IPA {riskRegression}

Explained variation for settings with binary, survival and competing risk outcome

Description

Index of Prediction Accuracy: General R^2 for binary outcome and right censored time to event (survival) outcome also with competing risks

Usage

rsquared(object,...)
IPA(object,...)
## Default S3 method:
rsquared(object,formula,newdata,times,cause,...)
## S3 method for class 'glm'
rsquared(object,formula,newdata,...)
## S3 method for class 'coxph'
rsquared(object,formula,newdata,times,...)
## S3 method for class 'CauseSpecificCox'
rsquared(object,formula,newdata,times,cause,...)
## Default S3 method:
IPA(object,formula,newdata,times,cause,...)
## S3 method for class 'glm'
IPA(object,formula,newdata,...)
## S3 method for class 'coxph'
IPA(object,formula,newdata,times,...)
## S3 method for class 'CauseSpecificCox'
IPA(object,formula,newdata,times,cause,...)

Arguments

object

Model for which we want IPA.

...

passed to riskRegression::Score

newdata

Optional validation data set in which to compute IPA

formula

Formula passed to Score. If not provided, try to use the formula of the call of object, if any.

cause

For competing risk models the event of interest

times

Vector of time points used as prediction horizon for the computation of Brier scores.

Details

IPA (R^2) is calculated based on the model's predicted risks. The Brier score of the model is compared to the Brier score of the null model.

Value

Data frame with explained variation values for the full model.

Author(s)

Thomas A. Gerds <[e-mail]>

See Also

Score

Examples

library(prodlim)
library(data.table)
# binary outcome
library(lava)
set.seed(18)
learndat <- sampleData(48,outcome="binary")
lr1 = glm(Y~X1+X2+X7+X9,data=learndat,family=binomial)
IPA(lr1)

## validation data
valdat=sampleData(94,outcome="binary")
IPA(lr1,newdata=valdat)

## predicted risks externally given
p1=predictRisk(lr1,newdata=valdat)
IPA(p1,formula=Y~1,valdat)

# survival
library(survival)
data(pbc)
pbc=na.omit(pbc)
pbctest=(1:NROW(pbc)) %in% sample(1:NROW(pbc),size=.632*NROW(pbc))
pbclearn=pbc[pbctest,]
cox1= coxph(Surv(time,status!=0)~age+sex+log(bili)+log(albumin)+log(protime),
      data=pbclearn,x=TRUE)

## same data
IPA(cox1,formula=Surv(time,status!=0)~1,times=1000)

## validation data
pbcval=pbc[!pbctest,]
IPA(cox1,formula=Surv(time,status!=0)~1,newdata=pbcval,times=1000)

## predicted risks externally given
p2=predictRisk(cox1,newdata=pbcval,times=1000)
IPA(cox1,formula=Surv(time,status!=0)~1,newdata=pbcval,times=1000)
 
# competing risks
data(Melanoma)
Melanomatest=(1:NROW(Melanoma)) %in% sample(1:NROW(Melanoma),size=.632*NROW(Melanoma))
Melanomalearn=Melanoma[Melanomatest,]
fit1 <- CSC(list(Hist(time,status)~sex,
                 Hist(time,status)~invasion+epicel+age),
                 data=Melanoma)
IPA(fit1,times=1000,cause=2)

## validation data
Melanomaval=Melanoma[!Melanomatest,]
IPA(fit1,formula=Hist(time,status)~1,newdata=Melanomaval,times=1000)

## predicted risks externally given
p3= predictRisk(fit1,cause=1,newdata=Melanomaval,times=1000)
IPA(p3,formula=Hist(time,status)~1,cause=1,newdata=Melanomaval,times=1000)