crossvalidation: Cross-Validation

In CoxFlexBoost: Boosting Flexible Cox Models (with Time-Varying Effects)

Description

Function for the estimation of mstop via cross-validation and (generic) functions to print or plot the resuls.

Usage

## S3 method for class 'cfboost'
cv(object, folds,
   grid = c(1:mstop(object, opt=FALSE)), ...)
## S3 method for class 'cv'
print(x, ...)
## S3 method for class 'cv'
plot(x, ylab = attr(x, "risk"), ylim = range(x),
     main = attr(x, "call"), ...)

Arguments

object	an object of class cfboost.
folds	a weight matrix with number of rows equal to the number of observations. The number of columns corresponds to the number of cross-validation runs.
grid	a vector of iterations the empirical risk is to be evaluated for. Per default the empirical risks for all iterations 1:mstop are computed and returned.
x	an object of class cv
ylab	A title for the y axis.
ylim	the y limits of the plot.
main	the main title of the plot.
...	additional arguments to be passed to plot.

Details

The number of boosting iterations is a hyper-parameter of the boosting algorithms. Cross-validated estimates of the empirical risk for different values of mstop (as given by grid) are computed, which are used to choose the appropriate number of boosting iterations to be applied.

Different forms of cross-validation can be applied, for example, 5-fold or 10-fold cross-validation. Bootstrapping is not implemented so far. The weights are defined via the folds matrix. A.t.m. they can only be used to specify a learning sample which consists of observations with weights == 1 and and an out-of-bag sample with weights == 0. The latter is used to determine the empirical risk (negative log likelihood).

If package multicore is available, cv runs in parallel on cores/processors available. The scheduling can be changed by the corresponding arguments of mclapply (via the dot arguments). No trace output is given when running in parallel.

Value

cv returns an object of class cv, which consists of a matrix of empirical risks and some further attributes.

See Also

cfboost for model fitting. See risk for methods to extract the inbag and out-of-bag risk and mstop for functions to extract the (optimal) stopping iteration (based on cross-validation or on the inbag and out-of-bag risk).

Examples

## fit a model with all observations first

## Not run: 
## (as this takes some minutes)
set.seed(1234)
## sample covariates first
X <- matrix(NA, nrow=400, ncol=3)
X[,1] <- runif(400, -1, 1)
X[,2] <- runif(400, -1, 1)
X[,3] <- runif(400, -1, 1)

## time-dependent hazard rate
lambda <- function(time, x){
   exp(0 * time + 0.7 * x[1] + x[2]^2)
}

## specify censoring function
cens_fct <- function(time, mean_cens){
  censor_time <- rexp(n = length(time), rate = 1/mean_cens)
  event <- (time <= censor_time)
  t_obs <- apply(cbind(time, censor_time), 1, min)
  return(cbind(t_obs, event))
}
daten <- rSurvTime(lambda, X, cens_fct, mean_cens = 5)

ctrl <- boost_control( mstop = 100, risk="none")
## fit (a simple) model
model <- cfboost(Surv(time, event) ~ bbs(x.1) + bbs(x.2) + bbs(x.3),
                 control = ctrl, data = daten)

## End(Not run)

## 5 -fold cross-validation

## Not run: 
## (as this takes some minutes)
n <- nrow(daten)
k <- 5
ntest <- floor(n / k)
cv5f <- matrix(c(rep(c(rep(0, ntest), rep(1, n)), k - 1),
                       rep(0, n * k - (k - 1) * (n + ntest))), nrow = n)
cvm <- cv(model, folds = cv5f)
print(cvm)
plot(cvm)
mstop(cvm)

## End(Not run)

CoxFlexBoost documentation built on May 2, 2019, 6:53 p.m.