# PseudoBoost

In a competing risk setting the CIF for the interesting risk can be
calculated at a grid of time. For this values a stagewise boosting
algortihm is performed, whereas the variable selection is coupled between
time points but the estimates may differ. Consequently this allows to
estimate time-varying effects, which can be smoothed over time. Addionally, confidence bands can be calculated.

## Installation
To install the development version for the package PseudoBoost, please start a current version of R and type (using devtools):

```{r,eval=FALSE}
# currently this can be done via github
install.packages("devtools") #if needed
devtools::install_github("danielazoeller/PseudoBoost")
```


## Program call example

For example, the call is as follows:

Firstly, we simulate data for 400 individuals with 10 binomially distributed covariates with 2 competing risks. These covarites do not have an effect on the second, uninteresting competing risk and only 4 have an effect on the first, interesting risk. This effect changes at t=0.1. Before this time point, only X1 and X2 have an effect, afterwards only X3 and X4.

For the call the status, the observations time and the observed covariates are needed.
```{r, eval=FALSE}
library(PseudoBoost)
n <- 400; p <- 10
beta <- c(c(1,-1),rep(0,p-2))
beta2 <- c(c(0,0,1,-1),rep(0,p-4))

xmat <- matrix(1*(rbinom(n*p,1,0.5)),n,p)
real.time <- -(log(runif(n)))/(1*exp(drop(xmat %*% beta)))
real.time2 <- -(log(runif(n)))/(1*exp(drop(xmat %*% beta2)))
real.time <- ifelse(real.time > 0.1,0.1+real.time2,real.time)
cens.time <- rexp(n,rate=1/1)
second.time <- rexp(n,rate=1/1)
status <- ifelse(real.time <= cens.time,1,0)
obs.time <- ifelse(real.time <= cens.time,real.time,cens.time)

status <- ifelse(obs.time <= second.time,status,2)
obs.time <- ifelse(obs.time <= second.time,obs.time,second.time)

```

To evaluate this data, we use a grid of time with 50 time points based on the quantils of the observed events of both risks.
```{r, eval=FALSE}
eval.times <- quantile(obs.time[status != 0],seq(from=0.1,to=0.9,length=50))
```

To calculated the estimates, the call using cross-validation with a maximum number of boostings steps of 250 and a penalty term of 0.1 is as follows:

```{r, eval=FALSE}
#Estimates - Unsmooth
res_unsmooth <- smoothPseudoBoost(as.data.frame(cbind(obs.time,status)),xmat,times=eval.times, stepno = 100,
                                        maxstepno = 250, nu = 0.1, cv = TRUE, multicore = TRUE,
                                        RepSmooth = 0, smooth_para = 0,
                                        trace = TRUE)


#Estimates - Smooth
res_unsmooth <- smoothPseudoBoost(as.data.frame(cbind(obs.time,status)),xmat,times=eval.times, stepno = 100,
                                        maxstepno = 250, nu = 0.1, cv = TRUE, multicore = TRUE,
                                        RepSmooth = 20, smooth_para = 0.001,
                                        trace = TRUE)
```

Confidence bands can be calculated, too. Note that this can be computaninally very expensive.

```{r, eval=FALSE}                                      
#Confidence bands - Unsmooth
res_cb_unsmooth <- cbSmoothPseudoBoost(as.data.frame(cbind(obs.time,status)),xmat,times=eval.times, stepno = 100,
                                        maxstepno = 250, nu = 0.1, cv = TRUE, multicore = TRUE,
                                        RepSmooth = 0, smooth_para = 0,
                                        trace = TRUE, RepCb=1000,cv_est=FALSE)

#Confidence bands - Smooth
res_cb_smooth <- cbSmoothPseudoBoost(as.data.frame(cbind(obs.time,status)),xmat,times=eval.times, stepno = 100,
                                        maxstepno = 250, nu = 0.1, cv = TRUE, multicore = TRUE,
                                        RepSmooth = 20, smooth_para = 0.001,
                                        trace = TRUE, RepCb=1000,cv_est=FALSE)
```

## Contact
For additional details regarding the functions of PseudoBoost, please consult the documentation or write an e-mail to [email protected].