighborhoods (i.e. overlapping strata)
mfit <- prodlim(Hist(time,status)~X1+X2,data=dat)
predict(mfit,times=c(-1,0,10,100,1000,10000),newdata=dat[18:21,])
#> $`X1=0, X2= 3.09`
#> [1] 1.00000000 1.00000000 0.04270648         NA         NA         NA
#> 
#> $`X1=0, X2= 1.00`
#> [1] 1.0000000 1.0000000 0.1070347        NA        NA        NA
#> 
#> $`X1=1, X2=-0.47`
#> [1] 1.0000000 1.0000000 0.3101857        NA        NA        NA
#> 
#> $`X1=0, X2= 0.50`
#> [1] 1.0000000 1.0000000 0.2267839        NA        NA        NA
#> 

## this can be requested in matrix form
predict(mfit,times=c(-1,0,10,100,1000,10000),newdata=dat[18:21,],mode="matrix")
#>                [,1] [,2]       [,3] [,4] [,5] [,6]
#> X1=0, X2= 3.09    1    1 0.04270648   NA   NA   NA
#> X1=0, X2= 1.00    1    1 0.10703473   NA   NA   NA
#> X1=1, X2=-0.47    1    1 0.31018568   NA   NA   NA
#> X1=0, X2= 0.50    1    1 0.22678391   NA   NA   NA

## and even transposed
predict(mfit,times=c(-1,0,10,100,1000,10000),newdata=dat[18:21,],mode="matrix",bytime=TRUE)
#>               [,1]      [,2]      [,3]      [,4]
#> t=-1    1.00000000 1.0000000 1.0000000 1.0000000
#> t=0     1.00000000 1.0000000 1.0000000 1.0000000
#> t=10    0.04270648 0.1070347 0.3101857 0.2267839
#> t=100           NA        NA        NA        NA
#> t=1000          NA        NA        NA        NA
#> t=10000         NA        NA        NA        NA