intPredErr {discSurv}

Integrated prediction error

Description

Computes the integrated prediction error curve for discrete survival models.

Usage

intPredErr(
  hazards,
  testTime,
  testEvent,
  trainTime,
  trainEvent,
  testDataLong,
  tmax = NULL
)

Arguments

hazards

Predicted discrete hazards in the test data ("numeric vector").

testTime

Discrete time intervals in short format of the test set ("integer vector").

testEvent

Events in short format in the test set ("binary vector").

trainTime

Discrete time intervals in short format of the training data set ("integer vector").

trainEvent

Events in short format in the training set ("binary vector").

testDataLong

Test data in long format("class data.frame"). The discrete survival function is calculated based on the predicted hazards. It is assumed that the data was preprocessed with a function with prefix "dataLong", see e. g. dataLong, dataLongTimeDep.

tmax

Gives the maximum time interval for which prediction errors are calculated ("integer vector"). It must be smaller than the maximum observed time in the training data of the object produced by function. The default setting NULL means, that all observed intervals are used.

Value

Integrated prediction error ("numeric vector").

Author(s)

Thomas Welchowski [e-mail]

References

Tutz G, Schmid M (2016). Modeling discrete time-to-event data. Springer Series in Statistics.

Gneiting T, Raftery AE (2007). “Strictly Proper Scoring Rules, Prediction, and Estimation.” Journal of the American Statistical Association, 102, 359-378.

See Also

predErrCurve, aggregate

Examples

##########################
# Example with cancer data

library(survival)
head(cancer)

# Data preparation and convertion to 30 intervals
cancerPrep <- cancer
cancerPrep$status <- cancerPrep$status-1
intLim <- quantile(cancerPrep$time, prob = seq(0, 1, length.out = 30))
intLim [length(intLim)] <- intLim [length(intLim)] + 1

# Cut discrete time in smaller number of intervals
cancerPrep <- contToDisc(dataShort = cancerPrep, timeColumn = "time", intervalLimits = intLim)

# Generate training and test data
set.seed(753)
TrainIndices <- sample (x = 1:dim(cancerPrep) [1], size = dim(cancerPrep) [1] * 0.75)
TrainCancer <- cancerPrep [TrainIndices, ]
TestCancer <- cancerPrep [-TrainIndices, ]
TrainCancer$timeDisc <- as.numeric(as.character(TrainCancer$timeDisc))
TestCancer$timeDisc <- as.numeric(as.character(TestCancer$timeDisc))

# Convert to long format
LongTrain <- dataLong(dataShort = TrainCancer, timeColumn = "timeDisc", eventColumn = "status",
timeAsFactor=FALSE)
LongTest <- dataLong(dataShort = TestCancer, timeColumn = "timeDisc", eventColumn = "status",
timeAsFactor=FALSE)
# Convert factors
LongTrain$timeInt <- as.numeric(as.character(LongTrain$timeInt))
LongTest$timeInt <- as.numeric(as.character(LongTest$timeInt))
LongTrain$sex <- factor(LongTrain$sex)
LongTest$sex <- factor(LongTest$sex)

# Estimate, for example, a generalized, additive model in discrete survival analysis
library(mgcv)
gamFit <- gam (formula = y ~ s(timeInt) + s(age) + sex + ph.ecog, data = LongTrain, 
family = binomial())
summary(gamFit)

# 1. Specification of predicted discrete hazards
# Estimate survival function of each person in the test data
testPredHaz <- predict(gamFit, newdata = LongTest, type = "response")

# 2. Calculate integrated prediction error
intPredErr(hazards = testPredHaz, 
testTime = TestCancer$timeDisc, testEvent = TestCancer$status, 
trainTime = TrainCancer$timeDisc, trainEvent = TrainCancer$status, 
testDataLong = LongTest)